btrfs: send: fix duplicated rmdir operations when using extrefs

Commit 29d6d30f ("Btrfs: send, don't send rmdir for same target
multiple times") has fixed an issue that a send stream contained a rmdir
operation for the same directory multiple times. After that fix we keep
track of the last directory for which we sent a rmdir operation and
compare with it before sending a rmdir for the parent inode of a deleted
hardlink we are processing. But there is still a corner case that in
between rmdir dir operations for the same inode we find deleted hardlinks
for other parent inodes, so tracking just the last inode for which we sent
a rmdir operation is not enough.

Hardlinks of a file in the same directory are stored in the same INODE_REF
item, but if the number of hardlinks is too large and can not fit in a
leaf, we use INODE_EXTREF items to store them. The key of an INODE_EXTREF
item is (inode_id, INODE_EXTREF, hash[name, parent ino]), so between two
hardlinks for the same parent directory, we can find others for other
parent directories. For example for the reproducer below we get the
following (from a btrfs inspect-internal dump-tree output):

    item 0 key (259 INODE_EXTREF 2309449) itemoff 16257 itemsize 26
            index 6925 parent 257 namelen 8 name: foo.6923
    item 1 key (259 INODE_EXTREF 2311350) itemoff 16231 itemsize 26
            index 6588 parent 258 namelen 8 name: foo.6587
    item 2 key (259 INODE_EXTREF 2457395) itemoff 16205 itemsize 26
            index 6611 parent 257 namelen 8 name: foo.6609
    (...)

So tracking the last directory's inode number does not work in this case
since we process a link for parent inode 257, then for 258 and then back
again for 257, and that second time we process a deleted link for 257 we
think we have not yet sent a rmdir operation.

Fix this by using a rbtree to keep track of all the directories for which
we have already sent rmdir operations, and add those directories to the
'check_dirs' ref list in process_recorded_refs() only if the directory is
not yet in the rbtree, otherwise skip it since it means we have already
sent a rmdir operation for that directory.

The following test script reproduces the problem:

  $ cat test.sh
  #!/bin/bash

  DEV=/dev/sdi
  MNT=/mnt/sdi

  mkfs.btrfs -f $DEV
  mount $DEV $MNT

  mkdir $MNT/a $MNT/b

  echo 123 > $MNT/a/foo
  for ((i = 1; i <= 1000; i++)); do
     ln $MNT/a/foo $MNT/a/foo.$i
     ln $MNT/a/foo $MNT/b/foo.$i
  done

  btrfs subvolume snapshot -r $MNT $MNT/snap1
  btrfs send $MNT/snap1 -f /tmp/base.send

  rm -r $MNT/a $MNT/b

  btrfs subvolume snapshot -r $MNT $MNT/snap2
  btrfs send -p $MNT/snap1 $MNT/snap2 -f /tmp/incremental.send

  umount $MNT
  mkfs.btrfs -f $DEV
  mount $DEV $MNT

  btrfs receive $MNT -f /tmp/base.send
  btrfs receive $MNT -f /tmp/incremental.send

  rm -f /tmp/base.send /tmp/incremental.send

  umount $MNT

When running it, it fails like this:

  $ ./test.sh
  (...)
  At subvol snap1
  At snapshot snap2
  ERROR: rmdir o257-9-0 failed: No such file or directory

CC: <stable@vger.kernel.org>
Reviewed-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Ting-Chang Hou <tchou@synology.com>
[ Updated changelog ]
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>