Revert "Revert "Revert "Fix unlink/xattr deadlock"""

This reverts commit 7973e46 which brings the basic flow of the
code back in line with the other ZFS implementations.  This
was possible due to the following related changes.

e89260a Directory xattr znodes hold a reference on their parent
6f9548c Fix deadlock in zfs_zget()
0a50679 Add zfs_iput_async() interface
4dd1893 Avoid 128K kmem allocations in mzap_upgrade()

Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Richard Yao <ryao@gentoo.org>
Closes openzfs#457
Closes openzfs#2058
Closes openzfs#2128
Closes openzfs#2240

module/zfs/zfs_dir.c

@@ -484,6 +484,57 @@ zfs_unlinked_add(znode_t *zp, dmu_tx_t *tx)
 	    zap_add_int(zsb->z_os, zsb->z_unlinkedobj, zp->z_id, tx));
 }
 
+/*
+ * Clean up any znodes that had no links when we either crashed or
+ * (force) umounted the file system.
+ */
+void
+zfs_unlinked_drain(zfs_sb_t *zsb)
+{
+	zap_cursor_t	zc;
+	zap_attribute_t zap;
+	dmu_object_info_t doi;
+	znode_t		*zp;
+	int		error;
+
+	/*
+	 * Iterate over the contents of the unlinked set.
+	 */
+	for (zap_cursor_init(&zc, zsb->z_os, zsb->z_unlinkedobj);
+	    zap_cursor_retrieve(&zc, &zap) == 0;
+	    zap_cursor_advance(&zc)) {
+
+		/*
+		 * See what kind of object we have in list
+		 */
+
+		error = dmu_object_info(zsb->z_os, zap.za_first_integer, &doi);
+		if (error != 0)
+			continue;
+
+		ASSERT((doi.doi_type == DMU_OT_PLAIN_FILE_CONTENTS) ||
+		    (doi.doi_type == DMU_OT_DIRECTORY_CONTENTS));
+		/*
+		 * We need to re-mark these list entries for deletion,
+		 * so we pull them back into core and set zp->z_unlinked.
+		 */
+		error = zfs_zget(zsb, zap.za_first_integer, &zp);
+
+		/*
+		 * We may pick up znodes that are already marked for deletion.
+		 * This could happen during the purge of an extended attribute
+		 * directory.  All we need to do is skip over them, since they
+		 * are already in the system marked z_unlinked.
+		 */
+		if (error != 0)
+			continue;
+
+		zp->z_unlinked = B_TRUE;
+		iput(ZTOI(zp));
+	}
+	zap_cursor_fini(&zc);
+}
+
 /*
  * Delete the entire contents of a directory.  Return a count
  * of the number of entries that could not be deleted. If we encounter
@@ -517,7 +568,8 @@ zfs_purgedir(znode_t *dzp)
 			continue;
 		}
 
-		ASSERT(S_ISREG(ZTOI(xzp)->i_mode)||S_ISLNK(ZTOI(xzp)->i_mode));
+		ASSERT(S_ISREG(ZTOI(xzp)->i_mode) ||
+		    S_ISLNK(ZTOI(xzp)->i_mode));
 
 		tx = dmu_tx_create(zsb->z_os);
 		dmu_tx_hold_sa(tx, dzp->z_sa_hdl, B_FALSE);
@@ -550,71 +602,6 @@ zfs_purgedir(znode_t *dzp)
 	return (skipped);
 }
 
-/*
- * Clean up any znodes that had no links when we either crashed or
- * (force) umounted the file system.
- */
-void
-zfs_unlinked_drain(zfs_sb_t *zsb)
-{
-	zap_cursor_t	zc;
-	zap_attribute_t zap;
-	dmu_object_info_t doi;
-	znode_t		*zp;
-	int		error;
-
-	/*
-	 * Iterate over the contents of the unlinked set.
-	 */
-	for (zap_cursor_init(&zc, zsb->z_os, zsb->z_unlinkedobj);
-	    zap_cursor_retrieve(&zc, &zap) == 0;
-	    zap_cursor_advance(&zc)) {
-
-		/*
-		 * See what kind of object we have in list
-		 */
-
-		error = dmu_object_info(zsb->z_os, zap.za_first_integer, &doi);
-		if (error != 0)
-			continue;
-
-		ASSERT((doi.doi_type == DMU_OT_PLAIN_FILE_CONTENTS) ||
-		    (doi.doi_type == DMU_OT_DIRECTORY_CONTENTS));
-		/*
-		 * We need to re-mark these list entries for deletion,
-		 * so we pull them back into core and set zp->z_unlinked.
-		 */
-		error = zfs_zget(zsb, zap.za_first_integer, &zp);
-
-		/*
-		 * We may pick up znodes that are already marked for deletion.
-		 * This could happen during the purge of an extended attribute
-		 * directory.  All we need to do is skip over them, since they
-		 * are already in the system marked z_unlinked.
-		 */
-		if (error != 0)
-			continue;
-
-		zp->z_unlinked = B_TRUE;
-
-		/*
-		 * If this is an attribute directory, purge its contents.
-		 */
-		if (S_ISDIR(ZTOI(zp)->i_mode) && (zp->z_pflags & ZFS_XATTR)) {
-			/*
-			 * We don't need to check the return value of
-			 * zfs_purgedir here, because zfs_rmnode will just
-			 * return this xattr directory to the unlinked set
-			 * until all of its xattrs are gone.
-			 */
-			(void) zfs_purgedir(zp);
-		}
-
-		iput(ZTOI(zp));
-	}
-	zap_cursor_fini(&zc);
-}
-
 void
 zfs_rmnode(znode_t *zp)
 {
@@ -624,7 +611,6 @@ zfs_rmnode(znode_t *zp)
 	dmu_tx_t	*tx;
 	uint64_t	acl_obj;
 	uint64_t	xattr_obj;
-	uint64_t	count;
 	int		error;
 
 	ASSERT(zp->z_links == 0);
@@ -634,27 +620,13 @@ zfs_rmnode(znode_t *zp)
 	 * If this is an attribute directory, purge its contents.
 	 */
 	if (S_ISDIR(ZTOI(zp)->i_mode) && (zp->z_pflags & ZFS_XATTR)) {
-		error = zap_count(os, zp->z_id, &count);
-		if (error) {
-			zfs_znode_dmu_fini(zp);
-			return;
-		}
-
-		if (count > 0) {
-			taskq_t *taskq;
-
+		if (zfs_purgedir(zp) != 0) {
 			/*
-			 * There are still directory entries in this xattr
-			 * directory.  Let zfs_unlinked_drain() deal with
-			 * them to avoid deadlocking this process in the
-			 * zfs_purgedir()->zfs_zget()->ilookup() callpath
-			 * on the xattr inode's I_FREEING bit.
+			 * Not enough space to delete some xattrs.
+			 * Leave it in the unlinked set.
 			 */
-			taskq = dsl_pool_iput_taskq(dmu_objset_pool(os));
-			taskq_dispatch(taskq, (task_func_t *)
-			    zfs_unlinked_drain, zsb, TQ_SLEEP);
-
 			zfs_znode_dmu_fini(zp);
+
 			return;
 		}
 	}