Merge branch 'for_linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tytso/ext4
authorLinus Torvalds <torvalds@linux-foundation.org>
Mon, 1 Aug 2011 23:56:03 +0000 (13:56 -1000)
committerLinus Torvalds <torvalds@linux-foundation.org>
Mon, 1 Aug 2011 23:56:03 +0000 (13:56 -1000)
* 'for_linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tytso/ext4: (60 commits)
  ext4: prevent memory leaks from ext4_mb_init_backend() on error path
  ext4: use EXT4_BAD_INO for buddy cache to avoid colliding with valid inode #
  ext4: use ext4_msg() instead of printk in mballoc
  ext4: use ext4_kvzalloc()/ext4_kvmalloc() for s_group_desc and s_group_info
  ext4: introduce ext4_kvmalloc(), ext4_kzalloc(), and ext4_kvfree()
  ext4: use the correct error exit path in ext4_init_inode_table()
  ext4: add missing kfree() on error return path in add_new_gdb()
  ext4: change umode_t in tracepoint headers to be an explicit __u16
  ext4: fix races in ext4_sync_parent()
  ext4: Fix overflow caused by missing cast in ext4_fallocate()
  ext4: add action of moving index in ext4_ext_rm_idx for Punch Hole
  ext4: simplify parameters of reserve_backup_gdb()
  ext4: simplify parameters of add_new_gdb()
  ext4: remove lock_buffer in bclean() and setup_new_group_blocks()
  ext4: simplify journal handling in setup_new_group_blocks()
  ext4: let setup_new_group_blocks() set multiple bits at a time
  ext4: fix a typo in ext4_group_extend()
  ext4: let ext4_group_add_blocks() handle 0 blocks quickly
  ext4: let ext4_group_add_blocks() return an error code
  ext4: rename ext4_add_groupblocks() to ext4_group_add_blocks()
  ...

Fix up conflict in fs/ext4/inode.c: commit aacfc19c626e ("fs: simplify
the blockdev_direct_IO prototype") had changed the ext4_ind_direct_IO()
function for the new simplified calling convention, while commit
dae1e52cb126 ("ext4: move ext4_ind_* functions from inode.c to
indirect.c") moved the function to another file.

1  2 
fs/ext4/ext4.h
fs/ext4/fsync.c
fs/ext4/indirect.c
fs/ext4/inode.c
fs/ext4/namei.c
include/trace/events/ext4.h

diff --cc fs/ext4/ext4.h
Simple merge
diff --cc fs/ext4/fsync.c
Simple merge
index 0000000,6c27111..b8602cd
mode 000000,100644..100644
--- /dev/null
@@@ -1,0 -1,1484 +1,1482 @@@
 -              ret = blockdev_direct_IO(rw, iocb, inode,
 -                               inode->i_sb->s_bdev, iov,
 -                               offset, nr_segs,
 -                               ext4_get_block, NULL);
+ /*
+  *  linux/fs/ext4/indirect.c
+  *
+  *  from
+  *
+  *  linux/fs/ext4/inode.c
+  *
+  * Copyright (C) 1992, 1993, 1994, 1995
+  * Remy Card (card@masi.ibp.fr)
+  * Laboratoire MASI - Institut Blaise Pascal
+  * Universite Pierre et Marie Curie (Paris VI)
+  *
+  *  from
+  *
+  *  linux/fs/minix/inode.c
+  *
+  *  Copyright (C) 1991, 1992  Linus Torvalds
+  *
+  *  Goal-directed block allocation by Stephen Tweedie
+  *    (sct@redhat.com), 1993, 1998
+  */
+ #include <linux/module.h>
+ #include "ext4_jbd2.h"
+ #include "truncate.h"
+ #include <trace/events/ext4.h>
+ typedef struct {
+       __le32  *p;
+       __le32  key;
+       struct buffer_head *bh;
+ } Indirect;
+ static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
+ {
+       p->key = *(p->p = v);
+       p->bh = bh;
+ }
+ /**
+  *    ext4_block_to_path - parse the block number into array of offsets
+  *    @inode: inode in question (we are only interested in its superblock)
+  *    @i_block: block number to be parsed
+  *    @offsets: array to store the offsets in
+  *    @boundary: set this non-zero if the referred-to block is likely to be
+  *           followed (on disk) by an indirect block.
+  *
+  *    To store the locations of file's data ext4 uses a data structure common
+  *    for UNIX filesystems - tree of pointers anchored in the inode, with
+  *    data blocks at leaves and indirect blocks in intermediate nodes.
+  *    This function translates the block number into path in that tree -
+  *    return value is the path length and @offsets[n] is the offset of
+  *    pointer to (n+1)th node in the nth one. If @block is out of range
+  *    (negative or too large) warning is printed and zero returned.
+  *
+  *    Note: function doesn't find node addresses, so no IO is needed. All
+  *    we need to know is the capacity of indirect blocks (taken from the
+  *    inode->i_sb).
+  */
+ /*
+  * Portability note: the last comparison (check that we fit into triple
+  * indirect block) is spelled differently, because otherwise on an
+  * architecture with 32-bit longs and 8Kb pages we might get into trouble
+  * if our filesystem had 8Kb blocks. We might use long long, but that would
+  * kill us on x86. Oh, well, at least the sign propagation does not matter -
+  * i_block would have to be negative in the very beginning, so we would not
+  * get there at all.
+  */
+ static int ext4_block_to_path(struct inode *inode,
+                             ext4_lblk_t i_block,
+                             ext4_lblk_t offsets[4], int *boundary)
+ {
+       int ptrs = EXT4_ADDR_PER_BLOCK(inode->i_sb);
+       int ptrs_bits = EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb);
+       const long direct_blocks = EXT4_NDIR_BLOCKS,
+               indirect_blocks = ptrs,
+               double_blocks = (1 << (ptrs_bits * 2));
+       int n = 0;
+       int final = 0;
+       if (i_block < direct_blocks) {
+               offsets[n++] = i_block;
+               final = direct_blocks;
+       } else if ((i_block -= direct_blocks) < indirect_blocks) {
+               offsets[n++] = EXT4_IND_BLOCK;
+               offsets[n++] = i_block;
+               final = ptrs;
+       } else if ((i_block -= indirect_blocks) < double_blocks) {
+               offsets[n++] = EXT4_DIND_BLOCK;
+               offsets[n++] = i_block >> ptrs_bits;
+               offsets[n++] = i_block & (ptrs - 1);
+               final = ptrs;
+       } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
+               offsets[n++] = EXT4_TIND_BLOCK;
+               offsets[n++] = i_block >> (ptrs_bits * 2);
+               offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
+               offsets[n++] = i_block & (ptrs - 1);
+               final = ptrs;
+       } else {
+               ext4_warning(inode->i_sb, "block %lu > max in inode %lu",
+                            i_block + direct_blocks +
+                            indirect_blocks + double_blocks, inode->i_ino);
+       }
+       if (boundary)
+               *boundary = final - 1 - (i_block & (ptrs - 1));
+       return n;
+ }
+ /**
+  *    ext4_get_branch - read the chain of indirect blocks leading to data
+  *    @inode: inode in question
+  *    @depth: depth of the chain (1 - direct pointer, etc.)
+  *    @offsets: offsets of pointers in inode/indirect blocks
+  *    @chain: place to store the result
+  *    @err: here we store the error value
+  *
+  *    Function fills the array of triples <key, p, bh> and returns %NULL
+  *    if everything went OK or the pointer to the last filled triple
+  *    (incomplete one) otherwise. Upon the return chain[i].key contains
+  *    the number of (i+1)-th block in the chain (as it is stored in memory,
+  *    i.e. little-endian 32-bit), chain[i].p contains the address of that
+  *    number (it points into struct inode for i==0 and into the bh->b_data
+  *    for i>0) and chain[i].bh points to the buffer_head of i-th indirect
+  *    block for i>0 and NULL for i==0. In other words, it holds the block
+  *    numbers of the chain, addresses they were taken from (and where we can
+  *    verify that chain did not change) and buffer_heads hosting these
+  *    numbers.
+  *
+  *    Function stops when it stumbles upon zero pointer (absent block)
+  *            (pointer to last triple returned, *@err == 0)
+  *    or when it gets an IO error reading an indirect block
+  *            (ditto, *@err == -EIO)
+  *    or when it reads all @depth-1 indirect blocks successfully and finds
+  *    the whole chain, all way to the data (returns %NULL, *err == 0).
+  *
+  *      Need to be called with
+  *      down_read(&EXT4_I(inode)->i_data_sem)
+  */
+ static Indirect *ext4_get_branch(struct inode *inode, int depth,
+                                ext4_lblk_t  *offsets,
+                                Indirect chain[4], int *err)
+ {
+       struct super_block *sb = inode->i_sb;
+       Indirect *p = chain;
+       struct buffer_head *bh;
+       *err = 0;
+       /* i_data is not going away, no lock needed */
+       add_chain(chain, NULL, EXT4_I(inode)->i_data + *offsets);
+       if (!p->key)
+               goto no_block;
+       while (--depth) {
+               bh = sb_getblk(sb, le32_to_cpu(p->key));
+               if (unlikely(!bh))
+                       goto failure;
+               if (!bh_uptodate_or_lock(bh)) {
+                       if (bh_submit_read(bh) < 0) {
+                               put_bh(bh);
+                               goto failure;
+                       }
+                       /* validate block references */
+                       if (ext4_check_indirect_blockref(inode, bh)) {
+                               put_bh(bh);
+                               goto failure;
+                       }
+               }
+               add_chain(++p, bh, (__le32 *)bh->b_data + *++offsets);
+               /* Reader: end */
+               if (!p->key)
+                       goto no_block;
+       }
+       return NULL;
+ failure:
+       *err = -EIO;
+ no_block:
+       return p;
+ }
+ /**
+  *    ext4_find_near - find a place for allocation with sufficient locality
+  *    @inode: owner
+  *    @ind: descriptor of indirect block.
+  *
+  *    This function returns the preferred place for block allocation.
+  *    It is used when heuristic for sequential allocation fails.
+  *    Rules are:
+  *      + if there is a block to the left of our position - allocate near it.
+  *      + if pointer will live in indirect block - allocate near that block.
+  *      + if pointer will live in inode - allocate in the same
+  *        cylinder group.
+  *
+  * In the latter case we colour the starting block by the callers PID to
+  * prevent it from clashing with concurrent allocations for a different inode
+  * in the same block group.   The PID is used here so that functionally related
+  * files will be close-by on-disk.
+  *
+  *    Caller must make sure that @ind is valid and will stay that way.
+  */
+ static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
+ {
+       struct ext4_inode_info *ei = EXT4_I(inode);
+       __le32 *start = ind->bh ? (__le32 *) ind->bh->b_data : ei->i_data;
+       __le32 *p;
+       /* Try to find previous block */
+       for (p = ind->p - 1; p >= start; p--) {
+               if (*p)
+                       return le32_to_cpu(*p);
+       }
+       /* No such thing, so let's try location of indirect block */
+       if (ind->bh)
+               return ind->bh->b_blocknr;
+       /*
+        * It is going to be referred to from the inode itself? OK, just put it
+        * into the same cylinder group then.
+        */
+       return ext4_inode_to_goal_block(inode);
+ }
+ /**
+  *    ext4_find_goal - find a preferred place for allocation.
+  *    @inode: owner
+  *    @block:  block we want
+  *    @partial: pointer to the last triple within a chain
+  *
+  *    Normally this function find the preferred place for block allocation,
+  *    returns it.
+  *    Because this is only used for non-extent files, we limit the block nr
+  *    to 32 bits.
+  */
+ static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
+                                  Indirect *partial)
+ {
+       ext4_fsblk_t goal;
+       /*
+        * XXX need to get goal block from mballoc's data structures
+        */
+       goal = ext4_find_near(inode, partial);
+       goal = goal & EXT4_MAX_BLOCK_FILE_PHYS;
+       return goal;
+ }
+ /**
+  *    ext4_blks_to_allocate - Look up the block map and count the number
+  *    of direct blocks need to be allocated for the given branch.
+  *
+  *    @branch: chain of indirect blocks
+  *    @k: number of blocks need for indirect blocks
+  *    @blks: number of data blocks to be mapped.
+  *    @blocks_to_boundary:  the offset in the indirect block
+  *
+  *    return the total number of blocks to be allocate, including the
+  *    direct and indirect blocks.
+  */
+ static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned int blks,
+                                int blocks_to_boundary)
+ {
+       unsigned int count = 0;
+       /*
+        * Simple case, [t,d]Indirect block(s) has not allocated yet
+        * then it's clear blocks on that path have not allocated
+        */
+       if (k > 0) {
+               /* right now we don't handle cross boundary allocation */
+               if (blks < blocks_to_boundary + 1)
+                       count += blks;
+               else
+                       count += blocks_to_boundary + 1;
+               return count;
+       }
+       count++;
+       while (count < blks && count <= blocks_to_boundary &&
+               le32_to_cpu(*(branch[0].p + count)) == 0) {
+               count++;
+       }
+       return count;
+ }
+ /**
+  *    ext4_alloc_blocks: multiple allocate blocks needed for a branch
+  *    @handle: handle for this transaction
+  *    @inode: inode which needs allocated blocks
+  *    @iblock: the logical block to start allocated at
+  *    @goal: preferred physical block of allocation
+  *    @indirect_blks: the number of blocks need to allocate for indirect
+  *                    blocks
+  *    @blks: number of desired blocks
+  *    @new_blocks: on return it will store the new block numbers for
+  *    the indirect blocks(if needed) and the first direct block,
+  *    @err: on return it will store the error code
+  *
+  *    This function will return the number of blocks allocated as
+  *    requested by the passed-in parameters.
+  */
+ static int ext4_alloc_blocks(handle_t *handle, struct inode *inode,
+                            ext4_lblk_t iblock, ext4_fsblk_t goal,
+                            int indirect_blks, int blks,
+                            ext4_fsblk_t new_blocks[4], int *err)
+ {
+       struct ext4_allocation_request ar;
+       int target, i;
+       unsigned long count = 0, blk_allocated = 0;
+       int index = 0;
+       ext4_fsblk_t current_block = 0;
+       int ret = 0;
+       /*
+        * Here we try to allocate the requested multiple blocks at once,
+        * on a best-effort basis.
+        * To build a branch, we should allocate blocks for
+        * the indirect blocks(if not allocated yet), and at least
+        * the first direct block of this branch.  That's the
+        * minimum number of blocks need to allocate(required)
+        */
+       /* first we try to allocate the indirect blocks */
+       target = indirect_blks;
+       while (target > 0) {
+               count = target;
+               /* allocating blocks for indirect blocks and direct blocks */
+               current_block = ext4_new_meta_blocks(handle, inode, goal,
+                                                    0, &count, err);
+               if (*err)
+                       goto failed_out;
+               if (unlikely(current_block + count > EXT4_MAX_BLOCK_FILE_PHYS)) {
+                       EXT4_ERROR_INODE(inode,
+                                        "current_block %llu + count %lu > %d!",
+                                        current_block, count,
+                                        EXT4_MAX_BLOCK_FILE_PHYS);
+                       *err = -EIO;
+                       goto failed_out;
+               }
+               target -= count;
+               /* allocate blocks for indirect blocks */
+               while (index < indirect_blks && count) {
+                       new_blocks[index++] = current_block++;
+                       count--;
+               }
+               if (count > 0) {
+                       /*
+                        * save the new block number
+                        * for the first direct block
+                        */
+                       new_blocks[index] = current_block;
+                       printk(KERN_INFO "%s returned more blocks than "
+                                               "requested\n", __func__);
+                       WARN_ON(1);
+                       break;
+               }
+       }
+       target = blks - count ;
+       blk_allocated = count;
+       if (!target)
+               goto allocated;
+       /* Now allocate data blocks */
+       memset(&ar, 0, sizeof(ar));
+       ar.inode = inode;
+       ar.goal = goal;
+       ar.len = target;
+       ar.logical = iblock;
+       if (S_ISREG(inode->i_mode))
+               /* enable in-core preallocation only for regular files */
+               ar.flags = EXT4_MB_HINT_DATA;
+       current_block = ext4_mb_new_blocks(handle, &ar, err);
+       if (unlikely(current_block + ar.len > EXT4_MAX_BLOCK_FILE_PHYS)) {
+               EXT4_ERROR_INODE(inode,
+                                "current_block %llu + ar.len %d > %d!",
+                                current_block, ar.len,
+                                EXT4_MAX_BLOCK_FILE_PHYS);
+               *err = -EIO;
+               goto failed_out;
+       }
+       if (*err && (target == blks)) {
+               /*
+                * if the allocation failed and we didn't allocate
+                * any blocks before
+                */
+               goto failed_out;
+       }
+       if (!*err) {
+               if (target == blks) {
+                       /*
+                        * save the new block number
+                        * for the first direct block
+                        */
+                       new_blocks[index] = current_block;
+               }
+               blk_allocated += ar.len;
+       }
+ allocated:
+       /* total number of blocks allocated for direct blocks */
+       ret = blk_allocated;
+       *err = 0;
+       return ret;
+ failed_out:
+       for (i = 0; i < index; i++)
+               ext4_free_blocks(handle, inode, NULL, new_blocks[i], 1, 0);
+       return ret;
+ }
+ /**
+  *    ext4_alloc_branch - allocate and set up a chain of blocks.
+  *    @handle: handle for this transaction
+  *    @inode: owner
+  *    @indirect_blks: number of allocated indirect blocks
+  *    @blks: number of allocated direct blocks
+  *    @goal: preferred place for allocation
+  *    @offsets: offsets (in the blocks) to store the pointers to next.
+  *    @branch: place to store the chain in.
+  *
+  *    This function allocates blocks, zeroes out all but the last one,
+  *    links them into chain and (if we are synchronous) writes them to disk.
+  *    In other words, it prepares a branch that can be spliced onto the
+  *    inode. It stores the information about that chain in the branch[], in
+  *    the same format as ext4_get_branch() would do. We are calling it after
+  *    we had read the existing part of chain and partial points to the last
+  *    triple of that (one with zero ->key). Upon the exit we have the same
+  *    picture as after the successful ext4_get_block(), except that in one
+  *    place chain is disconnected - *branch->p is still zero (we did not
+  *    set the last link), but branch->key contains the number that should
+  *    be placed into *branch->p to fill that gap.
+  *
+  *    If allocation fails we free all blocks we've allocated (and forget
+  *    their buffer_heads) and return the error value the from failed
+  *    ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
+  *    as described above and return 0.
+  */
+ static int ext4_alloc_branch(handle_t *handle, struct inode *inode,
+                            ext4_lblk_t iblock, int indirect_blks,
+                            int *blks, ext4_fsblk_t goal,
+                            ext4_lblk_t *offsets, Indirect *branch)
+ {
+       int blocksize = inode->i_sb->s_blocksize;
+       int i, n = 0;
+       int err = 0;
+       struct buffer_head *bh;
+       int num;
+       ext4_fsblk_t new_blocks[4];
+       ext4_fsblk_t current_block;
+       num = ext4_alloc_blocks(handle, inode, iblock, goal, indirect_blks,
+                               *blks, new_blocks, &err);
+       if (err)
+               return err;
+       branch[0].key = cpu_to_le32(new_blocks[0]);
+       /*
+        * metadata blocks and data blocks are allocated.
+        */
+       for (n = 1; n <= indirect_blks;  n++) {
+               /*
+                * Get buffer_head for parent block, zero it out
+                * and set the pointer to new one, then send
+                * parent to disk.
+                */
+               bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
+               if (unlikely(!bh)) {
+                       err = -EIO;
+                       goto failed;
+               }
+               branch[n].bh = bh;
+               lock_buffer(bh);
+               BUFFER_TRACE(bh, "call get_create_access");
+               err = ext4_journal_get_create_access(handle, bh);
+               if (err) {
+                       /* Don't brelse(bh) here; it's done in
+                        * ext4_journal_forget() below */
+                       unlock_buffer(bh);
+                       goto failed;
+               }
+               memset(bh->b_data, 0, blocksize);
+               branch[n].p = (__le32 *) bh->b_data + offsets[n];
+               branch[n].key = cpu_to_le32(new_blocks[n]);
+               *branch[n].p = branch[n].key;
+               if (n == indirect_blks) {
+                       current_block = new_blocks[n];
+                       /*
+                        * End of chain, update the last new metablock of
+                        * the chain to point to the new allocated
+                        * data blocks numbers
+                        */
+                       for (i = 1; i < num; i++)
+                               *(branch[n].p + i) = cpu_to_le32(++current_block);
+               }
+               BUFFER_TRACE(bh, "marking uptodate");
+               set_buffer_uptodate(bh);
+               unlock_buffer(bh);
+               BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
+               err = ext4_handle_dirty_metadata(handle, inode, bh);
+               if (err)
+                       goto failed;
+       }
+       *blks = num;
+       return err;
+ failed:
+       /* Allocation failed, free what we already allocated */
+       ext4_free_blocks(handle, inode, NULL, new_blocks[0], 1, 0);
+       for (i = 1; i <= n ; i++) {
+               /*
+                * branch[i].bh is newly allocated, so there is no
+                * need to revoke the block, which is why we don't
+                * need to set EXT4_FREE_BLOCKS_METADATA.
+                */
+               ext4_free_blocks(handle, inode, NULL, new_blocks[i], 1,
+                                EXT4_FREE_BLOCKS_FORGET);
+       }
+       for (i = n+1; i < indirect_blks; i++)
+               ext4_free_blocks(handle, inode, NULL, new_blocks[i], 1, 0);
+       ext4_free_blocks(handle, inode, NULL, new_blocks[i], num, 0);
+       return err;
+ }
+ /**
+  * ext4_splice_branch - splice the allocated branch onto inode.
+  * @handle: handle for this transaction
+  * @inode: owner
+  * @block: (logical) number of block we are adding
+  * @chain: chain of indirect blocks (with a missing link - see
+  *    ext4_alloc_branch)
+  * @where: location of missing link
+  * @num:   number of indirect blocks we are adding
+  * @blks:  number of direct blocks we are adding
+  *
+  * This function fills the missing link and does all housekeeping needed in
+  * inode (->i_blocks, etc.). In case of success we end up with the full
+  * chain to new block and return 0.
+  */
+ static int ext4_splice_branch(handle_t *handle, struct inode *inode,
+                             ext4_lblk_t block, Indirect *where, int num,
+                             int blks)
+ {
+       int i;
+       int err = 0;
+       ext4_fsblk_t current_block;
+       /*
+        * If we're splicing into a [td]indirect block (as opposed to the
+        * inode) then we need to get write access to the [td]indirect block
+        * before the splice.
+        */
+       if (where->bh) {
+               BUFFER_TRACE(where->bh, "get_write_access");
+               err = ext4_journal_get_write_access(handle, where->bh);
+               if (err)
+                       goto err_out;
+       }
+       /* That's it */
+       *where->p = where->key;
+       /*
+        * Update the host buffer_head or inode to point to more just allocated
+        * direct blocks blocks
+        */
+       if (num == 0 && blks > 1) {
+               current_block = le32_to_cpu(where->key) + 1;
+               for (i = 1; i < blks; i++)
+                       *(where->p + i) = cpu_to_le32(current_block++);
+       }
+       /* We are done with atomic stuff, now do the rest of housekeeping */
+       /* had we spliced it onto indirect block? */
+       if (where->bh) {
+               /*
+                * If we spliced it onto an indirect block, we haven't
+                * altered the inode.  Note however that if it is being spliced
+                * onto an indirect block at the very end of the file (the
+                * file is growing) then we *will* alter the inode to reflect
+                * the new i_size.  But that is not done here - it is done in
+                * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
+                */
+               jbd_debug(5, "splicing indirect only\n");
+               BUFFER_TRACE(where->bh, "call ext4_handle_dirty_metadata");
+               err = ext4_handle_dirty_metadata(handle, inode, where->bh);
+               if (err)
+                       goto err_out;
+       } else {
+               /*
+                * OK, we spliced it into the inode itself on a direct block.
+                */
+               ext4_mark_inode_dirty(handle, inode);
+               jbd_debug(5, "splicing direct\n");
+       }
+       return err;
+ err_out:
+       for (i = 1; i <= num; i++) {
+               /*
+                * branch[i].bh is newly allocated, so there is no
+                * need to revoke the block, which is why we don't
+                * need to set EXT4_FREE_BLOCKS_METADATA.
+                */
+               ext4_free_blocks(handle, inode, where[i].bh, 0, 1,
+                                EXT4_FREE_BLOCKS_FORGET);
+       }
+       ext4_free_blocks(handle, inode, NULL, le32_to_cpu(where[num].key),
+                        blks, 0);
+       return err;
+ }
+ /*
+  * The ext4_ind_map_blocks() function handles non-extents inodes
+  * (i.e., using the traditional indirect/double-indirect i_blocks
+  * scheme) for ext4_map_blocks().
+  *
+  * Allocation strategy is simple: if we have to allocate something, we will
+  * have to go the whole way to leaf. So let's do it before attaching anything
+  * to tree, set linkage between the newborn blocks, write them if sync is
+  * required, recheck the path, free and repeat if check fails, otherwise
+  * set the last missing link (that will protect us from any truncate-generated
+  * removals - all blocks on the path are immune now) and possibly force the
+  * write on the parent block.
+  * That has a nice additional property: no special recovery from the failed
+  * allocations is needed - we simply release blocks and do not touch anything
+  * reachable from inode.
+  *
+  * `handle' can be NULL if create == 0.
+  *
+  * return > 0, # of blocks mapped or allocated.
+  * return = 0, if plain lookup failed.
+  * return < 0, error case.
+  *
+  * The ext4_ind_get_blocks() function should be called with
+  * down_write(&EXT4_I(inode)->i_data_sem) if allocating filesystem
+  * blocks (i.e., flags has EXT4_GET_BLOCKS_CREATE set) or
+  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system
+  * blocks.
+  */
+ int ext4_ind_map_blocks(handle_t *handle, struct inode *inode,
+                       struct ext4_map_blocks *map,
+                       int flags)
+ {
+       int err = -EIO;
+       ext4_lblk_t offsets[4];
+       Indirect chain[4];
+       Indirect *partial;
+       ext4_fsblk_t goal;
+       int indirect_blks;
+       int blocks_to_boundary = 0;
+       int depth;
+       int count = 0;
+       ext4_fsblk_t first_block = 0;
+       trace_ext4_ind_map_blocks_enter(inode, map->m_lblk, map->m_len, flags);
+       J_ASSERT(!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)));
+       J_ASSERT(handle != NULL || (flags & EXT4_GET_BLOCKS_CREATE) == 0);
+       depth = ext4_block_to_path(inode, map->m_lblk, offsets,
+                                  &blocks_to_boundary);
+       if (depth == 0)
+               goto out;
+       partial = ext4_get_branch(inode, depth, offsets, chain, &err);
+       /* Simplest case - block found, no allocation needed */
+       if (!partial) {
+               first_block = le32_to_cpu(chain[depth - 1].key);
+               count++;
+               /*map more blocks*/
+               while (count < map->m_len && count <= blocks_to_boundary) {
+                       ext4_fsblk_t blk;
+                       blk = le32_to_cpu(*(chain[depth-1].p + count));
+                       if (blk == first_block + count)
+                               count++;
+                       else
+                               break;
+               }
+               goto got_it;
+       }
+       /* Next simple case - plain lookup or failed read of indirect block */
+       if ((flags & EXT4_GET_BLOCKS_CREATE) == 0 || err == -EIO)
+               goto cleanup;
+       /*
+        * Okay, we need to do block allocation.
+       */
+       goal = ext4_find_goal(inode, map->m_lblk, partial);
+       /* the number of blocks need to allocate for [d,t]indirect blocks */
+       indirect_blks = (chain + depth) - partial - 1;
+       /*
+        * Next look up the indirect map to count the totoal number of
+        * direct blocks to allocate for this branch.
+        */
+       count = ext4_blks_to_allocate(partial, indirect_blks,
+                                     map->m_len, blocks_to_boundary);
+       /*
+        * Block out ext4_truncate while we alter the tree
+        */
+       err = ext4_alloc_branch(handle, inode, map->m_lblk, indirect_blks,
+                               &count, goal,
+                               offsets + (partial - chain), partial);
+       /*
+        * The ext4_splice_branch call will free and forget any buffers
+        * on the new chain if there is a failure, but that risks using
+        * up transaction credits, especially for bitmaps where the
+        * credits cannot be returned.  Can we handle this somehow?  We
+        * may need to return -EAGAIN upwards in the worst case.  --sct
+        */
+       if (!err)
+               err = ext4_splice_branch(handle, inode, map->m_lblk,
+                                        partial, indirect_blks, count);
+       if (err)
+               goto cleanup;
+       map->m_flags |= EXT4_MAP_NEW;
+       ext4_update_inode_fsync_trans(handle, inode, 1);
+ got_it:
+       map->m_flags |= EXT4_MAP_MAPPED;
+       map->m_pblk = le32_to_cpu(chain[depth-1].key);
+       map->m_len = count;
+       if (count > blocks_to_boundary)
+               map->m_flags |= EXT4_MAP_BOUNDARY;
+       err = count;
+       /* Clean up and exit */
+       partial = chain + depth - 1;    /* the whole chain */
+ cleanup:
+       while (partial > chain) {
+               BUFFER_TRACE(partial->bh, "call brelse");
+               brelse(partial->bh);
+               partial--;
+       }
+ out:
+       trace_ext4_ind_map_blocks_exit(inode, map->m_lblk,
+                               map->m_pblk, map->m_len, err);
+       return err;
+ }
+ /*
+  * O_DIRECT for ext3 (or indirect map) based files
+  *
+  * If the O_DIRECT write will extend the file then add this inode to the
+  * orphan list.  So recovery will truncate it back to the original size
+  * if the machine crashes during the write.
+  *
+  * If the O_DIRECT write is intantiating holes inside i_size and the machine
+  * crashes then stale disk data _may_ be exposed inside the file. But current
+  * VFS code falls back into buffered path in that case so we are safe.
+  */
+ ssize_t ext4_ind_direct_IO(int rw, struct kiocb *iocb,
+                          const struct iovec *iov, loff_t offset,
+                          unsigned long nr_segs)
+ {
+       struct file *file = iocb->ki_filp;
+       struct inode *inode = file->f_mapping->host;
+       struct ext4_inode_info *ei = EXT4_I(inode);
+       handle_t *handle;
+       ssize_t ret;
+       int orphan = 0;
+       size_t count = iov_length(iov, nr_segs);
+       int retries = 0;
+       if (rw == WRITE) {
+               loff_t final_size = offset + count;
+               if (final_size > inode->i_size) {
+                       /* Credits for sb + inode write */
+                       handle = ext4_journal_start(inode, 2);
+                       if (IS_ERR(handle)) {
+                               ret = PTR_ERR(handle);
+                               goto out;
+                       }
+                       ret = ext4_orphan_add(handle, inode);
+                       if (ret) {
+                               ext4_journal_stop(handle);
+                               goto out;
+                       }
+                       orphan = 1;
+                       ei->i_disksize = inode->i_size;
+                       ext4_journal_stop(handle);
+               }
+       }
+ retry:
+       if (rw == READ && ext4_should_dioread_nolock(inode))
+               ret = __blockdev_direct_IO(rw, iocb, inode,
+                                inode->i_sb->s_bdev, iov,
+                                offset, nr_segs,
+                                ext4_get_block, NULL, NULL, 0);
+       else {
++              ret = blockdev_direct_IO(rw, iocb, inode, iov,
++                               offset, nr_segs, ext4_get_block);
+               if (unlikely((rw & WRITE) && ret < 0)) {
+                       loff_t isize = i_size_read(inode);
+                       loff_t end = offset + iov_length(iov, nr_segs);
+                       if (end > isize)
+                               ext4_truncate_failed_write(inode);
+               }
+       }
+       if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
+               goto retry;
+       if (orphan) {
+               int err;
+               /* Credits for sb + inode write */
+               handle = ext4_journal_start(inode, 2);
+               if (IS_ERR(handle)) {
+                       /* This is really bad luck. We've written the data
+                        * but cannot extend i_size. Bail out and pretend
+                        * the write failed... */
+                       ret = PTR_ERR(handle);
+                       if (inode->i_nlink)
+                               ext4_orphan_del(NULL, inode);
+                       goto out;
+               }
+               if (inode->i_nlink)
+                       ext4_orphan_del(handle, inode);
+               if (ret > 0) {
+                       loff_t end = offset + ret;
+                       if (end > inode->i_size) {
+                               ei->i_disksize = end;
+                               i_size_write(inode, end);
+                               /*
+                                * We're going to return a positive `ret'
+                                * here due to non-zero-length I/O, so there's
+                                * no way of reporting error returns from
+                                * ext4_mark_inode_dirty() to userspace.  So
+                                * ignore it.
+                                */
+                               ext4_mark_inode_dirty(handle, inode);
+                       }
+               }
+               err = ext4_journal_stop(handle);
+               if (ret == 0)
+                       ret = err;
+       }
+ out:
+       return ret;
+ }
+ /*
+  * Calculate the number of metadata blocks need to reserve
+  * to allocate a new block at @lblocks for non extent file based file
+  */
+ int ext4_ind_calc_metadata_amount(struct inode *inode, sector_t lblock)
+ {
+       struct ext4_inode_info *ei = EXT4_I(inode);
+       sector_t dind_mask = ~((sector_t)EXT4_ADDR_PER_BLOCK(inode->i_sb) - 1);
+       int blk_bits;
+       if (lblock < EXT4_NDIR_BLOCKS)
+               return 0;
+       lblock -= EXT4_NDIR_BLOCKS;
+       if (ei->i_da_metadata_calc_len &&
+           (lblock & dind_mask) == ei->i_da_metadata_calc_last_lblock) {
+               ei->i_da_metadata_calc_len++;
+               return 0;
+       }
+       ei->i_da_metadata_calc_last_lblock = lblock & dind_mask;
+       ei->i_da_metadata_calc_len = 1;
+       blk_bits = order_base_2(lblock);
+       return (blk_bits / EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb)) + 1;
+ }
+ int ext4_ind_trans_blocks(struct inode *inode, int nrblocks, int chunk)
+ {
+       int indirects;
+       /* if nrblocks are contiguous */
+       if (chunk) {
+               /*
+                * With N contiguous data blocks, we need at most
+                * N/EXT4_ADDR_PER_BLOCK(inode->i_sb) + 1 indirect blocks,
+                * 2 dindirect blocks, and 1 tindirect block
+                */
+               return DIV_ROUND_UP(nrblocks,
+                                   EXT4_ADDR_PER_BLOCK(inode->i_sb)) + 4;
+       }
+       /*
+        * if nrblocks are not contiguous, worse case, each block touch
+        * a indirect block, and each indirect block touch a double indirect
+        * block, plus a triple indirect block
+        */
+       indirects = nrblocks * 2 + 1;
+       return indirects;
+ }
+ /*
+  * Truncate transactions can be complex and absolutely huge.  So we need to
+  * be able to restart the transaction at a conventient checkpoint to make
+  * sure we don't overflow the journal.
+  *
+  * start_transaction gets us a new handle for a truncate transaction,
+  * and extend_transaction tries to extend the existing one a bit.  If
+  * extend fails, we need to propagate the failure up and restart the
+  * transaction in the top-level truncate loop. --sct
+  */
+ static handle_t *start_transaction(struct inode *inode)
+ {
+       handle_t *result;
+       result = ext4_journal_start(inode, ext4_blocks_for_truncate(inode));
+       if (!IS_ERR(result))
+               return result;
+       ext4_std_error(inode->i_sb, PTR_ERR(result));
+       return result;
+ }
+ /*
+  * Try to extend this transaction for the purposes of truncation.
+  *
+  * Returns 0 if we managed to create more room.  If we can't create more
+  * room, and the transaction must be restarted we return 1.
+  */
+ static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
+ {
+       if (!ext4_handle_valid(handle))
+               return 0;
+       if (ext4_handle_has_enough_credits(handle, EXT4_RESERVE_TRANS_BLOCKS+1))
+               return 0;
+       if (!ext4_journal_extend(handle, ext4_blocks_for_truncate(inode)))
+               return 0;
+       return 1;
+ }
+ /*
+  * Probably it should be a library function... search for first non-zero word
+  * or memcmp with zero_page, whatever is better for particular architecture.
+  * Linus?
+  */
+ static inline int all_zeroes(__le32 *p, __le32 *q)
+ {
+       while (p < q)
+               if (*p++)
+                       return 0;
+       return 1;
+ }
+ /**
+  *    ext4_find_shared - find the indirect blocks for partial truncation.
+  *    @inode:   inode in question
+  *    @depth:   depth of the affected branch
+  *    @offsets: offsets of pointers in that branch (see ext4_block_to_path)
+  *    @chain:   place to store the pointers to partial indirect blocks
+  *    @top:     place to the (detached) top of branch
+  *
+  *    This is a helper function used by ext4_truncate().
+  *
+  *    When we do truncate() we may have to clean the ends of several
+  *    indirect blocks but leave the blocks themselves alive. Block is
+  *    partially truncated if some data below the new i_size is referred
+  *    from it (and it is on the path to the first completely truncated
+  *    data block, indeed).  We have to free the top of that path along
+  *    with everything to the right of the path. Since no allocation
+  *    past the truncation point is possible until ext4_truncate()
+  *    finishes, we may safely do the latter, but top of branch may
+  *    require special attention - pageout below the truncation point
+  *    might try to populate it.
+  *
+  *    We atomically detach the top of branch from the tree, store the
+  *    block number of its root in *@top, pointers to buffer_heads of
+  *    partially truncated blocks - in @chain[].bh and pointers to
+  *    their last elements that should not be removed - in
+  *    @chain[].p. Return value is the pointer to last filled element
+  *    of @chain.
+  *
+  *    The work left to caller to do the actual freeing of subtrees:
+  *            a) free the subtree starting from *@top
+  *            b) free the subtrees whose roots are stored in
+  *                    (@chain[i].p+1 .. end of @chain[i].bh->b_data)
+  *            c) free the subtrees growing from the inode past the @chain[0].
+  *                    (no partially truncated stuff there).  */
+ static Indirect *ext4_find_shared(struct inode *inode, int depth,
+                                 ext4_lblk_t offsets[4], Indirect chain[4],
+                                 __le32 *top)
+ {
+       Indirect *partial, *p;
+       int k, err;
+       *top = 0;
+       /* Make k index the deepest non-null offset + 1 */
+       for (k = depth; k > 1 && !offsets[k-1]; k--)
+               ;
+       partial = ext4_get_branch(inode, k, offsets, chain, &err);
+       /* Writer: pointers */
+       if (!partial)
+               partial = chain + k-1;
+       /*
+        * If the branch acquired continuation since we've looked at it -
+        * fine, it should all survive and (new) top doesn't belong to us.
+        */
+       if (!partial->key && *partial->p)
+               /* Writer: end */
+               goto no_top;
+       for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
+               ;
+       /*
+        * OK, we've found the last block that must survive. The rest of our
+        * branch should be detached before unlocking. However, if that rest
+        * of branch is all ours and does not grow immediately from the inode
+        * it's easier to cheat and just decrement partial->p.
+        */
+       if (p == chain + k - 1 && p > chain) {
+               p->p--;
+       } else {
+               *top = *p->p;
+               /* Nope, don't do this in ext4.  Must leave the tree intact */
+ #if 0
+               *p->p = 0;
+ #endif
+       }
+       /* Writer: end */
+       while (partial > p) {
+               brelse(partial->bh);
+               partial--;
+       }
+ no_top:
+       return partial;
+ }
+ /*
+  * Zero a number of block pointers in either an inode or an indirect block.
+  * If we restart the transaction we must again get write access to the
+  * indirect block for further modification.
+  *
+  * We release `count' blocks on disk, but (last - first) may be greater
+  * than `count' because there can be holes in there.
+  *
+  * Return 0 on success, 1 on invalid block range
+  * and < 0 on fatal error.
+  */
+ static int ext4_clear_blocks(handle_t *handle, struct inode *inode,
+                            struct buffer_head *bh,
+                            ext4_fsblk_t block_to_free,
+                            unsigned long count, __le32 *first,
+                            __le32 *last)
+ {
+       __le32 *p;
+       int     flags = EXT4_FREE_BLOCKS_FORGET | EXT4_FREE_BLOCKS_VALIDATED;
+       int     err;
+       if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
+               flags |= EXT4_FREE_BLOCKS_METADATA;
+       if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), block_to_free,
+                                  count)) {
+               EXT4_ERROR_INODE(inode, "attempt to clear invalid "
+                                "blocks %llu len %lu",
+                                (unsigned long long) block_to_free, count);
+               return 1;
+       }
+       if (try_to_extend_transaction(handle, inode)) {
+               if (bh) {
+                       BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
+                       err = ext4_handle_dirty_metadata(handle, inode, bh);
+                       if (unlikely(err))
+                               goto out_err;
+               }
+               err = ext4_mark_inode_dirty(handle, inode);
+               if (unlikely(err))
+                       goto out_err;
+               err = ext4_truncate_restart_trans(handle, inode,
+                                       ext4_blocks_for_truncate(inode));
+               if (unlikely(err))
+                       goto out_err;
+               if (bh) {
+                       BUFFER_TRACE(bh, "retaking write access");
+                       err = ext4_journal_get_write_access(handle, bh);
+                       if (unlikely(err))
+                               goto out_err;
+               }
+       }
+       for (p = first; p < last; p++)
+               *p = 0;
+       ext4_free_blocks(handle, inode, NULL, block_to_free, count, flags);
+       return 0;
+ out_err:
+       ext4_std_error(inode->i_sb, err);
+       return err;
+ }
+ /**
+  * ext4_free_data - free a list of data blocks
+  * @handle:   handle for this transaction
+  * @inode:    inode we are dealing with
+  * @this_bh:  indirect buffer_head which contains *@first and *@last
+  * @first:    array of block numbers
+  * @last:     points immediately past the end of array
+  *
+  * We are freeing all blocks referred from that array (numbers are stored as
+  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
+  *
+  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
+  * blocks are contiguous then releasing them at one time will only affect one
+  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
+  * actually use a lot of journal space.
+  *
+  * @this_bh will be %NULL if @first and @last point into the inode's direct
+  * block pointers.
+  */
+ static void ext4_free_data(handle_t *handle, struct inode *inode,
+                          struct buffer_head *this_bh,
+                          __le32 *first, __le32 *last)
+ {
+       ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
+       unsigned long count = 0;            /* Number of blocks in the run */
+       __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
+                                              corresponding to
+                                              block_to_free */
+       ext4_fsblk_t nr;                    /* Current block # */
+       __le32 *p;                          /* Pointer into inode/ind
+                                              for current block */
+       int err = 0;
+       if (this_bh) {                          /* For indirect block */
+               BUFFER_TRACE(this_bh, "get_write_access");
+               err = ext4_journal_get_write_access(handle, this_bh);
+               /* Important: if we can't update the indirect pointers
+                * to the blocks, we can't free them. */
+               if (err)
+                       return;
+       }
+       for (p = first; p < last; p++) {
+               nr = le32_to_cpu(*p);
+               if (nr) {
+                       /* accumulate blocks to free if they're contiguous */
+                       if (count == 0) {
+                               block_to_free = nr;
+                               block_to_free_p = p;
+                               count = 1;
+                       } else if (nr == block_to_free + count) {
+                               count++;
+                       } else {
+                               err = ext4_clear_blocks(handle, inode, this_bh,
+                                                       block_to_free, count,
+                                                       block_to_free_p, p);
+                               if (err)
+                                       break;
+                               block_to_free = nr;
+                               block_to_free_p = p;
+                               count = 1;
+                       }
+               }
+       }
+       if (!err && count > 0)
+               err = ext4_clear_blocks(handle, inode, this_bh, block_to_free,
+                                       count, block_to_free_p, p);
+       if (err < 0)
+               /* fatal error */
+               return;
+       if (this_bh) {
+               BUFFER_TRACE(this_bh, "call ext4_handle_dirty_metadata");
+               /*
+                * The buffer head should have an attached journal head at this
+                * point. However, if the data is corrupted and an indirect
+                * block pointed to itself, it would have been detached when
+                * the block was cleared. Check for this instead of OOPSing.
+                */
+               if ((EXT4_JOURNAL(inode) == NULL) || bh2jh(this_bh))
+                       ext4_handle_dirty_metadata(handle, inode, this_bh);
+               else
+                       EXT4_ERROR_INODE(inode,
+                                        "circular indirect block detected at "
+                                        "block %llu",
+                               (unsigned long long) this_bh->b_blocknr);
+       }
+ }
+ /**
+  *    ext4_free_branches - free an array of branches
+  *    @handle: JBD handle for this transaction
+  *    @inode: inode we are dealing with
+  *    @parent_bh: the buffer_head which contains *@first and *@last
+  *    @first: array of block numbers
+  *    @last:  pointer immediately past the end of array
+  *    @depth: depth of the branches to free
+  *
+  *    We are freeing all blocks referred from these branches (numbers are
+  *    stored as little-endian 32-bit) and updating @inode->i_blocks
+  *    appropriately.
+  */
+ static void ext4_free_branches(handle_t *handle, struct inode *inode,
+                              struct buffer_head *parent_bh,
+                              __le32 *first, __le32 *last, int depth)
+ {
+       ext4_fsblk_t nr;
+       __le32 *p;
+       if (ext4_handle_is_aborted(handle))
+               return;
+       if (depth--) {
+               struct buffer_head *bh;
+               int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
+               p = last;
+               while (--p >= first) {
+                       nr = le32_to_cpu(*p);
+                       if (!nr)
+                               continue;               /* A hole */
+                       if (!ext4_data_block_valid(EXT4_SB(inode->i_sb),
+                                                  nr, 1)) {
+                               EXT4_ERROR_INODE(inode,
+                                                "invalid indirect mapped "
+                                                "block %lu (level %d)",
+                                                (unsigned long) nr, depth);
+                               break;
+                       }
+                       /* Go read the buffer for the next level down */
+                       bh = sb_bread(inode->i_sb, nr);
+                       /*
+                        * A read failure? Report error and clear slot
+                        * (should be rare).
+                        */
+                       if (!bh) {
+                               EXT4_ERROR_INODE_BLOCK(inode, nr,
+                                                      "Read failure");
+                               continue;
+                       }
+                       /* This zaps the entire block.  Bottom up. */
+                       BUFFER_TRACE(bh, "free child branches");
+                       ext4_free_branches(handle, inode, bh,
+                                       (__le32 *) bh->b_data,
+                                       (__le32 *) bh->b_data + addr_per_block,
+                                       depth);
+                       brelse(bh);
+                       /*
+                        * Everything below this this pointer has been
+                        * released.  Now let this top-of-subtree go.
+                        *
+                        * We want the freeing of this indirect block to be
+                        * atomic in the journal with the updating of the
+                        * bitmap block which owns it.  So make some room in
+                        * the journal.
+                        *
+                        * We zero the parent pointer *after* freeing its
+                        * pointee in the bitmaps, so if extend_transaction()
+                        * for some reason fails to put the bitmap changes and
+                        * the release into the same transaction, recovery
+                        * will merely complain about releasing a free block,
+                        * rather than leaking blocks.
+                        */
+                       if (ext4_handle_is_aborted(handle))
+                               return;
+                       if (try_to_extend_transaction(handle, inode)) {
+                               ext4_mark_inode_dirty(handle, inode);
+                               ext4_truncate_restart_trans(handle, inode,
+                                           ext4_blocks_for_truncate(inode));
+                       }
+                       /*
+                        * The forget flag here is critical because if
+                        * we are journaling (and not doing data
+                        * journaling), we have to make sure a revoke
+                        * record is written to prevent the journal
+                        * replay from overwriting the (former)
+                        * indirect block if it gets reallocated as a
+                        * data block.  This must happen in the same
+                        * transaction where the data blocks are
+                        * actually freed.
+                        */
+                       ext4_free_blocks(handle, inode, NULL, nr, 1,
+                                        EXT4_FREE_BLOCKS_METADATA|
+                                        EXT4_FREE_BLOCKS_FORGET);
+                       if (parent_bh) {
+                               /*
+                                * The block which we have just freed is
+                                * pointed to by an indirect block: journal it
+                                */
+                               BUFFER_TRACE(parent_bh, "get_write_access");
+                               if (!ext4_journal_get_write_access(handle,
+                                                                  parent_bh)){
+                                       *p = 0;
+                                       BUFFER_TRACE(parent_bh,
+                                       "call ext4_handle_dirty_metadata");
+                                       ext4_handle_dirty_metadata(handle,
+                                                                  inode,
+                                                                  parent_bh);
+                               }
+                       }
+               }
+       } else {
+               /* We have reached the bottom of the tree. */
+               BUFFER_TRACE(parent_bh, "free data blocks");
+               ext4_free_data(handle, inode, parent_bh, first, last);
+       }
+ }
+ void ext4_ind_truncate(struct inode *inode)
+ {
+       handle_t *handle;
+       struct ext4_inode_info *ei = EXT4_I(inode);
+       __le32 *i_data = ei->i_data;
+       int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
+       struct address_space *mapping = inode->i_mapping;
+       ext4_lblk_t offsets[4];
+       Indirect chain[4];
+       Indirect *partial;
+       __le32 nr = 0;
+       int n = 0;
+       ext4_lblk_t last_block, max_block;
+       unsigned blocksize = inode->i_sb->s_blocksize;
+       handle = start_transaction(inode);
+       if (IS_ERR(handle))
+               return;         /* AKPM: return what? */
+       last_block = (inode->i_size + blocksize-1)
+                                       >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
+       max_block = (EXT4_SB(inode->i_sb)->s_bitmap_maxbytes + blocksize-1)
+                                       >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
+       if (inode->i_size & (blocksize - 1))
+               if (ext4_block_truncate_page(handle, mapping, inode->i_size))
+                       goto out_stop;
+       if (last_block != max_block) {
+               n = ext4_block_to_path(inode, last_block, offsets, NULL);
+               if (n == 0)
+                       goto out_stop;  /* error */
+       }
+       /*
+        * OK.  This truncate is going to happen.  We add the inode to the
+        * orphan list, so that if this truncate spans multiple transactions,
+        * and we crash, we will resume the truncate when the filesystem
+        * recovers.  It also marks the inode dirty, to catch the new size.
+        *
+        * Implication: the file must always be in a sane, consistent
+        * truncatable state while each transaction commits.
+        */
+       if (ext4_orphan_add(handle, inode))
+               goto out_stop;
+       /*
+        * From here we block out all ext4_get_block() callers who want to
+        * modify the block allocation tree.
+        */
+       down_write(&ei->i_data_sem);
+       ext4_discard_preallocations(inode);
+       /*
+        * The orphan list entry will now protect us from any crash which
+        * occurs before the truncate completes, so it is now safe to propagate
+        * the new, shorter inode size (held for now in i_size) into the
+        * on-disk inode. We do this via i_disksize, which is the value which
+        * ext4 *really* writes onto the disk inode.
+        */
+       ei->i_disksize = inode->i_size;
+       if (last_block == max_block) {
+               /*
+                * It is unnecessary to free any data blocks if last_block is
+                * equal to the indirect block limit.
+                */
+               goto out_unlock;
+       } else if (n == 1) {            /* direct blocks */
+               ext4_free_data(handle, inode, NULL, i_data+offsets[0],
+                              i_data + EXT4_NDIR_BLOCKS);
+               goto do_indirects;
+       }
+       partial = ext4_find_shared(inode, n, offsets, chain, &nr);
+       /* Kill the top of shared branch (not detached) */
+       if (nr) {
+               if (partial == chain) {
+                       /* Shared branch grows from the inode */
+                       ext4_free_branches(handle, inode, NULL,
+                                          &nr, &nr+1, (chain+n-1) - partial);
+                       *partial->p = 0;
+                       /*
+                        * We mark the inode dirty prior to restart,
+                        * and prior to stop.  No need for it here.
+                        */
+               } else {
+                       /* Shared branch grows from an indirect block */
+                       BUFFER_TRACE(partial->bh, "get_write_access");
+                       ext4_free_branches(handle, inode, partial->bh,
+                                       partial->p,
+                                       partial->p+1, (chain+n-1) - partial);
+               }
+       }
+       /* Clear the ends of indirect blocks on the shared branch */
+       while (partial > chain) {
+               ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
+                                  (__le32*)partial->bh->b_data+addr_per_block,
+                                  (chain+n-1) - partial);
+               BUFFER_TRACE(partial->bh, "call brelse");
+               brelse(partial->bh);
+               partial--;
+       }
+ do_indirects:
+       /* Kill the remaining (whole) subtrees */
+       switch (offsets[0]) {
+       default:
+               nr = i_data[EXT4_IND_BLOCK];
+               if (nr) {
+                       ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
+                       i_data[EXT4_IND_BLOCK] = 0;
+               }
+       case EXT4_IND_BLOCK:
+               nr = i_data[EXT4_DIND_BLOCK];
+               if (nr) {
+                       ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
+                       i_data[EXT4_DIND_BLOCK] = 0;
+               }
+       case EXT4_DIND_BLOCK:
+               nr = i_data[EXT4_TIND_BLOCK];
+               if (nr) {
+                       ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
+                       i_data[EXT4_TIND_BLOCK] = 0;
+               }
+       case EXT4_TIND_BLOCK:
+               ;
+       }
+ out_unlock:
+       up_write(&ei->i_data_sem);
+       inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
+       ext4_mark_inode_dirty(handle, inode);
+       /*
+        * In a multi-transaction truncate, we only make the final transaction
+        * synchronous
+        */
+       if (IS_SYNC(inode))
+               ext4_handle_sync(handle);
+ out_stop:
+       /*
+        * If this was a simple ftruncate(), and the file will remain alive
+        * then we need to clear up the orphan record which we created above.
+        * However, if this was a real unlink then we were called by
+        * ext4_delete_inode(), and we allow that function to clean up the
+        * orphan info for us.
+        */
+       if (inode->i_nlink)
+               ext4_orphan_del(handle, inode);
+       ext4_journal_stop(handle);
+       trace_ext4_truncate_exit(inode);
+ }
diff --cc fs/ext4/inode.c
Simple merge
diff --cc fs/ext4/namei.c
Simple merge
Simple merge