Illumos #2703: add mechanism to report ZFS send progress
[zfs.git] / lib / libzfs / libzfs_sendrecv.c
index 5a2e2ae..9dbfb16 100644 (file)
  */
 
 /*
- * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
- * Use is subject to license terms.
+ * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2012 by Delphix. All rights reserved.
+ * Copyright (c) 2012 Pawel Jakub Dawidek <pawel@dawidek.net>.
+ * Copyright (c) 2012, Joyent, Inc. All rights reserved.
+ * All rights reserved
  */
 
 #include <assert.h>
 #include <ctype.h>
 #include <errno.h>
-#include <libdevinfo.h>
 #include <libintl.h>
 #include <stdio.h>
 #include <stdlib.h>
 #include <sys/mntent.h>
 #include <sys/mnttab.h>
 #include <sys/avl.h>
+#include <sys/debug.h>
 #include <stddef.h>
+#include <pthread.h>
+#include <umem.h>
+#include <time.h>
 
 #include <libzfs.h>
 
 #include "zfs_namecheck.h"
 #include "zfs_prop.h"
+#include "zfs_fletcher.h"
 #include "libzfs_impl.h"
+#include <sys/zio_checksum.h>
+#include <sys/ddt.h>
+#include <sys/socket.h>
+
+/* 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 *,
+    int, const char *, nvlist_t *, avl_tree_t *, char **, int, uint64_t *);
+
+static const zio_cksum_t zero_cksum = { { 0 } };
+
+typedef struct dedup_arg {
+       int     inputfd;
+       int     outputfd;
+       libzfs_handle_t  *dedup_hdl;
+} dedup_arg_t;
+
+typedef struct progress_arg {
+       zfs_handle_t *pa_zhp;
+       int pa_fd;
+       boolean_t pa_parsable;
+} progress_arg_t;
+
+typedef struct dataref {
+       uint64_t ref_guid;
+       uint64_t ref_object;
+       uint64_t ref_offset;
+} dataref_t;
+
+typedef struct dedup_entry {
+       struct dedup_entry      *dde_next;
+       zio_cksum_t dde_chksum;
+       uint64_t dde_prop;
+       dataref_t dde_ref;
+} dedup_entry_t;
+
+#define        MAX_DDT_PHYSMEM_PERCENT         20
+#define        SMALLEST_POSSIBLE_MAX_DDT_MB            128
+
+typedef struct dedup_table {
+       dedup_entry_t   **dedup_hash_array;
+       umem_cache_t    *ddecache;
+       uint64_t        max_ddt_size;  /* max dedup table size in bytes */
+       uint64_t        cur_ddt_size;  /* current dedup table size in bytes */
+       uint64_t        ddt_count;
+       int             numhashbits;
+       boolean_t       ddt_full;
+} dedup_table_t;
 
-#include <fletcher.c> /* XXX */
+static int
+high_order_bit(uint64_t n)
+{
+       int count;
+
+       for (count = 0; n != 0; count++)
+               n >>= 1;
+       return (count);
+}
+
+static size_t
+ssread(void *buf, size_t len, FILE *stream)
+{
+       size_t outlen;
+
+       if ((outlen = fread(buf, len, 1, stream)) == 0)
+               return (0);
+
+       return (outlen);
+}
+
+static void
+ddt_hash_append(libzfs_handle_t *hdl, dedup_table_t *ddt, dedup_entry_t **ddepp,
+    zio_cksum_t *cs, uint64_t prop, dataref_t *dr)
+{
+       dedup_entry_t   *dde;
+
+       if (ddt->cur_ddt_size >= ddt->max_ddt_size) {
+               if (ddt->ddt_full == B_FALSE) {
+                       zfs_error_aux(hdl, dgettext(TEXT_DOMAIN,
+                           "Dedup table full.  Deduplication will continue "
+                           "with existing table entries"));
+                       ddt->ddt_full = B_TRUE;
+               }
+               return;
+       }
+
+       if ((dde = umem_cache_alloc(ddt->ddecache, UMEM_DEFAULT))
+           != NULL) {
+               assert(*ddepp == NULL);
+               dde->dde_next = NULL;
+               dde->dde_chksum = *cs;
+               dde->dde_prop = prop;
+               dde->dde_ref = *dr;
+               *ddepp = dde;
+               ddt->cur_ddt_size += sizeof (dedup_entry_t);
+               ddt->ddt_count++;
+       }
+}
+
+/*
+ * Using the specified dedup table, do a lookup for an entry with
+ * the checksum cs.  If found, return the block's reference info
+ * in *dr. Otherwise, insert a new entry in the dedup table, using
+ * the reference information specified by *dr.
+ *
+ * return value:  true - entry was found
+ *               false - entry was not found
+ */
+static boolean_t
+ddt_update(libzfs_handle_t *hdl, dedup_table_t *ddt, zio_cksum_t *cs,
+    uint64_t prop, dataref_t *dr)
+{
+       uint32_t hashcode;
+       dedup_entry_t **ddepp;
+
+       hashcode = BF64_GET(cs->zc_word[0], 0, ddt->numhashbits);
+
+       for (ddepp = &(ddt->dedup_hash_array[hashcode]); *ddepp != NULL;
+           ddepp = &((*ddepp)->dde_next)) {
+               if (ZIO_CHECKSUM_EQUAL(((*ddepp)->dde_chksum), *cs) &&
+                   (*ddepp)->dde_prop == prop) {
+                       *dr = (*ddepp)->dde_ref;
+                       return (B_TRUE);
+               }
+       }
+       ddt_hash_append(hdl, ddt, ddepp, cs, prop, dr);
+       return (B_FALSE);
+}
+
+static int
+cksum_and_write(const void *buf, uint64_t len, zio_cksum_t *zc, int outfd)
+{
+       fletcher_4_incremental_native(buf, len, zc);
+       return (write(outfd, buf, len));
+}
+
+/*
+ * This function is started in a separate thread when the dedup option
+ * has been requested.  The main send thread determines the list of
+ * snapshots to be included in the send stream and makes the ioctl calls
+ * for each one.  But instead of having the ioctl send the output to the
+ * the output fd specified by the caller of zfs_send()), the
+ * ioctl is told to direct the output to a pipe, which is read by the
+ * alternate thread running THIS function.  This function does the
+ * dedup'ing by:
+ *  1. building a dedup table (the DDT)
+ *  2. doing checksums on each data block and inserting a record in the DDT
+ *  3. looking for matching checksums, and
+ *  4.  sending a DRR_WRITE_BYREF record instead of a write record whenever
+ *      a duplicate block is found.
+ * The output of this function then goes to the output fd requested
+ * by the caller of zfs_send().
+ */
+static void *
+cksummer(void *arg)
+{
+       dedup_arg_t *dda = arg;
+       char *buf = malloc(1<<20);
+       dmu_replay_record_t thedrr;
+       dmu_replay_record_t *drr = &thedrr;
+       struct drr_begin *drrb = &thedrr.drr_u.drr_begin;
+       struct drr_end *drre = &thedrr.drr_u.drr_end;
+       struct drr_object *drro = &thedrr.drr_u.drr_object;
+       struct drr_write *drrw = &thedrr.drr_u.drr_write;
+       struct drr_spill *drrs = &thedrr.drr_u.drr_spill;
+       FILE *ofp;
+       int outfd;
+       dmu_replay_record_t wbr_drr = {0};
+       struct drr_write_byref *wbr_drrr = &wbr_drr.drr_u.drr_write_byref;
+       dedup_table_t ddt;
+       zio_cksum_t stream_cksum;
+       uint64_t physmem = sysconf(_SC_PHYS_PAGES) * sysconf(_SC_PAGESIZE);
+       uint64_t numbuckets;
+
+       ddt.max_ddt_size =
+           MAX((physmem * MAX_DDT_PHYSMEM_PERCENT)/100,
+           SMALLEST_POSSIBLE_MAX_DDT_MB<<20);
+
+       numbuckets = ddt.max_ddt_size/(sizeof (dedup_entry_t));
+
+       /*
+        * numbuckets must be a power of 2.  Increase number to
+        * a power of 2 if necessary.
+        */
+       if (!ISP2(numbuckets))
+               numbuckets = 1 << high_order_bit(numbuckets);
+
+       ddt.dedup_hash_array = calloc(numbuckets, sizeof (dedup_entry_t *));
+       ddt.ddecache = umem_cache_create("dde", sizeof (dedup_entry_t), 0,
+           NULL, NULL, NULL, NULL, NULL, 0);
+       ddt.cur_ddt_size = numbuckets * sizeof (dedup_entry_t *);
+       ddt.numhashbits = high_order_bit(numbuckets) - 1;
+       ddt.ddt_full = B_FALSE;
+
+       /* Initialize the write-by-reference block. */
+       wbr_drr.drr_type = DRR_WRITE_BYREF;
+       wbr_drr.drr_payloadlen = 0;
+
+       outfd = dda->outputfd;
+       ofp = fdopen(dda->inputfd, "r");
+       while (ssread(drr, sizeof (dmu_replay_record_t), ofp) != 0) {
+
+               switch (drr->drr_type) {
+               case DRR_BEGIN:
+               {
+                       int     fflags;
+                       ZIO_SET_CHECKSUM(&stream_cksum, 0, 0, 0, 0);
+
+                       /* set the DEDUP feature flag for this stream */
+                       fflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo);
+                       fflags |= (DMU_BACKUP_FEATURE_DEDUP |
+                           DMU_BACKUP_FEATURE_DEDUPPROPS);
+                       DMU_SET_FEATUREFLAGS(drrb->drr_versioninfo, fflags);
+
+                       if (cksum_and_write(drr, sizeof (dmu_replay_record_t),
+                           &stream_cksum, outfd) == -1)
+                               goto out;
+                       if (DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) ==
+                           DMU_COMPOUNDSTREAM && drr->drr_payloadlen != 0) {
+                               int sz = drr->drr_payloadlen;
+
+                               if (sz > 1<<20) {
+                                       free(buf);
+                                       buf = malloc(sz);
+                               }
+                               (void) ssread(buf, sz, ofp);
+                               if (ferror(stdin))
+                                       perror("fread");
+                               if (cksum_and_write(buf, sz, &stream_cksum,
+                                   outfd) == -1)
+                                       goto out;
+                       }
+                       break;
+               }
+
+               case DRR_END:
+               {
+                       /* use the recalculated checksum */
+                       ZIO_SET_CHECKSUM(&drre->drr_checksum,
+                           stream_cksum.zc_word[0], stream_cksum.zc_word[1],
+                           stream_cksum.zc_word[2], stream_cksum.zc_word[3]);
+                       if ((write(outfd, drr,
+                           sizeof (dmu_replay_record_t))) == -1)
+                               goto out;
+                       break;
+               }
+
+               case DRR_OBJECT:
+               {
+                       if (cksum_and_write(drr, sizeof (dmu_replay_record_t),
+                           &stream_cksum, outfd) == -1)
+                               goto out;
+                       if (drro->drr_bonuslen > 0) {
+                               (void) ssread(buf,
+                                   P2ROUNDUP((uint64_t)drro->drr_bonuslen, 8),
+                                   ofp);
+                               if (cksum_and_write(buf,
+                                   P2ROUNDUP((uint64_t)drro->drr_bonuslen, 8),
+                                   &stream_cksum, outfd) == -1)
+                                       goto out;
+                       }
+                       break;
+               }
 
-static int zfs_receive_impl(libzfs_handle_t *, const char *, recvflags_t,
-    int, avl_tree_t *, char **);
+               case DRR_SPILL:
+               {
+                       if (cksum_and_write(drr, sizeof (dmu_replay_record_t),
+                           &stream_cksum, outfd) == -1)
+                               goto out;
+                       (void) ssread(buf, drrs->drr_length, ofp);
+                       if (cksum_and_write(buf, drrs->drr_length,
+                           &stream_cksum, outfd) == -1)
+                               goto out;
+                       break;
+               }
+
+               case DRR_FREEOBJECTS:
+               {
+                       if (cksum_and_write(drr, sizeof (dmu_replay_record_t),
+                           &stream_cksum, outfd) == -1)
+                               goto out;
+                       break;
+               }
+
+               case DRR_WRITE:
+               {
+                       dataref_t       dataref;
+
+                       (void) ssread(buf, drrw->drr_length, ofp);
+
+                       /*
+                        * Use the existing checksum if it's dedup-capable,
+                        * else calculate a SHA256 checksum for it.
+                        */
+
+                       if (ZIO_CHECKSUM_EQUAL(drrw->drr_key.ddk_cksum,
+                           zero_cksum) ||
+                           !DRR_IS_DEDUP_CAPABLE(drrw->drr_checksumflags)) {
+                               zio_cksum_t tmpsha256;
+
+                               zio_checksum_SHA256(buf,
+                                   drrw->drr_length, &tmpsha256);
+
+                               drrw->drr_key.ddk_cksum.zc_word[0] =
+                                   BE_64(tmpsha256.zc_word[0]);
+                               drrw->drr_key.ddk_cksum.zc_word[1] =
+                                   BE_64(tmpsha256.zc_word[1]);
+                               drrw->drr_key.ddk_cksum.zc_word[2] =
+                                   BE_64(tmpsha256.zc_word[2]);
+                               drrw->drr_key.ddk_cksum.zc_word[3] =
+                                   BE_64(tmpsha256.zc_word[3]);
+                               drrw->drr_checksumtype = ZIO_CHECKSUM_SHA256;
+                               drrw->drr_checksumflags = DRR_CHECKSUM_DEDUP;
+                       }
+
+                       dataref.ref_guid = drrw->drr_toguid;
+                       dataref.ref_object = drrw->drr_object;
+                       dataref.ref_offset = drrw->drr_offset;
+
+                       if (ddt_update(dda->dedup_hdl, &ddt,
+                           &drrw->drr_key.ddk_cksum, drrw->drr_key.ddk_prop,
+                           &dataref)) {
+                               /* block already present in stream */
+                               wbr_drrr->drr_object = drrw->drr_object;
+                               wbr_drrr->drr_offset = drrw->drr_offset;
+                               wbr_drrr->drr_length = drrw->drr_length;
+                               wbr_drrr->drr_toguid = drrw->drr_toguid;
+                               wbr_drrr->drr_refguid = dataref.ref_guid;
+                               wbr_drrr->drr_refobject =
+                                   dataref.ref_object;
+                               wbr_drrr->drr_refoffset =
+                                   dataref.ref_offset;
+
+                               wbr_drrr->drr_checksumtype =
+                                   drrw->drr_checksumtype;
+                               wbr_drrr->drr_checksumflags =
+                                   drrw->drr_checksumtype;
+                               wbr_drrr->drr_key.ddk_cksum =
+                                   drrw->drr_key.ddk_cksum;
+                               wbr_drrr->drr_key.ddk_prop =
+                                   drrw->drr_key.ddk_prop;
+
+                               if (cksum_and_write(&wbr_drr,
+                                   sizeof (dmu_replay_record_t), &stream_cksum,
+                                   outfd) == -1)
+                                       goto out;
+                       } else {
+                               /* block not previously seen */
+                               if (cksum_and_write(drr,
+                                   sizeof (dmu_replay_record_t), &stream_cksum,
+                                   outfd) == -1)
+                                       goto out;
+                               if (cksum_and_write(buf,
+                                   drrw->drr_length,
+                                   &stream_cksum, outfd) == -1)
+                                       goto out;
+                       }
+                       break;
+               }
+
+               case DRR_FREE:
+               {
+                       if (cksum_and_write(drr, sizeof (dmu_replay_record_t),
+                           &stream_cksum, outfd) == -1)
+                               goto out;
+                       break;
+               }
+
+               default:
+                       (void) printf("INVALID record type 0x%x\n",
+                           drr->drr_type);
+                       /* should never happen, so assert */
+                       assert(B_FALSE);
+               }
+       }
+out:
+       umem_cache_destroy(ddt.ddecache);
+       free(ddt.dedup_hash_array);
+       free(buf);
+       (void) fclose(ofp);
+
+       return (NULL);
+}
 
 /*
  * Routines for dealing with the AVL tree of fs-nvlists
@@ -113,6 +499,9 @@ fsavl_destroy(avl_tree_t *avl)
        free(avl);
 }
 
+/*
+ * Given an nvlist, produce an avl tree of snapshots, ordered by guid
+ */
 static avl_tree_t *
 fsavl_create(nvlist_t *fss)
 {
@@ -170,6 +559,7 @@ typedef struct send_data {
        nvlist_t *snapprops;
        const char *fromsnap;
        const char *tosnap;
+       boolean_t recursive;
 
        /*
         * The header nvlist is of the following format:
@@ -237,23 +627,50 @@ send_iterate_prop(zfs_handle_t *zhp, nvlist_t *nv)
                zfs_prop_t prop = zfs_name_to_prop(propname);
                nvlist_t *propnv;
 
-               if (!zfs_prop_user(propname) && zfs_prop_readonly(prop))
-                       continue;
+               if (!zfs_prop_user(propname)) {
+                       /*
+                        * Realistically, this should never happen.  However,
+                        * we want the ability to add DSL properties without
+                        * needing to make incompatible version changes.  We
+                        * need to ignore unknown properties to allow older
+                        * software to still send datasets containing these
+                        * properties, with the unknown properties elided.
+                        */
+                       if (prop == ZPROP_INVAL)
+                               continue;
+
+                       if (zfs_prop_readonly(prop))
+                               continue;
+               }
 
                verify(nvpair_value_nvlist(elem, &propnv) == 0);
-               if (prop == ZFS_PROP_QUOTA || prop == ZFS_PROP_RESERVATION) {
-                       /* these guys are modifyable, but have no source */
+               if (prop == ZFS_PROP_QUOTA || prop == ZFS_PROP_RESERVATION ||
+                   prop == ZFS_PROP_REFQUOTA ||
+                   prop == ZFS_PROP_REFRESERVATION) {
+                       char *source;
                        uint64_t value;
                        verify(nvlist_lookup_uint64(propnv,
                            ZPROP_VALUE, &value) == 0);
                        if (zhp->zfs_type == ZFS_TYPE_SNAPSHOT)
                                continue;
+                       /*
+                        * May have no source before SPA_VERSION_RECVD_PROPS,
+                        * but is still modifiable.
+                        */
+                       if (nvlist_lookup_string(propnv,
+                           ZPROP_SOURCE, &source) == 0) {
+                               if ((strcmp(source, zhp->zfs_name) != 0) &&
+                                   (strcmp(source,
+                                   ZPROP_SOURCE_VAL_RECVD) != 0))
+                                       continue;
+                       }
                } else {
                        char *source;
                        if (nvlist_lookup_string(propnv,
                            ZPROP_SOURCE, &source) != 0)
                                continue;
-                       if (strcmp(source, zhp->zfs_name) != 0)
+                       if ((strcmp(source, zhp->zfs_name) != 0) &&
+                           (strcmp(source, ZPROP_SOURCE_VAL_RECVD) != 0))
                                continue;
                }
 
@@ -272,12 +689,17 @@ send_iterate_prop(zfs_handle_t *zhp, nvlist_t *nv)
        }
 }
 
+/*
+ * recursively generate nvlists describing datasets.  See comment
+ * for the data structure send_data_t above for description of contents
+ * of the nvlist.
+ */
 static int
 send_iterate_fs(zfs_handle_t *zhp, void *arg)
 {
        send_data_t *sd = arg;
        nvlist_t *nvfs, *nv;
-       int rv;
+       int rv = 0;
        uint64_t parent_fromsnap_guid_save = sd->parent_fromsnap_guid;
        uint64_t guid = zhp->zfs_dmustats.dds_guid;
        char guidstring[64];
@@ -306,7 +728,7 @@ 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, send_iterate_snap, sd);
+       (void) zfs_iter_snapshots(zhp, B_FALSE, 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);
@@ -319,7 +741,8 @@ send_iterate_fs(zfs_handle_t *zhp, void *arg)
        nvlist_free(nvfs);
 
        /* iterate over children */
-       rv = zfs_iter_filesystems(zhp, send_iterate_fs, sd);
+       if (sd->recursive)
+               rv = zfs_iter_filesystems(zhp, send_iterate_fs, sd);
 
        sd->parent_fromsnap_guid = parent_fromsnap_guid_save;
 
@@ -329,7 +752,7 @@ send_iterate_fs(zfs_handle_t *zhp, void *arg)
 
 static int
 gather_nvlist(libzfs_handle_t *hdl, const char *fsname, const char *fromsnap,
-    const char *tosnap, nvlist_t **nvlp, avl_tree_t **avlp)
+    const char *tosnap, boolean_t recursive, nvlist_t **nvlp, avl_tree_t **avlp)
 {
        zfs_handle_t *zhp;
        send_data_t sd = { 0 };
@@ -342,6 +765,7 @@ gather_nvlist(libzfs_handle_t *hdl, const char *fsname, const char *fromsnap,
        VERIFY(0 == nvlist_alloc(&sd.fss, NV_UNIQUE_NAME, 0));
        sd.fromsnap = fromsnap;
        sd.tosnap = tosnap;
+       sd.recursive = recursive;
 
        if ((error = send_iterate_fs(zhp, &sd)) != 0) {
                nvlist_free(sd.fss);
@@ -362,114 +786,128 @@ gather_nvlist(libzfs_handle_t *hdl, const char *fsname, const char *fromsnap,
 }
 
 /*
- * Routines for dealing with the sorted snapshot functionality
+ * Routines specific to "zfs send"
  */
-typedef struct zfs_node {
-       zfs_handle_t    *zn_handle;
-       avl_node_t      zn_avlnode;
-} zfs_node_t;
-
-static int
-zfs_sort_snaps(zfs_handle_t *zhp, void *data)
-{
-       avl_tree_t *avl = data;
-       zfs_node_t *node = zfs_alloc(zhp->zfs_hdl, sizeof (zfs_node_t));
-
-       node->zn_handle = zhp;
-       avl_add(avl, node);
-       return (0);
-}
+typedef struct send_dump_data {
+       /* these are all just the short snapname (the part after the @) */
+       const char *fromsnap;
+       const char *tosnap;
+       char prevsnap[ZFS_MAXNAMELEN];
+       uint64_t prevsnap_obj;
+       boolean_t seenfrom, seento, replicate, doall, fromorigin;
+       boolean_t verbose, dryrun, parsable, progress;
+       int outfd;
+       boolean_t err;
+       nvlist_t *fss;
+       avl_tree_t *fsavl;
+       snapfilter_cb_t *filter_cb;
+       void *filter_cb_arg;
+       nvlist_t *debugnv;
+       char holdtag[ZFS_MAXNAMELEN];
+       int cleanup_fd;
+       uint64_t size;
+} send_dump_data_t;
 
-/* ARGSUSED */
 static int
-zfs_snapshot_compare(const void *larg, const void *rarg)
+estimate_ioctl(zfs_handle_t *zhp, uint64_t fromsnap_obj,
+    boolean_t fromorigin, uint64_t *sizep)
 {
-       zfs_handle_t *l = ((zfs_node_t *)larg)->zn_handle;
-       zfs_handle_t *r = ((zfs_node_t *)rarg)->zn_handle;
-       uint64_t lcreate, rcreate;
+       zfs_cmd_t zc = { "\0", "\0", "\0", "\0", 0 };
+       libzfs_handle_t *hdl = zhp->zfs_hdl;
 
-       /*
-        * Sort them according to creation time.  We use the hidden
-        * CREATETXG property to get an absolute ordering of snapshots.
-        */
-       lcreate = zfs_prop_get_int(l, ZFS_PROP_CREATETXG);
-       rcreate = zfs_prop_get_int(r, ZFS_PROP_CREATETXG);
+       assert(zhp->zfs_type == ZFS_TYPE_SNAPSHOT);
+       assert(fromsnap_obj == 0 || !fromorigin);
 
-       if (lcreate < rcreate)
-               return (-1);
-       else if (lcreate > rcreate)
-               return (+1);
-       else
-               return (0);
-}
+       (void) strlcpy(zc.zc_name, zhp->zfs_name, sizeof (zc.zc_name));
+       zc.zc_obj = fromorigin;
+       zc.zc_sendobj = zfs_prop_get_int(zhp, ZFS_PROP_OBJSETID);
+       zc.zc_fromobj = fromsnap_obj;
+       zc.zc_guid = 1;  /* estimate flag */
 
-static int
-zfs_iter_snapshots_sorted(zfs_handle_t *zhp, zfs_iter_f callback, void *data)
-{
-       int ret = 0;
-       zfs_node_t *node;
-       avl_tree_t avl;
-       void *cookie = NULL;
+       if (zfs_ioctl(zhp->zfs_hdl, ZFS_IOC_SEND, &zc) != 0) {
+               char errbuf[1024];
+               (void) snprintf(errbuf, sizeof (errbuf), dgettext(TEXT_DOMAIN,
+                   "warning: cannot estimate space for '%s'"), zhp->zfs_name);
 
-       avl_create(&avl, zfs_snapshot_compare,
-           sizeof (zfs_node_t), offsetof(zfs_node_t, zn_avlnode));
+               switch (errno) {
+               case EXDEV:
+                       zfs_error_aux(hdl, dgettext(TEXT_DOMAIN,
+                           "not an earlier snapshot from the same fs"));
+                       return (zfs_error(hdl, EZFS_CROSSTARGET, errbuf));
 
-       ret = zfs_iter_snapshots(zhp, zfs_sort_snaps, &avl);
+               case ENOENT:
+                       if (zfs_dataset_exists(hdl, zc.zc_name,
+                           ZFS_TYPE_SNAPSHOT)) {
+                               zfs_error_aux(hdl, dgettext(TEXT_DOMAIN,
+                                   "incremental source (@%s) does not exist"),
+                                   zc.zc_value);
+                       }
+                       return (zfs_error(hdl, EZFS_NOENT, errbuf));
 
-       for (node = avl_first(&avl); node != NULL; node = AVL_NEXT(&avl, node))
-               ret |= callback(node->zn_handle, data);
+               case EDQUOT:
+               case EFBIG:
+               case EIO:
+               case ENOLINK:
+               case ENOSPC:
+               case ENOSTR:
+               case ENXIO:
+               case EPIPE:
+               case ERANGE:
+               case EFAULT:
+               case EROFS:
+                       zfs_error_aux(hdl, strerror(errno));
+                       return (zfs_error(hdl, EZFS_BADBACKUP, errbuf));
 
-       while ((node = avl_destroy_nodes(&avl, &cookie)) != NULL)
-               free(node);
+               default:
+                       return (zfs_standard_error(hdl, errno, errbuf));
+               }
+       }
 
-       avl_destroy(&avl);
+       *sizep = zc.zc_objset_type;
 
-       return (ret);
+       return (0);
 }
 
 /*
- * Routines specific to "zfs send"
- */
-typedef struct send_dump_data {
-       /* these are all just the short snapname (the part after the @) */
-       const char *fromsnap;
-       const char *tosnap;
-       char lastsnap[ZFS_MAXNAMELEN];
-       boolean_t seenfrom, seento, replicate, doall, fromorigin;
-       boolean_t verbose;
-       int outfd;
-       boolean_t err;
-       nvlist_t *fss;
-       avl_tree_t *fsavl;
-} send_dump_data_t;
-
-/*
  * Dumps a backup of the given snapshot (incremental from fromsnap if it's not
  * NULL) to the file descriptor specified by outfd.
  */
 static int
-dump_ioctl(zfs_handle_t *zhp, const char *fromsnap, boolean_t fromorigin,
-    int outfd)
+dump_ioctl(zfs_handle_t *zhp, const char *fromsnap, uint64_t fromsnap_obj,
+    boolean_t fromorigin, int outfd, nvlist_t *debugnv)
 {
-       zfs_cmd_t zc = { 0 };
+       zfs_cmd_t zc = { "\0", "\0", "\0", "\0", 0 };
        libzfs_handle_t *hdl = zhp->zfs_hdl;
+       nvlist_t *thisdbg;
 
        assert(zhp->zfs_type == ZFS_TYPE_SNAPSHOT);
-       assert(fromsnap == NULL || fromsnap[0] == '\0' || !fromorigin);
+       assert(fromsnap_obj == 0 || !fromorigin);
 
        (void) strlcpy(zc.zc_name, zhp->zfs_name, sizeof (zc.zc_name));
-       if (fromsnap)
-               (void) strlcpy(zc.zc_value, fromsnap, sizeof (zc.zc_value));
        zc.zc_cookie = outfd;
        zc.zc_obj = fromorigin;
+       zc.zc_sendobj = zfs_prop_get_int(zhp, ZFS_PROP_OBJSETID);
+       zc.zc_fromobj = fromsnap_obj;
 
-       if (ioctl(zhp->zfs_hdl->libzfs_fd, ZFS_IOC_SEND, &zc) != 0) {
+       VERIFY(0 == nvlist_alloc(&thisdbg, NV_UNIQUE_NAME, 0));
+       if (fromsnap && fromsnap[0] != '\0') {
+               VERIFY(0 == nvlist_add_string(thisdbg,
+                   "fromsnap", fromsnap));
+       }
+
+       if (zfs_ioctl(zhp->zfs_hdl, ZFS_IOC_SEND, &zc) != 0) {
                char errbuf[1024];
                (void) snprintf(errbuf, sizeof (errbuf), dgettext(TEXT_DOMAIN,
                    "warning: cannot send '%s'"), zhp->zfs_name);
 
-               switch (errno) {
+               VERIFY(0 == nvlist_add_uint64(thisdbg, "error", errno));
+               if (debugnv) {
+                       VERIFY(0 == nvlist_add_nvlist(debugnv,
+                           zhp->zfs_name, thisdbg));
+               }
+               nvlist_free(thisdbg);
 
+               switch (errno) {
                case EXDEV:
                        zfs_error_aux(hdl, dgettext(TEXT_DOMAIN,
                            "not an earlier snapshot from the same fs"));
@@ -503,24 +941,127 @@ dump_ioctl(zfs_handle_t *zhp, const char *fromsnap, boolean_t fromorigin,
                }
        }
 
+       if (debugnv)
+               VERIFY(0 == nvlist_add_nvlist(debugnv, zhp->zfs_name, thisdbg));
+       nvlist_free(thisdbg);
+
        return (0);
 }
 
 static int
+hold_for_send(zfs_handle_t *zhp, send_dump_data_t *sdd)
+{
+       zfs_handle_t *pzhp;
+       int error = 0;
+       char *thissnap;
+
+       assert(zhp->zfs_type == ZFS_TYPE_SNAPSHOT);
+
+       if (sdd->dryrun)
+               return (0);
+
+       /*
+        * zfs_send() only opens a cleanup_fd for sends that need it,
+        * e.g. replication and doall.
+        */
+       if (sdd->cleanup_fd == -1)
+               return (0);
+
+       thissnap = strchr(zhp->zfs_name, '@') + 1;
+       *(thissnap - 1) = '\0';
+       pzhp = zfs_open(zhp->zfs_hdl, zhp->zfs_name, ZFS_TYPE_DATASET);
+       *(thissnap - 1) = '@';
+
+       /*
+        * It's OK if the parent no longer exists.  The send code will
+        * handle that error.
+        */
+       if (pzhp) {
+               error = zfs_hold(pzhp, thissnap, sdd->holdtag,
+                   B_FALSE, B_TRUE, B_TRUE, sdd->cleanup_fd,
+                   zfs_prop_get_int(zhp, ZFS_PROP_OBJSETID),
+                   zfs_prop_get_int(zhp, ZFS_PROP_CREATETXG));
+               zfs_close(pzhp);
+       }
+
+       return (error);
+}
+
+static void *
+send_progress_thread(void *arg)
+{
+       progress_arg_t *pa = arg;
+
+       zfs_cmd_t zc = { "\0", "\0", "\0", "\0", 0 };
+       zfs_handle_t *zhp = pa->pa_zhp;
+       libzfs_handle_t *hdl = zhp->zfs_hdl;
+       unsigned long long bytes;
+       char buf[16];
+
+       time_t t;
+       struct tm *tm;
+
+       assert(zhp->zfs_type == ZFS_TYPE_SNAPSHOT);
+       (void) strlcpy(zc.zc_name, zhp->zfs_name, sizeof (zc.zc_name));
+
+       if (!pa->pa_parsable)
+               (void) fprintf(stderr, "TIME        SENT   SNAPSHOT\n");
+
+       /*
+        * Print the progress from ZFS_IOC_SEND_PROGRESS every second.
+        */
+       for (;;) {
+               (void) sleep(1);
+
+               zc.zc_cookie = pa->pa_fd;
+               if (zfs_ioctl(hdl, ZFS_IOC_SEND_PROGRESS, &zc) != 0)
+                       return ((void *)-1);
+
+               (void) time(&t);
+               tm = localtime(&t);
+               bytes = zc.zc_cookie;
+
+               if (pa->pa_parsable) {
+                       (void) fprintf(stderr, "%02d:%02d:%02d\t%llu\t%s\n",
+                           tm->tm_hour, tm->tm_min, tm->tm_sec,
+                           bytes, zhp->zfs_name);
+               } else {
+                       zfs_nicenum(bytes, buf, sizeof (buf));
+                       (void) fprintf(stderr, "%02d:%02d:%02d   %5s   %s\n",
+                           tm->tm_hour, tm->tm_min, tm->tm_sec,
+                           buf, zhp->zfs_name);
+               }
+       }
+}
+
+static int
 dump_snapshot(zfs_handle_t *zhp, void *arg)
 {
        send_dump_data_t *sdd = arg;
-       const char *thissnap;
+       progress_arg_t pa = { 0 };
+       pthread_t tid;
+
+       char *thissnap;
        int err;
+       boolean_t isfromsnap, istosnap, fromorigin;
+       boolean_t exclude = B_FALSE;
 
        thissnap = strchr(zhp->zfs_name, '@') + 1;
+       isfromsnap = (sdd->fromsnap != NULL &&
+           strcmp(sdd->fromsnap, thissnap) == 0);
 
-       if (sdd->fromsnap && !sdd->seenfrom &&
-           strcmp(sdd->fromsnap, thissnap) == 0) {
-               sdd->seenfrom = B_TRUE;
-               (void) strcpy(sdd->lastsnap, thissnap);
+       if (!sdd->seenfrom && isfromsnap) {
+               err = hold_for_send(zhp, sdd);
+               if (err == 0) {
+                       sdd->seenfrom = B_TRUE;
+                       (void) strcpy(sdd->prevsnap, thissnap);
+                       sdd->prevsnap_obj = zfs_prop_get_int(zhp,
+                           ZFS_PROP_OBJSETID);
+               } else if (err == ENOENT) {
+                       err = 0;
+               }
                zfs_close(zhp);
-               return (0);
+               return (err);
        }
 
        if (sdd->seento || !sdd->seenfrom) {
@@ -528,20 +1069,122 @@ dump_snapshot(zfs_handle_t *zhp, void *arg)
                return (0);
        }
 
-       /* send it */
+       istosnap = (strcmp(sdd->tosnap, thissnap) == 0);
+       if (istosnap)
+               sdd->seento = B_TRUE;
+
+       if (!sdd->doall && !isfromsnap && !istosnap) {
+               if (sdd->replicate) {
+                       char *snapname;
+                       nvlist_t *snapprops;
+                       /*
+                        * Filter out all intermediate snapshots except origin
+                        * snapshots needed to replicate clones.
+                        */
+                       nvlist_t *nvfs = fsavl_find(sdd->fsavl,
+                           zhp->zfs_dmustats.dds_guid, &snapname);
+
+                       VERIFY(0 == nvlist_lookup_nvlist(nvfs,
+                           "snapprops", &snapprops));
+                       VERIFY(0 == nvlist_lookup_nvlist(snapprops,
+                           thissnap, &snapprops));
+                       exclude = !nvlist_exists(snapprops, "is_clone_origin");
+               } else {
+                       exclude = B_TRUE;
+               }
+       }
+
+       /*
+        * If a filter function exists, call it to determine whether
+        * this snapshot will be sent.
+        */
+       if (exclude || (sdd->filter_cb != NULL &&
+           sdd->filter_cb(zhp, sdd->filter_cb_arg) == B_FALSE)) {
+               /*
+                * This snapshot is filtered out.  Don't send it, and don't
+                * set prevsnap_obj, so it will be as if this snapshot didn't
+                * exist, and the next accepted snapshot will be sent as
+                * an incremental from the last accepted one, or as the
+                * first (and full) snapshot in the case of a replication,
+                * non-incremental send.
+                */
+               zfs_close(zhp);
+               return (0);
+       }
+
+       err = hold_for_send(zhp, sdd);
+       if (err) {
+               if (err == ENOENT)
+                       err = 0;
+               zfs_close(zhp);
+               return (err);
+       }
+
+       fromorigin = sdd->prevsnap[0] == '\0' &&
+           (sdd->fromorigin || sdd->replicate);
+
        if (sdd->verbose) {
-               (void) fprintf(stderr, "sending from @%s to %s\n",
-                   sdd->lastsnap, zhp->zfs_name);
+               uint64_t size;
+               err = estimate_ioctl(zhp, sdd->prevsnap_obj,
+                   fromorigin, &size);
+
+               if (sdd->parsable) {
+                       if (sdd->prevsnap[0] != '\0') {
+                               (void) fprintf(stderr, "incremental\t%s\t%s",
+                                   sdd->prevsnap, zhp->zfs_name);
+                       } else {
+                               (void) fprintf(stderr, "full\t%s",
+                                   zhp->zfs_name);
+                       }
+               } else {
+                       (void) fprintf(stderr, dgettext(TEXT_DOMAIN,
+                           "send from @%s to %s"),
+                           sdd->prevsnap, zhp->zfs_name);
+               }
+               if (err == 0) {
+                       if (sdd->parsable) {
+                               (void) fprintf(stderr, "\t%llu\n",
+                                   (longlong_t)size);
+                       } else {
+                               char buf[16];
+                               zfs_nicenum(size, buf, sizeof (buf));
+                               (void) fprintf(stderr, dgettext(TEXT_DOMAIN,
+                                   " estimated size is %s\n"), buf);
+                       }
+                       sdd->size += size;
+               } else {
+                       (void) fprintf(stderr, "\n");
+               }
        }
 
-       err = dump_ioctl(zhp, sdd->lastsnap,
-           sdd->lastsnap[0] == '\0' && (sdd->fromorigin || sdd->replicate),
-           sdd->outfd);
+       if (!sdd->dryrun) {
+               /*
+                * If progress reporting is requested, spawn a new thread to
+                * poll ZFS_IOC_SEND_PROGRESS at a regular interval.
+                */
+               if (sdd->progress) {
+                       pa.pa_zhp = zhp;
+                       pa.pa_fd = sdd->outfd;
+                       pa.pa_parsable = sdd->parsable;
+
+                       if ((err = pthread_create(&tid, NULL,
+                           send_progress_thread, &pa))) {
+                               zfs_close(zhp);
+                               return (err);
+                       }
+               }
 
-       if (!sdd->seento && strcmp(sdd->tosnap, thissnap) == 0)
-               sdd->seento = B_TRUE;
+               err = dump_ioctl(zhp, sdd->prevsnap, sdd->prevsnap_obj,
+                   fromorigin, sdd->outfd, sdd->debugnv);
+
+               if (sdd->progress) {
+                       (void) pthread_cancel(tid);
+                       (void) pthread_join(tid, NULL);
+               }
+       }
 
-       (void) strcpy(sdd->lastsnap, thissnap);
+       (void) strcpy(sdd->prevsnap, thissnap);
+       sdd->prevsnap_obj = zfs_prop_get_int(zhp, ZFS_PROP_OBJSETID);
        zfs_close(zhp);
        return (err);
 }
@@ -552,13 +1195,13 @@ dump_filesystem(zfs_handle_t *zhp, void *arg)
        int rv = 0;
        send_dump_data_t *sdd = arg;
        boolean_t missingfrom = B_FALSE;
-       zfs_cmd_t zc = { 0 };
+       zfs_cmd_t zc = { "\0", "\0", "\0", "\0", 0 };
 
        (void) snprintf(zc.zc_name, sizeof (zc.zc_name), "%s@%s",
            zhp->zfs_name, sdd->tosnap);
        if (ioctl(zhp->zfs_hdl->libzfs_fd, ZFS_IOC_OBJSET_STATS, &zc) != 0) {
-               (void) fprintf(stderr, "WARNING: "
-                   "could not send %s@%s: does not exist\n",
+               (void) fprintf(stderr, dgettext(TEXT_DOMAIN,
+                   "WARNING: could not send %s@%s: does not exist\n"),
                    zhp->zfs_name, sdd->tosnap);
                sdd->err = B_TRUE;
                return (0);
@@ -580,45 +1223,34 @@ dump_filesystem(zfs_handle_t *zhp, void *arg)
                }
        }
 
-       if (sdd->doall) {
-               sdd->seenfrom = sdd->seento = sdd->lastsnap[0] = 0;
-               if (sdd->fromsnap == NULL || missingfrom)
-                       sdd->seenfrom = B_TRUE;
+       sdd->seenfrom = sdd->seento = sdd->prevsnap[0] = 0;
+       sdd->prevsnap_obj = 0;
+       if (sdd->fromsnap == NULL || missingfrom)
+               sdd->seenfrom = B_TRUE;
 
-               rv = zfs_iter_snapshots_sorted(zhp, dump_snapshot, arg);
-               if (!sdd->seenfrom) {
-                       (void) fprintf(stderr,
-                           "WARNING: could not send %s@%s:\n"
-                           "incremental source (%s@%s) does not exist\n",
-                           zhp->zfs_name, sdd->tosnap,
-                           zhp->zfs_name, sdd->fromsnap);
-                       sdd->err = B_TRUE;
-               } else if (!sdd->seento) {
-                       (void) fprintf(stderr,
+       rv = zfs_iter_snapshots_sorted(zhp, dump_snapshot, arg);
+       if (!sdd->seenfrom) {
+               (void) fprintf(stderr, dgettext(TEXT_DOMAIN,
+                   "WARNING: could not send %s@%s:\n"
+                   "incremental source (%s@%s) does not exist\n"),
+                   zhp->zfs_name, sdd->tosnap,
+                   zhp->zfs_name, sdd->fromsnap);
+               sdd->err = B_TRUE;
+       } else if (!sdd->seento) {
+               if (sdd->fromsnap) {
+                       (void) fprintf(stderr, dgettext(TEXT_DOMAIN,
                            "WARNING: could not send %s@%s:\n"
                            "incremental source (%s@%s) "
-                           "is not earlier than it\n",
+                           "is not earlier than it\n"),
                            zhp->zfs_name, sdd->tosnap,
                            zhp->zfs_name, sdd->fromsnap);
-                       sdd->err = B_TRUE;
-               }
-       } else {
-               zfs_handle_t *snapzhp;
-               char snapname[ZFS_MAXNAMELEN];
-
-               (void) snprintf(snapname, sizeof (snapname), "%s@%s",
-                   zfs_get_name(zhp), sdd->tosnap);
-               snapzhp = zfs_open(zhp->zfs_hdl, snapname, ZFS_TYPE_SNAPSHOT);
-               if (snapzhp == NULL) {
-                       rv = -1;
                } else {
-                       rv = dump_ioctl(snapzhp,
-                           missingfrom ? NULL : sdd->fromsnap,
-                           sdd->fromorigin || missingfrom,
-                           sdd->outfd);
-                       sdd->seento = B_TRUE;
-                       zfs_close(snapzhp);
+                       (void) fprintf(stderr, dgettext(TEXT_DOMAIN,
+                           "WARNING: "
+                           "could not send %s@%s: does not exist\n"),
+                           zhp->zfs_name, sdd->tosnap);
                }
+               sdd->err = B_TRUE;
        }
 
        return (rv);
@@ -634,16 +1266,39 @@ dump_filesystems(zfs_handle_t *rzhp, void *arg)
        if (!sdd->replicate)
                return (dump_filesystem(rzhp, sdd));
 
+       /* Mark the clone origin snapshots. */
+       for (fspair = nvlist_next_nvpair(sdd->fss, NULL); fspair;
+           fspair = nvlist_next_nvpair(sdd->fss, fspair)) {
+               nvlist_t *nvfs;
+               uint64_t origin_guid = 0;
+
+               VERIFY(0 == nvpair_value_nvlist(fspair, &nvfs));
+               (void) nvlist_lookup_uint64(nvfs, "origin", &origin_guid);
+               if (origin_guid != 0) {
+                       char *snapname;
+                       nvlist_t *origin_nv = fsavl_find(sdd->fsavl,
+                           origin_guid, &snapname);
+                       if (origin_nv != NULL) {
+                               nvlist_t *snapprops;
+                               VERIFY(0 == nvlist_lookup_nvlist(origin_nv,
+                                   "snapprops", &snapprops));
+                               VERIFY(0 == nvlist_lookup_nvlist(snapprops,
+                                   snapname, &snapprops));
+                               VERIFY(0 == nvlist_add_boolean(
+                                   snapprops, "is_clone_origin"));
+                       }
+               }
+       }
 again:
        needagain = progress = B_FALSE;
        for (fspair = nvlist_next_nvpair(sdd->fss, NULL); fspair;
            fspair = nvlist_next_nvpair(sdd->fss, fspair)) {
-               nvlist_t *fslist;
+               nvlist_t *fslist, *parent_nv;
                char *fsname;
                zfs_handle_t *zhp;
                int err;
                uint64_t origin_guid = 0;
-               nvlist_t *origin_nv;
+               uint64_t parent_guid = 0;
 
                VERIFY(nvpair_value_nvlist(fspair, &fslist) == 0);
                if (nvlist_lookup_boolean(fslist, "sent") == 0)
@@ -651,16 +1306,30 @@ again:
 
                VERIFY(nvlist_lookup_string(fslist, "name", &fsname) == 0);
                (void) nvlist_lookup_uint64(fslist, "origin", &origin_guid);
+               (void) nvlist_lookup_uint64(fslist, "parentfromsnap",
+                   &parent_guid);
 
-               origin_nv = fsavl_find(sdd->fsavl, origin_guid, NULL);
-               if (origin_nv &&
-                   nvlist_lookup_boolean(origin_nv, "sent") == ENOENT) {
-                       /*
-                        * origin has not been sent yet;
-                        * skip this clone.
-                        */
-                       needagain = B_TRUE;
-                       continue;
+               if (parent_guid != 0) {
+                       parent_nv = fsavl_find(sdd->fsavl, parent_guid, NULL);
+                       if (!nvlist_exists(parent_nv, "sent")) {
+                               /* parent has not been sent; skip this one */
+                               needagain = B_TRUE;
+                               continue;
+                       }
+               }
+
+               if (origin_guid != 0) {
+                       nvlist_t *origin_nv = fsavl_find(sdd->fsavl,
+                           origin_guid, NULL);
+                       if (origin_nv != NULL &&
+                           !nvlist_exists(origin_nv, "sent")) {
+                               /*
+                                * origin has not been sent yet;
+                                * skip this clone.
+                                */
+                               needagain = B_TRUE;
+                               continue;
+                       }
                }
 
                zhp = zfs_open(rzhp->zfs_hdl, fsname, ZFS_TYPE_DATASET);
@@ -677,24 +1346,51 @@ again:
                assert(progress);
                goto again;
        }
+
+       /* clean out the sent flags in case we reuse this fss */
+       for (fspair = nvlist_next_nvpair(sdd->fss, NULL); fspair;
+           fspair = nvlist_next_nvpair(sdd->fss, fspair)) {
+               nvlist_t *fslist;
+
+               VERIFY(nvpair_value_nvlist(fspair, &fslist) == 0);
+               (void) nvlist_remove_all(fslist, "sent");
+       }
+
        return (0);
 }
 
 /*
- * Dumps a backup of tosnap, incremental from fromsnap if it isn't NULL.
- * If 'doall', dump all intermediate snaps.
- * If 'replicate', dump special header and do recursively.
+ * Generate a send stream for the dataset identified by the argument zhp.
+ *
+ * The content of the send stream is the snapshot identified by
+ * 'tosnap'.  Incremental streams are requested in two ways:
+ *     - from the snapshot identified by "fromsnap" (if non-null) or
+ *     - from the origin of the dataset identified by zhp, which must
+ *      be a clone.  In this case, "fromsnap" is null and "fromorigin"
+ *      is TRUE.
+ *
+ * The send stream is recursive (i.e. dumps a hierarchy of snapshots) and
+ * uses a special header (with a hdrtype field of DMU_COMPOUNDSTREAM)
+ * if "replicate" is set.  If "doall" is set, dump all the intermediate
+ * snapshots. The DMU_COMPOUNDSTREAM header is used in the "doall"
+ * case too. If "props" is set, send properties.
  */
 int
 zfs_send(zfs_handle_t *zhp, const char *fromsnap, const char *tosnap,
-    boolean_t replicate, boolean_t doall, boolean_t fromorigin,
-    boolean_t verbose, int outfd)
+    sendflags_t *flags, int outfd, snapfilter_cb_t filter_func,
+    void *cb_arg, nvlist_t **debugnvp)
 {
        char errbuf[1024];
        send_dump_data_t sdd = { 0 };
-       int err;
+       int err = 0;
        nvlist_t *fss = NULL;
        avl_tree_t *fsavl = NULL;
+       static uint64_t holdseq;
+       int spa_version;
+       pthread_t tid;
+       int pipefd[2];
+       dedup_arg_t dda = { 0 };
+       int featureflags = 0;
 
        (void) snprintf(errbuf, sizeof (errbuf), dgettext(TEXT_DOMAIN,
            "cannot send '%s'"), zhp->zfs_name);
@@ -705,15 +1401,41 @@ zfs_send(zfs_handle_t *zhp, const char *fromsnap, const char *tosnap,
                return (zfs_error(zhp->zfs_hdl, EZFS_NOENT, errbuf));
        }
 
-       if (replicate || doall) {
+       if (zhp->zfs_type == ZFS_TYPE_FILESYSTEM) {
+               uint64_t version;
+               version = zfs_prop_get_int(zhp, ZFS_PROP_VERSION);
+               if (version >= ZPL_VERSION_SA) {
+                       featureflags |= DMU_BACKUP_FEATURE_SA_SPILL;
+               }
+       }
+
+       if (flags->dedup && !flags->dryrun) {
+               featureflags |= (DMU_BACKUP_FEATURE_DEDUP |
+                   DMU_BACKUP_FEATURE_DEDUPPROPS);
+               if ((err = socketpair(AF_UNIX, SOCK_STREAM, 0, pipefd))) {
+                       zfs_error_aux(zhp->zfs_hdl, strerror(errno));
+                       return (zfs_error(zhp->zfs_hdl, EZFS_PIPEFAILED,
+                           errbuf));
+               }
+               dda.outputfd = outfd;
+               dda.inputfd = pipefd[1];
+               dda.dedup_hdl = zhp->zfs_hdl;
+               if ((err = pthread_create(&tid, NULL, cksummer, &dda))) {
+                       (void) close(pipefd[0]);
+                       (void) close(pipefd[1]);
+                       zfs_error_aux(zhp->zfs_hdl, strerror(errno));
+                       return (zfs_error(zhp->zfs_hdl,
+                           EZFS_THREADCREATEFAILED, errbuf));
+               }
+       }
+
+       if (flags->replicate || flags->doall || flags->props) {
                dmu_replay_record_t drr = { 0 };
                char *packbuf = NULL;
                size_t buflen = 0;
-               zio_cksum_t zc = { 0 };
+               zio_cksum_t zc = { { 0 } };
 
-               assert(fromsnap || doall);
-
-               if (replicate) {
+               if (flags->replicate || flags->props) {
                        nvlist_t *hdrnv;
 
                        VERIFY(0 == nvlist_alloc(&hdrnv, NV_UNIQUE_NAME, 0));
@@ -722,49 +1444,57 @@ zfs_send(zfs_handle_t *zhp, const char *fromsnap, const char *tosnap,
                                    "fromsnap", fromsnap));
                        }
                        VERIFY(0 == nvlist_add_string(hdrnv, "tosnap", tosnap));
+                       if (!flags->replicate) {
+                               VERIFY(0 == nvlist_add_boolean(hdrnv,
+                                   "not_recursive"));
+                       }
 
                        err = gather_nvlist(zhp->zfs_hdl, zhp->zfs_name,
-                           fromsnap, tosnap, &fss, &fsavl);
+                           fromsnap, tosnap, flags->replicate, &fss, &fsavl);
                        if (err)
-                               return (err);
+                               goto err_out;
                        VERIFY(0 == nvlist_add_nvlist(hdrnv, "fss", fss));
                        err = nvlist_pack(hdrnv, &packbuf, &buflen,
                            NV_ENCODE_XDR, 0);
-                       nvlist_free(hdrnv);
+                       if (debugnvp)
+                               *debugnvp = hdrnv;
+                       else
+                               nvlist_free(hdrnv);
                        if (err) {
                                fsavl_destroy(fsavl);
                                nvlist_free(fss);
-                               return (zfs_standard_error(zhp->zfs_hdl,
-                                   err, errbuf));
+                               goto stderr_out;
                        }
                }
 
-               /* write first begin record */
-               drr.drr_type = DRR_BEGIN;
-               drr.drr_u.drr_begin.drr_magic = DMU_BACKUP_MAGIC;
-               drr.drr_u.drr_begin.drr_version = DMU_BACKUP_HEADER_VERSION;
-               (void) snprintf(drr.drr_u.drr_begin.drr_toname,
-                   sizeof (drr.drr_u.drr_begin.drr_toname),
-                   "%s@%s", zhp->zfs_name, tosnap);
-               drr.drr_payloadlen = buflen;
-               fletcher_4_incremental_native(&drr, sizeof (drr), &zc);
-               err = write(outfd, &drr, sizeof (drr));
-
-               /* write header nvlist */
-               if (err != -1) {
-                       fletcher_4_incremental_native(packbuf, buflen, &zc);
-                       err = write(outfd, packbuf, buflen);
-               }
-               free(packbuf);
-               if (err == -1) {
-                       fsavl_destroy(fsavl);
-                       nvlist_free(fss);
-                       return (zfs_standard_error(zhp->zfs_hdl,
-                           errno, errbuf));
-               }
+               if (!flags->dryrun) {
+                       /* write first begin record */
+                       drr.drr_type = DRR_BEGIN;
+                       drr.drr_u.drr_begin.drr_magic = DMU_BACKUP_MAGIC;
+                       DMU_SET_STREAM_HDRTYPE(drr.drr_u.drr_begin.
+                           drr_versioninfo, DMU_COMPOUNDSTREAM);
+                       DMU_SET_FEATUREFLAGS(drr.drr_u.drr_begin.
+                           drr_versioninfo, featureflags);
+                       (void) snprintf(drr.drr_u.drr_begin.drr_toname,
+                           sizeof (drr.drr_u.drr_begin.drr_toname),
+                           "%s@%s", zhp->zfs_name, tosnap);
+                       drr.drr_payloadlen = buflen;
+                       err = cksum_and_write(&drr, sizeof (drr), &zc, outfd);
+
+                       /* write header nvlist */
+                       if (err != -1 && packbuf != NULL) {
+                               err = cksum_and_write(packbuf, buflen, &zc,
+                                   outfd);
+                       }
+                       free(packbuf);
+                       if (err == -1) {
+                               fsavl_destroy(fsavl);
+                               nvlist_free(fss);
+                               err = errno;
+                               goto stderr_out;
+                       }
 
-               /* write end record */
-               if (err != -1) {
+                       /* write end record */
                        bzero(&drr, sizeof (drr));
                        drr.drr_type = DRR_END;
                        drr.drr_u.drr_end.drr_checksum = zc;
@@ -772,27 +1502,93 @@ zfs_send(zfs_handle_t *zhp, const char *fromsnap, const char *tosnap,
                        if (err == -1) {
                                fsavl_destroy(fsavl);
                                nvlist_free(fss);
-                               return (zfs_standard_error(zhp->zfs_hdl,
-                                   errno, errbuf));
+                               err = errno;
+                               goto stderr_out;
                        }
+
+                       err = 0;
                }
        }
 
        /* dump each stream */
        sdd.fromsnap = fromsnap;
        sdd.tosnap = tosnap;
-       sdd.outfd = outfd;
-       sdd.replicate = replicate;
-       sdd.doall = doall;
-       sdd.fromorigin = fromorigin;
+       if (flags->dedup)
+               sdd.outfd = pipefd[0];
+       else
+               sdd.outfd = outfd;
+       sdd.replicate = flags->replicate;
+       sdd.doall = flags->doall;
+       sdd.fromorigin = flags->fromorigin;
        sdd.fss = fss;
        sdd.fsavl = fsavl;
-       sdd.verbose = verbose;
+       sdd.verbose = flags->verbose;
+       sdd.parsable = flags->parsable;
+       sdd.progress = flags->progress;
+       sdd.dryrun = flags->dryrun;
+       sdd.filter_cb = filter_func;
+       sdd.filter_cb_arg = cb_arg;
+       if (debugnvp)
+               sdd.debugnv = *debugnvp;
+
+       /*
+        * Some flags require that we place user holds on the datasets that are
+        * being sent so they don't get destroyed during the send. We can skip
+        * this step if the pool is imported read-only since the datasets cannot
+        * be destroyed.
+        */
+       if (!flags->dryrun && !zpool_get_prop_int(zfs_get_pool_handle(zhp),
+           ZPOOL_PROP_READONLY, NULL) &&
+           zfs_spa_version(zhp, &spa_version) == 0 &&
+           spa_version >= SPA_VERSION_USERREFS &&
+           (flags->doall || flags->replicate)) {
+               ++holdseq;
+               (void) snprintf(sdd.holdtag, sizeof (sdd.holdtag),
+                   ".send-%d-%llu", getpid(), (u_longlong_t)holdseq);
+               sdd.cleanup_fd = open(ZFS_DEV, O_RDWR);
+               if (sdd.cleanup_fd < 0) {
+                       err = errno;
+                       goto stderr_out;
+               }
+       } else {
+               sdd.cleanup_fd = -1;
+       }
+       if (flags->verbose) {
+               /*
+                * Do a verbose no-op dry run to get all the verbose output
+                * before generating any data.  Then do a non-verbose real
+                * run to generate the streams.
+                */
+               sdd.dryrun = B_TRUE;
+               err = dump_filesystems(zhp, &sdd);
+               sdd.dryrun = flags->dryrun;
+               sdd.verbose = B_FALSE;
+               if (flags->parsable) {
+                       (void) fprintf(stderr, "size\t%llu\n",
+                           (longlong_t)sdd.size);
+               } else {
+                       char buf[16];
+                       zfs_nicenum(sdd.size, buf, sizeof (buf));
+                       (void) fprintf(stderr, dgettext(TEXT_DOMAIN,
+                           "total estimated size is %s\n"), buf);
+               }
+       }
        err = dump_filesystems(zhp, &sdd);
        fsavl_destroy(fsavl);
        nvlist_free(fss);
 
-       if (replicate || doall) {
+       if (flags->dedup) {
+               (void) close(pipefd[0]);
+               (void) pthread_join(tid, NULL);
+       }
+
+       if (sdd.cleanup_fd != -1) {
+               VERIFY(0 == close(sdd.cleanup_fd));
+               sdd.cleanup_fd = -1;
+       }
+
+       if (!flags->dryrun && (flags->replicate || flags->doall ||
+           flags->props)) {
                /*
                 * write final end record.  NB: want to do this even if
                 * there was some error, because it might not be totally
@@ -807,6 +1603,18 @@ zfs_send(zfs_handle_t *zhp, const char *fromsnap, const char *tosnap,
        }
 
        return (err || sdd.err);
+
+stderr_out:
+       err = zfs_standard_error(zhp->zfs_hdl, err, errbuf);
+err_out:
+       if (sdd.cleanup_fd != -1)
+               VERIFY(0 == close(sdd.cleanup_fd));
+       if (flags->dedup) {
+               (void) pthread_cancel(tid);
+               (void) pthread_join(tid, NULL);
+               (void) close(pipefd[0]);
+       }
+       return (err);
 }
 
 /*
@@ -872,10 +1680,10 @@ recv_read_nvlist(libzfs_handle_t *hdl, int fd, int len, nvlist_t **nvp,
 
 static int
 recv_rename(libzfs_handle_t *hdl, const char *name, const char *tryname,
-    int baselen, char *newname, recvflags_t flags)
+    int baselen, char *newname, recvflags_t *flags)
 {
        static int seq;
-       zfs_cmd_t zc = { 0 };
+       zfs_cmd_t zc = { "\0", "\0", "\0", "\0", 0 };
        int err;
        prop_changelist_t *clp;
        zfs_handle_t *zhp;
@@ -884,7 +1692,7 @@ recv_rename(libzfs_handle_t *hdl, const char *name, const char *tryname,
        if (zhp == NULL)
                return (-1);
        clp = changelist_gather(zhp, ZFS_PROP_NAME, 0,
-           flags.force ? MS_FORCE : 0);
+           flags->force ? MS_FORCE : 0);
        zfs_close(zhp);
        if (clp == NULL)
                return (-1);
@@ -892,14 +1700,15 @@ recv_rename(libzfs_handle_t *hdl, const char *name, const char *tryname,
        if (err)
                return (err);
 
+       zc.zc_objset_type = DMU_OST_ZFS;
+       (void) strlcpy(zc.zc_name, name, sizeof (zc.zc_name));
+
        if (tryname) {
                (void) strcpy(newname, tryname);
 
-               zc.zc_objset_type = DMU_OST_ZFS;
-               (void) strlcpy(zc.zc_name, name, sizeof (zc.zc_name));
                (void) strlcpy(zc.zc_value, tryname, sizeof (zc.zc_value));
 
-               if (flags.verbose) {
+               if (flags->verbose) {
                        (void) printf("attempting rename %s to %s\n",
                            zc.zc_name, zc.zc_value);
                }
@@ -915,22 +1724,22 @@ recv_rename(libzfs_handle_t *hdl, const char *name, const char *tryname,
 
                (void) strncpy(newname, name, baselen);
                (void) snprintf(newname+baselen, ZFS_MAXNAMELEN-baselen,
-                   "recv-%u-%u", getpid(), seq);
+                   "recv-%ld-%u", (long) getpid(), seq);
                (void) strlcpy(zc.zc_value, newname, sizeof (zc.zc_value));
 
-               if (flags.verbose) {
+               if (flags->verbose) {
                        (void) printf("failed - trying rename %s to %s\n",
                            zc.zc_name, zc.zc_value);
                }
                err = ioctl(hdl->libzfs_fd, ZFS_IOC_RENAME, &zc);
                if (err == 0)
                        changelist_rename(clp, name, newname);
-               if (err && flags.verbose) {
+               if (err && flags->verbose) {
                        (void) printf("failed (%u) - "
                            "will try again on next pass\n", errno);
                }
                err = EAGAIN;
-       } else if (flags.verbose) {
+       } else if (flags->verbose) {
                if (err == 0)
                        (void) printf("success\n");
                else
@@ -945,18 +1754,24 @@ recv_rename(libzfs_handle_t *hdl, const char *name, const char *tryname,
 
 static int
 recv_destroy(libzfs_handle_t *hdl, const char *name, int baselen,
-    char *newname, recvflags_t flags)
+    char *newname, recvflags_t *flags)
 {
-       zfs_cmd_t zc = { 0 };
+       zfs_cmd_t zc = { "\0", "\0", "\0", "\0", 0 };
        int err = 0;
        prop_changelist_t *clp;
        zfs_handle_t *zhp;
+       boolean_t defer = B_FALSE;
+       int spa_version;
 
        zhp = zfs_open(hdl, name, ZFS_TYPE_DATASET);
        if (zhp == NULL)
                return (-1);
        clp = changelist_gather(zhp, ZFS_PROP_NAME, 0,
-           flags.force ? MS_FORCE : 0);
+           flags->force ? MS_FORCE : 0);
+       if (zfs_get_type(zhp) == ZFS_TYPE_SNAPSHOT &&
+           zfs_spa_version(zhp, &spa_version) == 0 &&
+           spa_version >= SPA_VERSION_USERREFS)
+               defer = B_TRUE;
        zfs_close(zhp);
        if (clp == NULL)
                return (-1);
@@ -965,14 +1780,14 @@ recv_destroy(libzfs_handle_t *hdl, const char *name, int baselen,
                return (err);
 
        zc.zc_objset_type = DMU_OST_ZFS;
+       zc.zc_defer_destroy = defer;
        (void) strlcpy(zc.zc_name, name, sizeof (zc.zc_name));
 
-       if (flags.verbose)
+       if (flags->verbose)
                (void) printf("attempting destroy %s\n", zc.zc_name);
        err = ioctl(hdl->libzfs_fd, ZFS_IOC_DESTROY, &zc);
-
        if (err == 0) {
-               if (flags.verbose)
+               if (flags->verbose)
                        (void) printf("success\n");
                changelist_remove(clp, zc.zc_name);
        }
@@ -980,8 +1795,14 @@ recv_destroy(libzfs_handle_t *hdl, const char *name, int baselen,
        (void) changelist_postfix(clp);
        changelist_free(clp);
 
-       if (err != 0)
+       /*
+        * Deferred destroy might destroy the snapshot or only mark it to be
+        * destroyed later, and it returns success in either case.
+        */
+       if (err != 0 || (defer && zfs_dataset_exists(hdl, name,
+           ZFS_TYPE_SNAPSHOT))) {
                err = recv_rename(hdl, name, NULL, baselen, newname, flags);
+       }
 
        return (err);
 }
@@ -989,6 +1810,7 @@ recv_destroy(libzfs_handle_t *hdl, const char *name, int baselen,
 typedef struct guid_to_name_data {
        uint64_t guid;
        char *name;
+       char *skip;
 } guid_to_name_data_t;
 
 static int
@@ -997,20 +1819,35 @@ guid_to_name_cb(zfs_handle_t *zhp, void *arg)
        guid_to_name_data_t *gtnd = arg;
        int err;
 
+       if (gtnd->skip != NULL &&
+           strcmp(zhp->zfs_name, gtnd->skip) == 0) {
+               return (0);
+       }
+
        if (zhp->zfs_dmustats.dds_guid == gtnd->guid) {
                (void) strcpy(gtnd->name, zhp->zfs_name);
+               zfs_close(zhp);
                return (EEXIST);
        }
+
        err = zfs_iter_children(zhp, guid_to_name_cb, gtnd);
        zfs_close(zhp);
        return (err);
 }
 
+/*
+ * Attempt to find the local dataset associated with this guid.  In the case of
+ * multiple matches, we attempt to find the "best" match by searching
+ * progressively larger portions of the hierarchy.  This allows one to send a
+ * tree of datasets individually and guarantee that we will find the source
+ * guid within that hierarchy, even if there are multiple matches elsewhere.
+ */
 static int
 guid_to_name(libzfs_handle_t *hdl, const char *parent, uint64_t guid,
     char *name)
 {
        /* exhaustive search all local snapshots */
+       char pname[ZFS_MAXNAMELEN];
        guid_to_name_data_t gtnd;
        int err = 0;
        zfs_handle_t *zhp;
@@ -1018,35 +1855,42 @@ guid_to_name(libzfs_handle_t *hdl, const char *parent, uint64_t guid,
 
        gtnd.guid = guid;
        gtnd.name = name;
+       gtnd.skip = NULL;
 
-       if (strchr(parent, '@') == NULL) {
-               zhp = make_dataset_handle(hdl, parent);
-               if (zhp != NULL) {
-                       err = zfs_iter_children(zhp, guid_to_name_cb, &gtnd);
-                       zfs_close(zhp);
-                       if (err == EEXIST)
-                               return (0);
-               }
-       }
+       (void) strlcpy(pname, parent, sizeof (pname));
 
-       cp = strchr(parent, '/');
-       if (cp)
+       /*
+        * Search progressively larger portions of the hierarchy.  This will
+        * select the "most local" version of the origin snapshot in the case
+        * that there are multiple matching snapshots in the system.
+        */
+       while ((cp = strrchr(pname, '/')) != NULL) {
+
+               /* Chop off the last component and open the parent */
                *cp = '\0';
-       zhp = make_dataset_handle(hdl, parent);
-       if (cp)
-               *cp = '/';
+               zhp = make_dataset_handle(hdl, pname);
+
+               if (zhp == NULL)
+                       continue;
 
-       if (zhp) {
                err = zfs_iter_children(zhp, guid_to_name_cb, &gtnd);
                zfs_close(zhp);
-       }
+               if (err == EEXIST)
+                       return (0);
 
-       return (err == EEXIST ? 0 : ENOENT);
+               /*
+                * Remember the dataset that we already searched, so we
+                * skip it next time through.
+                */
+               gtnd.skip = pname;
+       }
 
+       return (ENOENT);
 }
 
 /*
- * Return true if dataset guid1 is created before guid2.
+ * Return +1 if guid1 is before guid2, 0 if they are the same, and -1 if
+ * guid1 is after guid2.
  */
 static int
 created_before(libzfs_handle_t *hdl, avl_tree_t *avl,
@@ -1056,7 +1900,8 @@ created_before(libzfs_handle_t *hdl, avl_tree_t *avl,
        char *fsname, *snapname;
        char buf[ZFS_MAXNAMELEN];
        int rv;
-       zfs_node_t zn1, zn2;
+       zfs_handle_t *guid1hdl, *guid2hdl;
+       uint64_t create1, create2;
 
        if (guid2 == 0)
                return (0);
@@ -1066,50 +1911,62 @@ created_before(libzfs_handle_t *hdl, avl_tree_t *avl,
        nvfs = fsavl_find(avl, guid1, &snapname);
        VERIFY(0 == nvlist_lookup_string(nvfs, "name", &fsname));
        (void) snprintf(buf, sizeof (buf), "%s@%s", fsname, snapname);
-       zn1.zn_handle = zfs_open(hdl, buf, ZFS_TYPE_SNAPSHOT);
-       if (zn1.zn_handle == NULL)
+       guid1hdl = zfs_open(hdl, buf, ZFS_TYPE_SNAPSHOT);
+       if (guid1hdl == NULL)
                return (-1);
 
        nvfs = fsavl_find(avl, guid2, &snapname);
        VERIFY(0 == nvlist_lookup_string(nvfs, "name", &fsname));
        (void) snprintf(buf, sizeof (buf), "%s@%s", fsname, snapname);
-       zn2.zn_handle = zfs_open(hdl, buf, ZFS_TYPE_SNAPSHOT);
-       if (zn2.zn_handle == NULL) {
-               zfs_close(zn2.zn_handle);
+       guid2hdl = zfs_open(hdl, buf, ZFS_TYPE_SNAPSHOT);
+       if (guid2hdl == NULL) {
+               zfs_close(guid1hdl);
                return (-1);
        }
 
-       rv = (zfs_snapshot_compare(&zn1, &zn2) == -1);
+       create1 = zfs_prop_get_int(guid1hdl, ZFS_PROP_CREATETXG);
+       create2 = zfs_prop_get_int(guid2hdl, ZFS_PROP_CREATETXG);
+
+       if (create1 < create2)
+               rv = -1;
+       else if (create1 > create2)
+               rv = +1;
+       else
+               rv = 0;
 
-       zfs_close(zn1.zn_handle);
-       zfs_close(zn2.zn_handle);
+       zfs_close(guid1hdl);
+       zfs_close(guid2hdl);
 
        return (rv);
 }
 
 static int
 recv_incremental_replication(libzfs_handle_t *hdl, const char *tofs,
-    recvflags_t flags, nvlist_t *stream_nv, avl_tree_t *stream_avl)
+    recvflags_t *flags, nvlist_t *stream_nv, avl_tree_t *stream_avl,
+    nvlist_t *renamed)
 {
        nvlist_t *local_nv;
        avl_tree_t *local_avl;
        nvpair_t *fselem, *nextfselem;
-       char *tosnap, *fromsnap;
+       char *fromsnap;
        char newname[ZFS_MAXNAMELEN];
        int error;
-       boolean_t needagain, progress;
+       boolean_t needagain, progress, recursive;
+       char *s1, *s2;
 
        VERIFY(0 == nvlist_lookup_string(stream_nv, "fromsnap", &fromsnap));
-       VERIFY(0 == nvlist_lookup_string(stream_nv, "tosnap", &tosnap));
 
-       if (flags.dryrun)
+       recursive = (nvlist_lookup_boolean(stream_nv, "not_recursive") ==
+           ENOENT);
+
+       if (flags->dryrun)
                return (0);
 
 again:
        needagain = progress = B_FALSE;
 
        if ((error = gather_nvlist(hdl, tofs, fromsnap, NULL,
-           &local_nv, &local_avl)) != 0)
+           recursive, &local_nv, &local_avl)) != 0)
                return (error);
 
        /*
@@ -1158,11 +2015,11 @@ again:
                            stream_originguid, originguid)) {
                        case 1: {
                                /* promote it! */
-                               zfs_cmd_t zc = { 0 };
+                               zfs_cmd_t zc = { "\0", "\0", "\0", "\0", 0 };
                                nvlist_t *origin_nvfs;
                                char *origin_fsname;
 
-                               if (flags.verbose)
+                               if (flags->verbose)
                                        (void) printf("promoting %s\n", fsname);
 
                                origin_nvfs = fsavl_find(local_avl, originguid,
@@ -1210,7 +2067,7 @@ again:
                        if (found == NULL) {
                                char name[ZFS_MAXNAMELEN];
 
-                               if (!flags.force)
+                               if (!flags->force)
                                        continue;
 
                                (void) snprintf(name, sizeof (name), "%s@%s",
@@ -1230,9 +2087,9 @@ again:
                        if (0 == nvlist_lookup_nvlist(stream_nvfs, "snapprops",
                            &props) && 0 == nvlist_lookup_nvlist(props,
                            stream_snapname, &props)) {
-                               zfs_cmd_t zc = { 0 };
+                               zfs_cmd_t zc = { "\0", "\0", "\0", "\0", 0 };
 
-                               zc.zc_cookie = B_TRUE; /* clear current props */
+                               zc.zc_cookie = B_TRUE; /* received */
                                (void) snprintf(zc.zc_name, sizeof (zc.zc_name),
                                    "%s@%s", fsname, nvpair_name(snapelem));
                                if (zcmd_write_src_nvlist(hdl, &zc,
@@ -1268,7 +2125,7 @@ again:
 
                /* check for delete */
                if (stream_nvfs == NULL) {
-                       if (!flags.force)
+                       if (!flags->force)
                                continue;
 
                        error = recv_destroy(hdl, fsname, strlen(tofs)+1,
@@ -1280,10 +2137,13 @@ again:
                        continue;
                }
 
-               if (fromguid == 0 && flags.verbose) {
-                       (void) printf("local fs %s does not have fromsnap "
-                           "(%s in stream); must have been deleted locally; "
-                           "ignoring\n", fsname, fromsnap);
+               if (fromguid == 0) {
+                       if (flags->verbose) {
+                               (void) printf("local fs %s does not have "
+                                   "fromsnap (%s in stream); must have "
+                                   "been deleted locally; ignoring\n",
+                                   fsname, fromsnap);
+                       }
                        continue;
                }
 
@@ -1292,11 +2152,19 @@ again:
                VERIFY(0 == nvlist_lookup_uint64(stream_nvfs,
                    "parentfromsnap", &stream_parent_fromsnap_guid));
 
-               /* check for rename */
+               s1 = strrchr(fsname, '/');
+               s2 = strrchr(stream_fsname, '/');
+
+               /*
+                * 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.
+                */
                if ((stream_parent_fromsnap_guid != 0 &&
+                   parent_fromsnap_guid != 0 &&
                    stream_parent_fromsnap_guid != parent_fromsnap_guid) ||
-                   strcmp(strrchr(fsname, '/'),
-                   strrchr(stream_fsname, '/')) != 0) {
+                   ((flags->isprefix || strcmp(tofs, fsname) != 0) &&
+                   (s1 != NULL) && (s2 != NULL) && strcmp(s1, s2) != 0)) {
                        nvlist_t *parent;
                        char tryname[ZFS_MAXNAMELEN];
 
@@ -1318,14 +2186,22 @@ again:
                                    "%s%s", pname, strrchr(stream_fsname, '/'));
                        } else {
                                tryname[0] = '\0';
-                               if (flags.verbose) {
+                               if (flags->verbose) {
                                        (void) printf("local fs %s new parent "
                                            "not found\n", fsname);
                                }
                        }
 
+                       newname[0] = '\0';
+
                        error = recv_rename(hdl, fsname, tryname,
                            strlen(tofs)+1, newname, flags);
+
+                       if (renamed != NULL && newname[0] != '\0') {
+                               VERIFY(0 == nvlist_add_boolean(renamed,
+                                   newname));
+                       }
+
                        if (error)
                                needagain = B_TRUE;
                        else
@@ -1338,7 +2214,7 @@ again:
 
        if (needagain && progress) {
                /* do another pass to fix up temporary names */
-               if (flags.verbose)
+               if (flags->verbose)
                        (void) printf("another pass:\n");
                goto again;
        }
@@ -1348,58 +2224,59 @@ again:
 
 static int
 zfs_receive_package(libzfs_handle_t *hdl, int fd, const char *destname,
-    recvflags_t flags, dmu_replay_record_t *drr, zio_cksum_t *zc,
-    char **top_zfs)
+    recvflags_t *flags, dmu_replay_record_t *drr, zio_cksum_t *zc,
+    char **top_zfs, int cleanup_fd, uint64_t *action_handlep)
 {
        nvlist_t *stream_nv = NULL;
        avl_tree_t *stream_avl = NULL;
        char *fromsnap = NULL;
+       char *cp;
        char tofs[ZFS_MAXNAMELEN];
+       char sendfs[ZFS_MAXNAMELEN];
        char errbuf[1024];
        dmu_replay_record_t drre;
        int error;
        boolean_t anyerr = B_FALSE;
        boolean_t softerr = B_FALSE;
+       boolean_t recursive;
 
        (void) snprintf(errbuf, sizeof (errbuf), dgettext(TEXT_DOMAIN,
            "cannot receive"));
 
-       if (strchr(destname, '@')) {
-               zfs_error_aux(hdl, dgettext(TEXT_DOMAIN,
-                   "can not specify snapshot name for multi-snapshot stream"));
-               return (zfs_error(hdl, EZFS_BADSTREAM, errbuf));
-       }
-
        assert(drr->drr_type == DRR_BEGIN);
        assert(drr->drr_u.drr_begin.drr_magic == DMU_BACKUP_MAGIC);
-       assert(drr->drr_u.drr_begin.drr_version == DMU_BACKUP_HEADER_VERSION);
+       assert(DMU_GET_STREAM_HDRTYPE(drr->drr_u.drr_begin.drr_versioninfo) ==
+           DMU_COMPOUNDSTREAM);
 
        /*
         * Read in the nvlist from the stream.
         */
        if (drr->drr_payloadlen != 0) {
-               if (!flags.isprefix) {
-                       zfs_error_aux(hdl, dgettext(TEXT_DOMAIN,
-                           "must use -d to receive replication "
-                           "(send -R) stream"));
-                       return (zfs_error(hdl, EZFS_BADSTREAM, errbuf));
-               }
-
                error = recv_read_nvlist(hdl, fd, drr->drr_payloadlen,
-                   &stream_nv, flags.byteswap, zc);
+                   &stream_nv, flags->byteswap, zc);
                if (error) {
                        error = zfs_error(hdl, EZFS_BADSTREAM, errbuf);
                        goto out;
                }
        }
 
+       recursive = (nvlist_lookup_boolean(stream_nv, "not_recursive") ==
+           ENOENT);
+
+       if (recursive && strchr(destname, '@')) {
+               zfs_error_aux(hdl, dgettext(TEXT_DOMAIN,
+                   "cannot specify snapshot name for multi-snapshot stream"));
+               error = zfs_error(hdl, EZFS_BADSTREAM, errbuf);
+               goto out;
+       }
+
        /*
         * Read in the end record and verify checksum.
         */
        if (0 != (error = recv_read(hdl, fd, &drre, sizeof (drre),
-           flags.byteswap, NULL)))
+           flags->byteswap, NULL)))
                goto out;
-       if (flags.byteswap) {
+       if (flags->byteswap) {
                drre.drr_type = BSWAP_32(drre.drr_type);
                drre.drr_u.drr_end.drr_checksum.zc_word[0] =
                    BSWAP_64(drre.drr_u.drr_end.drr_checksum.zc_word[0]);
@@ -1436,21 +2313,73 @@ zfs_receive_package(libzfs_handle_t *hdl, int fd, const char *destname,
                }
 
                if (fromsnap != NULL) {
+                       nvlist_t *renamed = NULL;
+                       nvpair_t *pair = NULL;
+
                        (void) strlcpy(tofs, destname, ZFS_MAXNAMELEN);
-                       if (flags.isprefix) {
-                               int i = strcspn(drr->drr_u.drr_begin.drr_toname,
-                                   "/@");
+                       if (flags->isprefix) {
+                               struct drr_begin *drrb = &drr->drr_u.drr_begin;
+                               int i;
+
+                               if (flags->istail) {
+                                       cp = strrchr(drrb->drr_toname, '/');
+                                       if (cp == NULL) {
+                                               (void) strlcat(tofs, "/",
+                                                   ZFS_MAXNAMELEN);
+                                               i = 0;
+                                       } else {
+                                               i = (cp - drrb->drr_toname);
+                                       }
+                               } else {
+                                       i = strcspn(drrb->drr_toname, "/@");
+                               }
                                /* zfs_receive_one() will create_parents() */
-                               (void) strlcat(tofs,
-                                   &drr->drr_u.drr_begin.drr_toname[i],
+                               (void) strlcat(tofs, &drrb->drr_toname[i],
                                    ZFS_MAXNAMELEN);
                                *strchr(tofs, '@') = '\0';
                        }
-                       softerr = recv_incremental_replication(hdl, tofs,
-                           flags, stream_nv, stream_avl);
+
+                       if (recursive && !flags->dryrun && !flags->nomount) {
+                               VERIFY(0 == nvlist_alloc(&renamed,
+                                   NV_UNIQUE_NAME, 0));
+                       }
+
+                       softerr = recv_incremental_replication(hdl, tofs, flags,
+                           stream_nv, stream_avl, renamed);
+
+                       /* Unmount renamed filesystems before receiving. */
+                       while ((pair = nvlist_next_nvpair(renamed,
+                           pair)) != NULL) {
+                               zfs_handle_t *zhp;
+                               prop_changelist_t *clp = NULL;
+
+                               zhp = zfs_open(hdl, nvpair_name(pair),
+                                   ZFS_TYPE_FILESYSTEM);
+                               if (zhp != NULL) {
+                                       clp = changelist_gather(zhp,
+                                           ZFS_PROP_MOUNTPOINT, 0, 0);
+                                       zfs_close(zhp);
+                                       if (clp != NULL) {
+                                               softerr |=
+                                                   changelist_prefix(clp);
+                                               changelist_free(clp);
+                                       }
+                               }
+                       }
+
+                       nvlist_free(renamed);
                }
        }
 
+       /*
+        * Get the fs specified by the first path in the stream (the top level
+        * specified by 'zfs send') and pass it to each invocation of
+        * zfs_receive_one().
+        */
+       (void) strlcpy(sendfs, drr->drr_u.drr_begin.drr_toname,
+           ZFS_MAXNAMELEN);
+       if ((cp = strchr(sendfs, '@')) != NULL)
+               *cp = '\0';
 
        /* Finally, receive each contained stream */
        do {
@@ -1462,7 +2391,8 @@ zfs_receive_package(libzfs_handle_t *hdl, int fd, const char *destname,
                 * recv_skip() and return 0).
                 */
                error = zfs_receive_impl(hdl, destname, flags, fd,
-                   stream_avl, top_zfs);
+                   sendfs, stream_nv, stream_avl, top_zfs, cleanup_fd,
+                   action_handlep);
                if (error == ENODATA) {
                        error = 0;
                        break;
@@ -1476,7 +2406,7 @@ zfs_receive_package(libzfs_handle_t *hdl, int fd, const char *destname,
                 * renames again.
                 */
                softerr = recv_incremental_replication(hdl, tofs, flags,
-                   stream_nv, stream_avl);
+                   stream_nv, stream_avl, NULL);
        }
 
 out:
@@ -1490,11 +2420,28 @@ out:
        return (error);
 }
 
+static void
+trunc_prop_errs(int truncated)
+{
+       ASSERT(truncated != 0);
+
+       if (truncated == 1)
+               (void) fprintf(stderr, dgettext(TEXT_DOMAIN,
+                   "1 more property could not be set\n"));
+       else
+               (void) fprintf(stderr, dgettext(TEXT_DOMAIN,
+                   "%d more properties could not be set\n"), truncated);
+}
+
 static int
 recv_skip(libzfs_handle_t *hdl, int fd, boolean_t byteswap)
 {
        dmu_replay_record_t *drr;
        void *buf = malloc(1<<20);
+       char errbuf[1024];
+
+       (void) snprintf(errbuf, sizeof (errbuf), dgettext(TEXT_DOMAIN,
+           "cannot receive:"));
 
        /* XXX would be great to use lseek if possible... */
        drr = buf;
@@ -1507,7 +2454,11 @@ recv_skip(libzfs_handle_t *hdl, int fd, boolean_t byteswap)
                switch (drr->drr_type) {
                case DRR_BEGIN:
                        /* NB: not to be used on v2 stream packages */
-                       assert(drr->drr_payloadlen == 0);
+                       if (drr->drr_payloadlen != 0) {
+                               zfs_error_aux(hdl, dgettext(TEXT_DOMAIN,
+                                   "invalid substream header"));
+                               return (zfs_error(hdl, EZFS_BADSTREAM, errbuf));
+                       }
                        break;
 
                case DRR_END:
@@ -1533,13 +2484,23 @@ recv_skip(libzfs_handle_t *hdl, int fd, boolean_t byteswap)
                        (void) recv_read(hdl, fd, buf,
                            drr->drr_u.drr_write.drr_length, B_FALSE, NULL);
                        break;
-
+               case DRR_SPILL:
+                       if (byteswap) {
+                               drr->drr_u.drr_write.drr_length =
+                                   BSWAP_64(drr->drr_u.drr_spill.drr_length);
+                       }
+                       (void) recv_read(hdl, fd, buf,
+                           drr->drr_u.drr_spill.drr_length, B_FALSE, NULL);
+                       break;
+               case DRR_WRITE_BYREF:
                case DRR_FREEOBJECTS:
                case DRR_FREE:
                        break;
 
                default:
-                       assert(!"invalid record type");
+                       zfs_error_aux(hdl, dgettext(TEXT_DOMAIN,
+                           "invalid record type"));
+                       return (zfs_error(hdl, EZFS_BADSTREAM, errbuf));
                }
        }
 
@@ -1552,28 +2513,35 @@ 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,
-    dmu_replay_record_t *drr_noswap, avl_tree_t *stream_avl,
-    char **top_zfs)
+    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)
 {
-       zfs_cmd_t zc = { 0 };
+       zfs_cmd_t zc = { "\0", "\0", "\0", "\0", 0 };
        time_t begin_time;
-       int ioctl_err, ioctl_errno, err, choplen;
+       int ioctl_err, ioctl_errno, err;
        char *cp;
        struct drr_begin *drrb = &drr->drr_u.drr_begin;
        char errbuf[1024];
-       char chopprefix[ZFS_MAXNAMELEN];
+       char prop_errbuf[1024];
+       const char *chopprefix;
        boolean_t newfs = B_FALSE;
        boolean_t stream_wantsnewfs;
        uint64_t parent_snapguid = 0;
        prop_changelist_t *clp = NULL;
        nvlist_t *snapprops_nvlist = NULL;
+       zprop_errflags_t prop_errflags;
+       boolean_t recursive;
 
        begin_time = time(NULL);
 
        (void) snprintf(errbuf, sizeof (errbuf), dgettext(TEXT_DOMAIN,
            "cannot receive"));
 
+       recursive = (nvlist_lookup_boolean(stream_nv, "not_recursive") ==
+           ENOENT);
+
        if (stream_avl != NULL) {
                char *snapname;
                nvlist_t *fs = fsavl_find(stream_avl, drrb->drr_toguid,
@@ -1587,7 +2555,7 @@ zfs_receive_one(libzfs_handle_t *hdl, int infd, const char *tosnap,
                if (err)
                        VERIFY(0 == nvlist_alloc(&props, NV_UNIQUE_NAME, 0));
 
-               if (flags.canmountoff) {
+               if (flags->canmountoff) {
                        VERIFY(0 == nvlist_add_uint64(props,
                            zfs_prop_to_name(ZFS_PROP_CANMOUNT), 0));
                }
@@ -1604,6 +2572,8 @@ zfs_receive_one(libzfs_handle_t *hdl, int infd, const char *tosnap,
                        return (-1);
        }
 
+       cp = NULL;
+
        /*
         * Determine how much of the snapshot name stored in the stream
         * we are going to tack on to the name they specified on the
@@ -1612,38 +2582,77 @@ zfs_receive_one(libzfs_handle_t *hdl, int infd, const char *tosnap,
         * If they specified a snapshot, chop the entire name stored in
         * the stream.
         */
-       (void) strcpy(chopprefix, drrb->drr_toname);
-       if (flags.isprefix) {
+       if (flags->istail) {
+               /*
+                * A filesystem was specified with -e. We want to tack on only
+                * the tail of the sent snapshot path.
+                */
+               if (strchr(tosnap, '@')) {
+                       zfs_error_aux(hdl, dgettext(TEXT_DOMAIN, "invalid "
+                           "argument - snapshot not allowed with -e"));
+                       return (zfs_error(hdl, EZFS_INVALIDNAME, errbuf));
+               }
+
+               chopprefix = strrchr(sendfs, '/');
+
+               if (chopprefix == NULL) {
+                       /*
+                        * The tail is the poolname, so we need to
+                        * prepend a path separator.
+                        */
+                       int len = strlen(drrb->drr_toname);
+                       cp = malloc(len + 2);
+                       cp[0] = '/';
+                       (void) strcpy(&cp[1], drrb->drr_toname);
+                       chopprefix = cp;
+               } else {
+                       chopprefix = drrb->drr_toname + (chopprefix - sendfs);
+               }
+       } else if (flags->isprefix) {
                /*
-                * They specified a fs with -d, we want to tack on
-                * everything but the pool name stored in the stream
+                * A filesystem was specified with -d. We want to tack on
+                * everything but the first element of the sent snapshot path
+                * (all but the pool name).
                 */
                if (strchr(tosnap, '@')) {
                        zfs_error_aux(hdl, dgettext(TEXT_DOMAIN, "invalid "
                            "argument - snapshot not allowed with -d"));
                        return (zfs_error(hdl, EZFS_INVALIDNAME, errbuf));
                }
-               cp = strchr(chopprefix, '/');
-               if (cp == NULL)
-                       cp = strchr(chopprefix, '@');
-               *cp = '\0';
+
+               chopprefix = strchr(drrb->drr_toname, '/');
+               if (chopprefix == NULL)
+                       chopprefix = strchr(drrb->drr_toname, '@');
        } else if (strchr(tosnap, '@') == NULL) {
                /*
-                * If they specified a filesystem without -d, we want to
-                * tack on everything after the fs specified in the
-                * first name from the stream.
+                * If a filesystem was specified without -d or -e, we want to
+                * tack on everything after the fs specified by 'zfs send'.
                 */
-               cp = strchr(chopprefix, '@');
-               *cp = '\0';
+               chopprefix = drrb->drr_toname + strlen(sendfs);
+       } else {
+               /* A snapshot was specified as an exact path (no -d or -e). */
+               if (recursive) {
+                       zfs_error_aux(hdl, dgettext(TEXT_DOMAIN,
+                           "cannot specify snapshot name for multi-snapshot "
+                           "stream"));
+                       return (zfs_error(hdl, EZFS_BADSTREAM, errbuf));
+               }
+               chopprefix = drrb->drr_toname + strlen(drrb->drr_toname);
        }
-       choplen = strlen(chopprefix);
+
+       ASSERT(strstr(drrb->drr_toname, sendfs) == drrb->drr_toname);
+       ASSERT(chopprefix > drrb->drr_toname);
+       ASSERT(chopprefix <= drrb->drr_toname + strlen(drrb->drr_toname));
+       ASSERT(chopprefix[0] == '/' || chopprefix[0] == '@' ||
+           chopprefix[0] == '\0');
 
        /*
         * Determine name of destination snapshot, store in zc_value.
         */
+       (void) strcpy(zc.zc_top_ds, tosnap);
        (void) strcpy(zc.zc_value, tosnap);
-       (void) strncat(zc.zc_value, drrb->drr_toname+choplen,
-           sizeof (zc.zc_value));
+       (void) strlcat(zc.zc_value, chopprefix, sizeof (zc.zc_value));
+       free(cp);
        if (!zfs_name_valid(zc.zc_value, ZFS_TYPE_SNAPSHOT)) {
                zcmd_free_nvlists(&zc);
                return (zfs_error(hdl, EZFS_INVALIDNAME, errbuf));
@@ -1653,7 +2662,7 @@ zfs_receive_one(libzfs_handle_t *hdl, int infd, const char *tosnap,
         * Determine the name of the origin snapshot, store in zc_string.
         */
        if (drrb->drr_flags & DRR_FLAG_CLONE) {
-               if (guid_to_name(hdl, tosnap,
+               if (guid_to_name(hdl, zc.zc_value,
                    drrb->drr_fromguid, zc.zc_string) != 0) {
                        zcmd_free_nvlists(&zc);
                        zfs_error_aux(hdl, dgettext(TEXT_DOMAIN,
@@ -1661,11 +2670,11 @@ zfs_receive_one(libzfs_handle_t *hdl, int infd, const char *tosnap,
                            zc.zc_value);
                        return (zfs_error(hdl, EZFS_NOENT, errbuf));
                }
-               if (flags.verbose)
+               if (flags->verbose)
                        (void) printf("found clone origin %s\n", zc.zc_string);
        }
 
-       stream_wantsnewfs = (drrb->drr_fromguid == NULL ||
+       stream_wantsnewfs = (drrb->drr_fromguid == 0 ||
            (drrb->drr_flags & DRR_FLAG_CLONE));
 
        if (stream_wantsnewfs) {
@@ -1684,7 +2693,7 @@ zfs_receive_one(libzfs_handle_t *hdl, int infd, const char *tosnap,
                    !zfs_dataset_exists(hdl, zc.zc_name, ZFS_TYPE_DATASET)) {
                        char suffix[ZFS_MAXNAMELEN];
                        (void) strcpy(suffix, strrchr(zc.zc_value, '/'));
-                       if (guid_to_name(hdl, tosnap, parent_snapguid,
+                       if (guid_to_name(hdl, zc.zc_name, parent_snapguid,
                            zc.zc_value) == 0) {
                                *strchr(zc.zc_value, '@') = '\0';
                                (void) strcat(zc.zc_value, suffix);
@@ -1701,10 +2710,17 @@ zfs_receive_one(libzfs_handle_t *hdl, int infd, const char *tosnap,
                (void) strcpy(zc.zc_name, zc.zc_value);
                *strchr(zc.zc_name, '@') = '\0';
 
-               if (!zfs_dataset_exists(hdl, zc.zc_name, ZFS_TYPE_DATASET)) {
+               /*
+                * If the exact receive path was specified and this is the
+                * topmost path in the stream, then if the fs does not exist we
+                * should look no further.
+                */
+               if ((flags->isprefix || (*(chopprefix = drrb->drr_toname +
+                   strlen(sendfs)) != '\0' && *chopprefix != '@')) &&
+                   !zfs_dataset_exists(hdl, zc.zc_name, ZFS_TYPE_DATASET)) {
                        char snap[ZFS_MAXNAMELEN];
                        (void) strcpy(snap, strchr(zc.zc_value, '@'));
-                       if (guid_to_name(hdl, tosnap, drrb->drr_fromguid,
+                       if (guid_to_name(hdl, zc.zc_name, drrb->drr_fromguid,
                            zc.zc_value) == 0) {
                                *strchr(zc.zc_value, '@') = '\0';
                                (void) strcat(zc.zc_value, snap);
@@ -1717,6 +2733,7 @@ zfs_receive_one(libzfs_handle_t *hdl, int infd, const char *tosnap,
 
        if (zfs_dataset_exists(hdl, zc.zc_name, ZFS_TYPE_DATASET)) {
                zfs_handle_t *zhp;
+
                /*
                 * Destination fs exists.  Therefore this should either
                 * be an incremental, or the stream specifies a new fs
@@ -1724,9 +2741,8 @@ zfs_receive_one(libzfs_handle_t *hdl, int infd, const char *tosnap,
                 * away (and have therefore specified -F and removed any
                 * snapshots).
                 */
-
                if (stream_wantsnewfs) {
-                       if (!flags.force) {
+                       if (!flags->force) {
                                zcmd_free_nvlists(&zc);
                                zfs_error_aux(hdl, dgettext(TEXT_DOMAIN,
                                    "destination '%s' exists\n"
@@ -1762,21 +2778,23 @@ zfs_receive_one(libzfs_handle_t *hdl, int infd, const char *tosnap,
                        return (zfs_error(hdl, EZFS_EXISTS, errbuf));
                }
 
-               if (!flags.dryrun && zhp->zfs_type == ZFS_TYPE_FILESYSTEM &&
+               if (!flags->dryrun && zhp->zfs_type == ZFS_TYPE_FILESYSTEM &&
                    stream_wantsnewfs) {
                        /* We can't do online recv in this case */
                        clp = changelist_gather(zhp, ZFS_PROP_NAME, 0, 0);
                        if (clp == NULL) {
+                               zfs_close(zhp);
                                zcmd_free_nvlists(&zc);
                                return (-1);
                        }
                        if (changelist_prefix(clp) != 0) {
                                changelist_free(clp);
+                               zfs_close(zhp);
                                zcmd_free_nvlists(&zc);
                                return (-1);
                        }
                }
-               if (!flags.dryrun && zhp->zfs_type == ZFS_TYPE_VOLUME &&
+               if (!flags->dryrun && zhp->zfs_type == ZFS_TYPE_VOLUME &&
                    zvol_remove_link(hdl, zhp->zfs_name) != 0) {
                        zfs_close(zhp);
                        zcmd_free_nvlists(&zc);
@@ -1805,7 +2823,7 @@ zfs_receive_one(libzfs_handle_t *hdl, int infd, const char *tosnap,
                 */
                *cp = '\0';
 
-               if (flags.isprefix && !flags.dryrun &&
+               if (flags->isprefix && !flags->istail && !flags->dryrun &&
                    create_parents(hdl, zc.zc_value, strlen(tosnap)) != 0) {
                        zcmd_free_nvlists(&zc);
                        return (zfs_error(hdl, EZFS_BADRESTORE, errbuf));
@@ -1816,35 +2834,75 @@ zfs_receive_one(libzfs_handle_t *hdl, int infd, const char *tosnap,
 
        zc.zc_begin_record = drr_noswap->drr_u.drr_begin;
        zc.zc_cookie = infd;
-       zc.zc_guid = flags.force;
-       if (flags.verbose) {
+       zc.zc_guid = flags->force;
+       if (flags->verbose) {
                (void) printf("%s %s stream of %s into %s\n",
-                   flags.dryrun ? "would receive" : "receiving",
+                   flags->dryrun ? "would receive" : "receiving",
                    drrb->drr_fromguid ? "incremental" : "full",
                    drrb->drr_toname, zc.zc_value);
                (void) fflush(stdout);
        }
 
-       if (flags.dryrun) {
+       if (flags->dryrun) {
                zcmd_free_nvlists(&zc);
-               return (recv_skip(hdl, infd, flags.byteswap));
+               return (recv_skip(hdl, infd, flags->byteswap));
        }
 
+       zc.zc_nvlist_dst = (uint64_t)(uintptr_t)prop_errbuf;
+       zc.zc_nvlist_dst_size = sizeof (prop_errbuf);
+       zc.zc_cleanup_fd = cleanup_fd;
+       zc.zc_action_handle = *action_handlep;
+
        err = ioctl_err = zfs_ioctl(hdl, ZFS_IOC_RECV, &zc);
        ioctl_errno = errno;
+       prop_errflags = (zprop_errflags_t)zc.zc_obj;
+
+       if (err == 0) {
+               nvlist_t *prop_errors;
+               VERIFY(0 == nvlist_unpack((void *)(uintptr_t)zc.zc_nvlist_dst,
+                   zc.zc_nvlist_dst_size, &prop_errors, 0));
+
+               nvpair_t *prop_err = NULL;
+
+               while ((prop_err = nvlist_next_nvpair(prop_errors,
+                   prop_err)) != NULL) {
+                       char tbuf[1024];
+                       zfs_prop_t prop;
+                       int intval;
+
+                       prop = zfs_name_to_prop(nvpair_name(prop_err));
+                       (void) nvpair_value_int32(prop_err, &intval);
+                       if (strcmp(nvpair_name(prop_err),
+                           ZPROP_N_MORE_ERRORS) == 0) {
+                               trunc_prop_errs(intval);
+                               break;
+                       } else {
+                               (void) snprintf(tbuf, sizeof (tbuf),
+                                   dgettext(TEXT_DOMAIN,
+                                   "cannot receive %s property on %s"),
+                                   nvpair_name(prop_err), zc.zc_name);
+                               zfs_setprop_error(hdl, prop, intval, tbuf);
+                       }
+               }
+               nvlist_free(prop_errors);
+       }
+
+       zc.zc_nvlist_dst = 0;
+       zc.zc_nvlist_dst_size = 0;
        zcmd_free_nvlists(&zc);
 
        if (err == 0 && snapprops_nvlist) {
-               zfs_cmd_t zc2 = { 0 };
+               zfs_cmd_t zc2 = { "\0", "\0", "\0", "\0", 0 };
 
                (void) strcpy(zc2.zc_name, zc.zc_value);
+               zc2.zc_cookie = B_TRUE; /* received */
                if (zcmd_write_src_nvlist(hdl, &zc2, snapprops_nvlist) == 0) {
                        (void) zfs_ioctl(hdl, ZFS_IOC_SET_PROP, &zc2);
                        zcmd_free_nvlists(&zc2);
                }
        }
 
-       if (err && (ioctl_errno == ENOENT || ioctl_errno == ENODEV)) {
+       if (err && (ioctl_errno == ENOENT || ioctl_errno == EEXIST)) {
                /*
                 * It may be that this snapshot already exists,
                 * in which case we want to consume & ignore it
@@ -1852,7 +2910,7 @@ zfs_receive_one(libzfs_handle_t *hdl, int infd, const char *tosnap,
                 */
                avl_tree_t *local_avl;
                nvlist_t *local_nv, *fs;
-               char *cp = strchr(zc.zc_value, '@');
+               cp = strchr(zc.zc_value, '@');
 
                /*
                 * XXX Do this faster by just iterating over snaps in
@@ -1860,7 +2918,7 @@ zfs_receive_one(libzfs_handle_t *hdl, int infd, const char *tosnap,
                 * get a strange "does not exist" error message.
                 */
                *cp = '\0';
-               if (gather_nvlist(hdl, zc.zc_value, NULL, NULL,
+               if (gather_nvlist(hdl, zc.zc_value, NULL, NULL, B_FALSE,
                    &local_nv, &local_avl) == 0) {
                        *cp = '@';
                        fs = fsavl_find(local_avl, drrb->drr_toguid, NULL);
@@ -1868,18 +2926,17 @@ zfs_receive_one(libzfs_handle_t *hdl, int infd, const char *tosnap,
                        nvlist_free(local_nv);
 
                        if (fs != NULL) {
-                               if (flags.verbose) {
+                               if (flags->verbose) {
                                        (void) printf("snap %s already exists; "
                                            "ignoring\n", zc.zc_value);
                                }
-                               ioctl_err = recv_skip(hdl, infd,
-                                   flags.byteswap);
+                               err = ioctl_err = recv_skip(hdl, infd,
+                                   flags->byteswap);
                        }
                }
                *cp = '@';
        }
 
-
        if (ioctl_err != 0) {
                switch (ioctl_errno) {
                case ENODEV:
@@ -1918,17 +2975,25 @@ zfs_receive_one(libzfs_handle_t *hdl, int infd, const char *tosnap,
                            "invalid stream (checksum mismatch)"));
                        (void) zfs_error(hdl, EZFS_BADSTREAM, errbuf);
                        break;
+               case ENOTSUP:
+                       zfs_error_aux(hdl, dgettext(TEXT_DOMAIN,
+                           "pool must be upgraded to receive this stream."));
+                       (void) zfs_error(hdl, EZFS_BADVERSION, errbuf);
+                       break;
+               case EDQUOT:
+                       zfs_error_aux(hdl, dgettext(TEXT_DOMAIN,
+                           "destination %s space quota exceeded"), zc.zc_name);
+                       (void) zfs_error(hdl, EZFS_NOSPC, errbuf);
+                       break;
                default:
                        (void) zfs_standard_error(hdl, ioctl_errno, errbuf);
                }
        }
 
        /*
-        * Mount or recreate the /dev links for the target filesystem
-        * (if created, or if we tore them down to do an incremental
-        * restore), and the /dev links for the new snapshot (if
-        * created). Also mount any children of the target filesystem
-        * if we did an incremental receive.
+        * Mount the target filesystem (if created).  Also mount any
+        * children of the target filesystem if we did a replication
+        * receive (indicated by stream_avl being non-NULL).
         */
        cp = strchr(zc.zc_value, '@');
        if (cp && (ioctl_err == 0 || !newfs)) {
@@ -1944,7 +3009,7 @@ zfs_receive_one(libzfs_handle_t *hdl, int infd, const char *tosnap,
                                if (err == 0 && ioctl_err == 0)
                                        err = zvol_create_link(hdl,
                                            zc.zc_value);
-                       } else if (newfs) {
+                       } else if (newfs || stream_avl) {
                                /*
                                 * Track the first/top of hierarchy fs,
                                 * for mounting and sharing later.
@@ -1962,10 +3027,25 @@ zfs_receive_one(libzfs_handle_t *hdl, int infd, const char *tosnap,
                changelist_free(clp);
        }
 
+       if (prop_errflags & ZPROP_ERR_NOCLEAR) {
+               (void) fprintf(stderr, dgettext(TEXT_DOMAIN, "Warning: "
+                   "failed to clear unreceived properties on %s"),
+                   zc.zc_name);
+               (void) fprintf(stderr, "\n");
+       }
+       if (prop_errflags & ZPROP_ERR_NORESTORE) {
+               (void) fprintf(stderr, dgettext(TEXT_DOMAIN, "Warning: "
+                   "failed to restore original properties on %s"),
+                   zc.zc_name);
+               (void) fprintf(stderr, "\n");
+       }
+
        if (err || ioctl_err)
                return (-1);
 
-       if (flags.verbose) {
+       *action_handlep = zc.zc_action_handle;
+
+       if (flags->verbose) {
                char buf1[64];
                char buf2[64];
                uint64_t bytes = zc.zc_cookie;
@@ -1983,19 +3063,22 @@ 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,
-    int infd, avl_tree_t *stream_avl, char **top_zfs)
+zfs_receive_impl(libzfs_handle_t *hdl, const char *tosnap, 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)
 {
        int err;
        dmu_replay_record_t drr, drr_noswap;
        struct drr_begin *drrb = &drr.drr_u.drr_begin;
        char errbuf[1024];
-       zio_cksum_t zcksum = { 0 };
+       zio_cksum_t zcksum = { { 0 } };
+       uint64_t featureflags;
+       int hdrtype;
 
        (void) snprintf(errbuf, sizeof (errbuf), dgettext(TEXT_DOMAIN,
            "cannot receive"));
 
-       if (flags.isprefix &&
+       if (flags->isprefix &&
            !zfs_dataset_exists(hdl, tosnap, ZFS_TYPE_DATASET)) {
                zfs_error_aux(hdl, dgettext(TEXT_DOMAIN, "specified fs "
                    "(%s) does not exist"), tosnap);
@@ -2015,7 +3098,7 @@ zfs_receive_impl(libzfs_handle_t *hdl, const char *tosnap, recvflags_t flags,
        /* the kernel needs the non-byteswapped begin record */
        drr_noswap = drr;
 
-       flags.byteswap = B_FALSE;
+       flags->byteswap = B_FALSE;
        if (drrb->drr_magic == BSWAP_64(DMU_BACKUP_MAGIC)) {
                /*
                 * We computed the checksum in the wrong byteorder in
@@ -2023,12 +3106,12 @@ zfs_receive_impl(libzfs_handle_t *hdl, const char *tosnap, recvflags_t flags,
                 */
                bzero(&zcksum, sizeof (zio_cksum_t));
                fletcher_4_incremental_byteswap(&drr, sizeof (drr), &zcksum);
-               flags.byteswap = B_TRUE;
+               flags->byteswap = B_TRUE;
 
                drr.drr_type = BSWAP_32(drr.drr_type);
                drr.drr_payloadlen = BSWAP_32(drr.drr_payloadlen);
                drrb->drr_magic = BSWAP_64(drrb->drr_magic);
-               drrb->drr_version = BSWAP_64(drrb->drr_version);
+               drrb->drr_versioninfo = BSWAP_64(drrb->drr_versioninfo);
                drrb->drr_creation_time = BSWAP_64(drrb->drr_creation_time);
                drrb->drr_type = BSWAP_32(drrb->drr_type);
                drrb->drr_flags = BSWAP_32(drrb->drr_flags);
@@ -2042,23 +3125,45 @@ zfs_receive_impl(libzfs_handle_t *hdl, const char *tosnap, recvflags_t flags,
                return (zfs_error(hdl, EZFS_BADSTREAM, errbuf));
        }
 
+       featureflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo);
+       hdrtype = DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo);
+
+       if (!DMU_STREAM_SUPPORTED(featureflags) ||
+           (hdrtype != DMU_SUBSTREAM && hdrtype != DMU_COMPOUNDSTREAM)) {
+               zfs_error_aux(hdl, dgettext(TEXT_DOMAIN,
+                   "stream has unsupported feature, feature flags = %lx"),
+                   featureflags);
+               return (zfs_error(hdl, EZFS_BADSTREAM, errbuf));
+       }
+
        if (strchr(drrb->drr_toname, '@') == NULL) {
                zfs_error_aux(hdl, dgettext(TEXT_DOMAIN, "invalid "
                    "stream (bad snapshot name)"));
                return (zfs_error(hdl, EZFS_BADSTREAM, errbuf));
        }
 
-       if (drrb->drr_version == DMU_BACKUP_STREAM_VERSION) {
+       if (DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) == DMU_SUBSTREAM) {
+               char nonpackage_sendfs[ZFS_MAXNAMELEN];
+               if (sendfs == NULL) {
+                       /*
+                        * We were not called from zfs_receive_package(). Get
+                        * the fs specified by 'zfs send'.
+                        */
+                       char *cp;
+                       (void) strlcpy(nonpackage_sendfs,
+                           drr.drr_u.drr_begin.drr_toname, ZFS_MAXNAMELEN);
+                       if ((cp = strchr(nonpackage_sendfs, '@')) != NULL)
+                               *cp = '\0';
+                       sendfs = nonpackage_sendfs;
+               }
                return (zfs_receive_one(hdl, infd, tosnap, flags,
-                   &drr, &drr_noswap, stream_avl, top_zfs));
-       } else if (drrb->drr_version == DMU_BACKUP_HEADER_VERSION) {
-               return (zfs_receive_package(hdl, infd, tosnap, flags,
-                   &drr, &zcksum, top_zfs));
+                   &drr, &drr_noswap, sendfs, stream_nv, stream_avl,
+                   top_zfs, cleanup_fd, action_handlep));
        } else {
-               zfs_error_aux(hdl, dgettext(TEXT_DOMAIN,
-                   "stream is unsupported version %llu"),
-                   drrb->drr_version);
-               return (zfs_error(hdl, EZFS_BADSTREAM, errbuf));
+               assert(DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) ==
+                   DMU_COMPOUNDSTREAM);
+               return (zfs_receive_package(hdl, infd, tosnap, flags,
+                   &drr, &zcksum, top_zfs, cleanup_fd, action_handlep));
        }
 }
 
@@ -2069,15 +3174,23 @@ 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, recvflags_t *flags,
     int infd, avl_tree_t *stream_avl)
 {
        char *top_zfs = NULL;
        int err;
+       int cleanup_fd;
+       uint64_t action_handle = 0;
+
+       cleanup_fd = open(ZFS_DEV, O_RDWR);
+       VERIFY(cleanup_fd >= 0);
+
+       err = zfs_receive_impl(hdl, tosnap, flags, infd, NULL, NULL,
+           stream_avl, &top_zfs, cleanup_fd, &action_handle);
 
-       err = zfs_receive_impl(hdl, tosnap, flags, infd, stream_avl, &top_zfs);
+       VERIFY(0 == close(cleanup_fd));
 
-       if (err == 0 && !flags.nomount && top_zfs) {
+       if (err == 0 && !flags->nomount && top_zfs) {
                zfs_handle_t *zhp;
                prop_changelist_t *clp;