sys/contrib/openzfs/module/zfs/dmu_recv.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 https://opensource.org/licenses/CDDL-1.0.
  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, 2020 by Delphix. All rights reserved.
  25  * Copyright (c) 2014, Joyent, Inc. All rights reserved.
  26  * Copyright 2014 HybridCluster. All rights reserved.
  27  * Copyright (c) 2018, loli10K <ezomori.nozomu@gmail.com>. All rights reserved.
  28  * Copyright (c) 2019, Klara Inc.
  29  * Copyright (c) 2019, Allan Jude
  30  * Copyright (c) 2019 Datto Inc.
  31  * Copyright (c) 2022 Axcient.
  32  */
  33
  34 #include <sys/arc.h>
  35 #include <sys/spa_impl.h>
  36 #include <sys/dmu.h>
  37 #include <sys/dmu_impl.h>
  38 #include <sys/dmu_send.h>
  39 #include <sys/dmu_recv.h>
  40 #include <sys/dmu_tx.h>
  41 #include <sys/dbuf.h>
  42 #include <sys/dnode.h>
  43 #include <sys/zfs_context.h>
  44 #include <sys/dmu_objset.h>
  45 #include <sys/dmu_traverse.h>
  46 #include <sys/dsl_dataset.h>
  47 #include <sys/dsl_dir.h>
  48 #include <sys/dsl_prop.h>
  49 #include <sys/dsl_pool.h>
  50 #include <sys/dsl_synctask.h>
  51 #include <sys/zfs_ioctl.h>
  52 #include <sys/zap.h>
  53 #include <sys/zvol.h>
  54 #include <sys/zio_checksum.h>
  55 #include <sys/zfs_znode.h>
  56 #include <zfs_fletcher.h>
  57 #include <sys/avl.h>
  58 #include <sys/ddt.h>
  59 #include <sys/zfs_onexit.h>
  60 #include <sys/dsl_destroy.h>
  61 #include <sys/blkptr.h>
  62 #include <sys/dsl_bookmark.h>
  63 #include <sys/zfeature.h>
  64 #include <sys/bqueue.h>
  65 #include <sys/objlist.h>
  66 #ifdef _KERNEL
  67 #include <sys/zfs_vfsops.h>
  68 #endif
  69 #include <sys/zfs_file.h>
  70
  71 static uint_t zfs_recv_queue_length = SPA_MAXBLOCKSIZE;
  72 static uint_t zfs_recv_queue_ff = 20;
  73 static uint_t zfs_recv_write_batch_size = 1024 * 1024;
  74 static int zfs_recv_best_effort_corrective = 0;
  75
  76 static const void *const dmu_recv_tag = "dmu_recv_tag";
  77 const char *const recv_clone_name = "%recv";
  78
  79 typedef enum {
  80         ORNS_NO,
  81         ORNS_YES,
  82         ORNS_MAYBE
  83 } or_need_sync_t;
  84
  85 static int receive_read_payload_and_next_header(dmu_recv_cookie_t *ra, int len,
  86     void *buf);
  87
  88 struct receive_record_arg {
  89         dmu_replay_record_t header;
  90         void *payload; /* Pointer to a buffer containing the payload */
  91         /*
  92          * If the record is a WRITE or SPILL, pointer to the abd containing the
  93          * payload.
  94          */
  95         abd_t *abd;
  96         int payload_size;
  97         uint64_t bytes_read; /* bytes read from stream when record created */
  98         boolean_t eos_marker; /* Marks the end of the stream */
  99         bqueue_node_t node;
 100 };
 101
 102 struct receive_writer_arg {
 103         objset_t *os;
 104         boolean_t byteswap;
 105         bqueue_t q;
 106
 107         /*
 108          * These three members are used to signal to the main thread when
 109          * we're done.
 110          */
 111         kmutex_t mutex;
 112         kcondvar_t cv;
 113         boolean_t done;
 114
 115         int err;
 116         const char *tofs;
 117         boolean_t heal;
 118         boolean_t resumable;
 119         boolean_t raw;   /* DMU_BACKUP_FEATURE_RAW set */
 120         boolean_t spill; /* DRR_FLAG_SPILL_BLOCK set */
 121         boolean_t full;  /* this is a full send stream */
 122         uint64_t last_object;
 123         uint64_t last_offset;
 124         uint64_t max_object; /* highest object ID referenced in stream */
 125         uint64_t bytes_read; /* bytes read when current record created */
 126
 127         list_t write_batch;
 128
 129         /* Encryption parameters for the last received DRR_OBJECT_RANGE */
 130         boolean_t or_crypt_params_present;
 131         uint64_t or_firstobj;
 132         uint64_t or_numslots;
 133         uint8_t or_salt[ZIO_DATA_SALT_LEN];
 134         uint8_t or_iv[ZIO_DATA_IV_LEN];
 135         uint8_t or_mac[ZIO_DATA_MAC_LEN];
 136         boolean_t or_byteorder;
 137         zio_t *heal_pio;
 138
 139         /* Keep track of DRR_FREEOBJECTS right after DRR_OBJECT_RANGE */
 140         or_need_sync_t or_need_sync;
 141 };
 142
 143 typedef struct dmu_recv_begin_arg {
 144         const char *drba_origin;
 145         dmu_recv_cookie_t *drba_cookie;
 146         cred_t *drba_cred;
 147         proc_t *drba_proc;
 148         dsl_crypto_params_t *drba_dcp;
 149 } dmu_recv_begin_arg_t;
 150
 151 static void
 152 byteswap_record(dmu_replay_record_t *drr)
 153 {
 154 #define DO64(X) (drr->drr_u.X = BSWAP_64(drr->drr_u.X))
 155 #define DO32(X) (drr->drr_u.X = BSWAP_32(drr->drr_u.X))
 156         drr->drr_type = BSWAP_32(drr->drr_type);
 157         drr->drr_payloadlen = BSWAP_32(drr->drr_payloadlen);
 158
 159         switch (drr->drr_type) {
 160         case DRR_BEGIN:
 161                 DO64(drr_begin.drr_magic);
 162                 DO64(drr_begin.drr_versioninfo);
 163                 DO64(drr_begin.drr_creation_time);
 164                 DO32(drr_begin.drr_type);
 165                 DO32(drr_begin.drr_flags);
 166                 DO64(drr_begin.drr_toguid);
 167                 DO64(drr_begin.drr_fromguid);
 168                 break;
 169         case DRR_OBJECT:
 170                 DO64(drr_object.drr_object);
 171                 DO32(drr_object.drr_type);
 172                 DO32(drr_object.drr_bonustype);
 173                 DO32(drr_object.drr_blksz);
 174                 DO32(drr_object.drr_bonuslen);
 175                 DO32(drr_object.drr_raw_bonuslen);
 176                 DO64(drr_object.drr_toguid);
 177                 DO64(drr_object.drr_maxblkid);
 178                 break;
 179         case DRR_FREEOBJECTS:
 180                 DO64(drr_freeobjects.drr_firstobj);
 181                 DO64(drr_freeobjects.drr_numobjs);
 182                 DO64(drr_freeobjects.drr_toguid);
 183                 break;
 184         case DRR_WRITE:
 185                 DO64(drr_write.drr_object);
 186                 DO32(drr_write.drr_type);
 187                 DO64(drr_write.drr_offset);
 188                 DO64(drr_write.drr_logical_size);
 189                 DO64(drr_write.drr_toguid);
 190                 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_write.drr_key.ddk_cksum);
 191                 DO64(drr_write.drr_key.ddk_prop);
 192                 DO64(drr_write.drr_compressed_size);
 193                 break;
 194         case DRR_WRITE_EMBEDDED:
 195                 DO64(drr_write_embedded.drr_object);
 196                 DO64(drr_write_embedded.drr_offset);
 197                 DO64(drr_write_embedded.drr_length);
 198                 DO64(drr_write_embedded.drr_toguid);
 199                 DO32(drr_write_embedded.drr_lsize);
 200                 DO32(drr_write_embedded.drr_psize);
 201                 break;
 202         case DRR_FREE:
 203                 DO64(drr_free.drr_object);
 204                 DO64(drr_free.drr_offset);
 205                 DO64(drr_free.drr_length);
 206                 DO64(drr_free.drr_toguid);
 207                 break;
 208         case DRR_SPILL:
 209                 DO64(drr_spill.drr_object);
 210                 DO64(drr_spill.drr_length);
 211                 DO64(drr_spill.drr_toguid);
 212                 DO64(drr_spill.drr_compressed_size);
 213                 DO32(drr_spill.drr_type);
 214                 break;
 215         case DRR_OBJECT_RANGE:
 216                 DO64(drr_object_range.drr_firstobj);
 217                 DO64(drr_object_range.drr_numslots);
 218                 DO64(drr_object_range.drr_toguid);
 219                 break;
 220         case DRR_REDACT:
 221                 DO64(drr_redact.drr_object);
 222                 DO64(drr_redact.drr_offset);
 223                 DO64(drr_redact.drr_length);
 224                 DO64(drr_redact.drr_toguid);
 225                 break;
 226         case DRR_END:
 227                 DO64(drr_end.drr_toguid);
 228                 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_end.drr_checksum);
 229                 break;
 230         default:
 231                 break;
 232         }
 233
 234         if (drr->drr_type != DRR_BEGIN) {
 235                 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_checksum.drr_checksum);
 236         }
 237
 238 #undef DO64
 239 #undef DO32
 240 }
 241
 242 static boolean_t
 243 redact_snaps_contains(uint64_t *snaps, uint64_t num_snaps, uint64_t guid)
 244 {
 245         for (int i = 0; i < num_snaps; i++) {
 246                 if (snaps[i] == guid)
 247                         return (B_TRUE);
 248         }
 249         return (B_FALSE);
 250 }
 251
 252 /*
 253  * Check that the new stream we're trying to receive is redacted with respect to
 254  * a subset of the snapshots that the origin was redacted with respect to.  For
 255  * the reasons behind this, see the man page on redacted zfs sends and receives.
 256  */
 257 static boolean_t
 258 compatible_redact_snaps(uint64_t *origin_snaps, uint64_t origin_num_snaps,
 259     uint64_t *redact_snaps, uint64_t num_redact_snaps)
 260 {
 261         /*
 262          * Short circuit the comparison; if we are redacted with respect to
 263          * more snapshots than the origin, we can't be redacted with respect
 264          * to a subset.
 265          */
 266         if (num_redact_snaps > origin_num_snaps) {
 267                 return (B_FALSE);
 268         }
 269
 270         for (int i = 0; i < num_redact_snaps; i++) {
 271                 if (!redact_snaps_contains(origin_snaps, origin_num_snaps,
 272                     redact_snaps[i])) {
 273                         return (B_FALSE);
 274                 }
 275         }
 276         return (B_TRUE);
 277 }
 278
 279 static boolean_t
 280 redact_check(dmu_recv_begin_arg_t *drba, dsl_dataset_t *origin)
 281 {
 282         uint64_t *origin_snaps;
 283         uint64_t origin_num_snaps;
 284         dmu_recv_cookie_t *drc = drba->drba_cookie;
 285         struct drr_begin *drrb = drc->drc_drrb;
 286         int featureflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo);
 287         int err = 0;
 288         boolean_t ret = B_TRUE;
 289         uint64_t *redact_snaps;
 290         uint_t numredactsnaps;
 291
 292         /*
 293          * If this is a full send stream, we're safe no matter what.
 294          */
 295         if (drrb->drr_fromguid == 0)
 296                 return (ret);
 297
 298         VERIFY(dsl_dataset_get_uint64_array_feature(origin,
 299             SPA_FEATURE_REDACTED_DATASETS, &origin_num_snaps, &origin_snaps));
 300
 301         if (nvlist_lookup_uint64_array(drc->drc_begin_nvl,
 302             BEGINNV_REDACT_FROM_SNAPS, &redact_snaps, &numredactsnaps) ==
 303             0) {
 304                 /*
 305                  * If the send stream was sent from the redaction bookmark or
 306                  * the redacted version of the dataset, then we're safe.  Verify
 307                  * that this is from the a compatible redaction bookmark or
 308                  * redacted dataset.
 309                  */
 310                 if (!compatible_redact_snaps(origin_snaps, origin_num_snaps,
 311                     redact_snaps, numredactsnaps)) {
 312                         err = EINVAL;
 313                 }
 314         } else if (featureflags & DMU_BACKUP_FEATURE_REDACTED) {
 315                 /*
 316                  * If the stream is redacted, it must be redacted with respect
 317                  * to a subset of what the origin is redacted with respect to.
 318                  * See case number 2 in the zfs man page section on redacted zfs
 319                  * send.
 320                  */
 321                 err = nvlist_lookup_uint64_array(drc->drc_begin_nvl,
 322                     BEGINNV_REDACT_SNAPS, &redact_snaps, &numredactsnaps);
 323
 324                 if (err != 0 || !compatible_redact_snaps(origin_snaps,
 325                     origin_num_snaps, redact_snaps, numredactsnaps)) {
 326                         err = EINVAL;
 327                 }
 328         } else if (!redact_snaps_contains(origin_snaps, origin_num_snaps,
 329             drrb->drr_toguid)) {
 330                 /*
 331                  * If the stream isn't redacted but the origin is, this must be
 332                  * one of the snapshots the origin is redacted with respect to.
 333                  * See case number 1 in the zfs man page section on redacted zfs
 334                  * send.
 335                  */
 336                 err = EINVAL;
 337         }
 338
 339         if (err != 0)
 340                 ret = B_FALSE;
 341         return (ret);
 342 }
 343
 344 /*
 345  * If we previously received a stream with --large-block, we don't support
 346  * receiving an incremental on top of it without --large-block.  This avoids
 347  * forcing a read-modify-write or trying to re-aggregate a string of WRITE
 348  * records.
 349  */
 350 static int
 351 recv_check_large_blocks(dsl_dataset_t *ds, uint64_t featureflags)
 352 {
 353         if (dsl_dataset_feature_is_active(ds, SPA_FEATURE_LARGE_BLOCKS) &&
 354             !(featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS))
 355                 return (SET_ERROR(ZFS_ERR_STREAM_LARGE_BLOCK_MISMATCH));
 356         return (0);
 357 }
 358
 359 static int
 360 recv_begin_check_existing_impl(dmu_recv_begin_arg_t *drba, dsl_dataset_t *ds,
 361     uint64_t fromguid, uint64_t featureflags)
 362 {
 363         uint64_t obj;
 364         uint64_t children;
 365         int error;
 366         dsl_dataset_t *snap;
 367         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 368         boolean_t encrypted = ds->ds_dir->dd_crypto_obj != 0;
 369         boolean_t raw = (featureflags & DMU_BACKUP_FEATURE_RAW) != 0;
 370         boolean_t embed = (featureflags & DMU_BACKUP_FEATURE_EMBED_DATA) != 0;
 371
 372         /* Temporary clone name must not exist. */
 373         error = zap_lookup(dp->dp_meta_objset,
 374             dsl_dir_phys(ds->ds_dir)->dd_child_dir_zapobj, recv_clone_name,
 375             8, 1, &obj);
 376         if (error != ENOENT)
 377                 return (error == 0 ? SET_ERROR(EBUSY) : error);
 378
 379         /* Resume state must not be set. */
 380         if (dsl_dataset_has_resume_receive_state(ds))
 381                 return (SET_ERROR(EBUSY));
 382
 383         /* New snapshot name must not exist if we're not healing it. */
 384         error = zap_lookup(dp->dp_meta_objset,
 385             dsl_dataset_phys(ds)->ds_snapnames_zapobj,
 386             drba->drba_cookie->drc_tosnap, 8, 1, &obj);
 387         if (drba->drba_cookie->drc_heal) {
 388                 if (error != 0)
 389                         return (error);
 390         } else if (error != ENOENT) {
 391                 return (error == 0 ? SET_ERROR(EEXIST) : error);
 392         }
 393
 394         /* Must not have children if receiving a ZVOL. */
 395         error = zap_count(dp->dp_meta_objset,
 396             dsl_dir_phys(ds->ds_dir)->dd_child_dir_zapobj, &children);
 397         if (error != 0)
 398                 return (error);
 399         if (drba->drba_cookie->drc_drrb->drr_type != DMU_OST_ZFS &&
 400             children > 0)
 401                 return (SET_ERROR(ZFS_ERR_WRONG_PARENT));
 402
 403         /*
 404          * Check snapshot limit before receiving. We'll recheck again at the
 405          * end, but might as well abort before receiving if we're already over
 406          * the limit.
 407          *
 408          * Note that we do not check the file system limit with
 409          * dsl_dir_fscount_check because the temporary %clones don't count
 410          * against that limit.
 411          */
 412         error = dsl_fs_ss_limit_check(ds->ds_dir, 1, ZFS_PROP_SNAPSHOT_LIMIT,
 413             NULL, drba->drba_cred, drba->drba_proc);
 414         if (error != 0)
 415                 return (error);
 416
 417         if (drba->drba_cookie->drc_heal) {
 418                 /* Encryption is incompatible with embedded data. */
 419                 if (encrypted && embed)
 420                         return (SET_ERROR(EINVAL));
 421
 422                 /* Healing is not supported when in 'force' mode. */
 423                 if (drba->drba_cookie->drc_force)
 424                         return (SET_ERROR(EINVAL));
 425
 426                 /* Must have keys loaded if doing encrypted non-raw recv. */
 427                 if (encrypted && !raw) {
 428                         if (spa_keystore_lookup_key(dp->dp_spa, ds->ds_object,
 429                             NULL, NULL) != 0)
 430                                 return (SET_ERROR(EACCES));
 431                 }
 432
 433                 error = dsl_dataset_hold_obj(dp, obj, FTAG, &snap);
 434                 if (error != 0)
 435                         return (error);
 436
 437                 /*
 438                  * When not doing best effort corrective recv healing can only
 439                  * be done if the send stream is for the same snapshot as the
 440                  * one we are trying to heal.
 441                  */
 442                 if (zfs_recv_best_effort_corrective == 0 &&
 443                     drba->drba_cookie->drc_drrb->drr_toguid !=
 444                     dsl_dataset_phys(snap)->ds_guid) {
 445                         dsl_dataset_rele(snap, FTAG);
 446                         return (SET_ERROR(ENOTSUP));
 447                 }
 448                 dsl_dataset_rele(snap, FTAG);
 449         } else if (fromguid != 0) {
 450                 /* Sanity check the incremental recv */
 451                 uint64_t obj = dsl_dataset_phys(ds)->ds_prev_snap_obj;
 452
 453                 /* Can't perform a raw receive on top of a non-raw receive */
 454                 if (!encrypted && raw)
 455                         return (SET_ERROR(EINVAL));
 456
 457                 /* Encryption is incompatible with embedded data */
 458                 if (encrypted && embed)
 459                         return (SET_ERROR(EINVAL));
 460
 461                 /* Find snapshot in this dir that matches fromguid. */
 462                 while (obj != 0) {
 463                         error = dsl_dataset_hold_obj(dp, obj, FTAG,
 464                             &snap);
 465                         if (error != 0)
 466                                 return (SET_ERROR(ENODEV));
 467                         if (snap->ds_dir != ds->ds_dir) {
 468                                 dsl_dataset_rele(snap, FTAG);
 469                                 return (SET_ERROR(ENODEV));
 470                         }
 471                         if (dsl_dataset_phys(snap)->ds_guid == fromguid)
 472                                 break;
 473                         obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
 474                         dsl_dataset_rele(snap, FTAG);
 475                 }
 476                 if (obj == 0)
 477                         return (SET_ERROR(ENODEV));
 478
 479                 if (drba->drba_cookie->drc_force) {
 480                         drba->drba_cookie->drc_fromsnapobj = obj;
 481                 } else {
 482                         /*
 483                          * If we are not forcing, there must be no
 484                          * changes since fromsnap. Raw sends have an
 485                          * additional constraint that requires that
 486                          * no "noop" snapshots exist between fromsnap
 487                          * and tosnap for the IVset checking code to
 488                          * work properly.
 489                          */
 490                         if (dsl_dataset_modified_since_snap(ds, snap) ||
 491                             (raw &&
 492                             dsl_dataset_phys(ds)->ds_prev_snap_obj !=
 493                             snap->ds_object)) {
 494                                 dsl_dataset_rele(snap, FTAG);
 495                                 return (SET_ERROR(ETXTBSY));
 496                         }
 497                         drba->drba_cookie->drc_fromsnapobj =
 498                             ds->ds_prev->ds_object;
 499                 }
 500
 501                 if (dsl_dataset_feature_is_active(snap,
 502                     SPA_FEATURE_REDACTED_DATASETS) && !redact_check(drba,
 503                     snap)) {
 504                         dsl_dataset_rele(snap, FTAG);
 505                         return (SET_ERROR(EINVAL));
 506                 }
 507
 508                 error = recv_check_large_blocks(snap, featureflags);
 509                 if (error != 0) {
 510                         dsl_dataset_rele(snap, FTAG);
 511                         return (error);
 512                 }
 513
 514                 dsl_dataset_rele(snap, FTAG);
 515         } else {
 516                 /* If full and not healing then must be forced. */
 517                 if (!drba->drba_cookie->drc_force)
 518                         return (SET_ERROR(EEXIST));
 519
 520                 /*
 521                  * We don't support using zfs recv -F to blow away
 522                  * encrypted filesystems. This would require the
 523                  * dsl dir to point to the old encryption key and
 524                  * the new one at the same time during the receive.
 525                  */
 526                 if ((!encrypted && raw) || encrypted)
 527                         return (SET_ERROR(EINVAL));
 528
 529                 /*
 530                  * Perform the same encryption checks we would if
 531                  * we were creating a new dataset from scratch.
 532                  */
 533                 if (!raw) {
 534                         boolean_t will_encrypt;
 535
 536                         error = dmu_objset_create_crypt_check(
 537                             ds->ds_dir->dd_parent, drba->drba_dcp,
 538                             &will_encrypt);
 539                         if (error != 0)
 540                                 return (error);
 541
 542                         if (will_encrypt && embed)
 543                                 return (SET_ERROR(EINVAL));
 544                 }
 545         }
 546
 547         return (0);
 548 }
 549
 550 /*
 551  * Check that any feature flags used in the data stream we're receiving are
 552  * supported by the pool we are receiving into.
 553  *
 554  * Note that some of the features we explicitly check here have additional
 555  * (implicit) features they depend on, but those dependencies are enforced
 556  * through the zfeature_register() calls declaring the features that we
 557  * explicitly check.
 558  */
 559 static int
 560 recv_begin_check_feature_flags_impl(uint64_t featureflags, spa_t *spa)
 561 {
 562         /*
 563          * Check if there are any unsupported feature flags.
 564          */
 565         if (!DMU_STREAM_SUPPORTED(featureflags)) {
 566                 return (SET_ERROR(ZFS_ERR_UNKNOWN_SEND_STREAM_FEATURE));
 567         }
 568
 569         /* Verify pool version supports SA if SA_SPILL feature set */
 570         if ((featureflags & DMU_BACKUP_FEATURE_SA_SPILL) &&
 571             spa_version(spa) < SPA_VERSION_SA)
 572                 return (SET_ERROR(ENOTSUP));
 573
 574         /*
 575          * LZ4 compressed, ZSTD compressed, embedded, mooched, large blocks,
 576          * and large_dnodes in the stream can only be used if those pool
 577          * features are enabled because we don't attempt to decompress /
 578          * un-embed / un-mooch / split up the blocks / dnodes during the
 579          * receive process.
 580          */
 581         if ((featureflags & DMU_BACKUP_FEATURE_LZ4) &&
 582             !spa_feature_is_enabled(spa, SPA_FEATURE_LZ4_COMPRESS))
 583                 return (SET_ERROR(ENOTSUP));
 584         if ((featureflags & DMU_BACKUP_FEATURE_ZSTD) &&
 585             !spa_feature_is_enabled(spa, SPA_FEATURE_ZSTD_COMPRESS))
 586                 return (SET_ERROR(ENOTSUP));
 587         if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA) &&
 588             !spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA))
 589                 return (SET_ERROR(ENOTSUP));
 590         if ((featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
 591             !spa_feature_is_enabled(spa, SPA_FEATURE_LARGE_BLOCKS))
 592                 return (SET_ERROR(ENOTSUP));
 593         if ((featureflags & DMU_BACKUP_FEATURE_LARGE_DNODE) &&
 594             !spa_feature_is_enabled(spa, SPA_FEATURE_LARGE_DNODE))
 595                 return (SET_ERROR(ENOTSUP));
 596
 597         /*
 598          * Receiving redacted streams requires that redacted datasets are
 599          * enabled.
 600          */
 601         if ((featureflags & DMU_BACKUP_FEATURE_REDACTED) &&
 602             !spa_feature_is_enabled(spa, SPA_FEATURE_REDACTED_DATASETS))
 603                 return (SET_ERROR(ENOTSUP));
 604
 605         return (0);
 606 }
 607
 608 static int
 609 dmu_recv_begin_check(void *arg, dmu_tx_t *tx)
 610 {
 611         dmu_recv_begin_arg_t *drba = arg;
 612         dsl_pool_t *dp = dmu_tx_pool(tx);
 613         struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
 614         uint64_t fromguid = drrb->drr_fromguid;
 615         int flags = drrb->drr_flags;
 616         ds_hold_flags_t dsflags = DS_HOLD_FLAG_NONE;
 617         int error;
 618         uint64_t featureflags = drba->drba_cookie->drc_featureflags;
 619         dsl_dataset_t *ds;
 620         const char *tofs = drba->drba_cookie->drc_tofs;
 621
 622         /* already checked */
 623         ASSERT3U(drrb->drr_magic, ==, DMU_BACKUP_MAGIC);
 624         ASSERT(!(featureflags & DMU_BACKUP_FEATURE_RESUMING));
 625
 626         if (DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) ==
 627             DMU_COMPOUNDSTREAM ||
 628             drrb->drr_type >= DMU_OST_NUMTYPES ||
 629             ((flags & DRR_FLAG_CLONE) && drba->drba_origin == NULL))
 630                 return (SET_ERROR(EINVAL));
 631
 632         error = recv_begin_check_feature_flags_impl(featureflags, dp->dp_spa);
 633         if (error != 0)
 634                 return (error);
 635
 636         /* Resumable receives require extensible datasets */
 637         if (drba->drba_cookie->drc_resumable &&
 638             !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EXTENSIBLE_DATASET))
 639                 return (SET_ERROR(ENOTSUP));
 640
 641         if (featureflags & DMU_BACKUP_FEATURE_RAW) {
 642                 /* raw receives require the encryption feature */
 643                 if (!spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_ENCRYPTION))
 644                         return (SET_ERROR(ENOTSUP));
 645
 646                 /* embedded data is incompatible with encryption and raw recv */
 647                 if (featureflags & DMU_BACKUP_FEATURE_EMBED_DATA)
 648                         return (SET_ERROR(EINVAL));
 649
 650                 /* raw receives require spill block allocation flag */
 651                 if (!(flags & DRR_FLAG_SPILL_BLOCK))
 652                         return (SET_ERROR(ZFS_ERR_SPILL_BLOCK_FLAG_MISSING));
 653         } else {
 654                 /*
 655                  * We support unencrypted datasets below encrypted ones now,
 656                  * so add the DS_HOLD_FLAG_DECRYPT flag only if we are dealing
 657                  * with a dataset we may encrypt.
 658                  */
 659                 if (drba->drba_dcp == NULL ||
 660                     drba->drba_dcp->cp_crypt != ZIO_CRYPT_OFF) {
 661                         dsflags |= DS_HOLD_FLAG_DECRYPT;
 662                 }
 663         }
 664
 665         error = dsl_dataset_hold_flags(dp, tofs, dsflags, FTAG, &ds);
 666         if (error == 0) {
 667                 /* target fs already exists; recv into temp clone */
 668
 669                 /* Can't recv a clone into an existing fs */
 670                 if (flags & DRR_FLAG_CLONE || drba->drba_origin) {
 671                         dsl_dataset_rele_flags(ds, dsflags, FTAG);
 672                         return (SET_ERROR(EINVAL));
 673                 }
 674
 675                 error = recv_begin_check_existing_impl(drba, ds, fromguid,
 676                     featureflags);
 677                 dsl_dataset_rele_flags(ds, dsflags, FTAG);
 678         } else if (error == ENOENT) {
 679                 /* target fs does not exist; must be a full backup or clone */
 680                 char buf[ZFS_MAX_DATASET_NAME_LEN];
 681                 objset_t *os;
 682
 683                 /* healing recv must be done "into" an existing snapshot */
 684                 if (drba->drba_cookie->drc_heal == B_TRUE)
 685                         return (SET_ERROR(ENOTSUP));
 686
 687                 /*
 688                  * If it's a non-clone incremental, we are missing the
 689                  * target fs, so fail the recv.
 690                  */
 691                 if (fromguid != 0 && !((flags & DRR_FLAG_CLONE) ||
 692                     drba->drba_origin))
 693                         return (SET_ERROR(ENOENT));
 694
 695                 /*
 696                  * If we're receiving a full send as a clone, and it doesn't
 697                  * contain all the necessary free records and freeobject
 698                  * records, reject it.
 699                  */
 700                 if (fromguid == 0 && drba->drba_origin != NULL &&
 701                     !(flags & DRR_FLAG_FREERECORDS))
 702                         return (SET_ERROR(EINVAL));
 703
 704                 /* Open the parent of tofs */
 705                 ASSERT3U(strlen(tofs), <, sizeof (buf));
 706                 (void) strlcpy(buf, tofs, strrchr(tofs, '/') - tofs + 1);
 707                 error = dsl_dataset_hold(dp, buf, FTAG, &ds);
 708                 if (error != 0)
 709                         return (error);
 710
 711                 if ((featureflags & DMU_BACKUP_FEATURE_RAW) == 0 &&
 712                     drba->drba_origin == NULL) {
 713                         boolean_t will_encrypt;
 714
 715                         /*
 716                          * Check that we aren't breaking any encryption rules
 717                          * and that we have all the parameters we need to
 718                          * create an encrypted dataset if necessary. If we are
 719                          * making an encrypted dataset the stream can't have
 720                          * embedded data.
 721                          */
 722                         error = dmu_objset_create_crypt_check(ds->ds_dir,
 723                             drba->drba_dcp, &will_encrypt);
 724                         if (error != 0) {
 725                                 dsl_dataset_rele(ds, FTAG);
 726                                 return (error);
 727                         }
 728
 729                         if (will_encrypt &&
 730                             (featureflags & DMU_BACKUP_FEATURE_EMBED_DATA)) {
 731                                 dsl_dataset_rele(ds, FTAG);
 732                                 return (SET_ERROR(EINVAL));
 733                         }
 734                 }
 735
 736                 /*
 737                  * Check filesystem and snapshot limits before receiving. We'll
 738                  * recheck snapshot limits again at the end (we create the
 739                  * filesystems and increment those counts during begin_sync).
 740                  */
 741                 error = dsl_fs_ss_limit_check(ds->ds_dir, 1,
 742                     ZFS_PROP_FILESYSTEM_LIMIT, NULL,
 743                     drba->drba_cred, drba->drba_proc);
 744                 if (error != 0) {
 745                         dsl_dataset_rele(ds, FTAG);
 746                         return (error);
 747                 }
 748
 749                 error = dsl_fs_ss_limit_check(ds->ds_dir, 1,
 750                     ZFS_PROP_SNAPSHOT_LIMIT, NULL,
 751                     drba->drba_cred, drba->drba_proc);
 752                 if (error != 0) {
 753                         dsl_dataset_rele(ds, FTAG);
 754                         return (error);
 755                 }
 756
 757                 /* can't recv below anything but filesystems (eg. no ZVOLs) */
 758                 error = dmu_objset_from_ds(ds, &os);
 759                 if (error != 0) {
 760                         dsl_dataset_rele(ds, FTAG);
 761                         return (error);
 762                 }
 763                 if (dmu_objset_type(os) != DMU_OST_ZFS) {
 764                         dsl_dataset_rele(ds, FTAG);
 765                         return (SET_ERROR(ZFS_ERR_WRONG_PARENT));
 766                 }
 767
 768                 if (drba->drba_origin != NULL) {
 769                         dsl_dataset_t *origin;
 770                         error = dsl_dataset_hold_flags(dp, drba->drba_origin,
 771                             dsflags, FTAG, &origin);
 772                         if (error != 0) {
 773                                 dsl_dataset_rele(ds, FTAG);
 774                                 return (error);
 775                         }
 776                         if (!origin->ds_is_snapshot) {
 777                                 dsl_dataset_rele_flags(origin, dsflags, FTAG);
 778                                 dsl_dataset_rele(ds, FTAG);
 779                                 return (SET_ERROR(EINVAL));
 780                         }
 781                         if (dsl_dataset_phys(origin)->ds_guid != fromguid &&
 782                             fromguid != 0) {
 783                                 dsl_dataset_rele_flags(origin, dsflags, FTAG);
 784                                 dsl_dataset_rele(ds, FTAG);
 785                                 return (SET_ERROR(ENODEV));
 786                         }
 787
 788                         if (origin->ds_dir->dd_crypto_obj != 0 &&
 789                             (featureflags & DMU_BACKUP_FEATURE_EMBED_DATA)) {
 790                                 dsl_dataset_rele_flags(origin, dsflags, FTAG);
 791                                 dsl_dataset_rele(ds, FTAG);
 792                                 return (SET_ERROR(EINVAL));
 793                         }
 794
 795                         /*
 796                          * If the origin is redacted we need to verify that this
 797                          * send stream can safely be received on top of the
 798                          * origin.
 799                          */
 800                         if (dsl_dataset_feature_is_active(origin,
 801                             SPA_FEATURE_REDACTED_DATASETS)) {
 802                                 if (!redact_check(drba, origin)) {
 803                                         dsl_dataset_rele_flags(origin, dsflags,
 804                                             FTAG);
 805                                         dsl_dataset_rele_flags(ds, dsflags,
 806                                             FTAG);
 807                                         return (SET_ERROR(EINVAL));
 808                                 }
 809                         }
 810
 811                         error = recv_check_large_blocks(ds, featureflags);
 812                         if (error != 0) {
 813                                 dsl_dataset_rele_flags(origin, dsflags, FTAG);
 814                                 dsl_dataset_rele_flags(ds, dsflags, FTAG);
 815                                 return (error);
 816                         }
 817
 818                         dsl_dataset_rele_flags(origin, dsflags, FTAG);
 819                 }
 820
 821                 dsl_dataset_rele(ds, FTAG);
 822                 error = 0;
 823         }
 824         return (error);
 825 }
 826
 827 static void
 828 dmu_recv_begin_sync(void *arg, dmu_tx_t *tx)
 829 {
 830         dmu_recv_begin_arg_t *drba = arg;
 831         dsl_pool_t *dp = dmu_tx_pool(tx);
 832         objset_t *mos = dp->dp_meta_objset;
 833         dmu_recv_cookie_t *drc = drba->drba_cookie;
 834         struct drr_begin *drrb = drc->drc_drrb;
 835         const char *tofs = drc->drc_tofs;
 836         uint64_t featureflags = drc->drc_featureflags;
 837         dsl_dataset_t *ds, *newds;
 838         objset_t *os;
 839         uint64_t dsobj;
 840         ds_hold_flags_t dsflags = DS_HOLD_FLAG_NONE;
 841         int error;
 842         uint64_t crflags = 0;
 843         dsl_crypto_params_t dummy_dcp = { 0 };
 844         dsl_crypto_params_t *dcp = drba->drba_dcp;
 845
 846         if (drrb->drr_flags & DRR_FLAG_CI_DATA)
 847                 crflags |= DS_FLAG_CI_DATASET;
 848
 849         if ((featureflags & DMU_BACKUP_FEATURE_RAW) == 0)
 850                 dsflags |= DS_HOLD_FLAG_DECRYPT;
 851
 852         /*
 853          * Raw, non-incremental recvs always use a dummy dcp with
 854          * the raw cmd set. Raw incremental recvs do not use a dcp
 855          * since the encryption parameters are already set in stone.
 856          */
 857         if (dcp == NULL && drrb->drr_fromguid == 0 &&
 858             drba->drba_origin == NULL) {
 859                 ASSERT3P(dcp, ==, NULL);
 860                 dcp = &dummy_dcp;
 861
 862                 if (featureflags & DMU_BACKUP_FEATURE_RAW)
 863                         dcp->cp_cmd = DCP_CMD_RAW_RECV;
 864         }
 865
 866         error = dsl_dataset_hold_flags(dp, tofs, dsflags, FTAG, &ds);
 867         if (error == 0) {
 868                 /* Create temporary clone unless we're doing corrective recv */
 869                 dsl_dataset_t *snap = NULL;
 870
 871                 if (drba->drba_cookie->drc_fromsnapobj != 0) {
 872                         VERIFY0(dsl_dataset_hold_obj(dp,
 873                             drba->drba_cookie->drc_fromsnapobj, FTAG, &snap));
 874                         ASSERT3P(dcp, ==, NULL);
 875                 }
 876                 if (drc->drc_heal) {
 877                         /* When healing we want to use the provided snapshot */
 878                         VERIFY0(dsl_dataset_snap_lookup(ds, drc->drc_tosnap,
 879                             &dsobj));
 880                 } else {
 881                         dsobj = dsl_dataset_create_sync(ds->ds_dir,
 882                             recv_clone_name, snap, crflags, drba->drba_cred,
 883                             dcp, tx);
 884                 }
 885                 if (drba->drba_cookie->drc_fromsnapobj != 0)
 886                         dsl_dataset_rele(snap, FTAG);
 887                 dsl_dataset_rele_flags(ds, dsflags, FTAG);
 888         } else {
 889                 dsl_dir_t *dd;
 890                 const char *tail;
 891                 dsl_dataset_t *origin = NULL;
 892
 893                 VERIFY0(dsl_dir_hold(dp, tofs, FTAG, &dd, &tail));
 894
 895                 if (drba->drba_origin != NULL) {
 896                         VERIFY0(dsl_dataset_hold(dp, drba->drba_origin,
 897                             FTAG, &origin));
 898                         ASSERT3P(dcp, ==, NULL);
 899                 }
 900
 901                 /* Create new dataset. */
 902                 dsobj = dsl_dataset_create_sync(dd, strrchr(tofs, '/') + 1,
 903                     origin, crflags, drba->drba_cred, dcp, tx);
 904                 if (origin != NULL)
 905                         dsl_dataset_rele(origin, FTAG);
 906                 dsl_dir_rele(dd, FTAG);
 907                 drc->drc_newfs = B_TRUE;
 908         }
 909         VERIFY0(dsl_dataset_own_obj_force(dp, dsobj, dsflags, dmu_recv_tag,
 910             &newds));
 911         if (dsl_dataset_feature_is_active(newds,
 912             SPA_FEATURE_REDACTED_DATASETS)) {
 913                 /*
 914                  * If the origin dataset is redacted, the child will be redacted
 915                  * when we create it.  We clear the new dataset's
 916                  * redaction info; if it should be redacted, we'll fill
 917                  * in its information later.
 918                  */
 919                 dsl_dataset_deactivate_feature(newds,
 920                     SPA_FEATURE_REDACTED_DATASETS, tx);
 921         }
 922         VERIFY0(dmu_objset_from_ds(newds, &os));
 923
 924         if (drc->drc_resumable) {
 925                 dsl_dataset_zapify(newds, tx);
 926                 if (drrb->drr_fromguid != 0) {
 927                         VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_FROMGUID,
 928                             8, 1, &drrb->drr_fromguid, tx));
 929                 }
 930                 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_TOGUID,
 931                     8, 1, &drrb->drr_toguid, tx));
 932                 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_TONAME,
 933                     1, strlen(drrb->drr_toname) + 1, drrb->drr_toname, tx));
 934                 uint64_t one = 1;
 935                 uint64_t zero = 0;
 936                 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_OBJECT,
 937                     8, 1, &one, tx));
 938                 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_OFFSET,
 939                     8, 1, &zero, tx));
 940                 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_BYTES,
 941                     8, 1, &zero, tx));
 942                 if (featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) {
 943                         VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_LARGEBLOCK,
 944                             8, 1, &one, tx));
 945                 }
 946                 if (featureflags & DMU_BACKUP_FEATURE_EMBED_DATA) {
 947                         VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_EMBEDOK,
 948                             8, 1, &one, tx));
 949                 }
 950                 if (featureflags & DMU_BACKUP_FEATURE_COMPRESSED) {
 951                         VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_COMPRESSOK,
 952                             8, 1, &one, tx));
 953                 }
 954                 if (featureflags & DMU_BACKUP_FEATURE_RAW) {
 955                         VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_RAWOK,
 956                             8, 1, &one, tx));
 957                 }
 958
 959                 uint64_t *redact_snaps;
 960                 uint_t numredactsnaps;
 961                 if (nvlist_lookup_uint64_array(drc->drc_begin_nvl,
 962                     BEGINNV_REDACT_FROM_SNAPS, &redact_snaps,
 963                     &numredactsnaps) == 0) {
 964                         VERIFY0(zap_add(mos, dsobj,
 965                             DS_FIELD_RESUME_REDACT_BOOKMARK_SNAPS,
 966                             sizeof (*redact_snaps), numredactsnaps,
 967                             redact_snaps, tx));
 968                 }
 969         }
 970
 971         /*
 972          * Usually the os->os_encrypted value is tied to the presence of a
 973          * DSL Crypto Key object in the dd. However, that will not be received
 974          * until dmu_recv_stream(), so we set the value manually for now.
 975          */
 976         if (featureflags & DMU_BACKUP_FEATURE_RAW) {
 977                 os->os_encrypted = B_TRUE;
 978                 drba->drba_cookie->drc_raw = B_TRUE;
 979         }
 980
 981         if (featureflags & DMU_BACKUP_FEATURE_REDACTED) {
 982                 uint64_t *redact_snaps;
 983                 uint_t numredactsnaps;
 984                 VERIFY0(nvlist_lookup_uint64_array(drc->drc_begin_nvl,
 985                     BEGINNV_REDACT_SNAPS, &redact_snaps, &numredactsnaps));
 986                 dsl_dataset_activate_redaction(newds, redact_snaps,
 987                     numredactsnaps, tx);
 988         }
 989
 990         dmu_buf_will_dirty(newds->ds_dbuf, tx);
 991         dsl_dataset_phys(newds)->ds_flags |= DS_FLAG_INCONSISTENT;
 992
 993         /*
 994          * If we actually created a non-clone, we need to create the objset
 995          * in our new dataset. If this is a raw send we postpone this until
 996          * dmu_recv_stream() so that we can allocate the metadnode with the
 997          * properties from the DRR_BEGIN payload.
 998          */
 999         rrw_enter(&newds->ds_bp_rwlock, RW_READER, FTAG);
1000         if (BP_IS_HOLE(dsl_dataset_get_blkptr(newds)) &&
1001             (featureflags & DMU_BACKUP_FEATURE_RAW) == 0 &&
1002             !drc->drc_heal) {
1003                 (void) dmu_objset_create_impl(dp->dp_spa,
1004                     newds, dsl_dataset_get_blkptr(newds), drrb->drr_type, tx);
1005         }
1006         rrw_exit(&newds->ds_bp_rwlock, FTAG);
1007
1008         drba->drba_cookie->drc_ds = newds;
1009         drba->drba_cookie->drc_os = os;
1010
1011         spa_history_log_internal_ds(newds, "receive", tx, " ");
1012 }
1013
1014 static int
1015 dmu_recv_resume_begin_check(void *arg, dmu_tx_t *tx)
1016 {
1017         dmu_recv_begin_arg_t *drba = arg;
1018         dmu_recv_cookie_t *drc = drba->drba_cookie;
1019         dsl_pool_t *dp = dmu_tx_pool(tx);
1020         struct drr_begin *drrb = drc->drc_drrb;
1021         int error;
1022         ds_hold_flags_t dsflags = DS_HOLD_FLAG_NONE;
1023         dsl_dataset_t *ds;
1024         const char *tofs = drc->drc_tofs;
1025
1026         /* already checked */
1027         ASSERT3U(drrb->drr_magic, ==, DMU_BACKUP_MAGIC);
1028         ASSERT(drc->drc_featureflags & DMU_BACKUP_FEATURE_RESUMING);
1029
1030         if (DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) ==
1031             DMU_COMPOUNDSTREAM ||
1032             drrb->drr_type >= DMU_OST_NUMTYPES)
1033                 return (SET_ERROR(EINVAL));
1034
1035         /*
1036          * This is mostly a sanity check since we should have already done these
1037          * checks during a previous attempt to receive the data.
1038          */
1039         error = recv_begin_check_feature_flags_impl(drc->drc_featureflags,
1040             dp->dp_spa);
1041         if (error != 0)
1042                 return (error);
1043
1044         /* 6 extra bytes for /%recv */
1045         char recvname[ZFS_MAX_DATASET_NAME_LEN + 6];
1046
1047         (void) snprintf(recvname, sizeof (recvname), "%s/%s",
1048             tofs, recv_clone_name);
1049
1050         if (drc->drc_featureflags & DMU_BACKUP_FEATURE_RAW) {
1051                 /* raw receives require spill block allocation flag */
1052                 if (!(drrb->drr_flags & DRR_FLAG_SPILL_BLOCK))
1053                         return (SET_ERROR(ZFS_ERR_SPILL_BLOCK_FLAG_MISSING));
1054         } else {
1055                 dsflags |= DS_HOLD_FLAG_DECRYPT;
1056         }
1057
1058         boolean_t recvexist = B_TRUE;
1059         if (dsl_dataset_hold_flags(dp, recvname, dsflags, FTAG, &ds) != 0) {
1060                 /* %recv does not exist; continue in tofs */
1061                 recvexist = B_FALSE;
1062                 error = dsl_dataset_hold_flags(dp, tofs, dsflags, FTAG, &ds);
1063                 if (error != 0)
1064                         return (error);
1065         }
1066
1067         /*
1068          * Resume of full/newfs recv on existing dataset should be done with
1069          * force flag
1070          */
1071         if (recvexist && drrb->drr_fromguid == 0 && !drc->drc_force) {
1072                 dsl_dataset_rele_flags(ds, dsflags, FTAG);
1073                 return (SET_ERROR(ZFS_ERR_RESUME_EXISTS));
1074         }
1075
1076         /* check that ds is marked inconsistent */
1077         if (!DS_IS_INCONSISTENT(ds)) {
1078                 dsl_dataset_rele_flags(ds, dsflags, FTAG);
1079                 return (SET_ERROR(EINVAL));
1080         }
1081
1082         /* check that there is resuming data, and that the toguid matches */
1083         if (!dsl_dataset_is_zapified(ds)) {
1084                 dsl_dataset_rele_flags(ds, dsflags, FTAG);
1085                 return (SET_ERROR(EINVAL));
1086         }
1087         uint64_t val;
1088         error = zap_lookup(dp->dp_meta_objset, ds->ds_object,
1089             DS_FIELD_RESUME_TOGUID, sizeof (val), 1, &val);
1090         if (error != 0 || drrb->drr_toguid != val) {
1091                 dsl_dataset_rele_flags(ds, dsflags, FTAG);
1092                 return (SET_ERROR(EINVAL));
1093         }
1094
1095         /*
1096          * Check if the receive is still running.  If so, it will be owned.
1097          * Note that nothing else can own the dataset (e.g. after the receive
1098          * fails) because it will be marked inconsistent.
1099          */
1100         if (dsl_dataset_has_owner(ds)) {
1101                 dsl_dataset_rele_flags(ds, dsflags, FTAG);
1102                 return (SET_ERROR(EBUSY));
1103         }
1104
1105         /* There should not be any snapshots of this fs yet. */
1106         if (ds->ds_prev != NULL && ds->ds_prev->ds_dir == ds->ds_dir) {
1107                 dsl_dataset_rele_flags(ds, dsflags, FTAG);
1108                 return (SET_ERROR(EINVAL));
1109         }
1110
1111         /*
1112          * Note: resume point will be checked when we process the first WRITE
1113          * record.
1114          */
1115
1116         /* check that the origin matches */
1117         val = 0;
1118         (void) zap_lookup(dp->dp_meta_objset, ds->ds_object,
1119             DS_FIELD_RESUME_FROMGUID, sizeof (val), 1, &val);
1120         if (drrb->drr_fromguid != val) {
1121                 dsl_dataset_rele_flags(ds, dsflags, FTAG);
1122                 return (SET_ERROR(EINVAL));
1123         }
1124
1125         if (ds->ds_prev != NULL && drrb->drr_fromguid != 0)
1126                 drc->drc_fromsnapobj = ds->ds_prev->ds_object;
1127
1128         /*
1129          * If we're resuming, and the send is redacted, then the original send
1130          * must have been redacted, and must have been redacted with respect to
1131          * the same snapshots.
1132          */
1133         if (drc->drc_featureflags & DMU_BACKUP_FEATURE_REDACTED) {
1134                 uint64_t num_ds_redact_snaps;
1135                 uint64_t *ds_redact_snaps;
1136
1137                 uint_t num_stream_redact_snaps;
1138                 uint64_t *stream_redact_snaps;
1139
1140                 if (nvlist_lookup_uint64_array(drc->drc_begin_nvl,
1141                     BEGINNV_REDACT_SNAPS, &stream_redact_snaps,
1142                     &num_stream_redact_snaps) != 0) {
1143                         dsl_dataset_rele_flags(ds, dsflags, FTAG);
1144                         return (SET_ERROR(EINVAL));
1145                 }
1146
1147                 if (!dsl_dataset_get_uint64_array_feature(ds,
1148                     SPA_FEATURE_REDACTED_DATASETS, &num_ds_redact_snaps,
1149                     &ds_redact_snaps)) {
1150                         dsl_dataset_rele_flags(ds, dsflags, FTAG);
1151                         return (SET_ERROR(EINVAL));
1152                 }
1153
1154                 for (int i = 0; i < num_ds_redact_snaps; i++) {
1155                         if (!redact_snaps_contains(ds_redact_snaps,
1156                             num_ds_redact_snaps, stream_redact_snaps[i])) {
1157                                 dsl_dataset_rele_flags(ds, dsflags, FTAG);
1158                                 return (SET_ERROR(EINVAL));
1159                         }
1160                 }
1161         }
1162
1163         error = recv_check_large_blocks(ds, drc->drc_featureflags);
1164         if (error != 0) {
1165                 dsl_dataset_rele_flags(ds, dsflags, FTAG);
1166                 return (error);
1167         }
1168
1169         dsl_dataset_rele_flags(ds, dsflags, FTAG);
1170         return (0);
1171 }
1172
1173 static void
1174 dmu_recv_resume_begin_sync(void *arg, dmu_tx_t *tx)
1175 {
1176         dmu_recv_begin_arg_t *drba = arg;
1177         dsl_pool_t *dp = dmu_tx_pool(tx);
1178         const char *tofs = drba->drba_cookie->drc_tofs;
1179         uint64_t featureflags = drba->drba_cookie->drc_featureflags;
1180         dsl_dataset_t *ds;
1181         ds_hold_flags_t dsflags = DS_HOLD_FLAG_NONE;
1182         /* 6 extra bytes for /%recv */
1183         char recvname[ZFS_MAX_DATASET_NAME_LEN + 6];
1184
1185         (void) snprintf(recvname, sizeof (recvname), "%s/%s", tofs,
1186             recv_clone_name);
1187
1188         if (featureflags & DMU_BACKUP_FEATURE_RAW) {
1189                 drba->drba_cookie->drc_raw = B_TRUE;
1190         } else {
1191                 dsflags |= DS_HOLD_FLAG_DECRYPT;
1192         }
1193
1194         if (dsl_dataset_own_force(dp, recvname, dsflags, dmu_recv_tag, &ds)
1195             != 0) {
1196                 /* %recv does not exist; continue in tofs */
1197                 VERIFY0(dsl_dataset_own_force(dp, tofs, dsflags, dmu_recv_tag,
1198                     &ds));
1199                 drba->drba_cookie->drc_newfs = B_TRUE;
1200         }
1201
1202         ASSERT(DS_IS_INCONSISTENT(ds));
1203         rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
1204         ASSERT(!BP_IS_HOLE(dsl_dataset_get_blkptr(ds)) ||
1205             drba->drba_cookie->drc_raw);
1206         rrw_exit(&ds->ds_bp_rwlock, FTAG);
1207
1208         drba->drba_cookie->drc_ds = ds;
1209         VERIFY0(dmu_objset_from_ds(ds, &drba->drba_cookie->drc_os));
1210         drba->drba_cookie->drc_should_save = B_TRUE;
1211
1212         spa_history_log_internal_ds(ds, "resume receive", tx, " ");
1213 }
1214
1215 /*
1216  * NB: callers *MUST* call dmu_recv_stream() if dmu_recv_begin()
1217  * succeeds; otherwise we will leak the holds on the datasets.
1218  */
1219 int
1220 dmu_recv_begin(const char *tofs, const char *tosnap,
1221     dmu_replay_record_t *drr_begin, boolean_t force, boolean_t heal,
1222     boolean_t resumable, nvlist_t *localprops, nvlist_t *hidden_args,
1223     const char *origin, dmu_recv_cookie_t *drc, zfs_file_t *fp,
1224     offset_t *voffp)
1225 {
1226         dmu_recv_begin_arg_t drba = { 0 };
1227         int err = 0;
1228
1229         memset(drc, 0, sizeof (dmu_recv_cookie_t));
1230         drc->drc_drr_begin = drr_begin;
1231         drc->drc_drrb = &drr_begin->drr_u.drr_begin;
1232         drc->drc_tosnap = tosnap;
1233         drc->drc_tofs = tofs;
1234         drc->drc_force = force;
1235         drc->drc_heal = heal;
1236         drc->drc_resumable = resumable;
1237         drc->drc_cred = CRED();
1238         drc->drc_proc = curproc;
1239         drc->drc_clone = (origin != NULL);
1240
1241         if (drc->drc_drrb->drr_magic == BSWAP_64(DMU_BACKUP_MAGIC)) {
1242                 drc->drc_byteswap = B_TRUE;
1243                 (void) fletcher_4_incremental_byteswap(drr_begin,
1244                     sizeof (dmu_replay_record_t), &drc->drc_cksum);
1245                 byteswap_record(drr_begin);
1246         } else if (drc->drc_drrb->drr_magic == DMU_BACKUP_MAGIC) {
1247                 (void) fletcher_4_incremental_native(drr_begin,
1248                     sizeof (dmu_replay_record_t), &drc->drc_cksum);
1249         } else {
1250                 return (SET_ERROR(EINVAL));
1251         }
1252
1253         drc->drc_fp = fp;
1254         drc->drc_voff = *voffp;
1255         drc->drc_featureflags =
1256             DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo);
1257
1258         uint32_t payloadlen = drc->drc_drr_begin->drr_payloadlen;
1259
1260         /*
1261          * Since OpenZFS 2.0.0, we have enforced a 64MB limit in userspace
1262          * configurable via ZFS_SENDRECV_MAX_NVLIST. We enforce 256MB as a hard
1263          * upper limit. Systems with less than 1GB of RAM will see a lower
1264          * limit from `arc_all_memory() / 4`.
1265          */
1266         if (payloadlen > (MIN((1U << 28), arc_all_memory() / 4)))
1267                 return (E2BIG);
1268
1269
1270         if (payloadlen != 0) {
1271                 void *payload = vmem_alloc(payloadlen, KM_SLEEP);
1272                 /*
1273                  * For compatibility with recursive send streams, we don't do
1274                  * this here if the stream could be part of a package. Instead,
1275                  * we'll do it in dmu_recv_stream. If we pull the next header
1276                  * too early, and it's the END record, we break the `recv_skip`
1277                  * logic.
1278                  */
1279
1280                 err = receive_read_payload_and_next_header(drc, payloadlen,
1281                     payload);
1282                 if (err != 0) {
1283                         vmem_free(payload, payloadlen);
1284                         return (err);
1285                 }
1286                 err = nvlist_unpack(payload, payloadlen, &drc->drc_begin_nvl,
1287                     KM_SLEEP);
1288                 vmem_free(payload, payloadlen);
1289                 if (err != 0) {
1290                         kmem_free(drc->drc_next_rrd,
1291                             sizeof (*drc->drc_next_rrd));
1292                         return (err);
1293                 }
1294         }
1295
1296         if (drc->drc_drrb->drr_flags & DRR_FLAG_SPILL_BLOCK)
1297                 drc->drc_spill = B_TRUE;
1298
1299         drba.drba_origin = origin;
1300         drba.drba_cookie = drc;
1301         drba.drba_cred = CRED();
1302         drba.drba_proc = curproc;
1303
1304         if (drc->drc_featureflags & DMU_BACKUP_FEATURE_RESUMING) {
1305                 err = dsl_sync_task(tofs,
1306                     dmu_recv_resume_begin_check, dmu_recv_resume_begin_sync,
1307                     &drba, 5, ZFS_SPACE_CHECK_NORMAL);
1308         } else {
1309                 /*
1310                  * For non-raw, non-incremental, non-resuming receives the
1311                  * user can specify encryption parameters on the command line
1312                  * with "zfs recv -o". For these receives we create a dcp and
1313                  * pass it to the sync task. Creating the dcp will implicitly
1314                  * remove the encryption params from the localprops nvlist,
1315                  * which avoids errors when trying to set these normally
1316                  * read-only properties. Any other kind of receive that
1317                  * attempts to set these properties will fail as a result.
1318                  */
1319                 if ((DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo) &
1320                     DMU_BACKUP_FEATURE_RAW) == 0 &&
1321                     origin == NULL && drc->drc_drrb->drr_fromguid == 0) {
1322                         err = dsl_crypto_params_create_nvlist(DCP_CMD_NONE,
1323                             localprops, hidden_args, &drba.drba_dcp);
1324                 }
1325
1326                 if (err == 0) {
1327                         err = dsl_sync_task(tofs,
1328                             dmu_recv_begin_check, dmu_recv_begin_sync,
1329                             &drba, 5, ZFS_SPACE_CHECK_NORMAL);
1330                         dsl_crypto_params_free(drba.drba_dcp, !!err);
1331                 }
1332         }
1333
1334         if (err != 0) {
1335                 kmem_free(drc->drc_next_rrd, sizeof (*drc->drc_next_rrd));
1336                 nvlist_free(drc->drc_begin_nvl);
1337         }
1338         return (err);
1339 }
1340
1341 /*
1342  * Holds data need for corrective recv callback
1343  */
1344 typedef struct cr_cb_data {
1345         uint64_t size;
1346         zbookmark_phys_t zb;
1347         spa_t *spa;
1348 } cr_cb_data_t;
1349
1350 static void
1351 corrective_read_done(zio_t *zio)
1352 {
1353         cr_cb_data_t *data = zio->io_private;
1354         /* Corruption corrected; update error log if needed */
1355         if (zio->io_error == 0)
1356                 spa_remove_error(data->spa, &data->zb, &zio->io_bp->blk_birth);
1357         kmem_free(data, sizeof (cr_cb_data_t));
1358         abd_free(zio->io_abd);
1359 }
1360
1361 /*
1362  * zio_rewrite the data pointed to by bp with the data from the rrd's abd.
1363  */
1364 static int
1365 do_corrective_recv(struct receive_writer_arg *rwa, struct drr_write *drrw,
1366     struct receive_record_arg *rrd, blkptr_t *bp)
1367 {
1368         int err;
1369         zio_t *io;
1370         zbookmark_phys_t zb;
1371         dnode_t *dn;
1372         abd_t *abd = rrd->abd;
1373         zio_cksum_t bp_cksum = bp->blk_cksum;
1374         zio_flag_t flags = ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_RETRY |
1375             ZIO_FLAG_CANFAIL;
1376
1377         if (rwa->raw)
1378                 flags |= ZIO_FLAG_RAW;
1379
1380         err = dnode_hold(rwa->os, drrw->drr_object, FTAG, &dn);
1381         if (err != 0)
1382                 return (err);
1383         SET_BOOKMARK(&zb, dmu_objset_id(rwa->os), drrw->drr_object, 0,
1384             dbuf_whichblock(dn, 0, drrw->drr_offset));
1385         dnode_rele(dn, FTAG);
1386
1387         if (!rwa->raw && DRR_WRITE_COMPRESSED(drrw)) {
1388                 /* Decompress the stream data */
1389                 abd_t *dabd = abd_alloc_linear(
1390                     drrw->drr_logical_size, B_FALSE);
1391                 err = zio_decompress_data(drrw->drr_compressiontype,
1392                     abd, abd_to_buf(dabd), abd_get_size(abd),
1393                     abd_get_size(dabd), NULL);
1394
1395                 if (err != 0) {
1396                         abd_free(dabd);
1397                         return (err);
1398                 }
1399                 /* Swap in the newly decompressed data into the abd */
1400                 abd_free(abd);
1401                 abd = dabd;
1402         }
1403
1404         if (!rwa->raw && BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF) {
1405                 /* Recompress the data */
1406                 abd_t *cabd = abd_alloc_linear(BP_GET_PSIZE(bp),
1407                     B_FALSE);
1408                 void *buf = abd_to_buf(cabd);
1409                 uint64_t csize = zio_compress_data(BP_GET_COMPRESS(bp),
1410                     abd, &buf, abd_get_size(abd),
1411                     rwa->os->os_complevel);
1412                 abd_zero_off(cabd, csize, BP_GET_PSIZE(bp) - csize);
1413                 /* Swap in newly compressed data into the abd */
1414                 abd_free(abd);
1415                 abd = cabd;
1416                 flags |= ZIO_FLAG_RAW_COMPRESS;
1417         }
1418
1419         /*
1420          * The stream is not encrypted but the data on-disk is.
1421          * We need to re-encrypt the buf using the same
1422          * encryption type, salt, iv, and mac that was used to encrypt
1423          * the block previosly.
1424          */
1425         if (!rwa->raw && BP_USES_CRYPT(bp)) {
1426                 dsl_dataset_t *ds;
1427                 dsl_crypto_key_t *dck = NULL;
1428                 uint8_t salt[ZIO_DATA_SALT_LEN];
1429                 uint8_t iv[ZIO_DATA_IV_LEN];
1430                 uint8_t mac[ZIO_DATA_MAC_LEN];
1431                 boolean_t no_crypt = B_FALSE;
1432                 dsl_pool_t *dp = dmu_objset_pool(rwa->os);
1433                 abd_t *eabd = abd_alloc_linear(BP_GET_PSIZE(bp), B_FALSE);
1434
1435                 zio_crypt_decode_params_bp(bp, salt, iv);
1436                 zio_crypt_decode_mac_bp(bp, mac);
1437
1438                 dsl_pool_config_enter(dp, FTAG);
1439                 err = dsl_dataset_hold_flags(dp, rwa->tofs,
1440                     DS_HOLD_FLAG_DECRYPT, FTAG, &ds);
1441                 if (err != 0) {
1442                         dsl_pool_config_exit(dp, FTAG);
1443                         abd_free(eabd);
1444                         return (SET_ERROR(EACCES));
1445                 }
1446
1447                 /* Look up the key from the spa's keystore */
1448                 err = spa_keystore_lookup_key(rwa->os->os_spa,
1449                     zb.zb_objset, FTAG, &dck);
1450                 if (err != 0) {
1451                         dsl_dataset_rele_flags(ds, DS_HOLD_FLAG_DECRYPT,
1452                             FTAG);
1453                         dsl_pool_config_exit(dp, FTAG);
1454                         abd_free(eabd);
1455                         return (SET_ERROR(EACCES));
1456                 }
1457
1458                 err = zio_do_crypt_abd(B_TRUE, &dck->dck_key,
1459                     BP_GET_TYPE(bp), BP_SHOULD_BYTESWAP(bp), salt, iv,
1460                     mac, abd_get_size(abd), abd, eabd, &no_crypt);
1461
1462                 spa_keystore_dsl_key_rele(rwa->os->os_spa, dck, FTAG);
1463                 dsl_dataset_rele_flags(ds, DS_HOLD_FLAG_DECRYPT, FTAG);
1464                 dsl_pool_config_exit(dp, FTAG);
1465
1466                 ASSERT0(no_crypt);
1467                 if (err != 0) {
1468                         abd_free(eabd);
1469                         return (err);
1470                 }
1471                 /* Swap in the newly encrypted data into the abd */
1472                 abd_free(abd);
1473                 abd = eabd;
1474
1475                 /*
1476                  * We want to prevent zio_rewrite() from trying to
1477                  * encrypt the data again
1478                  */
1479                 flags |= ZIO_FLAG_RAW_ENCRYPT;
1480         }
1481         rrd->abd = abd;
1482
1483         io = zio_rewrite(NULL, rwa->os->os_spa, bp->blk_birth, bp, abd,
1484             BP_GET_PSIZE(bp), NULL, NULL, ZIO_PRIORITY_SYNC_WRITE, flags, &zb);
1485
1486         ASSERT(abd_get_size(abd) == BP_GET_LSIZE(bp) ||
1487             abd_get_size(abd) == BP_GET_PSIZE(bp));
1488
1489         /* compute new bp checksum value and make sure it matches the old one */
1490         zio_checksum_compute(io, BP_GET_CHECKSUM(bp), abd, abd_get_size(abd));
1491         if (!ZIO_CHECKSUM_EQUAL(bp_cksum, io->io_bp->blk_cksum)) {
1492                 zio_destroy(io);
1493                 if (zfs_recv_best_effort_corrective != 0)
1494                         return (0);
1495                 return (SET_ERROR(ECKSUM));
1496         }
1497
1498         /* Correct the corruption in place */
1499         err = zio_wait(io);
1500         if (err == 0) {
1501                 cr_cb_data_t *cb_data =
1502                     kmem_alloc(sizeof (cr_cb_data_t), KM_SLEEP);
1503                 cb_data->spa = rwa->os->os_spa;
1504                 cb_data->size = drrw->drr_logical_size;
1505                 cb_data->zb = zb;
1506                 /* Test if healing worked by re-reading the bp */
1507                 err = zio_wait(zio_read(rwa->heal_pio, rwa->os->os_spa, bp,
1508                     abd_alloc_for_io(drrw->drr_logical_size, B_FALSE),
1509                     drrw->drr_logical_size, corrective_read_done,
1510                     cb_data, ZIO_PRIORITY_ASYNC_READ, flags, NULL));
1511         }
1512         if (err != 0 && zfs_recv_best_effort_corrective != 0)
1513                 err = 0;
1514
1515         return (err);
1516 }
1517
1518 static int
1519 receive_read(dmu_recv_cookie_t *drc, int len, void *buf)
1520 {
1521         int done = 0;
1522
1523         /*
1524          * The code doesn't rely on this (lengths being multiples of 8).  See
1525          * comment in dump_bytes.
1526          */
1527         ASSERT(len % 8 == 0 ||
1528             (drc->drc_featureflags & DMU_BACKUP_FEATURE_RAW) != 0);
1529
1530         while (done < len) {
1531                 ssize_t resid = len - done;
1532                 zfs_file_t *fp = drc->drc_fp;
1533                 int err = zfs_file_read(fp, (char *)buf + done,
1534                     len - done, &resid);
1535                 if (err == 0 && resid == len - done) {
1536                         /*
1537                          * Note: ECKSUM or ZFS_ERR_STREAM_TRUNCATED indicates
1538                          * that the receive was interrupted and can
1539                          * potentially be resumed.
1540                          */
1541                         err = SET_ERROR(ZFS_ERR_STREAM_TRUNCATED);
1542                 }
1543                 drc->drc_voff += len - done - resid;
1544                 done = len - resid;
1545                 if (err != 0)
1546                         return (err);
1547         }
1548
1549         drc->drc_bytes_read += len;
1550
1551         ASSERT3U(done, ==, len);
1552         return (0);
1553 }
1554
1555 static inline uint8_t
1556 deduce_nblkptr(dmu_object_type_t bonus_type, uint64_t bonus_size)
1557 {
1558         if (bonus_type == DMU_OT_SA) {
1559                 return (1);
1560         } else {
1561                 return (1 +
1562                     ((DN_OLD_MAX_BONUSLEN -
1563                     MIN(DN_OLD_MAX_BONUSLEN, bonus_size)) >> SPA_BLKPTRSHIFT));
1564         }
1565 }
1566
1567 static void
1568 save_resume_state(struct receive_writer_arg *rwa,
1569     uint64_t object, uint64_t offset, dmu_tx_t *tx)
1570 {
1571         int txgoff = dmu_tx_get_txg(tx) & TXG_MASK;
1572
1573         if (!rwa->resumable)
1574                 return;
1575
1576         /*
1577          * We use ds_resume_bytes[] != 0 to indicate that we need to
1578          * update this on disk, so it must not be 0.
1579          */
1580         ASSERT(rwa->bytes_read != 0);
1581
1582         /*
1583          * We only resume from write records, which have a valid
1584          * (non-meta-dnode) object number.
1585          */
1586         ASSERT(object != 0);
1587
1588         /*
1589          * For resuming to work correctly, we must receive records in order,
1590          * sorted by object,offset.  This is checked by the callers, but
1591          * assert it here for good measure.
1592          */
1593         ASSERT3U(object, >=, rwa->os->os_dsl_dataset->ds_resume_object[txgoff]);
1594         ASSERT(object != rwa->os->os_dsl_dataset->ds_resume_object[txgoff] ||
1595             offset >= rwa->os->os_dsl_dataset->ds_resume_offset[txgoff]);
1596         ASSERT3U(rwa->bytes_read, >=,
1597             rwa->os->os_dsl_dataset->ds_resume_bytes[txgoff]);
1598
1599         rwa->os->os_dsl_dataset->ds_resume_object[txgoff] = object;
1600         rwa->os->os_dsl_dataset->ds_resume_offset[txgoff] = offset;
1601         rwa->os->os_dsl_dataset->ds_resume_bytes[txgoff] = rwa->bytes_read;
1602 }
1603
1604 static int
1605 receive_object_is_same_generation(objset_t *os, uint64_t object,
1606     dmu_object_type_t old_bonus_type, dmu_object_type_t new_bonus_type,
1607     const void *new_bonus, boolean_t *samegenp)
1608 {
1609         zfs_file_info_t zoi;
1610         int err;
1611
1612         dmu_buf_t *old_bonus_dbuf;
1613         err = dmu_bonus_hold(os, object, FTAG, &old_bonus_dbuf);
1614         if (err != 0)
1615                 return (err);
1616         err = dmu_get_file_info(os, old_bonus_type, old_bonus_dbuf->db_data,
1617             &zoi);
1618         dmu_buf_rele(old_bonus_dbuf, FTAG);
1619         if (err != 0)
1620                 return (err);
1621         uint64_t old_gen = zoi.zfi_generation;
1622
1623         err = dmu_get_file_info(os, new_bonus_type, new_bonus, &zoi);
1624         if (err != 0)
1625                 return (err);
1626         uint64_t new_gen = zoi.zfi_generation;
1627
1628         *samegenp = (old_gen == new_gen);
1629         return (0);
1630 }
1631
1632 static int
1633 receive_handle_existing_object(const struct receive_writer_arg *rwa,
1634     const struct drr_object *drro, const dmu_object_info_t *doi,
1635     const void *bonus_data,
1636     uint64_t *object_to_hold, uint32_t *new_blksz)
1637 {
1638         uint32_t indblksz = drro->drr_indblkshift ?
1639             1ULL << drro->drr_indblkshift : 0;
1640         int nblkptr = deduce_nblkptr(drro->drr_bonustype,
1641             drro->drr_bonuslen);
1642         uint8_t dn_slots = drro->drr_dn_slots != 0 ?
1643             drro->drr_dn_slots : DNODE_MIN_SLOTS;
1644         boolean_t do_free_range = B_FALSE;
1645         int err;
1646
1647         *object_to_hold = drro->drr_object;
1648
1649         /* nblkptr should be bounded by the bonus size and type */
1650         if (rwa->raw && nblkptr != drro->drr_nblkptr)
1651                 return (SET_ERROR(EINVAL));
1652
1653         /*
1654          * After the previous send stream, the sending system may
1655          * have freed this object, and then happened to re-allocate
1656          * this object number in a later txg. In this case, we are
1657          * receiving a different logical file, and the block size may
1658          * appear to be different.  i.e. we may have a different
1659          * block size for this object than what the send stream says.
1660          * In this case we need to remove the object's contents,
1661          * so that its structure can be changed and then its contents
1662          * entirely replaced by subsequent WRITE records.
1663          *
1664          * If this is a -L (--large-block) incremental stream, and
1665          * the previous stream was not -L, the block size may appear
1666          * to increase.  i.e. we may have a smaller block size for
1667          * this object than what the send stream says.  In this case
1668          * we need to keep the object's contents and block size
1669          * intact, so that we don't lose parts of the object's
1670          * contents that are not changed by this incremental send
1671          * stream.
1672          *
1673          * We can distinguish between the two above cases by using
1674          * the ZPL's generation number (see
1675          * receive_object_is_same_generation()).  However, we only
1676          * want to rely on the generation number when absolutely
1677          * necessary, because with raw receives, the generation is
1678          * encrypted.  We also want to minimize dependence on the
1679          * ZPL, so that other types of datasets can also be received
1680          * (e.g. ZVOLs, although note that ZVOLS currently do not
1681          * reallocate their objects or change their structure).
1682          * Therefore, we check a number of different cases where we
1683          * know it is safe to discard the object's contents, before
1684          * using the ZPL's generation number to make the above
1685          * distinction.
1686          */
1687         if (drro->drr_blksz != doi->doi_data_block_size) {
1688                 if (rwa->raw) {
1689                         /*
1690                          * RAW streams always have large blocks, so
1691                          * we are sure that the data is not needed
1692                          * due to changing --large-block to be on.
1693                          * Which is fortunate since the bonus buffer
1694                          * (which contains the ZPL generation) is
1695                          * encrypted, and the key might not be
1696                          * loaded.
1697                          */
1698                         do_free_range = B_TRUE;
1699                 } else if (rwa->full) {
1700                         /*
1701                          * This is a full send stream, so it always
1702                          * replaces what we have.  Even if the
1703                          * generation numbers happen to match, this
1704                          * can not actually be the same logical file.
1705                          * This is relevant when receiving a full
1706                          * send as a clone.
1707                          */
1708                         do_free_range = B_TRUE;
1709                 } else if (drro->drr_type !=
1710                     DMU_OT_PLAIN_FILE_CONTENTS ||
1711                     doi->doi_type != DMU_OT_PLAIN_FILE_CONTENTS) {
1712                         /*
1713                          * PLAIN_FILE_CONTENTS are the only type of
1714                          * objects that have ever been stored with
1715                          * large blocks, so we don't need the special
1716                          * logic below.  ZAP blocks can shrink (when
1717                          * there's only one block), so we don't want
1718                          * to hit the error below about block size
1719                          * only increasing.
1720                          */
1721                         do_free_range = B_TRUE;
1722                 } else if (doi->doi_max_offset <=
1723                     doi->doi_data_block_size) {
1724                         /*
1725                          * There is only one block.  We can free it,
1726                          * because its contents will be replaced by a
1727                          * WRITE record.  This can not be the no-L ->
1728                          * -L case, because the no-L case would have
1729                          * resulted in multiple blocks.  If we
1730                          * supported -L -> no-L, it would not be safe
1731                          * to free the file's contents.  Fortunately,
1732                          * that is not allowed (see
1733                          * recv_check_large_blocks()).
1734                          */
1735                         do_free_range = B_TRUE;
1736                 } else {
1737                         boolean_t is_same_gen;
1738                         err = receive_object_is_same_generation(rwa->os,
1739                             drro->drr_object, doi->doi_bonus_type,
1740                             drro->drr_bonustype, bonus_data, &is_same_gen);
1741                         if (err != 0)
1742                                 return (SET_ERROR(EINVAL));
1743
1744                         if (is_same_gen) {
1745                                 /*
1746                                  * This is the same logical file, and
1747                                  * the block size must be increasing.
1748                                  * It could only decrease if
1749                                  * --large-block was changed to be
1750                                  * off, which is checked in
1751                                  * recv_check_large_blocks().
1752                                  */
1753                                 if (drro->drr_blksz <=
1754                                     doi->doi_data_block_size)
1755                                         return (SET_ERROR(EINVAL));
1756                                 /*
1757                                  * We keep the existing blocksize and
1758                                  * contents.
1759                                  */
1760                                 *new_blksz =
1761                                     doi->doi_data_block_size;
1762                         } else {
1763                                 do_free_range = B_TRUE;
1764                         }
1765                 }
1766         }
1767
1768         /* nblkptr can only decrease if the object was reallocated */
1769         if (nblkptr < doi->doi_nblkptr)
1770                 do_free_range = B_TRUE;
1771
1772         /* number of slots can only change on reallocation */
1773         if (dn_slots != doi->doi_dnodesize >> DNODE_SHIFT)
1774                 do_free_range = B_TRUE;
1775
1776         /*
1777          * For raw sends we also check a few other fields to
1778          * ensure we are preserving the objset structure exactly
1779          * as it was on the receive side:
1780          *     - A changed indirect block size
1781          *     - A smaller nlevels
1782          */
1783         if (rwa->raw) {
1784                 if (indblksz != doi->doi_metadata_block_size)
1785                         do_free_range = B_TRUE;
1786                 if (drro->drr_nlevels < doi->doi_indirection)
1787                         do_free_range = B_TRUE;
1788         }
1789
1790         if (do_free_range) {
1791                 err = dmu_free_long_range(rwa->os, drro->drr_object,
1792                     0, DMU_OBJECT_END);
1793                 if (err != 0)
1794                         return (SET_ERROR(EINVAL));
1795         }
1796
1797         /*
1798          * The dmu does not currently support decreasing nlevels
1799          * or changing the number of dnode slots on an object. For
1800          * non-raw sends, this does not matter and the new object
1801          * can just use the previous one's nlevels. For raw sends,
1802          * however, the structure of the received dnode (including
1803          * nlevels and dnode slots) must match that of the send
1804          * side. Therefore, instead of using dmu_object_reclaim(),
1805          * we must free the object completely and call
1806          * dmu_object_claim_dnsize() instead.
1807          */
1808         if ((rwa->raw && drro->drr_nlevels < doi->doi_indirection) ||
1809             dn_slots != doi->doi_dnodesize >> DNODE_SHIFT) {
1810                 err = dmu_free_long_object(rwa->os, drro->drr_object);
1811                 if (err != 0)
1812                         return (SET_ERROR(EINVAL));
1813
1814                 txg_wait_synced(dmu_objset_pool(rwa->os), 0);
1815                 *object_to_hold = DMU_NEW_OBJECT;
1816         }
1817
1818         /*
1819          * For raw receives, free everything beyond the new incoming
1820          * maxblkid. Normally this would be done with a DRR_FREE
1821          * record that would come after this DRR_OBJECT record is
1822          * processed. However, for raw receives we manually set the
1823          * maxblkid from the drr_maxblkid and so we must first free
1824          * everything above that blkid to ensure the DMU is always
1825          * consistent with itself. We will never free the first block
1826          * of the object here because a maxblkid of 0 could indicate
1827          * an object with a single block or one with no blocks. This
1828          * free may be skipped when dmu_free_long_range() was called
1829          * above since it covers the entire object's contents.
1830          */
1831         if (rwa->raw && *object_to_hold != DMU_NEW_OBJECT && !do_free_range) {
1832                 err = dmu_free_long_range(rwa->os, drro->drr_object,
1833                     (drro->drr_maxblkid + 1) * doi->doi_data_block_size,
1834                     DMU_OBJECT_END);
1835                 if (err != 0)
1836                         return (SET_ERROR(EINVAL));
1837         }
1838         return (0);
1839 }
1840
1841 noinline static int
1842 receive_object(struct receive_writer_arg *rwa, struct drr_object *drro,
1843     void *data)
1844 {
1845         dmu_object_info_t doi;
1846         dmu_tx_t *tx;
1847         int err;
1848         uint32_t new_blksz = drro->drr_blksz;
1849         uint8_t dn_slots = drro->drr_dn_slots != 0 ?
1850             drro->drr_dn_slots : DNODE_MIN_SLOTS;
1851
1852         if (drro->drr_type == DMU_OT_NONE ||
1853             !DMU_OT_IS_VALID(drro->drr_type) ||
1854             !DMU_OT_IS_VALID(drro->drr_bonustype) ||
1855             drro->drr_checksumtype >= ZIO_CHECKSUM_FUNCTIONS ||
1856             drro->drr_compress >= ZIO_COMPRESS_FUNCTIONS ||
1857             P2PHASE(drro->drr_blksz, SPA_MINBLOCKSIZE) ||
1858             drro->drr_blksz < SPA_MINBLOCKSIZE ||
1859             drro->drr_blksz > spa_maxblocksize(dmu_objset_spa(rwa->os)) ||
1860             drro->drr_bonuslen >
1861             DN_BONUS_SIZE(spa_maxdnodesize(dmu_objset_spa(rwa->os))) ||
1862             dn_slots >
1863             (spa_maxdnodesize(dmu_objset_spa(rwa->os)) >> DNODE_SHIFT)) {
1864                 return (SET_ERROR(EINVAL));
1865         }
1866
1867         if (rwa->raw) {
1868                 /*
1869                  * We should have received a DRR_OBJECT_RANGE record
1870                  * containing this block and stored it in rwa.
1871                  */
1872                 if (drro->drr_object < rwa->or_firstobj ||
1873                     drro->drr_object >= rwa->or_firstobj + rwa->or_numslots ||
1874                     drro->drr_raw_bonuslen < drro->drr_bonuslen ||
1875                     drro->drr_indblkshift > SPA_MAXBLOCKSHIFT ||
1876                     drro->drr_nlevels > DN_MAX_LEVELS ||
1877                     drro->drr_nblkptr > DN_MAX_NBLKPTR ||
1878                     DN_SLOTS_TO_BONUSLEN(dn_slots) <
1879                     drro->drr_raw_bonuslen)
1880                         return (SET_ERROR(EINVAL));
1881         } else {
1882                 /*
1883                  * The DRR_OBJECT_SPILL flag is valid when the DRR_BEGIN
1884                  * record indicates this by setting DRR_FLAG_SPILL_BLOCK.
1885                  */
1886                 if (((drro->drr_flags & ~(DRR_OBJECT_SPILL))) ||
1887                     (!rwa->spill && DRR_OBJECT_HAS_SPILL(drro->drr_flags))) {
1888                         return (SET_ERROR(EINVAL));
1889                 }
1890
1891                 if (drro->drr_raw_bonuslen != 0 || drro->drr_nblkptr != 0 ||
1892                     drro->drr_indblkshift != 0 || drro->drr_nlevels != 0) {
1893                         return (SET_ERROR(EINVAL));
1894                 }
1895         }
1896
1897         err = dmu_object_info(rwa->os, drro->drr_object, &doi);
1898
1899         if (err != 0 && err != ENOENT && err != EEXIST)
1900                 return (SET_ERROR(EINVAL));
1901
1902         if (drro->drr_object > rwa->max_object)
1903                 rwa->max_object = drro->drr_object;
1904
1905         /*
1906          * If we are losing blkptrs or changing the block size this must
1907          * be a new file instance.  We must clear out the previous file
1908          * contents before we can change this type of metadata in the dnode.
1909          * Raw receives will also check that the indirect structure of the
1910          * dnode hasn't changed.
1911          */
1912         uint64_t object_to_hold;
1913         if (err == 0) {
1914                 err = receive_handle_existing_object(rwa, drro, &doi, data,
1915                     &object_to_hold, &new_blksz);
1916                 if (err != 0)
1917                         return (err);
1918         } else if (err == EEXIST) {
1919                 /*
1920                  * The object requested is currently an interior slot of a
1921                  * multi-slot dnode. This will be resolved when the next txg
1922                  * is synced out, since the send stream will have told us
1923                  * to free this slot when we freed the associated dnode
1924                  * earlier in the stream.
1925                  */
1926                 txg_wait_synced(dmu_objset_pool(rwa->os), 0);
1927
1928                 if (dmu_object_info(rwa->os, drro->drr_object, NULL) != ENOENT)
1929                         return (SET_ERROR(EINVAL));
1930
1931                 /* object was freed and we are about to allocate a new one */
1932                 object_to_hold = DMU_NEW_OBJECT;
1933         } else {
1934                 /*
1935                  * If the only record in this range so far was DRR_FREEOBJECTS
1936                  * with at least one actually freed object, it's possible that
1937                  * the block will now be converted to a hole. We need to wait
1938                  * for the txg to sync to prevent races.
1939                  */
1940                 if (rwa->or_need_sync == ORNS_YES)
1941                         txg_wait_synced(dmu_objset_pool(rwa->os), 0);
1942
1943                 /* object is free and we are about to allocate a new one */
1944                 object_to_hold = DMU_NEW_OBJECT;
1945         }
1946
1947         /* Only relevant for the first object in the range */
1948         rwa->or_need_sync = ORNS_NO;
1949
1950         /*
1951          * If this is a multi-slot dnode there is a chance that this
1952          * object will expand into a slot that is already used by
1953          * another object from the previous snapshot. We must free
1954          * these objects before we attempt to allocate the new dnode.
1955          */
1956         if (dn_slots > 1) {
1957                 boolean_t need_sync = B_FALSE;
1958
1959                 for (uint64_t slot = drro->drr_object + 1;
1960                     slot < drro->drr_object + dn_slots;
1961                     slot++) {
1962                         dmu_object_info_t slot_doi;
1963
1964                         err = dmu_object_info(rwa->os, slot, &slot_doi);
1965                         if (err == ENOENT || err == EEXIST)
1966                                 continue;
1967                         else if (err != 0)
1968                                 return (err);
1969
1970                         err = dmu_free_long_object(rwa->os, slot);
1971                         if (err != 0)
1972                                 return (err);
1973
1974                         need_sync = B_TRUE;
1975                 }
1976
1977                 if (need_sync)
1978                         txg_wait_synced(dmu_objset_pool(rwa->os), 0);
1979         }
1980
1981         tx = dmu_tx_create(rwa->os);
1982         dmu_tx_hold_bonus(tx, object_to_hold);
1983         dmu_tx_hold_write(tx, object_to_hold, 0, 0);
1984         err = dmu_tx_assign(tx, TXG_WAIT);
1985         if (err != 0) {
1986                 dmu_tx_abort(tx);
1987                 return (err);
1988         }
1989
1990         if (object_to_hold == DMU_NEW_OBJECT) {
1991                 /* Currently free, wants to be allocated */
1992                 err = dmu_object_claim_dnsize(rwa->os, drro->drr_object,
1993                     drro->drr_type, new_blksz,
1994                     drro->drr_bonustype, drro->drr_bonuslen,
1995                     dn_slots << DNODE_SHIFT, tx);
1996         } else if (drro->drr_type != doi.doi_type ||
1997             new_blksz != doi.doi_data_block_size ||
1998             drro->drr_bonustype != doi.doi_bonus_type ||
1999             drro->drr_bonuslen != doi.doi_bonus_size) {
2000                 /* Currently allocated, but with different properties */
2001                 err = dmu_object_reclaim_dnsize(rwa->os, drro->drr_object,
2002                     drro->drr_type, new_blksz,
2003                     drro->drr_bonustype, drro->drr_bonuslen,
2004                     dn_slots << DNODE_SHIFT, rwa->spill ?
2005                     DRR_OBJECT_HAS_SPILL(drro->drr_flags) : B_FALSE, tx);
2006         } else if (rwa->spill && !DRR_OBJECT_HAS_SPILL(drro->drr_flags)) {
2007                 /*
2008                  * Currently allocated, the existing version of this object
2009                  * may reference a spill block that is no longer allocated
2010                  * at the source and needs to be freed.
2011                  */
2012                 err = dmu_object_rm_spill(rwa->os, drro->drr_object, tx);
2013         }
2014
2015         if (err != 0) {
2016                 dmu_tx_commit(tx);
2017                 return (SET_ERROR(EINVAL));
2018         }
2019
2020         if (rwa->or_crypt_params_present) {
2021                 /*
2022                  * Set the crypt params for the buffer associated with this
2023                  * range of dnodes.  This causes the blkptr_t to have the
2024                  * same crypt params (byteorder, salt, iv, mac) as on the
2025                  * sending side.
2026                  *
2027                  * Since we are committing this tx now, it is possible for
2028                  * the dnode block to end up on-disk with the incorrect MAC,
2029                  * if subsequent objects in this block are received in a
2030                  * different txg.  However, since the dataset is marked as
2031                  * inconsistent, no code paths will do a non-raw read (or
2032                  * decrypt the block / verify the MAC). The receive code and
2033                  * scrub code can safely do raw reads and verify the
2034                  * checksum.  They don't need to verify the MAC.
2035                  */
2036                 dmu_buf_t *db = NULL;
2037                 uint64_t offset = rwa->or_firstobj * DNODE_MIN_SIZE;
2038
2039                 err = dmu_buf_hold_by_dnode(DMU_META_DNODE(rwa->os),
2040                     offset, FTAG, &db, DMU_READ_PREFETCH | DMU_READ_NO_DECRYPT);
2041                 if (err != 0) {
2042                         dmu_tx_commit(tx);
2043                         return (SET_ERROR(EINVAL));
2044                 }
2045
2046                 dmu_buf_set_crypt_params(db, rwa->or_byteorder,
2047                     rwa->or_salt, rwa->or_iv, rwa->or_mac, tx);
2048
2049                 dmu_buf_rele(db, FTAG);
2050
2051                 rwa->or_crypt_params_present = B_FALSE;
2052         }
2053
2054         dmu_object_set_checksum(rwa->os, drro->drr_object,
2055             drro->drr_checksumtype, tx);
2056         dmu_object_set_compress(rwa->os, drro->drr_object,
2057             drro->drr_compress, tx);
2058
2059         /* handle more restrictive dnode structuring for raw recvs */
2060         if (rwa->raw) {
2061                 /*
2062                  * Set the indirect block size, block shift, nlevels.
2063                  * This will not fail because we ensured all of the
2064                  * blocks were freed earlier if this is a new object.
2065                  * For non-new objects block size and indirect block
2066                  * shift cannot change and nlevels can only increase.
2067                  */
2068                 ASSERT3U(new_blksz, ==, drro->drr_blksz);
2069                 VERIFY0(dmu_object_set_blocksize(rwa->os, drro->drr_object,
2070                     drro->drr_blksz, drro->drr_indblkshift, tx));
2071                 VERIFY0(dmu_object_set_nlevels(rwa->os, drro->drr_object,
2072                     drro->drr_nlevels, tx));
2073
2074                 /*
2075                  * Set the maxblkid. This will always succeed because
2076                  * we freed all blocks beyond the new maxblkid above.
2077                  */
2078                 VERIFY0(dmu_object_set_maxblkid(rwa->os, drro->drr_object,
2079                     drro->drr_maxblkid, tx));
2080         }
2081
2082         if (data != NULL) {
2083                 dmu_buf_t *db;
2084                 dnode_t *dn;
2085                 uint32_t flags = DMU_READ_NO_PREFETCH;
2086
2087                 if (rwa->raw)
2088                         flags |= DMU_READ_NO_DECRYPT;
2089
2090                 VERIFY0(dnode_hold(rwa->os, drro->drr_object, FTAG, &dn));
2091                 VERIFY0(dmu_bonus_hold_by_dnode(dn, FTAG, &db, flags));
2092
2093                 dmu_buf_will_dirty(db, tx);
2094
2095                 ASSERT3U(db->db_size, >=, drro->drr_bonuslen);
2096                 memcpy(db->db_data, data, DRR_OBJECT_PAYLOAD_SIZE(drro));
2097
2098                 /*
2099                  * Raw bonus buffers have their byteorder determined by the
2100                  * DRR_OBJECT_RANGE record.
2101                  */
2102                 if (rwa->byteswap && !rwa->raw) {
2103                         dmu_object_byteswap_t byteswap =
2104                             DMU_OT_BYTESWAP(drro->drr_bonustype);
2105                         dmu_ot_byteswap[byteswap].ob_func(db->db_data,
2106                             DRR_OBJECT_PAYLOAD_SIZE(drro));
2107                 }
2108                 dmu_buf_rele(db, FTAG);
2109                 dnode_rele(dn, FTAG);
2110         }
2111         dmu_tx_commit(tx);
2112
2113         return (0);
2114 }
2115
2116 noinline static int
2117 receive_freeobjects(struct receive_writer_arg *rwa,
2118     struct drr_freeobjects *drrfo)
2119 {
2120         uint64_t obj;
2121         int next_err = 0;
2122
2123         if (drrfo->drr_firstobj + drrfo->drr_numobjs < drrfo->drr_firstobj)
2124                 return (SET_ERROR(EINVAL));
2125
2126         for (obj = drrfo->drr_firstobj == 0 ? 1 : drrfo->drr_firstobj;
2127             obj < drrfo->drr_firstobj + drrfo->drr_numobjs &&
2128             obj < DN_MAX_OBJECT && next_err == 0;
2129             next_err = dmu_object_next(rwa->os, &obj, FALSE, 0)) {
2130                 dmu_object_info_t doi;
2131                 int err;
2132
2133                 err = dmu_object_info(rwa->os, obj, &doi);
2134                 if (err == ENOENT)
2135                         continue;
2136                 else if (err != 0)
2137                         return (err);
2138
2139                 err = dmu_free_long_object(rwa->os, obj);
2140
2141                 if (err != 0)
2142                         return (err);
2143
2144                 if (rwa->or_need_sync == ORNS_MAYBE)
2145                         rwa->or_need_sync = ORNS_YES;
2146         }
2147         if (next_err != ESRCH)
2148                 return (next_err);
2149         return (0);
2150 }
2151
2152 /*
2153  * Note: if this fails, the caller will clean up any records left on the
2154  * rwa->write_batch list.
2155  */
2156 static int
2157 flush_write_batch_impl(struct receive_writer_arg *rwa)
2158 {
2159         dnode_t *dn;
2160         int err;
2161
2162         if (dnode_hold(rwa->os, rwa->last_object, FTAG, &dn) != 0)
2163                 return (SET_ERROR(EINVAL));
2164
2165         struct receive_record_arg *last_rrd = list_tail(&rwa->write_batch);
2166         struct drr_write *last_drrw = &last_rrd->header.drr_u.drr_write;
2167
2168         struct receive_record_arg *first_rrd = list_head(&rwa->write_batch);
2169         struct drr_write *first_drrw = &first_rrd->header.drr_u.drr_write;
2170
2171         ASSERT3U(rwa->last_object, ==, last_drrw->drr_object);
2172         ASSERT3U(rwa->last_offset, ==, last_drrw->drr_offset);
2173
2174         dmu_tx_t *tx = dmu_tx_create(rwa->os);
2175         dmu_tx_hold_write_by_dnode(tx, dn, first_drrw->drr_offset,
2176             last_drrw->drr_offset - first_drrw->drr_offset +
2177             last_drrw->drr_logical_size);
2178         err = dmu_tx_assign(tx, TXG_WAIT);
2179         if (err != 0) {
2180                 dmu_tx_abort(tx);
2181                 dnode_rele(dn, FTAG);
2182                 return (err);
2183         }
2184
2185         struct receive_record_arg *rrd;
2186         while ((rrd = list_head(&rwa->write_batch)) != NULL) {
2187                 struct drr_write *drrw = &rrd->header.drr_u.drr_write;
2188                 abd_t *abd = rrd->abd;
2189
2190                 ASSERT3U(drrw->drr_object, ==, rwa->last_object);
2191
2192                 if (drrw->drr_logical_size != dn->dn_datablksz) {
2193                         /*
2194                          * The WRITE record is larger than the object's block
2195                          * size.  We must be receiving an incremental
2196                          * large-block stream into a dataset that previously did
2197                          * a non-large-block receive.  Lightweight writes must
2198                          * be exactly one block, so we need to decompress the
2199                          * data (if compressed) and do a normal dmu_write().
2200                          */
2201                         ASSERT3U(drrw->drr_logical_size, >, dn->dn_datablksz);
2202                         if (DRR_WRITE_COMPRESSED(drrw)) {
2203                                 abd_t *decomp_abd =
2204                                     abd_alloc_linear(drrw->drr_logical_size,
2205                                     B_FALSE);
2206
2207                                 err = zio_decompress_data(
2208                                     drrw->drr_compressiontype,
2209                                     abd, abd_to_buf(decomp_abd),
2210                                     abd_get_size(abd),
2211                                     abd_get_size(decomp_abd), NULL);
2212
2213                                 if (err == 0) {
2214                                         dmu_write_by_dnode(dn,
2215                                             drrw->drr_offset,
2216                                             drrw->drr_logical_size,
2217                                             abd_to_buf(decomp_abd), tx);
2218                                 }
2219                                 abd_free(decomp_abd);
2220                         } else {
2221                                 dmu_write_by_dnode(dn,
2222                                     drrw->drr_offset,
2223                                     drrw->drr_logical_size,
2224                                     abd_to_buf(abd), tx);
2225                         }
2226                         if (err == 0)
2227                                 abd_free(abd);
2228                 } else {
2229                         zio_prop_t zp = {0};
2230                         dmu_write_policy(rwa->os, dn, 0, 0, &zp);
2231
2232                         zio_flag_t zio_flags = 0;
2233
2234                         if (rwa->raw) {
2235                                 zp.zp_encrypt = B_TRUE;
2236                                 zp.zp_compress = drrw->drr_compressiontype;
2237                                 zp.zp_byteorder = ZFS_HOST_BYTEORDER ^
2238                                     !!DRR_IS_RAW_BYTESWAPPED(drrw->drr_flags) ^
2239                                     rwa->byteswap;
2240                                 memcpy(zp.zp_salt, drrw->drr_salt,
2241                                     ZIO_DATA_SALT_LEN);
2242                                 memcpy(zp.zp_iv, drrw->drr_iv,
2243                                     ZIO_DATA_IV_LEN);
2244                                 memcpy(zp.zp_mac, drrw->drr_mac,
2245                                     ZIO_DATA_MAC_LEN);
2246                                 if (DMU_OT_IS_ENCRYPTED(zp.zp_type)) {
2247                                         zp.zp_nopwrite = B_FALSE;
2248                                         zp.zp_copies = MIN(zp.zp_copies,
2249                                             SPA_DVAS_PER_BP - 1);
2250                                 }
2251                                 zio_flags |= ZIO_FLAG_RAW;
2252                         } else if (DRR_WRITE_COMPRESSED(drrw)) {
2253                                 ASSERT3U(drrw->drr_compressed_size, >, 0);
2254                                 ASSERT3U(drrw->drr_logical_size, >=,
2255                                     drrw->drr_compressed_size);
2256                                 zp.zp_compress = drrw->drr_compressiontype;
2257                                 zio_flags |= ZIO_FLAG_RAW_COMPRESS;
2258                         } else if (rwa->byteswap) {
2259                                 /*
2260                                  * Note: compressed blocks never need to be
2261                                  * byteswapped, because WRITE records for
2262                                  * metadata blocks are never compressed. The
2263                                  * exception is raw streams, which are written
2264                                  * in the original byteorder, and the byteorder
2265                                  * bit is preserved in the BP by setting
2266                                  * zp_byteorder above.
2267                                  */
2268                                 dmu_object_byteswap_t byteswap =
2269                                     DMU_OT_BYTESWAP(drrw->drr_type);
2270                                 dmu_ot_byteswap[byteswap].ob_func(
2271                                     abd_to_buf(abd),
2272                                     DRR_WRITE_PAYLOAD_SIZE(drrw));
2273                         }
2274
2275                         /*
2276                          * Since this data can't be read until the receive
2277                          * completes, we can do a "lightweight" write for
2278                          * improved performance.
2279                          */
2280                         err = dmu_lightweight_write_by_dnode(dn,
2281                             drrw->drr_offset, abd, &zp, zio_flags, tx);
2282                 }
2283
2284                 if (err != 0) {
2285                         /*
2286                          * This rrd is left on the list, so the caller will
2287                          * free it (and the abd).
2288                          */
2289                         break;
2290                 }
2291
2292                 /*
2293                  * Note: If the receive fails, we want the resume stream to
2294                  * start with the same record that we last successfully
2295                  * received (as opposed to the next record), so that we can
2296                  * verify that we are resuming from the correct location.
2297                  */
2298                 save_resume_state(rwa, drrw->drr_object, drrw->drr_offset, tx);
2299
2300                 list_remove(&rwa->write_batch, rrd);
2301                 kmem_free(rrd, sizeof (*rrd));
2302         }
2303
2304         dmu_tx_commit(tx);
2305         dnode_rele(dn, FTAG);
2306         return (err);
2307 }
2308
2309 noinline static int
2310 flush_write_batch(struct receive_writer_arg *rwa)
2311 {
2312         if (list_is_empty(&rwa->write_batch))
2313                 return (0);
2314         int err = rwa->err;
2315         if (err == 0)
2316                 err = flush_write_batch_impl(rwa);
2317         if (err != 0) {
2318                 struct receive_record_arg *rrd;
2319                 while ((rrd = list_remove_head(&rwa->write_batch)) != NULL) {
2320                         abd_free(rrd->abd);
2321                         kmem_free(rrd, sizeof (*rrd));
2322                 }
2323         }
2324         ASSERT(list_is_empty(&rwa->write_batch));
2325         return (err);
2326 }
2327
2328 noinline static int
2329 receive_process_write_record(struct receive_writer_arg *rwa,
2330     struct receive_record_arg *rrd)
2331 {
2332         int err = 0;
2333
2334         ASSERT3U(rrd->header.drr_type, ==, DRR_WRITE);
2335         struct drr_write *drrw = &rrd->header.drr_u.drr_write;
2336
2337         if (drrw->drr_offset + drrw->drr_logical_size < drrw->drr_offset ||
2338             !DMU_OT_IS_VALID(drrw->drr_type))
2339                 return (SET_ERROR(EINVAL));
2340
2341         if (rwa->heal) {
2342                 blkptr_t *bp;
2343                 dmu_buf_t *dbp;
2344                 dnode_t *dn;
2345                 int flags = DB_RF_CANFAIL;
2346
2347                 if (rwa->raw)
2348                         flags |= DB_RF_NO_DECRYPT;
2349
2350                 if (rwa->byteswap) {
2351                         dmu_object_byteswap_t byteswap =
2352                             DMU_OT_BYTESWAP(drrw->drr_type);
2353                         dmu_ot_byteswap[byteswap].ob_func(abd_to_buf(rrd->abd),
2354                             DRR_WRITE_PAYLOAD_SIZE(drrw));
2355                 }
2356
2357                 err = dmu_buf_hold_noread(rwa->os, drrw->drr_object,
2358                     drrw->drr_offset, FTAG, &dbp);
2359                 if (err != 0)
2360                         return (err);
2361
2362                 /* Try to read the object to see if it needs healing */
2363                 err = dbuf_read((dmu_buf_impl_t *)dbp, NULL, flags);
2364                 /*
2365                  * We only try to heal when dbuf_read() returns a ECKSUMs.
2366                  * Other errors (even EIO) get returned to caller.
2367                  * EIO indicates that the device is not present/accessible,
2368                  * so writing to it will likely fail.
2369                  * If the block is healthy, we don't want to overwrite it
2370                  * unnecessarily.
2371                  */
2372                 if (err != ECKSUM) {
2373                         dmu_buf_rele(dbp, FTAG);
2374                         return (err);
2375                 }
2376                 dn = dmu_buf_dnode_enter(dbp);
2377                 /* Make sure the on-disk block and recv record sizes match */
2378                 if (drrw->drr_logical_size !=
2379                     dn->dn_datablkszsec << SPA_MINBLOCKSHIFT) {
2380                         err = ENOTSUP;
2381                         dmu_buf_dnode_exit(dbp);
2382                         dmu_buf_rele(dbp, FTAG);
2383                         return (err);
2384                 }
2385                 /* Get the block pointer for the corrupted block */
2386                 bp = dmu_buf_get_blkptr(dbp);
2387                 err = do_corrective_recv(rwa, drrw, rrd, bp);
2388                 dmu_buf_dnode_exit(dbp);
2389                 dmu_buf_rele(dbp, FTAG);
2390                 return (err);
2391         }
2392
2393         /*
2394          * For resuming to work, records must be in increasing order
2395          * by (object, offset).
2396          */
2397         if (drrw->drr_object < rwa->last_object ||
2398             (drrw->drr_object == rwa->last_object &&
2399             drrw->drr_offset < rwa->last_offset)) {
2400                 return (SET_ERROR(EINVAL));
2401         }
2402
2403         struct receive_record_arg *first_rrd = list_head(&rwa->write_batch);
2404         struct drr_write *first_drrw = &first_rrd->header.drr_u.drr_write;
2405         uint64_t batch_size =
2406             MIN(zfs_recv_write_batch_size, DMU_MAX_ACCESS / 2);
2407         if (first_rrd != NULL &&
2408             (drrw->drr_object != first_drrw->drr_object ||
2409             drrw->drr_offset >= first_drrw->drr_offset + batch_size)) {
2410                 err = flush_write_batch(rwa);
2411                 if (err != 0)
2412                         return (err);
2413         }
2414
2415         rwa->last_object = drrw->drr_object;
2416         rwa->last_offset = drrw->drr_offset;
2417
2418         if (rwa->last_object > rwa->max_object)
2419                 rwa->max_object = rwa->last_object;
2420
2421         list_insert_tail(&rwa->write_batch, rrd);
2422         /*
2423          * Return EAGAIN to indicate that we will use this rrd again,
2424          * so the caller should not free it
2425          */
2426         return (EAGAIN);
2427 }
2428
2429 static int
2430 receive_write_embedded(struct receive_writer_arg *rwa,
2431     struct drr_write_embedded *drrwe, void *data)
2432 {
2433         dmu_tx_t *tx;
2434         int err;
2435
2436         if (drrwe->drr_offset + drrwe->drr_length < drrwe->drr_offset)
2437                 return (SET_ERROR(EINVAL));
2438
2439         if (drrwe->drr_psize > BPE_PAYLOAD_SIZE)
2440                 return (SET_ERROR(EINVAL));
2441
2442         if (drrwe->drr_etype >= NUM_BP_EMBEDDED_TYPES)
2443                 return (SET_ERROR(EINVAL));
2444         if (drrwe->drr_compression >= ZIO_COMPRESS_FUNCTIONS)
2445                 return (SET_ERROR(EINVAL));
2446         if (rwa->raw)
2447                 return (SET_ERROR(EINVAL));
2448
2449         if (drrwe->drr_object > rwa->max_object)
2450                 rwa->max_object = drrwe->drr_object;
2451
2452         tx = dmu_tx_create(rwa->os);
2453
2454         dmu_tx_hold_write(tx, drrwe->drr_object,
2455             drrwe->drr_offset, drrwe->drr_length);
2456         err = dmu_tx_assign(tx, TXG_WAIT);
2457         if (err != 0) {
2458                 dmu_tx_abort(tx);
2459                 return (err);
2460         }
2461
2462         dmu_write_embedded(rwa->os, drrwe->drr_object,
2463             drrwe->drr_offset, data, drrwe->drr_etype,
2464             drrwe->drr_compression, drrwe->drr_lsize, drrwe->drr_psize,
2465             rwa->byteswap ^ ZFS_HOST_BYTEORDER, tx);
2466
2467         /* See comment in restore_write. */
2468         save_resume_state(rwa, drrwe->drr_object, drrwe->drr_offset, tx);
2469         dmu_tx_commit(tx);
2470         return (0);
2471 }
2472
2473 static int
2474 receive_spill(struct receive_writer_arg *rwa, struct drr_spill *drrs,
2475     abd_t *abd)
2476 {
2477         dmu_buf_t *db, *db_spill;
2478         int err;
2479
2480         if (drrs->drr_length < SPA_MINBLOCKSIZE ||
2481             drrs->drr_length > spa_maxblocksize(dmu_objset_spa(rwa->os)))
2482                 return (SET_ERROR(EINVAL));
2483
2484         /*
2485          * This is an unmodified spill block which was added to the stream
2486          * to resolve an issue with incorrectly removing spill blocks.  It
2487          * should be ignored by current versions of the code which support
2488          * the DRR_FLAG_SPILL_BLOCK flag.
2489          */
2490         if (rwa->spill && DRR_SPILL_IS_UNMODIFIED(drrs->drr_flags)) {
2491                 abd_free(abd);
2492                 return (0);
2493         }
2494
2495         if (rwa->raw) {
2496                 if (!DMU_OT_IS_VALID(drrs->drr_type) ||
2497                     drrs->drr_compressiontype >= ZIO_COMPRESS_FUNCTIONS ||
2498                     drrs->drr_compressed_size == 0)
2499                         return (SET_ERROR(EINVAL));
2500         }
2501
2502         if (dmu_object_info(rwa->os, drrs->drr_object, NULL) != 0)
2503                 return (SET_ERROR(EINVAL));
2504
2505         if (drrs->drr_object > rwa->max_object)
2506                 rwa->max_object = drrs->drr_object;
2507
2508         VERIFY0(dmu_bonus_hold(rwa->os, drrs->drr_object, FTAG, &db));
2509         if ((err = dmu_spill_hold_by_bonus(db, DMU_READ_NO_DECRYPT, FTAG,
2510             &db_spill)) != 0) {
2511                 dmu_buf_rele(db, FTAG);
2512                 return (err);
2513         }
2514
2515         dmu_tx_t *tx = dmu_tx_create(rwa->os);
2516
2517         dmu_tx_hold_spill(tx, db->db_object);
2518
2519         err = dmu_tx_assign(tx, TXG_WAIT);
2520         if (err != 0) {
2521                 dmu_buf_rele(db, FTAG);
2522                 dmu_buf_rele(db_spill, FTAG);
2523                 dmu_tx_abort(tx);
2524                 return (err);
2525         }
2526
2527         /*
2528          * Spill blocks may both grow and shrink.  When a change in size
2529          * occurs any existing dbuf must be updated to match the logical
2530          * size of the provided arc_buf_t.
2531          */
2532         if (db_spill->db_size != drrs->drr_length) {
2533                 dmu_buf_will_fill(db_spill, tx);
2534                 VERIFY0(dbuf_spill_set_blksz(db_spill,
2535                     drrs->drr_length, tx));
2536         }
2537
2538         arc_buf_t *abuf;
2539         if (rwa->raw) {
2540                 boolean_t byteorder = ZFS_HOST_BYTEORDER ^
2541                     !!DRR_IS_RAW_BYTESWAPPED(drrs->drr_flags) ^
2542                     rwa->byteswap;
2543
2544                 abuf = arc_loan_raw_buf(dmu_objset_spa(rwa->os),
2545                     drrs->drr_object, byteorder, drrs->drr_salt,
2546                     drrs->drr_iv, drrs->drr_mac, drrs->drr_type,
2547                     drrs->drr_compressed_size, drrs->drr_length,
2548                     drrs->drr_compressiontype, 0);
2549         } else {
2550                 abuf = arc_loan_buf(dmu_objset_spa(rwa->os),
2551                     DMU_OT_IS_METADATA(drrs->drr_type),
2552                     drrs->drr_length);
2553                 if (rwa->byteswap) {
2554                         dmu_object_byteswap_t byteswap =
2555                             DMU_OT_BYTESWAP(drrs->drr_type);
2556                         dmu_ot_byteswap[byteswap].ob_func(abd_to_buf(abd),
2557                             DRR_SPILL_PAYLOAD_SIZE(drrs));
2558                 }
2559         }
2560
2561         memcpy(abuf->b_data, abd_to_buf(abd), DRR_SPILL_PAYLOAD_SIZE(drrs));
2562         abd_free(abd);
2563         dbuf_assign_arcbuf((dmu_buf_impl_t *)db_spill, abuf, tx);
2564
2565         dmu_buf_rele(db, FTAG);
2566         dmu_buf_rele(db_spill, FTAG);
2567
2568         dmu_tx_commit(tx);
2569         return (0);
2570 }
2571
2572 noinline static int
2573 receive_free(struct receive_writer_arg *rwa, struct drr_free *drrf)
2574 {
2575         int err;
2576
2577         if (drrf->drr_length != -1ULL &&
2578             drrf->drr_offset + drrf->drr_length < drrf->drr_offset)
2579                 return (SET_ERROR(EINVAL));
2580
2581         if (dmu_object_info(rwa->os, drrf->drr_object, NULL) != 0)
2582                 return (SET_ERROR(EINVAL));
2583
2584         if (drrf->drr_object > rwa->max_object)
2585                 rwa->max_object = drrf->drr_object;
2586
2587         err = dmu_free_long_range(rwa->os, drrf->drr_object,
2588             drrf->drr_offset, drrf->drr_length);
2589
2590         return (err);
2591 }
2592
2593 static int
2594 receive_object_range(struct receive_writer_arg *rwa,
2595     struct drr_object_range *drror)
2596 {
2597         /*
2598          * By default, we assume this block is in our native format
2599          * (ZFS_HOST_BYTEORDER). We then take into account whether
2600          * the send stream is byteswapped (rwa->byteswap). Finally,
2601          * we need to byteswap again if this particular block was
2602          * in non-native format on the send side.
2603          */
2604         boolean_t byteorder = ZFS_HOST_BYTEORDER ^ rwa->byteswap ^
2605             !!DRR_IS_RAW_BYTESWAPPED(drror->drr_flags);
2606
2607         /*
2608          * Since dnode block sizes are constant, we should not need to worry
2609          * about making sure that the dnode block size is the same on the
2610          * sending and receiving sides for the time being. For non-raw sends,
2611          * this does not matter (and in fact we do not send a DRR_OBJECT_RANGE
2612          * record at all). Raw sends require this record type because the
2613          * encryption parameters are used to protect an entire block of bonus
2614          * buffers. If the size of dnode blocks ever becomes variable,
2615          * handling will need to be added to ensure that dnode block sizes
2616          * match on the sending and receiving side.
2617          */
2618         if (drror->drr_numslots != DNODES_PER_BLOCK ||
2619             P2PHASE(drror->drr_firstobj, DNODES_PER_BLOCK) != 0 ||
2620             !rwa->raw)
2621                 return (SET_ERROR(EINVAL));
2622
2623         if (drror->drr_firstobj > rwa->max_object)
2624                 rwa->max_object = drror->drr_firstobj;
2625
2626         /*
2627          * The DRR_OBJECT_RANGE handling must be deferred to receive_object()
2628          * so that the block of dnodes is not written out when it's empty,
2629          * and converted to a HOLE BP.
2630          */
2631         rwa->or_crypt_params_present = B_TRUE;
2632         rwa->or_firstobj = drror->drr_firstobj;
2633         rwa->or_numslots = drror->drr_numslots;
2634         memcpy(rwa->or_salt, drror->drr_salt, ZIO_DATA_SALT_LEN);
2635         memcpy(rwa->or_iv, drror->drr_iv, ZIO_DATA_IV_LEN);
2636         memcpy(rwa->or_mac, drror->drr_mac, ZIO_DATA_MAC_LEN);
2637         rwa->or_byteorder = byteorder;
2638
2639         rwa->or_need_sync = ORNS_MAYBE;
2640
2641         return (0);
2642 }
2643
2644 /*
2645  * Until we have the ability to redact large ranges of data efficiently, we
2646  * process these records as frees.
2647  */
2648 noinline static int
2649 receive_redact(struct receive_writer_arg *rwa, struct drr_redact *drrr)
2650 {
2651         struct drr_free drrf = {0};
2652         drrf.drr_length = drrr->drr_length;
2653         drrf.drr_object = drrr->drr_object;
2654         drrf.drr_offset = drrr->drr_offset;
2655         drrf.drr_toguid = drrr->drr_toguid;
2656         return (receive_free(rwa, &drrf));
2657 }
2658
2659 /* used to destroy the drc_ds on error */
2660 static void
2661 dmu_recv_cleanup_ds(dmu_recv_cookie_t *drc)
2662 {
2663         dsl_dataset_t *ds = drc->drc_ds;
2664         ds_hold_flags_t dsflags;
2665
2666         dsflags = (drc->drc_raw) ? DS_HOLD_FLAG_NONE : DS_HOLD_FLAG_DECRYPT;
2667         /*
2668          * Wait for the txg sync before cleaning up the receive. For
2669          * resumable receives, this ensures that our resume state has
2670          * been written out to disk. For raw receives, this ensures
2671          * that the user accounting code will not attempt to do anything
2672          * after we stopped receiving the dataset.
2673          */
2674         txg_wait_synced(ds->ds_dir->dd_pool, 0);
2675         ds->ds_objset->os_raw_receive = B_FALSE;
2676
2677         rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
2678         if (drc->drc_resumable && drc->drc_should_save &&
2679             !BP_IS_HOLE(dsl_dataset_get_blkptr(ds))) {
2680                 rrw_exit(&ds->ds_bp_rwlock, FTAG);
2681                 dsl_dataset_disown(ds, dsflags, dmu_recv_tag);
2682         } else {
2683                 char name[ZFS_MAX_DATASET_NAME_LEN];
2684                 rrw_exit(&ds->ds_bp_rwlock, FTAG);
2685                 dsl_dataset_name(ds, name);
2686                 dsl_dataset_disown(ds, dsflags, dmu_recv_tag);
2687                 if (!drc->drc_heal)
2688                         (void) dsl_destroy_head(name);
2689         }
2690 }
2691
2692 static void
2693 receive_cksum(dmu_recv_cookie_t *drc, int len, void *buf)
2694 {
2695         if (drc->drc_byteswap) {
2696                 (void) fletcher_4_incremental_byteswap(buf, len,
2697                     &drc->drc_cksum);
2698         } else {
2699                 (void) fletcher_4_incremental_native(buf, len, &drc->drc_cksum);
2700         }
2701 }
2702
2703 /*
2704  * Read the payload into a buffer of size len, and update the current record's
2705  * payload field.
2706  * Allocate drc->drc_next_rrd and read the next record's header into
2707  * drc->drc_next_rrd->header.
2708  * Verify checksum of payload and next record.
2709  */
2710 static int
2711 receive_read_payload_and_next_header(dmu_recv_cookie_t *drc, int len, void *buf)
2712 {
2713         int err;
2714
2715         if (len != 0) {
2716                 ASSERT3U(len, <=, SPA_MAXBLOCKSIZE);
2717                 err = receive_read(drc, len, buf);
2718                 if (err != 0)
2719                         return (err);
2720                 receive_cksum(drc, len, buf);
2721
2722                 /* note: rrd is NULL when reading the begin record's payload */
2723                 if (drc->drc_rrd != NULL) {
2724                         drc->drc_rrd->payload = buf;
2725                         drc->drc_rrd->payload_size = len;
2726                         drc->drc_rrd->bytes_read = drc->drc_bytes_read;
2727                 }
2728         } else {
2729                 ASSERT3P(buf, ==, NULL);
2730         }
2731
2732         drc->drc_prev_cksum = drc->drc_cksum;
2733
2734         drc->drc_next_rrd = kmem_zalloc(sizeof (*drc->drc_next_rrd), KM_SLEEP);
2735         err = receive_read(drc, sizeof (drc->drc_next_rrd->header),
2736             &drc->drc_next_rrd->header);
2737         drc->drc_next_rrd->bytes_read = drc->drc_bytes_read;
2738
2739         if (err != 0) {
2740                 kmem_free(drc->drc_next_rrd, sizeof (*drc->drc_next_rrd));
2741                 drc->drc_next_rrd = NULL;
2742                 return (err);
2743         }
2744         if (drc->drc_next_rrd->header.drr_type == DRR_BEGIN) {
2745                 kmem_free(drc->drc_next_rrd, sizeof (*drc->drc_next_rrd));
2746                 drc->drc_next_rrd = NULL;
2747                 return (SET_ERROR(EINVAL));
2748         }
2749
2750         /*
2751          * Note: checksum is of everything up to but not including the
2752          * checksum itself.
2753          */
2754         ASSERT3U(offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
2755             ==, sizeof (dmu_replay_record_t) - sizeof (zio_cksum_t));
2756         receive_cksum(drc,
2757             offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
2758             &drc->drc_next_rrd->header);
2759
2760         zio_cksum_t cksum_orig =
2761             drc->drc_next_rrd->header.drr_u.drr_checksum.drr_checksum;
2762         zio_cksum_t *cksump =
2763             &drc->drc_next_rrd->header.drr_u.drr_checksum.drr_checksum;
2764
2765         if (drc->drc_byteswap)
2766                 byteswap_record(&drc->drc_next_rrd->header);
2767
2768         if ((!ZIO_CHECKSUM_IS_ZERO(cksump)) &&
2769             !ZIO_CHECKSUM_EQUAL(drc->drc_cksum, *cksump)) {
2770                 kmem_free(drc->drc_next_rrd, sizeof (*drc->drc_next_rrd));
2771                 drc->drc_next_rrd = NULL;
2772                 return (SET_ERROR(ECKSUM));
2773         }
2774
2775         receive_cksum(drc, sizeof (cksum_orig), &cksum_orig);
2776
2777         return (0);
2778 }
2779
2780 /*
2781  * Issue the prefetch reads for any necessary indirect blocks.
2782  *
2783  * We use the object ignore list to tell us whether or not to issue prefetches
2784  * for a given object.  We do this for both correctness (in case the blocksize
2785  * of an object has changed) and performance (if the object doesn't exist, don't
2786  * needlessly try to issue prefetches).  We also trim the list as we go through
2787  * the stream to prevent it from growing to an unbounded size.
2788  *
2789  * The object numbers within will always be in sorted order, and any write
2790  * records we see will also be in sorted order, but they're not sorted with
2791  * respect to each other (i.e. we can get several object records before
2792  * receiving each object's write records).  As a result, once we've reached a
2793  * given object number, we can safely remove any reference to lower object
2794  * numbers in the ignore list. In practice, we receive up to 32 object records
2795  * before receiving write records, so the list can have up to 32 nodes in it.
2796  */
2797 static void
2798 receive_read_prefetch(dmu_recv_cookie_t *drc, uint64_t object, uint64_t offset,
2799     uint64_t length)
2800 {
2801         if (!objlist_exists(drc->drc_ignore_objlist, object)) {
2802                 dmu_prefetch(drc->drc_os, object, 1, offset, length,
2803                     ZIO_PRIORITY_SYNC_READ);
2804         }
2805 }
2806
2807 /*
2808  * Read records off the stream, issuing any necessary prefetches.
2809  */
2810 static int
2811 receive_read_record(dmu_recv_cookie_t *drc)
2812 {
2813         int err;
2814
2815         switch (drc->drc_rrd->header.drr_type) {
2816         case DRR_OBJECT:
2817         {
2818                 struct drr_object *drro =
2819                     &drc->drc_rrd->header.drr_u.drr_object;
2820                 uint32_t size = DRR_OBJECT_PAYLOAD_SIZE(drro);
2821                 void *buf = NULL;
2822                 dmu_object_info_t doi;
2823
2824                 if (size != 0)
2825                         buf = kmem_zalloc(size, KM_SLEEP);
2826
2827                 err = receive_read_payload_and_next_header(drc, size, buf);
2828                 if (err != 0) {
2829                         kmem_free(buf, size);
2830                         return (err);
2831                 }
2832                 err = dmu_object_info(drc->drc_os, drro->drr_object, &doi);
2833                 /*
2834                  * See receive_read_prefetch for an explanation why we're
2835                  * storing this object in the ignore_obj_list.
2836                  */
2837                 if (err == ENOENT || err == EEXIST ||
2838                     (err == 0 && doi.doi_data_block_size != drro->drr_blksz)) {
2839                         objlist_insert(drc->drc_ignore_objlist,
2840                             drro->drr_object);
2841                         err = 0;
2842                 }
2843                 return (err);
2844         }
2845         case DRR_FREEOBJECTS:
2846         {
2847                 err = receive_read_payload_and_next_header(drc, 0, NULL);
2848                 return (err);
2849         }
2850         case DRR_WRITE:
2851         {
2852                 struct drr_write *drrw = &drc->drc_rrd->header.drr_u.drr_write;
2853                 int size = DRR_WRITE_PAYLOAD_SIZE(drrw);
2854                 abd_t *abd = abd_alloc_linear(size, B_FALSE);
2855                 err = receive_read_payload_and_next_header(drc, size,
2856                     abd_to_buf(abd));
2857                 if (err != 0) {
2858                         abd_free(abd);
2859                         return (err);
2860                 }
2861                 drc->drc_rrd->abd = abd;
2862                 receive_read_prefetch(drc, drrw->drr_object, drrw->drr_offset,
2863                     drrw->drr_logical_size);
2864                 return (err);
2865         }
2866         case DRR_WRITE_EMBEDDED:
2867         {
2868                 struct drr_write_embedded *drrwe =
2869                     &drc->drc_rrd->header.drr_u.drr_write_embedded;
2870                 uint32_t size = P2ROUNDUP(drrwe->drr_psize, 8);
2871                 void *buf = kmem_zalloc(size, KM_SLEEP);
2872
2873                 err = receive_read_payload_and_next_header(drc, size, buf);
2874                 if (err != 0) {
2875                         kmem_free(buf, size);
2876                         return (err);
2877                 }
2878
2879                 receive_read_prefetch(drc, drrwe->drr_object, drrwe->drr_offset,
2880                     drrwe->drr_length);
2881                 return (err);
2882         }
2883         case DRR_FREE:
2884         case DRR_REDACT:
2885         {
2886                 /*
2887                  * It might be beneficial to prefetch indirect blocks here, but
2888                  * we don't really have the data to decide for sure.
2889                  */
2890                 err = receive_read_payload_and_next_header(drc, 0, NULL);
2891                 return (err);
2892         }
2893         case DRR_END:
2894         {
2895                 struct drr_end *drre = &drc->drc_rrd->header.drr_u.drr_end;
2896                 if (!ZIO_CHECKSUM_EQUAL(drc->drc_prev_cksum,
2897                     drre->drr_checksum))
2898                         return (SET_ERROR(ECKSUM));
2899                 return (0);
2900         }
2901         case DRR_SPILL:
2902         {
2903                 struct drr_spill *drrs = &drc->drc_rrd->header.drr_u.drr_spill;
2904                 int size = DRR_SPILL_PAYLOAD_SIZE(drrs);
2905                 abd_t *abd = abd_alloc_linear(size, B_FALSE);
2906                 err = receive_read_payload_and_next_header(drc, size,
2907                     abd_to_buf(abd));
2908                 if (err != 0)
2909                         abd_free(abd);
2910                 else
2911                         drc->drc_rrd->abd = abd;
2912                 return (err);
2913         }
2914         case DRR_OBJECT_RANGE:
2915         {
2916                 err = receive_read_payload_and_next_header(drc, 0, NULL);
2917                 return (err);
2918
2919         }
2920         default:
2921                 return (SET_ERROR(EINVAL));
2922         }
2923 }
2924
2925
2926
2927 static void
2928 dprintf_drr(struct receive_record_arg *rrd, int err)
2929 {
2930 #ifdef ZFS_DEBUG
2931         switch (rrd->header.drr_type) {
2932         case DRR_OBJECT:
2933         {
2934                 struct drr_object *drro = &rrd->header.drr_u.drr_object;
2935                 dprintf("drr_type = OBJECT obj = %llu type = %u "
2936                     "bonustype = %u blksz = %u bonuslen = %u cksumtype = %u "
2937                     "compress = %u dn_slots = %u err = %d\n",
2938                     (u_longlong_t)drro->drr_object, drro->drr_type,
2939                     drro->drr_bonustype, drro->drr_blksz, drro->drr_bonuslen,
2940                     drro->drr_checksumtype, drro->drr_compress,
2941                     drro->drr_dn_slots, err);
2942                 break;
2943         }
2944         case DRR_FREEOBJECTS:
2945         {
2946                 struct drr_freeobjects *drrfo =
2947                     &rrd->header.drr_u.drr_freeobjects;
2948                 dprintf("drr_type = FREEOBJECTS firstobj = %llu "
2949                     "numobjs = %llu err = %d\n",
2950                     (u_longlong_t)drrfo->drr_firstobj,
2951                     (u_longlong_t)drrfo->drr_numobjs, err);
2952                 break;
2953         }
2954         case DRR_WRITE:
2955         {
2956                 struct drr_write *drrw = &rrd->header.drr_u.drr_write;
2957                 dprintf("drr_type = WRITE obj = %llu type = %u offset = %llu "
2958                     "lsize = %llu cksumtype = %u flags = %u "
2959                     "compress = %u psize = %llu err = %d\n",
2960                     (u_longlong_t)drrw->drr_object, drrw->drr_type,
2961                     (u_longlong_t)drrw->drr_offset,
2962                     (u_longlong_t)drrw->drr_logical_size,
2963                     drrw->drr_checksumtype, drrw->drr_flags,
2964                     drrw->drr_compressiontype,
2965                     (u_longlong_t)drrw->drr_compressed_size, err);
2966                 break;
2967         }
2968         case DRR_WRITE_BYREF:
2969         {
2970                 struct drr_write_byref *drrwbr =
2971                     &rrd->header.drr_u.drr_write_byref;
2972                 dprintf("drr_type = WRITE_BYREF obj = %llu offset = %llu "
2973                     "length = %llu toguid = %llx refguid = %llx "
2974                     "refobject = %llu refoffset = %llu cksumtype = %u "
2975                     "flags = %u err = %d\n",
2976                     (u_longlong_t)drrwbr->drr_object,
2977                     (u_longlong_t)drrwbr->drr_offset,
2978                     (u_longlong_t)drrwbr->drr_length,
2979                     (u_longlong_t)drrwbr->drr_toguid,
2980                     (u_longlong_t)drrwbr->drr_refguid,
2981                     (u_longlong_t)drrwbr->drr_refobject,
2982                     (u_longlong_t)drrwbr->drr_refoffset,
2983                     drrwbr->drr_checksumtype, drrwbr->drr_flags, err);
2984                 break;
2985         }
2986         case DRR_WRITE_EMBEDDED:
2987         {
2988                 struct drr_write_embedded *drrwe =
2989                     &rrd->header.drr_u.drr_write_embedded;
2990                 dprintf("drr_type = WRITE_EMBEDDED obj = %llu offset = %llu "
2991                     "length = %llu compress = %u etype = %u lsize = %u "
2992                     "psize = %u err = %d\n",
2993                     (u_longlong_t)drrwe->drr_object,
2994                     (u_longlong_t)drrwe->drr_offset,
2995                     (u_longlong_t)drrwe->drr_length,
2996                     drrwe->drr_compression, drrwe->drr_etype,
2997                     drrwe->drr_lsize, drrwe->drr_psize, err);
2998                 break;
2999         }
3000         case DRR_FREE:
3001         {
3002                 struct drr_free *drrf = &rrd->header.drr_u.drr_free;
3003                 dprintf("drr_type = FREE obj = %llu offset = %llu "
3004                     "length = %lld err = %d\n",
3005                     (u_longlong_t)drrf->drr_object,
3006                     (u_longlong_t)drrf->drr_offset,
3007                     (longlong_t)drrf->drr_length,
3008                     err);
3009                 break;
3010         }
3011         case DRR_SPILL:
3012         {
3013                 struct drr_spill *drrs = &rrd->header.drr_u.drr_spill;
3014                 dprintf("drr_type = SPILL obj = %llu length = %llu "
3015                     "err = %d\n", (u_longlong_t)drrs->drr_object,
3016                     (u_longlong_t)drrs->drr_length, err);
3017                 break;
3018         }
3019         case DRR_OBJECT_RANGE:
3020         {
3021                 struct drr_object_range *drror =
3022                     &rrd->header.drr_u.drr_object_range;
3023                 dprintf("drr_type = OBJECT_RANGE firstobj = %llu "
3024                     "numslots = %llu flags = %u err = %d\n",
3025                     (u_longlong_t)drror->drr_firstobj,
3026                     (u_longlong_t)drror->drr_numslots,
3027                     drror->drr_flags, err);
3028                 break;
3029         }
3030         default:
3031                 return;
3032         }
3033 #endif
3034 }
3035
3036 /*
3037  * Commit the records to the pool.
3038  */
3039 static int
3040 receive_process_record(struct receive_writer_arg *rwa,
3041     struct receive_record_arg *rrd)
3042 {
3043         int err;
3044
3045         /* Processing in order, therefore bytes_read should be increasing. */
3046         ASSERT3U(rrd->bytes_read, >=, rwa->bytes_read);
3047         rwa->bytes_read = rrd->bytes_read;
3048
3049         /* We can only heal write records; other ones get ignored */
3050         if (rwa->heal && rrd->header.drr_type != DRR_WRITE) {
3051                 if (rrd->abd != NULL) {
3052                         abd_free(rrd->abd);
3053                         rrd->abd = NULL;
3054                 } else if (rrd->payload != NULL) {
3055                         kmem_free(rrd->payload, rrd->payload_size);
3056                         rrd->payload = NULL;
3057                 }
3058                 return (0);
3059         }
3060
3061         if (!rwa->heal && rrd->header.drr_type != DRR_WRITE) {
3062                 err = flush_write_batch(rwa);
3063                 if (err != 0) {
3064                         if (rrd->abd != NULL) {
3065                                 abd_free(rrd->abd);
3066                                 rrd->abd = NULL;
3067                                 rrd->payload = NULL;
3068                         } else if (rrd->payload != NULL) {
3069                                 kmem_free(rrd->payload, rrd->payload_size);
3070                                 rrd->payload = NULL;
3071                         }
3072
3073                         return (err);
3074                 }
3075         }
3076
3077         switch (rrd->header.drr_type) {
3078         case DRR_OBJECT:
3079         {
3080                 struct drr_object *drro = &rrd->header.drr_u.drr_object;
3081                 err = receive_object(rwa, drro, rrd->payload);
3082                 kmem_free(rrd->payload, rrd->payload_size);
3083                 rrd->payload = NULL;
3084                 break;
3085         }
3086         case DRR_FREEOBJECTS:
3087         {
3088                 struct drr_freeobjects *drrfo =
3089                     &rrd->header.drr_u.drr_freeobjects;
3090                 err = receive_freeobjects(rwa, drrfo);
3091                 break;
3092         }
3093         case DRR_WRITE:
3094         {
3095                 err = receive_process_write_record(rwa, rrd);
3096                 if (rwa->heal) {
3097                         /*
3098                          * If healing - always free the abd after processing
3099                          */
3100                         abd_free(rrd->abd);
3101                         rrd->abd = NULL;
3102                 } else if (err != EAGAIN) {
3103                         /*
3104                          * On success, a non-healing
3105                          * receive_process_write_record() returns
3106                          * EAGAIN to indicate that we do not want to free
3107                          * the rrd or arc_buf.
3108                          */
3109                         ASSERT(err != 0);
3110                         abd_free(rrd->abd);
3111                         rrd->abd = NULL;
3112                 }
3113                 break;
3114         }
3115         case DRR_WRITE_EMBEDDED:
3116         {
3117                 struct drr_write_embedded *drrwe =
3118                     &rrd->header.drr_u.drr_write_embedded;
3119                 err = receive_write_embedded(rwa, drrwe, rrd->payload);
3120                 kmem_free(rrd->payload, rrd->payload_size);
3121                 rrd->payload = NULL;
3122                 break;
3123         }
3124         case DRR_FREE:
3125         {
3126                 struct drr_free *drrf = &rrd->header.drr_u.drr_free;
3127                 err = receive_free(rwa, drrf);
3128                 break;
3129         }
3130         case DRR_SPILL:
3131         {
3132                 struct drr_spill *drrs = &rrd->header.drr_u.drr_spill;
3133                 err = receive_spill(rwa, drrs, rrd->abd);
3134                 if (err != 0)
3135                         abd_free(rrd->abd);
3136                 rrd->abd = NULL;
3137                 rrd->payload = NULL;
3138                 break;
3139         }
3140         case DRR_OBJECT_RANGE:
3141         {
3142                 struct drr_object_range *drror =
3143                     &rrd->header.drr_u.drr_object_range;
3144                 err = receive_object_range(rwa, drror);
3145                 break;
3146         }
3147         case DRR_REDACT:
3148         {
3149                 struct drr_redact *drrr = &rrd->header.drr_u.drr_redact;
3150                 err = receive_redact(rwa, drrr);
3151                 break;
3152         }
3153         default:
3154                 err = (SET_ERROR(EINVAL));
3155         }
3156
3157         if (err != 0)
3158                 dprintf_drr(rrd, err);
3159
3160         return (err);
3161 }
3162
3163 /*
3164  * dmu_recv_stream's worker thread; pull records off the queue, and then call
3165  * receive_process_record  When we're done, signal the main thread and exit.
3166  */
3167 static __attribute__((noreturn)) void
3168 receive_writer_thread(void *arg)
3169 {
3170         struct receive_writer_arg *rwa = arg;
3171         struct receive_record_arg *rrd;
3172         fstrans_cookie_t cookie = spl_fstrans_mark();
3173
3174         for (rrd = bqueue_dequeue(&rwa->q); !rrd->eos_marker;
3175             rrd = bqueue_dequeue(&rwa->q)) {
3176                 /*
3177                  * If there's an error, the main thread will stop putting things
3178                  * on the queue, but we need to clear everything in it before we
3179                  * can exit.
3180                  */
3181                 int err = 0;
3182                 if (rwa->err == 0) {
3183                         err = receive_process_record(rwa, rrd);
3184                 } else if (rrd->abd != NULL) {
3185                         abd_free(rrd->abd);
3186                         rrd->abd = NULL;
3187                         rrd->payload = NULL;
3188                 } else if (rrd->payload != NULL) {
3189                         kmem_free(rrd->payload, rrd->payload_size);
3190                         rrd->payload = NULL;
3191                 }
3192                 /*
3193                  * EAGAIN indicates that this record has been saved (on
3194                  * raw->write_batch), and will be used again, so we don't
3195                  * free it.
3196                  * When healing data we always need to free the record.
3197                  */
3198                 if (err != EAGAIN || rwa->heal) {
3199                         if (rwa->err == 0)
3200                                 rwa->err = err;
3201                         kmem_free(rrd, sizeof (*rrd));
3202                 }
3203         }
3204         kmem_free(rrd, sizeof (*rrd));
3205
3206         if (rwa->heal) {
3207                 zio_wait(rwa->heal_pio);
3208         } else {
3209                 int err = flush_write_batch(rwa);
3210                 if (rwa->err == 0)
3211                         rwa->err = err;
3212         }
3213         mutex_enter(&rwa->mutex);
3214         rwa->done = B_TRUE;
3215         cv_signal(&rwa->cv);
3216         mutex_exit(&rwa->mutex);
3217         spl_fstrans_unmark(cookie);
3218         thread_exit();
3219 }
3220
3221 static int
3222 resume_check(dmu_recv_cookie_t *drc, nvlist_t *begin_nvl)
3223 {
3224         uint64_t val;
3225         objset_t *mos = dmu_objset_pool(drc->drc_os)->dp_meta_objset;
3226         uint64_t dsobj = dmu_objset_id(drc->drc_os);
3227         uint64_t resume_obj, resume_off;
3228
3229         if (nvlist_lookup_uint64(begin_nvl,
3230             "resume_object", &resume_obj) != 0 ||
3231             nvlist_lookup_uint64(begin_nvl,
3232             "resume_offset", &resume_off) != 0) {
3233                 return (SET_ERROR(EINVAL));
3234         }
3235         VERIFY0(zap_lookup(mos, dsobj,
3236             DS_FIELD_RESUME_OBJECT, sizeof (val), 1, &val));
3237         if (resume_obj != val)
3238                 return (SET_ERROR(EINVAL));
3239         VERIFY0(zap_lookup(mos, dsobj,
3240             DS_FIELD_RESUME_OFFSET, sizeof (val), 1, &val));
3241         if (resume_off != val)
3242                 return (SET_ERROR(EINVAL));
3243
3244         return (0);
3245 }
3246
3247 /*
3248  * Read in the stream's records, one by one, and apply them to the pool.  There
3249  * are two threads involved; the thread that calls this function will spin up a
3250  * worker thread, read the records off the stream one by one, and issue
3251  * prefetches for any necessary indirect blocks.  It will then push the records
3252  * onto an internal blocking queue.  The worker thread will pull the records off
3253  * the queue, and actually write the data into the DMU.  This way, the worker
3254  * thread doesn't have to wait for reads to complete, since everything it needs
3255  * (the indirect blocks) will be prefetched.
3256  *
3257  * NB: callers *must* call dmu_recv_end() if this succeeds.
3258  */
3259 int
3260 dmu_recv_stream(dmu_recv_cookie_t *drc, offset_t *voffp)
3261 {
3262         int err = 0;
3263         struct receive_writer_arg *rwa = kmem_zalloc(sizeof (*rwa), KM_SLEEP);
3264
3265         if (dsl_dataset_has_resume_receive_state(drc->drc_ds)) {
3266                 uint64_t bytes = 0;
3267                 (void) zap_lookup(drc->drc_ds->ds_dir->dd_pool->dp_meta_objset,
3268                     drc->drc_ds->ds_object, DS_FIELD_RESUME_BYTES,
3269                     sizeof (bytes), 1, &bytes);
3270                 drc->drc_bytes_read += bytes;
3271         }
3272
3273         drc->drc_ignore_objlist = objlist_create();
3274
3275         /* these were verified in dmu_recv_begin */
3276         ASSERT3U(DMU_GET_STREAM_HDRTYPE(drc->drc_drrb->drr_versioninfo), ==,
3277             DMU_SUBSTREAM);
3278         ASSERT3U(drc->drc_drrb->drr_type, <, DMU_OST_NUMTYPES);
3279
3280         ASSERT(dsl_dataset_phys(drc->drc_ds)->ds_flags & DS_FLAG_INCONSISTENT);
3281         ASSERT0(drc->drc_os->os_encrypted &&
3282             (drc->drc_featureflags & DMU_BACKUP_FEATURE_EMBED_DATA));
3283
3284         /* handle DSL encryption key payload */
3285         if (drc->drc_featureflags & DMU_BACKUP_FEATURE_RAW) {
3286                 nvlist_t *keynvl = NULL;
3287
3288                 ASSERT(drc->drc_os->os_encrypted);
3289                 ASSERT(drc->drc_raw);
3290
3291                 err = nvlist_lookup_nvlist(drc->drc_begin_nvl, "crypt_keydata",
3292                     &keynvl);
3293                 if (err != 0)
3294                         goto out;
3295
3296                 if (!drc->drc_heal) {
3297                         /*
3298                          * If this is a new dataset we set the key immediately.
3299                          * Otherwise we don't want to change the key until we
3300                          * are sure the rest of the receive succeeded so we
3301                          * stash the keynvl away until then.
3302                          */
3303                         err = dsl_crypto_recv_raw(spa_name(drc->drc_os->os_spa),
3304                             drc->drc_ds->ds_object, drc->drc_fromsnapobj,
3305                             drc->drc_drrb->drr_type, keynvl, drc->drc_newfs);
3306                         if (err != 0)
3307                                 goto out;
3308                 }
3309
3310                 /* see comment in dmu_recv_end_sync() */
3311                 drc->drc_ivset_guid = 0;
3312                 (void) nvlist_lookup_uint64(keynvl, "to_ivset_guid",
3313                     &drc->drc_ivset_guid);
3314
3315                 if (!drc->drc_newfs)
3316                         drc->drc_keynvl = fnvlist_dup(keynvl);
3317         }
3318
3319         if (drc->drc_featureflags & DMU_BACKUP_FEATURE_RESUMING) {
3320                 err = resume_check(drc, drc->drc_begin_nvl);
3321                 if (err != 0)
3322                         goto out;
3323         }
3324
3325         /*
3326          * For compatibility with recursive send streams, we do this here,
3327          * rather than in dmu_recv_begin. If we pull the next header too
3328          * early, and it's the END record, we break the `recv_skip` logic.
3329          */
3330         if (drc->drc_drr_begin->drr_payloadlen == 0) {
3331                 err = receive_read_payload_and_next_header(drc, 0, NULL);
3332                 if (err != 0)
3333                         goto out;
3334         }
3335
3336         /*
3337          * If we failed before this point we will clean up any new resume
3338          * state that was created. Now that we've gotten past the initial
3339          * checks we are ok to retain that resume state.
3340          */
3341         drc->drc_should_save = B_TRUE;
3342
3343         (void) bqueue_init(&rwa->q, zfs_recv_queue_ff,
3344             MAX(zfs_recv_queue_length, 2 * zfs_max_recordsize),
3345             offsetof(struct receive_record_arg, node));
3346         cv_init(&rwa->cv, NULL, CV_DEFAULT, NULL);
3347         mutex_init(&rwa->mutex, NULL, MUTEX_DEFAULT, NULL);
3348         rwa->os = drc->drc_os;
3349         rwa->byteswap = drc->drc_byteswap;
3350         rwa->heal = drc->drc_heal;
3351         rwa->tofs = drc->drc_tofs;
3352         rwa->resumable = drc->drc_resumable;
3353         rwa->raw = drc->drc_raw;
3354         rwa->spill = drc->drc_spill;
3355         rwa->full = (drc->drc_drr_begin->drr_u.drr_begin.drr_fromguid == 0);
3356         rwa->os->os_raw_receive = drc->drc_raw;
3357         if (drc->drc_heal) {
3358                 rwa->heal_pio = zio_root(drc->drc_os->os_spa, NULL, NULL,
3359                     ZIO_FLAG_GODFATHER);
3360         }
3361         list_create(&rwa->write_batch, sizeof (struct receive_record_arg),
3362             offsetof(struct receive_record_arg, node.bqn_node));
3363
3364         (void) thread_create(NULL, 0, receive_writer_thread, rwa, 0, curproc,
3365             TS_RUN, minclsyspri);
3366         /*
3367          * We're reading rwa->err without locks, which is safe since we are the
3368          * only reader, and the worker thread is the only writer.  It's ok if we
3369          * miss a write for an iteration or two of the loop, since the writer
3370          * thread will keep freeing records we send it until we send it an eos
3371          * marker.
3372          *
3373          * We can leave this loop in 3 ways:  First, if rwa->err is
3374          * non-zero.  In that case, the writer thread will free the rrd we just
3375          * pushed.  Second, if  we're interrupted; in that case, either it's the
3376          * first loop and drc->drc_rrd was never allocated, or it's later, and
3377          * drc->drc_rrd has been handed off to the writer thread who will free
3378          * it.  Finally, if receive_read_record fails or we're at the end of the
3379          * stream, then we free drc->drc_rrd and exit.
3380          */
3381         while (rwa->err == 0) {
3382                 if (issig(JUSTLOOKING) && issig(FORREAL)) {
3383                         err = SET_ERROR(EINTR);
3384                         break;
3385                 }
3386
3387                 ASSERT3P(drc->drc_rrd, ==, NULL);
3388                 drc->drc_rrd = drc->drc_next_rrd;
3389                 drc->drc_next_rrd = NULL;
3390                 /* Allocates and loads header into drc->drc_next_rrd */
3391                 err = receive_read_record(drc);
3392
3393                 if (drc->drc_rrd->header.drr_type == DRR_END || err != 0) {
3394                         kmem_free(drc->drc_rrd, sizeof (*drc->drc_rrd));
3395                         drc->drc_rrd = NULL;
3396                         break;
3397                 }
3398
3399                 bqueue_enqueue(&rwa->q, drc->drc_rrd,
3400                     sizeof (struct receive_record_arg) +
3401                     drc->drc_rrd->payload_size);
3402                 drc->drc_rrd = NULL;
3403         }
3404
3405         ASSERT3P(drc->drc_rrd, ==, NULL);
3406         drc->drc_rrd = kmem_zalloc(sizeof (*drc->drc_rrd), KM_SLEEP);
3407         drc->drc_rrd->eos_marker = B_TRUE;
3408         bqueue_enqueue_flush(&rwa->q, drc->drc_rrd, 1);
3409
3410         mutex_enter(&rwa->mutex);
3411         while (!rwa->done) {
3412                 /*
3413                  * We need to use cv_wait_sig() so that any process that may
3414                  * be sleeping here can still fork.
3415                  */
3416                 (void) cv_wait_sig(&rwa->cv, &rwa->mutex);
3417         }
3418         mutex_exit(&rwa->mutex);
3419
3420         /*
3421          * If we are receiving a full stream as a clone, all object IDs which
3422          * are greater than the maximum ID referenced in the stream are
3423          * by definition unused and must be freed.
3424          */
3425         if (drc->drc_clone && drc->drc_drrb->drr_fromguid == 0) {
3426                 uint64_t obj = rwa->max_object + 1;
3427                 int free_err = 0;
3428                 int next_err = 0;
3429
3430                 while (next_err == 0) {
3431                         free_err = dmu_free_long_object(rwa->os, obj);
3432                         if (free_err != 0 && free_err != ENOENT)
3433                                 break;
3434
3435                         next_err = dmu_object_next(rwa->os, &obj, FALSE, 0);
3436                 }
3437
3438                 if (err == 0) {
3439                         if (free_err != 0 && free_err != ENOENT)
3440                                 err = free_err;
3441                         else if (next_err != ESRCH)
3442                                 err = next_err;
3443                 }
3444         }
3445
3446         cv_destroy(&rwa->cv);
3447         mutex_destroy(&rwa->mutex);
3448         bqueue_destroy(&rwa->q);
3449         list_destroy(&rwa->write_batch);
3450         if (err == 0)
3451                 err = rwa->err;
3452
3453 out:
3454         /*
3455          * If we hit an error before we started the receive_writer_thread
3456          * we need to clean up the next_rrd we create by processing the
3457          * DRR_BEGIN record.
3458          */
3459         if (drc->drc_next_rrd != NULL)
3460                 kmem_free(drc->drc_next_rrd, sizeof (*drc->drc_next_rrd));
3461
3462         /*
3463          * The objset will be invalidated by dmu_recv_end() when we do
3464          * dsl_dataset_clone_swap_sync_impl().
3465          */
3466         drc->drc_os = NULL;
3467
3468         kmem_free(rwa, sizeof (*rwa));
3469         nvlist_free(drc->drc_begin_nvl);
3470
3471         if (err != 0) {
3472                 /*
3473                  * Clean up references. If receive is not resumable,
3474                  * destroy what we created, so we don't leave it in
3475                  * the inconsistent state.
3476                  */
3477                 dmu_recv_cleanup_ds(drc);
3478                 nvlist_free(drc->drc_keynvl);
3479         }
3480
3481         objlist_destroy(drc->drc_ignore_objlist);
3482         drc->drc_ignore_objlist = NULL;
3483         *voffp = drc->drc_voff;
3484         return (err);
3485 }
3486
3487 static int
3488 dmu_recv_end_check(void *arg, dmu_tx_t *tx)
3489 {
3490         dmu_recv_cookie_t *drc = arg;
3491         dsl_pool_t *dp = dmu_tx_pool(tx);
3492         int error;
3493
3494         ASSERT3P(drc->drc_ds->ds_owner, ==, dmu_recv_tag);
3495
3496         if (drc->drc_heal) {
3497                 error = 0;
3498         } else if (!drc->drc_newfs) {
3499                 dsl_dataset_t *origin_head;
3500
3501                 error = dsl_dataset_hold(dp, drc->drc_tofs, FTAG, &origin_head);
3502                 if (error != 0)
3503                         return (error);
3504                 if (drc->drc_force) {
3505                         /*
3506                          * We will destroy any snapshots in tofs (i.e. before
3507                          * origin_head) that are after the origin (which is
3508                          * the snap before drc_ds, because drc_ds can not
3509                          * have any snaps of its own).
3510                          */
3511                         uint64_t obj;
3512
3513                         obj = dsl_dataset_phys(origin_head)->ds_prev_snap_obj;
3514                         while (obj !=
3515                             dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj) {
3516                                 dsl_dataset_t *snap;
3517                                 error = dsl_dataset_hold_obj(dp, obj, FTAG,
3518                                     &snap);
3519                                 if (error != 0)
3520                                         break;
3521                                 if (snap->ds_dir != origin_head->ds_dir)
3522                                         error = SET_ERROR(EINVAL);
3523                                 if (error == 0)  {
3524                                         error = dsl_destroy_snapshot_check_impl(
3525                                             snap, B_FALSE);
3526                                 }
3527                                 obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
3528                                 dsl_dataset_rele(snap, FTAG);
3529                                 if (error != 0)
3530                                         break;
3531                         }
3532                         if (error != 0) {
3533                                 dsl_dataset_rele(origin_head, FTAG);
3534                                 return (error);
3535                         }
3536                 }
3537                 if (drc->drc_keynvl != NULL) {
3538                         error = dsl_crypto_recv_raw_key_check(drc->drc_ds,
3539                             drc->drc_keynvl, tx);
3540                         if (error != 0) {
3541                                 dsl_dataset_rele(origin_head, FTAG);
3542                                 return (error);
3543                         }
3544                 }
3545
3546                 error = dsl_dataset_clone_swap_check_impl(drc->drc_ds,
3547                     origin_head, drc->drc_force, drc->drc_owner, tx);
3548                 if (error != 0) {
3549                         dsl_dataset_rele(origin_head, FTAG);
3550                         return (error);
3551                 }
3552                 error = dsl_dataset_snapshot_check_impl(origin_head,
3553                     drc->drc_tosnap, tx, B_TRUE, 1,
3554                     drc->drc_cred, drc->drc_proc);
3555                 dsl_dataset_rele(origin_head, FTAG);
3556                 if (error != 0)
3557                         return (error);
3558
3559                 error = dsl_destroy_head_check_impl(drc->drc_ds, 1);
3560         } else {
3561                 error = dsl_dataset_snapshot_check_impl(drc->drc_ds,
3562                     drc->drc_tosnap, tx, B_TRUE, 1,
3563                     drc->drc_cred, drc->drc_proc);
3564         }
3565         return (error);
3566 }
3567
3568 static void
3569 dmu_recv_end_sync(void *arg, dmu_tx_t *tx)
3570 {
3571         dmu_recv_cookie_t *drc = arg;
3572         dsl_pool_t *dp = dmu_tx_pool(tx);
3573         boolean_t encrypted = drc->drc_ds->ds_dir->dd_crypto_obj != 0;
3574         uint64_t newsnapobj = 0;
3575
3576         spa_history_log_internal_ds(drc->drc_ds, "finish receiving",
3577             tx, "snap=%s", drc->drc_tosnap);
3578         drc->drc_ds->ds_objset->os_raw_receive = B_FALSE;
3579
3580         if (drc->drc_heal) {
3581                 if (drc->drc_keynvl != NULL) {
3582                         nvlist_free(drc->drc_keynvl);
3583                         drc->drc_keynvl = NULL;
3584                 }
3585         } else if (!drc->drc_newfs) {
3586                 dsl_dataset_t *origin_head;
3587
3588                 VERIFY0(dsl_dataset_hold(dp, drc->drc_tofs, FTAG,
3589                     &origin_head));
3590
3591                 if (drc->drc_force) {
3592                         /*
3593                          * Destroy any snapshots of drc_tofs (origin_head)
3594                          * after the origin (the snap before drc_ds).
3595                          */
3596                         uint64_t obj;
3597
3598                         obj = dsl_dataset_phys(origin_head)->ds_prev_snap_obj;
3599                         while (obj !=
3600                             dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj) {
3601                                 dsl_dataset_t *snap;
3602                                 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG,
3603                                     &snap));
3604                                 ASSERT3P(snap->ds_dir, ==, origin_head->ds_dir);
3605                                 obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
3606                                 dsl_destroy_snapshot_sync_impl(snap,
3607                                     B_FALSE, tx);
3608                                 dsl_dataset_rele(snap, FTAG);
3609                         }
3610                 }
3611                 if (drc->drc_keynvl != NULL) {
3612                         dsl_crypto_recv_raw_key_sync(drc->drc_ds,
3613                             drc->drc_keynvl, tx);
3614                         nvlist_free(drc->drc_keynvl);
3615                         drc->drc_keynvl = NULL;
3616                 }
3617
3618                 VERIFY3P(drc->drc_ds->ds_prev, ==,
3619                     origin_head->ds_prev);
3620
3621                 dsl_dataset_clone_swap_sync_impl(drc->drc_ds,
3622                     origin_head, tx);
3623                 /*
3624                  * The objset was evicted by dsl_dataset_clone_swap_sync_impl,
3625                  * so drc_os is no longer valid.
3626                  */
3627                 drc->drc_os = NULL;
3628
3629                 dsl_dataset_snapshot_sync_impl(origin_head,
3630                     drc->drc_tosnap, tx);
3631
3632                 /* set snapshot's creation time and guid */
3633                 dmu_buf_will_dirty(origin_head->ds_prev->ds_dbuf, tx);
3634                 dsl_dataset_phys(origin_head->ds_prev)->ds_creation_time =
3635                     drc->drc_drrb->drr_creation_time;
3636                 dsl_dataset_phys(origin_head->ds_prev)->ds_guid =
3637                     drc->drc_drrb->drr_toguid;
3638                 dsl_dataset_phys(origin_head->ds_prev)->ds_flags &=
3639                     ~DS_FLAG_INCONSISTENT;
3640
3641                 dmu_buf_will_dirty(origin_head->ds_dbuf, tx);
3642                 dsl_dataset_phys(origin_head)->ds_flags &=
3643                     ~DS_FLAG_INCONSISTENT;
3644
3645                 newsnapobj =
3646                     dsl_dataset_phys(origin_head)->ds_prev_snap_obj;
3647
3648                 dsl_dataset_rele(origin_head, FTAG);
3649                 dsl_destroy_head_sync_impl(drc->drc_ds, tx);
3650
3651                 if (drc->drc_owner != NULL)
3652                         VERIFY3P(origin_head->ds_owner, ==, drc->drc_owner);
3653         } else {
3654                 dsl_dataset_t *ds = drc->drc_ds;
3655
3656                 dsl_dataset_snapshot_sync_impl(ds, drc->drc_tosnap, tx);
3657
3658                 /* set snapshot's creation time and guid */
3659                 dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
3660                 dsl_dataset_phys(ds->ds_prev)->ds_creation_time =
3661                     drc->drc_drrb->drr_creation_time;
3662                 dsl_dataset_phys(ds->ds_prev)->ds_guid =
3663                     drc->drc_drrb->drr_toguid;
3664                 dsl_dataset_phys(ds->ds_prev)->ds_flags &=
3665                     ~DS_FLAG_INCONSISTENT;
3666
3667                 dmu_buf_will_dirty(ds->ds_dbuf, tx);
3668                 dsl_dataset_phys(ds)->ds_flags &= ~DS_FLAG_INCONSISTENT;
3669                 if (dsl_dataset_has_resume_receive_state(ds)) {
3670                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3671                             DS_FIELD_RESUME_FROMGUID, tx);
3672                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3673                             DS_FIELD_RESUME_OBJECT, tx);
3674                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3675                             DS_FIELD_RESUME_OFFSET, tx);
3676                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3677                             DS_FIELD_RESUME_BYTES, tx);
3678                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3679                             DS_FIELD_RESUME_TOGUID, tx);
3680                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3681                             DS_FIELD_RESUME_TONAME, tx);
3682                         (void) zap_remove(dp->dp_meta_objset, ds->ds_object,
3683                             DS_FIELD_RESUME_REDACT_BOOKMARK_SNAPS, tx);
3684                 }
3685                 newsnapobj =
3686                     dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj;
3687         }
3688
3689         /*
3690          * If this is a raw receive, the crypt_keydata nvlist will include
3691          * a to_ivset_guid for us to set on the new snapshot. This value
3692          * will override the value generated by the snapshot code. However,
3693          * this value may not be present, because older implementations of
3694          * the raw send code did not include this value, and we are still
3695          * allowed to receive them if the zfs_disable_ivset_guid_check
3696          * tunable is set, in which case we will leave the newly-generated
3697          * value.
3698          */
3699         if (!drc->drc_heal && drc->drc_raw && drc->drc_ivset_guid != 0) {
3700                 dmu_object_zapify(dp->dp_meta_objset, newsnapobj,
3701                     DMU_OT_DSL_DATASET, tx);
3702                 VERIFY0(zap_update(dp->dp_meta_objset, newsnapobj,
3703                     DS_FIELD_IVSET_GUID, sizeof (uint64_t), 1,
3704                     &drc->drc_ivset_guid, tx));
3705         }
3706
3707         /*
3708          * Release the hold from dmu_recv_begin.  This must be done before
3709          * we return to open context, so that when we free the dataset's dnode
3710          * we can evict its bonus buffer. Since the dataset may be destroyed
3711          * at this point (and therefore won't have a valid pointer to the spa)
3712          * we release the key mapping manually here while we do have a valid
3713          * pointer, if it exists.
3714          */
3715         if (!drc->drc_raw && encrypted) {
3716                 (void) spa_keystore_remove_mapping(dmu_tx_pool(tx)->dp_spa,
3717                     drc->drc_ds->ds_object, drc->drc_ds);
3718         }
3719         dsl_dataset_disown(drc->drc_ds, 0, dmu_recv_tag);
3720         drc->drc_ds = NULL;
3721 }
3722
3723 static int dmu_recv_end_modified_blocks = 3;
3724
3725 static int
3726 dmu_recv_existing_end(dmu_recv_cookie_t *drc)
3727 {
3728 #ifdef _KERNEL
3729         /*
3730          * We will be destroying the ds; make sure its origin is unmounted if
3731          * necessary.
3732          */
3733         char name[ZFS_MAX_DATASET_NAME_LEN];
3734         dsl_dataset_name(drc->drc_ds, name);
3735         zfs_destroy_unmount_origin(name);
3736 #endif
3737
3738         return (dsl_sync_task(drc->drc_tofs,
3739             dmu_recv_end_check, dmu_recv_end_sync, drc,
3740             dmu_recv_end_modified_blocks, ZFS_SPACE_CHECK_NORMAL));
3741 }
3742
3743 static int
3744 dmu_recv_new_end(dmu_recv_cookie_t *drc)
3745 {
3746         return (dsl_sync_task(drc->drc_tofs,
3747             dmu_recv_end_check, dmu_recv_end_sync, drc,
3748             dmu_recv_end_modified_blocks, ZFS_SPACE_CHECK_NORMAL));
3749 }
3750
3751 int
3752 dmu_recv_end(dmu_recv_cookie_t *drc, void *owner)
3753 {
3754         int error;
3755
3756         drc->drc_owner = owner;
3757
3758         if (drc->drc_newfs)
3759                 error = dmu_recv_new_end(drc);
3760         else
3761                 error = dmu_recv_existing_end(drc);
3762
3763         if (error != 0) {
3764                 dmu_recv_cleanup_ds(drc);
3765                 nvlist_free(drc->drc_keynvl);
3766         } else if (!drc->drc_heal) {
3767                 if (drc->drc_newfs) {
3768                         zvol_create_minor(drc->drc_tofs);
3769                 }
3770                 char *snapname = kmem_asprintf("%s@%s",
3771                     drc->drc_tofs, drc->drc_tosnap);
3772                 zvol_create_minor(snapname);
3773                 kmem_strfree(snapname);
3774         }
3775         return (error);
3776 }
3777
3778 /*
3779  * Return TRUE if this objset is currently being received into.
3780  */
3781 boolean_t
3782 dmu_objset_is_receiving(objset_t *os)
3783 {
3784         return (os->os_dsl_dataset != NULL &&
3785             os->os_dsl_dataset->ds_owner == dmu_recv_tag);
3786 }
3787
3788 ZFS_MODULE_PARAM(zfs_recv, zfs_recv_, queue_length, UINT, ZMOD_RW,
3789         "Maximum receive queue length");
3790
3791 ZFS_MODULE_PARAM(zfs_recv, zfs_recv_, queue_ff, UINT, ZMOD_RW,
3792         "Receive queue fill fraction");
3793
3794 ZFS_MODULE_PARAM(zfs_recv, zfs_recv_, write_batch_size, UINT, ZMOD_RW,
3795         "Maximum amount of writes to batch into one transaction");
3796
3797 ZFS_MODULE_PARAM(zfs_recv, zfs_recv_, best_effort_corrective, INT, ZMOD_RW,
3798         "Ignore errors during corrective receive");
3799 /* END CSTYLED */