Just curious about the status of this patch. :) Let me know if you
need further information.

Thanks!

On Tue, Sep 7, 2021 at 10:44 AM Baptiste Lepers
[off-list ref] wrote:

No, they need to be between the reads to have an effect. See
https://www.kernel.org/doc/Documentation/memory-barriers.txt §SMP
BARRIER PAIRING ("When dealing with CPU-CPU interactions..."). You
will see that the barriers are always between the ordered reads and
not before.

I think that Paul, the barrier guru, can confirm that the barrier was
misplaced in the original code? :)


On Tue, Sep 7, 2021 at 10:43 AM Baptiste Lepers
[off-list ref] wrote:

quoted

On Mon, Sep 6, 2021 at 10:27 PM David Sterba [off-list ref] wrote:

quoted

On Mon, Sep 06, 2021 at 11:25:59AM +1000, Baptiste Lepers wrote:

quoted

Per comment, record_root_in_trans orders the writes of the root->state
and root->last_trans:
      set_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
      smp_wmb();
      root->last_trans = trans->transid;

But the barrier that enforces the order on the read side is misplaced:
     smp_rmb(); <-- misplaced
     if (root->last_trans == trans->transid &&
    <-- missing barrier here -->
            !test_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state))

This patches fixes the ordering and wraps the racy accesses with
READ_ONCE and WRITE_ONCE calls to avoid load/store tearing.

Fixes: 7585717f304f5 ("Btrfs: fix relocation races")
Signed-off-by: Baptiste Lepers <redacted>
---
 fs/btrfs/transaction.c | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/fs/btrfs/transaction.c b/fs/btrfs/transaction.c
index 14b9fdc8aaa9..a609222e6704 100644
--- a/fs/btrfs/transaction.c
+++ b/fs/btrfs/transaction.c

@@ -437,7 +437,7 @@ static int record_root_in_trans(struct btrfs_trans_handle *trans,
                                 (unsigned long)root->root_key.objectid,
                                 BTRFS_ROOT_TRANS_TAG);
              spin_unlock(&fs_info->fs_roots_radix_lock);
-             root->last_trans = trans->transid;
+             WRITE_ONCE(root->last_trans, trans->transid);

              /* this is pretty tricky.  We don't want to
               * take the relocation lock in btrfs_record_root_in_trans

@@ -489,7 +489,7 @@ int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
                             struct btrfs_root *root)
 {
      struct btrfs_fs_info *fs_info = root->fs_info;
-     int ret;
+     int ret, last_trans;

      if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
              return 0;

@@ -498,8 +498,9 @@ int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
       * see record_root_in_trans for comments about IN_TRANS_SETUP usage
       * and barriers
       */
+     last_trans = READ_ONCE(root->last_trans);
      smp_rmb();
-     if (root->last_trans == trans->transid &&
+     if (last_trans == trans->transid &&
          !test_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state))

Aren't the smp_rmb barriers supposed to be used before the condition?

No, they need to be between the reads to have an effect. See  https://www.kernel.org/doc/Documentation/memory-barriers.txt §SMP BARRIER PAIRING ("When dealing with CPU-CPU interactions..."). You will see that the barriers are always between the ordered reads and not before.

I think that Paul, the barrier guru, can confirm that the barrier was misplaced in the original code? :)