FreeBSD PR kern/172259: Fixes zfs receive errors

FreeBSD PR kern/172259: Fixes zfs receive errors caused by snapshot
replication being processed in a random order instead of creation
order.

Eliminates needless filesystem renames caused by removed parent
snapshots which subsequently causes many more errors.

PR:		kern/172259
Submitted by:	Steven Hartland
Reviewed by:	pjd (mentor)
Approved by:	pjd (mentor)
MFC after:	2 weeks

References:
  https://github.com/freebsd/freebsd/commit/4995789

Porting notes:

Minor whitespace fixes were made to conform with style requirements:

lib/libzfs/libzfs_sendrecv.c: 2269: indent by spaces instead of tabs
lib/libzfs/libzfs_sendrecv.c: 2270: indent by spaces instead of tabs

Ported-by: Richard Yao <[email protected]>
Signed-off-by: Brian Behlendorf <[email protected]>
Issue #2729

1 files changed, 29 insertions, 2 deletions

diff --git a/lib/libzfs/libzfs_sendrecv.c b/lib/libzfs/libzfs_sendrecv.c
index ebd8a7be2..3b4acc956 100644
--- a/lib/libzfs/libzfs_sendrecv.c
+++ b/lib/libzfs/libzfs_sendrecv.c
@@ -745,7 +745,7 @@ send_iterate_fs(zfs_handle_t *zhp, void *arg)
 	sd->parent_fromsnap_guid = 0;
 	VERIFY(0 == nvlist_alloc(&sd->parent_snaps, NV_UNIQUE_NAME, 0));
 	VERIFY(0 == nvlist_alloc(&sd->snapprops, NV_UNIQUE_NAME, 0));
-	(void) zfs_iter_snapshots(zhp, B_FALSE, send_iterate_snap, sd);
+	(void) zfs_iter_snapshots_sorted(zhp, send_iterate_snap, sd);
 	VERIFY(0 == nvlist_add_nvlist(nvfs, "snaps", sd->parent_snaps));
 	VERIFY(0 == nvlist_add_nvlist(nvfs, "snapprops", sd->snapprops));
 	nvlist_free(sd->parent_snaps);
@@ -2020,11 +2020,12 @@ recv_incremental_replication(libzfs_handle_t *hdl, const char *tofs,
     recvflags_t *flags, nvlist_t *stream_nv, avl_tree_t *stream_avl,
     nvlist_t *renamed)
 {
-	nvlist_t *local_nv;
+	nvlist_t *local_nv, *deleted = NULL;
 	avl_tree_t *local_avl;
 	nvpair_t *fselem, *nextfselem;
 	char *fromsnap;
 	char newname[ZFS_MAXNAMELEN];
+	char guidname[32];
 	int error;
 	boolean_t needagain, progress, recursive;
 	char *s1, *s2;
@@ -2040,6 +2041,8 @@ recv_incremental_replication(libzfs_handle_t *hdl, const char *tofs,
 again:
 	needagain = progress = B_FALSE;
 
+	VERIFY(0 == nvlist_alloc(&deleted, NV_UNIQUE_NAME, 0));
+
 	if ((error = gather_nvlist(hdl, tofs, fromsnap, NULL,
 	    recursive, &local_nv, &local_avl)) != 0)
 		return (error);
@@ -2154,6 +2157,8 @@ again:
 					needagain = B_TRUE;
 				else
 					progress = B_TRUE;
+				sprintf(guidname, "%lu", thisguid);
+				nvlist_add_boolean(deleted, guidname);
 				continue;
 			}
 
@@ -2209,6 +2214,8 @@ again:
 				needagain = B_TRUE;
 			else
 				progress = B_TRUE;
+			sprintf(guidname, "%lu", parent_fromsnap_guid);
+			nvlist_add_boolean(deleted, guidname);
 			continue;
 		}
 
@@ -2231,6 +2238,24 @@ again:
 		s2 = strrchr(stream_fsname, '/');
 
 		/*
+		 * Check if we're going to rename based on parent guid change
+		 * and the current parent guid was also deleted. If it was then
+		 * rename will fail and is likely unneeded, so avoid this and
+		 * force an early retry to determine the new
+		 * parent_fromsnap_guid.
+		 */
+		if (stream_parent_fromsnap_guid != 0 &&
+		    parent_fromsnap_guid != 0 &&
+		    stream_parent_fromsnap_guid != parent_fromsnap_guid) {
+			sprintf(guidname, "%lu", parent_fromsnap_guid);
+			if (nvlist_exists(deleted, guidname)) {
+				progress = B_TRUE;
+				needagain = B_TRUE;
+				goto doagain;
+			}
+		}
+
+		/*
 		 * Check for rename. If the exact receive path is specified, it
 		 * does not count as a rename, but we still need to check the
 		 * datasets beneath it.
@@ -2284,8 +2309,10 @@ again:
 		}
 	}
 
+doagain:
 	fsavl_destroy(local_avl);
 	nvlist_free(local_nv);
+	nvlist_free(deleted);
 
 	if (needagain && progress) {
 		/* do another pass to fix up temporary names */