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  */
  31
  32 #include <sys/dmu.h>
  33 #include <sys/dmu_impl.h>
  34 #include <sys/dmu_tx.h>
  35 #include <sys/dbuf.h>
  36 #include <sys/dnode.h>
  37 #include <sys/zfs_context.h>
  38 #include <sys/dmu_objset.h>
  39 #include <sys/dmu_traverse.h>
  40 #include <sys/dsl_dataset.h>
  41 #include <sys/dsl_dir.h>
  42 #include <sys/dsl_prop.h>
  43 #include <sys/dsl_pool.h>
  44 #include <sys/dsl_synctask.h>
  45 #include <sys/zfs_ioctl.h>
  46 #include <sys/zap.h>
  47 #include <sys/zio_checksum.h>
  48 #include <sys/zfs_znode.h>
  49 #include <zfs_fletcher.h>
  50 #include <sys/avl.h>
  51 #include <sys/ddt.h>
  52 #include <sys/zfs_onexit.h>
  53 #include <sys/dmu_send.h>
  54 #include <sys/dsl_destroy.h>
  55 #include <sys/blkptr.h>
  56 #include <sys/dsl_bookmark.h>
  57 #include <sys/zfeature.h>
  58 #include <sys/bqueue.h>
  59
  60 #ifdef __FreeBSD__
  61 #undef dump_write
  62 #define dump_write dmu_dump_write
  63 #endif
  64
  65 /* Set this tunable to TRUE to replace corrupt data with 0x2f5baddb10c */
  66 int zfs_send_corrupt_data = B_FALSE;
  67 int zfs_send_queue_length = 16 * 1024 * 1024;
  68 int zfs_recv_queue_length = 16 * 1024 * 1024;
  69 /* Set this tunable to FALSE to disable setting of DRR_FLAG_FREERECORDS */
  70 int zfs_send_set_freerecords_bit = B_TRUE;
  71
  72 #ifdef _KERNEL
  73 TUNABLE_INT("vfs.zfs.send_set_freerecords_bit", &zfs_send_set_freerecords_bit);
  74 #endif
  75
  76 static char *dmu_recv_tag = "dmu_recv_tag";
  77 const char *recv_clone_name = "%recv";
  78
  79 #define BP_SPAN(datablkszsec, indblkshift, level) \
  80         (((uint64_t)datablkszsec) << (SPA_MINBLOCKSHIFT + \
  81         (level) * (indblkshift - SPA_BLKPTRSHIFT)))
  82
  83 static void byteswap_record(dmu_replay_record_t *drr);
  84
  85 struct send_thread_arg {
  86         bqueue_t        q;
  87         dsl_dataset_t   *ds;            /* Dataset to traverse */
  88         uint64_t        fromtxg;        /* Traverse from this txg */
  89         int             flags;          /* flags to pass to traverse_dataset */
  90         int             error_code;
  91         boolean_t       cancel;
  92         zbookmark_phys_t resume;
  93 };
  94
  95 struct send_block_record {
  96         boolean_t               eos_marker; /* Marks the end of the stream */
  97         blkptr_t                bp;
  98         zbookmark_phys_t        zb;
  99         uint8_t                 indblkshift;
 100         uint16_t                datablkszsec;
 101         bqueue_node_t           ln;
 102 };
 103
 104 static int
 105 dump_bytes(dmu_sendarg_t *dsp, void *buf, int len)
 106 {
 107         dsl_dataset_t *ds = dmu_objset_ds(dsp->dsa_os);
 108         struct uio auio;
 109         struct iovec aiov;
 110
 111         /*
 112          * The code does not rely on this (len being a multiple of 8).  We keep
 113          * this assertion because of the corresponding assertion in
 114          * receive_read().  Keeping this assertion ensures that we do not
 115          * inadvertently break backwards compatibility (causing the assertion
 116          * in receive_read() to trigger on old software).
 117          *
 118          * Removing the assertions could be rolled into a new feature that uses
 119          * data that isn't 8-byte aligned; if the assertions were removed, a
 120          * feature flag would have to be added.
 121          */
 122
 123         ASSERT0(len % 8);
 124
 125         aiov.iov_base = buf;
 126         aiov.iov_len = len;
 127         auio.uio_iov = &aiov;
 128         auio.uio_iovcnt = 1;
 129         auio.uio_resid = len;
 130         auio.uio_segflg = UIO_SYSSPACE;
 131         auio.uio_rw = UIO_WRITE;
 132         auio.uio_offset = (off_t)-1;
 133         auio.uio_td = dsp->dsa_td;
 134 #ifdef _KERNEL
 135         if (dsp->dsa_fp->f_type == DTYPE_VNODE)
 136                 bwillwrite();
 137         dsp->dsa_err = fo_write(dsp->dsa_fp, &auio, dsp->dsa_td->td_ucred, 0,
 138             dsp->dsa_td);
 139 #else
 140         fprintf(stderr, "%s: returning EOPNOTSUPP\n", __func__);
 141         dsp->dsa_err = EOPNOTSUPP;
 142 #endif
 143         mutex_enter(&ds->ds_sendstream_lock);
 144         *dsp->dsa_off += len;
 145         mutex_exit(&ds->ds_sendstream_lock);
 146
 147         return (dsp->dsa_err);
 148 }
 149
 150 /*
 151  * For all record types except BEGIN, fill in the checksum (overlaid in
 152  * drr_u.drr_checksum.drr_checksum).  The checksum verifies everything
 153  * up to the start of the checksum itself.
 154  */
 155 static int
 156 dump_record(dmu_sendarg_t *dsp, void *payload, int payload_len)
 157 {
 158         ASSERT3U(offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
 159             ==, sizeof (dmu_replay_record_t) - sizeof (zio_cksum_t));
 160         fletcher_4_incremental_native(dsp->dsa_drr,
 161             offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
 162             &dsp->dsa_zc);
 163         if (dsp->dsa_drr->drr_type == DRR_BEGIN) {
 164                 dsp->dsa_sent_begin = B_TRUE;
 165         } else {
 166                 ASSERT(ZIO_CHECKSUM_IS_ZERO(&dsp->dsa_drr->drr_u.
 167                     drr_checksum.drr_checksum));
 168                 dsp->dsa_drr->drr_u.drr_checksum.drr_checksum = dsp->dsa_zc;
 169         }
 170         if (dsp->dsa_drr->drr_type == DRR_END) {
 171                 dsp->dsa_sent_end = B_TRUE;
 172         }
 173         fletcher_4_incremental_native(&dsp->dsa_drr->
 174             drr_u.drr_checksum.drr_checksum,
 175             sizeof (zio_cksum_t), &dsp->dsa_zc);
 176         if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)) != 0)
 177                 return (SET_ERROR(EINTR));
 178         if (payload_len != 0) {
 179                 fletcher_4_incremental_native(payload, payload_len,
 180                     &dsp->dsa_zc);
 181                 if (dump_bytes(dsp, payload, payload_len) != 0)
 182                         return (SET_ERROR(EINTR));
 183         }
 184         return (0);
 185 }
 186
 187 /*
 188  * Fill in the drr_free struct, or perform aggregation if the previous record is
 189  * also a free record, and the two are adjacent.
 190  *
 191  * Note that we send free records even for a full send, because we want to be
 192  * able to receive a full send as a clone, which requires a list of all the free
 193  * and freeobject records that were generated on the source.
 194  */
 195 static int
 196 dump_free(dmu_sendarg_t *dsp, uint64_t object, uint64_t offset,
 197     uint64_t length)
 198 {
 199         struct drr_free *drrf = &(dsp->dsa_drr->drr_u.drr_free);
 200
 201         /*
 202          * When we receive a free record, dbuf_free_range() assumes
 203          * that the receiving system doesn't have any dbufs in the range
 204          * being freed.  This is always true because there is a one-record
 205          * constraint: we only send one WRITE record for any given
 206          * object,offset.  We know that the one-record constraint is
 207          * true because we always send data in increasing order by
 208          * object,offset.
 209          *
 210          * If the increasing-order constraint ever changes, we should find
 211          * another way to assert that the one-record constraint is still
 212          * satisfied.
 213          */
 214         ASSERT(object > dsp->dsa_last_data_object ||
 215             (object == dsp->dsa_last_data_object &&
 216             offset > dsp->dsa_last_data_offset));
 217
 218         if (length != -1ULL && offset + length < offset)
 219                 length = -1ULL;
 220
 221         /*
 222          * If there is a pending op, but it's not PENDING_FREE, push it out,
 223          * since free block aggregation can only be done for blocks of the
 224          * same type (i.e., DRR_FREE records can only be aggregated with
 225          * other DRR_FREE records.  DRR_FREEOBJECTS records can only be
 226          * aggregated with other DRR_FREEOBJECTS records.
 227          */
 228         if (dsp->dsa_pending_op != PENDING_NONE &&
 229             dsp->dsa_pending_op != PENDING_FREE) {
 230                 if (dump_record(dsp, NULL, 0) != 0)
 231                         return (SET_ERROR(EINTR));
 232                 dsp->dsa_pending_op = PENDING_NONE;
 233         }
 234
 235         if (dsp->dsa_pending_op == PENDING_FREE) {
 236                 /*
 237                  * There should never be a PENDING_FREE if length is -1
 238                  * (because dump_dnode is the only place where this
 239                  * function is called with a -1, and only after flushing
 240                  * any pending record).
 241                  */
 242                 ASSERT(length != -1ULL);
 243                 /*
 244                  * Check to see whether this free block can be aggregated
 245                  * with pending one.
 246                  */
 247                 if (drrf->drr_object == object && drrf->drr_offset +
 248                     drrf->drr_length == offset) {
 249                         drrf->drr_length += length;
 250                         return (0);
 251                 } else {
 252                         /* not a continuation.  Push out pending record */
 253                         if (dump_record(dsp, NULL, 0) != 0)
 254                                 return (SET_ERROR(EINTR));
 255                         dsp->dsa_pending_op = PENDING_NONE;
 256                 }
 257         }
 258         /* create a FREE record and make it pending */
 259         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 260         dsp->dsa_drr->drr_type = DRR_FREE;
 261         drrf->drr_object = object;
 262         drrf->drr_offset = offset;
 263         drrf->drr_length = length;
 264         drrf->drr_toguid = dsp->dsa_toguid;
 265         if (length == -1ULL) {
 266                 if (dump_record(dsp, NULL, 0) != 0)
 267                         return (SET_ERROR(EINTR));
 268         } else {
 269                 dsp->dsa_pending_op = PENDING_FREE;
 270         }
 271
 272         return (0);
 273 }
 274
 275 static int
 276 dump_write(dmu_sendarg_t *dsp, dmu_object_type_t type,
 277     uint64_t object, uint64_t offset, int blksz, const blkptr_t *bp, void *data)
 278 {
 279         struct drr_write *drrw = &(dsp->dsa_drr->drr_u.drr_write);
 280
 281         /*
 282          * We send data in increasing object, offset order.
 283          * See comment in dump_free() for details.
 284          */
 285         ASSERT(object > dsp->dsa_last_data_object ||
 286             (object == dsp->dsa_last_data_object &&
 287             offset > dsp->dsa_last_data_offset));
 288         dsp->dsa_last_data_object = object;
 289         dsp->dsa_last_data_offset = offset + blksz - 1;
 290
 291         /*
 292          * If there is any kind of pending aggregation (currently either
 293          * a grouping of free objects or free blocks), push it out to
 294          * the stream, since aggregation can't be done across operations
 295          * of different types.
 296          */
 297         if (dsp->dsa_pending_op != PENDING_NONE) {
 298                 if (dump_record(dsp, NULL, 0) != 0)
 299                         return (SET_ERROR(EINTR));
 300                 dsp->dsa_pending_op = PENDING_NONE;
 301         }
 302         /* write a WRITE record */
 303         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 304         dsp->dsa_drr->drr_type = DRR_WRITE;
 305         drrw->drr_object = object;
 306         drrw->drr_type = type;
 307         drrw->drr_offset = offset;
 308         drrw->drr_length = blksz;
 309         drrw->drr_toguid = dsp->dsa_toguid;
 310         if (bp == NULL || BP_IS_EMBEDDED(bp)) {
 311                 /*
 312                  * There's no pre-computed checksum for partial-block
 313                  * writes or embedded BP's, so (like
 314                  * fletcher4-checkummed blocks) userland will have to
 315                  * compute a dedup-capable checksum itself.
 316                  */
 317                 drrw->drr_checksumtype = ZIO_CHECKSUM_OFF;
 318         } else {
 319                 drrw->drr_checksumtype = BP_GET_CHECKSUM(bp);
 320                 if (zio_checksum_table[drrw->drr_checksumtype].ci_flags &
 321                     ZCHECKSUM_FLAG_DEDUP)
 322                         drrw->drr_checksumflags |= DRR_CHECKSUM_DEDUP;
 323                 DDK_SET_LSIZE(&drrw->drr_key, BP_GET_LSIZE(bp));
 324                 DDK_SET_PSIZE(&drrw->drr_key, BP_GET_PSIZE(bp));
 325                 DDK_SET_COMPRESS(&drrw->drr_key, BP_GET_COMPRESS(bp));
 326                 drrw->drr_key.ddk_cksum = bp->blk_cksum;
 327         }
 328
 329         if (dump_record(dsp, data, blksz) != 0)
 330                 return (SET_ERROR(EINTR));
 331         return (0);
 332 }
 333
 334 static int
 335 dump_write_embedded(dmu_sendarg_t *dsp, uint64_t object, uint64_t offset,
 336     int blksz, const blkptr_t *bp)
 337 {
 338         char buf[BPE_PAYLOAD_SIZE];
 339         struct drr_write_embedded *drrw =
 340             &(dsp->dsa_drr->drr_u.drr_write_embedded);
 341
 342         if (dsp->dsa_pending_op != PENDING_NONE) {
 343                 if (dump_record(dsp, NULL, 0) != 0)
 344                         return (EINTR);
 345                 dsp->dsa_pending_op = PENDING_NONE;
 346         }
 347
 348         ASSERT(BP_IS_EMBEDDED(bp));
 349
 350         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 351         dsp->dsa_drr->drr_type = DRR_WRITE_EMBEDDED;
 352         drrw->drr_object = object;
 353         drrw->drr_offset = offset;
 354         drrw->drr_length = blksz;
 355         drrw->drr_toguid = dsp->dsa_toguid;
 356         drrw->drr_compression = BP_GET_COMPRESS(bp);
 357         drrw->drr_etype = BPE_GET_ETYPE(bp);
 358         drrw->drr_lsize = BPE_GET_LSIZE(bp);
 359         drrw->drr_psize = BPE_GET_PSIZE(bp);
 360
 361         decode_embedded_bp_compressed(bp, buf);
 362
 363         if (dump_record(dsp, buf, P2ROUNDUP(drrw->drr_psize, 8)) != 0)
 364                 return (EINTR);
 365         return (0);
 366 }
 367
 368 static int
 369 dump_spill(dmu_sendarg_t *dsp, uint64_t object, int blksz, void *data)
 370 {
 371         struct drr_spill *drrs = &(dsp->dsa_drr->drr_u.drr_spill);
 372
 373         if (dsp->dsa_pending_op != PENDING_NONE) {
 374                 if (dump_record(dsp, NULL, 0) != 0)
 375                         return (SET_ERROR(EINTR));
 376                 dsp->dsa_pending_op = PENDING_NONE;
 377         }
 378
 379         /* write a SPILL record */
 380         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 381         dsp->dsa_drr->drr_type = DRR_SPILL;
 382         drrs->drr_object = object;
 383         drrs->drr_length = blksz;
 384         drrs->drr_toguid = dsp->dsa_toguid;
 385
 386         if (dump_record(dsp, data, blksz) != 0)
 387                 return (SET_ERROR(EINTR));
 388         return (0);
 389 }
 390
 391 static int
 392 dump_freeobjects(dmu_sendarg_t *dsp, uint64_t firstobj, uint64_t numobjs)
 393 {
 394         struct drr_freeobjects *drrfo = &(dsp->dsa_drr->drr_u.drr_freeobjects);
 395
 396         /*
 397          * If there is a pending op, but it's not PENDING_FREEOBJECTS,
 398          * push it out, since free block aggregation can only be done for
 399          * blocks of the same type (i.e., DRR_FREE records can only be
 400          * aggregated with other DRR_FREE records.  DRR_FREEOBJECTS records
 401          * can only be aggregated with other DRR_FREEOBJECTS records.
 402          */
 403         if (dsp->dsa_pending_op != PENDING_NONE &&
 404             dsp->dsa_pending_op != PENDING_FREEOBJECTS) {
 405                 if (dump_record(dsp, NULL, 0) != 0)
 406                         return (SET_ERROR(EINTR));
 407                 dsp->dsa_pending_op = PENDING_NONE;
 408         }
 409         if (dsp->dsa_pending_op == PENDING_FREEOBJECTS) {
 410                 /*
 411                  * See whether this free object array can be aggregated
 412                  * with pending one
 413                  */
 414                 if (drrfo->drr_firstobj + drrfo->drr_numobjs == firstobj) {
 415                         drrfo->drr_numobjs += numobjs;
 416                         return (0);
 417                 } else {
 418                         /* can't be aggregated.  Push out pending record */
 419                         if (dump_record(dsp, NULL, 0) != 0)
 420                                 return (SET_ERROR(EINTR));
 421                         dsp->dsa_pending_op = PENDING_NONE;
 422                 }
 423         }
 424
 425         /* write a FREEOBJECTS record */
 426         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 427         dsp->dsa_drr->drr_type = DRR_FREEOBJECTS;
 428         drrfo->drr_firstobj = firstobj;
 429         drrfo->drr_numobjs = numobjs;
 430         drrfo->drr_toguid = dsp->dsa_toguid;
 431
 432         dsp->dsa_pending_op = PENDING_FREEOBJECTS;
 433
 434         return (0);
 435 }
 436
 437 static int
 438 dump_dnode(dmu_sendarg_t *dsp, uint64_t object, dnode_phys_t *dnp)
 439 {
 440         struct drr_object *drro = &(dsp->dsa_drr->drr_u.drr_object);
 441
 442         if (object < dsp->dsa_resume_object) {
 443                 /*
 444                  * Note: when resuming, we will visit all the dnodes in
 445                  * the block of dnodes that we are resuming from.  In
 446                  * this case it's unnecessary to send the dnodes prior to
 447                  * the one we are resuming from.  We should be at most one
 448                  * block's worth of dnodes behind the resume point.
 449                  */
 450                 ASSERT3U(dsp->dsa_resume_object - object, <,
 451                     1 << (DNODE_BLOCK_SHIFT - DNODE_SHIFT));
 452                 return (0);
 453         }
 454
 455         if (dnp == NULL || dnp->dn_type == DMU_OT_NONE)
 456                 return (dump_freeobjects(dsp, object, 1));
 457
 458         if (dsp->dsa_pending_op != PENDING_NONE) {
 459                 if (dump_record(dsp, NULL, 0) != 0)
 460                         return (SET_ERROR(EINTR));
 461                 dsp->dsa_pending_op = PENDING_NONE;
 462         }
 463
 464         /* write an OBJECT record */
 465         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 466         dsp->dsa_drr->drr_type = DRR_OBJECT;
 467         drro->drr_object = object;
 468         drro->drr_type = dnp->dn_type;
 469         drro->drr_bonustype = dnp->dn_bonustype;
 470         drro->drr_blksz = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT;
 471         drro->drr_bonuslen = dnp->dn_bonuslen;
 472         drro->drr_checksumtype = dnp->dn_checksum;
 473         drro->drr_compress = dnp->dn_compress;
 474         drro->drr_toguid = dsp->dsa_toguid;
 475
 476         if (!(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
 477             drro->drr_blksz > SPA_OLD_MAXBLOCKSIZE)
 478                 drro->drr_blksz = SPA_OLD_MAXBLOCKSIZE;
 479
 480         if (dump_record(dsp, DN_BONUS(dnp),
 481             P2ROUNDUP(dnp->dn_bonuslen, 8)) != 0) {
 482                 return (SET_ERROR(EINTR));
 483         }
 484
 485         /* Free anything past the end of the file. */
 486         if (dump_free(dsp, object, (dnp->dn_maxblkid + 1) *
 487             (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT), -1ULL) != 0)
 488                 return (SET_ERROR(EINTR));
 489         if (dsp->dsa_err != 0)
 490                 return (SET_ERROR(EINTR));
 491         return (0);
 492 }
 493
 494 static boolean_t
 495 backup_do_embed(dmu_sendarg_t *dsp, const blkptr_t *bp)
 496 {
 497         if (!BP_IS_EMBEDDED(bp))
 498                 return (B_FALSE);
 499
 500         /*
 501          * Compression function must be legacy, or explicitly enabled.
 502          */
 503         if ((BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_LEGACY_FUNCTIONS &&
 504             !(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_EMBED_DATA_LZ4)))
 505                 return (B_FALSE);
 506
 507         /*
 508          * Embed type must be explicitly enabled.
 509          */
 510         switch (BPE_GET_ETYPE(bp)) {
 511         case BP_EMBEDDED_TYPE_DATA:
 512                 if (dsp->dsa_featureflags & DMU_BACKUP_FEATURE_EMBED_DATA)
 513                         return (B_TRUE);
 514                 break;
 515         default:
 516                 return (B_FALSE);
 517         }
 518         return (B_FALSE);
 519 }
 520
 521 /*
 522  * This is the callback function to traverse_dataset that acts as the worker
 523  * thread for dmu_send_impl.
 524  */
 525 /*ARGSUSED*/
 526 static int
 527 send_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
 528     const zbookmark_phys_t *zb, const struct dnode_phys *dnp, void *arg)
 529 {
 530         struct send_thread_arg *sta = arg;
 531         struct send_block_record *record;
 532         uint64_t record_size;
 533         int err = 0;
 534
 535         ASSERT(zb->zb_object == DMU_META_DNODE_OBJECT ||
 536             zb->zb_object >= sta->resume.zb_object);
 537
 538         if (sta->cancel)
 539                 return (SET_ERROR(EINTR));
 540
 541         if (bp == NULL) {
 542                 ASSERT3U(zb->zb_level, ==, ZB_DNODE_LEVEL);
 543                 return (0);
 544         } else if (zb->zb_level < 0) {
 545                 return (0);
 546         }
 547
 548         record = kmem_zalloc(sizeof (struct send_block_record), KM_SLEEP);
 549         record->eos_marker = B_FALSE;
 550         record->bp = *bp;
 551         record->zb = *zb;
 552         record->indblkshift = dnp->dn_indblkshift;
 553         record->datablkszsec = dnp->dn_datablkszsec;
 554         record_size = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT;
 555         bqueue_enqueue(&sta->q, record, record_size);
 556
 557         return (err);
 558 }
 559
 560 /*
 561  * This function kicks off the traverse_dataset.  It also handles setting the
 562  * error code of the thread in case something goes wrong, and pushes the End of
 563  * Stream record when the traverse_dataset call has finished.  If there is no
 564  * dataset to traverse, the thread immediately pushes End of Stream marker.
 565  */
 566 static void
 567 send_traverse_thread(void *arg)
 568 {
 569         struct send_thread_arg *st_arg = arg;
 570         int err;
 571         struct send_block_record *data;
 572
 573         if (st_arg->ds != NULL) {
 574                 err = traverse_dataset_resume(st_arg->ds,
 575                     st_arg->fromtxg, &st_arg->resume,
 576                     st_arg->flags, send_cb, st_arg);
 577
 578                 if (err != EINTR)
 579                         st_arg->error_code = err;
 580         }
 581         data = kmem_zalloc(sizeof (*data), KM_SLEEP);
 582         data->eos_marker = B_TRUE;
 583         bqueue_enqueue(&st_arg->q, data, 1);
 584         thread_exit();
 585 }
 586
 587 /*
 588  * This function actually handles figuring out what kind of record needs to be
 589  * dumped, reading the data (which has hopefully been prefetched), and calling
 590  * the appropriate helper function.
 591  */
 592 static int
 593 do_dump(dmu_sendarg_t *dsa, struct send_block_record *data)
 594 {
 595         dsl_dataset_t *ds = dmu_objset_ds(dsa->dsa_os);
 596         const blkptr_t *bp = &data->bp;
 597         const zbookmark_phys_t *zb = &data->zb;
 598         uint8_t indblkshift = data->indblkshift;
 599         uint16_t dblkszsec = data->datablkszsec;
 600         spa_t *spa = ds->ds_dir->dd_pool->dp_spa;
 601         dmu_object_type_t type = bp ? BP_GET_TYPE(bp) : DMU_OT_NONE;
 602         int err = 0;
 603
 604         ASSERT3U(zb->zb_level, >=, 0);
 605
 606         ASSERT(zb->zb_object == DMU_META_DNODE_OBJECT ||
 607             zb->zb_object >= dsa->dsa_resume_object);
 608
 609         if (zb->zb_object != DMU_META_DNODE_OBJECT &&
 610             DMU_OBJECT_IS_SPECIAL(zb->zb_object)) {
 611                 return (0);
 612         } else if (BP_IS_HOLE(bp) &&
 613             zb->zb_object == DMU_META_DNODE_OBJECT) {
 614                 uint64_t span = BP_SPAN(dblkszsec, indblkshift, zb->zb_level);
 615                 uint64_t dnobj = (zb->zb_blkid * span) >> DNODE_SHIFT;
 616                 err = dump_freeobjects(dsa, dnobj, span >> DNODE_SHIFT);
 617         } else if (BP_IS_HOLE(bp)) {
 618                 uint64_t span = BP_SPAN(dblkszsec, indblkshift, zb->zb_level);
 619                 uint64_t offset = zb->zb_blkid * span;
 620                 err = dump_free(dsa, zb->zb_object, offset, span);
 621         } else if (zb->zb_level > 0 || type == DMU_OT_OBJSET) {
 622                 return (0);
 623         } else if (type == DMU_OT_DNODE) {
 624                 int blksz = BP_GET_LSIZE(bp);
 625                 arc_flags_t aflags = ARC_FLAG_WAIT;
 626                 arc_buf_t *abuf;
 627
 628                 ASSERT0(zb->zb_level);
 629
 630                 if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
 631                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
 632                     &aflags, zb) != 0)
 633                         return (SET_ERROR(EIO));
 634
 635                 dnode_phys_t *blk = abuf->b_data;
 636                 uint64_t dnobj = zb->zb_blkid * (blksz >> DNODE_SHIFT);
 637                 for (int i = 0; i < blksz >> DNODE_SHIFT; i++) {
 638                         err = dump_dnode(dsa, dnobj + i, blk + i);
 639                         if (err != 0)
 640                                 break;
 641                 }
 642                 arc_buf_destroy(abuf, &abuf);
 643         } else if (type == DMU_OT_SA) {
 644                 arc_flags_t aflags = ARC_FLAG_WAIT;
 645                 arc_buf_t *abuf;
 646                 int blksz = BP_GET_LSIZE(bp);
 647
 648                 if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
 649                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
 650                     &aflags, zb) != 0)
 651                         return (SET_ERROR(EIO));
 652
 653                 err = dump_spill(dsa, zb->zb_object, blksz, abuf->b_data);
 654                 arc_buf_destroy(abuf, &abuf);
 655         } else if (backup_do_embed(dsa, bp)) {
 656                 /* it's an embedded level-0 block of a regular object */
 657                 int blksz = dblkszsec << SPA_MINBLOCKSHIFT;
 658                 ASSERT0(zb->zb_level);
 659                 err = dump_write_embedded(dsa, zb->zb_object,
 660                     zb->zb_blkid * blksz, blksz, bp);
 661         } else {
 662                 /* it's a level-0 block of a regular object */
 663                 arc_flags_t aflags = ARC_FLAG_WAIT;
 664                 arc_buf_t *abuf;
 665                 int blksz = dblkszsec << SPA_MINBLOCKSHIFT;
 666                 uint64_t offset;
 667
 668                 ASSERT0(zb->zb_level);
 669                 ASSERT(zb->zb_object > dsa->dsa_resume_object ||
 670                     (zb->zb_object == dsa->dsa_resume_object &&
 671                     zb->zb_blkid * blksz >= dsa->dsa_resume_offset));
 672
 673                 if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
 674                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
 675                     &aflags, zb) != 0) {
 676                         if (zfs_send_corrupt_data) {
 677                                 /* Send a block filled with 0x"zfs badd bloc" */
 678                                 abuf = arc_alloc_buf(spa, blksz, &abuf,
 679                                     ARC_BUFC_DATA);
 680                                 uint64_t *ptr;
 681                                 for (ptr = abuf->b_data;
 682                                     (char *)ptr < (char *)abuf->b_data + blksz;
 683                                     ptr++)
 684                                         *ptr = 0x2f5baddb10cULL;
 685                         } else {
 686                                 return (SET_ERROR(EIO));
 687                         }
 688                 }
 689
 690                 offset = zb->zb_blkid * blksz;
 691
 692                 if (!(dsa->dsa_featureflags &
 693                     DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
 694                     blksz > SPA_OLD_MAXBLOCKSIZE) {
 695                         char *buf = abuf->b_data;
 696                         while (blksz > 0 && err == 0) {
 697                                 int n = MIN(blksz, SPA_OLD_MAXBLOCKSIZE);
 698                                 err = dump_write(dsa, type, zb->zb_object,
 699                                     offset, n, NULL, buf);
 700                                 offset += n;
 701                                 buf += n;
 702                                 blksz -= n;
 703                         }
 704                 } else {
 705                         err = dump_write(dsa, type, zb->zb_object,
 706                             offset, blksz, bp, abuf->b_data);
 707                 }
 708                 arc_buf_destroy(abuf, &abuf);
 709         }
 710
 711         ASSERT(err == 0 || err == EINTR);
 712         return (err);
 713 }
 714
 715 /*
 716  * Pop the new data off the queue, and free the old data.
 717  */
 718 static struct send_block_record *
 719 get_next_record(bqueue_t *bq, struct send_block_record *data)
 720 {
 721         struct send_block_record *tmp = bqueue_dequeue(bq);
 722         kmem_free(data, sizeof (*data));
 723         return (tmp);
 724 }
 725
 726 /*
 727  * Actually do the bulk of the work in a zfs send.
 728  *
 729  * Note: Releases dp using the specified tag.
 730  */
 731 static int
 732 dmu_send_impl(void *tag, dsl_pool_t *dp, dsl_dataset_t *to_ds,
 733     zfs_bookmark_phys_t *ancestor_zb,
 734     boolean_t is_clone, boolean_t embedok, boolean_t large_block_ok, int outfd,
 735     uint64_t resumeobj, uint64_t resumeoff,
 736 #ifdef illumos
 737     vnode_t *vp, offset_t *off)
 738 #else
 739     struct file *fp, offset_t *off)
 740 #endif
 741 {
 742         objset_t *os;
 743         dmu_replay_record_t *drr;
 744         dmu_sendarg_t *dsp;
 745         int err;
 746         uint64_t fromtxg = 0;
 747         uint64_t featureflags = 0;
 748         struct send_thread_arg to_arg = { 0 };
 749
 750         err = dmu_objset_from_ds(to_ds, &os);
 751         if (err != 0) {
 752                 dsl_pool_rele(dp, tag);
 753                 return (err);
 754         }
 755
 756         drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP);
 757         drr->drr_type = DRR_BEGIN;
 758         drr->drr_u.drr_begin.drr_magic = DMU_BACKUP_MAGIC;
 759         DMU_SET_STREAM_HDRTYPE(drr->drr_u.drr_begin.drr_versioninfo,
 760             DMU_SUBSTREAM);
 761
 762 #ifdef _KERNEL
 763         if (dmu_objset_type(os) == DMU_OST_ZFS) {
 764                 uint64_t version;
 765                 if (zfs_get_zplprop(os, ZFS_PROP_VERSION, &version) != 0) {
 766                         kmem_free(drr, sizeof (dmu_replay_record_t));
 767                         dsl_pool_rele(dp, tag);
 768                         return (SET_ERROR(EINVAL));
 769                 }
 770                 if (version >= ZPL_VERSION_SA) {
 771                         featureflags |= DMU_BACKUP_FEATURE_SA_SPILL;
 772                 }
 773         }
 774 #endif
 775
 776         if (large_block_ok && to_ds->ds_feature_inuse[SPA_FEATURE_LARGE_BLOCKS])
 777                 featureflags |= DMU_BACKUP_FEATURE_LARGE_BLOCKS;
 778         if (embedok &&
 779             spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA)) {
 780                 featureflags |= DMU_BACKUP_FEATURE_EMBED_DATA;
 781                 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS))
 782                         featureflags |= DMU_BACKUP_FEATURE_EMBED_DATA_LZ4;
 783         }
 784
 785         if (resumeobj != 0 || resumeoff != 0) {
 786                 featureflags |= DMU_BACKUP_FEATURE_RESUMING;
 787         }
 788
 789         DMU_SET_FEATUREFLAGS(drr->drr_u.drr_begin.drr_versioninfo,
 790             featureflags);
 791
 792         drr->drr_u.drr_begin.drr_creation_time =
 793             dsl_dataset_phys(to_ds)->ds_creation_time;
 794         drr->drr_u.drr_begin.drr_type = dmu_objset_type(os);
 795         if (is_clone)
 796                 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CLONE;
 797         drr->drr_u.drr_begin.drr_toguid = dsl_dataset_phys(to_ds)->ds_guid;
 798         if (dsl_dataset_phys(to_ds)->ds_flags & DS_FLAG_CI_DATASET)
 799                 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CI_DATA;
 800         if (zfs_send_set_freerecords_bit)
 801                 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_FREERECORDS;
 802
 803         if (ancestor_zb != NULL) {
 804                 drr->drr_u.drr_begin.drr_fromguid =
 805                     ancestor_zb->zbm_guid;
 806                 fromtxg = ancestor_zb->zbm_creation_txg;
 807         }
 808         dsl_dataset_name(to_ds, drr->drr_u.drr_begin.drr_toname);
 809         if (!to_ds->ds_is_snapshot) {
 810                 (void) strlcat(drr->drr_u.drr_begin.drr_toname, "@--head--",
 811                     sizeof (drr->drr_u.drr_begin.drr_toname));
 812         }
 813
 814         dsp = kmem_zalloc(sizeof (dmu_sendarg_t), KM_SLEEP);
 815
 816         dsp->dsa_drr = drr;
 817         dsp->dsa_outfd = outfd;
 818         dsp->dsa_proc = curproc;
 819         dsp->dsa_td = curthread;
 820         dsp->dsa_fp = fp;
 821         dsp->dsa_os = os;
 822         dsp->dsa_off = off;
 823         dsp->dsa_toguid = dsl_dataset_phys(to_ds)->ds_guid;
 824         dsp->dsa_pending_op = PENDING_NONE;
 825         dsp->dsa_featureflags = featureflags;
 826         dsp->dsa_resume_object = resumeobj;
 827         dsp->dsa_resume_offset = resumeoff;
 828
 829         mutex_enter(&to_ds->ds_sendstream_lock);
 830         list_insert_head(&to_ds->ds_sendstreams, dsp);
 831         mutex_exit(&to_ds->ds_sendstream_lock);
 832
 833         dsl_dataset_long_hold(to_ds, FTAG);
 834         dsl_pool_rele(dp, tag);
 835
 836         void *payload = NULL;
 837         size_t payload_len = 0;
 838         if (resumeobj != 0 || resumeoff != 0) {
 839                 dmu_object_info_t to_doi;
 840                 err = dmu_object_info(os, resumeobj, &to_doi);
 841                 if (err != 0)
 842                         goto out;
 843                 SET_BOOKMARK(&to_arg.resume, to_ds->ds_object, resumeobj, 0,
 844                     resumeoff / to_doi.doi_data_block_size);
 845
 846                 nvlist_t *nvl = fnvlist_alloc();
 847                 fnvlist_add_uint64(nvl, "resume_object", resumeobj);
 848                 fnvlist_add_uint64(nvl, "resume_offset", resumeoff);
 849                 payload = fnvlist_pack(nvl, &payload_len);
 850                 drr->drr_payloadlen = payload_len;
 851                 fnvlist_free(nvl);
 852         }
 853
 854         err = dump_record(dsp, payload, payload_len);
 855         fnvlist_pack_free(payload, payload_len);
 856         if (err != 0) {
 857                 err = dsp->dsa_err;
 858                 goto out;
 859         }
 860
 861         err = bqueue_init(&to_arg.q, zfs_send_queue_length,
 862             offsetof(struct send_block_record, ln));
 863         to_arg.error_code = 0;
 864         to_arg.cancel = B_FALSE;
 865         to_arg.ds = to_ds;
 866         to_arg.fromtxg = fromtxg;
 867         to_arg.flags = TRAVERSE_PRE | TRAVERSE_PREFETCH;
 868         (void) thread_create(NULL, 0, send_traverse_thread, &to_arg, 0, &p0,
 869             TS_RUN, minclsyspri);
 870
 871         struct send_block_record *to_data;
 872         to_data = bqueue_dequeue(&to_arg.q);
 873
 874         while (!to_data->eos_marker && err == 0) {
 875                 err = do_dump(dsp, to_data);
 876                 to_data = get_next_record(&to_arg.q, to_data);
 877                 if (issig(JUSTLOOKING) && issig(FORREAL))
 878                         err = EINTR;
 879         }
 880
 881         if (err != 0) {
 882                 to_arg.cancel = B_TRUE;
 883                 while (!to_data->eos_marker) {
 884                         to_data = get_next_record(&to_arg.q, to_data);
 885                 }
 886         }
 887         kmem_free(to_data, sizeof (*to_data));
 888
 889         bqueue_destroy(&to_arg.q);
 890
 891         if (err == 0 && to_arg.error_code != 0)
 892                 err = to_arg.error_code;
 893
 894         if (err != 0)
 895                 goto out;
 896
 897         if (dsp->dsa_pending_op != PENDING_NONE)
 898                 if (dump_record(dsp, NULL, 0) != 0)
 899                         err = SET_ERROR(EINTR);
 900
 901         if (err != 0) {
 902                 if (err == EINTR && dsp->dsa_err != 0)
 903                         err = dsp->dsa_err;
 904                 goto out;
 905         }
 906
 907         bzero(drr, sizeof (dmu_replay_record_t));
 908         drr->drr_type = DRR_END;
 909         drr->drr_u.drr_end.drr_checksum = dsp->dsa_zc;
 910         drr->drr_u.drr_end.drr_toguid = dsp->dsa_toguid;
 911
 912         if (dump_record(dsp, NULL, 0) != 0)
 913                 err = dsp->dsa_err;
 914
 915 out:
 916         mutex_enter(&to_ds->ds_sendstream_lock);
 917         list_remove(&to_ds->ds_sendstreams, dsp);
 918         mutex_exit(&to_ds->ds_sendstream_lock);
 919
 920         VERIFY(err != 0 || (dsp->dsa_sent_begin && dsp->dsa_sent_end));
 921
 922         kmem_free(drr, sizeof (dmu_replay_record_t));
 923         kmem_free(dsp, sizeof (dmu_sendarg_t));
 924
 925         dsl_dataset_long_rele(to_ds, FTAG);
 926
 927         return (err);
 928 }
 929
 930 int
 931 dmu_send_obj(const char *pool, uint64_t tosnap, uint64_t fromsnap,
 932     boolean_t embedok, boolean_t large_block_ok,
 933 #ifdef illumos
 934     int outfd, vnode_t *vp, offset_t *off)
 935 #else
 936     int outfd, struct file *fp, offset_t *off)
 937 #endif
 938 {
 939         dsl_pool_t *dp;
 940         dsl_dataset_t *ds;
 941         dsl_dataset_t *fromds = NULL;
 942         int err;
 943
 944         err = dsl_pool_hold(pool, FTAG, &dp);
 945         if (err != 0)
 946                 return (err);
 947
 948         err = dsl_dataset_hold_obj(dp, tosnap, FTAG, &ds);
 949         if (err != 0) {
 950                 dsl_pool_rele(dp, FTAG);
 951                 return (err);
 952         }
 953
 954         if (fromsnap != 0) {
 955                 zfs_bookmark_phys_t zb;
 956                 boolean_t is_clone;
 957
 958                 err = dsl_dataset_hold_obj(dp, fromsnap, FTAG, &fromds);
 959                 if (err != 0) {
 960                         dsl_dataset_rele(ds, FTAG);
 961                         dsl_pool_rele(dp, FTAG);
 962                         return (err);
 963                 }
 964                 if (!dsl_dataset_is_before(ds, fromds, 0))
 965                         err = SET_ERROR(EXDEV);
 966                 zb.zbm_creation_time =
 967                     dsl_dataset_phys(fromds)->ds_creation_time;
 968                 zb.zbm_creation_txg = dsl_dataset_phys(fromds)->ds_creation_txg;
 969                 zb.zbm_guid = dsl_dataset_phys(fromds)->ds_guid;
 970                 is_clone = (fromds->ds_dir != ds->ds_dir);
 971                 dsl_dataset_rele(fromds, FTAG);
 972                 err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone,
 973                     embedok, large_block_ok, outfd, 0, 0, fp, off);
 974         } else {
 975                 err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE,
 976                     embedok, large_block_ok, outfd, 0, 0, fp, off);
 977         }
 978         dsl_dataset_rele(ds, FTAG);
 979         return (err);
 980 }
 981
 982 int
 983 dmu_send(const char *tosnap, const char *fromsnap, boolean_t embedok,
 984     boolean_t large_block_ok, int outfd, uint64_t resumeobj, uint64_t resumeoff,
 985 #ifdef illumos
 986     vnode_t *vp, offset_t *off)
 987 #else
 988     struct file *fp, offset_t *off)
 989 #endif
 990 {
 991         dsl_pool_t *dp;
 992         dsl_dataset_t *ds;
 993         int err;
 994         boolean_t owned = B_FALSE;
 995
 996         if (fromsnap != NULL && strpbrk(fromsnap, "@#") == NULL)
 997                 return (SET_ERROR(EINVAL));
 998
 999         err = dsl_pool_hold(tosnap, FTAG, &dp);
1000         if (err != 0)
1001                 return (err);
1002
1003         if (strchr(tosnap, '@') == NULL && spa_writeable(dp->dp_spa)) {
1004                 /*
1005                  * We are sending a filesystem or volume.  Ensure
1006                  * that it doesn't change by owning the dataset.
1007                  */
1008                 err = dsl_dataset_own(dp, tosnap, FTAG, &ds);
1009                 owned = B_TRUE;
1010         } else {
1011                 err = dsl_dataset_hold(dp, tosnap, FTAG, &ds);
1012         }
1013         if (err != 0) {
1014                 dsl_pool_rele(dp, FTAG);
1015                 return (err);
1016         }
1017
1018         if (fromsnap != NULL) {
1019                 zfs_bookmark_phys_t zb;
1020                 boolean_t is_clone = B_FALSE;
1021                 int fsnamelen = strchr(tosnap, '@') - tosnap;
1022
1023                 /*
1024                  * If the fromsnap is in a different filesystem, then
1025                  * mark the send stream as a clone.
1026                  */
1027                 if (strncmp(tosnap, fromsnap, fsnamelen) != 0 ||
1028                     (fromsnap[fsnamelen] != '@' &&
1029                     fromsnap[fsnamelen] != '#')) {
1030                         is_clone = B_TRUE;
1031                 }
1032
1033                 if (strchr(fromsnap, '@')) {
1034                         dsl_dataset_t *fromds;
1035                         err = dsl_dataset_hold(dp, fromsnap, FTAG, &fromds);
1036                         if (err == 0) {
1037                                 if (!dsl_dataset_is_before(ds, fromds, 0))
1038                                         err = SET_ERROR(EXDEV);
1039                                 zb.zbm_creation_time =
1040                                     dsl_dataset_phys(fromds)->ds_creation_time;
1041                                 zb.zbm_creation_txg =
1042                                     dsl_dataset_phys(fromds)->ds_creation_txg;
1043                                 zb.zbm_guid = dsl_dataset_phys(fromds)->ds_guid;
1044                                 is_clone = (ds->ds_dir != fromds->ds_dir);
1045                                 dsl_dataset_rele(fromds, FTAG);
1046                         }
1047                 } else {
1048                         err = dsl_bookmark_lookup(dp, fromsnap, ds, &zb);
1049                 }
1050                 if (err != 0) {
1051                         dsl_dataset_rele(ds, FTAG);
1052                         dsl_pool_rele(dp, FTAG);
1053                         return (err);
1054                 }
1055                 err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone,
1056                     embedok, large_block_ok,
1057                     outfd, resumeobj, resumeoff, fp, off);
1058         } else {
1059                 err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE,
1060                     embedok, large_block_ok,
1061                     outfd, resumeobj, resumeoff, fp, off);
1062         }
1063         if (owned)
1064                 dsl_dataset_disown(ds, FTAG);
1065         else
1066                 dsl_dataset_rele(ds, FTAG);
1067         return (err);
1068 }
1069
1070 static int
1071 dmu_adjust_send_estimate_for_indirects(dsl_dataset_t *ds, uint64_t size,
1072     uint64_t *sizep)
1073 {
1074         int err;
1075         /*
1076          * Assume that space (both on-disk and in-stream) is dominated by
1077          * data.  We will adjust for indirect blocks and the copies property,
1078          * but ignore per-object space used (eg, dnodes and DRR_OBJECT records).
1079          */
1080
1081         /*
1082          * Subtract out approximate space used by indirect blocks.
1083          * Assume most space is used by data blocks (non-indirect, non-dnode).
1084          * Assume all blocks are recordsize.  Assume ditto blocks and
1085          * internal fragmentation counter out compression.
1086          *
1087          * Therefore, space used by indirect blocks is sizeof(blkptr_t) per
1088          * block, which we observe in practice.
1089          */
1090         uint64_t recordsize;
1091         err = dsl_prop_get_int_ds(ds, "recordsize", &recordsize);
1092         if (err != 0)
1093                 return (err);
1094         size -= size / recordsize * sizeof (blkptr_t);
1095
1096         /* Add in the space for the record associated with each block. */
1097         size += size / recordsize * sizeof (dmu_replay_record_t);
1098
1099         *sizep = size;
1100
1101         return (0);
1102 }
1103
1104 int
1105 dmu_send_estimate(dsl_dataset_t *ds, dsl_dataset_t *fromds, uint64_t *sizep)
1106 {
1107         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1108         int err;
1109         uint64_t size;
1110
1111         ASSERT(dsl_pool_config_held(dp));
1112
1113         /* tosnap must be a snapshot */
1114         if (!ds->ds_is_snapshot)
1115                 return (SET_ERROR(EINVAL));
1116
1117         /* fromsnap, if provided, must be a snapshot */
1118         if (fromds != NULL && !fromds->ds_is_snapshot)
1119                 return (SET_ERROR(EINVAL));
1120
1121         /*
1122          * fromsnap must be an earlier snapshot from the same fs as tosnap,
1123          * or the origin's fs.
1124          */
1125         if (fromds != NULL && !dsl_dataset_is_before(ds, fromds, 0))
1126                 return (SET_ERROR(EXDEV));
1127
1128         /* Get uncompressed size estimate of changed data. */
1129         if (fromds == NULL) {
1130                 size = dsl_dataset_phys(ds)->ds_uncompressed_bytes;
1131         } else {
1132                 uint64_t used, comp;
1133                 err = dsl_dataset_space_written(fromds, ds,
1134                     &used, &comp, &size);
1135                 if (err != 0)
1136                         return (err);
1137         }
1138
1139         err = dmu_adjust_send_estimate_for_indirects(ds, size, sizep);
1140         return (err);
1141 }
1142
1143 /*
1144  * Simple callback used to traverse the blocks of a snapshot and sum their
1145  * uncompressed size
1146  */
1147 /* ARGSUSED */
1148 static int
1149 dmu_calculate_send_traversal(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
1150     const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
1151 {
1152         uint64_t *spaceptr = arg;
1153         if (bp != NULL && !BP_IS_HOLE(bp)) {
1154                 *spaceptr += BP_GET_UCSIZE(bp);
1155         }
1156         return (0);
1157 }
1158
1159 /*
1160  * Given a desination snapshot and a TXG, calculate the approximate size of a
1161  * send stream sent from that TXG. from_txg may be zero, indicating that the
1162  * whole snapshot will be sent.
1163  */
1164 int
1165 dmu_send_estimate_from_txg(dsl_dataset_t *ds, uint64_t from_txg,
1166     uint64_t *sizep)
1167 {
1168         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1169         int err;
1170         uint64_t size = 0;
1171
1172         ASSERT(dsl_pool_config_held(dp));
1173
1174         /* tosnap must be a snapshot */
1175         if (!dsl_dataset_is_snapshot(ds))
1176                 return (SET_ERROR(EINVAL));
1177
1178         /* verify that from_txg is before the provided snapshot was taken */
1179         if (from_txg >= dsl_dataset_phys(ds)->ds_creation_txg) {
1180                 return (SET_ERROR(EXDEV));
1181         }
1182
1183         /*
1184          * traverse the blocks of the snapshot with birth times after
1185          * from_txg, summing their uncompressed size
1186          */
1187         err = traverse_dataset(ds, from_txg, TRAVERSE_POST,
1188             dmu_calculate_send_traversal, &size);
1189         if (err)
1190                 return (err);
1191
1192         err = dmu_adjust_send_estimate_for_indirects(ds, size, sizep);
1193         return (err);
1194 }
1195
1196 typedef struct dmu_recv_begin_arg {
1197         const char *drba_origin;
1198         dmu_recv_cookie_t *drba_cookie;
1199         cred_t *drba_cred;
1200         uint64_t drba_snapobj;
1201 } dmu_recv_begin_arg_t;
1202
1203 static int
1204 recv_begin_check_existing_impl(dmu_recv_begin_arg_t *drba, dsl_dataset_t *ds,
1205     uint64_t fromguid)
1206 {
1207         uint64_t val;
1208         int error;
1209         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1210
1211         /* temporary clone name must not exist */
1212         error = zap_lookup(dp->dp_meta_objset,
1213             dsl_dir_phys(ds->ds_dir)->dd_child_dir_zapobj, recv_clone_name,
1214             8, 1, &val);
1215         if (error != ENOENT)
1216                 return (error == 0 ? EBUSY : error);
1217
1218         /* new snapshot name must not exist */
1219         error = zap_lookup(dp->dp_meta_objset,
1220             dsl_dataset_phys(ds)->ds_snapnames_zapobj,
1221             drba->drba_cookie->drc_tosnap, 8, 1, &val);
1222         if (error != ENOENT)
1223                 return (error == 0 ? EEXIST : error);
1224
1225         /*
1226          * Check snapshot limit before receiving. We'll recheck again at the
1227          * end, but might as well abort before receiving if we're already over
1228          * the limit.
1229          *
1230          * Note that we do not check the file system limit with
1231          * dsl_dir_fscount_check because the temporary %clones don't count
1232          * against that limit.
1233          */
1234         error = dsl_fs_ss_limit_check(ds->ds_dir, 1, ZFS_PROP_SNAPSHOT_LIMIT,
1235             NULL, drba->drba_cred);
1236         if (error != 0)
1237                 return (error);
1238
1239         if (fromguid != 0) {
1240                 dsl_dataset_t *snap;
1241                 uint64_t obj = dsl_dataset_phys(ds)->ds_prev_snap_obj;
1242
1243                 /* Find snapshot in this dir that matches fromguid. */
1244                 while (obj != 0) {
1245                         error = dsl_dataset_hold_obj(dp, obj, FTAG,
1246                             &snap);
1247                         if (error != 0)
1248                                 return (SET_ERROR(ENODEV));
1249                         if (snap->ds_dir != ds->ds_dir) {
1250                                 dsl_dataset_rele(snap, FTAG);
1251                                 return (SET_ERROR(ENODEV));
1252                         }
1253                         if (dsl_dataset_phys(snap)->ds_guid == fromguid)
1254                                 break;
1255                         obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
1256                         dsl_dataset_rele(snap, FTAG);
1257                 }
1258                 if (obj == 0)
1259                         return (SET_ERROR(ENODEV));
1260
1261                 if (drba->drba_cookie->drc_force) {
1262                         drba->drba_snapobj = obj;
1263                 } else {
1264                         /*
1265                          * If we are not forcing, there must be no
1266                          * changes since fromsnap.
1267                          */
1268                         if (dsl_dataset_modified_since_snap(ds, snap)) {
1269                                 dsl_dataset_rele(snap, FTAG);
1270                                 return (SET_ERROR(ETXTBSY));
1271                         }
1272                         drba->drba_snapobj = ds->ds_prev->ds_object;
1273                 }
1274
1275                 dsl_dataset_rele(snap, FTAG);
1276         } else {
1277                 /* if full, then must be forced */
1278                 if (!drba->drba_cookie->drc_force)
1279                         return (SET_ERROR(EEXIST));
1280                 /* start from $ORIGIN@$ORIGIN, if supported */
1281                 drba->drba_snapobj = dp->dp_origin_snap != NULL ?
1282                     dp->dp_origin_snap->ds_object : 0;
1283         }
1284
1285         return (0);
1286
1287 }
1288
1289 static int
1290 dmu_recv_begin_check(void *arg, dmu_tx_t *tx)
1291 {
1292         dmu_recv_begin_arg_t *drba = arg;
1293         dsl_pool_t *dp = dmu_tx_pool(tx);
1294         struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
1295         uint64_t fromguid = drrb->drr_fromguid;
1296         int flags = drrb->drr_flags;
1297         int error;
1298         uint64_t featureflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo);
1299         dsl_dataset_t *ds;
1300         const char *tofs = drba->drba_cookie->drc_tofs;
1301
1302         /* already checked */
1303         ASSERT3U(drrb->drr_magic, ==, DMU_BACKUP_MAGIC);
1304         ASSERT(!(featureflags & DMU_BACKUP_FEATURE_RESUMING));
1305
1306         if (DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) ==
1307             DMU_COMPOUNDSTREAM ||
1308             drrb->drr_type >= DMU_OST_NUMTYPES ||
1309             ((flags & DRR_FLAG_CLONE) && drba->drba_origin == NULL))
1310                 return (SET_ERROR(EINVAL));
1311
1312         /* Verify pool version supports SA if SA_SPILL feature set */
1313         if ((featureflags & DMU_BACKUP_FEATURE_SA_SPILL) &&
1314             spa_version(dp->dp_spa) < SPA_VERSION_SA)
1315                 return (SET_ERROR(ENOTSUP));
1316
1317         if (drba->drba_cookie->drc_resumable &&
1318             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EXTENSIBLE_DATASET))
1319                 return (SET_ERROR(ENOTSUP));
1320
1321         /*
1322          * The receiving code doesn't know how to translate a WRITE_EMBEDDED
1323          * record to a plan WRITE record, so the pool must have the
1324          * EMBEDDED_DATA feature enabled if the stream has WRITE_EMBEDDED
1325          * records.  Same with WRITE_EMBEDDED records that use LZ4 compression.
1326          */
1327         if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA) &&
1328             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA))
1329                 return (SET_ERROR(ENOTSUP));
1330         if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA_LZ4) &&
1331             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS))
1332                 return (SET_ERROR(ENOTSUP));
1333
1334         /*
1335          * The receiving code doesn't know how to translate large blocks
1336          * to smaller ones, so the pool must have the LARGE_BLOCKS
1337          * feature enabled if the stream has LARGE_BLOCKS.
1338          */
1339         if ((featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
1340             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LARGE_BLOCKS))
1341                 return (SET_ERROR(ENOTSUP));
1342
1343         error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
1344         if (error == 0) {
1345                 /* target fs already exists; recv into temp clone */
1346
1347                 /* Can't recv a clone into an existing fs */
1348                 if (flags & DRR_FLAG_CLONE || drba->drba_origin) {
1349                         dsl_dataset_rele(ds, FTAG);
1350                         return (SET_ERROR(EINVAL));
1351                 }
1352
1353                 error = recv_begin_check_existing_impl(drba, ds, fromguid);
1354                 dsl_dataset_rele(ds, FTAG);
1355         } else if (error == ENOENT) {
1356                 /* target fs does not exist; must be a full backup or clone */
1357                 char buf[ZFS_MAX_DATASET_NAME_LEN];
1358
1359                 /*
1360                  * If it's a non-clone incremental, we are missing the
1361                  * target fs, so fail the recv.
1362                  */
1363                 if (fromguid != 0 && !(flags & DRR_FLAG_CLONE ||
1364                     drba->drba_origin))
1365                         return (SET_ERROR(ENOENT));
1366
1367                 /*
1368                  * If we're receiving a full send as a clone, and it doesn't
1369                  * contain all the necessary free records and freeobject
1370                  * records, reject it.
1371                  */
1372                 if (fromguid == 0 && drba->drba_origin &&
1373                     !(flags & DRR_FLAG_FREERECORDS))
1374                         return (SET_ERROR(EINVAL));
1375
1376                 /* Open the parent of tofs */
1377                 ASSERT3U(strlen(tofs), <, sizeof (buf));
1378                 (void) strlcpy(buf, tofs, strrchr(tofs, '/') - tofs + 1);
1379                 error = dsl_dataset_hold(dp, buf, FTAG, &ds);
1380                 if (error != 0)
1381                         return (error);
1382
1383                 /*
1384                  * Check filesystem and snapshot limits before receiving. We'll
1385                  * recheck snapshot limits again at the end (we create the
1386                  * filesystems and increment those counts during begin_sync).
1387                  */
1388                 error = dsl_fs_ss_limit_check(ds->ds_dir, 1,
1389                     ZFS_PROP_FILESYSTEM_LIMIT, NULL, drba->drba_cred);
1390                 if (error != 0) {
1391                         dsl_dataset_rele(ds, FTAG);
1392                         return (error);
1393                 }
1394
1395                 error = dsl_fs_ss_limit_check(ds->ds_dir, 1,
1396                     ZFS_PROP_SNAPSHOT_LIMIT, NULL, drba->drba_cred);
1397                 if (error != 0) {
1398                         dsl_dataset_rele(ds, FTAG);
1399                         return (error);
1400                 }
1401
1402                 if (drba->drba_origin != NULL) {
1403                         dsl_dataset_t *origin;
1404                         error = dsl_dataset_hold(dp, drba->drba_origin,
1405                             FTAG, &origin);
1406                         if (error != 0) {
1407                                 dsl_dataset_rele(ds, FTAG);
1408                                 return (error);
1409                         }
1410                         if (!origin->ds_is_snapshot) {
1411                                 dsl_dataset_rele(origin, FTAG);
1412                                 dsl_dataset_rele(ds, FTAG);
1413                                 return (SET_ERROR(EINVAL));
1414                         }
1415                         if (dsl_dataset_phys(origin)->ds_guid != fromguid &&
1416                             fromguid != 0) {
1417                                 dsl_dataset_rele(origin, FTAG);
1418                                 dsl_dataset_rele(ds, FTAG);
1419                                 return (SET_ERROR(ENODEV));
1420                         }
1421                         dsl_dataset_rele(origin, FTAG);
1422                 }
1423                 dsl_dataset_rele(ds, FTAG);
1424                 error = 0;
1425         }
1426         return (error);
1427 }
1428
1429 static void
1430 dmu_recv_begin_sync(void *arg, dmu_tx_t *tx)
1431 {
1432         dmu_recv_begin_arg_t *drba = arg;
1433         dsl_pool_t *dp = dmu_tx_pool(tx);
1434         objset_t *mos = dp->dp_meta_objset;
1435         struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
1436         const char *tofs = drba->drba_cookie->drc_tofs;
1437         dsl_dataset_t *ds, *newds;
1438         uint64_t dsobj;
1439         int error;
1440         uint64_t crflags = 0;
1441
1442         if (drrb->drr_flags & DRR_FLAG_CI_DATA)
1443                 crflags |= DS_FLAG_CI_DATASET;
1444
1445         error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
1446         if (error == 0) {
1447                 /* create temporary clone */
1448                 dsl_dataset_t *snap = NULL;
1449                 if (drba->drba_snapobj != 0) {
1450                         VERIFY0(dsl_dataset_hold_obj(dp,
1451                             drba->drba_snapobj, FTAG, &snap));
1452                 }
1453                 dsobj = dsl_dataset_create_sync(ds->ds_dir, recv_clone_name,
1454                     snap, crflags, drba->drba_cred, tx);
1455                 if (drba->drba_snapobj != 0)
1456                         dsl_dataset_rele(snap, FTAG);
1457                 dsl_dataset_rele(ds, FTAG);
1458         } else {
1459                 dsl_dir_t *dd;
1460                 const char *tail;
1461                 dsl_dataset_t *origin = NULL;
1462
1463                 VERIFY0(dsl_dir_hold(dp, tofs, FTAG, &dd, &tail));
1464
1465                 if (drba->drba_origin != NULL) {
1466                         VERIFY0(dsl_dataset_hold(dp, drba->drba_origin,
1467                             FTAG, &origin));
1468                 }
1469
1470                 /* Create new dataset. */
1471                 dsobj = dsl_dataset_create_sync(dd,
1472                     strrchr(tofs, '/') + 1,
1473                     origin, crflags, drba->drba_cred, tx);
1474                 if (origin != NULL)
1475                         dsl_dataset_rele(origin, FTAG);
1476                 dsl_dir_rele(dd, FTAG);
1477                 drba->drba_cookie->drc_newfs = B_TRUE;
1478         }
1479         VERIFY0(dsl_dataset_own_obj(dp, dsobj, dmu_recv_tag, &newds));
1480
1481         if (drba->drba_cookie->drc_resumable) {
1482                 dsl_dataset_zapify(newds, tx);
1483                 if (drrb->drr_fromguid != 0) {
1484                         VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_FROMGUID,
1485                             8, 1, &drrb->drr_fromguid, tx));
1486                 }
1487                 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_TOGUID,
1488                     8, 1, &drrb->drr_toguid, tx));
1489                 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_TONAME,
1490                     1, strlen(drrb->drr_toname) + 1, drrb->drr_toname, tx));
1491                 uint64_t one = 1;
1492                 uint64_t zero = 0;
1493                 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_OBJECT,
1494                     8, 1, &one, tx));
1495                 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_OFFSET,
1496                     8, 1, &zero, tx));
1497                 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_BYTES,
1498                     8, 1, &zero, tx));
1499                 if (DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo) &
1500                     DMU_BACKUP_FEATURE_EMBED_DATA) {
1501                         VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_EMBEDOK,
1502                             8, 1, &one, tx));
1503                 }
1504         }
1505
1506         dmu_buf_will_dirty(newds->ds_dbuf, tx);
1507         dsl_dataset_phys(newds)->ds_flags |= DS_FLAG_INCONSISTENT;
1508
1509         /*
1510          * If we actually created a non-clone, we need to create the
1511          * objset in our new dataset.
1512          */
1513         rrw_enter(&newds->ds_bp_rwlock, RW_READER, FTAG);
1514         if (BP_IS_HOLE(dsl_dataset_get_blkptr(newds))) {
1515                 (void) dmu_objset_create_impl(dp->dp_spa,
1516                     newds, dsl_dataset_get_blkptr(newds), drrb->drr_type, tx);
1517         }
1518         rrw_exit(&newds->ds_bp_rwlock, FTAG);
1519
1520         drba->drba_cookie->drc_ds = newds;
1521
1522         spa_history_log_internal_ds(newds, "receive", tx, "");
1523 }
1524
1525 static int
1526 dmu_recv_resume_begin_check(void *arg, dmu_tx_t *tx)
1527 {
1528         dmu_recv_begin_arg_t *drba = arg;
1529         dsl_pool_t *dp = dmu_tx_pool(tx);
1530         struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
1531         int error;
1532         uint64_t featureflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo);
1533         dsl_dataset_t *ds;
1534         const char *tofs = drba->drba_cookie->drc_tofs;
1535
1536         /* already checked */
1537         ASSERT3U(drrb->drr_magic, ==, DMU_BACKUP_MAGIC);
1538         ASSERT(featureflags & DMU_BACKUP_FEATURE_RESUMING);
1539
1540         if (DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) ==
1541             DMU_COMPOUNDSTREAM ||
1542             drrb->drr_type >= DMU_OST_NUMTYPES)
1543                 return (SET_ERROR(EINVAL));
1544
1545         /* Verify pool version supports SA if SA_SPILL feature set */
1546         if ((featureflags & DMU_BACKUP_FEATURE_SA_SPILL) &&
1547             spa_version(dp->dp_spa) < SPA_VERSION_SA)
1548                 return (SET_ERROR(ENOTSUP));
1549
1550         /*
1551          * The receiving code doesn't know how to translate a WRITE_EMBEDDED
1552          * record to a plain WRITE record, so the pool must have the
1553          * EMBEDDED_DATA feature enabled if the stream has WRITE_EMBEDDED
1554          * records.  Same with WRITE_EMBEDDED records that use LZ4 compression.
1555          */
1556         if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA) &&
1557             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA))
1558                 return (SET_ERROR(ENOTSUP));
1559         if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA_LZ4) &&
1560             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS))
1561                 return (SET_ERROR(ENOTSUP));
1562
1563         /* 6 extra bytes for /%recv */
1564         char recvname[ZFS_MAX_DATASET_NAME_LEN + 6];
1565
1566         (void) snprintf(recvname, sizeof (recvname), "%s/%s",
1567             tofs, recv_clone_name);
1568
1569         if (dsl_dataset_hold(dp, recvname, FTAG, &ds) != 0) {
1570                 /* %recv does not exist; continue in tofs */
1571                 error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
1572                 if (error != 0)
1573                         return (error);
1574         }
1575
1576         /* check that ds is marked inconsistent */
1577         if (!DS_IS_INCONSISTENT(ds)) {
1578                 dsl_dataset_rele(ds, FTAG);
1579                 return (SET_ERROR(EINVAL));
1580         }
1581
1582         /* check that there is resuming data, and that the toguid matches */
1583         if (!dsl_dataset_is_zapified(ds)) {
1584                 dsl_dataset_rele(ds, FTAG);
1585                 return (SET_ERROR(EINVAL));
1586         }
1587         uint64_t val;
1588         error = zap_lookup(dp->dp_meta_objset, ds->ds_object,
1589             DS_FIELD_RESUME_TOGUID, sizeof (val), 1, &val);
1590         if (error != 0 || drrb->drr_toguid != val) {
1591                 dsl_dataset_rele(ds, FTAG);
1592                 return (SET_ERROR(EINVAL));
1593         }
1594
1595         /*
1596          * Check if the receive is still running.  If so, it will be owned.
1597          * Note that nothing else can own the dataset (e.g. after the receive
1598          * fails) because it will be marked inconsistent.
1599          */
1600         if (dsl_dataset_has_owner(ds)) {
1601                 dsl_dataset_rele(ds, FTAG);
1602                 return (SET_ERROR(EBUSY));
1603         }
1604
1605         /* There should not be any snapshots of this fs yet. */
1606         if (ds->ds_prev != NULL && ds->ds_prev->ds_dir == ds->ds_dir) {
1607                 dsl_dataset_rele(ds, FTAG);
1608                 return (SET_ERROR(EINVAL));
1609         }
1610
1611         /*
1612          * Note: resume point will be checked when we process the first WRITE
1613          * record.
1614          */
1615
1616         /* check that the origin matches */
1617         val = 0;
1618         (void) zap_lookup(dp->dp_meta_objset, ds->ds_object,
1619             DS_FIELD_RESUME_FROMGUID, sizeof (val), 1, &val);
1620         if (drrb->drr_fromguid != val) {
1621                 dsl_dataset_rele(ds, FTAG);
1622                 return (SET_ERROR(EINVAL));
1623         }
1624
1625         dsl_dataset_rele(ds, FTAG);
1626         return (0);
1627 }
1628
1629 static void
1630 dmu_recv_resume_begin_sync(void *arg, dmu_tx_t *tx)
1631 {
1632         dmu_recv_begin_arg_t *drba = arg;
1633         dsl_pool_t *dp = dmu_tx_pool(tx);
1634         const char *tofs = drba->drba_cookie->drc_tofs;
1635         dsl_dataset_t *ds;
1636         uint64_t dsobj;
1637         /* 6 extra bytes for /%recv */
1638         char recvname[ZFS_MAX_DATASET_NAME_LEN + 6];
1639
1640         (void) snprintf(recvname, sizeof (recvname), "%s/%s",
1641             tofs, recv_clone_name);
1642
1643         if (dsl_dataset_hold(dp, recvname, FTAG, &ds) != 0) {
1644                 /* %recv does not exist; continue in tofs */
1645                 VERIFY0(dsl_dataset_hold(dp, tofs, FTAG, &ds));
1646                 drba->drba_cookie->drc_newfs = B_TRUE;
1647         }
1648
1649         /* clear the inconsistent flag so that we can own it */
1650         ASSERT(DS_IS_INCONSISTENT(ds));
1651         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1652         dsl_dataset_phys(ds)->ds_flags &= ~DS_FLAG_INCONSISTENT;
1653         dsobj = ds->ds_object;
1654         dsl_dataset_rele(ds, FTAG);
1655
1656         VERIFY0(dsl_dataset_own_obj(dp, dsobj, dmu_recv_tag, &ds));
1657
1658         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1659         dsl_dataset_phys(ds)->ds_flags |= DS_FLAG_INCONSISTENT;
1660
1661         rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
1662         ASSERT(!BP_IS_HOLE(dsl_dataset_get_blkptr(ds)));
1663         rrw_exit(&ds->ds_bp_rwlock, FTAG);
1664
1665         drba->drba_cookie->drc_ds = ds;
1666
1667         spa_history_log_internal_ds(ds, "resume receive", tx, "");
1668 }
1669
1670 /*
1671  * NB: callers *MUST* call dmu_recv_stream() if dmu_recv_begin()
1672  * succeeds; otherwise we will leak the holds on the datasets.
1673  */
1674 int
1675 dmu_recv_begin(char *tofs, char *tosnap, dmu_replay_record_t *drr_begin,
1676     boolean_t force, boolean_t resumable, char *origin, dmu_recv_cookie_t *drc)
1677 {
1678         dmu_recv_begin_arg_t drba = { 0 };
1679
1680         bzero(drc, sizeof (dmu_recv_cookie_t));
1681         drc->drc_drr_begin = drr_begin;
1682         drc->drc_drrb = &drr_begin->drr_u.drr_begin;
1683         drc->drc_tosnap = tosnap;
1684         drc->drc_tofs = tofs;
1685         drc->drc_force = force;
1686         drc->drc_resumable = resumable;
1687         drc->drc_cred = CRED();
1688
1689         if (drc->drc_drrb->drr_magic == BSWAP_64(DMU_BACKUP_MAGIC)) {
1690                 drc->drc_byteswap = B_TRUE;
1691                 fletcher_4_incremental_byteswap(drr_begin,
1692                     sizeof (dmu_replay_record_t), &drc->drc_cksum);
1693                 byteswap_record(drr_begin);
1694         } else if (drc->drc_drrb->drr_magic == DMU_BACKUP_MAGIC) {
1695                 fletcher_4_incremental_native(drr_begin,
1696                     sizeof (dmu_replay_record_t), &drc->drc_cksum);
1697         } else {
1698                 return (SET_ERROR(EINVAL));
1699         }
1700
1701         drba.drba_origin = origin;
1702         drba.drba_cookie = drc;
1703         drba.drba_cred = CRED();
1704
1705         if (DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo) &
1706             DMU_BACKUP_FEATURE_RESUMING) {
1707                 return (dsl_sync_task(tofs,
1708                     dmu_recv_resume_begin_check, dmu_recv_resume_begin_sync,
1709                     &drba, 5, ZFS_SPACE_CHECK_NORMAL));
1710         } else  {
1711                 return (dsl_sync_task(tofs,
1712                     dmu_recv_begin_check, dmu_recv_begin_sync,
1713                     &drba, 5, ZFS_SPACE_CHECK_NORMAL));
1714         }
1715 }
1716
1717 struct receive_record_arg {
1718         dmu_replay_record_t header;
1719         void *payload; /* Pointer to a buffer containing the payload */
1720         /*
1721          * If the record is a write, pointer to the arc_buf_t containing the
1722          * payload.
1723          */
1724         arc_buf_t *write_buf;
1725         int payload_size;
1726         uint64_t bytes_read; /* bytes read from stream when record created */
1727         boolean_t eos_marker; /* Marks the end of the stream */
1728         bqueue_node_t node;
1729 };
1730
1731 struct receive_writer_arg {
1732         objset_t *os;
1733         boolean_t byteswap;
1734         bqueue_t q;
1735
1736         /*
1737          * These three args are used to signal to the main thread that we're
1738          * done.
1739          */
1740         kmutex_t mutex;
1741         kcondvar_t cv;
1742         boolean_t done;
1743
1744         int err;
1745         /* A map from guid to dataset to help handle dedup'd streams. */
1746         avl_tree_t *guid_to_ds_map;
1747         boolean_t resumable;
1748         uint64_t last_object, last_offset;
1749         uint64_t bytes_read; /* bytes read when current record created */
1750 };
1751
1752 struct objlist {
1753         list_t list; /* List of struct receive_objnode. */
1754         /*
1755          * Last object looked up. Used to assert that objects are being looked
1756          * up in ascending order.
1757          */
1758         uint64_t last_lookup;
1759 };
1760
1761 struct receive_objnode {
1762         list_node_t node;
1763         uint64_t object;
1764 };
1765
1766 struct receive_arg  {
1767         objset_t *os;
1768         kthread_t *td;
1769         struct file *fp;
1770         uint64_t voff; /* The current offset in the stream */
1771         uint64_t bytes_read;
1772         /*
1773          * A record that has had its payload read in, but hasn't yet been handed
1774          * off to the worker thread.
1775          */
1776         struct receive_record_arg *rrd;
1777         /* A record that has had its header read in, but not its payload. */
1778         struct receive_record_arg *next_rrd;
1779         zio_cksum_t cksum;
1780         zio_cksum_t prev_cksum;
1781         int err;
1782         boolean_t byteswap;
1783         /* Sorted list of objects not to issue prefetches for. */
1784         struct objlist ignore_objlist;
1785 };
1786
1787 typedef struct guid_map_entry {
1788         uint64_t        guid;
1789         dsl_dataset_t   *gme_ds;
1790         avl_node_t      avlnode;
1791 } guid_map_entry_t;
1792
1793 static int
1794 guid_compare(const void *arg1, const void *arg2)
1795 {
1796         const guid_map_entry_t *gmep1 = arg1;
1797         const guid_map_entry_t *gmep2 = arg2;
1798
1799         if (gmep1->guid < gmep2->guid)
1800                 return (-1);
1801         else if (gmep1->guid > gmep2->guid)
1802                 return (1);
1803         return (0);
1804 }
1805
1806 static void
1807 free_guid_map_onexit(void *arg)
1808 {
1809         avl_tree_t *ca = arg;
1810         void *cookie = NULL;
1811         guid_map_entry_t *gmep;
1812
1813         while ((gmep = avl_destroy_nodes(ca, &cookie)) != NULL) {
1814                 dsl_dataset_long_rele(gmep->gme_ds, gmep);
1815                 dsl_dataset_rele(gmep->gme_ds, gmep);
1816                 kmem_free(gmep, sizeof (guid_map_entry_t));
1817         }
1818         avl_destroy(ca);
1819         kmem_free(ca, sizeof (avl_tree_t));
1820 }
1821
1822 static int
1823 restore_bytes(struct receive_arg *ra, void *buf, int len, off_t off, ssize_t *resid)
1824 {
1825         struct uio auio;
1826         struct iovec aiov;
1827         int error;
1828
1829         aiov.iov_base = buf;
1830         aiov.iov_len = len;
1831         auio.uio_iov = &aiov;
1832         auio.uio_iovcnt = 1;
1833         auio.uio_resid = len;
1834         auio.uio_segflg = UIO_SYSSPACE;
1835         auio.uio_rw = UIO_READ;
1836         auio.uio_offset = off;
1837         auio.uio_td = ra->td;
1838 #ifdef _KERNEL
1839         error = fo_read(ra->fp, &auio, ra->td->td_ucred, FOF_OFFSET, ra->td);
1840 #else
1841         fprintf(stderr, "%s: returning EOPNOTSUPP\n", __func__);
1842         error = EOPNOTSUPP;
1843 #endif
1844         *resid = auio.uio_resid;
1845         return (error);
1846 }
1847
1848 static int
1849 receive_read(struct receive_arg *ra, int len, void *buf)
1850 {
1851         int done = 0;
1852
1853         /*
1854          * The code doesn't rely on this (lengths being multiples of 8).  See
1855          * comment in dump_bytes.
1856          */
1857         ASSERT0(len % 8);
1858
1859         while (done < len) {
1860                 ssize_t resid;
1861
1862                 ra->err = restore_bytes(ra, buf + done,
1863                     len - done, ra->voff, &resid);
1864
1865                 if (resid == len - done) {
1866                         /*
1867                          * Note: ECKSUM indicates that the receive
1868                          * was interrupted and can potentially be resumed.
1869                          */
1870                         ra->err = SET_ERROR(ECKSUM);
1871                 }
1872                 ra->voff += len - done - resid;
1873                 done = len - resid;
1874                 if (ra->err != 0)
1875                         return (ra->err);
1876         }
1877
1878         ra->bytes_read += len;
1879
1880         ASSERT3U(done, ==, len);
1881         return (0);
1882 }
1883
1884 static void
1885 byteswap_record(dmu_replay_record_t *drr)
1886 {
1887 #define DO64(X) (drr->drr_u.X = BSWAP_64(drr->drr_u.X))
1888 #define DO32(X) (drr->drr_u.X = BSWAP_32(drr->drr_u.X))
1889         drr->drr_type = BSWAP_32(drr->drr_type);
1890         drr->drr_payloadlen = BSWAP_32(drr->drr_payloadlen);
1891
1892         switch (drr->drr_type) {
1893         case DRR_BEGIN:
1894                 DO64(drr_begin.drr_magic);
1895                 DO64(drr_begin.drr_versioninfo);
1896                 DO64(drr_begin.drr_creation_time);
1897                 DO32(drr_begin.drr_type);
1898                 DO32(drr_begin.drr_flags);
1899                 DO64(drr_begin.drr_toguid);
1900                 DO64(drr_begin.drr_fromguid);
1901                 break;
1902         case DRR_OBJECT:
1903                 DO64(drr_object.drr_object);
1904                 DO32(drr_object.drr_type);
1905                 DO32(drr_object.drr_bonustype);
1906                 DO32(drr_object.drr_blksz);
1907                 DO32(drr_object.drr_bonuslen);
1908                 DO64(drr_object.drr_toguid);
1909                 break;
1910         case DRR_FREEOBJECTS:
1911                 DO64(drr_freeobjects.drr_firstobj);
1912                 DO64(drr_freeobjects.drr_numobjs);
1913                 DO64(drr_freeobjects.drr_toguid);
1914                 break;
1915         case DRR_WRITE:
1916                 DO64(drr_write.drr_object);
1917                 DO32(drr_write.drr_type);
1918                 DO64(drr_write.drr_offset);
1919                 DO64(drr_write.drr_length);
1920                 DO64(drr_write.drr_toguid);
1921                 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_write.drr_key.ddk_cksum);
1922                 DO64(drr_write.drr_key.ddk_prop);
1923                 break;
1924         case DRR_WRITE_BYREF:
1925                 DO64(drr_write_byref.drr_object);
1926                 DO64(drr_write_byref.drr_offset);
1927                 DO64(drr_write_byref.drr_length);
1928                 DO64(drr_write_byref.drr_toguid);
1929                 DO64(drr_write_byref.drr_refguid);
1930                 DO64(drr_write_byref.drr_refobject);
1931                 DO64(drr_write_byref.drr_refoffset);
1932                 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_write_byref.
1933                     drr_key.ddk_cksum);
1934                 DO64(drr_write_byref.drr_key.ddk_prop);
1935                 break;
1936         case DRR_WRITE_EMBEDDED:
1937                 DO64(drr_write_embedded.drr_object);
1938                 DO64(drr_write_embedded.drr_offset);
1939                 DO64(drr_write_embedded.drr_length);
1940                 DO64(drr_write_embedded.drr_toguid);
1941                 DO32(drr_write_embedded.drr_lsize);
1942                 DO32(drr_write_embedded.drr_psize);
1943                 break;
1944         case DRR_FREE:
1945                 DO64(drr_free.drr_object);
1946                 DO64(drr_free.drr_offset);
1947                 DO64(drr_free.drr_length);
1948                 DO64(drr_free.drr_toguid);
1949                 break;
1950         case DRR_SPILL:
1951                 DO64(drr_spill.drr_object);
1952                 DO64(drr_spill.drr_length);
1953                 DO64(drr_spill.drr_toguid);
1954                 break;
1955         case DRR_END:
1956                 DO64(drr_end.drr_toguid);
1957                 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_end.drr_checksum);
1958                 break;
1959         }
1960
1961         if (drr->drr_type != DRR_BEGIN) {
1962                 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_checksum.drr_checksum);
1963         }
1964
1965 #undef DO64
1966 #undef DO32
1967 }
1968
1969 static inline uint8_t
1970 deduce_nblkptr(dmu_object_type_t bonus_type, uint64_t bonus_size)
1971 {
1972         if (bonus_type == DMU_OT_SA) {
1973                 return (1);
1974         } else {
1975                 return (1 +
1976                     ((DN_MAX_BONUSLEN - bonus_size) >> SPA_BLKPTRSHIFT));
1977         }
1978 }
1979
1980 static void
1981 save_resume_state(struct receive_writer_arg *rwa,
1982     uint64_t object, uint64_t offset, dmu_tx_t *tx)
1983 {
1984         int txgoff = dmu_tx_get_txg(tx) & TXG_MASK;
1985
1986         if (!rwa->resumable)
1987                 return;
1988
1989         /*
1990          * We use ds_resume_bytes[] != 0 to indicate that we need to
1991          * update this on disk, so it must not be 0.
1992          */
1993         ASSERT(rwa->bytes_read != 0);
1994
1995         /*
1996          * We only resume from write records, which have a valid
1997          * (non-meta-dnode) object number.
1998          */
1999         ASSERT(object != 0);
2000
2001         /*
2002          * For resuming to work correctly, we must receive records in order,
2003          * sorted by object,offset.  This is checked by the callers, but
2004          * assert it here for good measure.
2005          */
2006         ASSERT3U(object, >=, rwa->os->os_dsl_dataset->ds_resume_object[txgoff]);
2007         ASSERT(object != rwa->os->os_dsl_dataset->ds_resume_object[txgoff] ||
2008             offset >= rwa->os->os_dsl_dataset->ds_resume_offset[txgoff]);
2009         ASSERT3U(rwa->bytes_read, >=,
2010             rwa->os->os_dsl_dataset->ds_resume_bytes[txgoff]);
2011
2012         rwa->os->os_dsl_dataset->ds_resume_object[txgoff] = object;
2013         rwa->os->os_dsl_dataset->ds_resume_offset[txgoff] = offset;
2014         rwa->os->os_dsl_dataset->ds_resume_bytes[txgoff] = rwa->bytes_read;
2015 }
2016
2017 static int
2018 receive_object(struct receive_writer_arg *rwa, struct drr_object *drro,
2019     void *data)
2020 {
2021         dmu_object_info_t doi;
2022         dmu_tx_t *tx;
2023         uint64_t object;
2024         int err;
2025
2026         if (drro->drr_type == DMU_OT_NONE ||
2027             !DMU_OT_IS_VALID(drro->drr_type) ||
2028             !DMU_OT_IS_VALID(drro->drr_bonustype) ||
2029             drro->drr_checksumtype >= ZIO_CHECKSUM_FUNCTIONS ||
2030             drro->drr_compress >= ZIO_COMPRESS_FUNCTIONS ||
2031             P2PHASE(drro->drr_blksz, SPA_MINBLOCKSIZE) ||
2032             drro->drr_blksz < SPA_MINBLOCKSIZE ||
2033             drro->drr_blksz > spa_maxblocksize(dmu_objset_spa(rwa->os)) ||
2034             drro->drr_bonuslen > DN_MAX_BONUSLEN) {
2035                 return (SET_ERROR(EINVAL));
2036         }
2037
2038         err = dmu_object_info(rwa->os, drro->drr_object, &doi);
2039
2040         if (err != 0 && err != ENOENT)
2041                 return (SET_ERROR(EINVAL));
2042         object = err == 0 ? drro->drr_object : DMU_NEW_OBJECT;
2043
2044         /*
2045          * If we are losing blkptrs or changing the block size this must
2046          * be a new file instance.  We must clear out the previous file
2047          * contents before we can change this type of metadata in the dnode.
2048          */
2049         if (err == 0) {
2050                 int nblkptr;
2051
2052                 nblkptr = deduce_nblkptr(drro->drr_bonustype,
2053                     drro->drr_bonuslen);
2054
2055                 if (drro->drr_blksz != doi.doi_data_block_size ||
2056                     nblkptr < doi.doi_nblkptr) {
2057                         err = dmu_free_long_range(rwa->os, drro->drr_object,
2058                             0, DMU_OBJECT_END);
2059                         if (err != 0)
2060                                 return (SET_ERROR(EINVAL));
2061                 }
2062         }
2063
2064         tx = dmu_tx_create(rwa->os);
2065         dmu_tx_hold_bonus(tx, object);
2066         err = dmu_tx_assign(tx, TXG_WAIT);
2067         if (err != 0) {
2068                 dmu_tx_abort(tx);
2069                 return (err);
2070         }
2071
2072         if (object == DMU_NEW_OBJECT) {
2073                 /* currently free, want to be allocated */
2074                 err = dmu_object_claim(rwa->os, drro->drr_object,
2075                     drro->drr_type, drro->drr_blksz,
2076                     drro->drr_bonustype, drro->drr_bonuslen, tx);
2077         } else if (drro->drr_type != doi.doi_type ||
2078             drro->drr_blksz != doi.doi_data_block_size ||
2079             drro->drr_bonustype != doi.doi_bonus_type ||
2080             drro->drr_bonuslen != doi.doi_bonus_size) {
2081                 /* currently allocated, but with different properties */
2082                 err = dmu_object_reclaim(rwa->os, drro->drr_object,
2083                     drro->drr_type, drro->drr_blksz,
2084                     drro->drr_bonustype, drro->drr_bonuslen, tx);
2085         }
2086         if (err != 0) {
2087                 dmu_tx_commit(tx);
2088                 return (SET_ERROR(EINVAL));
2089         }
2090
2091         dmu_object_set_checksum(rwa->os, drro->drr_object,
2092             drro->drr_checksumtype, tx);
2093         dmu_object_set_compress(rwa->os, drro->drr_object,
2094             drro->drr_compress, tx);
2095
2096         if (data != NULL) {
2097                 dmu_buf_t *db;
2098
2099                 VERIFY0(dmu_bonus_hold(rwa->os, drro->drr_object, FTAG, &db));
2100                 dmu_buf_will_dirty(db, tx);
2101
2102                 ASSERT3U(db->db_size, >=, drro->drr_bonuslen);
2103                 bcopy(data, db->db_data, drro->drr_bonuslen);
2104                 if (rwa->byteswap) {
2105                         dmu_object_byteswap_t byteswap =
2106                             DMU_OT_BYTESWAP(drro->drr_bonustype);
2107                         dmu_ot_byteswap[byteswap].ob_func(db->db_data,
2108                             drro->drr_bonuslen);
2109                 }
2110                 dmu_buf_rele(db, FTAG);
2111         }
2112         dmu_tx_commit(tx);
2113
2114         return (0);
2115 }
2116
2117 /* ARGSUSED */
2118 static int
2119 receive_freeobjects(struct receive_writer_arg *rwa,
2120     struct drr_freeobjects *drrfo)
2121 {
2122         uint64_t obj;
2123         int next_err = 0;
2124
2125         if (drrfo->drr_firstobj + drrfo->drr_numobjs < drrfo->drr_firstobj)
2126                 return (SET_ERROR(EINVAL));
2127
2128         for (obj = drrfo->drr_firstobj;
2129             obj < drrfo->drr_firstobj + drrfo->drr_numobjs && next_err == 0;
2130             next_err = dmu_object_next(rwa->os, &obj, FALSE, 0)) {
2131                 int err;
2132
2133                 if (dmu_object_info(rwa->os, obj, NULL) != 0)
2134                         continue;
2135
2136                 err = dmu_free_long_object(rwa->os, obj);
2137                 if (err != 0)
2138                         return (err);
2139         }
2140         if (next_err != ESRCH)
2141                 return (next_err);
2142         return (0);
2143 }
2144
2145 static int
2146 receive_write(struct receive_writer_arg *rwa, struct drr_write *drrw,
2147     arc_buf_t *abuf)
2148 {
2149         dmu_tx_t *tx;
2150         int err;
2151
2152         if (drrw->drr_offset + drrw->drr_length < drrw->drr_offset ||
2153             !DMU_OT_IS_VALID(drrw->drr_type))
2154                 return (SET_ERROR(EINVAL));
2155
2156         /*
2157          * For resuming to work, records must be in increasing order
2158          * by (object, offset).
2159          */
2160         if (drrw->drr_object < rwa->last_object ||
2161             (drrw->drr_object == rwa->last_object &&
2162             drrw->drr_offset < rwa->last_offset)) {
2163                 return (SET_ERROR(EINVAL));
2164         }
2165         rwa->last_object = drrw->drr_object;
2166         rwa->last_offset = drrw->drr_offset;
2167
2168         if (dmu_object_info(rwa->os, drrw->drr_object, NULL) != 0)
2169                 return (SET_ERROR(EINVAL));
2170
2171         tx = dmu_tx_create(rwa->os);
2172
2173         dmu_tx_hold_write(tx, drrw->drr_object,
2174             drrw->drr_offset, drrw->drr_length);
2175         err = dmu_tx_assign(tx, TXG_WAIT);
2176         if (err != 0) {
2177                 dmu_tx_abort(tx);
2178                 return (err);
2179         }
2180         if (rwa->byteswap) {
2181                 dmu_object_byteswap_t byteswap =
2182                     DMU_OT_BYTESWAP(drrw->drr_type);
2183                 dmu_ot_byteswap[byteswap].ob_func(abuf->b_data,
2184                     drrw->drr_length);
2185         }
2186
2187         dmu_buf_t *bonus;
2188         if (dmu_bonus_hold(rwa->os, drrw->drr_object, FTAG, &bonus) != 0)
2189                 return (SET_ERROR(EINVAL));
2190         dmu_assign_arcbuf(bonus, drrw->drr_offset, abuf, tx);
2191
2192         /*
2193          * Note: If the receive fails, we want the resume stream to start
2194          * with the same record that we last successfully received (as opposed
2195          * to the next record), so that we can verify that we are
2196          * resuming from the correct location.
2197          */
2198         save_resume_state(rwa, drrw->drr_object, drrw->drr_offset, tx);
2199         dmu_tx_commit(tx);
2200         dmu_buf_rele(bonus, FTAG);
2201
2202         return (0);
2203 }
2204
2205 /*
2206  * Handle a DRR_WRITE_BYREF record.  This record is used in dedup'ed
2207  * streams to refer to a copy of the data that is already on the
2208  * system because it came in earlier in the stream.  This function
2209  * finds the earlier copy of the data, and uses that copy instead of
2210  * data from the stream to fulfill this write.
2211  */
2212 static int
2213 receive_write_byref(struct receive_writer_arg *rwa,
2214     struct drr_write_byref *drrwbr)
2215 {
2216         dmu_tx_t *tx;
2217         int err;
2218         guid_map_entry_t gmesrch;
2219         guid_map_entry_t *gmep;
2220         avl_index_t where;
2221         objset_t *ref_os = NULL;
2222         dmu_buf_t *dbp;
2223
2224         if (drrwbr->drr_offset + drrwbr->drr_length < drrwbr->drr_offset)
2225                 return (SET_ERROR(EINVAL));
2226
2227         /*
2228          * If the GUID of the referenced dataset is different from the
2229          * GUID of the target dataset, find the referenced dataset.
2230          */
2231         if (drrwbr->drr_toguid != drrwbr->drr_refguid) {
2232                 gmesrch.guid = drrwbr->drr_refguid;
2233                 if ((gmep = avl_find(rwa->guid_to_ds_map, &gmesrch,
2234                     &where)) == NULL) {
2235                         return (SET_ERROR(EINVAL));
2236                 }
2237                 if (dmu_objset_from_ds(gmep->gme_ds, &ref_os))
2238                         return (SET_ERROR(EINVAL));
2239         } else {
2240                 ref_os = rwa->os;
2241         }
2242
2243         err = dmu_buf_hold(ref_os, drrwbr->drr_refobject,
2244             drrwbr->drr_refoffset, FTAG, &dbp, DMU_READ_PREFETCH);
2245         if (err != 0)
2246                 return (err);
2247
2248         tx = dmu_tx_create(rwa->os);
2249
2250         dmu_tx_hold_write(tx, drrwbr->drr_object,
2251             drrwbr->drr_offset, drrwbr->drr_length);
2252         err = dmu_tx_assign(tx, TXG_WAIT);
2253         if (err != 0) {
2254                 dmu_tx_abort(tx);
2255                 return (err);
2256         }
2257         dmu_write(rwa->os, drrwbr->drr_object,
2258             drrwbr->drr_offset, drrwbr->drr_length, dbp->db_data, tx);
2259         dmu_buf_rele(dbp, FTAG);
2260
2261         /* See comment in restore_write. */
2262         save_resume_state(rwa, drrwbr->drr_object, drrwbr->drr_offset, tx);
2263         dmu_tx_commit(tx);
2264         return (0);
2265 }
2266
2267 static int
2268 receive_write_embedded(struct receive_writer_arg *rwa,
2269     struct drr_write_embedded *drrwe, void *data)
2270 {
2271         dmu_tx_t *tx;
2272         int err;
2273
2274         if (drrwe->drr_offset + drrwe->drr_length < drrwe->drr_offset)
2275                 return (EINVAL);
2276
2277         if (drrwe->drr_psize > BPE_PAYLOAD_SIZE)
2278                 return (EINVAL);
2279
2280         if (drrwe->drr_etype >= NUM_BP_EMBEDDED_TYPES)
2281                 return (EINVAL);
2282         if (drrwe->drr_compression >= ZIO_COMPRESS_FUNCTIONS)
2283                 return (EINVAL);
2284
2285         tx = dmu_tx_create(rwa->os);
2286
2287         dmu_tx_hold_write(tx, drrwe->drr_object,
2288             drrwe->drr_offset, drrwe->drr_length);
2289         err = dmu_tx_assign(tx, TXG_WAIT);
2290         if (err != 0) {
2291                 dmu_tx_abort(tx);
2292                 return (err);
2293         }
2294
2295         dmu_write_embedded(rwa->os, drrwe->drr_object,
2296             drrwe->drr_offset, data, drrwe->drr_etype,
2297             drrwe->drr_compression, drrwe->drr_lsize, drrwe->drr_psize,
2298             rwa->byteswap ^ ZFS_HOST_BYTEORDER, tx);
2299
2300         /* See comment in restore_write. */
2301         save_resume_state(rwa, drrwe->drr_object, drrwe->drr_offset, tx);
2302         dmu_tx_commit(tx);
2303         return (0);
2304 }
2305
2306 static int
2307 receive_spill(struct receive_writer_arg *rwa, struct drr_spill *drrs,
2308     void *data)
2309 {
2310         dmu_tx_t *tx;
2311         dmu_buf_t *db, *db_spill;
2312         int err;
2313
2314         if (drrs->drr_length < SPA_MINBLOCKSIZE ||
2315             drrs->drr_length > spa_maxblocksize(dmu_objset_spa(rwa->os)))
2316                 return (SET_ERROR(EINVAL));
2317
2318         if (dmu_object_info(rwa->os, drrs->drr_object, NULL) != 0)
2319                 return (SET_ERROR(EINVAL));
2320
2321         VERIFY0(dmu_bonus_hold(rwa->os, drrs->drr_object, FTAG, &db));
2322         if ((err = dmu_spill_hold_by_bonus(db, FTAG, &db_spill)) != 0) {
2323                 dmu_buf_rele(db, FTAG);
2324                 return (err);
2325         }
2326
2327         tx = dmu_tx_create(rwa->os);
2328
2329         dmu_tx_hold_spill(tx, db->db_object);
2330
2331         err = dmu_tx_assign(tx, TXG_WAIT);
2332         if (err != 0) {
2333                 dmu_buf_rele(db, FTAG);
2334                 dmu_buf_rele(db_spill, FTAG);
2335                 dmu_tx_abort(tx);
2336                 return (err);
2337         }
2338         dmu_buf_will_dirty(db_spill, tx);
2339
2340         if (db_spill->db_size < drrs->drr_length)
2341                 VERIFY(0 == dbuf_spill_set_blksz(db_spill,
2342                     drrs->drr_length, tx));
2343         bcopy(data, db_spill->db_data, drrs->drr_length);
2344
2345         dmu_buf_rele(db, FTAG);
2346         dmu_buf_rele(db_spill, FTAG);
2347
2348         dmu_tx_commit(tx);
2349         return (0);
2350 }
2351
2352 /* ARGSUSED */
2353 static int
2354 receive_free(struct receive_writer_arg *rwa, struct drr_free *drrf)
2355 {
2356         int err;
2357
2358         if (drrf->drr_length != -1ULL &&
2359             drrf->drr_offset + drrf->drr_length < drrf->drr_offset)
2360                 return (SET_ERROR(EINVAL));
2361
2362         if (dmu_object_info(rwa->os, drrf->drr_object, NULL) != 0)
2363                 return (SET_ERROR(EINVAL));
2364
2365         err = dmu_free_long_range(rwa->os, drrf->drr_object,
2366             drrf->drr_offset, drrf->drr_length);
2367
2368         return (err);
2369 }
2370
2371 /* used to destroy the drc_ds on error */
2372 static void
2373 dmu_recv_cleanup_ds(dmu_recv_cookie_t *drc)
2374 {
2375         if (drc->drc_resumable) {
2376                 /* wait for our resume state to be written to disk */
2377                 txg_wait_synced(drc->drc_ds->ds_dir->dd_pool, 0);
2378                 dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
2379         } else {
2380                 char name[ZFS_MAX_DATASET_NAME_LEN];
2381                 dsl_dataset_name(drc->drc_ds, name);
2382                 dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
2383                 (void) dsl_destroy_head(name);
2384         }
2385 }
2386
2387 static void
2388 receive_cksum(struct receive_arg *ra, int len, void *buf)
2389 {
2390         if (ra->byteswap) {
2391                 fletcher_4_incremental_byteswap(buf, len, &ra->cksum);
2392         } else {
2393                 fletcher_4_incremental_native(buf, len, &ra->cksum);
2394         }
2395 }
2396
2397 /*
2398  * Read the payload into a buffer of size len, and update the current record's
2399  * payload field.
2400  * Allocate ra->next_rrd and read the next record's header into
2401  * ra->next_rrd->header.
2402  * Verify checksum of payload and next record.
2403  */
2404 static int
2405 receive_read_payload_and_next_header(struct receive_arg *ra, int len, void *buf)
2406 {
2407         int err;
2408
2409         if (len != 0) {
2410                 ASSERT3U(len, <=, SPA_MAXBLOCKSIZE);
2411                 err = receive_read(ra, len, buf);
2412                 if (err != 0)
2413                         return (err);
2414                 receive_cksum(ra, len, buf);
2415
2416                 /* note: rrd is NULL when reading the begin record's payload */
2417                 if (ra->rrd != NULL) {
2418                         ra->rrd->payload = buf;
2419                         ra->rrd->payload_size = len;
2420                         ra->rrd->bytes_read = ra->bytes_read;
2421                 }
2422         }
2423
2424         ra->prev_cksum = ra->cksum;
2425
2426         ra->next_rrd = kmem_zalloc(sizeof (*ra->next_rrd), KM_SLEEP);
2427         err = receive_read(ra, sizeof (ra->next_rrd->header),
2428             &ra->next_rrd->header);
2429         ra->next_rrd->bytes_read = ra->bytes_read;
2430         if (err != 0) {
2431                 kmem_free(ra->next_rrd, sizeof (*ra->next_rrd));
2432                 ra->next_rrd = NULL;
2433                 return (err);
2434         }
2435         if (ra->next_rrd->header.drr_type == DRR_BEGIN) {
2436                 kmem_free(ra->next_rrd, sizeof (*ra->next_rrd));
2437                 ra->next_rrd = NULL;
2438                 return (SET_ERROR(EINVAL));
2439         }
2440
2441         /*
2442          * Note: checksum is of everything up to but not including the
2443          * checksum itself.
2444          */
2445         ASSERT3U(offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
2446             ==, sizeof (dmu_replay_record_t) - sizeof (zio_cksum_t));
2447         receive_cksum(ra,
2448             offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
2449             &ra->next_rrd->header);
2450
2451         zio_cksum_t cksum_orig =
2452             ra->next_rrd->header.drr_u.drr_checksum.drr_checksum;
2453         zio_cksum_t *cksump =
2454             &ra->next_rrd->header.drr_u.drr_checksum.drr_checksum;
2455
2456         if (ra->byteswap)
2457                 byteswap_record(&ra->next_rrd->header);
2458
2459         if ((!ZIO_CHECKSUM_IS_ZERO(cksump)) &&
2460             !ZIO_CHECKSUM_EQUAL(ra->cksum, *cksump)) {
2461                 kmem_free(ra->next_rrd, sizeof (*ra->next_rrd));
2462                 ra->next_rrd = NULL;
2463                 return (SET_ERROR(ECKSUM));
2464         }
2465
2466         receive_cksum(ra, sizeof (cksum_orig), &cksum_orig);
2467
2468         return (0);
2469 }
2470
2471 static void
2472 objlist_create(struct objlist *list)
2473 {
2474         list_create(&list->list, sizeof (struct receive_objnode),
2475             offsetof(struct receive_objnode, node));
2476         list->last_lookup = 0;
2477 }
2478
2479 static void
2480 objlist_destroy(struct objlist *list)
2481 {
2482         for (struct receive_objnode *n = list_remove_head(&list->list);
2483             n != NULL; n = list_remove_head(&list->list)) {
2484                 kmem_free(n, sizeof (*n));
2485         }
2486         list_destroy(&list->list);
2487 }
2488
2489 /*
2490  * This function looks through the objlist to see if the specified object number
2491  * is contained in the objlist.  In the process, it will remove all object
2492  * numbers in the list that are smaller than the specified object number.  Thus,
2493  * any lookup of an object number smaller than a previously looked up object
2494  * number will always return false; therefore, all lookups should be done in
2495  * ascending order.
2496  */
2497 static boolean_t
2498 objlist_exists(struct objlist *list, uint64_t object)
2499 {
2500         struct receive_objnode *node = list_head(&list->list);
2501         ASSERT3U(object, >=, list->last_lookup);
2502         list->last_lookup = object;
2503         while (node != NULL && node->object < object) {
2504                 VERIFY3P(node, ==, list_remove_head(&list->list));
2505                 kmem_free(node, sizeof (*node));
2506                 node = list_head(&list->list);
2507         }
2508         return (node != NULL && node->object == object);
2509 }
2510
2511 /*
2512  * The objlist is a list of object numbers stored in ascending order.  However,
2513  * the insertion of new object numbers does not seek out the correct location to
2514  * store a new object number; instead, it appends it to the list for simplicity.
2515  * Thus, any users must take care to only insert new object numbers in ascending
2516  * order.
2517  */
2518 static void
2519 objlist_insert(struct objlist *list, uint64_t object)
2520 {
2521         struct receive_objnode *node = kmem_zalloc(sizeof (*node), KM_SLEEP);
2522         node->object = object;
2523 #ifdef ZFS_DEBUG
2524         struct receive_objnode *last_object = list_tail(&list->list);
2525         uint64_t last_objnum = (last_object != NULL ? last_object->object : 0);
2526         ASSERT3U(node->object, >, last_objnum);
2527 #endif
2528         list_insert_tail(&list->list, node);
2529 }
2530
2531 /*
2532  * Issue the prefetch reads for any necessary indirect blocks.
2533  *
2534  * We use the object ignore list to tell us whether or not to issue prefetches
2535  * for a given object.  We do this for both correctness (in case the blocksize
2536  * of an object has changed) and performance (if the object doesn't exist, don't
2537  * needlessly try to issue prefetches).  We also trim the list as we go through
2538  * the stream to prevent it from growing to an unbounded size.
2539  *
2540  * The object numbers within will always be in sorted order, and any write
2541  * records we see will also be in sorted order, but they're not sorted with
2542  * respect to each other (i.e. we can get several object records before
2543  * receiving each object's write records).  As a result, once we've reached a
2544  * given object number, we can safely remove any reference to lower object
2545  * numbers in the ignore list. In practice, we receive up to 32 object records
2546  * before receiving write records, so the list can have up to 32 nodes in it.
2547  */
2548 /* ARGSUSED */
2549 static void
2550 receive_read_prefetch(struct receive_arg *ra,
2551     uint64_t object, uint64_t offset, uint64_t length)
2552 {
2553         if (!objlist_exists(&ra->ignore_objlist, object)) {
2554                 dmu_prefetch(ra->os, object, 1, offset, length,
2555                     ZIO_PRIORITY_SYNC_READ);
2556         }
2557 }
2558
2559 /*
2560  * Read records off the stream, issuing any necessary prefetches.
2561  */
2562 static int
2563 receive_read_record(struct receive_arg *ra)
2564 {
2565         int err;
2566
2567         switch (ra->rrd->header.drr_type) {
2568         case DRR_OBJECT:
2569         {
2570                 struct drr_object *drro = &ra->rrd->header.drr_u.drr_object;
2571                 uint32_t size = P2ROUNDUP(drro->drr_bonuslen, 8);
2572                 void *buf = kmem_zalloc(size, KM_SLEEP);
2573                 dmu_object_info_t doi;
2574                 err = receive_read_payload_and_next_header(ra, size, buf);
2575                 if (err != 0) {
2576                         kmem_free(buf, size);
2577                         return (err);
2578                 }
2579                 err = dmu_object_info(ra->os, drro->drr_object, &doi);
2580                 /*
2581                  * See receive_read_prefetch for an explanation why we're
2582                  * storing this object in the ignore_obj_list.
2583                  */
2584                 if (err == ENOENT ||
2585                     (err == 0 && doi.doi_data_block_size != drro->drr_blksz)) {
2586                         objlist_insert(&ra->ignore_objlist, drro->drr_object);
2587                         err = 0;
2588                 }
2589                 return (err);
2590         }
2591         case DRR_FREEOBJECTS:
2592         {
2593                 err = receive_read_payload_and_next_header(ra, 0, NULL);
2594                 return (err);
2595         }
2596         case DRR_WRITE:
2597         {
2598                 struct drr_write *drrw = &ra->rrd->header.drr_u.drr_write;
2599                 arc_buf_t *abuf = arc_loan_buf(dmu_objset_spa(ra->os),
2600                     drrw->drr_length);
2601
2602                 err = receive_read_payload_and_next_header(ra,
2603                     drrw->drr_length, abuf->b_data);
2604                 if (err != 0) {
2605                         dmu_return_arcbuf(abuf);
2606                         return (err);
2607                 }
2608                 ra->rrd->write_buf = abuf;
2609                 receive_read_prefetch(ra, drrw->drr_object, drrw->drr_offset,
2610                     drrw->drr_length);
2611                 return (err);
2612         }
2613         case DRR_WRITE_BYREF:
2614         {
2615                 struct drr_write_byref *drrwb =
2616                     &ra->rrd->header.drr_u.drr_write_byref;
2617                 err = receive_read_payload_and_next_header(ra, 0, NULL);
2618                 receive_read_prefetch(ra, drrwb->drr_object, drrwb->drr_offset,
2619                     drrwb->drr_length);
2620                 return (err);
2621         }
2622         case DRR_WRITE_EMBEDDED:
2623         {
2624                 struct drr_write_embedded *drrwe =
2625                     &ra->rrd->header.drr_u.drr_write_embedded;
2626                 uint32_t size = P2ROUNDUP(drrwe->drr_psize, 8);
2627                 void *buf = kmem_zalloc(size, KM_SLEEP);
2628
2629                 err = receive_read_payload_and_next_header(ra, size, buf);
2630                 if (err != 0) {
2631                         kmem_free(buf, size);
2632                         return (err);
2633                 }
2634
2635                 receive_read_prefetch(ra, drrwe->drr_object, drrwe->drr_offset,
2636                     drrwe->drr_length);
2637                 return (err);
2638         }
2639         case DRR_FREE:
2640         {
2641                 /*
2642                  * It might be beneficial to prefetch indirect blocks here, but
2643                  * we don't really have the data to decide for sure.
2644                  */
2645                 err = receive_read_payload_and_next_header(ra, 0, NULL);
2646                 return (err);
2647         }
2648         case DRR_END:
2649         {
2650                 struct drr_end *drre = &ra->rrd->header.drr_u.drr_end;
2651                 if (!ZIO_CHECKSUM_EQUAL(ra->prev_cksum, drre->drr_checksum))
2652                         return (SET_ERROR(ECKSUM));
2653                 return (0);
2654         }
2655         case DRR_SPILL:
2656         {
2657                 struct drr_spill *drrs = &ra->rrd->header.drr_u.drr_spill;
2658                 void *buf = kmem_zalloc(drrs->drr_length, KM_SLEEP);
2659                 err = receive_read_payload_and_next_header(ra, drrs->drr_length,
2660                     buf);
2661                 if (err != 0)
2662                         kmem_free(buf, drrs->drr_length);
2663                 return (err);
2664         }
2665         default:
2666                 return (SET_ERROR(EINVAL));
2667         }
2668 }
2669
2670 /*
2671  * Commit the records to the pool.
2672  */
2673 static int
2674 receive_process_record(struct receive_writer_arg *rwa,
2675     struct receive_record_arg *rrd)
2676 {
2677         int err;
2678
2679         /* Processing in order, therefore bytes_read should be increasing. */
2680         ASSERT3U(rrd->bytes_read, >=, rwa->bytes_read);
2681         rwa->bytes_read = rrd->bytes_read;
2682
2683         switch (rrd->header.drr_type) {
2684         case DRR_OBJECT:
2685         {
2686                 struct drr_object *drro = &rrd->header.drr_u.drr_object;
2687                 err = receive_object(rwa, drro, rrd->payload);
2688                 kmem_free(rrd->payload, rrd->payload_size);
2689                 rrd->payload = NULL;
2690                 return (err);
2691         }
2692         case DRR_FREEOBJECTS:
2693         {
2694                 struct drr_freeobjects *drrfo =
2695                     &rrd->header.drr_u.drr_freeobjects;
2696                 return (receive_freeobjects(rwa, drrfo));
2697         }
2698         case DRR_WRITE:
2699         {
2700                 struct drr_write *drrw = &rrd->header.drr_u.drr_write;
2701                 err = receive_write(rwa, drrw, rrd->write_buf);
2702                 /* if receive_write() is successful, it consumes the arc_buf */
2703                 if (err != 0)
2704                         dmu_return_arcbuf(rrd->write_buf);
2705                 rrd->write_buf = NULL;
2706                 rrd->payload = NULL;
2707                 return (err);
2708         }
2709         case DRR_WRITE_BYREF:
2710         {
2711                 struct drr_write_byref *drrwbr =
2712                     &rrd->header.drr_u.drr_write_byref;
2713                 return (receive_write_byref(rwa, drrwbr));
2714         }
2715         case DRR_WRITE_EMBEDDED:
2716         {
2717                 struct drr_write_embedded *drrwe =
2718                     &rrd->header.drr_u.drr_write_embedded;
2719                 err = receive_write_embedded(rwa, drrwe, rrd->payload);
2720                 kmem_free(rrd->payload, rrd->payload_size);
2721                 rrd->payload = NULL;
2722                 return (err);
2723         }
2724         case DRR_FREE:
2725         {
2726                 struct drr_free *drrf = &rrd->header.drr_u.drr_free;
2727                 return (receive_free(rwa, drrf));
2728         }
2729         case DRR_SPILL:
2730         {
2731                 struct drr_spill *drrs = &rrd->header.drr_u.drr_spill;
2732                 err = receive_spill(rwa, drrs, rrd->payload);
2733                 kmem_free(rrd->payload, rrd->payload_size);
2734                 rrd->payload = NULL;
2735                 return (err);
2736         }
2737         default:
2738                 return (SET_ERROR(EINVAL));
2739         }
2740 }
2741
2742 /*
2743  * dmu_recv_stream's worker thread; pull records off the queue, and then call
2744  * receive_process_record  When we're done, signal the main thread and exit.
2745  */
2746 static void
2747 receive_writer_thread(void *arg)
2748 {
2749         struct receive_writer_arg *rwa = arg;
2750         struct receive_record_arg *rrd;
2751         for (rrd = bqueue_dequeue(&rwa->q); !rrd->eos_marker;
2752             rrd = bqueue_dequeue(&rwa->q)) {
2753                 /*
2754                  * If there's an error, the main thread will stop putting things
2755                  * on the queue, but we need to clear everything in it before we
2756                  * can exit.
2757                  */
2758                 if (rwa->err == 0) {
2759                         rwa->err = receive_process_record(rwa, rrd);
2760                 } else if (rrd->write_buf != NULL) {
2761                         dmu_return_arcbuf(rrd->write_buf);
2762                         rrd->write_buf = NULL;
2763                         rrd->payload = NULL;
2764                 } else if (rrd->payload != NULL) {
2765                         kmem_free(rrd->payload, rrd->payload_size);
2766                         rrd->payload = NULL;
2767                 }
2768                 kmem_free(rrd, sizeof (*rrd));
2769         }
2770         kmem_free(rrd, sizeof (*rrd));
2771         mutex_enter(&rwa->mutex);
2772         rwa->done = B_TRUE;
2773         cv_signal(&rwa->cv);
2774         mutex_exit(&rwa->mutex);
2775         thread_exit();
2776 }
2777
2778 static int
2779 resume_check(struct receive_arg *ra, nvlist_t *begin_nvl)
2780 {
2781         uint64_t val;
2782         objset_t *mos = dmu_objset_pool(ra->os)->dp_meta_objset;
2783         uint64_t dsobj = dmu_objset_id(ra->os);
2784         uint64_t resume_obj, resume_off;
2785
2786         if (nvlist_lookup_uint64(begin_nvl,
2787             "resume_object", &resume_obj) != 0 ||
2788             nvlist_lookup_uint64(begin_nvl,
2789             "resume_offset", &resume_off) != 0) {
2790                 return (SET_ERROR(EINVAL));
2791         }
2792         VERIFY0(zap_lookup(mos, dsobj,
2793             DS_FIELD_RESUME_OBJECT, sizeof (val), 1, &val));
2794         if (resume_obj != val)
2795                 return (SET_ERROR(EINVAL));
2796         VERIFY0(zap_lookup(mos, dsobj,
2797             DS_FIELD_RESUME_OFFSET, sizeof (val), 1, &val));
2798         if (resume_off != val)
2799                 return (SET_ERROR(EINVAL));
2800
2801         return (0);
2802 }
2803
2804 /*
2805  * Read in the stream's records, one by one, and apply them to the pool.  There
2806  * are two threads involved; the thread that calls this function will spin up a
2807  * worker thread, read the records off the stream one by one, and issue
2808  * prefetches for any necessary indirect blocks.  It will then push the records
2809  * onto an internal blocking queue.  The worker thread will pull the records off
2810  * the queue, and actually write the data into the DMU.  This way, the worker
2811  * thread doesn't have to wait for reads to complete, since everything it needs
2812  * (the indirect blocks) will be prefetched.
2813  *
2814  * NB: callers *must* call dmu_recv_end() if this succeeds.
2815  */
2816 int
2817 dmu_recv_stream(dmu_recv_cookie_t *drc, struct file *fp, offset_t *voffp,
2818     int cleanup_fd, uint64_t *action_handlep)
2819 {
2820         int err = 0;
2821         struct receive_arg ra = { 0 };
2822         struct receive_writer_arg rwa = { 0 };
2823         int featureflags;
2824         nvlist_t *begin_nvl = NULL;
2825
2826         ra.byteswap = drc->drc_byteswap;
2827         ra.cksum = drc->drc_cksum;
2828         ra.td = curthread;
2829         ra.fp = fp;
2830         ra.voff = *voffp;
2831
2832         if (dsl_dataset_is_zapified(drc->drc_ds)) {
2833                 (void) zap_lookup(drc->drc_ds->ds_dir->dd_pool->dp_meta_objset,
2834                     drc->drc_ds->ds_object, DS_FIELD_RESUME_BYTES,
2835                     sizeof (ra.bytes_read), 1, &ra.bytes_read);
2836         }
2837
2838         objlist_create(&ra.ignore_objlist);
2839
2840         /* these were verified in dmu_recv_begin */
2841         ASSERT3U(DMU_GET_STREAM_HDRTYPE(drc->drc_drrb->drr_versioninfo), ==,
2842             DMU_SUBSTREAM);
2843         ASSERT3U(drc->drc_drrb->drr_type, <, DMU_OST_NUMTYPES);
2844
2845         /*
2846          * Open the objset we are modifying.
2847          */
2848         VERIFY0(dmu_objset_from_ds(drc->drc_ds, &ra.os));
2849
2850         ASSERT(dsl_dataset_phys(drc->drc_ds)->ds_flags & DS_FLAG_INCONSISTENT);
2851
2852         featureflags = DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo);
2853
2854         /* if this stream is dedup'ed, set up the avl tree for guid mapping */
2855         if (featureflags & DMU_BACKUP_FEATURE_DEDUP) {
2856                 minor_t minor;
2857
2858                 if (cleanup_fd == -1) {
2859                         ra.err = SET_ERROR(EBADF);
2860                         goto out;
2861                 }
2862                 ra.err = zfs_onexit_fd_hold(cleanup_fd, &minor);
2863                 if (ra.err != 0) {
2864                         cleanup_fd = -1;
2865                         goto out;
2866                 }
2867
2868                 if (*action_handlep == 0) {
2869                         rwa.guid_to_ds_map =
2870                             kmem_alloc(sizeof (avl_tree_t), KM_SLEEP);
2871                         avl_create(rwa.guid_to_ds_map, guid_compare,
2872                             sizeof (guid_map_entry_t),
2873                             offsetof(guid_map_entry_t, avlnode));
2874                         err = zfs_onexit_add_cb(minor,
2875                             free_guid_map_onexit, rwa.guid_to_ds_map,
2876                             action_handlep);
2877                         if (ra.err != 0)
2878                                 goto out;
2879                 } else {
2880                         err = zfs_onexit_cb_data(minor, *action_handlep,
2881                             (void **)&rwa.guid_to_ds_map);
2882                         if (ra.err != 0)
2883                                 goto out;
2884                 }
2885
2886                 drc->drc_guid_to_ds_map = rwa.guid_to_ds_map;
2887         }
2888
2889         uint32_t payloadlen = drc->drc_drr_begin->drr_payloadlen;
2890         void *payload = NULL;
2891         if (payloadlen != 0)
2892                 payload = kmem_alloc(payloadlen, KM_SLEEP);
2893
2894         err = receive_read_payload_and_next_header(&ra, payloadlen, payload);
2895         if (err != 0) {
2896                 if (payloadlen != 0)
2897                         kmem_free(payload, payloadlen);
2898                 goto out;
2899         }
2900         if (payloadlen != 0) {
2901                 err = nvlist_unpack(payload, payloadlen, &begin_nvl, KM_SLEEP);
2902                 kmem_free(payload, payloadlen);
2903                 if (err != 0)
2904                         goto out;
2905         }
2906
2907         if (featureflags & DMU_BACKUP_FEATURE_RESUMING) {
2908                 err = resume_check(&ra, begin_nvl);
2909                 if (err != 0)
2910                         goto out;
2911         }
2912
2913         (void) bqueue_init(&rwa.q, zfs_recv_queue_length,
2914             offsetof(struct receive_record_arg, node));
2915         cv_init(&rwa.cv, NULL, CV_DEFAULT, NULL);
2916         mutex_init(&rwa.mutex, NULL, MUTEX_DEFAULT, NULL);
2917         rwa.os = ra.os;
2918         rwa.byteswap = drc->drc_byteswap;
2919         rwa.resumable = drc->drc_resumable;
2920
2921         (void) thread_create(NULL, 0, receive_writer_thread, &rwa, 0, &p0,
2922             TS_RUN, minclsyspri);
2923         /*
2924          * We're reading rwa.err without locks, which is safe since we are the
2925          * only reader, and the worker thread is the only writer.  It's ok if we
2926          * miss a write for an iteration or two of the loop, since the writer
2927          * thread will keep freeing records we send it until we send it an eos
2928          * marker.
2929          *
2930          * We can leave this loop in 3 ways:  First, if rwa.err is
2931          * non-zero.  In that case, the writer thread will free the rrd we just
2932          * pushed.  Second, if  we're interrupted; in that case, either it's the
2933          * first loop and ra.rrd was never allocated, or it's later, and ra.rrd
2934          * has been handed off to the writer thread who will free it.  Finally,
2935          * if receive_read_record fails or we're at the end of the stream, then
2936          * we free ra.rrd and exit.
2937          */
2938         while (rwa.err == 0) {
2939                 if (issig(JUSTLOOKING) && issig(FORREAL)) {
2940                         err = SET_ERROR(EINTR);
2941                         break;
2942                 }
2943
2944                 ASSERT3P(ra.rrd, ==, NULL);
2945                 ra.rrd = ra.next_rrd;
2946                 ra.next_rrd = NULL;
2947                 /* Allocates and loads header into ra.next_rrd */
2948                 err = receive_read_record(&ra);
2949
2950                 if (ra.rrd->header.drr_type == DRR_END || err != 0) {
2951                         kmem_free(ra.rrd, sizeof (*ra.rrd));
2952                         ra.rrd = NULL;
2953                         break;
2954                 }
2955
2956                 bqueue_enqueue(&rwa.q, ra.rrd,
2957                     sizeof (struct receive_record_arg) + ra.rrd->payload_size);
2958                 ra.rrd = NULL;
2959         }
2960         if (ra.next_rrd == NULL)
2961                 ra.next_rrd = kmem_zalloc(sizeof (*ra.next_rrd), KM_SLEEP);
2962         ra.next_rrd->eos_marker = B_TRUE;
2963         bqueue_enqueue(&rwa.q, ra.next_rrd, 1);
2964
2965         mutex_enter(&rwa.mutex);
2966         while (!rwa.done) {
2967                 cv_wait(&rwa.cv, &rwa.mutex);
2968         }
2969         mutex_exit(&rwa.mutex);
2970
2971         cv_destroy(&rwa.cv);
2972         mutex_destroy(&rwa.mutex);
2973         bqueue_destroy(&rwa.q);
2974         if (err == 0)
2975                 err = rwa.err;
2976
2977 out:
2978         nvlist_free(begin_nvl);
2979         if ((featureflags & DMU_BACKUP_FEATURE_DEDUP) && (cleanup_fd != -1))
2980                 zfs_onexit_fd_rele(cleanup_fd);
2981
2982         if (err != 0) {
2983                 /*
2984                  * Clean up references. If receive is not resumable,
2985                  * destroy what we created, so we don't leave it in
2986                  * the inconsistent state.
2987                  */
2988                 dmu_recv_cleanup_ds(drc);
2989         }
2990
2991         *voffp = ra.voff;
2992         objlist_destroy(&ra.ignore_objlist);
2993         return (err);
2994 }
2995
2996 static int
2997 dmu_recv_end_check(void *arg, dmu_tx_t *tx)
2998 {
2999         dmu_recv_cookie_t *drc = arg;
3000         dsl_pool_t *dp = dmu_tx_pool(tx);
3001         int error;
3002
3003         ASSERT3P(drc->drc_ds->ds_owner, ==, dmu_recv_tag);
3004
3005         if (!drc->drc_newfs) {
3006                 dsl_dataset_t *origin_head;
3007
3008                 error = dsl_dataset_hold(dp, drc->drc_tofs, FTAG, &origin_head);
3009                 if (error != 0)
3010                         return (error);
3011                 if (drc->drc_force) {
3012                         /*
3013                          * We will destroy any snapshots in tofs (i.e. before
3014                          * origin_head) that are after the origin (which is
3015                          * the snap before drc_ds, because drc_ds can not
3016                          * have any snaps of its own).
3017                          */
3018                         uint64_t obj;
3019
3020                         obj = dsl_dataset_phys(origin_head)->ds_prev_snap_obj;
3021                         while (obj !=
3022                             dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj) {
3023                                 dsl_dataset_t *snap;
3024                                 error = dsl_dataset_hold_obj(dp, obj, FTAG,
3025                                     &snap);
3026                                 if (error != 0)
3027                                         break;
3028                                 if (snap->ds_dir != origin_head->ds_dir)
3029                                         error = SET_ERROR(EINVAL);
3030                                 if (error == 0)  {
3031                                         error = dsl_destroy_snapshot_check_impl(
3032                                             snap, B_FALSE);
3033                                 }
3034                                 obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
3035                                 dsl_dataset_rele(snap, FTAG);
3036                                 if (error != 0)
3037                                         break;
3038                         }
3039                         if (error != 0) {
3040                                 dsl_dataset_rele(origin_head, FTAG);
3041                                 return (error);
3042                         }
3043                 }
3044                 error = dsl_dataset_clone_swap_check_impl(drc->drc_ds,
3045                     origin_head, drc->drc_force, drc->drc_owner, tx);
3046                 if (error != 0) {
3047                         dsl_dataset_rele(origin_head, FTAG);
3048                         return (error);
3049                 }
3050                 error = dsl_dataset_snapshot_check_impl(origin_head,
3051                     drc->drc_tosnap, tx, B_TRUE, 1, drc->drc_cred);
3052                 dsl_dataset_rele(origin_head, FTAG);
3053                 if (error != 0)
3054                         return (error);
3055
3056                 error = dsl_destroy_head_check_impl(drc->drc_ds, 1);
3057         } else {
3058                 error = dsl_dataset_snapshot_check_impl(drc->drc_ds,
3059                     drc->drc_tosnap, tx, B_TRUE, 1, drc->drc_cred);
3060         }
3061         return (error);
3062 }
3063
3064 static void
3065 dmu_recv_end_sync(void *arg, dmu_tx_t *tx)
3066 {
3067         dmu_recv_cookie_t *drc = arg;
3068         dsl_pool_t *dp = dmu_tx_pool(tx);
3069
3070         spa_history_log_internal_ds(drc->drc_ds, "finish receiving",
3071             tx, "snap=%s", drc->drc_tosnap);
3072
3073         if (!drc->drc_newfs) {
3074                 dsl_dataset_t *origin_head;
3075
3076                 VERIFY0(dsl_dataset_hold(dp, drc->drc_tofs, FTAG,
3077                     &origin_head));
3078
3079                 if (drc->drc_force) {
3080                         /*
3081                          * Destroy any snapshots of drc_tofs (origin_head)
3082                          * after the origin (the snap before drc_ds).
3083                          */
3084                         uint64_t obj;
3085
3086                         obj = dsl_dataset_phys(origin_head)->ds_prev_snap_obj;
3087                         while (obj !=
3088                             dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj) {
3089                                 dsl_dataset_t *snap;
3090                                 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG,
3091                                     &snap));
3092                                 ASSERT3P(snap->ds_dir, ==, origin_head->ds_dir);
3093                                 obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
3094                                 dsl_destroy_snapshot_sync_impl(snap,
3095                                     B_FALSE, tx);
3096                                 dsl_dataset_rele(snap, FTAG);
3097                         }
3098                 }
3099                 VERIFY3P(drc->drc_ds->ds_prev, ==,
3100                     origin_head->ds_prev);
3101
3102                 dsl_dataset_clone_swap_sync_impl(drc->drc_ds,
3103                     origin_head, tx);
3104                 dsl_dataset_snapshot_sync_impl(origin_head,
3105                     drc->drc_tosnap, tx);
3106
3107                 /* set snapshot's creation time and guid */
3108                 dmu_buf_will_dirty(origin_head->ds_prev->ds_dbuf, tx);
3109                 dsl_dataset_phys(origin_head->ds_prev)->ds_creation_time =
3110                     drc->drc_drrb->drr_creation_time;
3111                 dsl_dataset_phys(origin_head->ds_prev)->ds_guid =
3112                     drc->drc_drrb->drr_toguid;
3113                 dsl_dataset_phys(origin_head->ds_prev)->ds_flags &=
3114                     ~DS_FLAG_INCONSISTENT;
3115
3116                 dmu_buf_will_dirty(origin_head->ds_dbuf, tx);
3117                 dsl_dataset_phys(origin_head)->ds_flags &=
3118                     ~DS_FLAG_INCONSISTENT;
3119
3120                 drc->drc_newsnapobj =
3121                     dsl_dataset_phys(origin_head)->ds_prev_snap_obj;
3122
3123                 dsl_dataset_rele(origin_head, FTAG);
3124                 dsl_destroy_head_sync_impl(drc->drc_ds, tx);
3125
3126                 if (drc->drc_owner != NULL)
3127                         VERIFY3P(origin_head->ds_owner, ==, drc->drc_owner);
3128         } else {
3129                 dsl_dataset_t *ds = drc->drc_ds;
3130
3131                 dsl_dataset_snapshot_sync_impl(ds, drc->drc_tosnap, tx);
3132
3133                 /* set snapshot's creation time and guid */
3134                 dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
3135                 dsl_dataset_phys(ds->ds_prev)->ds_creation_time =
3136                     drc->drc_drrb->drr_creation_time;
3137                 dsl_dataset_phys(ds->ds_prev)->ds_guid =
3138                     drc->drc_drrb->drr_toguid;
3139                 dsl_dataset_phys(ds->ds_prev)->ds_flags &=
3140                     ~DS_FLAG_INCONSISTENT;
3141
3142                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
3143                 dsl_dataset_phys(ds)->ds_flags &= ~DS_FLAG_INCONSISTENT;
3144                 if (dsl_dataset_has_resume_receive_state(ds)) {
3145                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3146                             DS_FIELD_RESUME_FROMGUID, tx);
3147                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3148                             DS_FIELD_RESUME_OBJECT, tx);
3149                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3150                             DS_FIELD_RESUME_OFFSET, tx);
3151                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3152                             DS_FIELD_RESUME_BYTES, tx);
3153                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3154                             DS_FIELD_RESUME_TOGUID, tx);
3155                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3156                             DS_FIELD_RESUME_TONAME, tx);
3157                 }
3158                 drc->drc_newsnapobj =
3159                     dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj;
3160         }
3161         /*
3162          * Release the hold from dmu_recv_begin.  This must be done before
3163          * we return to open context, so that when we free the dataset's dnode,
3164          * we can evict its bonus buffer.
3165          */
3166         dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
3167         drc->drc_ds = NULL;
3168 }
3169
3170 static int
3171 add_ds_to_guidmap(const char *name, avl_tree_t *guid_map, uint64_t snapobj)
3172 {
3173         dsl_pool_t *dp;
3174         dsl_dataset_t *snapds;
3175         guid_map_entry_t *gmep;
3176         int err;
3177
3178         ASSERT(guid_map != NULL);
3179
3180         err = dsl_pool_hold(name, FTAG, &dp);
3181         if (err != 0)
3182                 return (err);
3183         gmep = kmem_alloc(sizeof (*gmep), KM_SLEEP);
3184         err = dsl_dataset_hold_obj(dp, snapobj, gmep, &snapds);
3185         if (err == 0) {
3186                 gmep->guid = dsl_dataset_phys(snapds)->ds_guid;
3187                 gmep->gme_ds = snapds;
3188                 avl_add(guid_map, gmep);
3189                 dsl_dataset_long_hold(snapds, gmep);
3190         } else
3191                 kmem_free(gmep, sizeof (*gmep));
3192
3193         dsl_pool_rele(dp, FTAG);
3194         return (err);
3195 }
3196
3197 static int dmu_recv_end_modified_blocks = 3;
3198
3199 static int
3200 dmu_recv_existing_end(dmu_recv_cookie_t *drc)
3201 {
3202 #ifdef _KERNEL
3203         /*
3204          * We will be destroying the ds; make sure its origin is unmounted if
3205          * necessary.
3206          */
3207         char name[ZFS_MAX_DATASET_NAME_LEN];
3208         dsl_dataset_name(drc->drc_ds, name);
3209         zfs_destroy_unmount_origin(name);
3210 #endif
3211
3212         return (dsl_sync_task(drc->drc_tofs,
3213             dmu_recv_end_check, dmu_recv_end_sync, drc,
3214             dmu_recv_end_modified_blocks, ZFS_SPACE_CHECK_NORMAL));
3215 }
3216
3217 static int
3218 dmu_recv_new_end(dmu_recv_cookie_t *drc)
3219 {
3220         return (dsl_sync_task(drc->drc_tofs,
3221             dmu_recv_end_check, dmu_recv_end_sync, drc,
3222             dmu_recv_end_modified_blocks, ZFS_SPACE_CHECK_NORMAL));
3223 }
3224
3225 int
3226 dmu_recv_end(dmu_recv_cookie_t *drc, void *owner)
3227 {
3228         int error;
3229
3230         drc->drc_owner = owner;
3231
3232         if (drc->drc_newfs)
3233                 error = dmu_recv_new_end(drc);
3234         else
3235                 error = dmu_recv_existing_end(drc);
3236
3237         if (error != 0) {
3238                 dmu_recv_cleanup_ds(drc);
3239         } else if (drc->drc_guid_to_ds_map != NULL) {
3240                 (void) add_ds_to_guidmap(drc->drc_tofs,
3241                     drc->drc_guid_to_ds_map,
3242                     drc->drc_newsnapobj);
3243         }
3244         return (error);
3245 }
3246
3247 /*
3248  * Return TRUE if this objset is currently being received into.
3249  */
3250 boolean_t
3251 dmu_objset_is_receiving(objset_t *os)
3252 {
3253         return (os->os_dsl_dataset != NULL &&
3254             os->os_dsl_dataset->ds_owner == dmu_recv_tag);
3255 }