ext4: partial revert to fix double brelse WARNING()
authorTheodore Ts'o <tytso@mit.edu>
Sun, 8 Nov 2009 20:45:44 +0000 (15:45 -0500)
committerTheodore Ts'o <tytso@mit.edu>
Sun, 8 Nov 2009 20:45:44 +0000 (15:45 -0500)
This is a partial revert of commit 6487a9d (only the changes made to
fs/ext4/namei.c), since it is causing the following brelse()
double-free warning when running fsstress on a file system with 1k
blocksize and we run into a block allocation failure while converting
a single-block directory to a multi-block hash-tree indexed directory.

WARNING: at fs/buffer.c:1197 __brelse+0x2e/0x33()
Hardware name:
VFS: brelse: Trying to free free buffer
Modules linked in:
Pid: 2226, comm: jbd2/sdd-8 Not tainted 2.6.32-rc6-00577-g0003f55 #101
Call Trace:
 [<c01587fb>] warn_slowpath_common+0x65/0x95
 [<c0158869>] warn_slowpath_fmt+0x29/0x2c
 [<c021168e>] __brelse+0x2e/0x33
 [<c0288a9f>] jbd2_journal_refile_buffer+0x67/0x6c
 [<c028a9ed>] jbd2_journal_commit_transaction+0x319/0x14d8
 [<c0164d73>] ? try_to_del_timer_sync+0x58/0x60
 [<c0175bcc>] ? sched_clock_cpu+0x12a/0x13e
 [<c017f6b4>] ? trace_hardirqs_off+0xb/0xd
 [<c0175c1f>] ? cpu_clock+0x3f/0x5b
 [<c017f6ec>] ? lock_release_holdtime+0x36/0x137
 [<c0664ad0>] ? _spin_unlock_irqrestore+0x44/0x51
 [<c0180af3>] ? trace_hardirqs_on_caller+0x103/0x124
 [<c0180b1f>] ? trace_hardirqs_on+0xb/0xd
 [<c0164d73>] ? try_to_del_timer_sync+0x58/0x60
 [<c0290d1c>] kjournald2+0x11a/0x310
 [<c017118e>] ? autoremove_wake_function+0x0/0x38
 [<c0290c02>] ? kjournald2+0x0/0x310
 [<c0170ee6>] kthread+0x66/0x6b
 [<c0170e80>] ? kthread+0x0/0x6b
 [<c01251b3>] kernel_thread_helper+0x7/0x10
---[ end trace 5579351b86af61e3 ]---

Commit 6487a9d was an attempt some buffer head leaks in an ENOSPC
error path, but in some cases it actually results in an excess ENOSPC,
as shown above.  Fixing this means cleaning up who is responsible for
releasing the buffer heads from the callee to the caller of
add_dirent_to_buf().

Since that's a relatively complex change, and we're late in the rcX
development cycle, I'm reverting this now, and holding back a more
complete fix until after 2.6.32 ships.  We've lived with this
buffer_head leak on ENOSPC in ext3 and ext4 for a very long time; a
few more months won't kill us.

Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Cc: Curt Wohlgemuth <curtw@google.com>
fs/ext4/namei.c

index 7c8fe80bacdd93ef64b5cec3f300690099c56a37..6d2c1b897fc778bd6a08d680aef414a97ac5aa5e 100644 (file)
@@ -1518,12 +1518,8 @@ static int ext4_add_entry(handle_t *handle, struct dentry *dentry,
                        return retval;
 
                if (blocks == 1 && !dx_fallback &&
-                   EXT4_HAS_COMPAT_FEATURE(sb, EXT4_FEATURE_COMPAT_DIR_INDEX)) {
-                       retval = make_indexed_dir(handle, dentry, inode, bh);
-                       if (retval == -ENOSPC)
-                               brelse(bh);
-                       return retval;
-               }
+                   EXT4_HAS_COMPAT_FEATURE(sb, EXT4_FEATURE_COMPAT_DIR_INDEX))
+                       return make_indexed_dir(handle, dentry, inode, bh);
                brelse(bh);
        }
        bh = ext4_append(handle, dir, &block, &retval);
@@ -1532,10 +1528,7 @@ static int ext4_add_entry(handle_t *handle, struct dentry *dentry,
        de = (struct ext4_dir_entry_2 *) bh->b_data;
        de->inode = 0;
        de->rec_len = ext4_rec_len_to_disk(blocksize, blocksize);
-       retval = add_dirent_to_buf(handle, dentry, inode, de, bh);
-       if (retval == -ENOSPC)
-               brelse(bh);
-       return retval;
+       return add_dirent_to_buf(handle, dentry, inode, de, bh);
 }
 
 /*
@@ -1664,8 +1657,7 @@ static int ext4_dx_add_entry(handle_t *handle, struct dentry *dentry,
        if (!de)
                goto cleanup;
        err = add_dirent_to_buf(handle, dentry, inode, de, bh);
-       if (err != -ENOSPC)
-               bh = NULL;
+       bh = NULL;
        goto cleanup;
 
 journal_error: