26126 panic system rather than corrupting pool if we hit bug 26100

References:
  delphix/delphix-os@931c8aaab74b6412933d299890894262e2ef8380

Ported-by: Richard Yao <[email protected]>
Signed-off-by: Brian Behlendorf <[email protected]>
Closes #1650

4 files changed, 26 insertions, 10 deletions

diff --git a/module/zfs/bptree.c b/module/zfs/bptree.c
index a0c90cc4d..c03cb1f84 100644
--- a/module/zfs/bptree.c
+++ b/module/zfs/bptree.c
@@ -20,7 +20,7 @@
  */
 
 /*
- * Copyright (c) 2012 by Delphix. All rights reserved.
+ * Copyright (c) 2013 by Delphix. All rights reserved.
  */
 
 #include <sys/arc.h>
@@ -180,6 +180,7 @@ bptree_iterate(objset_t *os, uint64_t obj, boolean_t free, bptree_itor_t func,
 	err = 0;
 	for (i = ba.ba_phys->bt_begin; i < ba.ba_phys->bt_end; i++) {
 		bptree_entry_phys_t bte;
+		int flags = TRAVERSE_PREFETCH_METADATA | TRAVERSE_POST;
 
 		ASSERT(!free || i == ba.ba_phys->bt_begin);
 
@@ -188,13 +189,13 @@ bptree_iterate(objset_t *os, uint64_t obj, boolean_t free, bptree_itor_t func,
 		if (err != 0)
 			break;
 
+		if (zfs_recover)
+			flags |= TRAVERSE_HARD;
 		err = traverse_dataset_destroyed(os->os_spa, &bte.be_bp,
-		    bte.be_birth_txg, &bte.be_zb,
-		    TRAVERSE_PREFETCH_METADATA | TRAVERSE_POST,
+		    bte.be_birth_txg, &bte.be_zb, flags,
 		    bptree_visit_cb, &ba);
 		if (free) {
-			ASSERT(err == 0 || err == ERESTART);
-			if (err != 0) {
+			if (err == ERESTART) {
 				/* save bookmark for future resume */
 				ASSERT3U(bte.be_zb.zb_objset, ==,
 				    ZB_DESTROYED_OBJSET);
@@ -202,11 +203,21 @@ bptree_iterate(objset_t *os, uint64_t obj, boolean_t free, bptree_itor_t func,
 				dmu_write(os, obj, i * sizeof (bte),
 				    sizeof (bte), &bte, tx);
 				break;
-			} else {
-				ba.ba_phys->bt_begin++;
-				(void) dmu_free_range(os, obj,
-				    i * sizeof (bte), sizeof (bte), tx);
 			}
+			if (err != 0) {
+				/*
+				 * We can not properly handle an i/o
+				 * error, because the traversal code
+				 * does not know how to resume from an
+				 * arbitrary bookmark.
+				 */
+				zfs_panic_recover("error %u from "
+				    "traverse_dataset_destroyed()", err);
+			}
+
+			ba.ba_phys->bt_begin++;
+			(void) dmu_free_range(os, obj,
+			    i * sizeof (bte), sizeof (bte), tx);
 		}
 	}
 
diff --git a/module/zfs/dmu_traverse.c b/module/zfs/dmu_traverse.c
index ce306cbda..939dfe2fa 100644
--- a/module/zfs/dmu_traverse.c
+++ b/module/zfs/dmu_traverse.c
@@ -361,7 +361,7 @@ traverse_visitbp(traverse_data_t *td, const dnode_phys_t *dnp,
 		(void) arc_buf_remove_ref(buf, &buf);
 
 post:
-	if (err == 0 && lasterr == 0 && (td->td_flags & TRAVERSE_POST)) {
+	if (err == 0 && (td->td_flags & TRAVERSE_POST)) {
 		err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
 		if (err == ERESTART)
 			pause = B_TRUE;
diff --git a/module/zfs/dsl_scan.c b/module/zfs/dsl_scan.c
index eef509bda..354d5b1df 100644
--- a/module/zfs/dsl_scan.c
+++ b/module/zfs/dsl_scan.c
@@ -1334,6 +1334,9 @@ dsl_scan_free_should_pause(dsl_scan_t *scn)
 {
 	uint64_t elapsed_nanosecs;
 
+	if (zfs_recover)
+		return (B_FALSE);
+
 	elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
 	return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
 	    (NSEC2MSEC(elapsed_nanosecs) > zfs_free_min_time_ms &&
diff --git a/module/zfs/zfs_debug.c b/module/zfs/zfs_debug.c
index cd83e2392..55a18e839 100644
--- a/module/zfs/zfs_debug.c
+++ b/module/zfs/zfs_debug.c
@@ -41,6 +41,8 @@ int zfs_flags = 0;
  * zfs_recover can be set to nonzero to attempt to recover from
  * otherwise-fatal errors, typically caused by on-disk corruption.  When
  * set, calls to zfs_panic_recover() will turn into warning messages.
+ * This should only be used as a last resort, as it typically results
+ * in leaked space, or worse.
  */
 int zfs_recover = 0;