sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dmu_send.c

   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 (c) 2012, Martin Matuska <mm@FreeBSD.org>. All rights reserved.
  27  * Copyright 2014 HybridCluster. All rights reserved.
  28  * Copyright 2016 RackTop Systems.
  29  * Copyright (c) 2014 Integros [integros.com]
  30  * Copyright (c) 2018, loli10K <ezomori.nozomu@gmail.com>. All rights reserved.
  31  */
  32
  33 #include <sys/dmu.h>
  34 #include <sys/dmu_impl.h>
  35 #include <sys/dmu_tx.h>
  36 #include <sys/dbuf.h>
  37 #include <sys/dnode.h>
  38 #include <sys/zfs_context.h>
  39 #include <sys/dmu_objset.h>
  40 #include <sys/dmu_traverse.h>
  41 #include <sys/dsl_dataset.h>
  42 #include <sys/dsl_dir.h>
  43 #include <sys/dsl_prop.h>
  44 #include <sys/dsl_pool.h>
  45 #include <sys/dsl_synctask.h>
  46 #include <sys/zfs_ioctl.h>
  47 #include <sys/zap.h>
  48 #include <sys/zio_checksum.h>
  49 #include <sys/zfs_znode.h>
  50 #include <zfs_fletcher.h>
  51 #include <sys/avl.h>
  52 #include <sys/ddt.h>
  53 #include <sys/zfs_onexit.h>
  54 #include <sys/dmu_send.h>
  55 #include <sys/dsl_destroy.h>
  56 #include <sys/blkptr.h>
  57 #include <sys/dsl_bookmark.h>
  58 #include <sys/zfeature.h>
  59 #include <sys/bqueue.h>
  60 #ifdef __FreeBSD__
  61 #include <sys/zvol.h>
  62 #endif
  63
  64 #ifdef __FreeBSD__
  65 #undef dump_write
  66 #define dump_write dmu_dump_write
  67 #endif
  68
  69 /* Set this tunable to TRUE to replace corrupt data with 0x2f5baddb10c */
  70 int zfs_send_corrupt_data = B_FALSE;
  71 int zfs_send_queue_length = 16 * 1024 * 1024;
  72 int zfs_recv_queue_length = 16 * 1024 * 1024;
  73 /* Set this tunable to FALSE to disable setting of DRR_FLAG_FREERECORDS */
  74 int zfs_send_set_freerecords_bit = B_TRUE;
  75
  76 #ifdef _KERNEL
  77 TUNABLE_INT("vfs.zfs.send_set_freerecords_bit", &zfs_send_set_freerecords_bit);
  78 #endif
  79
  80 static char *dmu_recv_tag = "dmu_recv_tag";
  81 const char *recv_clone_name = "%recv";
  82
  83 /*
  84  * Use this to override the recordsize calculation for fast zfs send estimates.
  85  */
  86 uint64_t zfs_override_estimate_recordsize = 0;
  87
  88 #define BP_SPAN(datablkszsec, indblkshift, level) \
  89         (((uint64_t)datablkszsec) << (SPA_MINBLOCKSHIFT + \
  90         (level) * (indblkshift - SPA_BLKPTRSHIFT)))
  91
  92 static void byteswap_record(dmu_replay_record_t *drr);
  93
  94 struct send_thread_arg {
  95         bqueue_t        q;
  96         dsl_dataset_t   *ds;            /* Dataset to traverse */
  97         uint64_t        fromtxg;        /* Traverse from this txg */
  98         int             flags;          /* flags to pass to traverse_dataset */
  99         int             error_code;
 100         boolean_t       cancel;
 101         zbookmark_phys_t resume;
 102 };
 103
 104 struct send_block_record {
 105         boolean_t               eos_marker; /* Marks the end of the stream */
 106         blkptr_t                bp;
 107         zbookmark_phys_t        zb;
 108         uint8_t                 indblkshift;
 109         uint16_t                datablkszsec;
 110         bqueue_node_t           ln;
 111 };
 112
 113 static int
 114 dump_bytes(dmu_sendarg_t *dsp, void *buf, int len)
 115 {
 116         dsl_dataset_t *ds = dmu_objset_ds(dsp->dsa_os);
 117         struct uio auio;
 118         struct iovec aiov;
 119
 120         /*
 121          * The code does not rely on this (len being a multiple of 8).  We keep
 122          * this assertion because of the corresponding assertion in
 123          * receive_read().  Keeping this assertion ensures that we do not
 124          * inadvertently break backwards compatibility (causing the assertion
 125          * in receive_read() to trigger on old software).
 126          *
 127          * Removing the assertions could be rolled into a new feature that uses
 128          * data that isn't 8-byte aligned; if the assertions were removed, a
 129          * feature flag would have to be added.
 130          */
 131
 132         ASSERT0(len % 8);
 133
 134         aiov.iov_base = buf;
 135         aiov.iov_len = len;
 136         auio.uio_iov = &aiov;
 137         auio.uio_iovcnt = 1;
 138         auio.uio_resid = len;
 139         auio.uio_segflg = UIO_SYSSPACE;
 140         auio.uio_rw = UIO_WRITE;
 141         auio.uio_offset = (off_t)-1;
 142         auio.uio_td = dsp->dsa_td;
 143 #ifdef _KERNEL
 144         if (dsp->dsa_fp->f_type == DTYPE_VNODE)
 145                 bwillwrite();
 146         dsp->dsa_err = fo_write(dsp->dsa_fp, &auio, dsp->dsa_td->td_ucred, 0,
 147             dsp->dsa_td);
 148 #else
 149         fprintf(stderr, "%s: returning EOPNOTSUPP\n", __func__);
 150         dsp->dsa_err = EOPNOTSUPP;
 151 #endif
 152         mutex_enter(&ds->ds_sendstream_lock);
 153         *dsp->dsa_off += len;
 154         mutex_exit(&ds->ds_sendstream_lock);
 155
 156         return (dsp->dsa_err);
 157 }
 158
 159 /*
 160  * For all record types except BEGIN, fill in the checksum (overlaid in
 161  * drr_u.drr_checksum.drr_checksum).  The checksum verifies everything
 162  * up to the start of the checksum itself.
 163  */
 164 static int
 165 dump_record(dmu_sendarg_t *dsp, void *payload, int payload_len)
 166 {
 167         ASSERT3U(offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
 168             ==, sizeof (dmu_replay_record_t) - sizeof (zio_cksum_t));
 169         (void) fletcher_4_incremental_native(dsp->dsa_drr,
 170             offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
 171             &dsp->dsa_zc);
 172         if (dsp->dsa_drr->drr_type == DRR_BEGIN) {
 173                 dsp->dsa_sent_begin = B_TRUE;
 174         } else {
 175                 ASSERT(ZIO_CHECKSUM_IS_ZERO(&dsp->dsa_drr->drr_u.
 176                     drr_checksum.drr_checksum));
 177                 dsp->dsa_drr->drr_u.drr_checksum.drr_checksum = dsp->dsa_zc;
 178         }
 179         if (dsp->dsa_drr->drr_type == DRR_END) {
 180                 dsp->dsa_sent_end = B_TRUE;
 181         }
 182         (void) fletcher_4_incremental_native(&dsp->dsa_drr->
 183             drr_u.drr_checksum.drr_checksum,
 184             sizeof (zio_cksum_t), &dsp->dsa_zc);
 185         if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)) != 0)
 186                 return (SET_ERROR(EINTR));
 187         if (payload_len != 0) {
 188                 (void) fletcher_4_incremental_native(payload, payload_len,
 189                     &dsp->dsa_zc);
 190                 if (dump_bytes(dsp, payload, payload_len) != 0)
 191                         return (SET_ERROR(EINTR));
 192         }
 193         return (0);
 194 }
 195
 196 /*
 197  * Fill in the drr_free struct, or perform aggregation if the previous record is
 198  * also a free record, and the two are adjacent.
 199  *
 200  * Note that we send free records even for a full send, because we want to be
 201  * able to receive a full send as a clone, which requires a list of all the free
 202  * and freeobject records that were generated on the source.
 203  */
 204 static int
 205 dump_free(dmu_sendarg_t *dsp, uint64_t object, uint64_t offset,
 206     uint64_t length)
 207 {
 208         struct drr_free *drrf = &(dsp->dsa_drr->drr_u.drr_free);
 209
 210         /*
 211          * When we receive a free record, dbuf_free_range() assumes
 212          * that the receiving system doesn't have any dbufs in the range
 213          * being freed.  This is always true because there is a one-record
 214          * constraint: we only send one WRITE record for any given
 215          * object,offset.  We know that the one-record constraint is
 216          * true because we always send data in increasing order by
 217          * object,offset.
 218          *
 219          * If the increasing-order constraint ever changes, we should find
 220          * another way to assert that the one-record constraint is still
 221          * satisfied.
 222          */
 223         ASSERT(object > dsp->dsa_last_data_object ||
 224             (object == dsp->dsa_last_data_object &&
 225             offset > dsp->dsa_last_data_offset));
 226
 227         if (length != -1ULL && offset + length < offset)
 228                 length = -1ULL;
 229
 230         /*
 231          * If there is a pending op, but it's not PENDING_FREE, push it out,
 232          * since free block aggregation can only be done for blocks of the
 233          * same type (i.e., DRR_FREE records can only be aggregated with
 234          * other DRR_FREE records.  DRR_FREEOBJECTS records can only be
 235          * aggregated with other DRR_FREEOBJECTS records.
 236          */
 237         if (dsp->dsa_pending_op != PENDING_NONE &&
 238             dsp->dsa_pending_op != PENDING_FREE) {
 239                 if (dump_record(dsp, NULL, 0) != 0)
 240                         return (SET_ERROR(EINTR));
 241                 dsp->dsa_pending_op = PENDING_NONE;
 242         }
 243
 244         if (dsp->dsa_pending_op == PENDING_FREE) {
 245                 /*
 246                  * There should never be a PENDING_FREE if length is -1
 247                  * (because dump_dnode is the only place where this
 248                  * function is called with a -1, and only after flushing
 249                  * any pending record).
 250                  */
 251                 ASSERT(length != -1ULL);
 252                 /*
 253                  * Check to see whether this free block can be aggregated
 254                  * with pending one.
 255                  */
 256                 if (drrf->drr_object == object && drrf->drr_offset +
 257                     drrf->drr_length == offset) {
 258                         drrf->drr_length += length;
 259                         return (0);
 260                 } else {
 261                         /* not a continuation.  Push out pending record */
 262                         if (dump_record(dsp, NULL, 0) != 0)
 263                                 return (SET_ERROR(EINTR));
 264                         dsp->dsa_pending_op = PENDING_NONE;
 265                 }
 266         }
 267         /* create a FREE record and make it pending */
 268         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 269         dsp->dsa_drr->drr_type = DRR_FREE;
 270         drrf->drr_object = object;
 271         drrf->drr_offset = offset;
 272         drrf->drr_length = length;
 273         drrf->drr_toguid = dsp->dsa_toguid;
 274         if (length == -1ULL) {
 275                 if (dump_record(dsp, NULL, 0) != 0)
 276                         return (SET_ERROR(EINTR));
 277         } else {
 278                 dsp->dsa_pending_op = PENDING_FREE;
 279         }
 280
 281         return (0);
 282 }
 283
 284 static int
 285 dump_write(dmu_sendarg_t *dsp, dmu_object_type_t type,
 286     uint64_t object, uint64_t offset, int lsize, int psize, const blkptr_t *bp,
 287     void *data)
 288 {
 289         uint64_t payload_size;
 290         struct drr_write *drrw = &(dsp->dsa_drr->drr_u.drr_write);
 291
 292         /*
 293          * We send data in increasing object, offset order.
 294          * See comment in dump_free() for details.
 295          */
 296         ASSERT(object > dsp->dsa_last_data_object ||
 297             (object == dsp->dsa_last_data_object &&
 298             offset > dsp->dsa_last_data_offset));
 299         dsp->dsa_last_data_object = object;
 300         dsp->dsa_last_data_offset = offset + lsize - 1;
 301
 302         /*
 303          * If there is any kind of pending aggregation (currently either
 304          * a grouping of free objects or free blocks), push it out to
 305          * the stream, since aggregation can't be done across operations
 306          * of different types.
 307          */
 308         if (dsp->dsa_pending_op != PENDING_NONE) {
 309                 if (dump_record(dsp, NULL, 0) != 0)
 310                         return (SET_ERROR(EINTR));
 311                 dsp->dsa_pending_op = PENDING_NONE;
 312         }
 313         /* write a WRITE record */
 314         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 315         dsp->dsa_drr->drr_type = DRR_WRITE;
 316         drrw->drr_object = object;
 317         drrw->drr_type = type;
 318         drrw->drr_offset = offset;
 319         drrw->drr_toguid = dsp->dsa_toguid;
 320         drrw->drr_logical_size = lsize;
 321
 322         /* only set the compression fields if the buf is compressed */
 323         if (lsize != psize) {
 324                 ASSERT(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_COMPRESSED);
 325                 ASSERT(!BP_IS_EMBEDDED(bp));
 326                 ASSERT(!BP_SHOULD_BYTESWAP(bp));
 327                 ASSERT(!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)));
 328                 ASSERT3U(BP_GET_COMPRESS(bp), !=, ZIO_COMPRESS_OFF);
 329                 ASSERT3S(psize, >, 0);
 330                 ASSERT3S(lsize, >=, psize);
 331
 332                 drrw->drr_compressiontype = BP_GET_COMPRESS(bp);
 333                 drrw->drr_compressed_size = psize;
 334                 payload_size = drrw->drr_compressed_size;
 335         } else {
 336                 payload_size = drrw->drr_logical_size;
 337         }
 338
 339         if (bp == NULL || BP_IS_EMBEDDED(bp)) {
 340                 /*
 341                  * There's no pre-computed checksum for partial-block
 342                  * writes or embedded BP's, so (like
 343                  * fletcher4-checkummed blocks) userland will have to
 344                  * compute a dedup-capable checksum itself.
 345                  */
 346                 drrw->drr_checksumtype = ZIO_CHECKSUM_OFF;
 347         } else {
 348                 drrw->drr_checksumtype = BP_GET_CHECKSUM(bp);
 349                 if (zio_checksum_table[drrw->drr_checksumtype].ci_flags &
 350                     ZCHECKSUM_FLAG_DEDUP)
 351                         drrw->drr_checksumflags |= DRR_CHECKSUM_DEDUP;
 352                 DDK_SET_LSIZE(&drrw->drr_key, BP_GET_LSIZE(bp));
 353                 DDK_SET_PSIZE(&drrw->drr_key, BP_GET_PSIZE(bp));
 354                 DDK_SET_COMPRESS(&drrw->drr_key, BP_GET_COMPRESS(bp));
 355                 drrw->drr_key.ddk_cksum = bp->blk_cksum;
 356         }
 357
 358         if (dump_record(dsp, data, payload_size) != 0)
 359                 return (SET_ERROR(EINTR));
 360         return (0);
 361 }
 362
 363 static int
 364 dump_write_embedded(dmu_sendarg_t *dsp, uint64_t object, uint64_t offset,
 365     int blksz, const blkptr_t *bp)
 366 {
 367         char buf[BPE_PAYLOAD_SIZE];
 368         struct drr_write_embedded *drrw =
 369             &(dsp->dsa_drr->drr_u.drr_write_embedded);
 370
 371         if (dsp->dsa_pending_op != PENDING_NONE) {
 372                 if (dump_record(dsp, NULL, 0) != 0)
 373                         return (EINTR);
 374                 dsp->dsa_pending_op = PENDING_NONE;
 375         }
 376
 377         ASSERT(BP_IS_EMBEDDED(bp));
 378
 379         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 380         dsp->dsa_drr->drr_type = DRR_WRITE_EMBEDDED;
 381         drrw->drr_object = object;
 382         drrw->drr_offset = offset;
 383         drrw->drr_length = blksz;
 384         drrw->drr_toguid = dsp->dsa_toguid;
 385         drrw->drr_compression = BP_GET_COMPRESS(bp);
 386         drrw->drr_etype = BPE_GET_ETYPE(bp);
 387         drrw->drr_lsize = BPE_GET_LSIZE(bp);
 388         drrw->drr_psize = BPE_GET_PSIZE(bp);
 389
 390         decode_embedded_bp_compressed(bp, buf);
 391
 392         if (dump_record(dsp, buf, P2ROUNDUP(drrw->drr_psize, 8)) != 0)
 393                 return (EINTR);
 394         return (0);
 395 }
 396
 397 static int
 398 dump_spill(dmu_sendarg_t *dsp, uint64_t object, int blksz, void *data)
 399 {
 400         struct drr_spill *drrs = &(dsp->dsa_drr->drr_u.drr_spill);
 401
 402         if (dsp->dsa_pending_op != PENDING_NONE) {
 403                 if (dump_record(dsp, NULL, 0) != 0)
 404                         return (SET_ERROR(EINTR));
 405                 dsp->dsa_pending_op = PENDING_NONE;
 406         }
 407
 408         /* write a SPILL record */
 409         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 410         dsp->dsa_drr->drr_type = DRR_SPILL;
 411         drrs->drr_object = object;
 412         drrs->drr_length = blksz;
 413         drrs->drr_toguid = dsp->dsa_toguid;
 414
 415         if (dump_record(dsp, data, blksz) != 0)
 416                 return (SET_ERROR(EINTR));
 417         return (0);
 418 }
 419
 420 static int
 421 dump_freeobjects(dmu_sendarg_t *dsp, uint64_t firstobj, uint64_t numobjs)
 422 {
 423         struct drr_freeobjects *drrfo = &(dsp->dsa_drr->drr_u.drr_freeobjects);
 424
 425         /*
 426          * If there is a pending op, but it's not PENDING_FREEOBJECTS,
 427          * push it out, since free block aggregation can only be done for
 428          * blocks of the same type (i.e., DRR_FREE records can only be
 429          * aggregated with other DRR_FREE records.  DRR_FREEOBJECTS records
 430          * can only be aggregated with other DRR_FREEOBJECTS records.
 431          */
 432         if (dsp->dsa_pending_op != PENDING_NONE &&
 433             dsp->dsa_pending_op != PENDING_FREEOBJECTS) {
 434                 if (dump_record(dsp, NULL, 0) != 0)
 435                         return (SET_ERROR(EINTR));
 436                 dsp->dsa_pending_op = PENDING_NONE;
 437         }
 438         if (dsp->dsa_pending_op == PENDING_FREEOBJECTS) {
 439                 /*
 440                  * See whether this free object array can be aggregated
 441                  * with pending one
 442                  */
 443                 if (drrfo->drr_firstobj + drrfo->drr_numobjs == firstobj) {
 444                         drrfo->drr_numobjs += numobjs;
 445                         return (0);
 446                 } else {
 447                         /* can't be aggregated.  Push out pending record */
 448                         if (dump_record(dsp, NULL, 0) != 0)
 449                                 return (SET_ERROR(EINTR));
 450                         dsp->dsa_pending_op = PENDING_NONE;
 451                 }
 452         }
 453
 454         /* write a FREEOBJECTS record */
 455         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 456         dsp->dsa_drr->drr_type = DRR_FREEOBJECTS;
 457         drrfo->drr_firstobj = firstobj;
 458         drrfo->drr_numobjs = numobjs;
 459         drrfo->drr_toguid = dsp->dsa_toguid;
 460
 461         dsp->dsa_pending_op = PENDING_FREEOBJECTS;
 462
 463         return (0);
 464 }
 465
 466 static int
 467 dump_dnode(dmu_sendarg_t *dsp, uint64_t object, dnode_phys_t *dnp)
 468 {
 469         struct drr_object *drro = &(dsp->dsa_drr->drr_u.drr_object);
 470
 471         if (object < dsp->dsa_resume_object) {
 472                 /*
 473                  * Note: when resuming, we will visit all the dnodes in
 474                  * the block of dnodes that we are resuming from.  In
 475                  * this case it's unnecessary to send the dnodes prior to
 476                  * the one we are resuming from.  We should be at most one
 477                  * block's worth of dnodes behind the resume point.
 478                  */
 479                 ASSERT3U(dsp->dsa_resume_object - object, <,
 480                     1 << (DNODE_BLOCK_SHIFT - DNODE_SHIFT));
 481                 return (0);
 482         }
 483
 484         if (dnp == NULL || dnp->dn_type == DMU_OT_NONE)
 485                 return (dump_freeobjects(dsp, object, 1));
 486
 487         if (dsp->dsa_pending_op != PENDING_NONE) {
 488                 if (dump_record(dsp, NULL, 0) != 0)
 489                         return (SET_ERROR(EINTR));
 490                 dsp->dsa_pending_op = PENDING_NONE;
 491         }
 492
 493         /* write an OBJECT record */
 494         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 495         dsp->dsa_drr->drr_type = DRR_OBJECT;
 496         drro->drr_object = object;
 497         drro->drr_type = dnp->dn_type;
 498         drro->drr_bonustype = dnp->dn_bonustype;
 499         drro->drr_blksz = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT;
 500         drro->drr_bonuslen = dnp->dn_bonuslen;
 501         drro->drr_dn_slots = dnp->dn_extra_slots + 1;
 502         drro->drr_checksumtype = dnp->dn_checksum;
 503         drro->drr_compress = dnp->dn_compress;
 504         drro->drr_toguid = dsp->dsa_toguid;
 505
 506         if (!(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
 507             drro->drr_blksz > SPA_OLD_MAXBLOCKSIZE)
 508                 drro->drr_blksz = SPA_OLD_MAXBLOCKSIZE;
 509
 510         if (dump_record(dsp, DN_BONUS(dnp),
 511             P2ROUNDUP(dnp->dn_bonuslen, 8)) != 0) {
 512                 return (SET_ERROR(EINTR));
 513         }
 514
 515         /* Free anything past the end of the file. */
 516         if (dump_free(dsp, object, (dnp->dn_maxblkid + 1) *
 517             (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT), -1ULL) != 0)
 518                 return (SET_ERROR(EINTR));
 519         if (dsp->dsa_err != 0)
 520                 return (SET_ERROR(EINTR));
 521         return (0);
 522 }
 523
 524 static boolean_t
 525 backup_do_embed(dmu_sendarg_t *dsp, const blkptr_t *bp)
 526 {
 527         if (!BP_IS_EMBEDDED(bp))
 528                 return (B_FALSE);
 529
 530         /*
 531          * Compression function must be legacy, or explicitly enabled.
 532          */
 533         if ((BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_LEGACY_FUNCTIONS &&
 534             !(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_LZ4)))
 535                 return (B_FALSE);
 536
 537         /*
 538          * Embed type must be explicitly enabled.
 539          */
 540         switch (BPE_GET_ETYPE(bp)) {
 541         case BP_EMBEDDED_TYPE_DATA:
 542                 if (dsp->dsa_featureflags & DMU_BACKUP_FEATURE_EMBED_DATA)
 543                         return (B_TRUE);
 544                 break;
 545         default:
 546                 return (B_FALSE);
 547         }
 548         return (B_FALSE);
 549 }
 550
 551 /*
 552  * This is the callback function to traverse_dataset that acts as the worker
 553  * thread for dmu_send_impl.
 554  */
 555 /*ARGSUSED*/
 556 static int
 557 send_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
 558     const zbookmark_phys_t *zb, const struct dnode_phys *dnp, void *arg)
 559 {
 560         struct send_thread_arg *sta = arg;
 561         struct send_block_record *record;
 562         uint64_t record_size;
 563         int err = 0;
 564
 565         ASSERT(zb->zb_object == DMU_META_DNODE_OBJECT ||
 566             zb->zb_object >= sta->resume.zb_object);
 567
 568         if (sta->cancel)
 569                 return (SET_ERROR(EINTR));
 570
 571         if (bp == NULL) {
 572                 ASSERT3U(zb->zb_level, ==, ZB_DNODE_LEVEL);
 573                 return (0);
 574         } else if (zb->zb_level < 0) {
 575                 return (0);
 576         }
 577
 578         record = kmem_zalloc(sizeof (struct send_block_record), KM_SLEEP);
 579         record->eos_marker = B_FALSE;
 580         record->bp = *bp;
 581         record->zb = *zb;
 582         record->indblkshift = dnp->dn_indblkshift;
 583         record->datablkszsec = dnp->dn_datablkszsec;
 584         record_size = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT;
 585         bqueue_enqueue(&sta->q, record, record_size);
 586
 587         return (err);
 588 }
 589
 590 /*
 591  * This function kicks off the traverse_dataset.  It also handles setting the
 592  * error code of the thread in case something goes wrong, and pushes the End of
 593  * Stream record when the traverse_dataset call has finished.  If there is no
 594  * dataset to traverse, the thread immediately pushes End of Stream marker.
 595  */
 596 static void
 597 send_traverse_thread(void *arg)
 598 {
 599         struct send_thread_arg *st_arg = arg;
 600         int err;
 601         struct send_block_record *data;
 602
 603         if (st_arg->ds != NULL) {
 604                 err = traverse_dataset_resume(st_arg->ds,
 605                     st_arg->fromtxg, &st_arg->resume,
 606                     st_arg->flags, send_cb, st_arg);
 607
 608                 if (err != EINTR)
 609                         st_arg->error_code = err;
 610         }
 611         data = kmem_zalloc(sizeof (*data), KM_SLEEP);
 612         data->eos_marker = B_TRUE;
 613         bqueue_enqueue(&st_arg->q, data, 1);
 614         thread_exit();
 615 }
 616
 617 /*
 618  * This function actually handles figuring out what kind of record needs to be
 619  * dumped, reading the data (which has hopefully been prefetched), and calling
 620  * the appropriate helper function.
 621  */
 622 static int
 623 do_dump(dmu_sendarg_t *dsa, struct send_block_record *data)
 624 {
 625         dsl_dataset_t *ds = dmu_objset_ds(dsa->dsa_os);
 626         const blkptr_t *bp = &data->bp;
 627         const zbookmark_phys_t *zb = &data->zb;
 628         uint8_t indblkshift = data->indblkshift;
 629         uint16_t dblkszsec = data->datablkszsec;
 630         spa_t *spa = ds->ds_dir->dd_pool->dp_spa;
 631         dmu_object_type_t type = bp ? BP_GET_TYPE(bp) : DMU_OT_NONE;
 632         int err = 0;
 633
 634         ASSERT3U(zb->zb_level, >=, 0);
 635
 636         ASSERT(zb->zb_object == DMU_META_DNODE_OBJECT ||
 637             zb->zb_object >= dsa->dsa_resume_object);
 638
 639         if (zb->zb_object != DMU_META_DNODE_OBJECT &&
 640             DMU_OBJECT_IS_SPECIAL(zb->zb_object)) {
 641                 return (0);
 642         } else if (BP_IS_HOLE(bp) &&
 643             zb->zb_object == DMU_META_DNODE_OBJECT) {
 644                 uint64_t span = BP_SPAN(dblkszsec, indblkshift, zb->zb_level);
 645                 uint64_t dnobj = (zb->zb_blkid * span) >> DNODE_SHIFT;
 646                 err = dump_freeobjects(dsa, dnobj, span >> DNODE_SHIFT);
 647         } else if (BP_IS_HOLE(bp)) {
 648                 uint64_t span = BP_SPAN(dblkszsec, indblkshift, zb->zb_level);
 649                 uint64_t offset = zb->zb_blkid * span;
 650                 err = dump_free(dsa, zb->zb_object, offset, span);
 651         } else if (zb->zb_level > 0 || type == DMU_OT_OBJSET) {
 652                 return (0);
 653         } else if (type == DMU_OT_DNODE) {
 654                 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
 655                 arc_flags_t aflags = ARC_FLAG_WAIT;
 656                 arc_buf_t *abuf;
 657
 658                 ASSERT0(zb->zb_level);
 659
 660                 if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
 661                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
 662                     &aflags, zb) != 0)
 663                         return (SET_ERROR(EIO));
 664
 665                 dnode_phys_t *blk = abuf->b_data;
 666                 uint64_t dnobj = zb->zb_blkid * epb;
 667                 for (int i = 0; i < epb; i += blk[i].dn_extra_slots + 1) {
 668                         err = dump_dnode(dsa, dnobj + i, blk + i);
 669                         if (err != 0)
 670                                 break;
 671                 }
 672                 arc_buf_destroy(abuf, &abuf);
 673         } else if (type == DMU_OT_SA) {
 674                 arc_flags_t aflags = ARC_FLAG_WAIT;
 675                 arc_buf_t *abuf;
 676                 int blksz = BP_GET_LSIZE(bp);
 677
 678                 if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
 679                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
 680                     &aflags, zb) != 0)
 681                         return (SET_ERROR(EIO));
 682
 683                 err = dump_spill(dsa, zb->zb_object, blksz, abuf->b_data);
 684                 arc_buf_destroy(abuf, &abuf);
 685         } else if (backup_do_embed(dsa, bp)) {
 686                 /* it's an embedded level-0 block of a regular object */
 687                 int blksz = dblkszsec << SPA_MINBLOCKSHIFT;
 688                 ASSERT0(zb->zb_level);
 689                 err = dump_write_embedded(dsa, zb->zb_object,
 690                     zb->zb_blkid * blksz, blksz, bp);
 691         } else {
 692                 /* it's a level-0 block of a regular object */
 693                 arc_flags_t aflags = ARC_FLAG_WAIT;
 694                 arc_buf_t *abuf;
 695                 int blksz = dblkszsec << SPA_MINBLOCKSHIFT;
 696                 uint64_t offset;
 697
 698                 /*
 699                  * If we have large blocks stored on disk but the send flags
 700                  * don't allow us to send large blocks, we split the data from
 701                  * the arc buf into chunks.
 702                  */
 703                 boolean_t split_large_blocks = blksz > SPA_OLD_MAXBLOCKSIZE &&
 704                     !(dsa->dsa_featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS);
 705                 /*
 706                  * We should only request compressed data from the ARC if all
 707                  * the following are true:
 708                  *  - stream compression was requested
 709                  *  - we aren't splitting large blocks into smaller chunks
 710                  *  - the data won't need to be byteswapped before sending
 711                  *  - this isn't an embedded block
 712                  *  - this isn't metadata (if receiving on a different endian
 713                  *    system it can be byteswapped more easily)
 714                  */
 715                 boolean_t request_compressed =
 716                     (dsa->dsa_featureflags & DMU_BACKUP_FEATURE_COMPRESSED) &&
 717                     !split_large_blocks && !BP_SHOULD_BYTESWAP(bp) &&
 718                     !BP_IS_EMBEDDED(bp) && !DMU_OT_IS_METADATA(BP_GET_TYPE(bp));
 719
 720                 ASSERT0(zb->zb_level);
 721                 ASSERT(zb->zb_object > dsa->dsa_resume_object ||
 722                     (zb->zb_object == dsa->dsa_resume_object &&
 723                     zb->zb_blkid * blksz >= dsa->dsa_resume_offset));
 724
 725                 ASSERT0(zb->zb_level);
 726                 ASSERT(zb->zb_object > dsa->dsa_resume_object ||
 727                     (zb->zb_object == dsa->dsa_resume_object &&
 728                     zb->zb_blkid * blksz >= dsa->dsa_resume_offset));
 729
 730                 ASSERT3U(blksz, ==, BP_GET_LSIZE(bp));
 731
 732                 enum zio_flag zioflags = ZIO_FLAG_CANFAIL;
 733                 if (request_compressed)
 734                         zioflags |= ZIO_FLAG_RAW;
 735                 if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
 736                     ZIO_PRIORITY_ASYNC_READ, zioflags, &aflags, zb) != 0) {
 737                         if (zfs_send_corrupt_data) {
 738                                 /* Send a block filled with 0x"zfs badd bloc" */
 739                                 abuf = arc_alloc_buf(spa, &abuf, ARC_BUFC_DATA,
 740                                     blksz);
 741                                 uint64_t *ptr;
 742                                 for (ptr = abuf->b_data;
 743                                     (char *)ptr < (char *)abuf->b_data + blksz;
 744                                     ptr++)
 745                                         *ptr = 0x2f5baddb10cULL;
 746                         } else {
 747                                 return (SET_ERROR(EIO));
 748                         }
 749                 }
 750
 751                 offset = zb->zb_blkid * blksz;
 752
 753                 if (split_large_blocks) {
 754                         ASSERT3U(arc_get_compression(abuf), ==,
 755                             ZIO_COMPRESS_OFF);
 756                         char *buf = abuf->b_data;
 757                         while (blksz > 0 && err == 0) {
 758                                 int n = MIN(blksz, SPA_OLD_MAXBLOCKSIZE);
 759                                 err = dump_write(dsa, type, zb->zb_object,
 760                                     offset, n, n, NULL, buf);
 761                                 offset += n;
 762                                 buf += n;
 763                                 blksz -= n;
 764                         }
 765                 } else {
 766                         err = dump_write(dsa, type, zb->zb_object, offset,
 767                             blksz, arc_buf_size(abuf), bp, abuf->b_data);
 768                 }
 769                 arc_buf_destroy(abuf, &abuf);
 770         }
 771
 772         ASSERT(err == 0 || err == EINTR);
 773         return (err);
 774 }
 775
 776 /*
 777  * Pop the new data off the queue, and free the old data.
 778  */
 779 static struct send_block_record *
 780 get_next_record(bqueue_t *bq, struct send_block_record *data)
 781 {
 782         struct send_block_record *tmp = bqueue_dequeue(bq);
 783         kmem_free(data, sizeof (*data));
 784         return (tmp);
 785 }
 786
 787 /*
 788  * Actually do the bulk of the work in a zfs send.
 789  *
 790  * Note: Releases dp using the specified tag.
 791  */
 792 static int
 793 dmu_send_impl(void *tag, dsl_pool_t *dp, dsl_dataset_t *to_ds,
 794     zfs_bookmark_phys_t *ancestor_zb, boolean_t is_clone,
 795     boolean_t embedok, boolean_t large_block_ok, boolean_t compressok,
 796     int outfd, uint64_t resumeobj, uint64_t resumeoff,
 797 #ifdef illumos
 798     vnode_t *vp, offset_t *off)
 799 #else
 800     struct file *fp, offset_t *off)
 801 #endif
 802 {
 803         objset_t *os;
 804         dmu_replay_record_t *drr;
 805         dmu_sendarg_t *dsp;
 806         int err;
 807         uint64_t fromtxg = 0;
 808         uint64_t featureflags = 0;
 809         struct send_thread_arg to_arg = { 0 };
 810
 811         err = dmu_objset_from_ds(to_ds, &os);
 812         if (err != 0) {
 813                 dsl_pool_rele(dp, tag);
 814                 return (err);
 815         }
 816
 817         drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP);
 818         drr->drr_type = DRR_BEGIN;
 819         drr->drr_u.drr_begin.drr_magic = DMU_BACKUP_MAGIC;
 820         DMU_SET_STREAM_HDRTYPE(drr->drr_u.drr_begin.drr_versioninfo,
 821             DMU_SUBSTREAM);
 822
 823 #ifdef _KERNEL
 824         if (dmu_objset_type(os) == DMU_OST_ZFS) {
 825                 uint64_t version;
 826                 if (zfs_get_zplprop(os, ZFS_PROP_VERSION, &version) != 0) {
 827                         kmem_free(drr, sizeof (dmu_replay_record_t));
 828                         dsl_pool_rele(dp, tag);
 829                         return (SET_ERROR(EINVAL));
 830                 }
 831                 if (version >= ZPL_VERSION_SA) {
 832                         featureflags |= DMU_BACKUP_FEATURE_SA_SPILL;
 833                 }
 834         }
 835 #endif
 836
 837         if (large_block_ok && to_ds->ds_feature_inuse[SPA_FEATURE_LARGE_BLOCKS])
 838                 featureflags |= DMU_BACKUP_FEATURE_LARGE_BLOCKS;
 839         if (to_ds->ds_feature_inuse[SPA_FEATURE_LARGE_DNODE])
 840                 featureflags |= DMU_BACKUP_FEATURE_LARGE_DNODE;
 841         if (embedok &&
 842             spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA)) {
 843                 featureflags |= DMU_BACKUP_FEATURE_EMBED_DATA;
 844                 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS))
 845                         featureflags |= DMU_BACKUP_FEATURE_LZ4;
 846         }
 847         if (compressok) {
 848                 featureflags |= DMU_BACKUP_FEATURE_COMPRESSED;
 849         }
 850         if ((featureflags &
 851             (DMU_BACKUP_FEATURE_EMBED_DATA | DMU_BACKUP_FEATURE_COMPRESSED)) !=
 852             0 && spa_feature_is_active(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS)) {
 853                 featureflags |= DMU_BACKUP_FEATURE_LZ4;
 854         }
 855
 856         if (resumeobj != 0 || resumeoff != 0) {
 857                 featureflags |= DMU_BACKUP_FEATURE_RESUMING;
 858         }
 859
 860         DMU_SET_FEATUREFLAGS(drr->drr_u.drr_begin.drr_versioninfo,
 861             featureflags);
 862
 863         drr->drr_u.drr_begin.drr_creation_time =
 864             dsl_dataset_phys(to_ds)->ds_creation_time;
 865         drr->drr_u.drr_begin.drr_type = dmu_objset_type(os);
 866         if (is_clone)
 867                 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CLONE;
 868         drr->drr_u.drr_begin.drr_toguid = dsl_dataset_phys(to_ds)->ds_guid;
 869         if (dsl_dataset_phys(to_ds)->ds_flags & DS_FLAG_CI_DATASET)
 870                 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CI_DATA;
 871         if (zfs_send_set_freerecords_bit)
 872                 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_FREERECORDS;
 873
 874         if (ancestor_zb != NULL) {
 875                 drr->drr_u.drr_begin.drr_fromguid =
 876                     ancestor_zb->zbm_guid;
 877                 fromtxg = ancestor_zb->zbm_creation_txg;
 878         }
 879         dsl_dataset_name(to_ds, drr->drr_u.drr_begin.drr_toname);
 880         if (!to_ds->ds_is_snapshot) {
 881                 (void) strlcat(drr->drr_u.drr_begin.drr_toname, "@--head--",
 882                     sizeof (drr->drr_u.drr_begin.drr_toname));
 883         }
 884
 885         dsp = kmem_zalloc(sizeof (dmu_sendarg_t), KM_SLEEP);
 886
 887         dsp->dsa_drr = drr;
 888         dsp->dsa_outfd = outfd;
 889         dsp->dsa_proc = curproc;
 890         dsp->dsa_td = curthread;
 891         dsp->dsa_fp = fp;
 892         dsp->dsa_os = os;
 893         dsp->dsa_off = off;
 894         dsp->dsa_toguid = dsl_dataset_phys(to_ds)->ds_guid;
 895         dsp->dsa_pending_op = PENDING_NONE;
 896         dsp->dsa_featureflags = featureflags;
 897         dsp->dsa_resume_object = resumeobj;
 898         dsp->dsa_resume_offset = resumeoff;
 899
 900         mutex_enter(&to_ds->ds_sendstream_lock);
 901         list_insert_head(&to_ds->ds_sendstreams, dsp);
 902         mutex_exit(&to_ds->ds_sendstream_lock);
 903
 904         dsl_dataset_long_hold(to_ds, FTAG);
 905         dsl_pool_rele(dp, tag);
 906
 907         void *payload = NULL;
 908         size_t payload_len = 0;
 909         if (resumeobj != 0 || resumeoff != 0) {
 910                 dmu_object_info_t to_doi;
 911                 err = dmu_object_info(os, resumeobj, &to_doi);
 912                 if (err != 0)
 913                         goto out;
 914                 SET_BOOKMARK(&to_arg.resume, to_ds->ds_object, resumeobj, 0,
 915                     resumeoff / to_doi.doi_data_block_size);
 916
 917                 nvlist_t *nvl = fnvlist_alloc();
 918                 fnvlist_add_uint64(nvl, "resume_object", resumeobj);
 919                 fnvlist_add_uint64(nvl, "resume_offset", resumeoff);
 920                 payload = fnvlist_pack(nvl, &payload_len);
 921                 drr->drr_payloadlen = payload_len;
 922                 fnvlist_free(nvl);
 923         }
 924
 925         err = dump_record(dsp, payload, payload_len);
 926         fnvlist_pack_free(payload, payload_len);
 927         if (err != 0) {
 928                 err = dsp->dsa_err;
 929                 goto out;
 930         }
 931
 932         err = bqueue_init(&to_arg.q, zfs_send_queue_length,
 933             offsetof(struct send_block_record, ln));
 934         to_arg.error_code = 0;
 935         to_arg.cancel = B_FALSE;
 936         to_arg.ds = to_ds;
 937         to_arg.fromtxg = fromtxg;
 938         to_arg.flags = TRAVERSE_PRE | TRAVERSE_PREFETCH;
 939         (void) thread_create(NULL, 0, send_traverse_thread, &to_arg, 0, &p0,
 940             TS_RUN, minclsyspri);
 941
 942         struct send_block_record *to_data;
 943         to_data = bqueue_dequeue(&to_arg.q);
 944
 945         while (!to_data->eos_marker && err == 0) {
 946                 err = do_dump(dsp, to_data);
 947                 to_data = get_next_record(&to_arg.q, to_data);
 948                 if (issig(JUSTLOOKING) && issig(FORREAL))
 949                         err = EINTR;
 950         }
 951
 952         if (err != 0) {
 953                 to_arg.cancel = B_TRUE;
 954                 while (!to_data->eos_marker) {
 955                         to_data = get_next_record(&to_arg.q, to_data);
 956                 }
 957         }
 958         kmem_free(to_data, sizeof (*to_data));
 959
 960         bqueue_destroy(&to_arg.q);
 961
 962         if (err == 0 && to_arg.error_code != 0)
 963                 err = to_arg.error_code;
 964
 965         if (err != 0)
 966                 goto out;
 967
 968         if (dsp->dsa_pending_op != PENDING_NONE)
 969                 if (dump_record(dsp, NULL, 0) != 0)
 970                         err = SET_ERROR(EINTR);
 971
 972         if (err != 0) {
 973                 if (err == EINTR && dsp->dsa_err != 0)
 974                         err = dsp->dsa_err;
 975                 goto out;
 976         }
 977
 978         bzero(drr, sizeof (dmu_replay_record_t));
 979         drr->drr_type = DRR_END;
 980         drr->drr_u.drr_end.drr_checksum = dsp->dsa_zc;
 981         drr->drr_u.drr_end.drr_toguid = dsp->dsa_toguid;
 982
 983         if (dump_record(dsp, NULL, 0) != 0)
 984                 err = dsp->dsa_err;
 985
 986 out:
 987         mutex_enter(&to_ds->ds_sendstream_lock);
 988         list_remove(&to_ds->ds_sendstreams, dsp);
 989         mutex_exit(&to_ds->ds_sendstream_lock);
 990
 991         VERIFY(err != 0 || (dsp->dsa_sent_begin && dsp->dsa_sent_end));
 992
 993         kmem_free(drr, sizeof (dmu_replay_record_t));
 994         kmem_free(dsp, sizeof (dmu_sendarg_t));
 995
 996         dsl_dataset_long_rele(to_ds, FTAG);
 997
 998         return (err);
 999 }
1000
1001 int
1002 dmu_send_obj(const char *pool, uint64_t tosnap, uint64_t fromsnap,
1003     boolean_t embedok, boolean_t large_block_ok, boolean_t compressok,
1004 #ifdef illumos
1005     int outfd, vnode_t *vp, offset_t *off)
1006 #else
1007     int outfd, struct file *fp, offset_t *off)
1008 #endif
1009 {
1010         dsl_pool_t *dp;
1011         dsl_dataset_t *ds;
1012         dsl_dataset_t *fromds = NULL;
1013         int err;
1014
1015         err = dsl_pool_hold(pool, FTAG, &dp);
1016         if (err != 0)
1017                 return (err);
1018
1019         err = dsl_dataset_hold_obj(dp, tosnap, FTAG, &ds);
1020         if (err != 0) {
1021                 dsl_pool_rele(dp, FTAG);
1022                 return (err);
1023         }
1024
1025         if (fromsnap != 0) {
1026                 zfs_bookmark_phys_t zb;
1027                 boolean_t is_clone;
1028
1029                 err = dsl_dataset_hold_obj(dp, fromsnap, FTAG, &fromds);
1030                 if (err != 0) {
1031                         dsl_dataset_rele(ds, FTAG);
1032                         dsl_pool_rele(dp, FTAG);
1033                         return (err);
1034                 }
1035                 if (!dsl_dataset_is_before(ds, fromds, 0))
1036                         err = SET_ERROR(EXDEV);
1037                 zb.zbm_creation_time =
1038                     dsl_dataset_phys(fromds)->ds_creation_time;
1039                 zb.zbm_creation_txg = dsl_dataset_phys(fromds)->ds_creation_txg;
1040                 zb.zbm_guid = dsl_dataset_phys(fromds)->ds_guid;
1041                 is_clone = (fromds->ds_dir != ds->ds_dir);
1042                 dsl_dataset_rele(fromds, FTAG);
1043                 err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone,
1044                     embedok, large_block_ok, compressok, outfd, 0, 0, fp, off);
1045         } else {
1046                 err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE,
1047                     embedok, large_block_ok, compressok, outfd, 0, 0, fp, off);
1048         }
1049         dsl_dataset_rele(ds, FTAG);
1050         return (err);
1051 }
1052
1053 int
1054 dmu_send(const char *tosnap, const char *fromsnap, boolean_t embedok,
1055     boolean_t large_block_ok, boolean_t compressok, int outfd,
1056     uint64_t resumeobj, uint64_t resumeoff,
1057 #ifdef illumos
1058     vnode_t *vp, offset_t *off)
1059 #else
1060     struct file *fp, offset_t *off)
1061 #endif
1062 {
1063         dsl_pool_t *dp;
1064         dsl_dataset_t *ds;
1065         int err;
1066         boolean_t owned = B_FALSE;
1067
1068         if (fromsnap != NULL && strpbrk(fromsnap, "@#") == NULL)
1069                 return (SET_ERROR(EINVAL));
1070
1071         err = dsl_pool_hold(tosnap, FTAG, &dp);
1072         if (err != 0)
1073                 return (err);
1074
1075         if (strchr(tosnap, '@') == NULL && spa_writeable(dp->dp_spa)) {
1076                 /*
1077                  * We are sending a filesystem or volume.  Ensure
1078                  * that it doesn't change by owning the dataset.
1079                  */
1080                 err = dsl_dataset_own(dp, tosnap, FTAG, &ds);
1081                 owned = B_TRUE;
1082         } else {
1083                 err = dsl_dataset_hold(dp, tosnap, FTAG, &ds);
1084         }
1085         if (err != 0) {
1086                 dsl_pool_rele(dp, FTAG);
1087                 return (err);
1088         }
1089
1090         if (fromsnap != NULL) {
1091                 zfs_bookmark_phys_t zb;
1092                 boolean_t is_clone = B_FALSE;
1093                 int fsnamelen = strchr(tosnap, '@') - tosnap;
1094
1095                 /*
1096                  * If the fromsnap is in a different filesystem, then
1097                  * mark the send stream as a clone.
1098                  */
1099                 if (strncmp(tosnap, fromsnap, fsnamelen) != 0 ||
1100                     (fromsnap[fsnamelen] != '@' &&
1101                     fromsnap[fsnamelen] != '#')) {
1102                         is_clone = B_TRUE;
1103                 }
1104
1105                 if (strchr(fromsnap, '@')) {
1106                         dsl_dataset_t *fromds;
1107                         err = dsl_dataset_hold(dp, fromsnap, FTAG, &fromds);
1108                         if (err == 0) {
1109                                 if (!dsl_dataset_is_before(ds, fromds, 0))
1110                                         err = SET_ERROR(EXDEV);
1111                                 zb.zbm_creation_time =
1112                                     dsl_dataset_phys(fromds)->ds_creation_time;
1113                                 zb.zbm_creation_txg =
1114                                     dsl_dataset_phys(fromds)->ds_creation_txg;
1115                                 zb.zbm_guid = dsl_dataset_phys(fromds)->ds_guid;
1116                                 is_clone = (ds->ds_dir != fromds->ds_dir);
1117                                 dsl_dataset_rele(fromds, FTAG);
1118                         }
1119                 } else {
1120                         err = dsl_bookmark_lookup(dp, fromsnap, ds, &zb);
1121                 }
1122                 if (err != 0) {
1123                         dsl_dataset_rele(ds, FTAG);
1124                         dsl_pool_rele(dp, FTAG);
1125                         return (err);
1126                 }
1127                 err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone,
1128                     embedok, large_block_ok, compressok,
1129                     outfd, resumeobj, resumeoff, fp, off);
1130         } else {
1131                 err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE,
1132                     embedok, large_block_ok, compressok,
1133                     outfd, resumeobj, resumeoff, fp, off);
1134         }
1135         if (owned)
1136                 dsl_dataset_disown(ds, FTAG);
1137         else
1138                 dsl_dataset_rele(ds, FTAG);
1139         return (err);
1140 }
1141
1142 static int
1143 dmu_adjust_send_estimate_for_indirects(dsl_dataset_t *ds, uint64_t uncompressed,
1144     uint64_t compressed, boolean_t stream_compressed, uint64_t *sizep)
1145 {
1146         int err = 0;
1147         uint64_t size;
1148         /*
1149          * Assume that space (both on-disk and in-stream) is dominated by
1150          * data.  We will adjust for indirect blocks and the copies property,
1151          * but ignore per-object space used (eg, dnodes and DRR_OBJECT records).
1152          */
1153         uint64_t recordsize;
1154         uint64_t record_count;
1155         objset_t *os;
1156         VERIFY0(dmu_objset_from_ds(ds, &os));
1157
1158         /* Assume all (uncompressed) blocks are recordsize. */
1159         if (zfs_override_estimate_recordsize != 0) {
1160                 recordsize = zfs_override_estimate_recordsize;
1161         } else if (os->os_phys->os_type == DMU_OST_ZVOL) {
1162                 err = dsl_prop_get_int_ds(ds,
1163                     zfs_prop_to_name(ZFS_PROP_VOLBLOCKSIZE), &recordsize);
1164         } else {
1165                 err = dsl_prop_get_int_ds(ds,
1166                     zfs_prop_to_name(ZFS_PROP_RECORDSIZE), &recordsize);
1167         }
1168         if (err != 0)
1169                 return (err);
1170         record_count = uncompressed / recordsize;
1171
1172         /*
1173          * If we're estimating a send size for a compressed stream, use the
1174          * compressed data size to estimate the stream size. Otherwise, use the
1175          * uncompressed data size.
1176          */
1177         size = stream_compressed ? compressed : uncompressed;
1178
1179         /*
1180          * Subtract out approximate space used by indirect blocks.
1181          * Assume most space is used by data blocks (non-indirect, non-dnode).
1182          * Assume no ditto blocks or internal fragmentation.
1183          *
1184          * Therefore, space used by indirect blocks is sizeof(blkptr_t) per
1185          * block.
1186          */
1187         size -= record_count * sizeof (blkptr_t);
1188
1189         /* Add in the space for the record associated with each block. */
1190         size += record_count * sizeof (dmu_replay_record_t);
1191
1192         *sizep = size;
1193
1194         return (0);
1195 }
1196
1197 int
1198 dmu_send_estimate(dsl_dataset_t *ds, dsl_dataset_t *fromds,
1199     boolean_t stream_compressed, uint64_t *sizep)
1200 {
1201         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1202         int err;
1203         uint64_t uncomp, comp;
1204
1205         ASSERT(dsl_pool_config_held(dp));
1206
1207         /* tosnap must be a snapshot */
1208         if (!ds->ds_is_snapshot)
1209                 return (SET_ERROR(EINVAL));
1210
1211         /* fromsnap, if provided, must be a snapshot */
1212         if (fromds != NULL && !fromds->ds_is_snapshot)
1213                 return (SET_ERROR(EINVAL));
1214
1215         /*
1216          * fromsnap must be an earlier snapshot from the same fs as tosnap,
1217          * or the origin's fs.
1218          */
1219         if (fromds != NULL && !dsl_dataset_is_before(ds, fromds, 0))
1220                 return (SET_ERROR(EXDEV));
1221
1222         /* Get compressed and uncompressed size estimates of changed data. */
1223         if (fromds == NULL) {
1224                 uncomp = dsl_dataset_phys(ds)->ds_uncompressed_bytes;
1225                 comp = dsl_dataset_phys(ds)->ds_compressed_bytes;
1226         } else {
1227                 uint64_t used;
1228                 err = dsl_dataset_space_written(fromds, ds,
1229                     &used, &comp, &uncomp);
1230                 if (err != 0)
1231                         return (err);
1232         }
1233
1234         err = dmu_adjust_send_estimate_for_indirects(ds, uncomp, comp,
1235             stream_compressed, sizep);
1236         /*
1237          * Add the size of the BEGIN and END records to the estimate.
1238          */
1239         *sizep += 2 * sizeof (dmu_replay_record_t);
1240         return (err);
1241 }
1242
1243 struct calculate_send_arg {
1244         uint64_t uncompressed;
1245         uint64_t compressed;
1246 };
1247
1248 /*
1249  * Simple callback used to traverse the blocks of a snapshot and sum their
1250  * uncompressed and compressed sizes.
1251  */
1252 /* ARGSUSED */
1253 static int
1254 dmu_calculate_send_traversal(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
1255     const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
1256 {
1257         struct calculate_send_arg *space = arg;
1258         if (bp != NULL && !BP_IS_HOLE(bp)) {
1259                 space->uncompressed += BP_GET_UCSIZE(bp);
1260                 space->compressed += BP_GET_PSIZE(bp);
1261         }
1262         return (0);
1263 }
1264
1265 /*
1266  * Given a desination snapshot and a TXG, calculate the approximate size of a
1267  * send stream sent from that TXG. from_txg may be zero, indicating that the
1268  * whole snapshot will be sent.
1269  */
1270 int
1271 dmu_send_estimate_from_txg(dsl_dataset_t *ds, uint64_t from_txg,
1272     boolean_t stream_compressed, uint64_t *sizep)
1273 {
1274         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1275         int err;
1276         struct calculate_send_arg size = { 0 };
1277
1278         ASSERT(dsl_pool_config_held(dp));
1279
1280         /* tosnap must be a snapshot */
1281         if (!ds->ds_is_snapshot)
1282                 return (SET_ERROR(EINVAL));
1283
1284         /* verify that from_txg is before the provided snapshot was taken */
1285         if (from_txg >= dsl_dataset_phys(ds)->ds_creation_txg) {
1286                 return (SET_ERROR(EXDEV));
1287         }
1288
1289         /*
1290          * traverse the blocks of the snapshot with birth times after
1291          * from_txg, summing their uncompressed size
1292          */
1293         err = traverse_dataset(ds, from_txg, TRAVERSE_POST,
1294             dmu_calculate_send_traversal, &size);
1295         if (err)
1296                 return (err);
1297
1298         err = dmu_adjust_send_estimate_for_indirects(ds, size.uncompressed,
1299             size.compressed, stream_compressed, sizep);
1300         return (err);
1301 }
1302
1303 typedef struct dmu_recv_begin_arg {
1304         const char *drba_origin;
1305         dmu_recv_cookie_t *drba_cookie;
1306         cred_t *drba_cred;
1307         uint64_t drba_snapobj;
1308 } dmu_recv_begin_arg_t;
1309
1310 static int
1311 recv_begin_check_existing_impl(dmu_recv_begin_arg_t *drba, dsl_dataset_t *ds,
1312     uint64_t fromguid)
1313 {
1314         uint64_t val;
1315         uint64_t children;
1316         int error;
1317         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1318
1319         /* Temporary clone name must not exist. */
1320         error = zap_lookup(dp->dp_meta_objset,
1321             dsl_dir_phys(ds->ds_dir)->dd_child_dir_zapobj, recv_clone_name,
1322             8, 1, &val);
1323         if (error != ENOENT)
1324                 return (error == 0 ? SET_ERROR(EBUSY) : error);
1325
1326         /* Resume state must not be set. */
1327         if (dsl_dataset_has_resume_receive_state(ds))
1328                 return (SET_ERROR(EBUSY));
1329
1330         /* New snapshot name must not exist. */
1331         error = zap_lookup(dp->dp_meta_objset,
1332             dsl_dataset_phys(ds)->ds_snapnames_zapobj,
1333             drba->drba_cookie->drc_tosnap, 8, 1, &val);
1334         if (error != ENOENT)
1335                 return (error == 0 ? SET_ERROR(EEXIST) : error);
1336
1337         /* must not have children if receiving a ZVOL */
1338         error = zap_count(dp->dp_meta_objset,
1339             dsl_dir_phys(ds->ds_dir)->dd_child_dir_zapobj, &children);
1340         if (error != 0)
1341                 return (error);
1342         if (drba->drba_cookie->drc_drrb->drr_type != DMU_OST_ZFS &&
1343             children > 0)
1344                 return (SET_ERROR(ZFS_ERR_WRONG_PARENT));
1345
1346         /*
1347          * Check snapshot limit before receiving. We'll recheck again at the
1348          * end, but might as well abort before receiving if we're already over
1349          * the limit.
1350          *
1351          * Note that we do not check the file system limit with
1352          * dsl_dir_fscount_check because the temporary %clones don't count
1353          * against that limit.
1354          */
1355         error = dsl_fs_ss_limit_check(ds->ds_dir, 1, ZFS_PROP_SNAPSHOT_LIMIT,
1356             NULL, drba->drba_cred);
1357         if (error != 0)
1358                 return (error);
1359
1360         if (fromguid != 0) {
1361                 dsl_dataset_t *snap;
1362                 uint64_t obj = dsl_dataset_phys(ds)->ds_prev_snap_obj;
1363
1364                 /* Find snapshot in this dir that matches fromguid. */
1365                 while (obj != 0) {
1366                         error = dsl_dataset_hold_obj(dp, obj, FTAG,
1367                             &snap);
1368                         if (error != 0)
1369                                 return (SET_ERROR(ENODEV));
1370                         if (snap->ds_dir != ds->ds_dir) {
1371                                 dsl_dataset_rele(snap, FTAG);
1372                                 return (SET_ERROR(ENODEV));
1373                         }
1374                         if (dsl_dataset_phys(snap)->ds_guid == fromguid)
1375                                 break;
1376                         obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
1377                         dsl_dataset_rele(snap, FTAG);
1378                 }
1379                 if (obj == 0)
1380                         return (SET_ERROR(ENODEV));
1381
1382                 if (drba->drba_cookie->drc_force) {
1383                         drba->drba_snapobj = obj;
1384                 } else {
1385                         /*
1386                          * If we are not forcing, there must be no
1387                          * changes since fromsnap.
1388                          */
1389                         if (dsl_dataset_modified_since_snap(ds, snap)) {
1390                                 dsl_dataset_rele(snap, FTAG);
1391                                 return (SET_ERROR(ETXTBSY));
1392                         }
1393                         drba->drba_snapobj = ds->ds_prev->ds_object;
1394                 }
1395
1396                 dsl_dataset_rele(snap, FTAG);
1397         } else {
1398                 /* if full, then must be forced */
1399                 if (!drba->drba_cookie->drc_force)
1400                         return (SET_ERROR(EEXIST));
1401                 /* start from $ORIGIN@$ORIGIN, if supported */
1402                 drba->drba_snapobj = dp->dp_origin_snap != NULL ?
1403                     dp->dp_origin_snap->ds_object : 0;
1404         }
1405
1406         return (0);
1407
1408 }
1409
1410 static int
1411 dmu_recv_begin_check(void *arg, dmu_tx_t *tx)
1412 {
1413         dmu_recv_begin_arg_t *drba = arg;
1414         dsl_pool_t *dp = dmu_tx_pool(tx);
1415         struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
1416         uint64_t fromguid = drrb->drr_fromguid;
1417         int flags = drrb->drr_flags;
1418         int error;
1419         uint64_t featureflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo);
1420         dsl_dataset_t *ds;
1421         const char *tofs = drba->drba_cookie->drc_tofs;
1422
1423         /* already checked */
1424         ASSERT3U(drrb->drr_magic, ==, DMU_BACKUP_MAGIC);
1425         ASSERT(!(featureflags & DMU_BACKUP_FEATURE_RESUMING));
1426
1427         if (DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) ==
1428             DMU_COMPOUNDSTREAM ||
1429             drrb->drr_type >= DMU_OST_NUMTYPES ||
1430             ((flags & DRR_FLAG_CLONE) && drba->drba_origin == NULL))
1431                 return (SET_ERROR(EINVAL));
1432
1433         /* Verify pool version supports SA if SA_SPILL feature set */
1434         if ((featureflags & DMU_BACKUP_FEATURE_SA_SPILL) &&
1435             spa_version(dp->dp_spa) < SPA_VERSION_SA)
1436                 return (SET_ERROR(ENOTSUP));
1437
1438         if (drba->drba_cookie->drc_resumable &&
1439             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EXTENSIBLE_DATASET))
1440                 return (SET_ERROR(ENOTSUP));
1441
1442         /*
1443          * The receiving code doesn't know how to translate a WRITE_EMBEDDED
1444          * record to a plain WRITE record, so the pool must have the
1445          * EMBEDDED_DATA feature enabled if the stream has WRITE_EMBEDDED
1446          * records.  Same with WRITE_EMBEDDED records that use LZ4 compression.
1447          */
1448         if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA) &&
1449             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA))
1450                 return (SET_ERROR(ENOTSUP));
1451         if ((featureflags & DMU_BACKUP_FEATURE_LZ4) &&
1452             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS))
1453                 return (SET_ERROR(ENOTSUP));
1454
1455         /*
1456          * The receiving code doesn't know how to translate large blocks
1457          * to smaller ones, so the pool must have the LARGE_BLOCKS
1458          * feature enabled if the stream has LARGE_BLOCKS. Same with
1459          * large dnodes.
1460          */
1461         if ((featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
1462             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LARGE_BLOCKS))
1463                 return (SET_ERROR(ENOTSUP));
1464         if ((featureflags & DMU_BACKUP_FEATURE_LARGE_DNODE) &&
1465             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LARGE_DNODE))
1466                 return (SET_ERROR(ENOTSUP));
1467
1468         error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
1469         if (error == 0) {
1470                 /* target fs already exists; recv into temp clone */
1471
1472                 /* Can't recv a clone into an existing fs */
1473                 if (flags & DRR_FLAG_CLONE || drba->drba_origin) {
1474                         dsl_dataset_rele(ds, FTAG);
1475                         return (SET_ERROR(EINVAL));
1476                 }
1477
1478                 error = recv_begin_check_existing_impl(drba, ds, fromguid);
1479                 dsl_dataset_rele(ds, FTAG);
1480         } else if (error == ENOENT) {
1481                 /* target fs does not exist; must be a full backup or clone */
1482                 char buf[ZFS_MAX_DATASET_NAME_LEN];
1483                 objset_t *os;
1484
1485                 /*
1486                  * If it's a non-clone incremental, we are missing the
1487                  * target fs, so fail the recv.
1488                  */
1489                 if (fromguid != 0 && !(flags & DRR_FLAG_CLONE ||
1490                     drba->drba_origin))
1491                         return (SET_ERROR(ENOENT));
1492
1493                 /*
1494                  * If we're receiving a full send as a clone, and it doesn't
1495                  * contain all the necessary free records and freeobject
1496                  * records, reject it.
1497                  */
1498                 if (fromguid == 0 && drba->drba_origin &&
1499                     !(flags & DRR_FLAG_FREERECORDS))
1500                         return (SET_ERROR(EINVAL));
1501
1502                 /* Open the parent of tofs */
1503                 ASSERT3U(strlen(tofs), <, sizeof (buf));
1504                 (void) strlcpy(buf, tofs, strrchr(tofs, '/') - tofs + 1);
1505                 error = dsl_dataset_hold(dp, buf, FTAG, &ds);
1506                 if (error != 0)
1507                         return (error);
1508
1509                 /*
1510                  * Check filesystem and snapshot limits before receiving. We'll
1511                  * recheck snapshot limits again at the end (we create the
1512                  * filesystems and increment those counts during begin_sync).
1513                  */
1514                 error = dsl_fs_ss_limit_check(ds->ds_dir, 1,
1515                     ZFS_PROP_FILESYSTEM_LIMIT, NULL, drba->drba_cred);
1516                 if (error != 0) {
1517                         dsl_dataset_rele(ds, FTAG);
1518                         return (error);
1519                 }
1520
1521                 error = dsl_fs_ss_limit_check(ds->ds_dir, 1,
1522                     ZFS_PROP_SNAPSHOT_LIMIT, NULL, drba->drba_cred);
1523                 if (error != 0) {
1524                         dsl_dataset_rele(ds, FTAG);
1525                         return (error);
1526                 }
1527
1528                 /* can't recv below anything but filesystems (eg. no ZVOLs) */
1529                 error = dmu_objset_from_ds(ds, &os);
1530                 if (error != 0) {
1531                         dsl_dataset_rele(ds, FTAG);
1532                         return (error);
1533                 }
1534                 if (dmu_objset_type(os) != DMU_OST_ZFS) {
1535                         dsl_dataset_rele(ds, FTAG);
1536                         return (SET_ERROR(ZFS_ERR_WRONG_PARENT));
1537                 }
1538
1539                 if (drba->drba_origin != NULL) {
1540                         dsl_dataset_t *origin;
1541                         error = dsl_dataset_hold(dp, drba->drba_origin,
1542                             FTAG, &origin);
1543                         if (error != 0) {
1544                                 dsl_dataset_rele(ds, FTAG);
1545                                 return (error);
1546                         }
1547                         if (!origin->ds_is_snapshot) {
1548                                 dsl_dataset_rele(origin, FTAG);
1549                                 dsl_dataset_rele(ds, FTAG);
1550                                 return (SET_ERROR(EINVAL));
1551                         }
1552                         if (dsl_dataset_phys(origin)->ds_guid != fromguid &&
1553                             fromguid != 0) {
1554                                 dsl_dataset_rele(origin, FTAG);
1555                                 dsl_dataset_rele(ds, FTAG);
1556                                 return (SET_ERROR(ENODEV));
1557                         }
1558                         dsl_dataset_rele(origin, FTAG);
1559                 }
1560
1561                 dsl_dataset_rele(ds, FTAG);
1562                 error = 0;
1563         }
1564         return (error);
1565 }
1566
1567 static void
1568 dmu_recv_begin_sync(void *arg, dmu_tx_t *tx)
1569 {
1570         dmu_recv_begin_arg_t *drba = arg;
1571         dsl_pool_t *dp = dmu_tx_pool(tx);
1572         objset_t *mos = dp->dp_meta_objset;
1573         struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
1574         const char *tofs = drba->drba_cookie->drc_tofs;
1575         dsl_dataset_t *ds, *newds;
1576         uint64_t dsobj;
1577         int error;
1578         uint64_t crflags = 0;
1579
1580         if (drrb->drr_flags & DRR_FLAG_CI_DATA)
1581                 crflags |= DS_FLAG_CI_DATASET;
1582
1583         error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
1584         if (error == 0) {
1585                 /* create temporary clone */
1586                 dsl_dataset_t *snap = NULL;
1587                 if (drba->drba_snapobj != 0) {
1588                         VERIFY0(dsl_dataset_hold_obj(dp,
1589                             drba->drba_snapobj, FTAG, &snap));
1590                 }
1591                 dsobj = dsl_dataset_create_sync(ds->ds_dir, recv_clone_name,
1592                     snap, crflags, drba->drba_cred, tx);
1593                 if (drba->drba_snapobj != 0)
1594                         dsl_dataset_rele(snap, FTAG);
1595                 dsl_dataset_rele(ds, FTAG);
1596         } else {
1597                 dsl_dir_t *dd;
1598                 const char *tail;
1599                 dsl_dataset_t *origin = NULL;
1600
1601                 VERIFY0(dsl_dir_hold(dp, tofs, FTAG, &dd, &tail));
1602
1603                 if (drba->drba_origin != NULL) {
1604                         VERIFY0(dsl_dataset_hold(dp, drba->drba_origin,
1605                             FTAG, &origin));
1606                 }
1607
1608                 /* Create new dataset. */
1609                 dsobj = dsl_dataset_create_sync(dd,
1610                     strrchr(tofs, '/') + 1,
1611                     origin, crflags, drba->drba_cred, tx);
1612                 if (origin != NULL)
1613                         dsl_dataset_rele(origin, FTAG);
1614                 dsl_dir_rele(dd, FTAG);
1615                 drba->drba_cookie->drc_newfs = B_TRUE;
1616         }
1617         VERIFY0(dsl_dataset_own_obj(dp, dsobj, dmu_recv_tag, &newds));
1618
1619         if (drba->drba_cookie->drc_resumable) {
1620                 dsl_dataset_zapify(newds, tx);
1621                 if (drrb->drr_fromguid != 0) {
1622                         VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_FROMGUID,
1623                             8, 1, &drrb->drr_fromguid, tx));
1624                 }
1625                 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_TOGUID,
1626                     8, 1, &drrb->drr_toguid, tx));
1627                 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_TONAME,
1628                     1, strlen(drrb->drr_toname) + 1, drrb->drr_toname, tx));
1629                 uint64_t one = 1;
1630                 uint64_t zero = 0;
1631                 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_OBJECT,
1632                     8, 1, &one, tx));
1633                 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_OFFSET,
1634                     8, 1, &zero, tx));
1635                 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_BYTES,
1636                     8, 1, &zero, tx));
1637                 if (DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo) &
1638                     DMU_BACKUP_FEATURE_LARGE_BLOCKS) {
1639                         VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_LARGEBLOCK,
1640                             8, 1, &one, tx));
1641                 }
1642                 if (DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo) &
1643                     DMU_BACKUP_FEATURE_EMBED_DATA) {
1644                         VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_EMBEDOK,
1645                             8, 1, &one, tx));
1646                 }
1647                 if (DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo) &
1648                     DMU_BACKUP_FEATURE_COMPRESSED) {
1649                         VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_COMPRESSOK,
1650                             8, 1, &one, tx));
1651                 }
1652         }
1653
1654         dmu_buf_will_dirty(newds->ds_dbuf, tx);
1655         dsl_dataset_phys(newds)->ds_flags |= DS_FLAG_INCONSISTENT;
1656
1657         /*
1658          * If we actually created a non-clone, we need to create the
1659          * objset in our new dataset.
1660          */
1661         rrw_enter(&newds->ds_bp_rwlock, RW_READER, FTAG);
1662         if (BP_IS_HOLE(dsl_dataset_get_blkptr(newds))) {
1663                 (void) dmu_objset_create_impl(dp->dp_spa,
1664                     newds, dsl_dataset_get_blkptr(newds), drrb->drr_type, tx);
1665         }
1666         rrw_exit(&newds->ds_bp_rwlock, FTAG);
1667
1668         drba->drba_cookie->drc_ds = newds;
1669
1670         spa_history_log_internal_ds(newds, "receive", tx, "");
1671 }
1672
1673 static int
1674 dmu_recv_resume_begin_check(void *arg, dmu_tx_t *tx)
1675 {
1676         dmu_recv_begin_arg_t *drba = arg;
1677         dsl_pool_t *dp = dmu_tx_pool(tx);
1678         struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
1679         int error;
1680         uint64_t featureflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo);
1681         dsl_dataset_t *ds;
1682         const char *tofs = drba->drba_cookie->drc_tofs;
1683
1684         /* 6 extra bytes for /%recv */
1685         char recvname[ZFS_MAX_DATASET_NAME_LEN + 6];
1686
1687         /* already checked */
1688         ASSERT3U(drrb->drr_magic, ==, DMU_BACKUP_MAGIC);
1689         ASSERT(featureflags & DMU_BACKUP_FEATURE_RESUMING);
1690
1691         if (DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) ==
1692             DMU_COMPOUNDSTREAM ||
1693             drrb->drr_type >= DMU_OST_NUMTYPES)
1694                 return (SET_ERROR(EINVAL));
1695
1696         /* Verify pool version supports SA if SA_SPILL feature set */
1697         if ((featureflags & DMU_BACKUP_FEATURE_SA_SPILL) &&
1698             spa_version(dp->dp_spa) < SPA_VERSION_SA)
1699                 return (SET_ERROR(ENOTSUP));
1700
1701         /*
1702          * The receiving code doesn't know how to translate a WRITE_EMBEDDED
1703          * record to a plain WRITE record, so the pool must have the
1704          * EMBEDDED_DATA feature enabled if the stream has WRITE_EMBEDDED
1705          * records.  Same with WRITE_EMBEDDED records that use LZ4 compression.
1706          */
1707         if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA) &&
1708             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA))
1709                 return (SET_ERROR(ENOTSUP));
1710         if ((featureflags & DMU_BACKUP_FEATURE_LZ4) &&
1711             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS))
1712                 return (SET_ERROR(ENOTSUP));
1713
1714         /*
1715          * The receiving code doesn't know how to translate large blocks
1716          * to smaller ones, so the pool must have the LARGE_BLOCKS
1717          * feature enabled if the stream has LARGE_BLOCKS. Same with
1718          * large dnodes.
1719          */
1720         if ((featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
1721             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LARGE_BLOCKS))
1722                 return (SET_ERROR(ENOTSUP));
1723         if ((featureflags & DMU_BACKUP_FEATURE_LARGE_DNODE) &&
1724             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LARGE_DNODE))
1725                 return (SET_ERROR(ENOTSUP));
1726
1727         (void) snprintf(recvname, sizeof (recvname), "%s/%s",
1728             tofs, recv_clone_name);
1729
1730         if (dsl_dataset_hold(dp, recvname, FTAG, &ds) != 0) {
1731                 /* %recv does not exist; continue in tofs */
1732                 error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
1733                 if (error != 0)
1734                         return (error);
1735         }
1736
1737         /* check that ds is marked inconsistent */
1738         if (!DS_IS_INCONSISTENT(ds)) {
1739                 dsl_dataset_rele(ds, FTAG);
1740                 return (SET_ERROR(EINVAL));
1741         }
1742
1743         /* check that there is resuming data, and that the toguid matches */
1744         if (!dsl_dataset_is_zapified(ds)) {
1745                 dsl_dataset_rele(ds, FTAG);
1746                 return (SET_ERROR(EINVAL));
1747         }
1748         uint64_t val;
1749         error = zap_lookup(dp->dp_meta_objset, ds->ds_object,
1750             DS_FIELD_RESUME_TOGUID, sizeof (val), 1, &val);
1751         if (error != 0 || drrb->drr_toguid != val) {
1752                 dsl_dataset_rele(ds, FTAG);
1753                 return (SET_ERROR(EINVAL));
1754         }
1755
1756         /*
1757          * Check if the receive is still running.  If so, it will be owned.
1758          * Note that nothing else can own the dataset (e.g. after the receive
1759          * fails) because it will be marked inconsistent.
1760          */
1761         if (dsl_dataset_has_owner(ds)) {
1762                 dsl_dataset_rele(ds, FTAG);
1763                 return (SET_ERROR(EBUSY));
1764         }
1765
1766         /* There should not be any snapshots of this fs yet. */
1767         if (ds->ds_prev != NULL && ds->ds_prev->ds_dir == ds->ds_dir) {
1768                 dsl_dataset_rele(ds, FTAG);
1769                 return (SET_ERROR(EINVAL));
1770         }
1771
1772         /*
1773          * Note: resume point will be checked when we process the first WRITE
1774          * record.
1775          */
1776
1777         /* check that the origin matches */
1778         val = 0;
1779         (void) zap_lookup(dp->dp_meta_objset, ds->ds_object,
1780             DS_FIELD_RESUME_FROMGUID, sizeof (val), 1, &val);
1781         if (drrb->drr_fromguid != val) {
1782                 dsl_dataset_rele(ds, FTAG);
1783                 return (SET_ERROR(EINVAL));
1784         }
1785
1786         dsl_dataset_rele(ds, FTAG);
1787         return (0);
1788 }
1789
1790 static void
1791 dmu_recv_resume_begin_sync(void *arg, dmu_tx_t *tx)
1792 {
1793         dmu_recv_begin_arg_t *drba = arg;
1794         dsl_pool_t *dp = dmu_tx_pool(tx);
1795         const char *tofs = drba->drba_cookie->drc_tofs;
1796         dsl_dataset_t *ds;
1797         uint64_t dsobj;
1798         /* 6 extra bytes for /%recv */
1799         char recvname[ZFS_MAX_DATASET_NAME_LEN + 6];
1800
1801         (void) snprintf(recvname, sizeof (recvname), "%s/%s",
1802             tofs, recv_clone_name);
1803
1804         if (dsl_dataset_hold(dp, recvname, FTAG, &ds) != 0) {
1805                 /* %recv does not exist; continue in tofs */
1806                 VERIFY0(dsl_dataset_hold(dp, tofs, FTAG, &ds));
1807                 drba->drba_cookie->drc_newfs = B_TRUE;
1808         }
1809
1810         /* clear the inconsistent flag so that we can own it */
1811         ASSERT(DS_IS_INCONSISTENT(ds));
1812         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1813         dsl_dataset_phys(ds)->ds_flags &= ~DS_FLAG_INCONSISTENT;
1814         dsobj = ds->ds_object;
1815         dsl_dataset_rele(ds, FTAG);
1816
1817         VERIFY0(dsl_dataset_own_obj(dp, dsobj, dmu_recv_tag, &ds));
1818
1819         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1820         dsl_dataset_phys(ds)->ds_flags |= DS_FLAG_INCONSISTENT;
1821
1822         rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
1823         ASSERT(!BP_IS_HOLE(dsl_dataset_get_blkptr(ds)));
1824         rrw_exit(&ds->ds_bp_rwlock, FTAG);
1825
1826         drba->drba_cookie->drc_ds = ds;
1827
1828         spa_history_log_internal_ds(ds, "resume receive", tx, "");
1829 }
1830
1831 /*
1832  * NB: callers *MUST* call dmu_recv_stream() if dmu_recv_begin()
1833  * succeeds; otherwise we will leak the holds on the datasets.
1834  */
1835 int
1836 dmu_recv_begin(char *tofs, char *tosnap, dmu_replay_record_t *drr_begin,
1837     boolean_t force, boolean_t resumable, char *origin, dmu_recv_cookie_t *drc)
1838 {
1839         dmu_recv_begin_arg_t drba = { 0 };
1840
1841         bzero(drc, sizeof (dmu_recv_cookie_t));
1842         drc->drc_drr_begin = drr_begin;
1843         drc->drc_drrb = &drr_begin->drr_u.drr_begin;
1844         drc->drc_tosnap = tosnap;
1845         drc->drc_tofs = tofs;
1846         drc->drc_force = force;
1847         drc->drc_resumable = resumable;
1848         drc->drc_cred = CRED();
1849         drc->drc_clone = (origin != NULL);
1850
1851         if (drc->drc_drrb->drr_magic == BSWAP_64(DMU_BACKUP_MAGIC)) {
1852                 drc->drc_byteswap = B_TRUE;
1853                 (void) fletcher_4_incremental_byteswap(drr_begin,
1854                     sizeof (dmu_replay_record_t), &drc->drc_cksum);
1855                 byteswap_record(drr_begin);
1856         } else if (drc->drc_drrb->drr_magic == DMU_BACKUP_MAGIC) {
1857                 (void) fletcher_4_incremental_native(drr_begin,
1858                     sizeof (dmu_replay_record_t), &drc->drc_cksum);
1859         } else {
1860                 return (SET_ERROR(EINVAL));
1861         }
1862
1863         drba.drba_origin = origin;
1864         drba.drba_cookie = drc;
1865         drba.drba_cred = CRED();
1866
1867         if (DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo) &
1868             DMU_BACKUP_FEATURE_RESUMING) {
1869                 return (dsl_sync_task(tofs,
1870                     dmu_recv_resume_begin_check, dmu_recv_resume_begin_sync,
1871                     &drba, 5, ZFS_SPACE_CHECK_NORMAL));
1872         } else  {
1873                 return (dsl_sync_task(tofs,
1874                     dmu_recv_begin_check, dmu_recv_begin_sync,
1875                     &drba, 5, ZFS_SPACE_CHECK_NORMAL));
1876         }
1877 }
1878
1879 struct receive_record_arg {
1880         dmu_replay_record_t header;
1881         void *payload; /* Pointer to a buffer containing the payload */
1882         /*
1883          * If the record is a write, pointer to the arc_buf_t containing the
1884          * payload.
1885          */
1886         arc_buf_t *write_buf;
1887         int payload_size;
1888         uint64_t bytes_read; /* bytes read from stream when record created */
1889         boolean_t eos_marker; /* Marks the end of the stream */
1890         bqueue_node_t node;
1891 };
1892
1893 struct receive_writer_arg {
1894         objset_t *os;
1895         boolean_t byteswap;
1896         bqueue_t q;
1897
1898         /*
1899          * These three args are used to signal to the main thread that we're
1900          * done.
1901          */
1902         kmutex_t mutex;
1903         kcondvar_t cv;
1904         boolean_t done;
1905
1906         int err;
1907         /* A map from guid to dataset to help handle dedup'd streams. */
1908         avl_tree_t *guid_to_ds_map;
1909         boolean_t resumable;
1910         uint64_t last_object;
1911         uint64_t last_offset;
1912         uint64_t max_object; /* highest object ID referenced in stream */
1913         uint64_t bytes_read; /* bytes read when current record created */
1914 };
1915
1916 struct objlist {
1917         list_t list; /* List of struct receive_objnode. */
1918         /*
1919          * Last object looked up. Used to assert that objects are being looked
1920          * up in ascending order.
1921          */
1922         uint64_t last_lookup;
1923 };
1924
1925 struct receive_objnode {
1926         list_node_t node;
1927         uint64_t object;
1928 };
1929
1930 struct receive_arg {
1931         objset_t *os;
1932         kthread_t *td;
1933         struct file *fp;
1934         uint64_t voff; /* The current offset in the stream */
1935         uint64_t bytes_read;
1936         /*
1937          * A record that has had its payload read in, but hasn't yet been handed
1938          * off to the worker thread.
1939          */
1940         struct receive_record_arg *rrd;
1941         /* A record that has had its header read in, but not its payload. */
1942         struct receive_record_arg *next_rrd;
1943         zio_cksum_t cksum;
1944         zio_cksum_t prev_cksum;
1945         int err;
1946         boolean_t byteswap;
1947         /* Sorted list of objects not to issue prefetches for. */
1948         struct objlist ignore_objlist;
1949 };
1950
1951 typedef struct guid_map_entry {
1952         uint64_t        guid;
1953         dsl_dataset_t   *gme_ds;
1954         avl_node_t      avlnode;
1955 } guid_map_entry_t;
1956
1957 static int
1958 guid_compare(const void *arg1, const void *arg2)
1959 {
1960         const guid_map_entry_t *gmep1 = (const guid_map_entry_t *)arg1;
1961         const guid_map_entry_t *gmep2 = (const guid_map_entry_t *)arg2;
1962
1963         return (AVL_CMP(gmep1->guid, gmep2->guid));
1964 }
1965
1966 static void
1967 free_guid_map_onexit(void *arg)
1968 {
1969         avl_tree_t *ca = arg;
1970         void *cookie = NULL;
1971         guid_map_entry_t *gmep;
1972
1973         while ((gmep = avl_destroy_nodes(ca, &cookie)) != NULL) {
1974                 dsl_dataset_long_rele(gmep->gme_ds, gmep);
1975                 dsl_dataset_rele(gmep->gme_ds, gmep);
1976                 kmem_free(gmep, sizeof (guid_map_entry_t));
1977         }
1978         avl_destroy(ca);
1979         kmem_free(ca, sizeof (avl_tree_t));
1980 }
1981
1982 static int
1983 restore_bytes(struct receive_arg *ra, void *buf, int len, off_t off, ssize_t *resid)
1984 {
1985         struct uio auio;
1986         struct iovec aiov;
1987         int error;
1988
1989         aiov.iov_base = buf;
1990         aiov.iov_len = len;
1991         auio.uio_iov = &aiov;
1992         auio.uio_iovcnt = 1;
1993         auio.uio_resid = len;
1994         auio.uio_segflg = UIO_SYSSPACE;
1995         auio.uio_rw = UIO_READ;
1996         auio.uio_offset = off;
1997         auio.uio_td = ra->td;
1998 #ifdef _KERNEL
1999         error = fo_read(ra->fp, &auio, ra->td->td_ucred, FOF_OFFSET, ra->td);
2000 #else
2001         fprintf(stderr, "%s: returning EOPNOTSUPP\n", __func__);
2002         error = EOPNOTSUPP;
2003 #endif
2004         *resid = auio.uio_resid;
2005         return (error);
2006 }
2007
2008 static int
2009 receive_read(struct receive_arg *ra, int len, void *buf)
2010 {
2011         int done = 0;
2012
2013         /*
2014          * The code doesn't rely on this (lengths being multiples of 8).  See
2015          * comment in dump_bytes.
2016          */
2017         ASSERT0(len % 8);
2018
2019         while (done < len) {
2020                 ssize_t resid;
2021
2022                 ra->err = restore_bytes(ra, buf + done,
2023                     len - done, ra->voff, &resid);
2024
2025                 if (resid == len - done) {
2026                         /*
2027                          * Note: ECKSUM indicates that the receive
2028                          * was interrupted and can potentially be resumed.
2029                          */
2030                         ra->err = SET_ERROR(ECKSUM);
2031                 }
2032                 ra->voff += len - done - resid;
2033                 done = len - resid;
2034                 if (ra->err != 0)
2035                         return (ra->err);
2036         }
2037
2038         ra->bytes_read += len;
2039
2040         ASSERT3U(done, ==, len);
2041         return (0);
2042 }
2043
2044 noinline static void
2045 byteswap_record(dmu_replay_record_t *drr)
2046 {
2047 #define DO64(X) (drr->drr_u.X = BSWAP_64(drr->drr_u.X))
2048 #define DO32(X) (drr->drr_u.X = BSWAP_32(drr->drr_u.X))
2049         drr->drr_type = BSWAP_32(drr->drr_type);
2050         drr->drr_payloadlen = BSWAP_32(drr->drr_payloadlen);
2051
2052         switch (drr->drr_type) {
2053         case DRR_BEGIN:
2054                 DO64(drr_begin.drr_magic);
2055                 DO64(drr_begin.drr_versioninfo);
2056                 DO64(drr_begin.drr_creation_time);
2057                 DO32(drr_begin.drr_type);
2058                 DO32(drr_begin.drr_flags);
2059                 DO64(drr_begin.drr_toguid);
2060                 DO64(drr_begin.drr_fromguid);
2061                 break;
2062         case DRR_OBJECT:
2063                 DO64(drr_object.drr_object);
2064                 DO32(drr_object.drr_type);
2065                 DO32(drr_object.drr_bonustype);
2066                 DO32(drr_object.drr_blksz);
2067                 DO32(drr_object.drr_bonuslen);
2068                 DO64(drr_object.drr_toguid);
2069                 break;
2070         case DRR_FREEOBJECTS:
2071                 DO64(drr_freeobjects.drr_firstobj);
2072                 DO64(drr_freeobjects.drr_numobjs);
2073                 DO64(drr_freeobjects.drr_toguid);
2074                 break;
2075         case DRR_WRITE:
2076                 DO64(drr_write.drr_object);
2077                 DO32(drr_write.drr_type);
2078                 DO64(drr_write.drr_offset);
2079                 DO64(drr_write.drr_logical_size);
2080                 DO64(drr_write.drr_toguid);
2081                 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_write.drr_key.ddk_cksum);
2082                 DO64(drr_write.drr_key.ddk_prop);
2083                 DO64(drr_write.drr_compressed_size);
2084                 break;
2085         case DRR_WRITE_BYREF:
2086                 DO64(drr_write_byref.drr_object);
2087                 DO64(drr_write_byref.drr_offset);
2088                 DO64(drr_write_byref.drr_length);
2089                 DO64(drr_write_byref.drr_toguid);
2090                 DO64(drr_write_byref.drr_refguid);
2091                 DO64(drr_write_byref.drr_refobject);
2092                 DO64(drr_write_byref.drr_refoffset);
2093                 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_write_byref.
2094                     drr_key.ddk_cksum);
2095                 DO64(drr_write_byref.drr_key.ddk_prop);
2096                 break;
2097         case DRR_WRITE_EMBEDDED:
2098                 DO64(drr_write_embedded.drr_object);
2099                 DO64(drr_write_embedded.drr_offset);
2100                 DO64(drr_write_embedded.drr_length);
2101                 DO64(drr_write_embedded.drr_toguid);
2102                 DO32(drr_write_embedded.drr_lsize);
2103                 DO32(drr_write_embedded.drr_psize);
2104                 break;
2105         case DRR_FREE:
2106                 DO64(drr_free.drr_object);
2107                 DO64(drr_free.drr_offset);
2108                 DO64(drr_free.drr_length);
2109                 DO64(drr_free.drr_toguid);
2110                 break;
2111         case DRR_SPILL:
2112                 DO64(drr_spill.drr_object);
2113                 DO64(drr_spill.drr_length);
2114                 DO64(drr_spill.drr_toguid);
2115                 break;
2116         case DRR_END:
2117                 DO64(drr_end.drr_toguid);
2118                 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_end.drr_checksum);
2119                 break;
2120         }
2121
2122         if (drr->drr_type != DRR_BEGIN) {
2123                 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_checksum.drr_checksum);
2124         }
2125
2126 #undef DO64
2127 #undef DO32
2128 }
2129
2130 static inline uint8_t
2131 deduce_nblkptr(dmu_object_type_t bonus_type, uint64_t bonus_size)
2132 {
2133         if (bonus_type == DMU_OT_SA) {
2134                 return (1);
2135         } else {
2136                 return (1 +
2137                     ((DN_OLD_MAX_BONUSLEN -
2138                     MIN(DN_OLD_MAX_BONUSLEN, bonus_size)) >> SPA_BLKPTRSHIFT));
2139         }
2140 }
2141
2142 static void
2143 save_resume_state(struct receive_writer_arg *rwa,
2144     uint64_t object, uint64_t offset, dmu_tx_t *tx)
2145 {
2146         int txgoff = dmu_tx_get_txg(tx) & TXG_MASK;
2147
2148         if (!rwa->resumable)
2149                 return;
2150
2151         /*
2152          * We use ds_resume_bytes[] != 0 to indicate that we need to
2153          * update this on disk, so it must not be 0.
2154          */
2155         ASSERT(rwa->bytes_read != 0);
2156
2157         /*
2158          * We only resume from write records, which have a valid
2159          * (non-meta-dnode) object number.
2160          */
2161         ASSERT(object != 0);
2162
2163         /*
2164          * For resuming to work correctly, we must receive records in order,
2165          * sorted by object,offset.  This is checked by the callers, but
2166          * assert it here for good measure.
2167          */
2168         ASSERT3U(object, >=, rwa->os->os_dsl_dataset->ds_resume_object[txgoff]);
2169         ASSERT(object != rwa->os->os_dsl_dataset->ds_resume_object[txgoff] ||
2170             offset >= rwa->os->os_dsl_dataset->ds_resume_offset[txgoff]);
2171         ASSERT3U(rwa->bytes_read, >=,
2172             rwa->os->os_dsl_dataset->ds_resume_bytes[txgoff]);
2173
2174         rwa->os->os_dsl_dataset->ds_resume_object[txgoff] = object;
2175         rwa->os->os_dsl_dataset->ds_resume_offset[txgoff] = offset;
2176         rwa->os->os_dsl_dataset->ds_resume_bytes[txgoff] = rwa->bytes_read;
2177 }
2178
2179 noinline static int
2180 receive_object(struct receive_writer_arg *rwa, struct drr_object *drro,
2181     void *data)
2182 {
2183         dmu_object_info_t doi;
2184         dmu_tx_t *tx;
2185         uint64_t object;
2186         int err;
2187         uint8_t dn_slots = drro->drr_dn_slots != 0 ?
2188             drro->drr_dn_slots : DNODE_MIN_SLOTS;
2189
2190         if (drro->drr_type == DMU_OT_NONE ||
2191             !DMU_OT_IS_VALID(drro->drr_type) ||
2192             !DMU_OT_IS_VALID(drro->drr_bonustype) ||
2193             drro->drr_checksumtype >= ZIO_CHECKSUM_FUNCTIONS ||
2194             drro->drr_compress >= ZIO_COMPRESS_FUNCTIONS ||
2195             P2PHASE(drro->drr_blksz, SPA_MINBLOCKSIZE) ||
2196             drro->drr_blksz < SPA_MINBLOCKSIZE ||
2197             drro->drr_blksz > spa_maxblocksize(dmu_objset_spa(rwa->os)) ||
2198             drro->drr_bonuslen >
2199             DN_BONUS_SIZE(spa_maxdnodesize(dmu_objset_spa(rwa->os))) ||
2200             dn_slots >
2201             (spa_maxdnodesize(dmu_objset_spa(rwa->os)) >> DNODE_SHIFT)) {
2202                 return (SET_ERROR(EINVAL));
2203         }
2204
2205         err = dmu_object_info(rwa->os, drro->drr_object, &doi);
2206
2207         if (err != 0 && err != ENOENT && err != EEXIST)
2208                 return (SET_ERROR(EINVAL));
2209
2210         if (drro->drr_object > rwa->max_object)
2211                 rwa->max_object = drro->drr_object;
2212
2213         /*
2214          * If we are losing blkptrs or changing the block size this must
2215          * be a new file instance.  We must clear out the previous file
2216          * contents before we can change this type of metadata in the dnode.
2217          */
2218         if (err == 0) {
2219                 int nblkptr;
2220
2221                 object = drro->drr_object;
2222
2223                 nblkptr = deduce_nblkptr(drro->drr_bonustype,
2224                     drro->drr_bonuslen);
2225
2226                 if (drro->drr_blksz != doi.doi_data_block_size ||
2227                     nblkptr < doi.doi_nblkptr ||
2228                     dn_slots != doi.doi_dnodesize >> DNODE_SHIFT) {
2229                         err = dmu_free_long_range(rwa->os, drro->drr_object,
2230                             0, DMU_OBJECT_END);
2231                         if (err != 0)
2232                                 return (SET_ERROR(EINVAL));
2233                 }
2234         } else if (err == EEXIST) {
2235                 /*
2236                  * The object requested is currently an interior slot of a
2237                  * multi-slot dnode. This will be resolved when the next txg
2238                  * is synced out, since the send stream will have told us
2239                  * to free this slot when we freed the associated dnode
2240                  * earlier in the stream.
2241                  */
2242                 txg_wait_synced(dmu_objset_pool(rwa->os), 0);
2243                 object = drro->drr_object;
2244         } else {
2245                 /* object is free and we are about to allocate a new one */
2246                 object = DMU_NEW_OBJECT;
2247         }
2248
2249         /*
2250          * If this is a multi-slot dnode there is a chance that this
2251          * object will expand into a slot that is already used by
2252          * another object from the previous snapshot. We must free
2253          * these objects before we attempt to allocate the new dnode.
2254          */
2255         if (dn_slots > 1) {
2256                 boolean_t need_sync = B_FALSE;
2257
2258                 for (uint64_t slot = drro->drr_object + 1;
2259                     slot < drro->drr_object + dn_slots;
2260                     slot++) {
2261                         dmu_object_info_t slot_doi;
2262
2263                         err = dmu_object_info(rwa->os, slot, &slot_doi);
2264                         if (err == ENOENT || err == EEXIST)
2265                                 continue;
2266                         else if (err != 0)
2267                                 return (err);
2268
2269                         err = dmu_free_long_object(rwa->os, slot);
2270
2271                         if (err != 0)
2272                                 return (err);
2273
2274                         need_sync = B_TRUE;
2275                 }
2276
2277                 if (need_sync)
2278                         txg_wait_synced(dmu_objset_pool(rwa->os), 0);
2279         }
2280
2281         tx = dmu_tx_create(rwa->os);
2282         dmu_tx_hold_bonus(tx, object);
2283         err = dmu_tx_assign(tx, TXG_WAIT);
2284         if (err != 0) {
2285                 dmu_tx_abort(tx);
2286                 return (err);
2287         }
2288
2289         if (object == DMU_NEW_OBJECT) {
2290                 /* currently free, want to be allocated */
2291                 err = dmu_object_claim_dnsize(rwa->os, drro->drr_object,
2292                     drro->drr_type, drro->drr_blksz,
2293                     drro->drr_bonustype, drro->drr_bonuslen,
2294                     dn_slots << DNODE_SHIFT, tx);
2295         } else if (drro->drr_type != doi.doi_type ||
2296             drro->drr_blksz != doi.doi_data_block_size ||
2297             drro->drr_bonustype != doi.doi_bonus_type ||
2298             drro->drr_bonuslen != doi.doi_bonus_size ||
2299             drro->drr_dn_slots != (doi.doi_dnodesize >> DNODE_SHIFT)) {
2300                 /* currently allocated, but with different properties */
2301                 err = dmu_object_reclaim_dnsize(rwa->os, drro->drr_object,
2302                     drro->drr_type, drro->drr_blksz,
2303                     drro->drr_bonustype, drro->drr_bonuslen,
2304                     drro->drr_dn_slots << DNODE_SHIFT, tx);
2305         }
2306         if (err != 0) {
2307                 dmu_tx_commit(tx);
2308                 return (SET_ERROR(EINVAL));
2309         }
2310
2311         dmu_object_set_checksum(rwa->os, drro->drr_object,
2312             drro->drr_checksumtype, tx);
2313         dmu_object_set_compress(rwa->os, drro->drr_object,
2314             drro->drr_compress, tx);
2315
2316         if (data != NULL) {
2317                 dmu_buf_t *db;
2318
2319                 VERIFY0(dmu_bonus_hold(rwa->os, drro->drr_object, FTAG, &db));
2320                 dmu_buf_will_dirty(db, tx);
2321
2322                 ASSERT3U(db->db_size, >=, drro->drr_bonuslen);
2323                 bcopy(data, db->db_data, drro->drr_bonuslen);
2324                 if (rwa->byteswap) {
2325                         dmu_object_byteswap_t byteswap =
2326                             DMU_OT_BYTESWAP(drro->drr_bonustype);
2327                         dmu_ot_byteswap[byteswap].ob_func(db->db_data,
2328                             drro->drr_bonuslen);
2329                 }
2330                 dmu_buf_rele(db, FTAG);
2331         }
2332         dmu_tx_commit(tx);
2333
2334         return (0);
2335 }
2336
2337 /* ARGSUSED */
2338 noinline static int
2339 receive_freeobjects(struct receive_writer_arg *rwa,
2340     struct drr_freeobjects *drrfo)
2341 {
2342         uint64_t obj;
2343         int next_err = 0;
2344
2345         if (drrfo->drr_firstobj + drrfo->drr_numobjs < drrfo->drr_firstobj)
2346                 return (SET_ERROR(EINVAL));
2347
2348         for (obj = drrfo->drr_firstobj == 0 ? 1 : drrfo->drr_firstobj;
2349             obj < drrfo->drr_firstobj + drrfo->drr_numobjs && next_err == 0;
2350             next_err = dmu_object_next(rwa->os, &obj, FALSE, 0)) {
2351                 dmu_object_info_t doi;
2352                 int err;
2353
2354                 err = dmu_object_info(rwa->os, obj, NULL);
2355                 if (err == ENOENT)
2356                         continue;
2357                 else if (err != 0)
2358                         return (err);
2359
2360                 err = dmu_free_long_object(rwa->os, obj);
2361                 if (err != 0)
2362                         return (err);
2363
2364                 if (obj > rwa->max_object)
2365                         rwa->max_object = obj;
2366         }
2367         if (next_err != ESRCH)
2368                 return (next_err);
2369         return (0);
2370 }
2371
2372 noinline static int
2373 receive_write(struct receive_writer_arg *rwa, struct drr_write *drrw,
2374     arc_buf_t *abuf)
2375 {
2376         dmu_tx_t *tx;
2377         int err;
2378
2379         if (drrw->drr_offset + drrw->drr_logical_size < drrw->drr_offset ||
2380             !DMU_OT_IS_VALID(drrw->drr_type))
2381                 return (SET_ERROR(EINVAL));
2382
2383         /*
2384          * For resuming to work, records must be in increasing order
2385          * by (object, offset).
2386          */
2387         if (drrw->drr_object < rwa->last_object ||
2388             (drrw->drr_object == rwa->last_object &&
2389             drrw->drr_offset < rwa->last_offset)) {
2390                 return (SET_ERROR(EINVAL));
2391         }
2392         rwa->last_object = drrw->drr_object;
2393         rwa->last_offset = drrw->drr_offset;
2394
2395         if (rwa->last_object > rwa->max_object)
2396                 rwa->max_object = rwa->last_object;
2397
2398         if (dmu_object_info(rwa->os, drrw->drr_object, NULL) != 0)
2399                 return (SET_ERROR(EINVAL));
2400
2401         tx = dmu_tx_create(rwa->os);
2402         dmu_tx_hold_write(tx, drrw->drr_object,
2403             drrw->drr_offset, drrw->drr_logical_size);
2404         err = dmu_tx_assign(tx, TXG_WAIT);
2405         if (err != 0) {
2406                 dmu_tx_abort(tx);
2407                 return (err);
2408         }
2409         if (rwa->byteswap) {
2410                 dmu_object_byteswap_t byteswap =
2411                     DMU_OT_BYTESWAP(drrw->drr_type);
2412                 dmu_ot_byteswap[byteswap].ob_func(abuf->b_data,
2413                     DRR_WRITE_PAYLOAD_SIZE(drrw));
2414         }
2415
2416         /* use the bonus buf to look up the dnode in dmu_assign_arcbuf */
2417         dmu_buf_t *bonus;
2418         if (dmu_bonus_hold(rwa->os, drrw->drr_object, FTAG, &bonus) != 0)
2419                 return (SET_ERROR(EINVAL));
2420         dmu_assign_arcbuf(bonus, drrw->drr_offset, abuf, tx);
2421
2422         /*
2423          * Note: If the receive fails, we want the resume stream to start
2424          * with the same record that we last successfully received (as opposed
2425          * to the next record), so that we can verify that we are
2426          * resuming from the correct location.
2427          */
2428         save_resume_state(rwa, drrw->drr_object, drrw->drr_offset, tx);
2429         dmu_tx_commit(tx);
2430         dmu_buf_rele(bonus, FTAG);
2431
2432         return (0);
2433 }
2434
2435 /*
2436  * Handle a DRR_WRITE_BYREF record.  This record is used in dedup'ed
2437  * streams to refer to a copy of the data that is already on the
2438  * system because it came in earlier in the stream.  This function
2439  * finds the earlier copy of the data, and uses that copy instead of
2440  * data from the stream to fulfill this write.
2441  */
2442 static int
2443 receive_write_byref(struct receive_writer_arg *rwa,
2444     struct drr_write_byref *drrwbr)
2445 {
2446         dmu_tx_t *tx;
2447         int err;
2448         guid_map_entry_t gmesrch;
2449         guid_map_entry_t *gmep;
2450         avl_index_t where;
2451         objset_t *ref_os = NULL;
2452         dmu_buf_t *dbp;
2453
2454         if (drrwbr->drr_offset + drrwbr->drr_length < drrwbr->drr_offset)
2455                 return (SET_ERROR(EINVAL));
2456
2457         /*
2458          * If the GUID of the referenced dataset is different from the
2459          * GUID of the target dataset, find the referenced dataset.
2460          */
2461         if (drrwbr->drr_toguid != drrwbr->drr_refguid) {
2462                 gmesrch.guid = drrwbr->drr_refguid;
2463                 if ((gmep = avl_find(rwa->guid_to_ds_map, &gmesrch,
2464                     &where)) == NULL) {
2465                         return (SET_ERROR(EINVAL));
2466                 }
2467                 if (dmu_objset_from_ds(gmep->gme_ds, &ref_os))
2468                         return (SET_ERROR(EINVAL));
2469         } else {
2470                 ref_os = rwa->os;
2471         }
2472
2473         if (drrwbr->drr_object > rwa->max_object)
2474                 rwa->max_object = drrwbr->drr_object;
2475
2476         err = dmu_buf_hold(ref_os, drrwbr->drr_refobject,
2477             drrwbr->drr_refoffset, FTAG, &dbp, DMU_READ_PREFETCH);
2478         if (err != 0)
2479                 return (err);
2480
2481         tx = dmu_tx_create(rwa->os);
2482
2483         dmu_tx_hold_write(tx, drrwbr->drr_object,
2484             drrwbr->drr_offset, drrwbr->drr_length);
2485         err = dmu_tx_assign(tx, TXG_WAIT);
2486         if (err != 0) {
2487                 dmu_tx_abort(tx);
2488                 return (err);
2489         }
2490         dmu_write(rwa->os, drrwbr->drr_object,
2491             drrwbr->drr_offset, drrwbr->drr_length, dbp->db_data, tx);
2492         dmu_buf_rele(dbp, FTAG);
2493
2494         /* See comment in restore_write. */
2495         save_resume_state(rwa, drrwbr->drr_object, drrwbr->drr_offset, tx);
2496         dmu_tx_commit(tx);
2497         return (0);
2498 }
2499
2500 static int
2501 receive_write_embedded(struct receive_writer_arg *rwa,
2502     struct drr_write_embedded *drrwe, void *data)
2503 {
2504         dmu_tx_t *tx;
2505         int err;
2506
2507         if (drrwe->drr_offset + drrwe->drr_length < drrwe->drr_offset)
2508                 return (EINVAL);
2509
2510         if (drrwe->drr_psize > BPE_PAYLOAD_SIZE)
2511                 return (EINVAL);
2512
2513         if (drrwe->drr_etype >= NUM_BP_EMBEDDED_TYPES)
2514                 return (EINVAL);
2515         if (drrwe->drr_compression >= ZIO_COMPRESS_FUNCTIONS)
2516                 return (EINVAL);
2517
2518         if (drrwe->drr_object > rwa->max_object)
2519                 rwa->max_object = drrwe->drr_object;
2520
2521         tx = dmu_tx_create(rwa->os);
2522
2523         dmu_tx_hold_write(tx, drrwe->drr_object,
2524             drrwe->drr_offset, drrwe->drr_length);
2525         err = dmu_tx_assign(tx, TXG_WAIT);
2526         if (err != 0) {
2527                 dmu_tx_abort(tx);
2528                 return (err);
2529         }
2530
2531         dmu_write_embedded(rwa->os, drrwe->drr_object,
2532             drrwe->drr_offset, data, drrwe->drr_etype,
2533             drrwe->drr_compression, drrwe->drr_lsize, drrwe->drr_psize,
2534             rwa->byteswap ^ ZFS_HOST_BYTEORDER, tx);
2535
2536         /* See comment in restore_write. */
2537         save_resume_state(rwa, drrwe->drr_object, drrwe->drr_offset, tx);
2538         dmu_tx_commit(tx);
2539         return (0);
2540 }
2541
2542 static int
2543 receive_spill(struct receive_writer_arg *rwa, struct drr_spill *drrs,
2544     void *data)
2545 {
2546         dmu_tx_t *tx;
2547         dmu_buf_t *db, *db_spill;
2548         int err;
2549
2550         if (drrs->drr_length < SPA_MINBLOCKSIZE ||
2551             drrs->drr_length > spa_maxblocksize(dmu_objset_spa(rwa->os)))
2552                 return (SET_ERROR(EINVAL));
2553
2554         if (dmu_object_info(rwa->os, drrs->drr_object, NULL) != 0)
2555                 return (SET_ERROR(EINVAL));
2556
2557         if (drrs->drr_object > rwa->max_object)
2558                 rwa->max_object = drrs->drr_object;
2559
2560         VERIFY0(dmu_bonus_hold(rwa->os, drrs->drr_object, FTAG, &db));
2561         if ((err = dmu_spill_hold_by_bonus(db, FTAG, &db_spill)) != 0) {
2562                 dmu_buf_rele(db, FTAG);
2563                 return (err);
2564         }
2565
2566         tx = dmu_tx_create(rwa->os);
2567
2568         dmu_tx_hold_spill(tx, db->db_object);
2569
2570         err = dmu_tx_assign(tx, TXG_WAIT);
2571         if (err != 0) {
2572                 dmu_buf_rele(db, FTAG);
2573                 dmu_buf_rele(db_spill, FTAG);
2574                 dmu_tx_abort(tx);
2575                 return (err);
2576         }
2577         dmu_buf_will_dirty(db_spill, tx);
2578
2579         if (db_spill->db_size < drrs->drr_length)
2580                 VERIFY(0 == dbuf_spill_set_blksz(db_spill,
2581                     drrs->drr_length, tx));
2582         bcopy(data, db_spill->db_data, drrs->drr_length);
2583
2584         dmu_buf_rele(db, FTAG);
2585         dmu_buf_rele(db_spill, FTAG);
2586
2587         dmu_tx_commit(tx);
2588         return (0);
2589 }
2590
2591 /* ARGSUSED */
2592 noinline static int
2593 receive_free(struct receive_writer_arg *rwa, struct drr_free *drrf)
2594 {
2595         int err;
2596
2597         if (drrf->drr_length != -1ULL &&
2598             drrf->drr_offset + drrf->drr_length < drrf->drr_offset)
2599                 return (SET_ERROR(EINVAL));
2600
2601         if (dmu_object_info(rwa->os, drrf->drr_object, NULL) != 0)
2602                 return (SET_ERROR(EINVAL));
2603
2604         if (drrf->drr_object > rwa->max_object)
2605                 rwa->max_object = drrf->drr_object;
2606
2607         err = dmu_free_long_range(rwa->os, drrf->drr_object,
2608             drrf->drr_offset, drrf->drr_length);
2609
2610         return (err);
2611 }
2612
2613 /* used to destroy the drc_ds on error */
2614 static void
2615 dmu_recv_cleanup_ds(dmu_recv_cookie_t *drc)
2616 {
2617         if (drc->drc_resumable) {
2618                 /* wait for our resume state to be written to disk */
2619                 txg_wait_synced(drc->drc_ds->ds_dir->dd_pool, 0);
2620                 dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
2621         } else {
2622                 char name[ZFS_MAX_DATASET_NAME_LEN];
2623                 dsl_dataset_name(drc->drc_ds, name);
2624                 dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
2625                 (void) dsl_destroy_head(name);
2626         }
2627 }
2628
2629 static void
2630 receive_cksum(struct receive_arg *ra, int len, void *buf)
2631 {
2632         if (ra->byteswap) {
2633                 (void) fletcher_4_incremental_byteswap(buf, len, &ra->cksum);
2634         } else {
2635                 (void) fletcher_4_incremental_native(buf, len, &ra->cksum);
2636         }
2637 }
2638
2639 /*
2640  * Read the payload into a buffer of size len, and update the current record's
2641  * payload field.
2642  * Allocate ra->next_rrd and read the next record's header into
2643  * ra->next_rrd->header.
2644  * Verify checksum of payload and next record.
2645  */
2646 static int
2647 receive_read_payload_and_next_header(struct receive_arg *ra, int len, void *buf)
2648 {
2649         int err;
2650
2651         if (len != 0) {
2652                 ASSERT3U(len, <=, SPA_MAXBLOCKSIZE);
2653                 err = receive_read(ra, len, buf);
2654                 if (err != 0)
2655                         return (err);
2656                 receive_cksum(ra, len, buf);
2657
2658                 /* note: rrd is NULL when reading the begin record's payload */
2659                 if (ra->rrd != NULL) {
2660                         ra->rrd->payload = buf;
2661                         ra->rrd->payload_size = len;
2662                         ra->rrd->bytes_read = ra->bytes_read;
2663                 }
2664         }
2665
2666         ra->prev_cksum = ra->cksum;
2667
2668         ra->next_rrd = kmem_zalloc(sizeof (*ra->next_rrd), KM_SLEEP);
2669         err = receive_read(ra, sizeof (ra->next_rrd->header),
2670             &ra->next_rrd->header);
2671         ra->next_rrd->bytes_read = ra->bytes_read;
2672         if (err != 0) {
2673                 kmem_free(ra->next_rrd, sizeof (*ra->next_rrd));
2674                 ra->next_rrd = NULL;
2675                 return (err);
2676         }
2677         if (ra->next_rrd->header.drr_type == DRR_BEGIN) {
2678                 kmem_free(ra->next_rrd, sizeof (*ra->next_rrd));
2679                 ra->next_rrd = NULL;
2680                 return (SET_ERROR(EINVAL));
2681         }
2682
2683         /*
2684          * Note: checksum is of everything up to but not including the
2685          * checksum itself.
2686          */
2687         ASSERT3U(offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
2688             ==, sizeof (dmu_replay_record_t) - sizeof (zio_cksum_t));
2689         receive_cksum(ra,
2690             offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
2691             &ra->next_rrd->header);
2692
2693         zio_cksum_t cksum_orig =
2694             ra->next_rrd->header.drr_u.drr_checksum.drr_checksum;
2695         zio_cksum_t *cksump =
2696             &ra->next_rrd->header.drr_u.drr_checksum.drr_checksum;
2697
2698         if (ra->byteswap)
2699                 byteswap_record(&ra->next_rrd->header);
2700
2701         if ((!ZIO_CHECKSUM_IS_ZERO(cksump)) &&
2702             !ZIO_CHECKSUM_EQUAL(ra->cksum, *cksump)) {
2703                 kmem_free(ra->next_rrd, sizeof (*ra->next_rrd));
2704                 ra->next_rrd = NULL;
2705                 return (SET_ERROR(ECKSUM));
2706         }
2707
2708         receive_cksum(ra, sizeof (cksum_orig), &cksum_orig);
2709
2710         return (0);
2711 }
2712
2713 static void
2714 objlist_create(struct objlist *list)
2715 {
2716         list_create(&list->list, sizeof (struct receive_objnode),
2717             offsetof(struct receive_objnode, node));
2718         list->last_lookup = 0;
2719 }
2720
2721 static void
2722 objlist_destroy(struct objlist *list)
2723 {
2724         for (struct receive_objnode *n = list_remove_head(&list->list);
2725             n != NULL; n = list_remove_head(&list->list)) {
2726                 kmem_free(n, sizeof (*n));
2727         }
2728         list_destroy(&list->list);
2729 }
2730
2731 /*
2732  * This function looks through the objlist to see if the specified object number
2733  * is contained in the objlist.  In the process, it will remove all object
2734  * numbers in the list that are smaller than the specified object number.  Thus,
2735  * any lookup of an object number smaller than a previously looked up object
2736  * number will always return false; therefore, all lookups should be done in
2737  * ascending order.
2738  */
2739 static boolean_t
2740 objlist_exists(struct objlist *list, uint64_t object)
2741 {
2742         struct receive_objnode *node = list_head(&list->list);
2743         ASSERT3U(object, >=, list->last_lookup);
2744         list->last_lookup = object;
2745         while (node != NULL && node->object < object) {
2746                 VERIFY3P(node, ==, list_remove_head(&list->list));
2747                 kmem_free(node, sizeof (*node));
2748                 node = list_head(&list->list);
2749         }
2750         return (node != NULL && node->object == object);
2751 }
2752
2753 /*
2754  * The objlist is a list of object numbers stored in ascending order.  However,
2755  * the insertion of new object numbers does not seek out the correct location to
2756  * store a new object number; instead, it appends it to the list for simplicity.
2757  * Thus, any users must take care to only insert new object numbers in ascending
2758  * order.
2759  */
2760 static void
2761 objlist_insert(struct objlist *list, uint64_t object)
2762 {
2763         struct receive_objnode *node = kmem_zalloc(sizeof (*node), KM_SLEEP);
2764         node->object = object;
2765 #ifdef ZFS_DEBUG
2766         struct receive_objnode *last_object = list_tail(&list->list);
2767         uint64_t last_objnum = (last_object != NULL ? last_object->object : 0);
2768         ASSERT3U(node->object, >, last_objnum);
2769 #endif
2770         list_insert_tail(&list->list, node);
2771 }
2772
2773 /*
2774  * Issue the prefetch reads for any necessary indirect blocks.
2775  *
2776  * We use the object ignore list to tell us whether or not to issue prefetches
2777  * for a given object.  We do this for both correctness (in case the blocksize
2778  * of an object has changed) and performance (if the object doesn't exist, don't
2779  * needlessly try to issue prefetches).  We also trim the list as we go through
2780  * the stream to prevent it from growing to an unbounded size.
2781  *
2782  * The object numbers within will always be in sorted order, and any write
2783  * records we see will also be in sorted order, but they're not sorted with
2784  * respect to each other (i.e. we can get several object records before
2785  * receiving each object's write records).  As a result, once we've reached a
2786  * given object number, we can safely remove any reference to lower object
2787  * numbers in the ignore list. In practice, we receive up to 32 object records
2788  * before receiving write records, so the list can have up to 32 nodes in it.
2789  */
2790 /* ARGSUSED */
2791 static void
2792 receive_read_prefetch(struct receive_arg *ra,
2793     uint64_t object, uint64_t offset, uint64_t length)
2794 {
2795         if (!objlist_exists(&ra->ignore_objlist, object)) {
2796                 dmu_prefetch(ra->os, object, 1, offset, length,
2797                     ZIO_PRIORITY_SYNC_READ);
2798         }
2799 }
2800
2801 /*
2802  * Read records off the stream, issuing any necessary prefetches.
2803  */
2804 static int
2805 receive_read_record(struct receive_arg *ra)
2806 {
2807         int err;
2808
2809         switch (ra->rrd->header.drr_type) {
2810         case DRR_OBJECT:
2811         {
2812                 struct drr_object *drro = &ra->rrd->header.drr_u.drr_object;
2813                 uint32_t size = P2ROUNDUP(drro->drr_bonuslen, 8);
2814                 void *buf = kmem_zalloc(size, KM_SLEEP);
2815                 dmu_object_info_t doi;
2816                 err = receive_read_payload_and_next_header(ra, size, buf);
2817                 if (err != 0) {
2818                         kmem_free(buf, size);
2819                         return (err);
2820                 }
2821                 err = dmu_object_info(ra->os, drro->drr_object, &doi);
2822                 /*
2823                  * See receive_read_prefetch for an explanation why we're
2824                  * storing this object in the ignore_obj_list.
2825                  */
2826                 if (err == ENOENT ||
2827                     (err == 0 && doi.doi_data_block_size != drro->drr_blksz)) {
2828                         objlist_insert(&ra->ignore_objlist, drro->drr_object);
2829                         err = 0;
2830                 }
2831                 return (err);
2832         }
2833         case DRR_FREEOBJECTS:
2834         {
2835                 err = receive_read_payload_and_next_header(ra, 0, NULL);
2836                 return (err);
2837         }
2838         case DRR_WRITE:
2839         {
2840                 struct drr_write *drrw = &ra->rrd->header.drr_u.drr_write;
2841                 arc_buf_t *abuf;
2842                 boolean_t is_meta = DMU_OT_IS_METADATA(drrw->drr_type);
2843                 if (DRR_WRITE_COMPRESSED(drrw)) {
2844                         ASSERT3U(drrw->drr_compressed_size, >, 0);
2845                         ASSERT3U(drrw->drr_logical_size, >=,
2846                             drrw->drr_compressed_size);
2847                         ASSERT(!is_meta);
2848                         abuf = arc_loan_compressed_buf(
2849                             dmu_objset_spa(ra->os),
2850                             drrw->drr_compressed_size, drrw->drr_logical_size,
2851                             drrw->drr_compressiontype);
2852                 } else {
2853                         abuf = arc_loan_buf(dmu_objset_spa(ra->os),
2854                             is_meta, drrw->drr_logical_size);
2855                 }
2856
2857                 err = receive_read_payload_and_next_header(ra,
2858                     DRR_WRITE_PAYLOAD_SIZE(drrw), abuf->b_data);
2859                 if (err != 0) {
2860                         dmu_return_arcbuf(abuf);
2861                         return (err);
2862                 }
2863                 ra->rrd->write_buf = abuf;
2864                 receive_read_prefetch(ra, drrw->drr_object, drrw->drr_offset,
2865                     drrw->drr_logical_size);
2866                 return (err);
2867         }
2868         case DRR_WRITE_BYREF:
2869         {
2870                 struct drr_write_byref *drrwb =
2871                     &ra->rrd->header.drr_u.drr_write_byref;
2872                 err = receive_read_payload_and_next_header(ra, 0, NULL);
2873                 receive_read_prefetch(ra, drrwb->drr_object, drrwb->drr_offset,
2874                     drrwb->drr_length);
2875                 return (err);
2876         }
2877         case DRR_WRITE_EMBEDDED:
2878         {
2879                 struct drr_write_embedded *drrwe =
2880                     &ra->rrd->header.drr_u.drr_write_embedded;
2881                 uint32_t size = P2ROUNDUP(drrwe->drr_psize, 8);
2882                 void *buf = kmem_zalloc(size, KM_SLEEP);
2883
2884                 err = receive_read_payload_and_next_header(ra, size, buf);
2885                 if (err != 0) {
2886                         kmem_free(buf, size);
2887                         return (err);
2888                 }
2889
2890                 receive_read_prefetch(ra, drrwe->drr_object, drrwe->drr_offset,
2891                     drrwe->drr_length);
2892                 return (err);
2893         }
2894         case DRR_FREE:
2895         {
2896                 /*
2897                  * It might be beneficial to prefetch indirect blocks here, but
2898                  * we don't really have the data to decide for sure.
2899                  */
2900                 err = receive_read_payload_and_next_header(ra, 0, NULL);
2901                 return (err);
2902         }
2903         case DRR_END:
2904         {
2905                 struct drr_end *drre = &ra->rrd->header.drr_u.drr_end;
2906                 if (!ZIO_CHECKSUM_EQUAL(ra->prev_cksum, drre->drr_checksum))
2907                         return (SET_ERROR(ECKSUM));
2908                 return (0);
2909         }
2910         case DRR_SPILL:
2911         {
2912                 struct drr_spill *drrs = &ra->rrd->header.drr_u.drr_spill;
2913                 void *buf = kmem_zalloc(drrs->drr_length, KM_SLEEP);
2914                 err = receive_read_payload_and_next_header(ra, drrs->drr_length,
2915                     buf);
2916                 if (err != 0)
2917                         kmem_free(buf, drrs->drr_length);
2918                 return (err);
2919         }
2920         default:
2921                 return (SET_ERROR(EINVAL));
2922         }
2923 }
2924
2925 /*
2926  * Commit the records to the pool.
2927  */
2928 static int
2929 receive_process_record(struct receive_writer_arg *rwa,
2930     struct receive_record_arg *rrd)
2931 {
2932         int err;
2933
2934         /* Processing in order, therefore bytes_read should be increasing. */
2935         ASSERT3U(rrd->bytes_read, >=, rwa->bytes_read);
2936         rwa->bytes_read = rrd->bytes_read;
2937
2938         switch (rrd->header.drr_type) {
2939         case DRR_OBJECT:
2940         {
2941                 struct drr_object *drro = &rrd->header.drr_u.drr_object;
2942                 err = receive_object(rwa, drro, rrd->payload);
2943                 kmem_free(rrd->payload, rrd->payload_size);
2944                 rrd->payload = NULL;
2945                 return (err);
2946         }
2947         case DRR_FREEOBJECTS:
2948         {
2949                 struct drr_freeobjects *drrfo =
2950                     &rrd->header.drr_u.drr_freeobjects;
2951                 return (receive_freeobjects(rwa, drrfo));
2952         }
2953         case DRR_WRITE:
2954         {
2955                 struct drr_write *drrw = &rrd->header.drr_u.drr_write;
2956                 err = receive_write(rwa, drrw, rrd->write_buf);
2957                 /* if receive_write() is successful, it consumes the arc_buf */
2958                 if (err != 0)
2959                         dmu_return_arcbuf(rrd->write_buf);
2960                 rrd->write_buf = NULL;
2961                 rrd->payload = NULL;
2962                 return (err);
2963         }
2964         case DRR_WRITE_BYREF:
2965         {
2966                 struct drr_write_byref *drrwbr =
2967                     &rrd->header.drr_u.drr_write_byref;
2968                 return (receive_write_byref(rwa, drrwbr));
2969         }
2970         case DRR_WRITE_EMBEDDED:
2971         {
2972                 struct drr_write_embedded *drrwe =
2973                     &rrd->header.drr_u.drr_write_embedded;
2974                 err = receive_write_embedded(rwa, drrwe, rrd->payload);
2975                 kmem_free(rrd->payload, rrd->payload_size);
2976                 rrd->payload = NULL;
2977                 return (err);
2978         }
2979         case DRR_FREE:
2980         {
2981                 struct drr_free *drrf = &rrd->header.drr_u.drr_free;
2982                 return (receive_free(rwa, drrf));
2983         }
2984         case DRR_SPILL:
2985         {
2986                 struct drr_spill *drrs = &rrd->header.drr_u.drr_spill;
2987                 err = receive_spill(rwa, drrs, rrd->payload);
2988                 kmem_free(rrd->payload, rrd->payload_size);
2989                 rrd->payload = NULL;
2990                 return (err);
2991         }
2992         default:
2993                 return (SET_ERROR(EINVAL));
2994         }
2995 }
2996
2997 /*
2998  * dmu_recv_stream's worker thread; pull records off the queue, and then call
2999  * receive_process_record  When we're done, signal the main thread and exit.
3000  */
3001 static void
3002 receive_writer_thread(void *arg)
3003 {
3004         struct receive_writer_arg *rwa = arg;
3005         struct receive_record_arg *rrd;
3006         for (rrd = bqueue_dequeue(&rwa->q); !rrd->eos_marker;
3007             rrd = bqueue_dequeue(&rwa->q)) {
3008                 /*
3009                  * If there's an error, the main thread will stop putting things
3010                  * on the queue, but we need to clear everything in it before we
3011                  * can exit.
3012                  */
3013                 if (rwa->err == 0) {
3014                         rwa->err = receive_process_record(rwa, rrd);
3015                 } else if (rrd->write_buf != NULL) {
3016                         dmu_return_arcbuf(rrd->write_buf);
3017                         rrd->write_buf = NULL;
3018                         rrd->payload = NULL;
3019                 } else if (rrd->payload != NULL) {
3020                         kmem_free(rrd->payload, rrd->payload_size);
3021                         rrd->payload = NULL;
3022                 }
3023                 kmem_free(rrd, sizeof (*rrd));
3024         }
3025         kmem_free(rrd, sizeof (*rrd));
3026         mutex_enter(&rwa->mutex);
3027         rwa->done = B_TRUE;
3028         cv_signal(&rwa->cv);
3029         mutex_exit(&rwa->mutex);
3030         thread_exit();
3031 }
3032
3033 static int
3034 resume_check(struct receive_arg *ra, nvlist_t *begin_nvl)
3035 {
3036         uint64_t val;
3037         objset_t *mos = dmu_objset_pool(ra->os)->dp_meta_objset;
3038         uint64_t dsobj = dmu_objset_id(ra->os);
3039         uint64_t resume_obj, resume_off;
3040
3041         if (nvlist_lookup_uint64(begin_nvl,
3042             "resume_object", &resume_obj) != 0 ||
3043             nvlist_lookup_uint64(begin_nvl,
3044             "resume_offset", &resume_off) != 0) {
3045                 return (SET_ERROR(EINVAL));
3046         }
3047         VERIFY0(zap_lookup(mos, dsobj,
3048             DS_FIELD_RESUME_OBJECT, sizeof (val), 1, &val));
3049         if (resume_obj != val)
3050                 return (SET_ERROR(EINVAL));
3051         VERIFY0(zap_lookup(mos, dsobj,
3052             DS_FIELD_RESUME_OFFSET, sizeof (val), 1, &val));
3053         if (resume_off != val)
3054                 return (SET_ERROR(EINVAL));
3055
3056         return (0);
3057 }
3058
3059 /*
3060  * Read in the stream's records, one by one, and apply them to the pool.  There
3061  * are two threads involved; the thread that calls this function will spin up a
3062  * worker thread, read the records off the stream one by one, and issue
3063  * prefetches for any necessary indirect blocks.  It will then push the records
3064  * onto an internal blocking queue.  The worker thread will pull the records off
3065  * the queue, and actually write the data into the DMU.  This way, the worker
3066  * thread doesn't have to wait for reads to complete, since everything it needs
3067  * (the indirect blocks) will be prefetched.
3068  *
3069  * NB: callers *must* call dmu_recv_end() if this succeeds.
3070  */
3071 int
3072 dmu_recv_stream(dmu_recv_cookie_t *drc, struct file *fp, offset_t *voffp,
3073     int cleanup_fd, uint64_t *action_handlep)
3074 {
3075         int err = 0;
3076         struct receive_arg ra = { 0 };
3077         struct receive_writer_arg rwa = { 0 };
3078         int featureflags;
3079         nvlist_t *begin_nvl = NULL;
3080
3081         ra.byteswap = drc->drc_byteswap;
3082         ra.cksum = drc->drc_cksum;
3083         ra.td = curthread;
3084         ra.fp = fp;
3085         ra.voff = *voffp;
3086
3087         if (dsl_dataset_is_zapified(drc->drc_ds)) {
3088                 (void) zap_lookup(drc->drc_ds->ds_dir->dd_pool->dp_meta_objset,
3089                     drc->drc_ds->ds_object, DS_FIELD_RESUME_BYTES,
3090                     sizeof (ra.bytes_read), 1, &ra.bytes_read);
3091         }
3092
3093         objlist_create(&ra.ignore_objlist);
3094
3095         /* these were verified in dmu_recv_begin */
3096         ASSERT3U(DMU_GET_STREAM_HDRTYPE(drc->drc_drrb->drr_versioninfo), ==,
3097             DMU_SUBSTREAM);
3098         ASSERT3U(drc->drc_drrb->drr_type, <, DMU_OST_NUMTYPES);
3099
3100         /*
3101          * Open the objset we are modifying.
3102          */
3103         VERIFY0(dmu_objset_from_ds(drc->drc_ds, &ra.os));
3104
3105         ASSERT(dsl_dataset_phys(drc->drc_ds)->ds_flags & DS_FLAG_INCONSISTENT);
3106
3107         featureflags = DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo);
3108
3109         /* if this stream is dedup'ed, set up the avl tree for guid mapping */
3110         if (featureflags & DMU_BACKUP_FEATURE_DEDUP) {
3111                 minor_t minor;
3112
3113                 if (cleanup_fd == -1) {
3114                         ra.err = SET_ERROR(EBADF);
3115                         goto out;
3116                 }
3117                 ra.err = zfs_onexit_fd_hold(cleanup_fd, &minor);
3118                 if (ra.err != 0) {
3119                         cleanup_fd = -1;
3120                         goto out;
3121                 }
3122
3123                 if (*action_handlep == 0) {
3124                         rwa.guid_to_ds_map =
3125                             kmem_alloc(sizeof (avl_tree_t), KM_SLEEP);
3126                         avl_create(rwa.guid_to_ds_map, guid_compare,
3127                             sizeof (guid_map_entry_t),
3128                             offsetof(guid_map_entry_t, avlnode));
3129                         err = zfs_onexit_add_cb(minor,
3130                             free_guid_map_onexit, rwa.guid_to_ds_map,
3131                             action_handlep);
3132                         if (ra.err != 0)
3133                                 goto out;
3134                 } else {
3135                         err = zfs_onexit_cb_data(minor, *action_handlep,
3136                             (void **)&rwa.guid_to_ds_map);
3137                         if (ra.err != 0)
3138                                 goto out;
3139                 }
3140
3141                 drc->drc_guid_to_ds_map = rwa.guid_to_ds_map;
3142         }
3143
3144         uint32_t payloadlen = drc->drc_drr_begin->drr_payloadlen;
3145         void *payload = NULL;
3146         if (payloadlen != 0)
3147                 payload = kmem_alloc(payloadlen, KM_SLEEP);
3148
3149         err = receive_read_payload_and_next_header(&ra, payloadlen, payload);
3150         if (err != 0) {
3151                 if (payloadlen != 0)
3152                         kmem_free(payload, payloadlen);
3153                 goto out;
3154         }
3155         if (payloadlen != 0) {
3156                 err = nvlist_unpack(payload, payloadlen, &begin_nvl, KM_SLEEP);
3157                 kmem_free(payload, payloadlen);
3158                 if (err != 0)
3159                         goto out;
3160         }
3161
3162         if (featureflags & DMU_BACKUP_FEATURE_RESUMING) {
3163                 err = resume_check(&ra, begin_nvl);
3164                 if (err != 0)
3165                         goto out;
3166         }
3167
3168         (void) bqueue_init(&rwa.q, zfs_recv_queue_length,
3169             offsetof(struct receive_record_arg, node));
3170         cv_init(&rwa.cv, NULL, CV_DEFAULT, NULL);
3171         mutex_init(&rwa.mutex, NULL, MUTEX_DEFAULT, NULL);
3172         rwa.os = ra.os;
3173         rwa.byteswap = drc->drc_byteswap;
3174         rwa.resumable = drc->drc_resumable;
3175
3176         (void) thread_create(NULL, 0, receive_writer_thread, &rwa, 0, &p0,
3177             TS_RUN, minclsyspri);
3178         /*
3179          * We're reading rwa.err without locks, which is safe since we are the
3180          * only reader, and the worker thread is the only writer.  It's ok if we
3181          * miss a write for an iteration or two of the loop, since the writer
3182          * thread will keep freeing records we send it until we send it an eos
3183          * marker.
3184          *
3185          * We can leave this loop in 3 ways:  First, if rwa.err is
3186          * non-zero.  In that case, the writer thread will free the rrd we just
3187          * pushed.  Second, if  we're interrupted; in that case, either it's the
3188          * first loop and ra.rrd was never allocated, or it's later, and ra.rrd
3189          * has been handed off to the writer thread who will free it.  Finally,
3190          * if receive_read_record fails or we're at the end of the stream, then
3191          * we free ra.rrd and exit.
3192          */
3193         while (rwa.err == 0) {
3194                 if (issig(JUSTLOOKING) && issig(FORREAL)) {
3195                         err = SET_ERROR(EINTR);
3196                         break;
3197                 }
3198
3199                 ASSERT3P(ra.rrd, ==, NULL);
3200                 ra.rrd = ra.next_rrd;
3201                 ra.next_rrd = NULL;
3202                 /* Allocates and loads header into ra.next_rrd */
3203                 err = receive_read_record(&ra);
3204
3205                 if (ra.rrd->header.drr_type == DRR_END || err != 0) {
3206                         kmem_free(ra.rrd, sizeof (*ra.rrd));
3207                         ra.rrd = NULL;
3208                         break;
3209                 }
3210
3211                 bqueue_enqueue(&rwa.q, ra.rrd,
3212                     sizeof (struct receive_record_arg) + ra.rrd->payload_size);
3213                 ra.rrd = NULL;
3214         }
3215         if (ra.next_rrd == NULL)
3216                 ra.next_rrd = kmem_zalloc(sizeof (*ra.next_rrd), KM_SLEEP);
3217         ra.next_rrd->eos_marker = B_TRUE;
3218         bqueue_enqueue(&rwa.q, ra.next_rrd, 1);
3219
3220         mutex_enter(&rwa.mutex);
3221         while (!rwa.done) {
3222                 cv_wait(&rwa.cv, &rwa.mutex);
3223         }
3224         mutex_exit(&rwa.mutex);
3225
3226         /*
3227          * If we are receiving a full stream as a clone, all object IDs which
3228          * are greater than the maximum ID referenced in the stream are
3229          * by definition unused and must be freed. Note that it's possible that
3230          * we've resumed this send and the first record we received was the END
3231          * record. In that case, max_object would be 0, but we shouldn't start
3232          * freeing all objects from there; instead we should start from the
3233          * resumeobj.
3234          */
3235         if (drc->drc_clone && drc->drc_drrb->drr_fromguid == 0) {
3236                 uint64_t obj;
3237                 if (nvlist_lookup_uint64(begin_nvl, "resume_object", &obj) != 0)
3238                         obj = 0;
3239                 if (rwa.max_object > obj)
3240                         obj = rwa.max_object;
3241                 obj++;
3242                 int free_err = 0;
3243                 int next_err = 0;
3244
3245                 while (next_err == 0) {
3246                         free_err = dmu_free_long_object(rwa.os, obj);
3247                         if (free_err != 0 && free_err != ENOENT)
3248                                 break;
3249
3250                         next_err = dmu_object_next(rwa.os, &obj, FALSE, 0);
3251                 }
3252
3253                 if (err == 0) {
3254                         if (free_err != 0 && free_err != ENOENT)
3255                                 err = free_err;
3256                         else if (next_err != ESRCH)
3257                                 err = next_err;
3258                 }
3259         }
3260
3261         cv_destroy(&rwa.cv);
3262         mutex_destroy(&rwa.mutex);
3263         bqueue_destroy(&rwa.q);
3264         if (err == 0)
3265                 err = rwa.err;
3266
3267 out:
3268         nvlist_free(begin_nvl);
3269         if ((featureflags & DMU_BACKUP_FEATURE_DEDUP) && (cleanup_fd != -1))
3270                 zfs_onexit_fd_rele(cleanup_fd);
3271
3272         if (err != 0) {
3273                 /*
3274                  * Clean up references. If receive is not resumable,
3275                  * destroy what we created, so we don't leave it in
3276                  * the inconsistent state.
3277                  */
3278                 dmu_recv_cleanup_ds(drc);
3279         }
3280
3281         *voffp = ra.voff;
3282         objlist_destroy(&ra.ignore_objlist);
3283         return (err);
3284 }
3285
3286 static int
3287 dmu_recv_end_check(void *arg, dmu_tx_t *tx)
3288 {
3289         dmu_recv_cookie_t *drc = arg;
3290         dsl_pool_t *dp = dmu_tx_pool(tx);
3291         int error;
3292
3293         ASSERT3P(drc->drc_ds->ds_owner, ==, dmu_recv_tag);
3294
3295         if (!drc->drc_newfs) {
3296                 dsl_dataset_t *origin_head;
3297
3298                 error = dsl_dataset_hold(dp, drc->drc_tofs, FTAG, &origin_head);
3299                 if (error != 0)
3300                         return (error);
3301                 if (drc->drc_force) {
3302                         /*
3303                          * We will destroy any snapshots in tofs (i.e. before
3304                          * origin_head) that are after the origin (which is
3305                          * the snap before drc_ds, because drc_ds can not
3306                          * have any snaps of its own).
3307                          */
3308                         uint64_t obj;
3309
3310                         obj = dsl_dataset_phys(origin_head)->ds_prev_snap_obj;
3311                         while (obj !=
3312                             dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj) {
3313                                 dsl_dataset_t *snap;
3314                                 error = dsl_dataset_hold_obj(dp, obj, FTAG,
3315                                     &snap);
3316                                 if (error != 0)
3317                                         break;
3318                                 if (snap->ds_dir != origin_head->ds_dir)
3319                                         error = SET_ERROR(EINVAL);
3320                                 if (error == 0)  {
3321                                         error = dsl_destroy_snapshot_check_impl(
3322                                             snap, B_FALSE);
3323                                 }
3324                                 obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
3325                                 dsl_dataset_rele(snap, FTAG);
3326                                 if (error != 0)
3327                                         break;
3328                         }
3329                         if (error != 0) {
3330                                 dsl_dataset_rele(origin_head, FTAG);
3331                                 return (error);
3332                         }
3333                 }
3334                 error = dsl_dataset_clone_swap_check_impl(drc->drc_ds,
3335                     origin_head, drc->drc_force, drc->drc_owner, tx);
3336                 if (error != 0) {
3337                         dsl_dataset_rele(origin_head, FTAG);
3338                         return (error);
3339                 }
3340                 error = dsl_dataset_snapshot_check_impl(origin_head,
3341                     drc->drc_tosnap, tx, B_TRUE, 1, drc->drc_cred);
3342                 dsl_dataset_rele(origin_head, FTAG);
3343                 if (error != 0)
3344                         return (error);
3345
3346                 error = dsl_destroy_head_check_impl(drc->drc_ds, 1);
3347         } else {
3348                 error = dsl_dataset_snapshot_check_impl(drc->drc_ds,
3349                     drc->drc_tosnap, tx, B_TRUE, 1, drc->drc_cred);
3350         }
3351         return (error);
3352 }
3353
3354 static void
3355 dmu_recv_end_sync(void *arg, dmu_tx_t *tx)
3356 {
3357         dmu_recv_cookie_t *drc = arg;
3358         dsl_pool_t *dp = dmu_tx_pool(tx);
3359
3360         spa_history_log_internal_ds(drc->drc_ds, "finish receiving",
3361             tx, "snap=%s", drc->drc_tosnap);
3362
3363         if (!drc->drc_newfs) {
3364                 dsl_dataset_t *origin_head;
3365
3366                 VERIFY0(dsl_dataset_hold(dp, drc->drc_tofs, FTAG,
3367                     &origin_head));
3368
3369                 if (drc->drc_force) {
3370                         /*
3371                          * Destroy any snapshots of drc_tofs (origin_head)
3372                          * after the origin (the snap before drc_ds).
3373                          */
3374                         uint64_t obj;
3375
3376                         obj = dsl_dataset_phys(origin_head)->ds_prev_snap_obj;
3377                         while (obj !=
3378                             dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj) {
3379                                 dsl_dataset_t *snap;
3380                                 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG,
3381                                     &snap));
3382                                 ASSERT3P(snap->ds_dir, ==, origin_head->ds_dir);
3383                                 obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
3384                                 dsl_destroy_snapshot_sync_impl(snap,
3385                                     B_FALSE, tx);
3386                                 dsl_dataset_rele(snap, FTAG);
3387                         }
3388                 }
3389                 VERIFY3P(drc->drc_ds->ds_prev, ==,
3390                     origin_head->ds_prev);
3391
3392                 dsl_dataset_clone_swap_sync_impl(drc->drc_ds,
3393                     origin_head, tx);
3394                 dsl_dataset_snapshot_sync_impl(origin_head,
3395                     drc->drc_tosnap, tx);
3396
3397                 /* set snapshot's creation time and guid */
3398                 dmu_buf_will_dirty(origin_head->ds_prev->ds_dbuf, tx);
3399                 dsl_dataset_phys(origin_head->ds_prev)->ds_creation_time =
3400                     drc->drc_drrb->drr_creation_time;
3401                 dsl_dataset_phys(origin_head->ds_prev)->ds_guid =
3402                     drc->drc_drrb->drr_toguid;
3403                 dsl_dataset_phys(origin_head->ds_prev)->ds_flags &=
3404                     ~DS_FLAG_INCONSISTENT;
3405
3406                 dmu_buf_will_dirty(origin_head->ds_dbuf, tx);
3407                 dsl_dataset_phys(origin_head)->ds_flags &=
3408                     ~DS_FLAG_INCONSISTENT;
3409
3410                 drc->drc_newsnapobj =
3411                     dsl_dataset_phys(origin_head)->ds_prev_snap_obj;
3412
3413                 dsl_dataset_rele(origin_head, FTAG);
3414                 dsl_destroy_head_sync_impl(drc->drc_ds, tx);
3415
3416                 if (drc->drc_owner != NULL)
3417                         VERIFY3P(origin_head->ds_owner, ==, drc->drc_owner);
3418         } else {
3419                 dsl_dataset_t *ds = drc->drc_ds;
3420
3421                 dsl_dataset_snapshot_sync_impl(ds, drc->drc_tosnap, tx);
3422
3423                 /* set snapshot's creation time and guid */
3424                 dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
3425                 dsl_dataset_phys(ds->ds_prev)->ds_creation_time =
3426                     drc->drc_drrb->drr_creation_time;
3427                 dsl_dataset_phys(ds->ds_prev)->ds_guid =
3428                     drc->drc_drrb->drr_toguid;
3429                 dsl_dataset_phys(ds->ds_prev)->ds_flags &=
3430                     ~DS_FLAG_INCONSISTENT;
3431
3432                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
3433                 dsl_dataset_phys(ds)->ds_flags &= ~DS_FLAG_INCONSISTENT;
3434                 if (dsl_dataset_has_resume_receive_state(ds)) {
3435                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3436                             DS_FIELD_RESUME_FROMGUID, tx);
3437                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3438                             DS_FIELD_RESUME_OBJECT, tx);
3439                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3440                             DS_FIELD_RESUME_OFFSET, tx);
3441                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3442                             DS_FIELD_RESUME_BYTES, tx);
3443                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3444                             DS_FIELD_RESUME_TOGUID, tx);
3445                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3446                             DS_FIELD_RESUME_TONAME, tx);
3447                 }
3448                 drc->drc_newsnapobj =
3449                     dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj;
3450         }
3451
3452 #if defined(__FreeBSD__) && defined(_KERNEL)
3453         zvol_create_minors(dp->dp_spa, drc->drc_tofs);
3454 #endif
3455
3456         /*
3457          * Release the hold from dmu_recv_begin.  This must be done before
3458          * we return to open context, so that when we free the dataset's dnode,
3459          * we can evict its bonus buffer.
3460          */
3461         dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
3462         drc->drc_ds = NULL;
3463 }
3464
3465 static int
3466 add_ds_to_guidmap(const char *name, avl_tree_t *guid_map, uint64_t snapobj)
3467 {
3468         dsl_pool_t *dp;
3469         dsl_dataset_t *snapds;
3470         guid_map_entry_t *gmep;
3471         int err;
3472
3473         ASSERT(guid_map != NULL);
3474
3475         err = dsl_pool_hold(name, FTAG, &dp);
3476         if (err != 0)
3477                 return (err);
3478         gmep = kmem_alloc(sizeof (*gmep), KM_SLEEP);
3479         err = dsl_dataset_hold_obj(dp, snapobj, gmep, &snapds);
3480         if (err == 0) {
3481                 gmep->guid = dsl_dataset_phys(snapds)->ds_guid;
3482                 gmep->gme_ds = snapds;
3483                 avl_add(guid_map, gmep);
3484                 dsl_dataset_long_hold(snapds, gmep);
3485         } else
3486                 kmem_free(gmep, sizeof (*gmep));
3487
3488         dsl_pool_rele(dp, FTAG);
3489         return (err);
3490 }
3491
3492 static int dmu_recv_end_modified_blocks = 3;
3493
3494 static int
3495 dmu_recv_existing_end(dmu_recv_cookie_t *drc)
3496 {
3497 #ifdef _KERNEL
3498         /*
3499          * We will be destroying the ds; make sure its origin is unmounted if
3500          * necessary.
3501          */
3502         char name[ZFS_MAX_DATASET_NAME_LEN];
3503         dsl_dataset_name(drc->drc_ds, name);
3504         zfs_destroy_unmount_origin(name);
3505 #endif
3506
3507         return (dsl_sync_task(drc->drc_tofs,
3508             dmu_recv_end_check, dmu_recv_end_sync, drc,
3509             dmu_recv_end_modified_blocks, ZFS_SPACE_CHECK_NORMAL));
3510 }
3511
3512 static int
3513 dmu_recv_new_end(dmu_recv_cookie_t *drc)
3514 {
3515         return (dsl_sync_task(drc->drc_tofs,
3516             dmu_recv_end_check, dmu_recv_end_sync, drc,
3517             dmu_recv_end_modified_blocks, ZFS_SPACE_CHECK_NORMAL));
3518 }
3519
3520 int
3521 dmu_recv_end(dmu_recv_cookie_t *drc, void *owner)
3522 {
3523         int error;
3524
3525         drc->drc_owner = owner;
3526
3527         if (drc->drc_newfs)
3528                 error = dmu_recv_new_end(drc);
3529         else
3530                 error = dmu_recv_existing_end(drc);
3531
3532         if (error != 0) {
3533                 dmu_recv_cleanup_ds(drc);
3534         } else if (drc->drc_guid_to_ds_map != NULL) {
3535                 (void) add_ds_to_guidmap(drc->drc_tofs,
3536                     drc->drc_guid_to_ds_map,
3537                     drc->drc_newsnapobj);
3538         }
3539         return (error);
3540 }
3541
3542 /*
3543  * Return TRUE if this objset is currently being received into.
3544  */
3545 boolean_t
3546 dmu_objset_is_receiving(objset_t *os)
3547 {
3548         return (os->os_dsl_dataset != NULL &&
3549             os->os_dsl_dataset->ds_owner == dmu_recv_tag);
3550 }