Fix out-of-order ZIL txtype lost on hardlinked files

We should only call zil_remove_async when an object is removed. However,
in current implementation, it is called whenever TX_REMOVE is called. In
the case of hardlinked file, every unlink will generate TX_REMOVE and
causing operations to be dropped even when the object is not removed.

We fix this by only calling zil_remove_async when the file is fully
unlinked.

Reviewed-by: George Wilson <[email protected]>
Reviewed-by: Brian Behlendorf <[email protected]>
Reviewed-by: Prakash Surya <[email protected]>
Signed-off-by: Chunwei Chen <[email protected]>
Closes #8769
Closes #9061 

3 files changed, 18 insertions, 14 deletions

diff --git a/module/zfs/zfs_log.c b/module/zfs/zfs_log.c
index ad5b5cf30..622ce08ac 100644
--- a/module/zfs/zfs_log.c
+++ b/module/zfs/zfs_log.c
@@ -380,12 +380,14 @@ zfs_log_create(zilog_t *zilog, dmu_tx_t *tx, uint64_t txtype,
 	zil_itx_assign(zilog, itx, tx);
 }
 
+void zil_remove_async(zilog_t *zilog, uint64_t oid);
+
 /*
  * Handles both TX_REMOVE and TX_RMDIR transactions.
  */
 void
 zfs_log_remove(zilog_t *zilog, dmu_tx_t *tx, uint64_t txtype,
-    znode_t *dzp, char *name, uint64_t foid)
+    znode_t *dzp, char *name, uint64_t foid, boolean_t unlinked)
 {
 	itx_t *itx;
 	lr_remove_t *lr;
@@ -401,6 +403,17 @@ zfs_log_remove(zilog_t *zilog, dmu_tx_t *tx, uint64_t txtype,
 
 	itx->itx_oid = foid;
 
+	/*
+	 * Object ids can be re-instantiated in the next txg so
+	 * remove any async transactions to avoid future leaks.
+	 * This can happen if a fsync occurs on the re-instantiated
+	 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
+	 * the new file data and flushes a write record for the old object.
+	 */
+	if (unlinked) {
+		ASSERT((txtype & ~TX_CI) == TX_REMOVE);
+		zil_remove_async(zilog, foid);
+	}
 	zil_itx_assign(zilog, itx, tx);
 }
 
diff --git a/module/zfs/zfs_vnops.c b/module/zfs/zfs_vnops.c
index 595bebc78..1ad6f1588 100644
--- a/module/zfs/zfs_vnops.c
+++ b/module/zfs/zfs_vnops.c
@@ -1886,7 +1886,7 @@ top:
 	txtype = TX_REMOVE;
 	if (flags & FIGNORECASE)
 		txtype |= TX_CI;
-	zfs_log_remove(zilog, tx, txtype, dzp, name, obj);
+	zfs_log_remove(zilog, tx, txtype, dzp, name, obj, unlinked);
 
 	dmu_tx_commit(tx);
 out:
@@ -2219,7 +2219,8 @@ top:
 		uint64_t txtype = TX_RMDIR;
 		if (flags & FIGNORECASE)
 			txtype |= TX_CI;
-		zfs_log_remove(zilog, tx, txtype, dzp, name, ZFS_NO_OBJECT);
+		zfs_log_remove(zilog, tx, txtype, dzp, name, ZFS_NO_OBJECT,
+		    B_FALSE);
 	}
 
 	dmu_tx_commit(tx);
diff --git a/module/zfs/zil.c b/module/zfs/zil.c
index 71079c4a3..98678aa44 100644
--- a/module/zfs/zil.c
+++ b/module/zfs/zil.c
@@ -1875,7 +1875,7 @@ zil_aitx_compare(const void *x1, const void *x2)
 /*
  * Remove all async itx with the given oid.
  */
-static void
+void
 zil_remove_async(zilog_t *zilog, uint64_t oid)
 {
 	uint64_t otxg, txg;
@@ -1928,16 +1928,6 @@ zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
 	itxs_t *itxs, *clean = NULL;
 
 	/*
-	 * Object ids can be re-instantiated in the next txg so
-	 * remove any async transactions to avoid future leaks.
-	 * This can happen if a fsync occurs on the re-instantiated
-	 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
-	 * the new file data and flushes a write record for the old object.
-	 */
-	if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
-		zil_remove_async(zilog, itx->itx_oid);
-
-	/*
 	 * Ensure the data of a renamed file is committed before the rename.
 	 */
 	if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)