btrfs: fix race between quota disable and relocation

[ Upstream commit 8a4a0b2a3eaf75ca8854f856ef29690c12b2f531 ]

If we disable quotas while we have a relocation of a metadata block group
that has extents belonging to the quota root, we can cause the relocation
to fail with -ENOENT. This is because relocation builds backref nodes for
extents of the quota root and later needs to walk the backrefs and access
the quota root - however if in between a task disables quotas, it results
in deleting the quota root from the root tree (with btrfs_del_root(),
called from btrfs_quota_disable().

This can be sporadically triggered by test case btrfs/255 from fstests:

  $ ./check btrfs/255
  FSTYP         -- btrfs
  PLATFORM      -- Linux/x86_64 debian0 6.4.0-rc6-btrfs-next-134+ #1 SMP PREEMPT_DYNAMIC Thu Jun 15 11:59:28 WEST 2023
  MKFS_OPTIONS  -- /dev/sdc
  MOUNT_OPTIONS -- /dev/sdc /home/fdmanana/btrfs-tests/scratch_1

  btrfs/255 6s ... _check_dmesg: something found in dmesg (see /home/fdmanana/git/hub/xfstests/results//btrfs/255.dmesg)
  - output mismatch (see /home/fdmanana/git/hub/xfstests/results//btrfs/255.out.bad)
#      --- tests/btrfs/255.out	2023-03-02 21:47:53.876609426 +0000
#      +++ /home/fdmanana/git/hub/xfstests/results//btrfs/255.out.bad	2023-06-16 10:20:39.267563212 +0100
#      @@ -1,2 +1,4 @@
#       QA output created by 255
#      +ERROR: error during balancing '/home/fdmanana/btrfs-tests/scratch_1': No such file or directory
#      +There may be more info in syslog - try dmesg | tail
#       Silence is golden
#      ...
      (Run 'diff -u /home/fdmanana/git/hub/xfstests/tests/btrfs/255.out /home/fdmanana/git/hub/xfstests/results//btrfs/255.out.bad'  to see the entire diff)
  Ran: btrfs/255
  Failures: btrfs/255
  Failed 1 of 1 tests

To fix this make the quota disable operation take the cleaner mutex, as
relocation of a block group also takes this mutex. This is also what we
do when deleting a subvolume/snapshot, we take the cleaner mutex in the
cleaner kthread (at cleaner_kthread()) and then we call btrfs_del_root()
at btrfs_drop_snapshot() while under the protection of the cleaner mutex.

Fixes: bed92eae26cc ("Btrfs: qgroup implementation and prototypes")
CC: stable@vger.kernel.org # 5.4+
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>

1 files changed, 15 insertions, 3 deletions

diff --git a/fs/btrfs/qgroup.c b/fs/btrfs/qgroup.c
index d408d1dfde7c..d46a070275ff 100644
--- a/fs/btrfs/qgroup.c
+++ b/fs/btrfs/qgroup.c
@@ -1201,12 +1201,23 @@ int btrfs_quota_disable(struct btrfs_fs_info *fs_info)
 	int ret = 0;
 
 	/*
-	 * We need to have subvol_sem write locked, to prevent races between
-	 * concurrent tasks trying to disable quotas, because we will unlock
-	 * and relock qgroup_ioctl_lock across BTRFS_FS_QUOTA_ENABLED changes.
+	 * We need to have subvol_sem write locked to prevent races with
+	 * snapshot creation.
 	 */
 	lockdep_assert_held_write(&fs_info->subvol_sem);
 
+	/*
+	 * Lock the cleaner mutex to prevent races with concurrent relocation,
+	 * because relocation may be building backrefs for blocks of the quota
+	 * root while we are deleting the root. This is like dropping fs roots
+	 * of deleted snapshots/subvolumes, we need the same protection.
+	 *
+	 * This also prevents races between concurrent tasks trying to disable
+	 * quotas, because we will unlock and relock qgroup_ioctl_lock across
+	 * BTRFS_FS_QUOTA_ENABLED changes.
+	 */
+	mutex_lock(&fs_info->cleaner_mutex);
+
 	mutex_lock(&fs_info->qgroup_ioctl_lock);
 	if (!fs_info->quota_root)
 		goto out;
@@ -1287,6 +1298,7 @@ out:
 		btrfs_end_transaction(trans);
 	else if (trans)
 		ret = btrfs_end_transaction(trans);
+	mutex_unlock(&fs_info->cleaner_mutex);
 
 	return ret;
 }