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  */
  30
  31 #include <sys/dmu.h>
  32 #include <sys/dmu_impl.h>
  33 #include <sys/dmu_tx.h>
  34 #include <sys/dbuf.h>
  35 #include <sys/dnode.h>
  36 #include <sys/zfs_context.h>
  37 #include <sys/dmu_objset.h>
  38 #include <sys/dmu_traverse.h>
  39 #include <sys/dsl_dataset.h>
  40 #include <sys/dsl_dir.h>
  41 #include <sys/dsl_prop.h>
  42 #include <sys/dsl_pool.h>
  43 #include <sys/dsl_synctask.h>
  44 #include <sys/zfs_ioctl.h>
  45 #include <sys/zap.h>
  46 #include <sys/zio_checksum.h>
  47 #include <sys/zfs_znode.h>
  48 #include <zfs_fletcher.h>
  49 #include <sys/avl.h>
  50 #include <sys/ddt.h>
  51 #include <sys/zfs_onexit.h>
  52 #include <sys/dmu_send.h>
  53 #include <sys/dsl_destroy.h>
  54 #include <sys/blkptr.h>
  55 #include <sys/dsl_bookmark.h>
  56 #include <sys/zfeature.h>
  57 #include <sys/bqueue.h>
  58
  59 #ifdef __FreeBSD__
  60 #undef dump_write
  61 #define dump_write dmu_dump_write
  62 #endif
  63
  64 /* Set this tunable to TRUE to replace corrupt data with 0x2f5baddb10c */
  65 int zfs_send_corrupt_data = B_FALSE;
  66 int zfs_send_queue_length = 16 * 1024 * 1024;
  67 int zfs_recv_queue_length = 16 * 1024 * 1024;
  68 /* Set this tunable to FALSE to disable setting of DRR_FLAG_FREERECORDS */
  69 int zfs_send_set_freerecords_bit = B_TRUE;
  70
  71 #ifdef _KERNEL
  72 TUNABLE_INT("vfs.zfs.send_set_freerecords_bit", &zfs_send_set_freerecords_bit);
  73 #endif
  74
  75 static char *dmu_recv_tag = "dmu_recv_tag";
  76 const char *recv_clone_name = "%recv";
  77
  78 #define BP_SPAN(datablkszsec, indblkshift, level) \
  79         (((uint64_t)datablkszsec) << (SPA_MINBLOCKSHIFT + \
  80         (level) * (indblkshift - SPA_BLKPTRSHIFT)))
  81
  82 static void byteswap_record(dmu_replay_record_t *drr);
  83
  84 struct send_thread_arg {
  85         bqueue_t        q;
  86         dsl_dataset_t   *ds;            /* Dataset to traverse */
  87         uint64_t        fromtxg;        /* Traverse from this txg */
  88         int             flags;          /* flags to pass to traverse_dataset */
  89         int             error_code;
  90         boolean_t       cancel;
  91         zbookmark_phys_t resume;
  92 };
  93
  94 struct send_block_record {
  95         boolean_t               eos_marker; /* Marks the end of the stream */
  96         blkptr_t                bp;
  97         zbookmark_phys_t        zb;
  98         uint8_t                 indblkshift;
  99         uint16_t                datablkszsec;
 100         bqueue_node_t           ln;
 101 };
 102
 103 static int
 104 dump_bytes(dmu_sendarg_t *dsp, void *buf, int len)
 105 {
 106         dsl_dataset_t *ds = dmu_objset_ds(dsp->dsa_os);
 107         struct uio auio;
 108         struct iovec aiov;
 109         ASSERT0(len % 8);
 110
 111         aiov.iov_base = buf;
 112         aiov.iov_len = len;
 113         auio.uio_iov = &aiov;
 114         auio.uio_iovcnt = 1;
 115         auio.uio_resid = len;
 116         auio.uio_segflg = UIO_SYSSPACE;
 117         auio.uio_rw = UIO_WRITE;
 118         auio.uio_offset = (off_t)-1;
 119         auio.uio_td = dsp->dsa_td;
 120 #ifdef _KERNEL
 121         if (dsp->dsa_fp->f_type == DTYPE_VNODE)
 122                 bwillwrite();
 123         dsp->dsa_err = fo_write(dsp->dsa_fp, &auio, dsp->dsa_td->td_ucred, 0,
 124             dsp->dsa_td);
 125 #else
 126         fprintf(stderr, "%s: returning EOPNOTSUPP\n", __func__);
 127         dsp->dsa_err = EOPNOTSUPP;
 128 #endif
 129         mutex_enter(&ds->ds_sendstream_lock);
 130         *dsp->dsa_off += len;
 131         mutex_exit(&ds->ds_sendstream_lock);
 132
 133         return (dsp->dsa_err);
 134 }
 135
 136 /*
 137  * For all record types except BEGIN, fill in the checksum (overlaid in
 138  * drr_u.drr_checksum.drr_checksum).  The checksum verifies everything
 139  * up to the start of the checksum itself.
 140  */
 141 static int
 142 dump_record(dmu_sendarg_t *dsp, void *payload, int payload_len)
 143 {
 144         ASSERT3U(offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
 145             ==, sizeof (dmu_replay_record_t) - sizeof (zio_cksum_t));
 146         fletcher_4_incremental_native(dsp->dsa_drr,
 147             offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
 148             &dsp->dsa_zc);
 149         if (dsp->dsa_drr->drr_type != DRR_BEGIN) {
 150                 ASSERT(ZIO_CHECKSUM_IS_ZERO(&dsp->dsa_drr->drr_u.
 151                     drr_checksum.drr_checksum));
 152                 dsp->dsa_drr->drr_u.drr_checksum.drr_checksum = dsp->dsa_zc;
 153         }
 154         fletcher_4_incremental_native(&dsp->dsa_drr->
 155             drr_u.drr_checksum.drr_checksum,
 156             sizeof (zio_cksum_t), &dsp->dsa_zc);
 157         if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)) != 0)
 158                 return (SET_ERROR(EINTR));
 159         if (payload_len != 0) {
 160                 fletcher_4_incremental_native(payload, payload_len,
 161                     &dsp->dsa_zc);
 162                 if (dump_bytes(dsp, payload, payload_len) != 0)
 163                         return (SET_ERROR(EINTR));
 164         }
 165         return (0);
 166 }
 167
 168 /*
 169  * Fill in the drr_free struct, or perform aggregation if the previous record is
 170  * also a free record, and the two are adjacent.
 171  *
 172  * Note that we send free records even for a full send, because we want to be
 173  * able to receive a full send as a clone, which requires a list of all the free
 174  * and freeobject records that were generated on the source.
 175  */
 176 static int
 177 dump_free(dmu_sendarg_t *dsp, uint64_t object, uint64_t offset,
 178     uint64_t length)
 179 {
 180         struct drr_free *drrf = &(dsp->dsa_drr->drr_u.drr_free);
 181
 182         /*
 183          * When we receive a free record, dbuf_free_range() assumes
 184          * that the receiving system doesn't have any dbufs in the range
 185          * being freed.  This is always true because there is a one-record
 186          * constraint: we only send one WRITE record for any given
 187          * object,offset.  We know that the one-record constraint is
 188          * true because we always send data in increasing order by
 189          * object,offset.
 190          *
 191          * If the increasing-order constraint ever changes, we should find
 192          * another way to assert that the one-record constraint is still
 193          * satisfied.
 194          */
 195         ASSERT(object > dsp->dsa_last_data_object ||
 196             (object == dsp->dsa_last_data_object &&
 197             offset > dsp->dsa_last_data_offset));
 198
 199         if (length != -1ULL && offset + length < offset)
 200                 length = -1ULL;
 201
 202         /*
 203          * If there is a pending op, but it's not PENDING_FREE, push it out,
 204          * since free block aggregation can only be done for blocks of the
 205          * same type (i.e., DRR_FREE records can only be aggregated with
 206          * other DRR_FREE records.  DRR_FREEOBJECTS records can only be
 207          * aggregated with other DRR_FREEOBJECTS records.
 208          */
 209         if (dsp->dsa_pending_op != PENDING_NONE &&
 210             dsp->dsa_pending_op != PENDING_FREE) {
 211                 if (dump_record(dsp, NULL, 0) != 0)
 212                         return (SET_ERROR(EINTR));
 213                 dsp->dsa_pending_op = PENDING_NONE;
 214         }
 215
 216         if (dsp->dsa_pending_op == PENDING_FREE) {
 217                 /*
 218                  * There should never be a PENDING_FREE if length is -1
 219                  * (because dump_dnode is the only place where this
 220                  * function is called with a -1, and only after flushing
 221                  * any pending record).
 222                  */
 223                 ASSERT(length != -1ULL);
 224                 /*
 225                  * Check to see whether this free block can be aggregated
 226                  * with pending one.
 227                  */
 228                 if (drrf->drr_object == object && drrf->drr_offset +
 229                     drrf->drr_length == offset) {
 230                         drrf->drr_length += length;
 231                         return (0);
 232                 } else {
 233                         /* not a continuation.  Push out pending record */
 234                         if (dump_record(dsp, NULL, 0) != 0)
 235                                 return (SET_ERROR(EINTR));
 236                         dsp->dsa_pending_op = PENDING_NONE;
 237                 }
 238         }
 239         /* create a FREE record and make it pending */
 240         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 241         dsp->dsa_drr->drr_type = DRR_FREE;
 242         drrf->drr_object = object;
 243         drrf->drr_offset = offset;
 244         drrf->drr_length = length;
 245         drrf->drr_toguid = dsp->dsa_toguid;
 246         if (length == -1ULL) {
 247                 if (dump_record(dsp, NULL, 0) != 0)
 248                         return (SET_ERROR(EINTR));
 249         } else {
 250                 dsp->dsa_pending_op = PENDING_FREE;
 251         }
 252
 253         return (0);
 254 }
 255
 256 static int
 257 dump_write(dmu_sendarg_t *dsp, dmu_object_type_t type,
 258     uint64_t object, uint64_t offset, int blksz, const blkptr_t *bp, void *data)
 259 {
 260         struct drr_write *drrw = &(dsp->dsa_drr->drr_u.drr_write);
 261
 262         /*
 263          * We send data in increasing object, offset order.
 264          * See comment in dump_free() for details.
 265          */
 266         ASSERT(object > dsp->dsa_last_data_object ||
 267             (object == dsp->dsa_last_data_object &&
 268             offset > dsp->dsa_last_data_offset));
 269         dsp->dsa_last_data_object = object;
 270         dsp->dsa_last_data_offset = offset + blksz - 1;
 271
 272         /*
 273          * If there is any kind of pending aggregation (currently either
 274          * a grouping of free objects or free blocks), push it out to
 275          * the stream, since aggregation can't be done across operations
 276          * of different types.
 277          */
 278         if (dsp->dsa_pending_op != PENDING_NONE) {
 279                 if (dump_record(dsp, NULL, 0) != 0)
 280                         return (SET_ERROR(EINTR));
 281                 dsp->dsa_pending_op = PENDING_NONE;
 282         }
 283         /* write a WRITE record */
 284         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 285         dsp->dsa_drr->drr_type = DRR_WRITE;
 286         drrw->drr_object = object;
 287         drrw->drr_type = type;
 288         drrw->drr_offset = offset;
 289         drrw->drr_length = blksz;
 290         drrw->drr_toguid = dsp->dsa_toguid;
 291         if (bp == NULL || BP_IS_EMBEDDED(bp)) {
 292                 /*
 293                  * There's no pre-computed checksum for partial-block
 294                  * writes or embedded BP's, so (like
 295                  * fletcher4-checkummed blocks) userland will have to
 296                  * compute a dedup-capable checksum itself.
 297                  */
 298                 drrw->drr_checksumtype = ZIO_CHECKSUM_OFF;
 299         } else {
 300                 drrw->drr_checksumtype = BP_GET_CHECKSUM(bp);
 301                 if (zio_checksum_table[drrw->drr_checksumtype].ci_flags &
 302                     ZCHECKSUM_FLAG_DEDUP)
 303                         drrw->drr_checksumflags |= DRR_CHECKSUM_DEDUP;
 304                 DDK_SET_LSIZE(&drrw->drr_key, BP_GET_LSIZE(bp));
 305                 DDK_SET_PSIZE(&drrw->drr_key, BP_GET_PSIZE(bp));
 306                 DDK_SET_COMPRESS(&drrw->drr_key, BP_GET_COMPRESS(bp));
 307                 drrw->drr_key.ddk_cksum = bp->blk_cksum;
 308         }
 309
 310         if (dump_record(dsp, data, blksz) != 0)
 311                 return (SET_ERROR(EINTR));
 312         return (0);
 313 }
 314
 315 static int
 316 dump_write_embedded(dmu_sendarg_t *dsp, uint64_t object, uint64_t offset,
 317     int blksz, const blkptr_t *bp)
 318 {
 319         char buf[BPE_PAYLOAD_SIZE];
 320         struct drr_write_embedded *drrw =
 321             &(dsp->dsa_drr->drr_u.drr_write_embedded);
 322
 323         if (dsp->dsa_pending_op != PENDING_NONE) {
 324                 if (dump_record(dsp, NULL, 0) != 0)
 325                         return (EINTR);
 326                 dsp->dsa_pending_op = PENDING_NONE;
 327         }
 328
 329         ASSERT(BP_IS_EMBEDDED(bp));
 330
 331         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 332         dsp->dsa_drr->drr_type = DRR_WRITE_EMBEDDED;
 333         drrw->drr_object = object;
 334         drrw->drr_offset = offset;
 335         drrw->drr_length = blksz;
 336         drrw->drr_toguid = dsp->dsa_toguid;
 337         drrw->drr_compression = BP_GET_COMPRESS(bp);
 338         drrw->drr_etype = BPE_GET_ETYPE(bp);
 339         drrw->drr_lsize = BPE_GET_LSIZE(bp);
 340         drrw->drr_psize = BPE_GET_PSIZE(bp);
 341
 342         decode_embedded_bp_compressed(bp, buf);
 343
 344         if (dump_record(dsp, buf, P2ROUNDUP(drrw->drr_psize, 8)) != 0)
 345                 return (EINTR);
 346         return (0);
 347 }
 348
 349 static int
 350 dump_spill(dmu_sendarg_t *dsp, uint64_t object, int blksz, void *data)
 351 {
 352         struct drr_spill *drrs = &(dsp->dsa_drr->drr_u.drr_spill);
 353
 354         if (dsp->dsa_pending_op != PENDING_NONE) {
 355                 if (dump_record(dsp, NULL, 0) != 0)
 356                         return (SET_ERROR(EINTR));
 357                 dsp->dsa_pending_op = PENDING_NONE;
 358         }
 359
 360         /* write a SPILL record */
 361         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 362         dsp->dsa_drr->drr_type = DRR_SPILL;
 363         drrs->drr_object = object;
 364         drrs->drr_length = blksz;
 365         drrs->drr_toguid = dsp->dsa_toguid;
 366
 367         if (dump_record(dsp, data, blksz) != 0)
 368                 return (SET_ERROR(EINTR));
 369         return (0);
 370 }
 371
 372 static int
 373 dump_freeobjects(dmu_sendarg_t *dsp, uint64_t firstobj, uint64_t numobjs)
 374 {
 375         struct drr_freeobjects *drrfo = &(dsp->dsa_drr->drr_u.drr_freeobjects);
 376
 377         /*
 378          * If there is a pending op, but it's not PENDING_FREEOBJECTS,
 379          * push it out, since free block aggregation can only be done for
 380          * blocks of the same type (i.e., DRR_FREE records can only be
 381          * aggregated with other DRR_FREE records.  DRR_FREEOBJECTS records
 382          * can only be aggregated with other DRR_FREEOBJECTS records.
 383          */
 384         if (dsp->dsa_pending_op != PENDING_NONE &&
 385             dsp->dsa_pending_op != PENDING_FREEOBJECTS) {
 386                 if (dump_record(dsp, NULL, 0) != 0)
 387                         return (SET_ERROR(EINTR));
 388                 dsp->dsa_pending_op = PENDING_NONE;
 389         }
 390         if (dsp->dsa_pending_op == PENDING_FREEOBJECTS) {
 391                 /*
 392                  * See whether this free object array can be aggregated
 393                  * with pending one
 394                  */
 395                 if (drrfo->drr_firstobj + drrfo->drr_numobjs == firstobj) {
 396                         drrfo->drr_numobjs += numobjs;
 397                         return (0);
 398                 } else {
 399                         /* can't be aggregated.  Push out pending record */
 400                         if (dump_record(dsp, NULL, 0) != 0)
 401                                 return (SET_ERROR(EINTR));
 402                         dsp->dsa_pending_op = PENDING_NONE;
 403                 }
 404         }
 405
 406         /* write a FREEOBJECTS record */
 407         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 408         dsp->dsa_drr->drr_type = DRR_FREEOBJECTS;
 409         drrfo->drr_firstobj = firstobj;
 410         drrfo->drr_numobjs = numobjs;
 411         drrfo->drr_toguid = dsp->dsa_toguid;
 412
 413         dsp->dsa_pending_op = PENDING_FREEOBJECTS;
 414
 415         return (0);
 416 }
 417
 418 static int
 419 dump_dnode(dmu_sendarg_t *dsp, uint64_t object, dnode_phys_t *dnp)
 420 {
 421         struct drr_object *drro = &(dsp->dsa_drr->drr_u.drr_object);
 422
 423         if (object < dsp->dsa_resume_object) {
 424                 /*
 425                  * Note: when resuming, we will visit all the dnodes in
 426                  * the block of dnodes that we are resuming from.  In
 427                  * this case it's unnecessary to send the dnodes prior to
 428                  * the one we are resuming from.  We should be at most one
 429                  * block's worth of dnodes behind the resume point.
 430                  */
 431                 ASSERT3U(dsp->dsa_resume_object - object, <,
 432                     1 << (DNODE_BLOCK_SHIFT - DNODE_SHIFT));
 433                 return (0);
 434         }
 435
 436         if (dnp == NULL || dnp->dn_type == DMU_OT_NONE)
 437                 return (dump_freeobjects(dsp, object, 1));
 438
 439         if (dsp->dsa_pending_op != PENDING_NONE) {
 440                 if (dump_record(dsp, NULL, 0) != 0)
 441                         return (SET_ERROR(EINTR));
 442                 dsp->dsa_pending_op = PENDING_NONE;
 443         }
 444
 445         /* write an OBJECT record */
 446         bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
 447         dsp->dsa_drr->drr_type = DRR_OBJECT;
 448         drro->drr_object = object;
 449         drro->drr_type = dnp->dn_type;
 450         drro->drr_bonustype = dnp->dn_bonustype;
 451         drro->drr_blksz = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT;
 452         drro->drr_bonuslen = dnp->dn_bonuslen;
 453         drro->drr_checksumtype = dnp->dn_checksum;
 454         drro->drr_compress = dnp->dn_compress;
 455         drro->drr_toguid = dsp->dsa_toguid;
 456
 457         if (!(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
 458             drro->drr_blksz > SPA_OLD_MAXBLOCKSIZE)
 459                 drro->drr_blksz = SPA_OLD_MAXBLOCKSIZE;
 460
 461         if (dump_record(dsp, DN_BONUS(dnp),
 462             P2ROUNDUP(dnp->dn_bonuslen, 8)) != 0) {
 463                 return (SET_ERROR(EINTR));
 464         }
 465
 466         /* Free anything past the end of the file. */
 467         if (dump_free(dsp, object, (dnp->dn_maxblkid + 1) *
 468             (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT), -1ULL) != 0)
 469                 return (SET_ERROR(EINTR));
 470         if (dsp->dsa_err != 0)
 471                 return (SET_ERROR(EINTR));
 472         return (0);
 473 }
 474
 475 static boolean_t
 476 backup_do_embed(dmu_sendarg_t *dsp, const blkptr_t *bp)
 477 {
 478         if (!BP_IS_EMBEDDED(bp))
 479                 return (B_FALSE);
 480
 481         /*
 482          * Compression function must be legacy, or explicitly enabled.
 483          */
 484         if ((BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_LEGACY_FUNCTIONS &&
 485             !(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_EMBED_DATA_LZ4)))
 486                 return (B_FALSE);
 487
 488         /*
 489          * Embed type must be explicitly enabled.
 490          */
 491         switch (BPE_GET_ETYPE(bp)) {
 492         case BP_EMBEDDED_TYPE_DATA:
 493                 if (dsp->dsa_featureflags & DMU_BACKUP_FEATURE_EMBED_DATA)
 494                         return (B_TRUE);
 495                 break;
 496         default:
 497                 return (B_FALSE);
 498         }
 499         return (B_FALSE);
 500 }
 501
 502 /*
 503  * This is the callback function to traverse_dataset that acts as the worker
 504  * thread for dmu_send_impl.
 505  */
 506 /*ARGSUSED*/
 507 static int
 508 send_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
 509     const zbookmark_phys_t *zb, const struct dnode_phys *dnp, void *arg)
 510 {
 511         struct send_thread_arg *sta = arg;
 512         struct send_block_record *record;
 513         uint64_t record_size;
 514         int err = 0;
 515
 516         ASSERT(zb->zb_object == DMU_META_DNODE_OBJECT ||
 517             zb->zb_object >= sta->resume.zb_object);
 518
 519         if (sta->cancel)
 520                 return (SET_ERROR(EINTR));
 521
 522         if (bp == NULL) {
 523                 ASSERT3U(zb->zb_level, ==, ZB_DNODE_LEVEL);
 524                 return (0);
 525         } else if (zb->zb_level < 0) {
 526                 return (0);
 527         }
 528
 529         record = kmem_zalloc(sizeof (struct send_block_record), KM_SLEEP);
 530         record->eos_marker = B_FALSE;
 531         record->bp = *bp;
 532         record->zb = *zb;
 533         record->indblkshift = dnp->dn_indblkshift;
 534         record->datablkszsec = dnp->dn_datablkszsec;
 535         record_size = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT;
 536         bqueue_enqueue(&sta->q, record, record_size);
 537
 538         return (err);
 539 }
 540
 541 /*
 542  * This function kicks off the traverse_dataset.  It also handles setting the
 543  * error code of the thread in case something goes wrong, and pushes the End of
 544  * Stream record when the traverse_dataset call has finished.  If there is no
 545  * dataset to traverse, the thread immediately pushes End of Stream marker.
 546  */
 547 static void
 548 send_traverse_thread(void *arg)
 549 {
 550         struct send_thread_arg *st_arg = arg;
 551         int err;
 552         struct send_block_record *data;
 553
 554         if (st_arg->ds != NULL) {
 555                 err = traverse_dataset_resume(st_arg->ds,
 556                     st_arg->fromtxg, &st_arg->resume,
 557                     st_arg->flags, send_cb, st_arg);
 558
 559                 if (err != EINTR)
 560                         st_arg->error_code = err;
 561         }
 562         data = kmem_zalloc(sizeof (*data), KM_SLEEP);
 563         data->eos_marker = B_TRUE;
 564         bqueue_enqueue(&st_arg->q, data, 1);
 565         thread_exit();
 566 }
 567
 568 /*
 569  * This function actually handles figuring out what kind of record needs to be
 570  * dumped, reading the data (which has hopefully been prefetched), and calling
 571  * the appropriate helper function.
 572  */
 573 static int
 574 do_dump(dmu_sendarg_t *dsa, struct send_block_record *data)
 575 {
 576         dsl_dataset_t *ds = dmu_objset_ds(dsa->dsa_os);
 577         const blkptr_t *bp = &data->bp;
 578         const zbookmark_phys_t *zb = &data->zb;
 579         uint8_t indblkshift = data->indblkshift;
 580         uint16_t dblkszsec = data->datablkszsec;
 581         spa_t *spa = ds->ds_dir->dd_pool->dp_spa;
 582         dmu_object_type_t type = bp ? BP_GET_TYPE(bp) : DMU_OT_NONE;
 583         int err = 0;
 584
 585         ASSERT3U(zb->zb_level, >=, 0);
 586
 587         ASSERT(zb->zb_object == DMU_META_DNODE_OBJECT ||
 588             zb->zb_object >= dsa->dsa_resume_object);
 589
 590         if (zb->zb_object != DMU_META_DNODE_OBJECT &&
 591             DMU_OBJECT_IS_SPECIAL(zb->zb_object)) {
 592                 return (0);
 593         } else if (BP_IS_HOLE(bp) &&
 594             zb->zb_object == DMU_META_DNODE_OBJECT) {
 595                 uint64_t span = BP_SPAN(dblkszsec, indblkshift, zb->zb_level);
 596                 uint64_t dnobj = (zb->zb_blkid * span) >> DNODE_SHIFT;
 597                 err = dump_freeobjects(dsa, dnobj, span >> DNODE_SHIFT);
 598         } else if (BP_IS_HOLE(bp)) {
 599                 uint64_t span = BP_SPAN(dblkszsec, indblkshift, zb->zb_level);
 600                 uint64_t offset = zb->zb_blkid * span;
 601                 err = dump_free(dsa, zb->zb_object, offset, span);
 602         } else if (zb->zb_level > 0 || type == DMU_OT_OBJSET) {
 603                 return (0);
 604         } else if (type == DMU_OT_DNODE) {
 605                 int blksz = BP_GET_LSIZE(bp);
 606                 arc_flags_t aflags = ARC_FLAG_WAIT;
 607                 arc_buf_t *abuf;
 608
 609                 ASSERT0(zb->zb_level);
 610
 611                 if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
 612                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
 613                     &aflags, zb) != 0)
 614                         return (SET_ERROR(EIO));
 615
 616                 dnode_phys_t *blk = abuf->b_data;
 617                 uint64_t dnobj = zb->zb_blkid * (blksz >> DNODE_SHIFT);
 618                 for (int i = 0; i < blksz >> DNODE_SHIFT; i++) {
 619                         err = dump_dnode(dsa, dnobj + i, blk + i);
 620                         if (err != 0)
 621                                 break;
 622                 }
 623                 (void) arc_buf_remove_ref(abuf, &abuf);
 624         } else if (type == DMU_OT_SA) {
 625                 arc_flags_t aflags = ARC_FLAG_WAIT;
 626                 arc_buf_t *abuf;
 627                 int blksz = BP_GET_LSIZE(bp);
 628
 629                 if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
 630                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
 631                     &aflags, zb) != 0)
 632                         return (SET_ERROR(EIO));
 633
 634                 err = dump_spill(dsa, zb->zb_object, blksz, abuf->b_data);
 635                 (void) arc_buf_remove_ref(abuf, &abuf);
 636         } else if (backup_do_embed(dsa, bp)) {
 637                 /* it's an embedded level-0 block of a regular object */
 638                 int blksz = dblkszsec << SPA_MINBLOCKSHIFT;
 639                 ASSERT0(zb->zb_level);
 640                 err = dump_write_embedded(dsa, zb->zb_object,
 641                     zb->zb_blkid * blksz, blksz, bp);
 642         } else {
 643                 /* it's a level-0 block of a regular object */
 644                 arc_flags_t aflags = ARC_FLAG_WAIT;
 645                 arc_buf_t *abuf;
 646                 int blksz = dblkszsec << SPA_MINBLOCKSHIFT;
 647                 uint64_t offset;
 648
 649                 ASSERT0(zb->zb_level);
 650                 ASSERT(zb->zb_object > dsa->dsa_resume_object ||
 651                     (zb->zb_object == dsa->dsa_resume_object &&
 652                     zb->zb_blkid * blksz >= dsa->dsa_resume_offset));
 653
 654                 if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
 655                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
 656                     &aflags, zb) != 0) {
 657                         if (zfs_send_corrupt_data) {
 658                                 /* Send a block filled with 0x"zfs badd bloc" */
 659                                 abuf = arc_buf_alloc(spa, blksz, &abuf,
 660                                     ARC_BUFC_DATA);
 661                                 uint64_t *ptr;
 662                                 for (ptr = abuf->b_data;
 663                                     (char *)ptr < (char *)abuf->b_data + blksz;
 664                                     ptr++)
 665                                         *ptr = 0x2f5baddb10cULL;
 666                         } else {
 667                                 return (SET_ERROR(EIO));
 668                         }
 669                 }
 670
 671                 offset = zb->zb_blkid * blksz;
 672
 673                 if (!(dsa->dsa_featureflags &
 674                     DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
 675                     blksz > SPA_OLD_MAXBLOCKSIZE) {
 676                         char *buf = abuf->b_data;
 677                         while (blksz > 0 && err == 0) {
 678                                 int n = MIN(blksz, SPA_OLD_MAXBLOCKSIZE);
 679                                 err = dump_write(dsa, type, zb->zb_object,
 680                                     offset, n, NULL, buf);
 681                                 offset += n;
 682                                 buf += n;
 683                                 blksz -= n;
 684                         }
 685                 } else {
 686                         err = dump_write(dsa, type, zb->zb_object,
 687                             offset, blksz, bp, abuf->b_data);
 688                 }
 689                 (void) arc_buf_remove_ref(abuf, &abuf);
 690         }
 691
 692         ASSERT(err == 0 || err == EINTR);
 693         return (err);
 694 }
 695
 696 /*
 697  * Pop the new data off the queue, and free the old data.
 698  */
 699 static struct send_block_record *
 700 get_next_record(bqueue_t *bq, struct send_block_record *data)
 701 {
 702         struct send_block_record *tmp = bqueue_dequeue(bq);
 703         kmem_free(data, sizeof (*data));
 704         return (tmp);
 705 }
 706
 707 /*
 708  * Actually do the bulk of the work in a zfs send.
 709  *
 710  * Note: Releases dp using the specified tag.
 711  */
 712 static int
 713 dmu_send_impl(void *tag, dsl_pool_t *dp, dsl_dataset_t *to_ds,
 714     zfs_bookmark_phys_t *ancestor_zb,
 715     boolean_t is_clone, boolean_t embedok, boolean_t large_block_ok, int outfd,
 716     uint64_t resumeobj, uint64_t resumeoff,
 717 #ifdef illumos
 718     vnode_t *vp, offset_t *off)
 719 #else
 720     struct file *fp, offset_t *off)
 721 #endif
 722 {
 723         objset_t *os;
 724         dmu_replay_record_t *drr;
 725         dmu_sendarg_t *dsp;
 726         int err;
 727         uint64_t fromtxg = 0;
 728         uint64_t featureflags = 0;
 729         struct send_thread_arg to_arg = { 0 };
 730
 731         err = dmu_objset_from_ds(to_ds, &os);
 732         if (err != 0) {
 733                 dsl_pool_rele(dp, tag);
 734                 return (err);
 735         }
 736
 737         drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP);
 738         drr->drr_type = DRR_BEGIN;
 739         drr->drr_u.drr_begin.drr_magic = DMU_BACKUP_MAGIC;
 740         DMU_SET_STREAM_HDRTYPE(drr->drr_u.drr_begin.drr_versioninfo,
 741             DMU_SUBSTREAM);
 742
 743 #ifdef _KERNEL
 744         if (dmu_objset_type(os) == DMU_OST_ZFS) {
 745                 uint64_t version;
 746                 if (zfs_get_zplprop(os, ZFS_PROP_VERSION, &version) != 0) {
 747                         kmem_free(drr, sizeof (dmu_replay_record_t));
 748                         dsl_pool_rele(dp, tag);
 749                         return (SET_ERROR(EINVAL));
 750                 }
 751                 if (version >= ZPL_VERSION_SA) {
 752                         featureflags |= DMU_BACKUP_FEATURE_SA_SPILL;
 753                 }
 754         }
 755 #endif
 756
 757         if (large_block_ok && to_ds->ds_feature_inuse[SPA_FEATURE_LARGE_BLOCKS])
 758                 featureflags |= DMU_BACKUP_FEATURE_LARGE_BLOCKS;
 759         if (embedok &&
 760             spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA)) {
 761                 featureflags |= DMU_BACKUP_FEATURE_EMBED_DATA;
 762                 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS))
 763                         featureflags |= DMU_BACKUP_FEATURE_EMBED_DATA_LZ4;
 764         }
 765
 766         if (resumeobj != 0 || resumeoff != 0) {
 767                 featureflags |= DMU_BACKUP_FEATURE_RESUMING;
 768         }
 769
 770         DMU_SET_FEATUREFLAGS(drr->drr_u.drr_begin.drr_versioninfo,
 771             featureflags);
 772
 773         drr->drr_u.drr_begin.drr_creation_time =
 774             dsl_dataset_phys(to_ds)->ds_creation_time;
 775         drr->drr_u.drr_begin.drr_type = dmu_objset_type(os);
 776         if (is_clone)
 777                 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CLONE;
 778         drr->drr_u.drr_begin.drr_toguid = dsl_dataset_phys(to_ds)->ds_guid;
 779         if (dsl_dataset_phys(to_ds)->ds_flags & DS_FLAG_CI_DATASET)
 780                 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CI_DATA;
 781         if (zfs_send_set_freerecords_bit)
 782                 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_FREERECORDS;
 783
 784         if (ancestor_zb != NULL) {
 785                 drr->drr_u.drr_begin.drr_fromguid =
 786                     ancestor_zb->zbm_guid;
 787                 fromtxg = ancestor_zb->zbm_creation_txg;
 788         }
 789         dsl_dataset_name(to_ds, drr->drr_u.drr_begin.drr_toname);
 790         if (!to_ds->ds_is_snapshot) {
 791                 (void) strlcat(drr->drr_u.drr_begin.drr_toname, "@--head--",
 792                     sizeof (drr->drr_u.drr_begin.drr_toname));
 793         }
 794
 795         dsp = kmem_zalloc(sizeof (dmu_sendarg_t), KM_SLEEP);
 796
 797         dsp->dsa_drr = drr;
 798         dsp->dsa_outfd = outfd;
 799         dsp->dsa_proc = curproc;
 800         dsp->dsa_td = curthread;
 801         dsp->dsa_fp = fp;
 802         dsp->dsa_os = os;
 803         dsp->dsa_off = off;
 804         dsp->dsa_toguid = dsl_dataset_phys(to_ds)->ds_guid;
 805         dsp->dsa_pending_op = PENDING_NONE;
 806         dsp->dsa_featureflags = featureflags;
 807         dsp->dsa_resume_object = resumeobj;
 808         dsp->dsa_resume_offset = resumeoff;
 809
 810         mutex_enter(&to_ds->ds_sendstream_lock);
 811         list_insert_head(&to_ds->ds_sendstreams, dsp);
 812         mutex_exit(&to_ds->ds_sendstream_lock);
 813
 814         dsl_dataset_long_hold(to_ds, FTAG);
 815         dsl_pool_rele(dp, tag);
 816
 817         void *payload = NULL;
 818         size_t payload_len = 0;
 819         if (resumeobj != 0 || resumeoff != 0) {
 820                 dmu_object_info_t to_doi;
 821                 err = dmu_object_info(os, resumeobj, &to_doi);
 822                 if (err != 0)
 823                         goto out;
 824                 SET_BOOKMARK(&to_arg.resume, to_ds->ds_object, resumeobj, 0,
 825                     resumeoff / to_doi.doi_data_block_size);
 826
 827                 nvlist_t *nvl = fnvlist_alloc();
 828                 fnvlist_add_uint64(nvl, "resume_object", resumeobj);
 829                 fnvlist_add_uint64(nvl, "resume_offset", resumeoff);
 830                 payload = fnvlist_pack(nvl, &payload_len);
 831                 drr->drr_payloadlen = payload_len;
 832                 fnvlist_free(nvl);
 833         }
 834
 835         err = dump_record(dsp, payload, payload_len);
 836         fnvlist_pack_free(payload, payload_len);
 837         if (err != 0) {
 838                 err = dsp->dsa_err;
 839                 goto out;
 840         }
 841
 842         err = bqueue_init(&to_arg.q, zfs_send_queue_length,
 843             offsetof(struct send_block_record, ln));
 844         to_arg.error_code = 0;
 845         to_arg.cancel = B_FALSE;
 846         to_arg.ds = to_ds;
 847         to_arg.fromtxg = fromtxg;
 848         to_arg.flags = TRAVERSE_PRE | TRAVERSE_PREFETCH;
 849         (void) thread_create(NULL, 0, send_traverse_thread, &to_arg, 0, &p0,
 850             TS_RUN, minclsyspri);
 851
 852         struct send_block_record *to_data;
 853         to_data = bqueue_dequeue(&to_arg.q);
 854
 855         while (!to_data->eos_marker && err == 0) {
 856                 err = do_dump(dsp, to_data);
 857                 to_data = get_next_record(&to_arg.q, to_data);
 858                 if (issig(JUSTLOOKING) && issig(FORREAL))
 859                         err = EINTR;
 860         }
 861
 862         if (err != 0) {
 863                 to_arg.cancel = B_TRUE;
 864                 while (!to_data->eos_marker) {
 865                         to_data = get_next_record(&to_arg.q, to_data);
 866                 }
 867         }
 868         kmem_free(to_data, sizeof (*to_data));
 869
 870         bqueue_destroy(&to_arg.q);
 871
 872         if (err == 0 && to_arg.error_code != 0)
 873                 err = to_arg.error_code;
 874
 875         if (err != 0)
 876                 goto out;
 877
 878         if (dsp->dsa_pending_op != PENDING_NONE)
 879                 if (dump_record(dsp, NULL, 0) != 0)
 880                         err = SET_ERROR(EINTR);
 881
 882         if (err != 0) {
 883                 if (err == EINTR && dsp->dsa_err != 0)
 884                         err = dsp->dsa_err;
 885                 goto out;
 886         }
 887
 888         bzero(drr, sizeof (dmu_replay_record_t));
 889         drr->drr_type = DRR_END;
 890         drr->drr_u.drr_end.drr_checksum = dsp->dsa_zc;
 891         drr->drr_u.drr_end.drr_toguid = dsp->dsa_toguid;
 892
 893         if (dump_record(dsp, NULL, 0) != 0)
 894                 err = dsp->dsa_err;
 895
 896 out:
 897         mutex_enter(&to_ds->ds_sendstream_lock);
 898         list_remove(&to_ds->ds_sendstreams, dsp);
 899         mutex_exit(&to_ds->ds_sendstream_lock);
 900
 901         kmem_free(drr, sizeof (dmu_replay_record_t));
 902         kmem_free(dsp, sizeof (dmu_sendarg_t));
 903
 904         dsl_dataset_long_rele(to_ds, FTAG);
 905
 906         return (err);
 907 }
 908
 909 int
 910 dmu_send_obj(const char *pool, uint64_t tosnap, uint64_t fromsnap,
 911     boolean_t embedok, boolean_t large_block_ok,
 912 #ifdef illumos
 913     int outfd, vnode_t *vp, offset_t *off)
 914 #else
 915     int outfd, struct file *fp, offset_t *off)
 916 #endif
 917 {
 918         dsl_pool_t *dp;
 919         dsl_dataset_t *ds;
 920         dsl_dataset_t *fromds = NULL;
 921         int err;
 922
 923         err = dsl_pool_hold(pool, FTAG, &dp);
 924         if (err != 0)
 925                 return (err);
 926
 927         err = dsl_dataset_hold_obj(dp, tosnap, FTAG, &ds);
 928         if (err != 0) {
 929                 dsl_pool_rele(dp, FTAG);
 930                 return (err);
 931         }
 932
 933         if (fromsnap != 0) {
 934                 zfs_bookmark_phys_t zb;
 935                 boolean_t is_clone;
 936
 937                 err = dsl_dataset_hold_obj(dp, fromsnap, FTAG, &fromds);
 938                 if (err != 0) {
 939                         dsl_dataset_rele(ds, FTAG);
 940                         dsl_pool_rele(dp, FTAG);
 941                         return (err);
 942                 }
 943                 if (!dsl_dataset_is_before(ds, fromds, 0))
 944                         err = SET_ERROR(EXDEV);
 945                 zb.zbm_creation_time =
 946                     dsl_dataset_phys(fromds)->ds_creation_time;
 947                 zb.zbm_creation_txg = dsl_dataset_phys(fromds)->ds_creation_txg;
 948                 zb.zbm_guid = dsl_dataset_phys(fromds)->ds_guid;
 949                 is_clone = (fromds->ds_dir != ds->ds_dir);
 950                 dsl_dataset_rele(fromds, FTAG);
 951                 err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone,
 952                     embedok, large_block_ok, outfd, 0, 0, fp, off);
 953         } else {
 954                 err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE,
 955                     embedok, large_block_ok, outfd, 0, 0, fp, off);
 956         }
 957         dsl_dataset_rele(ds, FTAG);
 958         return (err);
 959 }
 960
 961 int
 962 dmu_send(const char *tosnap, const char *fromsnap, boolean_t embedok,
 963     boolean_t large_block_ok, int outfd, uint64_t resumeobj, uint64_t resumeoff,
 964 #ifdef illumos
 965     vnode_t *vp, offset_t *off)
 966 #else
 967     struct file *fp, offset_t *off)
 968 #endif
 969 {
 970         dsl_pool_t *dp;
 971         dsl_dataset_t *ds;
 972         int err;
 973         boolean_t owned = B_FALSE;
 974
 975         if (fromsnap != NULL && strpbrk(fromsnap, "@#") == NULL)
 976                 return (SET_ERROR(EINVAL));
 977
 978         err = dsl_pool_hold(tosnap, FTAG, &dp);
 979         if (err != 0)
 980                 return (err);
 981
 982         if (strchr(tosnap, '@') == NULL && spa_writeable(dp->dp_spa)) {
 983                 /*
 984                  * We are sending a filesystem or volume.  Ensure
 985                  * that it doesn't change by owning the dataset.
 986                  */
 987                 err = dsl_dataset_own(dp, tosnap, FTAG, &ds);
 988                 owned = B_TRUE;
 989         } else {
 990                 err = dsl_dataset_hold(dp, tosnap, FTAG, &ds);
 991         }
 992         if (err != 0) {
 993                 dsl_pool_rele(dp, FTAG);
 994                 return (err);
 995         }
 996
 997         if (fromsnap != NULL) {
 998                 zfs_bookmark_phys_t zb;
 999                 boolean_t is_clone = B_FALSE;
1000                 int fsnamelen = strchr(tosnap, '@') - tosnap;
1001
1002                 /*
1003                  * If the fromsnap is in a different filesystem, then
1004                  * mark the send stream as a clone.
1005                  */
1006                 if (strncmp(tosnap, fromsnap, fsnamelen) != 0 ||
1007                     (fromsnap[fsnamelen] != '@' &&
1008                     fromsnap[fsnamelen] != '#')) {
1009                         is_clone = B_TRUE;
1010                 }
1011
1012                 if (strchr(fromsnap, '@')) {
1013                         dsl_dataset_t *fromds;
1014                         err = dsl_dataset_hold(dp, fromsnap, FTAG, &fromds);
1015                         if (err == 0) {
1016                                 if (!dsl_dataset_is_before(ds, fromds, 0))
1017                                         err = SET_ERROR(EXDEV);
1018                                 zb.zbm_creation_time =
1019                                     dsl_dataset_phys(fromds)->ds_creation_time;
1020                                 zb.zbm_creation_txg =
1021                                     dsl_dataset_phys(fromds)->ds_creation_txg;
1022                                 zb.zbm_guid = dsl_dataset_phys(fromds)->ds_guid;
1023                                 is_clone = (ds->ds_dir != fromds->ds_dir);
1024                                 dsl_dataset_rele(fromds, FTAG);
1025                         }
1026                 } else {
1027                         err = dsl_bookmark_lookup(dp, fromsnap, ds, &zb);
1028                 }
1029                 if (err != 0) {
1030                         dsl_dataset_rele(ds, FTAG);
1031                         dsl_pool_rele(dp, FTAG);
1032                         return (err);
1033                 }
1034                 err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone,
1035                     embedok, large_block_ok,
1036                     outfd, resumeobj, resumeoff, fp, off);
1037         } else {
1038                 err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE,
1039                     embedok, large_block_ok,
1040                     outfd, resumeobj, resumeoff, fp, off);
1041         }
1042         if (owned)
1043                 dsl_dataset_disown(ds, FTAG);
1044         else
1045                 dsl_dataset_rele(ds, FTAG);
1046         return (err);
1047 }
1048
1049 static int
1050 dmu_adjust_send_estimate_for_indirects(dsl_dataset_t *ds, uint64_t size,
1051     uint64_t *sizep)
1052 {
1053         int err;
1054         /*
1055          * Assume that space (both on-disk and in-stream) is dominated by
1056          * data.  We will adjust for indirect blocks and the copies property,
1057          * but ignore per-object space used (eg, dnodes and DRR_OBJECT records).
1058          */
1059
1060         /*
1061          * Subtract out approximate space used by indirect blocks.
1062          * Assume most space is used by data blocks (non-indirect, non-dnode).
1063          * Assume all blocks are recordsize.  Assume ditto blocks and
1064          * internal fragmentation counter out compression.
1065          *
1066          * Therefore, space used by indirect blocks is sizeof(blkptr_t) per
1067          * block, which we observe in practice.
1068          */
1069         uint64_t recordsize;
1070         err = dsl_prop_get_int_ds(ds, "recordsize", &recordsize);
1071         if (err != 0)
1072                 return (err);
1073         size -= size / recordsize * sizeof (blkptr_t);
1074
1075         /* Add in the space for the record associated with each block. */
1076         size += size / recordsize * sizeof (dmu_replay_record_t);
1077
1078         *sizep = size;
1079
1080         return (0);
1081 }
1082
1083 int
1084 dmu_send_estimate(dsl_dataset_t *ds, dsl_dataset_t *fromds, uint64_t *sizep)
1085 {
1086         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1087         int err;
1088         uint64_t size;
1089
1090         ASSERT(dsl_pool_config_held(dp));
1091
1092         /* tosnap must be a snapshot */
1093         if (!ds->ds_is_snapshot)
1094                 return (SET_ERROR(EINVAL));
1095
1096         /* fromsnap, if provided, must be a snapshot */
1097         if (fromds != NULL && !fromds->ds_is_snapshot)
1098                 return (SET_ERROR(EINVAL));
1099
1100         /*
1101          * fromsnap must be an earlier snapshot from the same fs as tosnap,
1102          * or the origin's fs.
1103          */
1104         if (fromds != NULL && !dsl_dataset_is_before(ds, fromds, 0))
1105                 return (SET_ERROR(EXDEV));
1106
1107         /* Get uncompressed size estimate of changed data. */
1108         if (fromds == NULL) {
1109                 size = dsl_dataset_phys(ds)->ds_uncompressed_bytes;
1110         } else {
1111                 uint64_t used, comp;
1112                 err = dsl_dataset_space_written(fromds, ds,
1113                     &used, &comp, &size);
1114                 if (err != 0)
1115                         return (err);
1116         }
1117
1118         err = dmu_adjust_send_estimate_for_indirects(ds, size, sizep);
1119         return (err);
1120 }
1121
1122 /*
1123  * Simple callback used to traverse the blocks of a snapshot and sum their
1124  * uncompressed size
1125  */
1126 /* ARGSUSED */
1127 static int
1128 dmu_calculate_send_traversal(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
1129     const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
1130 {
1131         uint64_t *spaceptr = arg;
1132         if (bp != NULL && !BP_IS_HOLE(bp)) {
1133                 *spaceptr += BP_GET_UCSIZE(bp);
1134         }
1135         return (0);
1136 }
1137
1138 /*
1139  * Given a desination snapshot and a TXG, calculate the approximate size of a
1140  * send stream sent from that TXG. from_txg may be zero, indicating that the
1141  * whole snapshot will be sent.
1142  */
1143 int
1144 dmu_send_estimate_from_txg(dsl_dataset_t *ds, uint64_t from_txg,
1145     uint64_t *sizep)
1146 {
1147         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1148         int err;
1149         uint64_t size = 0;
1150
1151         ASSERT(dsl_pool_config_held(dp));
1152
1153         /* tosnap must be a snapshot */
1154         if (!dsl_dataset_is_snapshot(ds))
1155                 return (SET_ERROR(EINVAL));
1156
1157         /* verify that from_txg is before the provided snapshot was taken */
1158         if (from_txg >= dsl_dataset_phys(ds)->ds_creation_txg) {
1159                 return (SET_ERROR(EXDEV));
1160         }
1161
1162         /*
1163          * traverse the blocks of the snapshot with birth times after
1164          * from_txg, summing their uncompressed size
1165          */
1166         err = traverse_dataset(ds, from_txg, TRAVERSE_POST,
1167             dmu_calculate_send_traversal, &size);
1168         if (err)
1169                 return (err);
1170
1171         err = dmu_adjust_send_estimate_for_indirects(ds, size, sizep);
1172         return (err);
1173 }
1174
1175 typedef struct dmu_recv_begin_arg {
1176         const char *drba_origin;
1177         dmu_recv_cookie_t *drba_cookie;
1178         cred_t *drba_cred;
1179         uint64_t drba_snapobj;
1180 } dmu_recv_begin_arg_t;
1181
1182 static int
1183 recv_begin_check_existing_impl(dmu_recv_begin_arg_t *drba, dsl_dataset_t *ds,
1184     uint64_t fromguid)
1185 {
1186         uint64_t val;
1187         int error;
1188         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1189
1190         /* temporary clone name must not exist */
1191         error = zap_lookup(dp->dp_meta_objset,
1192             dsl_dir_phys(ds->ds_dir)->dd_child_dir_zapobj, recv_clone_name,
1193             8, 1, &val);
1194         if (error != ENOENT)
1195                 return (error == 0 ? EBUSY : error);
1196
1197         /* new snapshot name must not exist */
1198         error = zap_lookup(dp->dp_meta_objset,
1199             dsl_dataset_phys(ds)->ds_snapnames_zapobj,
1200             drba->drba_cookie->drc_tosnap, 8, 1, &val);
1201         if (error != ENOENT)
1202                 return (error == 0 ? EEXIST : error);
1203
1204         /*
1205          * Check snapshot limit before receiving. We'll recheck again at the
1206          * end, but might as well abort before receiving if we're already over
1207          * the limit.
1208          *
1209          * Note that we do not check the file system limit with
1210          * dsl_dir_fscount_check because the temporary %clones don't count
1211          * against that limit.
1212          */
1213         error = dsl_fs_ss_limit_check(ds->ds_dir, 1, ZFS_PROP_SNAPSHOT_LIMIT,
1214             NULL, drba->drba_cred);
1215         if (error != 0)
1216                 return (error);
1217
1218         if (fromguid != 0) {
1219                 dsl_dataset_t *snap;
1220                 uint64_t obj = dsl_dataset_phys(ds)->ds_prev_snap_obj;
1221
1222                 /* Find snapshot in this dir that matches fromguid. */
1223                 while (obj != 0) {
1224                         error = dsl_dataset_hold_obj(dp, obj, FTAG,
1225                             &snap);
1226                         if (error != 0)
1227                                 return (SET_ERROR(ENODEV));
1228                         if (snap->ds_dir != ds->ds_dir) {
1229                                 dsl_dataset_rele(snap, FTAG);
1230                                 return (SET_ERROR(ENODEV));
1231                         }
1232                         if (dsl_dataset_phys(snap)->ds_guid == fromguid)
1233                                 break;
1234                         obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
1235                         dsl_dataset_rele(snap, FTAG);
1236                 }
1237                 if (obj == 0)
1238                         return (SET_ERROR(ENODEV));
1239
1240                 if (drba->drba_cookie->drc_force) {
1241                         drba->drba_snapobj = obj;
1242                 } else {
1243                         /*
1244                          * If we are not forcing, there must be no
1245                          * changes since fromsnap.
1246                          */
1247                         if (dsl_dataset_modified_since_snap(ds, snap)) {
1248                                 dsl_dataset_rele(snap, FTAG);
1249                                 return (SET_ERROR(ETXTBSY));
1250                         }
1251                         drba->drba_snapobj = ds->ds_prev->ds_object;
1252                 }
1253
1254                 dsl_dataset_rele(snap, FTAG);
1255         } else {
1256                 /* if full, then must be forced */
1257                 if (!drba->drba_cookie->drc_force)
1258                         return (SET_ERROR(EEXIST));
1259                 /* start from $ORIGIN@$ORIGIN, if supported */
1260                 drba->drba_snapobj = dp->dp_origin_snap != NULL ?
1261                     dp->dp_origin_snap->ds_object : 0;
1262         }
1263
1264         return (0);
1265
1266 }
1267
1268 static int
1269 dmu_recv_begin_check(void *arg, dmu_tx_t *tx)
1270 {
1271         dmu_recv_begin_arg_t *drba = arg;
1272         dsl_pool_t *dp = dmu_tx_pool(tx);
1273         struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
1274         uint64_t fromguid = drrb->drr_fromguid;
1275         int flags = drrb->drr_flags;
1276         int error;
1277         uint64_t featureflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo);
1278         dsl_dataset_t *ds;
1279         const char *tofs = drba->drba_cookie->drc_tofs;
1280
1281         /* already checked */
1282         ASSERT3U(drrb->drr_magic, ==, DMU_BACKUP_MAGIC);
1283         ASSERT(!(featureflags & DMU_BACKUP_FEATURE_RESUMING));
1284
1285         if (DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) ==
1286             DMU_COMPOUNDSTREAM ||
1287             drrb->drr_type >= DMU_OST_NUMTYPES ||
1288             ((flags & DRR_FLAG_CLONE) && drba->drba_origin == NULL))
1289                 return (SET_ERROR(EINVAL));
1290
1291         /* Verify pool version supports SA if SA_SPILL feature set */
1292         if ((featureflags & DMU_BACKUP_FEATURE_SA_SPILL) &&
1293             spa_version(dp->dp_spa) < SPA_VERSION_SA)
1294                 return (SET_ERROR(ENOTSUP));
1295
1296         if (drba->drba_cookie->drc_resumable &&
1297             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EXTENSIBLE_DATASET))
1298                 return (SET_ERROR(ENOTSUP));
1299
1300         /*
1301          * The receiving code doesn't know how to translate a WRITE_EMBEDDED
1302          * record to a plan WRITE record, so the pool must have the
1303          * EMBEDDED_DATA feature enabled if the stream has WRITE_EMBEDDED
1304          * records.  Same with WRITE_EMBEDDED records that use LZ4 compression.
1305          */
1306         if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA) &&
1307             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA))
1308                 return (SET_ERROR(ENOTSUP));
1309         if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA_LZ4) &&
1310             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS))
1311                 return (SET_ERROR(ENOTSUP));
1312
1313         /*
1314          * The receiving code doesn't know how to translate large blocks
1315          * to smaller ones, so the pool must have the LARGE_BLOCKS
1316          * feature enabled if the stream has LARGE_BLOCKS.
1317          */
1318         if ((featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
1319             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LARGE_BLOCKS))
1320                 return (SET_ERROR(ENOTSUP));
1321
1322         error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
1323         if (error == 0) {
1324                 /* target fs already exists; recv into temp clone */
1325
1326                 /* Can't recv a clone into an existing fs */
1327                 if (flags & DRR_FLAG_CLONE || drba->drba_origin) {
1328                         dsl_dataset_rele(ds, FTAG);
1329                         return (SET_ERROR(EINVAL));
1330                 }
1331
1332                 error = recv_begin_check_existing_impl(drba, ds, fromguid);
1333                 dsl_dataset_rele(ds, FTAG);
1334         } else if (error == ENOENT) {
1335                 /* target fs does not exist; must be a full backup or clone */
1336                 char buf[MAXNAMELEN];
1337
1338                 /*
1339                  * If it's a non-clone incremental, we are missing the
1340                  * target fs, so fail the recv.
1341                  */
1342                 if (fromguid != 0 && !(flags & DRR_FLAG_CLONE ||
1343                     drba->drba_origin))
1344                         return (SET_ERROR(ENOENT));
1345
1346                 /*
1347                  * If we're receiving a full send as a clone, and it doesn't
1348                  * contain all the necessary free records and freeobject
1349                  * records, reject it.
1350                  */
1351                 if (fromguid == 0 && drba->drba_origin &&
1352                     !(flags & DRR_FLAG_FREERECORDS))
1353                         return (SET_ERROR(EINVAL));
1354
1355                 /* Open the parent of tofs */
1356                 ASSERT3U(strlen(tofs), <, MAXNAMELEN);
1357                 (void) strlcpy(buf, tofs, strrchr(tofs, '/') - tofs + 1);
1358                 error = dsl_dataset_hold(dp, buf, FTAG, &ds);
1359                 if (error != 0)
1360                         return (error);
1361
1362                 /*
1363                  * Check filesystem and snapshot limits before receiving. We'll
1364                  * recheck snapshot limits again at the end (we create the
1365                  * filesystems and increment those counts during begin_sync).
1366                  */
1367                 error = dsl_fs_ss_limit_check(ds->ds_dir, 1,
1368                     ZFS_PROP_FILESYSTEM_LIMIT, NULL, drba->drba_cred);
1369                 if (error != 0) {
1370                         dsl_dataset_rele(ds, FTAG);
1371                         return (error);
1372                 }
1373
1374                 error = dsl_fs_ss_limit_check(ds->ds_dir, 1,
1375                     ZFS_PROP_SNAPSHOT_LIMIT, NULL, drba->drba_cred);
1376                 if (error != 0) {
1377                         dsl_dataset_rele(ds, FTAG);
1378                         return (error);
1379                 }
1380
1381                 if (drba->drba_origin != NULL) {
1382                         dsl_dataset_t *origin;
1383                         error = dsl_dataset_hold(dp, drba->drba_origin,
1384                             FTAG, &origin);
1385                         if (error != 0) {
1386                                 dsl_dataset_rele(ds, FTAG);
1387                                 return (error);
1388                         }
1389                         if (!origin->ds_is_snapshot) {
1390                                 dsl_dataset_rele(origin, FTAG);
1391                                 dsl_dataset_rele(ds, FTAG);
1392                                 return (SET_ERROR(EINVAL));
1393                         }
1394                         if (dsl_dataset_phys(origin)->ds_guid != fromguid &&
1395                             fromguid != 0) {
1396                                 dsl_dataset_rele(origin, FTAG);
1397                                 dsl_dataset_rele(ds, FTAG);
1398                                 return (SET_ERROR(ENODEV));
1399                         }
1400                         dsl_dataset_rele(origin, FTAG);
1401                 }
1402                 dsl_dataset_rele(ds, FTAG);
1403                 error = 0;
1404         }
1405         return (error);
1406 }
1407
1408 static void
1409 dmu_recv_begin_sync(void *arg, dmu_tx_t *tx)
1410 {
1411         dmu_recv_begin_arg_t *drba = arg;
1412         dsl_pool_t *dp = dmu_tx_pool(tx);
1413         objset_t *mos = dp->dp_meta_objset;
1414         struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
1415         const char *tofs = drba->drba_cookie->drc_tofs;
1416         dsl_dataset_t *ds, *newds;
1417         uint64_t dsobj;
1418         int error;
1419         uint64_t crflags = 0;
1420
1421         if (drrb->drr_flags & DRR_FLAG_CI_DATA)
1422                 crflags |= DS_FLAG_CI_DATASET;
1423
1424         error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
1425         if (error == 0) {
1426                 /* create temporary clone */
1427                 dsl_dataset_t *snap = NULL;
1428                 if (drba->drba_snapobj != 0) {
1429                         VERIFY0(dsl_dataset_hold_obj(dp,
1430                             drba->drba_snapobj, FTAG, &snap));
1431                 }
1432                 dsobj = dsl_dataset_create_sync(ds->ds_dir, recv_clone_name,
1433                     snap, crflags, drba->drba_cred, tx);
1434                 if (drba->drba_snapobj != 0)
1435                         dsl_dataset_rele(snap, FTAG);
1436                 dsl_dataset_rele(ds, FTAG);
1437         } else {
1438                 dsl_dir_t *dd;
1439                 const char *tail;
1440                 dsl_dataset_t *origin = NULL;
1441
1442                 VERIFY0(dsl_dir_hold(dp, tofs, FTAG, &dd, &tail));
1443
1444                 if (drba->drba_origin != NULL) {
1445                         VERIFY0(dsl_dataset_hold(dp, drba->drba_origin,
1446                             FTAG, &origin));
1447                 }
1448
1449                 /* Create new dataset. */
1450                 dsobj = dsl_dataset_create_sync(dd,
1451                     strrchr(tofs, '/') + 1,
1452                     origin, crflags, drba->drba_cred, tx);
1453                 if (origin != NULL)
1454                         dsl_dataset_rele(origin, FTAG);
1455                 dsl_dir_rele(dd, FTAG);
1456                 drba->drba_cookie->drc_newfs = B_TRUE;
1457         }
1458         VERIFY0(dsl_dataset_own_obj(dp, dsobj, dmu_recv_tag, &newds));
1459
1460         if (drba->drba_cookie->drc_resumable) {
1461                 dsl_dataset_zapify(newds, tx);
1462                 if (drrb->drr_fromguid != 0) {
1463                         VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_FROMGUID,
1464                             8, 1, &drrb->drr_fromguid, tx));
1465                 }
1466                 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_TOGUID,
1467                     8, 1, &drrb->drr_toguid, tx));
1468                 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_TONAME,
1469                     1, strlen(drrb->drr_toname) + 1, drrb->drr_toname, tx));
1470                 uint64_t one = 1;
1471                 uint64_t zero = 0;
1472                 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_OBJECT,
1473                     8, 1, &one, tx));
1474                 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_OFFSET,
1475                     8, 1, &zero, tx));
1476                 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_BYTES,
1477                     8, 1, &zero, tx));
1478                 if (DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo) &
1479                     DMU_BACKUP_FEATURE_EMBED_DATA) {
1480                         VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_EMBEDOK,
1481                             8, 1, &one, tx));
1482                 }
1483         }
1484
1485         dmu_buf_will_dirty(newds->ds_dbuf, tx);
1486         dsl_dataset_phys(newds)->ds_flags |= DS_FLAG_INCONSISTENT;
1487
1488         /*
1489          * If we actually created a non-clone, we need to create the
1490          * objset in our new dataset.
1491          */
1492         if (BP_IS_HOLE(dsl_dataset_get_blkptr(newds))) {
1493                 (void) dmu_objset_create_impl(dp->dp_spa,
1494                     newds, dsl_dataset_get_blkptr(newds), drrb->drr_type, tx);
1495         }
1496
1497         drba->drba_cookie->drc_ds = newds;
1498
1499         spa_history_log_internal_ds(newds, "receive", tx, "");
1500 }
1501
1502 static int
1503 dmu_recv_resume_begin_check(void *arg, dmu_tx_t *tx)
1504 {
1505         dmu_recv_begin_arg_t *drba = arg;
1506         dsl_pool_t *dp = dmu_tx_pool(tx);
1507         struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
1508         int error;
1509         uint64_t featureflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo);
1510         dsl_dataset_t *ds;
1511         const char *tofs = drba->drba_cookie->drc_tofs;
1512
1513         /* already checked */
1514         ASSERT3U(drrb->drr_magic, ==, DMU_BACKUP_MAGIC);
1515         ASSERT(featureflags & DMU_BACKUP_FEATURE_RESUMING);
1516
1517         if (DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) ==
1518             DMU_COMPOUNDSTREAM ||
1519             drrb->drr_type >= DMU_OST_NUMTYPES)
1520                 return (SET_ERROR(EINVAL));
1521
1522         /* Verify pool version supports SA if SA_SPILL feature set */
1523         if ((featureflags & DMU_BACKUP_FEATURE_SA_SPILL) &&
1524             spa_version(dp->dp_spa) < SPA_VERSION_SA)
1525                 return (SET_ERROR(ENOTSUP));
1526
1527         /*
1528          * The receiving code doesn't know how to translate a WRITE_EMBEDDED
1529          * record to a plain WRITE record, so the pool must have the
1530          * EMBEDDED_DATA feature enabled if the stream has WRITE_EMBEDDED
1531          * records.  Same with WRITE_EMBEDDED records that use LZ4 compression.
1532          */
1533         if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA) &&
1534             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA))
1535                 return (SET_ERROR(ENOTSUP));
1536         if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA_LZ4) &&
1537             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS))
1538                 return (SET_ERROR(ENOTSUP));
1539
1540         char recvname[ZFS_MAXNAMELEN];
1541
1542         (void) snprintf(recvname, sizeof (recvname), "%s/%s",
1543             tofs, recv_clone_name);
1544
1545         if (dsl_dataset_hold(dp, recvname, FTAG, &ds) != 0) {
1546                 /* %recv does not exist; continue in tofs */
1547                 error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
1548                 if (error != 0)
1549                         return (error);
1550         }
1551
1552         /* check that ds is marked inconsistent */
1553         if (!DS_IS_INCONSISTENT(ds)) {
1554                 dsl_dataset_rele(ds, FTAG);
1555                 return (SET_ERROR(EINVAL));
1556         }
1557
1558         /* check that there is resuming data, and that the toguid matches */
1559         if (!dsl_dataset_is_zapified(ds)) {
1560                 dsl_dataset_rele(ds, FTAG);
1561                 return (SET_ERROR(EINVAL));
1562         }
1563         uint64_t val;
1564         error = zap_lookup(dp->dp_meta_objset, ds->ds_object,
1565             DS_FIELD_RESUME_TOGUID, sizeof (val), 1, &val);
1566         if (error != 0 || drrb->drr_toguid != val) {
1567                 dsl_dataset_rele(ds, FTAG);
1568                 return (SET_ERROR(EINVAL));
1569         }
1570
1571         /*
1572          * Check if the receive is still running.  If so, it will be owned.
1573          * Note that nothing else can own the dataset (e.g. after the receive
1574          * fails) because it will be marked inconsistent.
1575          */
1576         if (dsl_dataset_has_owner(ds)) {
1577                 dsl_dataset_rele(ds, FTAG);
1578                 return (SET_ERROR(EBUSY));
1579         }
1580
1581         /* There should not be any snapshots of this fs yet. */
1582         if (ds->ds_prev != NULL && ds->ds_prev->ds_dir == ds->ds_dir) {
1583                 dsl_dataset_rele(ds, FTAG);
1584                 return (SET_ERROR(EINVAL));
1585         }
1586
1587         /*
1588          * Note: resume point will be checked when we process the first WRITE
1589          * record.
1590          */
1591
1592         /* check that the origin matches */
1593         val = 0;
1594         (void) zap_lookup(dp->dp_meta_objset, ds->ds_object,
1595             DS_FIELD_RESUME_FROMGUID, sizeof (val), 1, &val);
1596         if (drrb->drr_fromguid != val) {
1597                 dsl_dataset_rele(ds, FTAG);
1598                 return (SET_ERROR(EINVAL));
1599         }
1600
1601         dsl_dataset_rele(ds, FTAG);
1602         return (0);
1603 }
1604
1605 static void
1606 dmu_recv_resume_begin_sync(void *arg, dmu_tx_t *tx)
1607 {
1608         dmu_recv_begin_arg_t *drba = arg;
1609         dsl_pool_t *dp = dmu_tx_pool(tx);
1610         const char *tofs = drba->drba_cookie->drc_tofs;
1611         dsl_dataset_t *ds;
1612         uint64_t dsobj;
1613         char recvname[ZFS_MAXNAMELEN];
1614
1615         (void) snprintf(recvname, sizeof (recvname), "%s/%s",
1616             tofs, recv_clone_name);
1617
1618         if (dsl_dataset_hold(dp, recvname, FTAG, &ds) != 0) {
1619                 /* %recv does not exist; continue in tofs */
1620                 VERIFY0(dsl_dataset_hold(dp, tofs, FTAG, &ds));
1621                 drba->drba_cookie->drc_newfs = B_TRUE;
1622         }
1623
1624         /* clear the inconsistent flag so that we can own it */
1625         ASSERT(DS_IS_INCONSISTENT(ds));
1626         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1627         dsl_dataset_phys(ds)->ds_flags &= ~DS_FLAG_INCONSISTENT;
1628         dsobj = ds->ds_object;
1629         dsl_dataset_rele(ds, FTAG);
1630
1631         VERIFY0(dsl_dataset_own_obj(dp, dsobj, dmu_recv_tag, &ds));
1632
1633         dmu_buf_will_dirty(ds->ds_dbuf, tx);
1634         dsl_dataset_phys(ds)->ds_flags |= DS_FLAG_INCONSISTENT;
1635
1636         ASSERT(!BP_IS_HOLE(dsl_dataset_get_blkptr(ds)));
1637
1638         drba->drba_cookie->drc_ds = ds;
1639
1640         spa_history_log_internal_ds(ds, "resume receive", tx, "");
1641 }
1642
1643 /*
1644  * NB: callers *MUST* call dmu_recv_stream() if dmu_recv_begin()
1645  * succeeds; otherwise we will leak the holds on the datasets.
1646  */
1647 int
1648 dmu_recv_begin(char *tofs, char *tosnap, dmu_replay_record_t *drr_begin,
1649     boolean_t force, boolean_t resumable, char *origin, dmu_recv_cookie_t *drc)
1650 {
1651         dmu_recv_begin_arg_t drba = { 0 };
1652
1653         bzero(drc, sizeof (dmu_recv_cookie_t));
1654         drc->drc_drr_begin = drr_begin;
1655         drc->drc_drrb = &drr_begin->drr_u.drr_begin;
1656         drc->drc_tosnap = tosnap;
1657         drc->drc_tofs = tofs;
1658         drc->drc_force = force;
1659         drc->drc_resumable = resumable;
1660         drc->drc_cred = CRED();
1661
1662         if (drc->drc_drrb->drr_magic == BSWAP_64(DMU_BACKUP_MAGIC)) {
1663                 drc->drc_byteswap = B_TRUE;
1664                 fletcher_4_incremental_byteswap(drr_begin,
1665                     sizeof (dmu_replay_record_t), &drc->drc_cksum);
1666                 byteswap_record(drr_begin);
1667         } else if (drc->drc_drrb->drr_magic == DMU_BACKUP_MAGIC) {
1668                 fletcher_4_incremental_native(drr_begin,
1669                     sizeof (dmu_replay_record_t), &drc->drc_cksum);
1670         } else {
1671                 return (SET_ERROR(EINVAL));
1672         }
1673
1674         drba.drba_origin = origin;
1675         drba.drba_cookie = drc;
1676         drba.drba_cred = CRED();
1677
1678         if (DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo) &
1679             DMU_BACKUP_FEATURE_RESUMING) {
1680                 return (dsl_sync_task(tofs,
1681                     dmu_recv_resume_begin_check, dmu_recv_resume_begin_sync,
1682                     &drba, 5, ZFS_SPACE_CHECK_NORMAL));
1683         } else  {
1684                 return (dsl_sync_task(tofs,
1685                     dmu_recv_begin_check, dmu_recv_begin_sync,
1686                     &drba, 5, ZFS_SPACE_CHECK_NORMAL));
1687         }
1688 }
1689
1690 struct receive_record_arg {
1691         dmu_replay_record_t header;
1692         void *payload; /* Pointer to a buffer containing the payload */
1693         /*
1694          * If the record is a write, pointer to the arc_buf_t containing the
1695          * payload.
1696          */
1697         arc_buf_t *write_buf;
1698         int payload_size;
1699         uint64_t bytes_read; /* bytes read from stream when record created */
1700         boolean_t eos_marker; /* Marks the end of the stream */
1701         bqueue_node_t node;
1702 };
1703
1704 struct receive_writer_arg {
1705         objset_t *os;
1706         boolean_t byteswap;
1707         bqueue_t q;
1708
1709         /*
1710          * These three args are used to signal to the main thread that we're
1711          * done.
1712          */
1713         kmutex_t mutex;
1714         kcondvar_t cv;
1715         boolean_t done;
1716
1717         int err;
1718         /* A map from guid to dataset to help handle dedup'd streams. */
1719         avl_tree_t *guid_to_ds_map;
1720         boolean_t resumable;
1721         uint64_t last_object, last_offset;
1722         uint64_t bytes_read; /* bytes read when current record created */
1723 };
1724
1725 struct objlist {
1726         list_t list; /* List of struct receive_objnode. */
1727         /*
1728          * Last object looked up. Used to assert that objects are being looked
1729          * up in ascending order.
1730          */
1731         uint64_t last_lookup;
1732 };
1733
1734 struct receive_objnode {
1735         list_node_t node;
1736         uint64_t object;
1737 };
1738
1739 struct receive_arg  {
1740         objset_t *os;
1741         kthread_t *td;
1742         struct file *fp;
1743         uint64_t voff; /* The current offset in the stream */
1744         uint64_t bytes_read;
1745         /*
1746          * A record that has had its payload read in, but hasn't yet been handed
1747          * off to the worker thread.
1748          */
1749         struct receive_record_arg *rrd;
1750         /* A record that has had its header read in, but not its payload. */
1751         struct receive_record_arg *next_rrd;
1752         zio_cksum_t cksum;
1753         zio_cksum_t prev_cksum;
1754         int err;
1755         boolean_t byteswap;
1756         /* Sorted list of objects not to issue prefetches for. */
1757         struct objlist ignore_objlist;
1758 };
1759
1760 typedef struct guid_map_entry {
1761         uint64_t        guid;
1762         dsl_dataset_t   *gme_ds;
1763         avl_node_t      avlnode;
1764 } guid_map_entry_t;
1765
1766 static int
1767 guid_compare(const void *arg1, const void *arg2)
1768 {
1769         const guid_map_entry_t *gmep1 = arg1;
1770         const guid_map_entry_t *gmep2 = arg2;
1771
1772         if (gmep1->guid < gmep2->guid)
1773                 return (-1);
1774         else if (gmep1->guid > gmep2->guid)
1775                 return (1);
1776         return (0);
1777 }
1778
1779 static void
1780 free_guid_map_onexit(void *arg)
1781 {
1782         avl_tree_t *ca = arg;
1783         void *cookie = NULL;
1784         guid_map_entry_t *gmep;
1785
1786         while ((gmep = avl_destroy_nodes(ca, &cookie)) != NULL) {
1787                 dsl_dataset_long_rele(gmep->gme_ds, gmep);
1788                 dsl_dataset_rele(gmep->gme_ds, gmep);
1789                 kmem_free(gmep, sizeof (guid_map_entry_t));
1790         }
1791         avl_destroy(ca);
1792         kmem_free(ca, sizeof (avl_tree_t));
1793 }
1794
1795 static int
1796 restore_bytes(struct receive_arg *ra, void *buf, int len, off_t off, ssize_t *resid)
1797 {
1798         struct uio auio;
1799         struct iovec aiov;
1800         int error;
1801
1802         aiov.iov_base = buf;
1803         aiov.iov_len = len;
1804         auio.uio_iov = &aiov;
1805         auio.uio_iovcnt = 1;
1806         auio.uio_resid = len;
1807         auio.uio_segflg = UIO_SYSSPACE;
1808         auio.uio_rw = UIO_READ;
1809         auio.uio_offset = off;
1810         auio.uio_td = ra->td;
1811 #ifdef _KERNEL
1812         error = fo_read(ra->fp, &auio, ra->td->td_ucred, FOF_OFFSET, ra->td);
1813 #else
1814         fprintf(stderr, "%s: returning EOPNOTSUPP\n", __func__);
1815         error = EOPNOTSUPP;
1816 #endif
1817         *resid = auio.uio_resid;
1818         return (error);
1819 }
1820
1821 static int
1822 receive_read(struct receive_arg *ra, int len, void *buf)
1823 {
1824         int done = 0;
1825
1826         /* some things will require 8-byte alignment, so everything must */
1827         ASSERT0(len % 8);
1828
1829         while (done < len) {
1830                 ssize_t resid;
1831
1832                 ra->err = restore_bytes(ra, buf + done,
1833                     len - done, ra->voff, &resid);
1834
1835                 if (resid == len - done) {
1836                         /*
1837                          * Note: ECKSUM indicates that the receive
1838                          * was interrupted and can potentially be resumed.
1839                          */
1840                         ra->err = SET_ERROR(ECKSUM);
1841                 }
1842                 ra->voff += len - done - resid;
1843                 done = len - resid;
1844                 if (ra->err != 0)
1845                         return (ra->err);
1846         }
1847
1848         ra->bytes_read += len;
1849
1850         ASSERT3U(done, ==, len);
1851         return (0);
1852 }
1853
1854 static void
1855 byteswap_record(dmu_replay_record_t *drr)
1856 {
1857 #define DO64(X) (drr->drr_u.X = BSWAP_64(drr->drr_u.X))
1858 #define DO32(X) (drr->drr_u.X = BSWAP_32(drr->drr_u.X))
1859         drr->drr_type = BSWAP_32(drr->drr_type);
1860         drr->drr_payloadlen = BSWAP_32(drr->drr_payloadlen);
1861
1862         switch (drr->drr_type) {
1863         case DRR_BEGIN:
1864                 DO64(drr_begin.drr_magic);
1865                 DO64(drr_begin.drr_versioninfo);
1866                 DO64(drr_begin.drr_creation_time);
1867                 DO32(drr_begin.drr_type);
1868                 DO32(drr_begin.drr_flags);
1869                 DO64(drr_begin.drr_toguid);
1870                 DO64(drr_begin.drr_fromguid);
1871                 break;
1872         case DRR_OBJECT:
1873                 DO64(drr_object.drr_object);
1874                 DO32(drr_object.drr_type);
1875                 DO32(drr_object.drr_bonustype);
1876                 DO32(drr_object.drr_blksz);
1877                 DO32(drr_object.drr_bonuslen);
1878                 DO64(drr_object.drr_toguid);
1879                 break;
1880         case DRR_FREEOBJECTS:
1881                 DO64(drr_freeobjects.drr_firstobj);
1882                 DO64(drr_freeobjects.drr_numobjs);
1883                 DO64(drr_freeobjects.drr_toguid);
1884                 break;
1885         case DRR_WRITE:
1886                 DO64(drr_write.drr_object);
1887                 DO32(drr_write.drr_type);
1888                 DO64(drr_write.drr_offset);
1889                 DO64(drr_write.drr_length);
1890                 DO64(drr_write.drr_toguid);
1891                 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_write.drr_key.ddk_cksum);
1892                 DO64(drr_write.drr_key.ddk_prop);
1893                 break;
1894         case DRR_WRITE_BYREF:
1895                 DO64(drr_write_byref.drr_object);
1896                 DO64(drr_write_byref.drr_offset);
1897                 DO64(drr_write_byref.drr_length);
1898                 DO64(drr_write_byref.drr_toguid);
1899                 DO64(drr_write_byref.drr_refguid);
1900                 DO64(drr_write_byref.drr_refobject);
1901                 DO64(drr_write_byref.drr_refoffset);
1902                 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_write_byref.
1903                     drr_key.ddk_cksum);
1904                 DO64(drr_write_byref.drr_key.ddk_prop);
1905                 break;
1906         case DRR_WRITE_EMBEDDED:
1907                 DO64(drr_write_embedded.drr_object);
1908                 DO64(drr_write_embedded.drr_offset);
1909                 DO64(drr_write_embedded.drr_length);
1910                 DO64(drr_write_embedded.drr_toguid);
1911                 DO32(drr_write_embedded.drr_lsize);
1912                 DO32(drr_write_embedded.drr_psize);
1913                 break;
1914         case DRR_FREE:
1915                 DO64(drr_free.drr_object);
1916                 DO64(drr_free.drr_offset);
1917                 DO64(drr_free.drr_length);
1918                 DO64(drr_free.drr_toguid);
1919                 break;
1920         case DRR_SPILL:
1921                 DO64(drr_spill.drr_object);
1922                 DO64(drr_spill.drr_length);
1923                 DO64(drr_spill.drr_toguid);
1924                 break;
1925         case DRR_END:
1926                 DO64(drr_end.drr_toguid);
1927                 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_end.drr_checksum);
1928                 break;
1929         }
1930
1931         if (drr->drr_type != DRR_BEGIN) {
1932                 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_checksum.drr_checksum);
1933         }
1934
1935 #undef DO64
1936 #undef DO32
1937 }
1938
1939 static inline uint8_t
1940 deduce_nblkptr(dmu_object_type_t bonus_type, uint64_t bonus_size)
1941 {
1942         if (bonus_type == DMU_OT_SA) {
1943                 return (1);
1944         } else {
1945                 return (1 +
1946                     ((DN_MAX_BONUSLEN - bonus_size) >> SPA_BLKPTRSHIFT));
1947         }
1948 }
1949
1950 static void
1951 save_resume_state(struct receive_writer_arg *rwa,
1952     uint64_t object, uint64_t offset, dmu_tx_t *tx)
1953 {
1954         int txgoff = dmu_tx_get_txg(tx) & TXG_MASK;
1955
1956         if (!rwa->resumable)
1957                 return;
1958
1959         /*
1960          * We use ds_resume_bytes[] != 0 to indicate that we need to
1961          * update this on disk, so it must not be 0.
1962          */
1963         ASSERT(rwa->bytes_read != 0);
1964
1965         /*
1966          * We only resume from write records, which have a valid
1967          * (non-meta-dnode) object number.
1968          */
1969         ASSERT(object != 0);
1970
1971         /*
1972          * For resuming to work correctly, we must receive records in order,
1973          * sorted by object,offset.  This is checked by the callers, but
1974          * assert it here for good measure.
1975          */
1976         ASSERT3U(object, >=, rwa->os->os_dsl_dataset->ds_resume_object[txgoff]);
1977         ASSERT(object != rwa->os->os_dsl_dataset->ds_resume_object[txgoff] ||
1978             offset >= rwa->os->os_dsl_dataset->ds_resume_offset[txgoff]);
1979         ASSERT3U(rwa->bytes_read, >=,
1980             rwa->os->os_dsl_dataset->ds_resume_bytes[txgoff]);
1981
1982         rwa->os->os_dsl_dataset->ds_resume_object[txgoff] = object;
1983         rwa->os->os_dsl_dataset->ds_resume_offset[txgoff] = offset;
1984         rwa->os->os_dsl_dataset->ds_resume_bytes[txgoff] = rwa->bytes_read;
1985 }
1986
1987 static int
1988 receive_object(struct receive_writer_arg *rwa, struct drr_object *drro,
1989     void *data)
1990 {
1991         dmu_object_info_t doi;
1992         dmu_tx_t *tx;
1993         uint64_t object;
1994         int err;
1995
1996         if (drro->drr_type == DMU_OT_NONE ||
1997             !DMU_OT_IS_VALID(drro->drr_type) ||
1998             !DMU_OT_IS_VALID(drro->drr_bonustype) ||
1999             drro->drr_checksumtype >= ZIO_CHECKSUM_FUNCTIONS ||
2000             drro->drr_compress >= ZIO_COMPRESS_FUNCTIONS ||
2001             P2PHASE(drro->drr_blksz, SPA_MINBLOCKSIZE) ||
2002             drro->drr_blksz < SPA_MINBLOCKSIZE ||
2003             drro->drr_blksz > spa_maxblocksize(dmu_objset_spa(rwa->os)) ||
2004             drro->drr_bonuslen > DN_MAX_BONUSLEN) {
2005                 return (SET_ERROR(EINVAL));
2006         }
2007
2008         err = dmu_object_info(rwa->os, drro->drr_object, &doi);
2009
2010         if (err != 0 && err != ENOENT)
2011                 return (SET_ERROR(EINVAL));
2012         object = err == 0 ? drro->drr_object : DMU_NEW_OBJECT;
2013
2014         /*
2015          * If we are losing blkptrs or changing the block size this must
2016          * be a new file instance.  We must clear out the previous file
2017          * contents before we can change this type of metadata in the dnode.
2018          */
2019         if (err == 0) {
2020                 int nblkptr;
2021
2022                 nblkptr = deduce_nblkptr(drro->drr_bonustype,
2023                     drro->drr_bonuslen);
2024
2025                 if (drro->drr_blksz != doi.doi_data_block_size ||
2026                     nblkptr < doi.doi_nblkptr) {
2027                         err = dmu_free_long_range(rwa->os, drro->drr_object,
2028                             0, DMU_OBJECT_END);
2029                         if (err != 0)
2030                                 return (SET_ERROR(EINVAL));
2031                 }
2032         }
2033
2034         tx = dmu_tx_create(rwa->os);
2035         dmu_tx_hold_bonus(tx, object);
2036         err = dmu_tx_assign(tx, TXG_WAIT);
2037         if (err != 0) {
2038                 dmu_tx_abort(tx);
2039                 return (err);
2040         }
2041
2042         if (object == DMU_NEW_OBJECT) {
2043                 /* currently free, want to be allocated */
2044                 err = dmu_object_claim(rwa->os, drro->drr_object,
2045                     drro->drr_type, drro->drr_blksz,
2046                     drro->drr_bonustype, drro->drr_bonuslen, tx);
2047         } else if (drro->drr_type != doi.doi_type ||
2048             drro->drr_blksz != doi.doi_data_block_size ||
2049             drro->drr_bonustype != doi.doi_bonus_type ||
2050             drro->drr_bonuslen != doi.doi_bonus_size) {
2051                 /* currently allocated, but with different properties */
2052                 err = dmu_object_reclaim(rwa->os, drro->drr_object,
2053                     drro->drr_type, drro->drr_blksz,
2054                     drro->drr_bonustype, drro->drr_bonuslen, tx);
2055         }
2056         if (err != 0) {
2057                 dmu_tx_commit(tx);
2058                 return (SET_ERROR(EINVAL));
2059         }
2060
2061         dmu_object_set_checksum(rwa->os, drro->drr_object,
2062             drro->drr_checksumtype, tx);
2063         dmu_object_set_compress(rwa->os, drro->drr_object,
2064             drro->drr_compress, tx);
2065
2066         if (data != NULL) {
2067                 dmu_buf_t *db;
2068
2069                 VERIFY0(dmu_bonus_hold(rwa->os, drro->drr_object, FTAG, &db));
2070                 dmu_buf_will_dirty(db, tx);
2071
2072                 ASSERT3U(db->db_size, >=, drro->drr_bonuslen);
2073                 bcopy(data, db->db_data, drro->drr_bonuslen);
2074                 if (rwa->byteswap) {
2075                         dmu_object_byteswap_t byteswap =
2076                             DMU_OT_BYTESWAP(drro->drr_bonustype);
2077                         dmu_ot_byteswap[byteswap].ob_func(db->db_data,
2078                             drro->drr_bonuslen);
2079                 }
2080                 dmu_buf_rele(db, FTAG);
2081         }
2082         dmu_tx_commit(tx);
2083
2084         return (0);
2085 }
2086
2087 /* ARGSUSED */
2088 static int
2089 receive_freeobjects(struct receive_writer_arg *rwa,
2090     struct drr_freeobjects *drrfo)
2091 {
2092         uint64_t obj;
2093         int next_err = 0;
2094
2095         if (drrfo->drr_firstobj + drrfo->drr_numobjs < drrfo->drr_firstobj)
2096                 return (SET_ERROR(EINVAL));
2097
2098         for (obj = drrfo->drr_firstobj;
2099             obj < drrfo->drr_firstobj + drrfo->drr_numobjs && next_err == 0;
2100             next_err = dmu_object_next(rwa->os, &obj, FALSE, 0)) {
2101                 int err;
2102
2103                 if (dmu_object_info(rwa->os, obj, NULL) != 0)
2104                         continue;
2105
2106                 err = dmu_free_long_object(rwa->os, obj);
2107                 if (err != 0)
2108                         return (err);
2109         }
2110         if (next_err != ESRCH)
2111                 return (next_err);
2112         return (0);
2113 }
2114
2115 static int
2116 receive_write(struct receive_writer_arg *rwa, struct drr_write *drrw,
2117     arc_buf_t *abuf)
2118 {
2119         dmu_tx_t *tx;
2120         int err;
2121
2122         if (drrw->drr_offset + drrw->drr_length < drrw->drr_offset ||
2123             !DMU_OT_IS_VALID(drrw->drr_type))
2124                 return (SET_ERROR(EINVAL));
2125
2126         /*
2127          * For resuming to work, records must be in increasing order
2128          * by (object, offset).
2129          */
2130         if (drrw->drr_object < rwa->last_object ||
2131             (drrw->drr_object == rwa->last_object &&
2132             drrw->drr_offset < rwa->last_offset)) {
2133                 return (SET_ERROR(EINVAL));
2134         }
2135         rwa->last_object = drrw->drr_object;
2136         rwa->last_offset = drrw->drr_offset;
2137
2138         if (dmu_object_info(rwa->os, drrw->drr_object, NULL) != 0)
2139                 return (SET_ERROR(EINVAL));
2140
2141         tx = dmu_tx_create(rwa->os);
2142
2143         dmu_tx_hold_write(tx, drrw->drr_object,
2144             drrw->drr_offset, drrw->drr_length);
2145         err = dmu_tx_assign(tx, TXG_WAIT);
2146         if (err != 0) {
2147                 dmu_tx_abort(tx);
2148                 return (err);
2149         }
2150         if (rwa->byteswap) {
2151                 dmu_object_byteswap_t byteswap =
2152                     DMU_OT_BYTESWAP(drrw->drr_type);
2153                 dmu_ot_byteswap[byteswap].ob_func(abuf->b_data,
2154                     drrw->drr_length);
2155         }
2156
2157         dmu_buf_t *bonus;
2158         if (dmu_bonus_hold(rwa->os, drrw->drr_object, FTAG, &bonus) != 0)
2159                 return (SET_ERROR(EINVAL));
2160         dmu_assign_arcbuf(bonus, drrw->drr_offset, abuf, tx);
2161
2162         /*
2163          * Note: If the receive fails, we want the resume stream to start
2164          * with the same record that we last successfully received (as opposed
2165          * to the next record), so that we can verify that we are
2166          * resuming from the correct location.
2167          */
2168         save_resume_state(rwa, drrw->drr_object, drrw->drr_offset, tx);
2169         dmu_tx_commit(tx);
2170         dmu_buf_rele(bonus, FTAG);
2171
2172         return (0);
2173 }
2174
2175 /*
2176  * Handle a DRR_WRITE_BYREF record.  This record is used in dedup'ed
2177  * streams to refer to a copy of the data that is already on the
2178  * system because it came in earlier in the stream.  This function
2179  * finds the earlier copy of the data, and uses that copy instead of
2180  * data from the stream to fulfill this write.
2181  */
2182 static int
2183 receive_write_byref(struct receive_writer_arg *rwa,
2184     struct drr_write_byref *drrwbr)
2185 {
2186         dmu_tx_t *tx;
2187         int err;
2188         guid_map_entry_t gmesrch;
2189         guid_map_entry_t *gmep;
2190         avl_index_t where;
2191         objset_t *ref_os = NULL;
2192         dmu_buf_t *dbp;
2193
2194         if (drrwbr->drr_offset + drrwbr->drr_length < drrwbr->drr_offset)
2195                 return (SET_ERROR(EINVAL));
2196
2197         /*
2198          * If the GUID of the referenced dataset is different from the
2199          * GUID of the target dataset, find the referenced dataset.
2200          */
2201         if (drrwbr->drr_toguid != drrwbr->drr_refguid) {
2202                 gmesrch.guid = drrwbr->drr_refguid;
2203                 if ((gmep = avl_find(rwa->guid_to_ds_map, &gmesrch,
2204                     &where)) == NULL) {
2205                         return (SET_ERROR(EINVAL));
2206                 }
2207                 if (dmu_objset_from_ds(gmep->gme_ds, &ref_os))
2208                         return (SET_ERROR(EINVAL));
2209         } else {
2210                 ref_os = rwa->os;
2211         }
2212
2213         err = dmu_buf_hold(ref_os, drrwbr->drr_refobject,
2214             drrwbr->drr_refoffset, FTAG, &dbp, DMU_READ_PREFETCH);
2215         if (err != 0)
2216                 return (err);
2217
2218         tx = dmu_tx_create(rwa->os);
2219
2220         dmu_tx_hold_write(tx, drrwbr->drr_object,
2221             drrwbr->drr_offset, drrwbr->drr_length);
2222         err = dmu_tx_assign(tx, TXG_WAIT);
2223         if (err != 0) {
2224                 dmu_tx_abort(tx);
2225                 return (err);
2226         }
2227         dmu_write(rwa->os, drrwbr->drr_object,
2228             drrwbr->drr_offset, drrwbr->drr_length, dbp->db_data, tx);
2229         dmu_buf_rele(dbp, FTAG);
2230
2231         /* See comment in restore_write. */
2232         save_resume_state(rwa, drrwbr->drr_object, drrwbr->drr_offset, tx);
2233         dmu_tx_commit(tx);
2234         return (0);
2235 }
2236
2237 static int
2238 receive_write_embedded(struct receive_writer_arg *rwa,
2239     struct drr_write_embedded *drrwe, void *data)
2240 {
2241         dmu_tx_t *tx;
2242         int err;
2243
2244         if (drrwe->drr_offset + drrwe->drr_length < drrwe->drr_offset)
2245                 return (EINVAL);
2246
2247         if (drrwe->drr_psize > BPE_PAYLOAD_SIZE)
2248                 return (EINVAL);
2249
2250         if (drrwe->drr_etype >= NUM_BP_EMBEDDED_TYPES)
2251                 return (EINVAL);
2252         if (drrwe->drr_compression >= ZIO_COMPRESS_FUNCTIONS)
2253                 return (EINVAL);
2254
2255         tx = dmu_tx_create(rwa->os);
2256
2257         dmu_tx_hold_write(tx, drrwe->drr_object,
2258             drrwe->drr_offset, drrwe->drr_length);
2259         err = dmu_tx_assign(tx, TXG_WAIT);
2260         if (err != 0) {
2261                 dmu_tx_abort(tx);
2262                 return (err);
2263         }
2264
2265         dmu_write_embedded(rwa->os, drrwe->drr_object,
2266             drrwe->drr_offset, data, drrwe->drr_etype,
2267             drrwe->drr_compression, drrwe->drr_lsize, drrwe->drr_psize,
2268             rwa->byteswap ^ ZFS_HOST_BYTEORDER, tx);
2269
2270         /* See comment in restore_write. */
2271         save_resume_state(rwa, drrwe->drr_object, drrwe->drr_offset, tx);
2272         dmu_tx_commit(tx);
2273         return (0);
2274 }
2275
2276 static int
2277 receive_spill(struct receive_writer_arg *rwa, struct drr_spill *drrs,
2278     void *data)
2279 {
2280         dmu_tx_t *tx;
2281         dmu_buf_t *db, *db_spill;
2282         int err;
2283
2284         if (drrs->drr_length < SPA_MINBLOCKSIZE ||
2285             drrs->drr_length > spa_maxblocksize(dmu_objset_spa(rwa->os)))
2286                 return (SET_ERROR(EINVAL));
2287
2288         if (dmu_object_info(rwa->os, drrs->drr_object, NULL) != 0)
2289                 return (SET_ERROR(EINVAL));
2290
2291         VERIFY0(dmu_bonus_hold(rwa->os, drrs->drr_object, FTAG, &db));
2292         if ((err = dmu_spill_hold_by_bonus(db, FTAG, &db_spill)) != 0) {
2293                 dmu_buf_rele(db, FTAG);
2294                 return (err);
2295         }
2296
2297         tx = dmu_tx_create(rwa->os);
2298
2299         dmu_tx_hold_spill(tx, db->db_object);
2300
2301         err = dmu_tx_assign(tx, TXG_WAIT);
2302         if (err != 0) {
2303                 dmu_buf_rele(db, FTAG);
2304                 dmu_buf_rele(db_spill, FTAG);
2305                 dmu_tx_abort(tx);
2306                 return (err);
2307         }
2308         dmu_buf_will_dirty(db_spill, tx);
2309
2310         if (db_spill->db_size < drrs->drr_length)
2311                 VERIFY(0 == dbuf_spill_set_blksz(db_spill,
2312                     drrs->drr_length, tx));
2313         bcopy(data, db_spill->db_data, drrs->drr_length);
2314
2315         dmu_buf_rele(db, FTAG);
2316         dmu_buf_rele(db_spill, FTAG);
2317
2318         dmu_tx_commit(tx);
2319         return (0);
2320 }
2321
2322 /* ARGSUSED */
2323 static int
2324 receive_free(struct receive_writer_arg *rwa, struct drr_free *drrf)
2325 {
2326         int err;
2327
2328         if (drrf->drr_length != -1ULL &&
2329             drrf->drr_offset + drrf->drr_length < drrf->drr_offset)
2330                 return (SET_ERROR(EINVAL));
2331
2332         if (dmu_object_info(rwa->os, drrf->drr_object, NULL) != 0)
2333                 return (SET_ERROR(EINVAL));
2334
2335         err = dmu_free_long_range(rwa->os, drrf->drr_object,
2336             drrf->drr_offset, drrf->drr_length);
2337
2338         return (err);
2339 }
2340
2341 /* used to destroy the drc_ds on error */
2342 static void
2343 dmu_recv_cleanup_ds(dmu_recv_cookie_t *drc)
2344 {
2345         if (drc->drc_resumable) {
2346                 /* wait for our resume state to be written to disk */
2347                 txg_wait_synced(drc->drc_ds->ds_dir->dd_pool, 0);
2348                 dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
2349         } else {
2350                 char name[MAXNAMELEN];
2351                 dsl_dataset_name(drc->drc_ds, name);
2352                 dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
2353                 (void) dsl_destroy_head(name);
2354         }
2355 }
2356
2357 static void
2358 receive_cksum(struct receive_arg *ra, int len, void *buf)
2359 {
2360         if (ra->byteswap) {
2361                 fletcher_4_incremental_byteswap(buf, len, &ra->cksum);
2362         } else {
2363                 fletcher_4_incremental_native(buf, len, &ra->cksum);
2364         }
2365 }
2366
2367 /*
2368  * Read the payload into a buffer of size len, and update the current record's
2369  * payload field.
2370  * Allocate ra->next_rrd and read the next record's header into
2371  * ra->next_rrd->header.
2372  * Verify checksum of payload and next record.
2373  */
2374 static int
2375 receive_read_payload_and_next_header(struct receive_arg *ra, int len, void *buf)
2376 {
2377         int err;
2378
2379         if (len != 0) {
2380                 ASSERT3U(len, <=, SPA_MAXBLOCKSIZE);
2381                 err = receive_read(ra, len, buf);
2382                 if (err != 0)
2383                         return (err);
2384                 receive_cksum(ra, len, buf);
2385
2386                 /* note: rrd is NULL when reading the begin record's payload */
2387                 if (ra->rrd != NULL) {
2388                         ra->rrd->payload = buf;
2389                         ra->rrd->payload_size = len;
2390                         ra->rrd->bytes_read = ra->bytes_read;
2391                 }
2392         }
2393
2394         ra->prev_cksum = ra->cksum;
2395
2396         ra->next_rrd = kmem_zalloc(sizeof (*ra->next_rrd), KM_SLEEP);
2397         err = receive_read(ra, sizeof (ra->next_rrd->header),
2398             &ra->next_rrd->header);
2399         ra->next_rrd->bytes_read = ra->bytes_read;
2400         if (err != 0) {
2401                 kmem_free(ra->next_rrd, sizeof (*ra->next_rrd));
2402                 ra->next_rrd = NULL;
2403                 return (err);
2404         }
2405         if (ra->next_rrd->header.drr_type == DRR_BEGIN) {
2406                 kmem_free(ra->next_rrd, sizeof (*ra->next_rrd));
2407                 ra->next_rrd = NULL;
2408                 return (SET_ERROR(EINVAL));
2409         }
2410
2411         /*
2412          * Note: checksum is of everything up to but not including the
2413          * checksum itself.
2414          */
2415         ASSERT3U(offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
2416             ==, sizeof (dmu_replay_record_t) - sizeof (zio_cksum_t));
2417         receive_cksum(ra,
2418             offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
2419             &ra->next_rrd->header);
2420
2421         zio_cksum_t cksum_orig =
2422             ra->next_rrd->header.drr_u.drr_checksum.drr_checksum;
2423         zio_cksum_t *cksump =
2424             &ra->next_rrd->header.drr_u.drr_checksum.drr_checksum;
2425
2426         if (ra->byteswap)
2427                 byteswap_record(&ra->next_rrd->header);
2428
2429         if ((!ZIO_CHECKSUM_IS_ZERO(cksump)) &&
2430             !ZIO_CHECKSUM_EQUAL(ra->cksum, *cksump)) {
2431                 kmem_free(ra->next_rrd, sizeof (*ra->next_rrd));
2432                 ra->next_rrd = NULL;
2433                 return (SET_ERROR(ECKSUM));
2434         }
2435
2436         receive_cksum(ra, sizeof (cksum_orig), &cksum_orig);
2437
2438         return (0);
2439 }
2440
2441 static void
2442 objlist_create(struct objlist *list)
2443 {
2444         list_create(&list->list, sizeof (struct receive_objnode),
2445             offsetof(struct receive_objnode, node));
2446         list->last_lookup = 0;
2447 }
2448
2449 static void
2450 objlist_destroy(struct objlist *list)
2451 {
2452         for (struct receive_objnode *n = list_remove_head(&list->list);
2453             n != NULL; n = list_remove_head(&list->list)) {
2454                 kmem_free(n, sizeof (*n));
2455         }
2456         list_destroy(&list->list);
2457 }
2458
2459 /*
2460  * This function looks through the objlist to see if the specified object number
2461  * is contained in the objlist.  In the process, it will remove all object
2462  * numbers in the list that are smaller than the specified object number.  Thus,
2463  * any lookup of an object number smaller than a previously looked up object
2464  * number will always return false; therefore, all lookups should be done in
2465  * ascending order.
2466  */
2467 static boolean_t
2468 objlist_exists(struct objlist *list, uint64_t object)
2469 {
2470         struct receive_objnode *node = list_head(&list->list);
2471         ASSERT3U(object, >=, list->last_lookup);
2472         list->last_lookup = object;
2473         while (node != NULL && node->object < object) {
2474                 VERIFY3P(node, ==, list_remove_head(&list->list));
2475                 kmem_free(node, sizeof (*node));
2476                 node = list_head(&list->list);
2477         }
2478         return (node != NULL && node->object == object);
2479 }
2480
2481 /*
2482  * The objlist is a list of object numbers stored in ascending order.  However,
2483  * the insertion of new object numbers does not seek out the correct location to
2484  * store a new object number; instead, it appends it to the list for simplicity.
2485  * Thus, any users must take care to only insert new object numbers in ascending
2486  * order.
2487  */
2488 static void
2489 objlist_insert(struct objlist *list, uint64_t object)
2490 {
2491         struct receive_objnode *node = kmem_zalloc(sizeof (*node), KM_SLEEP);
2492         node->object = object;
2493 #ifdef ZFS_DEBUG
2494         struct receive_objnode *last_object = list_tail(&list->list);
2495         uint64_t last_objnum = (last_object != NULL ? last_object->object : 0);
2496         ASSERT3U(node->object, >, last_objnum);
2497 #endif
2498         list_insert_tail(&list->list, node);
2499 }
2500
2501 /*
2502  * Issue the prefetch reads for any necessary indirect blocks.
2503  *
2504  * We use the object ignore list to tell us whether or not to issue prefetches
2505  * for a given object.  We do this for both correctness (in case the blocksize
2506  * of an object has changed) and performance (if the object doesn't exist, don't
2507  * needlessly try to issue prefetches).  We also trim the list as we go through
2508  * the stream to prevent it from growing to an unbounded size.
2509  *
2510  * The object numbers within will always be in sorted order, and any write
2511  * records we see will also be in sorted order, but they're not sorted with
2512  * respect to each other (i.e. we can get several object records before
2513  * receiving each object's write records).  As a result, once we've reached a
2514  * given object number, we can safely remove any reference to lower object
2515  * numbers in the ignore list. In practice, we receive up to 32 object records
2516  * before receiving write records, so the list can have up to 32 nodes in it.
2517  */
2518 /* ARGSUSED */
2519 static void
2520 receive_read_prefetch(struct receive_arg *ra,
2521     uint64_t object, uint64_t offset, uint64_t length)
2522 {
2523         if (!objlist_exists(&ra->ignore_objlist, object)) {
2524                 dmu_prefetch(ra->os, object, 1, offset, length,
2525                     ZIO_PRIORITY_SYNC_READ);
2526         }
2527 }
2528
2529 /*
2530  * Read records off the stream, issuing any necessary prefetches.
2531  */
2532 static int
2533 receive_read_record(struct receive_arg *ra)
2534 {
2535         int err;
2536
2537         switch (ra->rrd->header.drr_type) {
2538         case DRR_OBJECT:
2539         {
2540                 struct drr_object *drro = &ra->rrd->header.drr_u.drr_object;
2541                 uint32_t size = P2ROUNDUP(drro->drr_bonuslen, 8);
2542                 void *buf = kmem_zalloc(size, KM_SLEEP);
2543                 dmu_object_info_t doi;
2544                 err = receive_read_payload_and_next_header(ra, size, buf);
2545                 if (err != 0) {
2546                         kmem_free(buf, size);
2547                         return (err);
2548                 }
2549                 err = dmu_object_info(ra->os, drro->drr_object, &doi);
2550                 /*
2551                  * See receive_read_prefetch for an explanation why we're
2552                  * storing this object in the ignore_obj_list.
2553                  */
2554                 if (err == ENOENT ||
2555                     (err == 0 && doi.doi_data_block_size != drro->drr_blksz)) {
2556                         objlist_insert(&ra->ignore_objlist, drro->drr_object);
2557                         err = 0;
2558                 }
2559                 return (err);
2560         }
2561         case DRR_FREEOBJECTS:
2562         {
2563                 err = receive_read_payload_and_next_header(ra, 0, NULL);
2564                 return (err);
2565         }
2566         case DRR_WRITE:
2567         {
2568                 struct drr_write *drrw = &ra->rrd->header.drr_u.drr_write;
2569                 arc_buf_t *abuf = arc_loan_buf(dmu_objset_spa(ra->os),
2570                     drrw->drr_length);
2571
2572                 err = receive_read_payload_and_next_header(ra,
2573                     drrw->drr_length, abuf->b_data);
2574                 if (err != 0) {
2575                         dmu_return_arcbuf(abuf);
2576                         return (err);
2577                 }
2578                 ra->rrd->write_buf = abuf;
2579                 receive_read_prefetch(ra, drrw->drr_object, drrw->drr_offset,
2580                     drrw->drr_length);
2581                 return (err);
2582         }
2583         case DRR_WRITE_BYREF:
2584         {
2585                 struct drr_write_byref *drrwb =
2586                     &ra->rrd->header.drr_u.drr_write_byref;
2587                 err = receive_read_payload_and_next_header(ra, 0, NULL);
2588                 receive_read_prefetch(ra, drrwb->drr_object, drrwb->drr_offset,
2589                     drrwb->drr_length);
2590                 return (err);
2591         }
2592         case DRR_WRITE_EMBEDDED:
2593         {
2594                 struct drr_write_embedded *drrwe =
2595                     &ra->rrd->header.drr_u.drr_write_embedded;
2596                 uint32_t size = P2ROUNDUP(drrwe->drr_psize, 8);
2597                 void *buf = kmem_zalloc(size, KM_SLEEP);
2598
2599                 err = receive_read_payload_and_next_header(ra, size, buf);
2600                 if (err != 0) {
2601                         kmem_free(buf, size);
2602                         return (err);
2603                 }
2604
2605                 receive_read_prefetch(ra, drrwe->drr_object, drrwe->drr_offset,
2606                     drrwe->drr_length);
2607                 return (err);
2608         }
2609         case DRR_FREE:
2610         {
2611                 /*
2612                  * It might be beneficial to prefetch indirect blocks here, but
2613                  * we don't really have the data to decide for sure.
2614                  */
2615                 err = receive_read_payload_and_next_header(ra, 0, NULL);
2616                 return (err);
2617         }
2618         case DRR_END:
2619         {
2620                 struct drr_end *drre = &ra->rrd->header.drr_u.drr_end;
2621                 if (!ZIO_CHECKSUM_EQUAL(ra->prev_cksum, drre->drr_checksum))
2622                         return (SET_ERROR(ECKSUM));
2623                 return (0);
2624         }
2625         case DRR_SPILL:
2626         {
2627                 struct drr_spill *drrs = &ra->rrd->header.drr_u.drr_spill;
2628                 void *buf = kmem_zalloc(drrs->drr_length, KM_SLEEP);
2629                 err = receive_read_payload_and_next_header(ra, drrs->drr_length,
2630                     buf);
2631                 if (err != 0)
2632                         kmem_free(buf, drrs->drr_length);
2633                 return (err);
2634         }
2635         default:
2636                 return (SET_ERROR(EINVAL));
2637         }
2638 }
2639
2640 /*
2641  * Commit the records to the pool.
2642  */
2643 static int
2644 receive_process_record(struct receive_writer_arg *rwa,
2645     struct receive_record_arg *rrd)
2646 {
2647         int err;
2648
2649         /* Processing in order, therefore bytes_read should be increasing. */
2650         ASSERT3U(rrd->bytes_read, >=, rwa->bytes_read);
2651         rwa->bytes_read = rrd->bytes_read;
2652
2653         switch (rrd->header.drr_type) {
2654         case DRR_OBJECT:
2655         {
2656                 struct drr_object *drro = &rrd->header.drr_u.drr_object;
2657                 err = receive_object(rwa, drro, rrd->payload);
2658                 kmem_free(rrd->payload, rrd->payload_size);
2659                 rrd->payload = NULL;
2660                 return (err);
2661         }
2662         case DRR_FREEOBJECTS:
2663         {
2664                 struct drr_freeobjects *drrfo =
2665                     &rrd->header.drr_u.drr_freeobjects;
2666                 return (receive_freeobjects(rwa, drrfo));
2667         }
2668         case DRR_WRITE:
2669         {
2670                 struct drr_write *drrw = &rrd->header.drr_u.drr_write;
2671                 err = receive_write(rwa, drrw, rrd->write_buf);
2672                 /* if receive_write() is successful, it consumes the arc_buf */
2673                 if (err != 0)
2674                         dmu_return_arcbuf(rrd->write_buf);
2675                 rrd->write_buf = NULL;
2676                 rrd->payload = NULL;
2677                 return (err);
2678         }
2679         case DRR_WRITE_BYREF:
2680         {
2681                 struct drr_write_byref *drrwbr =
2682                     &rrd->header.drr_u.drr_write_byref;
2683                 return (receive_write_byref(rwa, drrwbr));
2684         }
2685         case DRR_WRITE_EMBEDDED:
2686         {
2687                 struct drr_write_embedded *drrwe =
2688                     &rrd->header.drr_u.drr_write_embedded;
2689                 err = receive_write_embedded(rwa, drrwe, rrd->payload);
2690                 kmem_free(rrd->payload, rrd->payload_size);
2691                 rrd->payload = NULL;
2692                 return (err);
2693         }
2694         case DRR_FREE:
2695         {
2696                 struct drr_free *drrf = &rrd->header.drr_u.drr_free;
2697                 return (receive_free(rwa, drrf));
2698         }
2699         case DRR_SPILL:
2700         {
2701                 struct drr_spill *drrs = &rrd->header.drr_u.drr_spill;
2702                 err = receive_spill(rwa, drrs, rrd->payload);
2703                 kmem_free(rrd->payload, rrd->payload_size);
2704                 rrd->payload = NULL;
2705                 return (err);
2706         }
2707         default:
2708                 return (SET_ERROR(EINVAL));
2709         }
2710 }
2711
2712 /*
2713  * dmu_recv_stream's worker thread; pull records off the queue, and then call
2714  * receive_process_record  When we're done, signal the main thread and exit.
2715  */
2716 static void
2717 receive_writer_thread(void *arg)
2718 {
2719         struct receive_writer_arg *rwa = arg;
2720         struct receive_record_arg *rrd;
2721         for (rrd = bqueue_dequeue(&rwa->q); !rrd->eos_marker;
2722             rrd = bqueue_dequeue(&rwa->q)) {
2723                 /*
2724                  * If there's an error, the main thread will stop putting things
2725                  * on the queue, but we need to clear everything in it before we
2726                  * can exit.
2727                  */
2728                 if (rwa->err == 0) {
2729                         rwa->err = receive_process_record(rwa, rrd);
2730                 } else if (rrd->write_buf != NULL) {
2731                         dmu_return_arcbuf(rrd->write_buf);
2732                         rrd->write_buf = NULL;
2733                         rrd->payload = NULL;
2734                 } else if (rrd->payload != NULL) {
2735                         kmem_free(rrd->payload, rrd->payload_size);
2736                         rrd->payload = NULL;
2737                 }
2738                 kmem_free(rrd, sizeof (*rrd));
2739         }
2740         kmem_free(rrd, sizeof (*rrd));
2741         mutex_enter(&rwa->mutex);
2742         rwa->done = B_TRUE;
2743         cv_signal(&rwa->cv);
2744         mutex_exit(&rwa->mutex);
2745         thread_exit();
2746 }
2747
2748 static int
2749 resume_check(struct receive_arg *ra, nvlist_t *begin_nvl)
2750 {
2751         uint64_t val;
2752         objset_t *mos = dmu_objset_pool(ra->os)->dp_meta_objset;
2753         uint64_t dsobj = dmu_objset_id(ra->os);
2754         uint64_t resume_obj, resume_off;
2755
2756         if (nvlist_lookup_uint64(begin_nvl,
2757             "resume_object", &resume_obj) != 0 ||
2758             nvlist_lookup_uint64(begin_nvl,
2759             "resume_offset", &resume_off) != 0) {
2760                 return (SET_ERROR(EINVAL));
2761         }
2762         VERIFY0(zap_lookup(mos, dsobj,
2763             DS_FIELD_RESUME_OBJECT, sizeof (val), 1, &val));
2764         if (resume_obj != val)
2765                 return (SET_ERROR(EINVAL));
2766         VERIFY0(zap_lookup(mos, dsobj,
2767             DS_FIELD_RESUME_OFFSET, sizeof (val), 1, &val));
2768         if (resume_off != val)
2769                 return (SET_ERROR(EINVAL));
2770
2771         return (0);
2772 }
2773
2774 /*
2775  * Read in the stream's records, one by one, and apply them to the pool.  There
2776  * are two threads involved; the thread that calls this function will spin up a
2777  * worker thread, read the records off the stream one by one, and issue
2778  * prefetches for any necessary indirect blocks.  It will then push the records
2779  * onto an internal blocking queue.  The worker thread will pull the records off
2780  * the queue, and actually write the data into the DMU.  This way, the worker
2781  * thread doesn't have to wait for reads to complete, since everything it needs
2782  * (the indirect blocks) will be prefetched.
2783  *
2784  * NB: callers *must* call dmu_recv_end() if this succeeds.
2785  */
2786 int
2787 dmu_recv_stream(dmu_recv_cookie_t *drc, struct file *fp, offset_t *voffp,
2788     int cleanup_fd, uint64_t *action_handlep)
2789 {
2790         int err = 0;
2791         struct receive_arg ra = { 0 };
2792         struct receive_writer_arg rwa = { 0 };
2793         int featureflags;
2794         nvlist_t *begin_nvl = NULL;
2795
2796         ra.byteswap = drc->drc_byteswap;
2797         ra.cksum = drc->drc_cksum;
2798         ra.td = curthread;
2799         ra.fp = fp;
2800         ra.voff = *voffp;
2801
2802         if (dsl_dataset_is_zapified(drc->drc_ds)) {
2803                 (void) zap_lookup(drc->drc_ds->ds_dir->dd_pool->dp_meta_objset,
2804                     drc->drc_ds->ds_object, DS_FIELD_RESUME_BYTES,
2805                     sizeof (ra.bytes_read), 1, &ra.bytes_read);
2806         }
2807
2808         objlist_create(&ra.ignore_objlist);
2809
2810         /* these were verified in dmu_recv_begin */
2811         ASSERT3U(DMU_GET_STREAM_HDRTYPE(drc->drc_drrb->drr_versioninfo), ==,
2812             DMU_SUBSTREAM);
2813         ASSERT3U(drc->drc_drrb->drr_type, <, DMU_OST_NUMTYPES);
2814
2815         /*
2816          * Open the objset we are modifying.
2817          */
2818         VERIFY0(dmu_objset_from_ds(drc->drc_ds, &ra.os));
2819
2820         ASSERT(dsl_dataset_phys(drc->drc_ds)->ds_flags & DS_FLAG_INCONSISTENT);
2821
2822         featureflags = DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo);
2823
2824         /* if this stream is dedup'ed, set up the avl tree for guid mapping */
2825         if (featureflags & DMU_BACKUP_FEATURE_DEDUP) {
2826                 minor_t minor;
2827
2828                 if (cleanup_fd == -1) {
2829                         ra.err = SET_ERROR(EBADF);
2830                         goto out;
2831                 }
2832                 ra.err = zfs_onexit_fd_hold(cleanup_fd, &minor);
2833                 if (ra.err != 0) {
2834                         cleanup_fd = -1;
2835                         goto out;
2836                 }
2837
2838                 if (*action_handlep == 0) {
2839                         rwa.guid_to_ds_map =
2840                             kmem_alloc(sizeof (avl_tree_t), KM_SLEEP);
2841                         avl_create(rwa.guid_to_ds_map, guid_compare,
2842                             sizeof (guid_map_entry_t),
2843                             offsetof(guid_map_entry_t, avlnode));
2844                         err = zfs_onexit_add_cb(minor,
2845                             free_guid_map_onexit, rwa.guid_to_ds_map,
2846                             action_handlep);
2847                         if (ra.err != 0)
2848                                 goto out;
2849                 } else {
2850                         err = zfs_onexit_cb_data(minor, *action_handlep,
2851                             (void **)&rwa.guid_to_ds_map);
2852                         if (ra.err != 0)
2853                                 goto out;
2854                 }
2855
2856                 drc->drc_guid_to_ds_map = rwa.guid_to_ds_map;
2857         }
2858
2859         uint32_t payloadlen = drc->drc_drr_begin->drr_payloadlen;
2860         void *payload = NULL;
2861         if (payloadlen != 0)
2862                 payload = kmem_alloc(payloadlen, KM_SLEEP);
2863
2864         err = receive_read_payload_and_next_header(&ra, payloadlen, payload);
2865         if (err != 0) {
2866                 if (payloadlen != 0)
2867                         kmem_free(payload, payloadlen);
2868                 goto out;
2869         }
2870         if (payloadlen != 0) {
2871                 err = nvlist_unpack(payload, payloadlen, &begin_nvl, KM_SLEEP);
2872                 kmem_free(payload, payloadlen);
2873                 if (err != 0)
2874                         goto out;
2875         }
2876
2877         if (featureflags & DMU_BACKUP_FEATURE_RESUMING) {
2878                 err = resume_check(&ra, begin_nvl);
2879                 if (err != 0)
2880                         goto out;
2881         }
2882
2883         (void) bqueue_init(&rwa.q, zfs_recv_queue_length,
2884             offsetof(struct receive_record_arg, node));
2885         cv_init(&rwa.cv, NULL, CV_DEFAULT, NULL);
2886         mutex_init(&rwa.mutex, NULL, MUTEX_DEFAULT, NULL);
2887         rwa.os = ra.os;
2888         rwa.byteswap = drc->drc_byteswap;
2889         rwa.resumable = drc->drc_resumable;
2890
2891         (void) thread_create(NULL, 0, receive_writer_thread, &rwa, 0, &p0,
2892             TS_RUN, minclsyspri);
2893         /*
2894          * We're reading rwa.err without locks, which is safe since we are the
2895          * only reader, and the worker thread is the only writer.  It's ok if we
2896          * miss a write for an iteration or two of the loop, since the writer
2897          * thread will keep freeing records we send it until we send it an eos
2898          * marker.
2899          *
2900          * We can leave this loop in 3 ways:  First, if rwa.err is
2901          * non-zero.  In that case, the writer thread will free the rrd we just
2902          * pushed.  Second, if  we're interrupted; in that case, either it's the
2903          * first loop and ra.rrd was never allocated, or it's later, and ra.rrd
2904          * has been handed off to the writer thread who will free it.  Finally,
2905          * if receive_read_record fails or we're at the end of the stream, then
2906          * we free ra.rrd and exit.
2907          */
2908         while (rwa.err == 0) {
2909                 if (issig(JUSTLOOKING) && issig(FORREAL)) {
2910                         err = SET_ERROR(EINTR);
2911                         break;
2912                 }
2913
2914                 ASSERT3P(ra.rrd, ==, NULL);
2915                 ra.rrd = ra.next_rrd;
2916                 ra.next_rrd = NULL;
2917                 /* Allocates and loads header into ra.next_rrd */
2918                 err = receive_read_record(&ra);
2919
2920                 if (ra.rrd->header.drr_type == DRR_END || err != 0) {
2921                         kmem_free(ra.rrd, sizeof (*ra.rrd));
2922                         ra.rrd = NULL;
2923                         break;
2924                 }
2925
2926                 bqueue_enqueue(&rwa.q, ra.rrd,
2927                     sizeof (struct receive_record_arg) + ra.rrd->payload_size);
2928                 ra.rrd = NULL;
2929         }
2930         if (ra.next_rrd == NULL)
2931                 ra.next_rrd = kmem_zalloc(sizeof (*ra.next_rrd), KM_SLEEP);
2932         ra.next_rrd->eos_marker = B_TRUE;
2933         bqueue_enqueue(&rwa.q, ra.next_rrd, 1);
2934
2935         mutex_enter(&rwa.mutex);
2936         while (!rwa.done) {
2937                 cv_wait(&rwa.cv, &rwa.mutex);
2938         }
2939         mutex_exit(&rwa.mutex);
2940
2941         cv_destroy(&rwa.cv);
2942         mutex_destroy(&rwa.mutex);
2943         bqueue_destroy(&rwa.q);
2944         if (err == 0)
2945                 err = rwa.err;
2946
2947 out:
2948         nvlist_free(begin_nvl);
2949         if ((featureflags & DMU_BACKUP_FEATURE_DEDUP) && (cleanup_fd != -1))
2950                 zfs_onexit_fd_rele(cleanup_fd);
2951
2952         if (err != 0) {
2953                 /*
2954                  * Clean up references. If receive is not resumable,
2955                  * destroy what we created, so we don't leave it in
2956                  * the inconsistent state.
2957                  */
2958                 dmu_recv_cleanup_ds(drc);
2959         }
2960
2961         *voffp = ra.voff;
2962         objlist_destroy(&ra.ignore_objlist);
2963         return (err);
2964 }
2965
2966 static int
2967 dmu_recv_end_check(void *arg, dmu_tx_t *tx)
2968 {
2969         dmu_recv_cookie_t *drc = arg;
2970         dsl_pool_t *dp = dmu_tx_pool(tx);
2971         int error;
2972
2973         ASSERT3P(drc->drc_ds->ds_owner, ==, dmu_recv_tag);
2974
2975         if (!drc->drc_newfs) {
2976                 dsl_dataset_t *origin_head;
2977
2978                 error = dsl_dataset_hold(dp, drc->drc_tofs, FTAG, &origin_head);
2979                 if (error != 0)
2980                         return (error);
2981                 if (drc->drc_force) {
2982                         /*
2983                          * We will destroy any snapshots in tofs (i.e. before
2984                          * origin_head) that are after the origin (which is
2985                          * the snap before drc_ds, because drc_ds can not
2986                          * have any snaps of its own).
2987                          */
2988                         uint64_t obj;
2989
2990                         obj = dsl_dataset_phys(origin_head)->ds_prev_snap_obj;
2991                         while (obj !=
2992                             dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj) {
2993                                 dsl_dataset_t *snap;
2994                                 error = dsl_dataset_hold_obj(dp, obj, FTAG,
2995                                     &snap);
2996                                 if (error != 0)
2997                                         break;
2998                                 if (snap->ds_dir != origin_head->ds_dir)
2999                                         error = SET_ERROR(EINVAL);
3000                                 if (error == 0)  {
3001                                         error = dsl_destroy_snapshot_check_impl(
3002                                             snap, B_FALSE);
3003                                 }
3004                                 obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
3005                                 dsl_dataset_rele(snap, FTAG);
3006                                 if (error != 0)
3007                                         break;
3008                         }
3009                         if (error != 0) {
3010                                 dsl_dataset_rele(origin_head, FTAG);
3011                                 return (error);
3012                         }
3013                 }
3014                 error = dsl_dataset_clone_swap_check_impl(drc->drc_ds,
3015                     origin_head, drc->drc_force, drc->drc_owner, tx);
3016                 if (error != 0) {
3017                         dsl_dataset_rele(origin_head, FTAG);
3018                         return (error);
3019                 }
3020                 error = dsl_dataset_snapshot_check_impl(origin_head,
3021                     drc->drc_tosnap, tx, B_TRUE, 1, drc->drc_cred);
3022                 dsl_dataset_rele(origin_head, FTAG);
3023                 if (error != 0)
3024                         return (error);
3025
3026                 error = dsl_destroy_head_check_impl(drc->drc_ds, 1);
3027         } else {
3028                 error = dsl_dataset_snapshot_check_impl(drc->drc_ds,
3029                     drc->drc_tosnap, tx, B_TRUE, 1, drc->drc_cred);
3030         }
3031         return (error);
3032 }
3033
3034 static void
3035 dmu_recv_end_sync(void *arg, dmu_tx_t *tx)
3036 {
3037         dmu_recv_cookie_t *drc = arg;
3038         dsl_pool_t *dp = dmu_tx_pool(tx);
3039
3040         spa_history_log_internal_ds(drc->drc_ds, "finish receiving",
3041             tx, "snap=%s", drc->drc_tosnap);
3042
3043         if (!drc->drc_newfs) {
3044                 dsl_dataset_t *origin_head;
3045
3046                 VERIFY0(dsl_dataset_hold(dp, drc->drc_tofs, FTAG,
3047                     &origin_head));
3048
3049                 if (drc->drc_force) {
3050                         /*
3051                          * Destroy any snapshots of drc_tofs (origin_head)
3052                          * after the origin (the snap before drc_ds).
3053                          */
3054                         uint64_t obj;
3055
3056                         obj = dsl_dataset_phys(origin_head)->ds_prev_snap_obj;
3057                         while (obj !=
3058                             dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj) {
3059                                 dsl_dataset_t *snap;
3060                                 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG,
3061                                     &snap));
3062                                 ASSERT3P(snap->ds_dir, ==, origin_head->ds_dir);
3063                                 obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
3064                                 dsl_destroy_snapshot_sync_impl(snap,
3065                                     B_FALSE, tx);
3066                                 dsl_dataset_rele(snap, FTAG);
3067                         }
3068                 }
3069                 VERIFY3P(drc->drc_ds->ds_prev, ==,
3070                     origin_head->ds_prev);
3071
3072                 dsl_dataset_clone_swap_sync_impl(drc->drc_ds,
3073                     origin_head, tx);
3074                 dsl_dataset_snapshot_sync_impl(origin_head,
3075                     drc->drc_tosnap, tx);
3076
3077                 /* set snapshot's creation time and guid */
3078                 dmu_buf_will_dirty(origin_head->ds_prev->ds_dbuf, tx);
3079                 dsl_dataset_phys(origin_head->ds_prev)->ds_creation_time =
3080                     drc->drc_drrb->drr_creation_time;
3081                 dsl_dataset_phys(origin_head->ds_prev)->ds_guid =
3082                     drc->drc_drrb->drr_toguid;
3083                 dsl_dataset_phys(origin_head->ds_prev)->ds_flags &=
3084                     ~DS_FLAG_INCONSISTENT;
3085
3086                 dmu_buf_will_dirty(origin_head->ds_dbuf, tx);
3087                 dsl_dataset_phys(origin_head)->ds_flags &=
3088                     ~DS_FLAG_INCONSISTENT;
3089
3090                 dsl_dataset_rele(origin_head, FTAG);
3091                 dsl_destroy_head_sync_impl(drc->drc_ds, tx);
3092
3093                 if (drc->drc_owner != NULL)
3094                         VERIFY3P(origin_head->ds_owner, ==, drc->drc_owner);
3095         } else {
3096                 dsl_dataset_t *ds = drc->drc_ds;
3097
3098                 dsl_dataset_snapshot_sync_impl(ds, drc->drc_tosnap, tx);
3099
3100                 /* set snapshot's creation time and guid */
3101                 dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
3102                 dsl_dataset_phys(ds->ds_prev)->ds_creation_time =
3103                     drc->drc_drrb->drr_creation_time;
3104                 dsl_dataset_phys(ds->ds_prev)->ds_guid =
3105                     drc->drc_drrb->drr_toguid;
3106                 dsl_dataset_phys(ds->ds_prev)->ds_flags &=
3107                     ~DS_FLAG_INCONSISTENT;
3108
3109                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
3110                 dsl_dataset_phys(ds)->ds_flags &= ~DS_FLAG_INCONSISTENT;
3111                 if (dsl_dataset_has_resume_receive_state(ds)) {
3112                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3113                             DS_FIELD_RESUME_FROMGUID, tx);
3114                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3115                             DS_FIELD_RESUME_OBJECT, tx);
3116                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3117                             DS_FIELD_RESUME_OFFSET, tx);
3118                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3119                             DS_FIELD_RESUME_BYTES, tx);
3120                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3121                             DS_FIELD_RESUME_TOGUID, tx);
3122                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3123                             DS_FIELD_RESUME_TONAME, tx);
3124                 }
3125         }
3126         drc->drc_newsnapobj = dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj;
3127         /*
3128          * Release the hold from dmu_recv_begin.  This must be done before
3129          * we return to open context, so that when we free the dataset's dnode,
3130          * we can evict its bonus buffer.
3131          */
3132         dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
3133         drc->drc_ds = NULL;
3134 }
3135
3136 static int
3137 add_ds_to_guidmap(const char *name, avl_tree_t *guid_map, uint64_t snapobj)
3138 {
3139         dsl_pool_t *dp;
3140         dsl_dataset_t *snapds;
3141         guid_map_entry_t *gmep;
3142         int err;
3143
3144         ASSERT(guid_map != NULL);
3145
3146         err = dsl_pool_hold(name, FTAG, &dp);
3147         if (err != 0)
3148                 return (err);
3149         gmep = kmem_alloc(sizeof (*gmep), KM_SLEEP);
3150         err = dsl_dataset_hold_obj(dp, snapobj, gmep, &snapds);
3151         if (err == 0) {
3152                 gmep->guid = dsl_dataset_phys(snapds)->ds_guid;
3153                 gmep->gme_ds = snapds;
3154                 avl_add(guid_map, gmep);
3155                 dsl_dataset_long_hold(snapds, gmep);
3156         } else
3157                 kmem_free(gmep, sizeof (*gmep));
3158
3159         dsl_pool_rele(dp, FTAG);
3160         return (err);
3161 }
3162
3163 static int dmu_recv_end_modified_blocks = 3;
3164
3165 static int
3166 dmu_recv_existing_end(dmu_recv_cookie_t *drc)
3167 {
3168         int error;
3169         char name[MAXNAMELEN];
3170
3171 #ifdef _KERNEL
3172         /*
3173          * We will be destroying the ds; make sure its origin is unmounted if
3174          * necessary.
3175          */
3176         dsl_dataset_name(drc->drc_ds, name);
3177         zfs_destroy_unmount_origin(name);
3178 #endif
3179
3180         error = dsl_sync_task(drc->drc_tofs,
3181             dmu_recv_end_check, dmu_recv_end_sync, drc,
3182             dmu_recv_end_modified_blocks, ZFS_SPACE_CHECK_NORMAL);
3183
3184         if (error != 0)
3185                 dmu_recv_cleanup_ds(drc);
3186         return (error);
3187 }
3188
3189 static int
3190 dmu_recv_new_end(dmu_recv_cookie_t *drc)
3191 {
3192         int error;
3193
3194         error = dsl_sync_task(drc->drc_tofs,
3195             dmu_recv_end_check, dmu_recv_end_sync, drc,
3196             dmu_recv_end_modified_blocks, ZFS_SPACE_CHECK_NORMAL);
3197
3198         if (error != 0) {
3199                 dmu_recv_cleanup_ds(drc);
3200         } else if (drc->drc_guid_to_ds_map != NULL) {
3201                 (void) add_ds_to_guidmap(drc->drc_tofs,
3202                     drc->drc_guid_to_ds_map,
3203                     drc->drc_newsnapobj);
3204         }
3205         return (error);
3206 }
3207
3208 int
3209 dmu_recv_end(dmu_recv_cookie_t *drc, void *owner)
3210 {
3211         drc->drc_owner = owner;
3212
3213         if (drc->drc_newfs)
3214                 return (dmu_recv_new_end(drc));
3215         else
3216                 return (dmu_recv_existing_end(drc));
3217 }
3218
3219 /*
3220  * Return TRUE if this objset is currently being received into.
3221  */
3222 boolean_t
3223 dmu_objset_is_receiving(objset_t *os)
3224 {
3225         return (os->os_dsl_dataset != NULL &&
3226             os->os_dsl_dataset->ds_owner == dmu_recv_tag);
3227 }