OpenZFS 4986 - receiving replication stream fails if any snapshot exceeds refquota

Authored by: Dan McDonald <[email protected]>
Reviewed by: John Kennedy <[email protected]>
Reviewed by: Matthew Ahrens <[email protected]>
Approved by: Gordon Ross <[email protected]>
Ported-by: Brian Behlendorf <[email protected]>

OpenZFS-issue: https://www.illumos.org/issues/4986
OpenZFS-commit: https://github.com/openzfs/openzfs/commit/5878fad

2 files changed, 121 insertions, 19 deletions

diff --git a/lib/libzfs/libzfs_sendrecv.c b/lib/libzfs/libzfs_sendrecv.c
index 1bf8bf49a..6231cf0e4 100644
--- a/lib/libzfs/libzfs_sendrecv.c
+++ b/lib/libzfs/libzfs_sendrecv.c
@@ -26,6 +26,7 @@
  * Copyright (c) 2012 Pawel Jakub Dawidek <[email protected]>.
  * All rights reserved
  * Copyright (c) 2013 Steven Hartland. All rights reserved.
+ * Copyright 2015, OmniTI Computer Consulting, Inc. All rights reserved.
  */
 
 #include <assert.h>
@@ -66,7 +67,7 @@ extern void zfs_setprop_error(libzfs_handle_t *, zfs_prop_t, int, char *);
 
 static int zfs_receive_impl(libzfs_handle_t *, const char *, const char *,
     recvflags_t *, int, const char *, nvlist_t *, avl_tree_t *, char **, int,
-    uint64_t *);
+    uint64_t *, const char *);
 static int guid_to_name(libzfs_handle_t *, const char *,
     uint64_t, boolean_t, char *);
 
@@ -2621,6 +2622,7 @@ zfs_receive_package(libzfs_handle_t *hdl, int fd, const char *destname,
 	nvlist_t *stream_nv = NULL;
 	avl_tree_t *stream_avl = NULL;
 	char *fromsnap = NULL;
+	char *sendsnap = NULL;
 	char *cp;
 	char tofs[ZFS_MAXNAMELEN];
 	char sendfs[ZFS_MAXNAMELEN];
@@ -2769,8 +2771,16 @@ zfs_receive_package(libzfs_handle_t *hdl, int fd, const char *destname,
 	 */
 	(void) strlcpy(sendfs, drr->drr_u.drr_begin.drr_toname,
 	    ZFS_MAXNAMELEN);
-	if ((cp = strchr(sendfs, '@')) != NULL)
+	if ((cp = strchr(sendfs, '@')) != NULL) {
 		*cp = '\0';
+		/*
+		 * Find the "sendsnap", the final snapshot in a replication
+		 * stream.  zfs_receive_one() handles certain errors
+		 * differently, depending on if the contained stream is the
+		 * last one or not.
+		 */
+		sendsnap = (cp + 1);
+	}
 
 	/* Finally, receive each contained stream */
 	do {
@@ -2783,7 +2793,7 @@ zfs_receive_package(libzfs_handle_t *hdl, int fd, const char *destname,
 		 */
 		error = zfs_receive_impl(hdl, destname, NULL, flags, fd,
 		    sendfs, stream_nv, stream_avl, top_zfs, cleanup_fd,
-		    action_handlep);
+		    action_handlep, sendsnap);
 		if (error == ENODATA) {
 			error = 0;
 			break;
@@ -2948,7 +2958,7 @@ zfs_receive_one(libzfs_handle_t *hdl, int infd, const char *tosnap,
     const char *originsnap, recvflags_t *flags, dmu_replay_record_t *drr,
     dmu_replay_record_t *drr_noswap, const char *sendfs, nvlist_t *stream_nv,
     avl_tree_t *stream_avl, char **top_zfs, int cleanup_fd,
-    uint64_t *action_handlep)
+    uint64_t *action_handlep, const char *finalsnap)
 {
 	zfs_cmd_t zc = {"\0"};
 	time_t begin_time;
@@ -2965,6 +2975,7 @@ zfs_receive_one(libzfs_handle_t *hdl, int infd, const char *tosnap,
 	nvlist_t *snapprops_nvlist = NULL;
 	zprop_errflags_t prop_errflags;
 	boolean_t recursive;
+	char *snapname = NULL;
 
 	begin_time = time(NULL);
 
@@ -2975,7 +2986,6 @@ zfs_receive_one(libzfs_handle_t *hdl, int infd, const char *tosnap,
 	    ENOENT);
 
 	if (stream_avl != NULL) {
-		char *snapname;
 		nvlist_t *fs = fsavl_find(stream_avl, drrb->drr_toguid,
 		    &snapname);
 		nvlist_t *props;
@@ -3316,7 +3326,21 @@ zfs_receive_one(libzfs_handle_t *hdl, int infd, const char *tosnap,
 			    ZPROP_N_MORE_ERRORS) == 0) {
 				trunc_prop_errs(intval);
 				break;
-			} else {
+			} else if (snapname == NULL || finalsnap == NULL ||
+			    strcmp(finalsnap, snapname) == 0 ||
+			    strcmp(nvpair_name(prop_err),
+			    zfs_prop_to_name(ZFS_PROP_REFQUOTA)) != 0) {
+				/*
+				 * Skip the special case of, for example,
+				 * "refquota", errors on intermediate
+				 * snapshots leading up to a final one.
+				 * That's why we have all of the checks above.
+				 *
+				 * See zfs_ioctl.c's extract_delay_props() for
+				 * a list of props which can fail on
+				 * intermediate snapshots, but shouldn't
+				 * affect the overall receive.
+				 */
 				(void) snprintf(tbuf, sizeof (tbuf),
 				    dgettext(TEXT_DOMAIN,
 				    "cannot receive %s property on %s"),
@@ -3501,7 +3525,7 @@ static int
 zfs_receive_impl(libzfs_handle_t *hdl, const char *tosnap,
     const char *originsnap, recvflags_t *flags, int infd, const char *sendfs,
     nvlist_t *stream_nv, avl_tree_t *stream_avl, char **top_zfs, int cleanup_fd,
-    uint64_t *action_handlep)
+    uint64_t *action_handlep, const char *finalsnap)
 {
 	int err;
 	dmu_replay_record_t drr, drr_noswap;
@@ -3597,10 +3621,11 @@ zfs_receive_impl(libzfs_handle_t *hdl, const char *tosnap,
 			if ((cp = strchr(nonpackage_sendfs, '@')) != NULL)
 				*cp = '\0';
 			sendfs = nonpackage_sendfs;
+			VERIFY(finalsnap == NULL);
 		}
 		return (zfs_receive_one(hdl, infd, tosnap, originsnap, flags,
 		    &drr, &drr_noswap, sendfs, stream_nv, stream_avl, top_zfs,
-		    cleanup_fd, action_handlep));
+		    cleanup_fd, action_handlep, finalsnap));
 	} else {
 		assert(DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) ==
 		    DMU_COMPOUNDSTREAM);
@@ -3678,7 +3703,7 @@ zfs_receive(libzfs_handle_t *hdl, const char *tosnap, nvlist_t *props,
 	VERIFY(cleanup_fd >= 0);
 
 	err = zfs_receive_impl(hdl, tosnap, originsnap, flags, infd, NULL, NULL,
-	    stream_avl, &top_zfs, cleanup_fd, &action_handle);
+	    stream_avl, &top_zfs, cleanup_fd, &action_handle, NULL);
 
 	VERIFY(0 == close(cleanup_fd));
 
diff --git a/module/zfs/zfs_ioctl.c b/module/zfs/zfs_ioctl.c
index 825e83847..dd3ce09e3 100644
--- a/module/zfs/zfs_ioctl.c
+++ b/module/zfs/zfs_ioctl.c
@@ -22,6 +22,7 @@
 /*
  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  * Portions Copyright 2011 Martin Matuska
+ * Copyright 2015, OmniTI Computer Consulting, Inc. All rights reserved.
  * Portions Copyright 2012 Pawel Jakub Dawidek <[email protected]>
  * Copyright (c) 2012, Joyent, Inc. All rights reserved.
  * Copyright 2015 Nexenta Systems, Inc.  All rights reserved.
@@ -3990,6 +3991,56 @@ next:
 	}
 }
 
+/*
+ * Extract properties that cannot be set PRIOR to the receipt of a dataset.
+ * For example, refquota cannot be set until after the receipt of a dataset,
+ * because in replication streams, an older/earlier snapshot may exceed the
+ * refquota.  We want to receive the older/earlier snapshot, but setting
+ * refquota pre-receipt will set the dsl's ACTUAL quota, which will prevent
+ * the older/earlier snapshot from being received (with EDQUOT).
+ *
+ * The ZFS test "zfs_receive_011_pos" demonstrates such a scenario.
+ *
+ * libzfs will need to be judicious handling errors encountered by props
+ * extracted by this function.
+ */
+static nvlist_t *
+extract_delay_props(nvlist_t *props)
+{
+	nvlist_t *delayprops;
+	nvpair_t *nvp, *tmp;
+	static const zfs_prop_t delayable[] = { ZFS_PROP_REFQUOTA, 0 };
+	int i;
+
+	VERIFY(nvlist_alloc(&delayprops, NV_UNIQUE_NAME, KM_SLEEP) == 0);
+
+	for (nvp = nvlist_next_nvpair(props, NULL); nvp != NULL;
+	    nvp = nvlist_next_nvpair(props, nvp)) {
+		/*
+		 * strcmp() is safe because zfs_prop_to_name() always returns
+		 * a bounded string.
+		 */
+		for (i = 0; delayable[i] != 0; i++) {
+			if (strcmp(zfs_prop_to_name(delayable[i]),
+			    nvpair_name(nvp)) == 0) {
+				break;
+			}
+		}
+		if (delayable[i] != 0) {
+			tmp = nvlist_prev_nvpair(props, nvp);
+			VERIFY(nvlist_add_nvpair(delayprops, nvp) == 0);
+			VERIFY(nvlist_remove_nvpair(props, nvp) == 0);
+			nvp = tmp;
+		}
+	}
+
+	if (nvlist_empty(delayprops)) {
+		nvlist_free(delayprops);
+		delayprops = NULL;
+	}
+	return (delayprops);
+}
+
 #ifdef	DEBUG
 static boolean_t zfs_ioc_recv_inject_err;
 #endif
@@ -4026,6 +4077,7 @@ zfs_ioc_recv(zfs_cmd_t *zc)
 	offset_t off;
 	nvlist_t *props = NULL; /* sent properties */
 	nvlist_t *origprops = NULL; /* existing properties */
+	nvlist_t *delayprops = NULL; /* sent properties applied post-receive */
 	char *origin = NULL;
 	char *tosnap;
 	char tofs[ZFS_MAXNAMELEN];
@@ -4106,21 +4158,12 @@ zfs_ioc_recv(zfs_cmd_t *zc)
 		props_error = dsl_prop_set_hasrecvd(tofs);
 
 		if (props_error == 0) {
+			delayprops = extract_delay_props(props);
 			(void) zfs_set_prop_nvlist(tofs, ZPROP_SRC_RECEIVED,
 			    props, errors);
 		}
 	}
 
-	if (zc->zc_nvlist_dst_size != 0 &&
-	    (nvlist_smush(errors, zc->zc_nvlist_dst_size) != 0 ||
-	    put_nvlist(zc, errors) != 0)) {
-		/*
-		 * Caller made zc->zc_nvlist_dst less than the minimum expected
-		 * size or supplied an invalid address.
-		 */
-		props_error = SET_ERROR(EINVAL);
-	}
-
 	off = fp->f_offset;
 	error = dmu_recv_stream(&drc, fp->f_vnode, &off, zc->zc_cleanup_fd,
 	    &zc->zc_action_handle);
@@ -4145,6 +4188,40 @@ zfs_ioc_recv(zfs_cmd_t *zc)
 		} else {
 			error = dmu_recv_end(&drc, NULL);
 		}
+
+		/* Set delayed properties now, after we're done receiving. */
+		if (delayprops != NULL && error == 0) {
+			(void) zfs_set_prop_nvlist(tofs, ZPROP_SRC_RECEIVED,
+			    delayprops, errors);
+		}
+	}
+
+	if (delayprops != NULL) {
+		/*
+		 * Merge delayed props back in with initial props, in case
+		 * we're DEBUG and zfs_ioc_recv_inject_err is set (which means
+		 * we have to make sure clear_received_props() includes
+		 * the delayed properties).
+		 *
+		 * Since zfs_ioc_recv_inject_err is only in DEBUG kernels,
+		 * using ASSERT() will be just like a VERIFY.
+		 */
+		ASSERT(nvlist_merge(props, delayprops, 0) == 0);
+		nvlist_free(delayprops);
+	}
+
+	/*
+	 * Now that all props, initial and delayed, are set, report the prop
+	 * errors to the caller.
+	 */
+	if (zc->zc_nvlist_dst_size != 0 &&
+	    (nvlist_smush(errors, zc->zc_nvlist_dst_size) != 0 ||
+	    put_nvlist(zc, errors) != 0)) {
+		/*
+		 * Caller made zc->zc_nvlist_dst less than the minimum expected
+		 * size or supplied an invalid address.
+		 */
+		props_error = SET_ERROR(EINVAL);
 	}
 
 	zc->zc_cookie = off - fp->f_offset;