The same as reiser4-for-2.6.22.patch plus a fix for file conversion related bug wich caused metadata corruption when REISER4_DEBUG is on. diff -urN linux-2.6.22.orig/arch/i386/lib/usercopy.c linux-2.6.22/arch/i386/lib/usercopy.c --- linux-2.6.22.orig/arch/i386/lib/usercopy.c 2007-07-21 00:32:46.973831675 +0400 +++ linux-2.6.22/arch/i386/lib/usercopy.c 2007-07-29 00:25:34.800676805 +0400 @@ -817,6 +817,7 @@ #endif return n; } +EXPORT_SYMBOL(__copy_from_user_ll_nocache); unsigned long __copy_from_user_ll_nocache_nozero(void *to, const void __user *from, unsigned long n) @@ -831,6 +832,7 @@ #endif return n; } +EXPORT_SYMBOL(__copy_from_user_ll_nocache_nozero); /** * copy_to_user: - Copy a block of data into user space. diff -urN linux-2.6.22.orig/Documentation/Changes linux-2.6.22/Documentation/Changes --- linux-2.6.22.orig/Documentation/Changes 2007-07-21 00:31:57.012856483 +0400 +++ linux-2.6.22/Documentation/Changes 2007-07-29 00:25:34.800676805 +0400 @@ -36,6 +36,7 @@ o e2fsprogs 1.29 # tune2fs o jfsutils 1.1.3 # fsck.jfs -V o reiserfsprogs 3.6.3 # reiserfsck -V 2>&1|grep reiserfsprogs +o reiser4progs 1.0.0 # fsck.reiser4 -V o xfsprogs 2.6.0 # xfs_db -V o pcmciautils 004 # pccardctl -V o quota-tools 3.09 # quota -V @@ -144,6 +145,13 @@ versions of mkreiserfs, resize_reiserfs, debugreiserfs and reiserfsck. These utils work on both i386 and alpha platforms. +Reiser4progs +------------ + +The reiser4progs package contains utilities for the reiser4 file system. +Detailed instructions are provided in the README file located at: +. + Xfsprogs -------- @@ -322,6 +330,10 @@ ------------- o +Reiser4progs +------------ +o + Xfsprogs -------- o diff -urN linux-2.6.22.orig/Documentation/filesystems/reiser4.txt linux-2.6.22/Documentation/filesystems/reiser4.txt --- linux-2.6.22.orig/Documentation/filesystems/reiser4.txt 1970-01-01 03:00:00.000000000 +0300 +++ linux-2.6.22/Documentation/filesystems/reiser4.txt 2007-07-29 00:25:34.800676805 +0400 @@ -0,0 +1,75 @@ +Reiser4 filesystem +================== +Reiser4 is a file system based on dancing tree algorithms, and is +described at http://www.namesys.com + + +References +========== +web page http://namesys.com/v4/v4.html +source code ftp://ftp.namesys.com/pub/reiser4-for-2.6/ +userland tools ftp://ftp.namesys.com/pub/reiser4progs/ +install page http://www.namesys.com/install_v4.html + +Compile options +=============== +Enable reiser4 debug mode + This checks everything imaginable while reiser4 + runs + +Mount options +============= +tmgr.atom_max_size=N + Atoms containing more than N blocks will be forced to commit. + N is decimal. + Default is nr_free_pagecache_pages() / 2 at mount time. + +tmgr.atom_max_age=N + Atoms older than N seconds will be forced to commit. N is decimal. + Default is 600. + +tmgr.atom_max_flushers=N + Limit of concurrent flushers for one atom. 0 means no limit. + Default is 0. + +tree.cbk_cache.nr_slots=N + Number of slots in the cbk cache. + +flush.relocate_threshold=N + If flush finds more than N adjacent dirty leaf-level blocks it + will force them to be relocated. + Default is 64. + +flush.relocate_distance=N + If flush finds can find a block allocation closer than at most + N from the preceder it will relocate to that position. + Default is 64. + +flush.scan_maxnodes=N + The maximum number of nodes to scan left on a level during + flush. + Default is 10000. + +optimal_io_size=N + Preferred IO size. This value is used to set st_blksize of + struct stat. + Default is 65536. + +bsdgroups + Turn on BSD-style gid assignment. + +32bittimes + By default file in reiser4 have 64 bit timestamps. Files + created when filesystem is mounted with 32bittimes mount + option will get 32 bit timestamps. + +mtflush + Turn off concurrent flushing. + +nopseudo + Disable pseudo files support. See + http://namesys.com/v4/pseudo.html for more about pseudo files. + +dont_load_bitmap + Don't load all bitmap blocks at mount time, it is useful for + machines with tiny RAM and large disks. diff -urN linux-2.6.22.orig/fs/fs-writeback.c linux-2.6.22/fs/fs-writeback.c --- linux-2.6.22.orig/fs/fs-writeback.c 2007-07-21 00:32:04.502801671 +0400 +++ linux-2.6.22/fs/fs-writeback.c 2007-07-29 00:25:34.808678876 +0400 @@ -296,8 +296,6 @@ * WB_SYNC_HOLD is a hack for sys_sync(): reattach the inode to sb->s_dirty so * that it can be located for waiting on in __writeback_single_inode(). * - * Called under inode_lock. - * * If `bdi' is non-zero then we're being asked to writeback a specific queue. * This function assumes that the blockdev superblock's inodes are backed by * a variety of queues, so all inodes are searched. For other superblocks, @@ -313,11 +311,13 @@ * on the writer throttling path, and we get decent balancing between many * throttled threads: we don't want them all piling up on __wait_on_inode. */ -static void -sync_sb_inodes(struct super_block *sb, struct writeback_control *wbc) +void +generic_sync_sb_inodes(struct super_block *sb, struct writeback_control *wbc) { const unsigned long start = jiffies; /* livelock avoidance */ + spin_lock(&inode_lock); + if (!wbc->for_kupdate || list_empty(&sb->s_io)) list_splice_init(&sb->s_dirty, &sb->s_io); @@ -397,8 +397,19 @@ if (wbc->nr_to_write <= 0) break; } + spin_unlock(&inode_lock); return; /* Leave any unwritten inodes on s_io */ } +EXPORT_SYMBOL(generic_sync_sb_inodes); + +static void +sync_sb_inodes(struct super_block *sb, struct writeback_control *wbc) +{ + if (sb->s_op->sync_inodes) + sb->s_op->sync_inodes(sb, wbc); + else + generic_sync_sb_inodes(sb, wbc); +} /* * Start writeback of dirty pagecache data against all unlocked inodes. @@ -439,11 +450,8 @@ * be unmounted by the time it is released. */ if (down_read_trylock(&sb->s_umount)) { - if (sb->s_root) { - spin_lock(&inode_lock); + if (sb->s_root) sync_sb_inodes(sb, wbc); - spin_unlock(&inode_lock); - } up_read(&sb->s_umount); } spin_lock(&sb_lock); @@ -481,9 +489,7 @@ (inodes_stat.nr_inodes - inodes_stat.nr_unused) + nr_dirty + nr_unstable; wbc.nr_to_write += wbc.nr_to_write / 2; /* Bit more for luck */ - spin_lock(&inode_lock); sync_sb_inodes(sb, &wbc); - spin_unlock(&inode_lock); } /* diff -urN linux-2.6.22.orig/fs/Kconfig linux-2.6.22/fs/Kconfig --- linux-2.6.22.orig/fs/Kconfig 2007-07-21 00:32:57.540575927 +0400 +++ linux-2.6.22/fs/Kconfig 2007-07-29 00:25:34.812679911 +0400 @@ -272,6 +272,8 @@ default y if EXT2_FS=y || EXT3_FS=y || EXT4DEV_FS=y default m if EXT2_FS=m || EXT3_FS=m || EXT4DEV_FS=m +source "fs/reiser4/Kconfig" + config REISERFS_FS tristate "Reiserfs support" help diff -urN linux-2.6.22.orig/fs/Makefile linux-2.6.22/fs/Makefile --- linux-2.6.22.orig/fs/Makefile 2007-07-21 00:32:57.544576967 +0400 +++ linux-2.6.22/fs/Makefile 2007-07-29 00:25:34.812679911 +0400 @@ -66,6 +66,7 @@ # Do not add any filesystems before this line obj-$(CONFIG_REISERFS_FS) += reiserfs/ +obj-$(CONFIG_REISER4_FS) += reiser4/ obj-$(CONFIG_EXT3_FS) += ext3/ # Before ext2 so root fs can be ext3 obj-$(CONFIG_EXT4DEV_FS) += ext4/ # Before ext2 so root fs can be ext4dev obj-$(CONFIG_JBD) += jbd/ diff -urN linux-2.6.22.orig/fs/reiser4/as_ops.c linux-2.6.22/fs/reiser4/as_ops.c --- linux-2.6.22.orig/fs/reiser4/as_ops.c 1970-01-01 03:00:00.000000000 +0300 +++ linux-2.6.22/fs/reiser4/as_ops.c 2007-07-29 00:25:34.816680947 +0400 @@ -0,0 +1,337 @@ +/* Copyright 2003 by Hans Reiser, licensing governed by reiser4/README */ + +/* Interface to VFS. Reiser4 address_space_operations are defined here. */ + +#include "forward.h" +#include "debug.h" +#include "dformat.h" +#include "coord.h" +#include "plugin/item/item.h" +#include "plugin/file/file.h" +#include "plugin/security/perm.h" +#include "plugin/disk_format/disk_format.h" +#include "plugin/plugin.h" +#include "plugin/plugin_set.h" +#include "plugin/object.h" +#include "txnmgr.h" +#include "jnode.h" +#include "znode.h" +#include "block_alloc.h" +#include "tree.h" +#include "vfs_ops.h" +#include "inode.h" +#include "page_cache.h" +#include "ktxnmgrd.h" +#include "super.h" +#include "reiser4.h" +#include "entd.h" + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +/* address space operations */ + +/** + * reiser4_set_page_dirty - set dirty bit, tag in page tree, dirty accounting + * @page: page to be dirtied + * + * Operation of struct address_space_operations. This implementation is used by + * unix and cryptcompress file plugins. + * + * This is called when reiser4 page gets dirtied outside of reiser4, for + * example, when dirty bit is moved from pte to physical page. + * + * Tags page in the mapping's page tree with special tag so that it is possible + * to do all the reiser4 specific work wrt dirty pages (jnode creation, + * capturing by an atom) later because it can not be done in the contexts where + * set_page_dirty is called. + */ +int reiser4_set_page_dirty(struct page *page) +{ + /* this page can be unformatted only */ + assert("vs-1734", (page->mapping && + page->mapping->host && + reiser4_get_super_fake(page->mapping->host->i_sb) != + page->mapping->host + && reiser4_get_cc_fake(page->mapping->host->i_sb) != + page->mapping->host + && reiser4_get_bitmap_fake(page->mapping->host->i_sb) != + page->mapping->host)); + + if (!TestSetPageDirty(page)) { + struct address_space *mapping = page->mapping; + + if (mapping) { + write_lock_irq(&mapping->tree_lock); + + /* check for race with truncate */ + if (page->mapping) { + assert("vs-1652", page->mapping == mapping); + if (mapping_cap_account_dirty(mapping)) + inc_zone_page_state(page, + NR_FILE_DIRTY); + radix_tree_tag_set(&mapping->page_tree, + page->index, + PAGECACHE_TAG_REISER4_MOVED); + } + write_unlock_irq(&mapping->tree_lock); + __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); + } + } + return 0; +} + +/* ->invalidatepage method for reiser4 */ + +/* + * this is called for each truncated page from + * truncate_inode_pages()->truncate_{complete,partial}_page(). + * + * At the moment of call, page is under lock, and outstanding io (if any) has + * completed. + */ + +/** + * reiser4_invalidatepage + * @page: page to invalidate + * @offset: starting offset for partial invalidation + * + */ +void reiser4_invalidatepage(struct page *page, unsigned long offset) +{ + int ret = 0; + reiser4_context *ctx; + struct inode *inode; + jnode *node; + + /* + * This is called to truncate file's page. + * + * Originally, reiser4 implemented truncate in a standard way + * (vmtruncate() calls ->invalidatepage() on all truncated pages + * first, then file system ->truncate() call-back is invoked). + * + * This lead to the problem when ->invalidatepage() was called on a + * page with jnode that was captured into atom in ASTAGE_PRE_COMMIT + * process. That is, truncate was bypassing transactions. To avoid + * this, try_capture_page_to_invalidate() call was added here. + * + * After many troubles with vmtruncate() based truncate (including + * races with flush, tail conversion, etc.) it was re-written in the + * top-to-bottom style: items are killed in reiser4_cut_tree_object() + * and pages belonging to extent are invalidated in kill_hook_extent(). + * So probably now additional call to capture is not needed here. + */ + + assert("nikita-3137", PageLocked(page)); + assert("nikita-3138", !PageWriteback(page)); + inode = page->mapping->host; + + /* + * ->invalidatepage() should only be called for the unformatted + * jnodes. Destruction of all other types of jnodes is performed + * separately. But, during some corner cases (like handling errors + * during mount) it is simpler to let ->invalidatepage to be called on + * them. Check for this, and do nothing. + */ + if (reiser4_get_super_fake(inode->i_sb) == inode) + return; + if (reiser4_get_cc_fake(inode->i_sb) == inode) + return; + if (reiser4_get_bitmap_fake(inode->i_sb) == inode) + return; + assert("vs-1426", PagePrivate(page)); + assert("vs-1427", + page->mapping == jnode_get_mapping(jnode_by_page(page))); + assert("", jprivate(page) != NULL); + assert("", ergo(inode_file_plugin(inode) != + file_plugin_by_id(CRYPTCOMPRESS_FILE_PLUGIN_ID), + offset == 0)); + + ctx = reiser4_init_context(inode->i_sb); + if (IS_ERR(ctx)) + return; + + node = jprivate(page); + spin_lock_jnode(node); + if (!(node->state & ((1 << JNODE_DIRTY) | (1<< JNODE_FLUSH_QUEUED) | + (1 << JNODE_WRITEBACK) | (1 << JNODE_OVRWR)))) { + /* there is not need to capture */ + jref(node); + JF_SET(node, JNODE_HEARD_BANSHEE); + page_clear_jnode(page, node); + reiser4_uncapture_jnode(node); + unhash_unformatted_jnode(node); + jput(node); + reiser4_exit_context(ctx); + return; + } + spin_unlock_jnode(node); + + /* capture page being truncated. */ + ret = try_capture_page_to_invalidate(page); + if (ret != 0) + warning("nikita-3141", "Cannot capture: %i", ret); + + if (offset == 0) { + /* remove jnode from transaction and detach it from page. */ + jref(node); + JF_SET(node, JNODE_HEARD_BANSHEE); + /* page cannot be detached from jnode concurrently, because it + * is locked */ + reiser4_uncapture_page(page); + + /* this detaches page from jnode, so that jdelete will not try + * to lock page which is already locked */ + spin_lock_jnode(node); + page_clear_jnode(page, node); + spin_unlock_jnode(node); + unhash_unformatted_jnode(node); + + jput(node); + } + + reiser4_exit_context(ctx); +} + +/* help function called from reiser4_releasepage(). It returns true if jnode + * can be detached from its page and page released. */ +int jnode_is_releasable(jnode * node /* node to check */ ) +{ + assert("nikita-2781", node != NULL); + assert_spin_locked(&(node->guard)); + assert_spin_locked(&(node->load)); + + /* is some thread is currently using jnode page, later cannot be + * detached */ + if (atomic_read(&node->d_count) != 0) { + return 0; + } + + assert("vs-1214", !jnode_is_loaded(node)); + + /* + * can only release page if real block number is assigned to it. Simple + * check for ->atom wouldn't do, because it is possible for node to be + * clean, not it atom yet, and still having fake block number. For + * example, node just created in jinit_new(). + */ + if (reiser4_blocknr_is_fake(jnode_get_block(node))) + return 0; + + /* + * pages prepared for write can not be released anyway, so avoid + * detaching jnode from the page + */ + if (JF_ISSET(node, JNODE_WRITE_PREPARED)) + return 0; + + /* + * dirty jnode cannot be released. It can however be submitted to disk + * as part of early flushing, but only after getting flush-prepped. + */ + if (JF_ISSET(node, JNODE_DIRTY)) + return 0; + + /* overwrite set is only written by log writer. */ + if (JF_ISSET(node, JNODE_OVRWR)) + return 0; + + /* jnode is already under writeback */ + if (JF_ISSET(node, JNODE_WRITEBACK)) + return 0; + + /* don't flush bitmaps or journal records */ + if (!jnode_is_znode(node) && !jnode_is_unformatted(node)) + return 0; + + return 1; +} + +/* + * ->releasepage method for reiser4 + * + * This is called by VM scanner when it comes across clean page. What we have + * to do here is to check whether page can really be released (freed that is) + * and if so, detach jnode from it and remove page from the page cache. + * + * Check for releasability is done by releasable() function. + */ +int reiser4_releasepage(struct page *page, gfp_t gfp UNUSED_ARG) +{ + jnode *node; + + assert("nikita-2257", PagePrivate(page)); + assert("nikita-2259", PageLocked(page)); + assert("nikita-2892", !PageWriteback(page)); + assert("nikita-3019", reiser4_schedulable()); + + /* NOTE-NIKITA: this can be called in the context of reiser4 call. It + is not clear what to do in this case. A lot of deadlocks seems be + possible. */ + + node = jnode_by_page(page); + assert("nikita-2258", node != NULL); + assert("reiser4-4", page->mapping != NULL); + assert("reiser4-5", page->mapping->host != NULL); + + if (PageDirty(page)) + return 0; + + /* extra page reference is used by reiser4 to protect + * jnode<->page link from this ->releasepage(). */ + if (page_count(page) > 3) + return 0; + + /* releasable() needs jnode lock, because it looks at the jnode fields + * and we need jload_lock here to avoid races with jload(). */ + spin_lock_jnode(node); + spin_lock(&(node->load)); + if (jnode_is_releasable(node)) { + struct address_space *mapping; + + mapping = page->mapping; + jref(node); + /* there is no need to synchronize against + * jnode_extent_write() here, because pages seen by + * jnode_extent_write() are !releasable(). */ + page_clear_jnode(page, node); + spin_unlock(&(node->load)); + spin_unlock_jnode(node); + + /* we are under memory pressure so release jnode also. */ + jput(node); + + return 1; + } else { + spin_unlock(&(node->load)); + spin_unlock_jnode(node); + assert("nikita-3020", reiser4_schedulable()); + return 0; + } +} + +/* Make Linus happy. + Local variables: + c-indentation-style: "K&R" + mode-name: "LC" + c-basic-offset: 8 + tab-width: 8 + fill-column: 120 + End: +*/ diff -urN linux-2.6.22.orig/fs/reiser4/block_alloc.c linux-2.6.22/fs/reiser4/block_alloc.c --- linux-2.6.22.orig/fs/reiser4/block_alloc.c 1970-01-01 03:00:00.000000000 +0300 +++ linux-2.6.22/fs/reiser4/block_alloc.c 2007-07-29 00:25:34.816680947 +0400 @@ -0,0 +1,1137 @@ +/* Copyright 2001, 2002, 2003 by Hans Reiser, licensing governed by reiser4/README */ + +#include "debug.h" +#include "dformat.h" +#include "plugin/plugin.h" +#include "txnmgr.h" +#include "znode.h" +#include "block_alloc.h" +#include "tree.h" +#include "super.h" + +#include /* for __u?? */ +#include /* for struct super_block */ +#include + +/* THE REISER4 DISK SPACE RESERVATION SCHEME. */ + +/* We need to be able to reserve enough disk space to ensure that an atomic + operation will have enough disk space to flush (see flush.c and + http://namesys.com/v4/v4.html) and commit it once it is started. + + In our design a call for reserving disk space may fail but not an actual + block allocation. + + All free blocks, already allocated blocks, and all kinds of reserved blocks + are counted in different per-fs block counters. + + A reiser4 super block's set of block counters currently is: + + free -- free blocks, + used -- already allocated blocks, + + grabbed -- initially reserved for performing an fs operation, those blocks + are taken from free blocks, then grabbed disk space leaks from grabbed + blocks counter to other counters like "fake allocated", "flush + reserved", "used", the rest of not used grabbed space is returned to + free space at the end of fs operation; + + fake allocated -- counts all nodes without real disk block numbers assigned, + we have separate accounting for formatted and unformatted + nodes (for easier debugging); + + flush reserved -- disk space needed for flushing and committing an atom. + Each dirty already allocated block could be written as a + part of atom's overwrite set or as a part of atom's + relocate set. In both case one additional block is needed, + it is used as a wandered block if we do overwrite or as a + new location for a relocated block. + + In addition, blocks in some states are counted on per-thread and per-atom + basis. A reiser4 context has a counter of blocks grabbed by this transaction + and the sb's grabbed blocks counter is a sum of grabbed blocks counter values + of each reiser4 context. Each reiser4 atom has a counter of "flush reserved" + blocks, which are reserved for flush processing and atom commit. */ + +/* AN EXAMPLE: suppose we insert new item to the reiser4 tree. We estimate + number of blocks to grab for most expensive case of balancing when the leaf + node we insert new item to gets split and new leaf node is allocated. + + So, we need to grab blocks for + + 1) one block for possible dirtying the node we insert an item to. That block + would be used for node relocation at flush time or for allocating of a + wandered one, it depends what will be a result (what set, relocate or + overwrite the node gets assigned to) of the node processing by the flush + algorithm. + + 2) one block for either allocating a new node, or dirtying of right or left + clean neighbor, only one case may happen. + + VS-FIXME-HANS: why can only one case happen? I would expect to see dirtying of left neighbor, right neighbor, current + node, and creation of new node. have I forgotten something? email me. + + These grabbed blocks are counted in both reiser4 context "grabbed blocks" + counter and in the fs-wide one (both ctx->grabbed_blocks and + sbinfo->blocks_grabbed get incremented by 2), sb's free blocks counter is + decremented by 2. + + Suppose both two blocks were spent for dirtying of an already allocated clean + node (one block went from "grabbed" to "flush reserved") and for new block + allocating (one block went from "grabbed" to "fake allocated formatted"). + + Inserting of a child pointer to the parent node caused parent node to be + split, the balancing code takes care about this grabbing necessary space + immediately by calling reiser4_grab with BA_RESERVED flag set which means + "can use the 5% reserved disk space". + + At this moment insertion completes and grabbed blocks (if they were not used) + should be returned to the free space counter. + + However the atom life-cycle is not completed. The atom had one "flush + reserved" block added by our insertion and the new fake allocated node is + counted as a "fake allocated formatted" one. The atom has to be fully + processed by flush before commit. Suppose that the flush moved the first, + already allocated node to the atom's overwrite list, the new fake allocated + node, obviously, went into the atom relocate set. The reiser4 flush + allocates the new node using one unit from "fake allocated formatted" + counter, the log writer uses one from "flush reserved" for wandered block + allocation. + + And, it is not the end. When the wandered block is deallocated after the + atom gets fully played (see wander.c for term description), the disk space + occupied for it is returned to free blocks. */ + +/* BLOCK NUMBERS */ + +/* Any reiser4 node has a block number assigned to it. We use these numbers for + indexing in hash tables, so if a block has not yet been assigned a location + on disk we need to give it a temporary fake block number. + + Current implementation of reiser4 uses 64-bit integers for block numbers. We + use highest bit in 64-bit block number to distinguish fake and real block + numbers. So, only 63 bits may be used to addressing of real device + blocks. That "fake" block numbers space is divided into subspaces of fake + block numbers for data blocks and for shadow (working) bitmap blocks. + + Fake block numbers for data blocks are generated by a cyclic counter, which + gets incremented after each real block allocation. We assume that it is + impossible to overload this counter during one transaction life. */ + +/* Initialize a blocknr hint. */ +void reiser4_blocknr_hint_init(reiser4_blocknr_hint * hint) +{ + memset(hint, 0, sizeof(reiser4_blocknr_hint)); +} + +/* Release any resources of a blocknr hint. */ +void reiser4_blocknr_hint_done(reiser4_blocknr_hint * hint UNUSED_ARG) +{ + /* No resources should be freed in current blocknr_hint implementation. */ +} + +/* see above for explanation of fake block number. */ +/* Audited by: green(2002.06.11) */ +int reiser4_blocknr_is_fake(const reiser4_block_nr * da) +{ + /* The reason for not simply returning result of '&' operation is that + while return value is (possibly 32bit) int, the reiser4_block_nr is + at least 64 bits long, and high bit (which is the only possible + non zero bit after the masking) would be stripped off */ + return (*da & REISER4_FAKE_BLOCKNR_BIT_MASK) ? 1 : 0; +} + +/* Static functions for / block counters + arithmetic. Mostly, they are isolated to not to code same assertions in + several places. */ +static void sub_from_ctx_grabbed(reiser4_context * ctx, __u64 count) +{ + BUG_ON(ctx->grabbed_blocks < count); + assert("zam-527", ctx->grabbed_blocks >= count); + ctx->grabbed_blocks -= count; +} + +static void add_to_ctx_grabbed(reiser4_context * ctx, __u64 count) +{ + ctx->grabbed_blocks += count; +} + +static void sub_from_sb_grabbed(reiser4_super_info_data * sbinfo, __u64 count) +{ + assert("zam-525", sbinfo->blocks_grabbed >= count); + sbinfo->blocks_grabbed -= count; +} + +/* Decrease the counter of block reserved for flush in super block. */ +static void +sub_from_sb_flush_reserved(reiser4_super_info_data * sbinfo, __u64 count) +{ + assert("vpf-291", sbinfo->blocks_flush_reserved >= count); + sbinfo->blocks_flush_reserved -= count; +} + +static void +sub_from_sb_fake_allocated(reiser4_super_info_data * sbinfo, __u64 count, + reiser4_ba_flags_t flags) +{ + if (flags & BA_FORMATTED) { + assert("zam-806", sbinfo->blocks_fake_allocated >= count); + sbinfo->blocks_fake_allocated -= count; + } else { + assert("zam-528", + sbinfo->blocks_fake_allocated_unformatted >= count); + sbinfo->blocks_fake_allocated_unformatted -= count; + } +} + +static void sub_from_sb_used(reiser4_super_info_data * sbinfo, __u64 count) +{ + assert("zam-530", + sbinfo->blocks_used >= count + sbinfo->min_blocks_used); + sbinfo->blocks_used -= count; +} + +static void +sub_from_cluster_reserved(reiser4_super_info_data * sbinfo, __u64 count) +{ + assert("edward-501", sbinfo->blocks_clustered >= count); + sbinfo->blocks_clustered -= count; +} + +/* Increase the counter of block reserved for flush in atom. */ +static void add_to_atom_flush_reserved_nolock(txn_atom * atom, __u32 count) +{ + assert("zam-772", atom != NULL); + assert_spin_locked(&(atom->alock)); + atom->flush_reserved += count; +} + +/* Decrease the counter of block reserved for flush in atom. */ +static void sub_from_atom_flush_reserved_nolock(txn_atom * atom, __u32 count) +{ + assert("zam-774", atom != NULL); + assert_spin_locked(&(atom->alock)); + assert("nikita-2790", atom->flush_reserved >= count); + atom->flush_reserved -= count; +} + +/* super block has 6 counters: free, used, grabbed, fake allocated + (formatted and unformatted) and flush reserved. Their sum must be + number of blocks on a device. This function checks this */ +int reiser4_check_block_counters(const struct super_block *super) +{ + __u64 sum; + + sum = reiser4_grabbed_blocks(super) + reiser4_free_blocks(super) + + reiser4_data_blocks(super) + reiser4_fake_allocated(super) + + reiser4_fake_allocated_unformatted(super) + reiser4_flush_reserved(super) + + reiser4_clustered_blocks(super); + if (reiser4_block_count(super) != sum) { + printk("super block counters: " + "used %llu, free %llu, " + "grabbed %llu, fake allocated (formatetd %llu, unformatted %llu), " + "reserved %llu, clustered %llu, sum %llu, must be (block count) %llu\n", + (unsigned long long)reiser4_data_blocks(super), + (unsigned long long)reiser4_free_blocks(super), + (unsigned long long)reiser4_grabbed_blocks(super), + (unsigned long long)reiser4_fake_allocated(super), + (unsigned long long) + reiser4_fake_allocated_unformatted(super), + (unsigned long long)reiser4_flush_reserved(super), + (unsigned long long)reiser4_clustered_blocks(super), + (unsigned long long)sum, + (unsigned long long)reiser4_block_count(super)); + return 0; + } + return 1; +} + +/* Adjust "working" free blocks counter for number of blocks we are going to + allocate. Record number of grabbed blocks in fs-wide and per-thread + counters. This function should be called before bitmap scanning or + allocating fake block numbers + + @super -- pointer to reiser4 super block; + @count -- number of blocks we reserve; + + @return -- 0 if success, -ENOSPC, if all + free blocks are preserved or already allocated. +*/ + +static int +reiser4_grab(reiser4_context * ctx, __u64 count, reiser4_ba_flags_t flags) +{ + __u64 free_blocks; + int ret = 0, use_reserved = flags & BA_RESERVED; + reiser4_super_info_data *sbinfo; + + assert("vs-1276", ctx == get_current_context()); + + /* Do not grab anything on ro-mounted fs. */ + if (rofs_super(ctx->super)) { + ctx->grab_enabled = 0; + return 0; + } + + sbinfo = get_super_private(ctx->super); + + spin_lock_reiser4_super(sbinfo); + + free_blocks = sbinfo->blocks_free; + + if ((use_reserved && free_blocks < count) || + (!use_reserved && free_blocks < count + sbinfo->blocks_reserved)) { + ret = RETERR(-ENOSPC); + goto unlock_and_ret; + } + + add_to_ctx_grabbed(ctx, count); + + sbinfo->blocks_grabbed += count; + sbinfo->blocks_free -= count; + +#if REISER4_DEBUG + if (ctx->grabbed_initially == 0) + ctx->grabbed_initially = count; +#endif + + assert("nikita-2986", reiser4_check_block_counters(ctx->super)); + + /* disable grab space in current context */ + ctx->grab_enabled = 0; + + unlock_and_ret: + spin_unlock_reiser4_super(sbinfo); + + return ret; +} + +int reiser4_grab_space(__u64 count, reiser4_ba_flags_t flags) +{ + int ret; + reiser4_context *ctx; + + assert("nikita-2964", ergo(flags & BA_CAN_COMMIT, + lock_stack_isclean(get_current_lock_stack + ()))); + ctx = get_current_context(); + if (!(flags & BA_FORCE) && !is_grab_enabled(ctx)) { + return 0; + } + + ret = reiser4_grab(ctx, count, flags); + if (ret == -ENOSPC) { + + /* Trying to commit the all transactions if BA_CAN_COMMIT flag present */ + if (flags & BA_CAN_COMMIT) { + txnmgr_force_commit_all(ctx->super, 0); + ctx->grab_enabled = 1; + ret = reiser4_grab(ctx, count, flags); + } + } + /* + * allocation from reserved pool cannot fail. This is severe error. + */ + assert("nikita-3005", ergo(flags & BA_RESERVED, ret == 0)); + return ret; +} + +/* + * SPACE RESERVED FOR UNLINK/TRUNCATE + * + * Unlink and truncate require space in transaction (to update stat data, at + * least). But we don't want rm(1) to fail with "No space on device" error. + * + * Solution is to reserve 5% of disk space for truncates and + * unlinks. Specifically, normal space grabbing requests don't grab space from + * reserved area. Only requests with BA_RESERVED bit in flags are allowed to + * drain it. Per super block delete mutex is used to allow only one + * thread at a time to grab from reserved area. + * + * Grabbing from reserved area should always be performed with BA_CAN_COMMIT + * flag. + * + */ + +int reiser4_grab_reserved(struct super_block *super, + __u64 count, reiser4_ba_flags_t flags) +{ + reiser4_super_info_data *sbinfo = get_super_private(super); + + assert("nikita-3175", flags & BA_CAN_COMMIT); + + /* Check the delete mutex already taken by us, we assume that + * reading of machine word is atomic. */ + if (sbinfo->delete_mutex_owner == current) { + if (reiser4_grab_space + (count, (flags | BA_RESERVED) & ~BA_CAN_COMMIT)) { + warning("zam-1003", + "nested call of grab_reserved fails count=(%llu)", + (unsigned long long)count); + reiser4_release_reserved(super); + return RETERR(-ENOSPC); + } + return 0; + } + + if (reiser4_grab_space(count, flags)) { + mutex_lock(&sbinfo->delete_mutex); + assert("nikita-2929", sbinfo->delete_mutex_owner == NULL); + sbinfo->delete_mutex_owner = current; + + if (reiser4_grab_space(count, flags | BA_RESERVED)) { + warning("zam-833", + "reserved space is not enough (%llu)", + (unsigned long long)count); + reiser4_release_reserved(super); + return RETERR(-ENOSPC); + } + } + return 0; +} + +void reiser4_release_reserved(struct super_block *super) +{ + reiser4_super_info_data *info; + + info = get_super_private(super); + if (info->delete_mutex_owner == current) { + info->delete_mutex_owner = NULL; + mutex_unlock(&info->delete_mutex); + } +} + +static reiser4_super_info_data *grabbed2fake_allocated_head(int count) +{ + reiser4_context *ctx; + reiser4_super_info_data *sbinfo; + + ctx = get_current_context(); + sub_from_ctx_grabbed(ctx, count); + + sbinfo = get_super_private(ctx->super); + spin_lock_reiser4_super(sbinfo); + + sub_from_sb_grabbed(sbinfo, count); + /* return sbinfo locked */ + return sbinfo; +} + +/* is called after @count fake block numbers are allocated and pointer to + those blocks are inserted into tree. */ +static void grabbed2fake_allocated_formatted(void) +{ + reiser4_super_info_data *sbinfo; + + sbinfo = grabbed2fake_allocated_head(1); + sbinfo->blocks_fake_allocated++; + + assert("vs-922", reiser4_check_block_counters(reiser4_get_current_sb())); + + spin_unlock_reiser4_super(sbinfo); +} + +/** + * grabbed2fake_allocated_unformatted + * @count: + * + */ +static void grabbed2fake_allocated_unformatted(int count) +{ + reiser4_super_info_data *sbinfo; + + sbinfo = grabbed2fake_allocated_head(count); + sbinfo->blocks_fake_allocated_unformatted += count; + + assert("vs-9221", reiser4_check_block_counters(reiser4_get_current_sb())); + + spin_unlock_reiser4_super(sbinfo); +} + +void grabbed2cluster_reserved(int count) +{ + reiser4_context *ctx; + reiser4_super_info_data *sbinfo; + + ctx = get_current_context(); + sub_from_ctx_grabbed(ctx, count); + + sbinfo = get_super_private(ctx->super); + spin_lock_reiser4_super(sbinfo); + + sub_from_sb_grabbed(sbinfo, count); + sbinfo->blocks_clustered += count; + + assert("edward-504", reiser4_check_block_counters(ctx->super)); + + spin_unlock_reiser4_super(sbinfo); +} + +void cluster_reserved2grabbed(int count) +{ + reiser4_context *ctx; + reiser4_super_info_data *sbinfo; + + ctx = get_current_context(); + + sbinfo = get_super_private(ctx->super); + spin_lock_reiser4_super(sbinfo); + + sub_from_cluster_reserved(sbinfo, count); + sbinfo->blocks_grabbed += count; + + assert("edward-505", reiser4_check_block_counters(ctx->super)); + + spin_unlock_reiser4_super(sbinfo); + add_to_ctx_grabbed(ctx, count); +} + +void cluster_reserved2free(int count) +{ + reiser4_context *ctx; + reiser4_super_info_data *sbinfo; + + ctx = get_current_context(); + sbinfo = get_super_private(ctx->super); + + cluster_reserved2grabbed(count); + grabbed2free(ctx, sbinfo, count); +} + +static DEFINE_SPINLOCK(fake_lock); +static reiser4_block_nr fake_gen = 0; + +/** + * assign_fake_blocknr + * @blocknr: + * @count: + * + * Obtain a fake block number for new node which will be used to refer to + * this newly allocated node until real allocation is done. + */ +static void assign_fake_blocknr(reiser4_block_nr *blocknr, int count) +{ + spin_lock(&fake_lock); + *blocknr = fake_gen; + fake_gen += count; + spin_unlock(&fake_lock); + + BUG_ON(*blocknr & REISER4_BLOCKNR_STATUS_BIT_MASK); + /**blocknr &= ~REISER4_BLOCKNR_STATUS_BIT_MASK;*/ + *blocknr |= REISER4_UNALLOCATED_STATUS_VALUE; + assert("zam-394", zlook(current_tree, blocknr) == NULL); +} + +int assign_fake_blocknr_formatted(reiser4_block_nr * blocknr) +{ + assign_fake_blocknr(blocknr, 1); + grabbed2fake_allocated_formatted(); + return 0; +} + +/** + * fake_blocknrs_unformatted + * @count: number of fake numbers to get + * + * Allocates @count fake block numbers which will be assigned to jnodes + */ +reiser4_block_nr fake_blocknr_unformatted(int count) +{ + reiser4_block_nr blocknr; + + assign_fake_blocknr(&blocknr, count); + grabbed2fake_allocated_unformatted(count); + + return blocknr; +} + +/* adjust sb block counters, if real (on-disk) block allocation immediately + follows grabbing of free disk space. */ +static void grabbed2used(reiser4_context *ctx, reiser4_super_info_data *sbinfo, + __u64 count) +{ + sub_from_ctx_grabbed(ctx, count); + + spin_lock_reiser4_super(sbinfo); + + sub_from_sb_grabbed(sbinfo, count); + sbinfo->blocks_used += count; + + assert("nikita-2679", reiser4_check_block_counters(ctx->super)); + + spin_unlock_reiser4_super(sbinfo); +} + +/* adjust sb block counters when @count unallocated blocks get mapped to disk */ +static void fake_allocated2used(reiser4_super_info_data *sbinfo, __u64 count, + reiser4_ba_flags_t flags) +{ + spin_lock_reiser4_super(sbinfo); + + sub_from_sb_fake_allocated(sbinfo, count, flags); + sbinfo->blocks_used += count; + + assert("nikita-2680", + reiser4_check_block_counters(reiser4_get_current_sb())); + + spin_unlock_reiser4_super(sbinfo); +} + +static void flush_reserved2used(txn_atom * atom, __u64 count) +{ + reiser4_super_info_data *sbinfo; + + assert("zam-787", atom != NULL); + assert_spin_locked(&(atom->alock)); + + sub_from_atom_flush_reserved_nolock(atom, (__u32) count); + + sbinfo = get_current_super_private(); + spin_lock_reiser4_super(sbinfo); + + sub_from_sb_flush_reserved(sbinfo, count); + sbinfo->blocks_used += count; + + assert("zam-789", + reiser4_check_block_counters(reiser4_get_current_sb())); + + spin_unlock_reiser4_super(sbinfo); +} + +/* update the per fs blocknr hint default value. */ +void +update_blocknr_hint_default(const struct super_block *s, + const reiser4_block_nr * block) +{ + reiser4_super_info_data *sbinfo = get_super_private(s); + + assert("nikita-3342", !reiser4_blocknr_is_fake(block)); + + spin_lock_reiser4_super(sbinfo); + if (*block < sbinfo->block_count) { + sbinfo->blocknr_hint_default = *block; + } else { + warning("zam-676", + "block number %llu is too large to be used in a blocknr hint\n", + (unsigned long long)*block); + dump_stack(); + DEBUGON(1); + } + spin_unlock_reiser4_super(sbinfo); +} + +/* get current value of the default blocknr hint. */ +void get_blocknr_hint_default(reiser4_block_nr * result) +{ + reiser4_super_info_data *sbinfo = get_current_super_private(); + + spin_lock_reiser4_super(sbinfo); + *result = sbinfo->blocknr_hint_default; + assert("zam-677", *result < sbinfo->block_count); + spin_unlock_reiser4_super(sbinfo); +} + +/* Allocate "real" disk blocks by calling a proper space allocation plugin + * method. Blocks are allocated in one contiguous disk region. The plugin + * independent part accounts blocks by subtracting allocated amount from grabbed + * or fake block counter and add the same amount to the counter of allocated + * blocks. + * + * @hint -- a reiser4 blocknr hint object which contains further block + * allocation hints and parameters (search start, a stage of block + * which will be mapped to disk, etc.), + * @blk -- an out parameter for the beginning of the allocated region, + * @len -- in/out parameter, it should contain the maximum number of allocated + * blocks, after block allocation completes, it contains the length of + * allocated disk region. + * @flags -- see reiser4_ba_flags_t description. + * + * @return -- 0 if success, error code otherwise. + */ +int +reiser4_alloc_blocks(reiser4_blocknr_hint * hint, reiser4_block_nr * blk, + reiser4_block_nr * len, reiser4_ba_flags_t flags) +{ + __u64 needed = *len; + reiser4_context *ctx; + reiser4_super_info_data *sbinfo; + int ret; + + assert("zam-986", hint != NULL); + + ctx = get_current_context(); + sbinfo = get_super_private(ctx->super); + + /* For write-optimized data we use default search start value, which is + * close to last write location. */ + if (flags & BA_USE_DEFAULT_SEARCH_START) { + get_blocknr_hint_default(&hint->blk); + } + + /* VITALY: allocator should grab this for internal/tx-lists/similar only. */ +/* VS-FIXME-HANS: why is this comment above addressed to vitaly (from vitaly)? */ + if (hint->block_stage == BLOCK_NOT_COUNTED) { + ret = reiser4_grab_space_force(*len, flags); + if (ret != 0) + return ret; + } + + ret = + sa_alloc_blocks(reiser4_get_space_allocator(ctx->super), + hint, (int)needed, blk, len); + + if (!ret) { + assert("zam-680", *blk < reiser4_block_count(ctx->super)); + assert("zam-681", + *blk + *len <= reiser4_block_count(ctx->super)); + + if (flags & BA_PERMANENT) { + /* we assume that current atom exists at this moment */ + txn_atom *atom = get_current_atom_locked(); + atom->nr_blocks_allocated += *len; + spin_unlock_atom(atom); + } + + switch (hint->block_stage) { + case BLOCK_NOT_COUNTED: + case BLOCK_GRABBED: + grabbed2used(ctx, sbinfo, *len); + break; + case BLOCK_UNALLOCATED: + fake_allocated2used(sbinfo, *len, flags); + break; + case BLOCK_FLUSH_RESERVED: + { + txn_atom *atom = get_current_atom_locked(); + flush_reserved2used(atom, *len); + spin_unlock_atom(atom); + } + break; + default: + impossible("zam-531", "wrong block stage"); + } + } else { + assert("zam-821", + ergo(hint->max_dist == 0 + && !hint->backward, ret != -ENOSPC)); + if (hint->block_stage == BLOCK_NOT_COUNTED) + grabbed2free(ctx, sbinfo, needed); + } + + return ret; +} + +/* used -> fake_allocated -> grabbed -> free */ + +/* adjust sb block counters when @count unallocated blocks get unmapped from + disk */ +static void +used2fake_allocated(reiser4_super_info_data * sbinfo, __u64 count, + int formatted) +{ + spin_lock_reiser4_super(sbinfo); + + if (formatted) + sbinfo->blocks_fake_allocated += count; + else + sbinfo->blocks_fake_allocated_unformatted += count; + + sub_from_sb_used(sbinfo, count); + + assert("nikita-2681", + reiser4_check_block_counters(reiser4_get_current_sb())); + + spin_unlock_reiser4_super(sbinfo); +} + +static void +used2flush_reserved(reiser4_super_info_data * sbinfo, txn_atom * atom, + __u64 count, reiser4_ba_flags_t flags UNUSED_ARG) +{ + assert("nikita-2791", atom != NULL); + assert_spin_locked(&(atom->alock)); + + add_to_atom_flush_reserved_nolock(atom, (__u32) count); + + spin_lock_reiser4_super(sbinfo); + + sbinfo->blocks_flush_reserved += count; + /*add_to_sb_flush_reserved(sbinfo, count); */ + sub_from_sb_used(sbinfo, count); + + assert("nikita-2681", + reiser4_check_block_counters(reiser4_get_current_sb())); + + spin_unlock_reiser4_super(sbinfo); +} + +/* disk space, virtually used by fake block numbers is counted as "grabbed" again. */ +static void +fake_allocated2grabbed(reiser4_context * ctx, reiser4_super_info_data * sbinfo, + __u64 count, reiser4_ba_flags_t flags) +{ + add_to_ctx_grabbed(ctx, count); + + spin_lock_reiser4_super(sbinfo); + + assert("nikita-2682", reiser4_check_block_counters(ctx->super)); + + sbinfo->blocks_grabbed += count; + sub_from_sb_fake_allocated(sbinfo, count, flags & BA_FORMATTED); + + assert("nikita-2683", reiser4_check_block_counters(ctx->super)); + + spin_unlock_reiser4_super(sbinfo); +} + +void fake_allocated2free(__u64 count, reiser4_ba_flags_t flags) +{ + reiser4_context *ctx; + reiser4_super_info_data *sbinfo; + + ctx = get_current_context(); + sbinfo = get_super_private(ctx->super); + + fake_allocated2grabbed(ctx, sbinfo, count, flags); + grabbed2free(ctx, sbinfo, count); +} + +void grabbed2free_mark(__u64 mark) +{ + reiser4_context *ctx; + reiser4_super_info_data *sbinfo; + + ctx = get_current_context(); + sbinfo = get_super_private(ctx->super); + + assert("nikita-3007", (__s64) mark >= 0); + assert("nikita-3006", ctx->grabbed_blocks >= mark); + grabbed2free(ctx, sbinfo, ctx->grabbed_blocks - mark); +} + +/** + * grabbed2free - adjust grabbed and free block counters + * @ctx: context to update grabbed block counter of + * @sbinfo: super block to update grabbed and free block counters of + * @count: number of blocks to adjust counters by + * + * Decreases context's and per filesystem's counters of grabbed + * blocks. Increases per filesystem's counter of free blocks. + */ +void grabbed2free(reiser4_context *ctx, reiser4_super_info_data *sbinfo, + __u64 count) +{ + sub_from_ctx_grabbed(ctx, count); + + spin_lock_reiser4_super(sbinfo); + + sub_from_sb_grabbed(sbinfo, count); + sbinfo->blocks_free += count; + assert("nikita-2684", reiser4_check_block_counters(ctx->super)); + + spin_unlock_reiser4_super(sbinfo); +} + +void grabbed2flush_reserved_nolock(txn_atom * atom, __u64 count) +{ + reiser4_context *ctx; + reiser4_super_info_data *sbinfo; + + assert("vs-1095", atom); + + ctx = get_current_context(); + sbinfo = get_super_private(ctx->super); + + sub_from_ctx_grabbed(ctx, count); + + add_to_atom_flush_reserved_nolock(atom, count); + + spin_lock_reiser4_super(sbinfo); + + sbinfo->blocks_flush_reserved += count; + sub_from_sb_grabbed(sbinfo, count); + + assert("vpf-292", reiser4_check_block_counters(ctx->super)); + + spin_unlock_reiser4_super(sbinfo); +} + +void grabbed2flush_reserved(__u64 count) +{ + txn_atom *atom = get_current_atom_locked(); + + grabbed2flush_reserved_nolock(atom, count); + + spin_unlock_atom(atom); +} + +void flush_reserved2grabbed(txn_atom * atom, __u64 count) +{ + reiser4_context *ctx; + reiser4_super_info_data *sbinfo; + + assert("nikita-2788", atom != NULL); + assert_spin_locked(&(atom->alock)); + + ctx = get_current_context(); + sbinfo = get_super_private(ctx->super); + + add_to_ctx_grabbed(ctx, count); + + sub_from_atom_flush_reserved_nolock(atom, (__u32) count); + + spin_lock_reiser4_super(sbinfo); + + sbinfo->blocks_grabbed += count; + sub_from_sb_flush_reserved(sbinfo, count); + + assert("vpf-292", reiser4_check_block_counters(ctx->super)); + + spin_unlock_reiser4_super(sbinfo); +} + +/** + * all_grabbed2free - releases all blocks grabbed in context + * + * Decreases context's and super block's grabbed block counters by number of + * blocks grabbed by current context and increases super block's free block + * counter correspondingly. + */ +void all_grabbed2free(void) +{ + reiser4_context *ctx = get_current_context(); + + grabbed2free(ctx, get_super_private(ctx->super), ctx->grabbed_blocks); +} + +/* adjust sb block counters if real (on-disk) blocks do not become unallocated + after freeing, @count blocks become "grabbed". */ +static void +used2grabbed(reiser4_context * ctx, reiser4_super_info_data * sbinfo, + __u64 count) +{ + add_to_ctx_grabbed(ctx, count); + + spin_lock_reiser4_super(sbinfo); + + sbinfo->blocks_grabbed += count; + sub_from_sb_used(sbinfo, count); + + assert("nikita-2685", reiser4_check_block_counters(ctx->super)); + + spin_unlock_reiser4_super(sbinfo); +} + +/* this used to be done through used2grabbed and grabbed2free*/ +static void used2free(reiser4_super_info_data * sbinfo, __u64 count) +{ + spin_lock_reiser4_super(sbinfo); + + sbinfo->blocks_free += count; + sub_from_sb_used(sbinfo, count); + + assert("nikita-2685", + reiser4_check_block_counters(reiser4_get_current_sb())); + + spin_unlock_reiser4_super(sbinfo); +} + +#if REISER4_DEBUG + +/* check "allocated" state of given block range */ +static void +reiser4_check_blocks(const reiser4_block_nr * start, + const reiser4_block_nr * len, int desired) +{ + sa_check_blocks(start, len, desired); +} + +/* check "allocated" state of given block */ +void reiser4_check_block(const reiser4_block_nr * block, int desired) +{ + const reiser4_block_nr one = 1; + + reiser4_check_blocks(block, &one, desired); +} + +#endif + +/* Blocks deallocation function may do an actual deallocation through space + plugin allocation or store deleted block numbers in atom's delete_set data + structure depend on @defer parameter. */ + +/* if BA_DEFER bit is not turned on, @target_stage means the stage of blocks which + will be deleted from WORKING bitmap. They might be just unmapped from disk, or + freed but disk space is still grabbed by current thread, or these blocks must + not be counted in any reiser4 sb block counters, see block_stage_t comment */ + +/* BA_FORMATTED bit is only used when BA_DEFER in not present: it is used to + distinguish blocks allocated for unformatted and formatted nodes */ + +int +reiser4_dealloc_blocks(const reiser4_block_nr * start, + const reiser4_block_nr * len, + block_stage_t target_stage, reiser4_ba_flags_t flags) +{ + txn_atom *atom = NULL; + int ret; + reiser4_context *ctx; + reiser4_super_info_data *sbinfo; + + ctx = get_current_context(); + sbinfo = get_super_private(ctx->super); + + if (REISER4_DEBUG) { + assert("zam-431", *len != 0); + assert("zam-432", *start != 0); + assert("zam-558", !reiser4_blocknr_is_fake(start)); + + spin_lock_reiser4_super(sbinfo); + assert("zam-562", *start < sbinfo->block_count); + spin_unlock_reiser4_super(sbinfo); + } + + if (flags & BA_DEFER) { + blocknr_set_entry *bsep = NULL; + + /* storing deleted block numbers in a blocknr set + datastructure for further actual deletion */ + do { + atom = get_current_atom_locked(); + assert("zam-430", atom != NULL); + + ret = + blocknr_set_add_extent(atom, &atom->delete_set, + &bsep, start, len); + + if (ret == -ENOMEM) + return ret; + + /* This loop might spin at most two times */ + } while (ret == -E_REPEAT); + + assert("zam-477", ret == 0); + assert("zam-433", atom != NULL); + + spin_unlock_atom(atom); + + } else { + assert("zam-425", get_current_super_private() != NULL); + sa_dealloc_blocks(reiser4_get_space_allocator(ctx->super), + *start, *len); + + if (flags & BA_PERMANENT) { + /* These blocks were counted as allocated, we have to revert it + * back if allocation is discarded. */ + txn_atom *atom = get_current_atom_locked(); + atom->nr_blocks_allocated -= *len; + spin_unlock_atom(atom); + } + + switch (target_stage) { + case BLOCK_NOT_COUNTED: + assert("vs-960", flags & BA_FORMATTED); + /* VITALY: This is what was grabbed for internal/tx-lists/similar only */ + used2free(sbinfo, *len); + break; + + case BLOCK_GRABBED: + used2grabbed(ctx, sbinfo, *len); + break; + + case BLOCK_UNALLOCATED: + used2fake_allocated(sbinfo, *len, flags & BA_FORMATTED); + break; + + case BLOCK_FLUSH_RESERVED:{ + txn_atom *atom; + + atom = get_current_atom_locked(); + used2flush_reserved(sbinfo, atom, *len, + flags & BA_FORMATTED); + spin_unlock_atom(atom); + break; + } + default: + impossible("zam-532", "wrong block stage"); + } + } + + return 0; +} + +/* wrappers for block allocator plugin methods */ +int reiser4_pre_commit_hook(void) +{ + assert("zam-502", get_current_super_private() != NULL); + sa_pre_commit_hook(); + return 0; +} + +/* an actor which applies delete set to block allocator data */ +static int +apply_dset(txn_atom * atom UNUSED_ARG, const reiser4_block_nr * a, + const reiser4_block_nr * b, void *data UNUSED_ARG) +{ + reiser4_context *ctx; + reiser4_super_info_data *sbinfo; + + __u64 len = 1; + + ctx = get_current_context(); + sbinfo = get_super_private(ctx->super); + + assert("zam-877", atom->stage >= ASTAGE_PRE_COMMIT); + assert("zam-552", sbinfo != NULL); + + if (b != NULL) + len = *b; + + if (REISER4_DEBUG) { + spin_lock_reiser4_super(sbinfo); + + assert("zam-554", *a < reiser4_block_count(ctx->super)); + assert("zam-555", *a + len <= reiser4_block_count(ctx->super)); + + spin_unlock_reiser4_super(sbinfo); + } + + sa_dealloc_blocks(&sbinfo->space_allocator, *a, len); + /* adjust sb block counters */ + used2free(sbinfo, len); + return 0; +} + +void reiser4_post_commit_hook(void) +{ + txn_atom *atom; + + atom = get_current_atom_locked(); + assert("zam-452", atom->stage == ASTAGE_POST_COMMIT); + spin_unlock_atom(atom); + + /* do the block deallocation which was deferred + until commit is done */ + blocknr_set_iterator(atom, &atom->delete_set, apply_dset, NULL, 1); + + assert("zam-504", get_current_super_private() != NULL); + sa_post_commit_hook(); +} + +void reiser4_post_write_back_hook(void) +{ + assert("zam-504", get_current_super_private() != NULL); + + sa_post_commit_hook(); +} + +/* + Local variables: + c-indentation-style: "K&R" + mode-name: "LC" + c-basic-offset: 8 + tab-width: 8 + fill-column: 120 + scroll-step: 1 + End: +*/ diff -urN linux-2.6.22.orig/fs/reiser4/block_alloc.h linux-2.6.22/fs/reiser4/block_alloc.h --- linux-2.6.22.orig/fs/reiser4/block_alloc.h 1970-01-01 03:00:00.000000000 +0300 +++ linux-2.6.22/fs/reiser4/block_alloc.h 2007-07-29 00:25:34.820681982 +0400 @@ -0,0 +1,175 @@ +/* Copyright 2002, 2003 by Hans Reiser, licensing governed by reiser4/README */ + +#if !defined (__FS_REISER4_BLOCK_ALLOC_H__) +#define __FS_REISER4_BLOCK_ALLOC_H__ + +#include "dformat.h" +#include "forward.h" + +#include /* for __u?? */ +#include + +/* Mask when is applied to given block number shows is that block number is a fake one */ +#define REISER4_FAKE_BLOCKNR_BIT_MASK 0x8000000000000000ULL +/* Mask which isolates a type of object this fake block number was assigned to */ +#define REISER4_BLOCKNR_STATUS_BIT_MASK 0xC000000000000000ULL + +/*result after applying the REISER4_BLOCKNR_STATUS_BIT_MASK should be compared + against these two values to understand is the object unallocated or bitmap + shadow object (WORKING BITMAP block, look at the plugin/space/bitmap.c) */ +#define REISER4_UNALLOCATED_STATUS_VALUE 0xC000000000000000ULL +#define REISER4_BITMAP_BLOCKS_STATUS_VALUE 0x8000000000000000ULL + +/* specification how block allocation was counted in sb block counters */ +typedef enum { + BLOCK_NOT_COUNTED = 0, /* reiser4 has no info about this block yet */ + BLOCK_GRABBED = 1, /* free space grabbed for further allocation + of this block */ + BLOCK_FLUSH_RESERVED = 2, /* block is reserved for flush needs. */ + BLOCK_UNALLOCATED = 3, /* block is used for existing in-memory object + ( unallocated formatted or unformatted + node) */ + BLOCK_ALLOCATED = 4 /* block is mapped to disk, real on-disk block + number assigned */ +} block_stage_t; + +/* a hint for block allocator */ +struct reiser4_blocknr_hint { + /* FIXME: I think we want to add a longterm lock on the bitmap block here. This + is to prevent jnode_flush() calls from interleaving allocations on the same + bitmap, once a hint is established. */ + + /* search start hint */ + reiser4_block_nr blk; + /* if not zero, it is a region size we search for free blocks in */ + reiser4_block_nr max_dist; + /* level for allocation, may be useful have branch-level and higher + write-optimized. */ + tree_level level; + /* block allocator assumes that blocks, which will be mapped to disk, + are in this specified block_stage */ + block_stage_t block_stage; + /* If direction = 1 allocate blocks in backward direction from the end + * of disk to the beginning of disk. */ + unsigned int backward:1; + +}; + +/* These flags control block allocation/deallocation behavior */ +enum reiser4_ba_flags { + /* do allocatations from reserved (5%) area */ + BA_RESERVED = (1 << 0), + + /* block allocator can do commit trying to recover free space */ + BA_CAN_COMMIT = (1 << 1), + + /* if operation will be applied to formatted block */ + BA_FORMATTED = (1 << 2), + + /* defer actual block freeing until transaction commit */ + BA_DEFER = (1 << 3), + + /* allocate blocks for permanent fs objects (formatted or unformatted), not + wandered of log blocks */ + BA_PERMANENT = (1 << 4), + + /* grab space even it was disabled */ + BA_FORCE = (1 << 5), + + /* use default start value for free blocks search. */ + BA_USE_DEFAULT_SEARCH_START = (1 << 6) +}; + +typedef enum reiser4_ba_flags reiser4_ba_flags_t; + +extern void reiser4_blocknr_hint_init(reiser4_blocknr_hint * hint); +extern void reiser4_blocknr_hint_done(reiser4_blocknr_hint * hint); +extern void update_blocknr_hint_default(const struct super_block *, + const reiser4_block_nr *); +extern void get_blocknr_hint_default(reiser4_block_nr *); + +extern reiser4_block_nr reiser4_fs_reserved_space(struct super_block *super); + +int assign_fake_blocknr_formatted(reiser4_block_nr *); +reiser4_block_nr fake_blocknr_unformatted(int); + +/* free -> grabbed -> fake_allocated -> used */ + +int reiser4_grab_space(__u64 count, reiser4_ba_flags_t flags); +void all_grabbed2free(void); +void grabbed2free(reiser4_context *, reiser4_super_info_data *, __u64 count); +void fake_allocated2free(__u64 count, reiser4_ba_flags_t flags); +void grabbed2flush_reserved_nolock(txn_atom * atom, __u64 count); +void grabbed2flush_reserved(__u64 count); +int reiser4_alloc_blocks(reiser4_blocknr_hint * hint, + reiser4_block_nr * start, + reiser4_block_nr * len, reiser4_ba_flags_t flags); +int reiser4_dealloc_blocks(const reiser4_block_nr *, + const reiser4_block_nr *, + block_stage_t, reiser4_ba_flags_t flags); + +static inline int reiser4_alloc_block(reiser4_blocknr_hint * hint, + reiser4_block_nr * start, + reiser4_ba_flags_t flags) +{ + reiser4_block_nr one = 1; + return reiser4_alloc_blocks(hint, start, &one, flags); +} + +static inline int reiser4_dealloc_block(const reiser4_block_nr * block, + block_stage_t stage, + reiser4_ba_flags_t flags) +{ + const reiser4_block_nr one = 1; + return reiser4_dealloc_blocks(block, &one, stage, flags); +} + +#define reiser4_grab_space_force(count, flags) \ + reiser4_grab_space(count, flags | BA_FORCE) + +extern void grabbed2free_mark(__u64 mark); +extern int reiser4_grab_reserved(struct super_block *, + __u64, reiser4_ba_flags_t); +extern void reiser4_release_reserved(struct super_block *super); + +/* grabbed -> fake_allocated */ + +/* fake_allocated -> used */ + +/* used -> fake_allocated -> grabbed -> free */ + +extern void flush_reserved2grabbed(txn_atom * atom, __u64 count); + +extern int reiser4_blocknr_is_fake(const reiser4_block_nr * da); + +extern void grabbed2cluster_reserved(int count); +extern void cluster_reserved2grabbed(int count); +extern void cluster_reserved2free(int count); + +extern int reiser4_check_block_counters(const struct super_block *); + +#if REISER4_DEBUG + +extern void reiser4_check_block(const reiser4_block_nr *, int); + +#else + +# define reiser4_check_block(beg, val) noop + +#endif + +extern int reiser4_pre_commit_hook(void); +extern void reiser4_post_commit_hook(void); +extern void reiser4_post_write_back_hook(void); + +#endif /* __FS_REISER4_BLOCK_ALLOC_H__ */ + +/* Make Linus happy. + Local variables: + c-indentation-style: "K&R" + mode-name: "LC" + c-basic-offset: 8 + tab-width: 8 + fill-column: 120 + End: +*/ diff -urN linux-2.6.22.orig/fs/reiser4/blocknrset.c linux-2.6.22/fs/reiser4/blocknrset.c --- linux-2.6.22.orig/fs/reiser4/blocknrset.c 1970-01-01 03:00:00.000000000 +0300 +++ linux-2.6.22/fs/reiser4/blocknrset.c 2007-07-29 00:25:34.820681982 +0400 @@ -0,0 +1,368 @@ +/* Copyright 2001, 2002, 2003 by Hans Reiser, licensing governed by reiser4/README */ + +/* This file contains code for various block number sets used by the atom to + track the deleted set and wandered block mappings. */ + +#include "debug.h" +#include "dformat.h" +#include "txnmgr.h" +#include "context.h" + +#include + +/* The proposed data structure for storing unordered block number sets is a + list of elements, each of which contains an array of block number or/and + array of block number pairs. That element called blocknr_set_entry is used + to store block numbers from the beginning and for extents from the end of + the data field (char data[...]). The ->nr_blocks and ->nr_pairs fields + count numbers of blocks and extents. + + +------------------- blocknr_set_entry->data ------------------+ + |block1|block2| ... ... |pair3|pair2|pair1| + +------------------------------------------------------------+ + + When current blocknr_set_entry is full, allocate a new one. */ + +/* Usage examples: blocknr sets are used in reiser4 for storing atom's delete + * set (single blocks and block extents), in that case blocknr pair represent an + * extent; atom's wandered map is also stored as a blocknr set, blocknr pairs + * there represent a (real block) -> (wandered block) mapping. */ + +/* Protection: blocknr sets belong to reiser4 atom, and + * their modifications are performed with the atom lock held */ + +/* The total size of a blocknr_set_entry. */ +#define BLOCKNR_SET_ENTRY_SIZE 128 + +/* The number of blocks that can fit the blocknr data area. */ +#define BLOCKNR_SET_ENTRIES_NUMBER \ + ((BLOCKNR_SET_ENTRY_SIZE - \ + 2 * sizeof (unsigned) - \ + sizeof(struct list_head)) / \ + sizeof(reiser4_block_nr)) + +/* An entry of the blocknr_set */ +struct blocknr_set_entry { + unsigned nr_singles; + unsigned nr_pairs; + struct list_head link; + reiser4_block_nr entries[BLOCKNR_SET_ENTRIES_NUMBER]; +}; + +/* A pair of blocks as recorded in the blocknr_set_entry data. */ +struct blocknr_pair { + reiser4_block_nr a; + reiser4_block_nr b; +}; + +/* Return the number of blocknr slots available in a blocknr_set_entry. */ +/* Audited by: green(2002.06.11) */ +static unsigned bse_avail(blocknr_set_entry * bse) +{ + unsigned used = bse->nr_singles + 2 * bse->nr_pairs; + + assert("jmacd-5088", BLOCKNR_SET_ENTRIES_NUMBER >= used); + cassert(sizeof(blocknr_set_entry) == BLOCKNR_SET_ENTRY_SIZE); + + return BLOCKNR_SET_ENTRIES_NUMBER - used; +} + +/* Initialize a blocknr_set_entry. */ +static void bse_init(blocknr_set_entry *bse) +{ + bse->nr_singles = 0; + bse->nr_pairs = 0; + INIT_LIST_HEAD(&bse->link); +} + +/* Allocate and initialize a blocknr_set_entry. */ +/* Audited by: green(2002.06.11) */ +static blocknr_set_entry *bse_alloc(void) +{ + blocknr_set_entry *e; + + if ((e = (blocknr_set_entry *) kmalloc(sizeof(blocknr_set_entry), + reiser4_ctx_gfp_mask_get())) == NULL) + return NULL; + + bse_init(e); + + return e; +} + +/* Free a blocknr_set_entry. */ +/* Audited by: green(2002.06.11) */ +static void bse_free(blocknr_set_entry * bse) +{ + kfree(bse); +} + +/* Add a block number to a blocknr_set_entry */ +/* Audited by: green(2002.06.11) */ +static void +bse_put_single(blocknr_set_entry * bse, const reiser4_block_nr * block) +{ + assert("jmacd-5099", bse_avail(bse) >= 1); + + bse->entries[bse->nr_singles++] = *block; +} + +/* Get a pair of block numbers */ +/* Audited by: green(2002.06.11) */ +static inline struct blocknr_pair *bse_get_pair(blocknr_set_entry * bse, + unsigned pno) +{ + assert("green-1", BLOCKNR_SET_ENTRIES_NUMBER >= 2 * (pno + 1)); + + return (struct blocknr_pair *) (bse->entries + + BLOCKNR_SET_ENTRIES_NUMBER - + 2 * (pno + 1)); +} + +/* Add a pair of block numbers to a blocknr_set_entry */ +/* Audited by: green(2002.06.11) */ +static void +bse_put_pair(blocknr_set_entry * bse, const reiser4_block_nr * a, + const reiser4_block_nr * b) +{ + struct blocknr_pair *pair; + + assert("jmacd-5100", bse_avail(bse) >= 2 && a != NULL && b != NULL); + + pair = bse_get_pair(bse, bse->nr_pairs++); + + pair->a = *a; + pair->b = *b; +} + +/* Add either a block or pair of blocks to the block number set. The first + blocknr (@a) must be non-NULL. If @b is NULL a single blocknr is added, if + @b is non-NULL a pair is added. The block number set belongs to atom, and + the call is made with the atom lock held. There may not be enough space in + the current blocknr_set_entry. If new_bsep points to a non-NULL + blocknr_set_entry then it will be added to the blocknr_set and new_bsep + will be set to NULL. If new_bsep contains NULL then the atom lock will be + released and a new bse will be allocated in new_bsep. E_REPEAT will be + returned with the atom unlocked for the operation to be tried again. If + the operation succeeds, 0 is returned. If new_bsep is non-NULL and not + used during the call, it will be freed automatically. */ +static int blocknr_set_add(txn_atom *atom, struct list_head *bset, + blocknr_set_entry **new_bsep, const reiser4_block_nr *a, + const reiser4_block_nr *b) +{ + blocknr_set_entry *bse; + unsigned entries_needed; + + assert("jmacd-5101", a != NULL); + + entries_needed = (b == NULL) ? 1 : 2; + if (list_empty(bset) || + bse_avail(list_entry(bset->next, blocknr_set_entry, link)) < entries_needed) { + /* See if a bse was previously allocated. */ + if (*new_bsep == NULL) { + spin_unlock_atom(atom); + *new_bsep = bse_alloc(); + return (*new_bsep != NULL) ? -E_REPEAT : + RETERR(-ENOMEM); + } + + /* Put it on the head of the list. */ + list_add(&((*new_bsep)->link), bset); + + *new_bsep = NULL; + } + + /* Add the single or pair. */ + bse = list_entry(bset->next, blocknr_set_entry, link); + if (b == NULL) { + bse_put_single(bse, a); + } else { + bse_put_pair(bse, a, b); + } + + /* If new_bsep is non-NULL then there was an allocation race, free this copy. */ + if (*new_bsep != NULL) { + bse_free(*new_bsep); + *new_bsep = NULL; + } + + return 0; +} + +/* Add an extent to the block set. If the length is 1, it is treated as a + single block (e.g., reiser4_set_add_block). */ +/* Audited by: green(2002.06.11) */ +/* Auditor note: Entire call chain cannot hold any spinlocks, because + kmalloc might schedule. The only exception is atom spinlock, which is + properly freed. */ +int +blocknr_set_add_extent(txn_atom * atom, + struct list_head * bset, + blocknr_set_entry ** new_bsep, + const reiser4_block_nr * start, + const reiser4_block_nr * len) +{ + assert("jmacd-5102", start != NULL && len != NULL && *len > 0); + return blocknr_set_add(atom, bset, new_bsep, start, + *len == 1 ? NULL : len); +} + +/* Add a block pair to the block set. It adds exactly a pair, which is checked + * by an assertion that both arguments are not null.*/ +/* Audited by: green(2002.06.11) */ +/* Auditor note: Entire call chain cannot hold any spinlocks, because + kmalloc might schedule. The only exception is atom spinlock, which is + properly freed. */ +int +blocknr_set_add_pair(txn_atom * atom, + struct list_head * bset, + blocknr_set_entry ** new_bsep, const reiser4_block_nr * a, + const reiser4_block_nr * b) +{ + assert("jmacd-5103", a != NULL && b != NULL); + return blocknr_set_add(atom, bset, new_bsep, a, b); +} + +/* Initialize a blocknr_set. */ +void blocknr_set_init(struct list_head *bset) +{ + INIT_LIST_HEAD(bset); +} + +/* Release the entries of a blocknr_set. */ +void blocknr_set_destroy(struct list_head *bset) +{ + blocknr_set_entry *bse; + + while (!list_empty(bset)) { + bse = list_entry(bset->next, blocknr_set_entry, link); + list_del_init(&bse->link); + bse_free(bse); + } +} + +/* Merge blocknr_set entries out of @from into @into. */ +/* Audited by: green(2002.06.11) */ +/* Auditor comments: This merge does not know if merged sets contain + blocks pairs (As for wandered sets) or extents, so it cannot really merge + overlapping ranges if there is some. So I believe it may lead to + some blocks being presented several times in one blocknr_set. To help + debugging such problems it might help to check for duplicate entries on + actual processing of this set. Testing this kind of stuff right here is + also complicated by the fact that these sets are not sorted and going + through whole set on each element addition is going to be CPU-heavy task */ +void blocknr_set_merge(struct list_head * from, struct list_head * into) +{ + blocknr_set_entry *bse_into = NULL; + + /* If @from is empty, no work to perform. */ + if (list_empty(from)) + return; + /* If @into is not empty, try merging partial-entries. */ + if (!list_empty(into)) { + + /* Neither set is empty, pop the front to members and try to combine them. */ + blocknr_set_entry *bse_from; + unsigned into_avail; + + bse_into = list_entry(into->next, blocknr_set_entry, link); + list_del_init(&bse_into->link); + bse_from = list_entry(from->next, blocknr_set_entry, link); + list_del_init(&bse_from->link); + + /* Combine singles. */ + for (into_avail = bse_avail(bse_into); + into_avail != 0 && bse_from->nr_singles != 0; + into_avail -= 1) { + bse_put_single(bse_into, + &bse_from->entries[--bse_from-> + nr_singles]); + } + + /* Combine pairs. */ + for (; into_avail > 1 && bse_from->nr_pairs != 0; + into_avail -= 2) { + struct blocknr_pair *pair = + bse_get_pair(bse_from, --bse_from->nr_pairs); + bse_put_pair(bse_into, &pair->a, &pair->b); + } + + /* If bse_from is empty, delete it now. */ + if (bse_avail(bse_from) == BLOCKNR_SET_ENTRIES_NUMBER) { + bse_free(bse_from); + } else { + /* Otherwise, bse_into is full or nearly full (e.g., + it could have one slot avail and bse_from has one + pair left). Push it back onto the list. bse_from + becomes bse_into, which will be the new partial. */ + list_add(&bse_into->link, into); + bse_into = bse_from; + } + } + + /* Splice lists together. */ + list_splice_init(from, into->prev); + + /* Add the partial entry back to the head of the list. */ + if (bse_into != NULL) + list_add(&bse_into->link, into); +} + +/* Iterate over all blocknr set elements. */ +int blocknr_set_iterator(txn_atom *atom, struct list_head *bset, + blocknr_set_actor_f actor, void *data, int delete) +{ + + blocknr_set_entry *entry; + + assert("zam-429", atom != NULL); + assert("zam-430", atom_is_protected(atom)); + assert("zam-431", bset != 0); + assert("zam-432", actor != NULL); + + entry = list_entry(bset->next, blocknr_set_entry, link); + while (bset != &entry->link) { + blocknr_set_entry *tmp = list_entry(entry->link.next, blocknr_set_entry, link); + unsigned int i; + int ret; + + for (i = 0; i < entry->nr_singles; i++) { + ret = actor(atom, &entry->entries[i], NULL, data); + + /* We can't break a loop if delete flag is set. */ + if (ret != 0 && !delete) + return ret; + } + + for (i = 0; i < entry->nr_pairs; i++) { + struct blocknr_pair *ab; + + ab = bse_get_pair(entry, i); + + ret = actor(atom, &ab->a, &ab->b, data); + + if (ret != 0 && !delete) + return ret; + } + + if (delete) { + list_del(&entry->link); + bse_free(entry); + } + + entry = tmp; + } + + return 0; +} + +/* + * Local variables: + * c-indentation-style: "K&R" + * mode-name: "LC" + * c-basic-offset: 8 + * tab-width: 8 + * fill-column: 79 + * scroll-step: 1 + * End: + */ diff -urN linux-2.6.22.orig/fs/reiser4/carry.c linux-2.6.22/fs/reiser4/carry.c --- linux-2.6.22.orig/fs/reiser4/carry.c 1970-01-01 03:00:00.000000000 +0300 +++ linux-2.6.22/fs/reiser4/carry.c 2007-07-29 00:25:34.820681982 +0400 @@ -0,0 +1,1391 @@ +/* Copyright 2001, 2002, 2003 by Hans Reiser, licensing governed by reiser4/README */ +/* Functions to "carry" tree modification(s) upward. */ +/* Tree is modified one level at a time. As we modify a level we accumulate a + set of changes that need to be propagated to the next level. We manage + node locking such that any searches that collide with carrying are + restarted, from the root if necessary. + + Insertion of a new item may result in items being moved among nodes and + this requires the delimiting key to be updated at the least common parent + of the nodes modified to preserve search tree invariants. Also, insertion + may require allocation of a new node. A pointer to the new node has to be + inserted into some node on the parent level, etc. + + Tree carrying is meant to be analogous to arithmetic carrying. + + A carry operation is always associated with some node (&carry_node). + + Carry process starts with some initial set of operations to be performed + and an initial set of already locked nodes. Operations are performed one + by one. Performing each single operation has following possible effects: + + - content of carry node associated with operation is modified + - new carry nodes are locked and involved into carry process on this level + - new carry operations are posted to the next level + + After all carry operations on this level are done, process is repeated for + the accumulated sequence on carry operations for the next level. This + starts by trying to lock (in left to right order) all carry nodes + associated with carry operations on the parent level. After this, we decide + whether more nodes are required on the left of already locked set. If so, + all locks taken on the parent level are released, new carry nodes are + added, and locking process repeats. + + It may happen that balancing process fails owing to unrecoverable error on + some of upper levels of a tree (possible causes are io error, failure to + allocate new node, etc.). In this case we should unmount the filesystem, + rebooting if it is the root, and possibly advise the use of fsck. + + USAGE: + + int some_tree_operation( znode *node, ... ) + { + // Allocate on a stack pool of carry objects: operations and nodes. + // Most carry processes will only take objects from here, without + // dynamic allocation. + +I feel uneasy about this pool. It adds to code complexity, I understand why it exists, but.... -Hans + + carry_pool pool; + carry_level lowest_level; + carry_op *op; + + init_carry_pool( &pool ); + init_carry_level( &lowest_level, &pool ); + + // operation may be one of: + // COP_INSERT --- insert new item into node + // COP_CUT --- remove part of or whole node + // COP_PASTE --- increase size of item + // COP_DELETE --- delete pointer from parent node + // COP_UPDATE --- update delimiting key in least + // common ancestor of two + + op = reiser4_post_carry( &lowest_level, operation, node, 0 ); + if( IS_ERR( op ) || ( op == NULL ) ) { + handle error + } else { + // fill in remaining fields in @op, according to carry.h:carry_op + result = carry( &lowest_level, NULL ); + } + done_carry_pool( &pool ); + } + + When you are implementing node plugin method that participates in carry + (shifting, insertion, deletion, etc.), do the following: + + int foo_node_method( znode *node, ..., carry_level *todo ) + { + carry_op *op; + + .... + + // note, that last argument to reiser4_post_carry() is non-null + // here, because @op is to be applied to the parent of @node, rather + // than to the @node itself as in the previous case. + + op = node_post_carry( todo, operation, node, 1 ); + // fill in remaining fields in @op, according to carry.h:carry_op + + .... + + } + + BATCHING: + + One of the main advantages of level-by-level balancing implemented here is + ability to batch updates on a parent level and to peform them more + efficiently as a result. + + Description To Be Done (TBD). + + DIFFICULTIES AND SUBTLE POINTS: + + 1. complex plumbing is required, because: + + a. effective allocation through pools is needed + + b. target of operation is not exactly known when operation is + posted. This is worked around through bitfields in &carry_node and + logic in lock_carry_node() + + c. of interaction with locking code: node should be added into sibling + list when pointer to it is inserted into its parent, which is some time + after node was created. Between these moments, node is somewhat in + suspended state and is only registered in the carry lists + + 2. whole balancing logic is implemented here, in particular, insertion + logic is coded in make_space(). + + 3. special cases like insertion (reiser4_add_tree_root()) or deletion + (reiser4_kill_tree_root()) of tree root and morphing of paste into insert + (insert_paste()) have to be handled. + + 4. there is non-trivial interdependency between allocation of new nodes + and almost everything else. This is mainly due to the (1.c) above. I shall + write about this later. + +*/ + +#include "forward.h" +#include "debug.h" +#include "key.h" +#include "coord.h" +#include "plugin/item/item.h" +#include "plugin/item/extent.h" +#include "plugin/node/node.h" +#include "jnode.h" +#include "znode.h" +#include "tree_mod.h" +#include "tree_walk.h" +#include "block_alloc.h" +#include "pool.h" +#include "tree.h" +#include "carry.h" +#include "carry_ops.h" +#include "super.h" +#include "reiser4.h" + +#include + +/* level locking/unlocking */ +static int lock_carry_level(carry_level * level); +static void unlock_carry_level(carry_level * level, int failure); +static void done_carry_level(carry_level * level); +static void unlock_carry_node(carry_level * level, carry_node * node, int fail); + +int lock_carry_node(carry_level * level, carry_node * node); +int lock_carry_node_tail(carry_node * node); + +/* carry processing proper */ +static int carry_on_level(carry_level * doing, carry_level * todo); + +static carry_op *add_op(carry_level * level, pool_ordering order, + carry_op * reference); + +/* handlers for carry operations. */ + +static void fatal_carry_error(carry_level * doing, int ecode); +static int add_new_root(carry_level * level, carry_node * node, znode * fake); + +static void print_level(const char *prefix, carry_level * level); + +#if REISER4_DEBUG +typedef enum { + CARRY_TODO, + CARRY_DOING +} carry_queue_state; +static int carry_level_invariant(carry_level * level, carry_queue_state state); +#endif + +/* main entry point for tree balancing. + + Tree carry performs operations from @doing and while doing so accumulates + information about operations to be performed on the next level ("carried" + to the parent level). Carried operations are performed, causing possibly + more operations to be carried upward etc. carry() takes care about + locking and pinning znodes while operating on them. + + For usage, see comment at the top of fs/reiser4/carry.c + +*/ +int reiser4_carry(carry_level * doing /* set of carry operations to be + * performed */ , + carry_level * done /* set of nodes, already performed + * at the previous level. + * NULL in most cases */) +{ + int result = 0; + /* queue of new requests */ + carry_level *todo; + ON_DEBUG(STORE_COUNTERS); + + assert("nikita-888", doing != NULL); + BUG_ON(done != NULL); + + todo = doing + 1; + init_carry_level(todo, doing->pool); + + /* queue of requests preformed on the previous level */ + done = todo + 1; + init_carry_level(done, doing->pool); + + /* iterate until there is nothing more to do */ + while (result == 0 && doing->ops_num > 0) { + carry_level *tmp; + + /* at this point @done is locked. */ + /* repeat lock/do/unlock while + + (1) lock_carry_level() fails due to deadlock avoidance, or + + (2) carry_on_level() decides that more nodes have to + be involved. + + (3) some unexpected error occurred while balancing on the + upper levels. In this case all changes are rolled back. + + */ + while (1) { + result = lock_carry_level(doing); + if (result == 0) { + /* perform operations from @doing and + accumulate new requests in @todo */ + result = carry_on_level(doing, todo); + if (result == 0) + break; + else if (result != -E_REPEAT || + !doing->restartable) { + warning("nikita-1043", + "Fatal error during carry: %i", + result); + print_level("done", done); + print_level("doing", doing); + print_level("todo", todo); + /* do some rough stuff like aborting + all pending transcrashes and thus + pushing tree back to the consistent + state. Alternatvely, just panic. + */ + fatal_carry_error(doing, result); + return result; + } + } else if (result != -E_REPEAT) { + fatal_carry_error(doing, result); + return result; + } + unlock_carry_level(doing, 1); + } + /* at this point @done can be safely unlocked */ + done_carry_level(done); + + /* cyclically shift queues */ + tmp = done; + done = doing; + doing = todo; + todo = tmp; + init_carry_level(todo, doing->pool); + + /* give other threads chance to run */ + reiser4_preempt_point(); + } + done_carry_level(done); + + /* all counters, but x_refs should remain the same. x_refs can change + owing to transaction manager */ + ON_DEBUG(CHECK_COUNTERS); + return result; +} + +/* perform carry operations on given level. + + Optimizations proposed by pooh: + + (1) don't lock all nodes from queue at the same time. Lock nodes lazily as + required; + + (2) unlock node if there are no more operations to be performed upon it and + node didn't add any operation to @todo. This can be implemented by + attaching to each node two counters: counter of operaions working on this + node and counter and operations carried upward from this node. + +*/ +static int carry_on_level(carry_level * doing /* queue of carry operations to + * do on this level */ , + carry_level * todo /* queue where new carry + * operations to be performed on + * the * parent level are + * accumulated during @doing + * processing. */ ) +{ + int result; + int (*f) (carry_op *, carry_level *, carry_level *); + carry_op *op; + carry_op *tmp_op; + + assert("nikita-1034", doing != NULL); + assert("nikita-1035", todo != NULL); + + /* @doing->nodes are locked. */ + + /* This function can be split into two phases: analysis and modification. + + Analysis calculates precisely what items should be moved between + nodes. This information is gathered in some structures attached to + each carry_node in a @doing queue. Analysis also determines whether + new nodes are to be allocated etc. + + After analysis is completed, actual modification is performed. Here + we can take advantage of "batch modification": if there are several + operations acting on the same node, modifications can be performed + more efficiently when batched together. + + Above is an optimization left for the future. + */ + /* Important, but delayed optimization: it's possible to batch + operations together and perform them more efficiently as a + result. For example, deletion of several neighboring items from a + node can be converted to a single ->cut() operation. + + Before processing queue, it should be scanned and "mergeable" + operations merged. + */ + result = 0; + for_all_ops(doing, op, tmp_op) { + carry_opcode opcode; + + assert("nikita-1041", op != NULL); + opcode = op->op; + assert("nikita-1042", op->op < COP_LAST_OP); + f = op_dispatch_table[op->op].handler; + result = f(op, doing, todo); + /* locking can fail with -E_REPEAT. Any different error is fatal + and will be handled by fatal_carry_error() sledgehammer. + */ + if (result != 0) + break; + } + if (result == 0) { + carry_plugin_info info; + carry_node *scan; + carry_node *tmp_scan; + + info.doing = doing; + info.todo = todo; + + assert("nikita-3002", + carry_level_invariant(doing, CARRY_DOING)); + for_all_nodes(doing, scan, tmp_scan) { + znode *node; + + node = reiser4_carry_real(scan); + assert("nikita-2547", node != NULL); + if (node_is_empty(node)) { + result = + node_plugin_by_node(node)-> + prepare_removal(node, &info); + if (result != 0) + break; + } + } + } + return result; +} + +/* post carry operation + + This is main function used by external carry clients: node layout plugins + and tree operations to create new carry operation to be performed on some + level. + + New operation will be included in the @level queue. To actually perform it, + call carry( level, ... ). This function takes write lock on @node. Carry + manages all its locks by itself, don't worry about this. + + This function adds operation and node at the end of the queue. It is up to + caller to guarantee proper ordering of node queue. + +*/ +carry_op * reiser4_post_carry(carry_level * level /* queue where new operation + * is to be posted at */ , + carry_opcode op /* opcode of operation */ , + znode * node /* node on which this operation + * will operate */ , + int apply_to_parent_p /* whether operation will + * operate directly on @node + * or on it parent. */) +{ + carry_op *result; + carry_node *child; + + assert("nikita-1046", level != NULL); + assert("nikita-1788", znode_is_write_locked(node)); + + result = add_op(level, POOLO_LAST, NULL); + if (IS_ERR(result)) + return result; + child = reiser4_add_carry(level, POOLO_LAST, NULL); + if (IS_ERR(child)) { + reiser4_pool_free(&level->pool->op_pool, &result->header); + return (carry_op *) child; + } + result->node = child; + result->op = op; + child->parent = apply_to_parent_p; + if (ZF_ISSET(node, JNODE_ORPHAN)) + child->left_before = 1; + child->node = node; + return result; +} + +/* initialize carry queue */ +void init_carry_level(carry_level * level /* level to initialize */ , + carry_pool * pool /* pool @level will allocate objects + * from */ ) +{ + assert("nikita-1045", level != NULL); + assert("nikita-967", pool != NULL); + + memset(level, 0, sizeof *level); + level->pool = pool; + + INIT_LIST_HEAD(&level->nodes); + INIT_LIST_HEAD(&level->ops); +} + +/* allocate carry pool and initialize pools within queue */ +carry_pool *init_carry_pool(int size) +{ + carry_pool *pool; + + assert("", size >= sizeof(carry_pool) + 3 * sizeof(carry_level)); + pool = kmalloc(size, reiser4_ctx_gfp_mask_get()); + if (pool == NULL) + return ERR_PTR(RETERR(-ENOMEM)); + + reiser4_init_pool(&pool->op_pool, sizeof(carry_op), CARRIES_POOL_SIZE, + (char *)pool->op); + reiser4_init_pool(&pool->node_pool, sizeof(carry_node), + NODES_LOCKED_POOL_SIZE, (char *)pool->node); + return pool; +} + +/* finish with queue pools */ +void done_carry_pool(carry_pool * pool /* pool to destroy */ ) +{ + reiser4_done_pool(&pool->op_pool); + reiser4_done_pool(&pool->node_pool); + kfree(pool); +} + +/* add new carry node to the @level. + + Returns pointer to the new carry node allocated from pool. It's up to + callers to maintain proper order in the @level. Assumption is that if carry + nodes on one level are already sorted and modifications are peroformed from + left to right, carry nodes added on the parent level will be ordered + automatically. To control ordering use @order and @reference parameters. + +*/ +carry_node *reiser4_add_carry_skip(carry_level * level /* &carry_level to add + * node to */ , + pool_ordering order /* where to insert: + * at the beginning of + * @level, + * before @reference, + * after @reference, + * at the end of @level + */ , + carry_node * reference/* reference node for + * insertion */) +{ + ON_DEBUG(carry_node * orig_ref = reference); + + if (order == POOLO_BEFORE) { + reference = find_left_carry(reference, level); + if (reference == NULL) + reference = list_entry(level->nodes.next, carry_node, + header.level_linkage); + else + reference = list_entry(reference->header.level_linkage.next, + carry_node, header.level_linkage); + } else if (order == POOLO_AFTER) { + reference = find_right_carry(reference, level); + if (reference == NULL) + reference = list_entry(level->nodes.prev, carry_node, + header.level_linkage); + else + reference = list_entry(reference->header.level_linkage.prev, + carry_node, header.level_linkage); + } + assert("nikita-2209", + ergo(orig_ref != NULL, + reiser4_carry_real(reference) == + reiser4_carry_real(orig_ref))); + return reiser4_add_carry(level, order, reference); +} + +carry_node *reiser4_add_carry(carry_level * level /* &carry_level to add node + * to */ , + pool_ordering order /* where to insert: at the + * beginning of @level, before + * @reference, after @reference, + * at the end of @level */ , + carry_node * reference /* reference node for + * insertion */ ) +{ + carry_node *result; + + result = + (carry_node *) reiser4_add_obj(&level->pool->node_pool, + &level->nodes, + order, &reference->header); + if (!IS_ERR(result) && (result != NULL)) + ++level->nodes_num; + return result; +} + +/* add new carry operation to the @level. + + Returns pointer to the new carry operations allocated from pool. It's up to + callers to maintain proper order in the @level. To control ordering use + @order and @reference parameters. + +*/ +static carry_op *add_op(carry_level * level /* &carry_level to add node to */ , + pool_ordering order /* where to insert: at the beginning of + * @level, before @reference, after + * @reference, at the end of @level */ , + carry_op * + reference /* reference node for insertion */ ) +{ + carry_op *result; + + result = + (carry_op *) reiser4_add_obj(&level->pool->op_pool, &level->ops, + order, &reference->header); + if (!IS_ERR(result) && (result != NULL)) + ++level->ops_num; + return result; +} + +/* Return node on the right of which @node was created. + + Each node is created on the right of some existing node (or it is new root, + which is special case not handled here). + + @node is new node created on some level, but not yet inserted into its + parent, it has corresponding bit (JNODE_ORPHAN) set in zstate. + +*/ +static carry_node *find_begetting_brother(carry_node * node /* node to start search + * from */ , + carry_level * kin UNUSED_ARG /* level to + * scan */ ) +{ + carry_node *scan; + + assert("nikita-1614", node != NULL); + assert("nikita-1615", kin != NULL); + assert("nikita-1616", LOCK_CNT_GTZ(rw_locked_tree)); + assert("nikita-1619", ergo(reiser4_carry_real(node) != NULL, + ZF_ISSET(reiser4_carry_real(node), + JNODE_ORPHAN))); + for (scan = node;; + scan = list_entry(scan->header.level_linkage.prev, carry_node, + header.level_linkage)) { + assert("nikita-1617", &kin->nodes != &scan->header.level_linkage); + if ((scan->node != node->node) && + !ZF_ISSET(scan->node, JNODE_ORPHAN)) { + assert("nikita-1618", reiser4_carry_real(scan) != NULL); + break; + } + } + return scan; +} + +static cmp_t +carry_node_cmp(carry_level * level, carry_node * n1, carry_node * n2) +{ + assert("nikita-2199", n1 != NULL); + assert("nikita-2200", n2 != NULL); + + if (n1 == n2) + return EQUAL_TO; + while (1) { + n1 = carry_node_next(n1); + if (carry_node_end(level, n1)) + return GREATER_THAN; + if (n1 == n2) + return LESS_THAN; + } + impossible("nikita-2201", "End of level reached"); +} + +carry_node *find_carry_node(carry_level * level, const znode * node) +{ + carry_node *scan; + carry_node *tmp_scan; + + assert("nikita-2202", level != NULL); + assert("nikita-2203", node != NULL); + + for_all_nodes(level, scan, tmp_scan) { + if (reiser4_carry_real(scan) == node) + return scan; + } + return NULL; +} + +znode *reiser4_carry_real(const carry_node * node) +{ + assert("nikita-3061", node != NULL); + + return node->lock_handle.node; +} + +carry_node *insert_carry_node(carry_level * doing, carry_level * todo, + const znode * node) +{ + carry_node *base; + carry_node *scan; + carry_node *tmp_scan; + carry_node *proj; + + base = find_carry_node(doing, node); + assert("nikita-2204", base != NULL); + + for_all_nodes(todo, scan, tmp_scan) { + proj = find_carry_node(doing, scan->node); + assert("nikita-2205", proj != NULL); + if (carry_node_cmp(doing, proj, base) != LESS_THAN) + break; + } + return scan; +} + +static carry_node *add_carry_atplace(carry_level * doing, carry_level * todo, + znode * node) +{ + carry_node *reference; + + assert("nikita-2994", doing != NULL); + assert("nikita-2995", todo != NULL); + assert("nikita-2996", node != NULL); + + reference = insert_carry_node(doing, todo, node); + assert("nikita-2997", reference != NULL); + + return reiser4_add_carry(todo, POOLO_BEFORE, reference); +} + +/* like reiser4_post_carry(), but designed to be called from node plugin methods. + This function is different from reiser4_post_carry() in that it finds proper + place to insert node in the queue. */ +carry_op *node_post_carry(carry_plugin_info * info /* carry parameters + * passed down to node + * plugin */ , + carry_opcode op /* opcode of operation */ , + znode * node /* node on which this + * operation will operate */ , + int apply_to_parent_p /* whether operation will + * operate directly on @node + * or on it parent. */ ) +{ + carry_op *result; + carry_node *child; + + assert("nikita-2207", info != NULL); + assert("nikita-2208", info->todo != NULL); + + if (info->doing == NULL) + return reiser4_post_carry(info->todo, op, node, + apply_to_parent_p); + + result = add_op(info->todo, POOLO_LAST, NULL); + if (IS_ERR(result)) + return result; + child = add_carry_atplace(info->doing, info->todo, node); + if (IS_ERR(child)) { + reiser4_pool_free(&info->todo->pool->op_pool, &result->header); + return (carry_op *) child; + } + result->node = child; + result->op = op; + child->parent = apply_to_parent_p; + if (ZF_ISSET(node, JNODE_ORPHAN)) + child->left_before = 1; + child->node = node; + return result; +} + +/* lock all carry nodes in @level */ +static int lock_carry_level(carry_level * level /* level to lock */ ) +{ + int result; + carry_node *node; + carry_node *tmp_node; + + assert("nikita-881", level != NULL); + assert("nikita-2229", carry_level_invariant(level, CARRY_TODO)); + + /* lock nodes from left to right */ + result = 0; + for_all_nodes(level, node, tmp_node) { + result = lock_carry_node(level, node); + if (result != 0) + break; + } + return result; +} + +/* Synchronize delimiting keys between @node and its left neighbor. + + To reduce contention on dk key and simplify carry code, we synchronize + delimiting keys only when carry ultimately leaves tree level (carrying + changes upward) and unlocks nodes at this level. + + This function first finds left neighbor of @node and then updates left + neighbor's right delimiting key to conincide with least key in @node. + +*/ + +ON_DEBUG(extern atomic_t delim_key_version; + ) + +static void sync_dkeys(znode * spot /* node to update */ ) +{ + reiser4_key pivot; + reiser4_tree *tree; + + assert("nikita-1610", spot != NULL); + assert("nikita-1612", LOCK_CNT_NIL(rw_locked_dk)); + + tree = znode_get_tree(spot); + read_lock_tree(tree); + write_lock_dk(tree); + + assert("nikita-2192", znode_is_loaded(spot)); + + /* sync left delimiting key of @spot with key in its leftmost item */ + if (node_is_empty(spot)) + pivot = *znode_get_rd_key(spot); + else + leftmost_key_in_node(spot, &pivot); + + znode_set_ld_key(spot, &pivot); + + /* there can be sequence of empty nodes pending removal on the left of + @spot. Scan them and update their left and right delimiting keys to + match left delimiting key of @spot. Also, update right delimiting + key of first non-empty left neighbor. + */ + while (1) { + if (!ZF_ISSET(spot, JNODE_LEFT_CONNECTED)) + break; + + spot = spot->left; + if (spot == NULL) + break; + + znode_set_rd_key(spot, &pivot); + /* don't sink into the domain of another balancing */ + if (!znode_is_write_locked(spot)) + break; + if (ZF_ISSET(spot, JNODE_HEARD_BANSHEE)) + znode_set_ld_key(spot, &pivot); + else + break; + } + + write_unlock_dk(tree); + read_unlock_tree(tree); +} + +/* unlock all carry nodes in @level */ +static void unlock_carry_level(carry_level * level /* level to unlock */ , + int failure /* true if unlocking owing to + * failure */ ) +{ + carry_node *node; + carry_node *tmp_node; + + assert("nikita-889", level != NULL); + + if (!failure) { + znode *spot; + + spot = NULL; + /* update delimiting keys */ + for_all_nodes(level, node, tmp_node) { + if (reiser4_carry_real(node) != spot) { + spot = reiser4_carry_real(node); + sync_dkeys(spot); + } + } + } + + /* nodes can be unlocked in arbitrary order. In preemptible + environment it's better to unlock in reverse order of locking, + though. + */ + for_all_nodes_back(level, node, tmp_node) { + /* all allocated nodes should be already linked to their + parents at this moment. */ + assert("nikita-1631", + ergo(!failure, !ZF_ISSET(reiser4_carry_real(node), + JNODE_ORPHAN))); + ON_DEBUG(check_dkeys(reiser4_carry_real(node))); + unlock_carry_node(level, node, failure); + } + level->new_root = NULL; +} + +/* finish with @level + + Unlock nodes and release all allocated resources */ +static void done_carry_level(carry_level * level /* level to finish */ ) +{ + carry_node *node; + carry_node *tmp_node; + carry_op *op; + carry_op *tmp_op; + + assert("nikita-1076", level != NULL); + + unlock_carry_level(level, 0); + for_all_nodes(level, node, tmp_node) { + assert("nikita-2113", list_empty_careful(&node->lock_handle.locks_link)); + assert("nikita-2114", list_empty_careful(&node->lock_handle.owners_link)); + reiser4_pool_free(&level->pool->node_pool, &node->header); + } + for_all_ops(level, op, tmp_op) + reiser4_pool_free(&level->pool->op_pool, &op->header); +} + +/* helper function to complete locking of carry node + + Finish locking of carry node. There are several ways in which new carry + node can be added into carry level and locked. Normal is through + lock_carry_node(), but also from find_{left|right}_neighbor(). This + function factors out common final part of all locking scenarios. It + supposes that @node -> lock_handle is lock handle for lock just taken and + fills ->real_node from this lock handle. + +*/ +int lock_carry_node_tail(carry_node * node /* node to complete locking of */ ) +{ + assert("nikita-1052", node != NULL); + assert("nikita-1187", reiser4_carry_real(node) != NULL); + assert("nikita-1188", !node->unlock); + + node->unlock = 1; + /* Load node content into memory and install node plugin by + looking at the node header. + + Most of the time this call is cheap because the node is + already in memory. + + Corresponding zrelse() is in unlock_carry_node() + */ + return zload(reiser4_carry_real(node)); +} + +/* lock carry node + + "Resolve" node to real znode, lock it and mark as locked. + This requires recursive locking of znodes. + + When operation is posted to the parent level, node it will be applied to is + not yet known. For example, when shifting data between two nodes, + delimiting has to be updated in parent or parents of nodes involved. But + their parents is not yet locked and, moreover said nodes can be reparented + by concurrent balancing. + + To work around this, carry operation is applied to special "carry node" + rather than to the znode itself. Carry node consists of some "base" or + "reference" znode and flags indicating how to get to the target of carry + operation (->real_node field of carry_node) from base. + +*/ +int lock_carry_node(carry_level * level /* level @node is in */ , + carry_node * node /* node to lock */ ) +{ + int result; + znode *reference_point; + lock_handle lh; + lock_handle tmp_lh; + reiser4_tree *tree; + + assert("nikita-887", level != NULL); + assert("nikita-882", node != NULL); + + result = 0; + reference_point = node->node; + init_lh(&lh); + init_lh(&tmp_lh); + if (node->left_before) { + /* handling of new nodes, allocated on the previous level: + + some carry ops were propably posted from the new node, but + this node neither has parent pointer set, nor is + connected. This will be done in ->create_hook() for + internal item. + + No then less, parent of new node has to be locked. To do + this, first go to the "left" in the carry order. This + depends on the decision to always allocate new node on the + right of existing one. + + Loop handles case when multiple nodes, all orphans, were + inserted. + + Strictly speaking, taking tree lock is not necessary here, + because all nodes scanned by loop in + find_begetting_brother() are write-locked by this thread, + and thus, their sibling linkage cannot change. + + */ + tree = znode_get_tree(reference_point); + read_lock_tree(tree); + reference_point = find_begetting_brother(node, level)->node; + read_unlock_tree(tree); + assert("nikita-1186", reference_point != NULL); + } + if (node->parent && (result == 0)) { + result = + reiser4_get_parent(&tmp_lh, reference_point, + ZNODE_WRITE_LOCK); + if (result != 0) { + ; /* nothing */ + } else if (znode_get_level(tmp_lh.node) == 0) { + assert("nikita-1347", znode_above_root(tmp_lh.node)); + result = add_new_root(level, node, tmp_lh.node); + if (result == 0) { + reference_point = level->new_root; + move_lh(&lh, &node->lock_handle); + } + } else if ((level->new_root != NULL) + && (level->new_root != + znode_parent_nolock(reference_point))) { + /* parent of node exists, but this level aready + created different new root, so */ + warning("nikita-1109", + /* it should be "radicis", but tradition is + tradition. do banshees read latin? */ + "hodie natus est radici frater"); + result = -EIO; + } else { + move_lh(&lh, &tmp_lh); + reference_point = lh.node; + } + } + if (node->left && (result == 0)) { + assert("nikita-1183", node->parent); + assert("nikita-883", reference_point != NULL); + result = + reiser4_get_left_neighbor(&tmp_lh, reference_point, + ZNODE_WRITE_LOCK, + GN_CAN_USE_UPPER_LEVELS); + if (result == 0) { + done_lh(&lh); + move_lh(&lh, &tmp_lh); + reference_point = lh.node; + } + } + if (!node->parent && !node->left && !node->left_before) { + result = + longterm_lock_znode(&lh, reference_point, ZNODE_WRITE_LOCK, + ZNODE_LOCK_HIPRI); + } + if (result == 0) { + move_lh(&node->lock_handle, &lh); + result = lock_carry_node_tail(node); + } + done_lh(&tmp_lh); + done_lh(&lh); + return result; +} + +/* release a lock on &carry_node. + + Release if necessary lock on @node. This opearion is pair of + lock_carry_node() and is idempotent: you can call it more than once on the + same node. + +*/ +static void +unlock_carry_node(carry_level * level, + carry_node * node /* node to be released */ , + int failure /* 0 if node is unlocked due + * to some error */ ) +{ + znode *real_node; + + assert("nikita-884", node != NULL); + + real_node = reiser4_carry_real(node); + /* pair to zload() in lock_carry_node_tail() */ + zrelse(real_node); + if (node->unlock && (real_node != NULL)) { + assert("nikita-899", real_node == node->lock_handle.node); + longterm_unlock_znode(&node->lock_handle); + } + if (failure) { + if (node->deallocate && (real_node != NULL)) { + /* free node in bitmap + + Prepare node for removal. Last zput() will finish + with it. + */ + ZF_SET(real_node, JNODE_HEARD_BANSHEE); + } + if (node->free) { + assert("nikita-2177", + list_empty_careful(&node->lock_handle.locks_link)); + assert("nikita-2112", + list_empty_careful(&node->lock_handle.owners_link)); + reiser4_pool_free(&level->pool->node_pool, + &node->header); + } + } +} + +/* fatal_carry_error() - all-catching error handling function + + It is possible that carry faces unrecoverable error, like unability to + insert pointer at the internal level. Our simple solution is just panic in + this situation. More sophisticated things like attempt to remount + file-system as read-only can be implemented without much difficlties. + + It is believed, that: + + 1. in stead of panicking, all current transactions can be aborted rolling + system back to the consistent state. + +Umm, if you simply panic without doing anything more at all, then all current +transactions are aborted and the system is rolled back to a consistent state, +by virtue of the design of the transactional mechanism. Well, wait, let's be +precise. If an internal node is corrupted on disk due to hardware failure, +then there may be no consistent state that can be rolled back to, so instead +we should say that it will rollback the transactions, which barring other +factors means rolling back to a consistent state. + +# Nikita: there is a subtle difference between panic and aborting +# transactions: machine doesn't reboot. Processes aren't killed. Processes +# don't using reiser4 (not that we care about such processes), or using other +# reiser4 mounts (about them we do care) will simply continue to run. With +# some luck, even application using aborted file system can survive: it will +# get some error, like EBADF, from each file descriptor on failed file system, +# but applications that do care about tolerance will cope with this (squid +# will). + +It would be a nice feature though to support rollback without rebooting +followed by remount, but this can wait for later versions. + + 2. once isolated transactions will be implemented it will be possible to + roll back offending transaction. + +2. is additional code complexity of inconsistent value (it implies that a broken tree should be kept in operation), so we must think about +it more before deciding if it should be done. -Hans + +*/ +static void fatal_carry_error(carry_level * doing UNUSED_ARG /* carry level + * where + * unrecoverable + * error + * occurred */ , + int ecode /* error code */ ) +{ + assert("nikita-1230", doing != NULL); + assert("nikita-1231", ecode < 0); + + reiser4_panic("nikita-1232", "Carry failed: %i", ecode); +} + +/* add new root to the tree + + This function itself only manages changes in carry structures and delegates + all hard work (allocation of znode for new root, changes of parent and + sibling pointers to the reiser4_add_tree_root(). + + Locking: old tree root is locked by carry at this point. Fake znode is also + locked. + +*/ +static int add_new_root(carry_level * level /* carry level in context of which + * operation is performed */ , + carry_node * node /* carry node for existing root */ , + znode * fake /* "fake" znode already locked by + * us */ ) +{ + int result; + + assert("nikita-1104", level != NULL); + assert("nikita-1105", node != NULL); + + assert("nikita-1403", znode_is_write_locked(node->node)); + assert("nikita-1404", znode_is_write_locked(fake)); + + /* trying to create new root. */ + /* @node is root and it's already locked by us. This + means that nobody else can be trying to add/remove + tree root right now. + */ + if (level->new_root == NULL) + level->new_root = reiser4_add_tree_root(node->node, fake); + if (!IS_ERR(level->new_root)) { + assert("nikita-1210", znode_is_root(level->new_root)); + node->deallocate = 1; + result = + longterm_lock_znode(&node->lock_handle, level->new_root, + ZNODE_WRITE_LOCK, ZNODE_LOCK_LOPRI); + if (result == 0) + zput(level->new_root); + } else { + result = PTR_ERR(level->new_root); + level->new_root = NULL; + } + return result; +} + +/* allocate new znode and add the operation that inserts the + pointer to it into the parent node into the todo level + + Allocate new znode, add it into carry queue and post into @todo queue + request to add pointer to new node into its parent. + + This is carry related routing that calls reiser4_new_node() to allocate new + node. +*/ +carry_node *add_new_znode(znode * brother /* existing left neighbor of new + * node */ , + carry_node * ref /* carry node after which new + * carry node is to be inserted + * into queue. This affects + * locking. */ , + carry_level * doing /* carry queue where new node is + * to be added */ , + carry_level * todo /* carry queue where COP_INSERT + * operation to add pointer to + * new node will ne added */ ) +{ + carry_node *fresh; + znode *new_znode; + carry_op *add_pointer; + carry_plugin_info info; + + assert("nikita-1048", brother != NULL); + assert("nikita-1049", todo != NULL); + + /* There is a lot of possible variations here: to what parent + new node will be attached and where. For simplicity, always + do the following: + + (1) new node and @brother will have the same parent. + + (2) new node is added on the right of @brother + + */ + + fresh = reiser4_add_carry_skip(doing, + ref ? POOLO_AFTER : POOLO_LAST, ref); + if (IS_ERR(fresh)) + return fresh; + + fresh->deallocate = 1; + fresh->free = 1; + + new_znode = reiser4_new_node(brother, znode_get_level(brother)); + if (IS_ERR(new_znode)) + /* @fresh will be deallocated automatically by error + handling code in the caller. */ + return (carry_node *) new_znode; + + /* new_znode returned znode with x_count 1. Caller has to decrease + it. make_space() does. */ + + ZF_SET(new_znode, JNODE_ORPHAN); + fresh->node = new_znode; + + while (ZF_ISSET(reiser4_carry_real(ref), JNODE_ORPHAN)) { + ref = carry_node_prev(ref); + assert("nikita-1606", !carry_node_end(doing, ref)); + } + + info.todo = todo; + info.doing = doing; + add_pointer = node_post_carry(&info, COP_INSERT, + reiser4_carry_real(ref), 1); + if (IS_ERR(add_pointer)) { + /* no need to deallocate @new_znode here: it will be + deallocated during carry error handling. */ + return (carry_node *) add_pointer; + } + + add_pointer->u.insert.type = COPT_CHILD; + add_pointer->u.insert.child = fresh; + add_pointer->u.insert.brother = brother; + /* initially new node spawns empty key range */ + write_lock_dk(znode_get_tree(brother)); + znode_set_ld_key(new_znode, + znode_set_rd_key(new_znode, + znode_get_rd_key(brother))); + write_unlock_dk(znode_get_tree(brother)); + return fresh; +} + +/* DEBUGGING FUNCTIONS. + + Probably we also should leave them on even when + debugging is turned off to print dumps at errors. +*/ +#if REISER4_DEBUG +static int carry_level_invariant(carry_level * level, carry_queue_state state) +{ + carry_node *node; + carry_node *tmp_node; + + if (level == NULL) + return 0; + + if (level->track_type != 0 && + level->track_type != CARRY_TRACK_NODE && + level->track_type != CARRY_TRACK_CHANGE) + return 0; + + /* check that nodes are in ascending order */ + for_all_nodes(level, node, tmp_node) { + znode *left; + znode *right; + + reiser4_key lkey; + reiser4_key rkey; + + if (node != carry_node_front(level)) { + if (state == CARRY_TODO) { + right = node->node; + left = carry_node_prev(node)->node; + } else { + right = reiser4_carry_real(node); + left = reiser4_carry_real(carry_node_prev(node)); + } + if (right == NULL || left == NULL) + continue; + if (node_is_empty(right) || node_is_empty(left)) + continue; + if (!keyle(leftmost_key_in_node(left, &lkey), + leftmost_key_in_node(right, &rkey))) { + warning("", "wrong key order"); + return 0; + } + } + } + return 1; +} +#endif + +/* get symbolic name for boolean */ +static const char *tf(int boolean /* truth value */ ) +{ + return boolean ? "t" : "f"; +} + +/* symbolic name for carry operation */ +static const char *carry_op_name(carry_opcode op /* carry opcode */ ) +{ + switch (op) { + case COP_INSERT: + return "COP_INSERT"; + case COP_DELETE: + return "COP_DELETE"; + case COP_CUT: + return "COP_CUT"; + case COP_PASTE: + return "COP_PASTE"; + case COP_UPDATE: + return "COP_UPDATE"; + case COP_EXTENT: + return "COP_EXTENT"; + case COP_INSERT_FLOW: + return "COP_INSERT_FLOW"; + default:{ + /* not mt safe, but who cares? */ + static char buf[20]; + + sprintf(buf, "unknown op: %x", op); + return buf; + } + } +} + +/* dump information about carry node */ +static void print_carry(const char *prefix /* prefix to print */ , + carry_node * node /* node to print */ ) +{ + if (node == NULL) { + printk("%s: null\n", prefix); + return; + } + printk + ("%s: %p parent: %s, left: %s, unlock: %s, free: %s, dealloc: %s\n", + prefix, node, tf(node->parent), tf(node->left), tf(node->unlock), + tf(node->free), tf(node->deallocate)); +} + +/* dump information about carry operation */ +static void print_op(const char *prefix /* prefix to print */ , + carry_op * op /* operation to print */ ) +{ + if (op == NULL) { + printk("%s: null\n", prefix); + return; + } + printk("%s: %p carry_opcode: %s\n", prefix, op, carry_op_name(op->op)); + print_carry("\tnode", op->node); + switch (op->op) { + case COP_INSERT: + case COP_PASTE: + print_coord("\tcoord", + op->u.insert.d ? op->u.insert.d->coord : NULL, 0); + reiser4_print_key("\tkey", + op->u.insert.d ? op->u.insert.d->key : NULL); + print_carry("\tchild", op->u.insert.child); + break; + case COP_DELETE: + print_carry("\tchild", op->u.delete.child); + break; + case COP_CUT: + if (op->u.cut_or_kill.is_cut) { + print_coord("\tfrom", + op->u.cut_or_kill.u.kill->params.from, 0); + print_coord("\tto", op->u.cut_or_kill.u.kill->params.to, + 0); + } else { + print_coord("\tfrom", + op->u.cut_or_kill.u.cut->params.from, 0); + print_coord("\tto", op->u.cut_or_kill.u.cut->params.to, + 0); + } + break; + case COP_UPDATE: + print_carry("\tleft", op->u.update.left); + break; + default: + /* do nothing */ + break; + } +} + +/* dump information about all nodes and operations in a @level */ +static void print_level(const char *prefix /* prefix to print */ , + carry_level * level /* level to print */ ) +{ + carry_node *node; + carry_node *tmp_node; + carry_op *op; + carry_op *tmp_op; + + if (level == NULL) { + printk("%s: null\n", prefix); + return; + } + printk("%s: %p, restartable: %s\n", + prefix, level, tf(level->restartable)); + + for_all_nodes(level, node, tmp_node) + print_carry("\tcarry node", node); + for_all_ops(level, op, tmp_op) + print_op("\tcarry op", op); +} + +/* Make Linus happy. + Local variables: + c-indentation-style: "K&R" + mode-name: "LC" + c-basic-offset: 8 + tab-width: 8 + fill-column: 120 + scroll-step: 1 + End: +*/ diff -urN linux-2.6.22.orig/fs/reiser4/carry.h linux-2.6.22/fs/reiser4/carry.h --- linux-2.6.22.orig/fs/reiser4/carry.h 1970-01-01 03:00:00.000000000 +0300 +++ linux-2.6.22/fs/reiser4/carry.h 2007-07-29 00:25:34.824683017 +0400 @@ -0,0 +1,442 @@ +/* Copyright 2001, 2002, 2003 by Hans Reiser, licensing governed by reiser4/README */ + +/* Functions and data types to "carry" tree modification(s) upward. + See fs/reiser4/carry.c for details. */ + +#if !defined( __FS_REISER4_CARRY_H__ ) +#define __FS_REISER4_CARRY_H__ + +#include "forward.h" +#include "debug.h" +#include "pool.h" +#include "znode.h" + +#include + +/* &carry_node - "location" of carry node. + + "location" of node that is involved or going to be involved into + carry process. Node where operation will be carried to on the + parent level cannot be recorded explicitly. Operation will be carried + usually to the parent of some node (where changes are performed at + the current level) or, to the left neighbor of its parent. But while + modifications are performed at the current level, parent may + change. So, we have to allow some indirection (or, positevly, + flexibility) in locating carry nodes. + +*/ +typedef struct carry_node { + /* pool linkage */ + struct reiser4_pool_header header; + + /* base node from which real_node is calculated. See + fs/reiser4/carry.c:lock_carry_node(). */ + znode *node; + + /* how to get ->real_node */ + /* to get ->real_node obtain parent of ->node */ + __u32 parent:1; + /* to get ->real_node obtain left neighbor of parent of + ->node */ + __u32 left:1; + __u32 left_before:1; + + /* locking */ + + /* this node was locked by carry process and should be + unlocked when carry leaves a level */ + __u32 unlock:1; + + /* disk block for this node was allocated by carry process and + should be deallocated when carry leaves a level */ + __u32 deallocate:1; + /* this carry node was allocated by carry process and should be + freed when carry leaves a level */ + __u32 free:1; + + /* type of lock we want to take on this node */ + lock_handle lock_handle; +} carry_node; + +/* &carry_opcode - elementary operations that can be carried upward + + Operations that carry() can handle. This list is supposed to be + expanded. + + Each carry operation (cop) is handled by appropriate function defined + in fs/reiser4/carry.c. For example COP_INSERT is handled by + fs/reiser4/carry.c:carry_insert() etc. These functions in turn + call plugins of nodes affected by operation to modify nodes' content + and to gather operations to be performed on the next level. + +*/ +typedef enum { + /* insert new item into node. */ + COP_INSERT, + /* delete pointer from parent node */ + COP_DELETE, + /* remove part of or whole node. */ + COP_CUT, + /* increase size of item. */ + COP_PASTE, + /* insert extent (that is sequence of unformatted nodes). */ + COP_EXTENT, + /* update delimiting key in least common ancestor of two + nodes. This is performed when items are moved between two + nodes. + */ + COP_UPDATE, + /* insert flow */ + COP_INSERT_FLOW, + COP_LAST_OP, +} carry_opcode; + +#define CARRY_FLOW_NEW_NODES_LIMIT 20 + +/* mode (or subtype) of COP_{INSERT|PASTE} operation. Specifies how target + item is determined. */ +typedef enum { + /* target item is one containing pointer to the ->child node */ + COPT_CHILD, + /* target item is given explicitly by @coord */ + COPT_ITEM_DATA, + /* target item is given by key */ + COPT_KEY, + /* see insert_paste_common() for more comments on this. */ + COPT_PASTE_RESTARTED, +} cop_insert_pos_type; + +/* flags to cut and delete */ +typedef enum { + /* don't kill node even if it became completely empty as results of + * cut. This is needed for eottl handling. See carry_extent() for + * details. */ + DELETE_RETAIN_EMPTY = (1 << 0) +} cop_delete_flag; + +/* + * carry() implements "lock handle tracking" feature. + * + * Callers supply carry with node where to perform initial operation and lock + * handle on this node. Trying to optimize node utilization carry may actually + * move insertion point to different node. Callers expect that lock handle + * will rebe transferred to the new node also. + * + */ +typedef enum { + /* transfer lock handle along with insertion point */ + CARRY_TRACK_CHANGE = 1, + /* acquire new lock handle to the node where insertion point is. This + * is used when carry() client doesn't initially possess lock handle + * on the insertion point node, for example, by extent insertion + * code. See carry_extent(). */ + CARRY_TRACK_NODE = 2 +} carry_track_type; + +/* data supplied to COP_{INSERT|PASTE} by callers */ +typedef struct carry_insert_data { + /* position where new item is to be inserted */ + coord_t *coord; + /* new item description */ + reiser4_item_data *data; + /* key of new item */ + const reiser4_key *key; +} carry_insert_data; + +/* cut and kill are similar, so carry_cut_data and carry_kill_data share the below structure of parameters */ +struct cut_kill_params { + /* coord where cut starts (inclusive) */ + coord_t *from; + /* coord where cut stops (inclusive, this item/unit will also be + * cut) */ + coord_t *to; + /* starting key. This is necessary when item and unit pos don't + * uniquely identify what portion or tree to remove. For example, this + * indicates what portion of extent unit will be affected. */ + const reiser4_key *from_key; + /* exclusive stop key */ + const reiser4_key *to_key; + /* if this is not NULL, smallest actually removed key is stored + * here. */ + reiser4_key *smallest_removed; + /* kill_node_content() is called for file truncate */ + int truncate; +}; + +struct carry_cut_data { + struct cut_kill_params params; +}; + +struct carry_kill_data { + struct cut_kill_params params; + /* parameter to be passed to the ->kill_hook() method of item + * plugin */ + /*void *iplug_params; *//* FIXME: unused currently */ + /* if not NULL---inode whose items are being removed. This is needed + * for ->kill_hook() of extent item to update VM structures when + * removing pages. */ + struct inode *inode; + /* sibling list maintenance is complicated by existence of eottl. When + * eottl whose left and right neighbors are formatted leaves is + * removed, one has to connect said leaves in the sibling list. This + * cannot be done when extent removal is just started as locking rules + * require sibling list update to happen atomically with removal of + * extent item. Therefore: 1. pointers to left and right neighbors + * have to be passed down to the ->kill_hook() of extent item, and + * 2. said neighbors have to be locked. */ + lock_handle *left; + lock_handle *right; + /* flags modifying behavior of kill. Currently, it may have DELETE_RETAIN_EMPTY set. */ + unsigned flags; + char *buf; +}; + +/* &carry_tree_op - operation to "carry" upward. + + Description of an operation we want to "carry" to the upper level of + a tree: e.g, when we insert something and there is not enough space + we allocate a new node and "carry" the operation of inserting a + pointer to the new node to the upper level, on removal of empty node, + we carry up operation of removing appropriate entry from parent. + + There are two types of carry ops: when adding or deleting node we + node at the parent level where appropriate modification has to be + performed is known in advance. When shifting items between nodes + (split, merge), delimiting key should be changed in the least common + parent of the nodes involved that is not known in advance. + + For the operations of the first type we store in &carry_op pointer to + the &carry_node at the parent level. For the operation of the second + type we store &carry_node or parents of the left and right nodes + modified and keep track of them upward until they coincide. + +*/ +typedef struct carry_op { + /* pool linkage */ + struct reiser4_pool_header header; + carry_opcode op; + /* node on which operation is to be performed: + + for insert, paste: node where new item is to be inserted + + for delete: node where pointer is to be deleted + + for cut: node to cut from + + for update: node where delimiting key is to be modified + + for modify: parent of modified node + + */ + carry_node *node; + union { + struct { + /* (sub-)type of insertion/paste. Taken from + cop_insert_pos_type. */ + __u8 type; + /* various operation flags. Taken from + cop_insert_flag. */ + __u8 flags; + carry_insert_data *d; + carry_node *child; + znode *brother; + } insert, paste, extent; + + struct { + int is_cut; + union { + carry_kill_data *kill; + carry_cut_data *cut; + } u; + } cut_or_kill; + + struct { + carry_node *left; + } update; + struct { + /* changed child */ + carry_node *child; + /* bitmask of changes. See &cop_modify_flag */ + __u32 flag; + } modify; + struct { + /* flags to deletion operation. Are taken from + cop_delete_flag */ + __u32 flags; + /* child to delete from parent. If this is + NULL, delete op->node. */ + carry_node *child; + } delete; + struct { + /* various operation flags. Taken from + cop_insert_flag. */ + __u32 flags; + flow_t *flow; + coord_t *insert_point; + reiser4_item_data *data; + /* flow insertion is limited by number of new blocks + added in that operation which do not get any data + but part of flow. This limit is set by macro + CARRY_FLOW_NEW_NODES_LIMIT. This field stores number + of nodes added already during one carry_flow */ + int new_nodes; + } insert_flow; + } u; +} carry_op; + +/* &carry_op_pool - preallocated pool of carry operations, and nodes */ +typedef struct carry_pool { + carry_op op[CARRIES_POOL_SIZE]; + struct reiser4_pool op_pool; + carry_node node[NODES_LOCKED_POOL_SIZE]; + struct reiser4_pool node_pool; +} carry_pool; + +/* &carry_tree_level - carry process on given level + + Description of balancing process on the given level. + + No need for locking here, as carry_tree_level is essentially per + thread thing (for now). + +*/ +struct carry_level { + /* this level may be restarted */ + __u32 restartable:1; + /* list of carry nodes on this level, ordered by key order */ + struct list_head nodes; + struct list_head ops; + /* pool where new objects are allocated from */ + carry_pool *pool; + int ops_num; + int nodes_num; + /* new root created on this level, if any */ + znode *new_root; + /* This is set by caller (insert_by_key(), rreiser4_esize_item(), etc.) + when they want ->tracked to automagically wander to the node where + insertion point moved after insert or paste. + */ + carry_track_type track_type; + /* lock handle supplied by user that we are tracking. See + above. */ + lock_handle *tracked; +}; + +/* information carry passes to plugin methods that may add new operations to + the @todo queue */ +struct carry_plugin_info { + carry_level *doing; + carry_level *todo; +}; + +int reiser4_carry(carry_level * doing, carry_level * done); + +carry_node *reiser4_add_carry(carry_level * level, pool_ordering order, + carry_node * reference); +carry_node *reiser4_add_carry_skip(carry_level * level, pool_ordering order, + carry_node * reference); + +extern carry_node *insert_carry_node(carry_level * doing, + carry_level * todo, const znode * node); + +extern carry_pool *init_carry_pool(int); +extern void done_carry_pool(carry_pool * pool); + +extern void init_carry_level(carry_level * level, carry_pool * pool); + +extern carry_op *reiser4_post_carry(carry_level * level, carry_opcode op, + znode * node, int apply_to_parent); +extern carry_op *node_post_carry(carry_plugin_info * info, carry_opcode op, + znode * node, int apply_to_parent_p); + +carry_node *add_new_znode(znode * brother, carry_node * reference, + carry_level * doing, carry_level * todo); + +carry_node *find_carry_node(carry_level * level, const znode * node); + +extern znode *reiser4_carry_real(const carry_node * node); + +/* helper macros to iterate over carry queues */ + +#define carry_node_next( node ) \ + list_entry((node)->header.level_linkage.next, carry_node, \ + header.level_linkage) + +#define carry_node_prev( node ) \ + list_entry((node)->header.level_linkage.prev, carry_node, \ + header.level_linkage) + +#define carry_node_front( level ) \ + list_entry((level)->nodes.next, carry_node, header.level_linkage) + +#define carry_node_back( level ) \ + list_entry((level)->nodes.prev, carry_node, header.level_linkage) + +#define carry_node_end( level, node ) \ + (&(level)->nodes == &(node)->header.level_linkage) + +/* macro to iterate over all operations in a @level */ +#define for_all_ops( level /* carry level (of type carry_level *) */, \ + op /* pointer to carry operation, modified by loop (of \ + * type carry_op *) */, \ + tmp /* pointer to carry operation (of type carry_op *), \ + * used to make iterator stable in the face of \ + * deletions from the level */ ) \ +for (op = list_entry(level->ops.next, carry_op, header.level_linkage), \ + tmp = list_entry(op->header.level_linkage.next, carry_op, header.level_linkage); \ + &op->header.level_linkage != &level->ops; \ + op = tmp, \ + tmp = list_entry(op->header.level_linkage.next, carry_op, header.level_linkage)) + +#if 0 +for( op = ( carry_op * ) pool_level_list_front( &level -> ops ), \ + tmp = ( carry_op * ) pool_level_list_next( &op -> header ) ; \ + ! pool_level_list_end( &level -> ops, &op -> header ) ; \ + op = tmp, tmp = ( carry_op * ) pool_level_list_next( &op -> header ) ) +#endif + +/* macro to iterate over all nodes in a @level */ \ +#define for_all_nodes( level /* carry level (of type carry_level *) */, \ + node /* pointer to carry node, modified by loop (of \ + * type carry_node *) */, \ + tmp /* pointer to carry node (of type carry_node *), \ + * used to make iterator stable in the face of * \ + * deletions from the level */ ) \ +for (node = list_entry(level->nodes.next, carry_node, header.level_linkage), \ + tmp = list_entry(node->header.level_linkage.next, carry_node, header.level_linkage); \ + &node->header.level_linkage != &level->nodes; \ + node = tmp, \ + tmp = list_entry(node->header.level_linkage.next, carry_node, header.level_linkage)) + +#if 0 +for( node = carry_node_front( level ), \ + tmp = carry_node_next( node ) ; ! carry_node_end( level, node ) ; \ + node = tmp, tmp = carry_node_next( node ) ) +#endif + +/* macro to iterate over all nodes in a @level in reverse order + + This is used, because nodes are unlocked in reversed order of locking */ +#define for_all_nodes_back( level /* carry level (of type carry_level *) */, \ + node /* pointer to carry node, modified by loop \ + * (of type carry_node *) */, \ + tmp /* pointer to carry node (of type carry_node \ + * *), used to make iterator stable in the \ + * face of deletions from the level */ ) \ +for( node = carry_node_back( level ), \ + tmp = carry_node_prev( node ) ; ! carry_node_end( level, node ) ; \ + node = tmp, tmp = carry_node_prev( node ) ) + +/* __FS_REISER4_CARRY_H__ */ +#endif + +/* Make Linus happy. + Local variables: + c-indentation-style: "K&R" + mode-name: "LC" + c-basic-offset: 8 + tab-width: 8 + fill-column: 120 + scroll-step: 1 + End: +*/ diff -urN linux-2.6.22.orig/fs/reiser4/carry_ops.c linux-2.6.22/fs/reiser4/carry_ops.c --- linux-2.6.22.orig/fs/reiser4/carry_ops.c 1970-01-01 03:00:00.000000000 +0300 +++ linux-2.6.22/fs/reiser4/carry_ops.c 2007-07-29 00:25:34.828684053 +0400 @@ -0,0 +1,2131 @@ +/* Copyright 2001, 2002, 2003 by Hans Reiser, licensing governed by reiser4/README */ + +/* implementation of carry operations */ + +#include "forward.h" +#include "debug.h" +#include "key.h" +#include "coord.h" +#include "plugin/item/item.h" +#include "plugin/node/node.h" +#include "jnode.h" +#include "znode.h" +#include "block_alloc.h" +#include "tree_walk.h" +#include "pool.h" +#include "tree_mod.h" +#include "carry.h" +#include "carry_ops.h" +#include "tree.h" +#include "super.h" +#include "reiser4.h" + +#include +#include + +static int carry_shift_data(sideof side, coord_t * insert_coord, znode * node, + carry_level * doing, carry_level * todo, + unsigned int including_insert_coord_p); + +extern int lock_carry_node(carry_level * level, carry_node * node); +extern int lock_carry_node_tail(carry_node * node); + +/* find left neighbor of a carry node + + Look for left neighbor of @node and add it to the @doing queue. See + comments in the body. + +*/ +static carry_node *find_left_neighbor(carry_op * op /* node to find left + * neighbor of */ , + carry_level * doing /* level to scan */ ) +{ + int result; + carry_node *node; + carry_node *left; + int flags; + reiser4_tree *tree; + + node = op->node; + + tree = current_tree; + read_lock_tree(tree); + /* first, check whether left neighbor is already in a @doing queue */ + if (reiser4_carry_real(node)->left != NULL) { + /* NOTE: there is locking subtlety here. Look into + * find_right_neighbor() for more info */ + if (find_carry_node(doing, + reiser4_carry_real(node)->left) != NULL) { + read_unlock_tree(tree); + left = node; + do { + left = list_entry(left->header.level_linkage.prev, + carry_node, header.level_linkage); + assert("nikita-3408", !carry_node_end(doing, + left)); + } while (reiser4_carry_real(left) == + reiser4_carry_real(node)); + return left; + } + } + read_unlock_tree(tree); + + left = reiser4_add_carry_skip(doing, POOLO_BEFORE, node); + if (IS_ERR(left)) + return left; + + left->node = node->node; + left->free = 1; + + flags = GN_TRY_LOCK; + if (!op->u.insert.flags & COPI_LOAD_LEFT) + flags |= GN_NO_ALLOC; + + /* then, feeling lucky, peek left neighbor in the cache. */ + result = reiser4_get_left_neighbor(&left->lock_handle, + reiser4_carry_real(node), + ZNODE_WRITE_LOCK, flags); + if (result == 0) { + /* ok, node found and locked. */ + result = lock_carry_node_tail(left); + if (result != 0) + left = ERR_PTR(result); + } else if (result == -E_NO_NEIGHBOR || result == -ENOENT) { + /* node is leftmost node in a tree, or neighbor wasn't in + cache, or there is an extent on the left. */ + reiser4_pool_free(&doing->pool->node_pool, &left->header); + left = NULL; + } else if (doing->restartable) { + /* if left neighbor is locked, and level is restartable, add + new node to @doing and restart. */ + assert("nikita-913", node->parent != 0); + assert("nikita-914", node->node != NULL); + left->left = 1; + left->free = 0; + left = ERR_PTR(-E_REPEAT); + } else { + /* left neighbor is locked, level cannot be restarted. Just + ignore left neighbor. */ + reiser4_pool_free(&doing->pool->node_pool, &left->header); + left = NULL; + } + return left; +} + +/* find right neighbor of a carry node + + Look for right neighbor of @node and add it to the @doing queue. See + comments in the body. + +*/ +static carry_node *find_right_neighbor(carry_op * op /* node to find right + * neighbor of */ , + carry_level * doing /* level to scan */ ) +{ + int result; + carry_node *node; + carry_node *right; + lock_handle lh; + int flags; + reiser4_tree *tree; + + init_lh(&lh); + + node = op->node; + + tree = current_tree; + read_lock_tree(tree); + /* first, check whether right neighbor is already in a @doing queue */ + if (reiser4_carry_real(node)->right != NULL) { + /* + * Tree lock is taken here anyway, because, even if _outcome_ + * of (find_carry_node() != NULL) doesn't depends on + * concurrent updates to ->right, find_carry_node() cannot + * work with second argument NULL. Hence, following comment is + * of historic importance only. + * + * Subtle: + * + * Q: why don't we need tree lock here, looking for the right + * neighbor? + * + * A: even if value of node->real_node->right were changed + * during find_carry_node() execution, outcome of execution + * wouldn't change, because (in short) other thread cannot add + * elements to the @doing, and if node->real_node->right + * already was in @doing, value of node->real_node->right + * couldn't change, because node cannot be inserted between + * locked neighbors. + */ + if (find_carry_node(doing, + reiser4_carry_real(node)->right) != NULL) { + read_unlock_tree(tree); + /* + * What we are doing here (this is also applicable to + * the find_left_neighbor()). + * + * tree_walk.c code requires that insertion of a + * pointer to a child, modification of parent pointer + * in the child, and insertion of the child into + * sibling list are atomic (see + * plugin/item/internal.c:create_hook_internal()). + * + * carry allocates new node long before pointer to it + * is inserted into parent and, actually, long before + * parent is even known. Such allocated-but-orphaned + * nodes are only trackable through carry level lists. + * + * Situation that is handled here is following: @node + * has valid ->right pointer, but there is + * allocated-but-orphaned node in the carry queue that + * is logically between @node and @node->right. Here + * we are searching for it. Critical point is that + * this is only possible if @node->right is also in + * the carry queue (this is checked above), because + * this is the only way new orphaned node could be + * inserted between them (before inserting new node, + * make_space() first tries to shift to the right, so, + * right neighbor will be locked and queued). + * + */ + right = node; + do { + right = list_entry(right->header.level_linkage.next, + carry_node, header.level_linkage); + assert("nikita-3408", !carry_node_end(doing, + right)); + } while (reiser4_carry_real(right) == + reiser4_carry_real(node)); + return right; + } + } + read_unlock_tree(tree); + + flags = GN_CAN_USE_UPPER_LEVELS; + if (!op->u.insert.flags & COPI_LOAD_RIGHT) + flags = GN_NO_ALLOC; + + /* then, try to lock right neighbor */ + init_lh(&lh); + result = reiser4_get_right_neighbor(&lh, + reiser4_carry_real(node), + ZNODE_WRITE_LOCK, flags); + if (result == 0) { + /* ok, node found and locked. */ + right = reiser4_add_carry_skip(doing, POOLO_AFTER, node); + if (!IS_ERR(right)) { + right->node = lh.node; + move_lh(&right->lock_handle, &lh); + right->free = 1; + result = lock_carry_node_tail(right); + if (result != 0) + right = ERR_PTR(result); + } + } else if ((result == -E_NO_NEIGHBOR) || (result == -ENOENT)) { + /* node is rightmost node in a tree, or neighbor wasn't in + cache, or there is an extent on the right. */ + right = NULL; + } else + right = ERR_PTR(result); + done_lh(&lh); + return right; +} + +/* how much free space in a @node is needed for @op + + How much space in @node is required for completion of @op, where @op is + insert or paste operation. +*/ +static unsigned int space_needed_for_op(znode * node /* znode data are + * inserted or + * pasted in */ , + carry_op * op /* carry + operation */ ) +{ + assert("nikita-919", op != NULL); + + switch (op->op) { + default: + impossible("nikita-1701", "Wrong opcode"); + case COP_INSERT: + return space_needed(node, NULL, op->u.insert.d->data, 1); + case COP_PASTE: + return space_needed(node, op->u.insert.d->coord, + op->u.insert.d->data, 0); + } +} + +/* how much space in @node is required to insert or paste @data at + @coord. */ +unsigned int space_needed(const znode * node /* node data are inserted or + * pasted in */ , + const coord_t * coord /* coord where data are + * inserted or pasted + * at */ , + const reiser4_item_data * data /* data to insert or + * paste */ , + int insertion /* non-0 is inserting, 0---paste */ ) +{ + int result; + item_plugin *iplug; + + assert("nikita-917", node != NULL); + assert("nikita-918", node_plugin_by_node(node) != NULL); + assert("vs-230", !insertion || (coord == NULL)); + + result = 0; + iplug = data->iplug; + if (iplug->b.estimate != NULL) { + /* ask item plugin how much space is needed to insert this + item */ + result += iplug->b.estimate(insertion ? NULL : coord, data); + } else { + /* reasonable default */ + result += data->length; + } + if (insertion) { + node_plugin *nplug; + + nplug = node->nplug; + /* and add node overhead */ + if (nplug->item_overhead != NULL) { + result += nplug->item_overhead(node, NULL); + } + } + return result; +} + +/* find &coord in parent where pointer to new child is to be stored. */ +static int find_new_child_coord(carry_op * op /* COP_INSERT carry operation to + * insert pointer to new + * child */ ) +{ + int result; + znode *node; + znode *child; + + assert("nikita-941", op != NULL); + assert("nikita-942", op->op == COP_INSERT); + + node = reiser4_carry_real(op->node); + assert("nikita-943", node != NULL); + assert("nikita-944", node_plugin_by_node(node) != NULL); + + child = reiser4_carry_real(op->u.insert.child); + result = + find_new_child_ptr(node, child, op->u.insert.brother, + op->u.insert.d->coord); + + build_child_ptr_data(child, op->u.insert.d->data); + return result; +} + +/* additional amount of free space in @node required to complete @op */ +static int free_space_shortage(znode * node /* node to check */ , + carry_op * op /* operation being performed */ ) +{ + assert("nikita-1061", node != NULL); + assert("nikita-1062", op != NULL); + + switch (op->op) { + default: + impossible("nikita-1702", "Wrong opcode"); + case COP_INSERT: + case COP_PASTE: + return space_needed_for_op(node, op) - znode_free_space(node); + case COP_EXTENT: + /* when inserting extent shift data around until insertion + point is utmost in the node. */ + if (coord_wrt(op->u.insert.d->coord) == COORD_INSIDE) + return +1; + else + return -1; + } +} + +/* helper function: update node pointer in operation after insertion + point was probably shifted into @target. */ +static znode *sync_op(carry_op * op, carry_node * target) +{ + znode *insertion_node; + + /* reget node from coord: shift might move insertion coord to + the neighbor */ + insertion_node = op->u.insert.d->coord->node; + /* if insertion point was actually moved into new node, + update carry node pointer in operation. */ + if (insertion_node != reiser4_carry_real(op->node)) { + op->node = target; + assert("nikita-2540", + reiser4_carry_real(target) == insertion_node); + } + assert("nikita-2541", + reiser4_carry_real(op->node) == op->u.insert.d->coord->node); + return insertion_node; +} + +/* + * complete make_space() call: update tracked lock handle if necessary. See + * comments for fs/reiser4/carry.h:carry_track_type + */ +static int +make_space_tail(carry_op * op, carry_level * doing, znode * orig_node) +{ + int result; + carry_track_type tracking; + znode *node; + + tracking = doing->track_type; + node = op->u.insert.d->coord->node; + + if (tracking == CARRY_TRACK_NODE || + (tracking == CARRY_TRACK_CHANGE && node != orig_node)) { + /* inserting or pasting into node different from + original. Update lock handle supplied by caller. */ + assert("nikita-1417", doing->tracked != NULL); + done_lh(doing->tracked); + init_lh(doing->tracked); + result = longterm_lock_znode(doing->tracked, node, + ZNODE_WRITE_LOCK, + ZNODE_LOCK_HIPRI); + } else + result = 0; + return result; +} + +/* This is insertion policy function. It shifts data to the left and right + neighbors of insertion coord and allocates new nodes until there is enough + free space to complete @op. + + See comments in the body. + + Assumes that the node format favors insertions at the right end of the node + as node40 does. + + See carry_flow() on detail about flow insertion +*/ +static int make_space(carry_op * op /* carry operation, insert or paste */ , + carry_level * doing /* current carry queue */ , + carry_level * todo /* carry queue on the parent level */ ) +{ + znode *node; + int result; + int not_enough_space; + int blk_alloc; + znode *orig_node; + __u32 flags; + + coord_t *coord; + + assert("nikita-890", op != NULL); + assert("nikita-891", todo != NULL); + assert("nikita-892", + op->op == COP_INSERT || + op->op == COP_PASTE || op->op == COP_EXTENT); + assert("nikita-1607", + reiser4_carry_real(op->node) == op->u.insert.d->coord->node); + + flags = op->u.insert.flags; + + /* NOTE check that new node can only be allocated after checking left + * and right neighbors. This is necessary for proper work of + * find_{left,right}_neighbor(). */ + assert("nikita-3410", ergo(flags & COPI_DONT_ALLOCATE, + flags & COPI_DONT_SHIFT_LEFT)); + assert("nikita-3411", ergo(flags & COPI_DONT_ALLOCATE, + flags & COPI_DONT_SHIFT_RIGHT)); + + coord = op->u.insert.d->coord; + orig_node = node = coord->node; + + assert("nikita-908", node != NULL); + assert("nikita-909", node_plugin_by_node(node) != NULL); + + result = 0; + /* If there is not enough space in a node, try to shift something to + the left neighbor. This is a bit tricky, as locking to the left is + low priority. This is handled by restart logic in carry(). + */ + not_enough_space = free_space_shortage(node, op); + if (not_enough_space <= 0) + /* it is possible that carry was called when there actually + was enough space in the node. For example, when inserting + leftmost item so that delimiting keys have to be updated. + */ + return make_space_tail(op, doing, orig_node); + if (!(flags & COPI_DONT_SHIFT_LEFT)) { + carry_node *left; + /* make note in statistics of an attempt to move + something into the left neighbor */ + left = find_left_neighbor(op, doing); + if (unlikely(IS_ERR(left))) { + if (PTR_ERR(left) == -E_REPEAT) + return -E_REPEAT; + else { + /* some error other than restart request + occurred. This shouldn't happen. Issue a + warning and continue as if left neighbor + weren't existing. + */ + warning("nikita-924", + "Error accessing left neighbor: %li", + PTR_ERR(left)); + } + } else if (left != NULL) { + + /* shift everything possible on the left of and + including insertion coord into the left neighbor */ + result = carry_shift_data(LEFT_SIDE, coord, + reiser4_carry_real(left), + doing, todo, + flags & COPI_GO_LEFT); + + /* reget node from coord: shift_left() might move + insertion coord to the left neighbor */ + node = sync_op(op, left); + + not_enough_space = free_space_shortage(node, op); + /* There is not enough free space in @node, but + may be, there is enough free space in + @left. Various balancing decisions are valid here. + The same for the shifiting to the right. + */ + } + } + /* If there still is not enough space, shift to the right */ + if (not_enough_space > 0 && !(flags & COPI_DONT_SHIFT_RIGHT)) { + carry_node *right; + + right = find_right_neighbor(op, doing); + if (IS_ERR(right)) { + warning("nikita-1065", + "Error accessing right neighbor: %li", + PTR_ERR(right)); + } else if (right != NULL) { + /* node containing insertion point, and its right + neighbor node are write locked by now. + + shift everything possible on the right of but + excluding insertion coord into the right neighbor + */ + result = carry_shift_data(RIGHT_SIDE, coord, + reiser4_carry_real(right), + doing, todo, + flags & COPI_GO_RIGHT); + /* reget node from coord: shift_right() might move + insertion coord to the right neighbor */ + node = sync_op(op, right); + not_enough_space = free_space_shortage(node, op); + } + } + /* If there is still not enough space, allocate new node(s). + + We try to allocate new blocks if COPI_DONT_ALLOCATE is not set in + the carry operation flags (currently this is needed during flush + only). + */ + for (blk_alloc = 0; + not_enough_space > 0 && result == 0 && blk_alloc < 2 && + !(flags & COPI_DONT_ALLOCATE); ++blk_alloc) { + carry_node *fresh; /* new node we are allocating */ + coord_t coord_shadow; /* remembered insertion point before + * shifting data into new node */ + carry_node *node_shadow; /* remembered insertion node before + * shifting */ + unsigned int gointo; /* whether insertion point should move + * into newly allocated node */ + + /* allocate new node on the right of @node. Znode and disk + fake block number for new node are allocated. + + add_new_znode() posts carry operation COP_INSERT with + COPT_CHILD option to the parent level to add + pointer to newly created node to its parent. + + Subtle point: if several new nodes are required to complete + insertion operation at this level, they will be inserted + into their parents in the order of creation, which means + that @node will be valid "cookie" at the time of insertion. + + */ + fresh = add_new_znode(node, op->node, doing, todo); + if (IS_ERR(fresh)) + return PTR_ERR(fresh); + + /* Try to shift into new node. */ + result = lock_carry_node(doing, fresh); + zput(reiser4_carry_real(fresh)); + if (result != 0) { + warning("nikita-947", + "Cannot lock new node: %i", result); + return result; + } + + /* both nodes are write locked by now. + + shift everything possible on the right of and + including insertion coord into the right neighbor. + */ + coord_dup(&coord_shadow, op->u.insert.d->coord); + node_shadow = op->node; + /* move insertion point into newly created node if: + + . insertion point is rightmost in the source node, or + . this is not the first node we are allocating in a row. + */ + gointo = + (blk_alloc > 0) || + coord_is_after_rightmost(op->u.insert.d->coord); + + if (gointo && + op->op == COP_PASTE && + coord_is_existing_item(op->u.insert.d->coord) && + is_solid_item((item_plugin_by_coord(op->u.insert.d->coord)))) { + /* paste into solid (atomic) item, which can contain + only one unit, so we need to shift it right, where + insertion point supposed to be */ + + assert("edward-1444", op->u.insert.d->data->iplug == + item_plugin_by_id(STATIC_STAT_DATA_ID)); + assert("edward-1445", + op->u.insert.d->data->length > + node_plugin_by_node(coord->node)->free_space + (coord->node)); + + op->u.insert.d->coord->between = BEFORE_UNIT; + } + + result = carry_shift_data(RIGHT_SIDE, coord, + reiser4_carry_real(fresh), + doing, todo, gointo); + /* if insertion point was actually moved into new node, + update carry node pointer in operation. */ + node = sync_op(op, fresh); + not_enough_space = free_space_shortage(node, op); + if ((not_enough_space > 0) && (node != coord_shadow.node)) { + /* there is not enough free in new node. Shift + insertion point back to the @shadow_node so that + next new node would be inserted between + @shadow_node and @fresh. + */ + coord_normalize(&coord_shadow); + coord_dup(coord, &coord_shadow); + node = coord->node; + op->node = node_shadow; + if (1 || (flags & COPI_STEP_BACK)) { + /* still not enough space?! Maybe there is + enough space in the source node (i.e., node + data are moved from) now. + */ + not_enough_space = + free_space_shortage(node, op); + } + } + } + if (not_enough_space > 0) { + if (!(flags & COPI_DONT_ALLOCATE)) + warning("nikita-948", "Cannot insert new item"); + result = -E_NODE_FULL; + } + assert("nikita-1622", ergo(result == 0, + reiser4_carry_real(op->node) == coord->node)); + assert("nikita-2616", coord == op->u.insert.d->coord); + if (result == 0) + result = make_space_tail(op, doing, orig_node); + return result; +} + +/* insert_paste_common() - common part of insert and paste operations + + This function performs common part of COP_INSERT and COP_PASTE. + + There are two ways in which insertion/paste can be requested: + + . by directly supplying reiser4_item_data. In this case, op -> + u.insert.type is set to COPT_ITEM_DATA. + + . by supplying child pointer to which is to inserted into parent. In this + case op -> u.insert.type == COPT_CHILD. + + . by supplying key of new item/unit. This is currently only used during + extent insertion + + This is required, because when new node is allocated we don't know at what + position pointer to it is to be stored in the parent. Actually, we don't + even know what its parent will be, because parent can be re-balanced + concurrently and new node re-parented, and because parent can be full and + pointer to the new node will go into some other node. + + insert_paste_common() resolves pointer to child node into position in the + parent by calling find_new_child_coord(), that fills + reiser4_item_data. After this, insertion/paste proceeds uniformly. + + Another complication is with finding free space during pasting. It may + happen that while shifting items to the neighbors and newly allocated + nodes, insertion coord can no longer be in the item we wanted to paste + into. At this point, paste becomes (morphs) into insert. Moreover free + space analysis has to be repeated, because amount of space required for + insertion is different from that of paste (item header overhead, etc). + + This function "unifies" different insertion modes (by resolving child + pointer or key into insertion coord), and then calls make_space() to free + enough space in the node by shifting data to the left and right and by + allocating new nodes if necessary. Carry operation knows amount of space + required for its completion. After enough free space is obtained, caller of + this function (carry_{insert,paste,etc.}) performs actual insertion/paste + by calling item plugin method. + +*/ +static int insert_paste_common(carry_op * op /* carry operation being + * performed */ , + carry_level * doing /* current carry level */ , + carry_level * todo /* next carry level */ , + carry_insert_data * cdata /* pointer to + * cdata */ , + coord_t * coord /* insertion/paste coord */ , + reiser4_item_data * data /* data to be + * inserted/pasted */ ) +{ + assert("nikita-981", op != NULL); + assert("nikita-980", todo != NULL); + assert("nikita-979", (op->op == COP_INSERT) || (op->op == COP_PASTE) + || (op->op == COP_EXTENT)); + + if (op->u.insert.type == COPT_PASTE_RESTARTED) { + /* nothing to do. Fall through to make_space(). */ + ; + } else if (op->u.insert.type == COPT_KEY) { + node_search_result intra_node; + znode *node; + /* Problem with doing batching at the lowest level, is that + operations here are given by coords where modification is + to be performed, and one modification can invalidate coords + of all following operations. + + So, we are implementing yet another type for operation that + will use (the only) "locator" stable across shifting of + data between nodes, etc.: key (COPT_KEY). + + This clause resolves key to the coord in the node. + + But node can change also. Probably some pieces have to be + added to the lock_carry_node(), to lock node by its key. + + */ + /* NOTE-NIKITA Lookup bias is fixed to FIND_EXACT. Complain + if you need something else. */ + op->u.insert.d->coord = coord; + node = reiser4_carry_real(op->node); + intra_node = node_plugin_by_node(node)->lookup + (node, op->u.insert.d->key, FIND_EXACT, + op->u.insert.d->coord); + if ((intra_node != NS_FOUND) && (intra_node != NS_NOT_FOUND)) { + warning("nikita-1715", "Intra node lookup failure: %i", + intra_node); + return intra_node; + } + } else if (op->u.insert.type == COPT_CHILD) { + /* if we are asked to insert pointer to the child into + internal node, first convert pointer to the child into + coord within parent node. + */ + znode *child; + int result; + + op->u.insert.d = cdata; + op->u.insert.d->coord = coord; + op->u.insert.d->data = data; + op->u.insert.d->coord->node = reiser4_carry_real(op->node); + result = find_new_child_coord(op); + child = reiser4_carry_real(op->u.insert.child); + if (result != NS_NOT_FOUND) { + warning("nikita-993", + "Cannot find a place for child pointer: %i", + result); + return result; + } + /* This only happens when we did multiple insertions at + the previous level, trying to insert single item and + it so happened, that insertion of pointers to all new + nodes before this one already caused parent node to + split (may be several times). + + I am going to come up with better solution. + + You are not expected to understand this. + -- v6root/usr/sys/ken/slp.c + + Basically, what happens here is the following: carry came + to the parent level and is about to insert internal item + pointing to the child node that it just inserted in the + level below. Position where internal item is to be inserted + was found by find_new_child_coord() above, but node of the + current carry operation (that is, parent node of child + inserted on the previous level), was determined earlier in + the lock_carry_level/lock_carry_node. It could so happen + that other carry operations already performed on the parent + level already split parent node, so that insertion point + moved into another node. Handle this by creating new carry + node for insertion point if necessary. + */ + if (reiser4_carry_real(op->node) != + op->u.insert.d->coord->node) { + pool_ordering direction; + znode *z1; + znode *z2; + reiser4_key k1; + reiser4_key k2; + + /* + * determine in what direction insertion point + * moved. Do this by comparing delimiting keys. + */ + z1 = op->u.insert.d->coord->node; + z2 = reiser4_carry_real(op->node); + if (keyle(leftmost_key_in_node(z1, &k1), + leftmost_key_in_node(z2, &k2))) + /* insertion point moved to the left */ + direction = POOLO_BEFORE; + else + /* insertion point moved to the right */ + direction = POOLO_AFTER; + + op->node = reiser4_add_carry_skip(doing, + direction, op->node); + if (IS_ERR(op->node)) + return PTR_ERR(op->node); + op->node->node = op->u.insert.d->coord->node; + op->node->free = 1; + result = lock_carry_node(doing, op->node); + if (result != 0) + return result; + } + + /* + * set up key of an item being inserted: we are inserting + * internal item and its key is (by the very definition of + * search tree) is leftmost key in the child node. + */ + write_lock_dk(znode_get_tree(child)); + op->u.insert.d->key = leftmost_key_in_node(child, + znode_get_ld_key(child)); + write_unlock_dk(znode_get_tree(child)); + op->u.insert.d->data->arg = op->u.insert.brother; + } else { + assert("vs-243", op->u.insert.d->coord != NULL); + op->u.insert.d->coord->node = reiser4_carry_real(op->node); + } + + /* find free space. */ + return make_space(op, doing, todo); +} + +/* handle carry COP_INSERT operation. + + Insert new item into node. New item can be given in one of two ways: + + - by passing &tree_coord and &reiser4_item_data as part of @op. This is + only applicable at the leaf/twig level. + + - by passing a child node pointer to which is to be inserted by this + operation. + +*/ +static int carry_insert(carry_op * op /* operation to perform */ , + carry_level * doing /* queue of operations @op + * is part of */ , + carry_level * todo /* queue where new operations + * are accumulated */ ) +{ + znode *node; + carry_insert_data cdata; + coord_t coord; + reiser4_item_data data; + carry_plugin_info info; + int result; + + assert("nikita-1036", op != NULL); + assert("nikita-1037", todo != NULL); + assert("nikita-1038", op->op == COP_INSERT); + + coord_init_zero(&coord); + + /* perform common functionality of insert and paste. */ + result = insert_paste_common(op, doing, todo, &cdata, &coord, &data); + if (result != 0) + return result; + + node = op->u.insert.d->coord->node; + assert("nikita-1039", node != NULL); + assert("nikita-1040", node_plugin_by_node(node) != NULL); + + assert("nikita-949", + space_needed_for_op(node, op) <= znode_free_space(node)); + + /* ask node layout to create new item. */ + info.doing = doing; + info.todo = todo; + result = node_plugin_by_node(node)->create_item + (op->u.insert.d->coord, op->u.insert.d->key, op->u.insert.d->data, + &info); + doing->restartable = 0; + znode_make_dirty(node); + + return result; +} + +/* + * Flow insertion code. COP_INSERT_FLOW is special tree operation that is + * supplied with a "flow" (that is, a stream of data) and inserts it into tree + * by slicing into multiple items. + */ + +#define flow_insert_point(op) ( ( op ) -> u.insert_flow.insert_point ) +#define flow_insert_flow(op) ( ( op ) -> u.insert_flow.flow ) +#define flow_insert_data(op) ( ( op ) -> u.insert_flow.data ) + +static size_t item_data_overhead(carry_op * op) +{ + if (flow_insert_data(op)->iplug->b.estimate == NULL) + return 0; + return (flow_insert_data(op)->iplug->b. + estimate(NULL /* estimate insertion */ , flow_insert_data(op)) - + flow_insert_data(op)->length); +} + +/* FIXME-VS: this is called several times during one make_flow_for_insertion + and it will always return the same result. Some optimization could be made + by calculating this value once at the beginning and passing it around. That + would reduce some flexibility in future changes +*/ +static int can_paste(coord_t *, const reiser4_key *, const reiser4_item_data *); +static size_t flow_insertion_overhead(carry_op * op) +{ + znode *node; + size_t insertion_overhead; + + node = flow_insert_point(op)->node; + insertion_overhead = 0; + if (node->nplug->item_overhead && + !can_paste(flow_insert_point(op), &flow_insert_flow(op)->key, + flow_insert_data(op))) + insertion_overhead = + node->nplug->item_overhead(node, NULL) + + item_data_overhead(op); + return insertion_overhead; +} + +/* how many bytes of flow does fit to the node */ +static int what_can_fit_into_node(carry_op * op) +{ + size_t free, overhead; + + overhead = flow_insertion_overhead(op); + free = znode_free_space(flow_insert_point(op)->node); + if (free <= overhead) + return 0; + free -= overhead; + /* FIXME: flow->length is loff_t only to not get overflowed in case of expandign truncate */ + if (free < op->u.insert_flow.flow->length) + return free; + return (int)op->u.insert_flow.flow->length; +} + +/* in make_space_for_flow_insertion we need to check either whether whole flow + fits into a node or whether minimal fraction of flow fits into a node */ +static int enough_space_for_whole_flow(carry_op * op) +{ + return (unsigned)what_can_fit_into_node(op) == + op->u.insert_flow.flow->length; +} + +#define MIN_FLOW_FRACTION 1 +static int enough_space_for_min_flow_fraction(carry_op * op) +{ + assert("vs-902", coord_is_after_rightmost(flow_insert_point(op))); + + return what_can_fit_into_node(op) >= MIN_FLOW_FRACTION; +} + +/* this returns 0 if left neighbor was obtained successfully and everything + upto insertion point including it were shifted and left neighbor still has + some free space to put minimal fraction of flow into it */ +static int +make_space_by_shift_left(carry_op * op, carry_level * doing, carry_level * todo) +{ + carry_node *left; + znode *orig; + + left = find_left_neighbor(op, doing); + if (unlikely(IS_ERR(left))) { + warning("vs-899", + "make_space_by_shift_left: " + "error accessing left neighbor: %li", PTR_ERR(left)); + return 1; + } + if (left == NULL) + /* left neighbor either does not exist or is unformatted + node */ + return 1; + + orig = flow_insert_point(op)->node; + /* try to shift content of node @orig from its head upto insert point + including insertion point into the left neighbor */ + carry_shift_data(LEFT_SIDE, flow_insert_point(op), + reiser4_carry_real(left), doing, todo, + 1 /* including insert point */); + if (reiser4_carry_real(left) != flow_insert_point(op)->node) { + /* insertion point did not move */ + return 1; + } + + /* insertion point is set after last item in the node */ + assert("vs-900", coord_is_after_rightmost(flow_insert_point(op))); + + if (!enough_space_for_min_flow_fraction(op)) { + /* insertion point node does not have enough free space to put + even minimal portion of flow into it, therefore, move + insertion point back to orig node (before first item) */ + coord_init_before_first_item(flow_insert_point(op), orig); + return 1; + } + + /* part of flow is to be written to the end of node */ + op->node = left; + return 0; +} + +/* this returns 0 if right neighbor was obtained successfully and everything to + the right of insertion point was shifted to it and node got enough free + space to put minimal fraction of flow into it */ +static int +make_space_by_shift_right(carry_op * op, carry_level * doing, + carry_level * todo) +{ + carry_node *right; + + right = find_right_neighbor(op, doing); + if (unlikely(IS_ERR(right))) { + warning("nikita-1065", "shift_right_excluding_insert_point: " + "error accessing right neighbor: %li", PTR_ERR(right)); + return 1; + } + if (right) { + /* shift everything possible on the right of but excluding + insertion coord into the right neighbor */ + carry_shift_data(RIGHT_SIDE, flow_insert_point(op), + reiser4_carry_real(right), doing, todo, + 0 /* not including insert point */); + } else { + /* right neighbor either does not exist or is unformatted + node */ + ; + } + if (coord_is_after_rightmost(flow_insert_point(op))) { + if (enough_space_for_min_flow_fraction(op)) { + /* part of flow is to be written to the end of node */ + return 0; + } + } + + /* new node is to be added if insert point node did not get enough + space for whole flow */ + return 1; +} + +/* this returns 0 when insert coord is set at the node end and fraction of flow + fits into that node */ +static int +make_space_by_new_nodes(carry_op * op, carry_level * doing, carry_level * todo) +{ + int result; + znode *node; + carry_node *new; + + node = flow_insert_point(op)->node; + + if (op->u.insert_flow.new_nodes == CARRY_FLOW_NEW_NODES_LIMIT) + return RETERR(-E_NODE_FULL); + /* add new node after insert point node */ + new = add_new_znode(node, op->node, doing, todo); + if (unlikely(IS_ERR(new))) { + return PTR_ERR(new); + } + result = lock_carry_node(doing, new); + zput(reiser4_carry_real(new)); + if (unlikely(result)) { + return result; + } + op->u.insert_flow.new_nodes++; + if (!coord_is_after_rightmost(flow_insert_point(op))) { + carry_shift_data(RIGHT_SIDE, flow_insert_point(op), + reiser4_carry_real(new), doing, todo, + 0 /* not including insert point */); + assert("vs-901", + coord_is_after_rightmost(flow_insert_point(op))); + + if (enough_space_for_min_flow_fraction(op)) { + return 0; + } + if (op->u.insert_flow.new_nodes == CARRY_FLOW_NEW_NODES_LIMIT) + return RETERR(-E_NODE_FULL); + + /* add one more new node */ + new = add_new_znode(node, op->node, doing, todo); + if (unlikely(IS_ERR(new))) { + return PTR_ERR(new); + } + result = lock_carry_node(doing, new); + zput(reiser4_carry_real(new)); + if (unlikely(result)) { + return result; + } + op->u.insert_flow.new_nodes++; + } + + /* move insertion point to new node */ + coord_init_before_first_item(flow_insert_point(op), + reiser4_carry_real(new)); + op->node = new; + return 0; +} + +static int +make_space_for_flow_insertion(carry_op * op, carry_level * doing, + carry_level * todo) +{ + __u32 flags = op->u.insert_flow.flags; + + if (enough_space_for_whole_flow(op)) { + /* whole flow fits into insert point node */ + return 0; + } + + if (!(flags & COPI_DONT_SHIFT_LEFT) + && (make_space_by_shift_left(op, doing, todo) == 0)) { + /* insert point is shifted to left neighbor of original insert + point node and is set after last unit in that node. It has + enough space to fit at least minimal fraction of flow. */ + return 0; + } + + if (enough_space_for_whole_flow(op)) { + /* whole flow fits into insert point node */ + return 0; + } + + if (!(flags & COPI_DONT_SHIFT_RIGHT) + && (make_space_by_shift_right(op, doing, todo) == 0)) { + /* insert point is still set to the same node, but there is + nothing to the right of insert point. */ + return 0; + } + + if (enough_space_for_whole_flow(op)) { + /* whole flow fits into insert point node */ + return 0; + } + + return make_space_by_new_nodes(op, doing, todo); +} + +/* implements COP_INSERT_FLOW operation */ +static int +carry_insert_flow(carry_op * op, carry_level * doing, carry_level * todo) +{ + int result; + flow_t *f; + coord_t *insert_point; + node_plugin *nplug; + carry_plugin_info info; + znode *orig_node; + lock_handle *orig_lh; + + f = op->u.insert_flow.flow; + result = 0; + + /* carry system needs this to work */ + info.doing = doing; + info.todo = todo; + + orig_node = flow_insert_point(op)->node; + orig_lh = doing->tracked; + + while (f->length) { + result = make_space_for_flow_insertion(op, doing, todo); + if (result) + break; + + insert_point = flow_insert_point(op); + nplug = node_plugin_by_node(insert_point->node); + + /* compose item data for insertion/pasting */ + flow_insert_data(op)->data = f->data; + flow_insert_data(op)->length = what_can_fit_into_node(op); + + if (can_paste(insert_point, &f->key, flow_insert_data(op))) { + /* insert point is set to item of file we are writing to and we have to append to it */ + assert("vs-903", insert_point->between == AFTER_UNIT); + nplug->change_item_size(insert_point, + flow_insert_data(op)->length); + flow_insert_data(op)->iplug->b.paste(insert_point, + flow_insert_data + (op), &info); + } else { + /* new item must be inserted */ + pos_in_node_t new_pos; + flow_insert_data(op)->length += item_data_overhead(op); + + /* FIXME-VS: this is because node40_create_item changes + insert_point for obscure reasons */ + switch (insert_point->between) { + case AFTER_ITEM: + new_pos = insert_point->item_pos + 1; + break; + case EMPTY_NODE: + new_pos = 0; + break; + case BEFORE_ITEM: + assert("vs-905", insert_point->item_pos == 0); + new_pos = 0; + break; + default: + impossible("vs-906", + "carry_insert_flow: invalid coord"); + new_pos = 0; + break; + } + + nplug->create_item(insert_point, &f->key, + flow_insert_data(op), &info); + coord_set_item_pos(insert_point, new_pos); + } + coord_init_after_item_end(insert_point); + doing->restartable = 0; + znode_make_dirty(insert_point->node); + + move_flow_forward(f, (unsigned)flow_insert_data(op)->length); + } + + if (orig_node != flow_insert_point(op)->node) { + /* move lock to new insert point */ + done_lh(orig_lh); + init_lh(orig_lh); + result = + longterm_lock_znode(orig_lh, flow_insert_point(op)->node, + ZNODE_WRITE_LOCK, ZNODE_LOCK_HIPRI); + } + + return result; +} + +/* implements COP_DELETE operation + + Remove pointer to @op -> u.delete.child from it's parent. + + This function also handles killing of a tree root is last pointer from it + was removed. This is complicated by our handling of "twig" level: root on + twig level is never killed. + +*/ +static int carry_delete(carry_op * op /* operation to be performed */ , + carry_level * doing UNUSED_ARG /* current carry + * level */ , + carry_level * todo /* next carry level */ ) +{ + int result; + coord_t coord; + coord_t coord2; + znode *parent; + znode *child; + carry_plugin_info info; + reiser4_tree *tree; + + /* + * This operation is called to delete internal item pointing to the + * child node that was removed by carry from the tree on the previous + * tree level. + */ + + assert("nikita-893", op != NULL); + assert("nikita-894", todo != NULL); + assert("nikita-895", op->op == COP_DELETE); + + coord_init_zero(&coord); + coord_init_zero(&coord2); + + parent = reiser4_carry_real(op->node); + child = op->u.delete.child ? + reiser4_carry_real(op->u.delete.child) : op->node->node; + tree = znode_get_tree(child); + read_lock_tree(tree); + + /* + * @parent was determined when carry entered parent level + * (lock_carry_level/lock_carry_node). Since then, actual parent of + * @child node could change due to other carry operations performed on + * the parent level. Check for this. + */ + + if (znode_parent(child) != parent) { + /* NOTE-NIKITA add stat counter for this. */ + parent = znode_parent(child); + assert("nikita-2581", find_carry_node(doing, parent)); + } + read_unlock_tree(tree); + + assert("nikita-1213", znode_get_level(parent) > LEAF_LEVEL); + + /* Twig level horrors: tree should be of height at least 2. So, last + pointer from the root at twig level is preserved even if child is + empty. This is ugly, but so it was architectured. + */ + + if (znode_is_root(parent) && + znode_get_level(parent) <= REISER4_MIN_TREE_HEIGHT && + node_num_items(parent) == 1) { + /* Delimiting key manipulations. */ + write_lock_dk(tree); + znode_set_ld_key(child, znode_set_ld_key(parent, reiser4_min_key())); + znode_set_rd_key(child, znode_set_rd_key(parent, reiser4_max_key())); + ZF_SET(child, JNODE_DKSET); + write_unlock_dk(tree); + + /* @child escaped imminent death! */ + ZF_CLR(child, JNODE_HEARD_BANSHEE); + return 0; + } + + /* convert child pointer to the coord_t */ + result = find_child_ptr(parent, child, &coord); + if (result != NS_FOUND) { + warning("nikita-994", "Cannot find child pointer: %i", result); + print_coord_content("coord", &coord); + return result; + } + + coord_dup(&coord2, &coord); + info.doing = doing; + info.todo = todo; + { + /* + * Actually kill internal item: prepare structure with + * arguments for ->cut_and_kill() method... + */ + + struct carry_kill_data kdata; + kdata.params.from = &coord; + kdata.params.to = &coord2; + kdata.params.from_key = NULL; + kdata.params.to_key = NULL; + kdata.params.smallest_removed = NULL; + kdata.params.truncate = 1; + kdata.flags = op->u.delete.flags; + kdata.inode = NULL; + kdata.left = NULL; + kdata.right = NULL; + kdata.buf = NULL; + /* ... and call it. */ + result = node_plugin_by_node(parent)->cut_and_kill(&kdata, + &info); + } + doing->restartable = 0; + + /* check whether root should be killed violently */ + if (znode_is_root(parent) && + /* don't kill roots at and lower than twig level */ + znode_get_level(parent) > REISER4_MIN_TREE_HEIGHT && + node_num_items(parent) == 1) { + result = reiser4_kill_tree_root(coord.node); + } + + return result < 0 ? : 0; +} + +/* implements COP_CUT opration + + Cuts part or whole content of node. + +*/ +static int carry_cut(carry_op * op /* operation to be performed */ , + carry_level * doing /* current carry level */ , + carry_level * todo /* next carry level */ ) +{ + int result; + carry_plugin_info info; + node_plugin *nplug; + + assert("nikita-896", op != NULL); + assert("nikita-897", todo != NULL); + assert("nikita-898", op->op == COP_CUT); + + info.doing = doing; + info.todo = todo; + + nplug = node_plugin_by_node(reiser4_carry_real(op->node)); + if (op->u.cut_or_kill.is_cut) + result = nplug->cut(op->u.cut_or_kill.u.cut, &info); + else + result = nplug->cut_and_kill(op->u.cut_or_kill.u.kill, &info); + + doing->restartable = 0; + return result < 0 ? : 0; +} + +/* helper function for carry_paste(): returns true if @op can be continued as + paste */ +static int +can_paste(coord_t * icoord, const reiser4_key * key, + const reiser4_item_data * data) +{ + coord_t circa; + item_plugin *new_iplug; + item_plugin *old_iplug; + int result = 0; /* to keep gcc shut */ + + assert("", icoord->between != AT_UNIT); + + /* obviously, one cannot paste when node is empty---there is nothing + to paste into. */ + if (node_is_empty(icoord->node)) + return 0; + /* if insertion point is at the middle of the item, then paste */ + if (!coord_is_between_items(icoord)) + return 1; + coord_dup(&circa, icoord); + circa.between = AT_UNIT; + + old_iplug = item_plugin_by_coord(&circa); + new_iplug = data->iplug; + + /* check whether we can paste to the item @icoord is "at" when we + ignore ->between field */ + if (old_iplug == new_iplug && item_can_contain_key(&circa, key, data)) { + result = 1; + } else if (icoord->between == BEFORE_UNIT + || icoord->between == BEFORE_ITEM) { + /* otherwise, try to glue to the item at the left, if any */ + coord_dup(&circa, icoord); + if (coord_set_to_left(&circa)) { + result = 0; + coord_init_before_item(icoord); + } else { + old_iplug = item_plugin_by_coord(&circa); + result = (old_iplug == new_iplug) + && item_can_contain_key(icoord, key, data); + if (result) { + coord_dup(icoord, &circa); + icoord->between = AFTER_UNIT; + } + } + } else if (icoord->between == AFTER_UNIT + || icoord->between == AFTER_ITEM) { + coord_dup(&circa, icoord); + /* otherwise, try to glue to the item at the right, if any */ + if (coord_set_to_right(&circa)) { + result = 0; + coord_init_after_item(icoord); + } else { + int (*cck) (const coord_t *, const reiser4_key *, + const reiser4_item_data *); + + old_iplug = item_plugin_by_coord(&circa); + + cck = old_iplug->b.can_contain_key; + if (cck == NULL) + /* item doesn't define ->can_contain_key + method? So it is not expandable. */ + result = 0; + else { + result = (old_iplug == new_iplug) + && cck(&circa /*icoord */ , key, data); + if (result) { + coord_dup(icoord, &circa); + icoord->between = BEFORE_UNIT; + } + } + } + } else + impossible("nikita-2513", "Nothing works"); + if (result) { + if (icoord->between == BEFORE_ITEM) { + assert("vs-912", icoord->unit_pos == 0); + icoord->between = BEFORE_UNIT; + } else if (icoord->between == AFTER_ITEM) { + coord_init_after_item_end(icoord); + } + } + return result; +} + +/* implements COP_PASTE operation + + Paste data into existing item. This is complicated by the fact that after + we shifted something to the left or right neighbors trying to free some + space, item we were supposed to paste into can be in different node than + insertion coord. If so, we are no longer doing paste, but insert. See + comments in insert_paste_common(). + +*/ +static int carry_paste(carry_op * op /* operation to be performed */ , + carry_level * doing UNUSED_ARG /* current carry + * level */ , + carry_level * todo /* next carry level */ ) +{ + znode *node; + carry_insert_data cdata; + coord_t dcoord; + reiser4_item_data data; + int result; + int real_size; + item_plugin *iplug; + carry_plugin_info info; + coord_t *coord; + + assert("nikita-982", op != NULL); + assert("nikita-983", todo != NULL); + assert("nikita-984", op->op == COP_PASTE); + + coord_init_zero(&dcoord); + + result = insert_paste_common(op, doing, todo, &cdata, &dcoord, &data); + if (result != 0) + return result; + + coord = op->u.insert.d->coord; + + /* handle case when op -> u.insert.coord doesn't point to the item + of required type. restart as insert. */ + if (!can_paste(coord, op->u.insert.d->key, op->u.insert.d->data)) { + op->op = COP_INSERT; + op->u.insert.type = COPT_PASTE_RESTARTED; + result = op_dispatch_table[COP_INSERT].handler(op, doing, todo); + + return result; + } + + node = coord->node; + iplug = item_plugin_by_coord(coord); + assert("nikita-992", iplug != NULL); + + assert("nikita-985", node != NULL); + assert("nikita-986", node_plugin_by_node(node) != NULL); + + assert("nikita-987", + space_needed_for_op(node, op) <= znode_free_space(node)); + + assert("nikita-1286", coord_is_existing_item(coord)); + + /* + * if item is expanded as a result of this operation, we should first + * change item size, than call ->b.paste item method. If item is + * shrunk, it should be done other way around: first call ->b.paste + * method, then reduce item size. + */ + + real_size = space_needed_for_op(node, op); + if (real_size > 0) + node->nplug->change_item_size(coord, real_size); + + doing->restartable = 0; + info.doing = doing; + info.todo = todo; + + result = iplug->b.paste(coord, op->u.insert.d->data, &info); + + if (real_size < 0) + node->nplug->change_item_size(coord, real_size); + + /* if we pasted at the beginning of the item, update item's key. */ + if (coord->unit_pos == 0 && coord->between != AFTER_UNIT) + node->nplug->update_item_key(coord, op->u.insert.d->key, &info); + + znode_make_dirty(node); + return result; +} + +/* handle carry COP_EXTENT operation. */ +static int carry_extent(carry_op * op /* operation to perform */ , + carry_level * doing /* queue of operations @op + * is part of */ , + carry_level * todo /* queue where new operations + * are accumulated */ ) +{ + znode *node; + carry_insert_data cdata; + coord_t coord; + reiser4_item_data data; + carry_op *delete_dummy; + carry_op *insert_extent; + int result; + carry_plugin_info info; + + assert("nikita-1751", op != NULL); + assert("nikita-1752", todo != NULL); + assert("nikita-1753", op->op == COP_EXTENT); + + /* extent insertion overview: + + extents live on the TWIG LEVEL, which is level one above the leaf + one. This complicates extent insertion logic somewhat: it may + happen (and going to happen all the time) that in logical key + ordering extent has to be placed between items I1 and I2, located + at the leaf level, but I1 and I2 are in the same formatted leaf + node N1. To insert extent one has to + + (1) reach node N1 and shift data between N1, its neighbors and + possibly newly allocated nodes until I1 and I2 fall into different + nodes. Since I1 and I2 are still neighboring items in logical key + order, they will be necessary utmost items in their respective + nodes. + + (2) After this new extent item is inserted into node on the twig + level. + + Fortunately this process can reuse almost all code from standard + insertion procedure (viz. make_space() and insert_paste_common()), + due to the following observation: make_space() only shifts data up + to and excluding or including insertion point. It never + "over-moves" through insertion point. Thus, one can use + make_space() to perform step (1). All required for this is just to + instruct free_space_shortage() to keep make_space() shifting data + until insertion point is at the node border. + + */ + + /* perform common functionality of insert and paste. */ + result = insert_paste_common(op, doing, todo, &cdata, &coord, &data); + if (result != 0) + return result; + + node = op->u.extent.d->coord->node; + assert("nikita-1754", node != NULL); + assert("nikita-1755", node_plugin_by_node(node) != NULL); + assert("nikita-1700", coord_wrt(op->u.extent.d->coord) != COORD_INSIDE); + + /* NOTE-NIKITA add some checks here. Not assertions, -EIO. Check that + extent fits between items. */ + + info.doing = doing; + info.todo = todo; + + /* there is another complication due to placement of extents on the + twig level: extents are "rigid" in the sense that key-range + occupied by extent cannot grow indefinitely to the right as it is + for the formatted leaf nodes. Because of this when search finds two + adjacent extents on the twig level, it has to "drill" to the leaf + level, creating new node. Here we are removing this node. + */ + if (node_is_empty(node)) { + delete_dummy = node_post_carry(&info, COP_DELETE, node, 1); + if (IS_ERR(delete_dummy)) + return PTR_ERR(delete_dummy); + delete_dummy->u.delete.child = NULL; + delete_dummy->u.delete.flags = DELETE_RETAIN_EMPTY; + ZF_SET(node, JNODE_HEARD_BANSHEE); + } + + /* proceed with inserting extent item into parent. We are definitely + inserting rather than pasting if we get that far. */ + insert_extent = node_post_carry(&info, COP_INSERT, node, 1); + if (IS_ERR(insert_extent)) + /* @delete_dummy will be automatically destroyed on the level + exiting */ + return PTR_ERR(insert_extent); + /* NOTE-NIKITA insertion by key is simplest option here. Another + possibility is to insert on the left or right of already existing + item. + */ + insert_extent->u.insert.type = COPT_KEY; + insert_extent->u.insert.d = op->u.extent.d; + assert("nikita-1719", op->u.extent.d->key != NULL); + insert_extent->u.insert.d->data->arg = op->u.extent.d->coord; + insert_extent->u.insert.flags = + znode_get_tree(node)->carry.new_extent_flags; + + /* + * if carry was asked to track lock handle we should actually track + * lock handle on the twig node rather than on the leaf where + * operation was started from. Transfer tracked lock handle. + */ + if (doing->track_type) { + assert("nikita-3242", doing->tracked != NULL); + assert("nikita-3244", todo->tracked == NULL); + todo->tracked = doing->tracked; + todo->track_type = CARRY_TRACK_NODE; + doing->tracked = NULL; + doing->track_type = 0; + } + + return 0; +} + +/* update key in @parent between pointers to @left and @right. + + Find coords of @left and @right and update delimiting key between them. + This is helper function called by carry_update(). Finds position of + internal item involved. Updates item key. Updates delimiting keys of child + nodes involved. +*/ +static int update_delimiting_key(znode * parent /* node key is updated + * in */ , + znode * left /* child of @parent */ , + znode * right /* child of @parent */ , + carry_level * doing /* current carry + * level */ , + carry_level * todo /* parent carry + * level */ , + const char **error_msg /* place to + * store error + * message */ ) +{ + coord_t left_pos; + coord_t right_pos; + int result; + reiser4_key ldkey; + carry_plugin_info info; + + assert("nikita-1177", right != NULL); + /* find position of right left child in a parent */ + result = find_child_ptr(parent, right, &right_pos); + if (result != NS_FOUND) { + *error_msg = "Cannot find position of right child"; + return result; + } + + if ((left != NULL) && !coord_is_leftmost_unit(&right_pos)) { + /* find position of the left child in a parent */ + result = find_child_ptr(parent, left, &left_pos); + if (result != NS_FOUND) { + *error_msg = "Cannot find position of left child"; + return result; + } + assert("nikita-1355", left_pos.node != NULL); + } else + left_pos.node = NULL; + + /* check that they are separated by exactly one key and are basically + sane */ + if (REISER4_DEBUG) { + if ((left_pos.node != NULL) + && !coord_is_existing_unit(&left_pos)) { + *error_msg = "Left child is bastard"; + return RETERR(-EIO); + } + if (!coord_is_existing_unit(&right_pos)) { + *error_msg = "Right child is bastard"; + return RETERR(-EIO); + } + if (left_pos.node != NULL && + !coord_are_neighbors(&left_pos, &right_pos)) { + *error_msg = "Children are not direct siblings"; + return RETERR(-EIO); + } + } + *error_msg = NULL; + + info.doing = doing; + info.todo = todo; + + /* + * If child node is not empty, new key of internal item is a key of + * leftmost item in the child node. If the child is empty, take its + * right delimiting key as a new key of the internal item. Precise key + * in the latter case is not important per se, because the child (and + * the internal item) are going to be killed shortly anyway, but we + * have to preserve correct order of keys in the parent node. + */ + + if (!ZF_ISSET(right, JNODE_HEARD_BANSHEE)) + leftmost_key_in_node(right, &ldkey); + else { + read_lock_dk(znode_get_tree(parent)); + ldkey = *znode_get_rd_key(right); + read_unlock_dk(znode_get_tree(parent)); + } + node_plugin_by_node(parent)->update_item_key(&right_pos, &ldkey, &info); + doing->restartable = 0; + znode_make_dirty(parent); + return 0; +} + +/* implements COP_UPDATE opration + + Update delimiting keys. + +*/ +static int carry_update(carry_op * op /* operation to be performed */ , + carry_level * doing /* current carry level */ , + carry_level * todo /* next carry level */ ) +{ + int result; + carry_node *missing UNUSED_ARG; + znode *left; + znode *right; + carry_node *lchild; + carry_node *rchild; + const char *error_msg; + reiser4_tree *tree; + + /* + * This operation is called to update key of internal item. This is + * necessary when carry shifted of cut data on the child + * level. Arguments of this operation are: + * + * @right --- child node. Operation should update key of internal + * item pointing to @right. + * + * @left --- left neighbor of @right. This parameter is optional. + */ + + assert("nikita-902", op != NULL); + assert("nikita-903", todo != NULL); + assert("nikita-904", op->op == COP_UPDATE); + + lchild = op->u.update.left; + rchild = op->node; + + if (lchild != NULL) { + assert("nikita-1001", lchild->parent); + assert("nikita-1003", !lchild->left); + left = reiser4_carry_real(lchild); + } else + left = NULL; + + tree = znode_get_tree(rchild->node); + read_lock_tree(tree); + right = znode_parent(rchild->node); + read_unlock_tree(tree); + + if (right != NULL) { + result = update_delimiting_key(right, + lchild ? lchild->node : NULL, + rchild->node, + doing, todo, &error_msg); + } else { + error_msg = "Cannot find node to update key in"; + result = RETERR(-EIO); + } + /* operation will be reposted to the next level by the + ->update_item_key() method of node plugin, if necessary. */ + + if (result != 0) { + warning("nikita-999", "Error updating delimiting key: %s (%i)", + error_msg ? : "", result); + } + return result; +} + +/* move items from @node during carry */ +static int carry_shift_data(sideof side /* in what direction to move data */ , + coord_t * insert_coord /* coord where new item + * is to be inserted */ , + znode * node /* node which data are moved from */ , + carry_level * doing /* active carry queue */ , + carry_level * todo /* carry queue where new + * operations are to be put + * in */ , + unsigned int including_insert_coord_p /* true if + * @insertion_coord + * can be moved */ ) +{ + int result; + znode *source; + carry_plugin_info info; + node_plugin *nplug; + + source = insert_coord->node; + + info.doing = doing; + info.todo = todo; + + nplug = node_plugin_by_node(node); + result = nplug->shift(insert_coord, node, + (side == LEFT_SIDE) ? SHIFT_LEFT : SHIFT_RIGHT, 0, + (int)including_insert_coord_p, &info); + /* the only error ->shift() method of node plugin can return is + -ENOMEM due to carry node/operation allocation. */ + assert("nikita-915", result >= 0 || result == -ENOMEM); + if (result > 0) { + /* + * if some number of bytes was actually shifted, mark nodes + * dirty, and carry level as non-restartable. + */ + doing->restartable = 0; + znode_make_dirty(source); + znode_make_dirty(node); + } + + assert("nikita-2077", coord_check(insert_coord)); + return 0; +} + +typedef carry_node *(*carry_iterator) (carry_node * node); +static carry_node *find_dir_carry(carry_node * node, carry_level * level, + carry_iterator iterator); + +static carry_node *pool_level_list_prev(carry_node *node) +{ + return list_entry(node->header.level_linkage.prev, carry_node, header.level_linkage); +} + +/* look for the left neighbor of given carry node in a carry queue. + + This is used by find_left_neighbor(), but I am not sure that this + really gives any advantage. More statistics required. + +*/ +carry_node *find_left_carry(carry_node * node /* node to find left neighbor + * of */ , + carry_level * level /* level to scan */ ) +{ + return find_dir_carry(node, level, + (carry_iterator) pool_level_list_prev); +} + +static carry_node *pool_level_list_next(carry_node *node) +{ + return list_entry(node->header.level_linkage.next, carry_node, header.level_linkage); +} + +/* look for the right neighbor of given carry node in a + carry queue. + + This is used by find_right_neighbor(), but I am not sure that this + really gives any advantage. More statistics required. + +*/ +carry_node *find_right_carry(carry_node * node /* node to find right neighbor + * of */ , + carry_level * level /* level to scan */ ) +{ + return find_dir_carry(node, level, + (carry_iterator) pool_level_list_next); +} + +/* look for the left or right neighbor of given carry node in a carry + queue. + + Helper function used by find_{left|right}_carry(). +*/ +static carry_node *find_dir_carry(carry_node * node /* node to start scanning + * from */ , + carry_level * level /* level to scan */ , + carry_iterator iterator /* operation to + * move to the next + * node */ ) +{ + carry_node *neighbor; + + assert("nikita-1059", node != NULL); + assert("nikita-1060", level != NULL); + + /* scan list of carry nodes on this list dir-ward, skipping all + carry nodes referencing the same znode. */ + neighbor = node; + while (1) { + neighbor = iterator(neighbor); + if (carry_node_end(level, neighbor)) + /* list head is reached */ + return NULL; + if (reiser4_carry_real(neighbor) != reiser4_carry_real(node)) + return neighbor; + } +} + +/* + * Memory reservation estimation. + * + * Carry process proceeds through tree levels upwards. Carry assumes that it + * takes tree in consistent state (e.g., that search tree invariants hold), + * and leaves tree consistent after it finishes. This means that when some + * error occurs carry cannot simply return if there are pending carry + * operations. Generic solution for this problem is carry-undo either as + * transaction manager feature (requiring checkpoints and isolation), or + * through some carry specific mechanism. + * + * Our current approach is to panic if carry hits an error while tree is + * inconsistent. Unfortunately -ENOMEM can easily be triggered. To work around + * this "memory reservation" mechanism was added. + * + * Memory reservation is implemented by perthread-pages.diff patch from + * core-patches. Its API is defined in + * + * int perthread_pages_reserve(int nrpages, gfp_t gfp); + * void perthread_pages_release(int nrpages); + * int perthread_pages_count(void); + * + * carry estimates its worst case memory requirements at the entry, reserved + * enough memory, and released unused pages before returning. + * + * Code below estimates worst case memory requirements for a given carry + * queue. This is dome by summing worst case memory requirements for each + * operation in the queue. + * + */ + +/* + * Memory memory requirements of many operations depends on the tree + * height. For example, item insertion requires new node to be inserted at + * each tree level in the worst case. What tree height should be used for + * estimation? Current tree height is wrong, because tree height can change + * between the time when estimation was done and the time when operation is + * actually performed. Maximal possible tree height (REISER4_MAX_ZTREE_HEIGHT) + * is also not desirable, because it would lead to the huge over-estimation + * all the time. Plausible solution is "capped tree height": if current tree + * height is less than some TREE_HEIGHT_CAP constant, capped tree height is + * TREE_HEIGHT_CAP, otherwise it's current tree height. Idea behind this is + * that if tree height is TREE_HEIGHT_CAP or larger, it's extremely unlikely + * to be increased even more during short interval of time. + */ +#define TREE_HEIGHT_CAP (5) + +/* return capped tree height for the @tree. See comment above. */ +static int cap_tree_height(reiser4_tree * tree) +{ + return max_t(int, tree->height, TREE_HEIGHT_CAP); +} + +/* return capped tree height for the current tree. */ +static int capped_height(void) +{ + return cap_tree_height(current_tree); +} + +/* return number of pages required to store given number of bytes */ +static int bytes_to_pages(int bytes) +{ + return (bytes + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT; +} + +/* how many pages are required to allocate znodes during item insertion. */ +static int carry_estimate_znodes(void) +{ + /* + * Note, that there we have some problem here: there is no way to + * reserve pages specifically for the given slab. This means that + * these pages can be hijacked for some other end. + */ + + /* in the worst case we need 3 new znode on each tree level */ + return bytes_to_pages(capped_height() * sizeof(znode) * 3); +} + +/* + * how many pages are required to load bitmaps. One bitmap per level. + */ +static int carry_estimate_bitmaps(void) +{ + if (reiser4_is_set(reiser4_get_current_sb(), REISER4_DONT_LOAD_BITMAP)) { + int bytes; + + bytes = capped_height() * (0 + /* bnode should be added, but its is private to + * bitmap.c, skip for now. */ + 2 * sizeof(jnode)); /* working and commit jnodes */ + return bytes_to_pages(bytes) + 2; /* and their contents */ + } else + /* bitmaps were pre-loaded during mount */ + return 0; +} + +/* worst case item insertion memory requirements */ +static int carry_estimate_insert(carry_op * op, carry_level * level) +{ + return carry_estimate_bitmaps() + carry_estimate_znodes() + 1 + /* new atom */ + capped_height() + /* new block on each level */ + 1 + /* and possibly extra new block at the leaf level */ + 3; /* loading of leaves into memory */ +} + +/* worst case item deletion memory requirements */ +static int carry_estimate_delete(carry_op * op, carry_level * level) +{ + return carry_estimate_bitmaps() + carry_estimate_znodes() + 1 + /* new atom */ + 3; /* loading of leaves into memory */ +} + +/* worst case tree cut memory requirements */ +static int carry_estimate_cut(carry_op * op, carry_level * level) +{ + return carry_estimate_bitmaps() + carry_estimate_znodes() + 1 + /* new atom */ + 3; /* loading of leaves into memory */ +} + +/* worst case memory requirements of pasting into item */ +static int carry_estimate_paste(carry_op * op, carry_level * level) +{ + return carry_estimate_bitmaps() + carry_estimate_znodes() + 1 + /* new atom */ + capped_height() + /* new block on each level */ + 1 + /* and possibly extra new block at the leaf level */ + 3; /* loading of leaves into memory */ +} + +/* worst case memory requirements of extent insertion */ +static int carry_estimate_extent(carry_op * op, carry_level * level) +{ + return carry_estimate_insert(op, level) + /* insert extent */ + carry_estimate_delete(op, level); /* kill leaf */ +} + +/* worst case memory requirements of key update */ +static int carry_estimate_update(carry_op * op, carry_level * level) +{ + return 0; +} + +/* worst case memory requirements of flow insertion */ +static int carry_estimate_insert_flow(carry_op * op, carry_level * level) +{ + int newnodes; + + newnodes = min(bytes_to_pages(op->u.insert_flow.flow->length), + CARRY_FLOW_NEW_NODES_LIMIT); + /* + * roughly estimate insert_flow as a sequence of insertions. + */ + return newnodes * carry_estimate_insert(op, level); +} + +/* This is dispatch table for carry operations. It can be trivially + abstracted into useful plugin: tunable balancing policy is a good + thing. */ +carry_op_handler op_dispatch_table[COP_LAST_OP] = { + [COP_INSERT] = { + .handler = carry_insert, + .estimate = carry_estimate_insert} + , + [COP_DELETE] = { + .handler = carry_delete, + .estimate = carry_estimate_delete} + , + [COP_CUT] = { + .handler = carry_cut, + .estimate = carry_estimate_cut} + , + [COP_PASTE] = { + .handler = carry_paste, + .estimate = carry_estimate_paste} + , + [COP_EXTENT] = { + .handler = carry_extent, + .estimate = carry_estimate_extent} + , + [COP_UPDATE] = { + .handler = carry_update, + .estimate = carry_estimate_update} + , + [COP_INSERT_FLOW] = { + .handler = carry_insert_flow, + .estimate = carry_estimate_insert_flow} +}; + +/* Make Linus happy. + Local variables: + c-indentation-style: "K&R" + mode-name: "LC" + c-basic-offset: 8 + tab-width: 8 + fill-column: 120 + scroll-step: 1 + End: +*/ diff -urN linux-2.6.22.orig/fs/reiser4/carry_ops.h linux-2.6.22/fs/reiser4/carry_ops.h --- linux-2.6.22.orig/fs/reiser4/carry_ops.h 1970-01-01 03:00:00.000000000 +0300 +++ linux-2.6.22/fs/reiser4/carry_ops.h 2007-07-29 00:25:34.828684053 +0400 @@ -0,0 +1,42 @@ +/* Copyright 2001, 2002, 2003 by Hans Reiser, licensing governed by reiser4/README */ + +/* implementation of carry operations. See carry_ops.c for details. */ + +#if !defined( __CARRY_OPS_H__ ) +#define __CARRY_OPS_H__ + +#include "forward.h" +#include "znode.h" +#include "carry.h" + +/* carry operation handlers */ +typedef struct carry_op_handler { + /* perform operation */ + int (*handler) (carry_op * op, carry_level * doing, carry_level * todo); + /* estimate memory requirements for @op */ + int (*estimate) (carry_op * op, carry_level * level); +} carry_op_handler; + +/* This is dispatch table for carry operations. It can be trivially + abstracted into useful plugin: tunable balancing policy is a good + thing. */ +extern carry_op_handler op_dispatch_table[COP_LAST_OP]; + +unsigned int space_needed(const znode * node, const coord_t * coord, + const reiser4_item_data * data, int inserting); +extern carry_node *find_left_carry(carry_node * node, carry_level * level); +extern carry_node *find_right_carry(carry_node * node, carry_level * level); + +/* __CARRY_OPS_H__ */ +#endif + +/* Make Linus happy. + Local variables: + c-indentation-style: "K&R" + mode-name: "LC" + c-basic-offset: 8 + tab-width: 8 + fill-column: 120 + scroll-step: 1 + End: +*/ diff -urN linux-2.6.22.orig/fs/reiser4/context.c linux-2.6.22/fs/reiser4/context.c --- linux-2.6.22.orig/fs/reiser4/context.c 1970-01-01 03:00:00.000000000 +0300 +++ linux-2.6.22/fs/reiser4/context.c 2007-07-29 00:25:34.832685088 +0400 @@ -0,0 +1,288 @@ +/* Copyright 2002, 2003 by Hans Reiser, licensing governed by reiser4/README */ + +/* Manipulation of reiser4_context */ + +/* + * global context used during system call. Variable of this type is allocated + * on the stack at the beginning of the reiser4 part of the system call and + * pointer to it is stored in the current->fs_context. This allows us to avoid + * passing pointer to current transaction and current lockstack (both in + * one-to-one mapping with threads) all over the call chain. + * + * It's kind of like those global variables the prof used to tell you not to + * use in CS1, except thread specific.;-) Nikita, this was a good idea. + * + * In some situations it is desirable to have ability to enter reiser4_context + * more than once for the same thread (nested contexts). For example, there + * are some functions that can be called either directly from VFS/VM or from + * already active reiser4 context (->writepage, for example). + * + * In such situations "child" context acts like dummy: all activity is + * actually performed in the top level context, and get_current_context() + * always returns top level context. + * Of course, reiser4_init_context()/reiser4_done_context() have to be properly + * nested any way. + * + * Note that there is an important difference between reiser4 uses + * ->fs_context and the way other file systems use it. Other file systems + * (ext3 and reiserfs) use ->fs_context only for the duration of _transaction_ + * (this is why ->fs_context was initially called ->journal_info). This means, + * that when ext3 or reiserfs finds that ->fs_context is not NULL on the entry + * to the file system, they assume that some transaction is already underway, + * and usually bail out, because starting nested transaction would most likely + * lead to the deadlock. This gives false positives with reiser4, because we + * set ->fs_context before starting transaction. + */ + +#include "debug.h" +#include "super.h" +#include "context.h" + +#include /* balance_dirty_pages() */ +#include + +static void _reiser4_init_context(reiser4_context * context, + struct super_block *super) +{ + memset(context, 0, sizeof(*context)); + + context->super = super; + context->magic = context_magic; + context->outer = current->journal_info; + current->journal_info = (void *)context; + context->nr_children = 0; + context->gfp_mask = GFP_KERNEL; + + init_lock_stack(&context->stack); + + reiser4_txn_begin(context); + + /* initialize head of tap list */ + INIT_LIST_HEAD(&context->taps); +#if REISER4_DEBUG + context->task = current; +#endif + grab_space_enable(); +} + +/* initialize context and bind it to the current thread + + This function should be called at the beginning of reiser4 part of + syscall. +*/ +reiser4_context * reiser4_init_context(struct super_block * super) +{ + reiser4_context *context; + + assert("nikita-2662", !in_interrupt() && !in_irq()); + assert("nikita-3357", super != NULL); + assert("nikita-3358", super->s_op == NULL || is_reiser4_super(super)); + + context = get_current_context_check(); + if (context && context->super == super) { + context = (reiser4_context *) current->journal_info; + context->nr_children++; + return context; + } + + context = kmalloc(sizeof(*context), GFP_KERNEL); + if (context == NULL) + return ERR_PTR(RETERR(-ENOMEM)); + + _reiser4_init_context(context, super); + return context; +} + +/* this is used in scan_mgr which is called with spinlock held and in + reiser4_fill_super magic */ +void init_stack_context(reiser4_context *context, struct super_block *super) +{ + assert("nikita-2662", !in_interrupt() && !in_irq()); + assert("nikita-3357", super != NULL); + assert("nikita-3358", super->s_op == NULL || is_reiser4_super(super)); + assert("vs-12", !is_in_reiser4_context()); + + _reiser4_init_context(context, super); + context->on_stack = 1; + return; +} + +/* cast lock stack embedded into reiser4 context up to its container */ +reiser4_context *get_context_by_lock_stack(lock_stack * owner) +{ + return container_of(owner, reiser4_context, stack); +} + +/* true if there is already _any_ reiser4 context for the current thread */ +int is_in_reiser4_context(void) +{ + reiser4_context *ctx; + + ctx = current->journal_info; + return ctx != NULL && ((unsigned long)ctx->magic) == context_magic; +} + +/* + * call balance dirty pages for the current context. + * + * File system is expected to call balance_dirty_pages_ratelimited() whenever + * it dirties a page. reiser4 does this for unformatted nodes (that is, during + * write---this covers vast majority of all dirty traffic), but we cannot do + * this immediately when formatted node is dirtied, because long term lock is + * usually held at that time. To work around this, dirtying of formatted node + * simply increases ->nr_marked_dirty counter in the current reiser4 + * context. When we are about to leave this context, + * balance_dirty_pages_ratelimited() is called, if necessary. + * + * This introduces another problem: sometimes we do not want to run + * balance_dirty_pages_ratelimited() when leaving a context, for example + * because some important lock (like ->i_mutex on the parent directory) is + * held. To achieve this, ->nobalance flag can be set in the current context. + */ +static void balance_dirty_pages_at(reiser4_context *context) +{ + reiser4_super_info_data *sbinfo = get_super_private(context->super); + + /* + * call balance_dirty_pages_ratelimited() to process formatted nodes + * dirtied during this system call. Do that only if we are not in mount + * and there were nodes dirtied in this context and we are not in + * writepage (to avoid deadlock) and not in pdflush + */ + if (sbinfo != NULL && sbinfo->fake != NULL && + context->nr_marked_dirty != 0 && + !(current->flags & PF_MEMALLOC) && + !current_is_pdflush()) + balance_dirty_pages_ratelimited(sbinfo->fake->i_mapping); +} + +/* release resources associated with context. + + This function should be called at the end of "session" with reiser4, + typically just before leaving reiser4 driver back to VFS. + + This is good place to put some degugging consistency checks, like that + thread released all locks and closed transcrash etc. + +*/ +static void reiser4_done_context(reiser4_context * context /* context being released */ ) +{ + assert("nikita-860", context != NULL); + assert("nikita-859", context->magic == context_magic); + assert("vs-646", (reiser4_context *) current->journal_info == context); + assert("zam-686", !in_interrupt() && !in_irq()); + + /* only do anything when leaving top-level reiser4 context. All nested + * contexts are just dummies. */ + if (context->nr_children == 0) { + assert("jmacd-673", context->trans == NULL); + assert("jmacd-1002", lock_stack_isclean(&context->stack)); + assert("nikita-1936", reiser4_no_counters_are_held()); + assert("nikita-2626", list_empty_careful(reiser4_taps_list())); + assert("zam-1004", ergo(get_super_private(context->super), + get_super_private(context->super)->delete_mutex_owner != + current)); + + /* release all grabbed but as yet unused blocks */ + if (context->grabbed_blocks != 0) + all_grabbed2free(); + + /* + * synchronize against longterm_unlock_znode(): + * wake_up_requestor() wakes up requestors without holding + * zlock (otherwise they will immediately bump into that lock + * after wake up on another CPU). To work around (rare) + * situation where requestor has been woken up asynchronously + * and managed to run until completion (and destroy its + * context and lock stack) before wake_up_requestor() called + * wake_up() on it, wake_up_requestor() synchronize on lock + * stack spin lock. It has actually been observed that spin + * lock _was_ locked at this point, because + * wake_up_requestor() took interrupt. + */ + spin_lock_stack(&context->stack); + spin_unlock_stack(&context->stack); + + assert("zam-684", context->nr_children == 0); + /* restore original ->fs_context value */ + current->journal_info = context->outer; + if (context->on_stack == 0) + kfree(context); + } else { + context->nr_children--; +#if REISER4_DEBUG + assert("zam-685", context->nr_children >= 0); +#endif + } +} + +/* + * exit reiser4 context. Call balance_dirty_pages_at() if necessary. Close + * transaction. Call done_context() to do context related book-keeping. + */ +void reiser4_exit_context(reiser4_context * context) +{ + assert("nikita-3021", reiser4_schedulable()); + + if (context->nr_children == 0) { + if (!context->nobalance) { + reiser4_txn_restart(context); + balance_dirty_pages_at(context); + } + + /* if filesystem is mounted with -o sync or -o dirsync - commit + transaction. FIXME: TXNH_DONT_COMMIT is used to avoid + commiting on exit_context when inode semaphore is held and + to have ktxnmgrd to do commit instead to get better + concurrent filesystem accesses. But, when one mounts with -o + sync, he cares more about reliability than about + performance. So, for now we have this simple mount -o sync + support. */ + if (context->super->s_flags & (MS_SYNCHRONOUS | MS_DIRSYNC)) { + txn_atom *atom; + + atom = get_current_atom_locked_nocheck(); + if (atom) { + atom->flags |= ATOM_FORCE_COMMIT; + context->trans->flags &= ~TXNH_DONT_COMMIT; + spin_unlock_atom(atom); + } + } + reiser4_txn_end(context); + } + reiser4_done_context(context); +} + +void reiser4_ctx_gfp_mask_set(void) +{ + reiser4_context *ctx; + + ctx = get_current_context(); + if (ctx->entd == 0 && + list_empty(&ctx->stack.locks) && + ctx->trans->atom == NULL) + ctx->gfp_mask = GFP_KERNEL; + else + ctx->gfp_mask = GFP_NOFS; +} + +void reiser4_ctx_gfp_mask_force (gfp_t mask) +{ + reiser4_context *ctx; + ctx = get_current_context(); + + assert("edward-1454", ctx != NULL); + + ctx->gfp_mask = mask; +} + +/* + * Local variables: + * c-indentation-style: "K&R" + * mode-name: "LC" + * c-basic-offset: 8 + * tab-width: 8 + * fill-column: 120 + * scroll-step: 1 + * End: + */ diff -urN linux-2.6.22.orig/fs/reiser4/context.h linux-2.6.22/fs/reiser4/context.h --- linux-2.6.22.orig/fs/reiser4/context.h 1970-01-01 03:00:00.000000000 +0300 +++ linux-2.6.22/fs/reiser4/context.h 2007-07-29 00:25:34.832685088 +0400 @@ -0,0 +1,228 @@ +/* Copyright 2001, 2002, 2003, 2004 by Hans Reiser, licensing governed by + * reiser4/README */ + +/* Reiser4 context. See context.c for details. */ + +#if !defined( __REISER4_CONTEXT_H__ ) +#define __REISER4_CONTEXT_H__ + +#include "forward.h" +#include "debug.h" +#include "dformat.h" +#include "tap.h" +#include "lock.h" + +#include /* for __u?? */ +#include /* for struct super_block */ +#include +#include /* for struct task_struct */ + +/* reiser4 per-thread context */ +struct reiser4_context { + /* magic constant. For identification of reiser4 contexts. */ + __u32 magic; + + /* current lock stack. See lock.[ch]. This is where list of all + locks taken by current thread is kept. This is also used in + deadlock detection. */ + lock_stack stack; + + /* current transcrash. */ + txn_handle *trans; + /* transaction handle embedded into reiser4_context. ->trans points + * here by default. */ + txn_handle trans_in_ctx; + + /* super block we are working with. To get the current tree + use &get_super_private (reiser4_get_current_sb ())->tree. */ + struct super_block *super; + + /* parent fs activation */ + struct fs_activation *outer; + + /* per-thread grabbed (for further allocation) blocks counter */ + reiser4_block_nr grabbed_blocks; + + /* list of taps currently monitored. See tap.c */ + struct list_head taps; + + /* grabbing space is enabled */ + unsigned int grab_enabled:1; + /* should be set when we are write dirty nodes to disk in jnode_flush or + * reiser4_write_logs() */ + unsigned int writeout_mode:1; + /* true, if current thread is an ent thread */ + unsigned int entd:1; + /* true, if balance_dirty_pages() should not be run when leaving this + * context. This is used to avoid lengthly balance_dirty_pages() + * operation when holding some important resource, like directory + * ->i_mutex */ + unsigned int nobalance:1; + + /* this bit is used on reiser4_done_context to decide whether context is + kmalloc-ed and has to be kfree-ed */ + unsigned int on_stack:1; + + /* count non-trivial jnode_set_dirty() calls */ + unsigned long nr_marked_dirty; + + /* reiser4_sync_inodes calls (via generic_sync_sb_inodes) + * reiser4_writepages for each of dirty inodes. Reiser4_writepages + * captures pages. When number of pages captured in one + * reiser4_sync_inodes reaches some threshold - some atoms get + * flushed */ + int nr_captured; + int nr_children; /* number of child contexts */ +#if REISER4_DEBUG + /* debugging information about reiser4 locks held by the current + * thread */ + reiser4_lock_cnt_info locks; + struct task_struct *task; /* so we can easily find owner of the stack */ + + /* + * disk space grabbing debugging support + */ + /* how many disk blocks were grabbed by the first call to + * reiser4_grab_space() in this context */ + reiser4_block_nr grabbed_initially; + + /* list of all threads doing flush currently */ + struct list_head flushers_link; + /* information about last error encountered by reiser4 */ + err_site err; +#endif + void *vp; + gfp_t gfp_mask; +}; + +extern reiser4_context *get_context_by_lock_stack(lock_stack *); + +/* Debugging helps. */ +#if REISER4_DEBUG +extern void print_contexts(void); +#endif + +#define current_tree (&(get_super_private(reiser4_get_current_sb())->tree)) +#define current_blocksize reiser4_get_current_sb()->s_blocksize +#define current_blocksize_bits reiser4_get_current_sb()->s_blocksize_bits + +extern reiser4_context *reiser4_init_context(struct super_block *); +extern void init_stack_context(reiser4_context *, struct super_block *); +extern void reiser4_exit_context(reiser4_context *); + +/* magic constant we store in reiser4_context allocated at the stack. Used to + catch accesses to staled or uninitialized contexts. */ +#define context_magic ((__u32) 0x4b1b5d0b) + +extern int is_in_reiser4_context(void); + +/* + * return reiser4_context for the thread @tsk + */ +static inline reiser4_context *get_context(const struct task_struct *tsk) +{ + assert("vs-1682", + ((reiser4_context *) tsk->journal_info)->magic == context_magic); + return (reiser4_context *) tsk->journal_info; +} + +/* + * return reiser4 context of the current thread, or NULL if there is none. + */ +static inline reiser4_context *get_current_context_check(void) +{ + if (is_in_reiser4_context()) + return get_context(current); + else + return NULL; +} + +static inline reiser4_context *get_current_context(void); /* __attribute__((const)); */ + +/* return context associated with current thread */ +static inline reiser4_context *get_current_context(void) +{ + return get_context(current); +} + +static inline gfp_t reiser4_ctx_gfp_mask_get(void) +{ + reiser4_context *ctx; + + ctx = get_current_context_check(); + return (ctx == NULL) ? GFP_KERNEL : ctx->gfp_mask; +} + +void reiser4_ctx_gfp_mask_set(void); +void reiser4_ctx_gfp_mask_force (gfp_t mask); + +/* + * true if current thread is in the write-out mode. Thread enters write-out + * mode during jnode_flush and reiser4_write_logs(). + */ +static inline int is_writeout_mode(void) +{ + return get_current_context()->writeout_mode; +} + +/* + * enter write-out mode + */ +static inline void writeout_mode_enable(void) +{ + assert("zam-941", !get_current_context()->writeout_mode); + get_current_context()->writeout_mode = 1; +} + +/* + * leave write-out mode + */ +static inline void writeout_mode_disable(void) +{ + assert("zam-942", get_current_context()->writeout_mode); + get_current_context()->writeout_mode = 0; +} + +static inline void grab_space_enable(void) +{ + get_current_context()->grab_enabled = 1; +} + +static inline void grab_space_disable(void) +{ + get_current_context()->grab_enabled = 0; +} + +static inline void grab_space_set_enabled(int enabled) +{ + get_current_context()->grab_enabled = enabled; +} + +static inline int is_grab_enabled(reiser4_context * ctx) +{ + return ctx->grab_enabled; +} + +/* mark transaction handle in @ctx as TXNH_DONT_COMMIT, so that no commit or + * flush would be performed when it is closed. This is necessary when handle + * has to be closed under some coarse semaphore, like i_mutex of + * directory. Commit will be performed by ktxnmgrd. */ +static inline void context_set_commit_async(reiser4_context * context) +{ + context->nobalance = 1; + context->trans->flags |= TXNH_DONT_COMMIT; +} + +/* __REISER4_CONTEXT_H__ */ +#endif + +/* Make Linus happy. + Local variables: + c-indentation-style: "K&R" + mode-name: "LC" + c-basic-offset: 8 + tab-width: 8 + fill-column: 120 + scroll-step: 1 + End: +*/ diff -urN linux-2.6.22.orig/fs/reiser4/coord.c linux-2.6.22/fs/reiser4/coord.c --- linux-2.6.22.orig/fs/reiser4/coord.c 1970-01-01 03:00:00.000000000 +0300 +++ linux-2.6.22/fs/reiser4/coord.c 2007-07-29 00:25:34.832685088 +0400 @@ -0,0 +1,935 @@ +/* Copyright 2001, 2002, 2003 by Hans Reiser, licensing governed by reiser4/README */ + +#include "forward.h" +#include "debug.h" +#include "dformat.h" +#include "tree.h" +#include "plugin/item/item.h" +#include "znode.h" +#include "coord.h" + +/* Internal constructor. */ +static inline void +coord_init_values(coord_t * coord, const znode * node, pos_in_node_t item_pos, + pos_in_node_t unit_pos, between_enum between) +{ + coord->node = (znode *) node; + coord_set_item_pos(coord, item_pos); + coord->unit_pos = unit_pos; + coord->between = between; + ON_DEBUG(coord->plug_v = 0); + ON_DEBUG(coord->body_v = 0); + + /*ON_TRACE (TRACE_COORDS, "init coord %p node %p: %u %u %s\n", coord, node, item_pos, unit_pos, coord_tween_tostring (between)); */ +} + +/* after shifting of node content, coord previously set properly may become + invalid, try to "normalize" it. */ +void coord_normalize(coord_t * coord) +{ + znode *node; + + node = coord->node; + assert("vs-683", node); + + coord_clear_iplug(coord); + + if (node_is_empty(node)) { + coord_init_first_unit(coord, node); + } else if ((coord->between == AFTER_ITEM) + || (coord->between == AFTER_UNIT)) { + return; + } else if (coord->item_pos == coord_num_items(coord) + && coord->between == BEFORE_ITEM) { + coord_dec_item_pos(coord); + coord->between = AFTER_ITEM; + } else if (coord->unit_pos == coord_num_units(coord) + && coord->between == BEFORE_UNIT) { + coord->unit_pos--; + coord->between = AFTER_UNIT; + } else if (coord->item_pos == coord_num_items(coord) + && coord->unit_pos == 0 && coord->between == BEFORE_UNIT) { + coord_dec_item_pos(coord); + coord->unit_pos = 0; + coord->between = AFTER_ITEM; + } +} + +/* Copy a coordinate. */ +void coord_dup(coord_t * coord, const coord_t * old_coord) +{ + assert("jmacd-9800", coord_check(old_coord)); + coord_dup_nocheck(coord, old_coord); +} + +/* Copy a coordinate without check. Useful when old_coord->node is not + loaded. As in cbk_tree_lookup -> connect_znode -> connect_one_side */ +void coord_dup_nocheck(coord_t * coord, const coord_t * old_coord) +{ + coord->node = old_coord->node; + coord_set_item_pos(coord, old_coord->item_pos); + coord->unit_pos = old_coord->unit_pos; + coord->between = old_coord->between; + coord->iplugid = old_coord->iplugid; + ON_DEBUG(coord->plug_v = old_coord->plug_v); + ON_DEBUG(coord->body_v = old_coord->body_v); +} + +/* Initialize an invalid coordinate. */ +void coord_init_invalid(coord_t * coord, const znode * node) +{ + coord_init_values(coord, node, 0, 0, INVALID_COORD); +} + +void coord_init_first_unit_nocheck(coord_t * coord, const znode * node) +{ + coord_init_values(coord, node, 0, 0, AT_UNIT); +} + +/* Initialize a coordinate to point at the first unit of the first item. If the node is + empty, it is positioned at the EMPTY_NODE. */ +void coord_init_first_unit(coord_t * coord, const znode * node) +{ + int is_empty = node_is_empty(node); + + coord_init_values(coord, node, 0, 0, (is_empty ? EMPTY_NODE : AT_UNIT)); + + assert("jmacd-9801", coord_check(coord)); +} + +/* Initialize a coordinate to point at the last unit of the last item. If the node is + empty, it is positioned at the EMPTY_NODE. */ +void coord_init_last_unit(coord_t * coord, const znode * node) +{ + int is_empty = node_is_empty(node); + + coord_init_values(coord, node, + (is_empty ? 0 : node_num_items(node) - 1), 0, + (is_empty ? EMPTY_NODE : AT_UNIT)); + if (!is_empty) + coord->unit_pos = coord_last_unit_pos(coord); + assert("jmacd-9802", coord_check(coord)); +} + +/* Initialize a coordinate to before the first item. If the node is empty, it is + positioned at the EMPTY_NODE. */ +void coord_init_before_first_item(coord_t * coord, const znode * node) +{ + int is_empty = node_is_empty(node); + + coord_init_values(coord, node, 0, 0, + (is_empty ? EMPTY_NODE : BEFORE_UNIT)); + + assert("jmacd-9803", coord_check(coord)); +} + +/* Initialize a coordinate to after the last item. If the node is empty, it is positioned + at the EMPTY_NODE. */ +void coord_init_after_last_item(coord_t * coord, const znode * node) +{ + int is_empty = node_is_empty(node); + + coord_init_values(coord, node, + (is_empty ? 0 : node_num_items(node) - 1), 0, + (is_empty ? EMPTY_NODE : AFTER_ITEM)); + + assert("jmacd-9804", coord_check(coord)); +} + +/* Initialize a coordinate to after last unit in the item. Coord must be set + already to existing item */ +void coord_init_after_item_end(coord_t * coord) +{ + coord->between = AFTER_UNIT; + coord->unit_pos = coord_last_unit_pos(coord); +} + +/* Initialize a coordinate to before the item. Coord must be set already to existing item */ +void coord_init_before_item(coord_t * coord) +{ + coord->unit_pos = 0; + coord->between = BEFORE_ITEM; +} + +/* Initialize a coordinate to after the item. Coord must be set already to existing item */ +void coord_init_after_item(coord_t * coord) +{ + coord->unit_pos = 0; + coord->between = AFTER_ITEM; +} + +/* Initialize a coordinate by 0s. Used in places where init_coord was used and + it was not clear how actually */ +void coord_init_zero(coord_t * coord) +{ + memset(coord, 0, sizeof(*coord)); +} + +/* Return the number of units at the present item. Asserts coord_is_existing_item(). */ +unsigned coord_num_units(const coord_t * coord) +{ + assert("jmacd-9806", coord_is_existing_item(coord)); + + return item_plugin_by_coord(coord)->b.nr_units(coord); +} + +/* Returns true if the coord was initializewd by coord_init_invalid (). */ +/* Audited by: green(2002.06.15) */ +int coord_is_invalid(const coord_t * coord) +{ + return coord->between == INVALID_COORD; +} + +/* Returns true if the coordinate is positioned at an existing item, not before or after + an item. It may be placed at, before, or after any unit within the item, whether + existing or not. */ +int coord_is_existing_item(const coord_t * coord) +{ + switch (coord->between) { + case EMPTY_NODE: + case BEFORE_ITEM: + case AFTER_ITEM: + case INVALID_COORD: + return 0; + + case BEFORE_UNIT: + case AT_UNIT: + case AFTER_UNIT: + return coord->item_pos < coord_num_items(coord); + } + + impossible("jmacd-9900", "unreachable coord: %p", coord); + return 0; +} + +/* Returns true if the coordinate is positioned at an existing unit, not before or after a + unit. */ +/* Audited by: green(2002.06.15) */ +int coord_is_existing_unit(const coord_t * coord) +{ + switch (coord->between) { + case EMPTY_NODE: + case BEFORE_UNIT: + case AFTER_UNIT: + case BEFORE_ITEM: + case AFTER_ITEM: + case INVALID_COORD: + return 0; + + case AT_UNIT: + return (coord->item_pos < coord_num_items(coord) + && coord->unit_pos < coord_num_units(coord)); + } + + impossible("jmacd-9902", "unreachable"); + return 0; +} + +/* Returns true if the coordinate is positioned at the first unit of the first item. Not + true for empty nodes nor coordinates positioned before the first item. */ +/* Audited by: green(2002.06.15) */ +int coord_is_leftmost_unit(const coord_t * coord) +{ + return (coord->between == AT_UNIT && coord->item_pos == 0 + && coord->unit_pos == 0); +} + +#if REISER4_DEBUG +/* For assertions only, checks for a valid coordinate. */ +int coord_check(const coord_t * coord) +{ + if (coord->node == NULL) { + return 0; + } + if (znode_above_root(coord->node)) + return 1; + + switch (coord->between) { + default: + case INVALID_COORD: + return 0; + case EMPTY_NODE: + if (!node_is_empty(coord->node)) { + return 0; + } + return coord->item_pos == 0 && coord->unit_pos == 0; + + case BEFORE_UNIT: + case AFTER_UNIT: + if (node_is_empty(coord->node) && (coord->item_pos == 0) + && (coord->unit_pos == 0)) + return 1; + case AT_UNIT: + break; + case AFTER_ITEM: + case BEFORE_ITEM: + /* before/after item should not set unit_pos. */ + if (coord->unit_pos != 0) { + return 0; + } + break; + } + + if (coord->item_pos >= node_num_items(coord->node)) { + return 0; + } + + /* FIXME-VS: we are going to check unit_pos. This makes no sense when + between is set either AFTER_ITEM or BEFORE_ITEM */ + if (coord->between == AFTER_ITEM || coord->between == BEFORE_ITEM) + return 1; + + if (coord_is_iplug_set(coord) && + coord->unit_pos > + item_plugin_by_coord(coord)->b.nr_units(coord) - 1) { + return 0; + } + return 1; +} +#endif + +/* Adjust coordinate boundaries based on the number of items prior to coord_next/prev. + Returns 1 if the new position is does not exist. */ +static int coord_adjust_items(coord_t * coord, unsigned items, int is_next) +{ + /* If the node is invalid, leave it. */ + if (coord->between == INVALID_COORD) { + return 1; + } + + /* If the node is empty, set it appropriately. */ + if (items == 0) { + coord->between = EMPTY_NODE; + coord_set_item_pos(coord, 0); + coord->unit_pos = 0; + return 1; + } + + /* If it was empty and it no longer is, set to BEFORE/AFTER_ITEM. */ + if (coord->between == EMPTY_NODE) { + coord->between = (is_next ? BEFORE_ITEM : AFTER_ITEM); + coord_set_item_pos(coord, 0); + coord->unit_pos = 0; + return 0; + } + + /* If the item_pos is out-of-range, set it appropriatly. */ + if (coord->item_pos >= items) { + coord->between = AFTER_ITEM; + coord_set_item_pos(coord, items - 1); + coord->unit_pos = 0; + /* If is_next, return 1 (can't go any further). */ + return is_next; + } + + return 0; +} + +/* Advances the coordinate by one unit to the right. If empty, no change. If + coord_is_rightmost_unit, advances to AFTER THE LAST ITEM. Returns 0 if new position is an + existing unit. */ +int coord_next_unit(coord_t * coord) +{ + unsigned items = coord_num_items(coord); + + if (coord_adjust_items(coord, items, 1) == 1) { + return 1; + } + + switch (coord->between) { + case BEFORE_UNIT: + /* Now it is positioned at the same unit. */ + coord->between = AT_UNIT; + return 0; + + case AFTER_UNIT: + case AT_UNIT: + /* If it was at or after a unit and there are more units in this item, + advance to the next one. */ + if (coord->unit_pos < coord_last_unit_pos(coord)) { + coord->unit_pos += 1; + coord->between = AT_UNIT; + return 0; + } + + /* Otherwise, it is crossing an item boundary and treated as if it was + after the current item. */ + coord->between = AFTER_ITEM; + coord->unit_pos = 0; + /* FALLTHROUGH */ + + case AFTER_ITEM: + /* Check for end-of-node. */ + if (coord->item_pos == items - 1) { + return 1; + } + + coord_inc_item_pos(coord); + coord->unit_pos = 0; + coord->between = AT_UNIT; + return 0; + + case BEFORE_ITEM: + /* The adjust_items checks ensure that we are valid here. */ + coord->unit_pos = 0; + coord->between = AT_UNIT; + return 0; + + case INVALID_COORD: + case EMPTY_NODE: + /* Handled in coord_adjust_items(). */ + break; + } + + impossible("jmacd-9902", "unreachable"); + return 0; +} + +/* Advances the coordinate by one item to the right. If empty, no change. If + coord_is_rightmost_unit, advances to AFTER THE LAST ITEM. Returns 0 if new position is + an existing item. */ +int coord_next_item(coord_t * coord) +{ + unsigned items = coord_num_items(coord); + + if (coord_adjust_items(coord, items, 1) == 1) { + return 1; + } + + switch (coord->between) { + case AFTER_UNIT: + case AT_UNIT: + case BEFORE_UNIT: + case AFTER_ITEM: + /* Check for end-of-node. */ + if (coord->item_pos == items - 1) { + coord->between = AFTER_ITEM; + coord->unit_pos = 0; + coord_clear_iplug(coord); + return 1; + } + + /* Anywhere in an item, go to the next one. */ + coord->between = AT_UNIT; + coord_inc_item_pos(coord); + coord->unit_pos = 0; + return 0; + + case BEFORE_ITEM: + /* The out-of-range check ensures that we are valid here. */ + coord->unit_pos = 0; + coord->between = AT_UNIT; + return 0; + case INVALID_COORD: + case EMPTY_NODE: + /* Handled in coord_adjust_items(). */ + break; + } + + impossible("jmacd-9903", "unreachable"); + return 0; +} + +/* Advances the coordinate by one unit to the left. If empty, no change. If + coord_is_leftmost_unit, advances to BEFORE THE FIRST ITEM. Returns 0 if new position + is an existing unit. */ +int coord_prev_unit(coord_t * coord) +{ + unsigned items = coord_num_items(coord); + + if (coord_adjust_items(coord, items, 0) == 1) { + return 1; + } + + switch (coord->between) { + case AT_UNIT: + case BEFORE_UNIT: + if (coord->unit_pos > 0) { + coord->unit_pos -= 1; + coord->between = AT_UNIT; + return 0; + } + + if (coord->item_pos == 0) { + coord->between = BEFORE_ITEM; + return 1; + } + + coord_dec_item_pos(coord); + coord->unit_pos = coord_last_unit_pos(coord); + coord->between = AT_UNIT; + return 0; + + case AFTER_UNIT: + /* What if unit_pos is out-of-range? */ + assert("jmacd-5442", + coord->unit_pos <= coord_last_unit_pos(coord)); + coord->between = AT_UNIT; + return 0; + + case BEFORE_ITEM: + if (coord->item_pos == 0) { + return 1; + } + + coord_dec_item_pos(coord); + /* FALLTHROUGH */ + + case AFTER_ITEM: + coord->between = AT_UNIT; + coord->unit_pos = coord_last_unit_pos(coord); + return 0; + + case INVALID_COORD: + case EMPTY_NODE: + break; + } + + impossible("jmacd-9904", "unreachable"); + return 0; +} + +/* Advances the coordinate by one item to the left. If empty, no change. If + coord_is_leftmost_unit, advances to BEFORE THE FIRST ITEM. Returns 0 if new position + is an existing item. */ +int coord_prev_item(coord_t * coord) +{ + unsigned items = coord_num_items(coord); + + if (coord_adjust_items(coord, items, 0) == 1) { + return 1; + } + + switch (coord->between) { + case AT_UNIT: + case AFTER_UNIT: + case BEFORE_UNIT: + case BEFORE_ITEM: + + if (coord->item_pos == 0) { + coord->between = BEFORE_ITEM; + coord->unit_pos = 0; + return 1; + } + + coord_dec_item_pos(coord); + coord->unit_pos = 0; + coord->between = AT_UNIT; + return 0; + + case AFTER_ITEM: + coord->between = AT_UNIT; + coord->unit_pos = 0; + return 0; + + case INVALID_COORD: + case EMPTY_NODE: + break; + } + + impossible("jmacd-9905", "unreachable"); + return 0; +} + +/* Calls either coord_init_first_unit or coord_init_last_unit depending on sideof argument. */ +void coord_init_sideof_unit(coord_t * coord, const znode * node, sideof dir) +{ + assert("jmacd-9821", dir == LEFT_SIDE || dir == RIGHT_SIDE); + if (dir == LEFT_SIDE) { + coord_init_first_unit(coord, node); + } else { + coord_init_last_unit(coord, node); + } +} + +/* Calls either coord_is_before_leftmost or coord_is_after_rightmost depending on sideof + argument. */ +/* Audited by: green(2002.06.15) */ +int coord_is_after_sideof_unit(coord_t * coord, sideof dir) +{ + assert("jmacd-9822", dir == LEFT_SIDE || dir == RIGHT_SIDE); + if (dir == LEFT_SIDE) { + return coord_is_before_leftmost(coord); + } else { + return coord_is_after_rightmost(coord); + } +} + +/* Calls either coord_next_unit or coord_prev_unit depending on sideof argument. */ +/* Audited by: green(2002.06.15) */ +int coord_sideof_unit(coord_t * coord, sideof dir) +{ + assert("jmacd-9823", dir == LEFT_SIDE || dir == RIGHT_SIDE); + if (dir == LEFT_SIDE) { + return coord_prev_unit(coord); + } else { + return coord_next_unit(coord); + } +} + +#if REISER4_DEBUG +int coords_equal(const coord_t * c1, const coord_t * c2) +{ + assert("nikita-2840", c1 != NULL); + assert("nikita-2841", c2 != NULL); + + return + c1->node == c2->node && + c1->item_pos == c2->item_pos && + c1->unit_pos == c2->unit_pos && c1->between == c2->between; +} +#endif /* REISER4_DEBUG */ + +/* If coord_is_after_rightmost return NCOORD_ON_THE_RIGHT, if coord_is_after_leftmost + return NCOORD_ON_THE_LEFT, otherwise return NCOORD_INSIDE. */ +/* Audited by: green(2002.06.15) */ +coord_wrt_node coord_wrt(const coord_t * coord) +{ + if (coord_is_before_leftmost(coord)) { + return COORD_ON_THE_LEFT; + } + + if (coord_is_after_rightmost(coord)) { + return COORD_ON_THE_RIGHT; + } + + return COORD_INSIDE; +} + +/* Returns true if the coordinate is positioned after the last item or after the last unit + of the last item or it is an empty node. */ +/* Audited by: green(2002.06.15) */ +int coord_is_after_rightmost(const coord_t * coord) +{ + assert("jmacd-7313", coord_check(coord)); + + switch (coord->between) { + case INVALID_COORD: + case AT_UNIT: + case BEFORE_UNIT: + case BEFORE_ITEM: + return 0; + + case EMPTY_NODE: + return 1; + + case AFTER_ITEM: + return (coord->item_pos == node_num_items(coord->node) - 1); + + case AFTER_UNIT: + return ((coord->item_pos == node_num_items(coord->node) - 1) && + coord->unit_pos == coord_last_unit_pos(coord)); + } + + impossible("jmacd-9908", "unreachable"); + return 0; +} + +/* Returns true if the coordinate is positioned before the first item or it is an empty + node. */ +int coord_is_before_leftmost(const coord_t * coord) +{ + /* FIXME-VS: coord_check requires node to be loaded whereas it is not + necessary to check if coord is set before leftmost + assert ("jmacd-7313", coord_check (coord)); */ + switch (coord->between) { + case INVALID_COORD: + case AT_UNIT: + case AFTER_ITEM: + case AFTER_UNIT: + return 0; + + case EMPTY_NODE: + return 1; + + case BEFORE_ITEM: + case BEFORE_UNIT: + return (coord->item_pos == 0) && (coord->unit_pos == 0); + } + + impossible("jmacd-9908", "unreachable"); + return 0; +} + +/* Returns true if the coordinate is positioned after a item, before a item, after the + last unit of an item, before the first unit of an item, or at an empty node. */ +/* Audited by: green(2002.06.15) */ +int coord_is_between_items(const coord_t * coord) +{ + assert("jmacd-7313", coord_check(coord)); + + switch (coord->between) { + case INVALID_COORD: + case AT_UNIT: + return 0; + + case AFTER_ITEM: + case BEFORE_ITEM: + case EMPTY_NODE: + return 1; + + case BEFORE_UNIT: + return coord->unit_pos == 0; + + case AFTER_UNIT: + return coord->unit_pos == coord_last_unit_pos(coord); + } + + impossible("jmacd-9908", "unreachable"); + return 0; +} + +#if REISER4_DEBUG +/* Returns true if the coordinates are positioned at adjacent units, regardless of + before-after or item boundaries. */ +int coord_are_neighbors(coord_t * c1, coord_t * c2) +{ + coord_t *left; + coord_t *right; + + assert("nikita-1241", c1 != NULL); + assert("nikita-1242", c2 != NULL); + assert("nikita-1243", c1->node == c2->node); + assert("nikita-1244", coord_is_existing_unit(c1)); + assert("nikita-1245", coord_is_existing_unit(c2)); + + left = right = NULL; + switch (coord_compare(c1, c2)) { + case COORD_CMP_ON_LEFT: + left = c1; + right = c2; + break; + case COORD_CMP_ON_RIGHT: + left = c2; + right = c1; + break; + case COORD_CMP_SAME: + return 0; + default: + wrong_return_value("nikita-1246", "compare_coords()"); + } + assert("vs-731", left && right); + if (left->item_pos == right->item_pos) { + return left->unit_pos + 1 == right->unit_pos; + } else if (left->item_pos + 1 == right->item_pos) { + return (left->unit_pos == coord_last_unit_pos(left)) + && (right->unit_pos == 0); + } else { + return 0; + } +} +#endif /* REISER4_DEBUG */ + +/* Assuming two coordinates are positioned in the same node, return COORD_CMP_ON_RIGHT, + COORD_CMP_ON_LEFT, or COORD_CMP_SAME depending on c1's position relative to c2. */ +/* Audited by: green(2002.06.15) */ +coord_cmp coord_compare(coord_t * c1, coord_t * c2) +{ + assert("vs-209", c1->node == c2->node); + assert("vs-194", coord_is_existing_unit(c1) + && coord_is_existing_unit(c2)); + + if (c1->item_pos > c2->item_pos) + return COORD_CMP_ON_RIGHT; + if (c1->item_pos < c2->item_pos) + return COORD_CMP_ON_LEFT; + if (c1->unit_pos > c2->unit_pos) + return COORD_CMP_ON_RIGHT; + if (c1->unit_pos < c2->unit_pos) + return COORD_CMP_ON_LEFT; + return COORD_CMP_SAME; +} + +/* If the coordinate is between items, shifts it to the right. Returns 0 on success and + non-zero if there is no position to the right. */ +int coord_set_to_right(coord_t * coord) +{ + unsigned items = coord_num_items(coord); + + if (coord_adjust_items(coord, items, 1) == 1) { + return 1; + } + + switch (coord->between) { + case AT_UNIT: + return 0; + + case BEFORE_ITEM: + case BEFORE_UNIT: + coord->between = AT_UNIT; + return 0; + + case AFTER_UNIT: + if (coord->unit_pos < coord_last_unit_pos(coord)) { + coord->unit_pos += 1; + coord->between = AT_UNIT; + return 0; + } else { + + coord->unit_pos = 0; + + if (coord->item_pos == items - 1) { + coord->between = AFTER_ITEM; + return 1; + } + + coord_inc_item_pos(coord); + coord->between = AT_UNIT; + return 0; + } + + case AFTER_ITEM: + if (coord->item_pos == items - 1) { + return 1; + } + + coord_inc_item_pos(coord); + coord->unit_pos = 0; + coord->between = AT_UNIT; + return 0; + + case EMPTY_NODE: + return 1; + + case INVALID_COORD: + break; + } + + impossible("jmacd-9920", "unreachable"); + return 0; +} + +/* If the coordinate is between items, shifts it to the left. Returns 0 on success and + non-zero if there is no position to the left. */ +int coord_set_to_left(coord_t * coord) +{ + unsigned items = coord_num_items(coord); + + if (coord_adjust_items(coord, items, 0) == 1) { + return 1; + } + + switch (coord->between) { + case AT_UNIT: + return 0; + + case AFTER_UNIT: + coord->between = AT_UNIT; + return 0; + + case AFTER_ITEM: + coord->between = AT_UNIT; + coord->unit_pos = coord_last_unit_pos(coord); + return 0; + + case BEFORE_UNIT: + if (coord->unit_pos > 0) { + coord->unit_pos -= 1; + coord->between = AT_UNIT; + return 0; + } else { + + if (coord->item_pos == 0) { + coord->between = BEFORE_ITEM; + return 1; + } + + coord->unit_pos = coord_last_unit_pos(coord); + coord_dec_item_pos(coord); + coord->between = AT_UNIT; + return 0; + } + + case BEFORE_ITEM: + if (coord->item_pos == 0) { + return 1; + } + + coord_dec_item_pos(coord); + coord->unit_pos = coord_last_unit_pos(coord); + coord->between = AT_UNIT; + return 0; + + case EMPTY_NODE: + return 1; + + case INVALID_COORD: + break; + } + + impossible("jmacd-9920", "unreachable"); + return 0; +} + +static const char *coord_tween_tostring(between_enum n) +{ + switch (n) { + case BEFORE_UNIT: + return "before unit"; + case BEFORE_ITEM: + return "before item"; + case AT_UNIT: + return "at unit"; + case AFTER_UNIT: + return "after unit"; + case AFTER_ITEM: + return "after item"; + case EMPTY_NODE: + return "empty node"; + case INVALID_COORD: + return "invalid"; + default: + { + static char buf[30]; + + sprintf(buf, "unknown: %i", n); + return buf; + } + } +} + +void print_coord(const char *mes, const coord_t * coord, int node) +{ + if (coord == NULL) { + printk("%s: null\n", mes); + return; + } + printk("%s: item_pos = %d, unit_pos %d, tween=%s, iplug=%d\n", + mes, coord->item_pos, coord->unit_pos, + coord_tween_tostring(coord->between), coord->iplugid); +} + +int +item_utmost_child_real_block(const coord_t * coord, sideof side, + reiser4_block_nr * blk) +{ + return item_plugin_by_coord(coord)->f.utmost_child_real_block(coord, + side, + blk); +} + +int item_utmost_child(const coord_t * coord, sideof side, jnode ** child) +{ + return item_plugin_by_coord(coord)->f.utmost_child(coord, side, child); +} + +/* @count bytes of flow @f got written, update correspondingly f->length, + f->data and f->key */ +void move_flow_forward(flow_t * f, unsigned count) +{ + if (f->data) + f->data += count; + f->length -= count; + set_key_offset(&f->key, get_key_offset(&f->key) + count); +} + +/* + Local variables: + c-indentation-style: "K&R" + mode-name: "LC" + c-basic-offset: 8 + tab-width: 8 + fill-column: 120 + scroll-step: 1 + End: +*/ diff -urN linux-2.6.22.orig/fs/reiser4/coord.h linux-2.6.22/fs/reiser4/coord.h --- linux-2.6.22.orig/fs/reiser4/coord.h 1970-01-01 03:00:00.000000000 +0300 +++ linux-2.6.22/fs/reiser4/coord.h 2007-07-29 00:25:34.832685088 +0400 @@ -0,0 +1,389 @@ +/* Copyright 2001, 2002, 2003 by Hans Reiser, licensing governed by reiser4/README */ + +/* Coords */ + +#if !defined( __REISER4_COORD_H__ ) +#define __REISER4_COORD_H__ + +#include "forward.h" +#include "debug.h" +#include "dformat.h" +#include "key.h" + +/* insertions happen between coords in the tree, so we need some means + of specifying the sense of betweenness. */ +typedef enum { + BEFORE_UNIT, /* Note: we/init_coord depends on this value being zero. */ + AT_UNIT, + AFTER_UNIT, + BEFORE_ITEM, + AFTER_ITEM, + INVALID_COORD, + EMPTY_NODE, +} between_enum; + +/* location of coord w.r.t. its node */ +typedef enum { + COORD_ON_THE_LEFT = -1, + COORD_ON_THE_RIGHT = +1, + COORD_INSIDE = 0 +} coord_wrt_node; + +typedef enum { + COORD_CMP_SAME = 0, COORD_CMP_ON_LEFT = -1, COORD_CMP_ON_RIGHT = +1 +} coord_cmp; + +struct coord { + /* node in a tree */ + /* 0 */ znode *node; + + /* position of item within node */ + /* 4 */ pos_in_node_t item_pos; + /* position of unit within item */ + /* 6 */ pos_in_node_t unit_pos; + /* optimization: plugin of item is stored in coord_t. Until this was + implemented, item_plugin_by_coord() was major CPU consumer. ->iplugid + is invalidated (set to 0xff) on each modification of ->item_pos, + and all such modifications are funneled through coord_*_item_pos() + functions below. + */ + /* 8 */ char iplugid; + /* position of coord w.r.t. to neighboring items and/or units. + Values are taken from &between_enum above. + */ + /* 9 */ char between; + /* padding. It will be added by the compiler anyway to conform to the + * C language alignment requirements. We keep it here to be on the + * safe side and to have a clear picture of the memory layout of this + * structure. */ + /* 10 */ __u16 pad; + /* 12 */ int offset; +#if REISER4_DEBUG + unsigned long plug_v; + unsigned long body_v; +#endif +}; + +#define INVALID_PLUGID ((char)((1 << 8) - 1)) +#define INVALID_OFFSET -1 + +static inline void coord_clear_iplug(coord_t * coord) +{ + assert("nikita-2835", coord != NULL); + coord->iplugid = INVALID_PLUGID; + coord->offset = INVALID_OFFSET; +} + +static inline int coord_is_iplug_set(const coord_t * coord) +{ + assert("nikita-2836", coord != NULL); + return coord->iplugid != INVALID_PLUGID; +} + +static inline void coord_set_item_pos(coord_t * coord, pos_in_node_t pos) +{ + assert("nikita-2478", coord != NULL); + coord->item_pos = pos; + coord_clear_iplug(coord); +} + +static inline void coord_dec_item_pos(coord_t * coord) +{ + assert("nikita-2480", coord != NULL); + --coord->item_pos; + coord_clear_iplug(coord); +} + +static inline void coord_inc_item_pos(coord_t * coord) +{ + assert("nikita-2481", coord != NULL); + ++coord->item_pos; + coord_clear_iplug(coord); +} + +static inline void coord_add_item_pos(coord_t * coord, int delta) +{ + assert("nikita-2482", coord != NULL); + coord->item_pos += delta; + coord_clear_iplug(coord); +} + +static inline void coord_invalid_item_pos(coord_t * coord) +{ + assert("nikita-2832", coord != NULL); + coord->item_pos = (unsigned short)~0; + coord_clear_iplug(coord); +} + +/* Reverse a direction. */ +static inline sideof sideof_reverse(sideof side) +{ + return side == LEFT_SIDE ? RIGHT_SIDE : LEFT_SIDE; +} + +/* NOTE: There is a somewhat odd mixture of the following opposed terms: + + "first" and "last" + "next" and "prev" + "before" and "after" + "leftmost" and "rightmost" + + But I think the chosen names are decent the way they are. +*/ + +/* COORD INITIALIZERS */ + +/* Initialize an invalid coordinate. */ +extern void coord_init_invalid(coord_t * coord, const znode * node); + +extern void coord_init_first_unit_nocheck(coord_t * coord, const znode * node); + +/* Initialize a coordinate to point at the first unit of the first item. If the node is + empty, it is positioned at the EMPTY_NODE. */ +extern void coord_init_first_unit(coord_t * coord, const znode * node); + +/* Initialize a coordinate to point at the last unit of the last item. If the node is + empty, it is positioned at the EMPTY_NODE. */ +extern void coord_init_last_unit(coord_t * coord, const znode * node); + +/* Initialize a coordinate to before the first item. If the node is empty, it is + positioned at the EMPTY_NODE. */ +extern void coord_init_before_first_item(coord_t * coord, const znode * node); + +/* Initialize a coordinate to after the last item. If the node is empty, it is positioned + at the EMPTY_NODE. */ +extern void coord_init_after_last_item(coord_t * coord, const znode * node); + +/* Initialize a coordinate to after last unit in the item. Coord must be set + already to existing item */ +void coord_init_after_item_end(coord_t * coord); + +/* Initialize a coordinate to before the item. Coord must be set already to existing item */ +void coord_init_before_item(coord_t *); +/* Initialize a coordinate to after the item. Coord must be set already to existing item */ +void coord_init_after_item(coord_t *); + +/* Calls either coord_init_first_unit or coord_init_last_unit depending on sideof argument. */ +extern void coord_init_sideof_unit(coord_t * coord, const znode * node, + sideof dir); + +/* Initialize a coordinate by 0s. Used in places where init_coord was used and + it was not clear how actually + FIXME-VS: added by vs (2002, june, 8) */ +extern void coord_init_zero(coord_t * coord); + +/* COORD METHODS */ + +/* after shifting of node content, coord previously set properly may become + invalid, try to "normalize" it. */ +void coord_normalize(coord_t * coord); + +/* Copy a coordinate. */ +extern void coord_dup(coord_t * coord, const coord_t * old_coord); + +/* Copy a coordinate without check. */ +void coord_dup_nocheck(coord_t * coord, const coord_t * old_coord); + +unsigned coord_num_units(const coord_t * coord); + +/* Return the last valid unit number at the present item (i.e., + coord_num_units() - 1). */ +static inline unsigned coord_last_unit_pos(const coord_t * coord) +{ + return coord_num_units(coord) - 1; +} + +#if REISER4_DEBUG +/* For assertions only, checks for a valid coordinate. */ +extern int coord_check(const coord_t * coord); + +extern unsigned long znode_times_locked(const znode * z); + +static inline void coord_update_v(coord_t * coord) +{ + coord->plug_v = coord->body_v = znode_times_locked(coord->node); +} +#endif + +extern int coords_equal(const coord_t * c1, const coord_t * c2); + +extern void print_coord(const char *mes, const coord_t * coord, int print_node); + +/* If coord_is_after_rightmost return NCOORD_ON_THE_RIGHT, if coord_is_after_leftmost + return NCOORD_ON_THE_LEFT, otherwise return NCOORD_INSIDE. */ +extern coord_wrt_node coord_wrt(const coord_t * coord); + +/* Returns true if the coordinates are positioned at adjacent units, regardless of + before-after or item boundaries. */ +extern int coord_are_neighbors(coord_t * c1, coord_t * c2); + +/* Assuming two coordinates are positioned in the same node, return NCOORD_CMP_ON_RIGHT, + NCOORD_CMP_ON_LEFT, or NCOORD_CMP_SAME depending on c1's position relative to c2. */ +extern coord_cmp coord_compare(coord_t * c1, coord_t * c2); + +/* COORD PREDICATES */ + +/* Returns true if the coord was initializewd by coord_init_invalid (). */ +extern int coord_is_invalid(const coord_t * coord); + +/* Returns true if the coordinate is positioned at an existing item, not before or after + an item. It may be placed at, before, or after any unit within the item, whether + existing or not. If this is true you can call methods of the item plugin. */ +extern int coord_is_existing_item(const coord_t * coord); + +/* Returns true if the coordinate is positioned after a item, before a item, after the + last unit of an item, before the first unit of an item, or at an empty node. */ +extern int coord_is_between_items(const coord_t * coord); + +/* Returns true if the coordinate is positioned at an existing unit, not before or after a + unit. */ +extern int coord_is_existing_unit(const coord_t * coord); + +/* Returns true if the coordinate is positioned at an empty node. */ +extern int coord_is_empty(const coord_t * coord); + +/* Returns true if the coordinate is positioned at the first unit of the first item. Not + true for empty nodes nor coordinates positioned before the first item. */ +extern int coord_is_leftmost_unit(const coord_t * coord); + +/* Returns true if the coordinate is positioned after the last item or after the last unit + of the last item or it is an empty node. */ +extern int coord_is_after_rightmost(const coord_t * coord); + +/* Returns true if the coordinate is positioned before the first item or it is an empty + node. */ +extern int coord_is_before_leftmost(const coord_t * coord); + +/* Calls either coord_is_before_leftmost or coord_is_after_rightmost depending on sideof + argument. */ +extern int coord_is_after_sideof_unit(coord_t * coord, sideof dir); + +/* COORD MODIFIERS */ + +/* Advances the coordinate by one unit to the right. If empty, no change. If + coord_is_rightmost_unit, advances to AFTER THE LAST ITEM. Returns 0 if new position is + an existing unit. */ +extern int coord_next_unit(coord_t * coord); + +/* Advances the coordinate by one item to the right. If empty, no change. If + coord_is_rightmost_unit, advances to AFTER THE LAST ITEM. Returns 0 if new position is + an existing item. */ +extern int coord_next_item(coord_t * coord); + +/* Advances the coordinate by one unit to the left. If empty, no change. If + coord_is_leftmost_unit, advances to BEFORE THE FIRST ITEM. Returns 0 if new position + is an existing unit. */ +extern int coord_prev_unit(coord_t * coord); + +/* Advances the coordinate by one item to the left. If empty, no change. If + coord_is_leftmost_unit, advances to BEFORE THE FIRST ITEM. Returns 0 if new position + is an existing item. */ +extern int coord_prev_item(coord_t * coord); + +/* If the coordinate is between items, shifts it to the right. Returns 0 on success and + non-zero if there is no position to the right. */ +extern int coord_set_to_right(coord_t * coord); + +/* If the coordinate is between items, shifts it to the left. Returns 0 on success and + non-zero if there is no position to the left. */ +extern int coord_set_to_left(coord_t * coord); + +/* If the coordinate is at an existing unit, set to after that unit. Returns 0 on success + and non-zero if the unit did not exist. */ +extern int coord_set_after_unit(coord_t * coord); + +/* Calls either coord_next_unit or coord_prev_unit depending on sideof argument. */ +extern int coord_sideof_unit(coord_t * coord, sideof dir); + +/* iterate over all units in @node */ +#define for_all_units( coord, node ) \ + for( coord_init_before_first_item( ( coord ), ( node ) ) ; \ + coord_next_unit( coord ) == 0 ; ) + +/* iterate over all items in @node */ +#define for_all_items( coord, node ) \ + for( coord_init_before_first_item( ( coord ), ( node ) ) ; \ + coord_next_item( coord ) == 0 ; ) + +/* COORD/ITEM METHODS */ + +extern int item_utmost_child_real_block(const coord_t * coord, sideof side, + reiser4_block_nr * blk); +extern int item_utmost_child(const coord_t * coord, sideof side, + jnode ** child); + +/* a flow is a sequence of bytes being written to or read from the tree. The + tree will slice the flow into items while storing it into nodes, but all of + that is hidden from anything outside the tree. */ + +struct flow { + reiser4_key key; /* key of start of flow's sequence of bytes */ + loff_t length; /* length of flow's sequence of bytes */ + char *data; /* start of flow's sequence of bytes */ + int user; /* if 1 data is user space, 0 - kernel space */ + rw_op op; /* NIKITA-FIXME-HANS: comment is where? */ +}; + +void move_flow_forward(flow_t * f, unsigned count); + +/* &reiser4_item_data - description of data to be inserted or pasted + + Q: articulate the reasons for the difference between this and flow. + + A: Becides flow we insert into tree other things: stat data, directory + entry, etc. To insert them into tree one has to provide this structure. If + one is going to insert flow - he can use insert_flow, where this structure + does not have to be created +*/ +struct reiser4_item_data { + /* actual data to be inserted. If NULL, ->create_item() will not + do xmemcpy itself, leaving this up to the caller. This can + save some amount of unnecessary memory copying, for example, + during insertion of stat data. + + */ + char *data; + /* 1 if 'char * data' contains pointer to user space and 0 if it is + kernel space */ + int user; + /* amount of data we are going to insert or paste */ + int length; + /* "Arg" is opaque data that is passed down to the + ->create_item() method of node layout, which in turn + hands it to the ->create_hook() of item being created. This + arg is currently used by: + + . ->create_hook() of internal item + (fs/reiser4/plugin/item/internal.c:internal_create_hook()), + . ->paste() method of directory item. + . ->create_hook() of extent item + + For internal item, this is left "brother" of new node being + inserted and it is used to add new node into sibling list + after parent to it was just inserted into parent. + + While ->arg does look somewhat of unnecessary compication, + it actually saves a lot of headache in many places, because + all data necessary to insert or paste new data into tree are + collected in one place, and this eliminates a lot of extra + argument passing and storing everywhere. + + */ + void *arg; + /* plugin of item we are inserting */ + item_plugin *iplug; +}; + +/* __REISER4_COORD_H__ */ +#endif + +/* Make Linus happy. + Local variables: + c-indentation-style: "K&R" + mode-name: "LC" + c-basic-offset: 8 + tab-width: 8 + fill-column: 120 + scroll-step: 1 + End: +*/ diff -urN linux-2.6.22.orig/fs/reiser4/debug.c linux-2.6.22/fs/reiser4/debug.c --- linux-2.6.22.orig/fs/reiser4/debug.c 1970-01-01 03:00:00.000000000 +0300 +++ linux-2.6.22/fs/reiser4/debug.c 2007-07-29 00:25:34.836686123 +0400 @@ -0,0 +1,308 @@ +/* Copyright 2001, 2002, 2003 by Hans Reiser, licensing governed by + * reiser4/README */ + +/* Debugging facilities. */ + +/* + * This file contains generic debugging functions used by reiser4. Roughly + * following: + * + * panicking: reiser4_do_panic(), reiser4_print_prefix(). + * + * locking: + * reiser4_schedulable(), reiser4_lock_counters(), print_lock_counters(), + * reiser4_no_counters_are_held(), reiser4_commit_check_locks() + * + * error code monitoring (see comment before RETERR macro): + * reiser4_return_err(), reiser4_report_err(). + * + * stack back-tracing: fill_backtrace() + * + * miscellaneous: reiser4_preempt_point(), call_on_each_assert(), + * reiser4_debugtrap(). + * + */ + +#include "reiser4.h" +#include "context.h" +#include "super.h" +#include "txnmgr.h" +#include "znode.h" + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#if 0 +#if REISER4_DEBUG +static void reiser4_report_err(void); +#else +#define reiser4_report_err() noop +#endif +#endif /* 0 */ + +/* + * global buffer where message given to reiser4_panic is formatted. + */ +static char panic_buf[REISER4_PANIC_MSG_BUFFER_SIZE]; + +/* + * lock protecting consistency of panic_buf under concurrent panics + */ +static DEFINE_SPINLOCK(panic_guard); + +/* Your best friend. Call it on each occasion. This is called by + fs/reiser4/debug.h:reiser4_panic(). */ +void reiser4_do_panic(const char *format /* format string */ , ... /* rest */ ) +{ + static int in_panic = 0; + va_list args; + + /* + * check for recursive panic. + */ + if (in_panic == 0) { + in_panic = 1; + + spin_lock(&panic_guard); + va_start(args, format); + vsnprintf(panic_buf, sizeof(panic_buf), format, args); + va_end(args); + printk(KERN_EMERG "reiser4 panicked cowardly: %s", panic_buf); + spin_unlock(&panic_guard); + + /* + * if kernel debugger is configured---drop in. Early dropping + * into kgdb is not always convenient, because panic message + * is not yet printed most of the times. But: + * + * (1) message can be extracted from printk_buf[] + * (declared static inside of printk()), and + * + * (2) sometimes serial/kgdb combo dies while printing + * long panic message, so it's more prudent to break into + * debugger earlier. + * + */ + DEBUGON(1); + } + /* to make gcc happy about noreturn attribute */ + panic("%s", panic_buf); +} + +#if 0 +void +reiser4_print_prefix(const char *level, int reperr, const char *mid, + const char *function, const char *file, int lineno) +{ + const char *comm; + int pid; + + if (unlikely(in_interrupt() || in_irq())) { + comm = "interrupt"; + pid = 0; + } else { + comm = current->comm; + pid = current->pid; + } + printk("%sreiser4[%.16s(%i)]: %s (%s:%i)[%s]:\n", + level, comm, pid, function, file, lineno, mid); + if (reperr) + reiser4_report_err(); +} +#endif /* 0 */ + +/* Preemption point: this should be called periodically during long running + operations (carry, allocate, and squeeze are best examples) */ +int reiser4_preempt_point(void) +{ + assert("nikita-3008", reiser4_schedulable()); + cond_resched(); + return signal_pending(current); +} + +#if REISER4_DEBUG +/* Debugging aid: return struct where information about locks taken by current + thread is accumulated. This can be used to formulate lock ordering + constraints and various assertions. + +*/ +reiser4_lock_cnt_info *reiser4_lock_counters(void) +{ + reiser4_context *ctx = get_current_context(); + assert("jmacd-1123", ctx != NULL); + return &ctx->locks; +} + +/* + * print human readable information about locks held by the reiser4 context. + */ +static void print_lock_counters(const char *prefix, + const reiser4_lock_cnt_info * info) +{ + printk("%s: jnode: %i, tree: %i (r:%i,w:%i), dk: %i (r:%i,w:%i)\n" + "jload: %i, " + "txnh: %i, atom: %i, stack: %i, txnmgr: %i, " + "ktxnmgrd: %i, fq: %i\n" + "inode: %i, " + "cbk_cache: %i (r:%i,w%i), " + "eflush: %i, " + "zlock: %i,\n" + "spin: %i, long: %i inode_sem: (r:%i,w:%i)\n" + "d: %i, x: %i, t: %i\n", prefix, + info->spin_locked_jnode, + info->rw_locked_tree, info->read_locked_tree, + info->write_locked_tree, + info->rw_locked_dk, info->read_locked_dk, info->write_locked_dk, + info->spin_locked_jload, + info->spin_locked_txnh, + info->spin_locked_atom, info->spin_locked_stack, + info->spin_locked_txnmgr, info->spin_locked_ktxnmgrd, + info->spin_locked_fq, + info->spin_locked_inode, + info->rw_locked_cbk_cache, + info->read_locked_cbk_cache, + info->write_locked_cbk_cache, + info->spin_locked_super_eflush, + info->spin_locked_zlock, + info->spin_locked, + info->long_term_locked_znode, + info->inode_sem_r, info->inode_sem_w, + info->d_refs, info->x_refs, info->t_refs); +} + +/* check that no spinlocks are held */ +int reiser4_schedulable(void) +{ + if (get_current_context_check() != NULL) { + if (!LOCK_CNT_NIL(spin_locked)) { + print_lock_counters("in atomic", reiser4_lock_counters()); + return 0; + } + } + might_sleep(); + return 1; +} +/* + * return true, iff no locks are held. + */ +int reiser4_no_counters_are_held(void) +{ + reiser4_lock_cnt_info *counters; + + counters = reiser4_lock_counters(); + return + (counters->spin_locked_zlock == 0) && + (counters->spin_locked_jnode == 0) && + (counters->rw_locked_tree == 0) && + (counters->read_locked_tree == 0) && + (counters->write_locked_tree == 0) && + (counters->rw_locked_dk == 0) && + (counters->read_locked_dk == 0) && + (counters->write_locked_dk == 0) && + (counters->spin_locked_txnh == 0) && + (counters->spin_locked_atom == 0) && + (counters->spin_locked_stack == 0) && + (counters->spin_locked_txnmgr == 0) && + (counters->spin_locked_inode == 0) && + (counters->spin_locked == 0) && + (counters->long_term_locked_znode == 0) && + (counters->inode_sem_r == 0) && + (counters->inode_sem_w == 0) && (counters->d_refs == 0); +} + +/* + * return true, iff transaction commit can be done under locks held by the + * current thread. + */ +int reiser4_commit_check_locks(void) +{ + reiser4_lock_cnt_info *counters; + int inode_sem_r; + int inode_sem_w; + int result; + + /* + * inode's read/write semaphore is the only reiser4 lock that can be + * held during commit. + */ + + counters = reiser4_lock_counters(); + inode_sem_r = counters->inode_sem_r; + inode_sem_w = counters->inode_sem_w; + + counters->inode_sem_r = counters->inode_sem_w = 0; + result = reiser4_no_counters_are_held(); + counters->inode_sem_r = inode_sem_r; + counters->inode_sem_w = inode_sem_w; + return result; +} + +/* + * fill "error site" in the current reiser4 context. See comment before RETERR + * macro for more details. + */ +void reiser4_return_err(int code, const char *file, int line) +{ + if (code < 0 && is_in_reiser4_context()) { + reiser4_context *ctx = get_current_context(); + + if (ctx != NULL) { + ctx->err.code = code; + ctx->err.file = file; + ctx->err.line = line; + } + } +} + +#if 0 +/* + * report error information recorder by reiser4_return_err(). + */ +static void reiser4_report_err(void) +{ + reiser4_context *ctx = get_current_context_check(); + + if (ctx != NULL) { + if (ctx->err.code != 0) { + printk("code: %i at %s:%i\n", + ctx->err.code, ctx->err.file, ctx->err.line); + } + } +} +#endif /* 0 */ + +#endif /* REISER4_DEBUG */ + +#if KERNEL_DEBUGGER + +/* + * this functions just drops into kernel debugger. It is a convenient place to + * put breakpoint in. + */ +void reiser4_debugtrap(void) +{ + /* do nothing. Put break point here. */ +#if defined(CONFIG_KGDB) && !defined(CONFIG_REISER4_FS_MODULE) + extern void breakpoint(void); + breakpoint(); +#endif +} +#endif + +/* Make Linus happy. + Local variables: + c-indentation-style: "K&R" + mode-name: "LC" + c-basic-offset: 8 + tab-width: 8 + fill-column: 120 + End: +*/ diff -urN linux-2.6.22.orig/fs/reiser4/debug.h linux-2.6.22/fs/reiser4/debug.h --- linux-2.6.22.orig/fs/reiser4/debug.h 1970-01-01 03:00:00.000000000 +0300 +++ linux-2.6.22/fs/reiser4/debug.h 2007-07-29 00:25:34.836686123 +0400 @@ -0,0 +1,350 @@ +/* Copyright 2001, 2002, 2003 by Hans Reiser, licensing governed by reiser4/README */ + +/* Declarations of debug macros. */ + +#if !defined( __FS_REISER4_DEBUG_H__ ) +#define __FS_REISER4_DEBUG_H__ + +#include "forward.h" +#include "reiser4.h" + +/* generic function to produce formatted output, decorating it with + whatever standard prefixes/postfixes we want. "Fun" is a function + that will be actually called, can be printk, panic etc. + This is for use by other debugging macros, not by users. */ +#define DCALL(lev, fun, reperr, label, format, ...) \ +({ \ + fun(lev "reiser4[%.16s(%i)]: %s (%s:%i)[%s]:\n" format "\n" , \ + current->comm, current->pid, __FUNCTION__, \ + __FILE__, __LINE__, label, ## __VA_ARGS__); \ +}) + +/* + * cause kernel to crash + */ +#define reiser4_panic(mid, format, ...) \ + DCALL("", reiser4_do_panic, 1, mid, format , ## __VA_ARGS__) + +/* print message with indication of current process, file, line and + function */ +#define reiser4_log(label, format, ...) \ + DCALL(KERN_DEBUG, printk, 0, label, format , ## __VA_ARGS__) + +/* Assertion checked during compilation. + If "cond" is false (0) we get duplicate case label in switch. + Use this to check something like famous + cassert (sizeof(struct reiserfs_journal_commit) == 4096) ; + in 3.x journal.c. If cassertion fails you get compiler error, + so no "maintainer-id". +*/ +#define cassert(cond) ({ switch(-1) { case (cond): case 0: break; } }) + +#define noop do {;} while(0) + +#if REISER4_DEBUG +/* version of info that only actually prints anything when _d_ebugging + is on */ +#define dinfo(format, ...) printk(format , ## __VA_ARGS__) +/* macro to catch logical errors. Put it into `default' clause of + switch() statement. */ +#define impossible(label, format, ...) \ + reiser4_panic(label, "impossible: " format , ## __VA_ARGS__) +/* assert assures that @cond is true. If it is not, reiser4_panic() is + called. Use this for checking logical consistency and _never_ call + this to check correctness of external data: disk blocks and user-input . */ +#define assert(label, cond) \ +({ \ + /* call_on_each_assert(); */ \ + if (cond) { \ + /* put negated check to avoid using !(cond) that would lose \ + * warnings for things like assert(a = b); */ \ + ; \ + } else { \ + DEBUGON(1); \ + reiser4_panic(label, "assertion failed: %s", #cond); \ + } \ +}) + +/* like assertion, but @expr is evaluated even if REISER4_DEBUG is off. */ +#define check_me( label, expr ) assert( label, ( expr ) ) + +#define ON_DEBUG( exp ) exp + +extern int reiser4_schedulable(void); +extern void call_on_each_assert(void); + +#else + +#define dinfo( format, args... ) noop +#define impossible( label, format, args... ) noop +#define assert( label, cond ) noop +#define check_me( label, expr ) ( ( void ) ( expr ) ) +#define ON_DEBUG( exp ) +#define reiser4_schedulable() might_sleep() + +/* REISER4_DEBUG */ +#endif + +#if REISER4_DEBUG +/* per-thread information about lock acquired by this thread. Used by lock + * ordering checking in spin_macros.h */ +typedef struct reiser4_lock_cnt_info { + int rw_locked_tree; + int read_locked_tree; + int write_locked_tree; + + int rw_locked_dk; + int read_locked_dk; + int write_locked_dk; + + int rw_locked_cbk_cache; + int read_locked_cbk_cache; + int write_locked_cbk_cache; + + int spin_locked_zlock; + int spin_locked_jnode; + int spin_locked_jload; + int spin_locked_txnh; + int spin_locked_atom; + int spin_locked_stack; + int spin_locked_txnmgr; + int spin_locked_ktxnmgrd; + int spin_locked_fq; + int spin_locked_inode; + int spin_locked_super_eflush; + int spin_locked; + int long_term_locked_znode; + + int inode_sem_r; + int inode_sem_w; + + int d_refs; + int x_refs; + int t_refs; +} reiser4_lock_cnt_info; + +extern struct reiser4_lock_cnt_info *reiser4_lock_counters(void); +#define IN_CONTEXT(a, b) (is_in_reiser4_context() ? (a) : (b)) + +/* increment lock-counter @counter, if present */ +#define LOCK_CNT_INC(counter) \ + IN_CONTEXT(++(reiser4_lock_counters()->counter), 0) + +/* decrement lock-counter @counter, if present */ +#define LOCK_CNT_DEC(counter) \ + IN_CONTEXT(--(reiser4_lock_counters()->counter), 0) + +/* check that lock-counter is zero. This is for use in assertions */ +#define LOCK_CNT_NIL(counter) \ + IN_CONTEXT(reiser4_lock_counters()->counter == 0, 1) + +/* check that lock-counter is greater than zero. This is for use in + * assertions */ +#define LOCK_CNT_GTZ(counter) \ + IN_CONTEXT(reiser4_lock_counters()->counter > 0, 1) +#define LOCK_CNT_LT(counter,n) \ + IN_CONTEXT(reiser4_lock_counters()->counter < n, 1) + +#else /* REISER4_DEBUG */ + +/* no-op versions on the above */ + +typedef struct reiser4_lock_cnt_info { +} reiser4_lock_cnt_info; + +#define reiser4_lock_counters() ((reiser4_lock_cnt_info *)NULL) +#define LOCK_CNT_INC(counter) noop +#define LOCK_CNT_DEC(counter) noop +#define LOCK_CNT_NIL(counter) (1) +#define LOCK_CNT_GTZ(counter) (1) +#define LOCK_CNT_LT(counter,n) (1) + +#endif /* REISER4_DEBUG */ + +#define assert_spin_not_locked(lock) BUG_ON(0) +#define assert_rw_write_locked(lock) BUG_ON(0) +#define assert_rw_read_locked(lock) BUG_ON(0) +#define assert_rw_locked(lock) BUG_ON(0) +#define assert_rw_not_write_locked(lock) BUG_ON(0) +#define assert_rw_not_read_locked(lock) BUG_ON(0) +#define assert_rw_not_locked(lock) BUG_ON(0) + +/* flags controlling debugging behavior. Are set through debug_flags=N mount + option. */ +typedef enum { + /* print a lot of information during panic. When this is on all jnodes + * are listed. This can be *very* large output. Usually you don't want + * this. Especially over serial line. */ + REISER4_VERBOSE_PANIC = 0x00000001, + /* print a lot of information during umount */ + REISER4_VERBOSE_UMOUNT = 0x00000002, + /* print gathered statistics on umount */ + REISER4_STATS_ON_UMOUNT = 0x00000004, + /* check node consistency */ + REISER4_CHECK_NODE = 0x00000008 +} reiser4_debug_flags; + +extern int is_in_reiser4_context(void); + +/* + * evaluate expression @e only if with reiser4 context + */ +#define ON_CONTEXT(e) do { \ + if(is_in_reiser4_context()) { \ + e; \ + } } while(0) + +/* + * evaluate expression @e only when within reiser4_context and debugging is + * on. + */ +#define ON_DEBUG_CONTEXT( e ) ON_DEBUG( ON_CONTEXT( e ) ) + +/* + * complain about unexpected function result and crash. Used in "default" + * branches of switch statements and alike to assert that invalid results are + * not silently ignored. + */ +#define wrong_return_value( label, function ) \ + impossible( label, "wrong return value from " function ) + +/* Issue different types of reiser4 messages to the console */ +#define warning( label, format, ... ) \ + DCALL( KERN_WARNING, \ + printk, 1, label, "WARNING: " format , ## __VA_ARGS__ ) +#define notice( label, format, ... ) \ + DCALL( KERN_NOTICE, \ + printk, 1, label, "NOTICE: " format , ## __VA_ARGS__ ) + +/* mark not yet implemented functionality */ +#define not_yet( label, format, ... ) \ + reiser4_panic( label, "NOT YET IMPLEMENTED: " format , ## __VA_ARGS__ ) + +extern void reiser4_do_panic(const char *format, ...) + __attribute__ ((noreturn, format(printf, 1, 2))); + +extern int reiser4_preempt_point(void); +extern void reiser4_print_stats(void); + +#if REISER4_DEBUG +extern int reiser4_no_counters_are_held(void); +extern int reiser4_commit_check_locks(void); +#else +#define reiser4_no_counters_are_held() (1) +#define reiser4_commit_check_locks() (1) +#endif + +/* true if @i is power-of-two. Useful for rate-limited warnings, etc. */ +#define IS_POW(i) \ +({ \ + typeof(i) __i; \ + \ + __i = (i); \ + !(__i & (__i - 1)); \ +}) + +#define KERNEL_DEBUGGER (1) + +#if KERNEL_DEBUGGER + +extern void reiser4_debugtrap(void); + +/* + * Check condition @cond and drop into kernel debugger (kgdb) if it's true. If + * kgdb is not compiled in, do nothing. + */ +#define DEBUGON(cond) \ +({ \ + if (unlikely(cond)) \ + reiser4_debugtrap(); \ +}) +#else +#define DEBUGON(cond) noop +#endif + +/* + * Error code tracing facility. (Idea is borrowed from XFS code.) + * + * Suppose some strange and/or unexpected code is returned from some function + * (for example, write(2) returns -EEXIST). It is possible to place a + * breakpoint in the reiser4_write(), but it is too late here. How to find out + * in what particular place -EEXIST was generated first? + * + * In reiser4 all places where actual error codes are produced (that is, + * statements of the form + * + * return -EFOO; // (1), or + * + * result = -EFOO; // (2) + * + * are replaced with + * + * return RETERR(-EFOO); // (1a), and + * + * result = RETERR(-EFOO); // (2a) respectively + * + * RETERR() macro fills a backtrace in reiser4_context. This back-trace is + * printed in error and warning messages. Moreover, it's possible to put a + * conditional breakpoint in reiser4_return_err (low-level function called + * by RETERR() to do the actual work) to break into debugger immediately + * when particular error happens. + * + */ + +#if REISER4_DEBUG + +/* + * data-type to store information about where error happened ("error site"). + */ +typedef struct err_site { + int code; /* error code */ + const char *file; /* source file, filled by __FILE__ */ + int line; /* source file line, filled by __LINE__ */ +} err_site; + +extern void reiser4_return_err(int code, const char *file, int line); + +/* + * fill &get_current_context()->err_site with error information. + */ +#define RETERR(code) \ +({ \ + typeof(code) __code; \ + \ + __code = (code); \ + reiser4_return_err(__code, __FILE__, __LINE__); \ + __code; \ +}) + +#else + +/* + * no-op versions of the above + */ + +typedef struct err_site { +} err_site; +#define RETERR(code) code +#endif + +#if REISER4_LARGE_KEY +/* + * conditionally compile arguments only if REISER4_LARGE_KEY is on. + */ +#define ON_LARGE_KEY(...) __VA_ARGS__ +#else +#define ON_LARGE_KEY(...) +#endif + +/* __FS_REISER4_DEBUG_H__ */ +#endif + +/* Make Linus happy. + Local variables: + c-indentation-style: "K&R" + mode-name: "LC" + c-basic-offset: 8 + tab-width: 8 + fill-column: 120 + End: +*/ diff -urN linux-2.6.22.orig/fs/reiser4/dformat.h linux-2.6.22/fs/reiser4/dformat.h --- linux-2.6.22.orig/fs/reiser4/dformat.h 1970-01-01 03:00:00.000000000 +0300 +++ linux-2.6.22/fs/reiser4/dformat.h 2007-07-29 00:25:34.836686123 +0400 @@ -0,0 +1,70 @@ +/* Copyright 2001, 2002, 2003 by Hans Reiser, licensing governed by reiser4/README */ + +/* Formats of on-disk data and conversion functions. */ + +/* put all item formats in the files describing the particular items, + our model is, everything you need to do to add an item to reiser4, + (excepting the changes to the plugin that uses the item which go + into the file defining that plugin), you put into one file. */ +/* Data on disk are stored in little-endian format. + To declare fields of on-disk structures, use d8, d16, d32 and d64. + d??tocpu() and cputod??() to convert. */ + +#if !defined( __FS_REISER4_DFORMAT_H__ ) +#define __FS_REISER4_DFORMAT_H__ + +#include +#include +#include + +typedef __u8 d8; +typedef __le16 d16; +typedef __le32 d32; +typedef __le64 d64; + +#define PACKED __attribute__((packed)) + +/* data-type for block number */ +typedef __u64 reiser4_block_nr; + +/* data-type for block number on disk, disk format */ +typedef __le64 reiser4_dblock_nr; + +/** + * disk_addr_eq - compare disk addresses + * @b1: pointer to block number ot compare + * @b2: pointer to block number ot compare + * + * Returns true if if disk addresses are the same + */ +static inline int disk_addr_eq(const reiser4_block_nr *b1, + const reiser4_block_nr * b2) +{ + assert("nikita-1033", b1 != NULL); + assert("nikita-1266", b2 != NULL); + + return !memcmp(b1, b2, sizeof *b1); +} + +/* structure of master reiser4 super block */ +typedef struct reiser4_master_sb { + char magic[16]; /* "ReIsEr4" */ + __le16 disk_plugin_id; /* id of disk layout plugin */ + __le16 blocksize; + char uuid[16]; /* unique id */ + char label[16]; /* filesystem label */ + __le64 diskmap; /* location of the diskmap. 0 if not present */ +} reiser4_master_sb; + +/* __FS_REISER4_DFORMAT_H__ */ +#endif + +/* + * Local variables: + * c-indentation-style: "K&R" + * mode-name: "LC" + * c-basic-offset: 8 + * tab-width: 8 + * fill-column: 79 + * End: + */ diff -urN linux-2.6.22.orig/fs/reiser4/dscale.c linux-2.6.22/fs/reiser4/dscale.c --- linux-2.6.22.orig/fs/reiser4/dscale.c 1970-01-01 03:00:00.000000000 +0300 +++ linux-2.6.22/fs/reiser4/dscale.c 2007-07-29 00:25:34.836686123 +0400 @@ -0,0 +1,174 @@ +/* Copyright 2001, 2002, 2003 by Hans Reiser, licensing governed by + * reiser4/README */ + +/* Scalable on-disk integers */ + +/* + * Various on-disk structures contain integer-like structures. Stat-data + * contain [yes, "data" is plural, check the dictionary] file size, link + * count; extent unit contains extent width etc. To accommodate for general + * case enough space is reserved to keep largest possible value. 64 bits in + * all cases above. But in overwhelming majority of cases numbers actually + * stored in these fields will be comparatively small and reserving 8 bytes is + * a waste of precious disk bandwidth. + * + * Scalable integers are one way to solve this problem. dscale_write() + * function stores __u64 value in the given area consuming from 1 to 9 bytes, + * depending on the magnitude of the value supplied. dscale_read() reads value + * previously stored by dscale_write(). + * + * dscale_write() produces format not completely unlike of UTF: two highest + * bits of the first byte are used to store "tag". One of 4 possible tag + * values is chosen depending on the number being encoded: + * + * 0 ... 0x3f => 0 [table 1] + * 0x40 ... 0x3fff => 1 + * 0x4000 ... 0x3fffffff => 2 + * 0x40000000 ... 0xffffffffffffffff => 3 + * + * (see dscale_range() function) + * + * Values in the range 0x40000000 ... 0xffffffffffffffff require 8 full bytes + * to be stored, so in this case there is no place in the first byte to store + * tag. For such values tag is stored in an extra 9th byte. + * + * As _highest_ bits are used for the test (which is natural) scaled integers + * are stored in BIG-ENDIAN format in contrast with the rest of reiser4 which + * uses LITTLE-ENDIAN. + * + */ + +#include "debug.h" +#include "dscale.h" + +/* return tag of scaled integer stored at @address */ +static int gettag(const unsigned char *address) +{ + /* tag is stored in two highest bits */ + return (*address) >> 6; +} + +/* clear tag from value. Clear tag embedded into @value. */ +static void cleartag(__u64 * value, int tag) +{ + /* + * W-w-what ?! + * + * Actually, this is rather simple: @value passed here was read by + * dscale_read(), converted from BIG-ENDIAN, and padded to __u64 by + * zeroes. Tag is still stored in the highest (arithmetically) + * non-zero bits of @value, but relative position of tag within __u64 + * depends on @tag. + * + * For example if @tag is 0, it's stored 2 highest bits of lowest + * byte, and its offset (counting from lowest bit) is 8 - 2 == 6 bits. + * + * If tag is 1, it's stored in two highest bits of 2nd lowest byte, + * and it's offset if (2 * 8) - 2 == 14 bits. + * + * See table 1 above for details. + * + * All these cases are captured by the formula: + */ + *value &= ~(3 << (((1 << tag) << 3) - 2)); + /* + * That is, clear two (3 == 0t11) bits at the offset + * + * 8 * (2 ^ tag) - 2, + * + * that is, two highest bits of (2 ^ tag)-th byte of @value. + */ +} + +/* return tag for @value. See table 1 above for details. */ +static int dscale_range(__u64 value) +{ + if (value > 0x3fffffff) + return 3; + if (value > 0x3fff) + return 2; + if (value > 0x3f) + return 1; + return 0; +} + +/* restore value stored at @adderss by dscale_write() and return number of + * bytes consumed */ +int dscale_read(unsigned char *address, __u64 * value) +{ + int tag; + + /* read tag */ + tag = gettag(address); + switch (tag) { + case 3: + /* In this case tag is stored in an extra byte, skip this byte + * and decode value stored in the next 8 bytes.*/ + *value = __be64_to_cpu(get_unaligned((__be64 *)(address + 1))); + /* worst case: 8 bytes for value itself plus one byte for + * tag. */ + return 9; + case 0: + *value = get_unaligned(address); + break; + case 1: + *value = __be16_to_cpu(get_unaligned((__be16 *)address)); + break; + case 2: + *value = __be32_to_cpu(get_unaligned((__be32 *)address)); + break; + default: + return RETERR(-EIO); + } + /* clear tag embedded into @value */ + cleartag(value, tag); + /* number of bytes consumed is (2 ^ tag)---see table 1. */ + return 1 << tag; +} + +/* store @value at @address and return number of bytes consumed */ +int dscale_write(unsigned char *address, __u64 value) +{ + int tag; + int shift; + __be64 v; + unsigned char *valarr; + + tag = dscale_range(value); + v = __cpu_to_be64(value); + valarr = (unsigned char *)&v; + shift = (tag == 3) ? 1 : 0; + memcpy(address + shift, valarr + sizeof v - (1 << tag), 1 << tag); + *address |= (tag << 6); + return shift + (1 << tag); +} + +/* number of bytes required to store @value */ +int dscale_bytes(__u64 value) +{ + int bytes; + + bytes = 1 << dscale_range(value); + if (bytes == 8) + ++bytes; + return bytes; +} + +/* returns true if @value and @other require the same number of bytes to be + * stored. Used by detect when data structure (like stat-data) has to be + * expanded or contracted. */ +int dscale_fit(__u64 value, __u64 other) +{ + return dscale_range(value) == dscale_range(other); +} + +/* Make Linus happy. + Local variables: + c-indentation-style: "K&R" + mode-name: "LC" + c-basic-offset: 8 + tab-width: 8 + fill-column: 120 + scroll-step: 1 + End: +*/ diff -urN linux-2.6.22.orig/fs/reiser4/dscale.h linux-2.6.22/fs/reiser4/dscale.h --- linux-2.6.22.orig/fs/reiser4/dscale.h 1970-01-01 03:00:00.000000000 +0300 +++ linux-2.6.22/fs/reiser4/dscale.h 2007-07-29 00:25:34.836686123 +0400 @@ -0,0 +1,27 @@ +/* Copyright 2001, 2002, 2003 by Hans Reiser, licensing governed by + * reiser4/README */ + +/* Scalable on-disk integers. See dscale.h for details. */ + +#if !defined( __FS_REISER4_DSCALE_H__ ) +#define __FS_REISER4_DSCALE_H__ + +#include "dformat.h" + +extern int dscale_read(unsigned char *address, __u64 * value); +extern int dscale_write(unsigned char *address, __u64 value); +extern int dscale_bytes(__u64 value); +extern int dscale_fit(__u64 value, __u64 other); + +/* __FS_REISER4_DSCALE_H__ */ +#endif + +/* Make Linus happy. + Local variables: + c-indentation-style: "K&R" + mode-name: "LC" + c-basic-offset: 8 + tab-width: 8 + fill-column: 120 + End: +*/ diff -urN linux-2.6.22.orig/fs/reiser4/entd.c linux-2.6.22/fs/reiser4/entd.c --- linux-2.6.22.orig/fs/reiser4/entd.c 1970-01-01 03:00:00.000000000 +0300 +++ linux-2.6.22/fs/reiser4/entd.c 2007-07-29 00:25:34.840687159 +0400 @@ -0,0 +1,335 @@ +/* Copyright 2003, 2004 by Hans Reiser, licensing governed by + * reiser4/README */ + +/* Ent daemon. */ + +#include "debug.h" +#include "txnmgr.h" +#include "tree.h" +#include "entd.h" +#include "super.h" +#include "context.h" +#include "reiser4.h" +#include "vfs_ops.h" +#include "page_cache.h" +#include "inode.h" + +#include /* struct task_struct */ +#include +#include +#include +#include /* INITIAL_JIFFIES */ +#include /* bdi_write_congested */ +#include +#include +#include + +#define DEF_PRIORITY 12 +#define MAX_ENTD_ITERS 10 + +static void entd_flush(struct super_block *, struct wbq *); +static int entd(void *arg); + +/* + * set ->comm field of end thread to make its state visible to the user level + */ +#define entd_set_comm(state) \ + snprintf(current->comm, sizeof(current->comm), \ + "ent:%s%s", super->s_id, (state)) + +/** + * reiser4_init_entd - initialize entd context and start kernel daemon + * @super: super block to start ent thread for + * + * Creates entd contexts, starts kernel thread and waits until it + * initializes. + */ +int reiser4_init_entd(struct super_block *super) +{ + entd_context *ctx; + + assert("nikita-3104", super != NULL); + + ctx = get_entd_context(super); + + memset(ctx, 0, sizeof *ctx); + spin_lock_init(&ctx->guard); + init_waitqueue_head(&ctx->wait); +#if REISER4_DEBUG + INIT_LIST_HEAD(&ctx->flushers_list); +#endif + /* lists of writepage requests */ + INIT_LIST_HEAD(&ctx->todo_list); + INIT_LIST_HEAD(&ctx->done_list); + /* start entd */ + ctx->tsk = kthread_run(entd, super, "ent:%s", super->s_id); + if (IS_ERR(ctx->tsk)) + return PTR_ERR(ctx->tsk); + return 0; +} + +static void put_wbq(struct wbq *rq) +{ + iput(rq->mapping->host); + complete(&rq->completion); +} + +/* ent should be locked */ +static struct wbq *__get_wbq(entd_context * ent) +{ + struct wbq *wbq; + + if (list_empty(&ent->todo_list)) + return NULL; + + ent->nr_todo_reqs --; + wbq = list_entry(ent->todo_list.next, struct wbq, link); + list_del_init(&wbq->link); + return wbq; +} + +/* ent thread function */ +static int entd(void *arg) +{ + struct super_block *super; + entd_context *ent; + int done = 0; + + super = arg; + /* do_fork() just copies task_struct into the new + thread. ->fs_context shouldn't be copied of course. This shouldn't + be a problem for the rest of the code though. + */ + current->journal_info = NULL; + + ent = get_entd_context(super); + + while (!done) { + try_to_freeze(); + + spin_lock(&ent->guard); + while (ent->nr_todo_reqs != 0) { + struct wbq *rq; + + assert("", list_empty(&ent->done_list)); + + /* take request from the queue head */ + rq = __get_wbq(ent); + assert("", rq != NULL); + ent->cur_request = rq; + spin_unlock(&ent->guard); + + entd_set_comm("!"); + entd_flush(super, rq); + + put_wbq(rq); + + /* + * wakeup all requestors and iput their inodes + */ + spin_lock(&ent->guard); + while (!list_empty(&ent->done_list)) { + rq = list_entry(ent->done_list.next, struct wbq, link); + list_del_init(&rq->link); + ent->nr_done_reqs --; + spin_unlock(&ent->guard); + assert("", rq->written == 1); + put_wbq(rq); + spin_lock(&ent->guard); + } + } + spin_unlock(&ent->guard); + + entd_set_comm("."); + + { + DEFINE_WAIT(__wait); + + do { + prepare_to_wait(&ent->wait, &__wait, TASK_INTERRUPTIBLE); + if (kthread_should_stop()) { + done = 1; + break; + } + if (ent->nr_todo_reqs != 0) + break; + schedule(); + } while (0); + finish_wait(&ent->wait, &__wait); + } + } + BUG_ON(ent->nr_todo_reqs != 0); + return 0; +} + +/** + * reiser4_done_entd - stop entd kernel thread + * @super: super block to stop ent thread for + * + * It is called on umount. Sends stop signal to entd and wait until it handles + * it. + */ +void reiser4_done_entd(struct super_block *super) +{ + entd_context *ent; + + assert("nikita-3103", super != NULL); + + ent = get_entd_context(super); + assert("zam-1055", ent->tsk != NULL); + kthread_stop(ent->tsk); +} + +/* called at the beginning of jnode_flush to register flusher thread with ent + * daemon */ +void reiser4_enter_flush(struct super_block *super) +{ + entd_context *ent; + + assert("zam-1029", super != NULL); + ent = get_entd_context(super); + + assert("zam-1030", ent != NULL); + + spin_lock(&ent->guard); + ent->flushers++; +#if REISER4_DEBUG + list_add(&get_current_context()->flushers_link, &ent->flushers_list); +#endif + spin_unlock(&ent->guard); +} + +/* called at the end of jnode_flush */ +void reiser4_leave_flush(struct super_block *super) +{ + entd_context *ent; + int wake_up_ent; + + assert("zam-1027", super != NULL); + ent = get_entd_context(super); + + assert("zam-1028", ent != NULL); + + spin_lock(&ent->guard); + ent->flushers--; + wake_up_ent = (ent->flushers == 0 && ent->nr_todo_reqs != 0); +#if REISER4_DEBUG + list_del_init(&get_current_context()->flushers_link); +#endif + spin_unlock(&ent->guard); + if (wake_up_ent) + wake_up(&ent->wait); +} + +#define ENTD_CAPTURE_APAGE_BURST SWAP_CLUSTER_MAX + +static void entd_flush(struct super_block *super, struct wbq *rq) +{ + reiser4_context ctx; + int tmp; + + init_stack_context(&ctx, super); + ctx.entd = 1; + ctx.gfp_mask = GFP_NOFS; + + rq->wbc->range_start = page_offset(rq->page); + rq->wbc->range_end = rq->wbc->range_start + + (ENTD_CAPTURE_APAGE_BURST << PAGE_CACHE_SHIFT); + tmp = rq->wbc->nr_to_write; + rq->mapping->a_ops->writepages(rq->mapping, rq->wbc); + + if (rq->wbc->nr_to_write > 0) { + rq->wbc->range_start = 0; + rq->wbc->range_end = LLONG_MAX; + generic_sync_sb_inodes(super, rq->wbc); + } + rq->wbc->nr_to_write = ENTD_CAPTURE_APAGE_BURST; + reiser4_writeout(super, rq->wbc); + + context_set_commit_async(&ctx); + reiser4_exit_context(&ctx); +} + +/** + * write_page_by_ent - ask entd thread to flush this page as part of slum + * @page: page to be written + * @wbc: writeback control passed to reiser4_writepage + * + * Creates a request, puts it on entd list of requests, wakeups entd if + * necessary, waits until entd completes with the request. + */ +int write_page_by_ent(struct page *page, struct writeback_control *wbc) +{ + struct super_block *sb; + struct inode *inode; + entd_context *ent; + struct wbq rq; + + assert("", PageLocked(page)); + assert("", page->mapping != NULL); + + sb = page->mapping->host->i_sb; + ent = get_entd_context(sb); + assert("", ent && ent->done == 0); + + /* + * we are going to unlock page and ask ent thread to write the + * page. Re-dirty page before unlocking so that if ent thread fails to + * write it - it will remain dirty + */ + reiser4_set_page_dirty_internal(page); + + /* + * pin inode in memory, unlock page, entd_flush will iput. We can not + * iput here becasue we can not allow delete_inode to be called here + */ + inode = igrab(page->mapping->host); + unlock_page(page); + if (inode == NULL) + /* inode is getting freed */ + return 0; + + /* init wbq */ + INIT_LIST_HEAD(&rq.link); + rq.magic = WBQ_MAGIC; + rq.wbc = wbc; + rq.page = page; + rq.mapping = inode->i_mapping; + rq.node = NULL; + rq.written = 0; + init_completion(&rq.completion); + + /* add request to entd's list of writepage requests */ + spin_lock(&ent->guard); + ent->nr_todo_reqs++; + list_add_tail(&rq.link, &ent->todo_list); + if (ent->nr_todo_reqs == 1) + wake_up(&ent->wait); + + spin_unlock(&ent->guard); + + /* wait until entd finishes */ + wait_for_completion(&rq.completion); + + if (rq.written) + /* Eventually ENTD has written the page to disk. */ + return 0; + return 0; +} + +int wbq_available(void) +{ + struct super_block *sb = reiser4_get_current_sb(); + entd_context *ent = get_entd_context(sb); + return ent->nr_todo_reqs; +} + +/* + * Local variables: + * c-indentation-style: "K&R" + * mode-name: "LC" + * c-basic-offset: 8 + * tab-width: 8 + * fill-column: 79 + * End: + */ diff -urN linux-2.6.22.orig/fs/reiser4/entd.h linux-2.6.22/fs/reiser4/entd.h --- linux-2.6.22.orig/fs/reiser4/entd.h 1970-01-01 03:00:00.000000000 +0300 +++ linux-2.6.22/fs/reiser4/entd.h 2007-07-29 00:25:34.840687159 +0400 @@ -0,0 +1,90 @@ +/* Copyright 2003 by Hans Reiser, licensing governed by reiser4/README */ + +/* Ent daemon. */ + +#ifndef __ENTD_H__ +#define __ENTD_H__ + +#include "context.h" + +#include +#include +#include +#include +#include /* for struct task_struct */ + +#define WBQ_MAGIC 0x7876dc76 + +/* write-back request. */ +struct wbq { + int magic; + struct list_head link; /* list head of this list is in entd context */ + struct writeback_control *wbc; + struct page *page; + struct address_space *mapping; + struct completion completion; + jnode *node; /* set if ent thread captured requested page */ + int written; /* set if ent thread wrote requested page */ +}; + +/* ent-thread context. This is used to synchronize starting/stopping ent + * threads. */ +typedef struct entd_context { + /* wait queue that ent thread waits on for more work. It's + * signaled by write_page_by_ent(). */ + wait_queue_head_t wait; + /* spinlock protecting other fields */ + spinlock_t guard; + /* ent thread */ + struct task_struct *tsk; + /* set to indicate that ent thread should leave. */ + int done; + /* counter of active flushers */ + int flushers; + /* + * when reiser4_writepage asks entd to write a page - it adds struct + * wbq to this list + */ + struct list_head todo_list; + /* number of elements on the above list */ + int nr_todo_reqs; + + struct wbq *cur_request; + /* + * when entd writes a page it moves write-back request from todo_list + * to done_list. This list is used at the end of entd iteration to + * wakeup requestors and iput inodes. + */ + struct list_head done_list; + /* number of elements on the above list */ + int nr_done_reqs; + +#if REISER4_DEBUG + /* list of all active flushers */ + struct list_head flushers_list; +#endif +} entd_context; + +extern int reiser4_init_entd(struct super_block *); +extern void reiser4_done_entd(struct super_block *); + +extern void reiser4_enter_flush(struct super_block *); +extern void reiser4_leave_flush(struct super_block *); + +extern int write_page_by_ent(struct page *, struct writeback_control *); +extern int wbq_available(void); +extern void ent_writes_page(struct super_block *, struct page *); + +extern jnode *get_jnode_by_wbq(struct super_block *, struct wbq *); +/* __ENTD_H__ */ +#endif + +/* Make Linus happy. + Local variables: + c-indentation-style: "K&R" + mode-name: "LC" + c-basic-offset: 8 + tab-width: 8 + fill-column: 120 + End: +*/ diff -urN linux-2.6.22.orig/fs/reiser4/eottl.c linux-2.6.22/fs/reiser4/eottl.c --- linux-2.6.22.orig/fs/reiser4/eottl.c 1970-01-01 03:00:00.000000000 +0300 +++ linux-2.6.22/fs/reiser4/eottl.c 2007-07-29 00:25:34.840687159 +0400 @@ -0,0 +1,509 @@ +/* Copyright 2001, 2002, 2003 by Hans Reiser, licensing governed by reiser4/README */ + +#include "forward.h" +#include "debug.h" +#include "key.h" +#include "coord.h" +#include "plugin/item/item.h" +#include "plugin/node/node.h" +#include "znode.h" +#include "block_alloc.h" +#include "tree_walk.h" +#include "tree_mod.h" +#include "carry.h" +#include "tree.h" +#include "super.h" + +#include /* for __u?? */ + +/* + * Extents on the twig level (EOTTL) handling. + * + * EOTTL poses some problems to the tree traversal, that are better explained + * by example. + * + * Suppose we have block B1 on the twig level with the following items: + * + * 0. internal item I0 with key (0:0:0:0) (locality, key-type, object-id, + * offset) + * 1. extent item E1 with key (1:4:100:0), having 10 blocks of 4k each + * 2. internal item I2 with key (10:0:0:0) + * + * We are trying to insert item with key (5:0:0:0). Lookup finds node B1, and + * then intra-node lookup is done. This lookup finished on the E1, because the + * key we are looking for is larger than the key of E1 and is smaller than key + * the of I2. + * + * Here search is stuck. + * + * After some thought it is clear what is wrong here: extents on the twig level + * break some basic property of the *search* tree (on the pretext, that they + * restore property of balanced tree). + * + * Said property is the following: if in the internal node of the search tree + * we have [ ... Key1 Pointer Key2 ... ] then, all data that are or will be + * keyed in the tree with the Key such that Key1 <= Key < Key2 are accessible + * through the Pointer. + * + * This is not true, when Pointer is Extent-Pointer, simply because extent + * cannot expand indefinitely to the right to include any item with + * + * Key1 <= Key <= Key2. + * + * For example, our E1 extent is only responsible for the data with keys + * + * (1:4:100:0) <= key <= (1:4:100:0xffffffffffffffff), and + * + * so, key range + * + * ( (1:4:100:0xffffffffffffffff), (10:0:0:0) ) + * + * is orphaned: there is no way to get there from the tree root. + * + * In other words, extent pointers are different than normal child pointers as + * far as search tree is concerned, and this creates such problems. + * + * Possible solution for this problem is to insert our item into node pointed + * to by I2. There are some problems through: + * + * (1) I2 can be in a different node. + * (2) E1 can be immediately followed by another extent E2. + * + * (1) is solved by calling reiser4_get_right_neighbor() and accounting + * for locks/coords as necessary. + * + * (2) is more complex. Solution here is to insert new empty leaf node and + * insert internal item between E1 and E2 pointing to said leaf node. This is + * further complicated by possibility that E2 is in a different node, etc. + * + * Problems: + * + * (1) if there was internal item I2 immediately on the right of an extent E1 + * we and we decided to insert new item S1 into node N2 pointed to by I2, then + * key of S1 will be less than smallest key in the N2. Normally, search key + * checks that key we are looking for is in the range of keys covered by the + * node key is being looked in. To work around of this situation, while + * preserving useful consistency check new flag CBK_TRUST_DK was added to the + * cbk falgs bitmask. This flag is automatically set on entrance to the + * coord_by_key() and is only cleared when we are about to enter situation + * described above. + * + * (2) If extent E1 is immediately followed by another extent E2 and we are + * searching for the key that is between E1 and E2 we only have to insert new + * empty leaf node when coord_by_key was called for insertion, rather than just + * for lookup. To distinguish these cases, new flag CBK_FOR_INSERT was added to + * the cbk falgs bitmask. This flag is automatically set by coord_by_key calls + * performed by insert_by_key() and friends. + * + * (3) Insertion of new empty leaf node (possibly) requires balancing. In any + * case it requires modification of node content which is only possible under + * write lock. It may well happen that we only have read lock on the node where + * new internal pointer is to be inserted (common case: lookup of non-existent + * stat-data that fells between two extents). If only read lock is held, tree + * traversal is restarted with lock_level modified so that next time we hit + * this problem, write lock will be held. Once we have write lock, balancing + * will be performed. + */ + +/** + * is_next_item_internal - check whether next item is internal + * @coord: coordinate of extent item in twig node + * @key: search key + * @lh: twig node lock handle + * + * Looks at the unit next to @coord. If it is an internal one - 1 is returned, + * @coord is set to that unit. If that unit is in right neighbor, @lh is moved + * to that node, @coord is set to its first unit. If next item is not internal + * or does not exist then 0 is returned, @coord and @lh are left unchanged. 2 + * is returned if search restart has to be done. + */ +static int +is_next_item_internal(coord_t *coord, const reiser4_key *key, + lock_handle *lh) +{ + coord_t next; + lock_handle rn; + int result; + + coord_dup(&next, coord); + if (coord_next_unit(&next) == 0) { + /* next unit is in this node */ + if (item_is_internal(&next)) { + coord_dup(coord, &next); + return 1; + } + assert("vs-3", item_is_extent(&next)); + return 0; + } + + /* + * next unit either does not exist or is in right neighbor. If it is in + * right neighbor we have to check right delimiting key because + * concurrent thread could get their first and insert item with a key + * smaller than @key + */ + read_lock_dk(current_tree); + result = keycmp(key, znode_get_rd_key(coord->node)); + read_unlock_dk(current_tree); + assert("vs-6", result != EQUAL_TO); + if (result == GREATER_THAN) + return 2; + + /* lock right neighbor */ + init_lh(&rn); + result = reiser4_get_right_neighbor(&rn, coord->node, + znode_is_wlocked(coord->node) ? + ZNODE_WRITE_LOCK : ZNODE_READ_LOCK, + GN_CAN_USE_UPPER_LEVELS); + if (result == -E_NO_NEIGHBOR) { + /* we are on the rightmost edge of the tree */ + done_lh(&rn); + return 0; + } + + if (result) { + assert("vs-4", result < 0); + done_lh(&rn); + return result; + } + + /* + * check whether concurrent thread managed to insert item with a key + * smaller than @key + */ + read_lock_dk(current_tree); + result = keycmp(key, znode_get_ld_key(rn.node)); + read_unlock_dk(current_tree); + assert("vs-6", result != EQUAL_TO); + if (result == GREATER_THAN) { + done_lh(&rn); + return 2; + } + + result = zload(rn.node); + if (result) { + assert("vs-5", result < 0); + done_lh(&rn); + return result; + } + + coord_init_first_unit(&next, rn.node); + if (item_is_internal(&next)) { + /* + * next unit is in right neighbor and it is an unit of internal + * item. Unlock coord->node. Move @lh to right neighbor. @coord + * is set to the first unit of right neighbor. + */ + coord_dup(coord, &next); + zrelse(rn.node); + done_lh(lh); + move_lh(lh, &rn); + return 1; + } + + /* + * next unit is unit of extent item. Return without chaning @lh and + * @coord. + */ + assert("vs-6", item_is_extent(&next)); + zrelse(rn.node); + done_lh(&rn); + return 0; +} + +/** + * rd_key - calculate key of an item next to the given one + * @coord: position in a node + * @key: storage for result key + * + * @coord is set between items or after the last item in a node. Calculate key + * of item to the right of @coord. + */ +static reiser4_key *rd_key(const coord_t *coord, reiser4_key *key) +{ + coord_t dup; + + assert("nikita-2281", coord_is_between_items(coord)); + coord_dup(&dup, coord); + + if (coord_set_to_right(&dup) == 0) + /* next item is in this node. Return its key. */ + unit_key_by_coord(&dup, key); + else { + /* + * next item either does not exist or is in right + * neighbor. Return znode's right delimiting key. + */ + read_lock_dk(current_tree); + *key = *znode_get_rd_key(coord->node); + read_unlock_dk(current_tree); + } + return key; +} + +/** + * add_empty_leaf - insert empty leaf between two extents + * @insert_coord: position in twig node between two extents + * @lh: twig node lock handle + * @key: left delimiting key of new node + * @rdkey: right delimiting key of new node + * + * Inserts empty leaf node between two extent items. It is necessary when we + * have to insert an item on leaf level between two extents (items on the twig + * level). + */ +static int +add_empty_leaf(coord_t *insert_coord, lock_handle *lh, + const reiser4_key *key, const reiser4_key *rdkey) +{ + int result; + carry_pool *pool; + carry_level *todo; + reiser4_item_data *item; + carry_insert_data *cdata; + carry_op *op; + znode *node; + reiser4_tree *tree; + + assert("vs-49827", znode_contains_key_lock(insert_coord->node, key)); + tree = znode_get_tree(insert_coord->node); + node = reiser4_new_node(insert_coord->node, LEAF_LEVEL); + if (IS_ERR(node)) + return PTR_ERR(node); + + /* setup delimiting keys for node being inserted */ + write_lock_dk(tree); + znode_set_ld_key(node, key); + znode_set_rd_key(node, rdkey); + ON_DEBUG(node->creator = current); + ON_DEBUG(node->first_key = *key); + write_unlock_dk(tree); + + ZF_SET(node, JNODE_ORPHAN); + + /* + * allocate carry_pool, 3 carry_level-s, reiser4_item_data and + * carry_insert_data + */ + pool = init_carry_pool(sizeof(*pool) + 3 * sizeof(*todo) + + sizeof(*item) + sizeof(*cdata)); + if (IS_ERR(pool)) + return PTR_ERR(pool); + todo = (carry_level *) (pool + 1); + init_carry_level(todo, pool); + + item = (reiser4_item_data *) (todo + 3); + cdata = (carry_insert_data *) (item + 1); + + op = reiser4_post_carry(todo, COP_INSERT, insert_coord->node, 0); + if (!IS_ERR(op)) { + cdata->coord = insert_coord; + cdata->key = key; + cdata->data = item; + op->u.insert.d = cdata; + op->u.insert.type = COPT_ITEM_DATA; + build_child_ptr_data(node, item); + item->arg = NULL; + /* have @insert_coord to be set at inserted item after + insertion is done */ + todo->track_type = CARRY_TRACK_CHANGE; + todo->tracked = lh; + + result = reiser4_carry(todo, NULL); + if (result == 0) { + /* + * pin node in memory. This is necessary for + * znode_make_dirty() below. + */ + result = zload(node); + if (result == 0) { + lock_handle local_lh; + + /* + * if we inserted new child into tree we have + * to mark it dirty so that flush will be able + * to process it. + */ + init_lh(&local_lh); + result = longterm_lock_znode(&local_lh, node, + ZNODE_WRITE_LOCK, + ZNODE_LOCK_LOPRI); + if (result == 0) { + znode_make_dirty(node); + + /* + * when internal item pointing to @node + * was inserted into twig node + * create_hook_internal did not connect + * it properly because its right + * neighbor was not known. Do it + * here + */ + write_lock_tree(tree); + assert("nikita-3312", + znode_is_right_connected(node)); + assert("nikita-2984", + node->right == NULL); + ZF_CLR(node, JNODE_RIGHT_CONNECTED); + write_unlock_tree(tree); + result = + connect_znode(insert_coord, node); + ON_DEBUG(if (result == 0) check_dkeys(node);); + + done_lh(lh); + move_lh(lh, &local_lh); + assert("vs-1676", node_is_empty(node)); + coord_init_first_unit(insert_coord, + node); + } else { + warning("nikita-3136", + "Cannot lock child"); + } + done_lh(&local_lh); + zrelse(node); + } + } + } else + result = PTR_ERR(op); + zput(node); + done_carry_pool(pool); + return result; +} + +/** + * handle_eottl - handle extent-on-the-twig-level cases in tree traversal + * @h: search handle + * @outcome: flag saying whether search has to restart or is done + * + * Handles search on twig level. If this function completes search itself then + * it returns 1. If search has to go one level down then 0 is returned. If + * error happens then LOOKUP_DONE is returned via @outcome and error code is saved + * in @h->result. + */ +int handle_eottl(cbk_handle *h, int *outcome) +{ + int result; + reiser4_key key; + coord_t *coord; + + coord = h->coord; + + if (h->level != TWIG_LEVEL || + (coord_is_existing_item(coord) && item_is_internal(coord))) { + /* Continue to traverse tree downward. */ + return 0; + } + + /* + * make sure that @h->coord is set to twig node and that it is either + * set to extent item or after extent item + */ + assert("vs-356", h->level == TWIG_LEVEL); + assert("vs-357", ( { + coord_t lcoord; + coord_dup(&lcoord, coord); + check_me("vs-733", coord_set_to_left(&lcoord) == 0); + item_is_extent(&lcoord); + } + )); + + if (*outcome == NS_FOUND) { + /* we have found desired key on twig level in extent item */ + h->result = CBK_COORD_FOUND; + *outcome = LOOKUP_DONE; + return 1; + } + + if (!(h->flags & CBK_FOR_INSERT)) { + /* tree traversal is not for insertion. Just return + CBK_COORD_NOTFOUND. */ + h->result = CBK_COORD_NOTFOUND; + *outcome = LOOKUP_DONE; + return 1; + } + + /* take a look at the item to the right of h -> coord */ + result = is_next_item_internal(coord, h->key, h->active_lh); + if (unlikely(result < 0)) { + h->error = "get_right_neighbor failed"; + h->result = result; + *outcome = LOOKUP_DONE; + return 1; + } + if (result == 0) { + /* + * item to the right is also an extent one. Allocate a new node + * and insert pointer to it after item h -> coord. + * + * This is a result of extents being located at the twig + * level. For explanation, see comment just above + * is_next_item_internal(). + */ + znode *loaded; + + if (cbk_lock_mode(h->level, h) != ZNODE_WRITE_LOCK) { + /* + * we got node read locked, restart coord_by_key to + * have write lock on twig level + */ + h->lock_level = TWIG_LEVEL; + h->lock_mode = ZNODE_WRITE_LOCK; + *outcome = LOOKUP_REST; + return 1; + } + + loaded = coord->node; + result = + add_empty_leaf(coord, h->active_lh, h->key, + rd_key(coord, &key)); + if (result) { + h->error = "could not add empty leaf"; + h->result = result; + *outcome = LOOKUP_DONE; + return 1; + } + /* added empty leaf is locked (h->active_lh), its parent node + is unlocked, h->coord is set as EMPTY */ + assert("vs-13", coord->between == EMPTY_NODE); + assert("vs-14", znode_is_write_locked(coord->node)); + assert("vs-15", + WITH_DATA(coord->node, node_is_empty(coord->node))); + assert("vs-16", jnode_is_leaf(ZJNODE(coord->node))); + assert("vs-17", coord->node == h->active_lh->node); + *outcome = LOOKUP_DONE; + h->result = CBK_COORD_NOTFOUND; + return 1; + } else if (result == 1) { + /* + * this is special case mentioned in the comment on + * tree.h:cbk_flags. We have found internal item immediately on + * the right of extent, and we are going to insert new item + * there. Key of item we are going to insert is smaller than + * leftmost key in the node pointed to by said internal item + * (otherwise search wouldn't come to the extent in the first + * place). + * + * This is a result of extents being located at the twig + * level. For explanation, see comment just above + * is_next_item_internal(). + */ + h->flags &= ~CBK_TRUST_DK; + } else { + assert("vs-8", result == 2); + *outcome = LOOKUP_REST; + return 1; + } + assert("vs-362", WITH_DATA(coord->node, item_is_internal(coord))); + return 0; +} + +/* + * Local variables: + * c-indentation-style: "K&R" + * mode-name: "LC" + * c-basic-offset: 8 + * tab-width: 8 + * fill-column: 120 + * scroll-step: 1 + * End: + */ diff -urN linux-2.6.22.orig/fs/reiser4/estimate.c linux-2.6.22/fs/reiser4/estimate.c --- linux-2.6.22.orig/fs/reiser4/estimate.c 1970-01-01 03:00:00.000000000 +0300 +++ linux-2.6.22/fs/reiser4/estimate.c 2007-07-29 00:25:34.840687159 +0400 @@ -0,0 +1,120 @@ +/* Copyright 2001, 2002, 2003 by Hans Reiser, licensing governed by reiser4/README */ + +#include "debug.h" +#include "dformat.h" +#include "tree.h" +#include "carry.h" +#include "inode.h" +#include "plugin/cluster.h" +#include "plugin/item/ctail.h" + +/* this returns how many nodes might get dirty and added nodes if @children nodes are dirtied + + Amount of internals which will get dirty or get allocated we estimate as 5% of the childs + 1 balancing. 1 balancing + is 2 neighbours, 2 new blocks and the current block on the leaf level, 2 neighbour nodes + the current (or 1 + neighbour and 1 new and the current) on twig level, 2 neighbour nodes on upper levels and 1 for a new root. So 5 for + leaf level, 3 for twig level, 2 on upper + 1 for root. + + Do not calculate the current node of the lowest level here - this is overhead only. + + children is almost always 1 here. Exception is flow insertion +*/ +static reiser4_block_nr +max_balance_overhead(reiser4_block_nr childen, tree_level tree_height) +{ + reiser4_block_nr ten_percent; + + ten_percent = ((103 * childen) >> 10); + + /* If we have too many balancings at the time, tree height can raise on more + then 1. Assume that if tree_height is 5, it can raise on 1 only. */ + return ((tree_height < 5 ? 5 : tree_height) * 2 + (4 + ten_percent)); +} + +/* this returns maximal possible number of nodes which can be modified plus number of new nodes which can be required to + perform insertion of one item into the tree */ +/* it is only called when tree height changes, or gets initialized */ +reiser4_block_nr calc_estimate_one_insert(tree_level height) +{ + return 1 + max_balance_overhead(1, height); +} + +reiser4_block_nr estimate_one_insert_item(reiser4_tree * tree) +{ + return tree->estimate_one_insert; +} + +/* this returns maximal possible number of nodes which can be modified plus number of new nodes which can be required to + perform insertion of one unit into an item in the tree */ +reiser4_block_nr estimate_one_insert_into_item(reiser4_tree * tree) +{ + /* estimate insert into item just like item insertion */ + return tree->estimate_one_insert; +} + +reiser4_block_nr estimate_one_item_removal(reiser4_tree * tree) +{ + /* on item removal reiser4 does not try to pack nodes more complact, so, only one node may be dirtied on leaf + level */ + return tree->estimate_one_insert; +} + +/* on leaf level insert_flow may add CARRY_FLOW_NEW_NODES_LIMIT new nodes and dirty 3 existing nodes (insert point and + both its neighbors). Max_balance_overhead should estimate number of blocks which may change/get added on internal + levels */ +reiser4_block_nr estimate_insert_flow(tree_level height) +{ + return 3 + CARRY_FLOW_NEW_NODES_LIMIT + max_balance_overhead(3 + + CARRY_FLOW_NEW_NODES_LIMIT, + height); +} + +/* returnes max number of nodes can be occupied by disk cluster */ +static reiser4_block_nr estimate_cluster(struct inode * inode, int unprepped) +{ + int per_cluster; + per_cluster = (unprepped ? 1 : cluster_nrpages(inode)); + return 3 + per_cluster + + max_balance_overhead(3 + per_cluster, + REISER4_MAX_ZTREE_HEIGHT); +} + +/* how many nodes might get dirty and added + during insertion of a disk cluster */ +reiser4_block_nr estimate_insert_cluster(struct inode * inode) +{ + return estimate_cluster(inode, 1); /* 24 */ +} + +/* how many nodes might get dirty and added + during update of a (prepped or unprepped) disk cluster */ +reiser4_block_nr estimate_update_cluster(struct inode * inode) +{ + return estimate_cluster(inode, 0); /* 44, for 64K-cluster */ +} + +/* How many nodes occupied by a disk cluster might get dirty. + Note that this estimation is not precise (i.e. disk cluster + can occupy more nodes). + Q: Why we don't use precise estimation? + A: 1.Because precise estimation is fairly bad: 65536 nodes + for 64K logical cluster, it means 256M of dead space on + a partition + 2.It is a very rare case when disk cluster occupies more + nodes then this estimation returns. +*/ +reiser4_block_nr estimate_dirty_cluster(struct inode * inode) +{ + return cluster_nrpages(inode) + 4; +} + +/* Make Linus happy. + Local variables: + c-indentation-style: "K&R" + mode-name: "LC" + c-basic-offset: 8 + tab-width: 8 + fill-column: 120 + scroll-step: 1 + End: +*/ diff -urN linux-2.6.22.orig/fs/reiser4/export_ops.c linux-2.6.22/fs/reiser4/export_ops.c --- linux-2.6.22.orig/fs/reiser4/export_ops.c 1970-01-01 03:00:00.000000000 +0300 +++ linux-2.6.22/fs/reiser4/export_ops.c 2007-07-29 00:25:34.840687159 +0400 @@ -0,0 +1,295 @@ +/* Copyright 2005 by Hans Reiser, licensing governed by + * reiser4/README */ + +#include "inode.h" +#include "plugin/plugin.h" + +/* + * Supported file-handle types + */ +typedef enum { + FH_WITH_PARENT = 0x10, /* file handle with parent */ + FH_WITHOUT_PARENT = 0x11 /* file handle without parent */ +} reiser4_fhtype; + +#define NFSERROR (255) + +/* initialize place-holder for object */ +static void object_on_wire_init(reiser4_object_on_wire *o) +{ + o->plugin = NULL; +} + +/* finish with @o */ +static void object_on_wire_done(reiser4_object_on_wire *o) +{ + if (o->plugin != NULL) + o->plugin->wire.done(o); +} + +/* + * read serialized object identity from @addr and store information about + * object in @obj. This is dual to encode_inode(). + */ +static char *decode_inode(struct super_block *s, char *addr, + reiser4_object_on_wire * obj) +{ + file_plugin *fplug; + + /* identifier of object plugin is stored in the first two bytes, + * followed by... */ + fplug = file_plugin_by_disk_id(reiser4_get_tree(s), (d16 *) addr); + if (fplug != NULL) { + addr += sizeof(d16); + obj->plugin = fplug; + assert("nikita-3520", fplug->wire.read != NULL); + /* plugin specific encoding of object identity. */ + addr = fplug->wire.read(addr, obj); + } else + addr = ERR_PTR(RETERR(-EINVAL)); + return addr; +} + +/** + * reiser4_decode_fh - decode_fh of export operations + * @super: super block + * @fh: nfsd file handle + * @len: length of file handle + * @fhtype: type of file handle + * @acceptable: acceptability testing function + * @context: argument for @acceptable + * + * Returns dentry referring to the same file as @fh. + */ +static struct dentry *reiser4_decode_fh(struct super_block *super, __u32 *fh, + int len, int fhtype, + int (*acceptable) (void *context, + struct dentry *de), + void *context) +{ + reiser4_context *ctx; + reiser4_object_on_wire object; + reiser4_object_on_wire parent; + char *addr; + int with_parent; + + ctx = reiser4_init_context(super); + if (IS_ERR(ctx)) + return (struct dentry *)ctx; + + assert("vs-1482", + fhtype == FH_WITH_PARENT || fhtype == FH_WITHOUT_PARENT); + + with_parent = (fhtype == FH_WITH_PARENT); + + addr = (char *)fh; + + object_on_wire_init(&object); + object_on_wire_init(&parent); + + addr = decode_inode(super, addr, &object); + if (!IS_ERR(addr)) { + if (with_parent) + addr = decode_inode(super, addr, &parent); + if (!IS_ERR(addr)) { + struct dentry *d; + typeof(super->s_export_op->find_exported_dentry) fn; + + fn = super->s_export_op->find_exported_dentry; + assert("nikita-3521", fn != NULL); + d = fn(super, &object, with_parent ? &parent : NULL, + acceptable, context); + if (d != NULL && !IS_ERR(d)) + /* FIXME check for -ENOMEM */ + reiser4_get_dentry_fsdata(d)->stateless = 1; + addr = (char *)d; + } + } + + object_on_wire_done(&object); + object_on_wire_done(&parent); + + reiser4_exit_context(ctx); + return (void *)addr; +} + +/* + * Object serialization support. + * + * To support knfsd file system provides export_operations that are used to + * construct and interpret NFS file handles. As a generalization of this, + * reiser4 object plugins have serialization support: it provides methods to + * create on-wire representation of identity of reiser4 object, and + * re-create/locate object given its on-wire identity. + * + */ + +/* + * return number of bytes that on-wire representation of @inode's identity + * consumes. + */ +static int encode_inode_size(struct inode *inode) +{ + assert("nikita-3514", inode != NULL); + assert("nikita-3515", inode_file_plugin(inode) != NULL); + assert("nikita-3516", inode_file_plugin(inode)->wire.size != NULL); + + return inode_file_plugin(inode)->wire.size(inode) + sizeof(d16); +} + +/* + * store on-wire representation of @inode's identity at the area beginning at + * @start. + */ +static char *encode_inode(struct inode *inode, char *start) +{ + assert("nikita-3517", inode != NULL); + assert("nikita-3518", inode_file_plugin(inode) != NULL); + assert("nikita-3519", inode_file_plugin(inode)->wire.write != NULL); + + /* + * first, store two-byte identifier of object plugin, then + */ + save_plugin_id(file_plugin_to_plugin(inode_file_plugin(inode)), + (d16 *) start); + start += sizeof(d16); + /* + * call plugin to serialize object's identity + */ + return inode_file_plugin(inode)->wire.write(inode, start); +} + +/* this returns number of 32 bit long numbers encoded in @lenp. 255 is + * returned if file handle can not be stored */ +/** + * reiser4_encode_fh - encode_fh of export operations + * @dentry: + * @fh: + * @lenp: + * @need_parent: + * + */ +static int +reiser4_encode_fh(struct dentry *dentry, __u32 *fh, int *lenp, + int need_parent) +{ + struct inode *inode; + struct inode *parent; + char *addr; + int need; + int delta; + int result; + reiser4_context *ctx; + + /* + * knfsd asks as to serialize object in @dentry, and, optionally its + * parent (if need_parent != 0). + * + * encode_inode() and encode_inode_size() is used to build + * representation of object and its parent. All hard work is done by + * object plugins. + */ + inode = dentry->d_inode; + parent = dentry->d_parent->d_inode; + + addr = (char *)fh; + + need = encode_inode_size(inode); + if (need < 0) + return NFSERROR; + if (need_parent) { + delta = encode_inode_size(parent); + if (delta < 0) + return NFSERROR; + need += delta; + } + + ctx = reiser4_init_context(dentry->d_inode->i_sb); + if (IS_ERR(ctx)) + return PTR_ERR(ctx); + + if (need <= sizeof(__u32) * (*lenp)) { + addr = encode_inode(inode, addr); + if (need_parent) + addr = encode_inode(parent, addr); + + /* store in lenp number of 32bit words required for file + * handle. */ + *lenp = (need + sizeof(__u32) - 1) >> 2; + result = need_parent ? FH_WITH_PARENT : FH_WITHOUT_PARENT; + } else + /* no enough space in file handle */ + result = NFSERROR; + reiser4_exit_context(ctx); + return result; +} + +/** + * reiser4_get_dentry_parent - get_parent of export operations + * @child: + * + */ +static struct dentry *reiser4_get_dentry_parent(struct dentry *child) +{ + struct inode *dir; + dir_plugin *dplug; + + assert("nikita-3527", child != NULL); + /* see comment in reiser4_get_dentry() about following assertion */ + assert("nikita-3528", is_in_reiser4_context()); + + dir = child->d_inode; + assert("nikita-3529", dir != NULL); + dplug = inode_dir_plugin(dir); + assert("nikita-3531", ergo(dplug != NULL, dplug->get_parent != NULL)); + if (dplug != NULL) + return dplug->get_parent(dir); + else + return ERR_PTR(RETERR(-ENOTDIR)); +} + +/** + * reiser4_get_dentry - get_dentry of export operations + * @super: + * @data: + * + * + */ +static struct dentry *reiser4_get_dentry(struct super_block *super, void *data) +{ + reiser4_object_on_wire *o; + + assert("nikita-3522", super != NULL); + assert("nikita-3523", data != NULL); + /* + * this is only supposed to be called by + * + * reiser4_decode_fh->find_exported_dentry + * + * so, reiser4_context should be here already. + */ + assert("nikita-3526", is_in_reiser4_context()); + + o = (reiser4_object_on_wire *)data; + assert("nikita-3524", o->plugin != NULL); + assert("nikita-3525", o->plugin->wire.get != NULL); + + return o->plugin->wire.get(super, o); +} + +struct export_operations reiser4_export_operations = { + .encode_fh = reiser4_encode_fh, + .decode_fh = reiser4_decode_fh, + .get_parent = reiser4_get_dentry_parent, + .get_dentry = reiser4_get_dentry +}; + +/* + * Local variables: + * c-indentation-style: "K&R" + * mode-name: "LC" + * c-basic-offset: 8 + * tab-width: 8 + * fill-column: 79 + * End: + */ diff -urN linux-2.6.22.orig/fs/reiser4/flush.c linux-2.6.22/fs/reiser4/flush.c --- linux-2.6.22.orig/fs/reiser4/flush.c 1970-01-01 03:00:00.000000000 +0300 +++ linux-2.6.22/fs/reiser4/flush.c 2007-07-29 00:25:34.000000000 +0400 @@ -0,0 +1,3625 @@ +/* Copyright 2001, 2002, 2003 by Hans Reiser, licensing governed by reiser4/README */ + +/* The design document for this file is at http://www.namesys.com/v4/v4.html. */ + +#include "forward.h" +#include "debug.h" +#include "dformat.h" +#include "key.h" +#include "coord.h" +#include "plugin/item/item.h" +#include "plugin/plugin.h" +#include "plugin/object.h" +#include "txnmgr.h" +#include "jnode.h" +#include "znode.h" +#include "block_alloc.h" +#include "tree_walk.h" +#include "carry.h" +#include "tree.h" +#include "vfs_ops.h" +#include "inode.h" +#include "page_cache.h" +#include "wander.h" +#include "super.h" +#include "entd.h" +#include "reiser4.h" +#include "flush.h" +#include "writeout.h" + +#include +#include /* for struct super_block */ +#include /* for struct page */ +#include /* for struct bio */ +#include +#include + +/* IMPLEMENTATION NOTES */ + +/* PARENT-FIRST: Some terminology: A parent-first traversal is a way of assigning a total + order to the nodes of the tree in which the parent is placed before its children, which + are ordered (recursively) in left-to-right order. When we speak of a "parent-first preceder", it + describes the node that "came before in forward parent-first order". When we speak of a + "parent-first follower", it describes the node that "comes next in parent-first + order" (alternatively the node that "came before in reverse parent-first order"). + + The following pseudo-code prints the nodes of a tree in forward parent-first order: + + void parent_first (node) + { + print_node (node); + if (node->level > leaf) { + for (i = 0; i < num_children; i += 1) { + parent_first (node->child[i]); + } + } + } +*/ + +/* JUST WHAT ARE WE TRYING TO OPTIMIZE, HERE? The idea is to optimize block allocation so + that a left-to-right scan of the tree's data (i.e., the leaves in left-to-right order) + can be accomplished with sequential reads, which results in reading nodes in their + parent-first order. This is a read-optimization aspect of the flush algorithm, and + there is also a write-optimization aspect, which is that we wish to make large + sequential writes to the disk by allocating or reallocating blocks so that they can be + written in sequence. Sometimes the read-optimization and write-optimization goals + conflict with each other, as we discuss in more detail below. +*/ + +/* STATE BITS: The flush code revolves around the state of the jnodes it covers. Here are + the relevant jnode->state bits and their relevence to flush: + + JNODE_DIRTY: If a node is dirty, it must be flushed. But in order to be written it + must be allocated first. In order to be considered allocated, the jnode must have + exactly one of { JNODE_OVRWR, JNODE_RELOC } set. These two bits are exclusive, and + all dirtied jnodes eventually have one of these bits set during each transaction. + + JNODE_CREATED: The node was freshly created in its transaction and has no previous + block address, so it is unconditionally assigned to be relocated, although this is + mainly for code-convenience. It is not being 'relocated' from anything, but in + almost every regard it is treated as part of the relocate set. The JNODE_CREATED bit + remains set even after JNODE_RELOC is set, so the actual relocate can be + distinguished from the created-and-allocated set easily: relocate-set members + (belonging to the preserve-set) have (JNODE_RELOC) set and created-set members which + have no previous location to preserve have (JNODE_RELOC | JNODE_CREATED) set. + + JNODE_OVRWR: The node belongs to atom's overwrite set. The flush algorithm made the + decision to maintain the pre-existing location for this node and it will be written + to the wandered-log. + + JNODE_RELOC: The flush algorithm made the decision to relocate this block (if it was + not created, see note above). A block with JNODE_RELOC set is eligible for + early-flushing and may be submitted during flush_empty_queues. When the JNODE_RELOC + bit is set on a znode, the parent node's internal item is modified and the znode is + rehashed. + + JNODE_SQUEEZABLE: Before shifting everything left, the flush algorithm scans the node + and calls plugin->f.squeeze() method for its items. By this technology we update disk + clusters of cryptcompress objects. Also if leftmost point that was found by flush scan + has this flag (races with write(), rare case) the flush algorythm makes the decision + to pass it to squalloc() in spite of its flushprepped status for squeezing, not for + repeated allocation. + + JNODE_FLUSH_QUEUED: This bit is set when a call to flush enters the jnode into its + flush queue. This means the jnode is not on any clean or dirty list, instead it is + moved to one of the flush queue (see flush_queue.h) object private list. This + prevents multiple concurrent flushes from attempting to start flushing from the + same node. + + (DEAD STATE BIT) JNODE_FLUSH_BUSY: This bit was set during the bottom-up + squeeze-and-allocate on a node while its children are actively being squeezed and + allocated. This flag was created to avoid submitting a write request for a node + while its children are still being allocated and squeezed. Then flush queue was + re-implemented to allow unlimited number of nodes be queued. This flag support was + commented out in source code because we decided that there was no reason to submit + queued nodes before jnode_flush() finishes. However, current code calls fq_write() + during a slum traversal and may submit "busy nodes" to disk. Probably we can + re-enable the JNODE_FLUSH_BUSY bit support in future. + + With these state bits, we describe a test used frequently in the code below, + jnode_is_flushprepped() (and the spin-lock-taking jnode_check_flushprepped()). The + test for "flushprepped" returns true if any of the following are true: + + - The node is not dirty + - The node has JNODE_RELOC set + - The node has JNODE_OVRWR set + + If either the node is not dirty or it has already been processed by flush (and assigned + JNODE_OVRWR or JNODE_RELOC), then it is prepped. If jnode_is_flushprepped() returns + true then flush has work to do on that node. +*/ + +/* FLUSH_PREP_ONCE_PER_TRANSACTION: Within a single transaction a node is never + flushprepped twice (unless an explicit call to flush_unprep is made as described in + detail below). For example a node is dirtied, allocated, and then early-flushed to + disk and set clean. Before the transaction commits, the page is dirtied again and, due + to memory pressure, the node is flushed again. The flush algorithm will not relocate + the node to a new disk location, it will simply write it to the same, previously + relocated position again. +*/ + +/* THE BOTTOM-UP VS. TOP-DOWN ISSUE: This code implements a bottom-up algorithm where we + start at a leaf node and allocate in parent-first order by iterating to the right. At + each step of the iteration, we check for the right neighbor. Before advancing to the + right neighbor, we check if the current position and the right neighbor share the same + parent. If they do not share the same parent, the parent is allocated before the right + neighbor. + + This process goes recursively up the tree and squeeze nodes level by level as long as + the right neighbor and the current position have different parents, then it allocates + the right-neighbors-with-different-parents on the way back down. This process is + described in more detail in flush_squalloc_changed_ancestor and the recursive function + squalloc_one_changed_ancestor. But the purpose here is not to discuss the + specifics of the bottom-up approach as it is to contrast the bottom-up and top-down + approaches. + + The top-down algorithm was implemented earlier (April-May 2002). In the top-down + approach, we find a starting point by scanning left along each level past dirty nodes, + then going up and repeating the process until the left node and the parent node are + clean. We then perform a parent-first traversal from the starting point, which makes + allocating in parent-first order trivial. After one subtree has been allocated in this + manner, we move to the right, try moving upward, then repeat the parent-first + traversal. + + Both approaches have problems that need to be addressed. Both are approximately the + same amount of code, but the bottom-up approach has advantages in the order it acquires + locks which, at the very least, make it the better approach. At first glance each one + makes the other one look simpler, so it is important to remember a few of the problems + with each one. + + Main problem with the top-down approach: When you encounter a clean child during the + parent-first traversal, what do you do? You would like to avoid searching through a + large tree of nodes just to find a few dirty leaves at the bottom, and there is not an + obvious solution. One of the advantages of the top-down approach is that during the + parent-first traversal you check every child of a parent to see if it is dirty. In + this way, the top-down approach easily handles the main problem of the bottom-up + approach: unallocated children. + + The unallocated children problem is that before writing a node to disk we must make + sure that all of its children are allocated. Otherwise, the writing the node means + extra I/O because the node will have to be written again when the child is finally + allocated. + + WE HAVE NOT YET ELIMINATED THE UNALLOCATED CHILDREN PROBLEM. Except for bugs, this + should not cause any file system corruption, it only degrades I/O performance because a + node may be written when it is sure to be written at least one more time in the same + transaction when the remaining children are allocated. What follows is a description + of how we will solve the problem. +*/ + +/* HANDLING UNALLOCATED CHILDREN: During flush we may allocate a parent node then, + proceeding in parent first order, allocate some of its left-children, then encounter a + clean child in the middle of the parent. We do not allocate the clean child, but there + may remain unallocated (dirty) children to the right of the clean child. If we were to + stop flushing at this moment and write everything to disk, the parent might still + contain unallocated children. + + We could try to allocate all the descendents of every node that we allocate, but this + is not necessary. Doing so could result in allocating the entire tree: if the root + node is allocated then every unallocated node would have to be allocated before + flushing. Actually, we do not have to write a node just because we allocate it. It is + possible to allocate but not write a node during flush, when it still has unallocated + children. However, this approach is probably not optimal for the following reason. + + The flush algorithm is designed to allocate nodes in parent-first order in an attempt + to optimize reads that occur in the same order. Thus we are read-optimizing for a + left-to-right scan through all the leaves in the system, and we are hoping to + write-optimize at the same time because those nodes will be written together in batch. + What happens, however, if we assign a block number to a node in its read-optimized + order but then avoid writing it because it has unallocated children? In that + situation, we lose out on the write-optimization aspect because a node will have to be + written again to the its location on the device, later, which likely means seeking back + to that location. + + So there are tradeoffs. We can choose either: + + A. Allocate all unallocated children to preserve both write-optimization and + read-optimization, but this is not always desirable because it may mean having to + allocate and flush very many nodes at once. + + B. Defer writing nodes with unallocated children, keep their read-optimized locations, + but sacrifice write-optimization because those nodes will be written again. + + C. Defer writing nodes with unallocated children, but do not keep their read-optimized + locations. Instead, choose to write-optimize them later, when they are written. To + facilitate this, we "undo" the read-optimized allocation that was given to the node so + that later it can be write-optimized, thus "unpreparing" the flush decision. This is a + case where we disturb the FLUSH_PREP_ONCE_PER_TRANSACTION rule described above. By a + call to flush_unprep() we will: if the node was wandered, unset the JNODE_OVRWR bit; + if the node was relocated, unset the JNODE_RELOC bit, non-deferred-deallocate its block + location, and set the JNODE_CREATED bit, effectively setting the node back to an + unallocated state. + + We will take the following approach in v4.0: for twig nodes we will always finish + allocating unallocated children (A). For nodes with (level > TWIG) we will defer + writing and choose write-optimization (C). + + To summarize, there are several parts to a solution that avoids the problem with + unallocated children: + + FIXME-ZAM: Still no one approach is implemented to eliminate the "UNALLOCATED CHILDREN" + problem because there was an experiment which was done showed that we have 1-2 nodes + with unallocated children for thousands of written nodes. The experiment was simple + like coping / deletion of linux kernel sources. However the problem can arise in more + complex tests. I think we have jnode_io_hook to insert a check for unallocated + children and see what kind of problem we have. + + 1. When flush reaches a stopping point (e.g., a clean node), it should continue calling + squeeze-and-allocate on any remaining unallocated children. FIXME: Difficulty to + implement: should be simple -- amounts to adding a while loop to jnode_flush, see + comments in that function. + + 2. When flush reaches flush_empty_queue(), some of the (level > TWIG) nodes may still + have unallocated children. If the twig level has unallocated children it is an + assertion failure. If a higher-level node has unallocated children, then it should be + explicitly de-allocated by a call to flush_unprep(). FIXME: Difficulty to implement: + should be simple. + + 3. (CPU-Optimization) Checking whether a node has unallocated children may consume more + CPU cycles than we would like, and it is possible (but medium complexity) to optimize + this somewhat in the case where large sub-trees are flushed. The following observation + helps: if both the left- and right-neighbor of a node are processed by the flush + algorithm then the node itself is guaranteed to have all of its children allocated. + However, the cost of this check may not be so expensive after all: it is not needed for + leaves and flush can guarantee this property for twigs. That leaves only (level > + TWIG) nodes that have to be checked, so this optimization only helps if at least three + (level > TWIG) nodes are flushed in one pass, and the savings will be very small unless + there are many more (level > TWIG) nodes. But if there are many (level > TWIG) nodes + then the number of blocks being written will be very large, so the savings may be + insignificant. That said, the idea is to maintain both the left and right edges of + nodes that are processed in flush. When flush_empty_queue() is called, a relatively + simple test will tell whether the (level > TWIG) node is on the edge. If it is on the + edge, the slow check is necessary, but if it is in the interior then it can be assumed + to have all of its children allocated. FIXME: medium complexity to implement, but + simple to verify given that we must have a slow check anyway. + + 4. (Optional) This part is optional, not for v4.0--flush should work independently of + whether this option is used or not. Called RAPID_SCAN, the idea is to amend the + left-scan operation to take unallocated children into account. Normally, the left-scan + operation goes left as long as adjacent nodes are dirty up until some large maximum + value (FLUSH_SCAN_MAXNODES) at which point it stops and begins flushing. But scan-left + may stop at a position where there are unallocated children to the left with the same + parent. When RAPID_SCAN is enabled, the ordinary scan-left operation stops after + FLUSH_RELOCATE_THRESHOLD, which is much smaller than FLUSH_SCAN_MAXNODES, then procedes + with a rapid scan. The rapid scan skips all the interior children of a node--if the + leftmost child of a twig is dirty, check its left neighbor (the rightmost child of the + twig to the left). If the left neighbor of the leftmost child is also dirty, then + continue the scan at the left twig and repeat. This option will cause flush to + allocate more twigs in a single pass, but it also has the potential to write many more + nodes than would otherwise be written without the RAPID_SCAN option. RAPID_SCAN + was partially implemented, code removed August 12, 2002 by JMACD. +*/ + +/* FLUSH CALLED ON NON-LEAF LEVEL. Most of our design considerations assume that the + starting point for flush is a leaf node, but actually the flush code cares very little + about whether or not this is true. It is possible that all the leaf nodes are flushed + and dirty parent nodes still remain, in which case jnode_flush() is called on a + non-leaf argument. Flush doesn't care--it treats the argument node as if it were a + leaf, even when it is not. This is a simple approach, and there may be a more optimal + policy but until a problem with this approach is discovered, simplest is probably best. + + NOTE: In this case, the ordering produced by flush is parent-first only if you ignore + the leaves. This is done as a matter of simplicity and there is only one (shaky) + justification. When an atom commits, it flushes all leaf level nodes first, followed + by twigs, and so on. With flushing done in this order, if flush is eventually called + on a non-leaf node it means that (somehow) we reached a point where all leaves are + clean and only internal nodes need to be flushed. If that it the case, then it means + there were no leaves that were the parent-first preceder/follower of the parent. This + is expected to be a rare case, which is why we do nothing special about it. However, + memory pressure may pass an internal node to flush when there are still dirty leaf + nodes that need to be flushed, which could prove our original assumptions + "inoperative". If this needs to be fixed, then scan_left/right should have + special checks for the non-leaf levels. For example, instead of passing from a node to + the left neighbor, it should pass from the node to the left neighbor's rightmost + descendent (if dirty). + +*/ + +/* UNIMPLEMENTED AS YET: REPACKING AND RESIZING. We walk the tree in 4MB-16MB chunks, dirtying everything and putting + it into a transaction. We tell the allocator to allocate the blocks as far as possible towards one end of the + logical device--the left (starting) end of the device if we are walking from left to right, the right end of the + device if we are walking from right to left. We then make passes in alternating directions, and as we do this the + device becomes sorted such that tree order and block number order fully correlate. + + Resizing is done by shifting everything either all the way to the left or all the way + to the right, and then reporting the last block. +*/ + +/* RELOCATE DECISIONS: The code makes a decision to relocate in several places. This + descibes the policy from the highest level: + + The FLUSH_RELOCATE_THRESHOLD parameter: If we count this many consecutive nodes on the + leaf level during flush-scan (right, left), then we unconditionally decide to relocate + leaf nodes. + + Otherwise, there are two contexts in which we make a decision to relocate: + + 1. The REVERSE PARENT-FIRST context: Implemented in reverse_relocate_test(). + During the initial stages of flush, after scan-right completes, we want to ask the + question: should we relocate this leaf node and thus dirty the parent node. Then if + the node is a leftmost child its parent is its own parent-first preceder, thus we repeat + the question at the next level up, and so on. In these cases we are moving in the + reverse-parent first direction. + + There is another case which is considered the reverse direction, which comes at the end + of a twig in reverse_relocate_end_of_twig(). As we finish processing a twig we may + reach a point where there is a clean twig to the right with a dirty leftmost child. In + this case, we may wish to relocate the child by testing if it should be relocated + relative to its parent. + + 2. The FORWARD PARENT-FIRST context: Testing for forward relocation is done in + allocate_znode. What distinguishes the forward parent-first case from the + reverse-parent first case is that the preceder has already been allocated in the + forward case, whereas in the reverse case we don't know what the preceder is until we + finish "going in reverse". That simplifies the forward case considerably, and there we + actually use the block allocator to determine whether, e.g., a block closer to the + preceder is available. +*/ + +/* SQUEEZE_LEFT_EDGE: Unimplemented idea for future consideration. The idea is, once we + finish scan-left and find a starting point, if the parent's left neighbor is dirty then + squeeze the parent's left neighbor and the parent. This may change the + flush-starting-node's parent. Repeat until the child's parent is stable. If the child + is a leftmost child, repeat this left-edge squeezing operation at the next level up. + Note that we cannot allocate extents during this or they will be out of parent-first + order. There is also some difficult coordinate maintenence issues. We can't do a tree + search to find coordinates again (because we hold locks), we have to determine them + from the two nodes being squeezed. Looks difficult, but has potential to increase + space utilization. */ + +/* Flush-scan helper functions. */ +static void scan_init(flush_scan * scan); +static void scan_done(flush_scan * scan); + +/* Flush-scan algorithm. */ +static int scan_left(flush_scan * scan, flush_scan * right, jnode * node, + unsigned limit); +static int scan_right(flush_scan * scan, jnode * node, unsigned limit); +static int scan_common(flush_scan * scan, flush_scan * other); +static int scan_formatted(flush_scan * scan); +static int scan_unformatted(flush_scan * scan, flush_scan * other); +static int scan_by_coord(flush_scan * scan); + +/* Initial flush-point ancestor allocation. */ +static int alloc_pos_and_ancestors(flush_pos_t * pos); +static int alloc_one_ancestor(const coord_t * coord, flush_pos_t * pos); +static int set_preceder(const coord_t * coord_in, flush_pos_t * pos); + +/* Main flush algorithm. Note on abbreviation: "squeeze and allocate" == "squalloc". */ +static int squalloc(flush_pos_t * pos); + +/* Flush squeeze implementation. */ +static int squeeze_right_non_twig(znode * left, znode * right); +static int shift_one_internal_unit(znode * left, znode * right); + +/* Flush reverse parent-first relocation routines. */ +static int reverse_relocate_if_close_enough(const reiser4_block_nr * pblk, + const reiser4_block_nr * nblk); +static int reverse_relocate_test(jnode * node, const coord_t * parent_coord, + flush_pos_t * pos); +static int reverse_relocate_check_dirty_parent(jnode * node, + const coord_t * parent_coord, + flush_pos_t * pos); + +/* Flush allocate write-queueing functions: */ +static int allocate_znode(znode * node, const coord_t * parent_coord, + flush_pos_t * pos); +static int allocate_znode_update(znode * node, const coord_t * parent_coord, + flush_pos_t * pos); +static int lock_parent_and_allocate_znode(znode *, flush_pos_t *); + +/* Flush helper functions: */ +static int jnode_lock_parent_coord(jnode * node, + coord_t * coord, + lock_handle * parent_lh, + load_count * parent_zh, + znode_lock_mode mode, int try); +static int neighbor_in_slum(znode * node, lock_handle * right_lock, sideof side, + znode_lock_mode mode, int check_dirty, int expected); +static int znode_same_parents(znode * a, znode * b); + +static int znode_check_flushprepped(znode * node) +{ + return jnode_check_flushprepped(ZJNODE(node)); +} + +/* Flush position functions */ +static void pos_init(flush_pos_t * pos); +static int pos_valid(flush_pos_t * pos); +static void pos_done(flush_pos_t * pos); +static int pos_stop(flush_pos_t * pos); + +/* check that @org is first jnode extent unit, if extent is unallocated, + * because all jnodes of unallocated extent are dirty and of the same atom. */ +#define checkchild(scan) \ +assert("nikita-3435", \ + ergo(scan->direction == LEFT_SIDE && \ + (scan->parent_coord.node->level == TWIG_LEVEL) && \ + jnode_is_unformatted(scan->node) && \ + extent_is_unallocated(&scan->parent_coord), \ + extent_unit_index(&scan->parent_coord) == index_jnode(scan->node))) + +/* This flush_cnt variable is used to track the number of concurrent flush operations, + useful for debugging. It is initialized in txnmgr.c out of laziness (because flush has + no static initializer function...) */ +ON_DEBUG(atomic_t flush_cnt; + ) + +/* check fs backing device for write congestion */ +static int check_write_congestion(void) +{ + struct super_block *sb; + struct backing_dev_info *bdi; + + sb = reiser4_get_current_sb(); + bdi = reiser4_get_super_fake(sb)->i_mapping->backing_dev_info; + return bdi_write_congested(bdi); +} + +/* conditionally write flush queue */ +static int write_prepped_nodes(flush_pos_t * pos) +{ + int ret; + + assert("zam-831", pos); + assert("zam-832", pos->fq); + + if (!(pos->flags & JNODE_FLUSH_WRITE_BLOCKS)) + return 0; + + if (check_write_congestion()) + return 0; + + ret = reiser4_write_fq(pos->fq, pos->nr_written, + WRITEOUT_SINGLE_STREAM | WRITEOUT_FOR_PAGE_RECLAIM); + return ret; +} + +/* Proper release all flush pos. resources then move flush position to new + locked node */ +static void move_flush_pos(flush_pos_t * pos, lock_handle * new_lock, + load_count * new_load, const coord_t * new_coord) +{ + assert("zam-857", new_lock->node == new_load->node); + + if (new_coord) { + assert("zam-858", new_coord->node == new_lock->node); + coord_dup(&pos->coord, new_coord); + } else { + coord_init_first_unit(&pos->coord, new_lock->node); + } + + if (pos->child) { + jput(pos->child); + pos->child = NULL; + } + + move_load_count(&pos->load, new_load); + done_lh(&pos->lock); + move_lh(&pos->lock, new_lock); +} + +/* delete empty node which link from the parent still exists. */ +static int delete_empty_node(znode * node) +{ + reiser4_key smallest_removed; + + assert("zam-1019", node != NULL); + assert("zam-1020", node_is_empty(node)); + assert("zam-1023", znode_is_wlocked(node)); + + return reiser4_delete_node(node, &smallest_removed, NULL, 1); +} + +/* Prepare flush position for alloc_pos_and_ancestors() and squalloc() */ +static int prepare_flush_pos(flush_pos_t * pos, jnode * org) +{ + int ret; + load_count load; + lock_handle lock; + + init_lh(&lock); + init_load_count(&load); + + if (jnode_is_znode(org)) { + ret = longterm_lock_znode(&lock, JZNODE(org), + ZNODE_WRITE_LOCK, ZNODE_LOCK_HIPRI); + if (ret) + return ret; + + ret = incr_load_count_znode(&load, JZNODE(org)); + if (ret) + return ret; + + pos->state = + (jnode_get_level(org) == + LEAF_LEVEL) ? POS_ON_LEAF : POS_ON_INTERNAL; + move_flush_pos(pos, &lock, &load, NULL); + } else { + coord_t parent_coord; + ret = jnode_lock_parent_coord(org, &parent_coord, &lock, + &load, ZNODE_WRITE_LOCK, 0); + if (ret) + goto done; + if (!item_is_extent(&parent_coord)) { + /* file was converted to tail, org became HB, we found internal + item */ + ret = -EAGAIN; + goto done; + } + + pos->state = POS_ON_EPOINT; + move_flush_pos(pos, &lock, &load, &parent_coord); + pos->child = jref(org); + if (extent_is_unallocated(&parent_coord) + && extent_unit_index(&parent_coord) != index_jnode(org)) { + /* @org is not first child of its parent unit. This may happen + because longerm lock of its parent node was released between + scan_left and scan_right. For now work around this having flush to repeat */ + ret = -EAGAIN; + } + } + + done: + done_load_count(&load); + done_lh(&lock); + return ret; +} + +/* TODO LIST (no particular order): */ +/* I have labelled most of the legitimate FIXME comments in this file with letters to + indicate which issue they relate to. There are a few miscellaneous FIXMEs with + specific names mentioned instead that need to be inspected/resolved. */ +/* B. There is an issue described in reverse_relocate_test having to do with an + imprecise is_preceder? check having to do with partially-dirty extents. The code that + sets preceder hints and computes the preceder is basically untested. Careful testing + needs to be done that preceder calculations are done correctly, since if it doesn't + affect correctness we will not catch this stuff during regular testing. */ +/* C. EINVAL, E_DEADLOCK, E_NO_NEIGHBOR, ENOENT handling. It is unclear which of these are + considered expected but unlikely conditions. Flush currently returns 0 (i.e., success + but no progress, i.e., restart) whenever it receives any of these in jnode_flush(). + Many of the calls that may produce one of these return values (i.e., + longterm_lock_znode, reiser4_get_parent, reiser4_get_neighbor, ...) check some of these + values themselves and, for instance, stop flushing instead of resulting in a restart. + If any of these results are true error conditions then flush will go into a busy-loop, + as we noticed during testing when a corrupt tree caused find_child_ptr to return + ENOENT. It needs careful thought and testing of corner conditions. +*/ +/* D. Atomicity of flush_prep against deletion and flush concurrency. Suppose a created + block is assigned a block number then early-flushed to disk. It is dirtied again and + flush is called again. Concurrently, that block is deleted, and the de-allocation of + its block number does not need to be deferred, since it is not part of the preserve set + (i.e., it didn't exist before the transaction). I think there may be a race condition + where flush writes the dirty, created block after the non-deferred deallocated block + number is re-allocated, making it possible to write deleted data on top of non-deleted + data. Its just a theory, but it needs to be thought out. */ +/* F. bio_alloc() failure is not handled gracefully. */ +/* G. Unallocated children. */ +/* H. Add a WANDERED_LIST to the atom to clarify the placement of wandered blocks. */ +/* I. Rename flush-scan to scan-point, (flush-pos to flush-point?) */ + +/* JNODE_FLUSH: MAIN ENTRY POINT */ +/* This is the main entry point for flushing a jnode and its dirty neighborhood (dirty + neighborhood is named "slum"). Jnode_flush() is called if reiser4 has to write dirty + blocks to disk, it happens when Linux VM decides to reduce number of dirty pages or as + a part of transaction commit. + + Our objective here is to prep and flush the slum the jnode belongs to. We want to + squish the slum together, and allocate the nodes in it as we squish because allocation + of children affects squishing of parents. + + The "argument" @node tells flush where to start. From there, flush finds the left edge + of the slum, and calls squalloc (in which nodes are squeezed and allocated). To find a + "better place" to start squalloc first we perform a flush_scan. + + Flush-scanning may be performed in both left and right directions, but for different + purposes. When scanning to the left, we are searching for a node that precedes a + sequence of parent-first-ordered nodes which we will then flush in parent-first order. + During flush-scanning, we also take the opportunity to count the number of consecutive + leaf nodes. If this number is past some threshold (FLUSH_RELOCATE_THRESHOLD), then we + make a decision to reallocate leaf nodes (thus favoring write-optimization). + + Since the flush argument node can be anywhere in a sequence of dirty leaves, there may + also be dirty nodes to the right of the argument. If the scan-left operation does not + count at least FLUSH_RELOCATE_THRESHOLD nodes then we follow it with a right-scan + operation to see whether there is, in fact, enough nodes to meet the relocate + threshold. Each right- and left-scan operation uses a single flush_scan object. + + After left-scan and possibly right-scan, we prepare a flush_position object with the + starting flush point or parent coordinate, which was determined using scan-left. + + Next we call the main flush routine, squalloc, which iterates along the + leaf level, squeezing and allocating nodes (and placing them into the flush queue). + + After squalloc returns we take extra steps to ensure that all the children + of the final twig node are allocated--this involves repeating squalloc + until we finish at a twig with no unallocated children. + + Finally, we call flush_empty_queue to submit write-requests to disk. If we encounter + any above-twig nodes during flush_empty_queue that still have unallocated children, we + flush_unprep them. + + Flush treats several "failure" cases as non-failures, essentially causing them to start + over. E_DEADLOCK is one example. FIXME:(C) EINVAL, E_NO_NEIGHBOR, ENOENT: these should + probably be handled properly rather than restarting, but there are a bunch of cases to + audit. +*/ + +static int +jnode_flush(jnode * node, long nr_to_write, long *nr_written, + flush_queue_t * fq, int flags) +{ + long ret = 0; + flush_scan *right_scan; + flush_scan *left_scan; + flush_pos_t *flush_pos; + int todo; + struct super_block *sb; + reiser4_super_info_data *sbinfo; + jnode *leftmost_in_slum = NULL; + + assert("jmacd-76619", lock_stack_isclean(get_current_lock_stack())); + assert("nikita-3022", reiser4_schedulable()); + + assert("nikita-3185", + get_current_super_private()->delete_mutex_owner != current); + + /* allocate right_scan, left_scan and flush_pos */ + right_scan = + kmalloc(2 * sizeof(*right_scan) + sizeof(*flush_pos), + reiser4_ctx_gfp_mask_get()); + if (right_scan == NULL) + return RETERR(-ENOMEM); + left_scan = right_scan + 1; + flush_pos = (flush_pos_t *) (left_scan + 1); + + sb = reiser4_get_current_sb(); + sbinfo = get_super_private(sb); + + /* Flush-concurrency debug code */ +#if REISER4_DEBUG + atomic_inc(&flush_cnt); +#endif + + reiser4_enter_flush(sb); + + /* Initialize a flush position. */ + pos_init(flush_pos); + + flush_pos->nr_written = nr_written; + flush_pos->fq = fq; + flush_pos->flags = flags; + flush_pos->nr_to_write = nr_to_write; + + scan_init(right_scan); + scan_init(left_scan); + + /* First scan left and remember the leftmost scan position. If the leftmost + position is unformatted we remember its parent_coord. We scan until counting + FLUSH_SCAN_MAXNODES. + + If starting @node is unformatted, at the beginning of left scan its + parent (twig level node, containing extent item) will be long term + locked and lock handle will be stored in the + @right_scan->parent_lock. This lock is used to start the rightward + scan without redoing the tree traversal (necessary to find parent) + and, hence, is kept during leftward scan. As a result, we have to + use try-lock when taking long term locks during the leftward scan. + */ + ret = scan_left(left_scan, right_scan, + node, sbinfo->flush.scan_maxnodes); + if (ret != 0) + goto failed; + + leftmost_in_slum = jref(left_scan->node); + scan_done(left_scan); + + /* Then possibly go right to decide if we will use a policy of relocating leaves. + This is only done if we did not scan past (and count) enough nodes during the + leftward scan. If we do scan right, we only care to go far enough to establish + that at least FLUSH_RELOCATE_THRESHOLD number of nodes are being flushed. The + scan limit is the difference between left_scan.count and the threshold. */ + + todo = sbinfo->flush.relocate_threshold - left_scan->count; + /* scan right is inherently deadlock prone, because we are + * (potentially) holding a lock on the twig node at this moment. + * FIXME: this is incorrect comment: lock is not held */ + if (todo > 0) { + ret = scan_right(right_scan, node, (unsigned)todo); + if (ret != 0) + goto failed; + } + + /* Only the right-scan count is needed, release any rightward locks right away. */ + scan_done(right_scan); + + /* ... and the answer is: we should relocate leaf nodes if at least + FLUSH_RELOCATE_THRESHOLD nodes were found. */ + flush_pos->leaf_relocate = JF_ISSET(node, JNODE_REPACK) || + (left_scan->count + right_scan->count >= + sbinfo->flush.relocate_threshold); + + /* Funny business here. We set the 'point' in the flush_position at prior to + starting squalloc regardless of whether the first point is + formatted or unformatted. Without this there would be an invariant, in the + rest of the code, that if the flush_position is unformatted then + flush_position->point is NULL and flush_position->parent_{lock,coord} is set, + and if the flush_position is formatted then flush_position->point is non-NULL + and no parent info is set. + + This seems lazy, but it makes the initial calls to reverse_relocate_test + (which ask "is it the pos->point the leftmost child of its parent") much easier + because we know the first child already. Nothing is broken by this, but the + reasoning is subtle. Holding an extra reference on a jnode during flush can + cause us to see nodes with HEARD_BANSHEE during squalloc, because nodes are not + removed from sibling lists until they have zero reference count. Flush would + never observe a HEARD_BANSHEE node on the left-edge of flush, nodes are only + deleted to the right. So if nothing is broken, why fix it? + + NOTE-NIKITA actually, flush can meet HEARD_BANSHEE node at any + point and in any moment, because of the concurrent file system + activity (for example, truncate). */ + + /* Check jnode state after flush_scan completed. Having a lock on this + node or its parent (in case of unformatted) helps us in case of + concurrent flushing. */ + if (jnode_check_flushprepped(leftmost_in_slum) + && !jnode_convertible(leftmost_in_slum)) { + ret = 0; + goto failed; + } + + /* Now setup flush_pos using scan_left's endpoint. */ + ret = prepare_flush_pos(flush_pos, leftmost_in_slum); + if (ret) + goto failed; + + if (znode_get_level(flush_pos->coord.node) == LEAF_LEVEL + && node_is_empty(flush_pos->coord.node)) { + znode *empty = flush_pos->coord.node; + + assert("zam-1022", !ZF_ISSET(empty, JNODE_HEARD_BANSHEE)); + ret = delete_empty_node(empty); + goto failed; + } + + if (jnode_check_flushprepped(leftmost_in_slum) + && !jnode_convertible(leftmost_in_slum)) { + ret = 0; + goto failed; + } + + /* Set pos->preceder and (re)allocate pos and its ancestors if it is needed */ + ret = alloc_pos_and_ancestors(flush_pos); + if (ret) + goto failed; + + /* Do the main rightward-bottom-up squeeze and allocate loop. */ + ret = squalloc(flush_pos); + pos_stop(flush_pos); + if (ret) + goto failed; + + /* FIXME_NFQUCMPD: Here, handle the twig-special case for unallocated children. + First, the pos_stop() and pos_valid() routines should be modified + so that pos_stop() sets a flush_position->stop flag to 1 without + releasing the current position immediately--instead release it in + pos_done(). This is a better implementation than the current one anyway. + + It is not clear that all fields of the flush_position should not be released, + but at the very least the parent_lock, parent_coord, and parent_load should + remain held because they are hold the last twig when pos_stop() is + called. + + When we reach this point in the code, if the parent_coord is set to after the + last item then we know that flush reached the end of a twig (and according to + the new flush queueing design, we will return now). If parent_coord is not + past the last item, we should check if the current twig has any unallocated + children to the right (we are not concerned with unallocated children to the + left--in that case the twig itself should not have been allocated). If the + twig has unallocated children to the right, set the parent_coord to that + position and then repeat the call to squalloc. + + Testing for unallocated children may be defined in two ways: if any internal + item has a fake block number, it is unallocated; if any extent item is + unallocated then all of its children are unallocated. But there is a more + aggressive approach: if there are any dirty children of the twig to the right + of the current position, we may wish to relocate those nodes now. Checking for + potential relocation is more expensive as it requires knowing whether there are + any dirty children that are not unallocated. The extent_needs_allocation + should be used after setting the correct preceder. + + When we reach the end of a twig at this point in the code, if the flush can + continue (when the queue is ready) it will need some information on the future + starting point. That should be stored away in the flush_handle using a seal, I + believe. Holding a jref() on the future starting point may break other code + that deletes that node. + */ + + /* FIXME_NFQUCMPD: Also, we don't want to do any flushing when flush is called + above the twig level. If the VM calls flush above the twig level, do nothing + and return (but figure out why this happens). The txnmgr should be modified to + only flush its leaf-level dirty list. This will do all the necessary squeeze + and allocate steps but leave unallocated branches and possibly unallocated + twigs (when the twig's leftmost child is not dirty). After flushing the leaf + level, the remaining unallocated nodes should be given write-optimized + locations. (Possibly, the remaining unallocated twigs should be allocated just + before their leftmost child.) + */ + + /* Any failure reaches this point. */ + failed: + + switch (ret) { + case -E_REPEAT: + case -EINVAL: + case -E_DEADLOCK: + case -E_NO_NEIGHBOR: + case -ENOENT: + /* FIXME(C): Except for E_DEADLOCK, these should probably be handled properly + in each case. They already are handled in many cases. */ + /* Something bad happened, but difficult to avoid... Try again! */ + ret = 0; + } + + if (leftmost_in_slum) + jput(leftmost_in_slum); + + pos_done(flush_pos); + scan_done(left_scan); + scan_done(right_scan); + kfree(right_scan); + + ON_DEBUG(atomic_dec(&flush_cnt)); + + reiser4_leave_flush(sb); + + return ret; +} + +/* The reiser4 flush subsystem can be turned into "rapid flush mode" means that + * flusher should submit all prepped nodes immediately without keeping them in + * flush queues for long time. The reason for rapid flush mode is to free + * memory as fast as possible. */ + +#if REISER4_USE_RAPID_FLUSH + +/** + * submit all prepped nodes if rapid flush mode is set, + * turn rapid flush mode off. + */ + +static int rapid_flush(flush_pos_t * pos) +{ + if (!wbq_available()) + return 0; + + return write_prepped_nodes(pos); +} + +#else + +#define rapid_flush(pos) (0) + +#endif /* REISER4_USE_RAPID_FLUSH */ + +static jnode *find_flush_start_jnode(jnode *start, txn_atom *atom, + flush_queue_t *fq, int *nr_queued, + int flags) +{ + jnode * node; + + if (start != NULL) { + spin_lock_jnode(start); + if (!jnode_is_flushprepped(start)) { + assert("zam-1056", start->atom == atom); + node = start; + goto enter; + } + spin_unlock_jnode(start); + } + /* + * In this loop we process all already prepped (RELOC or OVRWR) and dirtied again + * nodes. The atom spin lock is not released until all dirty nodes processed or + * not prepped node found in the atom dirty lists. + */ + while ((node = find_first_dirty_jnode(atom, flags))) { + spin_lock_jnode(node); + enter: + assert("zam-881", JF_ISSET(node, JNODE_DIRTY)); + assert("zam-898", !JF_ISSET(node, JNODE_OVRWR)); + + if (JF_ISSET(node, JNODE_WRITEBACK)) { + /* move node to the end of atom's writeback list */ + list_move_tail(&node->capture_link, ATOM_WB_LIST(atom)); + + /* + * jnode is not necessarily on dirty list: if it was dirtied when + * it was on flush queue - it does not get moved to dirty list + */ + ON_DEBUG(count_jnode(atom, node, NODE_LIST(node), + WB_LIST, 1)); + + } else if (jnode_is_znode(node) + && znode_above_root(JZNODE(node))) { + /* + * A special case for znode-above-root. The above-root (fake) + * znode is captured and dirtied when the tree height changes or + * when the root node is relocated. This causes atoms to fuse so + * that changes at the root are serialized. However, this node is + * never flushed. This special case used to be in lock.c to + * prevent the above-root node from ever being captured, but now + * that it is captured we simply prevent it from flushing. The + * log-writer code relies on this to properly log superblock + * modifications of the tree height. + */ + jnode_make_wander_nolock(node); + } else if (JF_ISSET(node, JNODE_RELOC)) { + queue_jnode(fq, node); + ++(*nr_queued); + } else + break; + + spin_unlock_jnode(node); + } + return node; +} + +/* Flush some nodes of current atom, usually slum, return -E_REPEAT if there are more nodes + * to flush, return 0 if atom's dirty lists empty and keep current atom locked, return + * other errors as they are. */ +int +flush_current_atom(int flags, long nr_to_write, long *nr_submitted, + txn_atom ** atom, jnode *start) +{ + reiser4_super_info_data *sinfo = get_current_super_private(); + flush_queue_t *fq = NULL; + jnode *node; + int nr_queued; + int ret; + + assert("zam-889", atom != NULL && *atom != NULL); + assert_spin_locked(&((*atom)->alock)); + assert("zam-892", get_current_context()->trans->atom == *atom); + + nr_to_write = LONG_MAX; + while (1) { + ret = reiser4_fq_by_atom(*atom, &fq); + if (ret != -E_REPEAT) + break; + *atom = get_current_atom_locked(); + } + if (ret) + return ret; + + assert_spin_locked(&((*atom)->alock)); + + /* parallel flushers limit */ + if (sinfo->tmgr.atom_max_flushers != 0) { + while ((*atom)->nr_flushers >= sinfo->tmgr.atom_max_flushers) { + /* An reiser4_atom_send_event() call is inside + reiser4_fq_put_nolock() which is called when flush is + finished and nr_flushers is decremented. */ + reiser4_atom_wait_event(*atom); + *atom = get_current_atom_locked(); + } + } + + /* count ourself as a flusher */ + (*atom)->nr_flushers++; + + writeout_mode_enable(); + + nr_queued = 0; + node = find_flush_start_jnode(start, *atom, fq, &nr_queued, flags); + + if (node == NULL) { + if (nr_queued == 0) { + (*atom)->nr_flushers--; + reiser4_fq_put_nolock(fq); + reiser4_atom_send_event(*atom); + /* current atom remains locked */ + writeout_mode_disable(); + return 0; + } + spin_unlock_atom(*atom); + } else { + jref(node); + BUG_ON((*atom)->super != node->tree->super); + spin_unlock_atom(*atom); + spin_unlock_jnode(node); + BUG_ON(nr_to_write == 0); + ret = jnode_flush(node, nr_to_write, nr_submitted, fq, flags); + jput(node); + } + + ret = + reiser4_write_fq(fq, nr_submitted, + WRITEOUT_SINGLE_STREAM | WRITEOUT_FOR_PAGE_RECLAIM); + + *atom = get_current_atom_locked(); + (*atom)->nr_flushers--; + reiser4_fq_put_nolock(fq); + reiser4_atom_send_event(*atom); + spin_unlock_atom(*atom); + + writeout_mode_disable(); + + if (ret == 0) + ret = -E_REPEAT; + + return ret; +} + +/* REVERSE PARENT-FIRST RELOCATION POLICIES */ + +/* This implements the is-it-close-enough-to-its-preceder? test for relocation in the + reverse parent-first relocate context. Here all we know is the preceder and the block + number. Since we are going in reverse, the preceder may still be relocated as well, so + we can't ask the block allocator "is there a closer block available to relocate?" here. + In the _forward_ parent-first relocate context (not here) we actually call the block + allocator to try and find a closer location. */ +static int +reverse_relocate_if_close_enough(const reiser4_block_nr * pblk, + const reiser4_block_nr * nblk) +{ + reiser4_block_nr dist; + + assert("jmacd-7710", *pblk != 0 && *nblk != 0); + assert("jmacd-7711", !reiser4_blocknr_is_fake(pblk)); + assert("jmacd-7712", !reiser4_blocknr_is_fake(nblk)); + + /* Distance is the absolute value. */ + dist = (*pblk > *nblk) ? (*pblk - *nblk) : (*nblk - *pblk); + + /* If the block is less than FLUSH_RELOCATE_DISTANCE blocks away from its preceder + block, do not relocate. */ + if (dist <= get_current_super_private()->flush.relocate_distance) { + return 0; + } + + return 1; +} + +/* This function is a predicate that tests for relocation. Always called in the + reverse-parent-first context, when we are asking whether the current node should be + relocated in order to expand the flush by dirtying the parent level (and thus + proceeding to flush that level). When traversing in the forward parent-first direction + (not here), relocation decisions are handled in two places: allocate_znode() and + extent_needs_allocation(). */ +static int +reverse_relocate_test(jnode * node, const coord_t * parent_coord, + flush_pos_t * pos) +{ + reiser4_block_nr pblk = 0; + reiser4_block_nr nblk = 0; + + assert("jmacd-8989", !jnode_is_root(node)); + + /* + * This function is called only from the + * reverse_relocate_check_dirty_parent() and only if the parent + * node is clean. This implies that the parent has the real (i.e., not + * fake) block number, and, so does the child, because otherwise the + * parent would be dirty. + */ + + /* New nodes are treated as if they are being relocated. */ + if (JF_ISSET (node, JNODE_CREATED) || + (pos->leaf_relocate && jnode_get_level(node) == LEAF_LEVEL)) { + return 1; + } + + /* Find the preceder. FIXME(B): When the child is an unformatted, previously + existing node, the coord may be leftmost even though the child is not the + parent-first preceder of the parent. If the first dirty node appears somewhere + in the middle of the first extent unit, this preceder calculation is wrong. + Needs more logic in here. */ + if (coord_is_leftmost_unit(parent_coord)) { + pblk = *znode_get_block(parent_coord->node); + } else { + pblk = pos->preceder.blk; + } + check_preceder(pblk); + + /* If (pblk == 0) then the preceder isn't allocated or isn't known: relocate. */ + if (pblk == 0) { + return 1; + } + + nblk = *jnode_get_block(node); + + if (reiser4_blocknr_is_fake(&nblk)) + /* child is unallocated, mark parent dirty */ + return 1; + + return reverse_relocate_if_close_enough(&pblk, &nblk); +} + +/* This function calls reverse_relocate_test to make a reverse-parent-first + relocation decision and then, if yes, it marks the parent dirty. */ +static int +reverse_relocate_check_dirty_parent(jnode * node, const coord_t * parent_coord, + flush_pos_t * pos) +{ + int ret; + + if (!JF_ISSET(ZJNODE(parent_coord->node), JNODE_DIRTY)) { + + ret = reverse_relocate_test(node, parent_coord, pos); + if (ret < 0) { + return ret; + } + + /* FIXME-ZAM + if parent is already relocated - we do not want to grab space, right? */ + if (ret == 1) { + int grabbed; + + grabbed = get_current_context()->grabbed_blocks; + if (reiser4_grab_space_force((__u64) 1, BA_RESERVED) != + 0) + reiser4_panic("umka-1250", + "No space left during flush."); + + assert("jmacd-18923", + znode_is_write_locked(parent_coord->node)); + znode_make_dirty(parent_coord->node); + grabbed2free_mark(grabbed); + } + } + + return 0; +} + +/* INITIAL ALLOCATE ANCESTORS STEP (REVERSE PARENT-FIRST ALLOCATION BEFORE FORWARD + PARENT-FIRST LOOP BEGINS) */ + +/* Get the leftmost child for given coord. */ +static int get_leftmost_child_of_unit(const coord_t * coord, jnode ** child) +{ + int ret; + + ret = item_utmost_child(coord, LEFT_SIDE, child); + + if (ret) + return ret; + + if (IS_ERR(*child)) + return PTR_ERR(*child); + + return 0; +} + +/* This step occurs after the left- and right-scans are completed, before starting the + forward parent-first traversal. Here we attempt to allocate ancestors of the starting + flush point, which means continuing in the reverse parent-first direction to the + parent, grandparent, and so on (as long as the child is a leftmost child). This + routine calls a recursive process, alloc_one_ancestor, which does the real work, + except there is special-case handling here for the first ancestor, which may be a twig. + At each level (here and alloc_one_ancestor), we check for relocation and then, if + the child is a leftmost child, repeat at the next level. On the way back down (the + recursion), we allocate the ancestors in parent-first order. */ +static int alloc_pos_and_ancestors(flush_pos_t * pos) +{ + int ret = 0; + lock_handle plock; + load_count pload; + coord_t pcoord; + + if (znode_check_flushprepped(pos->lock.node)) + return 0; + + coord_init_invalid(&pcoord, NULL); + init_lh(&plock); + init_load_count(&pload); + + if (pos->state == POS_ON_EPOINT) { + /* a special case for pos on twig level, where we already have + a lock on parent node. */ + /* The parent may not be dirty, in which case we should decide + whether to relocate the child now. If decision is made to + relocate the child, the parent is marked dirty. */ + ret = + reverse_relocate_check_dirty_parent(pos->child, &pos->coord, + pos); + if (ret) + goto exit; + + /* FIXME_NFQUCMPD: We only need to allocate the twig (if child + is leftmost) and the leaf/child, so recursion is not needed. + Levels above the twig will be allocated for + write-optimization before the transaction commits. */ + + /* Do the recursive step, allocating zero or more of our + * ancestors. */ + ret = alloc_one_ancestor(&pos->coord, pos); + + } else { + if (!znode_is_root(pos->lock.node)) { + /* all formatted nodes except tree root */ + ret = + reiser4_get_parent(&plock, pos->lock.node, + ZNODE_WRITE_LOCK); + if (ret) + goto exit; + + ret = incr_load_count_znode(&pload, plock.node); + if (ret) + goto exit; + + ret = + find_child_ptr(plock.node, pos->lock.node, &pcoord); + if (ret) + goto exit; + + ret = + reverse_relocate_check_dirty_parent(ZJNODE + (pos->lock. + node), &pcoord, + pos); + if (ret) + goto exit; + + ret = alloc_one_ancestor(&pcoord, pos); + if (ret) + goto exit; + } + + ret = allocate_znode(pos->lock.node, &pcoord, pos); + } + exit: + done_load_count(&pload); + done_lh(&plock); + return ret; +} + +/* This is the recursive step described in alloc_pos_and_ancestors, above. Ignoring the + call to set_preceder, which is the next function described, this checks if the + child is a leftmost child and returns if it is not. If the child is a leftmost child + it checks for relocation, possibly dirtying the parent. Then it performs the recursive + step. */ +static int alloc_one_ancestor(const coord_t * coord, flush_pos_t * pos) +{ + int ret = 0; + lock_handle alock; + load_count aload; + coord_t acoord; + + /* As we ascend at the left-edge of the region to flush, take this opportunity at + the twig level to find our parent-first preceder unless we have already set + it. */ + if (pos->preceder.blk == 0) { + ret = set_preceder(coord, pos); + if (ret != 0) + return ret; + } + + /* If the ancestor is clean or already allocated, or if the child is not a + leftmost child, stop going up, even leaving coord->node not flushprepped. */ + if (znode_check_flushprepped(coord->node) + || !coord_is_leftmost_unit(coord)) + return 0; + + init_lh(&alock); + init_load_count(&aload); + coord_init_invalid(&acoord, NULL); + + /* Only ascend to the next level if it is a leftmost child, but write-lock the + parent in case we will relocate the child. */ + if (!znode_is_root(coord->node)) { + + ret = + jnode_lock_parent_coord(ZJNODE(coord->node), &acoord, + &alock, &aload, ZNODE_WRITE_LOCK, + 0); + if (ret != 0) { + /* FIXME(C): check EINVAL, E_DEADLOCK */ + goto exit; + } + + ret = + reverse_relocate_check_dirty_parent(ZJNODE(coord->node), + &acoord, pos); + if (ret != 0) { + goto exit; + } + + /* Recursive call. */ + if (!znode_check_flushprepped(acoord.node)) { + ret = alloc_one_ancestor(&acoord, pos); + if (ret) + goto exit; + } + } + + /* Note: we call allocate with the parent write-locked (except at the root) in + case we relocate the child, in which case it will modify the parent during this + call. */ + ret = allocate_znode(coord->node, &acoord, pos); + + exit: + done_load_count(&aload); + done_lh(&alock); + return ret; +} + +/* During the reverse parent-first alloc_pos_and_ancestors process described above there is + a call to this function at the twig level. During alloc_pos_and_ancestors we may ask: + should this node be relocated (in reverse parent-first context)? We repeat this + process as long as the child is the leftmost child, eventually reaching an ancestor of + the flush point that is not a leftmost child. The preceder of that ancestors, which is + not a leftmost child, is actually on the leaf level. The preceder of that block is the + left-neighbor of the flush point. The preceder of that block is the rightmost child of + the twig on the left. So, when alloc_pos_and_ancestors passes upward through the twig + level, it stops momentarily to remember the block of the rightmost child of the twig on + the left and sets it to the flush_position's preceder_hint. + + There is one other place where we may set the flush_position's preceder hint, which is + during scan-left. +*/ +static int set_preceder(const coord_t * coord_in, flush_pos_t * pos) +{ + int ret; + coord_t coord; + lock_handle left_lock; + load_count left_load; + + coord_dup(&coord, coord_in); + + init_lh(&left_lock); + init_load_count(&left_load); + + /* FIXME(B): Same FIXME as in "Find the preceder" in reverse_relocate_test. + coord_is_leftmost_unit is not the right test if the unformatted child is in the + middle of the first extent unit. */ + if (!coord_is_leftmost_unit(&coord)) { + coord_prev_unit(&coord); + } else { + ret = + reiser4_get_left_neighbor(&left_lock, coord.node, + ZNODE_READ_LOCK, GN_SAME_ATOM); + if (ret) { + /* If we fail for any reason it doesn't matter because the + preceder is only a hint. We are low-priority at this point, so + this must be the case. */ + if (ret == -E_REPEAT || ret == -E_NO_NEIGHBOR || + ret == -ENOENT || ret == -EINVAL + || ret == -E_DEADLOCK) { + ret = 0; + } + goto exit; + } + + ret = incr_load_count_znode(&left_load, left_lock.node); + if (ret) + goto exit; + + coord_init_last_unit(&coord, left_lock.node); + } + + ret = + item_utmost_child_real_block(&coord, RIGHT_SIDE, + &pos->preceder.blk); + exit: + check_preceder(pos->preceder.blk); + done_load_count(&left_load); + done_lh(&left_lock); + return ret; +} + +/* MAIN SQUEEZE AND ALLOCATE LOOP (THREE BIG FUNCTIONS) */ + +/* This procedure implements the outer loop of the flush algorithm. To put this in + context, here is the general list of steps taken by the flush routine as a whole: + + 1. Scan-left + 2. Scan-right (maybe) + 3. Allocate initial flush position and its ancestors + 4. + 5. + 6. + + This procedure implements the loop in steps 4 through 6 in the above listing. + + Step 4: if the current flush position is an extent item (position on the twig level), + it allocates the extent (allocate_extent_item_in_place) then shifts to the next + coordinate. If the next coordinate's leftmost child needs flushprep, we will continue. + If the next coordinate is an internal item, we descend back to the leaf level, + otherwise we repeat a step #4 (labeled ALLOC_EXTENTS below). If the "next coordinate" + brings us past the end of the twig level, then we call + reverse_relocate_end_of_twig to possibly dirty the next (right) twig, prior to + step #5 which moves to the right. + + Step 5: calls squalloc_changed_ancestors, which initiates a recursive call up the + tree to allocate any ancestors of the next-right flush position that are not also + ancestors of the current position. Those ancestors (in top-down order) are the next in + parent-first order. We squeeze adjacent nodes on the way up until the right node and + current node share the same parent, then allocate on the way back down. Finally, this + step sets the flush position to the next-right node. Then repeat steps 4 and 5. +*/ + +/* SQUEEZE CODE */ + +/* squalloc_right_twig helper function, cut a range of extent items from + cut node to->node from the beginning up to coord @to. */ +static int squalloc_right_twig_cut(coord_t * to, reiser4_key * to_key, + znode * left) +{ + coord_t from; + reiser4_key from_key; + + coord_init_first_unit(&from, to->node); + item_key_by_coord(&from, &from_key); + + return cut_node_content(&from, to, &from_key, to_key, NULL); +} + +/* Copy as much of the leading extents from @right to @left, allocating + unallocated extents as they are copied. Returns SQUEEZE_TARGET_FULL or + SQUEEZE_SOURCE_EMPTY when no more can be shifted. If the next item is an + internal item it calls shift_one_internal_unit and may then return + SUBTREE_MOVED. */ +static int squeeze_right_twig(znode * left, znode * right, flush_pos_t * pos) +{ + int ret = SUBTREE_MOVED; + coord_t coord; /* used to iterate over items */ + reiser4_key stop_key; + + assert("jmacd-2008", !node_is_empty(right)); + coord_init_first_unit(&coord, right); + + /* FIXME: can be optimized to cut once */ + while (!node_is_empty(coord.node) && item_is_extent(&coord)) { + ON_DEBUG(void *vp); + + assert("vs-1468", coord_is_leftmost_unit(&coord)); + ON_DEBUG(vp = shift_check_prepare(left, coord.node)); + + /* stop_key is used to find what was copied and what to cut */ + stop_key = *reiser4_min_key(); + ret = squalloc_extent(left, &coord, pos, &stop_key); + if (ret != SQUEEZE_CONTINUE) { + ON_DEBUG(kfree(vp)); + break; + } + assert("vs-1465", !keyeq(&stop_key, reiser4_min_key())); + + /* Helper function to do the cutting. */ + set_key_offset(&stop_key, get_key_offset(&stop_key) - 1); + check_me("vs-1466", + squalloc_right_twig_cut(&coord, &stop_key, left) == 0); + + ON_DEBUG(shift_check(vp, left, coord.node)); + } + + if (node_is_empty(coord.node)) + ret = SQUEEZE_SOURCE_EMPTY; + + if (ret == SQUEEZE_TARGET_FULL) { + goto out; + } + + if (node_is_empty(right)) { + /* The whole right node was copied into @left. */ + assert("vs-464", ret == SQUEEZE_SOURCE_EMPTY); + goto out; + } + + coord_init_first_unit(&coord, right); + + if (!item_is_internal(&coord)) { + /* we do not want to squeeze anything else to left neighbor because "slum" + is over */ + ret = SQUEEZE_TARGET_FULL; + goto out; + } + assert("jmacd-433", item_is_internal(&coord)); + + /* Shift an internal unit. The child must be allocated before shifting any more + extents, so we stop here. */ + ret = shift_one_internal_unit(left, right); + + out: + assert("jmacd-8612", ret < 0 || ret == SQUEEZE_TARGET_FULL + || ret == SUBTREE_MOVED || ret == SQUEEZE_SOURCE_EMPTY); + + if (ret == SQUEEZE_TARGET_FULL) { + /* We submit prepped nodes here and expect that this @left twig + * will not be modified again during this jnode_flush() call. */ + int ret1; + + /* NOTE: seems like io is done under long term locks. */ + ret1 = write_prepped_nodes(pos); + if (ret1 < 0) + return ret1; + } + + return ret; +} + +#if REISER4_DEBUG +static void item_convert_invariant(flush_pos_t * pos) +{ + assert("edward-1225", coord_is_existing_item(&pos->coord)); + if (chaining_data_present(pos)) { + item_plugin *iplug = item_convert_plug(pos); + + assert("edward-1000", + iplug == item_plugin_by_coord(&pos->coord)); + assert("edward-1001", iplug->f.convert != NULL); + } else + assert("edward-1226", pos->child == NULL); +} +#else + +#define item_convert_invariant(pos) noop + +#endif + +/* Scan node items starting from the first one and apply for each + item its flush ->convert() method (if any). This method may + resize/kill the item so the tree will be changed. +*/ +static int convert_node(flush_pos_t * pos, znode * node) +{ + int ret = 0; + item_plugin *iplug; + + assert("edward-304", pos != NULL); + assert("edward-305", pos->child == NULL); + assert("edward-475", znode_convertible(node)); + assert("edward-669", znode_is_wlocked(node)); + assert("edward-1210", !node_is_empty(node)); + + if (znode_get_level(node) != LEAF_LEVEL) + /* unsupported */ + goto exit; + + coord_init_first_unit(&pos->coord, node); + + while (1) { + ret = 0; + coord_set_to_left(&pos->coord); + item_convert_invariant(pos); + + iplug = item_plugin_by_coord(&pos->coord); + assert("edward-844", iplug != NULL); + + if (iplug->f.convert) { + ret = iplug->f.convert(pos); + if (ret) + goto exit; + } + assert("edward-307", pos->child == NULL); + + if (coord_next_item(&pos->coord)) { + /* node is over */ + + if (!chaining_data_present(pos)) + /* finished this node */ + break; + if (should_chain_next_node(pos)) { + /* go to next node */ + move_chaining_data(pos, 0 /* to next node */ ); + break; + } + /* repeat this node */ + move_chaining_data(pos, 1 /* this node */ ); + continue; + } + /* Node is not over. + Check if there is attached convert data. + If so roll one item position back and repeat + on this node + */ + if (chaining_data_present(pos)) { + + if (iplug != item_plugin_by_coord(&pos->coord)) + set_item_convert_count(pos, 0); + + ret = coord_prev_item(&pos->coord); + assert("edward-1003", !ret); + + move_chaining_data(pos, 1 /* this node */ ); + } + } + JF_CLR(ZJNODE(node), JNODE_CONVERTIBLE); + znode_make_dirty(node); + exit: + assert("edward-1004", !ret); + return ret; +} + +/* Squeeze and allocate the right neighbor. This is called after @left and + its current children have been squeezed and allocated already. This + procedure's job is to squeeze and items from @right to @left. + + If at the leaf level, use the shift_everything_left memcpy-optimized + version of shifting (squeeze_right_leaf). + + If at the twig level, extents are allocated as they are shifted from @right + to @left (squalloc_right_twig). + + At any other level, shift one internal item and return to the caller + (squalloc_parent_first) so that the shifted-subtree can be processed in + parent-first order. + + When unit of internal item is moved, squeezing stops and SUBTREE_MOVED is + returned. When all content of @right is squeezed, SQUEEZE_SOURCE_EMPTY is + returned. If nothing can be moved into @left anymore, SQUEEZE_TARGET_FULL + is returned. +*/ + +static int squeeze_right_neighbor(flush_pos_t * pos, znode * left, + znode * right) +{ + int ret; + + /* FIXME it is possible to see empty hasn't-heard-banshee node in a + * tree owing to error (for example, ENOSPC) in write */ + /* assert("jmacd-9321", !node_is_empty(left)); */ + assert("jmacd-9322", !node_is_empty(right)); + assert("jmacd-9323", znode_get_level(left) == znode_get_level(right)); + + switch (znode_get_level(left)) { + case TWIG_LEVEL: + /* Shift with extent allocating until either an internal item + is encountered or everything is shifted or no free space + left in @left */ + ret = squeeze_right_twig(left, right, pos); + break; + + default: + /* All other levels can use shift_everything until we implement per-item + flush plugins. */ + ret = squeeze_right_non_twig(left, right); + break; + } + + assert("jmacd-2011", (ret < 0 || + ret == SQUEEZE_SOURCE_EMPTY + || ret == SQUEEZE_TARGET_FULL + || ret == SUBTREE_MOVED)); + return ret; +} + +static int squeeze_right_twig_and_advance_coord(flush_pos_t * pos, + znode * right) +{ + int ret; + + ret = squeeze_right_twig(pos->lock.node, right, pos); + if (ret < 0) + return ret; + if (ret > 0) { + coord_init_after_last_item(&pos->coord, pos->lock.node); + return ret; + } + + coord_init_last_unit(&pos->coord, pos->lock.node); + return 0; +} + +/* forward declaration */ +static int squalloc_upper_levels(flush_pos_t *, znode *, znode *); + +/* do a fast check for "same parents" condition before calling + * squalloc_upper_levels() */ +static inline int check_parents_and_squalloc_upper_levels(flush_pos_t * pos, + znode * left, + znode * right) +{ + if (znode_same_parents(left, right)) + return 0; + + return squalloc_upper_levels(pos, left, right); +} + +/* Check whether the parent of given @right node needs to be processes + ((re)allocated) prior to processing of the child. If @left and @right do not + share at least the parent of the @right is after the @left but before the + @right in parent-first order, we have to (re)allocate it before the @right + gets (re)allocated. */ +static int squalloc_upper_levels(flush_pos_t * pos, znode * left, znode * right) +{ + int ret; + + lock_handle left_parent_lock; + lock_handle right_parent_lock; + + load_count left_parent_load; + load_count right_parent_load; + + init_lh(&left_parent_lock); + init_lh(&right_parent_lock); + + init_load_count(&left_parent_load); + init_load_count(&right_parent_load); + + ret = reiser4_get_parent(&left_parent_lock, left, ZNODE_WRITE_LOCK); + if (ret) + goto out; + + ret = reiser4_get_parent(&right_parent_lock, right, ZNODE_WRITE_LOCK); + if (ret) + goto out; + + /* Check for same parents */ + if (left_parent_lock.node == right_parent_lock.node) + goto out; + + if (znode_check_flushprepped(right_parent_lock.node)) { + /* Keep parent-first order. In the order, the right parent node stands + before the @right node. If it is already allocated, we set the + preceder (next block search start point) to its block number, @right + node should be allocated after it. + + However, preceder is set only if the right parent is on twig level. + The explanation is the following: new branch nodes are allocated over + already allocated children while the tree grows, it is difficult to + keep tree ordered, we assume that only leaves and twings are correctly + allocated. So, only twigs are used as a preceder for allocating of the + rest of the slum. */ + if (znode_get_level(right_parent_lock.node) == TWIG_LEVEL) { + pos->preceder.blk = + *znode_get_block(right_parent_lock.node); + check_preceder(pos->preceder.blk); + } + goto out; + } + + ret = incr_load_count_znode(&left_parent_load, left_parent_lock.node); + if (ret) + goto out; + + ret = incr_load_count_znode(&right_parent_load, right_parent_lock.node); + if (ret) + goto out; + + ret = + squeeze_right_neighbor(pos, left_parent_lock.node, + right_parent_lock.node); + /* We stop if error. We stop if some items/units were shifted (ret == 0) + * and thus @right changed its parent. It means we have not process + * right_parent node prior to processing of @right. Positive return + * values say that shifting items was not happen because of "empty + * source" or "target full" conditions. */ + if (ret <= 0) + goto out; + + /* parent(@left) and parent(@right) may have different parents also. We + * do a recursive call for checking that. */ + ret = + check_parents_and_squalloc_upper_levels(pos, left_parent_lock.node, + right_parent_lock.node); + if (ret) + goto out; + + /* allocate znode when going down */ + ret = lock_parent_and_allocate_znode(right_parent_lock.node, pos); + + out: + done_load_count(&left_parent_load); + done_load_count(&right_parent_load); + + done_lh(&left_parent_lock); + done_lh(&right_parent_lock); + + return ret; +} + +/* Check the leftmost child "flushprepped" status, also returns true if child + * node was not found in cache. */ +static int leftmost_child_of_unit_check_flushprepped(const coord_t * coord) +{ + int ret; + int prepped; + + jnode *child; + + ret = get_leftmost_child_of_unit(coord, &child); + + if (ret) + return ret; + + if (child) { + prepped = jnode_check_flushprepped(child); + jput(child); + } else { + /* We consider not existing child as a node which slum + processing should not continue to. Not cached node is clean, + so it is flushprepped. */ + prepped = 1; + } + + return prepped; +} + +/* (re)allocate znode with automated getting parent node */ +static int lock_parent_and_allocate_znode(znode * node, flush_pos_t * pos) +{ + int ret; + lock_handle parent_lock; + load_count parent_load; + coord_t pcoord; + + assert("zam-851", znode_is_write_locked(node)); + + init_lh(&parent_lock); + init_load_count(&parent_load); + + ret = reiser4_get_parent(&parent_lock, node, ZNODE_WRITE_LOCK); + if (ret) + goto out; + + ret = incr_load_count_znode(&parent_load, parent_lock.node); + if (ret) + goto out; + + ret = find_child_ptr(parent_lock.node, node, &pcoord); + if (ret) + goto out; + + ret = allocate_znode(node, &pcoord, pos); + + out: + done_load_count(&parent_load); + done_lh(&parent_lock); + return ret; +} + +/* Process nodes on leaf level until unformatted node or rightmost node in the + * slum reached. */ +static int handle_pos_on_formatted(flush_pos_t * pos) +{ + int ret; + lock_handle right_lock; + load_count right_load; + + init_lh(&right_lock); + init_load_count(&right_load); + + if (should_convert_node(pos, pos->lock.node)) { + ret = convert_node(pos, pos->lock.node); + if (ret) + return ret; + } + + while (1) { + int expected; + expected = should_convert_next_node(pos); + ret = neighbor_in_slum(pos->lock.node, &right_lock, RIGHT_SIDE, + ZNODE_WRITE_LOCK, !expected, expected); + if (ret) { + if (expected) + warning("edward-1495", + "Expected neighbor not found (ret = %d). Fsck?", + ret); + break; + } + + /* we don't prep(allocate) nodes for flushing twice. This can be suboptimal, or it + * can be optimal. For now we choose to live with the risk that it will + * be suboptimal because it would be quite complex to code it to be + * smarter. */ + if (znode_check_flushprepped(right_lock.node) + && !znode_convertible(right_lock.node)) { + assert("edward-1005", !should_convert_next_node(pos)); + pos_stop(pos); + break; + } + + ret = incr_load_count_znode(&right_load, right_lock.node); + if (ret) + break; + if (should_convert_node(pos, right_lock.node)) { + ret = convert_node(pos, right_lock.node); + if (ret) + break; + if (node_is_empty(right_lock.node)) { + /* node became empty after converting, repeat */ + done_load_count(&right_load); + done_lh(&right_lock); + continue; + } + } + + /* squeeze _before_ going upward. */ + ret = + squeeze_right_neighbor(pos, pos->lock.node, + right_lock.node); + if (ret < 0) + break; + + if (znode_check_flushprepped(right_lock.node)) { + if (should_convert_next_node(pos)) { + /* in spite of flushprepped status of the node, + its right slum neighbor should be converted */ + assert("edward-953", convert_data(pos)); + assert("edward-954", item_convert_data(pos)); + + if (node_is_empty(right_lock.node)) { + done_load_count(&right_load); + done_lh(&right_lock); + } else + move_flush_pos(pos, &right_lock, + &right_load, NULL); + continue; + } + pos_stop(pos); + break; + } + + if (node_is_empty(right_lock.node)) { + /* repeat if right node was squeezed completely */ + done_load_count(&right_load); + done_lh(&right_lock); + continue; + } + + /* parent(right_lock.node) has to be processed before + * (right_lock.node) due to "parent-first" allocation order. */ + ret = + check_parents_and_squalloc_upper_levels(pos, pos->lock.node, + right_lock.node); + if (ret) + break; + /* (re)allocate _after_ going upward */ + ret = lock_parent_and_allocate_znode(right_lock.node, pos); + if (ret) + break; + if (should_terminate_squalloc(pos)) { + set_item_convert_count(pos, 0); + break; + } + + /* advance the flush position to the right neighbor */ + move_flush_pos(pos, &right_lock, &right_load, NULL); + + ret = rapid_flush(pos); + if (ret) + break; + } + check_convert_info(pos); + done_load_count(&right_load); + done_lh(&right_lock); + + /* This function indicates via pos whether to stop or go to twig or continue on current + * level. */ + return ret; + +} + +/* Process nodes on leaf level until unformatted node or rightmost node in the + * slum reached. */ +static int handle_pos_on_leaf(flush_pos_t * pos) +{ + int ret; + + assert("zam-845", pos->state == POS_ON_LEAF); + + ret = handle_pos_on_formatted(pos); + + if (ret == -E_NO_NEIGHBOR) { + /* cannot get right neighbor, go process extents. */ + pos->state = POS_TO_TWIG; + return 0; + } + + return ret; +} + +/* Process slum on level > 1 */ +static int handle_pos_on_internal(flush_pos_t * pos) +{ + assert("zam-850", pos->state == POS_ON_INTERNAL); + return handle_pos_on_formatted(pos); +} + +/* check whether squalloc should stop before processing given extent */ +static int squalloc_extent_should_stop(flush_pos_t * pos) +{ + assert("zam-869", item_is_extent(&pos->coord)); + + /* pos->child is a jnode handle_pos_on_extent() should start with in + * stead of the first child of the first extent unit. */ + if (pos->child) { + int prepped; + + assert("vs-1383", jnode_is_unformatted(pos->child)); + prepped = jnode_check_flushprepped(pos->child); + pos->pos_in_unit = + jnode_get_index(pos->child) - + extent_unit_index(&pos->coord); + assert("vs-1470", + pos->pos_in_unit < extent_unit_width(&pos->coord)); + assert("nikita-3434", + ergo(extent_is_unallocated(&pos->coord), + pos->pos_in_unit == 0)); + jput(pos->child); + pos->child = NULL; + + return prepped; + } + + pos->pos_in_unit = 0; + if (extent_is_unallocated(&pos->coord)) + return 0; + + return leftmost_child_of_unit_check_flushprepped(&pos->coord); +} + +/* Handle the case when regular reiser4 tree (znodes connected one to its + * neighbors by sibling pointers) is interrupted on leaf level by one or more + * unformatted nodes. By having a lock on twig level and use extent code + * routines to process unformatted nodes we swim around an irregular part of + * reiser4 tree. */ +static int handle_pos_on_twig(flush_pos_t * pos) +{ + int ret; + + assert("zam-844", pos->state == POS_ON_EPOINT); + assert("zam-843", item_is_extent(&pos->coord)); + + /* We decide should we continue slum processing with current extent + unit: if leftmost child of current extent unit is flushprepped + (i.e. clean or already processed by flush) we stop squalloc(). There + is a fast check for unallocated extents which we assume contain all + not flushprepped nodes. */ + /* FIXME: Here we implement simple check, we are only looking on the + leftmost child. */ + ret = squalloc_extent_should_stop(pos); + if (ret != 0) { + pos_stop(pos); + return ret; + } + + while (pos_valid(pos) && coord_is_existing_unit(&pos->coord) + && item_is_extent(&pos->coord)) { + ret = reiser4_alloc_extent(pos); + if (ret) { + break; + } + coord_next_unit(&pos->coord); + } + + if (coord_is_after_rightmost(&pos->coord)) { + pos->state = POS_END_OF_TWIG; + return 0; + } + if (item_is_internal(&pos->coord)) { + pos->state = POS_TO_LEAF; + return 0; + } + + assert("zam-860", item_is_extent(&pos->coord)); + + /* "slum" is over */ + pos->state = POS_INVALID; + return 0; +} + +/* When we about to return flush position from twig to leaf level we can process + * the right twig node or move position to the leaf. This processes right twig + * if it is possible and jump to leaf level if not. */ +static int handle_pos_end_of_twig(flush_pos_t * pos) +{ + int ret; + lock_handle right_lock; + load_count right_load; + coord_t at_right; + jnode *child = NULL; + + assert("zam-848", pos->state == POS_END_OF_TWIG); + assert("zam-849", coord_is_after_rightmost(&pos->coord)); + + init_lh(&right_lock); + init_load_count(&right_load); + + /* We get a lock on the right twig node even it is not dirty because + * slum continues or discontinues on leaf level not on next twig. This + * lock on the right twig is needed for getting its leftmost child. */ + ret = + reiser4_get_right_neighbor(&right_lock, pos->lock.node, + ZNODE_WRITE_LOCK, GN_SAME_ATOM); + if (ret) + goto out; + + ret = incr_load_count_znode(&right_load, right_lock.node); + if (ret) + goto out; + + /* right twig could be not dirty */ + if (JF_ISSET(ZJNODE(right_lock.node), JNODE_DIRTY)) { + /* If right twig node is dirty we always attempt to squeeze it + * content to the left... */ + became_dirty: + ret = + squeeze_right_twig_and_advance_coord(pos, right_lock.node); + if (ret <= 0) { + /* pos->coord is on internal item, go to leaf level, or + * we have an error which will be caught in squalloc() */ + pos->state = POS_TO_LEAF; + goto out; + } + + /* If right twig was squeezed completely we wave to re-lock + * right twig. now it is done through the top-level squalloc + * routine. */ + if (node_is_empty(right_lock.node)) + goto out; + + /* ... and prep it if it is not yet prepped */ + if (!znode_check_flushprepped(right_lock.node)) { + /* As usual, process parent before ... */ + ret = + check_parents_and_squalloc_upper_levels(pos, + pos->lock. + node, + right_lock. + node); + if (ret) + goto out; + + /* ... processing the child */ + ret = + lock_parent_and_allocate_znode(right_lock.node, + pos); + if (ret) + goto out; + } + } else { + coord_init_first_unit(&at_right, right_lock.node); + + /* check first child of next twig, should we continue there ? */ + ret = get_leftmost_child_of_unit(&at_right, &child); + if (ret || child == NULL || jnode_check_flushprepped(child)) { + pos_stop(pos); + goto out; + } + + /* check clean twig for possible relocation */ + if (!znode_check_flushprepped(right_lock.node)) { + ret = + reverse_relocate_check_dirty_parent(child, + &at_right, pos); + if (ret) + goto out; + if (JF_ISSET(ZJNODE(right_lock.node), JNODE_DIRTY)) + goto became_dirty; + } + } + + assert("zam-875", znode_check_flushprepped(right_lock.node)); + + /* Update the preceder by a block number of just processed right twig + * node. The code above could miss the preceder updating because + * allocate_znode() could not be called for this node. */ + pos->preceder.blk = *znode_get_block(right_lock.node); + check_preceder(pos->preceder.blk); + + coord_init_first_unit(&at_right, right_lock.node); + assert("zam-868", coord_is_existing_unit(&at_right)); + + pos->state = item_is_extent(&at_right) ? POS_ON_EPOINT : POS_TO_LEAF; + move_flush_pos(pos, &right_lock, &right_load, &at_right); + + out: + done_load_count(&right_load); + done_lh(&right_lock); + + if (child) + jput(child); + + return ret; +} + +/* Move the pos->lock to leaf node pointed by pos->coord, check should we + * continue there. */ +static int handle_pos_to_leaf(flush_pos_t * pos) +{ + int ret; + lock_handle child_lock; + load_count child_load; + jnode *child; + + assert("zam-846", pos->state == POS_TO_LEAF); + assert("zam-847", item_is_internal(&pos->coord)); + + init_lh(&child_lock); + init_load_count(&child_load); + + ret = get_leftmost_child_of_unit(&pos->coord, &child); + if (ret) + return ret; + if (child == NULL) { + pos_stop(pos); + return 0; + } + + if (jnode_check_flushprepped(child)) { + pos->state = POS_INVALID; + goto out; + } + + ret = + longterm_lock_znode(&child_lock, JZNODE(child), ZNODE_WRITE_LOCK, + ZNODE_LOCK_LOPRI); + if (ret) + goto out; + + ret = incr_load_count_znode(&child_load, JZNODE(child)); + if (ret) + goto out; + + ret = allocate_znode(JZNODE(child), &pos->coord, pos); + if (ret) + goto out; + + /* move flush position to leaf level */ + pos->state = POS_ON_LEAF; + move_flush_pos(pos, &child_lock, &child_load, NULL); + + if (node_is_empty(JZNODE(child))) { + ret = delete_empty_node(JZNODE(child)); + pos->state = POS_INVALID; + } + out: + done_load_count(&child_load); + done_lh(&child_lock); + jput(child); + + return ret; +} + +/* move pos from leaf to twig, and move lock from leaf to twig. */ +/* Move pos->lock to upper (twig) level */ +static int handle_pos_to_twig(flush_pos_t * pos) +{ + int ret; + + lock_handle parent_lock; + load_count parent_load; + coord_t pcoord; + + assert("zam-852", pos->state == POS_TO_TWIG); + + init_lh(&parent_lock); + init_load_count(&parent_load); + + ret = + reiser4_get_parent(&parent_lock, pos->lock.node, ZNODE_WRITE_LOCK); + if (ret) + goto out; + + ret = incr_load_count_znode(&parent_load, parent_lock.node); + if (ret) + goto out; + + ret = find_child_ptr(parent_lock.node, pos->lock.node, &pcoord); + if (ret) + goto out; + + assert("zam-870", item_is_internal(&pcoord)); + coord_next_item(&pcoord); + + if (coord_is_after_rightmost(&pcoord)) + pos->state = POS_END_OF_TWIG; + else if (item_is_extent(&pcoord)) + pos->state = POS_ON_EPOINT; + else { + /* Here we understand that getting -E_NO_NEIGHBOR in + * handle_pos_on_leaf() was because of just a reaching edge of + * slum */ + pos_stop(pos); + goto out; + } + + move_flush_pos(pos, &parent_lock, &parent_load, &pcoord); + + out: + done_load_count(&parent_load); + done_lh(&parent_lock); + + return ret; +} + +typedef int (*pos_state_handle_t) (flush_pos_t *); +static pos_state_handle_t flush_pos_handlers[] = { + /* process formatted nodes on leaf level, keep lock on a leaf node */ + [POS_ON_LEAF] = handle_pos_on_leaf, + /* process unformatted nodes, keep lock on twig node, pos->coord points to extent currently + * being processed */ + [POS_ON_EPOINT] = handle_pos_on_twig, + /* move a lock from leaf node to its parent for further processing of unformatted nodes */ + [POS_TO_TWIG] = handle_pos_to_twig, + /* move a lock from twig to leaf level when a processing of unformatted nodes finishes, + * pos->coord points to the leaf node we jump to */ + [POS_TO_LEAF] = handle_pos_to_leaf, + /* after processing last extent in the twig node, attempting to shift items from the twigs + * right neighbor and process them while shifting */ + [POS_END_OF_TWIG] = handle_pos_end_of_twig, + /* process formatted nodes on internal level, keep lock on an internal node */ + [POS_ON_INTERNAL] = handle_pos_on_internal +}; + +/* Advance flush position horizontally, prepare for flushing ((re)allocate, squeeze, + * encrypt) nodes and their ancestors in "parent-first" order */ +static int squalloc(flush_pos_t * pos) +{ + int ret = 0; + + /* maybe needs to be made a case statement with handle_pos_on_leaf as first case, for + * greater CPU efficiency? Measure and see.... -Hans */ + while (pos_valid(pos)) { + ret = flush_pos_handlers[pos->state] (pos); + if (ret < 0) + break; + + ret = rapid_flush(pos); + if (ret) + break; + } + + /* any positive value or -E_NO_NEIGHBOR are legal return codes for handle_pos* + routines, -E_NO_NEIGHBOR means that slum edge was reached */ + if (ret > 0 || ret == -E_NO_NEIGHBOR) + ret = 0; + + return ret; +} + +static void update_ldkey(znode * node) +{ + reiser4_key ldkey; + + assert_rw_write_locked(&(znode_get_tree(node)->dk_lock)); + if (node_is_empty(node)) + return; + + znode_set_ld_key(node, leftmost_key_in_node(node, &ldkey)); +} + +/* this is to be called after calling of shift node's method to shift data from @right to + @left. It sets left delimiting keys of @left and @right to keys of first items of @left + and @right correspondingly and sets right delimiting key of @left to first key of @right */ +static void update_znode_dkeys(znode * left, znode * right) +{ + assert_rw_write_locked(&(znode_get_tree(right)->dk_lock)); + assert("vs-1629", (znode_is_write_locked(left) && + znode_is_write_locked(right))); + + /* we need to update left delimiting of left if it was empty before shift */ + update_ldkey(left); + update_ldkey(right); + if (node_is_empty(right)) + znode_set_rd_key(left, znode_get_rd_key(right)); + else + znode_set_rd_key(left, znode_get_ld_key(right)); +} + +/* try to shift everything from @right to @left. If everything was shifted - + @right is removed from the tree. Result is the number of bytes shifted. */ +static int +shift_everything_left(znode * right, znode * left, carry_level * todo) +{ + coord_t from; + node_plugin *nplug; + carry_plugin_info info; + + coord_init_after_last_item(&from, right); + + nplug = node_plugin_by_node(right); + info.doing = NULL; + info.todo = todo; + return nplug->shift(&from, left, SHIFT_LEFT, + 1 /* delete @right if it becomes empty */ , + 1 + /* move coord @from to node @left if everything will be shifted */ + , + &info); +} + +/* Shift as much as possible from @right to @left using the memcpy-optimized + shift_everything_left. @left and @right are formatted neighboring nodes on + leaf level. */ +static int squeeze_right_non_twig(znode * left, znode * right) +{ + int ret; + carry_pool *pool; + carry_level *todo; + + assert("nikita-2246", znode_get_level(left) == znode_get_level(right)); + + if (!JF_ISSET(ZJNODE(left), JNODE_DIRTY) || + !JF_ISSET(ZJNODE(right), JNODE_DIRTY)) + return SQUEEZE_TARGET_FULL; + + pool = init_carry_pool(sizeof(*pool) + 3 * sizeof(*todo)); + if (IS_ERR(pool)) + return PTR_ERR(pool); + todo = (carry_level *) (pool + 1); + init_carry_level(todo, pool); + + ret = shift_everything_left(right, left, todo); + if (ret > 0) { + /* something was shifted */ + reiser4_tree *tree; + __u64 grabbed; + + znode_make_dirty(left); + znode_make_dirty(right); + + /* update delimiting keys of nodes which participated in + shift. FIXME: it would be better to have this in shift + node's operation. But it can not be done there. Nobody + remembers why, though */ + tree = znode_get_tree(left); + write_lock_dk(tree); + update_znode_dkeys(left, right); + write_unlock_dk(tree); + + /* Carry is called to update delimiting key and, maybe, to remove empty + node. */ + grabbed = get_current_context()->grabbed_blocks; + ret = reiser4_grab_space_force(tree->height, BA_RESERVED); + assert("nikita-3003", ret == 0); /* reserved space is exhausted. Ask Hans. */ + ret = reiser4_carry(todo, NULL /* previous level */ ); + grabbed2free_mark(grabbed); + } else { + /* Shifting impossible, we return appropriate result code */ + ret = + node_is_empty(right) ? SQUEEZE_SOURCE_EMPTY : + SQUEEZE_TARGET_FULL; + } + + done_carry_pool(pool); + + return ret; +} + +#if REISER4_DEBUG +static int sibling_link_is_ok(const znode *left, const znode *right) +{ + int result; + + read_lock_tree(znode_get_tree(left)); + result = (left->right == right && left == right->left); + read_unlock_tree(znode_get_tree(left)); + return result; +} +#endif + +/* Shift first unit of first item if it is an internal one. Return + SQUEEZE_TARGET_FULL if it fails to shift an item, otherwise return + SUBTREE_MOVED. */ +static int shift_one_internal_unit(znode * left, znode * right) +{ + int ret; + carry_pool *pool; + carry_level *todo; + coord_t *coord; + carry_plugin_info *info; + int size, moved; + + assert("nikita-2247", znode_get_level(left) == znode_get_level(right)); + assert("nikita-2435", znode_is_write_locked(left)); + assert("nikita-2436", znode_is_write_locked(right)); + assert("nikita-2434", sibling_link_is_ok(left, right)); + + pool = init_carry_pool(sizeof(*pool) + 3 * sizeof(*todo) + + sizeof(*coord) + sizeof(*info) +#if REISER4_DEBUG + + sizeof(*coord) + 2 * sizeof(reiser4_key) +#endif + ); + if (IS_ERR(pool)) + return PTR_ERR(pool); + todo = (carry_level *) (pool + 1); + init_carry_level(todo, pool); + + coord = (coord_t *) (todo + 3); + coord_init_first_unit(coord, right); + info = (carry_plugin_info *) (coord + 1); + +#if REISER4_DEBUG + if (!node_is_empty(left)) { + coord_t *last; + reiser4_key *right_key; + reiser4_key *left_key; + + last = (coord_t *) (info + 1); + right_key = (reiser4_key *) (last + 1); + left_key = right_key + 1; + coord_init_last_unit(last, left); + + assert("nikita-2463", + keyle(item_key_by_coord(last, left_key), + item_key_by_coord(coord, right_key))); + } +#endif + + assert("jmacd-2007", item_is_internal(coord)); + + size = item_length_by_coord(coord); + info->todo = todo; + info->doing = NULL; + + ret = node_plugin_by_node(left)->shift(coord, left, SHIFT_LEFT, + 1 + /* delete @right if it becomes empty */ + , + 0 + /* do not move coord @coord to node @left */ + , + info); + + /* If shift returns positive, then we shifted the item. */ + assert("vs-423", ret <= 0 || size == ret); + moved = (ret > 0); + + if (moved) { + /* something was moved */ + reiser4_tree *tree; + int grabbed; + + znode_make_dirty(left); + znode_make_dirty(right); + tree = znode_get_tree(left); + write_lock_dk(tree); + update_znode_dkeys(left, right); + write_unlock_dk(tree); + + /* reserve space for delimiting keys after shifting */ + grabbed = get_current_context()->grabbed_blocks; + ret = reiser4_grab_space_force(tree->height, BA_RESERVED); + assert("nikita-3003", ret == 0); /* reserved space is exhausted. Ask Hans. */ + + ret = reiser4_carry(todo, NULL /* previous level */ ); + grabbed2free_mark(grabbed); + } + + done_carry_pool(pool); + + if (ret != 0) { + /* Shift or carry operation failed. */ + assert("jmacd-7325", ret < 0); + return ret; + } + + return moved ? SUBTREE_MOVED : SQUEEZE_TARGET_FULL; +} + +/* Make the final relocate/wander decision during forward parent-first squalloc for a + znode. For unformatted nodes this is done in plugin/item/extent.c:extent_needs_allocation(). */ +static int +allocate_znode_loaded(znode * node, + const coord_t * parent_coord, flush_pos_t * pos) +{ + int ret; + reiser4_super_info_data *sbinfo = get_current_super_private(); + /* FIXME(D): We have the node write-locked and should have checked for ! + allocated() somewhere before reaching this point, but there can be a race, so + this assertion is bogus. */ + assert("jmacd-7987", !jnode_check_flushprepped(ZJNODE(node))); + assert("jmacd-7988", znode_is_write_locked(node)); + assert("jmacd-7989", coord_is_invalid(parent_coord) + || znode_is_write_locked(parent_coord->node)); + + if (ZF_ISSET(node, JNODE_REPACK) || ZF_ISSET(node, JNODE_CREATED) || + znode_is_root(node) || + /* We have enough nodes to relocate no matter what. */ + (pos->leaf_relocate != 0 && znode_get_level(node) == LEAF_LEVEL)) { + /* No need to decide with new nodes, they are treated the same as + relocate. If the root node is dirty, relocate. */ + if (pos->preceder.blk == 0) { + /* preceder is unknown and we have decided to relocate node -- + using of default value for search start is better than search + from block #0. */ + get_blocknr_hint_default(&pos->preceder.blk); + check_preceder(pos->preceder.blk); + } + + goto best_reloc; + + } else if (pos->preceder.blk == 0) { + /* If we don't know the preceder, leave it where it is. */ + jnode_make_wander(ZJNODE(node)); + } else { + /* Make a decision based on block distance. */ + reiser4_block_nr dist; + reiser4_block_nr nblk = *znode_get_block(node); + + assert("jmacd-6172", !reiser4_blocknr_is_fake(&nblk)); + assert("jmacd-6173", !reiser4_blocknr_is_fake(&pos->preceder.blk)); + assert("jmacd-6174", pos->preceder.blk != 0); + + if (pos->preceder.blk == nblk - 1) { + /* Ideal. */ + jnode_make_wander(ZJNODE(node)); + } else { + + dist = + (nblk < + pos->preceder.blk) ? (pos->preceder.blk - + nblk) : (nblk - + pos->preceder.blk); + + /* See if we can find a closer block (forward direction only). */ + pos->preceder.max_dist = + min((reiser4_block_nr) sbinfo->flush. + relocate_distance, dist); + pos->preceder.level = znode_get_level(node); + + ret = allocate_znode_update(node, parent_coord, pos); + + pos->preceder.max_dist = 0; + + if (ret && (ret != -ENOSPC)) + return ret; + + if (ret == 0) { + /* Got a better allocation. */ + znode_make_reloc(node, pos->fq); + } else if (dist < sbinfo->flush.relocate_distance) { + /* The present allocation is good enough. */ + jnode_make_wander(ZJNODE(node)); + } else { + /* Otherwise, try to relocate to the best position. */ + best_reloc: + ret = + allocate_znode_update(node, parent_coord, + pos); + if (ret != 0) + return ret; + + /* set JNODE_RELOC bit _after_ node gets allocated */ + znode_make_reloc(node, pos->fq); + } + } + } + + /* This is the new preceder. */ + pos->preceder.blk = *znode_get_block(node); + check_preceder(pos->preceder.blk); + pos->alloc_cnt += 1; + + assert("jmacd-4277", !reiser4_blocknr_is_fake(&pos->preceder.blk)); + + return 0; +} + +static int +allocate_znode(znode * node, const coord_t * parent_coord, flush_pos_t * pos) +{ + /* + * perform znode allocation with znode pinned in memory to avoid races + * with asynchronous emergency flush (which plays with + * JNODE_FLUSH_RESERVED bit). + */ + return WITH_DATA(node, allocate_znode_loaded(node, parent_coord, pos)); +} + +/* A subroutine of allocate_znode, this is called first to see if there is a close + position to relocate to. It may return ENOSPC if there is no close position. If there + is no close position it may not relocate. This takes care of updating the parent node + with the relocated block address. */ +static int +allocate_znode_update(znode * node, const coord_t * parent_coord, + flush_pos_t * pos) +{ + int ret; + reiser4_block_nr blk; + lock_handle uber_lock; + int flush_reserved_used = 0; + int grabbed; + reiser4_context *ctx; + reiser4_super_info_data *sbinfo; + + init_lh(&uber_lock); + + ctx = get_current_context(); + sbinfo = get_super_private(ctx->super); + + grabbed = ctx->grabbed_blocks; + + /* discard e-flush allocation */ + ret = zload(node); + if (ret) + return ret; + + if (ZF_ISSET(node, JNODE_CREATED)) { + assert("zam-816", reiser4_blocknr_is_fake(znode_get_block(node))); + pos->preceder.block_stage = BLOCK_UNALLOCATED; + } else { + pos->preceder.block_stage = BLOCK_GRABBED; + + /* The disk space for relocating the @node is already reserved in "flush reserved" + * counter if @node is leaf, otherwise we grab space using BA_RESERVED (means grab + * space from whole disk not from only 95%). */ + if (znode_get_level(node) == LEAF_LEVEL) { + /* + * earlier (during do_jnode_make_dirty()) we decided + * that @node can possibly go into overwrite set and + * reserved block for its wandering location. + */ + txn_atom *atom = get_current_atom_locked(); + assert("nikita-3449", + ZF_ISSET(node, JNODE_FLUSH_RESERVED)); + flush_reserved2grabbed(atom, (__u64) 1); + spin_unlock_atom(atom); + /* + * we are trying to move node into relocate + * set. Allocation of relocated position "uses" + * reserved block. + */ + ZF_CLR(node, JNODE_FLUSH_RESERVED); + flush_reserved_used = 1; + } else { + ret = reiser4_grab_space_force((__u64) 1, BA_RESERVED); + if (ret != 0) + goto exit; + } + } + + /* We may do not use 5% of reserved disk space here and flush will not pack tightly. */ + ret = reiser4_alloc_block(&pos->preceder, &blk, + BA_FORMATTED | BA_PERMANENT); + if (ret) + goto exit; + + if (!ZF_ISSET(node, JNODE_CREATED) && + (ret = + reiser4_dealloc_block(znode_get_block(node), 0, + BA_DEFER | BA_FORMATTED))) + goto exit; + + if (likely(!znode_is_root(node))) { + item_plugin *iplug; + + iplug = item_plugin_by_coord(parent_coord); + assert("nikita-2954", iplug->f.update != NULL); + iplug->f.update(parent_coord, &blk); + + znode_make_dirty(parent_coord->node); + + } else { + reiser4_tree *tree = znode_get_tree(node); + znode *uber; + + /* We take a longterm lock on the fake node in order to change + the root block number. This may cause atom fusion. */ + ret = get_uber_znode(tree, ZNODE_WRITE_LOCK, ZNODE_LOCK_HIPRI, + &uber_lock); + /* The fake node cannot be deleted, and we must have priority + here, and may not be confused with ENOSPC. */ + assert("jmacd-74412", + ret != -EINVAL && ret != -E_DEADLOCK && ret != -ENOSPC); + + if (ret) + goto exit; + + uber = uber_lock.node; + + write_lock_tree(tree); + tree->root_block = blk; + write_unlock_tree(tree); + + znode_make_dirty(uber); + } + + ret = znode_rehash(node, &blk); + exit: + if (ret) { + /* Get flush reserved block back if something fails, because + * callers assume that on error block wasn't relocated and its + * flush reserved block wasn't used. */ + if (flush_reserved_used) { + /* + * ok, we failed to move node into relocate + * set. Restore status quo. + */ + grabbed2flush_reserved((__u64) 1); + ZF_SET(node, JNODE_FLUSH_RESERVED); + } + } + zrelse(node); + done_lh(&uber_lock); + grabbed2free_mark(grabbed); + return ret; +} + +/* JNODE INTERFACE */ + +/* Lock a node (if formatted) and then get its parent locked, set the child's + coordinate in the parent. If the child is the root node, the above_root + znode is returned but the coord is not set. This function may cause atom + fusion, but it is only used for read locks (at this point) and therefore + fusion only occurs when the parent is already dirty. */ +/* Hans adds this note: remember to ask how expensive this operation is vs. storing parent + pointer in jnodes. */ +static int +jnode_lock_parent_coord(jnode * node, + coord_t * coord, + lock_handle * parent_lh, + load_count * parent_zh, + znode_lock_mode parent_mode, int try) +{ + int ret; + + assert("edward-53", jnode_is_unformatted(node) || jnode_is_znode(node)); + assert("edward-54", jnode_is_unformatted(node) + || znode_is_any_locked(JZNODE(node))); + + if (!jnode_is_znode(node)) { + reiser4_key key; + tree_level stop_level = TWIG_LEVEL; + lookup_bias bias = FIND_EXACT; + + assert("edward-168", !(jnode_get_type(node) == JNODE_BITMAP)); + + /* The case when node is not znode, but can have parent coord + (unformatted node, node which represents cluster page, + etc..). Generate a key for the appropriate entry, search + in the tree using coord_by_key, which handles locking for + us. */ + + /* + * nothing is locked at this moment, so, nothing prevents + * concurrent truncate from removing jnode from inode. To + * prevent this spin-lock jnode. jnode can be truncated just + * after call to the jnode_build_key(), but this is ok, + * because coord_by_key() will just fail to find appropriate + * extent. + */ + spin_lock_jnode(node); + if (!JF_ISSET(node, JNODE_HEARD_BANSHEE)) { + jnode_build_key(node, &key); + ret = 0; + } else + ret = RETERR(-ENOENT); + spin_unlock_jnode(node); + + if (ret != 0) + return ret; + + if (jnode_is_cluster_page(node)) + stop_level = LEAF_LEVEL; + + assert("jmacd-1812", coord != NULL); + + ret = coord_by_key(jnode_get_tree(node), &key, coord, parent_lh, + parent_mode, bias, stop_level, stop_level, + CBK_UNIQUE, NULL /*ra_info */ ); + switch (ret) { + case CBK_COORD_NOTFOUND: + assert("edward-1038", + ergo(jnode_is_cluster_page(node), + JF_ISSET(node, JNODE_HEARD_BANSHEE))); + if (!JF_ISSET(node, JNODE_HEARD_BANSHEE)) + warning("nikita-3177", "Parent not found"); + return ret; + case CBK_COORD_FOUND: + if (coord->between != AT_UNIT) { + /* FIXME: comment needed */ + done_lh(parent_lh); + if (!JF_ISSET(node, JNODE_HEARD_BANSHEE)) { + warning("nikita-3178", + "Found but not happy: %i", + coord->between); + } + return RETERR(-ENOENT); + } + ret = incr_load_count_znode(parent_zh, parent_lh->node); + if (ret != 0) + return ret; + /* if (jnode_is_cluster_page(node)) { + races with write() are possible + check_child_cluster (parent_lh->node); + } + */ + break; + default: + return ret; + } + + } else { + int flags; + znode *z; + + z = JZNODE(node); + /* Formatted node case: */ + assert("jmacd-2061", !znode_is_root(z)); + + flags = GN_ALLOW_NOT_CONNECTED; + if (try) + flags |= GN_TRY_LOCK; + + ret = + reiser4_get_parent_flags(parent_lh, z, parent_mode, flags); + if (ret != 0) + /* -E_REPEAT is ok here, it is handled by the caller. */ + return ret; + + /* Make the child's position "hint" up-to-date. (Unless above + root, which caller must check.) */ + if (coord != NULL) { + + ret = incr_load_count_znode(parent_zh, parent_lh->node); + if (ret != 0) { + warning("jmacd-976812386", + "incr_load_count_znode failed: %d", + ret); + return ret; + } + + ret = find_child_ptr(parent_lh->node, z, coord); + if (ret != 0) { + warning("jmacd-976812", + "find_child_ptr failed: %d", ret); + return ret; + } + } + } + + return 0; +} + +/* Get the (locked) next neighbor of a znode which is dirty and a member of the same atom. + If there is no next neighbor or the neighbor is not in memory or if there is a + neighbor but it is not dirty or not in the same atom, -E_NO_NEIGHBOR is returned. + In some cases the slum may include nodes which are not dirty, if so @check_dirty should be 0 */ +static int neighbor_in_slum(znode * node, /* starting point */ + lock_handle * lock, /* lock on starting point */ + sideof side, /* left or right direction we seek the next node in */ + znode_lock_mode mode, /* kind of lock we want */ + int check_dirty, /* true if the neighbor should be dirty */ + int use_upper_levels /* get neighbor by going though + upper levels */) +{ + int ret; + int flags; + + assert("jmacd-6334", znode_is_connected(node)); + + flags = GN_SAME_ATOM | (side == LEFT_SIDE ? GN_GO_LEFT : 0); + if (use_upper_levels) + flags |= GN_CAN_USE_UPPER_LEVELS; + + ret = reiser4_get_neighbor(lock, node, mode, flags); + if (ret) { + /* May return -ENOENT or -E_NO_NEIGHBOR. */ + /* FIXME(C): check EINVAL, E_DEADLOCK */ + if (ret == -ENOENT) { + ret = RETERR(-E_NO_NEIGHBOR); + } + return ret; + } + if (!check_dirty) + return 0; + /* Check dirty bit of locked znode, no races here */ + if (JF_ISSET(ZJNODE(lock->node), JNODE_DIRTY)) + return 0; + + done_lh(lock); + return RETERR(-E_NO_NEIGHBOR); +} + +/* Return true if two znodes have the same parent. This is called with both nodes + write-locked (for squeezing) so no tree lock is needed. */ +static int znode_same_parents(znode * a, znode * b) +{ + int result; + + assert("jmacd-7011", znode_is_write_locked(a)); + assert("jmacd-7012", znode_is_write_locked(b)); + + /* We lock the whole tree for this check.... I really don't like whole tree + * locks... -Hans */ + read_lock_tree(znode_get_tree(a)); + result = (znode_parent(a) == znode_parent(b)); + read_unlock_tree(znode_get_tree(a)); + return result; +} + +/* FLUSH SCAN */ + +/* Initialize the flush_scan data structure. */ +static void scan_init(flush_scan * scan) +{ + memset(scan, 0, sizeof(*scan)); + init_lh(&scan->node_lock); + init_lh(&scan->parent_lock); + init_load_count(&scan->parent_load); + init_load_count(&scan->node_load); + coord_init_invalid(&scan->parent_coord, NULL); +} + +/* Release any resources held by the flush scan, e.g., release locks, free memory, etc. */ +static void scan_done(flush_scan * scan) +{ + done_load_count(&scan->node_load); + if (scan->node != NULL) { + jput(scan->node); + scan->node = NULL; + } + done_load_count(&scan->parent_load); + done_lh(&scan->parent_lock); + done_lh(&scan->node_lock); +} + +/* Returns true if flush scanning is finished. */ +int reiser4_scan_finished(flush_scan * scan) +{ + return scan->stop || (scan->direction == RIGHT_SIDE && + scan->count >= scan->max_count); +} + +/* Return true if the scan should continue to the @tonode. True if the node meets the + same_slum_check condition. If not, deref the "left" node and stop the scan. */ +int reiser4_scan_goto(flush_scan * scan, jnode * tonode) +{ + int go = same_slum_check(scan->node, tonode, 1, 0); + + if (!go) { + scan->stop = 1; + jput(tonode); + } + + return go; +} + +/* Set the current scan->node, refcount it, increment count by the @add_count (number to + count, e.g., skipped unallocated nodes), deref previous current, and copy the current + parent coordinate. */ +int +scan_set_current(flush_scan * scan, jnode * node, unsigned add_count, + const coord_t * parent) +{ + /* Release the old references, take the new reference. */ + done_load_count(&scan->node_load); + + if (scan->node != NULL) { + jput(scan->node); + } + scan->node = node; + scan->count += add_count; + + /* This next stmt is somewhat inefficient. The reiser4_scan_extent() code could + delay this update step until it finishes and update the parent_coord only once. + It did that before, but there was a bug and this was the easiest way to make it + correct. */ + if (parent != NULL) { + coord_dup(&scan->parent_coord, parent); + } + + /* Failure may happen at the incr_load_count call, but the caller can assume the reference + is safely taken. */ + return incr_load_count_jnode(&scan->node_load, node); +} + +/* Return true if scanning in the leftward direction. */ +int reiser4_scanning_left(flush_scan * scan) +{ + return scan->direction == LEFT_SIDE; +} + +/* Performs leftward scanning starting from either kind of node. Counts the starting + node. The right-scan object is passed in for the left-scan in order to copy the parent + of an unformatted starting position. This way we avoid searching for the unformatted + node's parent when scanning in each direction. If we search for the parent once it is + set in both scan objects. The limit parameter tells flush-scan when to stop. + + Rapid scanning is used only during scan_left, where we are interested in finding the + 'leftpoint' where we begin flushing. We are interested in stopping at the left child + of a twig that does not have a dirty left neighbor. THIS IS A SPECIAL CASE. The + problem is finding a way to flush only those nodes without unallocated children, and it + is difficult to solve in the bottom-up flushing algorithm we are currently using. The + problem can be solved by scanning left at every level as we go upward, but this would + basically bring us back to using a top-down allocation strategy, which we already tried + (see BK history from May 2002), and has a different set of problems. The top-down + strategy makes avoiding unallocated children easier, but makes it difficult to + propertly flush dirty children with clean parents that would otherwise stop the + top-down flush, only later to dirty the parent once the children are flushed. So we + solve the problem in the bottom-up algorithm with a special case for twigs and leaves + only. + + The first step in solving the problem is this rapid leftward scan. After we determine + that there are at least enough nodes counted to qualify for FLUSH_RELOCATE_THRESHOLD we + are no longer interested in the exact count, we are only interested in finding a the + best place to start the flush. We could choose one of two possibilities: + + 1. Stop at the leftmost child (of a twig) that does not have a dirty left neighbor. + This requires checking one leaf per rapid-scan twig + + 2. Stop at the leftmost child (of a twig) where there are no dirty children of the twig + to the left. This requires checking possibly all of the in-memory children of each + twig during the rapid scan. + + For now we implement the first policy. +*/ +static int +scan_left(flush_scan * scan, flush_scan * right, jnode * node, unsigned limit) +{ + int ret = 0; + + scan->max_count = limit; + scan->direction = LEFT_SIDE; + + ret = scan_set_current(scan, jref(node), 1, NULL); + if (ret != 0) { + return ret; + } + + ret = scan_common(scan, right); + if (ret != 0) { + return ret; + } + + /* Before rapid scanning, we need a lock on scan->node so that we can get its + parent, only if formatted. */ + if (jnode_is_znode(scan->node)) { + ret = longterm_lock_znode(&scan->node_lock, JZNODE(scan->node), + ZNODE_WRITE_LOCK, ZNODE_LOCK_LOPRI); + } + + /* Rapid_scan would go here (with limit set to FLUSH_RELOCATE_THRESHOLD). */ + return ret; +} + +/* Performs rightward scanning... Does not count the starting node. The limit parameter + is described in scan_left. If the starting node is unformatted then the + parent_coord was already set during scan_left. The rapid_after parameter is not used + during right-scanning. + + scan_right is only called if the scan_left operation does not count at least + FLUSH_RELOCATE_THRESHOLD nodes for flushing. Otherwise, the limit parameter is set to + the difference between scan-left's count and FLUSH_RELOCATE_THRESHOLD, meaning + scan-right counts as high as FLUSH_RELOCATE_THRESHOLD and then stops. */ +static int scan_right(flush_scan * scan, jnode * node, unsigned limit) +{ + int ret; + + scan->max_count = limit; + scan->direction = RIGHT_SIDE; + + ret = scan_set_current(scan, jref(node), 0, NULL); + if (ret != 0) { + return ret; + } + + return scan_common(scan, NULL); +} + +/* Common code to perform left or right scanning. */ +static int scan_common(flush_scan * scan, flush_scan * other) +{ + int ret; + + assert("nikita-2376", scan->node != NULL); + assert("edward-54", jnode_is_unformatted(scan->node) + || jnode_is_znode(scan->node)); + + /* Special case for starting at an unformatted node. Optimization: we only want + to search for the parent (which requires a tree traversal) once. Obviously, we + shouldn't have to call it once for the left scan and once for the right scan. + For this reason, if we search for the parent during scan-left we then duplicate + the coord/lock/load into the scan-right object. */ + if (jnode_is_unformatted(scan->node)) { + ret = scan_unformatted(scan, other); + if (ret != 0) + return ret; + } + /* This loop expects to start at a formatted position and performs chaining of + formatted regions */ + while (!reiser4_scan_finished(scan)) { + + ret = scan_formatted(scan); + if (ret != 0) { + return ret; + } + } + + return 0; +} + +static int scan_unformatted(flush_scan * scan, flush_scan * other) +{ + int ret = 0; + int try = 0; + + if (!coord_is_invalid(&scan->parent_coord)) + goto scan; + + /* set parent coord from */ + if (!jnode_is_unformatted(scan->node)) { + /* formatted position */ + + lock_handle lock; + assert("edward-301", jnode_is_znode(scan->node)); + init_lh(&lock); + + /* + * when flush starts from unformatted node, first thing it + * does is tree traversal to find formatted parent of starting + * node. This parent is then kept lock across scans to the + * left and to the right. This means that during scan to the + * left we cannot take left-ward lock, because this is + * dead-lock prone. So, if we are scanning to the left and + * there is already lock held by this thread, + * jnode_lock_parent_coord() should use try-lock. + */ + try = reiser4_scanning_left(scan) + && !lock_stack_isclean(get_current_lock_stack()); + /* Need the node locked to get the parent lock, We have to + take write lock since there is at least one call path + where this znode is already write-locked by us. */ + ret = + longterm_lock_znode(&lock, JZNODE(scan->node), + ZNODE_WRITE_LOCK, + reiser4_scanning_left(scan) ? + ZNODE_LOCK_LOPRI : + ZNODE_LOCK_HIPRI); + if (ret != 0) + /* EINVAL or E_DEADLOCK here mean... try again! At this point we've + scanned too far and can't back out, just start over. */ + return ret; + + ret = jnode_lock_parent_coord(scan->node, + &scan->parent_coord, + &scan->parent_lock, + &scan->parent_load, + ZNODE_WRITE_LOCK, try); + + /* FIXME(C): check EINVAL, E_DEADLOCK */ + done_lh(&lock); + if (ret == -E_REPEAT) { + scan->stop = 1; + return 0; + } + if (ret) + return ret; + + } else { + /* unformatted position */ + + ret = + jnode_lock_parent_coord(scan->node, &scan->parent_coord, + &scan->parent_lock, + &scan->parent_load, + ZNODE_WRITE_LOCK, try); + + if (IS_CBKERR(ret)) + return ret; + + if (ret == CBK_COORD_NOTFOUND) + /* FIXME(C): check EINVAL, E_DEADLOCK */ + return ret; + + /* parent was found */ + assert("jmacd-8661", other != NULL); + /* Duplicate the reference into the other flush_scan. */ + coord_dup(&other->parent_coord, &scan->parent_coord); + copy_lh(&other->parent_lock, &scan->parent_lock); + copy_load_count(&other->parent_load, &scan->parent_load); + } + scan: + return scan_by_coord(scan); +} + +/* Performs left- or rightward scanning starting from a formatted node. Follow left + pointers under tree lock as long as: + + - node->left/right is non-NULL + - node->left/right is connected, dirty + - node->left/right belongs to the same atom + - scan has not reached maximum count +*/ +static int scan_formatted(flush_scan * scan) +{ + int ret; + znode *neighbor = NULL; + + assert("jmacd-1401", !reiser4_scan_finished(scan)); + + do { + znode *node = JZNODE(scan->node); + + /* Node should be connected, but if not stop the scan. */ + if (!znode_is_connected(node)) { + scan->stop = 1; + break; + } + + /* Lock the tree, check-for and reference the next sibling. */ + read_lock_tree(znode_get_tree(node)); + + /* It may be that a node is inserted or removed between a node and its + left sibling while the tree lock is released, but the flush-scan count + does not need to be precise. Thus, we release the tree lock as soon as + we get the neighboring node. */ + neighbor = + reiser4_scanning_left(scan) ? node->left : node->right; + if (neighbor != NULL) { + zref(neighbor); + } + + read_unlock_tree(znode_get_tree(node)); + + /* If neighbor is NULL at the leaf level, need to check for an unformatted + sibling using the parent--break in any case. */ + if (neighbor == NULL) { + break; + } + + /* Check the condition for going left, break if it is not met. This also + releases (jputs) the neighbor if false. */ + if (!reiser4_scan_goto(scan, ZJNODE(neighbor))) { + break; + } + + /* Advance the flush_scan state to the left, repeat. */ + ret = scan_set_current(scan, ZJNODE(neighbor), 1, NULL); + if (ret != 0) { + return ret; + } + + } while (!reiser4_scan_finished(scan)); + + /* If neighbor is NULL then we reached the end of a formatted region, or else the + sibling is out of memory, now check for an extent to the left (as long as + LEAF_LEVEL). */ + if (neighbor != NULL || jnode_get_level(scan->node) != LEAF_LEVEL + || reiser4_scan_finished(scan)) { + scan->stop = 1; + return 0; + } + /* Otherwise, calls scan_by_coord for the right(left)most item of the + left(right) neighbor on the parent level, then possibly continue. */ + + coord_init_invalid(&scan->parent_coord, NULL); + return scan_unformatted(scan, NULL); +} + +/* NOTE-EDWARD: + This scans adjacent items of the same type and calls scan flush plugin for each one. + Performs left(right)ward scanning starting from a (possibly) unformatted node. If we start + from unformatted node, then we continue only if the next neighbor is also unformatted. + When called from scan_formatted, we skip first iteration (to make sure that + right(left)most item of the left(right) neighbor on the parent level is of the same + type and set appropriate coord). */ +static int scan_by_coord(flush_scan * scan) +{ + int ret = 0; + int scan_this_coord; + lock_handle next_lock; + load_count next_load; + coord_t next_coord; + jnode *child; + item_plugin *iplug; + + init_lh(&next_lock); + init_load_count(&next_load); + scan_this_coord = (jnode_is_unformatted(scan->node) ? 1 : 0); + + /* set initial item id */ + iplug = item_plugin_by_coord(&scan->parent_coord); + + for (; !reiser4_scan_finished(scan); scan_this_coord = 1) { + if (scan_this_coord) { + /* Here we expect that unit is scannable. it would not be so due + * to race with extent->tail conversion. */ + if (iplug->f.scan == NULL) { + scan->stop = 1; + ret = -E_REPEAT; + /* skip the check at the end. */ + goto race; + } + + ret = iplug->f.scan(scan); + if (ret != 0) + goto exit; + + if (reiser4_scan_finished(scan)) { + checkchild(scan); + break; + } + } else { + /* the same race against truncate as above is possible + * here, it seems */ + + /* NOTE-JMACD: In this case, apply the same end-of-node logic but don't scan + the first coordinate. */ + assert("jmacd-1231", + item_is_internal(&scan->parent_coord)); + } + + if (iplug->f.utmost_child == NULL + || znode_get_level(scan->parent_coord.node) != TWIG_LEVEL) { + /* stop this coord and continue on parrent level */ + ret = + scan_set_current(scan, + ZJNODE(zref + (scan->parent_coord.node)), + 1, NULL); + if (ret != 0) + goto exit; + break; + } + + /* Either way, the invariant is that scan->parent_coord is set to the + parent of scan->node. Now get the next unit. */ + coord_dup(&next_coord, &scan->parent_coord); + coord_sideof_unit(&next_coord, scan->direction); + + /* If off-the-end of the twig, try the next twig. */ + if (coord_is_after_sideof_unit(&next_coord, scan->direction)) { + /* We take the write lock because we may start flushing from this + * coordinate. */ + ret = neighbor_in_slum(next_coord.node, + &next_lock, + scan->direction, + ZNODE_WRITE_LOCK, + 1 /* check dirty */, + 0 /* don't go though upper + levels */); + if (ret == -E_NO_NEIGHBOR) { + scan->stop = 1; + ret = 0; + break; + } + + if (ret != 0) { + goto exit; + } + + ret = incr_load_count_znode(&next_load, next_lock.node); + if (ret != 0) { + goto exit; + } + + coord_init_sideof_unit(&next_coord, next_lock.node, + sideof_reverse(scan->direction)); + } + + iplug = item_plugin_by_coord(&next_coord); + + /* Get the next child. */ + ret = + iplug->f.utmost_child(&next_coord, + sideof_reverse(scan->direction), + &child); + if (ret != 0) + goto exit; + /* If the next child is not in memory, or, item_utmost_child + failed (due to race with unlink, most probably), stop + here. */ + if (child == NULL || IS_ERR(child)) { + scan->stop = 1; + checkchild(scan); + break; + } + + assert("nikita-2374", jnode_is_unformatted(child) + || jnode_is_znode(child)); + + /* See if it is dirty, part of the same atom. */ + if (!reiser4_scan_goto(scan, child)) { + checkchild(scan); + break; + } + + /* If so, make this child current. */ + ret = scan_set_current(scan, child, 1, &next_coord); + if (ret != 0) + goto exit; + + /* Now continue. If formatted we release the parent lock and return, then + proceed. */ + if (jnode_is_znode(child)) + break; + + /* Otherwise, repeat the above loop with next_coord. */ + if (next_load.node != NULL) { + done_lh(&scan->parent_lock); + move_lh(&scan->parent_lock, &next_lock); + move_load_count(&scan->parent_load, &next_load); + } + } + + assert("jmacd-6233", + reiser4_scan_finished(scan) || jnode_is_znode(scan->node)); + exit: + checkchild(scan); + race: /* skip the above check */ + if (jnode_is_znode(scan->node)) { + done_lh(&scan->parent_lock); + done_load_count(&scan->parent_load); + } + + done_load_count(&next_load); + done_lh(&next_lock); + return ret; +} + +/* FLUSH POS HELPERS */ + +/* Initialize the fields of a flush_position. */ +static void pos_init(flush_pos_t * pos) +{ + memset(pos, 0, sizeof *pos); + + pos->state = POS_INVALID; + coord_init_invalid(&pos->coord, NULL); + init_lh(&pos->lock); + init_load_count(&pos->load); + + reiser4_blocknr_hint_init(&pos->preceder); +} + +/* The flush loop inside squalloc periodically checks pos_valid to + determine when "enough flushing" has been performed. This will return true until one + of the following conditions is met: + + 1. the number of flush-queued nodes has reached the kernel-supplied "int *nr_to_flush" + parameter, meaning we have flushed as many blocks as the kernel requested. When + flushing to commit, this parameter is NULL. + + 2. pos_stop() is called because squalloc discovers that the "next" node in the + flush order is either non-existant, not dirty, or not in the same atom. +*/ + +static int pos_valid(flush_pos_t * pos) +{ + return pos->state != POS_INVALID; +} + +/* Release any resources of a flush_position. Called when jnode_flush finishes. */ +static void pos_done(flush_pos_t * pos) +{ + pos_stop(pos); + reiser4_blocknr_hint_done(&pos->preceder); + if (convert_data(pos)) + free_convert_data(pos); +} + +/* Reset the point and parent. Called during flush subroutines to terminate the + squalloc loop. */ +static int pos_stop(flush_pos_t * pos) +{ + pos->state = POS_INVALID; + done_lh(&pos->lock); + done_load_count(&pos->load); + coord_init_invalid(&pos->coord, NULL); + + if (pos->child) { + jput(pos->child); + pos->child = NULL; + } + + return 0; +} + +/* Return the flush_position's block allocator hint. */ +reiser4_blocknr_hint *reiser4_pos_hint(flush_pos_t * pos) +{ + return &pos->preceder; +} + +flush_queue_t * reiser4_pos_fq(flush_pos_t * pos) +{ + return pos->fq; +} + +/* Make Linus happy. + Local variables: + c-indentation-style: "K&R" + mode-name: "LC" + c-basic-offset: 8 + tab-width: 8 + fill-column: 90 + LocalWords: preceder + End: +*/ diff -urN linux-2.6.22.orig/fs/reiser4/flush.h linux-2.6.22/fs/reiser4/flush.h --- linux-2.6.22.orig/fs/reiser4/flush.h 1970-01-01 03:00:00.000000000 +0300 +++ linux-2.6.22/fs/reiser4/flush.h 2007-07-29 00:25:34.864693371 +0400 @@ -0,0 +1,295 @@ +/* Copyright 2002, 2003 by Hans Reiser, licensing governed by reiser4/README */ + +/* DECLARATIONS: */ + +#if !defined(__REISER4_FLUSH_H__) +#define __REISER4_FLUSH_H__ + +#include "plugin/cluster.h" + +/* The flush_scan data structure maintains the state of an in-progress flush-scan on a + single level of the tree. A flush-scan is used for counting the number of adjacent + nodes to flush, which is used to determine whether we should relocate, and it is also + used to find a starting point for flush. A flush-scan object can scan in both right + and left directions via the scan_left() and scan_right() interfaces. The + right- and left-variations are similar but perform different functions. When scanning + left we (optionally perform rapid scanning and then) longterm-lock the endpoint node. + When scanning right we are simply counting the number of adjacent, dirty nodes. */ +struct flush_scan { + + /* The current number of nodes scanned on this level. */ + unsigned count; + + /* There may be a maximum number of nodes for a scan on any single level. When + going leftward, max_count is determined by FLUSH_SCAN_MAXNODES (see reiser4.h) */ + unsigned max_count; + + /* Direction: Set to one of the sideof enumeration: { LEFT_SIDE, RIGHT_SIDE }. */ + sideof direction; + + /* Initially @stop is set to false then set true once some condition stops the + search (e.g., we found a clean node before reaching max_count or we found a + node belonging to another atom). */ + int stop; + + /* The current scan position. If @node is non-NULL then its reference count has + been incremented to reflect this reference. */ + jnode *node; + + /* A handle for zload/zrelse of current scan position node. */ + load_count node_load; + + /* During left-scan, if the final position (a.k.a. endpoint node) is formatted the + node is locked using this lock handle. The endpoint needs to be locked for + transfer to the flush_position object after scanning finishes. */ + lock_handle node_lock; + + /* When the position is unformatted, its parent, coordinate, and parent + zload/zrelse handle. */ + lock_handle parent_lock; + coord_t parent_coord; + load_count parent_load; + + /* The block allocator preceder hint. Sometimes flush_scan determines what the + preceder is and if so it sets it here, after which it is copied into the + flush_position. Otherwise, the preceder is computed later. */ + reiser4_block_nr preceder_blk; +}; + +struct convert_item_info { + dc_item_stat d_cur; /* disk cluster state of the current item */ + dc_item_stat d_next; /* disk cluster state of the next slum item */ + struct inode *inode; + flow_t flow; +}; + +struct convert_info { + int count; /* for squalloc terminating */ + item_plugin *iplug; /* current item plugin */ + struct convert_item_info *itm; /* current item info */ + struct cluster_handle clust; /* transform cluster */ +}; + +typedef enum flush_position_state { + POS_INVALID, /* Invalid or stopped pos, do not continue slum + * processing */ + POS_ON_LEAF, /* pos points to already prepped, locked formatted node at + * leaf level */ + POS_ON_EPOINT, /* pos keeps a lock on twig level, "coord" field is used + * to traverse unformatted nodes */ + POS_TO_LEAF, /* pos is being moved to leaf level */ + POS_TO_TWIG, /* pos is being moved to twig level */ + POS_END_OF_TWIG, /* special case of POS_ON_TWIG, when coord is after + * rightmost unit of the current twig */ + POS_ON_INTERNAL /* same as POS_ON_LEAF, but points to internal node */ +} flushpos_state_t; + +/* An encapsulation of the current flush point and all the parameters that are passed + through the entire squeeze-and-allocate stage of the flush routine. A single + flush_position object is constructed after left- and right-scanning finishes. */ +struct flush_position { + flushpos_state_t state; + + coord_t coord; /* coord to traverse unformatted nodes */ + lock_handle lock; /* current lock we hold */ + load_count load; /* load status for current locked formatted node */ + + jnode *child; /* for passing a reference to unformatted child + * across pos state changes */ + + reiser4_blocknr_hint preceder; /* The flush 'hint' state. */ + int leaf_relocate; /* True if enough leaf-level nodes were + * found to suggest a relocate policy. */ + int alloc_cnt; /* The number of nodes allocated during squeeze and allococate. */ + int prep_or_free_cnt; /* The number of nodes prepared for write (allocate) or squeezed and freed. */ + flush_queue_t *fq; + long *nr_written; /* number of nodes submitted to disk */ + int flags; /* a copy of jnode_flush flags argument */ + + znode *prev_twig; /* previous parent pointer value, used to catch + * processing of new twig node */ + struct convert_info *sq; /* convert info */ + + unsigned long pos_in_unit; /* for extents only. Position + within an extent unit of first + jnode of slum */ + long nr_to_write; /* number of unformatted nodes to handle on flush */ +}; + +static inline int item_convert_count(flush_pos_t * pos) +{ + return pos->sq->count; +} +static inline void inc_item_convert_count(flush_pos_t * pos) +{ + pos->sq->count++; +} +static inline void set_item_convert_count(flush_pos_t * pos, int count) +{ + pos->sq->count = count; +} +static inline item_plugin *item_convert_plug(flush_pos_t * pos) +{ + return pos->sq->iplug; +} + +static inline struct convert_info *convert_data(flush_pos_t * pos) +{ + return pos->sq; +} + +static inline struct convert_item_info *item_convert_data(flush_pos_t * pos) +{ + assert("edward-955", convert_data(pos)); + return pos->sq->itm; +} + +static inline struct tfm_cluster * tfm_cluster_sq(flush_pos_t * pos) +{ + return &pos->sq->clust.tc; +} + +static inline struct tfm_stream * tfm_stream_sq(flush_pos_t * pos, + tfm_stream_id id) +{ + assert("edward-854", pos->sq != NULL); + return get_tfm_stream(tfm_cluster_sq(pos), id); +} + +static inline int chaining_data_present(flush_pos_t * pos) +{ + return convert_data(pos) && item_convert_data(pos); +} + +/* Returns true if next node contains next item of the disk cluster + so item convert data should be moved to the right slum neighbor. +*/ +static inline int should_chain_next_node(flush_pos_t * pos) +{ + int result = 0; + + assert("edward-1007", chaining_data_present(pos)); + + switch (item_convert_data(pos)->d_next) { + case DC_CHAINED_ITEM: + result = 1; + break; + case DC_AFTER_CLUSTER: + break; + default: + impossible("edward-1009", "bad state of next slum item"); + } + return result; +} + +/* update item state in a disk cluster to assign conversion mode */ +static inline void +move_chaining_data(flush_pos_t * pos, int this_node /* where is next item */ ) +{ + + assert("edward-1010", chaining_data_present(pos)); + + if (this_node == 0) { + /* next item is on the right neighbor */ + assert("edward-1011", + item_convert_data(pos)->d_cur == DC_FIRST_ITEM || + item_convert_data(pos)->d_cur == DC_CHAINED_ITEM); + assert("edward-1012", + item_convert_data(pos)->d_next == DC_CHAINED_ITEM); + + item_convert_data(pos)->d_cur = DC_CHAINED_ITEM; + item_convert_data(pos)->d_next = DC_INVALID_STATE; + } else { + /* next item is on the same node */ + assert("edward-1013", + item_convert_data(pos)->d_cur == DC_FIRST_ITEM || + item_convert_data(pos)->d_cur == DC_CHAINED_ITEM); + assert("edward-1227", + item_convert_data(pos)->d_next == DC_AFTER_CLUSTER || + item_convert_data(pos)->d_next == DC_INVALID_STATE); + + item_convert_data(pos)->d_cur = DC_AFTER_CLUSTER; + item_convert_data(pos)->d_next = DC_INVALID_STATE; + } +} + +static inline int should_convert_node(flush_pos_t * pos, znode * node) +{ + return znode_convertible(node); +} + +/* true if there is attached convert item info */ +static inline int should_convert_next_node(flush_pos_t * pos) +{ + return convert_data(pos) && item_convert_data(pos); +} + +#define SQUALLOC_THRESHOLD 256 + +static inline int should_terminate_squalloc(flush_pos_t * pos) +{ + return convert_data(pos) && + !item_convert_data(pos) && + item_convert_count(pos) >= SQUALLOC_THRESHOLD; +} + +#if 1 +#define check_convert_info(pos) \ +do { \ + if (unlikely(should_convert_next_node(pos))){ \ + warning("edward-1006", "unprocessed chained data"); \ + printk("d_cur = %d, d_next = %d, flow.len = %llu\n", \ + item_convert_data(pos)->d_cur, \ + item_convert_data(pos)->d_next, \ + item_convert_data(pos)->flow.length); \ + printk("inode %llu, size = %llu, cluster %lu\n", \ + (unsigned long long)get_inode_oid \ + (item_convert_data(pos)->inode), \ + i_size_read(item_convert_data(pos)->inode), \ + convert_data(pos)->clust.index); \ + } \ +} while (0) +#else +#define check_convert_info(pos) +#endif /* REISER4_DEBUG */ + +void free_convert_data(flush_pos_t * pos); +/* used in extent.c */ +int scan_set_current(flush_scan * scan, jnode * node, unsigned add_size, + const coord_t * parent); +int reiser4_scan_finished(flush_scan * scan); +int reiser4_scanning_left(flush_scan * scan); +int reiser4_scan_goto(flush_scan * scan, jnode * tonode); +txn_atom *atom_locked_by_fq(flush_queue_t * fq); +int reiser4_alloc_extent(flush_pos_t *flush_pos); +squeeze_result squalloc_extent(znode *left, const coord_t *, flush_pos_t *, + reiser4_key *stop_key); +extern int reiser4_init_fqs(void); +extern void reiser4_done_fqs(void); + +#if REISER4_DEBUG + +extern void reiser4_check_fq(const txn_atom *atom); +extern atomic_t flush_cnt; + +#define check_preceder(blk) \ +assert("nikita-2588", blk < reiser4_block_count(reiser4_get_current_sb())); +extern void check_pos(flush_pos_t * pos); +#else +#define check_preceder(b) noop +#define check_pos(pos) noop +#endif + +/* __REISER4_FLUSH_H__ */ +#endif + +/* Make Linus happy. + Local variables: + c-indentation-style: "K&R" + mode-name: "LC" + c-basic-offset: 8 + tab-width: 8 + fill-column: 90 + LocalWords: preceder + End: +*/ diff -urN linux-2.6.22.orig/fs/reiser4/flush_queue.c linux-2.6.22/fs/reiser4/flush_queue.c --- linux-2.6.22.orig/fs/reiser4/flush_queue.c 1970-01-01 03:00:00.000000000 +0300 +++ linux-2.6.22/fs/reiser4/flush_queue.c 2007-07-29 00:25:34.864693371 +0400 @@ -0,0 +1,680 @@ +/* Copyright 2001, 2002, 2003 by Hans Reiser, licensing governed by reiser4/README */ + +#include "debug.h" +#include "super.h" +#include "txnmgr.h" +#include "jnode.h" +#include "znode.h" +#include "page_cache.h" +#include "wander.h" +#include "vfs_ops.h" +#include "writeout.h" +#include "flush.h" + +#include +#include +#include +#include +#include + +/* A flush queue object is an accumulator for keeping jnodes prepared + by the jnode_flush() function for writing to disk. Those "queued" jnodes are + kept on the flush queue until memory pressure or atom commit asks + flush queues to write some or all from their jnodes. */ + +/* + LOCKING: + + fq->guard spin lock protects fq->atom pointer and nothing else. fq->prepped + list protected by atom spin lock. fq->prepped list uses the following + locking: + + two ways to protect fq->prepped list for read-only list traversal: + + 1. atom spin-lock atom. + 2. fq is IN_USE, atom->nr_running_queues increased. + + and one for list modification: + + 1. atom is spin-locked and one condition is true: fq is IN_USE or + atom->nr_running_queues == 0. + + The deadlock-safe order for flush queues and atoms is: first lock atom, then + lock flush queue, then lock jnode. +*/ + +#define fq_in_use(fq) ((fq)->state & FQ_IN_USE) +#define fq_ready(fq) (!fq_in_use(fq)) + +#define mark_fq_in_use(fq) do { (fq)->state |= FQ_IN_USE; } while (0) +#define mark_fq_ready(fq) do { (fq)->state &= ~FQ_IN_USE; } while (0) + +/* get lock on atom from locked flush queue object */ +static txn_atom *atom_locked_by_fq_nolock(flush_queue_t * fq) +{ + /* This code is similar to jnode_get_atom(), look at it for the + * explanation. */ + txn_atom *atom; + + assert_spin_locked(&(fq->guard)); + + while (1) { + atom = fq->atom; + if (atom == NULL) + break; + + if (spin_trylock_atom(atom)) + break; + + atomic_inc(&atom->refcount); + spin_unlock(&(fq->guard)); + spin_lock_atom(atom); + spin_lock(&(fq->guard)); + + if (fq->atom == atom) { + atomic_dec(&atom->refcount); + break; + } + + spin_unlock(&(fq->guard)); + atom_dec_and_unlock(atom); + spin_lock(&(fq->guard)); + } + + return atom; +} + +txn_atom *atom_locked_by_fq(flush_queue_t * fq) +{ + txn_atom *atom; + + spin_lock(&(fq->guard)); + atom = atom_locked_by_fq_nolock(fq); + spin_unlock(&(fq->guard)); + return atom; +} + +static void init_fq(flush_queue_t * fq) +{ + memset(fq, 0, sizeof *fq); + + atomic_set(&fq->nr_submitted, 0); + + INIT_LIST_HEAD(ATOM_FQ_LIST(fq)); + + init_waitqueue_head(&fq->wait); + spin_lock_init(&fq->guard); +} + +/* slab for flush queues */ +static struct kmem_cache *fq_slab; + +/** + * reiser4_init_fqs - create flush queue cache + * + * Initializes slab cache of flush queues. It is part of reiser4 module + * initialization. + */ +int reiser4_init_fqs(void) +{ + fq_slab = kmem_cache_create("fq", + sizeof(flush_queue_t), + 0, SLAB_HWCACHE_ALIGN, NULL, NULL); + if (fq_slab == NULL) + return RETERR(-ENOMEM); + return 0; +} + +/** + * reiser4_done_fqs - delete flush queue cache + * + * This is called on reiser4 module unloading or system shutdown. + */ +void reiser4_done_fqs(void) +{ + destroy_reiser4_cache(&fq_slab); +} + +/* create new flush queue object */ +static flush_queue_t *create_fq(gfp_t gfp) +{ + flush_queue_t *fq; + + fq = kmem_cache_alloc(fq_slab, gfp); + if (fq) + init_fq(fq); + + return fq; +} + +/* adjust atom's and flush queue's counters of queued nodes */ +static void count_enqueued_node(flush_queue_t * fq) +{ + ON_DEBUG(fq->atom->num_queued++); +} + +static void count_dequeued_node(flush_queue_t * fq) +{ + assert("zam-993", fq->atom->num_queued > 0); + ON_DEBUG(fq->atom->num_queued--); +} + +/* attach flush queue object to the atom */ +static void attach_fq(txn_atom *atom, flush_queue_t *fq) +{ + assert_spin_locked(&(atom->alock)); + list_add(&fq->alink, &atom->flush_queues); + fq->atom = atom; + ON_DEBUG(atom->nr_flush_queues++); +} + +static void detach_fq(flush_queue_t * fq) +{ + assert_spin_locked(&(fq->atom->alock)); + + spin_lock(&(fq->guard)); + list_del_init(&fq->alink); + assert("vs-1456", fq->atom->nr_flush_queues > 0); + ON_DEBUG(fq->atom->nr_flush_queues--); + fq->atom = NULL; + spin_unlock(&(fq->guard)); +} + +/* destroy flush queue object */ +static void done_fq(flush_queue_t * fq) +{ + assert("zam-763", list_empty_careful(ATOM_FQ_LIST(fq))); + assert("zam-766", atomic_read(&fq->nr_submitted) == 0); + + kmem_cache_free(fq_slab, fq); +} + +/* */ +static void mark_jnode_queued(flush_queue_t * fq, jnode * node) +{ + JF_SET(node, JNODE_FLUSH_QUEUED); + count_enqueued_node(fq); +} + +/* Putting jnode into the flush queue. Both atom and jnode should be + spin-locked. */ +void queue_jnode(flush_queue_t * fq, jnode * node) +{ + assert_spin_locked(&(node->guard)); + assert("zam-713", node->atom != NULL); + assert_spin_locked(&(node->atom->alock)); + assert("zam-716", fq->atom != NULL); + assert("zam-717", fq->atom == node->atom); + assert("zam-907", fq_in_use(fq)); + + assert("zam-714", JF_ISSET(node, JNODE_DIRTY)); + assert("zam-826", JF_ISSET(node, JNODE_RELOC)); + assert("vs-1481", !JF_ISSET(node, JNODE_FLUSH_QUEUED)); + assert("vs-1481", NODE_LIST(node) != FQ_LIST); + + mark_jnode_queued(fq, node); + list_move_tail(&node->capture_link, ATOM_FQ_LIST(fq)); + + ON_DEBUG(count_jnode(node->atom, node, NODE_LIST(node), + FQ_LIST, 1)); +} + +/* repeatable process for waiting io completion on a flush queue object */ +static int wait_io(flush_queue_t * fq, int *nr_io_errors) +{ + assert("zam-738", fq->atom != NULL); + assert_spin_locked(&(fq->atom->alock)); + assert("zam-736", fq_in_use(fq)); + assert("zam-911", list_empty_careful(ATOM_FQ_LIST(fq))); + + if (atomic_read(&fq->nr_submitted) != 0) { + struct super_block *super; + + spin_unlock_atom(fq->atom); + + assert("nikita-3013", reiser4_schedulable()); + + super = reiser4_get_current_sb(); + + /* FIXME: this is instead of blk_run_queues() */ + blk_run_address_space(reiser4_get_super_fake(super)->i_mapping); + + if (!(super->s_flags & MS_RDONLY)) + wait_event(fq->wait, atomic_read(&fq->nr_submitted) == 0); + + /* Ask the caller to re-acquire the locks and call this + function again. Note: this technique is commonly used in + the txnmgr code. */ + return -E_REPEAT; + } + + *nr_io_errors += atomic_read(&fq->nr_errors); + return 0; +} + +/* wait on I/O completion, re-submit dirty nodes to write */ +static int finish_fq(flush_queue_t * fq, int *nr_io_errors) +{ + int ret; + txn_atom *atom = fq->atom; + + assert("zam-801", atom != NULL); + assert_spin_locked(&(atom->alock)); + assert("zam-762", fq_in_use(fq)); + + ret = wait_io(fq, nr_io_errors); + if (ret) + return ret; + + detach_fq(fq); + done_fq(fq); + + reiser4_atom_send_event(atom); + + return 0; +} + +/* wait for all i/o for given atom to be completed, actually do one iteration + on that and return -E_REPEAT if there more iterations needed */ +static int finish_all_fq(txn_atom * atom, int *nr_io_errors) +{ + flush_queue_t *fq; + + assert_spin_locked(&(atom->alock)); + + if (list_empty_careful(&atom->flush_queues)) + return 0; + + list_for_each_entry(fq, &atom->flush_queues, alink) { + if (fq_ready(fq)) { + int ret; + + mark_fq_in_use(fq); + assert("vs-1247", fq->owner == NULL); + ON_DEBUG(fq->owner = current); + ret = finish_fq(fq, nr_io_errors); + + if (*nr_io_errors) + reiser4_handle_error(); + + if (ret) { + reiser4_fq_put(fq); + return ret; + } + + spin_unlock_atom(atom); + + return -E_REPEAT; + } + } + + /* All flush queues are in use; atom remains locked */ + return -EBUSY; +} + +/* wait all i/o for current atom */ +int current_atom_finish_all_fq(void) +{ + txn_atom *atom; + int nr_io_errors = 0; + int ret = 0; + + do { + while (1) { + atom = get_current_atom_locked(); + ret = finish_all_fq(atom, &nr_io_errors); + if (ret != -EBUSY) + break; + reiser4_atom_wait_event(atom); + } + } while (ret == -E_REPEAT); + + /* we do not need locked atom after this function finishes, SUCCESS or + -EBUSY are two return codes when atom remains locked after + finish_all_fq */ + if (!ret) + spin_unlock_atom(atom); + + assert_spin_not_locked(&(atom->alock)); + + if (ret) + return ret; + + if (nr_io_errors) + return RETERR(-EIO); + + return 0; +} + +/* change node->atom field for all jnode from given list */ +static void +scan_fq_and_update_atom_ref(struct list_head *list, txn_atom *atom) +{ + jnode *cur; + + list_for_each_entry(cur, list, capture_link) { + spin_lock_jnode(cur); + cur->atom = atom; + spin_unlock_jnode(cur); + } +} + +/* support for atom fusion operation */ +void reiser4_fuse_fq(txn_atom *to, txn_atom *from) +{ + flush_queue_t *fq; + + assert_spin_locked(&(to->alock)); + assert_spin_locked(&(from->alock)); + + list_for_each_entry(fq, &from->flush_queues, alink) { + scan_fq_and_update_atom_ref(ATOM_FQ_LIST(fq), to); + spin_lock(&(fq->guard)); + fq->atom = to; + spin_unlock(&(fq->guard)); + } + + list_splice_init(&from->flush_queues, to->flush_queues.prev); + +#if REISER4_DEBUG + to->num_queued += from->num_queued; + to->nr_flush_queues += from->nr_flush_queues; + from->nr_flush_queues = 0; +#endif +} + +#if REISER4_DEBUG +int atom_fq_parts_are_clean(txn_atom * atom) +{ + assert("zam-915", atom != NULL); + return list_empty_careful(&atom->flush_queues); +} +#endif +/* Bio i/o completion routine for reiser4 write operations. */ +static int +end_io_handler(struct bio *bio, unsigned int bytes_done UNUSED_ARG, + int err) +{ + int i; + int nr_errors = 0; + flush_queue_t *fq; + + assert("zam-958", bio->bi_rw & WRITE); + + /* i/o op. is not fully completed */ + if (bio->bi_size != 0) + return 1; + + if (err == -EOPNOTSUPP) + set_bit(BIO_EOPNOTSUPP, &bio->bi_flags); + + /* we expect that bio->private is set to NULL or fq object which is used + * for synchronization and error counting. */ + fq = bio->bi_private; + /* Check all elements of io_vec for correct write completion. */ + for (i = 0; i < bio->bi_vcnt; i += 1) { + struct page *pg = bio->bi_io_vec[i].bv_page; + + if (!test_bit(BIO_UPTODATE, &bio->bi_flags)) { + SetPageError(pg); + nr_errors++; + } + + { + /* jnode WRITEBACK ("write is in progress bit") is + * atomically cleared here. */ + jnode *node; + + assert("zam-736", pg != NULL); + assert("zam-736", PagePrivate(pg)); + node = jprivate(pg); + + JF_CLR(node, JNODE_WRITEBACK); + } + + end_page_writeback(pg); + page_cache_release(pg); + } + + if (fq) { + /* count i/o error in fq object */ + atomic_add(nr_errors, &fq->nr_errors); + + /* If all write requests registered in this "fq" are done we up + * the waiter. */ + if (atomic_sub_and_test(bio->bi_vcnt, &fq->nr_submitted)) + wake_up(&fq->wait); + } + + bio_put(bio); + return 0; +} + +/* Count I/O requests which will be submitted by @bio in given flush queues + @fq */ +void add_fq_to_bio(flush_queue_t * fq, struct bio *bio) +{ + bio->bi_private = fq; + bio->bi_end_io = end_io_handler; + + if (fq) + atomic_add(bio->bi_vcnt, &fq->nr_submitted); +} + +/* Move all queued nodes out from @fq->prepped list. */ +static void release_prepped_list(flush_queue_t * fq) +{ + txn_atom *atom; + + assert("zam-904", fq_in_use(fq)); + atom = atom_locked_by_fq(fq); + + while (!list_empty(ATOM_FQ_LIST(fq))) { + jnode *cur; + + cur = list_entry(ATOM_FQ_LIST(fq)->next, jnode, capture_link); + list_del_init(&cur->capture_link); + + count_dequeued_node(fq); + spin_lock_jnode(cur); + assert("nikita-3154", !JF_ISSET(cur, JNODE_OVRWR)); + assert("nikita-3154", JF_ISSET(cur, JNODE_RELOC)); + assert("nikita-3154", JF_ISSET(cur, JNODE_FLUSH_QUEUED)); + JF_CLR(cur, JNODE_FLUSH_QUEUED); + + if (JF_ISSET(cur, JNODE_DIRTY)) { + list_add_tail(&cur->capture_link, + ATOM_DIRTY_LIST(atom, jnode_get_level(cur))); + ON_DEBUG(count_jnode(atom, cur, FQ_LIST, + DIRTY_LIST, 1)); + } else { + list_add_tail(&cur->capture_link, ATOM_CLEAN_LIST(atom)); + ON_DEBUG(count_jnode(atom, cur, FQ_LIST, + CLEAN_LIST, 1)); + } + + spin_unlock_jnode(cur); + } + + if (--atom->nr_running_queues == 0) + reiser4_atom_send_event(atom); + + spin_unlock_atom(atom); +} + +/* Submit write requests for nodes on the already filled flush queue @fq. + + @fq: flush queue object which contains jnodes we can (and will) write. + @return: number of submitted blocks (>=0) if success, otherwise -- an error + code (<0). */ +int reiser4_write_fq(flush_queue_t * fq, long *nr_submitted, int flags) +{ + int ret; + txn_atom *atom; + + while (1) { + atom = atom_locked_by_fq(fq); + assert("zam-924", atom); + /* do not write fq in parallel. */ + if (atom->nr_running_queues == 0 + || !(flags & WRITEOUT_SINGLE_STREAM)) + break; + reiser4_atom_wait_event(atom); + } + + atom->nr_running_queues++; + spin_unlock_atom(atom); + + ret = write_jnode_list(ATOM_FQ_LIST(fq), fq, nr_submitted, flags); + release_prepped_list(fq); + + return ret; +} + +/* Getting flush queue object for exclusive use by one thread. May require + several iterations which is indicated by -E_REPEAT return code. + + This function does not contain code for obtaining an atom lock because an + atom lock is obtained by different ways in different parts of reiser4, + usually it is current atom, but we need a possibility for getting fq for the + atom of given jnode. */ +static int fq_by_atom_gfp(txn_atom *atom, flush_queue_t **new_fq, gfp_t gfp) +{ + flush_queue_t *fq; + + assert_spin_locked(&(atom->alock)); + + fq = list_entry(atom->flush_queues.next, flush_queue_t, alink); + while (&atom->flush_queues != &fq->alink) { + spin_lock(&(fq->guard)); + + if (fq_ready(fq)) { + mark_fq_in_use(fq); + assert("vs-1246", fq->owner == NULL); + ON_DEBUG(fq->owner = current); + spin_unlock(&(fq->guard)); + + if (*new_fq) + done_fq(*new_fq); + + *new_fq = fq; + + return 0; + } + + spin_unlock(&(fq->guard)); + + fq = list_entry(fq->alink.next, flush_queue_t, alink); + } + + /* Use previously allocated fq object */ + if (*new_fq) { + mark_fq_in_use(*new_fq); + assert("vs-1248", (*new_fq)->owner == 0); + ON_DEBUG((*new_fq)->owner = current); + attach_fq(atom, *new_fq); + + return 0; + } + + spin_unlock_atom(atom); + + *new_fq = create_fq(gfp); + + if (*new_fq == NULL) + return RETERR(-ENOMEM); + + return RETERR(-E_REPEAT); +} + +int reiser4_fq_by_atom(txn_atom * atom, flush_queue_t ** new_fq) +{ + return fq_by_atom_gfp(atom, new_fq, reiser4_ctx_gfp_mask_get()); +} + +/* A wrapper around reiser4_fq_by_atom for getting a flush queue + object for current atom, if success fq->atom remains locked. */ +flush_queue_t *get_fq_for_current_atom(void) +{ + flush_queue_t *fq = NULL; + txn_atom *atom; + int ret; + + do { + atom = get_current_atom_locked(); + ret = reiser4_fq_by_atom(atom, &fq); + } while (ret == -E_REPEAT); + + if (ret) + return ERR_PTR(ret); + return fq; +} + +/* Releasing flush queue object after exclusive use */ +void reiser4_fq_put_nolock(flush_queue_t *fq) +{ + assert("zam-747", fq->atom != NULL); + assert("zam-902", list_empty_careful(ATOM_FQ_LIST(fq))); + mark_fq_ready(fq); + assert("vs-1245", fq->owner == current); + ON_DEBUG(fq->owner = NULL); +} + +void reiser4_fq_put(flush_queue_t * fq) +{ + txn_atom *atom; + + spin_lock(&(fq->guard)); + atom = atom_locked_by_fq_nolock(fq); + + assert("zam-746", atom != NULL); + + reiser4_fq_put_nolock(fq); + reiser4_atom_send_event(atom); + + spin_unlock(&(fq->guard)); + spin_unlock_atom(atom); +} + +/* A part of atom object initialization related to the embedded flush queue + list head */ + +void init_atom_fq_parts(txn_atom *atom) +{ + INIT_LIST_HEAD(&atom->flush_queues); +} + +#if REISER4_DEBUG + +void reiser4_check_fq(const txn_atom *atom) +{ + /* check number of nodes on all atom's flush queues */ + flush_queue_t *fq; + int count; + struct list_head *pos; + + count = 0; + list_for_each_entry(fq, &atom->flush_queues, alink) { + spin_lock(&(fq->guard)); + /* calculate number of jnodes on fq' list of prepped jnodes */ + list_for_each(pos, ATOM_FQ_LIST(fq)) + count++; + spin_unlock(&(fq->guard)); + } + if (count != atom->fq) + warning("", "fq counter %d, real %d\n", atom->fq, count); + +} + +#endif + +/* + * Local variables: + * c-indentation-style: "K&R" + * mode-name: "LC" + * c-basic-offset: 8 + * tab-width: 8 + * fill-column: 79 + * scroll-step: 1 + * End: + */ diff -urN linux-2.6.22.orig/fs/reiser4/forward.h linux-2.6.22/fs/reiser4/forward.h --- linux-2.6.22.orig/fs/reiser4/forward.h 1970-01-01 03:00:00.000000000 +0300 +++ linux-2.6.22/fs/reiser4/forward.h 2007-07-29 00:25:34.864693371 +0400 @@ -0,0 +1,252 @@ +/* Copyright 2001, 2002, 2003 by Hans Reiser, licensing governed by reiser4/README */ + +/* Forward declarations. Thank you Kernighan. */ + +#if !defined( __REISER4_FORWARD_H__ ) +#define __REISER4_FORWARD_H__ + +#include +#include + +typedef struct zlock zlock; +typedef struct lock_stack lock_stack; +typedef struct lock_handle lock_handle; +typedef struct znode znode; +typedef struct flow flow_t; +typedef struct coord coord_t; +typedef struct tree_access_pointer tap_t; +typedef struct reiser4_object_create_data reiser4_object_create_data; +typedef union reiser4_plugin reiser4_plugin; +typedef __u16 reiser4_plugin_id; +typedef __u64 reiser4_plugin_groups; +typedef struct item_plugin item_plugin; +typedef struct jnode_plugin jnode_plugin; +typedef struct reiser4_item_data reiser4_item_data; +typedef union reiser4_key reiser4_key; +typedef struct reiser4_tree reiser4_tree; +typedef struct carry_cut_data carry_cut_data; +typedef struct carry_kill_data carry_kill_data; +typedef struct carry_tree_op carry_tree_op; +typedef struct carry_tree_node carry_tree_node; +typedef struct carry_plugin_info carry_plugin_info; +typedef struct reiser4_journal reiser4_journal; +typedef struct txn_atom txn_atom; +typedef struct txn_handle txn_handle; +typedef struct txn_mgr txn_mgr; +typedef struct reiser4_dir_entry_desc reiser4_dir_entry_desc; +typedef struct reiser4_context reiser4_context; +typedef struct carry_level carry_level; +typedef struct blocknr_set_entry blocknr_set_entry; +/* super_block->s_fs_info points to this */ +typedef struct reiser4_super_info_data reiser4_super_info_data; +/* next two objects are fields of reiser4_super_info_data */ +typedef struct reiser4_oid_allocator reiser4_oid_allocator; +typedef struct reiser4_space_allocator reiser4_space_allocator; + +typedef struct flush_scan flush_scan; +typedef struct flush_position flush_pos_t; + +typedef unsigned short pos_in_node_t; +#define MAX_POS_IN_NODE 65535 + +typedef struct jnode jnode; +typedef struct reiser4_blocknr_hint reiser4_blocknr_hint; + +typedef struct uf_coord uf_coord_t; +typedef struct hint hint_t; + +typedef struct ktxnmgrd_context ktxnmgrd_context; + +struct inode; +struct page; +struct file; +struct dentry; +struct super_block; + +/* return values of coord_by_key(). cbk == coord_by_key */ +typedef enum { + CBK_COORD_FOUND = 0, + CBK_COORD_NOTFOUND = -ENOENT, +} lookup_result; + +/* results of lookup with directory file */ +typedef enum { + FILE_NAME_FOUND = 0, + FILE_NAME_NOTFOUND = -ENOENT, + FILE_IO_ERROR = -EIO, /* FIXME: it seems silly to have special OOM, IO_ERROR return codes for each search. */ + FILE_OOM = -ENOMEM /* FIXME: it seems silly to have special OOM, IO_ERROR return codes for each search. */ +} file_lookup_result; + +/* behaviors of lookup. If coord we are looking for is actually in a tree, + both coincide. */ +typedef enum { + /* search exactly for the coord with key given */ + FIND_EXACT, + /* search for coord with the maximal key not greater than one + given */ + FIND_MAX_NOT_MORE_THAN /*LEFT_SLANT_BIAS */ +} lookup_bias; + +typedef enum { + /* number of leaf level of the tree + The fake root has (tree_level=0). */ + LEAF_LEVEL = 1, + + /* number of level one above leaf level of the tree. + + It is supposed that internal tree used by reiser4 to store file + system data and meta data will have height 2 initially (when + created by mkfs). + */ + TWIG_LEVEL = 2, +} tree_level; + +/* The "real" maximum ztree height is the 0-origin size of any per-level + array, since the zero'th level is not used. */ +#define REAL_MAX_ZTREE_HEIGHT (REISER4_MAX_ZTREE_HEIGHT-LEAF_LEVEL) + +/* enumeration of possible mutual position of item and coord. This enum is + return type of ->is_in_item() item plugin method which see. */ +typedef enum { + /* coord is on the left of an item */ + IP_ON_THE_LEFT, + /* coord is inside item */ + IP_INSIDE, + /* coord is inside item, but to the right of the rightmost unit of + this item */ + IP_RIGHT_EDGE, + /* coord is on the right of an item */ + IP_ON_THE_RIGHT +} interposition; + +/* type of lock to acquire on znode before returning it to caller */ +typedef enum { + ZNODE_NO_LOCK = 0, + ZNODE_READ_LOCK = 1, + ZNODE_WRITE_LOCK = 2, +} znode_lock_mode; + +/* type of lock request */ +typedef enum { + ZNODE_LOCK_LOPRI = 0, + ZNODE_LOCK_HIPRI = (1 << 0), + + /* By setting the ZNODE_LOCK_NONBLOCK flag in a lock request the call to longterm_lock_znode will not sleep + waiting for the lock to become available. If the lock is unavailable, reiser4_znode_lock will immediately + return the value -E_REPEAT. */ + ZNODE_LOCK_NONBLOCK = (1 << 1), + /* An option for longterm_lock_znode which prevents atom fusion */ + ZNODE_LOCK_DONT_FUSE = (1 << 2) +} znode_lock_request; + +typedef enum { READ_OP = 0, WRITE_OP = 1 } rw_op; + +/* used to specify direction of shift. These must be -1 and 1 */ +typedef enum { + SHIFT_LEFT = 1, + SHIFT_RIGHT = -1 +} shift_direction; + +typedef enum { + LEFT_SIDE, + RIGHT_SIDE +} sideof; + +#define round_up( value, order ) \ + ( ( typeof( value ) )( ( ( long ) ( value ) + ( order ) - 1U ) & \ + ~( ( order ) - 1 ) ) ) + +/* values returned by squalloc_right_neighbor and its auxiliary functions */ +typedef enum { + /* unit of internal item is moved */ + SUBTREE_MOVED = 0, + /* nothing else can be squeezed into left neighbor */ + SQUEEZE_TARGET_FULL = 1, + /* all content of node is squeezed into its left neighbor */ + SQUEEZE_SOURCE_EMPTY = 2, + /* one more item is copied (this is only returned by + allocate_and_copy_extent to squalloc_twig)) */ + SQUEEZE_CONTINUE = 3 +} squeeze_result; + +/* Do not change items ids. If you do - there will be format change */ +typedef enum { + STATIC_STAT_DATA_ID = 0x0, + SIMPLE_DIR_ENTRY_ID = 0x1, + COMPOUND_DIR_ID = 0x2, + NODE_POINTER_ID = 0x3, + EXTENT_POINTER_ID = 0x5, + FORMATTING_ID = 0x6, + CTAIL_ID = 0x7, + BLACK_BOX_ID = 0x8, + LAST_ITEM_ID = 0x9 +} item_id; + +/* Flags passed to jnode_flush() to allow it to distinguish default settings based on + whether commit() was called or VM memory pressure was applied. */ +typedef enum { + /* submit flush queue to disk at jnode_flush completion */ + JNODE_FLUSH_WRITE_BLOCKS = 1, + + /* flush is called for commit */ + JNODE_FLUSH_COMMIT = 2, + /* not implemented */ + JNODE_FLUSH_MEMORY_FORMATTED = 4, + + /* not implemented */ + JNODE_FLUSH_MEMORY_UNFORMATTED = 8, +} jnode_flush_flags; + +/* Flags to insert/paste carry operations. Currently they only used in + flushing code, but in future, they can be used to optimize for repetitive + accesses. */ +typedef enum { + /* carry is not allowed to shift data to the left when trying to find + free space */ + COPI_DONT_SHIFT_LEFT = (1 << 0), + /* carry is not allowed to shift data to the right when trying to find + free space */ + COPI_DONT_SHIFT_RIGHT = (1 << 1), + /* carry is not allowed to allocate new node(s) when trying to find + free space */ + COPI_DONT_ALLOCATE = (1 << 2), + /* try to load left neighbor if its not in a cache */ + COPI_LOAD_LEFT = (1 << 3), + /* try to load right neighbor if its not in a cache */ + COPI_LOAD_RIGHT = (1 << 4), + /* shift insertion point to the left neighbor */ + COPI_GO_LEFT = (1 << 5), + /* shift insertion point to the right neighbor */ + COPI_GO_RIGHT = (1 << 6), + /* try to step back into original node if insertion into new node + fails after shifting data there. */ + COPI_STEP_BACK = (1 << 7) +} cop_insert_flag; + +typedef enum { + SAFE_UNLINK, /* safe-link for unlink */ + SAFE_TRUNCATE /* safe-link for truncate */ +} reiser4_safe_link_t; + +/* this is to show on which list of atom jnode is */ +typedef enum { + NOT_CAPTURED, + DIRTY_LIST, + CLEAN_LIST, + FQ_LIST, + WB_LIST, + OVRWR_LIST +} atom_list; + +/* __REISER4_FORWARD_H__ */ +#endif + +/* Make Linus happy. + Local variables: + c-indentation-style: "K&R" + mode-name: "LC" + c-basic-offset: 8 + tab-width: 8 + fill-column: 120 + End: +*/ diff -urN linux-2.6.22.orig/fs/reiser4/fsdata.c linux-2.6.22/fs/reiser4/fsdata.c --- linux-2.6.22.orig/fs/reiser4/fsdata.c 1970-01-01 03:00:00.000000000 +0300 +++ linux-2.6.22/fs/reiser4/fsdata.c 2007-07-29 00:25:34.868694406 +0400 @@ -0,0 +1,808 @@ +/* Copyright 2001, 2002, 2003, 2004, 2005 by Hans Reiser, licensing governed by + * reiser4/README */ + +#include "fsdata.h" +#include "inode.h" + + +/* cache or dir_cursors */ +static struct kmem_cache *d_cursor_cache; +static struct shrinker *d_cursor_shrinker; + +/* list of unused cursors */ +static LIST_HEAD(cursor_cache); + +/* number of cursors in list of ununsed cursors */ +static unsigned long d_cursor_unused = 0; + +/* spinlock protecting manipulations with dir_cursor's hash table and lists */ +DEFINE_SPINLOCK(d_lock); + +static reiser4_file_fsdata *create_fsdata(struct file *file); +static int file_is_stateless(struct file *file); +static void free_fsdata(reiser4_file_fsdata *fsdata); +static void kill_cursor(dir_cursor *); + +/** + * d_cursor_shrink - shrink callback for cache of dir_cursor-s + * @nr: number of objects to free + * @mask: GFP mask + * + * Shrinks d_cursor_cache. Scan LRU list of unused cursors, freeing requested + * number. Return number of still freeable cursors. + */ +static int d_cursor_shrink(int nr, gfp_t mask) +{ + if (nr != 0) { + dir_cursor *scan; + int killed; + + killed = 0; + spin_lock(&d_lock); + while (!list_empty(&cursor_cache)) { + scan = list_entry(cursor_cache.next, dir_cursor, alist); + assert("nikita-3567", scan->ref == 0); + kill_cursor(scan); + ++killed; + --nr; + if (nr == 0) + break; + } + spin_unlock(&d_lock); + } + return d_cursor_unused; +} + +/** + * reiser4_init_d_cursor - create d_cursor cache + * + * Initializes slab cache of d_cursors. It is part of reiser4 module + * initialization. + */ +int reiser4_init_d_cursor(void) +{ + d_cursor_cache = kmem_cache_create("d_cursor", sizeof(dir_cursor), 0, + SLAB_HWCACHE_ALIGN, NULL, NULL); + if (d_cursor_cache == NULL) + return RETERR(-ENOMEM); + + /* + * actually, d_cursors are "priceless", because there is no way to + * recover information stored in them. On the other hand, we don't + * want to consume all kernel memory by them. As a compromise, just + * assign higher "seeks" value to d_cursor cache, so that it will be + * shrunk only if system is really tight on memory. + */ + d_cursor_shrinker = set_shrinker(DEFAULT_SEEKS << 3, + d_cursor_shrink); + if (d_cursor_shrinker == NULL) { + destroy_reiser4_cache(&d_cursor_cache); + d_cursor_cache = NULL; + return RETERR(-ENOMEM); + } + return 0; +} + +/** + * reiser4_done_d_cursor - delete d_cursor cache and d_cursor shrinker + * + * This is called on reiser4 module unloading or system shutdown. + */ +void reiser4_done_d_cursor(void) +{ + BUG_ON(d_cursor_shrinker == NULL); + remove_shrinker(d_cursor_shrinker); + d_cursor_shrinker = NULL; + + destroy_reiser4_cache(&d_cursor_cache); +} + +#define D_CURSOR_TABLE_SIZE (256) + +static inline unsigned long +d_cursor_hash(d_cursor_hash_table *table, const struct d_cursor_key *key) +{ + assert("nikita-3555", IS_POW(D_CURSOR_TABLE_SIZE)); + return (key->oid + key->cid) & (D_CURSOR_TABLE_SIZE - 1); +} + +static inline int d_cursor_eq(const struct d_cursor_key *k1, + const struct d_cursor_key *k2) +{ + return k1->cid == k2->cid && k1->oid == k2->oid; +} + +/* + * define functions to manipulate reiser4 super block's hash table of + * dir_cursors + */ +#define KMALLOC(size) kmalloc((size), reiser4_ctx_gfp_mask_get()) +#define KFREE(ptr, size) kfree(ptr) +TYPE_SAFE_HASH_DEFINE(d_cursor, + dir_cursor, + struct d_cursor_key, + key, hash, d_cursor_hash, d_cursor_eq); +#undef KFREE +#undef KMALLOC + +/** + * reiser4_init_super_d_info - initialize per-super-block d_cursor resources + * @super: super block to initialize + * + * Initializes per-super-block d_cursor's hash table and radix tree. It is part + * of mount. + */ +int reiser4_init_super_d_info(struct super_block *super) +{ + struct d_cursor_info *p; + + p = &get_super_private(super)->d_info; + + INIT_RADIX_TREE(&p->tree, reiser4_ctx_gfp_mask_get()); + return d_cursor_hash_init(&p->table, D_CURSOR_TABLE_SIZE); +} + +/** + * reiser4_done_super_d_info - release per-super-block d_cursor resources + * @super: super block being umounted + * + * It is called on umount. Kills all directory cursors attached to suoer block. + */ +void reiser4_done_super_d_info(struct super_block *super) +{ + struct d_cursor_info *d_info; + dir_cursor *cursor, *next; + + d_info = &get_super_private(super)->d_info; + for_all_in_htable(&d_info->table, d_cursor, cursor, next) + kill_cursor(cursor); + + BUG_ON(d_info->tree.rnode != NULL); + d_cursor_hash_done(&d_info->table); +} + +/** + * kill_cursor - free dir_cursor and reiser4_file_fsdata attached to it + * @cursor: cursor to free + * + * Removes reiser4_file_fsdata attached to @cursor from readdir list of + * reiser4_inode, frees that reiser4_file_fsdata. Removes @cursor from from + * indices, hash table, list of unused cursors and frees it. + */ +static void kill_cursor(dir_cursor *cursor) +{ + unsigned long index; + + assert("nikita-3566", cursor->ref == 0); + assert("nikita-3572", cursor->fsdata != NULL); + + index = (unsigned long)cursor->key.oid; + list_del_init(&cursor->fsdata->dir.linkage); + free_fsdata(cursor->fsdata); + cursor->fsdata = NULL; + + if (list_empty_careful(&cursor->list)) + /* this is last cursor for a file. Kill radix-tree entry */ + radix_tree_delete(&cursor->info->tree, index); + else { + void **slot; + + /* + * there are other cursors for the same oid. + */ + + /* + * if radix tree point to the cursor being removed, re-target + * radix tree slot to the next cursor in the (non-empty as was + * checked above) element of the circular list of all cursors + * for this oid. + */ + slot = radix_tree_lookup_slot(&cursor->info->tree, index); + assert("nikita-3571", *slot != NULL); + if (*slot == cursor) + *slot = list_entry(cursor->list.next, dir_cursor, list); + /* remove cursor from circular list */ + list_del_init(&cursor->list); + } + /* remove cursor from the list of unused cursors */ + list_del_init(&cursor->alist); + /* remove cursor from the hash table */ + d_cursor_hash_remove(&cursor->info->table, cursor); + /* and free it */ + kmem_cache_free(d_cursor_cache, cursor); + --d_cursor_unused; +} + +/* possible actions that can be performed on all cursors for the given file */ +enum cursor_action { + /* + * load all detached state: this is called when stat-data is loaded + * from the disk to recover information about all pending readdirs + */ + CURSOR_LOAD, + /* + * detach all state from inode, leaving it in the cache. This is called + * when inode is removed form the memory by memory pressure + */ + CURSOR_DISPOSE, + /* + * detach cursors from the inode, and free them. This is called when + * inode is destroyed + */ + CURSOR_KILL +}; + +/* + * return d_cursor data for the file system @inode is in. + */ +static inline struct d_cursor_info *d_info(struct inode *inode) +{ + return &get_super_private(inode->i_sb)->d_info; +} + +/* + * lookup d_cursor in the per-super-block radix tree. + */ +static inline dir_cursor *lookup(struct d_cursor_info * info, + unsigned long index) +{ + return (dir_cursor *) radix_tree_lookup(&info->tree, index); +} + +/* + * attach @cursor to the radix tree. There may be multiple cursors for the + * same oid, they are chained into circular list. + */ +static void bind_cursor(dir_cursor * cursor, unsigned long index) +{ + dir_cursor *head; + + head = lookup(cursor->info, index); + if (head == NULL) { + /* this is the first cursor for this index */ + INIT_LIST_HEAD(&cursor->list); + radix_tree_insert(&cursor->info->tree, index, cursor); + } else { + /* some cursor already exists. Chain ours */ + list_add(&cursor->list, &head->list); + } +} + +/* + * detach fsdata (if detachable) from file descriptor, and put cursor on the + * "unused" list. Called when file descriptor is not longer in active use. + */ +static void clean_fsdata(struct file *file) +{ + dir_cursor *cursor; + reiser4_file_fsdata *fsdata; + + assert("nikita-3570", file_is_stateless(file)); + + fsdata = (reiser4_file_fsdata *) file->private_data; + if (fsdata != NULL) { + cursor = fsdata->cursor; + if (cursor != NULL) { + spin_lock(&d_lock); + --cursor->ref; + if (cursor->ref == 0) { + list_add_tail(&cursor->alist, &cursor_cache); + ++d_cursor_unused; + } + spin_unlock(&d_lock); + file->private_data = NULL; + } + } +} + +/* + * global counter used to generate "client ids". These ids are encoded into + * high bits of fpos. + */ +static __u32 cid_counter = 0; +#define CID_SHIFT (20) +#define CID_MASK (0xfffffull) + +static void free_file_fsdata_nolock(struct file *); + +/** + * insert_cursor - allocate file_fsdata, insert cursor to tree and hash table + * @cursor: + * @file: + * @inode: + * + * Allocates reiser4_file_fsdata, attaches it to @cursor, inserts cursor to + * reiser4 super block's hash table and radix tree. + add detachable readdir + * state to the @f + */ +static int insert_cursor(dir_cursor *cursor, struct file *file, + struct inode *inode) +{ + int result; + reiser4_file_fsdata *fsdata; + + memset(cursor, 0, sizeof *cursor); + + /* this is either first call to readdir, or rewind. Anyway, create new + * cursor. */ + fsdata = create_fsdata(NULL); + if (fsdata != NULL) { + result = radix_tree_preload(reiser4_ctx_gfp_mask_get()); + if (result == 0) { + struct d_cursor_info *info; + oid_t oid; + + info = d_info(inode); + oid = get_inode_oid(inode); + /* cid occupies higher 12 bits of f->f_pos. Don't + * allow it to become negative: this confuses + * nfsd_readdir() */ + cursor->key.cid = (++cid_counter) & 0x7ff; + cursor->key.oid = oid; + cursor->fsdata = fsdata; + cursor->info = info; + cursor->ref = 1; + + spin_lock_inode(inode); + /* install cursor as @f's private_data, discarding old + * one if necessary */ +#if REISER4_DEBUG + if (file->private_data) + warning("", "file has fsdata already"); +#endif + clean_fsdata(file); + free_file_fsdata_nolock(file); + file->private_data = fsdata; + fsdata->cursor = cursor; + spin_unlock_inode(inode); + spin_lock(&d_lock); + /* insert cursor into hash table */ + d_cursor_hash_insert(&info->table, cursor); + /* and chain it into radix-tree */ + bind_cursor(cursor, (unsigned long)oid); + spin_unlock(&d_lock); + radix_tree_preload_end(); + file->f_pos = ((__u64) cursor->key.cid) << CID_SHIFT; + } + } else + result = RETERR(-ENOMEM); + return result; +} + +/** + * process_cursors - do action on each cursor attached to inode + * @inode: + * @act: action to do + * + * Finds all cursors of @inode in reiser4's super block radix tree of cursors + * and performs action specified by @act on each of cursors. + */ +static void process_cursors(struct inode *inode, enum cursor_action act) +{ + oid_t oid; + dir_cursor *start; + struct list_head *head; + reiser4_context *ctx; + struct d_cursor_info *info; + + /* this can be called by + * + * kswapd->...->prune_icache->..reiser4_destroy_inode + * + * without reiser4_context + */ + ctx = reiser4_init_context(inode->i_sb); + if (IS_ERR(ctx)) { + warning("vs-23", "failed to init context"); + return; + } + + assert("nikita-3558", inode != NULL); + + info = d_info(inode); + oid = get_inode_oid(inode); + spin_lock_inode(inode); + head = get_readdir_list(inode); + spin_lock(&d_lock); + /* find any cursor for this oid: reference to it is hanging of radix + * tree */ + start = lookup(info, (unsigned long)oid); + if (start != NULL) { + dir_cursor *scan; + reiser4_file_fsdata *fsdata; + + /* process circular list of cursors for this oid */ + scan = start; + do { + dir_cursor *next; + + next = list_entry(scan->list.next, dir_cursor, list); + fsdata = scan->fsdata; + assert("nikita-3557", fsdata != NULL); + if (scan->key.oid == oid) { + switch (act) { + case CURSOR_DISPOSE: + list_del_init(&fsdata->dir.linkage); + break; + case CURSOR_LOAD: + list_add(&fsdata->dir.linkage, head); + break; + case CURSOR_KILL: + kill_cursor(scan); + break; + } + } + if (scan == next) + /* last cursor was just killed */ + break; + scan = next; + } while (scan != start); + } + spin_unlock(&d_lock); + /* check that we killed 'em all */ + assert("nikita-3568", + ergo(act == CURSOR_KILL, + list_empty_careful(get_readdir_list(inode)))); + assert("nikita-3569", + ergo(act == CURSOR_KILL, lookup(info, oid) == NULL)); + spin_unlock_inode(inode); + reiser4_exit_context(ctx); +} + +/** + * reiser4_dispose_cursors - removes cursors from inode's list + * @inode: inode to dispose cursors of + * + * For each of cursors corresponding to @inode - removes reiser4_file_fsdata + * attached to cursor from inode's readdir list. This is called when inode is + * removed from the memory by memory pressure. + */ +void reiser4_dispose_cursors(struct inode *inode) +{ + process_cursors(inode, CURSOR_DISPOSE); +} + +/** + * reiser4_load_cursors - attach cursors to inode + * @inode: inode to load cursors to + * + * For each of cursors corresponding to @inode - attaches reiser4_file_fsdata + * attached to cursor to inode's readdir list. This is done when inode is + * loaded into memory. + */ +void reiser4_load_cursors(struct inode *inode) +{ + process_cursors(inode, CURSOR_LOAD); +} + +/** + * reiser4_kill_cursors - kill all inode cursors + * @inode: inode to kill cursors of + * + * Frees all cursors for this inode. This is called when inode is destroyed. + */ +void reiser4_kill_cursors(struct inode *inode) +{ + process_cursors(inode, CURSOR_KILL); +} + +/** + * file_is_stateless - + * @file: + * + * true, if file descriptor @f is created by NFS server by "demand" to serve + * one file system operation. This means that there may be "detached state" + * for underlying inode. + */ +static int file_is_stateless(struct file *file) +{ + return reiser4_get_dentry_fsdata(file->f_dentry)->stateless; +} + +/** + * reiser4_get_dir_fpos - + * @dir: + * + * Calculates ->fpos from user-supplied cookie. Normally it is dir->f_pos, but + * in the case of stateless directory operation (readdir-over-nfs), client id + * was encoded in the high bits of cookie and should me masked off. + */ +loff_t reiser4_get_dir_fpos(struct file *dir) +{ + if (file_is_stateless(dir)) + return dir->f_pos & CID_MASK; + else + return dir->f_pos; +} + +/** + * reiser4_attach_fsdata - try to attach fsdata + * @file: + * @inode: + * + * Finds or creates cursor for readdir-over-nfs. + */ +int reiser4_attach_fsdata(struct file *file, struct inode *inode) +{ + loff_t pos; + int result; + dir_cursor *cursor; + + /* + * we are serialized by inode->i_mutex + */ + if (!file_is_stateless(file)) + return 0; + + pos = file->f_pos; + result = 0; + if (pos == 0) { + /* + * first call to readdir (or rewind to the beginning of + * directory) + */ + cursor = kmem_cache_alloc(d_cursor_cache, + reiser4_ctx_gfp_mask_get()); + if (cursor != NULL) + result = insert_cursor(cursor, file, inode); + else + result = RETERR(-ENOMEM); + } else { + /* try to find existing cursor */ + struct d_cursor_key key; + + key.cid = pos >> CID_SHIFT; + key.oid = get_inode_oid(inode); + spin_lock(&d_lock); + cursor = d_cursor_hash_find(&d_info(inode)->table, &key); + if (cursor != NULL) { + /* cursor was found */ + if (cursor->ref == 0) { + /* move it from unused list */ + list_del_init(&cursor->alist); + --d_cursor_unused; + } + ++cursor->ref; + } + spin_unlock(&d_lock); + if (cursor != NULL) { + spin_lock_inode(inode); + assert("nikita-3556", cursor->fsdata->back == NULL); + clean_fsdata(file); + free_file_fsdata_nolock(file); + file->private_data = cursor->fsdata; + spin_unlock_inode(inode); + } + } + return result; +} + +/** + * reiser4_detach_fsdata - ??? + * @file: + * + * detach fsdata, if necessary + */ +void reiser4_detach_fsdata(struct file *file) +{ + struct inode *inode; + + if (!file_is_stateless(file)) + return; + + inode = file->f_dentry->d_inode; + spin_lock_inode(inode); + clean_fsdata(file); + spin_unlock_inode(inode); +} + +/* slab for reiser4_dentry_fsdata */ +static struct kmem_cache *dentry_fsdata_cache; + +/** + * reiser4_init_dentry_fsdata - create cache of dentry_fsdata + * + * Initializes slab cache of structures attached to denty->d_fsdata. It is + * part of reiser4 module initialization. + */ +int reiser4_init_dentry_fsdata(void) +{ + dentry_fsdata_cache = kmem_cache_create("dentry_fsdata", + sizeof(struct reiser4_dentry_fsdata), + 0, + SLAB_HWCACHE_ALIGN | + SLAB_RECLAIM_ACCOUNT, NULL, + NULL); + if (dentry_fsdata_cache == NULL) + return RETERR(-ENOMEM); + return 0; +} + +/** + * reiser4_done_dentry_fsdata - delete cache of dentry_fsdata + * + * This is called on reiser4 module unloading or system shutdown. + */ +void reiser4_done_dentry_fsdata(void) +{ + destroy_reiser4_cache(&dentry_fsdata_cache); +} + +/** + * reiser4_get_dentry_fsdata - get fs-specific dentry data + * @dentry: queried dentry + * + * Allocates if necessary and returns per-dentry data that we attach to each + * dentry. + */ +struct reiser4_dentry_fsdata *reiser4_get_dentry_fsdata(struct dentry *dentry) +{ + assert("nikita-1365", dentry != NULL); + + if (dentry->d_fsdata == NULL) { + dentry->d_fsdata = kmem_cache_alloc(dentry_fsdata_cache, + reiser4_ctx_gfp_mask_get()); + if (dentry->d_fsdata == NULL) + return ERR_PTR(RETERR(-ENOMEM)); + memset(dentry->d_fsdata, 0, + sizeof(struct reiser4_dentry_fsdata)); + } + return dentry->d_fsdata; +} + +/** + * reiser4_free_dentry_fsdata - detach and free dentry_fsdata + * @dentry: dentry to free fsdata of + * + * Detaches and frees fs-specific dentry data + */ +void reiser4_free_dentry_fsdata(struct dentry *dentry) +{ + if (dentry->d_fsdata != NULL) { + kmem_cache_free(dentry_fsdata_cache, dentry->d_fsdata); + dentry->d_fsdata = NULL; + } +} + +/* slab for reiser4_file_fsdata */ +static struct kmem_cache *file_fsdata_cache; + +/** + * reiser4_init_file_fsdata - create cache of reiser4_file_fsdata + * + * Initializes slab cache of structures attached to file->private_data. It is + * part of reiser4 module initialization. + */ +int reiser4_init_file_fsdata(void) +{ + file_fsdata_cache = kmem_cache_create("file_fsdata", + sizeof(reiser4_file_fsdata), + 0, + SLAB_HWCACHE_ALIGN | + SLAB_RECLAIM_ACCOUNT, NULL, NULL); + if (file_fsdata_cache == NULL) + return RETERR(-ENOMEM); + return 0; +} + +/** + * reiser4_done_file_fsdata - delete cache of reiser4_file_fsdata + * + * This is called on reiser4 module unloading or system shutdown. + */ +void reiser4_done_file_fsdata(void) +{ + destroy_reiser4_cache(&file_fsdata_cache); +} + +/** + * create_fsdata - allocate and initialize reiser4_file_fsdata + * @file: what to create file_fsdata for, may be NULL + * + * Allocates and initializes reiser4_file_fsdata structure. + */ +static reiser4_file_fsdata *create_fsdata(struct file *file) +{ + reiser4_file_fsdata *fsdata; + + fsdata = kmem_cache_alloc(file_fsdata_cache, + reiser4_ctx_gfp_mask_get()); + if (fsdata != NULL) { + memset(fsdata, 0, sizeof *fsdata); + fsdata->ra1.max_window_size = VM_MAX_READAHEAD * 1024; + fsdata->back = file; + INIT_LIST_HEAD(&fsdata->dir.linkage); + } + return fsdata; +} + +/** + * free_fsdata - free reiser4_file_fsdata + * @fsdata: object to free + * + * Dual to create_fsdata(). Free reiser4_file_fsdata. + */ +static void free_fsdata(reiser4_file_fsdata *fsdata) +{ + BUG_ON(fsdata == NULL); + kmem_cache_free(file_fsdata_cache, fsdata); +} + +/** + * reiser4_get_file_fsdata - get fs-specific file data + * @file: queried file + * + * Returns fs-specific data of @file. If it is NULL, allocates it and attaches + * to @file. + */ +reiser4_file_fsdata *reiser4_get_file_fsdata(struct file *file) +{ + assert("nikita-1603", file != NULL); + + if (file->private_data == NULL) { + reiser4_file_fsdata *fsdata; + struct inode *inode; + + fsdata = create_fsdata(file); + if (fsdata == NULL) + return ERR_PTR(RETERR(-ENOMEM)); + + inode = file->f_dentry->d_inode; + spin_lock_inode(inode); + if (file->private_data == NULL) { + file->private_data = fsdata; + fsdata = NULL; + } + spin_unlock_inode(inode); + if (fsdata != NULL) + /* other thread initialized ->fsdata */ + kmem_cache_free(file_fsdata_cache, fsdata); + } + assert("nikita-2665", file->private_data != NULL); + return file->private_data; +} + +/** + * free_file_fsdata_nolock - detach and free reiser4_file_fsdata + * @file: + * + * Detaches reiser4_file_fsdata from @file, removes reiser4_file_fsdata from + * readdir list, frees if it is not linked to d_cursor object. + */ +static void free_file_fsdata_nolock(struct file *file) +{ + reiser4_file_fsdata *fsdata; + + assert("", spin_inode_is_locked(file->f_dentry->d_inode)); + fsdata = file->private_data; + if (fsdata != NULL) { + list_del_init(&fsdata->dir.linkage); + if (fsdata->cursor == NULL) + free_fsdata(fsdata); + } + file->private_data = NULL; +} + +/** + * reiser4_free_file_fsdata - detach from struct file and free reiser4_file_fsdata + * @file: + * + * Spinlocks inode and calls free_file_fsdata_nolock to do the work. + */ +void reiser4_free_file_fsdata(struct file *file) +{ + spin_lock_inode(file->f_dentry->d_inode); + free_file_fsdata_nolock(file); + spin_unlock_inode(file->f_dentry->d_inode); +} + +/* + * Local variables: + * c-indentation-style: "K&R" + * mode-name: "LC" + * c-basic-offset: 8 + * tab-width: 8 + * fill-column: 79 + * End: + */ diff -urN linux-2.6.22.orig/fs/reiser4/fsdata.h linux-2.6.22/fs/reiser4/fsdata.h --- linux-2.6.22.orig/fs/reiser4/fsdata.h 1970-01-01 03:00:00.000000000 +0300 +++ linux-2.6.22/fs/reiser4/fsdata.h 2007-07-29 00:25:34.868694406 +0400 @@ -0,0 +1,205 @@ +/* Copyright 2001, 2002, 2003 by Hans Reiser, licensing governed by + * reiser4/README */ + +#if !defined( __REISER4_FSDATA_H__ ) +#define __REISER4_FSDATA_H__ + +#include "debug.h" +#include "kassign.h" +#include "seal.h" +#include "type_safe_hash.h" +#include "plugin/file/file.h" +#include "readahead.h" + +/* + * comment about reiser4_dentry_fsdata + * + * + */ + +/* + * locking: fields of per file descriptor readdir_pos and ->f_pos are + * protected by ->i_mutex on inode. Under this lock following invariant + * holds: + * + * file descriptor is "looking" at the entry_no-th directory entry from + * the beginning of directory. This entry has key dir_entry_key and is + * pos-th entry with duplicate-key sequence. + * + */ + +/* logical position within directory */ +struct dir_pos { + /* key of directory entry (actually, part of a key sufficient to + identify directory entry) */ + de_id dir_entry_key; + /* ordinal number of directory entry among all entries with the same + key. (Starting from 0.) */ + unsigned pos; +}; + +struct readdir_pos { + /* f_pos corresponding to this readdir position */ + __u64 fpos; + /* logical position within directory */ + struct dir_pos position; + /* logical number of directory entry within + directory */ + __u64 entry_no; +}; + +/* + * this is used to speed up lookups for directory entry: on initial call to + * ->lookup() seal and coord of directory entry (if found, that is) are stored + * in struct dentry and reused later to avoid tree traversals. + */ +struct de_location { + /* seal covering directory entry */ + seal_t entry_seal; + /* coord of directory entry */ + coord_t entry_coord; + /* ordinal number of directory entry among all entries with the same + key. (Starting from 0.) */ + int pos; +}; + +/** + * reiser4_dentry_fsdata - reiser4-specific data attached to dentries + * + * This is allocated dynamically and released in d_op->d_release() + * + * Currently it only contains cached location (hint) of directory entry, but + * it is expected that other information will be accumulated here. + */ +struct reiser4_dentry_fsdata { + /* + * here will go fields filled by ->lookup() to speedup next + * create/unlink, like blocknr of znode with stat-data, or key of + * stat-data. + */ + struct de_location dec; + int stateless; /* created through reiser4_decode_fh, needs special + * treatment in readdir. */ +}; + +extern int reiser4_init_dentry_fsdata(void); +extern void reiser4_done_dentry_fsdata(void); +extern struct reiser4_dentry_fsdata *reiser4_get_dentry_fsdata(struct dentry *); +extern void reiser4_free_dentry_fsdata(struct dentry *dentry); + +/** + * reiser4_file_fsdata - reiser4-specific data attached to file->private_data + * + * This is allocated dynamically and released in inode->i_fop->release + */ +typedef struct reiser4_file_fsdata { + /* + * pointer back to the struct file which this reiser4_file_fsdata is + * part of + */ + struct file *back; + /* detached cursor for stateless readdir. */ + struct dir_cursor *cursor; + /* + * We need both directory and regular file parts here, because there + * are file system objects that are files and directories. + */ + struct { + /* + * position in directory. It is updated each time directory is + * modified + */ + struct readdir_pos readdir; + /* head of this list is reiser4_inode->lists.readdir_list */ + struct list_head linkage; + } dir; + /* hints to speed up operations with regular files: read and write. */ + struct { + hint_t hint; + } reg; + struct reiser4_file_ra_state ra1; + +} reiser4_file_fsdata; + +extern int reiser4_init_file_fsdata(void); +extern void reiser4_done_file_fsdata(void); +extern reiser4_file_fsdata *reiser4_get_file_fsdata(struct file *); +extern void reiser4_free_file_fsdata(struct file *); + +/* + * d_cursor is reiser4_file_fsdata not attached to struct file. d_cursors are + * used to address problem reiser4 has with readdir accesses via NFS. See + * plugin/file_ops_readdir.c for more details. + */ +struct d_cursor_key{ + __u16 cid; + __u64 oid; +}; + +/* + * define structures d_cursor_hash_table d_cursor_hash_link which are used to + * maintain hash table of dir_cursor-s in reiser4's super block + */ +typedef struct dir_cursor dir_cursor; +TYPE_SAFE_HASH_DECLARE(d_cursor, dir_cursor); + +struct dir_cursor { + int ref; + reiser4_file_fsdata *fsdata; + + /* link to reiser4 super block hash table of cursors */ + d_cursor_hash_link hash; + + /* + * this is to link cursors to reiser4 super block's radix tree of + * cursors if there are more than one cursor of the same objectid + */ + struct list_head list; + struct d_cursor_key key; + struct d_cursor_info *info; + /* list of unused cursors */ + struct list_head alist; +}; + +extern int reiser4_init_d_cursor(void); +extern void reiser4_done_d_cursor(void); + +extern int reiser4_init_super_d_info(struct super_block *); +extern void reiser4_done_super_d_info(struct super_block *); + +extern loff_t reiser4_get_dir_fpos(struct file *); +extern int reiser4_attach_fsdata(struct file *, struct inode *); +extern void reiser4_detach_fsdata(struct file *); + +/* these are needed for "stateless" readdir. See plugin/file_ops_readdir.c for + more details */ +void reiser4_dispose_cursors(struct inode *inode); +void reiser4_load_cursors(struct inode *inode); +void reiser4_kill_cursors(struct inode *inode); +void reiser4_adjust_dir_file(struct inode *dir, const struct dentry *de, + int offset, int adj); + +/* + * this structure is embedded to reise4_super_info_data. It maintains d_cursors + * (detached readdir state). See plugin/file_ops_readdir.c for more details. + */ +struct d_cursor_info { + d_cursor_hash_table table; + struct radix_tree_root tree; +}; + +/* spinlock protecting readdir cursors */ +extern spinlock_t d_lock; + +/* __REISER4_FSDATA_H__ */ +#endif + +/* + * Local variables: + * c-indentation-style: "K&R" + * mode-name: "LC" + * c-basic-offset: 8 + * tab-width: 8 + * fill-column: 120 + * End: + */ diff -urN linux-2.6.22.orig/fs/reiser4/init_super.c linux-2.6.22/fs/reiser4/init_super.c --- linux-2.6.22.orig/fs/reiser4/init_super.c 1970-01-01 03:00:00.000000000 +0300 +++ linux-2.6.22/fs/reiser4/init_super.c 2007-07-29 00:25:34.868694406 +0400 @@ -0,0 +1,752 @@ +/* Copyright by Hans Reiser, 2003 */ + +#include "super.h" +#include "inode.h" +#include "plugin/plugin_set.h" + +#include + +/** + * init_fs_info - allocate reiser4 specific super block + * @super: super block of filesystem + * + * Allocates and initialize reiser4_super_info_data, attaches it to + * super->s_fs_info, initializes structures maintaining d_cursor-s. + */ +int reiser4_init_fs_info(struct super_block *super) +{ + reiser4_super_info_data *sbinfo; + + sbinfo = kmalloc(sizeof(reiser4_super_info_data), + reiser4_ctx_gfp_mask_get()); + if (!sbinfo) + return RETERR(-ENOMEM); + + super->s_fs_info = sbinfo; + super->s_op = NULL; + memset(sbinfo, 0, sizeof(*sbinfo)); + + ON_DEBUG(INIT_LIST_HEAD(&sbinfo->all_jnodes)); + ON_DEBUG(spin_lock_init(&sbinfo->all_guard)); + + mutex_init(&sbinfo->delete_mutex); + spin_lock_init(&(sbinfo->guard)); + + /* initialize per-super-block d_cursor resources */ + reiser4_init_super_d_info(super); + + return 0; +} + +/** + * reiser4_done_fs_info - free reiser4 specific super block + * @super: super block of filesystem + * + * Performs some sanity checks, releases structures maintaining d_cursor-s, + * frees reiser4_super_info_data. + */ +void reiser4_done_fs_info(struct super_block *super) +{ + assert("zam-990", super->s_fs_info != NULL); + + /* release per-super-block d_cursor resources */ + reiser4_done_super_d_info(super); + + /* make sure that there are not jnodes already */ + assert("", list_empty(&get_super_private(super)->all_jnodes)); + assert("", get_current_context()->trans->atom == NULL); + reiser4_check_block_counters(super); + kfree(super->s_fs_info); + super->s_fs_info = NULL; +} + +/* type of option parseable by parse_option() */ +typedef enum { + /* value of option is arbitrary string */ + OPT_STRING, + + /* + * option specifies bit in a bitmask. When option is set - bit in + * sbinfo->fs_flags is set. Examples are bsdgroups, 32bittimes, mtflush, + * dont_load_bitmap, atomic_write. + */ + OPT_BIT, + + /* + * value of option should conform to sprintf() format. Examples are + * tmgr.atom_max_size=N, tmgr.atom_max_age=N + */ + OPT_FORMAT, + + /* + * option can take one of predefined values. Example is onerror=panic or + * onerror=remount-ro + */ + OPT_ONEOF, +} opt_type_t; + +#if 0 +struct opt_bitmask_bit { + const char *bit_name; + int bit_nr; +}; +#endif + +/* description of option parseable by parse_option() */ +struct opt_desc { + /* option name. + + parsed portion of string has a form "name=value". + */ + const char *name; + /* type of option */ + opt_type_t type; + union { + /* where to store value of string option (type == OPT_STRING) */ + char **string; + /* description of bits for bit option (type == OPT_BIT) */ + struct { + int nr; + void *addr; + } bit; + /* description of format and targets for format option (type + == OPT_FORMAT) */ + struct { + const char *format; + int nr_args; + void *arg1; + void *arg2; + void *arg3; + void *arg4; + } f; + struct { + int *result; + const char *list[10]; + } oneof; + struct { + void *addr; + int nr_bits; + //struct opt_bitmask_bit *bits; + } bitmask; + } u; +}; + +/** + * parse_option - parse one option + * @opt_strin: starting point of parsing + * @opt: option description + * + * foo=bar, + * ^ ^ ^ + * | | +-- replaced to '\0' + * | +-- val_start + * +-- opt_string + * Figures out option type and handles option correspondingly. + */ +static int parse_option(char *opt_string, struct opt_desc *opt) +{ + char *val_start; + int result; + const char *err_msg; + + /* NOTE-NIKITA think about using lib/cmdline.c functions here. */ + + val_start = strchr(opt_string, '='); + if (val_start != NULL) { + *val_start = '\0'; + ++val_start; + } + + err_msg = NULL; + result = 0; + switch (opt->type) { + case OPT_STRING: + if (val_start == NULL) { + err_msg = "String arg missing"; + result = RETERR(-EINVAL); + } else + *opt->u.string = val_start; + break; + case OPT_BIT: + if (val_start != NULL) + err_msg = "Value ignored"; + else + set_bit(opt->u.bit.nr, opt->u.bit.addr); + break; + case OPT_FORMAT: + if (val_start == NULL) { + err_msg = "Formatted arg missing"; + result = RETERR(-EINVAL); + break; + } + if (sscanf(val_start, opt->u.f.format, + opt->u.f.arg1, opt->u.f.arg2, opt->u.f.arg3, + opt->u.f.arg4) != opt->u.f.nr_args) { + err_msg = "Wrong conversion"; + result = RETERR(-EINVAL); + } + break; + case OPT_ONEOF: + { + int i = 0; + + if (val_start == NULL) { + err_msg = "Value is missing"; + result = RETERR(-EINVAL); + break; + } + err_msg = "Wrong option value"; + result = RETERR(-EINVAL); + while (opt->u.oneof.list[i]) { + if (!strcmp(opt->u.oneof.list[i], val_start)) { + result = 0; + err_msg = NULL; + *opt->u.oneof.result = i; + break; + } + i++; + } + break; + } + default: + wrong_return_value("nikita-2100", "opt -> type"); + break; + } + if (err_msg != NULL) { + warning("nikita-2496", "%s when parsing option \"%s%s%s\"", + err_msg, opt->name, val_start ? "=" : "", + val_start ? : ""); + } + return result; +} + +/** + * parse_options - parse reiser4 mount options + * @opt_string: starting point + * @opts: array of option description + * @nr_opts: number of elements in @opts + * + * Parses comma separated list of reiser4 mount options. + */ +static int parse_options(char *opt_string, struct opt_desc *opts, int nr_opts) +{ + int result; + + result = 0; + while ((result == 0) && opt_string && *opt_string) { + int j; + char *next; + + next = strchr(opt_string, ','); + if (next != NULL) { + *next = '\0'; + ++next; + } + for (j = 0; j < nr_opts; ++j) { + if (!strncmp(opt_string, opts[j].name, + strlen(opts[j].name))) { + result = parse_option(opt_string, &opts[j]); + break; + } + } + if (j == nr_opts) { + warning("nikita-2307", "Unrecognized option: \"%s\"", + opt_string); + /* traditionally, -EINVAL is returned on wrong mount + option */ + result = RETERR(-EINVAL); + } + opt_string = next; + } + return result; +} + +#define NUM_OPT( label, fmt, addr ) \ + { \ + .name = ( label ), \ + .type = OPT_FORMAT, \ + .u = { \ + .f = { \ + .format = ( fmt ), \ + .nr_args = 1, \ + .arg1 = ( addr ), \ + .arg2 = NULL, \ + .arg3 = NULL, \ + .arg4 = NULL \ + } \ + } \ + } + +#define SB_FIELD_OPT( field, fmt ) NUM_OPT( #field, fmt, &sbinfo -> field ) + +#define BIT_OPT(label, bitnr) \ + { \ + .name = label, \ + .type = OPT_BIT, \ + .u = { \ + .bit = { \ + .nr = bitnr, \ + .addr = &sbinfo->fs_flags \ + } \ + } \ + } + +#define MAX_NR_OPTIONS (30) + +/** + * reiser4_init_super_data - initialize reiser4 private super block + * @super: super block to initialize + * @opt_string: list of reiser4 mount options + * + * Sets various reiser4 parameters to default values. Parses mount options and + * overwrites default settings. + */ +int reiser4_init_super_data(struct super_block *super, char *opt_string) +{ + int result; + struct opt_desc *opts, *p; + reiser4_super_info_data *sbinfo = get_super_private(super); + + /* initialize super, export, dentry operations */ + sbinfo->ops.super = reiser4_super_operations; + sbinfo->ops.export = reiser4_export_operations; + sbinfo->ops.dentry = reiser4_dentry_operations; + super->s_op = &sbinfo->ops.super; + super->s_export_op = &sbinfo->ops.export; + + /* initialize transaction manager parameters to default values */ + sbinfo->tmgr.atom_max_size = totalram_pages / 4; + sbinfo->tmgr.atom_max_age = REISER4_ATOM_MAX_AGE / HZ; + sbinfo->tmgr.at