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