Illumos 5960, 5925

5960 zfs recv should prefetch indirect blocks
5925 zfs receive -o origin=
Reviewed by: Prakash Surya <prakash.surya@delphix.com>
Reviewed by: Matthew Ahrens <mahrens@delphix.com>

References:
  https://www.illumos.org/issues/5960
  https://www.illumos.org/issues/5925
  https://github.com/illumos/illumos-gate/commit/a2cdcdd

Porting notes:
- [lib/libzfs/libzfs_sendrecv.c]
  - b8864a2 Fix gcc cast warnings
  - 325f023 Add linux kernel device support
  - 5c3f61e Increase Linux pipe buffer size on 'zfs receive'
- [module/zfs/zfs_vnops.c]
  - 3558fd7 Prototype/structure update for Linux
  - c12e3a5 Restructure zfs_readdir() to fix regressions
- [module/zfs/zvol.c]
  - Function @zvol_map_block() isn't needed in ZoL
  - 9965059 Prefetch start and end of volumes
- [module/zfs/dmu.c]
  - Fixed ISO C90 - mixed declarations and code
  - Function dmu_prefetch() 'int i' is initialized before
    the following code block (c90 vs. c99)
- [module/zfs/dbuf.c]
  - fc5bb51 Fix stack dbuf_hold_impl()
  - 9b67f60 Illumos 4757, 4913
  - 34229a2 Reduce stack usage for recursive traverse_visitbp()
- [module/zfs/dmu_send.c]
  - Fixed ISO C90 - mixed declarations and code
  - b58986e Use large stacks when available
  - 241b541 Illumos 5959 - clean up per-dataset feature count code
  - 77aef6f Use vmem_alloc() for nvlists
  - 00b4602 Add linux kernel memory support

Ported-by: kernelOfTruth kerneloftruth@gmail.com
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>

cmd/zdb/zdb.c

@@ -2489,6 +2489,9 @@ zdb_blkptr_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
 	dmu_object_type_t type;
 	boolean_t is_metadata;
 
+	if (bp == NULL)
+		return (0);
+
 	if (dump_opt['b'] >= 5 && bp->blk_birth > 0) {
 		char blkbuf[BP_SPRINTF_LEN];
 		snprintf_blkptr(blkbuf, sizeof (blkbuf), bp);
@@ -2985,7 +2988,7 @@ zdb_ddt_add_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
 	avl_index_t where;
 	zdb_ddt_entry_t *zdde, zdde_search;
 
-	if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
+	if (bp == NULL || BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
 		return (0);
 
 	if (dump_opt['S'] > 1 && zb->zb_level == ZB_ROOT_LEVEL) {

cmd/zfs/zfs_main.c

@@ -250,7 +250,8 @@ get_usage(zfs_help_t idx)
 	case HELP_RECEIVE:
 		return (gettext("\treceive [-vnFu] <filesystem|volume|"
 		    "snapshot>\n"
-		"\treceive [-vnFu] [-d | -e] <filesystem>\n"));
+		    "\treceive [-vnFu] [-o origin=<snapshot>] [-d | -e] "
+		    "<filesystem>\n"));
 	case HELP_RENAME:
 		return (gettext("\trename [-f] <filesystem|volume|snapshot> "
 		    "<filesystem|volume|snapshot>\n"
@@ -793,7 +794,7 @@ zfs_do_create(int argc, char **argv)
 				nomem();
 			break;
 		case 'o':
-			if (parseprop(props, optarg))
+			if (parseprop(props, optarg) != 0)
 				goto error;
 			break;
 		case 's':
@@ -3622,7 +3623,7 @@ zfs_do_snapshot(int argc, char **argv)
 	while ((c = getopt(argc, argv, "ro:")) != -1) {
 		switch (c) {
 		case 'o':
-			if (parseprop(props, optarg))
+			if (parseprop(props, optarg) != 0)
 				return (1);
 			break;
 		case 'r':
@@ -3881,10 +3882,19 @@ zfs_do_receive(int argc, char **argv)
 {
 	int c, err;
 	recvflags_t flags = { 0 };
+	nvlist_t *props;
+	nvpair_t *nvp = NULL;
+
+	if (nvlist_alloc(&props, NV_UNIQUE_NAME, 0) != 0)
+		nomem();
 
 	/* check options */
-	while ((c = getopt(argc, argv, ":denuvF")) != -1) {
+	while ((c = getopt(argc, argv, ":o:denuvF")) != -1) {
 		switch (c) {
+		case 'o':
+			if (parseprop(props, optarg) != 0)
+				return (1);
+			break;
 		case 'd':
 			flags.isprefix = B_TRUE;
 			break;
@@ -3929,6 +3939,13 @@ zfs_do_receive(int argc, char **argv)
 		usage(B_FALSE);
 	}
 
+	while ((nvp = nvlist_next_nvpair(props, nvp))) {
+		if (strcmp(nvpair_name(nvp), "origin") != 0) {
+			(void) fprintf(stderr, gettext("invalid option"));
+			usage(B_FALSE);
+		}
+	}
+
 	if (isatty(STDIN_FILENO)) {
 		(void) fprintf(stderr,
 		    gettext("Error: Backup stream can not be read "
@@ -3937,7 +3954,7 @@ zfs_do_receive(int argc, char **argv)
 		return (1);
 	}
 
-	err = zfs_receive(g_zfs, argv[0], &flags, STDIN_FILENO, NULL);
+	err = zfs_receive(g_zfs, argv[0], props, &flags, STDIN_FILENO, NULL);
 
 	return (err != 0);
 }

cmd/ztest/ztest.c

@@ -3728,7 +3728,8 @@ ztest_dmu_read_write(ztest_ds_t *zd, uint64_t id)
 	 */
 	n = ztest_random(regions) * stride + ztest_random(width);
 	s = 1 + ztest_random(2 * width - 1);
-	dmu_prefetch(os, bigobj, n * chunksize, s * chunksize);
+	dmu_prefetch(os, bigobj, 0, n * chunksize, s * chunksize,
+	    ZIO_PRIORITY_SYNC_READ);
 
 	/*
 	 * Pick a random index and compute the offsets into packobj and bigobj.
@@ -5930,8 +5931,10 @@ ztest_run(ztest_shared_t *zs)
 	 * Right before closing the pool, kick off a bunch of async I/O;
 	 * spa_close() should wait for it to complete.
 	 */
-	for (object = 1; object < 50; object++)
+	for (object = 1; object < 50; object++) {
-		dmu_prefetch(spa->spa_meta_objset, object, 0, 1ULL << 20);
+		dmu_prefetch(spa->spa_meta_objset, object, 0, 0, 1ULL << 20,
+		    ZIO_PRIORITY_SYNC_READ);
+	}
 
 	/* Verify that at least one commit cb was called in a timely fashion */
 	if (zc_cb_counter >= ZTEST_COMMIT_CB_MIN_REG)

include/libzfs.h

@@ -678,8 +678,8 @@ typedef struct recvflags {
 	boolean_t nomount;
 } recvflags_t;
 
-extern int zfs_receive(libzfs_handle_t *, const char *, recvflags_t *,
+extern int zfs_receive(libzfs_handle_t *, const char *, nvlist_t *,
-    int, avl_tree_t *);
+    recvflags_t *, int, avl_tree_t *);
 
 typedef enum diff_flags {
 	ZFS_DIFF_PARSEABLE = 0x1,

include/sys/Makefile.am

@@ -9,6 +9,7 @@ COMMON_H = \
 	$(top_srcdir)/include/sys/bplist.h \
 	$(top_srcdir)/include/sys/bpobj.h \
 	$(top_srcdir)/include/sys/bptree.h \
+	$(top_srcdir)/include/sys/bqueue.h \
 	$(top_srcdir)/include/sys/dbuf.h \
 	$(top_srcdir)/include/sys/ddt.h \
 	$(top_srcdir)/include/sys/dmu.h \
@@ -96,6 +97,7 @@ COMMON_H = \
 	$(top_srcdir)/include/sys/zio_compress.h \
 	$(top_srcdir)/include/sys/zio.h \
 	$(top_srcdir)/include/sys/zio_impl.h \
+	$(top_srcdir)/include/sys/zio_priority.h \
 	$(top_srcdir)/include/sys/zrlock.h
 
 KERNEL_H = \

include/sys/bqueue.h

@@ -0,0 +1,54 @@
+/*
+ * CDDL HEADER START
+ *
+ * This file and its contents are supplied under the terms of the
+ * Common Development and Distribution License ("CDDL"), version 1.0.
+ * You may only use this file in accordance with the terms of version
+ * 1.0 of the CDDL.
+ *
+ * A full copy of the text of the CDDL should have accompanied this
+ * source.  A copy of the CDDL is also available via the Internet at
+ * http://www.illumos.org/license/CDDL.
+ *
+ * CDDL HEADER END
+ */
+/*
+ * Copyright (c) 2014 by Delphix. All rights reserved.
+ */
+
+#ifndef	_BQUEUE_H
+#define	_BQUEUE_H
+
+#ifdef	__cplusplus
+extern "C" {
+#endif
+
+#include	<sys/zfs_context.h>
+
+typedef struct bqueue {
+	list_t bq_list;
+	kmutex_t bq_lock;
+	kcondvar_t bq_add_cv;
+	kcondvar_t bq_pop_cv;
+	uint64_t bq_size;
+	uint64_t bq_maxsize;
+	size_t bq_node_offset;
+} bqueue_t;
+
+typedef struct bqueue_node {
+	list_node_t bqn_node;
+	uint64_t bqn_size;
+} bqueue_node_t;
+
+
+int bqueue_init(bqueue_t *, uint64_t, size_t);
+void bqueue_destroy(bqueue_t *);
+void bqueue_enqueue(bqueue_t *, void *, uint64_t);
+void *bqueue_dequeue(bqueue_t *);
+boolean_t bqueue_empty(bqueue_t *);
+
+#ifdef	__cplusplus
+}
+#endif
+
+#endif	/* _BQUEUE_H */

include/sys/dbuf.h

@@ -261,8 +261,7 @@ typedef struct dbuf_hash_table {
 	kmutex_t hash_mutexes[DBUF_MUTEXES];
 } dbuf_hash_table_t;
 
-
+uint64_t dbuf_whichblock(struct dnode *di, int64_t level, uint64_t offset);
-uint64_t dbuf_whichblock(struct dnode *di, uint64_t offset);
 
 void dbuf_create_bonus(struct dnode *dn);
 int dbuf_spill_set_blksz(dmu_buf_t *db, uint64_t blksz, dmu_tx_t *tx);
@@ -272,10 +271,12 @@ void dbuf_rm_spill(struct dnode *dn, dmu_tx_t *tx);
 dmu_buf_impl_t *dbuf_hold(struct dnode *dn, uint64_t blkid, void *tag);
 dmu_buf_impl_t *dbuf_hold_level(struct dnode *dn, int level, uint64_t blkid,
     void *tag);
-int dbuf_hold_impl(struct dnode *dn, uint8_t level, uint64_t blkid, int create,
+int dbuf_hold_impl(struct dnode *dn, uint8_t level, uint64_t blkid,
+    boolean_t fail_sparse, boolean_t fail_uncached,
     void *tag, dmu_buf_impl_t **dbp);
 
-void dbuf_prefetch(struct dnode *dn, uint64_t blkid, zio_priority_t prio);
+void dbuf_prefetch(struct dnode *dn, int64_t level, uint64_t blkid,
+    zio_priority_t prio, arc_flags_t aflags);
 
 void dbuf_add_ref(dmu_buf_impl_t *db, void *tag);
 boolean_t dbuf_try_add_ref(dmu_buf_t *db, objset_t *os, uint64_t obj,

include/sys/dmu.h

@@ -44,6 +44,7 @@
 #include <sys/inttypes.h>
 #include <sys/cred.h>
 #include <sys/fs/zfs.h>
+#include <sys/zio_priority.h>
 #include <sys/uio.h>
 
 #ifdef	__cplusplus
@@ -737,8 +738,8 @@ extern int zfs_max_recordsize;
 /*
  * Asynchronously try to read in the data.
  */
-void dmu_prefetch(objset_t *os, uint64_t object, uint64_t offset,
+void dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset,
-    uint64_t len);
+	uint64_t len, enum zio_priority pri);
 
 typedef struct dmu_object_info {
 	/* All sizes are in bytes unless otherwise indicated. */

include/sys/dsl_dataset.h

@@ -21,7 +21,7 @@
 /*
  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  * Copyright (c) 2011, 2015 by Delphix. All rights reserved.
- * Copyright (c) 2013, Joyent, Inc. All rights reserved.
+ * Copyright (c) 2011, 2014 by Delphix. All rights reserved.
  * Copyright (c) 2013 Steven Hartland. All rights reserved.
  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
  */

include/sys/zfs_context.h

@@ -160,8 +160,18 @@ extern int aok;
 
 /*
  * DTrace SDT probes have different signatures in userland than they do in
- * kernel.  If they're being used in kernel code, re-define them out of
+ * the kernel.  If they're being used in kernel code, re-define them out of
  * existence for their counterparts in libzpool.
+ *
+ * Here's an example of how to use the set-error probes in userland:
+ * zfs$target:::set-error /arg0 == EBUSY/ {stack();}
+ *
+ * Here's an example of how to use DTRACE_PROBE probes in userland:
+ * If there is a probe declared as follows:
+ * DTRACE_PROBE2(zfs__probe_name, uint64_t, blkid, dnode_t *, dn);
+ * Then you can use it as follows:
+ * zfs$target:::probe2 /copyinstr(arg0) == "zfs__probe_name"/
+ *     {printf("%u %p\n", arg1, arg2);}
  */
 
 #ifdef DTRACE_PROBE

include/sys/zio.h

@@ -29,6 +29,7 @@
 #ifndef _ZIO_H
 #define	_ZIO_H
 
+#include <sys/zio_priority.h>
 #include <sys/zfs_context.h>
 #include <sys/spa.h>
 #include <sys/txg.h>
@@ -147,17 +148,6 @@ enum zio_compress {
 #define	ZIO_FAILURE_MODE_CONTINUE	1
 #define	ZIO_FAILURE_MODE_PANIC		2
 
-typedef enum zio_priority {
-	ZIO_PRIORITY_SYNC_READ,
-	ZIO_PRIORITY_SYNC_WRITE,	/* ZIL */
-	ZIO_PRIORITY_ASYNC_READ,	/* prefetch */
-	ZIO_PRIORITY_ASYNC_WRITE,	/* spa_sync() */
-	ZIO_PRIORITY_SCRUB,		/* asynchronous scrub/resilver reads */
-	ZIO_PRIORITY_NUM_QUEUEABLE,
-
-	ZIO_PRIORITY_NOW		/* non-queued i/os (e.g. free) */
-} zio_priority_t;
-
 enum zio_flag {
 	/*
 	 * Flags inherited by gang, ddt, and vdev children,
@@ -262,6 +252,7 @@ extern const char *zio_type_name[ZIO_TYPES];
  * Root blocks (objset_phys_t) are object 0, level -1:  <objset, 0, -1, 0>.
  * ZIL blocks are bookmarked <objset, 0, -2, blkid == ZIL sequence number>.
  * dmu_sync()ed ZIL data blocks are bookmarked <objset, object, -2, blkid>.
+ * dnode visit bookmarks are <objset, object id of dnode, -3, 0>.
  *
  * Note: this structure is called a bookmark because its original purpose
  * was to remember where to resume a pool-wide traverse.
@@ -294,6 +285,9 @@ struct zbookmark_phys {
 #define	ZB_ZIL_OBJECT		(0ULL)
 #define	ZB_ZIL_LEVEL		(-2LL)
 
+#define	ZB_DNODE_LEVEL		(-3LL)
+#define	ZB_DNODE_BLKID		(0ULL)
+
 #define	ZB_IS_ZERO(zb)						\
 	((zb)->zb_objset == 0 && (zb)->zb_object == 0 &&	\
 	(zb)->zb_level == 0 && (zb)->zb_blkid == 0)
@@ -599,8 +593,10 @@ extern void zfs_ereport_post_checksum(spa_t *spa, vdev_t *vd,
 extern void spa_handle_ignored_writes(spa_t *spa);
 
 /* zbookmark_phys functions */
-boolean_t zbookmark_is_before(const struct dnode_phys *dnp,
+boolean_t zbookmark_subtree_completed(const struct dnode_phys *dnp,
-    const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2);
+    const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block);
+int zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2,
+    uint8_t ibs2, const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2);
 
 #ifdef	__cplusplus
 }

include/sys/zio_checksum.h

@@ -44,7 +44,7 @@ typedef const struct zio_checksum_info {
 	zio_checksum_func_t *ci_func[2]; /* checksum function per byteorder */
 	int		ci_correctable;	/* number of correctable bits	*/
 	int		ci_eck;		/* uses zio embedded checksum? */
-	int		ci_dedup;	/* strong enough for dedup? */
+	boolean_t	ci_dedup;	/* strong enough for dedup? */
 	char		*ci_name;	/* descriptive name */
 } zio_checksum_info_t;
 

include/sys/zio_priority.h

@@ -0,0 +1,40 @@
+/*
+ * CDDL HEADER START
+ *
+ * This file and its contents are supplied under the terms of the
+ * Common Development and Distribution License ("CDDL"), version 1.0.
+ * You may only use this file in accordance with the terms of version
+ * 1.0 of the CDDL.
+ *
+ * A full copy of the text of the CDDL should have accompanied this
+ * source.  A copy of the CDDL is also available via the Internet at
+ * http://www.illumos.org/license/CDDL.
+ *
+ * CDDL HEADER END
+ */
+/*
+ * Copyright (c) 2014 by Delphix. All rights reserved.
+ */
+#ifndef	_ZIO_PRIORITY_H
+#define	_ZIO_PRIORITY_H
+
+#ifdef	__cplusplus
+extern "C" {
+#endif
+
+typedef enum zio_priority {
+	ZIO_PRIORITY_SYNC_READ,
+	ZIO_PRIORITY_SYNC_WRITE,	/* ZIL */
+	ZIO_PRIORITY_ASYNC_READ,	/* prefetch */
+	ZIO_PRIORITY_ASYNC_WRITE,	/* spa_sync() */
+	ZIO_PRIORITY_SCRUB,		/* asynchronous scrub/resilver reads */
+	ZIO_PRIORITY_NUM_QUEUEABLE,
+
+	ZIO_PRIORITY_NOW		/* non-queued i/os (e.g. free) */
+} zio_priority_t;
+
+#ifdef	__cplusplus
+}
+#endif
+
+#endif	/* _ZIO_PRIORITY_H */

lib/libzfs/libzfs_pool.c

@@ -3529,7 +3529,7 @@ zpool_vdev_name(libzfs_handle_t *hdl, zpool_handle_t *zhp, nvlist_t *nv,
 }
 
 static int
-zbookmark_compare(const void *a, const void *b)
+zbookmark_mem_compare(const void *a, const void *b)
 {
 	return (memcmp(a, b, sizeof (zbookmark_phys_t)));
 }
@@ -3592,7 +3592,7 @@ zpool_get_errlog(zpool_handle_t *zhp, nvlist_t **nverrlistp)
 	    zc.zc_nvlist_dst_size;
 	count -= zc.zc_nvlist_dst_size;
 
-	qsort(zb, count, sizeof (zbookmark_phys_t), zbookmark_compare);
+	qsort(zb, count, sizeof (zbookmark_phys_t), zbookmark_mem_compare);
 
 	verify(nvlist_alloc(nverrlistp, 0, KM_SLEEP) == 0);
 

lib/libzfs/libzfs_sendrecv.c

@@ -63,8 +63,9 @@
 /* in libzfs_dataset.c */
 extern void zfs_setprop_error(libzfs_handle_t *, zfs_prop_t, int, char *);
 
-static int zfs_receive_impl(libzfs_handle_t *, const char *, recvflags_t *,
+static int zfs_receive_impl(libzfs_handle_t *, const char *, const char *,
-    int, const char *, nvlist_t *, avl_tree_t *, char **, int, uint64_t *);
+    recvflags_t *, int, const char *, nvlist_t *, avl_tree_t *, char **, int,
+    uint64_t *);
 
 static const zio_cksum_t zero_cksum = { { 0 } };
 
@@ -2523,7 +2524,7 @@ zfs_receive_package(libzfs_handle_t *hdl, int fd, const char *destname,
 		 * zfs_receive_one() will take care of it (ie,
 		 * recv_skip() and return 0).
 		 */
-		error = zfs_receive_impl(hdl, destname, flags, fd,
+		error = zfs_receive_impl(hdl, destname, NULL, flags, fd,
 		    sendfs, stream_nv, stream_avl, top_zfs, cleanup_fd,
 		    action_handlep);
 		if (error == ENODATA) {
@@ -2656,9 +2657,9 @@ recv_skip(libzfs_handle_t *hdl, int fd, boolean_t byteswap)
  */
 static int
 zfs_receive_one(libzfs_handle_t *hdl, int infd, const char *tosnap,
-    recvflags_t *flags, dmu_replay_record_t *drr,
+    const char *originsnap, recvflags_t *flags, dmu_replay_record_t *drr,
-    dmu_replay_record_t *drr_noswap, const char *sendfs,
+    dmu_replay_record_t *drr_noswap, const char *sendfs, nvlist_t *stream_nv,
-    nvlist_t *stream_nv, avl_tree_t *stream_avl, char **top_zfs, int cleanup_fd,
+    avl_tree_t *stream_avl, char **top_zfs, int cleanup_fd,
     uint64_t *action_handlep)
 {
 	zfs_cmd_t zc = {"\0"};
@@ -2808,10 +2809,15 @@ zfs_receive_one(libzfs_handle_t *hdl, int infd, const char *tosnap,
 		}
 		if (flags->verbose)
 			(void) printf("found clone origin %s\n", zc.zc_string);
+	} else if (originsnap) {
+		(void) strncpy(zc.zc_string, originsnap, ZFS_MAXNAMELEN);
+		if (flags->verbose)
+			(void) printf("using provided clone origin %s\n",
+			    zc.zc_string);
 	}
 
 	stream_wantsnewfs = (drrb->drr_fromguid == 0 ||
-	    (drrb->drr_flags & DRR_FLAG_CLONE));
+	    (drrb->drr_flags & DRR_FLAG_CLONE) || originsnap);
 
 	if (stream_wantsnewfs) {
 		/*
@@ -3189,9 +3195,10 @@ zfs_receive_one(libzfs_handle_t *hdl, int infd, const char *tosnap,
 }
 
 static int
-zfs_receive_impl(libzfs_handle_t *hdl, const char *tosnap, recvflags_t *flags,
+zfs_receive_impl(libzfs_handle_t *hdl, const char *tosnap,
-    int infd, const char *sendfs, nvlist_t *stream_nv, avl_tree_t *stream_avl,
+    const char *originsnap, recvflags_t *flags, int infd, const char *sendfs,
-    char **top_zfs, int cleanup_fd, uint64_t *action_handlep)
+    nvlist_t *stream_nv, avl_tree_t *stream_avl, char **top_zfs, int cleanup_fd,
+    uint64_t *action_handlep)
 {
 	int err;
 	dmu_replay_record_t drr, drr_noswap;
@@ -3210,6 +3217,12 @@ zfs_receive_impl(libzfs_handle_t *hdl, const char *tosnap, recvflags_t *flags,
 		    "(%s) does not exist"), tosnap);
 		return (zfs_error(hdl, EZFS_NOENT, errbuf));
 	}
+	if (originsnap &&
+	    !zfs_dataset_exists(hdl, originsnap, ZFS_TYPE_DATASET)) {
+		zfs_error_aux(hdl, dgettext(TEXT_DOMAIN, "specified origin fs "
+		    "(%s) does not exist"), originsnap);
+		return (zfs_error(hdl, EZFS_NOENT, errbuf));
+	}
 
 	/* read in the BEGIN record */
 	if (0 != (err = recv_read(hdl, infd, &drr, sizeof (drr), B_FALSE,
@@ -3282,14 +3295,14 @@ zfs_receive_impl(libzfs_handle_t *hdl, const char *tosnap, recvflags_t *flags,
 				*cp = '\0';
 			sendfs = nonpackage_sendfs;
 		}
-		return (zfs_receive_one(hdl, infd, tosnap, flags,
+		return (zfs_receive_one(hdl, infd, tosnap, originsnap, flags,
-		    &drr, &drr_noswap, sendfs, stream_nv, stream_avl,
+		    &drr, &drr_noswap, sendfs, stream_nv, stream_avl, top_zfs,
-		    top_zfs, cleanup_fd, action_handlep));
+		    cleanup_fd, action_handlep));
 	} else {
 		assert(DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) ==
 		    DMU_COMPOUNDSTREAM);
-		return (zfs_receive_package(hdl, infd, tosnap, flags,
+		return (zfs_receive_package(hdl, infd, tosnap, flags, &drr,
-		    &drr, &zcksum, top_zfs, cleanup_fd, action_handlep));
+		    &zcksum, top_zfs, cleanup_fd, action_handlep));
 	}
 }
 
@@ -3300,14 +3313,15 @@ zfs_receive_impl(libzfs_handle_t *hdl, const char *tosnap, recvflags_t *flags,
  * (-1 will override -2).
  */
 int
-zfs_receive(libzfs_handle_t *hdl, const char *tosnap, recvflags_t *flags,
+zfs_receive(libzfs_handle_t *hdl, const char *tosnap, nvlist_t *props,
-    int infd, avl_tree_t *stream_avl)
+    recvflags_t *flags, int infd, avl_tree_t *stream_avl)
 {
 	char *top_zfs = NULL;
 	int err;
 	int cleanup_fd;
 	uint64_t action_handle = 0;
 	struct stat sb;
+	char *originsnap = NULL;
 
 	/*
 	 * The only way fstat can fail is if we do not have a valid file
@@ -3350,10 +3364,16 @@ zfs_receive(libzfs_handle_t *hdl, const char *tosnap, recvflags_t *flags,
 	}
 #endif /* __linux__ */
 
+	if (props) {
+		err = nvlist_lookup_string(props, "origin", &originsnap);
+		if (err && err != ENOENT)
+			return (err);
+	}
+
 	cleanup_fd = open(ZFS_DEV, O_RDWR);
 	VERIFY(cleanup_fd >= 0);
 
-	err = zfs_receive_impl(hdl, tosnap, flags, infd, NULL, NULL,
+	err = zfs_receive_impl(hdl, tosnap, originsnap, flags, infd, NULL, NULL,
 	    stream_avl, &top_zfs, cleanup_fd, &action_handle);
 
 	VERIFY(0 == close(cleanup_fd));

lib/libzpool/Makefile.am

@@ -32,6 +32,7 @@ KERNEL_C = \
 	bplist.c \
 	bpobj.c \
 	bptree.c \
+	bqueue.c \
 	dbuf.c \
 	dbuf_stats.c \
 	ddt.c \

man/man8/zfs.8

@@ -184,12 +184,12 @@ zfs \- configures ZFS file systems
 
 .LP
 .nf
-\fBzfs\fR \fBreceive | recv\fR [\fB-vnFu\fR] \fIfilesystem\fR|\fIvolume\fR|\fIsnapshot\fR
+\fBzfs\fR \fBreceive\fR [\fB-vnFu\fR] [\fB-o origin\fR=\fIsnapshot\fR] \fIfilesystem\fR|\fIvolume\fR|\fIsnapshot\fR
 .fi
 
 .LP
 .nf
-\fBzfs\fR \fBreceive | recv\fR [\fB-vnFu\fR] [\fB-d\fR|\fB-e\fR] \fIfilesystem\fR
+\fBzfs\fR \fBreceive\fR [\fB-vnFu\fR] [\fB-d\fR|\fB-e\fR] [\fB-o origin\fR=\fIsnapshot\fR] \fIfilesystem\fR
 .fi
 
 .LP
@@ -2929,11 +2929,11 @@ then the receiving system must have that feature enabled as well. See
 .ne 2
 .mk
 .na
-\fB\fBzfs receive\fR [\fB-vnFu\fR] \fIfilesystem\fR|\fIvolume\fR|\fIsnapshot\fR\fR
+\fB\fBzfs receive\fR [\fB-vnFu\fR] [\fB-o origin\fR=\fIsnapshot\fR] \fIfilesystem\fR|\fIvolume\fR|\fIsnapshot\fR\fR
 .ad
 .br
 .na
-\fB\fBzfs receive\fR [\fB-vnFu\fR] [\fB-d\fR|\fB-e\fR] \fIfilesystem\fR\fR
+\fB\fBzfs receive\fR [\fB-vnFu\fR] [\fB-d\fR|\fB-e\fR] [\fB-o origin\fR=\fIsnapshot\fR] \fIfilesystem\fR\fR
 .ad
 .sp .6
 .RS 4n
@@ -3001,6 +3001,17 @@ Print verbose information about the stream and the time required to perform the
 Do not actually receive the stream. This can be useful in conjunction with the \fB-v\fR option to verify the name the receive operation would use.
 .RE
 
+.sp
+.ne 2
+.mk
+.na
+\fB\fB-o\fR \fBorigin\fR=\fIsnapshot\fR
+.ad
+.sp .6
+.RS 4n
+Forces the stream to be received as a clone of the given snapshot. This is only valid if the stream is an incremental stream whose source is the same as the provided origin.
+.RE
+
 .sp
 .ne 2
 .mk

module/zfs/Makefile.in

@@ -14,6 +14,7 @@ $(MODULE)-objs += bpobj.o
 $(MODULE)-objs += dbuf.o
 $(MODULE)-objs += dbuf_stats.o
 $(MODULE)-objs += bptree.o
+$(MODULE)-objs += bqueue.o
 $(MODULE)-objs += ddt.o
 $(MODULE)-objs += ddt_zap.o
 $(MODULE)-objs += dmu.o

module/zfs/bptree.c

@@ -156,7 +156,7 @@ bptree_visit_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
 	int err;
 	struct bptree_args *ba = arg;
 
-	if (BP_IS_HOLE(bp))
+	if (bp == NULL || BP_IS_HOLE(bp))
 		return (0);
 
 	err = ba->ba_func(ba->ba_arg, bp, ba->ba_tx);

module/zfs/bqueue.c

@@ -0,0 +1,111 @@
+/*
+ * CDDL HEADER START
+ *
+ * This file and its contents are supplied under the terms of the
+ * Common Development and Distribution License ("CDDL"), version 1.0.
+ * You may only use this file in accordance with the terms of version
+ * 1.0 of the CDDL.
+ *
+ * A full copy of the text of the CDDL should have accompanied this
+ * source.  A copy of the CDDL is also available via the Internet at
+ * http://www.illumos.org/license/CDDL.
+ *
+ * CDDL HEADER END
+ */
+/*
+ * Copyright (c) 2014 by Delphix. All rights reserved.
+ */
+
+#include	<sys/bqueue.h>
+#include	<sys/zfs_context.h>
+
+static inline bqueue_node_t *
+obj2node(bqueue_t *q, void *data)
+{
+	return ((bqueue_node_t *)((char *)data + q->bq_node_offset));
+}
+
+/*
+ * Initialize a blocking queue  The maximum capacity of the queue is set to
+ * size.  Types that want to be stored in a bqueue must contain a bqueue_node_t,
+ * and offset should give its offset from the start of the struct.  Return 0 on
+ * success, or -1 on failure.
+ */
+int
+bqueue_init(bqueue_t *q, uint64_t size, size_t node_offset)
+{
+	list_create(&q->bq_list, node_offset + sizeof (bqueue_node_t),
+	    node_offset + offsetof(bqueue_node_t, bqn_node));
+	cv_init(&q->bq_add_cv, NULL, CV_DEFAULT, NULL);
+	cv_init(&q->bq_pop_cv, NULL, CV_DEFAULT, NULL);
+	mutex_init(&q->bq_lock, NULL, MUTEX_DEFAULT, NULL);
+	q->bq_node_offset = node_offset;
+	q->bq_size = 0;
+	q->bq_maxsize = size;
+	return (0);
+}
+
+/*
+ * Destroy a blocking queue.  This function asserts that there are no
+ * elements in the queue, and no one is blocked on the condition
+ * variables.
+ */
+void
+bqueue_destroy(bqueue_t *q)
+{
+	ASSERT0(q->bq_size);
+	cv_destroy(&q->bq_add_cv);
+	cv_destroy(&q->bq_pop_cv);
+	mutex_destroy(&q->bq_lock);
+	list_destroy(&q->bq_list);
+}
+
+/*
+ * Add data to q, consuming size units of capacity.  If there is insufficient
+ * capacity to consume size units, block until capacity exists.  Asserts size is
+ * > 0.
+ */
+void
+bqueue_enqueue(bqueue_t *q, void *data, uint64_t item_size)
+{
+	ASSERT3U(item_size, >, 0);
+	ASSERT3U(item_size, <, q->bq_maxsize);
+	mutex_enter(&q->bq_lock);
+	obj2node(q, data)->bqn_size = item_size;
+	while (q->bq_size + item_size > q->bq_maxsize) {
+		cv_wait(&q->bq_add_cv, &q->bq_lock);
+	}
+	q->bq_size += item_size;
+	list_insert_tail(&q->bq_list, data);
+	cv_signal(&q->bq_pop_cv);
+	mutex_exit(&q->bq_lock);
+}
+/*
+ * Take the first element off of q.  If there are no elements on the queue, wait
+ * until one is put there.  Return the removed element.
+ */
+void *
+bqueue_dequeue(bqueue_t *q)
+{
+	void *ret;
+	uint64_t item_size;
+	mutex_enter(&q->bq_lock);
+	while (q->bq_size == 0) {
+		cv_wait(&q->bq_pop_cv, &q->bq_lock);
+	}
+	ret = list_remove_head(&q->bq_list);
+	item_size = obj2node(q, ret)->bqn_size;
+	q->bq_size -= item_size;
+	mutex_exit(&q->bq_lock);
+	cv_signal(&q->bq_add_cv);
+	return (ret);
+}
+
+/*
+ * Returns true if the space used is 0.
+ */
+boolean_t
+bqueue_empty(bqueue_t *q)
+{
+	return (q->bq_size == 0);
+}

module/zfs/dbuf.c

@@ -51,7 +51,8 @@ struct dbuf_hold_impl_data {
 	dnode_t *dh_dn;
 	uint8_t dh_level;
 	uint64_t dh_blkid;
-	int dh_fail_sparse;
+	boolean_t dh_fail_sparse;
+	boolean_t dh_fail_uncached;
 	void *dh_tag;
 	dmu_buf_impl_t **dh_dbp;
 	/* Local variables */
@@ -65,7 +66,8 @@ struct dbuf_hold_impl_data {
 };
 
 static void __dbuf_hold_impl_init(struct dbuf_hold_impl_data *dh,
-    dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse,
+    dnode_t *dn, uint8_t level, uint64_t blkid, boolean_t fail_sparse,
+	boolean_t fail_uncached,
 	void *tag, dmu_buf_impl_t **dbp, int depth);
 static int __dbuf_hold_impl(struct dbuf_hold_impl_data *dh);
 
@@ -604,11 +606,35 @@ dbuf_loan_arcbuf(dmu_buf_impl_t *db)
 	return (abuf);
 }
 
+/*
+ * Calculate which level n block references the data at the level 0 offset
+ * provided.
+ */
 uint64_t
-dbuf_whichblock(dnode_t *dn, uint64_t offset)
+dbuf_whichblock(dnode_t *dn, int64_t level, uint64_t offset)
 {
-	if (dn->dn_datablkshift) {
+	if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) {
-		return (offset >> dn->dn_datablkshift);
+		/*
+		 * The level n blkid is equal to the level 0 blkid divided by
+		 * the number of level 0s in a level n block.
+		 *
+		 * The level 0 blkid is offset >> datablkshift =
+		 * offset / 2^datablkshift.
+		 *
+		 * The number of level 0s in a level n is the number of block
+		 * pointers in an indirect block, raised to the power of level.
+		 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
+		 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
+		 *
+		 * Thus, the level n blkid is: offset /
+		 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT)))
+		 * = offset / 2^(datablkshift + level *
+		 *   (indblkshift - SPA_BLKPTRSHIFT))
+		 * = offset >> (datablkshift + level *
+		 *   (indblkshift - SPA_BLKPTRSHIFT))
+		 */
+		return (offset >> (dn->dn_datablkshift + level *
+		    (dn->dn_indblkshift - SPA_BLKPTRSHIFT)));
 	} else {
 		ASSERT3U(offset, <, dn->dn_datablksz);
 		return (0);
@@ -1786,6 +1812,12 @@ dbuf_clear(dmu_buf_impl_t *db)
 		dbuf_rele(parent, db);
 }
 
+/*
+ * Note: While bpp will always be updated if the function returns success,
+ * parentp will not be updated if the dnode does not have dn_dbuf filled in;
+ * this happens when the dnode is the meta-dnode, or a userused or groupused
+ * object.
+ */
 __attribute__((always_inline))
 static inline int
 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
@@ -1828,11 +1860,11 @@ dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
 		/* this block is referenced from an indirect block */
 		int err;
 		if (dh == NULL) {
-			err = dbuf_hold_impl(dn, level+1, blkid >> epbs,
+			err = dbuf_hold_impl(dn, level+1,
-					fail_sparse, NULL, parentp);
+			    blkid >> epbs, fail_sparse, FALSE, NULL, parentp);
 		} else {
 			__dbuf_hold_impl_init(dh + 1, dn, dh->dh_level + 1,
-					blkid >> epbs, fail_sparse, NULL,
+			    blkid >> epbs, fail_sparse, FALSE, NULL,
 			    parentp, dh->dh_depth + 1);
 			err = __dbuf_hold_impl(dh + 1);
 		}
@@ -2011,11 +2043,102 @@ dbuf_destroy(dmu_buf_impl_t *db)
 	arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
 }
 
-void
+typedef struct dbuf_prefetch_arg {
-dbuf_prefetch(dnode_t *dn, uint64_t blkid, zio_priority_t prio)
+	spa_t *dpa_spa;	/* The spa to issue the prefetch in. */
+	zbookmark_phys_t dpa_zb; /* The target block to prefetch. */
+	int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */
+	int dpa_curlevel; /* The current level that we're reading */
+	zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */
+	zio_t *dpa_zio; /* The parent zio_t for all prefetches. */
+	arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */
+} dbuf_prefetch_arg_t;
+
+/*
+ * Actually issue the prefetch read for the block given.
+ */
+static void
+dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp)
 {
-	dmu_buf_impl_t *db = NULL;
+	arc_flags_t aflags;
-	blkptr_t *bp = NULL;
+	if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
+		return;
+
+	aflags = dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
+
+	ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
+	ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level);
+	ASSERT(dpa->dpa_zio != NULL);
+	(void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL,
+	    dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
+	    &aflags, &dpa->dpa_zb);
+}
+
+/*
+ * Called when an indirect block above our prefetch target is read in.  This
+ * will either read in the next indirect block down the tree or issue the actual
+ * prefetch if the next block down is our target.
+ */
+static void
+dbuf_prefetch_indirect_done(zio_t *zio, arc_buf_t *abuf, void *private)
+{
+	dbuf_prefetch_arg_t *dpa = private;
+	uint64_t nextblkid;
+	blkptr_t *bp;
+
+	ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel);
+	ASSERT3S(dpa->dpa_curlevel, >, 0);
+	if (zio != NULL) {
+		ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel);
+		ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size);
+		ASSERT3P(zio->io_spa, ==, dpa->dpa_spa);
+	}
+
+	dpa->dpa_curlevel--;
+
+	nextblkid = dpa->dpa_zb.zb_blkid >>
+	    (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level));
+	bp = ((blkptr_t *)abuf->b_data) +
+	    P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs);
+	if (BP_IS_HOLE(bp) || (zio != NULL && zio->io_error != 0)) {
+		kmem_free(dpa, sizeof (*dpa));
+	} else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) {
+		ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid);
+		dbuf_issue_final_prefetch(dpa, bp);
+		kmem_free(dpa, sizeof (*dpa));
+	} else {
+		arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
+		zbookmark_phys_t zb;
+
+		ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
+
+		SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset,
+		    dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid);
+
+		(void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
+		    bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio,
+		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
+		    &iter_aflags, &zb);
+	}
+	(void) arc_buf_remove_ref(abuf, private);
+}
+
+/*
+ * Issue prefetch reads for the given block on the given level.  If the indirect
+ * blocks above that block are not in memory, we will read them in
+ * asynchronously.  As a result, this call never blocks waiting for a read to
+ * complete.
+ */
+void
+dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio,
+    arc_flags_t aflags)
+{
+	blkptr_t bp;
+	int epbs, nlevels, curlevel;
+	uint64_t curblkid;
+	dmu_buf_impl_t *db;
+	zio_t *pio;
+	dbuf_prefetch_arg_t *dpa;
+	dsl_dataset_t *ds;
 
 	ASSERT(blkid != DMU_BONUS_BLKID);
 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
@@ -2023,35 +2146,104 @@ dbuf_prefetch(dnode_t *dn, uint64_t blkid, zio_priority_t prio)
 	if (dnode_block_freed(dn, blkid))
 		return;
 
-	/* dbuf_find() returns with db_mtx held */
-	if ((db = dbuf_find(dn->dn_objset, dn->dn_object, 0, blkid))) {
 	/*
-		 * This dbuf is already in the cache.  We assume that
+	 * This dnode hasn't been written to disk yet, so there's nothing to
-		 * it is already CACHED, or else about to be either
+	 * prefetch.
-		 * read or filled.
 	 */
+	nlevels = dn->dn_phys->dn_nlevels;
+	if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0)
+		return;
+
+	epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
+	if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level))
+		return;
+
+	db = dbuf_find(dn->dn_objset, dn->dn_object,
+	    level, blkid);
+	if (db != NULL) {
 		mutex_exit(&db->db_mtx);
+		/*
+		 * This dbuf already exists.  It is either CACHED, or
+		 * (we assume) about to be read or filled.
+		 */
 		return;
 	}
 
-	if (dbuf_findbp(dn, 0, blkid, TRUE, &db, &bp, NULL) == 0) {
+	/*
-		if (bp && !BP_IS_HOLE(bp) && !BP_IS_EMBEDDED(bp)) {
+	 * Find the closest ancestor (indirect block) of the target block
-			dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
+	 * that is present in the cache.  In this indirect block, we will
-			arc_flags_t aflags =
+	 * find the bp that is at curlevel, curblkid.
-			    ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
+	 */
+	curlevel = level;
+	curblkid = blkid;
+	while (curlevel < nlevels - 1) {
+		int parent_level = curlevel + 1;
+		uint64_t parent_blkid = curblkid >> epbs;
+		dmu_buf_impl_t *db;
+
+		if (dbuf_hold_impl(dn, parent_level, parent_blkid,
+		    FALSE, TRUE, FTAG, &db) == 0) {
+			blkptr_t *bpp = db->db_buf->b_data;
+			bp = bpp[P2PHASE(curblkid, 1 << epbs)];
+			dbuf_rele(db, FTAG);
+			break;
+		}
+
+		curlevel = parent_level;
+		curblkid = parent_blkid;
+	}
+
+	if (curlevel == nlevels - 1) {
+		/* No cached indirect blocks found. */
+		ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr);
+		bp = dn->dn_phys->dn_blkptr[curblkid];
+	}
+	if (BP_IS_HOLE(&bp))
+		return;
+
+	ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp));
+
+	pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL,
+	    ZIO_FLAG_CANFAIL);
+
+	dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP);
+	ds = dn->dn_objset->os_dsl_dataset;
+	SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
+	    dn->dn_object, level, blkid);
+	dpa->dpa_curlevel = curlevel;
+	dpa->dpa_prio = prio;
+	dpa->dpa_aflags = aflags;
+	dpa->dpa_spa = dn->dn_objset->os_spa;
+	dpa->dpa_epbs = epbs;
+	dpa->dpa_zio = pio;
+
+	/*
+	 * If we have the indirect just above us, no need to do the asynchronous
+	 * prefetch chain; we'll just run the last step ourselves.  If we're at
+	 * a higher level, though, we want to issue the prefetches for all the
+	 * indirect blocks asynchronously, so we can go on with whatever we were
+	 * doing.
+	 */
+	if (curlevel == level) {
+		ASSERT3U(curblkid, ==, blkid);
+		dbuf_issue_final_prefetch(dpa, &bp);
+		kmem_free(dpa, sizeof (*dpa));
+	} else {
+		arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
 		zbookmark_phys_t zb;
 
-			SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
+		SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
-			    dn->dn_object, 0, blkid);
+		    dn->dn_object, curlevel, curblkid);
-
+		(void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
-			(void) arc_read(NULL, dn->dn_objset->os_spa,
+		    &bp, dbuf_prefetch_indirect_done, dpa, prio,
-			    bp, NULL, NULL, prio,
 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
-			    &aflags, &zb);
+		    &iter_aflags, &zb);
-		}
-		if (db)
-			dbuf_rele(db, NULL);
 	}
+	/*
+	 * We use pio here instead of dpa_zio since it's possible that
+	 * dpa may have already been freed.
+	 */
+	zio_nowait(pio);
 }
 
 #define	DBUF_HOLD_IMPL_MAX_DEPTH	20
@@ -2079,6 +2271,9 @@ top:
 	if (dh->dh_db == NULL) {
 		dh->dh_bp = NULL;
 
+		if (dh->dh_fail_uncached)
+			return (SET_ERROR(ENOENT));
+
 		ASSERT3P(dh->dh_parent, ==, NULL);
 		dh->dh_err = dbuf_findbp(dh->dh_dn, dh->dh_level, dh->dh_blkid,
 					dh->dh_fail_sparse, &dh->dh_parent,
@@ -2099,6 +2294,11 @@ top:
 					dh->dh_parent, dh->dh_bp);
 	}
 
+	if (dh->dh_fail_uncached && dh->dh_db->db_state != DB_CACHED) {
+		mutex_exit(&dh->dh_db->db_mtx);
+		return (SET_ERROR(ENOENT));
+	}
+
 	if (dh->dh_db->db_buf && refcount_is_zero(&dh->dh_db->db_holds)) {
 		arc_buf_add_ref(dh->dh_db->db_buf, dh->dh_db);
 		if (dh->dh_db->db_buf->b_data == NULL) {
@@ -2159,7 +2359,8 @@ top:
  * on the stack for 20 levels of recursion.
  */
 int
-dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse,
+dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid,
+    boolean_t fail_sparse, boolean_t fail_uncached,
     void *tag, dmu_buf_impl_t **dbp)
 {
 	struct dbuf_hold_impl_data *dh;
@@ -2167,7 +2368,8 @@ dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse,
 
 	dh = kmem_zalloc(sizeof (struct dbuf_hold_impl_data) *
 	    DBUF_HOLD_IMPL_MAX_DEPTH, KM_SLEEP);
-	__dbuf_hold_impl_init(dh, dn, level, blkid, fail_sparse, tag, dbp, 0);
+	__dbuf_hold_impl_init(dh, dn, level, blkid, fail_sparse,
+		fail_uncached, tag, dbp, 0);
 
 	error = __dbuf_hold_impl(dh);
 
@@ -2179,13 +2381,17 @@ dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse,
 
 static void
 __dbuf_hold_impl_init(struct dbuf_hold_impl_data *dh,
-    dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse,
+    dnode_t *dn, uint8_t level, uint64_t blkid,
+	boolean_t fail_sparse, boolean_t fail_uncached,
 	void *tag, dmu_buf_impl_t **dbp, int depth)
 {
 	dh->dh_dn = dn;
 	dh->dh_level = level;
 	dh->dh_blkid = blkid;
+
 	dh->dh_fail_sparse = fail_sparse;
+	dh->dh_fail_uncached = fail_uncached;
+
 	dh->dh_tag = tag;
 	dh->dh_dbp = dbp;
 	dh->dh_depth = depth;
@@ -2194,16 +2400,14 @@ __dbuf_hold_impl_init(struct dbuf_hold_impl_data *dh,
 dmu_buf_impl_t *
 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
 {
-	dmu_buf_impl_t *db;
+	return (dbuf_hold_level(dn, 0, blkid, tag));
-	int err = dbuf_hold_impl(dn, 0, blkid, FALSE, tag, &db);
-	return (err ? NULL : db);
 }
 
 dmu_buf_impl_t *
 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
 {
 	dmu_buf_impl_t *db;
-	int err = dbuf_hold_impl(dn, level, blkid, FALSE, tag, &db);
+	int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db);
 	return (err ? NULL : db);
 }
 
@@ -2531,8 +2735,8 @@ dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
 		if (parent == NULL) {
 			mutex_exit(&db->db_mtx);
 			rw_enter(&dn->dn_struct_rwlock, RW_READER);
-			(void) dbuf_hold_impl(dn, db->db_level+1,
+			parent = dbuf_hold_level(dn, db->db_level + 1,
-			    db->db_blkid >> epbs, FALSE, db, &parent);
+			    db->db_blkid >> epbs, db);
 			rw_exit(&dn->dn_struct_rwlock);
 			mutex_enter(&db->db_mtx);
 			db->db_parent = parent;

module/zfs/dmu.c

@@ -138,7 +138,7 @@ dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset,
 	err = dnode_hold(os, object, FTAG, &dn);
 	if (err)
 		return (err);
-	blkid = dbuf_whichblock(dn, offset);
+	blkid = dbuf_whichblock(dn, 0, offset);
 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
 	db = dbuf_hold(dn, blkid, tag);
 	rw_exit(&dn->dn_struct_rwlock);
@@ -421,7 +421,7 @@ dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
 	dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
 
 	zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
-	blkid = dbuf_whichblock(dn, offset);
+	blkid = dbuf_whichblock(dn, 0, offset);
 	for (i = 0; i < nblks; i++) {
 		dmu_buf_impl_t *db = dbuf_hold(dn, blkid+i, tag);
 		if (db == NULL) {
@@ -522,17 +522,16 @@ dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
 }
 
 /*
- * Issue prefetch i/os for the given blocks.
+ * Issue prefetch i/os for the given blocks.  If level is greater than 0, the
+ * indirect blocks prefeteched will be those that point to the blocks containing
+ * the data starting at offset, and continuing to offset + len.
  *
- * Note: The assumption is that we *know* these blocks will be needed
+ * Note that if the indirect blocks above the blocks being prefetched are not in
- * almost immediately.  Therefore, the prefetch i/os will be issued at
+ * cache, they will be asychronously read in.
- * ZIO_PRIORITY_SYNC_READ
- *
- * Note: indirect blocks and other metadata will be read synchronously,
- * causing this function to block if they are not already cached.
  */
 void
-dmu_prefetch(objset_t *os, uint64_t object, uint64_t offset, uint64_t len)
+dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset,
+    uint64_t len, zio_priority_t pri)
 {
 	dnode_t *dn;
 	uint64_t blkid;
@@ -548,8 +547,9 @@ dmu_prefetch(objset_t *os, uint64_t object, uint64_t offset, uint64_t len)
 			return;
 
 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
-		blkid = dbuf_whichblock(dn, object * sizeof (dnode_phys_t));
+		blkid = dbuf_whichblock(dn, level,
-		dbuf_prefetch(dn, blkid, ZIO_PRIORITY_SYNC_READ);
+		    object * sizeof (dnode_phys_t));
+		dbuf_prefetch(dn, level, blkid, pri, 0);
 		rw_exit(&dn->dn_struct_rwlock);
 		return;
 	}
@@ -564,10 +564,16 @@ dmu_prefetch(objset_t *os, uint64_t object, uint64_t offset, uint64_t len)
 		return;
 
 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
-	if (dn->dn_datablkshift) {
+	/*
-		int blkshift = dn->dn_datablkshift;
+	 * offset + len - 1 is the last byte we want to prefetch for, and offset
-		nblks = (P2ROUNDUP(offset + len, 1 << blkshift) -
+	 * is the first.  Then dbuf_whichblk(dn, level, off + len - 1) is the
-		    P2ALIGN(offset, 1 << blkshift)) >> blkshift;
+	 * last block we want to prefetch, and dbuf_whichblock(dn, level,
+	 * offset)  is the first.  Then the number we need to prefetch is the
+	 * last - first + 1.
+	 */
+	if (level > 0 || dn->dn_datablkshift != 0) {
+		nblks = dbuf_whichblock(dn, level, offset + len - 1) -
+		    dbuf_whichblock(dn, level, offset) + 1;
 	} else {
 		nblks = (offset < dn->dn_datablksz);
 	}
@@ -575,9 +581,9 @@ dmu_prefetch(objset_t *os, uint64_t object, uint64_t offset, uint64_t len)
 	if (nblks != 0) {
 		int i;
 
-		blkid = dbuf_whichblock(dn, offset);
+		blkid = dbuf_whichblock(dn, level, offset);
 		for (i = 0; i < nblks; i++)
-			dbuf_prefetch(dn, blkid + i, ZIO_PRIORITY_SYNC_READ);
+			dbuf_prefetch(dn, level, blkid + i, pri, 0);
 	}
 
 	rw_exit(&dn->dn_struct_rwlock);
@@ -1293,7 +1299,7 @@ dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
 	DB_DNODE_ENTER(dbuf);
 	dn = DB_DNODE(dbuf);
 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
-	blkid = dbuf_whichblock(dn, offset);
+	blkid = dbuf_whichblock(dn, 0, offset);
 	VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
 	rw_exit(&dn->dn_struct_rwlock);
 	DB_DNODE_EXIT(dbuf);

module/zfs/dmu_diff.c

@@ -115,7 +115,7 @@ diff_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
 	if (issig(JUSTLOOKING) && issig(FORREAL))
 		return (SET_ERROR(EINTR));
 
-	if (zb->zb_object != DMU_META_DNODE_OBJECT)
+	if (bp == NULL || zb->zb_object != DMU_META_DNODE_OBJECT)
 		return (0);
 
 	if (BP_IS_HOLE(bp)) {

module/zfs/dmu_object.c

@@ -148,6 +148,11 @@ dmu_object_free(objset_t *os, uint64_t object, dmu_tx_t *tx)
 	return (0);
 }
 
+/*
+ * Return (in *objectp) the next object which is allocated (or a hole)
+ * after *object, taking into account only objects that may have been modified
+ * after the specified txg.
+ */
 int
 dmu_object_next(objset_t *os, uint64_t *objectp, boolean_t hole, uint64_t txg)
 {

module/zfs/dmu_traverse.c

@@ -157,7 +157,7 @@ resume_skip_check(traverse_data_t *td, const dnode_phys_t *dnp,
 		 * If we already visited this bp & everything below,
 		 * don't bother doing it again.
 		 */
-		if (zbookmark_is_before(dnp, zb, td->td_resume))
+		if (zbookmark_subtree_completed(dnp, zb, td->td_resume))
 			return (RESUME_SKIP_ALL);
 
 		/*
@@ -428,6 +428,17 @@ traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp,
 	int j, err = 0;
 	zbookmark_phys_t czb;
 
+	if (td->td_flags & TRAVERSE_PRE) {
+		SET_BOOKMARK(&czb, objset, object, ZB_DNODE_LEVEL,
+		    ZB_DNODE_BLKID);
+		err = td->td_func(td->td_spa, NULL, NULL, &czb, dnp,
+		    td->td_arg);
+		if (err == TRAVERSE_VISIT_NO_CHILDREN)
+			return (0);
+		if (err != 0)
+			return (err);
+	}
+
 	for (j = 0; j < dnp->dn_nblkptr; j++) {
 		SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
 		err = traverse_visitbp(td, dnp, &dnp->dn_blkptr[j], &czb);
@@ -435,10 +446,21 @@ traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp,
 			break;
 	}
 
-	if (err == 0 && dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
+	if (err == 0 && (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
 		SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
 		err = traverse_visitbp(td, dnp, &dnp->dn_spill, &czb);
 	}
+
+	if (err == 0 && (td->td_flags & TRAVERSE_POST)) {
+		SET_BOOKMARK(&czb, objset, object, ZB_DNODE_LEVEL,
+		    ZB_DNODE_BLKID);
+		err = td->td_func(td->td_spa, NULL, NULL, &czb, dnp,
+		    td->td_arg);
+		if (err == TRAVERSE_VISIT_NO_CHILDREN)
+			return (0);
+		if (err != 0)
+			return (err);
+	}
 	return (err);
 }
 
@@ -451,6 +473,8 @@ traverse_prefetcher(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
 	arc_flags_t aflags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
 
 	ASSERT(pfd->pd_bytes_fetched >= 0);
+	if (bp == NULL)
+		return (0);
 	if (pfd->pd_cancel)
 		return (SET_ERROR(EINTR));
 

module/zfs/dmu_tx.c

@@ -332,7 +332,8 @@ dmu_tx_count_write(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
 			dmu_buf_impl_t *db;
 
 			rw_enter(&dn->dn_struct_rwlock, RW_READER);
-			err = dbuf_hold_impl(dn, 0, start, FALSE, FTAG, &db);
+			err = dbuf_hold_impl(dn, 0, start,
+			    FALSE, FALSE, FTAG, &db);
 			rw_exit(&dn->dn_struct_rwlock);
 
 			if (err) {
@@ -533,7 +534,8 @@ dmu_tx_count_free(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
 		blkoff = P2PHASE(blkid, epb);
 		tochk = MIN(epb - blkoff, nblks);
 
-		err = dbuf_hold_impl(dn, 1, blkid >> epbs, FALSE, FTAG, &dbuf);
+		err = dbuf_hold_impl(dn, 1, blkid >> epbs,
+		    FALSE, FALSE, FTAG, &dbuf);
 		if (err) {
 			txh->txh_tx->tx_err = err;
 			break;

module/zfs/dmu_zfetch.c

@@ -293,7 +293,8 @@ dmu_zfetch_fetch(dnode_t *dn, uint64_t blkid, uint64_t nblks)
 	fetchsz = dmu_zfetch_fetchsz(dn, blkid, nblks);
 
 	for (i = 0; i < fetchsz; i++) {
-		dbuf_prefetch(dn, blkid + i, ZIO_PRIORITY_ASYNC_READ);
+		dbuf_prefetch(dn, 0, blkid + i, ZIO_PRIORITY_ASYNC_READ,
+		    ARC_FLAG_PREFETCH);
 	}
 
 	return (fetchsz);

module/zfs/dnode.c

@@ -1112,7 +1112,7 @@ dnode_hold_impl(objset_t *os, uint64_t object, int flag,
 		drop_struct_lock = TRUE;
 	}
 
-	blk = dbuf_whichblock(mdn, object * sizeof (dnode_phys_t));
+	blk = dbuf_whichblock(mdn, 0, object * sizeof (dnode_phys_t));
 
 	db = dbuf_hold(mdn, blk, FTAG);
 	if (drop_struct_lock)
@@ -1409,7 +1409,7 @@ dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
 		goto fail;
 
 	/* resize the old block */
-	err = dbuf_hold_impl(dn, 0, 0, TRUE, FTAG, &db);
+	err = dbuf_hold_impl(dn, 0, 0, TRUE, FALSE, FTAG, &db);
 	if (err == 0)
 		dbuf_new_size(db, size, tx);
 	else if (err != ENOENT)
@@ -1582,8 +1582,8 @@ dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
 		ASSERT3U(blkoff + head, ==, blksz);
 		if (len < head)
 			head = len;
-		if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, off), TRUE,
+		if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off),
-		    FTAG, &db) == 0) {
+		    TRUE, FALSE, FTAG, &db) == 0) {
 			caddr_t data;
 
 			/* don't dirty if it isn't on disk and isn't dirty */
@@ -1620,8 +1620,8 @@ dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
 	if (tail) {
 		if (len < tail)
 			tail = len;
-		if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, off+len),
+		if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off+len),
-		    TRUE, FTAG, &db) == 0) {
+		    TRUE, FALSE, FTAG, &db) == 0) {
 			/* don't dirty if not on disk and not dirty */
 			if (db->db_last_dirty ||
 			    (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
@@ -1875,8 +1875,8 @@ dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
 		epb = dn->dn_phys->dn_nblkptr;
 		data = dn->dn_phys->dn_blkptr;
 	} else {
-		uint64_t blkid = dbuf_whichblock(dn, *offset) >> (epbs * lvl);
+		uint64_t blkid = dbuf_whichblock(dn, lvl, *offset);
-		error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FTAG, &db);
+		error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FALSE, FTAG, &db);
 		if (error) {
 			if (error != ENOENT)
 				return (error);

module/zfs/dnode_sync.c

@@ -192,7 +192,7 @@ free_verify(dmu_buf_impl_t *db, uint64_t start, uint64_t end, dmu_tx_t *tx)
 
 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
 		err = dbuf_hold_impl(dn, db->db_level-1,
-		    (db->db_blkid << epbs) + i, TRUE, FTAG, &child);
+		    (db->db_blkid << epbs) + i, TRUE, FALSE, FTAG, &child);
 		rw_exit(&dn->dn_struct_rwlock);
 		if (err == ENOENT)
 			continue;
@@ -288,7 +288,7 @@ free_children(dmu_buf_impl_t *db, uint64_t blkid, uint64_t nblks,
 				continue;
 			rw_enter(&dn->dn_struct_rwlock, RW_READER);
 			VERIFY0(dbuf_hold_impl(dn, db->db_level - 1,
-			    i, B_TRUE, FTAG, &subdb));
+			    i, TRUE, FALSE, FTAG, &subdb));
 			rw_exit(&dn->dn_struct_rwlock);
 			ASSERT3P(bp, ==, subdb->db_blkptr);
 
@@ -362,7 +362,7 @@ dnode_sync_free_range_impl(dnode_t *dn, uint64_t blkid, uint64_t nblks,
 				continue;
 			rw_enter(&dn->dn_struct_rwlock, RW_READER);
 			VERIFY0(dbuf_hold_impl(dn, dnlevel - 1, i,
-			    TRUE, FTAG, &db));
+			    TRUE, FALSE, FTAG, &db));
 			rw_exit(&dn->dn_struct_rwlock);
 
 			free_children(db, blkid, nblks, tx);

module/zfs/dsl_dataset.c

@@ -547,6 +547,7 @@ dsl_dataset_hold(dsl_pool_t *dp, const char *name,
 	const char *snapname;
 	uint64_t obj;
 	int err = 0;
+	dsl_dataset_t *ds;
 
 	err = dsl_dir_hold(dp, name, FTAG, &dd, &snapname);
 	if (err != 0)
@@ -555,36 +556,37 @@ dsl_dataset_hold(dsl_pool_t *dp, const char *name,
 	ASSERT(dsl_pool_config_held(dp));
 	obj = dsl_dir_phys(dd)->dd_head_dataset_obj;
 	if (obj != 0)
-		err = dsl_dataset_hold_obj(dp, obj, tag, dsp);
+		err = dsl_dataset_hold_obj(dp, obj, tag, &ds);
 	else
 		err = SET_ERROR(ENOENT);
 
 	/* we may be looking for a snapshot */
 	if (err == 0 && snapname != NULL) {
-		dsl_dataset_t *ds;
+		dsl_dataset_t *snap_ds;
 
 		if (*snapname++ != '@') {
-			dsl_dataset_rele(*dsp, tag);
+			dsl_dataset_rele(ds, tag);
 			dsl_dir_rele(dd, FTAG);
 			return (SET_ERROR(ENOENT));
 		}
 
 		dprintf("looking for snapshot '%s'\n", snapname);
-		err = dsl_dataset_snap_lookup(*dsp, snapname, &obj);
+		err = dsl_dataset_snap_lookup(ds, snapname, &obj);
 		if (err == 0)
-			err = dsl_dataset_hold_obj(dp, obj, tag, &ds);
+			err = dsl_dataset_hold_obj(dp, obj, tag, &snap_ds);
-		dsl_dataset_rele(*dsp, tag);
+		dsl_dataset_rele(ds, tag);
 
 		if (err == 0) {
-			mutex_enter(&ds->ds_lock);
+			mutex_enter(&snap_ds->ds_lock);
-			if (ds->ds_snapname[0] == 0)
+			if (snap_ds->ds_snapname[0] == 0)
-				(void) strlcpy(ds->ds_snapname, snapname,
+				(void) strlcpy(snap_ds->ds_snapname, snapname,
-				    sizeof (ds->ds_snapname));
+				    sizeof (snap_ds->ds_snapname));
-			mutex_exit(&ds->ds_lock);
+			mutex_exit(&snap_ds->ds_lock);
+			ds = snap_ds;
+		}
+	}
+	if (err == 0)
 		*dsp = ds;
-		}
-	}
-
 	dsl_dir_rele(dd, FTAG);
 	return (err);
 }

module/zfs/dsl_destroy.c

@@ -560,7 +560,7 @@ kill_blkptr(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
 	struct killarg *ka = arg;
 	dmu_tx_t *tx = ka->tx;
 
-	if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
+	if (bp == NULL || BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
 		return (0);
 
 	if (zb->zb_level == ZB_ZIL_LEVEL) {

module/zfs/dsl_scan.c

@@ -619,7 +619,8 @@ dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
 		 * If we already visited this bp & everything below (in
 		 * a prior txg sync), don't bother doing it again.
 		 */
-		if (zbookmark_is_before(dnp, zb, &scn->scn_phys.scn_bookmark))
+		if (zbookmark_subtree_completed(dnp, zb,
+		    &scn->scn_phys.scn_bookmark))
 			return (B_TRUE);
 
 		/*

module/zfs/spa.c

@@ -1921,7 +1921,7 @@ spa_load_verify_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
 	size_t size;
 	void *data;
 
-	if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
+	if (bp == NULL || BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
 		return (0);
 	/*
 	 * Note: normally this routine will not be called if

module/zfs/space_map.c

@@ -76,8 +76,8 @@ space_map_load(space_map_t *sm, range_tree_t *rt, maptype_t maptype)
 
 	mutex_exit(sm->sm_lock);
 	if (end > bufsize) {
-		dmu_prefetch(sm->sm_os, space_map_object(sm), bufsize,
+		dmu_prefetch(sm->sm_os, space_map_object(sm), 0, bufsize,
-		    end - bufsize);
+		    end - bufsize, ZIO_PRIORITY_SYNC_READ);
 	}
 	mutex_enter(sm->sm_lock);
 

module/zfs/zap.c

@@ -162,8 +162,9 @@ zap_table_grow(zap_t *zap, zap_table_phys_t *tbl,
 		newblk = zap_allocate_blocks(zap, tbl->zt_numblks * 2);
 		tbl->zt_nextblk = newblk;
 		ASSERT0(tbl->zt_blks_copied);
-		dmu_prefetch(zap->zap_objset, zap->zap_object,
+		dmu_prefetch(zap->zap_objset, zap->zap_object, 0,
-		    tbl->zt_blk << bs, tbl->zt_numblks << bs);
+		    tbl->zt_blk << bs, tbl->zt_numblks << bs,
+		    ZIO_PRIORITY_SYNC_READ);
 	}
 
 	/*
@@ -949,7 +950,8 @@ fzap_prefetch(zap_name_t *zn)
 	if (zap_idx_to_blk(zap, idx, &blk) != 0)
 		return;
 	bs = FZAP_BLOCK_SHIFT(zap);
-	dmu_prefetch(zap->zap_objset, zap->zap_object, blk << bs, 1 << bs);
+	dmu_prefetch(zap->zap_objset, zap->zap_object, 0, blk << bs, 1 << bs,
+	    ZIO_PRIORITY_SYNC_READ);
 }
 
 /*
@@ -1295,9 +1297,10 @@ fzap_get_stats(zap_t *zap, zap_stats_t *zs)
 	} else {
 		int b;
 
-		dmu_prefetch(zap->zap_objset, zap->zap_object,
+		dmu_prefetch(zap->zap_objset, zap->zap_object, 0,
 		    zap_f_phys(zap)->zap_ptrtbl.zt_blk << bs,
-		    zap_f_phys(zap)->zap_ptrtbl.zt_numblks << bs);
+		    zap_f_phys(zap)->zap_ptrtbl.zt_numblks << bs,
+		    ZIO_PRIORITY_SYNC_READ);
 
 		for (b = 0; b < zap_f_phys(zap)->zap_ptrtbl.zt_numblks;
 		    b++) {

module/zfs/zfs_vfsops.c

@@ -20,7 +20,7 @@
  */
 /*
  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
- * Copyright (c) 2013 by Delphix. All rights reserved.
+ * Copyright (c) 2012, 2014 by Delphix. All rights reserved.
  */
 
 /* Portions Copyright 2010 Robert Milkowski */

module/zfs/zfs_vnops.c

@@ -2118,7 +2118,8 @@ zfs_readdir(struct inode *ip, struct dir_context *ctx, cred_t *cr)
 
 		/* Prefetch znode */
 		if (prefetch) {
-			dmu_prefetch(os, objnum, 0, 0);
+			dmu_prefetch(os, objnum, 0, 0, 0,
+			    ZIO_PRIORITY_SYNC_READ);
 		}
 
 		/*

module/zfs/zio.c

@@ -63,6 +63,9 @@ int zio_delay_max = ZIO_DELAY_MAX;
 #define	ZIO_PIPELINE_CONTINUE		0x100
 #define	ZIO_PIPELINE_STOP		0x101
 
+#define	BP_SPANB(indblkshift, level) \
+	(((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
+#define	COMPARE_META_LEVEL	0x80000000ul
 /*
  * The following actions directly effect the spa's sync-to-convergence logic.
  * The values below define the sync pass when we start performing the action.
@@ -3450,39 +3453,129 @@ static zio_pipe_stage_t *zio_pipeline[] = {
 	zio_done
 };
 
-/* dnp is the dnode for zb1->zb_object */
-boolean_t
-zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_phys_t *zb1,
-    const zbookmark_phys_t *zb2)
-{
-	uint64_t zb1nextL0, zb2thisobj;
 
-	ASSERT(zb1->zb_objset == zb2->zb_objset);
+
-	ASSERT(zb2->zb_level == 0);
+
+/*
+ * Compare two zbookmark_phys_t's to see which we would reach first in a
+ * pre-order traversal of the object tree.
+ *
+ * This is simple in every case aside from the meta-dnode object. For all other
+ * objects, we traverse them in order (object 1 before object 2, and so on).
+ * However, all of these objects are traversed while traversing object 0, since
+ * the data it points to is the list of objects.  Thus, we need to convert to a
+ * canonical representation so we can compare meta-dnode bookmarks to
+ * non-meta-dnode bookmarks.
+ *
+ * We do this by calculating "equivalents" for each field of the zbookmark.
+ * zbookmarks outside of the meta-dnode use their own object and level, and
+ * calculate the level 0 equivalent (the first L0 blkid that is contained in the
+ * blocks this bookmark refers to) by multiplying their blkid by their span
+ * (the number of L0 blocks contained within one block at their level).
+ * zbookmarks inside the meta-dnode calculate their object equivalent
+ * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
+ * level + 1<<31 (any value larger than a level could ever be) for their level.
+ * This causes them to always compare before a bookmark in their object
+ * equivalent, compare appropriately to bookmarks in other objects, and to
+ * compare appropriately to other bookmarks in the meta-dnode.
+ */
+int
+zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
+    const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
+{
+	/*
+	 * These variables represent the "equivalent" values for the zbookmark,
+	 * after converting zbookmarks inside the meta dnode to their
+	 * normal-object equivalents.
+	 */
+	uint64_t zb1obj, zb2obj;
+	uint64_t zb1L0, zb2L0;
+	uint64_t zb1level, zb2level;
+
+	if (zb1->zb_object == zb2->zb_object &&
+	    zb1->zb_level == zb2->zb_level &&
+	    zb1->zb_blkid == zb2->zb_blkid)
+		return (0);
+
+	/*
+	 * BP_SPANB calculates the span in blocks.
+	 */
+	zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
+	zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
+
+	if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
+		zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
+		zb1L0 = 0;
+		zb1level = zb1->zb_level + COMPARE_META_LEVEL;
+	} else {
+		zb1obj = zb1->zb_object;
+		zb1level = zb1->zb_level;
+	}
+
+	if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
+		zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
+		zb2L0 = 0;
+		zb2level = zb2->zb_level + COMPARE_META_LEVEL;
+	} else {
+		zb2obj = zb2->zb_object;
+		zb2level = zb2->zb_level;
+	}
+
+	/* Now that we have a canonical representation, do the comparison. */
+	if (zb1obj != zb2obj)
+		return (zb1obj < zb2obj ? -1 : 1);
+	else if (zb1L0 != zb2L0)
+		return (zb1L0 < zb2L0 ? -1 : 1);
+	else if (zb1level != zb2level)
+		return (zb1level > zb2level ? -1 : 1);
+	/*
+	 * This can (theoretically) happen if the bookmarks have the same object
+	 * and level, but different blkids, if the block sizes are not the same.
+	 * There is presently no way to change the indirect block sizes
+	 */
+	return (0);
+}
+
+/*
+ *  This function checks the following: given that last_block is the place that
+ *  our traversal stopped last time, does that guarantee that we've visited
+ *  every node under subtree_root?  Therefore, we can't just use the raw output
+ *  of zbookmark_compare.  We have to pass in a modified version of
+ *  subtree_root; by incrementing the block id, and then checking whether
+ *  last_block is before or equal to that, we can tell whether or not having
+ *  visited last_block implies that all of subtree_root's children have been
+ *  visited.
+ */
+boolean_t
+zbookmark_subtree_completed(const dnode_phys_t *dnp,
+    const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
+{
+	zbookmark_phys_t mod_zb = *subtree_root;
+	mod_zb.zb_blkid++;
+	ASSERT(last_block->zb_level == 0);
 
 	/* The objset_phys_t isn't before anything. */
 	if (dnp == NULL)
 		return (B_FALSE);
 
-	zb1nextL0 = (zb1->zb_blkid + 1) <<
+	/*
-	    ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
+	 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
-
+	 * data block size in sectors, because that variable is only used if
-	zb2thisobj = zb2->zb_object ? zb2->zb_object :
+	 * the bookmark refers to a block in the meta-dnode.  Since we don't
-	    zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
+	 * know without examining it what object it refers to, and there's no
-
+	 * harm in passing in this value in other cases, we always pass it in.
-	if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
+	 *
-		uint64_t nextobj = zb1nextL0 *
+	 * We pass in 0 for the indirect block size shift because zb2 must be
-		    (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
+	 * level 0.  The indirect block size is only used to calculate the span
-		return (nextobj <= zb2thisobj);
+	 * of the bookmark, but since the bookmark must be level 0, the span is
-	}
+	 * always 1, so the math works out.
-
+	 *
-	if (zb1->zb_object < zb2thisobj)
+	 * If you make changes to how the zbookmark_compare code works, be sure
-		return (B_TRUE);
+	 * to make sure that this code still works afterwards.
-	if (zb1->zb_object > zb2thisobj)
+	 */
-		return (B_FALSE);
+	return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
-	if (zb2->zb_object == DMU_META_DNODE_OBJECT)
+	    1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,
-		return (B_FALSE);
+	    last_block) <= 0);
-	return (zb1nextL0 <= zb2->zb_blkid);
 }
 
 #if defined(_KERNEL) && defined(HAVE_SPL)

module/zfs/zvol.c

@@ -1397,8 +1397,9 @@ __zvol_create_minor(const char *name, boolean_t ignore_snapdev)
 	 */
 	len = MIN(MAX(zvol_prefetch_bytes, 0), SPA_MAXBLOCKSIZE);
 	if (len > 0) {
-		dmu_prefetch(os, ZVOL_OBJ, 0, len);
+		dmu_prefetch(os, ZVOL_OBJ, 0, 0, len, ZIO_PRIORITY_SYNC_READ);
-		dmu_prefetch(os, ZVOL_OBJ, volsize - len, len);
+		dmu_prefetch(os, ZVOL_OBJ, 0, volsize - len, len,
+			ZIO_PRIORITY_SYNC_READ);
 	}
 
 	zv->zv_objset = NULL;