Btrfs: fix file loss on log replay after renaming a file and fsync

We have two cases where we end up deleting a file at log replay time
when we should not. For this to happen the file must have been renamed
and a directory inode must have been fsynced/logged.

Two examples that exercise these two cases are listed below.

  Case 1)

  $ mkfs.btrfs -f /dev/sdb
  $ mount /dev/sdb /mnt
  $ mkdir -p /mnt/a/b
  $ mkdir /mnt/c
  $ touch /mnt/a/b/foo
  $ sync
  $ mv /mnt/a/b/foo /mnt/c/
  # Create file bar just to make sure the fsync on directory a/ does
  # something and it's not a no-op.
  $ touch /mnt/a/bar
  $ xfs_io -c "fsync" /mnt/a
  < power fail / crash >

  The next time the filesystem is mounted, the log replay procedure
  deletes file foo.

  Case 2)

  $ mkfs.btrfs -f /dev/sdb
  $ mount /dev/sdb /mnt
  $ mkdir /mnt/a
  $ mkdir /mnt/b
  $ mkdir /mnt/c
  $ touch /mnt/a/foo
  $ ln /mnt/a/foo /mnt/b/foo_link
  $ touch /mnt/b/bar
  $ sync
  $ unlink /mnt/b/foo_link
  $ mv /mnt/b/bar /mnt/c/
  $ xfs_io -c "fsync" /mnt/a/foo
  < power fail / crash >

  The next time the filesystem is mounted, the log replay procedure
  deletes file bar.

The reason why the files are deleted is because when we log inodes
other then the fsync target inode, we ignore their last_unlink_trans
value and leave the log without enough information to later replay the
rename operations. So we need to look at the last_unlink_trans values
and fallback to a transaction commit if they are greater than the
id of the last committed transaction.

So fix this by looking at the last_unlink_trans values and fallback to
transaction commits when needed. Also, when logging other inodes (for
case 1 we logged descendants of the fsync target inode while for case 2
we logged ascendants) we need to care about concurrent tasks updating
the last_unlink_trans of inodes we are logging (which was already an
existing problem in check_parent_dirs_for_sync()). Since we can not
acquire their inode mutex (vfs' struct inode ->i_mutex), as that causes
deadlocks with other concurrent operations that acquire the i_mutex of
2 inodes (other fsyncs or renames for example), we need to serialize on
the log_mutex of the inode we are logging. A task setting a new value for
an inode's last_unlink_trans must acquire the inode's log_mutex and it
must do this update before doing the actual unlink operation (which is
already the case except when deleting a snapshot). Conversely the task
logging the inode must first log the inode and then check the inode's
last_unlink_trans value while holding its log_mutex, as if its value is
not greater then the id of the last committed transaction it means it
logged a safe state of the inode's items, while if its value is not
smaller then the id of the last committed transaction it means the inode
state it has logged might not be safe (the concurrent task might have
just updated last_unlink_trans but hasn't done yet the unlink operation)
and therefore a transaction commit must be done.

Test cases for xfstests follow in separate patches.

Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>