1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  23  * Copyright (c) 2011, 2015 by Delphix. All rights reserved.
  24  * Copyright (c) 2014, Joyent, Inc. All rights reserved.
  25  * Copyright 2014 HybridCluster. All rights reserved.
  26  * Copyright 2017 Nexenta Systems, Inc. All rights reserved.
  27  * Copyright 2016 RackTop Systems.
  28  * Copyright (c) 2014 Integros [integros.com]
  29  */
  30 
  31 #include <sys/dmu.h>
  32 #include <sys/dmu_impl.h>
  33 #include <sys/dmu_tx.h>
  34 #include <sys/dbuf.h>
  35 #include <sys/dnode.h>
  36 #include <sys/zfs_context.h>
  37 #include <sys/dmu_objset.h>
  38 #include <sys/dmu_traverse.h>
  39 #include <sys/dsl_dataset.h>
  40 #include <sys/dsl_dir.h>
  41 #include <sys/dsl_prop.h>
  42 #include <sys/dsl_pool.h>
  43 #include <sys/dsl_synctask.h>
  44 #include <sys/zfs_ioctl.h>
  45 #include <sys/zap.h>
  46 #include <sys/zio_checksum.h>
  47 #include <sys/zfs_znode.h>
  48 #include <zfs_fletcher.h>
  49 #include <sys/avl.h>
  50 #include <sys/ddt.h>
  51 #include <sys/zfs_onexit.h>
  52 #include <sys/dmu_send.h>
  53 #include <sys/dsl_destroy.h>
  54 #include <sys/blkptr.h>
  55 #include <sys/dsl_bookmark.h>
  56 #include <sys/zfeature.h>
  57 #include <sys/autosnap.h>
  58 #include <sys/bqueue.h>
  59 
  60 #include "zfs_errno.h"
  61 
  62 /* Set this tunable to TRUE to replace corrupt data with 0x2f5baddb10c */
  63 int zfs_send_corrupt_data = B_FALSE;
  64 int zfs_send_queue_length = 16 * 1024 * 1024;
  65 int zfs_recv_queue_length = 16 * 1024 * 1024;
  66 /* Set this tunable to FALSE to disable setting of DRR_FLAG_FREERECORDS */
  67 int zfs_send_set_freerecords_bit = B_TRUE;
  68 
  69 static char *dmu_recv_tag = "dmu_recv_tag";
  70 const char *recv_clone_name = "%recv";
  71 
  72 #define BP_SPAN(datablkszsec, indblkshift, level) \
  73         (((uint64_t)datablkszsec) << (SPA_MINBLOCKSHIFT + \
  74         (level) * (indblkshift - SPA_BLKPTRSHIFT)))
  75 
  76 static void byteswap_record(dmu_replay_record_t *drr);
  77 
  78 struct send_thread_arg {
  79         bqueue_t        q;
  80         dsl_dataset_t   *ds;            /* Dataset to traverse */
  81         uint64_t        fromtxg;        /* Traverse from this txg */
  82         int             flags;          /* flags to pass to traverse_dataset */
  83         int             error_code;
  84         boolean_t       cancel;
  85         zbookmark_phys_t resume;
  86 };
  87 
  88 struct send_block_record {
  89         boolean_t               eos_marker; /* Marks the end of the stream */
  90         blkptr_t                bp;
  91         zbookmark_phys_t        zb;
  92         uint8_t                 indblkshift;
  93         uint16_t                datablkszsec;
  94         bqueue_node_t           ln;
  95 };
  96 
  97 static int
  98 dump_bytes(dmu_sendarg_t *dsp, void *buf, int len)
  99 {
 100         dsl_dataset_t *ds = dmu_objset_ds(dsp->dsa_os);
 101         ssize_t resid; /* have to get resid to get detailed errno */
 102 
 103         /*
 104          * The code does not rely on this (len being a multiple of 8).  We keep
 105          * this assertion because of the corresponding assertion in
 106          * receive_read().  Keeping this assertion ensures that we do not
 107          * inadvertently break backwards compatibility (causing the assertion
 108          * in receive_read() to trigger on old software).
 109          *
 110          * Removing the assertions could be rolled into a new feature that uses
 111          * data that isn't 8-byte aligned; if the assertions were removed, a
 112          * feature flag would have to be added.
 113          */
 114 
 115         ASSERT0(len % 8);
 116         ASSERT(buf != NULL);
 117 
 118         dsp->dsa_err = 0;
 119         if (!dsp->sendsize) {
 120                 /* if vp is NULL, then the send is from krrp */
 121                 if (dsp->dsa_vp != NULL) {
 122                         dsp->dsa_err = vn_rdwr(UIO_WRITE, dsp->dsa_vp,
 123                             (caddr_t)buf, len,
 124                             0, UIO_SYSSPACE, FAPPEND, RLIM64_INFINITY,
 125                             CRED(), &resid);
 126                 } else {
 127                         ASSERT(dsp->dsa_krrp_task != NULL);
 128                         dsp->dsa_err = dmu_krrp_buffer_write(buf, len,
 129                             dsp->dsa_krrp_task);
 130                 }
 131         }
 132         mutex_enter(&ds->ds_sendstream_lock);
 133         *dsp->dsa_off += len;
 134         mutex_exit(&ds->ds_sendstream_lock);
 135 
 136         return (dsp->dsa_err);
 137 }
 138 
 139 static int
 140 dump_bytes_with_checksum(dmu_sendarg_t *dsp, void *buf, int len)
 141 {
 142         if (!dsp->sendsize && (dsp->dsa_krrp_task == NULL ||
 143             dsp->dsa_krrp_task->buffer_args.force_cksum)) {
 144                 (void) fletcher_4_incremental_native(buf, len, &dsp->dsa_zc);
 145         }
 146 
 147         return (dump_bytes(dsp, buf, len));
 148 }
 149 
 150 /*
 151  * For all record types except BEGIN, fill in the checksum (overlaid in
 152  * drr_u.drr_checksum.drr_checksum).  The checksum verifies everything
 153  * up to the start of the checksum itself.
 154  */
 155 static int
 156 dump_record(dmu_sendarg_t *dsp, void *payload, int payload_len)
 157 {
 158         boolean_t do_checksum = (dsp->dsa_krrp_task == NULL ||
 159             dsp->dsa_krrp_task->buffer_args.force_cksum);
 160 
 161         ASSERT3U(offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
 162             ==, sizeof (dmu_replay_record_t) - sizeof (zio_cksum_t));
 163 
 164         if (dsp->dsa_drr->drr_type == DRR_BEGIN) {
 165                 dsp->dsa_sent_begin = B_TRUE;
 166         }
 167 
 168         if (dsp->dsa_drr->drr_type == DRR_END) {
 169                 dsp->dsa_sent_end = B_TRUE;
 170         }
 171 
 172         if (!dsp->sendsize && do_checksum) {
 173                 (void) fletcher_4_incremental_native(dsp->dsa_drr,
 174                     offsetof(dmu_replay_record_t,
 175                     drr_u.drr_checksum.drr_checksum),
 176                     &dsp->dsa_zc);
 177                 if (dsp->dsa_drr->drr_type != DRR_BEGIN) {
 178                         ASSERT(ZIO_CHECKSUM_IS_ZERO(&dsp->dsa_drr->drr_u.
 179                             drr_checksum.drr_checksum));
 180                         dsp->dsa_drr->drr_u.drr_checksum.drr_checksum =
 181                             dsp->dsa_zc;
 182                 }
 183 
 184                 (void) fletcher_4_incremental_native(&dsp->dsa_drr->
 185                     drr_u.drr_checksum.drr_checksum,
 186                     sizeof (zio_cksum_t), &dsp->dsa_zc);
 187         }
 188 
 189         if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)) != 0)
 190                 return (SET_ERROR(EINTR));
 191         if (payload_len != 0) {
 192                 if (dump_bytes_with_checksum(dsp, payload, payload_len) != 0)
 193                         return (SET_ERROR(EINTR));
 194         }
 195         return (0);
 196 }
 197 
 198 /*
 199  * Fill in the drr_free struct, or perform aggregation if the previous record is
 200  * also a free record, and the two are adjacent.
 201  *
 202  * Note that we send free records even for a full send, because we want to be
 203  * able to receive a full send as a clone, which requires a list of all the free
 204  * and freeobject records that were generated on the source.
 205  */
 206 static int
 207 dump_free(dmu_sendarg_t *dsp, uint64_t object, uint64_t offset,
 208     uint64_t length)
 209 {
 210         struct drr_free *drrf = &(dsp->dsa_drr->drr_u.drr_free);
 211 
 212         /*
 213          * When we receive a free record, dbuf_free_range() assumes
 214          * that the receiving system doesn't have any dbufs in the range
 215          * being freed.  This is always true because there is a one-record
 216          * constraint: we only send one WRITE record for any given
 217          * object,offset.  We know that the one-record constraint is
 218          * true because we always send data in increasing order by
 219          * object,offset.
 220          *
 221          * If the increasing-order constraint ever changes, we should find
 222          * another way to assert that the one-record constraint is still
 223          * satisfied.
 224          */
 225         ASSERT(object > dsp->dsa_last_data_object ||
 226             (object == dsp->dsa_last_data_object &&
 227             offset > dsp->dsa_last_data_offset));
 228 
 229         if (length != -1ULL && offset + length < offset)
 230                 length = -1ULL;
 231 
 232         /*
 233          * If there is a pending op, but it's not PENDING_FREE, push it out,
 234          * since free block aggregation can only be done for blocks of the
 235          * same type (i.e., DRR_FREE records can only be aggregated with
 236          * other DRR_FREE records.  DRR_FREEOBJECTS records can only be
 237          * aggregated with other DRR_FREEOBJECTS records.
 238          */
 239         if (dsp->dsa_pending_op != PENDING_NONE &&
 240             dsp->dsa_pending_op != PENDING_FREE) {
 241                 if (dump_record(dsp, NULL, 0) != 0)
 242                         return (SET_ERROR(EINTR));
 243                 dsp->dsa_pending_op = PENDING_NONE;
 244         }
 245 
 246         if (dsp->dsa_pending_op == PENDING_FREE) {
 247                 /*
 248                  * There should never be a PENDING_FREE if length is -1
 249                  * (because dump_dnode is the only place where this
 250                  * function is called with a -1, and only after flushing
 251                  * any pending record).
 252                  */
 253                 ASSERT(length != -1ULL);
 254                 /*
 255                  * Check to see whether this free block can be aggregated
 256                  * with pending one.
 257                  */
 258                 if (drrf->drr_object == object && drrf->drr_offset +
 259                     drrf->drr_length == offset) {
 260                         drrf->drr_length += length;
 261                         return (0);
 262                 } else {
 263                         /* not a continuation.  Push out pending record */
 264                         if (dump_record(dsp, NULL, 0) != 0)
 265                                 return (SET_ERROR(EINTR));
 266                         dsp->dsa_pending_op = PENDING_NONE;
 267                 }
 268         }
 269         /* create a FREE record and make it pending */
 270         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 271         dsp->dsa_drr->drr_type = DRR_FREE;
 272         drrf->drr_object = object;
 273         drrf->drr_offset = offset;
 274         drrf->drr_length = length;
 275         drrf->drr_toguid = dsp->dsa_toguid;
 276         if (length == -1ULL) {
 277                 if (dump_record(dsp, NULL, 0) != 0)
 278                         return (SET_ERROR(EINTR));
 279         } else {
 280                 dsp->dsa_pending_op = PENDING_FREE;
 281         }
 282 
 283         return (0);
 284 }
 285 
 286 static int
 287 dump_write(dmu_sendarg_t *dsp, dmu_object_type_t type,
 288     uint64_t object, uint64_t offset, int lsize, int psize, const blkptr_t *bp,
 289     void *data)
 290 {
 291         uint64_t payload_size;
 292         struct drr_write *drrw = &(dsp->dsa_drr->drr_u.drr_write);
 293 
 294         /*
 295          * We send data in increasing object, offset order.
 296          * See comment in dump_free() for details.
 297          */
 298         ASSERT(object > dsp->dsa_last_data_object ||
 299             (object == dsp->dsa_last_data_object &&
 300             offset > dsp->dsa_last_data_offset));
 301         dsp->dsa_last_data_object = object;
 302         dsp->dsa_last_data_offset = offset + lsize - 1;
 303 
 304         /*
 305          * If there is any kind of pending aggregation (currently either
 306          * a grouping of free objects or free blocks), push it out to
 307          * the stream, since aggregation can't be done across operations
 308          * of different types.
 309          */
 310         if (dsp->dsa_pending_op != PENDING_NONE) {
 311                 if (dump_record(dsp, NULL, 0) != 0)
 312                         return (SET_ERROR(EINTR));
 313                 dsp->dsa_pending_op = PENDING_NONE;
 314         }
 315         /* write a WRITE record */
 316         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 317         dsp->dsa_drr->drr_type = DRR_WRITE;
 318         drrw->drr_object = object;
 319         drrw->drr_type = type;
 320         drrw->drr_offset = offset;
 321         drrw->drr_toguid = dsp->dsa_toguid;
 322         drrw->drr_logical_size = lsize;
 323 
 324         /* only set the compression fields if the buf is compressed */
 325         if (lsize != psize) {
 326                 ASSERT(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_COMPRESSED);
 327                 ASSERT(!BP_IS_EMBEDDED(bp));
 328                 ASSERT(!BP_SHOULD_BYTESWAP(bp));
 329                 ASSERT(!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)));
 330                 ASSERT3U(BP_GET_COMPRESS(bp), !=, ZIO_COMPRESS_OFF);
 331                 ASSERT3S(psize, >, 0);
 332                 ASSERT3S(lsize, >=, psize);
 333 
 334                 drrw->drr_compressiontype = BP_GET_COMPRESS(bp);
 335                 drrw->drr_compressed_size = psize;
 336                 payload_size = drrw->drr_compressed_size;
 337         } else {
 338                 payload_size = drrw->drr_logical_size;
 339         }
 340 
 341         if (bp == NULL || BP_IS_EMBEDDED(bp)) {
 342                 /*
 343                  * There's no pre-computed checksum for partial-block
 344                  * writes or embedded BP's, so (like
 345                  * fletcher4-checkummed blocks) userland will have to
 346                  * compute a dedup-capable checksum itself.
 347                  */
 348                 drrw->drr_checksumtype = ZIO_CHECKSUM_OFF;
 349         } else {
 350                 drrw->drr_checksumtype = BP_GET_CHECKSUM(bp);
 351                 if (zio_checksum_table[drrw->drr_checksumtype].ci_flags &
 352                     ZCHECKSUM_FLAG_DEDUP)
 353                         drrw->drr_checksumflags |= DRR_CHECKSUM_DEDUP;
 354                 DDK_SET_LSIZE(&drrw->drr_key, BP_GET_LSIZE(bp));
 355                 DDK_SET_PSIZE(&drrw->drr_key, BP_GET_PSIZE(bp));
 356                 DDK_SET_COMPRESS(&drrw->drr_key, BP_GET_COMPRESS(bp));
 357                 drrw->drr_key.ddk_cksum = bp->blk_cksum;
 358         }
 359 
 360         if (dump_record(dsp, data, payload_size) != 0)
 361                 return (SET_ERROR(EINTR));
 362         return (0);
 363 }
 364 
 365 static int
 366 dump_write_embedded(dmu_sendarg_t *dsp, uint64_t object, uint64_t offset,
 367     int blksz, const blkptr_t *bp)
 368 {
 369         char buf[BPE_PAYLOAD_SIZE];
 370         struct drr_write_embedded *drrw =
 371             &(dsp->dsa_drr->drr_u.drr_write_embedded);
 372 
 373         if (dsp->dsa_pending_op != PENDING_NONE) {
 374                 if (dump_record(dsp, NULL, 0) != 0)
 375                         return (EINTR);
 376                 dsp->dsa_pending_op = PENDING_NONE;
 377         }
 378 
 379         ASSERT(BP_IS_EMBEDDED(bp));
 380 
 381         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 382         dsp->dsa_drr->drr_type = DRR_WRITE_EMBEDDED;
 383         drrw->drr_object = object;
 384         drrw->drr_offset = offset;
 385         drrw->drr_length = blksz;
 386         drrw->drr_toguid = dsp->dsa_toguid;
 387         drrw->drr_compression = BP_GET_COMPRESS(bp);
 388         drrw->drr_etype = BPE_GET_ETYPE(bp);
 389         drrw->drr_lsize = BPE_GET_LSIZE(bp);
 390         drrw->drr_psize = BPE_GET_PSIZE(bp);
 391 
 392         decode_embedded_bp_compressed(bp, buf);
 393 
 394         if (dump_record(dsp, buf, P2ROUNDUP(drrw->drr_psize, 8)) != 0)
 395                 return (EINTR);
 396         return (0);
 397 }
 398 
 399 static int
 400 dump_spill(dmu_sendarg_t *dsp, uint64_t object,
 401     const blkptr_t *bp, const zbookmark_phys_t *zb)
 402 {
 403         int rc = 0;
 404         struct drr_spill *drrs = &(dsp->dsa_drr->drr_u.drr_spill);
 405         enum arc_flags aflags = ARC_FLAG_WAIT;
 406         int blksz = BP_GET_LSIZE(bp);
 407         arc_buf_t *abuf;
 408 
 409         if (dsp->dsa_pending_op != PENDING_NONE) {
 410                 if (dump_record(dsp, NULL, 0) != 0)
 411                         return (SET_ERROR(EINTR));
 412                 dsp->dsa_pending_op = PENDING_NONE;
 413         }
 414 
 415         /* write a SPILL record */
 416         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 417         dsp->dsa_drr->drr_type = DRR_SPILL;
 418         drrs->drr_object = object;
 419         drrs->drr_length = blksz;
 420         drrs->drr_toguid = dsp->dsa_toguid;
 421 
 422         if (dump_record(dsp, NULL, 0))
 423                 return (SET_ERROR(EINTR));
 424 
 425         /*
 426          * if dsa_krrp task is not NULL, then the send is from krrp and we can
 427          * try to bypass copying data to an intermediate buffer.
 428          */
 429         if (!dsp->sendsize && dsp->dsa_krrp_task != NULL) {
 430                 rc = dmu_krrp_direct_arc_read(dsp->dsa_os->os_spa,
 431                     dsp->dsa_krrp_task, &dsp->dsa_zc, bp);
 432                 /*
 433                  * rc == 0 means that we successfully copy
 434                  * the data directly from ARC to krrp buffer
 435                  * rc != 0 && rc != EINTR means that we cannot
 436                  * zerocopy the data and need to use slow-path
 437                  */
 438                 if (rc == 0 || rc == EINTR)
 439                         return (rc);
 440 
 441                 ASSERT3U(rc, ==, ENODATA);
 442         }
 443 
 444         if (arc_read(NULL, dsp->dsa_os->os_spa, bp, arc_getbuf_func, &abuf,
 445             ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
 446             &aflags, zb) != 0)
 447                 return (SET_ERROR(EIO));
 448 
 449         rc = dump_bytes_with_checksum(dsp, abuf->b_data, blksz);
 450         arc_buf_destroy(abuf, &abuf);
 451         if (rc != 0)
 452                 return (SET_ERROR(EINTR));
 453 
 454         return (0);
 455 }
 456 
 457 static int
 458 dump_freeobjects(dmu_sendarg_t *dsp, uint64_t firstobj, uint64_t numobjs)
 459 {
 460         struct drr_freeobjects *drrfo = &(dsp->dsa_drr->drr_u.drr_freeobjects);
 461 
 462         /*
 463          * If there is a pending op, but it's not PENDING_FREEOBJECTS,
 464          * push it out, since free block aggregation can only be done for
 465          * blocks of the same type (i.e., DRR_FREE records can only be
 466          * aggregated with other DRR_FREE records.  DRR_FREEOBJECTS records
 467          * can only be aggregated with other DRR_FREEOBJECTS records.
 468          */
 469         if (dsp->dsa_pending_op != PENDING_NONE &&
 470             dsp->dsa_pending_op != PENDING_FREEOBJECTS) {
 471                 if (dump_record(dsp, NULL, 0) != 0)
 472                         return (SET_ERROR(EINTR));
 473                 dsp->dsa_pending_op = PENDING_NONE;
 474         }
 475         if (dsp->dsa_pending_op == PENDING_FREEOBJECTS) {
 476                 /*
 477                  * See whether this free object array can be aggregated
 478                  * with pending one
 479                  */
 480                 if (drrfo->drr_firstobj + drrfo->drr_numobjs == firstobj) {
 481                         drrfo->drr_numobjs += numobjs;
 482                         return (0);
 483                 } else {
 484                         /* can't be aggregated.  Push out pending record */
 485                         if (dump_record(dsp, NULL, 0) != 0)
 486                                 return (SET_ERROR(EINTR));
 487                         dsp->dsa_pending_op = PENDING_NONE;
 488                 }
 489         }
 490 
 491         /* write a FREEOBJECTS record */
 492         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 493         dsp->dsa_drr->drr_type = DRR_FREEOBJECTS;
 494         drrfo->drr_firstobj = firstobj;
 495         drrfo->drr_numobjs = numobjs;
 496         drrfo->drr_toguid = dsp->dsa_toguid;
 497 
 498         dsp->dsa_pending_op = PENDING_FREEOBJECTS;
 499 
 500         return (0);
 501 }
 502 
 503 static int
 504 dump_dnode(dmu_sendarg_t *dsp, uint64_t object, dnode_phys_t *dnp)
 505 {
 506         struct drr_object *drro = &(dsp->dsa_drr->drr_u.drr_object);
 507 
 508         if (object < dsp->dsa_resume_object) {
 509                 /*
 510                  * Note: when resuming, we will visit all the dnodes in
 511                  * the block of dnodes that we are resuming from.  In
 512                  * this case it's unnecessary to send the dnodes prior to
 513                  * the one we are resuming from.  We should be at most one
 514                  * block's worth of dnodes behind the resume point.
 515                  */
 516                 ASSERT3U(dsp->dsa_resume_object - object, <,
 517                     1 << (DNODE_BLOCK_SHIFT - DNODE_SHIFT));
 518                 return (0);
 519         }
 520 
 521         if (dnp == NULL || dnp->dn_type == DMU_OT_NONE)
 522                 return (dump_freeobjects(dsp, object, 1));
 523 
 524         if (dsp->dsa_pending_op != PENDING_NONE) {
 525                 if (dump_record(dsp, NULL, 0) != 0)
 526                         return (SET_ERROR(EINTR));
 527                 dsp->dsa_pending_op = PENDING_NONE;
 528         }
 529 
 530         /* write an OBJECT record */
 531         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 532         dsp->dsa_drr->drr_type = DRR_OBJECT;
 533         drro->drr_object = object;
 534         drro->drr_type = dnp->dn_type;
 535         drro->drr_bonustype = dnp->dn_bonustype;
 536         drro->drr_blksz = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT;
 537         drro->drr_bonuslen = dnp->dn_bonuslen;
 538         drro->drr_checksumtype = dnp->dn_checksum;
 539         drro->drr_compress = dnp->dn_compress;
 540         drro->drr_toguid = dsp->dsa_toguid;
 541 
 542         if (!(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
 543             drro->drr_blksz > SPA_OLD_MAXBLOCKSIZE)
 544                 drro->drr_blksz = SPA_OLD_MAXBLOCKSIZE;
 545 
 546         if (dump_record(dsp, DN_BONUS(dnp),
 547             P2ROUNDUP(dnp->dn_bonuslen, 8)) != 0) {
 548                 return (SET_ERROR(EINTR));
 549         }
 550 
 551         /* Free anything past the end of the file. */
 552         if (dump_free(dsp, object, (dnp->dn_maxblkid + 1) *
 553             (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT), -1ULL) != 0)
 554                 return (SET_ERROR(EINTR));
 555         if (dsp->dsa_err != 0)
 556                 return (SET_ERROR(EINTR));
 557         return (0);
 558 }
 559 
 560 static boolean_t
 561 backup_do_embed(dmu_sendarg_t *dsp, const blkptr_t *bp)
 562 {
 563         if (!BP_IS_EMBEDDED(bp))
 564                 return (B_FALSE);
 565 
 566         /*
 567          * Compression function must be legacy, or explicitly enabled.
 568          */
 569         if ((BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_LEGACY_FUNCTIONS &&
 570             !(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_LZ4)))
 571                 return (B_FALSE);
 572 
 573         /*
 574          * Embed type must be explicitly enabled.
 575          */
 576         switch (BPE_GET_ETYPE(bp)) {
 577         case BP_EMBEDDED_TYPE_DATA:
 578                 if (dsp->dsa_featureflags & DMU_BACKUP_FEATURE_EMBED_DATA)
 579                         return (B_TRUE);
 580                 break;
 581         default:
 582                 return (B_FALSE);
 583         }
 584         return (B_FALSE);
 585 }
 586 
 587 /*
 588  * This is the callback function to traverse_dataset that acts as the worker
 589  * thread for dmu_send_impl.
 590  */
 591 /*ARGSUSED*/
 592 static int
 593 send_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
 594     const zbookmark_phys_t *zb, const struct dnode_phys *dnp, void *arg)
 595 {
 596         struct send_thread_arg *sta = arg;
 597         struct send_block_record *record;
 598         uint64_t record_size;
 599         int err = 0;
 600 
 601         ASSERT(zb->zb_object == DMU_META_DNODE_OBJECT ||
 602             zb->zb_object >= sta->resume.zb_object);
 603 
 604         if (sta->cancel)
 605                 return (SET_ERROR(EINTR));
 606 
 607         if (bp == NULL) {
 608                 ASSERT3U(zb->zb_level, ==, ZB_DNODE_LEVEL);
 609                 return (0);
 610         } else if (zb->zb_level < 0) {
 611                 return (0);
 612         }
 613 
 614         record = kmem_zalloc(sizeof (struct send_block_record), KM_SLEEP);
 615         record->eos_marker = B_FALSE;
 616         record->bp = *bp;
 617         record->zb = *zb;
 618         record->indblkshift = dnp->dn_indblkshift;
 619         record->datablkszsec = dnp->dn_datablkszsec;
 620         record_size = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT;
 621         bqueue_enqueue(&sta->q, record, record_size);
 622 
 623         return (err);
 624 }
 625 
 626 /*
 627  * This function kicks off the traverse_dataset.  It also handles setting the
 628  * error code of the thread in case something goes wrong, and pushes the End of
 629  * Stream record when the traverse_dataset call has finished.  If there is no
 630  * dataset to traverse, the thread immediately pushes End of Stream marker.
 631  */
 632 static void
 633 send_traverse_thread(void *arg)
 634 {
 635         struct send_thread_arg *st_arg = arg;
 636         int err;
 637         struct send_block_record *data;
 638 
 639         if (st_arg->ds != NULL) {
 640                 err = traverse_dataset_resume(st_arg->ds,
 641                     st_arg->fromtxg, &st_arg->resume,
 642                     st_arg->flags, send_cb, st_arg);
 643 
 644                 if (err != EINTR)
 645                         st_arg->error_code = err;
 646         }
 647         data = kmem_zalloc(sizeof (*data), KM_SLEEP);
 648         data->eos_marker = B_TRUE;
 649         bqueue_enqueue(&st_arg->q, data, 1);
 650         thread_exit();
 651 }
 652 
 653 /*
 654  * This function actually handles figuring out what kind of record needs to be
 655  * dumped, reading the data (which has hopefully been prefetched), and calling
 656  * the appropriate helper function.
 657  */
 658 static int
 659 do_dump(dmu_sendarg_t *dsa, struct send_block_record *data)
 660 {
 661         dsl_dataset_t *ds = dmu_objset_ds(dsa->dsa_os);
 662         const blkptr_t *bp = &data->bp;
 663         const zbookmark_phys_t *zb = &data->zb;
 664         uint8_t indblkshift = data->indblkshift;
 665         uint16_t dblkszsec = data->datablkszsec;
 666         spa_t *spa = ds->ds_dir->dd_pool->dp_spa;
 667         dmu_object_type_t type = bp ? BP_GET_TYPE(bp) : DMU_OT_NONE;
 668         int err = 0;
 669 
 670         ASSERT3U(zb->zb_level, >=, 0);
 671 
 672         ASSERT(zb->zb_object == DMU_META_DNODE_OBJECT ||
 673             zb->zb_object >= dsa->dsa_resume_object);
 674 
 675         if (zb->zb_object != DMU_META_DNODE_OBJECT &&
 676             DMU_OBJECT_IS_SPECIAL(zb->zb_object)) {
 677                 return (0);
 678         } else if (BP_IS_HOLE(bp) &&
 679             zb->zb_object == DMU_META_DNODE_OBJECT) {
 680                 uint64_t span = BP_SPAN(dblkszsec, indblkshift, zb->zb_level);
 681                 uint64_t dnobj = (zb->zb_blkid * span) >> DNODE_SHIFT;
 682                 err = dump_freeobjects(dsa, dnobj, span >> DNODE_SHIFT);
 683         } else if (BP_IS_HOLE(bp)) {
 684                 uint64_t span = BP_SPAN(dblkszsec, indblkshift, zb->zb_level);
 685                 uint64_t offset = zb->zb_blkid * span;
 686                 err = dump_free(dsa, zb->zb_object, offset, span);
 687         } else if (zb->zb_level > 0 || type == DMU_OT_OBJSET) {
 688                 return (0);
 689         } else if (type == DMU_OT_DNODE) {
 690                 int blksz = BP_GET_LSIZE(bp);
 691                 arc_flags_t aflags = ARC_FLAG_WAIT;
 692                 arc_buf_t *abuf;
 693 
 694                 ASSERT0(zb->zb_level);
 695 
 696                 if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
 697                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
 698                     &aflags, zb) != 0)
 699                         return (SET_ERROR(EIO));
 700 
 701                 dnode_phys_t *blk = abuf->b_data;
 702                 uint64_t dnobj = zb->zb_blkid * (blksz >> DNODE_SHIFT);
 703                 for (int i = 0; i < blksz >> DNODE_SHIFT; i++) {
 704                         err = dump_dnode(dsa, dnobj + i, blk + i);
 705                         if (err != 0)
 706                                 break;
 707                 }
 708                 arc_buf_destroy(abuf, &abuf);
 709         } else if (type == DMU_OT_SA) {
 710                 /*
 711                  * The upstream code has arc_read() call here, but we moved
 712                  * it to dump_spill() since we want to take advantage of
 713                  * zero copy of the buffer if possible
 714                  */
 715                 err = dump_spill(dsa, zb->zb_object, bp, zb);
 716         } else if (backup_do_embed(dsa, bp)) {
 717                 /* it's an embedded level-0 block of a regular object */
 718                 int blksz = dblkszsec << SPA_MINBLOCKSHIFT;
 719                 ASSERT0(zb->zb_level);
 720                 err = dump_write_embedded(dsa, zb->zb_object,
 721                     zb->zb_blkid * blksz, blksz, bp);
 722         } else {
 723                 /* it's a level-0 block of a regular object */
 724                 arc_flags_t aflags = ARC_FLAG_WAIT;
 725                 arc_buf_t *abuf;
 726                 int blksz = dblkszsec << SPA_MINBLOCKSHIFT;
 727                 uint64_t offset;
 728 
 729                 /*
 730                  * If we have large blocks stored on disk but the send flags
 731                  * don't allow us to send large blocks, we split the data from
 732                  * the arc buf into chunks.
 733                  */
 734                 boolean_t split_large_blocks = blksz > SPA_OLD_MAXBLOCKSIZE &&
 735                     !(dsa->dsa_featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS);
 736                 /*
 737                  * We should only request compressed data from the ARC if all
 738                  * the following are true:
 739                  *  - stream compression was requested
 740                  *  - we aren't splitting large blocks into smaller chunks
 741                  *  - the data won't need to be byteswapped before sending
 742                  *  - this isn't an embedded block
 743                  *  - this isn't metadata (if receiving on a different endian
 744                  *    system it can be byteswapped more easily)
 745                  */
 746                 boolean_t request_compressed =
 747                     (dsa->dsa_featureflags & DMU_BACKUP_FEATURE_COMPRESSED) &&
 748                     !split_large_blocks && !BP_SHOULD_BYTESWAP(bp) &&
 749                     !BP_IS_EMBEDDED(bp) && !DMU_OT_IS_METADATA(BP_GET_TYPE(bp));
 750 
 751                 ASSERT0(zb->zb_level);
 752                 ASSERT(zb->zb_object > dsa->dsa_resume_object ||
 753                     (zb->zb_object == dsa->dsa_resume_object &&
 754                     zb->zb_blkid * blksz >= dsa->dsa_resume_offset));
 755 
 756                 ASSERT3U(blksz, ==, BP_GET_LSIZE(bp));
 757 
 758                 enum zio_flag zioflags = ZIO_FLAG_CANFAIL;
 759                 if (request_compressed)
 760                         zioflags |= ZIO_FLAG_RAW;
 761                 if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
 762                     ZIO_PRIORITY_ASYNC_READ, zioflags, &aflags, zb) != 0) {
 763                         if (zfs_send_corrupt_data) {
 764                                 /* Send a block filled with 0x"zfs badd bloc" */
 765                                 abuf = arc_alloc_buf(spa, &abuf, ARC_BUFC_DATA,
 766                                     blksz);
 767                                 uint64_t *ptr;
 768                                 for (ptr = abuf->b_data;
 769                                     (char *)ptr < (char *)abuf->b_data + blksz;
 770                                     ptr++)
 771                                         *ptr = 0x2f5baddb10cULL;
 772                         } else {
 773                                 return (SET_ERROR(EIO));
 774                         }
 775                 }
 776 
 777                 offset = zb->zb_blkid * blksz;
 778 
 779                 if (split_large_blocks) {
 780                         ASSERT3U(arc_get_compression(abuf), ==,
 781                             ZIO_COMPRESS_OFF);
 782                         char *buf = abuf->b_data;
 783                         while (blksz > 0 && err == 0) {
 784                                 int n = MIN(blksz, SPA_OLD_MAXBLOCKSIZE);
 785                                 err = dump_write(dsa, type, zb->zb_object,
 786                                     offset, n, n, NULL, buf);
 787                                 offset += n;
 788                                 blksz -= n;
 789                         }
 790                 } else {
 791                         err = dump_write(dsa, type, zb->zb_object, offset,
 792                             blksz, arc_buf_size(abuf), bp, abuf->b_data);
 793                 }
 794                 arc_buf_destroy(abuf, &abuf);
 795         }
 796 
 797         ASSERT(err == 0 || err == EINTR);
 798         return (err);
 799 }
 800 
 801 /*
 802  * Pop the new data off the queue, and free the old data.
 803  */
 804 static struct send_block_record *
 805 get_next_record(bqueue_t *bq, struct send_block_record *data)
 806 {
 807         struct send_block_record *tmp = bqueue_dequeue(bq);
 808         kmem_free(data, sizeof (*data));
 809         return (tmp);
 810 }
 811 
 812 /*
 813  * Actually do the bulk of the work in a zfs send.
 814  *
 815  * Note: Releases dp using the specified tag.
 816  */
 817 static int
 818 dmu_send_impl_ss(void *tag, dsl_pool_t *dp, dsl_dataset_t *to_ds,
 819     zfs_bookmark_phys_t *ancestor_zb, boolean_t is_clone,
 820     boolean_t embedok, boolean_t large_block_ok, boolean_t compressok,
 821     int outfd, uint64_t resumeobj, uint64_t resumeoff, vnode_t *vp,
 822     offset_t *off, boolean_t sendsize, dmu_krrp_task_t *krrp_task)
 823 {
 824         objset_t *os;
 825         dmu_replay_record_t *drr;
 826         dmu_sendarg_t *dsp;
 827         int err;
 828         uint64_t fromtxg = 0;
 829         uint64_t featureflags = 0;
 830         struct send_thread_arg to_arg = { 0 };
 831 
 832         err = dmu_objset_from_ds(to_ds, &os);
 833         if (err != 0) {
 834                 dsl_pool_rele(dp, tag);
 835                 return (err);
 836         }
 837 
 838         drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP);
 839         drr->drr_type = DRR_BEGIN;
 840         drr->drr_u.drr_begin.drr_magic = DMU_BACKUP_MAGIC;
 841         DMU_SET_STREAM_HDRTYPE(drr->drr_u.drr_begin.drr_versioninfo,
 842             DMU_SUBSTREAM);
 843 
 844 #ifdef _KERNEL
 845         if (dmu_objset_type(os) == DMU_OST_ZFS) {
 846                 uint64_t version;
 847                 if (zfs_get_zplprop(os, ZFS_PROP_VERSION, &version) != 0) {
 848                         kmem_free(drr, sizeof (dmu_replay_record_t));
 849                         dsl_pool_rele(dp, tag);
 850                         return (SET_ERROR(EINVAL));
 851                 }
 852                 if (version >= ZPL_VERSION_SA) {
 853                         featureflags |= DMU_BACKUP_FEATURE_SA_SPILL;
 854                 }
 855         }
 856 #endif
 857 
 858         if (large_block_ok && to_ds->ds_feature_inuse[SPA_FEATURE_LARGE_BLOCKS])
 859                 featureflags |= DMU_BACKUP_FEATURE_LARGE_BLOCKS;
 860         if (embedok &&
 861             spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA)) {
 862                 featureflags |= DMU_BACKUP_FEATURE_EMBED_DATA;
 863                 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS))
 864                         featureflags |= DMU_BACKUP_FEATURE_LZ4;
 865         }
 866         if (compressok) {
 867                 featureflags |= DMU_BACKUP_FEATURE_COMPRESSED;
 868         }
 869         if ((featureflags &
 870             (DMU_BACKUP_FEATURE_EMBED_DATA | DMU_BACKUP_FEATURE_COMPRESSED)) !=
 871             0 && spa_feature_is_active(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS)) {
 872                 featureflags |= DMU_BACKUP_FEATURE_LZ4;
 873         }
 874 
 875         if (resumeobj != 0 || resumeoff != 0) {
 876                 featureflags |= DMU_BACKUP_FEATURE_RESUMING;
 877         }
 878 
 879         DMU_SET_FEATUREFLAGS(drr->drr_u.drr_begin.drr_versioninfo,
 880             featureflags);
 881 
 882         drr->drr_u.drr_begin.drr_creation_time =
 883             dsl_dataset_phys(to_ds)->ds_creation_time;
 884         drr->drr_u.drr_begin.drr_type = dmu_objset_type(os);
 885         if (is_clone)
 886                 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CLONE;
 887         drr->drr_u.drr_begin.drr_toguid = dsl_dataset_phys(to_ds)->ds_guid;
 888         if (dsl_dataset_phys(to_ds)->ds_flags & DS_FLAG_CI_DATASET)
 889                 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CI_DATA;
 890         if (zfs_send_set_freerecords_bit)
 891                 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_FREERECORDS;
 892 
 893         if (ancestor_zb != NULL) {
 894                 drr->drr_u.drr_begin.drr_fromguid =
 895                     ancestor_zb->zbm_guid;
 896                 fromtxg = ancestor_zb->zbm_creation_txg;
 897         }
 898         dsl_dataset_name(to_ds, drr->drr_u.drr_begin.drr_toname);
 899         if (!to_ds->ds_is_snapshot) {
 900                 (void) strlcat(drr->drr_u.drr_begin.drr_toname, "@--head--",
 901                     sizeof (drr->drr_u.drr_begin.drr_toname));
 902         }
 903 
 904         dsp = kmem_zalloc(sizeof (dmu_sendarg_t), KM_SLEEP);
 905 
 906         dsp->dsa_drr = drr;
 907         dsp->dsa_vp = vp;
 908         dsp->dsa_outfd = outfd;
 909         dsp->dsa_proc = curproc;
 910         dsp->dsa_os = os;
 911         dsp->dsa_off = off;
 912         dsp->dsa_toguid = dsl_dataset_phys(to_ds)->ds_guid;
 913         dsp->dsa_krrp_task = krrp_task;
 914         dsp->dsa_pending_op = PENDING_NONE;
 915         dsp->dsa_featureflags = featureflags;
 916         dsp->sendsize = sendsize;
 917         dsp->dsa_resume_object = resumeobj;
 918         dsp->dsa_resume_offset = resumeoff;
 919 
 920         mutex_enter(&to_ds->ds_sendstream_lock);
 921         list_insert_head(&to_ds->ds_sendstreams, dsp);
 922         mutex_exit(&to_ds->ds_sendstream_lock);
 923 
 924         dsl_dataset_long_hold(to_ds, FTAG);
 925         dsl_pool_rele(dp, tag);
 926 
 927         void *payload = NULL;
 928         size_t payload_len = 0;
 929         if (resumeobj != 0 || resumeoff != 0) {
 930                 dmu_object_info_t to_doi;
 931                 err = dmu_object_info(os, resumeobj, &to_doi);
 932                 if (err != 0)
 933                         goto out;
 934                 SET_BOOKMARK(&to_arg.resume, to_ds->ds_object, resumeobj, 0,
 935                     resumeoff / to_doi.doi_data_block_size);
 936 
 937                 nvlist_t *nvl = fnvlist_alloc();
 938                 fnvlist_add_uint64(nvl, "resume_object", resumeobj);
 939                 fnvlist_add_uint64(nvl, "resume_offset", resumeoff);
 940                 payload = fnvlist_pack(nvl, &payload_len);
 941                 drr->drr_payloadlen = payload_len;
 942                 fnvlist_free(nvl);
 943         }
 944 
 945         err = dump_record(dsp, payload, payload_len);
 946         fnvlist_pack_free(payload, payload_len);
 947         if (err != 0) {
 948                 err = dsp->dsa_err;
 949                 goto out;
 950         }
 951 
 952         err = bqueue_init(&to_arg.q, zfs_send_queue_length,
 953             offsetof(struct send_block_record, ln));
 954         to_arg.error_code = 0;
 955         to_arg.cancel = B_FALSE;
 956         to_arg.ds = to_ds;
 957         to_arg.fromtxg = fromtxg;
 958         to_arg.flags = TRAVERSE_PRE | TRAVERSE_PREFETCH;
 959         (void) thread_create(NULL, 0, send_traverse_thread, &to_arg, 0, curproc,
 960             TS_RUN, minclsyspri);
 961 
 962         struct send_block_record *to_data;
 963         to_data = bqueue_dequeue(&to_arg.q);
 964 
 965         while (!to_data->eos_marker && err == 0) {
 966                 err = do_dump(dsp, to_data);
 967                 to_data = get_next_record(&to_arg.q, to_data);
 968                 if (vp != NULL && issig(JUSTLOOKING) && issig(FORREAL))
 969                         err = EINTR;
 970         }
 971 
 972         if (err != 0) {
 973                 to_arg.cancel = B_TRUE;
 974                 while (!to_data->eos_marker) {
 975                         to_data = get_next_record(&to_arg.q, to_data);
 976                 }
 977         }
 978         kmem_free(to_data, sizeof (*to_data));
 979 
 980         bqueue_destroy(&to_arg.q);
 981 
 982         if (err == 0 && to_arg.error_code != 0)
 983                 err = to_arg.error_code;
 984 
 985         if (err != 0)
 986                 goto out;
 987 
 988         if (dsp->dsa_pending_op != PENDING_NONE)
 989                 if (dump_record(dsp, NULL, 0) != 0)
 990                         err = SET_ERROR(EINTR);
 991 
 992         if (err != 0) {
 993                 if (err == EINTR && dsp->dsa_err != 0)
 994                         err = dsp->dsa_err;
 995                 goto out;
 996         }
 997 
 998         bzero(drr, sizeof (dmu_replay_record_t));
 999         drr->drr_type = DRR_END;
1000         drr->drr_u.drr_end.drr_checksum = dsp->dsa_zc;
1001         drr->drr_u.drr_end.drr_toguid = dsp->dsa_toguid;
1002 
1003         if (dump_record(dsp, NULL, 0) != 0)
1004                 err = dsp->dsa_err;
1005 
1006 out:
1007         mutex_enter(&to_ds->ds_sendstream_lock);
1008         list_remove(&to_ds->ds_sendstreams, dsp);
1009         mutex_exit(&to_ds->ds_sendstream_lock);
1010 
1011         VERIFY(err != 0 || (dsp->dsa_sent_begin && dsp->dsa_sent_end));
1012 
1013         kmem_free(drr, sizeof (dmu_replay_record_t));
1014         kmem_free(dsp, sizeof (dmu_sendarg_t));
1015 
1016         dsl_dataset_long_rele(to_ds, FTAG);
1017 
1018         return (err);
1019 }
1020 
1021 int
1022 dmu_send_impl(void *tag, dsl_pool_t *dp, dsl_dataset_t *to_ds,
1023     zfs_bookmark_phys_t *ancestor_zb, boolean_t is_clone, boolean_t embedok,
1024     boolean_t large_block_ok, boolean_t compressok, int outfd,
1025     uint64_t resumeobj, uint64_t resumeoff, vnode_t *vp, offset_t *off,
1026     dmu_krrp_task_t *krrp_task)
1027 {
1028         return (dmu_send_impl_ss(tag, dp, to_ds, ancestor_zb, is_clone,
1029             embedok, large_block_ok, compressok, outfd, resumeobj, resumeoff,
1030             vp, off, B_FALSE, krrp_task));
1031 }
1032 
1033 int
1034 dmu_send_obj(const char *pool, uint64_t tosnap, uint64_t fromsnap,
1035     boolean_t embedok, boolean_t large_block_ok, boolean_t compressok,
1036     int outfd, vnode_t *vp, offset_t *off, boolean_t sendsize)
1037 {
1038         dsl_pool_t *dp;
1039         dsl_dataset_t *ds;
1040         dsl_dataset_t *fromds = NULL;
1041         int err;
1042 
1043         err = dsl_pool_hold(pool, FTAG, &dp);
1044         if (err != 0)
1045                 return (err);
1046 
1047         err = dsl_dataset_hold_obj(dp, tosnap, FTAG, &ds);
1048         if (err != 0) {
1049                 dsl_pool_rele(dp, FTAG);
1050                 return (err);
1051         }
1052 
1053         if (fromsnap != 0) {
1054                 zfs_bookmark_phys_t zb;
1055                 boolean_t is_clone;
1056 
1057                 err = dsl_dataset_hold_obj(dp, fromsnap, FTAG, &fromds);
1058                 if (err != 0) {
1059                         dsl_dataset_rele(ds, FTAG);
1060                         dsl_pool_rele(dp, FTAG);
1061                         return (err);
1062                 }
1063                 if (!dsl_dataset_is_before(ds, fromds, 0))
1064                         err = SET_ERROR(EXDEV);
1065                 zb.zbm_creation_time =
1066                     dsl_dataset_phys(fromds)->ds_creation_time;
1067                 zb.zbm_creation_txg = dsl_dataset_phys(fromds)->ds_creation_txg;
1068                 zb.zbm_guid = dsl_dataset_phys(fromds)->ds_guid;
1069                 is_clone = (fromds->ds_dir != ds->ds_dir);
1070                 dsl_dataset_rele(fromds, FTAG);
1071                 err = dmu_send_impl_ss(FTAG, dp, ds, &zb, is_clone,
1072                     embedok, large_block_ok, compressok, outfd, 0, 0, vp, off,
1073                         sendsize, NULL);
1074         } else {
1075                 err = dmu_send_impl_ss(FTAG, dp, ds, NULL, B_FALSE,
1076                     embedok, large_block_ok, compressok, outfd, 0, 0, vp, off,
1077                         sendsize, NULL);
1078         }
1079         dsl_dataset_rele(ds, FTAG);
1080         return (err);
1081 }
1082 
1083 int
1084 dmu_send(const char *tosnap, const char *fromsnap, boolean_t embedok,
1085     boolean_t large_block_ok, boolean_t compressok, int outfd,
1086     uint64_t resumeobj, uint64_t resumeoff,
1087     vnode_t *vp, offset_t *off)
1088 {
1089         dsl_pool_t *dp;
1090         dsl_dataset_t *ds;
1091         int err;
1092         boolean_t owned = B_FALSE;
1093 
1094         if (fromsnap != NULL && strpbrk(fromsnap, "@#") == NULL)
1095                 return (SET_ERROR(EINVAL));
1096 
1097         err = dsl_pool_hold(tosnap, FTAG, &dp);
1098         if (err != 0)
1099                 return (err);
1100 
1101         if (strchr(tosnap, '@') == NULL && spa_writeable(dp->dp_spa)) {
1102                 /*
1103                  * We are sending a filesystem or volume.  Ensure
1104                  * that it doesn't change by owning the dataset.
1105                  */
1106                 err = dsl_dataset_own(dp, tosnap, FTAG, &ds);
1107                 owned = B_TRUE;
1108         } else {
1109                 err = dsl_dataset_hold(dp, tosnap, FTAG, &ds);
1110         }
1111         if (err != 0) {
1112                 dsl_pool_rele(dp, FTAG);
1113                 return (err);
1114         }
1115 
1116         if (fromsnap != NULL) {
1117                 zfs_bookmark_phys_t zb;
1118                 boolean_t is_clone = B_FALSE;
1119                 int fsnamelen = strchr(tosnap, '@') - tosnap;
1120 
1121                 /*
1122                  * If the fromsnap is in a different filesystem, then
1123                  * mark the send stream as a clone.
1124                  */
1125                 if (strncmp(tosnap, fromsnap, fsnamelen) != 0 ||
1126                     (fromsnap[fsnamelen] != '@' &&
1127                     fromsnap[fsnamelen] != '#')) {
1128                         is_clone = B_TRUE;
1129                 }
1130 
1131                 if (strchr(fromsnap, '@')) {
1132                         dsl_dataset_t *fromds;
1133                         err = dsl_dataset_hold(dp, fromsnap, FTAG, &fromds);
1134                         if (err == 0) {
1135                                 if (!dsl_dataset_is_before(ds, fromds, 0))
1136                                         err = SET_ERROR(EXDEV);
1137                                 zb.zbm_creation_time =
1138                                     dsl_dataset_phys(fromds)->ds_creation_time;
1139                                 zb.zbm_creation_txg =
1140                                     dsl_dataset_phys(fromds)->ds_creation_txg;
1141                                 zb.zbm_guid = dsl_dataset_phys(fromds)->ds_guid;
1142                                 is_clone = (ds->ds_dir != fromds->ds_dir);
1143                                 dsl_dataset_rele(fromds, FTAG);
1144                         }
1145                 } else {
1146                         err = dsl_bookmark_lookup(dp, fromsnap, ds, &zb);
1147                 }
1148                 if (err != 0) {
1149                         dsl_dataset_rele(ds, FTAG);
1150                         dsl_pool_rele(dp, FTAG);
1151                         return (err);
1152                 }
1153                 err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone,
1154                     embedok, large_block_ok, compressok, outfd,
1155                     resumeobj, resumeoff, vp, off, NULL);
1156         } else {
1157                 err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE,
1158                     embedok, large_block_ok, compressok, outfd,
1159                     resumeobj, resumeoff, vp, off, NULL);
1160         }
1161         if (owned)
1162                 dsl_dataset_disown(ds, FTAG);
1163         else
1164                 dsl_dataset_rele(ds, FTAG);
1165         return (err);
1166 }
1167 
1168 static int
1169 dmu_adjust_send_estimate_for_indirects(dsl_dataset_t *ds, uint64_t uncompressed,
1170     uint64_t compressed, boolean_t stream_compressed, uint64_t *sizep)
1171 {
1172         int err;
1173         uint64_t size;
1174         /*
1175          * Assume that space (both on-disk and in-stream) is dominated by
1176          * data.  We will adjust for indirect blocks and the copies property,
1177          * but ignore per-object space used (eg, dnodes and DRR_OBJECT records).
1178          */
1179         uint64_t recordsize;
1180         uint64_t record_count;
1181         objset_t *os;
1182         VERIFY0(dmu_objset_from_ds(ds, &os));
1183 
1184         /* Assume all (uncompressed) blocks are recordsize. */
1185         if (os->os_phys->os_type == DMU_OST_ZVOL) {
1186                 err = dsl_prop_get_int_ds(ds,
1187                     zfs_prop_to_name(ZFS_PROP_VOLBLOCKSIZE), &recordsize);
1188         } else {
1189                 err = dsl_prop_get_int_ds(ds,
1190                     zfs_prop_to_name(ZFS_PROP_RECORDSIZE), &recordsize);
1191         }
1192         if (err != 0)
1193                 return (err);
1194         record_count = uncompressed / recordsize;
1195 
1196         /*
1197          * If we're estimating a send size for a compressed stream, use the
1198          * compressed data size to estimate the stream size. Otherwise, use the
1199          * uncompressed data size.
1200          */
1201         size = stream_compressed ? compressed : uncompressed;
1202 
1203         /*
1204          * Subtract out approximate space used by indirect blocks.
1205          * Assume most space is used by data blocks (non-indirect, non-dnode).
1206          * Assume no ditto blocks or internal fragmentation.
1207          *
1208          * Therefore, space used by indirect blocks is sizeof(blkptr_t) per
1209          * block.
1210          */
1211         size -= record_count * sizeof (blkptr_t);
1212 
1213         /* Add in the space for the record associated with each block. */
1214         size += record_count * sizeof (dmu_replay_record_t);
1215 
1216         *sizep = size;
1217 
1218         return (0);
1219 }
1220 
1221 int
1222 dmu_send_estimate(dsl_dataset_t *ds, dsl_dataset_t *fromds,
1223     boolean_t stream_compressed, uint64_t *sizep)
1224 {
1225         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1226         int err;
1227         uint64_t uncomp, comp;
1228 
1229         ASSERT(dsl_pool_config_held(dp));
1230 
1231         /* tosnap must be a snapshot */
1232         if (!ds->ds_is_snapshot)
1233                 return (SET_ERROR(EINVAL));
1234 
1235         /* fromsnap, if provided, must be a snapshot */
1236         if (fromds != NULL && !fromds->ds_is_snapshot)
1237                 return (SET_ERROR(EINVAL));
1238 
1239         /*
1240          * fromsnap must be an earlier snapshot from the same fs as tosnap,
1241          * or the origin's fs.
1242          */
1243         if (fromds != NULL && !dsl_dataset_is_before(ds, fromds, 0))
1244                 return (SET_ERROR(EXDEV));
1245 
1246         /* Get compressed and uncompressed size estimates of changed data. */
1247         if (fromds == NULL) {
1248                 uncomp = dsl_dataset_phys(ds)->ds_uncompressed_bytes;
1249                 comp = dsl_dataset_phys(ds)->ds_compressed_bytes;
1250         } else {
1251                 uint64_t used;
1252                 err = dsl_dataset_space_written(fromds, ds,
1253                     &used, &comp, &uncomp);
1254                 if (err != 0)
1255                         return (err);
1256         }
1257 
1258         err = dmu_adjust_send_estimate_for_indirects(ds, uncomp, comp,
1259             stream_compressed, sizep);
1260         /*
1261          * Add the size of the BEGIN and END records to the estimate.
1262          */
1263         *sizep += 2 * sizeof (dmu_replay_record_t);
1264         return (err);
1265 }
1266 
1267 struct calculate_send_arg {
1268         uint64_t uncompressed;
1269         uint64_t compressed;
1270 };
1271 
1272 /*
1273  * Simple callback used to traverse the blocks of a snapshot and sum their
1274  * uncompressed and compressed sizes.
1275  */
1276 /* ARGSUSED */
1277 static int
1278 dmu_calculate_send_traversal(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
1279     const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
1280 {
1281         struct calculate_send_arg *space = arg;
1282         if (bp != NULL && !BP_IS_HOLE(bp)) {
1283                 space->uncompressed += BP_GET_UCSIZE(bp);
1284                 space->compressed += BP_GET_PSIZE(bp);
1285         }
1286         return (0);
1287 }
1288 
1289 /*
1290  * Given a desination snapshot and a TXG, calculate the approximate size of a
1291  * send stream sent from that TXG. from_txg may be zero, indicating that the
1292  * whole snapshot will be sent.
1293  */
1294 int
1295 dmu_send_estimate_from_txg(dsl_dataset_t *ds, uint64_t from_txg,
1296     boolean_t stream_compressed, uint64_t *sizep)
1297 {
1298         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1299         int err;
1300         struct calculate_send_arg size = { 0 };
1301 
1302         ASSERT(dsl_pool_config_held(dp));
1303 
1304         /* tosnap must be a snapshot */
1305         if (!ds->ds_is_snapshot)
1306                 return (SET_ERROR(EINVAL));
1307 
1308         /* verify that from_txg is before the provided snapshot was taken */
1309         if (from_txg >= dsl_dataset_phys(ds)->ds_creation_txg) {
1310                 return (SET_ERROR(EXDEV));
1311         }
1312 
1313         /*
1314          * traverse the blocks of the snapshot with birth times after
1315          * from_txg, summing their uncompressed size
1316          */
1317         err = traverse_dataset(ds, from_txg, TRAVERSE_POST,
1318             dmu_calculate_send_traversal, &size);
1319         if (err)
1320                 return (err);
1321 
1322         err = dmu_adjust_send_estimate_for_indirects(ds, size.uncompressed,
1323             size.compressed, stream_compressed, sizep);
1324         return (err);
1325 }
1326 
1327 typedef struct dmu_recv_begin_arg {
1328         const char *drba_origin;
1329         dmu_recv_cookie_t *drba_cookie;
1330         cred_t *drba_cred;
1331         uint64_t drba_snapobj;
1332 } dmu_recv_begin_arg_t;
1333 
1334 static int
1335 recv_begin_check_existing_impl(dmu_recv_begin_arg_t *drba, dsl_dataset_t *ds,
1336     uint64_t fromguid, dmu_tx_t *tx)
1337 {
1338         uint64_t val;
1339         int error;
1340         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1341 
1342         if (dmu_tx_is_syncing(tx)) {
1343                 /* temporary clone name must not exist */
1344                 error = zap_lookup(dp->dp_meta_objset,
1345                     dsl_dir_phys(ds->ds_dir)->dd_child_dir_zapobj,
1346                     recv_clone_name, 8, 1, &val);
1347                 if (error == 0) {
1348                         dsl_dataset_t *tds;
1349 
1350                         /* check that if it is currently used */
1351                         error = dsl_dataset_own_obj(dp, val, FTAG, &tds);
1352                         if (!error) {
1353                                 char name[ZFS_MAX_DATASET_NAME_LEN];
1354 
1355                                 dsl_dataset_name(tds, name);
1356                                 dsl_dataset_disown(tds, FTAG);
1357 
1358                                 error = dsl_dataset_hold(dp, name, FTAG, &tds);
1359                                 if (!error) {
1360                                         dsl_destroy_head_sync_impl(tds, tx);
1361                                         dsl_dataset_rele(tds, FTAG);
1362                                         error = ENOENT;
1363                                 }
1364                         } else {
1365                                 error = 0;
1366                         }
1367                 }
1368                 if (error != ENOENT) {
1369                         return (error == 0 ?
1370                             SET_ERROR(EBUSY) : SET_ERROR(error));
1371                 }
1372         }
1373 
1374         /* new snapshot name must not exist */
1375         error = zap_lookup(dp->dp_meta_objset,
1376             dsl_dataset_phys(ds)->ds_snapnames_zapobj,
1377             drba->drba_cookie->drc_tosnap, 8, 1, &val);
1378         if (error != ENOENT)
1379                 return (error == 0 ? SET_ERROR(EEXIST) : SET_ERROR(error));
1380 
1381         /*
1382          * Check snapshot limit before receiving. We'll recheck again at the
1383          * end, but might as well abort before receiving if we're already over
1384          * the limit.
1385          *
1386          * Note that we do not check the file system limit with
1387          * dsl_dir_fscount_check because the temporary %clones don't count
1388          * against that limit.
1389          */
1390         error = dsl_fs_ss_limit_check(ds->ds_dir, 1, ZFS_PROP_SNAPSHOT_LIMIT,
1391             NULL, drba->drba_cred);
1392         if (error != 0)
1393                 return (error);
1394 
1395         if (fromguid != 0) {
1396                 dsl_dataset_t *snap;
1397                 uint64_t obj = dsl_dataset_phys(ds)->ds_prev_snap_obj;
1398 
1399                 /* Find snapshot in this dir that matches fromguid. */
1400                 while (obj != 0) {
1401                         error = dsl_dataset_hold_obj(dp, obj, FTAG,
1402                             &snap);
1403                         if (error != 0)
1404                                 return (SET_ERROR(ENODEV));
1405                         if (snap->ds_dir != ds->ds_dir) {
1406                                 dsl_dataset_rele(snap, FTAG);
1407                                 return (SET_ERROR(ENODEV));
1408                         }
1409                         if (dsl_dataset_phys(snap)->ds_guid == fromguid)
1410                                 break;
1411                         obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
1412                         dsl_dataset_rele(snap, FTAG);
1413                 }
1414                 if (obj == 0)
1415                         return (SET_ERROR(ENODEV));
1416 
1417                 if (drba->drba_cookie->drc_force) {
1418                         drba->drba_snapobj = obj;
1419                 } else {
1420                         /*
1421                          * If we are not forcing, there must be no
1422                          * changes since fromsnap.
1423                          */
1424                         if (dsl_dataset_modified_since_snap(ds, snap)) {
1425                                 dsl_dataset_rele(snap, FTAG);
1426                                 return (SET_ERROR(ETXTBSY));
1427                         }
1428                         drba->drba_snapobj = ds->ds_prev->ds_object;
1429                 }
1430 
1431                 dsl_dataset_rele(snap, FTAG);
1432         } else {
1433                 /* if full, then must be forced */
1434                 if (!drba->drba_cookie->drc_force)
1435                         return (SET_ERROR(EEXIST));
1436                 /* start from $ORIGIN@$ORIGIN, if supported */
1437                 drba->drba_snapobj = dp->dp_origin_snap != NULL ?
1438                     dp->dp_origin_snap->ds_object : 0;
1439         }
1440 
1441         return (0);
1442 
1443 }
1444 
1445 static int
1446 dmu_recv_begin_check(void *arg, dmu_tx_t *tx)
1447 {
1448         dmu_recv_begin_arg_t *drba = arg;
1449         dsl_pool_t *dp = dmu_tx_pool(tx);
1450         struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
1451         uint64_t fromguid = drrb->drr_fromguid;
1452         int flags = drrb->drr_flags;
1453         int error;
1454         uint64_t featureflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo);
1455         dsl_dataset_t *ds;
1456         const char *tofs = drba->drba_cookie->drc_tofs;
1457 
1458         /* already checked */
1459         ASSERT3U(drrb->drr_magic, ==, DMU_BACKUP_MAGIC);
1460         ASSERT(!(featureflags & DMU_BACKUP_FEATURE_RESUMING));
1461 
1462         if (DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) ==
1463             DMU_COMPOUNDSTREAM ||
1464             drrb->drr_type >= DMU_OST_NUMTYPES ||
1465             ((flags & DRR_FLAG_CLONE) && drba->drba_origin == NULL))
1466                 return (SET_ERROR(EINVAL));
1467 
1468         /* Verify pool version supports SA if SA_SPILL feature set */
1469         if ((featureflags & DMU_BACKUP_FEATURE_SA_SPILL) &&
1470             spa_version(dp->dp_spa) < SPA_VERSION_SA)
1471                 return (SET_ERROR(ENOTSUP));
1472 
1473         if (drba->drba_cookie->drc_resumable &&
1474             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EXTENSIBLE_DATASET))
1475                 return (SET_ERROR(ENOTSUP));
1476 
1477         /*
1478          * The receiving code doesn't know how to translate a WRITE_EMBEDDED
1479          * record to a plain WRITE record, so the pool must have the
1480          * EMBEDDED_DATA feature enabled if the stream has WRITE_EMBEDDED
1481          * records.  Same with WRITE_EMBEDDED records that use LZ4 compression.
1482          */
1483         if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA) &&
1484             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA))
1485                 return (SET_ERROR(ENOTSUP));
1486         if ((featureflags & DMU_BACKUP_FEATURE_LZ4) &&
1487             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS))
1488                 return (SET_ERROR(ENOTSUP));
1489 
1490         /*
1491          * The receiving code doesn't know how to translate large blocks
1492          * to smaller ones, so the pool must have the LARGE_BLOCKS
1493          * feature enabled if the stream has LARGE_BLOCKS.
1494          */
1495         if ((featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
1496             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LARGE_BLOCKS))
1497                 return (SET_ERROR(ENOTSUP));
1498 
1499         error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
1500         if (error == 0) {
1501                 /* target fs already exists; recv into temp clone */
1502 
1503                 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_WBC)) {
1504                         objset_t *os = NULL;
1505 
1506                         error = dmu_objset_from_ds(ds, &os);
1507                         if (error) {
1508                                 dsl_dataset_rele(ds, FTAG);
1509                                 return (error);
1510                         }
1511 
1512                         /* Recv is impossible into DS that uses WBC */
1513                         if (os->os_wbc_mode != ZFS_WBC_MODE_OFF) {
1514                                 dsl_dataset_rele(ds, FTAG);
1515                                 return (SET_ERROR(EKZFS_WBCNOTSUP));
1516                         }
1517                 }
1518 
1519                 /* Can't recv a clone into an existing fs */
1520                 if (flags & DRR_FLAG_CLONE || drba->drba_origin) {
1521                         dsl_dataset_rele(ds, FTAG);
1522                         return (SET_ERROR(EINVAL));
1523                 }
1524 
1525                 error = recv_begin_check_existing_impl(drba, ds, fromguid, tx);
1526                 dsl_dataset_rele(ds, FTAG);
1527         } else if (error == ENOENT) {
1528                 /* target fs does not exist; must be a full backup or clone */
1529                 char buf[ZFS_MAX_DATASET_NAME_LEN];
1530 
1531                 /*
1532                  * If it's a non-clone incremental, we are missing the
1533                  * target fs, so fail the recv.
1534                  */
1535                 if (fromguid != 0 && !(flags & DRR_FLAG_CLONE ||
1536                     drba->drba_origin))
1537                         return (SET_ERROR(ENOENT));
1538 
1539                 /*
1540                  * If we're receiving a full send as a clone, and it doesn't
1541                  * contain all the necessary free records and freeobject
1542                  * records, reject it.
1543                  */
1544                 if (fromguid == 0 && drba->drba_origin &&
1545                     !(flags & DRR_FLAG_FREERECORDS))
1546                         return (SET_ERROR(EINVAL));
1547 
1548                 /* Open the parent of tofs */
1549                 ASSERT3U(strlen(tofs), <, sizeof (buf));
1550                 (void) strlcpy(buf, tofs, strrchr(tofs, '/') - tofs + 1);
1551                 error = dsl_dataset_hold(dp, buf, FTAG, &ds);
1552                 if (error != 0)
1553                         return (error);
1554 
1555                 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_WBC)) {
1556                         objset_t *os = NULL;
1557 
1558                         error = dmu_objset_from_ds(ds, &os);
1559                         if (error) {
1560                                 dsl_dataset_rele(ds, FTAG);
1561                                 return (error);
1562                         }
1563 
1564                         /* Recv is impossible into DS that uses WBC */
1565                         if (os->os_wbc_mode != ZFS_WBC_MODE_OFF) {
1566                                 dsl_dataset_rele(ds, FTAG);
1567                                 return (SET_ERROR(EKZFS_WBCNOTSUP));
1568                         }
1569                 }
1570 
1571                 /*
1572                  * Check filesystem and snapshot limits before receiving. We'll
1573                  * recheck snapshot limits again at the end (we create the
1574                  * filesystems and increment those counts during begin_sync).
1575                  */
1576                 error = dsl_fs_ss_limit_check(ds->ds_dir, 1,
1577                     ZFS_PROP_FILESYSTEM_LIMIT, NULL, drba->drba_cred);
1578                 if (error != 0) {
1579                         dsl_dataset_rele(ds, FTAG);
1580                         return (error);
1581                 }
1582 
1583                 error = dsl_fs_ss_limit_check(ds->ds_dir, 1,
1584                     ZFS_PROP_SNAPSHOT_LIMIT, NULL, drba->drba_cred);
1585                 if (error != 0) {
1586                         dsl_dataset_rele(ds, FTAG);
1587                         return (error);
1588                 }
1589 
1590                 if (drba->drba_origin != NULL) {
1591                         dsl_dataset_t *origin;
1592                         error = dsl_dataset_hold(dp, drba->drba_origin,
1593                             FTAG, &origin);
1594                         if (error != 0) {
1595                                 dsl_dataset_rele(ds, FTAG);
1596                                 return (error);
1597                         }
1598                         if (!origin->ds_is_snapshot) {
1599                                 dsl_dataset_rele(origin, FTAG);
1600                                 dsl_dataset_rele(ds, FTAG);
1601                                 return (SET_ERROR(EINVAL));
1602                         }
1603                         if (dsl_dataset_phys(origin)->ds_guid != fromguid &&
1604                             fromguid != 0) {
1605                                 dsl_dataset_rele(origin, FTAG);
1606                                 dsl_dataset_rele(ds, FTAG);
1607                                 return (SET_ERROR(ENODEV));
1608                         }
1609                         dsl_dataset_rele(origin, FTAG);
1610                 }
1611                 dsl_dataset_rele(ds, FTAG);
1612                 error = 0;
1613         }
1614         return (error);
1615 }
1616 
1617 static void
1618 dmu_recv_begin_sync(void *arg, dmu_tx_t *tx)
1619 {
1620         dmu_recv_begin_arg_t *drba = arg;
1621         dsl_pool_t *dp = dmu_tx_pool(tx);
1622         objset_t *mos = dp->dp_meta_objset;
1623         struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
1624         const char *tofs = drba->drba_cookie->drc_tofs;
1625         dsl_dataset_t *ds, *newds;
1626         uint64_t dsobj;
1627         int error;
1628         uint64_t crflags = 0;
1629 
1630         if (drrb->drr_flags & DRR_FLAG_CI_DATA)
1631                 crflags |= DS_FLAG_CI_DATASET;
1632 
1633         error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
1634         if (error == 0) {
1635                 /* create temporary clone */
1636                 dsl_dataset_t *snap = NULL;
1637                 if (drba->drba_snapobj != 0) {
1638                         VERIFY0(dsl_dataset_hold_obj(dp,
1639                             drba->drba_snapobj, FTAG, &snap));
1640                 }
1641                 dsobj = dsl_dataset_create_sync(ds->ds_dir, recv_clone_name,
1642                     snap, crflags, drba->drba_cred, tx);
1643                 if (drba->drba_snapobj != 0)
1644                         dsl_dataset_rele(snap, FTAG);
1645                 dsl_dataset_rele(ds, FTAG);
1646         } else {
1647                 dsl_dir_t *dd;
1648                 const char *tail;
1649                 dsl_dataset_t *origin = NULL;
1650 
1651                 VERIFY0(dsl_dir_hold(dp, tofs, FTAG, &dd, &tail));
1652 
1653                 if (drba->drba_origin != NULL) {
1654                         VERIFY0(dsl_dataset_hold(dp, drba->drba_origin,
1655                             FTAG, &origin));
1656                 }
1657 
1658                 /* Create new dataset. */
1659                 dsobj = dsl_dataset_create_sync(dd,
1660                     strrchr(tofs, '/') + 1,
1661                     origin, crflags, drba->drba_cred, tx);
1662                 if (origin != NULL)
1663                         dsl_dataset_rele(origin, FTAG);
1664                 dsl_dir_rele(dd, FTAG);
1665                 drba->drba_cookie->drc_newfs = B_TRUE;
1666         }
1667         VERIFY0(dsl_dataset_own_obj(dp, dsobj, dmu_recv_tag, &newds));
1668 
1669         if (drba->drba_cookie->drc_resumable) {
1670                 dsl_dataset_zapify(newds, tx);
1671                 if (drrb->drr_fromguid != 0) {
1672                         VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_FROMGUID,
1673                             8, 1, &drrb->drr_fromguid, tx));
1674                 }
1675                 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_TOGUID,
1676                     8, 1, &drrb->drr_toguid, tx));
1677                 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_TONAME,
1678                     1, strlen(drrb->drr_toname) + 1, drrb->drr_toname, tx));
1679                 uint64_t one = 1;
1680                 uint64_t zero = 0;
1681                 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_OBJECT,
1682                     8, 1, &one, tx));
1683                 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_OFFSET,
1684                     8, 1, &zero, tx));
1685                 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_BYTES,
1686                     8, 1, &zero, tx));
1687                 if (DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo) &
1688                     DMU_BACKUP_FEATURE_LARGE_BLOCKS) {
1689                         VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_LARGEBLOCK,
1690                             8, 1, &one, tx));
1691                 }
1692                 if (DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo) &
1693                     DMU_BACKUP_FEATURE_EMBED_DATA) {
1694                         VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_EMBEDOK,
1695                             8, 1, &one, tx));
1696                 }
1697                 if (DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo) &
1698                     DMU_BACKUP_FEATURE_COMPRESSED) {
1699                         VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_COMPRESSOK,
1700                             8, 1, &one, tx));
1701                 }
1702         }
1703 
1704         dmu_buf_will_dirty(newds->ds_dbuf, tx);
1705         dsl_dataset_phys(newds)->ds_flags |= DS_FLAG_INCONSISTENT;
1706 
1707         /*
1708          * If we actually created a non-clone, we need to create the
1709          * objset in our new dataset.
1710          */
1711         rrw_enter(&newds->ds_bp_rwlock, RW_READER, FTAG);
1712         if (BP_IS_HOLE(dsl_dataset_get_blkptr(newds))) {
1713                 (void) dmu_objset_create_impl(dp->dp_spa,
1714                     newds, dsl_dataset_get_blkptr(newds), drrb->drr_type, tx);
1715         }
1716         rrw_exit(&newds->ds_bp_rwlock, FTAG);
1717 
1718         drba->drba_cookie->drc_ds = newds;
1719 
1720         spa_history_log_internal_ds(newds, "receive", tx, "");
1721 }
1722 
1723 static int
1724 dmu_recv_resume_begin_check(void *arg, dmu_tx_t *tx)
1725 {
1726         dmu_recv_begin_arg_t *drba = arg;
1727         dsl_pool_t *dp = dmu_tx_pool(tx);
1728         struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
1729         int error;
1730         uint64_t featureflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo);
1731         dsl_dataset_t *ds;
1732         const char *tofs = drba->drba_cookie->drc_tofs;
1733 
1734         /* already checked */
1735         ASSERT3U(drrb->drr_magic, ==, DMU_BACKUP_MAGIC);
1736         ASSERT(featureflags & DMU_BACKUP_FEATURE_RESUMING);
1737 
1738         if (DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) ==
1739             DMU_COMPOUNDSTREAM ||
1740             drrb->drr_type >= DMU_OST_NUMTYPES)
1741                 return (SET_ERROR(EINVAL));
1742 
1743         /* Verify pool version supports SA if SA_SPILL feature set */
1744         if ((featureflags & DMU_BACKUP_FEATURE_SA_SPILL) &&
1745             spa_version(dp->dp_spa) < SPA_VERSION_SA)
1746                 return (SET_ERROR(ENOTSUP));
1747 
1748         /*
1749          * The receiving code doesn't know how to translate a WRITE_EMBEDDED
1750          * record to a plain WRITE record, so the pool must have the
1751          * EMBEDDED_DATA feature enabled if the stream has WRITE_EMBEDDED
1752          * records.  Same with WRITE_EMBEDDED records that use LZ4 compression.
1753          */
1754         if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA) &&
1755             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA))
1756                 return (SET_ERROR(ENOTSUP));
1757         if ((featureflags & DMU_BACKUP_FEATURE_LZ4) &&
1758             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS))
1759                 return (SET_ERROR(ENOTSUP));
1760 
1761         /* 6 extra bytes for /%recv */
1762         char recvname[ZFS_MAX_DATASET_NAME_LEN + 6];
1763 
1764         (void) snprintf(recvname, sizeof (recvname), "%s/%s",
1765             tofs, recv_clone_name);
1766 
1767         if (dsl_dataset_hold(dp, recvname, FTAG, &ds) != 0) {
1768                 /* %recv does not exist; continue in tofs */
1769                 error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
1770                 if (error != 0)
1771                         return (error);
1772         }
1773 
1774         /* check that ds is marked inconsistent */
1775         if (!DS_IS_INCONSISTENT(ds)) {
1776                 dsl_dataset_rele(ds, FTAG);
1777                 return (SET_ERROR(EINVAL));
1778         }
1779 
1780         /* check that there is resuming data, and that the toguid matches */
1781         if (!dsl_dataset_is_zapified(ds)) {
1782                 dsl_dataset_rele(ds, FTAG);
1783                 return (SET_ERROR(EINVAL));
1784         }
1785         uint64_t val = 0;
1786         error = zap_lookup(dp->dp_meta_objset, ds->ds_object,
1787             DS_FIELD_RESUME_TOGUID, sizeof (val), 1, &val);
1788         if (error != 0 || drrb->drr_toguid != val) {
1789                 dsl_dataset_rele(ds, FTAG);
1790                 return (SET_ERROR(EINVAL));
1791         }
1792 
1793         /*
1794          * Check if the receive is still running.  If so, it will be owned.
1795          * Note that nothing else can own the dataset (e.g. after the receive
1796          * fails) because it will be marked inconsistent.
1797          */
1798         if (dsl_dataset_has_owner(ds)) {
1799                 dsl_dataset_rele(ds, FTAG);
1800                 return (SET_ERROR(EBUSY));
1801         }
1802 
1803         /* There should not be any snapshots of this fs yet. */
1804         if (ds->ds_prev != NULL && ds->ds_prev->ds_dir == ds->ds_dir) {
1805                 dsl_dataset_rele(ds, FTAG);
1806                 return (SET_ERROR(EINVAL));
1807         }
1808 
1809         /*
1810          * Note: resume point will be checked when we process the first WRITE
1811          * record.
1812          */
1813 
1814         /* check that the origin matches */
1815         val = 0;
1816         (void) zap_lookup(dp->dp_meta_objset, ds->ds_object,
1817             DS_FIELD_RESUME_FROMGUID, sizeof (val), 1, &val);
1818         if (drrb->drr_fromguid != val) {
1819                 dsl_dataset_rele(ds, FTAG);
1820                 return (SET_ERROR(EINVAL));
1821         }
1822 
1823         dsl_dataset_rele(ds, FTAG);
1824         return (0);
1825 }
1826 
1827 static void
1828 dmu_recv_resume_begin_sync(void *arg, dmu_tx_t *tx)
1829 {
1830         dmu_recv_begin_arg_t *drba = arg;
1831         dsl_pool_t *dp = dmu_tx_pool(tx);
1832         const char *tofs = drba->drba_cookie->drc_tofs;
1833         dsl_dataset_t *ds;
1834         uint64_t dsobj;
1835         /* 6 extra bytes for /%recv */
1836         char recvname[ZFS_MAX_DATASET_NAME_LEN + 6];
1837 
1838         (void) snprintf(recvname, sizeof (recvname), "%s/%s",
1839             tofs, recv_clone_name);
1840 
1841         if (dsl_dataset_hold(dp, recvname, FTAG, &ds) != 0) {
1842                 /* %recv does not exist; continue in tofs */
1843                 VERIFY0(dsl_dataset_hold(dp, tofs, FTAG, &ds));
1844                 drba->drba_cookie->drc_newfs = B_TRUE;
1845         }
1846 
1847         /* clear the inconsistent flag so that we can own it */
1848         ASSERT(DS_IS_INCONSISTENT(ds));
1849         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1850         dsl_dataset_phys(ds)->ds_flags &= ~DS_FLAG_INCONSISTENT;
1851         dsobj = ds->ds_object;
1852         dsl_dataset_rele(ds, FTAG);
1853 
1854         VERIFY0(dsl_dataset_own_obj(dp, dsobj, dmu_recv_tag, &ds));
1855 
1856         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1857         dsl_dataset_phys(ds)->ds_flags |= DS_FLAG_INCONSISTENT;
1858 
1859         rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
1860         ASSERT(!BP_IS_HOLE(dsl_dataset_get_blkptr(ds)));
1861         rrw_exit(&ds->ds_bp_rwlock, FTAG);
1862 
1863         drba->drba_cookie->drc_ds = ds;
1864 
1865         spa_history_log_internal_ds(ds, "resume receive", tx, "");
1866 }
1867 
1868 /*
1869  * NB: callers *MUST* call dmu_recv_stream() if dmu_recv_begin()
1870  * succeeds; otherwise we will leak the holds on the datasets.
1871  */
1872 int
1873 dmu_recv_begin(char *tofs, char *tosnap, dmu_replay_record_t *drr_begin,
1874     boolean_t force, boolean_t resumable, boolean_t force_cksum,
1875     char *origin, dmu_recv_cookie_t *drc)
1876 {
1877         dmu_recv_begin_arg_t drba = { 0 };
1878 
1879         bzero(drc, sizeof (dmu_recv_cookie_t));
1880         drc->drc_drr_begin = drr_begin;
1881         drc->drc_drrb = &drr_begin->drr_u.drr_begin;
1882         drc->drc_tosnap = tosnap;
1883         drc->drc_tofs = tofs;
1884         drc->drc_force = force;
1885         drc->drc_resumable = resumable;
1886         drc->drc_cred = CRED();
1887 
1888         if (drc->drc_drrb->drr_magic == BSWAP_64(DMU_BACKUP_MAGIC)) {
1889                 drc->drc_byteswap = B_TRUE;
1890 
1891                 /* on-wire checksum can be disabled for krrp */
1892                 if (force_cksum) {
1893                         (void) fletcher_4_incremental_byteswap(drr_begin,
1894                             sizeof (dmu_replay_record_t), &drc->drc_cksum);
1895                         byteswap_record(drr_begin);
1896                 }
1897         } else if (drc->drc_drrb->drr_magic == DMU_BACKUP_MAGIC) {
1898                 /* on-wire checksum can be disabled for krrp */
1899                 if (force_cksum) {
1900                         (void) fletcher_4_incremental_native(drr_begin,
1901                             sizeof (dmu_replay_record_t), &drc->drc_cksum);
1902                 }
1903         } else {
1904                 return (SET_ERROR(EINVAL));
1905         }
1906 
1907         drba.drba_origin = origin;
1908         drba.drba_cookie = drc;
1909         drba.drba_cred = CRED();
1910 
1911         if (DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo) &
1912             DMU_BACKUP_FEATURE_RESUMING) {
1913                 return (dsl_sync_task(tofs,
1914                     dmu_recv_resume_begin_check, dmu_recv_resume_begin_sync,
1915                     &drba, 5, ZFS_SPACE_CHECK_NORMAL));
1916         } else  {
1917                 return (dsl_sync_task(tofs,
1918                     dmu_recv_begin_check, dmu_recv_begin_sync,
1919                     &drba, 5, ZFS_SPACE_CHECK_NORMAL));
1920         }
1921 }
1922 
1923 struct receive_record_arg {
1924         dmu_replay_record_t header;
1925         void *payload; /* Pointer to a buffer containing the payload */
1926         /*
1927          * If the record is a write, pointer to the arc_buf_t containing the
1928          * payload.
1929          */
1930         arc_buf_t *write_buf;
1931         int payload_size;
1932         uint64_t bytes_read; /* bytes read from stream when record created */
1933         boolean_t eos_marker; /* Marks the end of the stream */
1934         bqueue_node_t node;
1935 };
1936 
1937 struct receive_writer_arg {
1938         objset_t *os;
1939         boolean_t byteswap;
1940         bqueue_t q;
1941 
1942         /*
1943          * These three args are used to signal to the main thread that we're
1944          * done.
1945          */
1946         kmutex_t mutex;
1947         kcondvar_t cv;
1948         boolean_t done;
1949 
1950         int err;
1951         /* A map from guid to dataset to help handle dedup'd streams. */
1952         avl_tree_t *guid_to_ds_map;
1953         boolean_t resumable;
1954         uint64_t last_object, last_offset;
1955         uint64_t bytes_read; /* bytes read when current record created */
1956 };
1957 
1958 struct objlist {
1959         list_t list; /* List of struct receive_objnode. */
1960         /*
1961          * Last object looked up. Used to assert that objects are being looked
1962          * up in ascending order.
1963          */
1964         uint64_t last_lookup;
1965 };
1966 
1967 struct receive_objnode {
1968         list_node_t node;
1969         uint64_t object;
1970 };
1971 
1972 struct receive_arg {
1973         objset_t *os;
1974         vnode_t *vp; /* The vnode to read the stream from */
1975         uint64_t voff; /* The current offset in the stream */
1976         uint64_t bytes_read;
1977         /*
1978          * A record that has had its payload read in, but hasn't yet been handed
1979          * off to the worker thread.
1980          */
1981         struct receive_record_arg *rrd;
1982         /* A record that has had its header read in, but not its payload. */
1983         struct receive_record_arg *next_rrd;
1984         zio_cksum_t cksum;
1985         zio_cksum_t prev_cksum;
1986         dmu_krrp_task_t *krrp_task;
1987         int err;
1988         boolean_t byteswap;
1989         /* Sorted list of objects not to issue prefetches for. */
1990         struct objlist ignore_objlist;
1991 };
1992 
1993 typedef struct guid_map_entry {
1994         uint64_t        guid;
1995         dsl_dataset_t   *gme_ds;
1996         avl_node_t      avlnode;
1997 } guid_map_entry_t;
1998 
1999 static int
2000 guid_compare(const void *arg1, const void *arg2)
2001 {
2002         const guid_map_entry_t *gmep1 = arg1;
2003         const guid_map_entry_t *gmep2 = arg2;
2004 
2005         if (gmep1->guid < gmep2->guid)
2006                 return (-1);
2007         else if (gmep1->guid > gmep2->guid)
2008                 return (1);
2009         return (0);
2010 }
2011 
2012 static void
2013 free_guid_map_onexit(void *arg)
2014 {
2015         avl_tree_t *ca = arg;
2016         void *cookie = NULL;
2017         guid_map_entry_t *gmep;
2018 
2019         while ((gmep = avl_destroy_nodes(ca, &cookie)) != NULL) {
2020                 dsl_dataset_long_rele(gmep->gme_ds, gmep);
2021                 dsl_dataset_rele(gmep->gme_ds, gmep);
2022                 kmem_free(gmep, sizeof (guid_map_entry_t));
2023         }
2024         avl_destroy(ca);
2025         kmem_free(ca, sizeof (avl_tree_t));
2026 }
2027 
2028 static int
2029 receive_read(struct receive_arg *ra, int len, void *buf)
2030 {
2031         int done = 0;
2032 
2033         /*
2034          * The code doesn't rely on this (lengths being multiples of 8).  See
2035          * comment in dump_bytes.
2036          */
2037         ASSERT0(len % 8);
2038 
2039         /*
2040          * if vp is NULL, then the send is from krrp and we can try to bypass
2041          * copying data to an intermediate buffer.
2042          */
2043         if (ra->vp != NULL) {
2044                 while (done < len) {
2045                         ssize_t resid = 0;
2046 
2047                         ra->err = vn_rdwr(UIO_READ, ra->vp,
2048                             (char *)buf + done, len - done,
2049                             ra->voff, UIO_SYSSPACE, FAPPEND,
2050                             RLIM64_INFINITY, CRED(), &resid);
2051                         if (resid == len - done) {
2052                                 /*
2053                                  * Note: ECKSUM indicates that the receive was
2054                                  * interrupted and can potentially be resumed.
2055                                  */
2056                                 ra->err = SET_ERROR(ECKSUM);
2057                         }
2058                         ra->voff += len - done - resid;
2059                         done = len - resid;
2060                         if (ra->err != 0)
2061                                 return (ra->err);
2062                 }
2063         } else {
2064                 ASSERT(ra->krrp_task != NULL);
2065                 ra->err = dmu_krrp_buffer_read(buf, len, ra->krrp_task);
2066                 if (ra->err != 0)
2067                         return (ra->err);
2068 
2069                 done = len;
2070         }
2071 
2072         ra->bytes_read += len;
2073 
2074         ASSERT3U(done, ==, len);
2075         return (0);
2076 }
2077 
2078 static void
2079 byteswap_record(dmu_replay_record_t *drr)
2080 {
2081 #define DO64(X) (drr->drr_u.X = BSWAP_64(drr->drr_u.X))
2082 #define DO32(X) (drr->drr_u.X = BSWAP_32(drr->drr_u.X))
2083         drr->drr_type = BSWAP_32(drr->drr_type);
2084         drr->drr_payloadlen = BSWAP_32(drr->drr_payloadlen);
2085 
2086         switch (drr->drr_type) {
2087         case DRR_BEGIN:
2088                 DO64(drr_begin.drr_magic);
2089                 DO64(drr_begin.drr_versioninfo);
2090                 DO64(drr_begin.drr_creation_time);
2091                 DO32(drr_begin.drr_type);
2092                 DO32(drr_begin.drr_flags);
2093                 DO64(drr_begin.drr_toguid);
2094                 DO64(drr_begin.drr_fromguid);
2095                 break;
2096         case DRR_OBJECT:
2097                 DO64(drr_object.drr_object);
2098                 DO32(drr_object.drr_type);
2099                 DO32(drr_object.drr_bonustype);
2100                 DO32(drr_object.drr_blksz);
2101                 DO32(drr_object.drr_bonuslen);
2102                 DO64(drr_object.drr_toguid);
2103                 break;
2104         case DRR_FREEOBJECTS:
2105                 DO64(drr_freeobjects.drr_firstobj);
2106                 DO64(drr_freeobjects.drr_numobjs);
2107                 DO64(drr_freeobjects.drr_toguid);
2108                 break;
2109         case DRR_WRITE:
2110                 DO64(drr_write.drr_object);
2111                 DO32(drr_write.drr_type);
2112                 DO64(drr_write.drr_offset);
2113                 DO64(drr_write.drr_logical_size);
2114                 DO64(drr_write.drr_toguid);
2115                 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_write.drr_key.ddk_cksum);
2116                 DO64(drr_write.drr_key.ddk_prop);
2117                 DO64(drr_write.drr_compressed_size);
2118                 break;
2119         case DRR_WRITE_BYREF:
2120                 DO64(drr_write_byref.drr_object);
2121                 DO64(drr_write_byref.drr_offset);
2122                 DO64(drr_write_byref.drr_length);
2123                 DO64(drr_write_byref.drr_toguid);
2124                 DO64(drr_write_byref.drr_refguid);
2125                 DO64(drr_write_byref.drr_refobject);
2126                 DO64(drr_write_byref.drr_refoffset);
2127                 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_write_byref.
2128                     drr_key.ddk_cksum);
2129                 DO64(drr_write_byref.drr_key.ddk_prop);
2130                 break;
2131         case DRR_WRITE_EMBEDDED:
2132                 DO64(drr_write_embedded.drr_object);
2133                 DO64(drr_write_embedded.drr_offset);
2134                 DO64(drr_write_embedded.drr_length);
2135                 DO64(drr_write_embedded.drr_toguid);
2136                 DO32(drr_write_embedded.drr_lsize);
2137                 DO32(drr_write_embedded.drr_psize);
2138                 break;
2139         case DRR_FREE:
2140                 DO64(drr_free.drr_object);
2141                 DO64(drr_free.drr_offset);
2142                 DO64(drr_free.drr_length);
2143                 DO64(drr_free.drr_toguid);
2144                 break;
2145         case DRR_SPILL:
2146                 DO64(drr_spill.drr_object);
2147                 DO64(drr_spill.drr_length);
2148                 DO64(drr_spill.drr_toguid);
2149                 break;
2150         case DRR_END:
2151                 DO64(drr_end.drr_toguid);
2152                 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_end.drr_checksum);
2153                 break;
2154         }
2155 
2156         if (drr->drr_type != DRR_BEGIN) {
2157                 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_checksum.drr_checksum);
2158         }
2159 
2160 #undef DO64
2161 #undef DO32
2162 }
2163 
2164 static inline uint8_t
2165 deduce_nblkptr(dmu_object_type_t bonus_type, uint64_t bonus_size)
2166 {
2167         if (bonus_type == DMU_OT_SA) {
2168                 return (1);
2169         } else {
2170                 return (1 +
2171                     ((DN_MAX_BONUSLEN - bonus_size) >> SPA_BLKPTRSHIFT));
2172         }
2173 }
2174 
2175 static void
2176 save_resume_state(struct receive_writer_arg *rwa,
2177     uint64_t object, uint64_t offset, dmu_tx_t *tx)
2178 {
2179         int txgoff = dmu_tx_get_txg(tx) & TXG_MASK;
2180 
2181         if (!rwa->resumable)
2182                 return;
2183 
2184         /*
2185          * We use ds_resume_bytes[] != 0 to indicate that we need to
2186          * update this on disk, so it must not be 0.
2187          */
2188         ASSERT(rwa->bytes_read != 0);
2189 
2190         /*
2191          * We only resume from write records, which have a valid
2192          * (non-meta-dnode) object number.
2193          */
2194         ASSERT(object != 0);
2195 
2196         /*
2197          * For resuming to work correctly, we must receive records in order,
2198          * sorted by object,offset.  This is checked by the callers, but
2199          * assert it here for good measure.
2200          */
2201         ASSERT3U(object, >=, rwa->os->os_dsl_dataset->ds_resume_object[txgoff]);
2202         ASSERT(object != rwa->os->os_dsl_dataset->ds_resume_object[txgoff] ||
2203             offset >= rwa->os->os_dsl_dataset->ds_resume_offset[txgoff]);
2204         ASSERT3U(rwa->bytes_read, >=,
2205             rwa->os->os_dsl_dataset->ds_resume_bytes[txgoff]);
2206 
2207         rwa->os->os_dsl_dataset->ds_resume_object[txgoff] = object;
2208         rwa->os->os_dsl_dataset->ds_resume_offset[txgoff] = offset;
2209         rwa->os->os_dsl_dataset->ds_resume_bytes[txgoff] = rwa->bytes_read;
2210 }
2211 
2212 static int
2213 receive_object(struct receive_writer_arg *rwa, struct drr_object *drro,
2214     void *data)
2215 {
2216         dmu_object_info_t doi;
2217         dmu_tx_t *tx;
2218         uint64_t object;
2219         int err;
2220 
2221         if (drro->drr_type == DMU_OT_NONE ||
2222             !DMU_OT_IS_VALID(drro->drr_type) ||
2223             !DMU_OT_IS_VALID(drro->drr_bonustype) ||
2224             drro->drr_checksumtype >= ZIO_CHECKSUM_FUNCTIONS ||
2225             drro->drr_compress >= ZIO_COMPRESS_FUNCTIONS ||
2226             P2PHASE(drro->drr_blksz, SPA_MINBLOCKSIZE) ||
2227             drro->drr_blksz < SPA_MINBLOCKSIZE ||
2228             drro->drr_blksz > spa_maxblocksize(dmu_objset_spa(rwa->os)) ||
2229             drro->drr_bonuslen > DN_MAX_BONUSLEN) {
2230                 return (SET_ERROR(EINVAL));
2231         }
2232 
2233         err = dmu_object_info(rwa->os, drro->drr_object, &doi);
2234 
2235         if (err != 0 && err != ENOENT)
2236                 return (SET_ERROR(EINVAL));
2237         object = err == 0 ? drro->drr_object : DMU_NEW_OBJECT;
2238 
2239         /*
2240          * If we are losing blkptrs or changing the block size this must
2241          * be a new file instance.  We must clear out the previous file
2242          * contents before we can change this type of metadata in the dnode.
2243          */
2244         if (err == 0) {
2245                 int nblkptr;
2246 
2247                 nblkptr = deduce_nblkptr(drro->drr_bonustype,
2248                     drro->drr_bonuslen);
2249 
2250                 if (drro->drr_blksz != doi.doi_data_block_size ||
2251                     nblkptr < doi.doi_nblkptr) {
2252                         err = dmu_free_long_range(rwa->os, drro->drr_object,
2253                             0, DMU_OBJECT_END);
2254                         if (err != 0)
2255                                 return (SET_ERROR(EINVAL));
2256                 }
2257         }
2258 
2259         tx = dmu_tx_create(rwa->os);
2260         dmu_tx_hold_bonus(tx, object);
2261         err = dmu_tx_assign(tx, TXG_WAIT);
2262         if (err != 0) {
2263                 dmu_tx_abort(tx);
2264                 return (err);
2265         }
2266 
2267         if (object == DMU_NEW_OBJECT) {
2268                 /* currently free, want to be allocated */
2269                 err = dmu_object_claim(rwa->os, drro->drr_object,
2270                     drro->drr_type, drro->drr_blksz,
2271                     drro->drr_bonustype, drro->drr_bonuslen, tx);
2272         } else if (drro->drr_type != doi.doi_type ||
2273             drro->drr_blksz != doi.doi_data_block_size ||
2274             drro->drr_bonustype != doi.doi_bonus_type ||
2275             drro->drr_bonuslen != doi.doi_bonus_size) {
2276                 /* currently allocated, but with different properties */
2277                 err = dmu_object_reclaim(rwa->os, drro->drr_object,
2278                     drro->drr_type, drro->drr_blksz,
2279                     drro->drr_bonustype, drro->drr_bonuslen, tx);
2280         }
2281         if (err != 0) {
2282                 dmu_tx_commit(tx);
2283                 return (SET_ERROR(EINVAL));
2284         }
2285 
2286         dmu_object_set_checksum(rwa->os, drro->drr_object,
2287             drro->drr_checksumtype, tx);
2288         dmu_object_set_compress(rwa->os, drro->drr_object,
2289             drro->drr_compress, tx);
2290 
2291         if (data != NULL) {
2292                 dmu_buf_t *db;
2293 
2294                 VERIFY0(dmu_bonus_hold(rwa->os, drro->drr_object, FTAG, &db));
2295                 dmu_buf_will_dirty(db, tx);
2296 
2297                 ASSERT3U(db->db_size, >=, drro->drr_bonuslen);
2298                 bcopy(data, db->db_data, drro->drr_bonuslen);
2299                 if (rwa->byteswap) {
2300                         dmu_object_byteswap_t byteswap =
2301                             DMU_OT_BYTESWAP(drro->drr_bonustype);
2302                         dmu_ot_byteswap[byteswap].ob_func(db->db_data,
2303                             drro->drr_bonuslen);
2304                 }
2305                 dmu_buf_rele(db, FTAG);
2306         }
2307         dmu_tx_commit(tx);
2308 
2309         return (0);
2310 }
2311 
2312 /* ARGSUSED */
2313 static int
2314 receive_freeobjects(struct receive_writer_arg *rwa,
2315     struct drr_freeobjects *drrfo)
2316 {
2317         uint64_t obj;
2318         int next_err = 0;
2319 
2320         if (drrfo->drr_firstobj + drrfo->drr_numobjs < drrfo->drr_firstobj)
2321                 return (SET_ERROR(EINVAL));
2322 
2323         for (obj = drrfo->drr_firstobj;
2324             obj < drrfo->drr_firstobj + drrfo->drr_numobjs && next_err == 0;
2325             next_err = dmu_object_next(rwa->os, &obj, FALSE, 0)) {
2326                 int err;
2327 
2328                 if (dmu_object_info(rwa->os, obj, NULL) != 0)
2329                         continue;
2330 
2331                 err = dmu_free_long_object(rwa->os, obj);
2332                 if (err != 0)
2333                         return (err);
2334         }
2335         if (next_err != ESRCH)
2336                 return (next_err);
2337         return (0);
2338 }
2339 
2340 static int
2341 receive_write(struct receive_writer_arg *rwa, struct drr_write *drrw,
2342     arc_buf_t *abuf)
2343 {
2344         dmu_tx_t *tx;
2345         int err;
2346 
2347         if (drrw->drr_offset + drrw->drr_logical_size < drrw->drr_offset ||
2348             !DMU_OT_IS_VALID(drrw->drr_type))
2349                 return (SET_ERROR(EINVAL));
2350 
2351         /*
2352          * For resuming to work, records must be in increasing order
2353          * by (object, offset).
2354          */
2355         if (drrw->drr_object < rwa->last_object ||
2356             (drrw->drr_object == rwa->last_object &&
2357             drrw->drr_offset < rwa->last_offset)) {
2358                 return (SET_ERROR(EINVAL));
2359         }
2360         rwa->last_object = drrw->drr_object;
2361         rwa->last_offset = drrw->drr_offset;
2362 
2363         if (dmu_object_info(rwa->os, drrw->drr_object, NULL) != 0)
2364                 return (SET_ERROR(EINVAL));
2365 
2366         tx = dmu_tx_create(rwa->os);
2367 
2368         dmu_tx_hold_write(tx, drrw->drr_object,
2369             drrw->drr_offset, drrw->drr_logical_size);
2370         err = dmu_tx_assign(tx, TXG_WAIT);
2371         if (err != 0) {
2372                 dmu_tx_abort(tx);
2373                 return (err);
2374         }
2375 
2376         if (rwa->byteswap) {
2377                 dmu_object_byteswap_t byteswap =
2378                     DMU_OT_BYTESWAP(drrw->drr_type);
2379                 dmu_ot_byteswap[byteswap].ob_func(abuf->b_data,
2380                     DRR_WRITE_PAYLOAD_SIZE(drrw));
2381         }
2382 
2383         /* use the bonus buf to look up the dnode in dmu_assign_arcbuf */
2384         dmu_buf_t *bonus;
2385         if (dmu_bonus_hold(rwa->os, drrw->drr_object, FTAG, &bonus) != 0)
2386                 return (SET_ERROR(EINVAL));
2387         dmu_assign_arcbuf(bonus, drrw->drr_offset, abuf, tx);
2388 
2389         /*
2390          * Note: If the receive fails, we want the resume stream to start
2391          * with the same record that we last successfully received (as opposed
2392          * to the next record), so that we can verify that we are
2393          * resuming from the correct location.
2394          */
2395         save_resume_state(rwa, drrw->drr_object, drrw->drr_offset, tx);
2396         dmu_tx_commit(tx);
2397         dmu_buf_rele(bonus, FTAG);
2398 
2399         return (0);
2400 }
2401 
2402 /*
2403  * Handle a DRR_WRITE_BYREF record.  This record is used in dedup'ed
2404  * streams to refer to a copy of the data that is already on the
2405  * system because it came in earlier in the stream.  This function
2406  * finds the earlier copy of the data, and uses that copy instead of
2407  * data from the stream to fulfill this write.
2408  */
2409 static int
2410 receive_write_byref(struct receive_writer_arg *rwa,
2411     struct drr_write_byref *drrwbr)
2412 {
2413         dmu_tx_t *tx;
2414         int err;
2415         guid_map_entry_t gmesrch;
2416         guid_map_entry_t *gmep;
2417         avl_index_t where;
2418         objset_t *ref_os = NULL;
2419         dmu_buf_t *dbp;
2420 
2421         if (drrwbr->drr_offset + drrwbr->drr_length < drrwbr->drr_offset)
2422                 return (SET_ERROR(EINVAL));
2423 
2424         /*
2425          * If the GUID of the referenced dataset is different from the
2426          * GUID of the target dataset, find the referenced dataset.
2427          */
2428         if (drrwbr->drr_toguid != drrwbr->drr_refguid) {
2429                 gmesrch.guid = drrwbr->drr_refguid;
2430                 if ((gmep = avl_find(rwa->guid_to_ds_map, &gmesrch,
2431                     &where)) == NULL) {
2432                         return (SET_ERROR(EINVAL));
2433                 }
2434                 if (dmu_objset_from_ds(gmep->gme_ds, &ref_os))
2435                         return (SET_ERROR(EINVAL));
2436         } else {
2437                 ref_os = rwa->os;
2438         }
2439 
2440         err = dmu_buf_hold(ref_os, drrwbr->drr_refobject,
2441             drrwbr->drr_refoffset, FTAG, &dbp, DMU_READ_PREFETCH);
2442         if (err != 0)
2443                 return (err);
2444 
2445         tx = dmu_tx_create(rwa->os);
2446 
2447         dmu_tx_hold_write(tx, drrwbr->drr_object,
2448             drrwbr->drr_offset, drrwbr->drr_length);
2449         err = dmu_tx_assign(tx, TXG_WAIT);
2450         if (err != 0) {
2451                 dmu_tx_abort(tx);
2452                 return (err);
2453         }
2454         dmu_write(rwa->os, drrwbr->drr_object,
2455             drrwbr->drr_offset, drrwbr->drr_length, dbp->db_data, tx);
2456         dmu_buf_rele(dbp, FTAG);
2457 
2458         /* See comment in restore_write. */
2459         save_resume_state(rwa, drrwbr->drr_object, drrwbr->drr_offset, tx);
2460         dmu_tx_commit(tx);
2461         return (0);
2462 }
2463 
2464 static int
2465 receive_write_embedded(struct receive_writer_arg *rwa,
2466     struct drr_write_embedded *drrwe, void *data)
2467 {
2468         dmu_tx_t *tx;
2469         int err;
2470 
2471         if (drrwe->drr_offset + drrwe->drr_length < drrwe->drr_offset)
2472                 return (EINVAL);
2473 
2474         if (drrwe->drr_psize > BPE_PAYLOAD_SIZE)
2475                 return (EINVAL);
2476 
2477         if (drrwe->drr_etype >= NUM_BP_EMBEDDED_TYPES)
2478                 return (EINVAL);
2479         if (drrwe->drr_compression >= ZIO_COMPRESS_FUNCTIONS)
2480                 return (EINVAL);
2481 
2482         tx = dmu_tx_create(rwa->os);
2483 
2484         dmu_tx_hold_write(tx, drrwe->drr_object,
2485             drrwe->drr_offset, drrwe->drr_length);
2486         err = dmu_tx_assign(tx, TXG_WAIT);
2487         if (err != 0) {
2488                 dmu_tx_abort(tx);
2489                 return (err);
2490         }
2491 
2492         dmu_write_embedded(rwa->os, drrwe->drr_object,
2493             drrwe->drr_offset, data, drrwe->drr_etype,
2494             drrwe->drr_compression, drrwe->drr_lsize, drrwe->drr_psize,
2495             rwa->byteswap ^ ZFS_HOST_BYTEORDER, tx);
2496 
2497         /* See comment in restore_write. */
2498         save_resume_state(rwa, drrwe->drr_object, drrwe->drr_offset, tx);
2499         dmu_tx_commit(tx);
2500         return (0);
2501 }
2502 
2503 static int
2504 receive_spill(struct receive_writer_arg *rwa, struct drr_spill *drrs,
2505     void *data)
2506 {
2507         dmu_tx_t *tx;
2508         dmu_buf_t *db, *db_spill;
2509         int err;
2510 
2511         if (drrs->drr_length < SPA_MINBLOCKSIZE ||
2512             drrs->drr_length > spa_maxblocksize(dmu_objset_spa(rwa->os)))
2513                 return (SET_ERROR(EINVAL));
2514 
2515         if (dmu_object_info(rwa->os, drrs->drr_object, NULL) != 0)
2516                 return (SET_ERROR(EINVAL));
2517 
2518         VERIFY0(dmu_bonus_hold(rwa->os, drrs->drr_object, FTAG, &db));
2519         if ((err = dmu_spill_hold_by_bonus(db, FTAG, &db_spill)) != 0) {
2520                 dmu_buf_rele(db, FTAG);
2521                 return (err);
2522         }
2523 
2524         tx = dmu_tx_create(rwa->os);
2525 
2526         dmu_tx_hold_spill(tx, db->db_object);
2527 
2528         err = dmu_tx_assign(tx, TXG_WAIT);
2529         if (err != 0) {
2530                 dmu_buf_rele(db, FTAG);
2531                 dmu_buf_rele(db_spill, FTAG);
2532                 dmu_tx_abort(tx);
2533                 return (err);
2534         }
2535         dmu_buf_will_dirty(db_spill, tx);
2536 
2537         if (db_spill->db_size < drrs->drr_length)
2538                 VERIFY(0 == dbuf_spill_set_blksz(db_spill,
2539                     drrs->drr_length, tx));
2540         bcopy(data, db_spill->db_data, drrs->drr_length);
2541 
2542         dmu_buf_rele(db, FTAG);
2543         dmu_buf_rele(db_spill, FTAG);
2544 
2545         dmu_tx_commit(tx);
2546         return (0);
2547 }
2548 
2549 /* ARGSUSED */
2550 static int
2551 receive_free(struct receive_writer_arg *rwa, struct drr_free *drrf)
2552 {
2553         int err;
2554 
2555         if (drrf->drr_length != -1ULL &&
2556             drrf->drr_offset + drrf->drr_length < drrf->drr_offset)
2557                 return (SET_ERROR(EINVAL));
2558 
2559         if (dmu_object_info(rwa->os, drrf->drr_object, NULL) != 0)
2560                 return (SET_ERROR(EINVAL));
2561 
2562         err = dmu_free_long_range(rwa->os, drrf->drr_object,
2563             drrf->drr_offset, drrf->drr_length);
2564 
2565         return (err);
2566 }
2567 
2568 /* used to destroy the drc_ds on error */
2569 static void
2570 dmu_recv_cleanup_ds(dmu_recv_cookie_t *drc)
2571 {
2572         if (drc->drc_resumable) {
2573                 /* wait for our resume state to be written to disk */
2574                 txg_wait_synced(drc->drc_ds->ds_dir->dd_pool, 0);
2575                 dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
2576         } else {
2577                 char name[ZFS_MAX_DATASET_NAME_LEN];
2578                 dsl_dataset_name(drc->drc_ds, name);
2579                 dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
2580                 (void) dsl_destroy_head(name);
2581         }
2582 }
2583 
2584 static void
2585 receive_cksum(struct receive_arg *ra, int len, void *buf)
2586 {
2587         if (ra->byteswap) {
2588                 (void) fletcher_4_incremental_byteswap(buf, len, &ra->cksum);
2589         } else {
2590                 (void) fletcher_4_incremental_native(buf, len, &ra->cksum);
2591         }
2592 }
2593 
2594 /*
2595  * Read the payload into a buffer of size len, and update the current record's
2596  * payload field.
2597  * Allocate ra->next_rrd and read the next record's header into
2598  * ra->next_rrd->header.
2599  * Verify checksum of payload and next record.
2600  */
2601 static int
2602 receive_read_payload_and_next_header(struct receive_arg *ra, int len, void *buf)
2603 {
2604         int err;
2605         boolean_t checksum_enable = (ra->krrp_task == NULL ||
2606             ra->krrp_task->buffer_args.force_cksum);
2607 
2608         if (len != 0) {
2609                 ASSERT3U(len, <=, SPA_MAXBLOCKSIZE);
2610                 err = receive_read(ra, len, buf);
2611                 if (err != 0)
2612                         return (err);
2613                 receive_cksum(ra, len, buf);
2614 
2615                 /* note: rrd is NULL when reading the begin record's payload */
2616                 if (ra->rrd != NULL) {
2617                         ra->rrd->payload = buf;
2618                         ra->rrd->payload_size = len;
2619                         ra->rrd->bytes_read = ra->bytes_read;
2620                 }
2621         }
2622 
2623         ra->prev_cksum = ra->cksum;
2624 
2625         ra->next_rrd = kmem_zalloc(sizeof (*ra->next_rrd), KM_SLEEP);
2626         err = receive_read(ra, sizeof (ra->next_rrd->header),
2627             &ra->next_rrd->header);
2628         ra->next_rrd->bytes_read = ra->bytes_read;
2629         if (err != 0) {
2630                 kmem_free(ra->next_rrd, sizeof (*ra->next_rrd));
2631                 ra->next_rrd = NULL;
2632                 return (err);
2633         }
2634         if (ra->next_rrd->header.drr_type == DRR_BEGIN) {
2635                 kmem_free(ra->next_rrd, sizeof (*ra->next_rrd));
2636                 ra->next_rrd = NULL;
2637                 return (SET_ERROR(EINVAL));
2638         }
2639 
2640         if (checksum_enable) {
2641                 /*
2642                  * Note: checksum is of everything up to but not including the
2643                  * checksum itself.
2644                  */
2645                 ASSERT3U(offsetof(dmu_replay_record_t,
2646                     drr_u.drr_checksum.drr_checksum),
2647                     ==, sizeof (dmu_replay_record_t) - sizeof (zio_cksum_t));
2648                 receive_cksum(ra,
2649                     offsetof(dmu_replay_record_t,
2650                     drr_u.drr_checksum.drr_checksum),
2651                     &ra->next_rrd->header);
2652 
2653                 zio_cksum_t cksum_orig =
2654                     ra->next_rrd->header.drr_u.drr_checksum.drr_checksum;
2655                 zio_cksum_t *cksump =
2656                     &ra->next_rrd->header.drr_u.drr_checksum.drr_checksum;
2657 
2658                 if (ra->byteswap)
2659                         byteswap_record(&ra->next_rrd->header);
2660 
2661                 if ((!ZIO_CHECKSUM_IS_ZERO(cksump)) &&
2662                     !ZIO_CHECKSUM_EQUAL(ra->cksum, *cksump)) {
2663                         kmem_free(ra->next_rrd, sizeof (*ra->next_rrd));
2664                         ra->next_rrd = NULL;
2665                         return (SET_ERROR(ECKSUM));
2666                 }
2667 
2668                 receive_cksum(ra, sizeof (cksum_orig), &cksum_orig);
2669         }
2670 
2671         return (0);
2672 }
2673 
2674 static void
2675 objlist_create(struct objlist *list)
2676 {
2677         list_create(&list->list, sizeof (struct receive_objnode),
2678             offsetof(struct receive_objnode, node));
2679         list->last_lookup = 0;
2680 }
2681 
2682 static void
2683 objlist_destroy(struct objlist *list)
2684 {
2685         for (struct receive_objnode *n = list_remove_head(&list->list);
2686             n != NULL; n = list_remove_head(&list->list)) {
2687                 kmem_free(n, sizeof (*n));
2688         }
2689         list_destroy(&list->list);
2690 }
2691 
2692 /*
2693  * This function looks through the objlist to see if the specified object number
2694  * is contained in the objlist.  In the process, it will remove all object
2695  * numbers in the list that are smaller than the specified object number.  Thus,
2696  * any lookup of an object number smaller than a previously looked up object
2697  * number will always return false; therefore, all lookups should be done in
2698  * ascending order.
2699  */
2700 static boolean_t
2701 objlist_exists(struct objlist *list, uint64_t object)
2702 {
2703         struct receive_objnode *node = list_head(&list->list);
2704         ASSERT3U(object, >=, list->last_lookup);
2705         list->last_lookup = object;
2706         while (node != NULL && node->object < object) {
2707                 VERIFY3P(node, ==, list_remove_head(&list->list));
2708                 kmem_free(node, sizeof (*node));
2709                 node = list_head(&list->list);
2710         }
2711         return (node != NULL && node->object == object);
2712 }
2713 
2714 /*
2715  * The objlist is a list of object numbers stored in ascending order.  However,
2716  * the insertion of new object numbers does not seek out the correct location to
2717  * store a new object number; instead, it appends it to the list for simplicity.
2718  * Thus, any users must take care to only insert new object numbers in ascending
2719  * order.
2720  */
2721 static void
2722 objlist_insert(struct objlist *list, uint64_t object)
2723 {
2724         struct receive_objnode *node = kmem_zalloc(sizeof (*node), KM_SLEEP);
2725         node->object = object;
2726 #ifdef ZFS_DEBUG
2727         struct receive_objnode *last_object = list_tail(&list->list);
2728         uint64_t last_objnum = (last_object != NULL ? last_object->object : 0);
2729         ASSERT3U(node->object, >, last_objnum);
2730 #endif
2731         list_insert_tail(&list->list, node);
2732 }
2733 
2734 /*
2735  * Issue the prefetch reads for any necessary indirect blocks.
2736  *
2737  * We use the object ignore list to tell us whether or not to issue prefetches
2738  * for a given object.  We do this for both correctness (in case the blocksize
2739  * of an object has changed) and performance (if the object doesn't exist, don't
2740  * needlessly try to issue prefetches).  We also trim the list as we go through
2741  * the stream to prevent it from growing to an unbounded size.
2742  *
2743  * The object numbers within will always be in sorted order, and any write
2744  * records we see will also be in sorted order, but they're not sorted with
2745  * respect to each other (i.e. we can get several object records before
2746  * receiving each object's write records).  As a result, once we've reached a
2747  * given object number, we can safely remove any reference to lower object
2748  * numbers in the ignore list. In practice, we receive up to 32 object records
2749  * before receiving write records, so the list can have up to 32 nodes in it.
2750  */
2751 /* ARGSUSED */
2752 static void
2753 receive_read_prefetch(struct receive_arg *ra,
2754     uint64_t object, uint64_t offset, uint64_t length)
2755 {
2756         if (!objlist_exists(&ra->ignore_objlist, object)) {
2757                 dmu_prefetch(ra->os, object, 1, offset, length,
2758                     ZIO_PRIORITY_SYNC_READ);
2759         }
2760 }
2761 
2762 /*
2763  * Read records off the stream, issuing any necessary prefetches.
2764  */
2765 static int
2766 receive_read_record(struct receive_arg *ra)
2767 {
2768         int err;
2769 
2770         switch (ra->rrd->header.drr_type) {
2771         case DRR_OBJECT:
2772         {
2773                 struct drr_object *drro = &ra->rrd->header.drr_u.drr_object;
2774                 uint32_t size = P2ROUNDUP(drro->drr_bonuslen, 8);
2775                 void *buf = kmem_zalloc(size, KM_SLEEP);
2776                 dmu_object_info_t doi;
2777                 err = receive_read_payload_and_next_header(ra, size, buf);
2778                 if (err != 0) {
2779                         kmem_free(buf, size);
2780                         return (err);
2781                 }
2782                 err = dmu_object_info(ra->os, drro->drr_object, &doi);
2783                 /*
2784                  * See receive_read_prefetch for an explanation why we're
2785                  * storing this object in the ignore_obj_list.
2786                  */
2787                 if (err == ENOENT ||
2788                     (err == 0 && doi.doi_data_block_size != drro->drr_blksz)) {
2789                         objlist_insert(&ra->ignore_objlist, drro->drr_object);
2790                         err = 0;
2791                 }
2792                 return (err);
2793         }
2794         case DRR_FREEOBJECTS:
2795         {
2796                 err = receive_read_payload_and_next_header(ra, 0, NULL);
2797                 return (err);
2798         }
2799         case DRR_WRITE:
2800         {
2801                 struct drr_write *drrw = &ra->rrd->header.drr_u.drr_write;
2802                 arc_buf_t *abuf;
2803                 boolean_t is_meta = DMU_OT_IS_METADATA(drrw->drr_type);
2804                 if (DRR_WRITE_COMPRESSED(drrw)) {
2805                         ASSERT3U(drrw->drr_compressed_size, >, 0);
2806                         ASSERT3U(drrw->drr_logical_size, >=,
2807                             drrw->drr_compressed_size);
2808                         ASSERT(!is_meta);
2809                         abuf = arc_loan_compressed_buf(
2810                             dmu_objset_spa(ra->os),
2811                             drrw->drr_compressed_size, drrw->drr_logical_size,
2812                             drrw->drr_compressiontype);
2813                 } else {
2814                         abuf = arc_loan_buf(dmu_objset_spa(ra->os),
2815                             is_meta, drrw->drr_logical_size);
2816                 }
2817 
2818                 err = receive_read_payload_and_next_header(ra,
2819                     DRR_WRITE_PAYLOAD_SIZE(drrw), abuf->b_data);
2820                 if (err != 0) {
2821                         dmu_return_arcbuf(abuf);
2822                         return (err);
2823                 }
2824                 ra->rrd->write_buf = abuf;
2825                 receive_read_prefetch(ra, drrw->drr_object, drrw->drr_offset,
2826                     drrw->drr_logical_size);
2827                 return (err);
2828         }
2829         case DRR_WRITE_BYREF:
2830         {
2831                 struct drr_write_byref *drrwb =
2832                     &ra->rrd->header.drr_u.drr_write_byref;
2833                 err = receive_read_payload_and_next_header(ra, 0, NULL);
2834                 receive_read_prefetch(ra, drrwb->drr_object, drrwb->drr_offset,
2835                     drrwb->drr_length);
2836                 return (err);
2837         }
2838         case DRR_WRITE_EMBEDDED:
2839         {
2840                 struct drr_write_embedded *drrwe =
2841                     &ra->rrd->header.drr_u.drr_write_embedded;
2842                 uint32_t size = P2ROUNDUP(drrwe->drr_psize, 8);
2843                 void *buf = kmem_zalloc(size, KM_SLEEP);
2844 
2845                 err = receive_read_payload_and_next_header(ra, size, buf);
2846                 if (err != 0) {
2847                         kmem_free(buf, size);
2848                         return (err);
2849                 }
2850 
2851                 receive_read_prefetch(ra, drrwe->drr_object, drrwe->drr_offset,
2852                     drrwe->drr_length);
2853                 return (err);
2854         }
2855         case DRR_FREE:
2856         {
2857                 /*
2858                  * It might be beneficial to prefetch indirect blocks here, but
2859                  * we don't really have the data to decide for sure.
2860                  */
2861                 err = receive_read_payload_and_next_header(ra, 0, NULL);
2862                 return (err);
2863         }
2864         case DRR_END:
2865         {
2866                 if (ra->krrp_task == NULL ||
2867                     ra->krrp_task->buffer_args.force_cksum) {
2868                         struct drr_end *drre = &ra->rrd->header.drr_u.drr_end;
2869                         if (!ZIO_CHECKSUM_EQUAL(ra->prev_cksum,
2870                             drre->drr_checksum))
2871                                 return (SET_ERROR(ECKSUM));
2872                 }
2873                 return (0);
2874         }
2875         case DRR_SPILL:
2876         {
2877                 struct drr_spill *drrs = &ra->rrd->header.drr_u.drr_spill;
2878                 void *buf = kmem_zalloc(drrs->drr_length, KM_SLEEP);
2879                 err = receive_read_payload_and_next_header(ra, drrs->drr_length,
2880                     buf);
2881                 if (err != 0)
2882                         kmem_free(buf, drrs->drr_length);
2883                 return (err);
2884         }
2885         default:
2886                 return (SET_ERROR(EINVAL));
2887         }
2888 }
2889 
2890 /*
2891  * Commit the records to the pool.
2892  */
2893 static int
2894 receive_process_record(struct receive_writer_arg *rwa,
2895     struct receive_record_arg *rrd)
2896 {
2897         int err;
2898 
2899         /* Processing in order, therefore bytes_read should be increasing. */
2900         ASSERT3U(rrd->bytes_read, >=, rwa->bytes_read);
2901         rwa->bytes_read = rrd->bytes_read;
2902 
2903         switch (rrd->header.drr_type) {
2904         case DRR_OBJECT:
2905         {
2906                 struct drr_object *drro = &rrd->header.drr_u.drr_object;
2907                 err = receive_object(rwa, drro, rrd->payload);
2908                 kmem_free(rrd->payload, rrd->payload_size);
2909                 rrd->payload = NULL;
2910                 return (err);
2911         }
2912         case DRR_FREEOBJECTS:
2913         {
2914                 struct drr_freeobjects *drrfo =
2915                     &rrd->header.drr_u.drr_freeobjects;
2916                 return (receive_freeobjects(rwa, drrfo));
2917         }
2918         case DRR_WRITE:
2919         {
2920                 struct drr_write *drrw = &rrd->header.drr_u.drr_write;
2921                 err = receive_write(rwa, drrw, rrd->write_buf);
2922                 /* if receive_write() is successful, it consumes the arc_buf */
2923                 if (err != 0)
2924                         dmu_return_arcbuf(rrd->write_buf);
2925                 rrd->write_buf = NULL;
2926                 rrd->payload = NULL;
2927                 return (err);
2928         }
2929         case DRR_WRITE_BYREF:
2930         {
2931                 struct drr_write_byref *drrwbr =
2932                     &rrd->header.drr_u.drr_write_byref;
2933                 return (receive_write_byref(rwa, drrwbr));
2934         }
2935         case DRR_WRITE_EMBEDDED:
2936         {
2937                 struct drr_write_embedded *drrwe =
2938                     &rrd->header.drr_u.drr_write_embedded;
2939                 err = receive_write_embedded(rwa, drrwe, rrd->payload);
2940                 kmem_free(rrd->payload, rrd->payload_size);
2941                 rrd->payload = NULL;
2942                 return (err);
2943         }
2944         case DRR_FREE:
2945         {
2946                 struct drr_free *drrf = &rrd->header.drr_u.drr_free;
2947                 return (receive_free(rwa, drrf));
2948         }
2949         case DRR_SPILL:
2950         {
2951                 struct drr_spill *drrs = &rrd->header.drr_u.drr_spill;
2952                 err = receive_spill(rwa, drrs, rrd->payload);
2953                 kmem_free(rrd->payload, rrd->payload_size);
2954                 rrd->payload = NULL;
2955                 return (err);
2956         }
2957         default:
2958                 return (SET_ERROR(EINVAL));
2959         }
2960 }
2961 
2962 /*
2963  * dmu_recv_stream's worker thread; pull records off the queue, and then call
2964  * receive_process_record  When we're done, signal the main thread and exit.
2965  */
2966 static void
2967 receive_writer_thread(void *arg)
2968 {
2969         struct receive_writer_arg *rwa = arg;
2970         struct receive_record_arg *rrd;
2971         for (rrd = bqueue_dequeue(&rwa->q); !rrd->eos_marker;
2972             rrd = bqueue_dequeue(&rwa->q)) {
2973                 /*
2974                  * If there's an error, the main thread will stop putting things
2975                  * on the queue, but we need to clear everything in it before we
2976                  * can exit.
2977                  */
2978                 if (rwa->err == 0) {
2979                         rwa->err = receive_process_record(rwa, rrd);
2980                 } else if (rrd->write_buf != NULL) {
2981                         dmu_return_arcbuf(rrd->write_buf);
2982                         rrd->write_buf = NULL;
2983                         rrd->payload = NULL;
2984                 } else if (rrd->payload != NULL) {
2985                         kmem_free(rrd->payload, rrd->payload_size);
2986                         rrd->payload = NULL;
2987                 }
2988                 kmem_free(rrd, sizeof (*rrd));
2989         }
2990         kmem_free(rrd, sizeof (*rrd));
2991         mutex_enter(&rwa->mutex);
2992         rwa->done = B_TRUE;
2993         cv_signal(&rwa->cv);
2994         mutex_exit(&rwa->mutex);
2995         thread_exit();
2996 }
2997 
2998 static int
2999 resume_check(struct receive_arg *ra, nvlist_t *begin_nvl)
3000 {
3001         uint64_t val;
3002         objset_t *mos = dmu_objset_pool(ra->os)->dp_meta_objset;
3003         uint64_t dsobj = dmu_objset_id(ra->os);
3004         uint64_t resume_obj, resume_off;
3005 
3006         if (nvlist_lookup_uint64(begin_nvl,
3007             "resume_object", &resume_obj) != 0 ||
3008             nvlist_lookup_uint64(begin_nvl,
3009             "resume_offset", &resume_off) != 0) {
3010                 return (SET_ERROR(EINVAL));
3011         }
3012         VERIFY0(zap_lookup(mos, dsobj,
3013             DS_FIELD_RESUME_OBJECT, sizeof (val), 1, &val));
3014         if (resume_obj != val)
3015                 return (SET_ERROR(EINVAL));
3016         VERIFY0(zap_lookup(mos, dsobj,
3017             DS_FIELD_RESUME_OFFSET, sizeof (val), 1, &val));
3018         if (resume_off != val)
3019                 return (SET_ERROR(EINVAL));
3020 
3021         return (0);
3022 }
3023 
3024 /*
3025  * Read in the stream's records, one by one, and apply them to the pool.  There
3026  * are two threads involved; the thread that calls this function will spin up a
3027  * worker thread, read the records off the stream one by one, and issue
3028  * prefetches for any necessary indirect blocks.  It will then push the records
3029  * onto an internal blocking queue.  The worker thread will pull the records off
3030  * the queue, and actually write the data into the DMU.  This way, the worker
3031  * thread doesn't have to wait for reads to complete, since everything it needs
3032  * (the indirect blocks) will be prefetched.
3033  *
3034  * NB: callers *must* call dmu_recv_end() if this succeeds.
3035  */
3036 int
3037 dmu_recv_stream(dmu_recv_cookie_t *drc, vnode_t *vp, offset_t *voffp,
3038     int cleanup_fd, uint64_t *action_handlep, dmu_krrp_task_t *krrp_task)
3039 {
3040         int err = 0;
3041         struct receive_arg ra = { 0 };
3042         struct receive_writer_arg rwa = { 0 };
3043         int featureflags;
3044         nvlist_t *begin_nvl = NULL;
3045 
3046         ra.byteswap = drc->drc_byteswap;
3047         ra.cksum = drc->drc_cksum;
3048         ra.vp = vp;
3049         ra.voff = *voffp;
3050         ra.krrp_task = krrp_task;
3051 
3052         if (dsl_dataset_is_zapified(drc->drc_ds)) {
3053                 (void) zap_lookup(drc->drc_ds->ds_dir->dd_pool->dp_meta_objset,
3054                     drc->drc_ds->ds_object, DS_FIELD_RESUME_BYTES,
3055                     sizeof (ra.bytes_read), 1, &ra.bytes_read);
3056         }
3057 
3058         objlist_create(&ra.ignore_objlist);
3059 
3060         /* these were verified in dmu_recv_begin */
3061         ASSERT3U(DMU_GET_STREAM_HDRTYPE(drc->drc_drrb->drr_versioninfo), ==,
3062             DMU_SUBSTREAM);
3063         ASSERT3U(drc->drc_drrb->drr_type, <, DMU_OST_NUMTYPES);
3064 
3065         /*
3066          * Open the objset we are modifying.
3067          */
3068         VERIFY0(dmu_objset_from_ds(drc->drc_ds, &ra.os));
3069 
3070         ASSERT(dsl_dataset_phys(drc->drc_ds)->ds_flags & DS_FLAG_INCONSISTENT);
3071 
3072         featureflags = DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo);
3073 
3074         /* if this stream is dedup'ed, set up the avl tree for guid mapping */
3075         if (featureflags & DMU_BACKUP_FEATURE_DEDUP) {
3076                 minor_t minor;
3077 
3078                 if (cleanup_fd == -1) {
3079                         ra.err = SET_ERROR(EBADF);
3080                         goto out;
3081                 }
3082                 ra.err = zfs_onexit_fd_hold(cleanup_fd, &minor);
3083                 if (ra.err != 0) {
3084                         cleanup_fd = -1;
3085                         goto out;
3086                 }
3087 
3088                 if (*action_handlep == 0) {
3089                         rwa.guid_to_ds_map =
3090                             kmem_alloc(sizeof (avl_tree_t), KM_SLEEP);
3091                         avl_create(rwa.guid_to_ds_map, guid_compare,
3092                             sizeof (guid_map_entry_t),
3093                             offsetof(guid_map_entry_t, avlnode));
3094                         err = zfs_onexit_add_cb(minor,
3095                             free_guid_map_onexit, rwa.guid_to_ds_map,
3096                             action_handlep);
3097                         if (ra.err != 0)
3098                                 goto out;
3099                 } else {
3100                         err = zfs_onexit_cb_data(minor, *action_handlep,
3101                             (void **)&rwa.guid_to_ds_map);
3102                         if (ra.err != 0)
3103                                 goto out;
3104                 }
3105 
3106                 drc->drc_guid_to_ds_map = rwa.guid_to_ds_map;
3107         }
3108 
3109         uint32_t payloadlen = drc->drc_drr_begin->drr_payloadlen;
3110         void *payload = NULL;
3111         if (payloadlen != 0)
3112                 payload = kmem_alloc(payloadlen, KM_SLEEP);
3113 
3114         err = receive_read_payload_and_next_header(&ra, payloadlen, payload);
3115         if (err != 0) {
3116                 if (payloadlen != 0)
3117                         kmem_free(payload, payloadlen);
3118                 goto out;
3119         }
3120         if (payloadlen != 0) {
3121                 err = nvlist_unpack(payload, payloadlen, &begin_nvl, KM_SLEEP);
3122                 kmem_free(payload, payloadlen);
3123                 if (err != 0)
3124                         goto out;
3125         }
3126 
3127         if (featureflags & DMU_BACKUP_FEATURE_RESUMING) {
3128                 err = resume_check(&ra, begin_nvl);
3129                 if (err != 0)
3130                         goto out;
3131         }
3132 
3133         (void) bqueue_init(&rwa.q, zfs_recv_queue_length,
3134             offsetof(struct receive_record_arg, node));
3135         cv_init(&rwa.cv, NULL, CV_DEFAULT, NULL);
3136         mutex_init(&rwa.mutex, NULL, MUTEX_DEFAULT, NULL);
3137         rwa.os = ra.os;
3138         rwa.byteswap = drc->drc_byteswap;
3139         rwa.resumable = drc->drc_resumable;
3140 
3141         (void) thread_create(NULL, 0, receive_writer_thread, &rwa, 0, curproc,
3142             TS_RUN, minclsyspri);
3143         /*
3144          * We're reading rwa.err without locks, which is safe since we are the
3145          * only reader, and the worker thread is the only writer.  It's ok if we
3146          * miss a write for an iteration or two of the loop, since the writer
3147          * thread will keep freeing records we send it until we send it an eos
3148          * marker.
3149          *
3150          * We can leave this loop in 3 ways:  First, if rwa.err is
3151          * non-zero.  In that case, the writer thread will free the rrd we just
3152          * pushed.  Second, if  we're interrupted; in that case, either it's the
3153          * first loop and ra.rrd was never allocated, or it's later, and ra.rrd
3154          * has been handed off to the writer thread who will free it.  Finally,
3155          * if receive_read_record fails or we're at the end of the stream, then
3156          * we free ra.rrd and exit.
3157          */
3158         while (rwa.err == 0) {
3159                 if (vp && issig(JUSTLOOKING) && issig(FORREAL)) {
3160                         err = SET_ERROR(EINTR);
3161                         break;
3162                 }
3163 
3164                 ASSERT3P(ra.rrd, ==, NULL);
3165                 ra.rrd = ra.next_rrd;
3166                 ra.next_rrd = NULL;
3167                 /* Allocates and loads header into ra.next_rrd */
3168                 err = receive_read_record(&ra);
3169 
3170                 if (ra.rrd->header.drr_type == DRR_END || err != 0) {
3171                         kmem_free(ra.rrd, sizeof (*ra.rrd));
3172                         ra.rrd = NULL;
3173                         break;
3174                 }
3175 
3176                 bqueue_enqueue(&rwa.q, ra.rrd,
3177                     sizeof (struct receive_record_arg) + ra.rrd->payload_size);
3178                 ra.rrd = NULL;
3179         }
3180         if (ra.next_rrd == NULL)
3181                 ra.next_rrd = kmem_zalloc(sizeof (*ra.next_rrd), KM_SLEEP);
3182         ra.next_rrd->eos_marker = B_TRUE;
3183         bqueue_enqueue(&rwa.q, ra.next_rrd, 1);
3184 
3185         mutex_enter(&rwa.mutex);
3186         while (!rwa.done) {
3187                 cv_wait(&rwa.cv, &rwa.mutex);
3188         }
3189         mutex_exit(&rwa.mutex);
3190 
3191         cv_destroy(&rwa.cv);
3192         mutex_destroy(&rwa.mutex);
3193         bqueue_destroy(&rwa.q);
3194         if (err == 0)
3195                 err = rwa.err;
3196 
3197 out:
3198         nvlist_free(begin_nvl);
3199         if ((featureflags & DMU_BACKUP_FEATURE_DEDUP) && (cleanup_fd != -1))
3200                 zfs_onexit_fd_rele(cleanup_fd);
3201 
3202         if (err != 0) {
3203                 /*
3204                  * Clean up references. If receive is not resumable,
3205                  * destroy what we created, so we don't leave it in
3206                  * the inconsistent state.
3207                  */
3208                 dmu_recv_cleanup_ds(drc);
3209         }
3210 
3211         *voffp = ra.voff;
3212         objlist_destroy(&ra.ignore_objlist);
3213         return (err);
3214 }
3215 
3216 static int
3217 dmu_recv_end_check(void *arg, dmu_tx_t *tx)
3218 {
3219         dmu_recv_cookie_t *drc = arg;
3220         dsl_pool_t *dp = dmu_tx_pool(tx);
3221         int error;
3222 
3223         ASSERT3P(drc->drc_ds->ds_owner, ==, dmu_recv_tag);
3224 
3225         if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_WBC)) {
3226                 objset_t *os = NULL;
3227 
3228                 error  = dmu_objset_from_ds(drc->drc_ds, &os);
3229                 if (error)
3230                         return (error);
3231 
3232                 /* Recv is impossible into DS that uses WBC */
3233                 if (os->os_wbc_mode != ZFS_WBC_MODE_OFF)
3234                         return (SET_ERROR(EKZFS_WBCNOTSUP));
3235         }
3236 
3237         if (!drc->drc_newfs) {
3238                 dsl_dataset_t *origin_head;
3239 
3240                 error = dsl_dataset_hold(dp, drc->drc_tofs, FTAG, &origin_head);
3241                 if (error != 0)
3242                         return (error);
3243                 if (drc->drc_force) {
3244                         /*
3245                          * We will destroy any snapshots in tofs (i.e. before
3246                          * origin_head) that are after the origin (which is
3247                          * the snap before drc_ds, because drc_ds can not
3248                          * have any snaps of its own).
3249                          */
3250                         uint64_t obj;
3251 
3252                         obj = dsl_dataset_phys(origin_head)->ds_prev_snap_obj;
3253                         while (obj !=
3254                             dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj) {
3255                                 dsl_dataset_t *snap;
3256                                 error = dsl_dataset_hold_obj(dp, obj, FTAG,
3257                                     &snap);
3258                                 if (error != 0)
3259                                         break;
3260                                 if (snap->ds_dir != origin_head->ds_dir)
3261                                         error = SET_ERROR(EINVAL);
3262                                 if (error == 0)  {
3263                                         error = dsl_destroy_snapshot_check_impl(
3264                                             snap, B_FALSE);
3265                                 }
3266                                 obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
3267                                 dsl_dataset_rele(snap, FTAG);
3268                                 if (error != 0)
3269                                         break;
3270                         }
3271                         if (error != 0) {
3272                                 dsl_dataset_rele(origin_head, FTAG);
3273                                 return (error);
3274                         }
3275                 }
3276                 error = dsl_dataset_clone_swap_check_impl(drc->drc_ds,
3277                     origin_head, drc->drc_force, drc->drc_owner, tx);
3278                 if (error != 0) {
3279                         dsl_dataset_rele(origin_head, FTAG);
3280                         return (error);
3281                 }
3282                 error = dsl_dataset_snapshot_check_impl(origin_head,
3283                     drc->drc_tosnap, tx, B_TRUE, 1, drc->drc_cred);
3284                 dsl_dataset_rele(origin_head, FTAG);
3285                 if (error != 0)
3286                         return (error);
3287 
3288                 error = dsl_destroy_head_check_impl(drc->drc_ds, 1);
3289         } else {
3290                 error = dsl_dataset_snapshot_check_impl(drc->drc_ds,
3291                     drc->drc_tosnap, tx, B_TRUE, 1, drc->drc_cred);
3292         }
3293 
3294         if (dmu_tx_is_syncing(tx) && drc->drc_krrp_task != NULL) {
3295                 const char *token =
3296                     drc->drc_krrp_task->buffer_args.to_ds;
3297                 const char *cookie = drc->drc_krrp_task->cookie;
3298                 dsl_pool_t *dp = tx->tx_pool;
3299 
3300                 if (*token != '\0') {
3301                         error = zap_update(dp->dp_meta_objset,
3302                             DMU_POOL_DIRECTORY_OBJECT, token, 1,
3303                             strlen(cookie) + 1, cookie, tx);
3304                 }
3305         }
3306         return (error);
3307 }
3308 
3309 static void
3310 dmu_recv_end_sync(void *arg, dmu_tx_t *tx)
3311 {
3312         dmu_recv_cookie_t *drc = arg;
3313         dsl_pool_t *dp = dmu_tx_pool(tx);
3314 
3315         spa_history_log_internal_ds(drc->drc_ds, "finish receiving",
3316             tx, "snap=%s", drc->drc_tosnap);
3317 
3318         if (!drc->drc_newfs) {
3319                 dsl_dataset_t *origin_head;
3320 
3321                 VERIFY0(dsl_dataset_hold(dp, drc->drc_tofs, FTAG,
3322                     &origin_head));
3323 
3324                 if (drc->drc_force) {
3325                         /*
3326                          * Destroy any snapshots of drc_tofs (origin_head)
3327                          * after the origin (the snap before drc_ds).
3328                          */
3329                         uint64_t obj;
3330 
3331                         obj = dsl_dataset_phys(origin_head)->ds_prev_snap_obj;
3332                         while (obj !=
3333                             dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj) {
3334                                 dsl_dataset_t *snap;
3335                                 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG,
3336                                     &snap));
3337                                 ASSERT3P(snap->ds_dir, ==, origin_head->ds_dir);
3338                                 obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
3339                                 dsl_destroy_snapshot_sync_impl(snap,
3340                                     B_FALSE, tx);
3341                                 dsl_dataset_rele(snap, FTAG);
3342                         }
3343                 }
3344                 VERIFY3P(drc->drc_ds->ds_prev, ==,
3345                     origin_head->ds_prev);
3346 
3347                 dsl_dataset_clone_swap_sync_impl(drc->drc_ds,
3348                     origin_head, tx);
3349                 dsl_dataset_snapshot_sync_impl(origin_head,
3350                     drc->drc_tosnap, tx);
3351 
3352                 /* set snapshot's creation time and guid */
3353                 dmu_buf_will_dirty(origin_head->ds_prev->ds_dbuf, tx);
3354                 dsl_dataset_phys(origin_head->ds_prev)->ds_creation_time =
3355                     drc->drc_drrb->drr_creation_time;
3356                 dsl_dataset_phys(origin_head->ds_prev)->ds_guid =
3357                     drc->drc_drrb->drr_toguid;
3358                 dsl_dataset_phys(origin_head->ds_prev)->ds_flags &=
3359                     ~DS_FLAG_INCONSISTENT;
3360 
3361                 dmu_buf_will_dirty(origin_head->ds_dbuf, tx);
3362                 dsl_dataset_phys(origin_head)->ds_flags &=
3363                     ~DS_FLAG_INCONSISTENT;
3364 
3365                 drc->drc_newsnapobj =
3366                     dsl_dataset_phys(origin_head)->ds_prev_snap_obj;
3367 
3368                 dsl_dataset_rele(origin_head, FTAG);
3369                 dsl_destroy_head_sync_impl(drc->drc_ds, tx);
3370 
3371                 if (drc->drc_owner != NULL)
3372                         VERIFY3P(origin_head->ds_owner, ==, drc->drc_owner);
3373         } else {
3374                 dsl_dataset_t *ds = drc->drc_ds;
3375 
3376                 dsl_dataset_snapshot_sync_impl(ds, drc->drc_tosnap, tx);
3377 
3378                 /* set snapshot's creation time and guid */
3379                 dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
3380                 dsl_dataset_phys(ds->ds_prev)->ds_creation_time =
3381                     drc->drc_drrb->drr_creation_time;
3382                 dsl_dataset_phys(ds->ds_prev)->ds_guid =
3383                     drc->drc_drrb->drr_toguid;
3384                 dsl_dataset_phys(ds->ds_prev)->ds_flags &=
3385                     ~DS_FLAG_INCONSISTENT;
3386 
3387                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
3388                 dsl_dataset_phys(ds)->ds_flags &= ~DS_FLAG_INCONSISTENT;
3389                 if (dsl_dataset_has_resume_receive_state(ds)) {
3390                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3391                             DS_FIELD_RESUME_FROMGUID, tx);
3392                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3393                             DS_FIELD_RESUME_OBJECT, tx);
3394                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3395                             DS_FIELD_RESUME_OFFSET, tx);
3396                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3397                             DS_FIELD_RESUME_BYTES, tx);
3398                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3399                             DS_FIELD_RESUME_TOGUID, tx);
3400                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3401                             DS_FIELD_RESUME_TONAME, tx);
3402                 }
3403                 drc->drc_newsnapobj =
3404                     dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj;
3405         }
3406         /*
3407          * Release the hold from dmu_recv_begin.  This must be done before
3408          * we return to open context, so that when we free the dataset's dnode,
3409          * we can evict its bonus buffer.
3410          */
3411         dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
3412         drc->drc_ds = NULL;
3413 }
3414 
3415 static int
3416 add_ds_to_guidmap(const char *name, avl_tree_t *guid_map, uint64_t snapobj)
3417 {
3418         dsl_pool_t *dp;
3419         dsl_dataset_t *snapds;
3420         guid_map_entry_t *gmep;
3421         int err;
3422 
3423         ASSERT(guid_map != NULL);
3424 
3425         err = dsl_pool_hold(name, FTAG, &dp);
3426         if (err != 0)
3427                 return (err);
3428         gmep = kmem_alloc(sizeof (*gmep), KM_SLEEP);
3429         err = dsl_dataset_hold_obj(dp, snapobj, gmep, &snapds);
3430         if (err == 0) {
3431                 gmep->guid = dsl_dataset_phys(snapds)->ds_guid;
3432                 gmep->gme_ds = snapds;
3433                 avl_add(guid_map, gmep);
3434                 dsl_dataset_long_hold(snapds, gmep);
3435         } else {
3436                 kmem_free(gmep, sizeof (*gmep));
3437         }
3438 
3439         dsl_pool_rele(dp, FTAG);
3440         return (err);
3441 }
3442 
3443 static int dmu_recv_end_modified_blocks = 3;
3444 
3445 static int
3446 dmu_recv_existing_end(dmu_recv_cookie_t *drc)
3447 {
3448 #ifdef _KERNEL
3449         /*
3450          * We will be destroying the ds; make sure its origin is unmounted if
3451          * necessary.
3452          */
3453         char name[ZFS_MAX_DATASET_NAME_LEN];
3454         dsl_dataset_name(drc->drc_ds, name);
3455         zfs_destroy_unmount_origin(name);
3456 #endif
3457 
3458         return (dsl_sync_task(drc->drc_tofs,
3459             dmu_recv_end_check, dmu_recv_end_sync, drc,
3460             dmu_recv_end_modified_blocks, ZFS_SPACE_CHECK_NORMAL));
3461 }
3462 
3463 static int
3464 dmu_recv_new_end(dmu_recv_cookie_t *drc)
3465 {
3466         return (dsl_sync_task(drc->drc_tofs,
3467             dmu_recv_end_check, dmu_recv_end_sync, drc,
3468             dmu_recv_end_modified_blocks, ZFS_SPACE_CHECK_NORMAL));
3469 }
3470 
3471 int
3472 dmu_recv_end(dmu_recv_cookie_t *drc, void *owner)
3473 {
3474         int error;
3475 
3476         drc->drc_owner = owner;
3477 
3478         if (drc->drc_newfs)
3479                 error = dmu_recv_new_end(drc);
3480         else
3481                 error = dmu_recv_existing_end(drc);
3482 
3483         if (error != 0) {
3484                 dmu_recv_cleanup_ds(drc);
3485         } else if (drc->drc_guid_to_ds_map != NULL) {
3486                 (void) add_ds_to_guidmap(drc->drc_tofs,
3487                     drc->drc_guid_to_ds_map,
3488                     drc->drc_newsnapobj);
3489         }
3490         return (error);
3491 }
3492 
3493 /*
3494  * Return TRUE if this objset is currently being received into.
3495  */
3496 boolean_t
3497 dmu_objset_is_receiving(objset_t *os)
3498 {
3499         return (os->os_dsl_dataset != NULL &&
3500             os->os_dsl_dataset->ds_owner == dmu_recv_tag);
3501 }