ext4: fix unjournaled inode bitmap modification
authorEric Sandeen <sandeen@redhat.com>
Mon, 29 Oct 2012 02:24:57 +0000 (22:24 -0400)
committerTheodore Ts'o <tytso@mit.edu>
Mon, 29 Oct 2012 02:24:57 +0000 (22:24 -0400)
commit 119c0d4460b001e44b41dcf73dc6ee794b98bd31 changed
ext4_new_inode() such that the inode bitmap was being modified
outside a transaction, which could lead to corruption, and was
discovered when journal_checksum found a bad checksum in the
journal during log replay.

Nix ran into this when using the journal_async_commit mount
option, which enables journal checksumming.  The ensuing
journal replay failures due to the bad checksums led to
filesystem corruption reported as the now infamous
"Apparent serious progressive ext4 data corruption bug"

[ Changed by tytso to only call ext4_journal_get_write_access() only
  when we're fairly certain that we're going to allocate the inode. ]

I've tested this by mounting with journal_checksum and
running fsstress then dropping power; I've also tested by
hacking DM to create snapshots w/o first quiescing, which
allows me to test journal replay repeatedly w/o actually
power-cycling the box.  Without the patch I hit a journal
checksum error every time.  With this fix it survives
many iterations.

Reported-by: Nix <nix@esperi.org.uk>
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Cc: stable@vger.kernel.org
fs/ext4/ialloc.c

index 4facdd2..3a100e7 100644 (file)
@@ -725,6 +725,10 @@ repeat_in_this_group:
                                   "inode=%lu", ino + 1);
                        continue;
                }
+               BUFFER_TRACE(inode_bitmap_bh, "get_write_access");
+               err = ext4_journal_get_write_access(handle, inode_bitmap_bh);
+               if (err)
+                       goto fail;
                ext4_lock_group(sb, group);
                ret2 = ext4_test_and_set_bit(ino, inode_bitmap_bh->b_data);
                ext4_unlock_group(sb, group);
@@ -738,6 +742,11 @@ repeat_in_this_group:
        goto out;
 
 got:
+       BUFFER_TRACE(inode_bitmap_bh, "call ext4_handle_dirty_metadata");
+       err = ext4_handle_dirty_metadata(handle, NULL, inode_bitmap_bh);
+       if (err)
+               goto fail;
+
        /* We may have to initialize the block bitmap if it isn't already */
        if (ext4_has_group_desc_csum(sb) &&
            gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
@@ -771,11 +780,6 @@ got:
                        goto fail;
        }
 
-       BUFFER_TRACE(inode_bitmap_bh, "get_write_access");
-       err = ext4_journal_get_write_access(handle, inode_bitmap_bh);
-       if (err)
-               goto fail;
-
        BUFFER_TRACE(group_desc_bh, "get_write_access");
        err = ext4_journal_get_write_access(handle, group_desc_bh);
        if (err)
@@ -823,11 +827,6 @@ got:
        }
        ext4_unlock_group(sb, group);
 
-       BUFFER_TRACE(inode_bitmap_bh, "call ext4_handle_dirty_metadata");
-       err = ext4_handle_dirty_metadata(handle, NULL, inode_bitmap_bh);
-       if (err)
-               goto fail;
-
        BUFFER_TRACE(group_desc_bh, "call ext4_handle_dirty_metadata");
        err = ext4_handle_dirty_metadata(handle, NULL, group_desc_bh);
        if (err)