1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21 /* 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23 * Copyright 2011 Nexenta Systems, Inc. All rights reserved. 24 * Copyright (c) 2011, 2015 by Delphix. All rights reserved. 25 * Copyright (c) 2014, Joyent, Inc. All rights reserved. 26 * Copyright 2014 HybridCluster. All rights reserved. 27 * Copyright 2016 RackTop Systems. 28 * Copyright (c) 2014 Integros [integros.com] 29 */ 30 31 #include <sys/dmu.h> 32 #include <sys/dmu_impl.h> 33 #include <sys/dmu_tx.h> 34 #include <sys/dbuf.h> 35 #include <sys/dnode.h> 36 #include <sys/zfs_context.h> 37 #include <sys/dmu_objset.h> 38 #include <sys/dmu_traverse.h> 39 #include <sys/dsl_dataset.h> 40 #include <sys/dsl_dir.h> 41 #include <sys/dsl_prop.h> 42 #include <sys/dsl_pool.h> 43 #include <sys/dsl_synctask.h> 44 #include <sys/zfs_ioctl.h> 45 #include <sys/zap.h> 46 #include <sys/zio_checksum.h> 47 #include <sys/zfs_znode.h> 48 #include <zfs_fletcher.h> 49 #include <sys/avl.h> 50 #include <sys/ddt.h> 51 #include <sys/zfs_onexit.h> 52 #include <sys/dmu_send.h> 53 #include <sys/dsl_destroy.h> 54 #include <sys/blkptr.h> 55 #include <sys/dsl_bookmark.h> 56 #include <sys/zfeature.h> 57 #include <sys/bqueue.h> 58 59 /* Set this tunable to TRUE to replace corrupt data with 0x2f5baddb10c */ 60 int zfs_send_corrupt_data = B_FALSE; 61 int zfs_send_queue_length = 16 * 1024 * 1024; 62 int zfs_recv_queue_length = 16 * 1024 * 1024; 63 /* Set this tunable to FALSE to disable setting of DRR_FLAG_FREERECORDS */ 64 int zfs_send_set_freerecords_bit = B_TRUE; 65 66 static char *dmu_recv_tag = "dmu_recv_tag"; 67 const char *recv_clone_name = "%recv"; 68 69 #define BP_SPAN(datablkszsec, indblkshift, level) \ 70 (((uint64_t)datablkszsec) << (SPA_MINBLOCKSHIFT + \ 71 (level) * (indblkshift - SPA_BLKPTRSHIFT))) 72 73 static void byteswap_record(dmu_replay_record_t *drr); 74 75 struct send_thread_arg { 76 bqueue_t q; 77 dsl_dataset_t *ds; /* Dataset to traverse */ 78 uint64_t fromtxg; /* Traverse from this txg */ 79 int flags; /* flags to pass to traverse_dataset */ 80 int error_code; 81 boolean_t cancel; 82 zbookmark_phys_t resume; 83 }; 84 85 struct send_block_record { 86 boolean_t eos_marker; /* Marks the end of the stream */ 87 blkptr_t bp; 88 zbookmark_phys_t zb; 89 uint8_t indblkshift; 90 uint16_t datablkszsec; 91 bqueue_node_t ln; 92 }; 93 94 static int 95 dump_bytes(dmu_sendarg_t *dsp, void *buf, int len) 96 { 97 dsl_dataset_t *ds = dmu_objset_ds(dsp->dsa_os); 98 ssize_t resid; /* have to get resid to get detailed errno */ 99 100 /* 101 * The code does not rely on this (len being a multiple of 8). We keep 102 * this assertion because of the corresponding assertion in 103 * receive_read(). Keeping this assertion ensures that we do not 104 * inadvertently break backwards compatibility (causing the assertion 105 * in receive_read() to trigger on old software). 106 * 107 * Removing the assertions could be rolled into a new feature that uses 108 * data that isn't 8-byte aligned; if the assertions were removed, a 109 * feature flag would have to be added. 110 */ 111 112 ASSERT0(len % 8); 113 114 dsp->dsa_err = vn_rdwr(UIO_WRITE, dsp->dsa_vp, 115 (caddr_t)buf, len, 116 0, UIO_SYSSPACE, FAPPEND, RLIM64_INFINITY, CRED(), &resid); 117 118 mutex_enter(&ds->ds_sendstream_lock); 119 *dsp->dsa_off += len; 120 mutex_exit(&ds->ds_sendstream_lock); 121 122 return (dsp->dsa_err); 123 } 124 125 /* 126 * For all record types except BEGIN, fill in the checksum (overlaid in 127 * drr_u.drr_checksum.drr_checksum). The checksum verifies everything 128 * up to the start of the checksum itself. 129 */ 130 static int 131 dump_record(dmu_sendarg_t *dsp, void *payload, int payload_len) 132 { 133 ASSERT3U(offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum), 134 ==, sizeof (dmu_replay_record_t) - sizeof (zio_cksum_t)); 135 fletcher_4_incremental_native(dsp->dsa_drr, 136 offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum), 137 &dsp->dsa_zc); 138 if (dsp->dsa_drr->drr_type == DRR_BEGIN) { 139 dsp->dsa_sent_begin = B_TRUE; 140 } else { 141 ASSERT(ZIO_CHECKSUM_IS_ZERO(&dsp->dsa_drr->drr_u. 142 drr_checksum.drr_checksum)); 143 dsp->dsa_drr->drr_u.drr_checksum.drr_checksum = dsp->dsa_zc; 144 } 145 if (dsp->dsa_drr->drr_type == DRR_END) { 146 dsp->dsa_sent_end = B_TRUE; 147 } 148 fletcher_4_incremental_native(&dsp->dsa_drr-> 149 drr_u.drr_checksum.drr_checksum, 150 sizeof (zio_cksum_t), &dsp->dsa_zc); 151 if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)) != 0) 152 return (SET_ERROR(EINTR)); 153 if (payload_len != 0) { 154 fletcher_4_incremental_native(payload, payload_len, 155 &dsp->dsa_zc); 156 if (dump_bytes(dsp, payload, payload_len) != 0) 157 return (SET_ERROR(EINTR)); 158 } 159 return (0); 160 } 161 162 /* 163 * Fill in the drr_free struct, or perform aggregation if the previous record is 164 * also a free record, and the two are adjacent. 165 * 166 * Note that we send free records even for a full send, because we want to be 167 * able to receive a full send as a clone, which requires a list of all the free 168 * and freeobject records that were generated on the source. 169 */ 170 static int 171 dump_free(dmu_sendarg_t *dsp, uint64_t object, uint64_t offset, 172 uint64_t length) 173 { 174 struct drr_free *drrf = &(dsp->dsa_drr->drr_u.drr_free); 175 176 /* 177 * When we receive a free record, dbuf_free_range() assumes 178 * that the receiving system doesn't have any dbufs in the range 179 * being freed. This is always true because there is a one-record 180 * constraint: we only send one WRITE record for any given 181 * object,offset. We know that the one-record constraint is 182 * true because we always send data in increasing order by 183 * object,offset. 184 * 185 * If the increasing-order constraint ever changes, we should find 186 * another way to assert that the one-record constraint is still 187 * satisfied. 188 */ 189 ASSERT(object > dsp->dsa_last_data_object || 190 (object == dsp->dsa_last_data_object && 191 offset > dsp->dsa_last_data_offset)); 192 193 if (length != -1ULL && offset + length < offset) 194 length = -1ULL; 195 196 /* 197 * If there is a pending op, but it's not PENDING_FREE, push it out, 198 * since free block aggregation can only be done for blocks of the 199 * same type (i.e., DRR_FREE records can only be aggregated with 200 * other DRR_FREE records. DRR_FREEOBJECTS records can only be 201 * aggregated with other DRR_FREEOBJECTS records. 202 */ 203 if (dsp->dsa_pending_op != PENDING_NONE && 204 dsp->dsa_pending_op != PENDING_FREE) { 205 if (dump_record(dsp, NULL, 0) != 0) 206 return (SET_ERROR(EINTR)); 207 dsp->dsa_pending_op = PENDING_NONE; 208 } 209 210 if (dsp->dsa_pending_op == PENDING_FREE) { 211 /* 212 * There should never be a PENDING_FREE if length is -1 213 * (because dump_dnode is the only place where this 214 * function is called with a -1, and only after flushing 215 * any pending record). 216 */ 217 ASSERT(length != -1ULL); 218 /* 219 * Check to see whether this free block can be aggregated 220 * with pending one. 221 */ 222 if (drrf->drr_object == object && drrf->drr_offset + 223 drrf->drr_length == offset) { 224 drrf->drr_length += length; 225 return (0); 226 } else { 227 /* not a continuation. Push out pending record */ 228 if (dump_record(dsp, NULL, 0) != 0) 229 return (SET_ERROR(EINTR)); 230 dsp->dsa_pending_op = PENDING_NONE; 231 } 232 } 233 /* create a FREE record and make it pending */ 234 bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t)); 235 dsp->dsa_drr->drr_type = DRR_FREE; 236 drrf->drr_object = object; 237 drrf->drr_offset = offset; 238 drrf->drr_length = length; 239 drrf->drr_toguid = dsp->dsa_toguid; 240 if (length == -1ULL) { 241 if (dump_record(dsp, NULL, 0) != 0) 242 return (SET_ERROR(EINTR)); 243 } else { 244 dsp->dsa_pending_op = PENDING_FREE; 245 } 246 247 return (0); 248 } 249 250 static int 251 dump_write(dmu_sendarg_t *dsp, dmu_object_type_t type, 252 uint64_t object, uint64_t offset, int lsize, int psize, const blkptr_t *bp, 253 void *data) 254 { 255 uint64_t payload_size; 256 struct drr_write *drrw = &(dsp->dsa_drr->drr_u.drr_write); 257 258 /* 259 * We send data in increasing object, offset order. 260 * See comment in dump_free() for details. 261 */ 262 ASSERT(object > dsp->dsa_last_data_object || 263 (object == dsp->dsa_last_data_object && 264 offset > dsp->dsa_last_data_offset)); 265 dsp->dsa_last_data_object = object; 266 dsp->dsa_last_data_offset = offset + lsize - 1; 267 268 /* 269 * If there is any kind of pending aggregation (currently either 270 * a grouping of free objects or free blocks), push it out to 271 * the stream, since aggregation can't be done across operations 272 * of different types. 273 */ 274 if (dsp->dsa_pending_op != PENDING_NONE) { 275 if (dump_record(dsp, NULL, 0) != 0) 276 return (SET_ERROR(EINTR)); 277 dsp->dsa_pending_op = PENDING_NONE; 278 } 279 /* write a WRITE record */ 280 bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t)); 281 dsp->dsa_drr->drr_type = DRR_WRITE; 282 drrw->drr_object = object; 283 drrw->drr_type = type; 284 drrw->drr_offset = offset; 285 drrw->drr_toguid = dsp->dsa_toguid; 286 drrw->drr_logical_size = lsize; 287 288 /* only set the compression fields if the buf is compressed */ 289 if (lsize != psize) { 290 ASSERT(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_COMPRESSED); 291 ASSERT(!BP_IS_EMBEDDED(bp)); 292 ASSERT(!BP_SHOULD_BYTESWAP(bp)); 293 ASSERT(!DMU_OT_IS_METADATA(BP_GET_TYPE(bp))); 294 ASSERT3U(BP_GET_COMPRESS(bp), !=, ZIO_COMPRESS_OFF); 295 ASSERT3S(psize, >, 0); 296 ASSERT3S(lsize, >=, psize); 297 298 drrw->drr_compressiontype = BP_GET_COMPRESS(bp); 299 drrw->drr_compressed_size = psize; 300 payload_size = drrw->drr_compressed_size; 301 } else { 302 payload_size = drrw->drr_logical_size; 303 } 304 305 if (bp == NULL || BP_IS_EMBEDDED(bp)) { 306 /* 307 * There's no pre-computed checksum for partial-block 308 * writes or embedded BP's, so (like 309 * fletcher4-checkummed blocks) userland will have to 310 * compute a dedup-capable checksum itself. 311 */ 312 drrw->drr_checksumtype = ZIO_CHECKSUM_OFF; 313 } else { 314 drrw->drr_checksumtype = BP_GET_CHECKSUM(bp); 315 if (zio_checksum_table[drrw->drr_checksumtype].ci_flags & 316 ZCHECKSUM_FLAG_DEDUP) 317 drrw->drr_checksumflags |= DRR_CHECKSUM_DEDUP; 318 DDK_SET_LSIZE(&drrw->drr_key, BP_GET_LSIZE(bp)); 319 DDK_SET_PSIZE(&drrw->drr_key, BP_GET_PSIZE(bp)); 320 DDK_SET_COMPRESS(&drrw->drr_key, BP_GET_COMPRESS(bp)); 321 drrw->drr_key.ddk_cksum = bp->blk_cksum; 322 } 323 324 if (dump_record(dsp, data, payload_size) != 0) 325 return (SET_ERROR(EINTR)); 326 return (0); 327 } 328 329 static int 330 dump_write_embedded(dmu_sendarg_t *dsp, uint64_t object, uint64_t offset, 331 int blksz, const blkptr_t *bp) 332 { 333 char buf[BPE_PAYLOAD_SIZE]; 334 struct drr_write_embedded *drrw = 335 &(dsp->dsa_drr->drr_u.drr_write_embedded); 336 337 if (dsp->dsa_pending_op != PENDING_NONE) { 338 if (dump_record(dsp, NULL, 0) != 0) 339 return (EINTR); 340 dsp->dsa_pending_op = PENDING_NONE; 341 } 342 343 ASSERT(BP_IS_EMBEDDED(bp)); 344 345 bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t)); 346 dsp->dsa_drr->drr_type = DRR_WRITE_EMBEDDED; 347 drrw->drr_object = object; 348 drrw->drr_offset = offset; 349 drrw->drr_length = blksz; 350 drrw->drr_toguid = dsp->dsa_toguid; 351 drrw->drr_compression = BP_GET_COMPRESS(bp); 352 drrw->drr_etype = BPE_GET_ETYPE(bp); 353 drrw->drr_lsize = BPE_GET_LSIZE(bp); 354 drrw->drr_psize = BPE_GET_PSIZE(bp); 355 356 decode_embedded_bp_compressed(bp, buf); 357 358 if (dump_record(dsp, buf, P2ROUNDUP(drrw->drr_psize, 8)) != 0) 359 return (EINTR); 360 return (0); 361 } 362 363 static int 364 dump_spill(dmu_sendarg_t *dsp, uint64_t object, int blksz, void *data) 365 { 366 struct drr_spill *drrs = &(dsp->dsa_drr->drr_u.drr_spill); 367 368 if (dsp->dsa_pending_op != PENDING_NONE) { 369 if (dump_record(dsp, NULL, 0) != 0) 370 return (SET_ERROR(EINTR)); 371 dsp->dsa_pending_op = PENDING_NONE; 372 } 373 374 /* write a SPILL record */ 375 bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t)); 376 dsp->dsa_drr->drr_type = DRR_SPILL; 377 drrs->drr_object = object; 378 drrs->drr_length = blksz; 379 drrs->drr_toguid = dsp->dsa_toguid; 380 381 if (dump_record(dsp, data, blksz) != 0) 382 return (SET_ERROR(EINTR)); 383 return (0); 384 } 385 386 static int 387 dump_freeobjects(dmu_sendarg_t *dsp, uint64_t firstobj, uint64_t numobjs) 388 { 389 struct drr_freeobjects *drrfo = &(dsp->dsa_drr->drr_u.drr_freeobjects); 390 391 /* 392 * If there is a pending op, but it's not PENDING_FREEOBJECTS, 393 * push it out, since free block aggregation can only be done for 394 * blocks of the same type (i.e., DRR_FREE records can only be 395 * aggregated with other DRR_FREE records. DRR_FREEOBJECTS records 396 * can only be aggregated with other DRR_FREEOBJECTS records. 397 */ 398 if (dsp->dsa_pending_op != PENDING_NONE && 399 dsp->dsa_pending_op != PENDING_FREEOBJECTS) { 400 if (dump_record(dsp, NULL, 0) != 0) 401 return (SET_ERROR(EINTR)); 402 dsp->dsa_pending_op = PENDING_NONE; 403 } 404 if (dsp->dsa_pending_op == PENDING_FREEOBJECTS) { 405 /* 406 * See whether this free object array can be aggregated 407 * with pending one 408 */ 409 if (drrfo->drr_firstobj + drrfo->drr_numobjs == firstobj) { 410 drrfo->drr_numobjs += numobjs; 411 return (0); 412 } else { 413 /* can't be aggregated. Push out pending record */ 414 if (dump_record(dsp, NULL, 0) != 0) 415 return (SET_ERROR(EINTR)); 416 dsp->dsa_pending_op = PENDING_NONE; 417 } 418 } 419 420 /* write a FREEOBJECTS record */ 421 bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t)); 422 dsp->dsa_drr->drr_type = DRR_FREEOBJECTS; 423 drrfo->drr_firstobj = firstobj; 424 drrfo->drr_numobjs = numobjs; 425 drrfo->drr_toguid = dsp->dsa_toguid; 426 427 dsp->dsa_pending_op = PENDING_FREEOBJECTS; 428 429 return (0); 430 } 431 432 static int 433 dump_dnode(dmu_sendarg_t *dsp, uint64_t object, dnode_phys_t *dnp) 434 { 435 struct drr_object *drro = &(dsp->dsa_drr->drr_u.drr_object); 436 437 if (object < dsp->dsa_resume_object) { 438 /* 439 * Note: when resuming, we will visit all the dnodes in 440 * the block of dnodes that we are resuming from. In 441 * this case it's unnecessary to send the dnodes prior to 442 * the one we are resuming from. We should be at most one 443 * block's worth of dnodes behind the resume point. 444 */ 445 ASSERT3U(dsp->dsa_resume_object - object, <, 446 1 << (DNODE_BLOCK_SHIFT - DNODE_SHIFT)); 447 return (0); 448 } 449 450 if (dnp == NULL || dnp->dn_type == DMU_OT_NONE) 451 return (dump_freeobjects(dsp, object, 1)); 452 453 if (dsp->dsa_pending_op != PENDING_NONE) { 454 if (dump_record(dsp, NULL, 0) != 0) 455 return (SET_ERROR(EINTR)); 456 dsp->dsa_pending_op = PENDING_NONE; 457 } 458 459 /* write an OBJECT record */ 460 bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t)); 461 dsp->dsa_drr->drr_type = DRR_OBJECT; 462 drro->drr_object = object; 463 drro->drr_type = dnp->dn_type; 464 drro->drr_bonustype = dnp->dn_bonustype; 465 drro->drr_blksz = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT; 466 drro->drr_bonuslen = dnp->dn_bonuslen; 467 drro->drr_checksumtype = dnp->dn_checksum; 468 drro->drr_compress = dnp->dn_compress; 469 drro->drr_toguid = dsp->dsa_toguid; 470 471 if (!(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) && 472 drro->drr_blksz > SPA_OLD_MAXBLOCKSIZE) 473 drro->drr_blksz = SPA_OLD_MAXBLOCKSIZE; 474 475 if (dump_record(dsp, DN_BONUS(dnp), 476 P2ROUNDUP(dnp->dn_bonuslen, 8)) != 0) { 477 return (SET_ERROR(EINTR)); 478 } 479 480 /* Free anything past the end of the file. */ 481 if (dump_free(dsp, object, (dnp->dn_maxblkid + 1) * 482 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT), -1ULL) != 0) 483 return (SET_ERROR(EINTR)); 484 if (dsp->dsa_err != 0) 485 return (SET_ERROR(EINTR)); 486 return (0); 487 } 488 489 static boolean_t 490 backup_do_embed(dmu_sendarg_t *dsp, const blkptr_t *bp) 491 { 492 if (!BP_IS_EMBEDDED(bp)) 493 return (B_FALSE); 494 495 /* 496 * Compression function must be legacy, or explicitly enabled. 497 */ 498 if ((BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_LEGACY_FUNCTIONS && 499 !(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_LZ4))) 500 return (B_FALSE); 501 502 /* 503 * Embed type must be explicitly enabled. 504 */ 505 switch (BPE_GET_ETYPE(bp)) { 506 case BP_EMBEDDED_TYPE_DATA: 507 if (dsp->dsa_featureflags & DMU_BACKUP_FEATURE_EMBED_DATA) 508 return (B_TRUE); 509 break; 510 default: 511 return (B_FALSE); 512 } 513 return (B_FALSE); 514 } 515 516 /* 517 * This is the callback function to traverse_dataset that acts as the worker 518 * thread for dmu_send_impl. 519 */ 520 /*ARGSUSED*/ 521 static int 522 send_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp, 523 const zbookmark_phys_t *zb, const struct dnode_phys *dnp, void *arg) 524 { 525 struct send_thread_arg *sta = arg; 526 struct send_block_record *record; 527 uint64_t record_size; 528 int err = 0; 529 530 ASSERT(zb->zb_object == DMU_META_DNODE_OBJECT || 531 zb->zb_object >= sta->resume.zb_object); 532 533 if (sta->cancel) 534 return (SET_ERROR(EINTR)); 535 536 if (bp == NULL) { 537 ASSERT3U(zb->zb_level, ==, ZB_DNODE_LEVEL); 538 return (0); 539 } else if (zb->zb_level < 0) { 540 return (0); 541 } 542 543 record = kmem_zalloc(sizeof (struct send_block_record), KM_SLEEP); 544 record->eos_marker = B_FALSE; 545 record->bp = *bp; 546 record->zb = *zb; 547 record->indblkshift = dnp->dn_indblkshift; 548 record->datablkszsec = dnp->dn_datablkszsec; 549 record_size = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT; 550 bqueue_enqueue(&sta->q, record, record_size); 551 552 return (err); 553 } 554 555 /* 556 * This function kicks off the traverse_dataset. It also handles setting the 557 * error code of the thread in case something goes wrong, and pushes the End of 558 * Stream record when the traverse_dataset call has finished. If there is no 559 * dataset to traverse, the thread immediately pushes End of Stream marker. 560 */ 561 static void 562 send_traverse_thread(void *arg) 563 { 564 struct send_thread_arg *st_arg = arg; 565 int err; 566 struct send_block_record *data; 567 568 if (st_arg->ds != NULL) { 569 err = traverse_dataset_resume(st_arg->ds, 570 st_arg->fromtxg, &st_arg->resume, 571 st_arg->flags, send_cb, st_arg); 572 573 if (err != EINTR) 574 st_arg->error_code = err; 575 } 576 data = kmem_zalloc(sizeof (*data), KM_SLEEP); 577 data->eos_marker = B_TRUE; 578 bqueue_enqueue(&st_arg->q, data, 1); 579 thread_exit(); 580 } 581 582 /* 583 * This function actually handles figuring out what kind of record needs to be 584 * dumped, reading the data (which has hopefully been prefetched), and calling 585 * the appropriate helper function. 586 */ 587 static int 588 do_dump(dmu_sendarg_t *dsa, struct send_block_record *data) 589 { 590 dsl_dataset_t *ds = dmu_objset_ds(dsa->dsa_os); 591 const blkptr_t *bp = &data->bp; 592 const zbookmark_phys_t *zb = &data->zb; 593 uint8_t indblkshift = data->indblkshift; 594 uint16_t dblkszsec = data->datablkszsec; 595 spa_t *spa = ds->ds_dir->dd_pool->dp_spa; 596 dmu_object_type_t type = bp ? BP_GET_TYPE(bp) : DMU_OT_NONE; 597 int err = 0; 598 599 ASSERT3U(zb->zb_level, >=, 0); 600 601 ASSERT(zb->zb_object == DMU_META_DNODE_OBJECT || 602 zb->zb_object >= dsa->dsa_resume_object); 603 604 if (zb->zb_object != DMU_META_DNODE_OBJECT && 605 DMU_OBJECT_IS_SPECIAL(zb->zb_object)) { 606 return (0); 607 } else if (BP_IS_HOLE(bp) && 608 zb->zb_object == DMU_META_DNODE_OBJECT) { 609 uint64_t span = BP_SPAN(dblkszsec, indblkshift, zb->zb_level); 610 uint64_t dnobj = (zb->zb_blkid * span) >> DNODE_SHIFT; 611 err = dump_freeobjects(dsa, dnobj, span >> DNODE_SHIFT); 612 } else if (BP_IS_HOLE(bp)) { 613 uint64_t span = BP_SPAN(dblkszsec, indblkshift, zb->zb_level); 614 uint64_t offset = zb->zb_blkid * span; 615 err = dump_free(dsa, zb->zb_object, offset, span); 616 } else if (zb->zb_level > 0 || type == DMU_OT_OBJSET) { 617 return (0); 618 } else if (type == DMU_OT_DNODE) { 619 int blksz = BP_GET_LSIZE(bp); 620 arc_flags_t aflags = ARC_FLAG_WAIT; 621 arc_buf_t *abuf; 622 623 ASSERT0(zb->zb_level); 624 625 if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf, 626 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, 627 &aflags, zb) != 0) 628 return (SET_ERROR(EIO)); 629 630 dnode_phys_t *blk = abuf->b_data; 631 uint64_t dnobj = zb->zb_blkid * (blksz >> DNODE_SHIFT); 632 for (int i = 0; i < blksz >> DNODE_SHIFT; i++) { 633 err = dump_dnode(dsa, dnobj + i, blk + i); 634 if (err != 0) 635 break; 636 } 637 arc_buf_destroy(abuf, &abuf); 638 } else if (type == DMU_OT_SA) { 639 arc_flags_t aflags = ARC_FLAG_WAIT; 640 arc_buf_t *abuf; 641 int blksz = BP_GET_LSIZE(bp); 642 643 if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf, 644 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, 645 &aflags, zb) != 0) 646 return (SET_ERROR(EIO)); 647 648 err = dump_spill(dsa, zb->zb_object, blksz, abuf->b_data); 649 arc_buf_destroy(abuf, &abuf); 650 } else if (backup_do_embed(dsa, bp)) { 651 /* it's an embedded level-0 block of a regular object */ 652 int blksz = dblkszsec << SPA_MINBLOCKSHIFT; 653 ASSERT0(zb->zb_level); 654 err = dump_write_embedded(dsa, zb->zb_object, 655 zb->zb_blkid * blksz, blksz, bp); 656 } else { 657 /* it's a level-0 block of a regular object */ 658 arc_flags_t aflags = ARC_FLAG_WAIT; 659 arc_buf_t *abuf; 660 int blksz = dblkszsec << SPA_MINBLOCKSHIFT; 661 uint64_t offset; 662 663 /* 664 * If we have large blocks stored on disk but the send flags 665 * don't allow us to send large blocks, we split the data from 666 * the arc buf into chunks. 667 */ 668 boolean_t split_large_blocks = blksz > SPA_OLD_MAXBLOCKSIZE && 669 !(dsa->dsa_featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS); 670 /* 671 * We should only request compressed data from the ARC if all 672 * the following are true: 673 * - stream compression was requested 674 * - we aren't splitting large blocks into smaller chunks 675 * - the data won't need to be byteswapped before sending 676 * - this isn't an embedded block 677 * - this isn't metadata (if receiving on a different endian 678 * system it can be byteswapped more easily) 679 */ 680 boolean_t request_compressed = 681 (dsa->dsa_featureflags & DMU_BACKUP_FEATURE_COMPRESSED) && 682 !split_large_blocks && !BP_SHOULD_BYTESWAP(bp) && 683 !BP_IS_EMBEDDED(bp) && !DMU_OT_IS_METADATA(BP_GET_TYPE(bp)); 684 685 ASSERT0(zb->zb_level); 686 ASSERT(zb->zb_object > dsa->dsa_resume_object || 687 (zb->zb_object == dsa->dsa_resume_object && 688 zb->zb_blkid * blksz >= dsa->dsa_resume_offset)); 689 690 ASSERT0(zb->zb_level); 691 ASSERT(zb->zb_object > dsa->dsa_resume_object || 692 (zb->zb_object == dsa->dsa_resume_object && 693 zb->zb_blkid * blksz >= dsa->dsa_resume_offset)); 694 695 ASSERT3U(blksz, ==, BP_GET_LSIZE(bp)); 696 697 enum zio_flag zioflags = ZIO_FLAG_CANFAIL; 698 if (request_compressed) 699 zioflags |= ZIO_FLAG_RAW; 700 if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf, 701 ZIO_PRIORITY_ASYNC_READ, zioflags, &aflags, zb) != 0) { 702 if (zfs_send_corrupt_data) { 703 /* Send a block filled with 0x"zfs badd bloc" */ 704 abuf = arc_alloc_buf(spa, &abuf, ARC_BUFC_DATA, 705 blksz); 706 uint64_t *ptr; 707 for (ptr = abuf->b_data; 708 (char *)ptr < (char *)abuf->b_data + blksz; 709 ptr++) 710 *ptr = 0x2f5baddb10cULL; 711 } else { 712 return (SET_ERROR(EIO)); 713 } 714 } 715 716 offset = zb->zb_blkid * blksz; 717 718 if (split_large_blocks) { 719 ASSERT3U(arc_get_compression(abuf), ==, 720 ZIO_COMPRESS_OFF); 721 char *buf = abuf->b_data; 722 while (blksz > 0 && err == 0) { 723 int n = MIN(blksz, SPA_OLD_MAXBLOCKSIZE); 724 err = dump_write(dsa, type, zb->zb_object, 725 offset, n, n, NULL, buf); 726 offset += n; 727 buf += n; 728 blksz -= n; 729 } 730 } else { 731 err = dump_write(dsa, type, zb->zb_object, offset, 732 blksz, arc_buf_size(abuf), bp, abuf->b_data); 733 } 734 arc_buf_destroy(abuf, &abuf); 735 } 736 737 ASSERT(err == 0 || err == EINTR); 738 return (err); 739 } 740 741 /* 742 * Pop the new data off the queue, and free the old data. 743 */ 744 static struct send_block_record * 745 get_next_record(bqueue_t *bq, struct send_block_record *data) 746 { 747 struct send_block_record *tmp = bqueue_dequeue(bq); 748 kmem_free(data, sizeof (*data)); 749 return (tmp); 750 } 751 752 /* 753 * Actually do the bulk of the work in a zfs send. 754 * 755 * Note: Releases dp using the specified tag. 756 */ 757 static int 758 dmu_send_impl(void *tag, dsl_pool_t *dp, dsl_dataset_t *to_ds, 759 zfs_bookmark_phys_t *ancestor_zb, boolean_t is_clone, 760 boolean_t embedok, boolean_t large_block_ok, boolean_t compressok, 761 int outfd, uint64_t resumeobj, uint64_t resumeoff, 762 vnode_t *vp, offset_t *off) 763 { 764 objset_t *os; 765 dmu_replay_record_t *drr; 766 dmu_sendarg_t *dsp; 767 int err; 768 uint64_t fromtxg = 0; 769 uint64_t featureflags = 0; 770 struct send_thread_arg to_arg = { 0 }; 771 772 err = dmu_objset_from_ds(to_ds, &os); 773 if (err != 0) { 774 dsl_pool_rele(dp, tag); 775 return (err); 776 } 777 778 drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP); 779 drr->drr_type = DRR_BEGIN; 780 drr->drr_u.drr_begin.drr_magic = DMU_BACKUP_MAGIC; 781 DMU_SET_STREAM_HDRTYPE(drr->drr_u.drr_begin.drr_versioninfo, 782 DMU_SUBSTREAM); 783 784 #ifdef _KERNEL 785 if (dmu_objset_type(os) == DMU_OST_ZFS) { 786 uint64_t version; 787 if (zfs_get_zplprop(os, ZFS_PROP_VERSION, &version) != 0) { 788 kmem_free(drr, sizeof (dmu_replay_record_t)); 789 dsl_pool_rele(dp, tag); 790 return (SET_ERROR(EINVAL)); 791 } 792 if (version >= ZPL_VERSION_SA) { 793 featureflags |= DMU_BACKUP_FEATURE_SA_SPILL; 794 } 795 } 796 #endif 797 798 if (large_block_ok && to_ds->ds_feature_inuse[SPA_FEATURE_LARGE_BLOCKS]) 799 featureflags |= DMU_BACKUP_FEATURE_LARGE_BLOCKS; 800 if (embedok && 801 spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA)) { 802 featureflags |= DMU_BACKUP_FEATURE_EMBED_DATA; 803 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS)) 804 featureflags |= DMU_BACKUP_FEATURE_LZ4; 805 } 806 if (compressok) { 807 featureflags |= DMU_BACKUP_FEATURE_COMPRESSED; 808 } 809 if ((featureflags & 810 (DMU_BACKUP_FEATURE_EMBED_DATA | DMU_BACKUP_FEATURE_COMPRESSED)) != 811 0 && spa_feature_is_active(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS)) { 812 featureflags |= DMU_BACKUP_FEATURE_LZ4; 813 } 814 815 if (resumeobj != 0 || resumeoff != 0) { 816 featureflags |= DMU_BACKUP_FEATURE_RESUMING; 817 } 818 819 DMU_SET_FEATUREFLAGS(drr->drr_u.drr_begin.drr_versioninfo, 820 featureflags); 821 822 drr->drr_u.drr_begin.drr_creation_time = 823 dsl_dataset_phys(to_ds)->ds_creation_time; 824 drr->drr_u.drr_begin.drr_type = dmu_objset_type(os); 825 if (is_clone) 826 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CLONE; 827 drr->drr_u.drr_begin.drr_toguid = dsl_dataset_phys(to_ds)->ds_guid; 828 if (dsl_dataset_phys(to_ds)->ds_flags & DS_FLAG_CI_DATASET) 829 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CI_DATA; 830 if (zfs_send_set_freerecords_bit) 831 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_FREERECORDS; 832 833 if (ancestor_zb != NULL) { 834 drr->drr_u.drr_begin.drr_fromguid = 835 ancestor_zb->zbm_guid; 836 fromtxg = ancestor_zb->zbm_creation_txg; 837 } 838 dsl_dataset_name(to_ds, drr->drr_u.drr_begin.drr_toname); 839 if (!to_ds->ds_is_snapshot) { 840 (void) strlcat(drr->drr_u.drr_begin.drr_toname, "@--head--", 841 sizeof (drr->drr_u.drr_begin.drr_toname)); 842 } 843 844 dsp = kmem_zalloc(sizeof (dmu_sendarg_t), KM_SLEEP); 845 846 dsp->dsa_drr = drr; 847 dsp->dsa_vp = vp; 848 dsp->dsa_outfd = outfd; 849 dsp->dsa_proc = curproc; 850 dsp->dsa_os = os; 851 dsp->dsa_off = off; 852 dsp->dsa_toguid = dsl_dataset_phys(to_ds)->ds_guid; 853 dsp->dsa_pending_op = PENDING_NONE; 854 dsp->dsa_featureflags = featureflags; 855 dsp->dsa_resume_object = resumeobj; 856 dsp->dsa_resume_offset = resumeoff; 857 858 mutex_enter(&to_ds->ds_sendstream_lock); 859 list_insert_head(&to_ds->ds_sendstreams, dsp); 860 mutex_exit(&to_ds->ds_sendstream_lock); 861 862 dsl_dataset_long_hold(to_ds, FTAG); 863 dsl_pool_rele(dp, tag); 864 865 void *payload = NULL; 866 size_t payload_len = 0; 867 if (resumeobj != 0 || resumeoff != 0) { 868 dmu_object_info_t to_doi; 869 err = dmu_object_info(os, resumeobj, &to_doi); 870 if (err != 0) 871 goto out; 872 SET_BOOKMARK(&to_arg.resume, to_ds->ds_object, resumeobj, 0, 873 resumeoff / to_doi.doi_data_block_size); 874 875 nvlist_t *nvl = fnvlist_alloc(); 876 fnvlist_add_uint64(nvl, "resume_object", resumeobj); 877 fnvlist_add_uint64(nvl, "resume_offset", resumeoff); 878 payload = fnvlist_pack(nvl, &payload_len); 879 drr->drr_payloadlen = payload_len; 880 fnvlist_free(nvl); 881 } 882 883 err = dump_record(dsp, payload, payload_len); 884 fnvlist_pack_free(payload, payload_len); 885 if (err != 0) { 886 err = dsp->dsa_err; 887 goto out; 888 } 889 890 err = bqueue_init(&to_arg.q, zfs_send_queue_length, 891 offsetof(struct send_block_record, ln)); 892 to_arg.error_code = 0; 893 to_arg.cancel = B_FALSE; 894 to_arg.ds = to_ds; 895 to_arg.fromtxg = fromtxg; 896 to_arg.flags = TRAVERSE_PRE | TRAVERSE_PREFETCH; 897 (void) thread_create(NULL, 0, send_traverse_thread, &to_arg, 0, curproc, 898 TS_RUN, minclsyspri); 899 900 struct send_block_record *to_data; 901 to_data = bqueue_dequeue(&to_arg.q); 902 903 while (!to_data->eos_marker && err == 0) { 904 err = do_dump(dsp, to_data); 905 to_data = get_next_record(&to_arg.q, to_data); 906 if (issig(JUSTLOOKING) && issig(FORREAL)) 907 err = EINTR; 908 } 909 910 if (err != 0) { 911 to_arg.cancel = B_TRUE; 912 while (!to_data->eos_marker) { 913 to_data = get_next_record(&to_arg.q, to_data); 914 } 915 } 916 kmem_free(to_data, sizeof (*to_data)); 917 918 bqueue_destroy(&to_arg.q); 919 920 if (err == 0 && to_arg.error_code != 0) 921 err = to_arg.error_code; 922 923 if (err != 0) 924 goto out; 925 926 if (dsp->dsa_pending_op != PENDING_NONE) 927 if (dump_record(dsp, NULL, 0) != 0) 928 err = SET_ERROR(EINTR); 929 930 if (err != 0) { 931 if (err == EINTR && dsp->dsa_err != 0) 932 err = dsp->dsa_err; 933 goto out; 934 } 935 936 bzero(drr, sizeof (dmu_replay_record_t)); 937 drr->drr_type = DRR_END; 938 drr->drr_u.drr_end.drr_checksum = dsp->dsa_zc; 939 drr->drr_u.drr_end.drr_toguid = dsp->dsa_toguid; 940 941 if (dump_record(dsp, NULL, 0) != 0) 942 err = dsp->dsa_err; 943 944 out: 945 mutex_enter(&to_ds->ds_sendstream_lock); 946 list_remove(&to_ds->ds_sendstreams, dsp); 947 mutex_exit(&to_ds->ds_sendstream_lock); 948 949 VERIFY(err != 0 || (dsp->dsa_sent_begin && dsp->dsa_sent_end)); 950 951 kmem_free(drr, sizeof (dmu_replay_record_t)); 952 kmem_free(dsp, sizeof (dmu_sendarg_t)); 953 954 dsl_dataset_long_rele(to_ds, FTAG); 955 956 return (err); 957 } 958 959 int 960 dmu_send_obj(const char *pool, uint64_t tosnap, uint64_t fromsnap, 961 boolean_t embedok, boolean_t large_block_ok, boolean_t compressok, 962 int outfd, vnode_t *vp, offset_t *off) 963 { 964 dsl_pool_t *dp; 965 dsl_dataset_t *ds; 966 dsl_dataset_t *fromds = NULL; 967 int err; 968 969 err = dsl_pool_hold(pool, FTAG, &dp); 970 if (err != 0) 971 return (err); 972 973 err = dsl_dataset_hold_obj(dp, tosnap, FTAG, &ds); 974 if (err != 0) { 975 dsl_pool_rele(dp, FTAG); 976 return (err); 977 } 978 979 if (fromsnap != 0) { 980 zfs_bookmark_phys_t zb; 981 boolean_t is_clone; 982 983 err = dsl_dataset_hold_obj(dp, fromsnap, FTAG, &fromds); 984 if (err != 0) { 985 dsl_dataset_rele(ds, FTAG); 986 dsl_pool_rele(dp, FTAG); 987 return (err); 988 } 989 if (!dsl_dataset_is_before(ds, fromds, 0)) 990 err = SET_ERROR(EXDEV); 991 zb.zbm_creation_time = 992 dsl_dataset_phys(fromds)->ds_creation_time; 993 zb.zbm_creation_txg = dsl_dataset_phys(fromds)->ds_creation_txg; 994 zb.zbm_guid = dsl_dataset_phys(fromds)->ds_guid; 995 is_clone = (fromds->ds_dir != ds->ds_dir); 996 dsl_dataset_rele(fromds, FTAG); 997 err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone, 998 embedok, large_block_ok, compressok, outfd, 0, 0, vp, off); 999 } else { 1000 err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE, 1001 embedok, large_block_ok, compressok, outfd, 0, 0, vp, off); 1002 } 1003 dsl_dataset_rele(ds, FTAG); 1004 return (err); 1005 } 1006 1007 int 1008 dmu_send(const char *tosnap, const char *fromsnap, boolean_t embedok, 1009 boolean_t large_block_ok, boolean_t compressok, int outfd, 1010 uint64_t resumeobj, uint64_t resumeoff, 1011 vnode_t *vp, offset_t *off) 1012 { 1013 dsl_pool_t *dp; 1014 dsl_dataset_t *ds; 1015 int err; 1016 boolean_t owned = B_FALSE; 1017 1018 if (fromsnap != NULL && strpbrk(fromsnap, "@#") == NULL) 1019 return (SET_ERROR(EINVAL)); 1020 1021 err = dsl_pool_hold(tosnap, FTAG, &dp); 1022 if (err != 0) 1023 return (err); 1024 1025 if (strchr(tosnap, '@') == NULL && spa_writeable(dp->dp_spa)) { 1026 /* 1027 * We are sending a filesystem or volume. Ensure 1028 * that it doesn't change by owning the dataset. 1029 */ 1030 err = dsl_dataset_own(dp, tosnap, FTAG, &ds); 1031 owned = B_TRUE; 1032 } else { 1033 err = dsl_dataset_hold(dp, tosnap, FTAG, &ds); 1034 } 1035 if (err != 0) { 1036 dsl_pool_rele(dp, FTAG); 1037 return (err); 1038 } 1039 1040 if (fromsnap != NULL) { 1041 zfs_bookmark_phys_t zb; 1042 boolean_t is_clone = B_FALSE; 1043 int fsnamelen = strchr(tosnap, '@') - tosnap; 1044 1045 /* 1046 * If the fromsnap is in a different filesystem, then 1047 * mark the send stream as a clone. 1048 */ 1049 if (strncmp(tosnap, fromsnap, fsnamelen) != 0 || 1050 (fromsnap[fsnamelen] != '@' && 1051 fromsnap[fsnamelen] != '#')) { 1052 is_clone = B_TRUE; 1053 } 1054 1055 if (strchr(fromsnap, '@')) { 1056 dsl_dataset_t *fromds; 1057 err = dsl_dataset_hold(dp, fromsnap, FTAG, &fromds); 1058 if (err == 0) { 1059 if (!dsl_dataset_is_before(ds, fromds, 0)) 1060 err = SET_ERROR(EXDEV); 1061 zb.zbm_creation_time = 1062 dsl_dataset_phys(fromds)->ds_creation_time; 1063 zb.zbm_creation_txg = 1064 dsl_dataset_phys(fromds)->ds_creation_txg; 1065 zb.zbm_guid = dsl_dataset_phys(fromds)->ds_guid; 1066 is_clone = (ds->ds_dir != fromds->ds_dir); 1067 dsl_dataset_rele(fromds, FTAG); 1068 } 1069 } else { 1070 err = dsl_bookmark_lookup(dp, fromsnap, ds, &zb); 1071 } 1072 if (err != 0) { 1073 dsl_dataset_rele(ds, FTAG); 1074 dsl_pool_rele(dp, FTAG); 1075 return (err); 1076 } 1077 err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone, 1078 embedok, large_block_ok, compressok, 1079 outfd, resumeobj, resumeoff, vp, off); 1080 } else { 1081 err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE, 1082 embedok, large_block_ok, compressok, 1083 outfd, resumeobj, resumeoff, vp, off); 1084 } 1085 if (owned) 1086 dsl_dataset_disown(ds, FTAG); 1087 else 1088 dsl_dataset_rele(ds, FTAG); 1089 return (err); 1090 } 1091 1092 static int 1093 dmu_adjust_send_estimate_for_indirects(dsl_dataset_t *ds, uint64_t uncompressed, 1094 uint64_t compressed, boolean_t stream_compressed, uint64_t *sizep) 1095 { 1096 int err; 1097 uint64_t size; 1098 /* 1099 * Assume that space (both on-disk and in-stream) is dominated by 1100 * data. We will adjust for indirect blocks and the copies property, 1101 * but ignore per-object space used (eg, dnodes and DRR_OBJECT records). 1102 */ 1103 uint64_t recordsize; 1104 uint64_t record_count; 1105 1106 /* Assume all (uncompressed) blocks are recordsize. */ 1107 err = dsl_prop_get_int_ds(ds, zfs_prop_to_name(ZFS_PROP_RECORDSIZE), 1108 &recordsize); 1109 if (err != 0) 1110 return (err); 1111 record_count = uncompressed / recordsize; 1112 1113 /* 1114 * If we're estimating a send size for a compressed stream, use the 1115 * compressed data size to estimate the stream size. Otherwise, use the 1116 * uncompressed data size. 1117 */ 1118 size = stream_compressed ? compressed : uncompressed; 1119 1120 /* 1121 * Subtract out approximate space used by indirect blocks. 1122 * Assume most space is used by data blocks (non-indirect, non-dnode). 1123 * Assume no ditto blocks or internal fragmentation. 1124 * 1125 * Therefore, space used by indirect blocks is sizeof(blkptr_t) per 1126 * block. 1127 */ 1128 size -= record_count * sizeof (blkptr_t); 1129 1130 /* Add in the space for the record associated with each block. */ 1131 size += record_count * sizeof (dmu_replay_record_t); 1132 1133 *sizep = size; 1134 1135 return (0); 1136 } 1137 1138 int 1139 dmu_send_estimate(dsl_dataset_t *ds, dsl_dataset_t *fromds, 1140 boolean_t stream_compressed, uint64_t *sizep) 1141 { 1142 dsl_pool_t *dp = ds->ds_dir->dd_pool; 1143 int err; 1144 uint64_t uncomp, comp; 1145 1146 ASSERT(dsl_pool_config_held(dp)); 1147 1148 /* tosnap must be a snapshot */ 1149 if (!ds->ds_is_snapshot) 1150 return (SET_ERROR(EINVAL)); 1151 1152 /* fromsnap, if provided, must be a snapshot */ 1153 if (fromds != NULL && !fromds->ds_is_snapshot) 1154 return (SET_ERROR(EINVAL)); 1155 1156 /* 1157 * fromsnap must be an earlier snapshot from the same fs as tosnap, 1158 * or the origin's fs. 1159 */ 1160 if (fromds != NULL && !dsl_dataset_is_before(ds, fromds, 0)) 1161 return (SET_ERROR(EXDEV)); 1162 1163 /* Get compressed and uncompressed size estimates of changed data. */ 1164 if (fromds == NULL) { 1165 uncomp = dsl_dataset_phys(ds)->ds_uncompressed_bytes; 1166 comp = dsl_dataset_phys(ds)->ds_compressed_bytes; 1167 } else { 1168 uint64_t used; 1169 err = dsl_dataset_space_written(fromds, ds, 1170 &used, &comp, &uncomp); 1171 if (err != 0) 1172 return (err); 1173 } 1174 1175 err = dmu_adjust_send_estimate_for_indirects(ds, uncomp, comp, 1176 stream_compressed, sizep); 1177 return (err); 1178 } 1179 1180 struct calculate_send_arg { 1181 uint64_t uncompressed; 1182 uint64_t compressed; 1183 }; 1184 1185 /* 1186 * Simple callback used to traverse the blocks of a snapshot and sum their 1187 * uncompressed and compressed sizes. 1188 */ 1189 /* ARGSUSED */ 1190 static int 1191 dmu_calculate_send_traversal(spa_t *spa, zilog_t *zilog, const blkptr_t *bp, 1192 const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg) 1193 { 1194 struct calculate_send_arg *space = arg; 1195 if (bp != NULL && !BP_IS_HOLE(bp)) { 1196 space->uncompressed += BP_GET_UCSIZE(bp); 1197 space->compressed += BP_GET_PSIZE(bp); 1198 } 1199 return (0); 1200 } 1201 1202 /* 1203 * Given a desination snapshot and a TXG, calculate the approximate size of a 1204 * send stream sent from that TXG. from_txg may be zero, indicating that the 1205 * whole snapshot will be sent. 1206 */ 1207 int 1208 dmu_send_estimate_from_txg(dsl_dataset_t *ds, uint64_t from_txg, 1209 boolean_t stream_compressed, uint64_t *sizep) 1210 { 1211 dsl_pool_t *dp = ds->ds_dir->dd_pool; 1212 int err; 1213 struct calculate_send_arg size = { 0 }; 1214 1215 ASSERT(dsl_pool_config_held(dp)); 1216 1217 /* tosnap must be a snapshot */ 1218 if (!ds->ds_is_snapshot) 1219 return (SET_ERROR(EINVAL)); 1220 1221 /* verify that from_txg is before the provided snapshot was taken */ 1222 if (from_txg >= dsl_dataset_phys(ds)->ds_creation_txg) { 1223 return (SET_ERROR(EXDEV)); 1224 } 1225 1226 /* 1227 * traverse the blocks of the snapshot with birth times after 1228 * from_txg, summing their uncompressed size 1229 */ 1230 err = traverse_dataset(ds, from_txg, TRAVERSE_POST, 1231 dmu_calculate_send_traversal, &size); 1232 if (err) 1233 return (err); 1234 1235 err = dmu_adjust_send_estimate_for_indirects(ds, size.uncompressed, 1236 size.compressed, stream_compressed, sizep); 1237 return (err); 1238 } 1239 1240 typedef struct dmu_recv_begin_arg { 1241 const char *drba_origin; 1242 dmu_recv_cookie_t *drba_cookie; 1243 cred_t *drba_cred; 1244 uint64_t drba_snapobj; 1245 } dmu_recv_begin_arg_t; 1246 1247 static int 1248 recv_begin_check_existing_impl(dmu_recv_begin_arg_t *drba, dsl_dataset_t *ds, 1249 uint64_t fromguid) 1250 { 1251 uint64_t val; 1252 int error; 1253 dsl_pool_t *dp = ds->ds_dir->dd_pool; 1254 1255 /* temporary clone name must not exist */ 1256 error = zap_lookup(dp->dp_meta_objset, 1257 dsl_dir_phys(ds->ds_dir)->dd_child_dir_zapobj, recv_clone_name, 1258 8, 1, &val); 1259 if (error != ENOENT) 1260 return (error == 0 ? EBUSY : error); 1261 1262 /* new snapshot name must not exist */ 1263 error = zap_lookup(dp->dp_meta_objset, 1264 dsl_dataset_phys(ds)->ds_snapnames_zapobj, 1265 drba->drba_cookie->drc_tosnap, 8, 1, &val); 1266 if (error != ENOENT) 1267 return (error == 0 ? EEXIST : error); 1268 1269 /* 1270 * Check snapshot limit before receiving. We'll recheck again at the 1271 * end, but might as well abort before receiving if we're already over 1272 * the limit. 1273 * 1274 * Note that we do not check the file system limit with 1275 * dsl_dir_fscount_check because the temporary %clones don't count 1276 * against that limit. 1277 */ 1278 error = dsl_fs_ss_limit_check(ds->ds_dir, 1, ZFS_PROP_SNAPSHOT_LIMIT, 1279 NULL, drba->drba_cred); 1280 if (error != 0) 1281 return (error); 1282 1283 if (fromguid != 0) { 1284 dsl_dataset_t *snap; 1285 uint64_t obj = dsl_dataset_phys(ds)->ds_prev_snap_obj; 1286 1287 /* Find snapshot in this dir that matches fromguid. */ 1288 while (obj != 0) { 1289 error = dsl_dataset_hold_obj(dp, obj, FTAG, 1290 &snap); 1291 if (error != 0) 1292 return (SET_ERROR(ENODEV)); 1293 if (snap->ds_dir != ds->ds_dir) { 1294 dsl_dataset_rele(snap, FTAG); 1295 return (SET_ERROR(ENODEV)); 1296 } 1297 if (dsl_dataset_phys(snap)->ds_guid == fromguid) 1298 break; 1299 obj = dsl_dataset_phys(snap)->ds_prev_snap_obj; 1300 dsl_dataset_rele(snap, FTAG); 1301 } 1302 if (obj == 0) 1303 return (SET_ERROR(ENODEV)); 1304 1305 if (drba->drba_cookie->drc_force) { 1306 drba->drba_snapobj = obj; 1307 } else { 1308 /* 1309 * If we are not forcing, there must be no 1310 * changes since fromsnap. 1311 */ 1312 if (dsl_dataset_modified_since_snap(ds, snap)) { 1313 dsl_dataset_rele(snap, FTAG); 1314 return (SET_ERROR(ETXTBSY)); 1315 } 1316 drba->drba_snapobj = ds->ds_prev->ds_object; 1317 } 1318 1319 dsl_dataset_rele(snap, FTAG); 1320 } else { 1321 /* if full, then must be forced */ 1322 if (!drba->drba_cookie->drc_force) 1323 return (SET_ERROR(EEXIST)); 1324 /* start from $ORIGIN@$ORIGIN, if supported */ 1325 drba->drba_snapobj = dp->dp_origin_snap != NULL ? 1326 dp->dp_origin_snap->ds_object : 0; 1327 } 1328 1329 return (0); 1330 1331 } 1332 1333 static int 1334 dmu_recv_begin_check(void *arg, dmu_tx_t *tx) 1335 { 1336 dmu_recv_begin_arg_t *drba = arg; 1337 dsl_pool_t *dp = dmu_tx_pool(tx); 1338 struct drr_begin *drrb = drba->drba_cookie->drc_drrb; 1339 uint64_t fromguid = drrb->drr_fromguid; 1340 int flags = drrb->drr_flags; 1341 int error; 1342 uint64_t featureflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo); 1343 dsl_dataset_t *ds; 1344 const char *tofs = drba->drba_cookie->drc_tofs; 1345 1346 /* already checked */ 1347 ASSERT3U(drrb->drr_magic, ==, DMU_BACKUP_MAGIC); 1348 ASSERT(!(featureflags & DMU_BACKUP_FEATURE_RESUMING)); 1349 1350 if (DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) == 1351 DMU_COMPOUNDSTREAM || 1352 drrb->drr_type >= DMU_OST_NUMTYPES || 1353 ((flags & DRR_FLAG_CLONE) && drba->drba_origin == NULL)) 1354 return (SET_ERROR(EINVAL)); 1355 1356 /* Verify pool version supports SA if SA_SPILL feature set */ 1357 if ((featureflags & DMU_BACKUP_FEATURE_SA_SPILL) && 1358 spa_version(dp->dp_spa) < SPA_VERSION_SA) 1359 return (SET_ERROR(ENOTSUP)); 1360 1361 if (drba->drba_cookie->drc_resumable && 1362 !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EXTENSIBLE_DATASET)) 1363 return (SET_ERROR(ENOTSUP)); 1364 1365 /* 1366 * The receiving code doesn't know how to translate a WRITE_EMBEDDED 1367 * record to a plain WRITE record, so the pool must have the 1368 * EMBEDDED_DATA feature enabled if the stream has WRITE_EMBEDDED 1369 * records. Same with WRITE_EMBEDDED records that use LZ4 compression. 1370 */ 1371 if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA) && 1372 !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA)) 1373 return (SET_ERROR(ENOTSUP)); 1374 if ((featureflags & DMU_BACKUP_FEATURE_LZ4) && 1375 !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS)) 1376 return (SET_ERROR(ENOTSUP)); 1377 1378 /* 1379 * The receiving code doesn't know how to translate large blocks 1380 * to smaller ones, so the pool must have the LARGE_BLOCKS 1381 * feature enabled if the stream has LARGE_BLOCKS. 1382 */ 1383 if ((featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) && 1384 !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LARGE_BLOCKS)) 1385 return (SET_ERROR(ENOTSUP)); 1386 1387 error = dsl_dataset_hold(dp, tofs, FTAG, &ds); 1388 if (error == 0) { 1389 /* target fs already exists; recv into temp clone */ 1390 1391 /* Can't recv a clone into an existing fs */ 1392 if (flags & DRR_FLAG_CLONE || drba->drba_origin) { 1393 dsl_dataset_rele(ds, FTAG); 1394 return (SET_ERROR(EINVAL)); 1395 } 1396 1397 error = recv_begin_check_existing_impl(drba, ds, fromguid); 1398 dsl_dataset_rele(ds, FTAG); 1399 } else if (error == ENOENT) { 1400 /* target fs does not exist; must be a full backup or clone */ 1401 char buf[ZFS_MAX_DATASET_NAME_LEN]; 1402 1403 /* 1404 * If it's a non-clone incremental, we are missing the 1405 * target fs, so fail the recv. 1406 */ 1407 if (fromguid != 0 && !(flags & DRR_FLAG_CLONE || 1408 drba->drba_origin)) 1409 return (SET_ERROR(ENOENT)); 1410 1411 /* 1412 * If we're receiving a full send as a clone, and it doesn't 1413 * contain all the necessary free records and freeobject 1414 * records, reject it. 1415 */ 1416 if (fromguid == 0 && drba->drba_origin && 1417 !(flags & DRR_FLAG_FREERECORDS)) 1418 return (SET_ERROR(EINVAL)); 1419 1420 /* Open the parent of tofs */ 1421 ASSERT3U(strlen(tofs), <, sizeof (buf)); 1422 (void) strlcpy(buf, tofs, strrchr(tofs, '/') - tofs + 1); 1423 error = dsl_dataset_hold(dp, buf, FTAG, &ds); 1424 if (error != 0) 1425 return (error); 1426 1427 /* 1428 * Check filesystem and snapshot limits before receiving. We'll 1429 * recheck snapshot limits again at the end (we create the 1430 * filesystems and increment those counts during begin_sync). 1431 */ 1432 error = dsl_fs_ss_limit_check(ds->ds_dir, 1, 1433 ZFS_PROP_FILESYSTEM_LIMIT, NULL, drba->drba_cred); 1434 if (error != 0) { 1435 dsl_dataset_rele(ds, FTAG); 1436 return (error); 1437 } 1438 1439 error = dsl_fs_ss_limit_check(ds->ds_dir, 1, 1440 ZFS_PROP_SNAPSHOT_LIMIT, NULL, drba->drba_cred); 1441 if (error != 0) { 1442 dsl_dataset_rele(ds, FTAG); 1443 return (error); 1444 } 1445 1446 if (drba->drba_origin != NULL) { 1447 dsl_dataset_t *origin; 1448 error = dsl_dataset_hold(dp, drba->drba_origin, 1449 FTAG, &origin); 1450 if (error != 0) { 1451 dsl_dataset_rele(ds, FTAG); 1452 return (error); 1453 } 1454 if (!origin->ds_is_snapshot) { 1455 dsl_dataset_rele(origin, FTAG); 1456 dsl_dataset_rele(ds, FTAG); 1457 return (SET_ERROR(EINVAL)); 1458 } 1459 if (dsl_dataset_phys(origin)->ds_guid != fromguid && 1460 fromguid != 0) { 1461 dsl_dataset_rele(origin, FTAG); 1462 dsl_dataset_rele(ds, FTAG); 1463 return (SET_ERROR(ENODEV)); 1464 } 1465 dsl_dataset_rele(origin, FTAG); 1466 } 1467 dsl_dataset_rele(ds, FTAG); 1468 error = 0; 1469 } 1470 return (error); 1471 } 1472 1473 static void 1474 dmu_recv_begin_sync(void *arg, dmu_tx_t *tx) 1475 { 1476 dmu_recv_begin_arg_t *drba = arg; 1477 dsl_pool_t *dp = dmu_tx_pool(tx); 1478 objset_t *mos = dp->dp_meta_objset; 1479 struct drr_begin *drrb = drba->drba_cookie->drc_drrb; 1480 const char *tofs = drba->drba_cookie->drc_tofs; 1481 dsl_dataset_t *ds, *newds; 1482 uint64_t dsobj; 1483 int error; 1484 uint64_t crflags = 0; 1485 1486 if (drrb->drr_flags & DRR_FLAG_CI_DATA) 1487 crflags |= DS_FLAG_CI_DATASET; 1488 1489 error = dsl_dataset_hold(dp, tofs, FTAG, &ds); 1490 if (error == 0) { 1491 /* create temporary clone */ 1492 dsl_dataset_t *snap = NULL; 1493 if (drba->drba_snapobj != 0) { 1494 VERIFY0(dsl_dataset_hold_obj(dp, 1495 drba->drba_snapobj, FTAG, &snap)); 1496 } 1497 dsobj = dsl_dataset_create_sync(ds->ds_dir, recv_clone_name, 1498 snap, crflags, drba->drba_cred, tx); 1499 if (drba->drba_snapobj != 0) 1500 dsl_dataset_rele(snap, FTAG); 1501 dsl_dataset_rele(ds, FTAG); 1502 } else { 1503 dsl_dir_t *dd; 1504 const char *tail; 1505 dsl_dataset_t *origin = NULL; 1506 1507 VERIFY0(dsl_dir_hold(dp, tofs, FTAG, &dd, &tail)); 1508 1509 if (drba->drba_origin != NULL) { 1510 VERIFY0(dsl_dataset_hold(dp, drba->drba_origin, 1511 FTAG, &origin)); 1512 } 1513 1514 /* Create new dataset. */ 1515 dsobj = dsl_dataset_create_sync(dd, 1516 strrchr(tofs, '/') + 1, 1517 origin, crflags, drba->drba_cred, tx); 1518 if (origin != NULL) 1519 dsl_dataset_rele(origin, FTAG); 1520 dsl_dir_rele(dd, FTAG); 1521 drba->drba_cookie->drc_newfs = B_TRUE; 1522 } 1523 VERIFY0(dsl_dataset_own_obj(dp, dsobj, dmu_recv_tag, &newds)); 1524 1525 if (drba->drba_cookie->drc_resumable) { 1526 dsl_dataset_zapify(newds, tx); 1527 if (drrb->drr_fromguid != 0) { 1528 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_FROMGUID, 1529 8, 1, &drrb->drr_fromguid, tx)); 1530 } 1531 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_TOGUID, 1532 8, 1, &drrb->drr_toguid, tx)); 1533 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_TONAME, 1534 1, strlen(drrb->drr_toname) + 1, drrb->drr_toname, tx)); 1535 uint64_t one = 1; 1536 uint64_t zero = 0; 1537 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_OBJECT, 1538 8, 1, &one, tx)); 1539 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_OFFSET, 1540 8, 1, &zero, tx)); 1541 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_BYTES, 1542 8, 1, &zero, tx)); 1543 if (DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo) & 1544 DMU_BACKUP_FEATURE_LARGE_BLOCKS) { 1545 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_LARGEBLOCK, 1546 8, 1, &one, tx)); 1547 } 1548 if (DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo) & 1549 DMU_BACKUP_FEATURE_EMBED_DATA) { 1550 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_EMBEDOK, 1551 8, 1, &one, tx)); 1552 } 1553 if (DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo) & 1554 DMU_BACKUP_FEATURE_COMPRESSED) { 1555 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_COMPRESSOK, 1556 8, 1, &one, tx)); 1557 } 1558 } 1559 1560 dmu_buf_will_dirty(newds->ds_dbuf, tx); 1561 dsl_dataset_phys(newds)->ds_flags |= DS_FLAG_INCONSISTENT; 1562 1563 /* 1564 * If we actually created a non-clone, we need to create the 1565 * objset in our new dataset. 1566 */ 1567 rrw_enter(&newds->ds_bp_rwlock, RW_READER, FTAG); 1568 if (BP_IS_HOLE(dsl_dataset_get_blkptr(newds))) { 1569 (void) dmu_objset_create_impl(dp->dp_spa, 1570 newds, dsl_dataset_get_blkptr(newds), drrb->drr_type, tx); 1571 } 1572 rrw_exit(&newds->ds_bp_rwlock, FTAG); 1573 1574 drba->drba_cookie->drc_ds = newds; 1575 1576 spa_history_log_internal_ds(newds, "receive", tx, ""); 1577 } 1578 1579 static int 1580 dmu_recv_resume_begin_check(void *arg, dmu_tx_t *tx) 1581 { 1582 dmu_recv_begin_arg_t *drba = arg; 1583 dsl_pool_t *dp = dmu_tx_pool(tx); 1584 struct drr_begin *drrb = drba->drba_cookie->drc_drrb; 1585 int error; 1586 uint64_t featureflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo); 1587 dsl_dataset_t *ds; 1588 const char *tofs = drba->drba_cookie->drc_tofs; 1589 1590 /* already checked */ 1591 ASSERT3U(drrb->drr_magic, ==, DMU_BACKUP_MAGIC); 1592 ASSERT(featureflags & DMU_BACKUP_FEATURE_RESUMING); 1593 1594 if (DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) == 1595 DMU_COMPOUNDSTREAM || 1596 drrb->drr_type >= DMU_OST_NUMTYPES) 1597 return (SET_ERROR(EINVAL)); 1598 1599 /* Verify pool version supports SA if SA_SPILL feature set */ 1600 if ((featureflags & DMU_BACKUP_FEATURE_SA_SPILL) && 1601 spa_version(dp->dp_spa) < SPA_VERSION_SA) 1602 return (SET_ERROR(ENOTSUP)); 1603 1604 /* 1605 * The receiving code doesn't know how to translate a WRITE_EMBEDDED 1606 * record to a plain WRITE record, so the pool must have the 1607 * EMBEDDED_DATA feature enabled if the stream has WRITE_EMBEDDED 1608 * records. Same with WRITE_EMBEDDED records that use LZ4 compression. 1609 */ 1610 if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA) && 1611 !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA)) 1612 return (SET_ERROR(ENOTSUP)); 1613 if ((featureflags & DMU_BACKUP_FEATURE_LZ4) && 1614 !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS)) 1615 return (SET_ERROR(ENOTSUP)); 1616 1617 /* 6 extra bytes for /%recv */ 1618 char recvname[ZFS_MAX_DATASET_NAME_LEN + 6]; 1619 1620 (void) snprintf(recvname, sizeof (recvname), "%s/%s", 1621 tofs, recv_clone_name); 1622 1623 if (dsl_dataset_hold(dp, recvname, FTAG, &ds) != 0) { 1624 /* %recv does not exist; continue in tofs */ 1625 error = dsl_dataset_hold(dp, tofs, FTAG, &ds); 1626 if (error != 0) 1627 return (error); 1628 } 1629 1630 /* check that ds is marked inconsistent */ 1631 if (!DS_IS_INCONSISTENT(ds)) { 1632 dsl_dataset_rele(ds, FTAG); 1633 return (SET_ERROR(EINVAL)); 1634 } 1635 1636 /* check that there is resuming data, and that the toguid matches */ 1637 if (!dsl_dataset_is_zapified(ds)) { 1638 dsl_dataset_rele(ds, FTAG); 1639 return (SET_ERROR(EINVAL)); 1640 } 1641 uint64_t val; 1642 error = zap_lookup(dp->dp_meta_objset, ds->ds_object, 1643 DS_FIELD_RESUME_TOGUID, sizeof (val), 1, &val); 1644 if (error != 0 || drrb->drr_toguid != val) { 1645 dsl_dataset_rele(ds, FTAG); 1646 return (SET_ERROR(EINVAL)); 1647 } 1648 1649 /* 1650 * Check if the receive is still running. If so, it will be owned. 1651 * Note that nothing else can own the dataset (e.g. after the receive 1652 * fails) because it will be marked inconsistent. 1653 */ 1654 if (dsl_dataset_has_owner(ds)) { 1655 dsl_dataset_rele(ds, FTAG); 1656 return (SET_ERROR(EBUSY)); 1657 } 1658 1659 /* There should not be any snapshots of this fs yet. */ 1660 if (ds->ds_prev != NULL && ds->ds_prev->ds_dir == ds->ds_dir) { 1661 dsl_dataset_rele(ds, FTAG); 1662 return (SET_ERROR(EINVAL)); 1663 } 1664 1665 /* 1666 * Note: resume point will be checked when we process the first WRITE 1667 * record. 1668 */ 1669 1670 /* check that the origin matches */ 1671 val = 0; 1672 (void) zap_lookup(dp->dp_meta_objset, ds->ds_object, 1673 DS_FIELD_RESUME_FROMGUID, sizeof (val), 1, &val); 1674 if (drrb->drr_fromguid != val) { 1675 dsl_dataset_rele(ds, FTAG); 1676 return (SET_ERROR(EINVAL)); 1677 } 1678 1679 dsl_dataset_rele(ds, FTAG); 1680 return (0); 1681 } 1682 1683 static void 1684 dmu_recv_resume_begin_sync(void *arg, dmu_tx_t *tx) 1685 { 1686 dmu_recv_begin_arg_t *drba = arg; 1687 dsl_pool_t *dp = dmu_tx_pool(tx); 1688 const char *tofs = drba->drba_cookie->drc_tofs; 1689 dsl_dataset_t *ds; 1690 uint64_t dsobj; 1691 /* 6 extra bytes for /%recv */ 1692 char recvname[ZFS_MAX_DATASET_NAME_LEN + 6]; 1693 1694 (void) snprintf(recvname, sizeof (recvname), "%s/%s", 1695 tofs, recv_clone_name); 1696 1697 if (dsl_dataset_hold(dp, recvname, FTAG, &ds) != 0) { 1698 /* %recv does not exist; continue in tofs */ 1699 VERIFY0(dsl_dataset_hold(dp, tofs, FTAG, &ds)); 1700 drba->drba_cookie->drc_newfs = B_TRUE; 1701 } 1702 1703 /* clear the inconsistent flag so that we can own it */ 1704 ASSERT(DS_IS_INCONSISTENT(ds)); 1705 dmu_buf_will_dirty(ds->ds_dbuf, tx); 1706 dsl_dataset_phys(ds)->ds_flags &= ~DS_FLAG_INCONSISTENT; 1707 dsobj = ds->ds_object; 1708 dsl_dataset_rele(ds, FTAG); 1709 1710 VERIFY0(dsl_dataset_own_obj(dp, dsobj, dmu_recv_tag, &ds)); 1711 1712 dmu_buf_will_dirty(ds->ds_dbuf, tx); 1713 dsl_dataset_phys(ds)->ds_flags |= DS_FLAG_INCONSISTENT; 1714 1715 rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG); 1716 ASSERT(!BP_IS_HOLE(dsl_dataset_get_blkptr(ds))); 1717 rrw_exit(&ds->ds_bp_rwlock, FTAG); 1718 1719 drba->drba_cookie->drc_ds = ds; 1720 1721 spa_history_log_internal_ds(ds, "resume receive", tx, ""); 1722 } 1723 1724 /* 1725 * NB: callers *MUST* call dmu_recv_stream() if dmu_recv_begin() 1726 * succeeds; otherwise we will leak the holds on the datasets. 1727 */ 1728 int 1729 dmu_recv_begin(char *tofs, char *tosnap, dmu_replay_record_t *drr_begin, 1730 boolean_t force, boolean_t resumable, char *origin, dmu_recv_cookie_t *drc) 1731 { 1732 dmu_recv_begin_arg_t drba = { 0 }; 1733 1734 bzero(drc, sizeof (dmu_recv_cookie_t)); 1735 drc->drc_drr_begin = drr_begin; 1736 drc->drc_drrb = &drr_begin->drr_u.drr_begin; 1737 drc->drc_tosnap = tosnap; 1738 drc->drc_tofs = tofs; 1739 drc->drc_force = force; 1740 drc->drc_resumable = resumable; 1741 drc->drc_cred = CRED(); 1742 1743 if (drc->drc_drrb->drr_magic == BSWAP_64(DMU_BACKUP_MAGIC)) { 1744 drc->drc_byteswap = B_TRUE; 1745 fletcher_4_incremental_byteswap(drr_begin, 1746 sizeof (dmu_replay_record_t), &drc->drc_cksum); 1747 byteswap_record(drr_begin); 1748 } else if (drc->drc_drrb->drr_magic == DMU_BACKUP_MAGIC) { 1749 fletcher_4_incremental_native(drr_begin, 1750 sizeof (dmu_replay_record_t), &drc->drc_cksum); 1751 } else { 1752 return (SET_ERROR(EINVAL)); 1753 } 1754 1755 drba.drba_origin = origin; 1756 drba.drba_cookie = drc; 1757 drba.drba_cred = CRED(); 1758 1759 if (DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo) & 1760 DMU_BACKUP_FEATURE_RESUMING) { 1761 return (dsl_sync_task(tofs, 1762 dmu_recv_resume_begin_check, dmu_recv_resume_begin_sync, 1763 &drba, 5, ZFS_SPACE_CHECK_NORMAL)); 1764 } else { 1765 return (dsl_sync_task(tofs, 1766 dmu_recv_begin_check, dmu_recv_begin_sync, 1767 &drba, 5, ZFS_SPACE_CHECK_NORMAL)); 1768 } 1769 } 1770 1771 struct receive_record_arg { 1772 dmu_replay_record_t header; 1773 void *payload; /* Pointer to a buffer containing the payload */ 1774 /* 1775 * If the record is a write, pointer to the arc_buf_t containing the 1776 * payload. 1777 */ 1778 arc_buf_t *write_buf; 1779 int payload_size; 1780 uint64_t bytes_read; /* bytes read from stream when record created */ 1781 boolean_t eos_marker; /* Marks the end of the stream */ 1782 bqueue_node_t node; 1783 }; 1784 1785 struct receive_writer_arg { 1786 objset_t *os; 1787 boolean_t byteswap; 1788 bqueue_t q; 1789 1790 /* 1791 * These three args are used to signal to the main thread that we're 1792 * done. 1793 */ 1794 kmutex_t mutex; 1795 kcondvar_t cv; 1796 boolean_t done; 1797 1798 int err; 1799 /* A map from guid to dataset to help handle dedup'd streams. */ 1800 avl_tree_t *guid_to_ds_map; 1801 boolean_t resumable; 1802 uint64_t last_object, last_offset; 1803 uint64_t bytes_read; /* bytes read when current record created */ 1804 }; 1805 1806 struct objlist { 1807 list_t list; /* List of struct receive_objnode. */ 1808 /* 1809 * Last object looked up. Used to assert that objects are being looked 1810 * up in ascending order. 1811 */ 1812 uint64_t last_lookup; 1813 }; 1814 1815 struct receive_objnode { 1816 list_node_t node; 1817 uint64_t object; 1818 }; 1819 1820 struct receive_arg { 1821 objset_t *os; 1822 vnode_t *vp; /* The vnode to read the stream from */ 1823 uint64_t voff; /* The current offset in the stream */ 1824 uint64_t bytes_read; 1825 /* 1826 * A record that has had its payload read in, but hasn't yet been handed 1827 * off to the worker thread. 1828 */ 1829 struct receive_record_arg *rrd; 1830 /* A record that has had its header read in, but not its payload. */ 1831 struct receive_record_arg *next_rrd; 1832 zio_cksum_t cksum; 1833 zio_cksum_t prev_cksum; 1834 int err; 1835 boolean_t byteswap; 1836 /* Sorted list of objects not to issue prefetches for. */ 1837 struct objlist ignore_objlist; 1838 }; 1839 1840 typedef struct guid_map_entry { 1841 uint64_t guid; 1842 dsl_dataset_t *gme_ds; 1843 avl_node_t avlnode; 1844 } guid_map_entry_t; 1845 1846 static int 1847 guid_compare(const void *arg1, const void *arg2) 1848 { 1849 const guid_map_entry_t *gmep1 = arg1; 1850 const guid_map_entry_t *gmep2 = arg2; 1851 1852 if (gmep1->guid < gmep2->guid) 1853 return (-1); 1854 else if (gmep1->guid > gmep2->guid) 1855 return (1); 1856 return (0); 1857 } 1858 1859 static void 1860 free_guid_map_onexit(void *arg) 1861 { 1862 avl_tree_t *ca = arg; 1863 void *cookie = NULL; 1864 guid_map_entry_t *gmep; 1865 1866 while ((gmep = avl_destroy_nodes(ca, &cookie)) != NULL) { 1867 dsl_dataset_long_rele(gmep->gme_ds, gmep); 1868 dsl_dataset_rele(gmep->gme_ds, gmep); 1869 kmem_free(gmep, sizeof (guid_map_entry_t)); 1870 } 1871 avl_destroy(ca); 1872 kmem_free(ca, sizeof (avl_tree_t)); 1873 } 1874 1875 static int 1876 receive_read(struct receive_arg *ra, int len, void *buf) 1877 { 1878 int done = 0; 1879 1880 /* 1881 * The code doesn't rely on this (lengths being multiples of 8). See 1882 * comment in dump_bytes. 1883 */ 1884 ASSERT0(len % 8); 1885 1886 while (done < len) { 1887 ssize_t resid; 1888 1889 ra->err = vn_rdwr(UIO_READ, ra->vp, 1890 (char *)buf + done, len - done, 1891 ra->voff, UIO_SYSSPACE, FAPPEND, 1892 RLIM64_INFINITY, CRED(), &resid); 1893 1894 if (resid == len - done) { 1895 /* 1896 * Note: ECKSUM indicates that the receive 1897 * was interrupted and can potentially be resumed. 1898 */ 1899 ra->err = SET_ERROR(ECKSUM); 1900 } 1901 ra->voff += len - done - resid; 1902 done = len - resid; 1903 if (ra->err != 0) 1904 return (ra->err); 1905 } 1906 1907 ra->bytes_read += len; 1908 1909 ASSERT3U(done, ==, len); 1910 return (0); 1911 } 1912 1913 static void 1914 byteswap_record(dmu_replay_record_t *drr) 1915 { 1916 #define DO64(X) (drr->drr_u.X = BSWAP_64(drr->drr_u.X)) 1917 #define DO32(X) (drr->drr_u.X = BSWAP_32(drr->drr_u.X)) 1918 drr->drr_type = BSWAP_32(drr->drr_type); 1919 drr->drr_payloadlen = BSWAP_32(drr->drr_payloadlen); 1920 1921 switch (drr->drr_type) { 1922 case DRR_BEGIN: 1923 DO64(drr_begin.drr_magic); 1924 DO64(drr_begin.drr_versioninfo); 1925 DO64(drr_begin.drr_creation_time); 1926 DO32(drr_begin.drr_type); 1927 DO32(drr_begin.drr_flags); 1928 DO64(drr_begin.drr_toguid); 1929 DO64(drr_begin.drr_fromguid); 1930 break; 1931 case DRR_OBJECT: 1932 DO64(drr_object.drr_object); 1933 DO32(drr_object.drr_type); 1934 DO32(drr_object.drr_bonustype); 1935 DO32(drr_object.drr_blksz); 1936 DO32(drr_object.drr_bonuslen); 1937 DO64(drr_object.drr_toguid); 1938 break; 1939 case DRR_FREEOBJECTS: 1940 DO64(drr_freeobjects.drr_firstobj); 1941 DO64(drr_freeobjects.drr_numobjs); 1942 DO64(drr_freeobjects.drr_toguid); 1943 break; 1944 case DRR_WRITE: 1945 DO64(drr_write.drr_object); 1946 DO32(drr_write.drr_type); 1947 DO64(drr_write.drr_offset); 1948 DO64(drr_write.drr_logical_size); 1949 DO64(drr_write.drr_toguid); 1950 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_write.drr_key.ddk_cksum); 1951 DO64(drr_write.drr_key.ddk_prop); 1952 DO64(drr_write.drr_compressed_size); 1953 break; 1954 case DRR_WRITE_BYREF: 1955 DO64(drr_write_byref.drr_object); 1956 DO64(drr_write_byref.drr_offset); 1957 DO64(drr_write_byref.drr_length); 1958 DO64(drr_write_byref.drr_toguid); 1959 DO64(drr_write_byref.drr_refguid); 1960 DO64(drr_write_byref.drr_refobject); 1961 DO64(drr_write_byref.drr_refoffset); 1962 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_write_byref. 1963 drr_key.ddk_cksum); 1964 DO64(drr_write_byref.drr_key.ddk_prop); 1965 break; 1966 case DRR_WRITE_EMBEDDED: 1967 DO64(drr_write_embedded.drr_object); 1968 DO64(drr_write_embedded.drr_offset); 1969 DO64(drr_write_embedded.drr_length); 1970 DO64(drr_write_embedded.drr_toguid); 1971 DO32(drr_write_embedded.drr_lsize); 1972 DO32(drr_write_embedded.drr_psize); 1973 break; 1974 case DRR_FREE: 1975 DO64(drr_free.drr_object); 1976 DO64(drr_free.drr_offset); 1977 DO64(drr_free.drr_length); 1978 DO64(drr_free.drr_toguid); 1979 break; 1980 case DRR_SPILL: 1981 DO64(drr_spill.drr_object); 1982 DO64(drr_spill.drr_length); 1983 DO64(drr_spill.drr_toguid); 1984 break; 1985 case DRR_END: 1986 DO64(drr_end.drr_toguid); 1987 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_end.drr_checksum); 1988 break; 1989 } 1990 1991 if (drr->drr_type != DRR_BEGIN) { 1992 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_checksum.drr_checksum); 1993 } 1994 1995 #undef DO64 1996 #undef DO32 1997 } 1998 1999 static inline uint8_t 2000 deduce_nblkptr(dmu_object_type_t bonus_type, uint64_t bonus_size) 2001 { 2002 if (bonus_type == DMU_OT_SA) { 2003 return (1); 2004 } else { 2005 return (1 + 2006 ((DN_MAX_BONUSLEN - bonus_size) >> SPA_BLKPTRSHIFT)); 2007 } 2008 } 2009 2010 static void 2011 save_resume_state(struct receive_writer_arg *rwa, 2012 uint64_t object, uint64_t offset, dmu_tx_t *tx) 2013 { 2014 int txgoff = dmu_tx_get_txg(tx) & TXG_MASK; 2015 2016 if (!rwa->resumable) 2017 return; 2018 2019 /* 2020 * We use ds_resume_bytes[] != 0 to indicate that we need to 2021 * update this on disk, so it must not be 0. 2022 */ 2023 ASSERT(rwa->bytes_read != 0); 2024 2025 /* 2026 * We only resume from write records, which have a valid 2027 * (non-meta-dnode) object number. 2028 */ 2029 ASSERT(object != 0); 2030 2031 /* 2032 * For resuming to work correctly, we must receive records in order, 2033 * sorted by object,offset. This is checked by the callers, but 2034 * assert it here for good measure. 2035 */ 2036 ASSERT3U(object, >=, rwa->os->os_dsl_dataset->ds_resume_object[txgoff]); 2037 ASSERT(object != rwa->os->os_dsl_dataset->ds_resume_object[txgoff] || 2038 offset >= rwa->os->os_dsl_dataset->ds_resume_offset[txgoff]); 2039 ASSERT3U(rwa->bytes_read, >=, 2040 rwa->os->os_dsl_dataset->ds_resume_bytes[txgoff]); 2041 2042 rwa->os->os_dsl_dataset->ds_resume_object[txgoff] = object; 2043 rwa->os->os_dsl_dataset->ds_resume_offset[txgoff] = offset; 2044 rwa->os->os_dsl_dataset->ds_resume_bytes[txgoff] = rwa->bytes_read; 2045 } 2046 2047 static int 2048 receive_object(struct receive_writer_arg *rwa, struct drr_object *drro, 2049 void *data) 2050 { 2051 dmu_object_info_t doi; 2052 dmu_tx_t *tx; 2053 uint64_t object; 2054 int err; 2055 2056 if (drro->drr_type == DMU_OT_NONE || 2057 !DMU_OT_IS_VALID(drro->drr_type) || 2058 !DMU_OT_IS_VALID(drro->drr_bonustype) || 2059 drro->drr_checksumtype >= ZIO_CHECKSUM_FUNCTIONS || 2060 drro->drr_compress >= ZIO_COMPRESS_FUNCTIONS || 2061 P2PHASE(drro->drr_blksz, SPA_MINBLOCKSIZE) || 2062 drro->drr_blksz < SPA_MINBLOCKSIZE || 2063 drro->drr_blksz > spa_maxblocksize(dmu_objset_spa(rwa->os)) || 2064 drro->drr_bonuslen > DN_MAX_BONUSLEN) { 2065 return (SET_ERROR(EINVAL)); 2066 } 2067 2068 err = dmu_object_info(rwa->os, drro->drr_object, &doi); 2069 2070 if (err != 0 && err != ENOENT) 2071 return (SET_ERROR(EINVAL)); 2072 object = err == 0 ? drro->drr_object : DMU_NEW_OBJECT; 2073 2074 /* 2075 * If we are losing blkptrs or changing the block size this must 2076 * be a new file instance. We must clear out the previous file 2077 * contents before we can change this type of metadata in the dnode. 2078 */ 2079 if (err == 0) { 2080 int nblkptr; 2081 2082 nblkptr = deduce_nblkptr(drro->drr_bonustype, 2083 drro->drr_bonuslen); 2084 2085 if (drro->drr_blksz != doi.doi_data_block_size || 2086 nblkptr < doi.doi_nblkptr) { 2087 err = dmu_free_long_range(rwa->os, drro->drr_object, 2088 0, DMU_OBJECT_END); 2089 if (err != 0) 2090 return (SET_ERROR(EINVAL)); 2091 } 2092 } 2093 2094 tx = dmu_tx_create(rwa->os); 2095 dmu_tx_hold_bonus(tx, object); 2096 err = dmu_tx_assign(tx, TXG_WAIT); 2097 if (err != 0) { 2098 dmu_tx_abort(tx); 2099 return (err); 2100 } 2101 2102 if (object == DMU_NEW_OBJECT) { 2103 /* currently free, want to be allocated */ 2104 err = dmu_object_claim(rwa->os, drro->drr_object, 2105 drro->drr_type, drro->drr_blksz, 2106 drro->drr_bonustype, drro->drr_bonuslen, tx); 2107 } else if (drro->drr_type != doi.doi_type || 2108 drro->drr_blksz != doi.doi_data_block_size || 2109 drro->drr_bonustype != doi.doi_bonus_type || 2110 drro->drr_bonuslen != doi.doi_bonus_size) { 2111 /* currently allocated, but with different properties */ 2112 err = dmu_object_reclaim(rwa->os, drro->drr_object, 2113 drro->drr_type, drro->drr_blksz, 2114 drro->drr_bonustype, drro->drr_bonuslen, tx); 2115 } 2116 if (err != 0) { 2117 dmu_tx_commit(tx); 2118 return (SET_ERROR(EINVAL)); 2119 } 2120 2121 dmu_object_set_checksum(rwa->os, drro->drr_object, 2122 drro->drr_checksumtype, tx); 2123 dmu_object_set_compress(rwa->os, drro->drr_object, 2124 drro->drr_compress, tx); 2125 2126 if (data != NULL) { 2127 dmu_buf_t *db; 2128 2129 VERIFY0(dmu_bonus_hold(rwa->os, drro->drr_object, FTAG, &db)); 2130 dmu_buf_will_dirty(db, tx); 2131 2132 ASSERT3U(db->db_size, >=, drro->drr_bonuslen); 2133 bcopy(data, db->db_data, drro->drr_bonuslen); 2134 if (rwa->byteswap) { 2135 dmu_object_byteswap_t byteswap = 2136 DMU_OT_BYTESWAP(drro->drr_bonustype); 2137 dmu_ot_byteswap[byteswap].ob_func(db->db_data, 2138 drro->drr_bonuslen); 2139 } 2140 dmu_buf_rele(db, FTAG); 2141 } 2142 dmu_tx_commit(tx); 2143 2144 return (0); 2145 } 2146 2147 /* ARGSUSED */ 2148 static int 2149 receive_freeobjects(struct receive_writer_arg *rwa, 2150 struct drr_freeobjects *drrfo) 2151 { 2152 uint64_t obj; 2153 int next_err = 0; 2154 2155 if (drrfo->drr_firstobj + drrfo->drr_numobjs < drrfo->drr_firstobj) 2156 return (SET_ERROR(EINVAL)); 2157 2158 for (obj = drrfo->drr_firstobj; 2159 obj < drrfo->drr_firstobj + drrfo->drr_numobjs && next_err == 0; 2160 next_err = dmu_object_next(rwa->os, &obj, FALSE, 0)) { 2161 int err; 2162 2163 if (dmu_object_info(rwa->os, obj, NULL) != 0) 2164 continue; 2165 2166 err = dmu_free_long_object(rwa->os, obj); 2167 if (err != 0) 2168 return (err); 2169 } 2170 if (next_err != ESRCH) 2171 return (next_err); 2172 return (0); 2173 } 2174 2175 static int 2176 receive_write(struct receive_writer_arg *rwa, struct drr_write *drrw, 2177 arc_buf_t *abuf) 2178 { 2179 dmu_tx_t *tx; 2180 int err; 2181 2182 if (drrw->drr_offset + drrw->drr_logical_size < drrw->drr_offset || 2183 !DMU_OT_IS_VALID(drrw->drr_type)) 2184 return (SET_ERROR(EINVAL)); 2185 2186 /* 2187 * For resuming to work, records must be in increasing order 2188 * by (object, offset). 2189 */ 2190 if (drrw->drr_object < rwa->last_object || 2191 (drrw->drr_object == rwa->last_object && 2192 drrw->drr_offset < rwa->last_offset)) { 2193 return (SET_ERROR(EINVAL)); 2194 } 2195 rwa->last_object = drrw->drr_object; 2196 rwa->last_offset = drrw->drr_offset; 2197 2198 if (dmu_object_info(rwa->os, drrw->drr_object, NULL) != 0) 2199 return (SET_ERROR(EINVAL)); 2200 2201 tx = dmu_tx_create(rwa->os); 2202 2203 dmu_tx_hold_write(tx, drrw->drr_object, 2204 drrw->drr_offset, drrw->drr_logical_size); 2205 err = dmu_tx_assign(tx, TXG_WAIT); 2206 if (err != 0) { 2207 dmu_tx_abort(tx); 2208 return (err); 2209 } 2210 if (rwa->byteswap) { 2211 dmu_object_byteswap_t byteswap = 2212 DMU_OT_BYTESWAP(drrw->drr_type); 2213 dmu_ot_byteswap[byteswap].ob_func(abuf->b_data, 2214 DRR_WRITE_PAYLOAD_SIZE(drrw)); 2215 } 2216 2217 /* use the bonus buf to look up the dnode in dmu_assign_arcbuf */ 2218 dmu_buf_t *bonus; 2219 if (dmu_bonus_hold(rwa->os, drrw->drr_object, FTAG, &bonus) != 0) 2220 return (SET_ERROR(EINVAL)); 2221 dmu_assign_arcbuf(bonus, drrw->drr_offset, abuf, tx); 2222 2223 /* 2224 * Note: If the receive fails, we want the resume stream to start 2225 * with the same record that we last successfully received (as opposed 2226 * to the next record), so that we can verify that we are 2227 * resuming from the correct location. 2228 */ 2229 save_resume_state(rwa, drrw->drr_object, drrw->drr_offset, tx); 2230 dmu_tx_commit(tx); 2231 dmu_buf_rele(bonus, FTAG); 2232 2233 return (0); 2234 } 2235 2236 /* 2237 * Handle a DRR_WRITE_BYREF record. This record is used in dedup'ed 2238 * streams to refer to a copy of the data that is already on the 2239 * system because it came in earlier in the stream. This function 2240 * finds the earlier copy of the data, and uses that copy instead of 2241 * data from the stream to fulfill this write. 2242 */ 2243 static int 2244 receive_write_byref(struct receive_writer_arg *rwa, 2245 struct drr_write_byref *drrwbr) 2246 { 2247 dmu_tx_t *tx; 2248 int err; 2249 guid_map_entry_t gmesrch; 2250 guid_map_entry_t *gmep; 2251 avl_index_t where; 2252 objset_t *ref_os = NULL; 2253 dmu_buf_t *dbp; 2254 2255 if (drrwbr->drr_offset + drrwbr->drr_length < drrwbr->drr_offset) 2256 return (SET_ERROR(EINVAL)); 2257 2258 /* 2259 * If the GUID of the referenced dataset is different from the 2260 * GUID of the target dataset, find the referenced dataset. 2261 */ 2262 if (drrwbr->drr_toguid != drrwbr->drr_refguid) { 2263 gmesrch.guid = drrwbr->drr_refguid; 2264 if ((gmep = avl_find(rwa->guid_to_ds_map, &gmesrch, 2265 &where)) == NULL) { 2266 return (SET_ERROR(EINVAL)); 2267 } 2268 if (dmu_objset_from_ds(gmep->gme_ds, &ref_os)) 2269 return (SET_ERROR(EINVAL)); 2270 } else { 2271 ref_os = rwa->os; 2272 } 2273 2274 err = dmu_buf_hold(ref_os, drrwbr->drr_refobject, 2275 drrwbr->drr_refoffset, FTAG, &dbp, DMU_READ_PREFETCH); 2276 if (err != 0) 2277 return (err); 2278 2279 tx = dmu_tx_create(rwa->os); 2280 2281 dmu_tx_hold_write(tx, drrwbr->drr_object, 2282 drrwbr->drr_offset, drrwbr->drr_length); 2283 err = dmu_tx_assign(tx, TXG_WAIT); 2284 if (err != 0) { 2285 dmu_tx_abort(tx); 2286 return (err); 2287 } 2288 dmu_write(rwa->os, drrwbr->drr_object, 2289 drrwbr->drr_offset, drrwbr->drr_length, dbp->db_data, tx); 2290 dmu_buf_rele(dbp, FTAG); 2291 2292 /* See comment in restore_write. */ 2293 save_resume_state(rwa, drrwbr->drr_object, drrwbr->drr_offset, tx); 2294 dmu_tx_commit(tx); 2295 return (0); 2296 } 2297 2298 static int 2299 receive_write_embedded(struct receive_writer_arg *rwa, 2300 struct drr_write_embedded *drrwe, void *data) 2301 { 2302 dmu_tx_t *tx; 2303 int err; 2304 2305 if (drrwe->drr_offset + drrwe->drr_length < drrwe->drr_offset) 2306 return (EINVAL); 2307 2308 if (drrwe->drr_psize > BPE_PAYLOAD_SIZE) 2309 return (EINVAL); 2310 2311 if (drrwe->drr_etype >= NUM_BP_EMBEDDED_TYPES) 2312 return (EINVAL); 2313 if (drrwe->drr_compression >= ZIO_COMPRESS_FUNCTIONS) 2314 return (EINVAL); 2315 2316 tx = dmu_tx_create(rwa->os); 2317 2318 dmu_tx_hold_write(tx, drrwe->drr_object, 2319 drrwe->drr_offset, drrwe->drr_length); 2320 err = dmu_tx_assign(tx, TXG_WAIT); 2321 if (err != 0) { 2322 dmu_tx_abort(tx); 2323 return (err); 2324 } 2325 2326 dmu_write_embedded(rwa->os, drrwe->drr_object, 2327 drrwe->drr_offset, data, drrwe->drr_etype, 2328 drrwe->drr_compression, drrwe->drr_lsize, drrwe->drr_psize, 2329 rwa->byteswap ^ ZFS_HOST_BYTEORDER, tx); 2330 2331 /* See comment in restore_write. */ 2332 save_resume_state(rwa, drrwe->drr_object, drrwe->drr_offset, tx); 2333 dmu_tx_commit(tx); 2334 return (0); 2335 } 2336 2337 static int 2338 receive_spill(struct receive_writer_arg *rwa, struct drr_spill *drrs, 2339 void *data) 2340 { 2341 dmu_tx_t *tx; 2342 dmu_buf_t *db, *db_spill; 2343 int err; 2344 2345 if (drrs->drr_length < SPA_MINBLOCKSIZE || 2346 drrs->drr_length > spa_maxblocksize(dmu_objset_spa(rwa->os))) 2347 return (SET_ERROR(EINVAL)); 2348 2349 if (dmu_object_info(rwa->os, drrs->drr_object, NULL) != 0) 2350 return (SET_ERROR(EINVAL)); 2351 2352 VERIFY0(dmu_bonus_hold(rwa->os, drrs->drr_object, FTAG, &db)); 2353 if ((err = dmu_spill_hold_by_bonus(db, FTAG, &db_spill)) != 0) { 2354 dmu_buf_rele(db, FTAG); 2355 return (err); 2356 } 2357 2358 tx = dmu_tx_create(rwa->os); 2359 2360 dmu_tx_hold_spill(tx, db->db_object); 2361 2362 err = dmu_tx_assign(tx, TXG_WAIT); 2363 if (err != 0) { 2364 dmu_buf_rele(db, FTAG); 2365 dmu_buf_rele(db_spill, FTAG); 2366 dmu_tx_abort(tx); 2367 return (err); 2368 } 2369 dmu_buf_will_dirty(db_spill, tx); 2370 2371 if (db_spill->db_size < drrs->drr_length) 2372 VERIFY(0 == dbuf_spill_set_blksz(db_spill, 2373 drrs->drr_length, tx)); 2374 bcopy(data, db_spill->db_data, drrs->drr_length); 2375 2376 dmu_buf_rele(db, FTAG); 2377 dmu_buf_rele(db_spill, FTAG); 2378 2379 dmu_tx_commit(tx); 2380 return (0); 2381 } 2382 2383 /* ARGSUSED */ 2384 static int 2385 receive_free(struct receive_writer_arg *rwa, struct drr_free *drrf) 2386 { 2387 int err; 2388 2389 if (drrf->drr_length != -1ULL && 2390 drrf->drr_offset + drrf->drr_length < drrf->drr_offset) 2391 return (SET_ERROR(EINVAL)); 2392 2393 if (dmu_object_info(rwa->os, drrf->drr_object, NULL) != 0) 2394 return (SET_ERROR(EINVAL)); 2395 2396 err = dmu_free_long_range(rwa->os, drrf->drr_object, 2397 drrf->drr_offset, drrf->drr_length); 2398 2399 return (err); 2400 } 2401 2402 /* used to destroy the drc_ds on error */ 2403 static void 2404 dmu_recv_cleanup_ds(dmu_recv_cookie_t *drc) 2405 { 2406 if (drc->drc_resumable) { 2407 /* wait for our resume state to be written to disk */ 2408 txg_wait_synced(drc->drc_ds->ds_dir->dd_pool, 0); 2409 dsl_dataset_disown(drc->drc_ds, dmu_recv_tag); 2410 } else { 2411 char name[ZFS_MAX_DATASET_NAME_LEN]; 2412 dsl_dataset_name(drc->drc_ds, name); 2413 dsl_dataset_disown(drc->drc_ds, dmu_recv_tag); 2414 (void) dsl_destroy_head(name); 2415 } 2416 } 2417 2418 static void 2419 receive_cksum(struct receive_arg *ra, int len, void *buf) 2420 { 2421 if (ra->byteswap) { 2422 fletcher_4_incremental_byteswap(buf, len, &ra->cksum); 2423 } else { 2424 fletcher_4_incremental_native(buf, len, &ra->cksum); 2425 } 2426 } 2427 2428 /* 2429 * Read the payload into a buffer of size len, and update the current record's 2430 * payload field. 2431 * Allocate ra->next_rrd and read the next record's header into 2432 * ra->next_rrd->header. 2433 * Verify checksum of payload and next record. 2434 */ 2435 static int 2436 receive_read_payload_and_next_header(struct receive_arg *ra, int len, void *buf) 2437 { 2438 int err; 2439 2440 if (len != 0) { 2441 ASSERT3U(len, <=, SPA_MAXBLOCKSIZE); 2442 err = receive_read(ra, len, buf); 2443 if (err != 0) 2444 return (err); 2445 receive_cksum(ra, len, buf); 2446 2447 /* note: rrd is NULL when reading the begin record's payload */ 2448 if (ra->rrd != NULL) { 2449 ra->rrd->payload = buf; 2450 ra->rrd->payload_size = len; 2451 ra->rrd->bytes_read = ra->bytes_read; 2452 } 2453 } 2454 2455 ra->prev_cksum = ra->cksum; 2456 2457 ra->next_rrd = kmem_zalloc(sizeof (*ra->next_rrd), KM_SLEEP); 2458 err = receive_read(ra, sizeof (ra->next_rrd->header), 2459 &ra->next_rrd->header); 2460 ra->next_rrd->bytes_read = ra->bytes_read; 2461 if (err != 0) { 2462 kmem_free(ra->next_rrd, sizeof (*ra->next_rrd)); 2463 ra->next_rrd = NULL; 2464 return (err); 2465 } 2466 if (ra->next_rrd->header.drr_type == DRR_BEGIN) { 2467 kmem_free(ra->next_rrd, sizeof (*ra->next_rrd)); 2468 ra->next_rrd = NULL; 2469 return (SET_ERROR(EINVAL)); 2470 } 2471 2472 /* 2473 * Note: checksum is of everything up to but not including the 2474 * checksum itself. 2475 */ 2476 ASSERT3U(offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum), 2477 ==, sizeof (dmu_replay_record_t) - sizeof (zio_cksum_t)); 2478 receive_cksum(ra, 2479 offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum), 2480 &ra->next_rrd->header); 2481 2482 zio_cksum_t cksum_orig = 2483 ra->next_rrd->header.drr_u.drr_checksum.drr_checksum; 2484 zio_cksum_t *cksump = 2485 &ra->next_rrd->header.drr_u.drr_checksum.drr_checksum; 2486 2487 if (ra->byteswap) 2488 byteswap_record(&ra->next_rrd->header); 2489 2490 if ((!ZIO_CHECKSUM_IS_ZERO(cksump)) && 2491 !ZIO_CHECKSUM_EQUAL(ra->cksum, *cksump)) { 2492 kmem_free(ra->next_rrd, sizeof (*ra->next_rrd)); 2493 ra->next_rrd = NULL; 2494 return (SET_ERROR(ECKSUM)); 2495 } 2496 2497 receive_cksum(ra, sizeof (cksum_orig), &cksum_orig); 2498 2499 return (0); 2500 } 2501 2502 static void 2503 objlist_create(struct objlist *list) 2504 { 2505 list_create(&list->list, sizeof (struct receive_objnode), 2506 offsetof(struct receive_objnode, node)); 2507 list->last_lookup = 0; 2508 } 2509 2510 static void 2511 objlist_destroy(struct objlist *list) 2512 { 2513 for (struct receive_objnode *n = list_remove_head(&list->list); 2514 n != NULL; n = list_remove_head(&list->list)) { 2515 kmem_free(n, sizeof (*n)); 2516 } 2517 list_destroy(&list->list); 2518 } 2519 2520 /* 2521 * This function looks through the objlist to see if the specified object number 2522 * is contained in the objlist. In the process, it will remove all object 2523 * numbers in the list that are smaller than the specified object number. Thus, 2524 * any lookup of an object number smaller than a previously looked up object 2525 * number will always return false; therefore, all lookups should be done in 2526 * ascending order. 2527 */ 2528 static boolean_t 2529 objlist_exists(struct objlist *list, uint64_t object) 2530 { 2531 struct receive_objnode *node = list_head(&list->list); 2532 ASSERT3U(object, >=, list->last_lookup); 2533 list->last_lookup = object; 2534 while (node != NULL && node->object < object) { 2535 VERIFY3P(node, ==, list_remove_head(&list->list)); 2536 kmem_free(node, sizeof (*node)); 2537 node = list_head(&list->list); 2538 } 2539 return (node != NULL && node->object == object); 2540 } 2541 2542 /* 2543 * The objlist is a list of object numbers stored in ascending order. However, 2544 * the insertion of new object numbers does not seek out the correct location to 2545 * store a new object number; instead, it appends it to the list for simplicity. 2546 * Thus, any users must take care to only insert new object numbers in ascending 2547 * order. 2548 */ 2549 static void 2550 objlist_insert(struct objlist *list, uint64_t object) 2551 { 2552 struct receive_objnode *node = kmem_zalloc(sizeof (*node), KM_SLEEP); 2553 node->object = object; 2554 #ifdef ZFS_DEBUG 2555 struct receive_objnode *last_object = list_tail(&list->list); 2556 uint64_t last_objnum = (last_object != NULL ? last_object->object : 0); 2557 ASSERT3U(node->object, >, last_objnum); 2558 #endif 2559 list_insert_tail(&list->list, node); 2560 } 2561 2562 /* 2563 * Issue the prefetch reads for any necessary indirect blocks. 2564 * 2565 * We use the object ignore list to tell us whether or not to issue prefetches 2566 * for a given object. We do this for both correctness (in case the blocksize 2567 * of an object has changed) and performance (if the object doesn't exist, don't 2568 * needlessly try to issue prefetches). We also trim the list as we go through 2569 * the stream to prevent it from growing to an unbounded size. 2570 * 2571 * The object numbers within will always be in sorted order, and any write 2572 * records we see will also be in sorted order, but they're not sorted with 2573 * respect to each other (i.e. we can get several object records before 2574 * receiving each object's write records). As a result, once we've reached a 2575 * given object number, we can safely remove any reference to lower object 2576 * numbers in the ignore list. In practice, we receive up to 32 object records 2577 * before receiving write records, so the list can have up to 32 nodes in it. 2578 */ 2579 /* ARGSUSED */ 2580 static void 2581 receive_read_prefetch(struct receive_arg *ra, 2582 uint64_t object, uint64_t offset, uint64_t length) 2583 { 2584 if (!objlist_exists(&ra->ignore_objlist, object)) { 2585 dmu_prefetch(ra->os, object, 1, offset, length, 2586 ZIO_PRIORITY_SYNC_READ); 2587 } 2588 } 2589 2590 /* 2591 * Read records off the stream, issuing any necessary prefetches. 2592 */ 2593 static int 2594 receive_read_record(struct receive_arg *ra) 2595 { 2596 int err; 2597 2598 switch (ra->rrd->header.drr_type) { 2599 case DRR_OBJECT: 2600 { 2601 struct drr_object *drro = &ra->rrd->header.drr_u.drr_object; 2602 uint32_t size = P2ROUNDUP(drro->drr_bonuslen, 8); 2603 void *buf = kmem_zalloc(size, KM_SLEEP); 2604 dmu_object_info_t doi; 2605 err = receive_read_payload_and_next_header(ra, size, buf); 2606 if (err != 0) { 2607 kmem_free(buf, size); 2608 return (err); 2609 } 2610 err = dmu_object_info(ra->os, drro->drr_object, &doi); 2611 /* 2612 * See receive_read_prefetch for an explanation why we're 2613 * storing this object in the ignore_obj_list. 2614 */ 2615 if (err == ENOENT || 2616 (err == 0 && doi.doi_data_block_size != drro->drr_blksz)) { 2617 objlist_insert(&ra->ignore_objlist, drro->drr_object); 2618 err = 0; 2619 } 2620 return (err); 2621 } 2622 case DRR_FREEOBJECTS: 2623 { 2624 err = receive_read_payload_and_next_header(ra, 0, NULL); 2625 return (err); 2626 } 2627 case DRR_WRITE: 2628 { 2629 struct drr_write *drrw = &ra->rrd->header.drr_u.drr_write; 2630 arc_buf_t *abuf; 2631 boolean_t is_meta = DMU_OT_IS_METADATA(drrw->drr_type); 2632 if (DRR_WRITE_COMPRESSED(drrw)) { 2633 ASSERT3U(drrw->drr_compressed_size, >, 0); 2634 ASSERT3U(drrw->drr_logical_size, >=, 2635 drrw->drr_compressed_size); 2636 ASSERT(!is_meta); 2637 abuf = arc_loan_compressed_buf( 2638 dmu_objset_spa(ra->os), 2639 drrw->drr_compressed_size, drrw->drr_logical_size, 2640 drrw->drr_compressiontype); 2641 } else { 2642 abuf = arc_loan_buf(dmu_objset_spa(ra->os), 2643 is_meta, drrw->drr_logical_size); 2644 } 2645 2646 err = receive_read_payload_and_next_header(ra, 2647 DRR_WRITE_PAYLOAD_SIZE(drrw), abuf->b_data); 2648 if (err != 0) { 2649 dmu_return_arcbuf(abuf); 2650 return (err); 2651 } 2652 ra->rrd->write_buf = abuf; 2653 receive_read_prefetch(ra, drrw->drr_object, drrw->drr_offset, 2654 drrw->drr_logical_size); 2655 return (err); 2656 } 2657 case DRR_WRITE_BYREF: 2658 { 2659 struct drr_write_byref *drrwb = 2660 &ra->rrd->header.drr_u.drr_write_byref; 2661 err = receive_read_payload_and_next_header(ra, 0, NULL); 2662 receive_read_prefetch(ra, drrwb->drr_object, drrwb->drr_offset, 2663 drrwb->drr_length); 2664 return (err); 2665 } 2666 case DRR_WRITE_EMBEDDED: 2667 { 2668 struct drr_write_embedded *drrwe = 2669 &ra->rrd->header.drr_u.drr_write_embedded; 2670 uint32_t size = P2ROUNDUP(drrwe->drr_psize, 8); 2671 void *buf = kmem_zalloc(size, KM_SLEEP); 2672 2673 err = receive_read_payload_and_next_header(ra, size, buf); 2674 if (err != 0) { 2675 kmem_free(buf, size); 2676 return (err); 2677 } 2678 2679 receive_read_prefetch(ra, drrwe->drr_object, drrwe->drr_offset, 2680 drrwe->drr_length); 2681 return (err); 2682 } 2683 case DRR_FREE: 2684 { 2685 /* 2686 * It might be beneficial to prefetch indirect blocks here, but 2687 * we don't really have the data to decide for sure. 2688 */ 2689 err = receive_read_payload_and_next_header(ra, 0, NULL); 2690 return (err); 2691 } 2692 case DRR_END: 2693 { 2694 struct drr_end *drre = &ra->rrd->header.drr_u.drr_end; 2695 if (!ZIO_CHECKSUM_EQUAL(ra->prev_cksum, drre->drr_checksum)) 2696 return (SET_ERROR(ECKSUM)); 2697 return (0); 2698 } 2699 case DRR_SPILL: 2700 { 2701 struct drr_spill *drrs = &ra->rrd->header.drr_u.drr_spill; 2702 void *buf = kmem_zalloc(drrs->drr_length, KM_SLEEP); 2703 err = receive_read_payload_and_next_header(ra, drrs->drr_length, 2704 buf); 2705 if (err != 0) 2706 kmem_free(buf, drrs->drr_length); 2707 return (err); 2708 } 2709 default: 2710 return (SET_ERROR(EINVAL)); 2711 } 2712 } 2713 2714 /* 2715 * Commit the records to the pool. 2716 */ 2717 static int 2718 receive_process_record(struct receive_writer_arg *rwa, 2719 struct receive_record_arg *rrd) 2720 { 2721 int err; 2722 2723 /* Processing in order, therefore bytes_read should be increasing. */ 2724 ASSERT3U(rrd->bytes_read, >=, rwa->bytes_read); 2725 rwa->bytes_read = rrd->bytes_read; 2726 2727 switch (rrd->header.drr_type) { 2728 case DRR_OBJECT: 2729 { 2730 struct drr_object *drro = &rrd->header.drr_u.drr_object; 2731 err = receive_object(rwa, drro, rrd->payload); 2732 kmem_free(rrd->payload, rrd->payload_size); 2733 rrd->payload = NULL; 2734 return (err); 2735 } 2736 case DRR_FREEOBJECTS: 2737 { 2738 struct drr_freeobjects *drrfo = 2739 &rrd->header.drr_u.drr_freeobjects; 2740 return (receive_freeobjects(rwa, drrfo)); 2741 } 2742 case DRR_WRITE: 2743 { 2744 struct drr_write *drrw = &rrd->header.drr_u.drr_write; 2745 err = receive_write(rwa, drrw, rrd->write_buf); 2746 /* if receive_write() is successful, it consumes the arc_buf */ 2747 if (err != 0) 2748 dmu_return_arcbuf(rrd->write_buf); 2749 rrd->write_buf = NULL; 2750 rrd->payload = NULL; 2751 return (err); 2752 } 2753 case DRR_WRITE_BYREF: 2754 { 2755 struct drr_write_byref *drrwbr = 2756 &rrd->header.drr_u.drr_write_byref; 2757 return (receive_write_byref(rwa, drrwbr)); 2758 } 2759 case DRR_WRITE_EMBEDDED: 2760 { 2761 struct drr_write_embedded *drrwe = 2762 &rrd->header.drr_u.drr_write_embedded; 2763 err = receive_write_embedded(rwa, drrwe, rrd->payload); 2764 kmem_free(rrd->payload, rrd->payload_size); 2765 rrd->payload = NULL; 2766 return (err); 2767 } 2768 case DRR_FREE: 2769 { 2770 struct drr_free *drrf = &rrd->header.drr_u.drr_free; 2771 return (receive_free(rwa, drrf)); 2772 } 2773 case DRR_SPILL: 2774 { 2775 struct drr_spill *drrs = &rrd->header.drr_u.drr_spill; 2776 err = receive_spill(rwa, drrs, rrd->payload); 2777 kmem_free(rrd->payload, rrd->payload_size); 2778 rrd->payload = NULL; 2779 return (err); 2780 } 2781 default: 2782 return (SET_ERROR(EINVAL)); 2783 } 2784 } 2785 2786 /* 2787 * dmu_recv_stream's worker thread; pull records off the queue, and then call 2788 * receive_process_record When we're done, signal the main thread and exit. 2789 */ 2790 static void 2791 receive_writer_thread(void *arg) 2792 { 2793 struct receive_writer_arg *rwa = arg; 2794 struct receive_record_arg *rrd; 2795 for (rrd = bqueue_dequeue(&rwa->q); !rrd->eos_marker; 2796 rrd = bqueue_dequeue(&rwa->q)) { 2797 /* 2798 * If there's an error, the main thread will stop putting things 2799 * on the queue, but we need to clear everything in it before we 2800 * can exit. 2801 */ 2802 if (rwa->err == 0) { 2803 rwa->err = receive_process_record(rwa, rrd); 2804 } else if (rrd->write_buf != NULL) { 2805 dmu_return_arcbuf(rrd->write_buf); 2806 rrd->write_buf = NULL; 2807 rrd->payload = NULL; 2808 } else if (rrd->payload != NULL) { 2809 kmem_free(rrd->payload, rrd->payload_size); 2810 rrd->payload = NULL; 2811 } 2812 kmem_free(rrd, sizeof (*rrd)); 2813 } 2814 kmem_free(rrd, sizeof (*rrd)); 2815 mutex_enter(&rwa->mutex); 2816 rwa->done = B_TRUE; 2817 cv_signal(&rwa->cv); 2818 mutex_exit(&rwa->mutex); 2819 thread_exit(); 2820 } 2821 2822 static int 2823 resume_check(struct receive_arg *ra, nvlist_t *begin_nvl) 2824 { 2825 uint64_t val; 2826 objset_t *mos = dmu_objset_pool(ra->os)->dp_meta_objset; 2827 uint64_t dsobj = dmu_objset_id(ra->os); 2828 uint64_t resume_obj, resume_off; 2829 2830 if (nvlist_lookup_uint64(begin_nvl, 2831 "resume_object", &resume_obj) != 0 || 2832 nvlist_lookup_uint64(begin_nvl, 2833 "resume_offset", &resume_off) != 0) { 2834 return (SET_ERROR(EINVAL)); 2835 } 2836 VERIFY0(zap_lookup(mos, dsobj, 2837 DS_FIELD_RESUME_OBJECT, sizeof (val), 1, &val)); 2838 if (resume_obj != val) 2839 return (SET_ERROR(EINVAL)); 2840 VERIFY0(zap_lookup(mos, dsobj, 2841 DS_FIELD_RESUME_OFFSET, sizeof (val), 1, &val)); 2842 if (resume_off != val) 2843 return (SET_ERROR(EINVAL)); 2844 2845 return (0); 2846 } 2847 2848 /* 2849 * Read in the stream's records, one by one, and apply them to the pool. There 2850 * are two threads involved; the thread that calls this function will spin up a 2851 * worker thread, read the records off the stream one by one, and issue 2852 * prefetches for any necessary indirect blocks. It will then push the records 2853 * onto an internal blocking queue. The worker thread will pull the records off 2854 * the queue, and actually write the data into the DMU. This way, the worker 2855 * thread doesn't have to wait for reads to complete, since everything it needs 2856 * (the indirect blocks) will be prefetched. 2857 * 2858 * NB: callers *must* call dmu_recv_end() if this succeeds. 2859 */ 2860 int 2861 dmu_recv_stream(dmu_recv_cookie_t *drc, vnode_t *vp, offset_t *voffp, 2862 int cleanup_fd, uint64_t *action_handlep) 2863 { 2864 int err = 0; 2865 struct receive_arg ra = { 0 }; 2866 struct receive_writer_arg rwa = { 0 }; 2867 int featureflags; 2868 nvlist_t *begin_nvl = NULL; 2869 2870 ra.byteswap = drc->drc_byteswap; 2871 ra.cksum = drc->drc_cksum; 2872 ra.vp = vp; 2873 ra.voff = *voffp; 2874 2875 if (dsl_dataset_is_zapified(drc->drc_ds)) { 2876 (void) zap_lookup(drc->drc_ds->ds_dir->dd_pool->dp_meta_objset, 2877 drc->drc_ds->ds_object, DS_FIELD_RESUME_BYTES, 2878 sizeof (ra.bytes_read), 1, &ra.bytes_read); 2879 } 2880 2881 objlist_create(&ra.ignore_objlist); 2882 2883 /* these were verified in dmu_recv_begin */ 2884 ASSERT3U(DMU_GET_STREAM_HDRTYPE(drc->drc_drrb->drr_versioninfo), ==, 2885 DMU_SUBSTREAM); 2886 ASSERT3U(drc->drc_drrb->drr_type, <, DMU_OST_NUMTYPES); 2887 2888 /* 2889 * Open the objset we are modifying. 2890 */ 2891 VERIFY0(dmu_objset_from_ds(drc->drc_ds, &ra.os)); 2892 2893 ASSERT(dsl_dataset_phys(drc->drc_ds)->ds_flags & DS_FLAG_INCONSISTENT); 2894 2895 featureflags = DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo); 2896 2897 /* if this stream is dedup'ed, set up the avl tree for guid mapping */ 2898 if (featureflags & DMU_BACKUP_FEATURE_DEDUP) { 2899 minor_t minor; 2900 2901 if (cleanup_fd == -1) { 2902 ra.err = SET_ERROR(EBADF); 2903 goto out; 2904 } 2905 ra.err = zfs_onexit_fd_hold(cleanup_fd, &minor); 2906 if (ra.err != 0) { 2907 cleanup_fd = -1; 2908 goto out; 2909 } 2910 2911 if (*action_handlep == 0) { 2912 rwa.guid_to_ds_map = 2913 kmem_alloc(sizeof (avl_tree_t), KM_SLEEP); 2914 avl_create(rwa.guid_to_ds_map, guid_compare, 2915 sizeof (guid_map_entry_t), 2916 offsetof(guid_map_entry_t, avlnode)); 2917 err = zfs_onexit_add_cb(minor, 2918 free_guid_map_onexit, rwa.guid_to_ds_map, 2919 action_handlep); 2920 if (ra.err != 0) 2921 goto out; 2922 } else { 2923 err = zfs_onexit_cb_data(minor, *action_handlep, 2924 (void **)&rwa.guid_to_ds_map); 2925 if (ra.err != 0) 2926 goto out; 2927 } 2928 2929 drc->drc_guid_to_ds_map = rwa.guid_to_ds_map; 2930 } 2931 2932 uint32_t payloadlen = drc->drc_drr_begin->drr_payloadlen; 2933 void *payload = NULL; 2934 if (payloadlen != 0) 2935 payload = kmem_alloc(payloadlen, KM_SLEEP); 2936 2937 err = receive_read_payload_and_next_header(&ra, payloadlen, payload); 2938 if (err != 0) { 2939 if (payloadlen != 0) 2940 kmem_free(payload, payloadlen); 2941 goto out; 2942 } 2943 if (payloadlen != 0) { 2944 err = nvlist_unpack(payload, payloadlen, &begin_nvl, KM_SLEEP); 2945 kmem_free(payload, payloadlen); 2946 if (err != 0) 2947 goto out; 2948 } 2949 2950 if (featureflags & DMU_BACKUP_FEATURE_RESUMING) { 2951 err = resume_check(&ra, begin_nvl); 2952 if (err != 0) 2953 goto out; 2954 } 2955 2956 (void) bqueue_init(&rwa.q, zfs_recv_queue_length, 2957 offsetof(struct receive_record_arg, node)); 2958 cv_init(&rwa.cv, NULL, CV_DEFAULT, NULL); 2959 mutex_init(&rwa.mutex, NULL, MUTEX_DEFAULT, NULL); 2960 rwa.os = ra.os; 2961 rwa.byteswap = drc->drc_byteswap; 2962 rwa.resumable = drc->drc_resumable; 2963 2964 (void) thread_create(NULL, 0, receive_writer_thread, &rwa, 0, curproc, 2965 TS_RUN, minclsyspri); 2966 /* 2967 * We're reading rwa.err without locks, which is safe since we are the 2968 * only reader, and the worker thread is the only writer. It's ok if we 2969 * miss a write for an iteration or two of the loop, since the writer 2970 * thread will keep freeing records we send it until we send it an eos 2971 * marker. 2972 * 2973 * We can leave this loop in 3 ways: First, if rwa.err is 2974 * non-zero. In that case, the writer thread will free the rrd we just 2975 * pushed. Second, if we're interrupted; in that case, either it's the 2976 * first loop and ra.rrd was never allocated, or it's later, and ra.rrd 2977 * has been handed off to the writer thread who will free it. Finally, 2978 * if receive_read_record fails or we're at the end of the stream, then 2979 * we free ra.rrd and exit. 2980 */ 2981 while (rwa.err == 0) { 2982 if (issig(JUSTLOOKING) && issig(FORREAL)) { 2983 err = SET_ERROR(EINTR); 2984 break; 2985 } 2986 2987 ASSERT3P(ra.rrd, ==, NULL); 2988 ra.rrd = ra.next_rrd; 2989 ra.next_rrd = NULL; 2990 /* Allocates and loads header into ra.next_rrd */ 2991 err = receive_read_record(&ra); 2992 2993 if (ra.rrd->header.drr_type == DRR_END || err != 0) { 2994 kmem_free(ra.rrd, sizeof (*ra.rrd)); 2995 ra.rrd = NULL; 2996 break; 2997 } 2998 2999 bqueue_enqueue(&rwa.q, ra.rrd, 3000 sizeof (struct receive_record_arg) + ra.rrd->payload_size); 3001 ra.rrd = NULL; 3002 } 3003 if (ra.next_rrd == NULL) 3004 ra.next_rrd = kmem_zalloc(sizeof (*ra.next_rrd), KM_SLEEP); 3005 ra.next_rrd->eos_marker = B_TRUE; 3006 bqueue_enqueue(&rwa.q, ra.next_rrd, 1); 3007 3008 mutex_enter(&rwa.mutex); 3009 while (!rwa.done) { 3010 cv_wait(&rwa.cv, &rwa.mutex); 3011 } 3012 mutex_exit(&rwa.mutex); 3013 3014 cv_destroy(&rwa.cv); 3015 mutex_destroy(&rwa.mutex); 3016 bqueue_destroy(&rwa.q); 3017 if (err == 0) 3018 err = rwa.err; 3019 3020 out: 3021 nvlist_free(begin_nvl); 3022 if ((featureflags & DMU_BACKUP_FEATURE_DEDUP) && (cleanup_fd != -1)) 3023 zfs_onexit_fd_rele(cleanup_fd); 3024 3025 if (err != 0) { 3026 /* 3027 * Clean up references. If receive is not resumable, 3028 * destroy what we created, so we don't leave it in 3029 * the inconsistent state. 3030 */ 3031 dmu_recv_cleanup_ds(drc); 3032 } 3033 3034 *voffp = ra.voff; 3035 objlist_destroy(&ra.ignore_objlist); 3036 return (err); 3037 } 3038 3039 static int 3040 dmu_recv_end_check(void *arg, dmu_tx_t *tx) 3041 { 3042 dmu_recv_cookie_t *drc = arg; 3043 dsl_pool_t *dp = dmu_tx_pool(tx); 3044 int error; 3045 3046 ASSERT3P(drc->drc_ds->ds_owner, ==, dmu_recv_tag); 3047 3048 if (!drc->drc_newfs) { 3049 dsl_dataset_t *origin_head; 3050 3051 error = dsl_dataset_hold(dp, drc->drc_tofs, FTAG, &origin_head); 3052 if (error != 0) 3053 return (error); 3054 if (drc->drc_force) { 3055 /* 3056 * We will destroy any snapshots in tofs (i.e. before 3057 * origin_head) that are after the origin (which is 3058 * the snap before drc_ds, because drc_ds can not 3059 * have any snaps of its own). 3060 */ 3061 uint64_t obj; 3062 3063 obj = dsl_dataset_phys(origin_head)->ds_prev_snap_obj; 3064 while (obj != 3065 dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj) { 3066 dsl_dataset_t *snap; 3067 error = dsl_dataset_hold_obj(dp, obj, FTAG, 3068 &snap); 3069 if (error != 0) 3070 break; 3071 if (snap->ds_dir != origin_head->ds_dir) 3072 error = SET_ERROR(EINVAL); 3073 if (error == 0) { 3074 error = dsl_destroy_snapshot_check_impl( 3075 snap, B_FALSE); 3076 } 3077 obj = dsl_dataset_phys(snap)->ds_prev_snap_obj; 3078 dsl_dataset_rele(snap, FTAG); 3079 if (error != 0) 3080 break; 3081 } 3082 if (error != 0) { 3083 dsl_dataset_rele(origin_head, FTAG); 3084 return (error); 3085 } 3086 } 3087 error = dsl_dataset_clone_swap_check_impl(drc->drc_ds, 3088 origin_head, drc->drc_force, drc->drc_owner, tx); 3089 if (error != 0) { 3090 dsl_dataset_rele(origin_head, FTAG); 3091 return (error); 3092 } 3093 error = dsl_dataset_snapshot_check_impl(origin_head, 3094 drc->drc_tosnap, tx, B_TRUE, 1, drc->drc_cred); 3095 dsl_dataset_rele(origin_head, FTAG); 3096 if (error != 0) 3097 return (error); 3098 3099 error = dsl_destroy_head_check_impl(drc->drc_ds, 1); 3100 } else { 3101 error = dsl_dataset_snapshot_check_impl(drc->drc_ds, 3102 drc->drc_tosnap, tx, B_TRUE, 1, drc->drc_cred); 3103 } 3104 return (error); 3105 } 3106 3107 static void 3108 dmu_recv_end_sync(void *arg, dmu_tx_t *tx) 3109 { 3110 dmu_recv_cookie_t *drc = arg; 3111 dsl_pool_t *dp = dmu_tx_pool(tx); 3112 3113 spa_history_log_internal_ds(drc->drc_ds, "finish receiving", 3114 tx, "snap=%s", drc->drc_tosnap); 3115 3116 if (!drc->drc_newfs) { 3117 dsl_dataset_t *origin_head; 3118 3119 VERIFY0(dsl_dataset_hold(dp, drc->drc_tofs, FTAG, 3120 &origin_head)); 3121 3122 if (drc->drc_force) { 3123 /* 3124 * Destroy any snapshots of drc_tofs (origin_head) 3125 * after the origin (the snap before drc_ds). 3126 */ 3127 uint64_t obj; 3128 3129 obj = dsl_dataset_phys(origin_head)->ds_prev_snap_obj; 3130 while (obj != 3131 dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj) { 3132 dsl_dataset_t *snap; 3133 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG, 3134 &snap)); 3135 ASSERT3P(snap->ds_dir, ==, origin_head->ds_dir); 3136 obj = dsl_dataset_phys(snap)->ds_prev_snap_obj; 3137 dsl_destroy_snapshot_sync_impl(snap, 3138 B_FALSE, tx); 3139 dsl_dataset_rele(snap, FTAG); 3140 } 3141 } 3142 VERIFY3P(drc->drc_ds->ds_prev, ==, 3143 origin_head->ds_prev); 3144 3145 dsl_dataset_clone_swap_sync_impl(drc->drc_ds, 3146 origin_head, tx); 3147 dsl_dataset_snapshot_sync_impl(origin_head, 3148 drc->drc_tosnap, tx); 3149 3150 /* set snapshot's creation time and guid */ 3151 dmu_buf_will_dirty(origin_head->ds_prev->ds_dbuf, tx); 3152 dsl_dataset_phys(origin_head->ds_prev)->ds_creation_time = 3153 drc->drc_drrb->drr_creation_time; 3154 dsl_dataset_phys(origin_head->ds_prev)->ds_guid = 3155 drc->drc_drrb->drr_toguid; 3156 dsl_dataset_phys(origin_head->ds_prev)->ds_flags &= 3157 ~DS_FLAG_INCONSISTENT; 3158 3159 dmu_buf_will_dirty(origin_head->ds_dbuf, tx); 3160 dsl_dataset_phys(origin_head)->ds_flags &= 3161 ~DS_FLAG_INCONSISTENT; 3162 3163 drc->drc_newsnapobj = 3164 dsl_dataset_phys(origin_head)->ds_prev_snap_obj; 3165 3166 dsl_dataset_rele(origin_head, FTAG); 3167 dsl_destroy_head_sync_impl(drc->drc_ds, tx); 3168 3169 if (drc->drc_owner != NULL) 3170 VERIFY3P(origin_head->ds_owner, ==, drc->drc_owner); 3171 } else { 3172 dsl_dataset_t *ds = drc->drc_ds; 3173 3174 dsl_dataset_snapshot_sync_impl(ds, drc->drc_tosnap, tx); 3175 3176 /* set snapshot's creation time and guid */ 3177 dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx); 3178 dsl_dataset_phys(ds->ds_prev)->ds_creation_time = 3179 drc->drc_drrb->drr_creation_time; 3180 dsl_dataset_phys(ds->ds_prev)->ds_guid = 3181 drc->drc_drrb->drr_toguid; 3182 dsl_dataset_phys(ds->ds_prev)->ds_flags &= 3183 ~DS_FLAG_INCONSISTENT; 3184 3185 dmu_buf_will_dirty(ds->ds_dbuf, tx); 3186 dsl_dataset_phys(ds)->ds_flags &= ~DS_FLAG_INCONSISTENT; 3187 if (dsl_dataset_has_resume_receive_state(ds)) { 3188 (void) zap_remove(dp->dp_meta_objset, ds->ds_object, 3189 DS_FIELD_RESUME_FROMGUID, tx); 3190 (void) zap_remove(dp->dp_meta_objset, ds->ds_object, 3191 DS_FIELD_RESUME_OBJECT, tx); 3192 (void) zap_remove(dp->dp_meta_objset, ds->ds_object, 3193 DS_FIELD_RESUME_OFFSET, tx); 3194 (void) zap_remove(dp->dp_meta_objset, ds->ds_object, 3195 DS_FIELD_RESUME_BYTES, tx); 3196 (void) zap_remove(dp->dp_meta_objset, ds->ds_object, 3197 DS_FIELD_RESUME_TOGUID, tx); 3198 (void) zap_remove(dp->dp_meta_objset, ds->ds_object, 3199 DS_FIELD_RESUME_TONAME, tx); 3200 } 3201 drc->drc_newsnapobj = 3202 dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj; 3203 } 3204 /* 3205 * Release the hold from dmu_recv_begin. This must be done before 3206 * we return to open context, so that when we free the dataset's dnode, 3207 * we can evict its bonus buffer. 3208 */ 3209 dsl_dataset_disown(drc->drc_ds, dmu_recv_tag); 3210 drc->drc_ds = NULL; 3211 } 3212 3213 static int 3214 add_ds_to_guidmap(const char *name, avl_tree_t *guid_map, uint64_t snapobj) 3215 { 3216 dsl_pool_t *dp; 3217 dsl_dataset_t *snapds; 3218 guid_map_entry_t *gmep; 3219 int err; 3220 3221 ASSERT(guid_map != NULL); 3222 3223 err = dsl_pool_hold(name, FTAG, &dp); 3224 if (err != 0) 3225 return (err); 3226 gmep = kmem_alloc(sizeof (*gmep), KM_SLEEP); 3227 err = dsl_dataset_hold_obj(dp, snapobj, gmep, &snapds); 3228 if (err == 0) { 3229 gmep->guid = dsl_dataset_phys(snapds)->ds_guid; 3230 gmep->gme_ds = snapds; 3231 avl_add(guid_map, gmep); 3232 dsl_dataset_long_hold(snapds, gmep); 3233 } else { 3234 kmem_free(gmep, sizeof (*gmep)); 3235 } 3236 3237 dsl_pool_rele(dp, FTAG); 3238 return (err); 3239 } 3240 3241 static int dmu_recv_end_modified_blocks = 3; 3242 3243 static int 3244 dmu_recv_existing_end(dmu_recv_cookie_t *drc) 3245 { 3246 #ifdef _KERNEL 3247 /* 3248 * We will be destroying the ds; make sure its origin is unmounted if 3249 * necessary. 3250 */ 3251 char name[ZFS_MAX_DATASET_NAME_LEN]; 3252 dsl_dataset_name(drc->drc_ds, name); 3253 zfs_destroy_unmount_origin(name); 3254 #endif 3255 3256 return (dsl_sync_task(drc->drc_tofs, 3257 dmu_recv_end_check, dmu_recv_end_sync, drc, 3258 dmu_recv_end_modified_blocks, ZFS_SPACE_CHECK_NORMAL)); 3259 } 3260 3261 static int 3262 dmu_recv_new_end(dmu_recv_cookie_t *drc) 3263 { 3264 return (dsl_sync_task(drc->drc_tofs, 3265 dmu_recv_end_check, dmu_recv_end_sync, drc, 3266 dmu_recv_end_modified_blocks, ZFS_SPACE_CHECK_NORMAL)); 3267 } 3268 3269 int 3270 dmu_recv_end(dmu_recv_cookie_t *drc, void *owner) 3271 { 3272 int error; 3273 3274 drc->drc_owner = owner; 3275 3276 if (drc->drc_newfs) 3277 error = dmu_recv_new_end(drc); 3278 else 3279 error = dmu_recv_existing_end(drc); 3280 3281 if (error != 0) { 3282 dmu_recv_cleanup_ds(drc); 3283 } else if (drc->drc_guid_to_ds_map != NULL) { 3284 (void) add_ds_to_guidmap(drc->drc_tofs, 3285 drc->drc_guid_to_ds_map, 3286 drc->drc_newsnapobj); 3287 } 3288 return (error); 3289 } 3290 3291 /* 3292 * Return TRUE if this objset is currently being received into. 3293 */ 3294 boolean_t 3295 dmu_objset_is_receiving(objset_t *os) 3296 { 3297 return (os->os_dsl_dataset != NULL && 3298 os->os_dsl_dataset->ds_owner == dmu_recv_tag); 3299 } 3300