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 2011 Nexenta Systems, Inc. All rights reserved.
  24  * Copyright (c) 2011, 2015 by Delphix. All rights reserved.
  25  * Copyright (c) 2014, Joyent, Inc. All rights reserved.
  26  * Copyright 2014 HybridCluster. All rights reserved.
  27  */
  28 
  29 #include <sys/dmu.h>
  30 #include <sys/dmu_impl.h>
  31 #include <sys/dmu_tx.h>
  32 #include <sys/dbuf.h>
  33 #include <sys/dnode.h>
  34 #include <sys/zfs_context.h>
  35 #include <sys/dmu_objset.h>
  36 #include <sys/dmu_traverse.h>
  37 #include <sys/dsl_dataset.h>
  38 #include <sys/dsl_dir.h>
  39 #include <sys/dsl_prop.h>
  40 #include <sys/dsl_pool.h>
  41 #include <sys/dsl_synctask.h>
  42 #include <sys/zfs_ioctl.h>
  43 #include <sys/zap.h>
  44 #include <sys/zio_checksum.h>
  45 #include <sys/zfs_znode.h>
  46 #include <zfs_fletcher.h>
  47 #include <sys/avl.h>
  48 #include <sys/ddt.h>
  49 #include <sys/zfs_onexit.h>
  50 #include <sys/dmu_send.h>
  51 #include <sys/dsl_destroy.h>
  52 #include <sys/blkptr.h>
  53 #include <sys/dsl_bookmark.h>
  54 #include <sys/zfeature.h>
  55 #include <sys/bqueue.h>
  56 
  57 /* Set this tunable to TRUE to replace corrupt data with 0x2f5baddb10c */
  58 int zfs_send_corrupt_data = B_FALSE;
  59 int zfs_send_queue_length = 16 * 1024 * 1024;
  60 int zfs_recv_queue_length = 16 * 1024 * 1024;
  61 
  62 static char *dmu_recv_tag = "dmu_recv_tag";
  63 const char *recv_clone_name = "%recv";
  64 
  65 #define BP_SPAN(datablkszsec, indblkshift, level) \
  66         (((uint64_t)datablkszsec) << (SPA_MINBLOCKSHIFT + \
  67         (level) * (indblkshift - SPA_BLKPTRSHIFT)))
  68 
  69 static void byteswap_record(dmu_replay_record_t *drr);
  70 
  71 struct send_thread_arg {
  72         bqueue_t        q;
  73         dsl_dataset_t   *ds;            /* Dataset to traverse */
  74         uint64_t        fromtxg;        /* Traverse from this txg */
  75         int             flags;          /* flags to pass to traverse_dataset */
  76         int             error_code;
  77         boolean_t       cancel;
  78         zbookmark_phys_t resume;
  79 };
  80 
  81 struct send_block_record {
  82         boolean_t               eos_marker; /* Marks the end of the stream */
  83         blkptr_t                bp;
  84         zbookmark_phys_t        zb;
  85         uint8_t                 indblkshift;
  86         uint16_t                datablkszsec;
  87         bqueue_node_t           ln;
  88 };
  89 
  90 static int
  91 dump_bytes(dmu_sendarg_t *dsp, void *buf, int len)
  92 {
  93         dsl_dataset_t *ds = dmu_objset_ds(dsp->dsa_os);
  94         ssize_t resid; /* have to get resid to get detailed errno */
  95         ASSERT0(len % 8);
  96 
  97         dsp->dsa_err = vn_rdwr(UIO_WRITE, dsp->dsa_vp,
  98             (caddr_t)buf, len,
  99             0, UIO_SYSSPACE, FAPPEND, RLIM64_INFINITY, CRED(), &resid);
 100 
 101         mutex_enter(&ds->ds_sendstream_lock);
 102         *dsp->dsa_off += len;
 103         mutex_exit(&ds->ds_sendstream_lock);
 104 
 105         return (dsp->dsa_err);
 106 }
 107 
 108 /*
 109  * For all record types except BEGIN, fill in the checksum (overlaid in
 110  * drr_u.drr_checksum.drr_checksum).  The checksum verifies everything
 111  * up to the start of the checksum itself.
 112  */
 113 static int
 114 dump_record(dmu_sendarg_t *dsp, void *payload, int payload_len)
 115 {
 116         ASSERT3U(offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
 117             ==, sizeof (dmu_replay_record_t) - sizeof (zio_cksum_t));
 118         fletcher_4_incremental_native(dsp->dsa_drr,
 119             offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
 120             &dsp->dsa_zc);
 121         if (dsp->dsa_drr->drr_type != DRR_BEGIN) {
 122                 ASSERT(ZIO_CHECKSUM_IS_ZERO(&dsp->dsa_drr->drr_u.
 123                     drr_checksum.drr_checksum));
 124                 dsp->dsa_drr->drr_u.drr_checksum.drr_checksum = dsp->dsa_zc;
 125         }
 126         fletcher_4_incremental_native(&dsp->dsa_drr->
 127             drr_u.drr_checksum.drr_checksum,
 128             sizeof (zio_cksum_t), &dsp->dsa_zc);
 129         if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)) != 0)
 130                 return (SET_ERROR(EINTR));
 131         if (payload_len != 0) {
 132                 fletcher_4_incremental_native(payload, payload_len,
 133                     &dsp->dsa_zc);
 134                 if (dump_bytes(dsp, payload, payload_len) != 0)
 135                         return (SET_ERROR(EINTR));
 136         }
 137         return (0);
 138 }
 139 
 140 /*
 141  * Fill in the drr_free struct, or perform aggregation if the previous record is
 142  * also a free record, and the two are adjacent.
 143  *
 144  * Note that we send free records even for a full send, because we want to be
 145  * able to receive a full send as a clone, which requires a list of all the free
 146  * and freeobject records that were generated on the source.
 147  */
 148 static int
 149 dump_free(dmu_sendarg_t *dsp, uint64_t object, uint64_t offset,
 150     uint64_t length)
 151 {
 152         struct drr_free *drrf = &(dsp->dsa_drr->drr_u.drr_free);
 153 
 154         /*
 155          * When we receive a free record, dbuf_free_range() assumes
 156          * that the receiving system doesn't have any dbufs in the range
 157          * being freed.  This is always true because there is a one-record
 158          * constraint: we only send one WRITE record for any given
 159          * object,offset.  We know that the one-record constraint is
 160          * true because we always send data in increasing order by
 161          * object,offset.
 162          *
 163          * If the increasing-order constraint ever changes, we should find
 164          * another way to assert that the one-record constraint is still
 165          * satisfied.
 166          */
 167         ASSERT(object > dsp->dsa_last_data_object ||
 168             (object == dsp->dsa_last_data_object &&
 169             offset > dsp->dsa_last_data_offset));
 170 
 171         if (length != -1ULL && offset + length < offset)
 172                 length = -1ULL;
 173 
 174         /*
 175          * If there is a pending op, but it's not PENDING_FREE, push it out,
 176          * since free block aggregation can only be done for blocks of the
 177          * same type (i.e., DRR_FREE records can only be aggregated with
 178          * other DRR_FREE records.  DRR_FREEOBJECTS records can only be
 179          * aggregated with other DRR_FREEOBJECTS records.
 180          */
 181         if (dsp->dsa_pending_op != PENDING_NONE &&
 182             dsp->dsa_pending_op != PENDING_FREE) {
 183                 if (dump_record(dsp, NULL, 0) != 0)
 184                         return (SET_ERROR(EINTR));
 185                 dsp->dsa_pending_op = PENDING_NONE;
 186         }
 187 
 188         if (dsp->dsa_pending_op == PENDING_FREE) {
 189                 /*
 190                  * There should never be a PENDING_FREE if length is -1
 191                  * (because dump_dnode is the only place where this
 192                  * function is called with a -1, and only after flushing
 193                  * any pending record).
 194                  */
 195                 ASSERT(length != -1ULL);
 196                 /*
 197                  * Check to see whether this free block can be aggregated
 198                  * with pending one.
 199                  */
 200                 if (drrf->drr_object == object && drrf->drr_offset +
 201                     drrf->drr_length == offset) {
 202                         drrf->drr_length += length;
 203                         return (0);
 204                 } else {
 205                         /* not a continuation.  Push out pending record */
 206                         if (dump_record(dsp, NULL, 0) != 0)
 207                                 return (SET_ERROR(EINTR));
 208                         dsp->dsa_pending_op = PENDING_NONE;
 209                 }
 210         }
 211         /* create a FREE record and make it pending */
 212         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 213         dsp->dsa_drr->drr_type = DRR_FREE;
 214         drrf->drr_object = object;
 215         drrf->drr_offset = offset;
 216         drrf->drr_length = length;
 217         drrf->drr_toguid = dsp->dsa_toguid;
 218         if (length == -1ULL) {
 219                 if (dump_record(dsp, NULL, 0) != 0)
 220                         return (SET_ERROR(EINTR));
 221         } else {
 222                 dsp->dsa_pending_op = PENDING_FREE;
 223         }
 224 
 225         return (0);
 226 }
 227 
 228 static int
 229 dump_write(dmu_sendarg_t *dsp, dmu_object_type_t type,
 230     uint64_t object, uint64_t offset, int blksz, const blkptr_t *bp, void *data)
 231 {
 232         struct drr_write *drrw = &(dsp->dsa_drr->drr_u.drr_write);
 233 
 234         /*
 235          * We send data in increasing object, offset order.
 236          * See comment in dump_free() for details.
 237          */
 238         ASSERT(object > dsp->dsa_last_data_object ||
 239             (object == dsp->dsa_last_data_object &&
 240             offset > dsp->dsa_last_data_offset));
 241         dsp->dsa_last_data_object = object;
 242         dsp->dsa_last_data_offset = offset + blksz - 1;
 243 
 244         /*
 245          * If there is any kind of pending aggregation (currently either
 246          * a grouping of free objects or free blocks), push it out to
 247          * the stream, since aggregation can't be done across operations
 248          * of different types.
 249          */
 250         if (dsp->dsa_pending_op != PENDING_NONE) {
 251                 if (dump_record(dsp, NULL, 0) != 0)
 252                         return (SET_ERROR(EINTR));
 253                 dsp->dsa_pending_op = PENDING_NONE;
 254         }
 255         /* write a WRITE record */
 256         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 257         dsp->dsa_drr->drr_type = DRR_WRITE;
 258         drrw->drr_object = object;
 259         drrw->drr_type = type;
 260         drrw->drr_offset = offset;
 261         drrw->drr_length = blksz;
 262         drrw->drr_toguid = dsp->dsa_toguid;
 263         if (bp == NULL || BP_IS_EMBEDDED(bp)) {
 264                 /*
 265                  * There's no pre-computed checksum for partial-block
 266                  * writes or embedded BP's, so (like
 267                  * fletcher4-checkummed blocks) userland will have to
 268                  * compute a dedup-capable checksum itself.
 269                  */
 270                 drrw->drr_checksumtype = ZIO_CHECKSUM_OFF;
 271         } else {
 272                 drrw->drr_checksumtype = BP_GET_CHECKSUM(bp);
 273                 if (zio_checksum_table[drrw->drr_checksumtype].ci_flags &
 274                     ZCHECKSUM_FLAG_DEDUP)
 275                         drrw->drr_checksumflags |= DRR_CHECKSUM_DEDUP;
 276                 DDK_SET_LSIZE(&drrw->drr_key, BP_GET_LSIZE(bp));
 277                 DDK_SET_PSIZE(&drrw->drr_key, BP_GET_PSIZE(bp));
 278                 DDK_SET_COMPRESS(&drrw->drr_key, BP_GET_COMPRESS(bp));
 279                 drrw->drr_key.ddk_cksum = bp->blk_cksum;
 280         }
 281 
 282         if (dump_record(dsp, data, blksz) != 0)
 283                 return (SET_ERROR(EINTR));
 284         return (0);
 285 }
 286 
 287 static int
 288 dump_write_embedded(dmu_sendarg_t *dsp, uint64_t object, uint64_t offset,
 289     int blksz, const blkptr_t *bp)
 290 {
 291         char buf[BPE_PAYLOAD_SIZE];
 292         struct drr_write_embedded *drrw =
 293             &(dsp->dsa_drr->drr_u.drr_write_embedded);
 294 
 295         if (dsp->dsa_pending_op != PENDING_NONE) {
 296                 if (dump_record(dsp, NULL, 0) != 0)
 297                         return (EINTR);
 298                 dsp->dsa_pending_op = PENDING_NONE;
 299         }
 300 
 301         ASSERT(BP_IS_EMBEDDED(bp));
 302 
 303         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 304         dsp->dsa_drr->drr_type = DRR_WRITE_EMBEDDED;
 305         drrw->drr_object = object;
 306         drrw->drr_offset = offset;
 307         drrw->drr_length = blksz;
 308         drrw->drr_toguid = dsp->dsa_toguid;
 309         drrw->drr_compression = BP_GET_COMPRESS(bp);
 310         drrw->drr_etype = BPE_GET_ETYPE(bp);
 311         drrw->drr_lsize = BPE_GET_LSIZE(bp);
 312         drrw->drr_psize = BPE_GET_PSIZE(bp);
 313 
 314         decode_embedded_bp_compressed(bp, buf);
 315 
 316         if (dump_record(dsp, buf, P2ROUNDUP(drrw->drr_psize, 8)) != 0)
 317                 return (EINTR);
 318         return (0);
 319 }
 320 
 321 static int
 322 dump_spill(dmu_sendarg_t *dsp, uint64_t object, int blksz, void *data)
 323 {
 324         struct drr_spill *drrs = &(dsp->dsa_drr->drr_u.drr_spill);
 325 
 326         if (dsp->dsa_pending_op != PENDING_NONE) {
 327                 if (dump_record(dsp, NULL, 0) != 0)
 328                         return (SET_ERROR(EINTR));
 329                 dsp->dsa_pending_op = PENDING_NONE;
 330         }
 331 
 332         /* write a SPILL record */
 333         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 334         dsp->dsa_drr->drr_type = DRR_SPILL;
 335         drrs->drr_object = object;
 336         drrs->drr_length = blksz;
 337         drrs->drr_toguid = dsp->dsa_toguid;
 338 
 339         if (dump_record(dsp, data, blksz) != 0)
 340                 return (SET_ERROR(EINTR));
 341         return (0);
 342 }
 343 
 344 static int
 345 dump_freeobjects(dmu_sendarg_t *dsp, uint64_t firstobj, uint64_t numobjs)
 346 {
 347         struct drr_freeobjects *drrfo = &(dsp->dsa_drr->drr_u.drr_freeobjects);
 348 
 349         /*
 350          * If there is a pending op, but it's not PENDING_FREEOBJECTS,
 351          * push it out, since free block aggregation can only be done for
 352          * blocks of the same type (i.e., DRR_FREE records can only be
 353          * aggregated with other DRR_FREE records.  DRR_FREEOBJECTS records
 354          * can only be aggregated with other DRR_FREEOBJECTS records.
 355          */
 356         if (dsp->dsa_pending_op != PENDING_NONE &&
 357             dsp->dsa_pending_op != PENDING_FREEOBJECTS) {
 358                 if (dump_record(dsp, NULL, 0) != 0)
 359                         return (SET_ERROR(EINTR));
 360                 dsp->dsa_pending_op = PENDING_NONE;
 361         }
 362         if (dsp->dsa_pending_op == PENDING_FREEOBJECTS) {
 363                 /*
 364                  * See whether this free object array can be aggregated
 365                  * with pending one
 366                  */
 367                 if (drrfo->drr_firstobj + drrfo->drr_numobjs == firstobj) {
 368                         drrfo->drr_numobjs += numobjs;
 369                         return (0);
 370                 } else {
 371                         /* can't be aggregated.  Push out pending record */
 372                         if (dump_record(dsp, NULL, 0) != 0)
 373                                 return (SET_ERROR(EINTR));
 374                         dsp->dsa_pending_op = PENDING_NONE;
 375                 }
 376         }
 377 
 378         /* write a FREEOBJECTS record */
 379         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 380         dsp->dsa_drr->drr_type = DRR_FREEOBJECTS;
 381         drrfo->drr_firstobj = firstobj;
 382         drrfo->drr_numobjs = numobjs;
 383         drrfo->drr_toguid = dsp->dsa_toguid;
 384 
 385         dsp->dsa_pending_op = PENDING_FREEOBJECTS;
 386 
 387         return (0);
 388 }
 389 
 390 static int
 391 dump_dnode(dmu_sendarg_t *dsp, uint64_t object, dnode_phys_t *dnp)
 392 {
 393         struct drr_object *drro = &(dsp->dsa_drr->drr_u.drr_object);
 394 
 395         if (object < dsp->dsa_resume_object) {
 396                 /*
 397                  * Note: when resuming, we will visit all the dnodes in
 398                  * the block of dnodes that we are resuming from.  In
 399                  * this case it's unnecessary to send the dnodes prior to
 400                  * the one we are resuming from.  We should be at most one
 401                  * block's worth of dnodes behind the resume point.
 402                  */
 403                 ASSERT3U(dsp->dsa_resume_object - object, <,
 404                     1 << (DNODE_BLOCK_SHIFT - DNODE_SHIFT));
 405                 return (0);
 406         }
 407 
 408         if (dnp == NULL || dnp->dn_type == DMU_OT_NONE)
 409                 return (dump_freeobjects(dsp, object, 1));
 410 
 411         if (dsp->dsa_pending_op != PENDING_NONE) {
 412                 if (dump_record(dsp, NULL, 0) != 0)
 413                         return (SET_ERROR(EINTR));
 414                 dsp->dsa_pending_op = PENDING_NONE;
 415         }
 416 
 417         /* write an OBJECT record */
 418         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 419         dsp->dsa_drr->drr_type = DRR_OBJECT;
 420         drro->drr_object = object;
 421         drro->drr_type = dnp->dn_type;
 422         drro->drr_bonustype = dnp->dn_bonustype;
 423         drro->drr_blksz = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT;
 424         drro->drr_bonuslen = dnp->dn_bonuslen;
 425         drro->drr_checksumtype = dnp->dn_checksum;
 426         drro->drr_compress = dnp->dn_compress;
 427         drro->drr_toguid = dsp->dsa_toguid;
 428 
 429         if (!(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
 430             drro->drr_blksz > SPA_OLD_MAXBLOCKSIZE)
 431                 drro->drr_blksz = SPA_OLD_MAXBLOCKSIZE;
 432 
 433         if (dump_record(dsp, DN_BONUS(dnp),
 434             P2ROUNDUP(dnp->dn_bonuslen, 8)) != 0) {
 435                 return (SET_ERROR(EINTR));
 436         }
 437 
 438         /* Free anything past the end of the file. */
 439         if (dump_free(dsp, object, (dnp->dn_maxblkid + 1) *
 440             (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT), -1ULL) != 0)
 441                 return (SET_ERROR(EINTR));
 442         if (dsp->dsa_err != 0)
 443                 return (SET_ERROR(EINTR));
 444         return (0);
 445 }
 446 
 447 static boolean_t
 448 backup_do_embed(dmu_sendarg_t *dsp, const blkptr_t *bp)
 449 {
 450         if (!BP_IS_EMBEDDED(bp))
 451                 return (B_FALSE);
 452 
 453         /*
 454          * Compression function must be legacy, or explicitly enabled.
 455          */
 456         if ((BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_LEGACY_FUNCTIONS &&
 457             !(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_EMBED_DATA_LZ4)))
 458                 return (B_FALSE);
 459 
 460         /*
 461          * Embed type must be explicitly enabled.
 462          */
 463         switch (BPE_GET_ETYPE(bp)) {
 464         case BP_EMBEDDED_TYPE_DATA:
 465                 if (dsp->dsa_featureflags & DMU_BACKUP_FEATURE_EMBED_DATA)
 466                         return (B_TRUE);
 467                 break;
 468         default:
 469                 return (B_FALSE);
 470         }
 471         return (B_FALSE);
 472 }
 473 
 474 /*
 475  * This is the callback function to traverse_dataset that acts as the worker
 476  * thread for dmu_send_impl.
 477  */
 478 /*ARGSUSED*/
 479 static int
 480 send_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
 481     const zbookmark_phys_t *zb, const struct dnode_phys *dnp, void *arg)
 482 {
 483         struct send_thread_arg *sta = arg;
 484         struct send_block_record *record;
 485         uint64_t record_size;
 486         int err = 0;
 487 
 488         ASSERT(zb->zb_object == DMU_META_DNODE_OBJECT ||
 489             zb->zb_object >= sta->resume.zb_object);
 490 
 491         if (sta->cancel)
 492                 return (SET_ERROR(EINTR));
 493 
 494         if (bp == NULL) {
 495                 ASSERT3U(zb->zb_level, ==, ZB_DNODE_LEVEL);
 496                 return (0);
 497         } else if (zb->zb_level < 0) {
 498                 return (0);
 499         }
 500 
 501         record = kmem_zalloc(sizeof (struct send_block_record), KM_SLEEP);
 502         record->eos_marker = B_FALSE;
 503         record->bp = *bp;
 504         record->zb = *zb;
 505         record->indblkshift = dnp->dn_indblkshift;
 506         record->datablkszsec = dnp->dn_datablkszsec;
 507         record_size = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT;
 508         bqueue_enqueue(&sta->q, record, record_size);
 509 
 510         return (err);
 511 }
 512 
 513 /*
 514  * This function kicks off the traverse_dataset.  It also handles setting the
 515  * error code of the thread in case something goes wrong, and pushes the End of
 516  * Stream record when the traverse_dataset call has finished.  If there is no
 517  * dataset to traverse, the thread immediately pushes End of Stream marker.
 518  */
 519 static void
 520 send_traverse_thread(void *arg)
 521 {
 522         struct send_thread_arg *st_arg = arg;
 523         int err;
 524         struct send_block_record *data;
 525 
 526         if (st_arg->ds != NULL) {
 527                 err = traverse_dataset_resume(st_arg->ds,
 528                     st_arg->fromtxg, &st_arg->resume,
 529                     st_arg->flags, send_cb, st_arg);
 530 
 531                 if (err != EINTR)
 532                         st_arg->error_code = err;
 533         }
 534         data = kmem_zalloc(sizeof (*data), KM_SLEEP);
 535         data->eos_marker = B_TRUE;
 536         bqueue_enqueue(&st_arg->q, data, 1);
 537 }
 538 
 539 /*
 540  * This function actually handles figuring out what kind of record needs to be
 541  * dumped, reading the data (which has hopefully been prefetched), and calling
 542  * the appropriate helper function.
 543  */
 544 static int
 545 do_dump(dmu_sendarg_t *dsa, struct send_block_record *data)
 546 {
 547         dsl_dataset_t *ds = dmu_objset_ds(dsa->dsa_os);
 548         const blkptr_t *bp = &data->bp;
 549         const zbookmark_phys_t *zb = &data->zb;
 550         uint8_t indblkshift = data->indblkshift;
 551         uint16_t dblkszsec = data->datablkszsec;
 552         spa_t *spa = ds->ds_dir->dd_pool->dp_spa;
 553         dmu_object_type_t type = bp ? BP_GET_TYPE(bp) : DMU_OT_NONE;
 554         int err = 0;
 555 
 556         ASSERT3U(zb->zb_level, >=, 0);
 557 
 558         ASSERT(zb->zb_object == DMU_META_DNODE_OBJECT ||
 559             zb->zb_object >= dsa->dsa_resume_object);
 560 
 561         if (zb->zb_object != DMU_META_DNODE_OBJECT &&
 562             DMU_OBJECT_IS_SPECIAL(zb->zb_object)) {
 563                 return (0);
 564         } else if (BP_IS_HOLE(bp) &&
 565             zb->zb_object == DMU_META_DNODE_OBJECT) {
 566                 uint64_t span = BP_SPAN(dblkszsec, indblkshift, zb->zb_level);
 567                 uint64_t dnobj = (zb->zb_blkid * span) >> DNODE_SHIFT;
 568                 err = dump_freeobjects(dsa, dnobj, span >> DNODE_SHIFT);
 569         } else if (BP_IS_HOLE(bp)) {
 570                 uint64_t span = BP_SPAN(dblkszsec, indblkshift, zb->zb_level);
 571                 uint64_t offset = zb->zb_blkid * span;
 572                 err = dump_free(dsa, zb->zb_object, offset, span);
 573         } else if (zb->zb_level > 0 || type == DMU_OT_OBJSET) {
 574                 return (0);
 575         } else if (type == DMU_OT_DNODE) {
 576                 int blksz = BP_GET_LSIZE(bp);
 577                 arc_flags_t aflags = ARC_FLAG_WAIT;
 578                 arc_buf_t *abuf;
 579 
 580                 ASSERT0(zb->zb_level);
 581 
 582                 if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
 583                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
 584                     &aflags, zb) != 0)
 585                         return (SET_ERROR(EIO));
 586 
 587                 dnode_phys_t *blk = abuf->b_data;
 588                 uint64_t dnobj = zb->zb_blkid * (blksz >> DNODE_SHIFT);
 589                 for (int i = 0; i < blksz >> DNODE_SHIFT; i++) {
 590                         err = dump_dnode(dsa, dnobj + i, blk + i);
 591                         if (err != 0)
 592                                 break;
 593                 }
 594                 (void) arc_buf_remove_ref(abuf, &abuf);
 595         } else if (type == DMU_OT_SA) {
 596                 arc_flags_t aflags = ARC_FLAG_WAIT;
 597                 arc_buf_t *abuf;
 598                 int blksz = BP_GET_LSIZE(bp);
 599 
 600                 if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
 601                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
 602                     &aflags, zb) != 0)
 603                         return (SET_ERROR(EIO));
 604 
 605                 err = dump_spill(dsa, zb->zb_object, blksz, abuf->b_data);
 606                 (void) arc_buf_remove_ref(abuf, &abuf);
 607         } else if (backup_do_embed(dsa, bp)) {
 608                 /* it's an embedded level-0 block of a regular object */
 609                 int blksz = dblkszsec << SPA_MINBLOCKSHIFT;
 610                 ASSERT0(zb->zb_level);
 611                 err = dump_write_embedded(dsa, zb->zb_object,
 612                     zb->zb_blkid * blksz, blksz, bp);
 613         } else {
 614                 /* it's a level-0 block of a regular object */
 615                 arc_flags_t aflags = ARC_FLAG_WAIT;
 616                 arc_buf_t *abuf;
 617                 int blksz = dblkszsec << SPA_MINBLOCKSHIFT;
 618                 uint64_t offset;
 619 
 620                 ASSERT0(zb->zb_level);
 621                 ASSERT(zb->zb_object > dsa->dsa_resume_object ||
 622                     (zb->zb_object == dsa->dsa_resume_object &&
 623                     zb->zb_blkid * blksz >= dsa->dsa_resume_offset));
 624 
 625                 if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
 626                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
 627                     &aflags, zb) != 0) {
 628                         if (zfs_send_corrupt_data) {
 629                                 /* Send a block filled with 0x"zfs badd bloc" */
 630                                 abuf = arc_buf_alloc(spa, blksz, &abuf,
 631                                     ARC_BUFC_DATA);
 632                                 uint64_t *ptr;
 633                                 for (ptr = abuf->b_data;
 634                                     (char *)ptr < (char *)abuf->b_data + blksz;
 635                                     ptr++)
 636                                         *ptr = 0x2f5baddb10cULL;
 637                         } else {
 638                                 return (SET_ERROR(EIO));
 639                         }
 640                 }
 641 
 642                 offset = zb->zb_blkid * blksz;
 643 
 644                 if (!(dsa->dsa_featureflags &
 645                     DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
 646                     blksz > SPA_OLD_MAXBLOCKSIZE) {
 647                         char *buf = abuf->b_data;
 648                         while (blksz > 0 && err == 0) {
 649                                 int n = MIN(blksz, SPA_OLD_MAXBLOCKSIZE);
 650                                 err = dump_write(dsa, type, zb->zb_object,
 651                                     offset, n, NULL, buf);
 652                                 offset += n;
 653                                 buf += n;
 654                                 blksz -= n;
 655                         }
 656                 } else {
 657                         err = dump_write(dsa, type, zb->zb_object,
 658                             offset, blksz, bp, abuf->b_data);
 659                 }
 660                 (void) arc_buf_remove_ref(abuf, &abuf);
 661         }
 662 
 663         ASSERT(err == 0 || err == EINTR);
 664         return (err);
 665 }
 666 
 667 /*
 668  * Pop the new data off the queue, and free the old data.
 669  */
 670 static struct send_block_record *
 671 get_next_record(bqueue_t *bq, struct send_block_record *data)
 672 {
 673         struct send_block_record *tmp = bqueue_dequeue(bq);
 674         kmem_free(data, sizeof (*data));
 675         return (tmp);
 676 }
 677 
 678 /*
 679  * Actually do the bulk of the work in a zfs send.
 680  *
 681  * Note: Releases dp using the specified tag.
 682  */
 683 static int
 684 dmu_send_impl(void *tag, dsl_pool_t *dp, dsl_dataset_t *to_ds,
 685     zfs_bookmark_phys_t *ancestor_zb,
 686     boolean_t is_clone, boolean_t embedok, boolean_t large_block_ok, int outfd,
 687     uint64_t resumeobj, uint64_t resumeoff,
 688     vnode_t *vp, offset_t *off)
 689 {
 690         objset_t *os;
 691         dmu_replay_record_t *drr;
 692         dmu_sendarg_t *dsp;
 693         int err;
 694         uint64_t fromtxg = 0;
 695         uint64_t featureflags = 0;
 696         struct send_thread_arg to_arg = { 0 };
 697 
 698         err = dmu_objset_from_ds(to_ds, &os);
 699         if (err != 0) {
 700                 dsl_pool_rele(dp, tag);
 701                 return (err);
 702         }
 703 
 704         drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP);
 705         drr->drr_type = DRR_BEGIN;
 706         drr->drr_u.drr_begin.drr_magic = DMU_BACKUP_MAGIC;
 707         DMU_SET_STREAM_HDRTYPE(drr->drr_u.drr_begin.drr_versioninfo,
 708             DMU_SUBSTREAM);
 709 
 710 #ifdef _KERNEL
 711         if (dmu_objset_type(os) == DMU_OST_ZFS) {
 712                 uint64_t version;
 713                 if (zfs_get_zplprop(os, ZFS_PROP_VERSION, &version) != 0) {
 714                         kmem_free(drr, sizeof (dmu_replay_record_t));
 715                         dsl_pool_rele(dp, tag);
 716                         return (SET_ERROR(EINVAL));
 717                 }
 718                 if (version >= ZPL_VERSION_SA) {
 719                         featureflags |= DMU_BACKUP_FEATURE_SA_SPILL;
 720                 }
 721         }
 722 #endif
 723 
 724         if (large_block_ok && to_ds->ds_feature_inuse[SPA_FEATURE_LARGE_BLOCKS])
 725                 featureflags |= DMU_BACKUP_FEATURE_LARGE_BLOCKS;
 726         if (embedok &&
 727             spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA)) {
 728                 featureflags |= DMU_BACKUP_FEATURE_EMBED_DATA;
 729                 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS))
 730                         featureflags |= DMU_BACKUP_FEATURE_EMBED_DATA_LZ4;
 731         }
 732 
 733         if (resumeobj != 0 || resumeoff != 0) {
 734                 featureflags |= DMU_BACKUP_FEATURE_RESUMING;
 735         }
 736 
 737         DMU_SET_FEATUREFLAGS(drr->drr_u.drr_begin.drr_versioninfo,
 738             featureflags);
 739 
 740         drr->drr_u.drr_begin.drr_creation_time =
 741             dsl_dataset_phys(to_ds)->ds_creation_time;
 742         drr->drr_u.drr_begin.drr_type = dmu_objset_type(os);
 743         if (is_clone)
 744                 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CLONE;
 745         drr->drr_u.drr_begin.drr_toguid = dsl_dataset_phys(to_ds)->ds_guid;
 746         if (dsl_dataset_phys(to_ds)->ds_flags & DS_FLAG_CI_DATASET)
 747                 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CI_DATA;
 748         drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_FREERECORDS;
 749 
 750         if (ancestor_zb != NULL) {
 751                 drr->drr_u.drr_begin.drr_fromguid =
 752                     ancestor_zb->zbm_guid;
 753                 fromtxg = ancestor_zb->zbm_creation_txg;
 754         }
 755         dsl_dataset_name(to_ds, drr->drr_u.drr_begin.drr_toname);
 756         if (!to_ds->ds_is_snapshot) {
 757                 (void) strlcat(drr->drr_u.drr_begin.drr_toname, "@--head--",
 758                     sizeof (drr->drr_u.drr_begin.drr_toname));
 759         }
 760 
 761         dsp = kmem_zalloc(sizeof (dmu_sendarg_t), KM_SLEEP);
 762 
 763         dsp->dsa_drr = drr;
 764         dsp->dsa_vp = vp;
 765         dsp->dsa_outfd = outfd;
 766         dsp->dsa_proc = curproc;
 767         dsp->dsa_os = os;
 768         dsp->dsa_off = off;
 769         dsp->dsa_toguid = dsl_dataset_phys(to_ds)->ds_guid;
 770         dsp->dsa_pending_op = PENDING_NONE;
 771         dsp->dsa_featureflags = featureflags;
 772         dsp->dsa_resume_object = resumeobj;
 773         dsp->dsa_resume_offset = resumeoff;
 774 
 775         mutex_enter(&to_ds->ds_sendstream_lock);
 776         list_insert_head(&to_ds->ds_sendstreams, dsp);
 777         mutex_exit(&to_ds->ds_sendstream_lock);
 778 
 779         dsl_dataset_long_hold(to_ds, FTAG);
 780         dsl_pool_rele(dp, tag);
 781 
 782         void *payload = NULL;
 783         size_t payload_len = 0;
 784         if (resumeobj != 0 || resumeoff != 0) {
 785                 dmu_object_info_t to_doi;
 786                 err = dmu_object_info(os, resumeobj, &to_doi);
 787                 if (err != 0)
 788                         goto out;
 789                 SET_BOOKMARK(&to_arg.resume, to_ds->ds_object, resumeobj, 0,
 790                     resumeoff / to_doi.doi_data_block_size);
 791 
 792                 nvlist_t *nvl = fnvlist_alloc();
 793                 fnvlist_add_uint64(nvl, "resume_object", resumeobj);
 794                 fnvlist_add_uint64(nvl, "resume_offset", resumeoff);
 795                 payload = fnvlist_pack(nvl, &payload_len);
 796                 drr->drr_payloadlen = payload_len;
 797                 fnvlist_free(nvl);
 798         }
 799 
 800         err = dump_record(dsp, payload, payload_len);
 801         fnvlist_pack_free(payload, payload_len);
 802         if (err != 0) {
 803                 err = dsp->dsa_err;
 804                 goto out;
 805         }
 806 
 807         err = bqueue_init(&to_arg.q, zfs_send_queue_length,
 808             offsetof(struct send_block_record, ln));
 809         to_arg.error_code = 0;
 810         to_arg.cancel = B_FALSE;
 811         to_arg.ds = to_ds;
 812         to_arg.fromtxg = fromtxg;
 813         to_arg.flags = TRAVERSE_PRE | TRAVERSE_PREFETCH;
 814         (void) thread_create(NULL, 0, send_traverse_thread, &to_arg, 0, curproc,
 815             TS_RUN, minclsyspri);
 816 
 817         struct send_block_record *to_data;
 818         to_data = bqueue_dequeue(&to_arg.q);
 819 
 820         while (!to_data->eos_marker && err == 0) {
 821                 err = do_dump(dsp, to_data);
 822                 to_data = get_next_record(&to_arg.q, to_data);
 823                 if (issig(JUSTLOOKING) && issig(FORREAL))
 824                         err = EINTR;
 825         }
 826 
 827         if (err != 0) {
 828                 to_arg.cancel = B_TRUE;
 829                 while (!to_data->eos_marker) {
 830                         to_data = get_next_record(&to_arg.q, to_data);
 831                 }
 832         }
 833         kmem_free(to_data, sizeof (*to_data));
 834 
 835         bqueue_destroy(&to_arg.q);
 836 
 837         if (err == 0 && to_arg.error_code != 0)
 838                 err = to_arg.error_code;
 839 
 840         if (err != 0)
 841                 goto out;
 842 
 843         if (dsp->dsa_pending_op != PENDING_NONE)
 844                 if (dump_record(dsp, NULL, 0) != 0)
 845                         err = SET_ERROR(EINTR);
 846 
 847         if (err != 0) {
 848                 if (err == EINTR && dsp->dsa_err != 0)
 849                         err = dsp->dsa_err;
 850                 goto out;
 851         }
 852 
 853         bzero(drr, sizeof (dmu_replay_record_t));
 854         drr->drr_type = DRR_END;
 855         drr->drr_u.drr_end.drr_checksum = dsp->dsa_zc;
 856         drr->drr_u.drr_end.drr_toguid = dsp->dsa_toguid;
 857 
 858         if (dump_record(dsp, NULL, 0) != 0)
 859                 err = dsp->dsa_err;
 860 
 861 out:
 862         mutex_enter(&to_ds->ds_sendstream_lock);
 863         list_remove(&to_ds->ds_sendstreams, dsp);
 864         mutex_exit(&to_ds->ds_sendstream_lock);
 865 
 866         kmem_free(drr, sizeof (dmu_replay_record_t));
 867         kmem_free(dsp, sizeof (dmu_sendarg_t));
 868 
 869         dsl_dataset_long_rele(to_ds, FTAG);
 870 
 871         return (err);
 872 }
 873 
 874 int
 875 dmu_send_obj(const char *pool, uint64_t tosnap, uint64_t fromsnap,
 876     boolean_t embedok, boolean_t large_block_ok,
 877     int outfd, vnode_t *vp, offset_t *off)
 878 {
 879         dsl_pool_t *dp;
 880         dsl_dataset_t *ds;
 881         dsl_dataset_t *fromds = NULL;
 882         int err;
 883 
 884         err = dsl_pool_hold(pool, FTAG, &dp);
 885         if (err != 0)
 886                 return (err);
 887 
 888         err = dsl_dataset_hold_obj(dp, tosnap, FTAG, &ds);
 889         if (err != 0) {
 890                 dsl_pool_rele(dp, FTAG);
 891                 return (err);
 892         }
 893 
 894         if (fromsnap != 0) {
 895                 zfs_bookmark_phys_t zb;
 896                 boolean_t is_clone;
 897 
 898                 err = dsl_dataset_hold_obj(dp, fromsnap, FTAG, &fromds);
 899                 if (err != 0) {
 900                         dsl_dataset_rele(ds, FTAG);
 901                         dsl_pool_rele(dp, FTAG);
 902                         return (err);
 903                 }
 904                 if (!dsl_dataset_is_before(ds, fromds, 0))
 905                         err = SET_ERROR(EXDEV);
 906                 zb.zbm_creation_time =
 907                     dsl_dataset_phys(fromds)->ds_creation_time;
 908                 zb.zbm_creation_txg = dsl_dataset_phys(fromds)->ds_creation_txg;
 909                 zb.zbm_guid = dsl_dataset_phys(fromds)->ds_guid;
 910                 is_clone = (fromds->ds_dir != ds->ds_dir);
 911                 dsl_dataset_rele(fromds, FTAG);
 912                 err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone,
 913                     embedok, large_block_ok, outfd, 0, 0, vp, off);
 914         } else {
 915                 err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE,
 916                     embedok, large_block_ok, outfd, 0, 0, vp, off);
 917         }
 918         dsl_dataset_rele(ds, FTAG);
 919         return (err);
 920 }
 921 
 922 int
 923 dmu_send(const char *tosnap, const char *fromsnap, boolean_t embedok,
 924     boolean_t large_block_ok, int outfd, uint64_t resumeobj, uint64_t resumeoff,
 925     vnode_t *vp, offset_t *off)
 926 {
 927         dsl_pool_t *dp;
 928         dsl_dataset_t *ds;
 929         int err;
 930         boolean_t owned = B_FALSE;
 931 
 932         if (fromsnap != NULL && strpbrk(fromsnap, "@#") == NULL)
 933                 return (SET_ERROR(EINVAL));
 934 
 935         err = dsl_pool_hold(tosnap, FTAG, &dp);
 936         if (err != 0)
 937                 return (err);
 938 
 939         if (strchr(tosnap, '@') == NULL && spa_writeable(dp->dp_spa)) {
 940                 /*
 941                  * We are sending a filesystem or volume.  Ensure
 942                  * that it doesn't change by owning the dataset.
 943                  */
 944                 err = dsl_dataset_own(dp, tosnap, FTAG, &ds);
 945                 owned = B_TRUE;
 946         } else {
 947                 err = dsl_dataset_hold(dp, tosnap, FTAG, &ds);
 948         }
 949         if (err != 0) {
 950                 dsl_pool_rele(dp, FTAG);
 951                 return (err);
 952         }
 953 
 954         if (fromsnap != NULL) {
 955                 zfs_bookmark_phys_t zb;
 956                 boolean_t is_clone = B_FALSE;
 957                 int fsnamelen = strchr(tosnap, '@') - tosnap;
 958 
 959                 /*
 960                  * If the fromsnap is in a different filesystem, then
 961                  * mark the send stream as a clone.
 962                  */
 963                 if (strncmp(tosnap, fromsnap, fsnamelen) != 0 ||
 964                     (fromsnap[fsnamelen] != '@' &&
 965                     fromsnap[fsnamelen] != '#')) {
 966                         is_clone = B_TRUE;
 967                 }
 968 
 969                 if (strchr(fromsnap, '@')) {
 970                         dsl_dataset_t *fromds;
 971                         err = dsl_dataset_hold(dp, fromsnap, FTAG, &fromds);
 972                         if (err == 0) {
 973                                 if (!dsl_dataset_is_before(ds, fromds, 0))
 974                                         err = SET_ERROR(EXDEV);
 975                                 zb.zbm_creation_time =
 976                                     dsl_dataset_phys(fromds)->ds_creation_time;
 977                                 zb.zbm_creation_txg =
 978                                     dsl_dataset_phys(fromds)->ds_creation_txg;
 979                                 zb.zbm_guid = dsl_dataset_phys(fromds)->ds_guid;
 980                                 is_clone = (ds->ds_dir != fromds->ds_dir);
 981                                 dsl_dataset_rele(fromds, FTAG);
 982                         }
 983                 } else {
 984                         err = dsl_bookmark_lookup(dp, fromsnap, ds, &zb);
 985                 }
 986                 if (err != 0) {
 987                         dsl_dataset_rele(ds, FTAG);
 988                         dsl_pool_rele(dp, FTAG);
 989                         return (err);
 990                 }
 991                 err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone,
 992                     embedok, large_block_ok,
 993                     outfd, resumeobj, resumeoff, vp, off);
 994         } else {
 995                 err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE,
 996                     embedok, large_block_ok,
 997                     outfd, resumeobj, resumeoff, vp, off);
 998         }
 999         if (owned)
1000                 dsl_dataset_disown(ds, FTAG);
1001         else
1002                 dsl_dataset_rele(ds, FTAG);
1003         return (err);
1004 }
1005 
1006 static int
1007 dmu_adjust_send_estimate_for_indirects(dsl_dataset_t *ds, uint64_t size,
1008     uint64_t *sizep)
1009 {
1010         int err;
1011         /*
1012          * Assume that space (both on-disk and in-stream) is dominated by
1013          * data.  We will adjust for indirect blocks and the copies property,
1014          * but ignore per-object space used (eg, dnodes and DRR_OBJECT records).
1015          */
1016 
1017         /*
1018          * Subtract out approximate space used by indirect blocks.
1019          * Assume most space is used by data blocks (non-indirect, non-dnode).
1020          * Assume all blocks are recordsize.  Assume ditto blocks and
1021          * internal fragmentation counter out compression.
1022          *
1023          * Therefore, space used by indirect blocks is sizeof(blkptr_t) per
1024          * block, which we observe in practice.
1025          */
1026         uint64_t recordsize;
1027         err = dsl_prop_get_int_ds(ds, "recordsize", &recordsize);
1028         if (err != 0)
1029                 return (err);
1030         size -= size / recordsize * sizeof (blkptr_t);
1031 
1032         /* Add in the space for the record associated with each block. */
1033         size += size / recordsize * sizeof (dmu_replay_record_t);
1034 
1035         *sizep = size;
1036 
1037         return (0);
1038 }
1039 
1040 int
1041 dmu_send_estimate(dsl_dataset_t *ds, dsl_dataset_t *fromds, uint64_t *sizep)
1042 {
1043         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1044         int err;
1045         uint64_t size;
1046 
1047         ASSERT(dsl_pool_config_held(dp));
1048 
1049         /* tosnap must be a snapshot */
1050         if (!ds->ds_is_snapshot)
1051                 return (SET_ERROR(EINVAL));
1052 
1053         /* fromsnap, if provided, must be a snapshot */
1054         if (fromds != NULL && !fromds->ds_is_snapshot)
1055                 return (SET_ERROR(EINVAL));
1056 
1057         /*
1058          * fromsnap must be an earlier snapshot from the same fs as tosnap,
1059          * or the origin's fs.
1060          */
1061         if (fromds != NULL && !dsl_dataset_is_before(ds, fromds, 0))
1062                 return (SET_ERROR(EXDEV));
1063 
1064         /* Get uncompressed size estimate of changed data. */
1065         if (fromds == NULL) {
1066                 size = dsl_dataset_phys(ds)->ds_uncompressed_bytes;
1067         } else {
1068                 uint64_t used, comp;
1069                 err = dsl_dataset_space_written(fromds, ds,
1070                     &used, &comp, &size);
1071                 if (err != 0)
1072                         return (err);
1073         }
1074 
1075         err = dmu_adjust_send_estimate_for_indirects(ds, size, sizep);
1076         return (err);
1077 }
1078 
1079 /*
1080  * Simple callback used to traverse the blocks of a snapshot and sum their
1081  * uncompressed size
1082  */
1083 /* ARGSUSED */
1084 static int
1085 dmu_calculate_send_traversal(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
1086     const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
1087 {
1088         uint64_t *spaceptr = arg;
1089         if (bp != NULL && !BP_IS_HOLE(bp)) {
1090                 *spaceptr += BP_GET_UCSIZE(bp);
1091         }
1092         return (0);
1093 }
1094 
1095 /*
1096  * Given a desination snapshot and a TXG, calculate the approximate size of a
1097  * send stream sent from that TXG. from_txg may be zero, indicating that the
1098  * whole snapshot will be sent.
1099  */
1100 int
1101 dmu_send_estimate_from_txg(dsl_dataset_t *ds, uint64_t from_txg,
1102     uint64_t *sizep)
1103 {
1104         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1105         int err;
1106         uint64_t size = 0;
1107 
1108         ASSERT(dsl_pool_config_held(dp));
1109 
1110         /* tosnap must be a snapshot */
1111         if (!dsl_dataset_is_snapshot(ds))
1112                 return (SET_ERROR(EINVAL));
1113 
1114         /* verify that from_txg is before the provided snapshot was taken */
1115         if (from_txg >= dsl_dataset_phys(ds)->ds_creation_txg) {
1116                 return (SET_ERROR(EXDEV));
1117         }
1118 
1119         /*
1120          * traverse the blocks of the snapshot with birth times after
1121          * from_txg, summing their uncompressed size
1122          */
1123         err = traverse_dataset(ds, from_txg, TRAVERSE_POST,
1124             dmu_calculate_send_traversal, &size);
1125         if (err)
1126                 return (err);
1127 
1128         err = dmu_adjust_send_estimate_for_indirects(ds, size, sizep);
1129         return (err);
1130 }
1131 
1132 typedef struct dmu_recv_begin_arg {
1133         const char *drba_origin;
1134         dmu_recv_cookie_t *drba_cookie;
1135         cred_t *drba_cred;
1136         uint64_t drba_snapobj;
1137 } dmu_recv_begin_arg_t;
1138 
1139 static int
1140 recv_begin_check_existing_impl(dmu_recv_begin_arg_t *drba, dsl_dataset_t *ds,
1141     uint64_t fromguid)
1142 {
1143         uint64_t val;
1144         int error;
1145         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1146 
1147         /* temporary clone name must not exist */
1148         error = zap_lookup(dp->dp_meta_objset,
1149             dsl_dir_phys(ds->ds_dir)->dd_child_dir_zapobj, recv_clone_name,
1150             8, 1, &val);
1151         if (error != ENOENT)
1152                 return (error == 0 ? EBUSY : error);
1153 
1154         /* new snapshot name must not exist */
1155         error = zap_lookup(dp->dp_meta_objset,
1156             dsl_dataset_phys(ds)->ds_snapnames_zapobj,
1157             drba->drba_cookie->drc_tosnap, 8, 1, &val);
1158         if (error != ENOENT)
1159                 return (error == 0 ? EEXIST : error);
1160 
1161         /*
1162          * Check snapshot limit before receiving. We'll recheck again at the
1163          * end, but might as well abort before receiving if we're already over
1164          * the limit.
1165          *
1166          * Note that we do not check the file system limit with
1167          * dsl_dir_fscount_check because the temporary %clones don't count
1168          * against that limit.
1169          */
1170         error = dsl_fs_ss_limit_check(ds->ds_dir, 1, ZFS_PROP_SNAPSHOT_LIMIT,
1171             NULL, drba->drba_cred);
1172         if (error != 0)
1173                 return (error);
1174 
1175         if (fromguid != 0) {
1176                 dsl_dataset_t *snap;
1177                 uint64_t obj = dsl_dataset_phys(ds)->ds_prev_snap_obj;
1178 
1179                 /* Find snapshot in this dir that matches fromguid. */
1180                 while (obj != 0) {
1181                         error = dsl_dataset_hold_obj(dp, obj, FTAG,
1182                             &snap);
1183                         if (error != 0)
1184                                 return (SET_ERROR(ENODEV));
1185                         if (snap->ds_dir != ds->ds_dir) {
1186                                 dsl_dataset_rele(snap, FTAG);
1187                                 return (SET_ERROR(ENODEV));
1188                         }
1189                         if (dsl_dataset_phys(snap)->ds_guid == fromguid)
1190                                 break;
1191                         obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
1192                         dsl_dataset_rele(snap, FTAG);
1193                 }
1194                 if (obj == 0)
1195                         return (SET_ERROR(ENODEV));
1196 
1197                 if (drba->drba_cookie->drc_force) {
1198                         drba->drba_snapobj = obj;
1199                 } else {
1200                         /*
1201                          * If we are not forcing, there must be no
1202                          * changes since fromsnap.
1203                          */
1204                         if (dsl_dataset_modified_since_snap(ds, snap)) {
1205                                 dsl_dataset_rele(snap, FTAG);
1206                                 return (SET_ERROR(ETXTBSY));
1207                         }
1208                         drba->drba_snapobj = ds->ds_prev->ds_object;
1209                 }
1210 
1211                 dsl_dataset_rele(snap, FTAG);
1212         } else {
1213                 /* if full, then must be forced */
1214                 if (!drba->drba_cookie->drc_force)
1215                         return (SET_ERROR(EEXIST));
1216                 /* start from $ORIGIN@$ORIGIN, if supported */
1217                 drba->drba_snapobj = dp->dp_origin_snap != NULL ?
1218                     dp->dp_origin_snap->ds_object : 0;
1219         }
1220 
1221         return (0);
1222 
1223 }
1224 
1225 static int
1226 dmu_recv_begin_check(void *arg, dmu_tx_t *tx)
1227 {
1228         dmu_recv_begin_arg_t *drba = arg;
1229         dsl_pool_t *dp = dmu_tx_pool(tx);
1230         struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
1231         uint64_t fromguid = drrb->drr_fromguid;
1232         int flags = drrb->drr_flags;
1233         int error;
1234         uint64_t featureflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo);
1235         dsl_dataset_t *ds;
1236         const char *tofs = drba->drba_cookie->drc_tofs;
1237 
1238         /* already checked */
1239         ASSERT3U(drrb->drr_magic, ==, DMU_BACKUP_MAGIC);
1240         ASSERT(!(featureflags & DMU_BACKUP_FEATURE_RESUMING));
1241 
1242         if (DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) ==
1243             DMU_COMPOUNDSTREAM ||
1244             drrb->drr_type >= DMU_OST_NUMTYPES ||
1245             ((flags & DRR_FLAG_CLONE) && drba->drba_origin == NULL))
1246                 return (SET_ERROR(EINVAL));
1247 
1248         /* Verify pool version supports SA if SA_SPILL feature set */
1249         if ((featureflags & DMU_BACKUP_FEATURE_SA_SPILL) &&
1250             spa_version(dp->dp_spa) < SPA_VERSION_SA)
1251                 return (SET_ERROR(ENOTSUP));
1252 
1253         if (drba->drba_cookie->drc_resumable &&
1254             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EXTENSIBLE_DATASET))
1255                 return (SET_ERROR(ENOTSUP));
1256 
1257         /*
1258          * The receiving code doesn't know how to translate a WRITE_EMBEDDED
1259          * record to a plan WRITE record, so the pool must have the
1260          * EMBEDDED_DATA feature enabled if the stream has WRITE_EMBEDDED
1261          * records.  Same with WRITE_EMBEDDED records that use LZ4 compression.
1262          */
1263         if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA) &&
1264             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA))
1265                 return (SET_ERROR(ENOTSUP));
1266         if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA_LZ4) &&
1267             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS))
1268                 return (SET_ERROR(ENOTSUP));
1269 
1270         /*
1271          * The receiving code doesn't know how to translate large blocks
1272          * to smaller ones, so the pool must have the LARGE_BLOCKS
1273          * feature enabled if the stream has LARGE_BLOCKS.
1274          */
1275         if ((featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
1276             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LARGE_BLOCKS))
1277                 return (SET_ERROR(ENOTSUP));
1278 
1279         error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
1280         if (error == 0) {
1281                 /* target fs already exists; recv into temp clone */
1282 
1283                 /* Can't recv a clone into an existing fs */
1284                 if (flags & DRR_FLAG_CLONE || drba->drba_origin) {
1285                         dsl_dataset_rele(ds, FTAG);
1286                         return (SET_ERROR(EINVAL));
1287                 }
1288 
1289                 error = recv_begin_check_existing_impl(drba, ds, fromguid);
1290                 dsl_dataset_rele(ds, FTAG);
1291         } else if (error == ENOENT) {
1292                 /* target fs does not exist; must be a full backup or clone */
1293                 char buf[MAXNAMELEN];
1294 
1295                 /*
1296                  * If it's a non-clone incremental, we are missing the
1297                  * target fs, so fail the recv.
1298                  */
1299                 if (fromguid != 0 && !(flags & DRR_FLAG_CLONE ||
1300                     drba->drba_origin))
1301                         return (SET_ERROR(ENOENT));
1302 
1303                 /*
1304                  * If we're receiving a full send as a clone, and it doesn't
1305                  * contain all the necessary free records and freeobject
1306                  * records, reject it.
1307                  */
1308                 if (fromguid == 0 && drba->drba_origin &&
1309                     !(flags & DRR_FLAG_FREERECORDS))
1310                         return (SET_ERROR(EINVAL));
1311 
1312                 /* Open the parent of tofs */
1313                 ASSERT3U(strlen(tofs), <, MAXNAMELEN);
1314                 (void) strlcpy(buf, tofs, strrchr(tofs, '/') - tofs + 1);
1315                 error = dsl_dataset_hold(dp, buf, FTAG, &ds);
1316                 if (error != 0)
1317                         return (error);
1318 
1319                 /*
1320                  * Check filesystem and snapshot limits before receiving. We'll
1321                  * recheck snapshot limits again at the end (we create the
1322                  * filesystems and increment those counts during begin_sync).
1323                  */
1324                 error = dsl_fs_ss_limit_check(ds->ds_dir, 1,
1325                     ZFS_PROP_FILESYSTEM_LIMIT, NULL, drba->drba_cred);
1326                 if (error != 0) {
1327                         dsl_dataset_rele(ds, FTAG);
1328                         return (error);
1329                 }
1330 
1331                 error = dsl_fs_ss_limit_check(ds->ds_dir, 1,
1332                     ZFS_PROP_SNAPSHOT_LIMIT, NULL, drba->drba_cred);
1333                 if (error != 0) {
1334                         dsl_dataset_rele(ds, FTAG);
1335                         return (error);
1336                 }
1337 
1338                 if (drba->drba_origin != NULL) {
1339                         dsl_dataset_t *origin;
1340                         error = dsl_dataset_hold(dp, drba->drba_origin,
1341                             FTAG, &origin);
1342                         if (error != 0) {
1343                                 dsl_dataset_rele(ds, FTAG);
1344                                 return (error);
1345                         }
1346                         if (!origin->ds_is_snapshot) {
1347                                 dsl_dataset_rele(origin, FTAG);
1348                                 dsl_dataset_rele(ds, FTAG);
1349                                 return (SET_ERROR(EINVAL));
1350                         }
1351                         if (dsl_dataset_phys(origin)->ds_guid != fromguid &&
1352                             fromguid != 0) {
1353                                 dsl_dataset_rele(origin, FTAG);
1354                                 dsl_dataset_rele(ds, FTAG);
1355                                 return (SET_ERROR(ENODEV));
1356                         }
1357                         dsl_dataset_rele(origin, FTAG);
1358                 }
1359                 dsl_dataset_rele(ds, FTAG);
1360                 error = 0;
1361         }
1362         return (error);
1363 }
1364 
1365 static void
1366 dmu_recv_begin_sync(void *arg, dmu_tx_t *tx)
1367 {
1368         dmu_recv_begin_arg_t *drba = arg;
1369         dsl_pool_t *dp = dmu_tx_pool(tx);
1370         objset_t *mos = dp->dp_meta_objset;
1371         struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
1372         const char *tofs = drba->drba_cookie->drc_tofs;
1373         dsl_dataset_t *ds, *newds;
1374         uint64_t dsobj;
1375         int error;
1376         uint64_t crflags = 0;
1377 
1378         if (drrb->drr_flags & DRR_FLAG_CI_DATA)
1379                 crflags |= DS_FLAG_CI_DATASET;
1380 
1381         error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
1382         if (error == 0) {
1383                 /* create temporary clone */
1384                 dsl_dataset_t *snap = NULL;
1385                 if (drba->drba_snapobj != 0) {
1386                         VERIFY0(dsl_dataset_hold_obj(dp,
1387                             drba->drba_snapobj, FTAG, &snap));
1388                 }
1389                 dsobj = dsl_dataset_create_sync(ds->ds_dir, recv_clone_name,
1390                     snap, crflags, drba->drba_cred, tx);
1391                 if (drba->drba_snapobj != 0)
1392                         dsl_dataset_rele(snap, FTAG);
1393                 dsl_dataset_rele(ds, FTAG);
1394         } else {
1395                 dsl_dir_t *dd;
1396                 const char *tail;
1397                 dsl_dataset_t *origin = NULL;
1398 
1399                 VERIFY0(dsl_dir_hold(dp, tofs, FTAG, &dd, &tail));
1400 
1401                 if (drba->drba_origin != NULL) {
1402                         VERIFY0(dsl_dataset_hold(dp, drba->drba_origin,
1403                             FTAG, &origin));
1404                 }
1405 
1406                 /* Create new dataset. */
1407                 dsobj = dsl_dataset_create_sync(dd,
1408                     strrchr(tofs, '/') + 1,
1409                     origin, crflags, drba->drba_cred, tx);
1410                 if (origin != NULL)
1411                         dsl_dataset_rele(origin, FTAG);
1412                 dsl_dir_rele(dd, FTAG);
1413                 drba->drba_cookie->drc_newfs = B_TRUE;
1414         }
1415         VERIFY0(dsl_dataset_own_obj(dp, dsobj, dmu_recv_tag, &newds));
1416 
1417         if (drba->drba_cookie->drc_resumable) {
1418                 dsl_dataset_zapify(newds, tx);
1419                 if (drrb->drr_fromguid != 0) {
1420                         VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_FROMGUID,
1421                             8, 1, &drrb->drr_fromguid, tx));
1422                 }
1423                 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_TOGUID,
1424                     8, 1, &drrb->drr_toguid, tx));
1425                 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_TONAME,
1426                     1, strlen(drrb->drr_toname) + 1, drrb->drr_toname, tx));
1427                 uint64_t one = 1;
1428                 uint64_t zero = 0;
1429                 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_OBJECT,
1430                     8, 1, &one, tx));
1431                 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_OFFSET,
1432                     8, 1, &zero, tx));
1433                 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_BYTES,
1434                     8, 1, &zero, tx));
1435                 if (DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo) &
1436                     DMU_BACKUP_FEATURE_EMBED_DATA) {
1437                         VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_EMBEDOK,
1438                             8, 1, &one, tx));
1439                 }
1440         }
1441 
1442         dmu_buf_will_dirty(newds->ds_dbuf, tx);
1443         dsl_dataset_phys(newds)->ds_flags |= DS_FLAG_INCONSISTENT;
1444 
1445         /*
1446          * If we actually created a non-clone, we need to create the
1447          * objset in our new dataset.
1448          */
1449         if (BP_IS_HOLE(dsl_dataset_get_blkptr(newds))) {
1450                 (void) dmu_objset_create_impl(dp->dp_spa,
1451                     newds, dsl_dataset_get_blkptr(newds), drrb->drr_type, tx);
1452         }
1453 
1454         drba->drba_cookie->drc_ds = newds;
1455 
1456         spa_history_log_internal_ds(newds, "receive", tx, "");
1457 }
1458 
1459 static int
1460 dmu_recv_resume_begin_check(void *arg, dmu_tx_t *tx)
1461 {
1462         dmu_recv_begin_arg_t *drba = arg;
1463         dsl_pool_t *dp = dmu_tx_pool(tx);
1464         struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
1465         int error;
1466         uint64_t featureflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo);
1467         dsl_dataset_t *ds;
1468         const char *tofs = drba->drba_cookie->drc_tofs;
1469 
1470         /* already checked */
1471         ASSERT3U(drrb->drr_magic, ==, DMU_BACKUP_MAGIC);
1472         ASSERT(featureflags & DMU_BACKUP_FEATURE_RESUMING);
1473 
1474         if (DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) ==
1475             DMU_COMPOUNDSTREAM ||
1476             drrb->drr_type >= DMU_OST_NUMTYPES)
1477                 return (SET_ERROR(EINVAL));
1478 
1479         /* Verify pool version supports SA if SA_SPILL feature set */
1480         if ((featureflags & DMU_BACKUP_FEATURE_SA_SPILL) &&
1481             spa_version(dp->dp_spa) < SPA_VERSION_SA)
1482                 return (SET_ERROR(ENOTSUP));
1483 
1484         /*
1485          * The receiving code doesn't know how to translate a WRITE_EMBEDDED
1486          * record to a plain WRITE record, so the pool must have the
1487          * EMBEDDED_DATA feature enabled if the stream has WRITE_EMBEDDED
1488          * records.  Same with WRITE_EMBEDDED records that use LZ4 compression.
1489          */
1490         if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA) &&
1491             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA))
1492                 return (SET_ERROR(ENOTSUP));
1493         if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA_LZ4) &&
1494             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS))
1495                 return (SET_ERROR(ENOTSUP));
1496 
1497         char recvname[ZFS_MAXNAMELEN];
1498 
1499         (void) snprintf(recvname, sizeof (recvname), "%s/%s",
1500             tofs, recv_clone_name);
1501 
1502         if (dsl_dataset_hold(dp, recvname, FTAG, &ds) != 0) {
1503                 /* %recv does not exist; continue in tofs */
1504                 error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
1505                 if (error != 0)
1506                         return (error);
1507         }
1508 
1509         /* check that ds is marked inconsistent */
1510         if (!DS_IS_INCONSISTENT(ds)) {
1511                 dsl_dataset_rele(ds, FTAG);
1512                 return (SET_ERROR(EINVAL));
1513         }
1514 
1515         /* check that there is resuming data, and that the toguid matches */
1516         if (!dsl_dataset_is_zapified(ds)) {
1517                 dsl_dataset_rele(ds, FTAG);
1518                 return (SET_ERROR(EINVAL));
1519         }
1520         uint64_t val;
1521         error = zap_lookup(dp->dp_meta_objset, ds->ds_object,
1522             DS_FIELD_RESUME_TOGUID, sizeof (val), 1, &val);
1523         if (error != 0 || drrb->drr_toguid != val) {
1524                 dsl_dataset_rele(ds, FTAG);
1525                 return (SET_ERROR(EINVAL));
1526         }
1527 
1528         /*
1529          * Check if the receive is still running.  If so, it will be owned.
1530          * Note that nothing else can own the dataset (e.g. after the receive
1531          * fails) because it will be marked inconsistent.
1532          */
1533         if (dsl_dataset_has_owner(ds)) {
1534                 dsl_dataset_rele(ds, FTAG);
1535                 return (SET_ERROR(EBUSY));
1536         }
1537 
1538         /* There should not be any snapshots of this fs yet. */
1539         if (ds->ds_prev != NULL && ds->ds_prev->ds_dir == ds->ds_dir) {
1540                 dsl_dataset_rele(ds, FTAG);
1541                 return (SET_ERROR(EINVAL));
1542         }
1543 
1544         /*
1545          * Note: resume point will be checked when we process the first WRITE
1546          * record.
1547          */
1548 
1549         /* check that the origin matches */
1550         val = 0;
1551         (void) zap_lookup(dp->dp_meta_objset, ds->ds_object,
1552             DS_FIELD_RESUME_FROMGUID, sizeof (val), 1, &val);
1553         if (drrb->drr_fromguid != val) {
1554                 dsl_dataset_rele(ds, FTAG);
1555                 return (SET_ERROR(EINVAL));
1556         }
1557 
1558         dsl_dataset_rele(ds, FTAG);
1559         return (0);
1560 }
1561 
1562 static void
1563 dmu_recv_resume_begin_sync(void *arg, dmu_tx_t *tx)
1564 {
1565         dmu_recv_begin_arg_t *drba = arg;
1566         dsl_pool_t *dp = dmu_tx_pool(tx);
1567         const char *tofs = drba->drba_cookie->drc_tofs;
1568         dsl_dataset_t *ds;
1569         uint64_t dsobj;
1570         char recvname[ZFS_MAXNAMELEN];
1571 
1572         (void) snprintf(recvname, sizeof (recvname), "%s/%s",
1573             tofs, recv_clone_name);
1574 
1575         if (dsl_dataset_hold(dp, recvname, FTAG, &ds) != 0) {
1576                 /* %recv does not exist; continue in tofs */
1577                 VERIFY0(dsl_dataset_hold(dp, tofs, FTAG, &ds));
1578                 drba->drba_cookie->drc_newfs = B_TRUE;
1579         }
1580 
1581         /* clear the inconsistent flag so that we can own it */
1582         ASSERT(DS_IS_INCONSISTENT(ds));
1583         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1584         dsl_dataset_phys(ds)->ds_flags &= ~DS_FLAG_INCONSISTENT;
1585         dsobj = ds->ds_object;
1586         dsl_dataset_rele(ds, FTAG);
1587 
1588         VERIFY0(dsl_dataset_own_obj(dp, dsobj, dmu_recv_tag, &ds));
1589 
1590         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1591         dsl_dataset_phys(ds)->ds_flags |= DS_FLAG_INCONSISTENT;
1592 
1593         ASSERT(!BP_IS_HOLE(dsl_dataset_get_blkptr(ds)));
1594 
1595         drba->drba_cookie->drc_ds = ds;
1596 
1597         spa_history_log_internal_ds(ds, "resume receive", tx, "");
1598 }
1599 
1600 /*
1601  * NB: callers *MUST* call dmu_recv_stream() if dmu_recv_begin()
1602  * succeeds; otherwise we will leak the holds on the datasets.
1603  */
1604 int
1605 dmu_recv_begin(char *tofs, char *tosnap, dmu_replay_record_t *drr_begin,
1606     boolean_t force, boolean_t resumable, char *origin, dmu_recv_cookie_t *drc)
1607 {
1608         dmu_recv_begin_arg_t drba = { 0 };
1609 
1610         bzero(drc, sizeof (dmu_recv_cookie_t));
1611         drc->drc_drr_begin = drr_begin;
1612         drc->drc_drrb = &drr_begin->drr_u.drr_begin;
1613         drc->drc_tosnap = tosnap;
1614         drc->drc_tofs = tofs;
1615         drc->drc_force = force;
1616         drc->drc_resumable = resumable;
1617         drc->drc_cred = CRED();
1618 
1619         if (drc->drc_drrb->drr_magic == BSWAP_64(DMU_BACKUP_MAGIC)) {
1620                 drc->drc_byteswap = B_TRUE;
1621                 fletcher_4_incremental_byteswap(drr_begin,
1622                     sizeof (dmu_replay_record_t), &drc->drc_cksum);
1623                 byteswap_record(drr_begin);
1624         } else if (drc->drc_drrb->drr_magic == DMU_BACKUP_MAGIC) {
1625                 fletcher_4_incremental_native(drr_begin,
1626                     sizeof (dmu_replay_record_t), &drc->drc_cksum);
1627         } else {
1628                 return (SET_ERROR(EINVAL));
1629         }
1630 
1631         drba.drba_origin = origin;
1632         drba.drba_cookie = drc;
1633         drba.drba_cred = CRED();
1634 
1635         if (DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo) &
1636             DMU_BACKUP_FEATURE_RESUMING) {
1637                 return (dsl_sync_task(tofs,
1638                     dmu_recv_resume_begin_check, dmu_recv_resume_begin_sync,
1639                     &drba, 5, ZFS_SPACE_CHECK_NORMAL));
1640         } else  {
1641                 return (dsl_sync_task(tofs,
1642                     dmu_recv_begin_check, dmu_recv_begin_sync,
1643                     &drba, 5, ZFS_SPACE_CHECK_NORMAL));
1644         }
1645 }
1646 
1647 struct receive_record_arg {
1648         dmu_replay_record_t header;
1649         void *payload; /* Pointer to a buffer containing the payload */
1650         /*
1651          * If the record is a write, pointer to the arc_buf_t containing the
1652          * payload.
1653          */
1654         arc_buf_t *write_buf;
1655         int payload_size;
1656         uint64_t bytes_read; /* bytes read from stream when record created */
1657         boolean_t eos_marker; /* Marks the end of the stream */
1658         bqueue_node_t node;
1659 };
1660 
1661 struct receive_writer_arg {
1662         objset_t *os;
1663         boolean_t byteswap;
1664         bqueue_t q;
1665 
1666         /*
1667          * These three args are used to signal to the main thread that we're
1668          * done.
1669          */
1670         kmutex_t mutex;
1671         kcondvar_t cv;
1672         boolean_t done;
1673 
1674         int err;
1675         /* A map from guid to dataset to help handle dedup'd streams. */
1676         avl_tree_t *guid_to_ds_map;
1677         boolean_t resumable;
1678         uint64_t last_object, last_offset;
1679         uint64_t bytes_read; /* bytes read when current record created */
1680 };
1681 
1682 struct objlist {
1683         list_t list; /* List of struct receive_objnode. */
1684         /*
1685          * Last object looked up. Used to assert that objects are being looked
1686          * up in ascending order.
1687          */
1688         uint64_t last_lookup;
1689 };
1690 
1691 struct receive_objnode {
1692         list_node_t node;
1693         uint64_t object;
1694 };
1695 
1696 struct receive_arg  {
1697         objset_t *os;
1698         vnode_t *vp; /* The vnode to read the stream from */
1699         uint64_t voff; /* The current offset in the stream */
1700         uint64_t bytes_read;
1701         /*
1702          * A record that has had its payload read in, but hasn't yet been handed
1703          * off to the worker thread.
1704          */
1705         struct receive_record_arg *rrd;
1706         /* A record that has had its header read in, but not its payload. */
1707         struct receive_record_arg *next_rrd;
1708         zio_cksum_t cksum;
1709         zio_cksum_t prev_cksum;
1710         int err;
1711         boolean_t byteswap;
1712         /* Sorted list of objects not to issue prefetches for. */
1713         struct objlist ignore_objlist;
1714 };
1715 
1716 typedef struct guid_map_entry {
1717         uint64_t        guid;
1718         dsl_dataset_t   *gme_ds;
1719         avl_node_t      avlnode;
1720 } guid_map_entry_t;
1721 
1722 static int
1723 guid_compare(const void *arg1, const void *arg2)
1724 {
1725         const guid_map_entry_t *gmep1 = arg1;
1726         const guid_map_entry_t *gmep2 = arg2;
1727 
1728         if (gmep1->guid < gmep2->guid)
1729                 return (-1);
1730         else if (gmep1->guid > gmep2->guid)
1731                 return (1);
1732         return (0);
1733 }
1734 
1735 static void
1736 free_guid_map_onexit(void *arg)
1737 {
1738         avl_tree_t *ca = arg;
1739         void *cookie = NULL;
1740         guid_map_entry_t *gmep;
1741 
1742         while ((gmep = avl_destroy_nodes(ca, &cookie)) != NULL) {
1743                 dsl_dataset_long_rele(gmep->gme_ds, gmep);
1744                 dsl_dataset_rele(gmep->gme_ds, gmep);
1745                 kmem_free(gmep, sizeof (guid_map_entry_t));
1746         }
1747         avl_destroy(ca);
1748         kmem_free(ca, sizeof (avl_tree_t));
1749 }
1750 
1751 static int
1752 receive_read(struct receive_arg *ra, int len, void *buf)
1753 {
1754         int done = 0;
1755 
1756         /* some things will require 8-byte alignment, so everything must */
1757         ASSERT0(len % 8);
1758 
1759         while (done < len) {
1760                 ssize_t resid;
1761 
1762                 ra->err = vn_rdwr(UIO_READ, ra->vp,
1763                     (char *)buf + done, len - done,
1764                     ra->voff, UIO_SYSSPACE, FAPPEND,
1765                     RLIM64_INFINITY, CRED(), &resid);
1766 
1767                 if (resid == len - done) {
1768                         /*
1769                          * Note: ECKSUM indicates that the receive
1770                          * was interrupted and can potentially be resumed.
1771                          */
1772                         ra->err = SET_ERROR(ECKSUM);
1773                 }
1774                 ra->voff += len - done - resid;
1775                 done = len - resid;
1776                 if (ra->err != 0)
1777                         return (ra->err);
1778         }
1779 
1780         ra->bytes_read += len;
1781 
1782         ASSERT3U(done, ==, len);
1783         return (0);
1784 }
1785 
1786 static void
1787 byteswap_record(dmu_replay_record_t *drr)
1788 {
1789 #define DO64(X) (drr->drr_u.X = BSWAP_64(drr->drr_u.X))
1790 #define DO32(X) (drr->drr_u.X = BSWAP_32(drr->drr_u.X))
1791         drr->drr_type = BSWAP_32(drr->drr_type);
1792         drr->drr_payloadlen = BSWAP_32(drr->drr_payloadlen);
1793 
1794         switch (drr->drr_type) {
1795         case DRR_BEGIN:
1796                 DO64(drr_begin.drr_magic);
1797                 DO64(drr_begin.drr_versioninfo);
1798                 DO64(drr_begin.drr_creation_time);
1799                 DO32(drr_begin.drr_type);
1800                 DO32(drr_begin.drr_flags);
1801                 DO64(drr_begin.drr_toguid);
1802                 DO64(drr_begin.drr_fromguid);
1803                 break;
1804         case DRR_OBJECT:
1805                 DO64(drr_object.drr_object);
1806                 DO32(drr_object.drr_type);
1807                 DO32(drr_object.drr_bonustype);
1808                 DO32(drr_object.drr_blksz);
1809                 DO32(drr_object.drr_bonuslen);
1810                 DO64(drr_object.drr_toguid);
1811                 break;
1812         case DRR_FREEOBJECTS:
1813                 DO64(drr_freeobjects.drr_firstobj);
1814                 DO64(drr_freeobjects.drr_numobjs);
1815                 DO64(drr_freeobjects.drr_toguid);
1816                 break;
1817         case DRR_WRITE:
1818                 DO64(drr_write.drr_object);
1819                 DO32(drr_write.drr_type);
1820                 DO64(drr_write.drr_offset);
1821                 DO64(drr_write.drr_length);
1822                 DO64(drr_write.drr_toguid);
1823                 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_write.drr_key.ddk_cksum);
1824                 DO64(drr_write.drr_key.ddk_prop);
1825                 break;
1826         case DRR_WRITE_BYREF:
1827                 DO64(drr_write_byref.drr_object);
1828                 DO64(drr_write_byref.drr_offset);
1829                 DO64(drr_write_byref.drr_length);
1830                 DO64(drr_write_byref.drr_toguid);
1831                 DO64(drr_write_byref.drr_refguid);
1832                 DO64(drr_write_byref.drr_refobject);
1833                 DO64(drr_write_byref.drr_refoffset);
1834                 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_write_byref.
1835                     drr_key.ddk_cksum);
1836                 DO64(drr_write_byref.drr_key.ddk_prop);
1837                 break;
1838         case DRR_WRITE_EMBEDDED:
1839                 DO64(drr_write_embedded.drr_object);
1840                 DO64(drr_write_embedded.drr_offset);
1841                 DO64(drr_write_embedded.drr_length);
1842                 DO64(drr_write_embedded.drr_toguid);
1843                 DO32(drr_write_embedded.drr_lsize);
1844                 DO32(drr_write_embedded.drr_psize);
1845                 break;
1846         case DRR_FREE:
1847                 DO64(drr_free.drr_object);
1848                 DO64(drr_free.drr_offset);
1849                 DO64(drr_free.drr_length);
1850                 DO64(drr_free.drr_toguid);
1851                 break;
1852         case DRR_SPILL:
1853                 DO64(drr_spill.drr_object);
1854                 DO64(drr_spill.drr_length);
1855                 DO64(drr_spill.drr_toguid);
1856                 break;
1857         case DRR_END:
1858                 DO64(drr_end.drr_toguid);
1859                 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_end.drr_checksum);
1860                 break;
1861         }
1862 
1863         if (drr->drr_type != DRR_BEGIN) {
1864                 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_checksum.drr_checksum);
1865         }
1866 
1867 #undef DO64
1868 #undef DO32
1869 }
1870 
1871 static inline uint8_t
1872 deduce_nblkptr(dmu_object_type_t bonus_type, uint64_t bonus_size)
1873 {
1874         if (bonus_type == DMU_OT_SA) {
1875                 return (1);
1876         } else {
1877                 return (1 +
1878                     ((DN_MAX_BONUSLEN - bonus_size) >> SPA_BLKPTRSHIFT));
1879         }
1880 }
1881 
1882 static void
1883 save_resume_state(struct receive_writer_arg *rwa,
1884     uint64_t object, uint64_t offset, dmu_tx_t *tx)
1885 {
1886         int txgoff = dmu_tx_get_txg(tx) & TXG_MASK;
1887 
1888         if (!rwa->resumable)
1889                 return;
1890 
1891         /*
1892          * We use ds_resume_bytes[] != 0 to indicate that we need to
1893          * update this on disk, so it must not be 0.
1894          */
1895         ASSERT(rwa->bytes_read != 0);
1896 
1897         /*
1898          * We only resume from write records, which have a valid
1899          * (non-meta-dnode) object number.
1900          */
1901         ASSERT(object != 0);
1902 
1903         /*
1904          * For resuming to work correctly, we must receive records in order,
1905          * sorted by object,offset.  This is checked by the callers, but
1906          * assert it here for good measure.
1907          */
1908         ASSERT3U(object, >=, rwa->os->os_dsl_dataset->ds_resume_object[txgoff]);
1909         ASSERT(object != rwa->os->os_dsl_dataset->ds_resume_object[txgoff] ||
1910             offset >= rwa->os->os_dsl_dataset->ds_resume_offset[txgoff]);
1911         ASSERT3U(rwa->bytes_read, >=,
1912             rwa->os->os_dsl_dataset->ds_resume_bytes[txgoff]);
1913 
1914         rwa->os->os_dsl_dataset->ds_resume_object[txgoff] = object;
1915         rwa->os->os_dsl_dataset->ds_resume_offset[txgoff] = offset;
1916         rwa->os->os_dsl_dataset->ds_resume_bytes[txgoff] = rwa->bytes_read;
1917 }
1918 
1919 static int
1920 receive_object(struct receive_writer_arg *rwa, struct drr_object *drro,
1921     void *data)
1922 {
1923         dmu_object_info_t doi;
1924         dmu_tx_t *tx;
1925         uint64_t object;
1926         int err;
1927 
1928         if (drro->drr_type == DMU_OT_NONE ||
1929             !DMU_OT_IS_VALID(drro->drr_type) ||
1930             !DMU_OT_IS_VALID(drro->drr_bonustype) ||
1931             drro->drr_checksumtype >= ZIO_CHECKSUM_FUNCTIONS ||
1932             drro->drr_compress >= ZIO_COMPRESS_FUNCTIONS ||
1933             P2PHASE(drro->drr_blksz, SPA_MINBLOCKSIZE) ||
1934             drro->drr_blksz < SPA_MINBLOCKSIZE ||
1935             drro->drr_blksz > spa_maxblocksize(dmu_objset_spa(rwa->os)) ||
1936             drro->drr_bonuslen > DN_MAX_BONUSLEN) {
1937                 return (SET_ERROR(EINVAL));
1938         }
1939 
1940         err = dmu_object_info(rwa->os, drro->drr_object, &doi);
1941 
1942         if (err != 0 && err != ENOENT)
1943                 return (SET_ERROR(EINVAL));
1944         object = err == 0 ? drro->drr_object : DMU_NEW_OBJECT;
1945 
1946         /*
1947          * If we are losing blkptrs or changing the block size this must
1948          * be a new file instance.  We must clear out the previous file
1949          * contents before we can change this type of metadata in the dnode.
1950          */
1951         if (err == 0) {
1952                 int nblkptr;
1953 
1954                 nblkptr = deduce_nblkptr(drro->drr_bonustype,
1955                     drro->drr_bonuslen);
1956 
1957                 if (drro->drr_blksz != doi.doi_data_block_size ||
1958                     nblkptr < doi.doi_nblkptr) {
1959                         err = dmu_free_long_range(rwa->os, drro->drr_object,
1960                             0, DMU_OBJECT_END);
1961                         if (err != 0)
1962                                 return (SET_ERROR(EINVAL));
1963                 }
1964         }
1965 
1966         tx = dmu_tx_create(rwa->os);
1967         dmu_tx_hold_bonus(tx, object);
1968         err = dmu_tx_assign(tx, TXG_WAIT);
1969         if (err != 0) {
1970                 dmu_tx_abort(tx);
1971                 return (err);
1972         }
1973 
1974         if (object == DMU_NEW_OBJECT) {
1975                 /* currently free, want to be allocated */
1976                 err = dmu_object_claim(rwa->os, drro->drr_object,
1977                     drro->drr_type, drro->drr_blksz,
1978                     drro->drr_bonustype, drro->drr_bonuslen, tx);
1979         } else if (drro->drr_type != doi.doi_type ||
1980             drro->drr_blksz != doi.doi_data_block_size ||
1981             drro->drr_bonustype != doi.doi_bonus_type ||
1982             drro->drr_bonuslen != doi.doi_bonus_size) {
1983                 /* currently allocated, but with different properties */
1984                 err = dmu_object_reclaim(rwa->os, drro->drr_object,
1985                     drro->drr_type, drro->drr_blksz,
1986                     drro->drr_bonustype, drro->drr_bonuslen, tx);
1987         }
1988         if (err != 0) {
1989                 dmu_tx_commit(tx);
1990                 return (SET_ERROR(EINVAL));
1991         }
1992 
1993         dmu_object_set_checksum(rwa->os, drro->drr_object,
1994             drro->drr_checksumtype, tx);
1995         dmu_object_set_compress(rwa->os, drro->drr_object,
1996             drro->drr_compress, tx);
1997 
1998         if (data != NULL) {
1999                 dmu_buf_t *db;
2000 
2001                 VERIFY0(dmu_bonus_hold(rwa->os, drro->drr_object, FTAG, &db));
2002                 dmu_buf_will_dirty(db, tx);
2003 
2004                 ASSERT3U(db->db_size, >=, drro->drr_bonuslen);
2005                 bcopy(data, db->db_data, drro->drr_bonuslen);
2006                 if (rwa->byteswap) {
2007                         dmu_object_byteswap_t byteswap =
2008                             DMU_OT_BYTESWAP(drro->drr_bonustype);
2009                         dmu_ot_byteswap[byteswap].ob_func(db->db_data,
2010                             drro->drr_bonuslen);
2011                 }
2012                 dmu_buf_rele(db, FTAG);
2013         }
2014         dmu_tx_commit(tx);
2015 
2016         return (0);
2017 }
2018 
2019 /* ARGSUSED */
2020 static int
2021 receive_freeobjects(struct receive_writer_arg *rwa,
2022     struct drr_freeobjects *drrfo)
2023 {
2024         uint64_t obj;
2025         int next_err = 0;
2026 
2027         if (drrfo->drr_firstobj + drrfo->drr_numobjs < drrfo->drr_firstobj)
2028                 return (SET_ERROR(EINVAL));
2029 
2030         for (obj = drrfo->drr_firstobj;
2031             obj < drrfo->drr_firstobj + drrfo->drr_numobjs && next_err == 0;
2032             next_err = dmu_object_next(rwa->os, &obj, FALSE, 0)) {
2033                 int err;
2034 
2035                 if (dmu_object_info(rwa->os, obj, NULL) != 0)
2036                         continue;
2037 
2038                 err = dmu_free_long_object(rwa->os, obj);
2039                 if (err != 0)
2040                         return (err);
2041         }
2042         if (next_err != ESRCH)
2043                 return (next_err);
2044         return (0);
2045 }
2046 
2047 static int
2048 receive_write(struct receive_writer_arg *rwa, struct drr_write *drrw,
2049     arc_buf_t *abuf)
2050 {
2051         dmu_tx_t *tx;
2052         int err;
2053 
2054         if (drrw->drr_offset + drrw->drr_length < drrw->drr_offset ||
2055             !DMU_OT_IS_VALID(drrw->drr_type))
2056                 return (SET_ERROR(EINVAL));
2057 
2058         /*
2059          * For resuming to work, records must be in increasing order
2060          * by (object, offset).
2061          */
2062         if (drrw->drr_object < rwa->last_object ||
2063             (drrw->drr_object == rwa->last_object &&
2064             drrw->drr_offset < rwa->last_offset)) {
2065                 return (SET_ERROR(EINVAL));
2066         }
2067         rwa->last_object = drrw->drr_object;
2068         rwa->last_offset = drrw->drr_offset;
2069 
2070         if (dmu_object_info(rwa->os, drrw->drr_object, NULL) != 0)
2071                 return (SET_ERROR(EINVAL));
2072 
2073         tx = dmu_tx_create(rwa->os);
2074 
2075         dmu_tx_hold_write(tx, drrw->drr_object,
2076             drrw->drr_offset, drrw->drr_length);
2077         err = dmu_tx_assign(tx, TXG_WAIT);
2078         if (err != 0) {
2079                 dmu_tx_abort(tx);
2080                 return (err);
2081         }
2082         if (rwa->byteswap) {
2083                 dmu_object_byteswap_t byteswap =
2084                     DMU_OT_BYTESWAP(drrw->drr_type);
2085                 dmu_ot_byteswap[byteswap].ob_func(abuf->b_data,
2086                     drrw->drr_length);
2087         }
2088 
2089         dmu_buf_t *bonus;
2090         if (dmu_bonus_hold(rwa->os, drrw->drr_object, FTAG, &bonus) != 0)
2091                 return (SET_ERROR(EINVAL));
2092         dmu_assign_arcbuf(bonus, drrw->drr_offset, abuf, tx);
2093 
2094         /*
2095          * Note: If the receive fails, we want the resume stream to start
2096          * with the same record that we last successfully received (as opposed
2097          * to the next record), so that we can verify that we are
2098          * resuming from the correct location.
2099          */
2100         save_resume_state(rwa, drrw->drr_object, drrw->drr_offset, tx);
2101         dmu_tx_commit(tx);
2102         dmu_buf_rele(bonus, FTAG);
2103 
2104         return (0);
2105 }
2106 
2107 /*
2108  * Handle a DRR_WRITE_BYREF record.  This record is used in dedup'ed
2109  * streams to refer to a copy of the data that is already on the
2110  * system because it came in earlier in the stream.  This function
2111  * finds the earlier copy of the data, and uses that copy instead of
2112  * data from the stream to fulfill this write.
2113  */
2114 static int
2115 receive_write_byref(struct receive_writer_arg *rwa,
2116     struct drr_write_byref *drrwbr)
2117 {
2118         dmu_tx_t *tx;
2119         int err;
2120         guid_map_entry_t gmesrch;
2121         guid_map_entry_t *gmep;
2122         avl_index_t where;
2123         objset_t *ref_os = NULL;
2124         dmu_buf_t *dbp;
2125 
2126         if (drrwbr->drr_offset + drrwbr->drr_length < drrwbr->drr_offset)
2127                 return (SET_ERROR(EINVAL));
2128 
2129         /*
2130          * If the GUID of the referenced dataset is different from the
2131          * GUID of the target dataset, find the referenced dataset.
2132          */
2133         if (drrwbr->drr_toguid != drrwbr->drr_refguid) {
2134                 gmesrch.guid = drrwbr->drr_refguid;
2135                 if ((gmep = avl_find(rwa->guid_to_ds_map, &gmesrch,
2136                     &where)) == NULL) {
2137                         return (SET_ERROR(EINVAL));
2138                 }
2139                 if (dmu_objset_from_ds(gmep->gme_ds, &ref_os))
2140                         return (SET_ERROR(EINVAL));
2141         } else {
2142                 ref_os = rwa->os;
2143         }
2144 
2145         err = dmu_buf_hold(ref_os, drrwbr->drr_refobject,
2146             drrwbr->drr_refoffset, FTAG, &dbp, DMU_READ_PREFETCH);
2147         if (err != 0)
2148                 return (err);
2149 
2150         tx = dmu_tx_create(rwa->os);
2151 
2152         dmu_tx_hold_write(tx, drrwbr->drr_object,
2153             drrwbr->drr_offset, drrwbr->drr_length);
2154         err = dmu_tx_assign(tx, TXG_WAIT);
2155         if (err != 0) {
2156                 dmu_tx_abort(tx);
2157                 return (err);
2158         }
2159         dmu_write(rwa->os, drrwbr->drr_object,
2160             drrwbr->drr_offset, drrwbr->drr_length, dbp->db_data, tx);
2161         dmu_buf_rele(dbp, FTAG);
2162 
2163         /* See comment in restore_write. */
2164         save_resume_state(rwa, drrwbr->drr_object, drrwbr->drr_offset, tx);
2165         dmu_tx_commit(tx);
2166         return (0);
2167 }
2168 
2169 static int
2170 receive_write_embedded(struct receive_writer_arg *rwa,
2171     struct drr_write_embedded *drrwe, void *data)
2172 {
2173         dmu_tx_t *tx;
2174         int err;
2175 
2176         if (drrwe->drr_offset + drrwe->drr_length < drrwe->drr_offset)
2177                 return (EINVAL);
2178 
2179         if (drrwe->drr_psize > BPE_PAYLOAD_SIZE)
2180                 return (EINVAL);
2181 
2182         if (drrwe->drr_etype >= NUM_BP_EMBEDDED_TYPES)
2183                 return (EINVAL);
2184         if (drrwe->drr_compression >= ZIO_COMPRESS_FUNCTIONS)
2185                 return (EINVAL);
2186 
2187         tx = dmu_tx_create(rwa->os);
2188 
2189         dmu_tx_hold_write(tx, drrwe->drr_object,
2190             drrwe->drr_offset, drrwe->drr_length);
2191         err = dmu_tx_assign(tx, TXG_WAIT);
2192         if (err != 0) {
2193                 dmu_tx_abort(tx);
2194                 return (err);
2195         }
2196 
2197         dmu_write_embedded(rwa->os, drrwe->drr_object,
2198             drrwe->drr_offset, data, drrwe->drr_etype,
2199             drrwe->drr_compression, drrwe->drr_lsize, drrwe->drr_psize,
2200             rwa->byteswap ^ ZFS_HOST_BYTEORDER, tx);
2201 
2202         /* See comment in restore_write. */
2203         save_resume_state(rwa, drrwe->drr_object, drrwe->drr_offset, tx);
2204         dmu_tx_commit(tx);
2205         return (0);
2206 }
2207 
2208 static int
2209 receive_spill(struct receive_writer_arg *rwa, struct drr_spill *drrs,
2210     void *data)
2211 {
2212         dmu_tx_t *tx;
2213         dmu_buf_t *db, *db_spill;
2214         int err;
2215 
2216         if (drrs->drr_length < SPA_MINBLOCKSIZE ||
2217             drrs->drr_length > spa_maxblocksize(dmu_objset_spa(rwa->os)))
2218                 return (SET_ERROR(EINVAL));
2219 
2220         if (dmu_object_info(rwa->os, drrs->drr_object, NULL) != 0)
2221                 return (SET_ERROR(EINVAL));
2222 
2223         VERIFY0(dmu_bonus_hold(rwa->os, drrs->drr_object, FTAG, &db));
2224         if ((err = dmu_spill_hold_by_bonus(db, FTAG, &db_spill)) != 0) {
2225                 dmu_buf_rele(db, FTAG);
2226                 return (err);
2227         }
2228 
2229         tx = dmu_tx_create(rwa->os);
2230 
2231         dmu_tx_hold_spill(tx, db->db_object);
2232 
2233         err = dmu_tx_assign(tx, TXG_WAIT);
2234         if (err != 0) {
2235                 dmu_buf_rele(db, FTAG);
2236                 dmu_buf_rele(db_spill, FTAG);
2237                 dmu_tx_abort(tx);
2238                 return (err);
2239         }
2240         dmu_buf_will_dirty(db_spill, tx);
2241 
2242         if (db_spill->db_size < drrs->drr_length)
2243                 VERIFY(0 == dbuf_spill_set_blksz(db_spill,
2244                     drrs->drr_length, tx));
2245         bcopy(data, db_spill->db_data, drrs->drr_length);
2246 
2247         dmu_buf_rele(db, FTAG);
2248         dmu_buf_rele(db_spill, FTAG);
2249 
2250         dmu_tx_commit(tx);
2251         return (0);
2252 }
2253 
2254 /* ARGSUSED */
2255 static int
2256 receive_free(struct receive_writer_arg *rwa, struct drr_free *drrf)
2257 {
2258         int err;
2259 
2260         if (drrf->drr_length != -1ULL &&
2261             drrf->drr_offset + drrf->drr_length < drrf->drr_offset)
2262                 return (SET_ERROR(EINVAL));
2263 
2264         if (dmu_object_info(rwa->os, drrf->drr_object, NULL) != 0)
2265                 return (SET_ERROR(EINVAL));
2266 
2267         err = dmu_free_long_range(rwa->os, drrf->drr_object,
2268             drrf->drr_offset, drrf->drr_length);
2269 
2270         return (err);
2271 }
2272 
2273 /* used to destroy the drc_ds on error */
2274 static void
2275 dmu_recv_cleanup_ds(dmu_recv_cookie_t *drc)
2276 {
2277         if (drc->drc_resumable) {
2278                 /* wait for our resume state to be written to disk */
2279                 txg_wait_synced(drc->drc_ds->ds_dir->dd_pool, 0);
2280                 dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
2281         } else {
2282                 char name[MAXNAMELEN];
2283                 dsl_dataset_name(drc->drc_ds, name);
2284                 dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
2285                 (void) dsl_destroy_head(name);
2286         }
2287 }
2288 
2289 static void
2290 receive_cksum(struct receive_arg *ra, int len, void *buf)
2291 {
2292         if (ra->byteswap) {
2293                 fletcher_4_incremental_byteswap(buf, len, &ra->cksum);
2294         } else {
2295                 fletcher_4_incremental_native(buf, len, &ra->cksum);
2296         }
2297 }
2298 
2299 /*
2300  * Read the payload into a buffer of size len, and update the current record's
2301  * payload field.
2302  * Allocate ra->next_rrd and read the next record's header into
2303  * ra->next_rrd->header.
2304  * Verify checksum of payload and next record.
2305  */
2306 static int
2307 receive_read_payload_and_next_header(struct receive_arg *ra, int len, void *buf)
2308 {
2309         int err;
2310 
2311         if (len != 0) {
2312                 ASSERT3U(len, <=, SPA_MAXBLOCKSIZE);
2313                 err = receive_read(ra, len, buf);
2314                 if (err != 0)
2315                         return (err);
2316                 receive_cksum(ra, len, buf);
2317 
2318                 /* note: rrd is NULL when reading the begin record's payload */
2319                 if (ra->rrd != NULL) {
2320                         ra->rrd->payload = buf;
2321                         ra->rrd->payload_size = len;
2322                         ra->rrd->bytes_read = ra->bytes_read;
2323                 }
2324         }
2325 
2326         ra->prev_cksum = ra->cksum;
2327 
2328         ra->next_rrd = kmem_zalloc(sizeof (*ra->next_rrd), KM_SLEEP);
2329         err = receive_read(ra, sizeof (ra->next_rrd->header),
2330             &ra->next_rrd->header);
2331         ra->next_rrd->bytes_read = ra->bytes_read;
2332         if (err != 0) {
2333                 kmem_free(ra->next_rrd, sizeof (*ra->next_rrd));
2334                 ra->next_rrd = NULL;
2335                 return (err);
2336         }
2337         if (ra->next_rrd->header.drr_type == DRR_BEGIN) {
2338                 kmem_free(ra->next_rrd, sizeof (*ra->next_rrd));
2339                 ra->next_rrd = NULL;
2340                 return (SET_ERROR(EINVAL));
2341         }
2342 
2343         /*
2344          * Note: checksum is of everything up to but not including the
2345          * checksum itself.
2346          */
2347         ASSERT3U(offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
2348             ==, sizeof (dmu_replay_record_t) - sizeof (zio_cksum_t));
2349         receive_cksum(ra,
2350             offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
2351             &ra->next_rrd->header);
2352 
2353         zio_cksum_t cksum_orig =
2354             ra->next_rrd->header.drr_u.drr_checksum.drr_checksum;
2355         zio_cksum_t *cksump =
2356             &ra->next_rrd->header.drr_u.drr_checksum.drr_checksum;
2357 
2358         if (ra->byteswap)
2359                 byteswap_record(&ra->next_rrd->header);
2360 
2361         if ((!ZIO_CHECKSUM_IS_ZERO(cksump)) &&
2362             !ZIO_CHECKSUM_EQUAL(ra->cksum, *cksump)) {
2363                 kmem_free(ra->next_rrd, sizeof (*ra->next_rrd));
2364                 ra->next_rrd = NULL;
2365                 return (SET_ERROR(ECKSUM));
2366         }
2367 
2368         receive_cksum(ra, sizeof (cksum_orig), &cksum_orig);
2369 
2370         return (0);
2371 }
2372 
2373 static void
2374 objlist_create(struct objlist *list)
2375 {
2376         list_create(&list->list, sizeof (struct receive_objnode),
2377             offsetof(struct receive_objnode, node));
2378         list->last_lookup = 0;
2379 }
2380 
2381 static void
2382 objlist_destroy(struct objlist *list)
2383 {
2384         for (struct receive_objnode *n = list_remove_head(&list->list);
2385             n != NULL; n = list_remove_head(&list->list)) {
2386                 kmem_free(n, sizeof (*n));
2387         }
2388         list_destroy(&list->list);
2389 }
2390 
2391 /*
2392  * This function looks through the objlist to see if the specified object number
2393  * is contained in the objlist.  In the process, it will remove all object
2394  * numbers in the list that are smaller than the specified object number.  Thus,
2395  * any lookup of an object number smaller than a previously looked up object
2396  * number will always return false; therefore, all lookups should be done in
2397  * ascending order.
2398  */
2399 static boolean_t
2400 objlist_exists(struct objlist *list, uint64_t object)
2401 {
2402         struct receive_objnode *node = list_head(&list->list);
2403         ASSERT3U(object, >=, list->last_lookup);
2404         list->last_lookup = object;
2405         while (node != NULL && node->object < object) {
2406                 VERIFY3P(node, ==, list_remove_head(&list->list));
2407                 kmem_free(node, sizeof (*node));
2408                 node = list_head(&list->list);
2409         }
2410         return (node != NULL && node->object == object);
2411 }
2412 
2413 /*
2414  * The objlist is a list of object numbers stored in ascending order.  However,
2415  * the insertion of new object numbers does not seek out the correct location to
2416  * store a new object number; instead, it appends it to the list for simplicity.
2417  * Thus, any users must take care to only insert new object numbers in ascending
2418  * order.
2419  */
2420 static void
2421 objlist_insert(struct objlist *list, uint64_t object)
2422 {
2423         struct receive_objnode *node = kmem_zalloc(sizeof (*node), KM_SLEEP);
2424         node->object = object;
2425 #ifdef ZFS_DEBUG
2426         struct receive_objnode *last_object = list_tail(&list->list);
2427         uint64_t last_objnum = (last_object != NULL ? last_object->object : 0);
2428         ASSERT3U(node->object, >, last_objnum);
2429 #endif
2430         list_insert_tail(&list->list, node);
2431 }
2432 
2433 /*
2434  * Issue the prefetch reads for any necessary indirect blocks.
2435  *
2436  * We use the object ignore list to tell us whether or not to issue prefetches
2437  * for a given object.  We do this for both correctness (in case the blocksize
2438  * of an object has changed) and performance (if the object doesn't exist, don't
2439  * needlessly try to issue prefetches).  We also trim the list as we go through
2440  * the stream to prevent it from growing to an unbounded size.
2441  *
2442  * The object numbers within will always be in sorted order, and any write
2443  * records we see will also be in sorted order, but they're not sorted with
2444  * respect to each other (i.e. we can get several object records before
2445  * receiving each object's write records).  As a result, once we've reached a
2446  * given object number, we can safely remove any reference to lower object
2447  * numbers in the ignore list. In practice, we receive up to 32 object records
2448  * before receiving write records, so the list can have up to 32 nodes in it.
2449  */
2450 /* ARGSUSED */
2451 static void
2452 receive_read_prefetch(struct receive_arg *ra,
2453     uint64_t object, uint64_t offset, uint64_t length)
2454 {
2455         if (!objlist_exists(&ra->ignore_objlist, object)) {
2456                 dmu_prefetch(ra->os, object, 1, offset, length,
2457                     ZIO_PRIORITY_SYNC_READ);
2458         }
2459 }
2460 
2461 /*
2462  * Read records off the stream, issuing any necessary prefetches.
2463  */
2464 static int
2465 receive_read_record(struct receive_arg *ra)
2466 {
2467         int err;
2468 
2469         switch (ra->rrd->header.drr_type) {
2470         case DRR_OBJECT:
2471         {
2472                 struct drr_object *drro = &ra->rrd->header.drr_u.drr_object;
2473                 uint32_t size = P2ROUNDUP(drro->drr_bonuslen, 8);
2474                 void *buf = kmem_zalloc(size, KM_SLEEP);
2475                 dmu_object_info_t doi;
2476                 err = receive_read_payload_and_next_header(ra, size, buf);
2477                 if (err != 0) {
2478                         kmem_free(buf, size);
2479                         return (err);
2480                 }
2481                 err = dmu_object_info(ra->os, drro->drr_object, &doi);
2482                 /*
2483                  * See receive_read_prefetch for an explanation why we're
2484                  * storing this object in the ignore_obj_list.
2485                  */
2486                 if (err == ENOENT ||
2487                     (err == 0 && doi.doi_data_block_size != drro->drr_blksz)) {
2488                         objlist_insert(&ra->ignore_objlist, drro->drr_object);
2489                         err = 0;
2490                 }
2491                 return (err);
2492         }
2493         case DRR_FREEOBJECTS:
2494         {
2495                 err = receive_read_payload_and_next_header(ra, 0, NULL);
2496                 return (err);
2497         }
2498         case DRR_WRITE:
2499         {
2500                 struct drr_write *drrw = &ra->rrd->header.drr_u.drr_write;
2501                 arc_buf_t *abuf = arc_loan_buf(dmu_objset_spa(ra->os),
2502                     drrw->drr_length);
2503 
2504                 err = receive_read_payload_and_next_header(ra,
2505                     drrw->drr_length, abuf->b_data);
2506                 if (err != 0) {
2507                         dmu_return_arcbuf(abuf);
2508                         return (err);
2509                 }
2510                 ra->rrd->write_buf = abuf;
2511                 receive_read_prefetch(ra, drrw->drr_object, drrw->drr_offset,
2512                     drrw->drr_length);
2513                 return (err);
2514         }
2515         case DRR_WRITE_BYREF:
2516         {
2517                 struct drr_write_byref *drrwb =
2518                     &ra->rrd->header.drr_u.drr_write_byref;
2519                 err = receive_read_payload_and_next_header(ra, 0, NULL);
2520                 receive_read_prefetch(ra, drrwb->drr_object, drrwb->drr_offset,
2521                     drrwb->drr_length);
2522                 return (err);
2523         }
2524         case DRR_WRITE_EMBEDDED:
2525         {
2526                 struct drr_write_embedded *drrwe =
2527                     &ra->rrd->header.drr_u.drr_write_embedded;
2528                 uint32_t size = P2ROUNDUP(drrwe->drr_psize, 8);
2529                 void *buf = kmem_zalloc(size, KM_SLEEP);
2530 
2531                 err = receive_read_payload_and_next_header(ra, size, buf);
2532                 if (err != 0) {
2533                         kmem_free(buf, size);
2534                         return (err);
2535                 }
2536 
2537                 receive_read_prefetch(ra, drrwe->drr_object, drrwe->drr_offset,
2538                     drrwe->drr_length);
2539                 return (err);
2540         }
2541         case DRR_FREE:
2542         {
2543                 /*
2544                  * It might be beneficial to prefetch indirect blocks here, but
2545                  * we don't really have the data to decide for sure.
2546                  */
2547                 err = receive_read_payload_and_next_header(ra, 0, NULL);
2548                 return (err);
2549         }
2550         case DRR_END:
2551         {
2552                 struct drr_end *drre = &ra->rrd->header.drr_u.drr_end;
2553                 if (!ZIO_CHECKSUM_EQUAL(ra->prev_cksum, drre->drr_checksum))
2554                         return (SET_ERROR(ECKSUM));
2555                 return (0);
2556         }
2557         case DRR_SPILL:
2558         {
2559                 struct drr_spill *drrs = &ra->rrd->header.drr_u.drr_spill;
2560                 void *buf = kmem_zalloc(drrs->drr_length, KM_SLEEP);
2561                 err = receive_read_payload_and_next_header(ra, drrs->drr_length,
2562                     buf);
2563                 if (err != 0)
2564                         kmem_free(buf, drrs->drr_length);
2565                 return (err);
2566         }
2567         default:
2568                 return (SET_ERROR(EINVAL));
2569         }
2570 }
2571 
2572 /*
2573  * Commit the records to the pool.
2574  */
2575 static int
2576 receive_process_record(struct receive_writer_arg *rwa,
2577     struct receive_record_arg *rrd)
2578 {
2579         int err;
2580 
2581         /* Processing in order, therefore bytes_read should be increasing. */
2582         ASSERT3U(rrd->bytes_read, >=, rwa->bytes_read);
2583         rwa->bytes_read = rrd->bytes_read;
2584 
2585         switch (rrd->header.drr_type) {
2586         case DRR_OBJECT:
2587         {
2588                 struct drr_object *drro = &rrd->header.drr_u.drr_object;
2589                 err = receive_object(rwa, drro, rrd->payload);
2590                 kmem_free(rrd->payload, rrd->payload_size);
2591                 rrd->payload = NULL;
2592                 return (err);
2593         }
2594         case DRR_FREEOBJECTS:
2595         {
2596                 struct drr_freeobjects *drrfo =
2597                     &rrd->header.drr_u.drr_freeobjects;
2598                 return (receive_freeobjects(rwa, drrfo));
2599         }
2600         case DRR_WRITE:
2601         {
2602                 struct drr_write *drrw = &rrd->header.drr_u.drr_write;
2603                 err = receive_write(rwa, drrw, rrd->write_buf);
2604                 /* if receive_write() is successful, it consumes the arc_buf */
2605                 if (err != 0)
2606                         dmu_return_arcbuf(rrd->write_buf);
2607                 rrd->write_buf = NULL;
2608                 rrd->payload = NULL;
2609                 return (err);
2610         }
2611         case DRR_WRITE_BYREF:
2612         {
2613                 struct drr_write_byref *drrwbr =
2614                     &rrd->header.drr_u.drr_write_byref;
2615                 return (receive_write_byref(rwa, drrwbr));
2616         }
2617         case DRR_WRITE_EMBEDDED:
2618         {
2619                 struct drr_write_embedded *drrwe =
2620                     &rrd->header.drr_u.drr_write_embedded;
2621                 err = receive_write_embedded(rwa, drrwe, rrd->payload);
2622                 kmem_free(rrd->payload, rrd->payload_size);
2623                 rrd->payload = NULL;
2624                 return (err);
2625         }
2626         case DRR_FREE:
2627         {
2628                 struct drr_free *drrf = &rrd->header.drr_u.drr_free;
2629                 return (receive_free(rwa, drrf));
2630         }
2631         case DRR_SPILL:
2632         {
2633                 struct drr_spill *drrs = &rrd->header.drr_u.drr_spill;
2634                 err = receive_spill(rwa, drrs, rrd->payload);
2635                 kmem_free(rrd->payload, rrd->payload_size);
2636                 rrd->payload = NULL;
2637                 return (err);
2638         }
2639         default:
2640                 return (SET_ERROR(EINVAL));
2641         }
2642 }
2643 
2644 /*
2645  * dmu_recv_stream's worker thread; pull records off the queue, and then call
2646  * receive_process_record  When we're done, signal the main thread and exit.
2647  */
2648 static void
2649 receive_writer_thread(void *arg)
2650 {
2651         struct receive_writer_arg *rwa = arg;
2652         struct receive_record_arg *rrd;
2653         for (rrd = bqueue_dequeue(&rwa->q); !rrd->eos_marker;
2654             rrd = bqueue_dequeue(&rwa->q)) {
2655                 /*
2656                  * If there's an error, the main thread will stop putting things
2657                  * on the queue, but we need to clear everything in it before we
2658                  * can exit.
2659                  */
2660                 if (rwa->err == 0) {
2661                         rwa->err = receive_process_record(rwa, rrd);
2662                 } else if (rrd->write_buf != NULL) {
2663                         dmu_return_arcbuf(rrd->write_buf);
2664                         rrd->write_buf = NULL;
2665                         rrd->payload = NULL;
2666                 } else if (rrd->payload != NULL) {
2667                         kmem_free(rrd->payload, rrd->payload_size);
2668                         rrd->payload = NULL;
2669                 }
2670                 kmem_free(rrd, sizeof (*rrd));
2671         }
2672         kmem_free(rrd, sizeof (*rrd));
2673         mutex_enter(&rwa->mutex);
2674         rwa->done = B_TRUE;
2675         cv_signal(&rwa->cv);
2676         mutex_exit(&rwa->mutex);
2677 }
2678 
2679 static int
2680 resume_check(struct receive_arg *ra, nvlist_t *begin_nvl)
2681 {
2682         uint64_t val;
2683         objset_t *mos = dmu_objset_pool(ra->os)->dp_meta_objset;
2684         uint64_t dsobj = dmu_objset_id(ra->os);
2685         uint64_t resume_obj, resume_off;
2686 
2687         if (nvlist_lookup_uint64(begin_nvl,
2688             "resume_object", &resume_obj) != 0 ||
2689             nvlist_lookup_uint64(begin_nvl,
2690             "resume_offset", &resume_off) != 0) {
2691                 return (SET_ERROR(EINVAL));
2692         }
2693         VERIFY0(zap_lookup(mos, dsobj,
2694             DS_FIELD_RESUME_OBJECT, sizeof (val), 1, &val));
2695         if (resume_obj != val)
2696                 return (SET_ERROR(EINVAL));
2697         VERIFY0(zap_lookup(mos, dsobj,
2698             DS_FIELD_RESUME_OFFSET, sizeof (val), 1, &val));
2699         if (resume_off != val)
2700                 return (SET_ERROR(EINVAL));
2701 
2702         return (0);
2703 }
2704 
2705 /*
2706  * Read in the stream's records, one by one, and apply them to the pool.  There
2707  * are two threads involved; the thread that calls this function will spin up a
2708  * worker thread, read the records off the stream one by one, and issue
2709  * prefetches for any necessary indirect blocks.  It will then push the records
2710  * onto an internal blocking queue.  The worker thread will pull the records off
2711  * the queue, and actually write the data into the DMU.  This way, the worker
2712  * thread doesn't have to wait for reads to complete, since everything it needs
2713  * (the indirect blocks) will be prefetched.
2714  *
2715  * NB: callers *must* call dmu_recv_end() if this succeeds.
2716  */
2717 int
2718 dmu_recv_stream(dmu_recv_cookie_t *drc, vnode_t *vp, offset_t *voffp,
2719     int cleanup_fd, uint64_t *action_handlep)
2720 {
2721         int err = 0;
2722         struct receive_arg ra = { 0 };
2723         struct receive_writer_arg rwa = { 0 };
2724         int featureflags;
2725         nvlist_t *begin_nvl = NULL;
2726 
2727         ra.byteswap = drc->drc_byteswap;
2728         ra.cksum = drc->drc_cksum;
2729         ra.vp = vp;
2730         ra.voff = *voffp;
2731 
2732         if (dsl_dataset_is_zapified(drc->drc_ds)) {
2733                 (void) zap_lookup(drc->drc_ds->ds_dir->dd_pool->dp_meta_objset,
2734                     drc->drc_ds->ds_object, DS_FIELD_RESUME_BYTES,
2735                     sizeof (ra.bytes_read), 1, &ra.bytes_read);
2736         }
2737 
2738         objlist_create(&ra.ignore_objlist);
2739 
2740         /* these were verified in dmu_recv_begin */
2741         ASSERT3U(DMU_GET_STREAM_HDRTYPE(drc->drc_drrb->drr_versioninfo), ==,
2742             DMU_SUBSTREAM);
2743         ASSERT3U(drc->drc_drrb->drr_type, <, DMU_OST_NUMTYPES);
2744 
2745         /*
2746          * Open the objset we are modifying.
2747          */
2748         VERIFY0(dmu_objset_from_ds(drc->drc_ds, &ra.os));
2749 
2750         ASSERT(dsl_dataset_phys(drc->drc_ds)->ds_flags & DS_FLAG_INCONSISTENT);
2751 
2752         featureflags = DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo);
2753 
2754         /* if this stream is dedup'ed, set up the avl tree for guid mapping */
2755         if (featureflags & DMU_BACKUP_FEATURE_DEDUP) {
2756                 minor_t minor;
2757 
2758                 if (cleanup_fd == -1) {
2759                         ra.err = SET_ERROR(EBADF);
2760                         goto out;
2761                 }
2762                 ra.err = zfs_onexit_fd_hold(cleanup_fd, &minor);
2763                 if (ra.err != 0) {
2764                         cleanup_fd = -1;
2765                         goto out;
2766                 }
2767 
2768                 if (*action_handlep == 0) {
2769                         rwa.guid_to_ds_map =
2770                             kmem_alloc(sizeof (avl_tree_t), KM_SLEEP);
2771                         avl_create(rwa.guid_to_ds_map, guid_compare,
2772                             sizeof (guid_map_entry_t),
2773                             offsetof(guid_map_entry_t, avlnode));
2774                         err = zfs_onexit_add_cb(minor,
2775                             free_guid_map_onexit, rwa.guid_to_ds_map,
2776                             action_handlep);
2777                         if (ra.err != 0)
2778                                 goto out;
2779                 } else {
2780                         err = zfs_onexit_cb_data(minor, *action_handlep,
2781                             (void **)&rwa.guid_to_ds_map);
2782                         if (ra.err != 0)
2783                                 goto out;
2784                 }
2785 
2786                 drc->drc_guid_to_ds_map = rwa.guid_to_ds_map;
2787         }
2788 
2789         uint32_t payloadlen = drc->drc_drr_begin->drr_payloadlen;
2790         void *payload = NULL;
2791         if (payloadlen != 0)
2792                 payload = kmem_alloc(payloadlen, KM_SLEEP);
2793 
2794         err = receive_read_payload_and_next_header(&ra, payloadlen, payload);
2795         if (err != 0) {
2796                 if (payloadlen != 0)
2797                         kmem_free(payload, payloadlen);
2798                 goto out;
2799         }
2800         if (payloadlen != 0) {
2801                 err = nvlist_unpack(payload, payloadlen, &begin_nvl, KM_SLEEP);
2802                 kmem_free(payload, payloadlen);
2803                 if (err != 0)
2804                         goto out;
2805         }
2806 
2807         if (featureflags & DMU_BACKUP_FEATURE_RESUMING) {
2808                 err = resume_check(&ra, begin_nvl);
2809                 if (err != 0)
2810                         goto out;
2811         }
2812 
2813         (void) bqueue_init(&rwa.q, zfs_recv_queue_length,
2814             offsetof(struct receive_record_arg, node));
2815         cv_init(&rwa.cv, NULL, CV_DEFAULT, NULL);
2816         mutex_init(&rwa.mutex, NULL, MUTEX_DEFAULT, NULL);
2817         rwa.os = ra.os;
2818         rwa.byteswap = drc->drc_byteswap;
2819         rwa.resumable = drc->drc_resumable;
2820 
2821         (void) thread_create(NULL, 0, receive_writer_thread, &rwa, 0, curproc,
2822             TS_RUN, minclsyspri);
2823         /*
2824          * We're reading rwa.err without locks, which is safe since we are the
2825          * only reader, and the worker thread is the only writer.  It's ok if we
2826          * miss a write for an iteration or two of the loop, since the writer
2827          * thread will keep freeing records we send it until we send it an eos
2828          * marker.
2829          *
2830          * We can leave this loop in 3 ways:  First, if rwa.err is
2831          * non-zero.  In that case, the writer thread will free the rrd we just
2832          * pushed.  Second, if  we're interrupted; in that case, either it's the
2833          * first loop and ra.rrd was never allocated, or it's later, and ra.rrd
2834          * has been handed off to the writer thread who will free it.  Finally,
2835          * if receive_read_record fails or we're at the end of the stream, then
2836          * we free ra.rrd and exit.
2837          */
2838         while (rwa.err == 0) {
2839                 if (issig(JUSTLOOKING) && issig(FORREAL)) {
2840                         err = SET_ERROR(EINTR);
2841                         break;
2842                 }
2843 
2844                 ASSERT3P(ra.rrd, ==, NULL);
2845                 ra.rrd = ra.next_rrd;
2846                 ra.next_rrd = NULL;
2847                 /* Allocates and loads header into ra.next_rrd */
2848                 err = receive_read_record(&ra);
2849 
2850                 if (ra.rrd->header.drr_type == DRR_END || err != 0) {
2851                         kmem_free(ra.rrd, sizeof (*ra.rrd));
2852                         ra.rrd = NULL;
2853                         break;
2854                 }
2855 
2856                 bqueue_enqueue(&rwa.q, ra.rrd,
2857                     sizeof (struct receive_record_arg) + ra.rrd->payload_size);
2858                 ra.rrd = NULL;
2859         }
2860         if (ra.next_rrd == NULL)
2861                 ra.next_rrd = kmem_zalloc(sizeof (*ra.next_rrd), KM_SLEEP);
2862         ra.next_rrd->eos_marker = B_TRUE;
2863         bqueue_enqueue(&rwa.q, ra.next_rrd, 1);
2864 
2865         mutex_enter(&rwa.mutex);
2866         while (!rwa.done) {
2867                 cv_wait(&rwa.cv, &rwa.mutex);
2868         }
2869         mutex_exit(&rwa.mutex);
2870 
2871         cv_destroy(&rwa.cv);
2872         mutex_destroy(&rwa.mutex);
2873         bqueue_destroy(&rwa.q);
2874         if (err == 0)
2875                 err = rwa.err;
2876 
2877 out:
2878         nvlist_free(begin_nvl);
2879         if ((featureflags & DMU_BACKUP_FEATURE_DEDUP) && (cleanup_fd != -1))
2880                 zfs_onexit_fd_rele(cleanup_fd);
2881 
2882         if (err != 0) {
2883                 /*
2884                  * Clean up references. If receive is not resumable,
2885                  * destroy what we created, so we don't leave it in
2886                  * the inconsistent state.
2887                  */
2888                 dmu_recv_cleanup_ds(drc);
2889         }
2890 
2891         *voffp = ra.voff;
2892         objlist_destroy(&ra.ignore_objlist);
2893         return (err);
2894 }
2895 
2896 static int
2897 dmu_recv_end_check(void *arg, dmu_tx_t *tx)
2898 {
2899         dmu_recv_cookie_t *drc = arg;
2900         dsl_pool_t *dp = dmu_tx_pool(tx);
2901         int error;
2902 
2903         ASSERT3P(drc->drc_ds->ds_owner, ==, dmu_recv_tag);
2904 
2905         if (!drc->drc_newfs) {
2906                 dsl_dataset_t *origin_head;
2907 
2908                 error = dsl_dataset_hold(dp, drc->drc_tofs, FTAG, &origin_head);
2909                 if (error != 0)
2910                         return (error);
2911                 if (drc->drc_force) {
2912                         /*
2913                          * We will destroy any snapshots in tofs (i.e. before
2914                          * origin_head) that are after the origin (which is
2915                          * the snap before drc_ds, because drc_ds can not
2916                          * have any snaps of its own).
2917                          */
2918                         uint64_t obj;
2919 
2920                         obj = dsl_dataset_phys(origin_head)->ds_prev_snap_obj;
2921                         while (obj !=
2922                             dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj) {
2923                                 dsl_dataset_t *snap;
2924                                 error = dsl_dataset_hold_obj(dp, obj, FTAG,
2925                                     &snap);
2926                                 if (error != 0)
2927                                         break;
2928                                 if (snap->ds_dir != origin_head->ds_dir)
2929                                         error = SET_ERROR(EINVAL);
2930                                 if (error == 0)  {
2931                                         error = dsl_destroy_snapshot_check_impl(
2932                                             snap, B_FALSE);
2933                                 }
2934                                 obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
2935                                 dsl_dataset_rele(snap, FTAG);
2936                                 if (error != 0)
2937                                         break;
2938                         }
2939                         if (error != 0) {
2940                                 dsl_dataset_rele(origin_head, FTAG);
2941                                 return (error);
2942                         }
2943                 }
2944                 error = dsl_dataset_clone_swap_check_impl(drc->drc_ds,
2945                     origin_head, drc->drc_force, drc->drc_owner, tx);
2946                 if (error != 0) {
2947                         dsl_dataset_rele(origin_head, FTAG);
2948                         return (error);
2949                 }
2950                 error = dsl_dataset_snapshot_check_impl(origin_head,
2951                     drc->drc_tosnap, tx, B_TRUE, 1, drc->drc_cred);
2952                 dsl_dataset_rele(origin_head, FTAG);
2953                 if (error != 0)
2954                         return (error);
2955 
2956                 error = dsl_destroy_head_check_impl(drc->drc_ds, 1);
2957         } else {
2958                 error = dsl_dataset_snapshot_check_impl(drc->drc_ds,
2959                     drc->drc_tosnap, tx, B_TRUE, 1, drc->drc_cred);
2960         }
2961         return (error);
2962 }
2963 
2964 static void
2965 dmu_recv_end_sync(void *arg, dmu_tx_t *tx)
2966 {
2967         dmu_recv_cookie_t *drc = arg;
2968         dsl_pool_t *dp = dmu_tx_pool(tx);
2969 
2970         spa_history_log_internal_ds(drc->drc_ds, "finish receiving",
2971             tx, "snap=%s", drc->drc_tosnap);
2972 
2973         if (!drc->drc_newfs) {
2974                 dsl_dataset_t *origin_head;
2975 
2976                 VERIFY0(dsl_dataset_hold(dp, drc->drc_tofs, FTAG,
2977                     &origin_head));
2978 
2979                 if (drc->drc_force) {
2980                         /*
2981                          * Destroy any snapshots of drc_tofs (origin_head)
2982                          * after the origin (the snap before drc_ds).
2983                          */
2984                         uint64_t obj;
2985 
2986                         obj = dsl_dataset_phys(origin_head)->ds_prev_snap_obj;
2987                         while (obj !=
2988                             dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj) {
2989                                 dsl_dataset_t *snap;
2990                                 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG,
2991                                     &snap));
2992                                 ASSERT3P(snap->ds_dir, ==, origin_head->ds_dir);
2993                                 obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
2994                                 dsl_destroy_snapshot_sync_impl(snap,
2995                                     B_FALSE, tx);
2996                                 dsl_dataset_rele(snap, FTAG);
2997                         }
2998                 }
2999                 VERIFY3P(drc->drc_ds->ds_prev, ==,
3000                     origin_head->ds_prev);
3001 
3002                 dsl_dataset_clone_swap_sync_impl(drc->drc_ds,
3003                     origin_head, tx);
3004                 dsl_dataset_snapshot_sync_impl(origin_head,
3005                     drc->drc_tosnap, tx);
3006 
3007                 /* set snapshot's creation time and guid */
3008                 dmu_buf_will_dirty(origin_head->ds_prev->ds_dbuf, tx);
3009                 dsl_dataset_phys(origin_head->ds_prev)->ds_creation_time =
3010                     drc->drc_drrb->drr_creation_time;
3011                 dsl_dataset_phys(origin_head->ds_prev)->ds_guid =
3012                     drc->drc_drrb->drr_toguid;
3013                 dsl_dataset_phys(origin_head->ds_prev)->ds_flags &=
3014                     ~DS_FLAG_INCONSISTENT;
3015 
3016                 dmu_buf_will_dirty(origin_head->ds_dbuf, tx);
3017                 dsl_dataset_phys(origin_head)->ds_flags &=
3018                     ~DS_FLAG_INCONSISTENT;
3019 
3020                 dsl_dataset_rele(origin_head, FTAG);
3021                 dsl_destroy_head_sync_impl(drc->drc_ds, tx);
3022 
3023                 if (drc->drc_owner != NULL)
3024                         VERIFY3P(origin_head->ds_owner, ==, drc->drc_owner);
3025         } else {
3026                 dsl_dataset_t *ds = drc->drc_ds;
3027 
3028                 dsl_dataset_snapshot_sync_impl(ds, drc->drc_tosnap, tx);
3029 
3030                 /* set snapshot's creation time and guid */
3031                 dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
3032                 dsl_dataset_phys(ds->ds_prev)->ds_creation_time =
3033                     drc->drc_drrb->drr_creation_time;
3034                 dsl_dataset_phys(ds->ds_prev)->ds_guid =
3035                     drc->drc_drrb->drr_toguid;
3036                 dsl_dataset_phys(ds->ds_prev)->ds_flags &=
3037                     ~DS_FLAG_INCONSISTENT;
3038 
3039                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
3040                 dsl_dataset_phys(ds)->ds_flags &= ~DS_FLAG_INCONSISTENT;
3041                 if (dsl_dataset_has_resume_receive_state(ds)) {
3042                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3043                             DS_FIELD_RESUME_FROMGUID, tx);
3044                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3045                             DS_FIELD_RESUME_OBJECT, tx);
3046                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3047                             DS_FIELD_RESUME_OFFSET, tx);
3048                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3049                             DS_FIELD_RESUME_BYTES, tx);
3050                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3051                             DS_FIELD_RESUME_TOGUID, tx);
3052                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3053                             DS_FIELD_RESUME_TONAME, tx);
3054                 }
3055         }
3056         drc->drc_newsnapobj = dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj;
3057         /*
3058          * Release the hold from dmu_recv_begin.  This must be done before
3059          * we return to open context, so that when we free the dataset's dnode,
3060          * we can evict its bonus buffer.
3061          */
3062         dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
3063         drc->drc_ds = NULL;
3064 }
3065 
3066 static int
3067 add_ds_to_guidmap(const char *name, avl_tree_t *guid_map, uint64_t snapobj)
3068 {
3069         dsl_pool_t *dp;
3070         dsl_dataset_t *snapds;
3071         guid_map_entry_t *gmep;
3072         int err;
3073 
3074         ASSERT(guid_map != NULL);
3075 
3076         err = dsl_pool_hold(name, FTAG, &dp);
3077         if (err != 0)
3078                 return (err);
3079         gmep = kmem_alloc(sizeof (*gmep), KM_SLEEP);
3080         err = dsl_dataset_hold_obj(dp, snapobj, gmep, &snapds);
3081         if (err == 0) {
3082                 gmep->guid = dsl_dataset_phys(snapds)->ds_guid;
3083                 gmep->gme_ds = snapds;
3084                 avl_add(guid_map, gmep);
3085                 dsl_dataset_long_hold(snapds, gmep);
3086         } else {
3087                 kmem_free(gmep, sizeof (*gmep));
3088         }
3089 
3090         dsl_pool_rele(dp, FTAG);
3091         return (err);
3092 }
3093 
3094 static int dmu_recv_end_modified_blocks = 3;
3095 
3096 static int
3097 dmu_recv_existing_end(dmu_recv_cookie_t *drc)
3098 {
3099         int error;
3100         char name[MAXNAMELEN];
3101 
3102 #ifdef _KERNEL
3103         /*
3104          * We will be destroying the ds; make sure its origin is unmounted if
3105          * necessary.
3106          */
3107         dsl_dataset_name(drc->drc_ds, name);
3108         zfs_destroy_unmount_origin(name);
3109 #endif
3110 
3111         error = dsl_sync_task(drc->drc_tofs,
3112             dmu_recv_end_check, dmu_recv_end_sync, drc,
3113             dmu_recv_end_modified_blocks, ZFS_SPACE_CHECK_NORMAL);
3114 
3115         if (error != 0)
3116                 dmu_recv_cleanup_ds(drc);
3117         return (error);
3118 }
3119 
3120 static int
3121 dmu_recv_new_end(dmu_recv_cookie_t *drc)
3122 {
3123         int error;
3124 
3125         error = dsl_sync_task(drc->drc_tofs,
3126             dmu_recv_end_check, dmu_recv_end_sync, drc,
3127             dmu_recv_end_modified_blocks, ZFS_SPACE_CHECK_NORMAL);
3128 
3129         if (error != 0) {
3130                 dmu_recv_cleanup_ds(drc);
3131         } else if (drc->drc_guid_to_ds_map != NULL) {
3132                 (void) add_ds_to_guidmap(drc->drc_tofs,
3133                     drc->drc_guid_to_ds_map,
3134                     drc->drc_newsnapobj);
3135         }
3136         return (error);
3137 }
3138 
3139 int
3140 dmu_recv_end(dmu_recv_cookie_t *drc, void *owner)
3141 {
3142         drc->drc_owner = owner;
3143 
3144         if (drc->drc_newfs)
3145                 return (dmu_recv_new_end(drc));
3146         else
3147                 return (dmu_recv_existing_end(drc));
3148 }
3149 
3150 /*
3151  * Return TRUE if this objset is currently being received into.
3152  */
3153 boolean_t
3154 dmu_objset_is_receiving(objset_t *os)
3155 {
3156         return (os->os_dsl_dataset != NULL &&
3157             os->os_dsl_dataset->ds_owner == dmu_recv_tag);
3158 }