Use zap_count instead of cached z_size for unlink

As a performance optimization Lustre does not strictly update
the SA_ZPL_SIZE when adding/removing from non-directory entries.
This results in entries which cannot be removed through the ZPL
layer even though the ZAP is empty and safe to remove.

Resolve this issue by checking the zap_count() directly instead
on relying on the cached SA_ZPL_SIZE.  Micro-benchmarks show no
significant performance impact due to the additional overhead
of using zap_count().

Reviewed-by: Olaf Faaland <[email protected]>
Reviewed-by: Giuseppe Di Natale <[email protected]>
Signed-off-by: Alex Zhuravlev <[email protected]>
Signed-off-by: Brian Behlendorf <[email protected]>
Closes #7019

1 files changed, 15 insertions, 1 deletions

diff --git a/module/zfs/zfs_dir.c b/module/zfs/zfs_dir.c
index c6ee30291..9a8bbccd9 100644
--- a/module/zfs/zfs_dir.c
+++ b/module/zfs/zfs_dir.c
@@ -977,11 +977,25 @@ zfs_link_destroy(zfs_dirlock_t *dl, znode_t *zp, dmu_tx_t *tx, int flag,
  * Indicate whether the directory is empty.  Works with or without z_lock
  * held, but can only be consider a hint in the latter case.  Returns true
  * if only "." and ".." remain and there's no work in progress.
+ *
+ * The internal ZAP size, rather than zp->z_size, needs to be checked since
+ * some consumers (Lustre) do not strictly maintain an accurate SA_ZPL_SIZE.
  */
 boolean_t
 zfs_dirempty(znode_t *dzp)
 {
-	return (dzp->z_size == 2 && dzp->z_dirlocks == 0);
+	zfsvfs_t *zfsvfs = ZTOZSB(dzp);
+	uint64_t count;
+	int error;
+
+	if (dzp->z_dirlocks != NULL)
+		return (B_FALSE);
+
+	error = zap_count(zfsvfs->z_os, dzp->z_id, &count);
+	if (error != 0 || count != 0)
+		return (B_FALSE);
+
+	return (B_TRUE);
 }
 
 int