queue-4.14/btrfs-fix-fsync-not-persisting-changed-attributes-of-a-directory.patch

   1 From 60d9f50308e5df19bc18c2fefab0eba4a843900a Mon Sep 17 00:00:00 2001
   2 From: Filipe Manana <fdmanana@suse.com>
   3 Date: Thu, 16 May 2019 15:48:55 +0100
   4 Subject: Btrfs: fix fsync not persisting changed attributes of a directory
   5
   6 From: Filipe Manana <fdmanana@suse.com>
   7
   8 commit 60d9f50308e5df19bc18c2fefab0eba4a843900a upstream.
   9
  10 While logging an inode we follow its ancestors and for each one we mark
  11 it as logged in the current transaction, even if we have not logged it.
  12 As a consequence if we change an attribute of an ancestor, such as the
  13 UID or GID for example, and then explicitly fsync it, we end up not
  14 logging the inode at all despite returning success to user space, which
  15 results in the attribute being lost if a power failure happens after
  16 the fsync.
  17
  18 Sample reproducer:
  19
  20   $ mkfs.btrfs -f /dev/sdb
  21   $ mount /dev/sdb /mnt
  22
  23   $ mkdir /mnt/dir
  24   $ chown 6007:6007 /mnt/dir
  25
  26   $ sync
  27
  28   $ chown 9003:9003 /mnt/dir
  29   $ touch /mnt/dir/file
  30   $ xfs_io -c fsync /mnt/dir/file
  31
  32   # fsync our directory after fsync'ing the new file, should persist the
  33   # new values for the uid and gid.
  34   $ xfs_io -c fsync /mnt/dir
  35
  36   <power failure>
  37
  38   $ mount /dev/sdb /mnt
  39   $ stat -c %u:%g /mnt/dir
  40   6007:6007
  41
  42     --> should be 9003:9003, the uid and gid were not persisted, despite
  43         the explicit fsync on the directory prior to the power failure
  44
  45 Fix this by not updating the logged_trans field of ancestor inodes when
  46 logging an inode, since we have not logged them. Let only future calls to
  47 btrfs_log_inode() to mark inodes as logged.
  48
  49 This could be triggered by my recent fsync fuzz tester for fstests, for
  50 which an fstests patch exists titled "fstests: generic, fsync fuzz tester
  51 with fsstress".
  52
  53 Fixes: 12fcfd22fe5b ("Btrfs: tree logging unlink/rename fixes")
  54 CC: stable@vger.kernel.org # 4.4+
  55 Signed-off-by: Filipe Manana <fdmanana@suse.com>
  56 Signed-off-by: David Sterba <dsterba@suse.com>
  57 Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
  58
  59 ---
  60  fs/btrfs/tree-log.c |   12 ------------
  61  1 file changed, 12 deletions(-)
  62
  63 --- a/fs/btrfs/tree-log.c
  64 +++ b/fs/btrfs/tree-log.c
  65 @@ -5332,7 +5332,6 @@ static noinline int check_parent_dirs_fo
  66  {
  67         int ret = 0;
  68         struct dentry *old_parent = NULL;
  69 -       struct btrfs_inode *orig_inode = inode;
  70
  71         /*
  72          * for regular files, if its inode is already on disk, we don't
  73 @@ -5352,16 +5351,6 @@ static noinline int check_parent_dirs_fo
  74         }
  75
  76         while (1) {
  77 -               /*
  78 -                * If we are logging a directory then we start with our inode,
  79 -                * not our parent's inode, so we need to skip setting the
  80 -                * logged_trans so that further down in the log code we don't
  81 -                * think this inode has already been logged.
  82 -                */
  83 -               if (inode != orig_inode)
  84 -                       inode->logged_trans = trans->transid;
  85 -               smp_mb();
  86 -
  87                 if (btrfs_must_commit_transaction(trans, inode)) {
  88                         ret = 1;
  89                         break;
  90 @@ -6091,7 +6080,6 @@ void btrfs_record_unlink_dir(struct btrf
  91          * if this directory was already logged any new
  92          * names for this file/dir will get recorded
  93          */
  94 -       smp_mb();
  95         if (dir->logged_trans == trans->transid)
  96                 return;
  97