dmu_send.c revision 307285
1/* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21/* 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23 * Copyright 2011 Nexenta Systems, Inc. All rights reserved. 24 * Copyright (c) 2011, 2015 by Delphix. All rights reserved. 25 * Copyright (c) 2014, Joyent, Inc. All rights reserved. 26 * Copyright (c) 2012, Martin Matuska <mm@FreeBSD.org>. All rights reserved. 27 * Copyright 2014 HybridCluster. All rights reserved. 28 * Copyright 2016 RackTop Systems. 29 * Copyright (c) 2014 Integros [integros.com] 30 */ 31 32#include <sys/dmu.h> 33#include <sys/dmu_impl.h> 34#include <sys/dmu_tx.h> 35#include <sys/dbuf.h> 36#include <sys/dnode.h> 37#include <sys/zfs_context.h> 38#include <sys/dmu_objset.h> 39#include <sys/dmu_traverse.h> 40#include <sys/dsl_dataset.h> 41#include <sys/dsl_dir.h> 42#include <sys/dsl_prop.h> 43#include <sys/dsl_pool.h> 44#include <sys/dsl_synctask.h> 45#include <sys/zfs_ioctl.h> 46#include <sys/zap.h> 47#include <sys/zio_checksum.h> 48#include <sys/zfs_znode.h> 49#include <zfs_fletcher.h> 50#include <sys/avl.h> 51#include <sys/ddt.h> 52#include <sys/zfs_onexit.h> 53#include <sys/dmu_send.h> 54#include <sys/dsl_destroy.h> 55#include <sys/blkptr.h> 56#include <sys/dsl_bookmark.h> 57#include <sys/zfeature.h> 58#include <sys/bqueue.h> 59 60#ifdef __FreeBSD__ 61#undef dump_write 62#define dump_write dmu_dump_write 63#endif 64 65/* Set this tunable to TRUE to replace corrupt data with 0x2f5baddb10c */ 66int zfs_send_corrupt_data = B_FALSE; 67int zfs_send_queue_length = 16 * 1024 * 1024; 68int zfs_recv_queue_length = 16 * 1024 * 1024; 69/* Set this tunable to FALSE to disable setting of DRR_FLAG_FREERECORDS */ 70int zfs_send_set_freerecords_bit = B_TRUE; 71 72#ifdef _KERNEL 73TUNABLE_INT("vfs.zfs.send_set_freerecords_bit", &zfs_send_set_freerecords_bit); 74#endif 75 76static char *dmu_recv_tag = "dmu_recv_tag"; 77const char *recv_clone_name = "%recv"; 78 79#define BP_SPAN(datablkszsec, indblkshift, level) \ 80 (((uint64_t)datablkszsec) << (SPA_MINBLOCKSHIFT + \ 81 (level) * (indblkshift - SPA_BLKPTRSHIFT))) 82 83static void byteswap_record(dmu_replay_record_t *drr); 84 85struct send_thread_arg { 86 bqueue_t q; 87 dsl_dataset_t *ds; /* Dataset to traverse */ 88 uint64_t fromtxg; /* Traverse from this txg */ 89 int flags; /* flags to pass to traverse_dataset */ 90 int error_code; 91 boolean_t cancel; 92 zbookmark_phys_t resume; 93}; 94 95struct send_block_record { 96 boolean_t eos_marker; /* Marks the end of the stream */ 97 blkptr_t bp; 98 zbookmark_phys_t zb; 99 uint8_t indblkshift; 100 uint16_t datablkszsec; 101 bqueue_node_t ln; 102}; 103 104static int 105dump_bytes(dmu_sendarg_t *dsp, void *buf, int len) 106{ 107 dsl_dataset_t *ds = dmu_objset_ds(dsp->dsa_os); 108 struct uio auio; 109 struct iovec aiov; 110 111 /* 112 * The code does not rely on this (len being a multiple of 8). We keep 113 * this assertion because of the corresponding assertion in 114 * receive_read(). Keeping this assertion ensures that we do not 115 * inadvertently break backwards compatibility (causing the assertion 116 * in receive_read() to trigger on old software). 117 * 118 * Removing the assertions could be rolled into a new feature that uses 119 * data that isn't 8-byte aligned; if the assertions were removed, a 120 * feature flag would have to be added. 121 */ 122 123 ASSERT0(len % 8); 124 125 aiov.iov_base = buf; 126 aiov.iov_len = len; 127 auio.uio_iov = &aiov; 128 auio.uio_iovcnt = 1; 129 auio.uio_resid = len; 130 auio.uio_segflg = UIO_SYSSPACE; 131 auio.uio_rw = UIO_WRITE; 132 auio.uio_offset = (off_t)-1; 133 auio.uio_td = dsp->dsa_td; 134#ifdef _KERNEL 135 if (dsp->dsa_fp->f_type == DTYPE_VNODE) 136 bwillwrite(); 137 dsp->dsa_err = fo_write(dsp->dsa_fp, &auio, dsp->dsa_td->td_ucred, 0, 138 dsp->dsa_td); 139#else 140 fprintf(stderr, "%s: returning EOPNOTSUPP\n", __func__); 141 dsp->dsa_err = EOPNOTSUPP; 142#endif 143 mutex_enter(&ds->ds_sendstream_lock); 144 *dsp->dsa_off += len; 145 mutex_exit(&ds->ds_sendstream_lock); 146 147 return (dsp->dsa_err); 148} 149 150/* 151 * For all record types except BEGIN, fill in the checksum (overlaid in 152 * drr_u.drr_checksum.drr_checksum). The checksum verifies everything 153 * up to the start of the checksum itself. 154 */ 155static int 156dump_record(dmu_sendarg_t *dsp, void *payload, int payload_len) 157{ 158 ASSERT3U(offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum), 159 ==, sizeof (dmu_replay_record_t) - sizeof (zio_cksum_t)); 160 fletcher_4_incremental_native(dsp->dsa_drr, 161 offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum), 162 &dsp->dsa_zc); 163 if (dsp->dsa_drr->drr_type == DRR_BEGIN) { 164 dsp->dsa_sent_begin = B_TRUE; 165 } else { 166 ASSERT(ZIO_CHECKSUM_IS_ZERO(&dsp->dsa_drr->drr_u. 167 drr_checksum.drr_checksum)); 168 dsp->dsa_drr->drr_u.drr_checksum.drr_checksum = dsp->dsa_zc; 169 } 170 if (dsp->dsa_drr->drr_type == DRR_END) { 171 dsp->dsa_sent_end = B_TRUE; 172 } 173 fletcher_4_incremental_native(&dsp->dsa_drr-> 174 drr_u.drr_checksum.drr_checksum, 175 sizeof (zio_cksum_t), &dsp->dsa_zc); 176 if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)) != 0) 177 return (SET_ERROR(EINTR)); 178 if (payload_len != 0) { 179 fletcher_4_incremental_native(payload, payload_len, 180 &dsp->dsa_zc); 181 if (dump_bytes(dsp, payload, payload_len) != 0) 182 return (SET_ERROR(EINTR)); 183 } 184 return (0); 185} 186 187/* 188 * Fill in the drr_free struct, or perform aggregation if the previous record is 189 * also a free record, and the two are adjacent. 190 * 191 * Note that we send free records even for a full send, because we want to be 192 * able to receive a full send as a clone, which requires a list of all the free 193 * and freeobject records that were generated on the source. 194 */ 195static int 196dump_free(dmu_sendarg_t *dsp, uint64_t object, uint64_t offset, 197 uint64_t length) 198{ 199 struct drr_free *drrf = &(dsp->dsa_drr->drr_u.drr_free); 200 201 /* 202 * When we receive a free record, dbuf_free_range() assumes 203 * that the receiving system doesn't have any dbufs in the range 204 * being freed. This is always true because there is a one-record 205 * constraint: we only send one WRITE record for any given 206 * object,offset. We know that the one-record constraint is 207 * true because we always send data in increasing order by 208 * object,offset. 209 * 210 * If the increasing-order constraint ever changes, we should find 211 * another way to assert that the one-record constraint is still 212 * satisfied. 213 */ 214 ASSERT(object > dsp->dsa_last_data_object || 215 (object == dsp->dsa_last_data_object && 216 offset > dsp->dsa_last_data_offset)); 217 218 if (length != -1ULL && offset + length < offset) 219 length = -1ULL; 220 221 /* 222 * If there is a pending op, but it's not PENDING_FREE, push it out, 223 * since free block aggregation can only be done for blocks of the 224 * same type (i.e., DRR_FREE records can only be aggregated with 225 * other DRR_FREE records. DRR_FREEOBJECTS records can only be 226 * aggregated with other DRR_FREEOBJECTS records. 227 */ 228 if (dsp->dsa_pending_op != PENDING_NONE && 229 dsp->dsa_pending_op != PENDING_FREE) { 230 if (dump_record(dsp, NULL, 0) != 0) 231 return (SET_ERROR(EINTR)); 232 dsp->dsa_pending_op = PENDING_NONE; 233 } 234 235 if (dsp->dsa_pending_op == PENDING_FREE) { 236 /* 237 * There should never be a PENDING_FREE if length is -1 238 * (because dump_dnode is the only place where this 239 * function is called with a -1, and only after flushing 240 * any pending record). 241 */ 242 ASSERT(length != -1ULL); 243 /* 244 * Check to see whether this free block can be aggregated 245 * with pending one. 246 */ 247 if (drrf->drr_object == object && drrf->drr_offset + 248 drrf->drr_length == offset) { 249 drrf->drr_length += length; 250 return (0); 251 } else { 252 /* not a continuation. Push out pending record */ 253 if (dump_record(dsp, NULL, 0) != 0) 254 return (SET_ERROR(EINTR)); 255 dsp->dsa_pending_op = PENDING_NONE; 256 } 257 } 258 /* create a FREE record and make it pending */ 259 bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t)); 260 dsp->dsa_drr->drr_type = DRR_FREE; 261 drrf->drr_object = object; 262 drrf->drr_offset = offset; 263 drrf->drr_length = length; 264 drrf->drr_toguid = dsp->dsa_toguid; 265 if (length == -1ULL) { 266 if (dump_record(dsp, NULL, 0) != 0) 267 return (SET_ERROR(EINTR)); 268 } else { 269 dsp->dsa_pending_op = PENDING_FREE; 270 } 271 272 return (0); 273} 274 275static int 276dump_write(dmu_sendarg_t *dsp, dmu_object_type_t type, 277 uint64_t object, uint64_t offset, int blksz, const blkptr_t *bp, void *data) 278{ 279 struct drr_write *drrw = &(dsp->dsa_drr->drr_u.drr_write); 280 281 /* 282 * We send data in increasing object, offset order. 283 * See comment in dump_free() for details. 284 */ 285 ASSERT(object > dsp->dsa_last_data_object || 286 (object == dsp->dsa_last_data_object && 287 offset > dsp->dsa_last_data_offset)); 288 dsp->dsa_last_data_object = object; 289 dsp->dsa_last_data_offset = offset + blksz - 1; 290 291 /* 292 * If there is any kind of pending aggregation (currently either 293 * a grouping of free objects or free blocks), push it out to 294 * the stream, since aggregation can't be done across operations 295 * of different types. 296 */ 297 if (dsp->dsa_pending_op != PENDING_NONE) { 298 if (dump_record(dsp, NULL, 0) != 0) 299 return (SET_ERROR(EINTR)); 300 dsp->dsa_pending_op = PENDING_NONE; 301 } 302 /* write a WRITE record */ 303 bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t)); 304 dsp->dsa_drr->drr_type = DRR_WRITE; 305 drrw->drr_object = object; 306 drrw->drr_type = type; 307 drrw->drr_offset = offset; 308 drrw->drr_length = blksz; 309 drrw->drr_toguid = dsp->dsa_toguid; 310 if (bp == NULL || BP_IS_EMBEDDED(bp)) { 311 /* 312 * There's no pre-computed checksum for partial-block 313 * writes or embedded BP's, so (like 314 * fletcher4-checkummed blocks) userland will have to 315 * compute a dedup-capable checksum itself. 316 */ 317 drrw->drr_checksumtype = ZIO_CHECKSUM_OFF; 318 } else { 319 drrw->drr_checksumtype = BP_GET_CHECKSUM(bp); 320 if (zio_checksum_table[drrw->drr_checksumtype].ci_flags & 321 ZCHECKSUM_FLAG_DEDUP) 322 drrw->drr_checksumflags |= DRR_CHECKSUM_DEDUP; 323 DDK_SET_LSIZE(&drrw->drr_key, BP_GET_LSIZE(bp)); 324 DDK_SET_PSIZE(&drrw->drr_key, BP_GET_PSIZE(bp)); 325 DDK_SET_COMPRESS(&drrw->drr_key, BP_GET_COMPRESS(bp)); 326 drrw->drr_key.ddk_cksum = bp->blk_cksum; 327 } 328 329 if (dump_record(dsp, data, blksz) != 0) 330 return (SET_ERROR(EINTR)); 331 return (0); 332} 333 334static int 335dump_write_embedded(dmu_sendarg_t *dsp, uint64_t object, uint64_t offset, 336 int blksz, const blkptr_t *bp) 337{ 338 char buf[BPE_PAYLOAD_SIZE]; 339 struct drr_write_embedded *drrw = 340 &(dsp->dsa_drr->drr_u.drr_write_embedded); 341 342 if (dsp->dsa_pending_op != PENDING_NONE) { 343 if (dump_record(dsp, NULL, 0) != 0) 344 return (EINTR); 345 dsp->dsa_pending_op = PENDING_NONE; 346 } 347 348 ASSERT(BP_IS_EMBEDDED(bp)); 349 350 bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t)); 351 dsp->dsa_drr->drr_type = DRR_WRITE_EMBEDDED; 352 drrw->drr_object = object; 353 drrw->drr_offset = offset; 354 drrw->drr_length = blksz; 355 drrw->drr_toguid = dsp->dsa_toguid; 356 drrw->drr_compression = BP_GET_COMPRESS(bp); 357 drrw->drr_etype = BPE_GET_ETYPE(bp); 358 drrw->drr_lsize = BPE_GET_LSIZE(bp); 359 drrw->drr_psize = BPE_GET_PSIZE(bp); 360 361 decode_embedded_bp_compressed(bp, buf); 362 363 if (dump_record(dsp, buf, P2ROUNDUP(drrw->drr_psize, 8)) != 0) 364 return (EINTR); 365 return (0); 366} 367 368static int 369dump_spill(dmu_sendarg_t *dsp, uint64_t object, int blksz, void *data) 370{ 371 struct drr_spill *drrs = &(dsp->dsa_drr->drr_u.drr_spill); 372 373 if (dsp->dsa_pending_op != PENDING_NONE) { 374 if (dump_record(dsp, NULL, 0) != 0) 375 return (SET_ERROR(EINTR)); 376 dsp->dsa_pending_op = PENDING_NONE; 377 } 378 379 /* write a SPILL record */ 380 bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t)); 381 dsp->dsa_drr->drr_type = DRR_SPILL; 382 drrs->drr_object = object; 383 drrs->drr_length = blksz; 384 drrs->drr_toguid = dsp->dsa_toguid; 385 386 if (dump_record(dsp, data, blksz) != 0) 387 return (SET_ERROR(EINTR)); 388 return (0); 389} 390 391static int 392dump_freeobjects(dmu_sendarg_t *dsp, uint64_t firstobj, uint64_t numobjs) 393{ 394 struct drr_freeobjects *drrfo = &(dsp->dsa_drr->drr_u.drr_freeobjects); 395 396 /* 397 * If there is a pending op, but it's not PENDING_FREEOBJECTS, 398 * push it out, since free block aggregation can only be done for 399 * blocks of the same type (i.e., DRR_FREE records can only be 400 * aggregated with other DRR_FREE records. DRR_FREEOBJECTS records 401 * can only be aggregated with other DRR_FREEOBJECTS records. 402 */ 403 if (dsp->dsa_pending_op != PENDING_NONE && 404 dsp->dsa_pending_op != PENDING_FREEOBJECTS) { 405 if (dump_record(dsp, NULL, 0) != 0) 406 return (SET_ERROR(EINTR)); 407 dsp->dsa_pending_op = PENDING_NONE; 408 } 409 if (dsp->dsa_pending_op == PENDING_FREEOBJECTS) { 410 /* 411 * See whether this free object array can be aggregated 412 * with pending one 413 */ 414 if (drrfo->drr_firstobj + drrfo->drr_numobjs == firstobj) { 415 drrfo->drr_numobjs += numobjs; 416 return (0); 417 } else { 418 /* can't be aggregated. Push out pending record */ 419 if (dump_record(dsp, NULL, 0) != 0) 420 return (SET_ERROR(EINTR)); 421 dsp->dsa_pending_op = PENDING_NONE; 422 } 423 } 424 425 /* write a FREEOBJECTS record */ 426 bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t)); 427 dsp->dsa_drr->drr_type = DRR_FREEOBJECTS; 428 drrfo->drr_firstobj = firstobj; 429 drrfo->drr_numobjs = numobjs; 430 drrfo->drr_toguid = dsp->dsa_toguid; 431 432 dsp->dsa_pending_op = PENDING_FREEOBJECTS; 433 434 return (0); 435} 436 437static int 438dump_dnode(dmu_sendarg_t *dsp, uint64_t object, dnode_phys_t *dnp) 439{ 440 struct drr_object *drro = &(dsp->dsa_drr->drr_u.drr_object); 441 442 if (object < dsp->dsa_resume_object) { 443 /* 444 * Note: when resuming, we will visit all the dnodes in 445 * the block of dnodes that we are resuming from. In 446 * this case it's unnecessary to send the dnodes prior to 447 * the one we are resuming from. We should be at most one 448 * block's worth of dnodes behind the resume point. 449 */ 450 ASSERT3U(dsp->dsa_resume_object - object, <, 451 1 << (DNODE_BLOCK_SHIFT - DNODE_SHIFT)); 452 return (0); 453 } 454 455 if (dnp == NULL || dnp->dn_type == DMU_OT_NONE) 456 return (dump_freeobjects(dsp, object, 1)); 457 458 if (dsp->dsa_pending_op != PENDING_NONE) { 459 if (dump_record(dsp, NULL, 0) != 0) 460 return (SET_ERROR(EINTR)); 461 dsp->dsa_pending_op = PENDING_NONE; 462 } 463 464 /* write an OBJECT record */ 465 bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t)); 466 dsp->dsa_drr->drr_type = DRR_OBJECT; 467 drro->drr_object = object; 468 drro->drr_type = dnp->dn_type; 469 drro->drr_bonustype = dnp->dn_bonustype; 470 drro->drr_blksz = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT; 471 drro->drr_bonuslen = dnp->dn_bonuslen; 472 drro->drr_checksumtype = dnp->dn_checksum; 473 drro->drr_compress = dnp->dn_compress; 474 drro->drr_toguid = dsp->dsa_toguid; 475 476 if (!(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) && 477 drro->drr_blksz > SPA_OLD_MAXBLOCKSIZE) 478 drro->drr_blksz = SPA_OLD_MAXBLOCKSIZE; 479 480 if (dump_record(dsp, DN_BONUS(dnp), 481 P2ROUNDUP(dnp->dn_bonuslen, 8)) != 0) { 482 return (SET_ERROR(EINTR)); 483 } 484 485 /* Free anything past the end of the file. */ 486 if (dump_free(dsp, object, (dnp->dn_maxblkid + 1) * 487 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT), -1ULL) != 0) 488 return (SET_ERROR(EINTR)); 489 if (dsp->dsa_err != 0) 490 return (SET_ERROR(EINTR)); 491 return (0); 492} 493 494static boolean_t 495backup_do_embed(dmu_sendarg_t *dsp, const blkptr_t *bp) 496{ 497 if (!BP_IS_EMBEDDED(bp)) 498 return (B_FALSE); 499 500 /* 501 * Compression function must be legacy, or explicitly enabled. 502 */ 503 if ((BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_LEGACY_FUNCTIONS && 504 !(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_EMBED_DATA_LZ4))) 505 return (B_FALSE); 506 507 /* 508 * Embed type must be explicitly enabled. 509 */ 510 switch (BPE_GET_ETYPE(bp)) { 511 case BP_EMBEDDED_TYPE_DATA: 512 if (dsp->dsa_featureflags & DMU_BACKUP_FEATURE_EMBED_DATA) 513 return (B_TRUE); 514 break; 515 default: 516 return (B_FALSE); 517 } 518 return (B_FALSE); 519} 520 521/* 522 * This is the callback function to traverse_dataset that acts as the worker 523 * thread for dmu_send_impl. 524 */ 525/*ARGSUSED*/ 526static int 527send_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp, 528 const zbookmark_phys_t *zb, const struct dnode_phys *dnp, void *arg) 529{ 530 struct send_thread_arg *sta = arg; 531 struct send_block_record *record; 532 uint64_t record_size; 533 int err = 0; 534 535 ASSERT(zb->zb_object == DMU_META_DNODE_OBJECT || 536 zb->zb_object >= sta->resume.zb_object); 537 538 if (sta->cancel) 539 return (SET_ERROR(EINTR)); 540 541 if (bp == NULL) { 542 ASSERT3U(zb->zb_level, ==, ZB_DNODE_LEVEL); 543 return (0); 544 } else if (zb->zb_level < 0) { 545 return (0); 546 } 547 548 record = kmem_zalloc(sizeof (struct send_block_record), KM_SLEEP); 549 record->eos_marker = B_FALSE; 550 record->bp = *bp; 551 record->zb = *zb; 552 record->indblkshift = dnp->dn_indblkshift; 553 record->datablkszsec = dnp->dn_datablkszsec; 554 record_size = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT; 555 bqueue_enqueue(&sta->q, record, record_size); 556 557 return (err); 558} 559 560/* 561 * This function kicks off the traverse_dataset. It also handles setting the 562 * error code of the thread in case something goes wrong, and pushes the End of 563 * Stream record when the traverse_dataset call has finished. If there is no 564 * dataset to traverse, the thread immediately pushes End of Stream marker. 565 */ 566static void 567send_traverse_thread(void *arg) 568{ 569 struct send_thread_arg *st_arg = arg; 570 int err; 571 struct send_block_record *data; 572 573 if (st_arg->ds != NULL) { 574 err = traverse_dataset_resume(st_arg->ds, 575 st_arg->fromtxg, &st_arg->resume, 576 st_arg->flags, send_cb, st_arg); 577 578 if (err != EINTR) 579 st_arg->error_code = err; 580 } 581 data = kmem_zalloc(sizeof (*data), KM_SLEEP); 582 data->eos_marker = B_TRUE; 583 bqueue_enqueue(&st_arg->q, data, 1); 584 thread_exit(); 585} 586 587/* 588 * This function actually handles figuring out what kind of record needs to be 589 * dumped, reading the data (which has hopefully been prefetched), and calling 590 * the appropriate helper function. 591 */ 592static int 593do_dump(dmu_sendarg_t *dsa, struct send_block_record *data) 594{ 595 dsl_dataset_t *ds = dmu_objset_ds(dsa->dsa_os); 596 const blkptr_t *bp = &data->bp; 597 const zbookmark_phys_t *zb = &data->zb; 598 uint8_t indblkshift = data->indblkshift; 599 uint16_t dblkszsec = data->datablkszsec; 600 spa_t *spa = ds->ds_dir->dd_pool->dp_spa; 601 dmu_object_type_t type = bp ? BP_GET_TYPE(bp) : DMU_OT_NONE; 602 int err = 0; 603 604 ASSERT3U(zb->zb_level, >=, 0); 605 606 ASSERT(zb->zb_object == DMU_META_DNODE_OBJECT || 607 zb->zb_object >= dsa->dsa_resume_object); 608 609 if (zb->zb_object != DMU_META_DNODE_OBJECT && 610 DMU_OBJECT_IS_SPECIAL(zb->zb_object)) { 611 return (0); 612 } else if (BP_IS_HOLE(bp) && 613 zb->zb_object == DMU_META_DNODE_OBJECT) { 614 uint64_t span = BP_SPAN(dblkszsec, indblkshift, zb->zb_level); 615 uint64_t dnobj = (zb->zb_blkid * span) >> DNODE_SHIFT; 616 err = dump_freeobjects(dsa, dnobj, span >> DNODE_SHIFT); 617 } else if (BP_IS_HOLE(bp)) { 618 uint64_t span = BP_SPAN(dblkszsec, indblkshift, zb->zb_level); 619 uint64_t offset = zb->zb_blkid * span; 620 err = dump_free(dsa, zb->zb_object, offset, span); 621 } else if (zb->zb_level > 0 || type == DMU_OT_OBJSET) { 622 return (0); 623 } else if (type == DMU_OT_DNODE) { 624 int blksz = BP_GET_LSIZE(bp); 625 arc_flags_t aflags = ARC_FLAG_WAIT; 626 arc_buf_t *abuf; 627 628 ASSERT0(zb->zb_level); 629 630 if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf, 631 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, 632 &aflags, zb) != 0) 633 return (SET_ERROR(EIO)); 634 635 dnode_phys_t *blk = abuf->b_data; 636 uint64_t dnobj = zb->zb_blkid * (blksz >> DNODE_SHIFT); 637 for (int i = 0; i < blksz >> DNODE_SHIFT; i++) { 638 err = dump_dnode(dsa, dnobj + i, blk + i); 639 if (err != 0) 640 break; 641 } 642 arc_buf_destroy(abuf, &abuf); 643 } else if (type == DMU_OT_SA) { 644 arc_flags_t aflags = ARC_FLAG_WAIT; 645 arc_buf_t *abuf; 646 int blksz = BP_GET_LSIZE(bp); 647 648 if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf, 649 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, 650 &aflags, zb) != 0) 651 return (SET_ERROR(EIO)); 652 653 err = dump_spill(dsa, zb->zb_object, blksz, abuf->b_data); 654 arc_buf_destroy(abuf, &abuf); 655 } else if (backup_do_embed(dsa, bp)) { 656 /* it's an embedded level-0 block of a regular object */ 657 int blksz = dblkszsec << SPA_MINBLOCKSHIFT; 658 ASSERT0(zb->zb_level); 659 err = dump_write_embedded(dsa, zb->zb_object, 660 zb->zb_blkid * blksz, blksz, bp); 661 } else { 662 /* it's a level-0 block of a regular object */ 663 arc_flags_t aflags = ARC_FLAG_WAIT; 664 arc_buf_t *abuf; 665 int blksz = dblkszsec << SPA_MINBLOCKSHIFT; 666 uint64_t offset; 667 668 ASSERT0(zb->zb_level); 669 ASSERT(zb->zb_object > dsa->dsa_resume_object || 670 (zb->zb_object == dsa->dsa_resume_object && 671 zb->zb_blkid * blksz >= dsa->dsa_resume_offset)); 672 673 if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf, 674 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, 675 &aflags, zb) != 0) { 676 if (zfs_send_corrupt_data) { 677 /* Send a block filled with 0x"zfs badd bloc" */ 678 abuf = arc_alloc_buf(spa, blksz, &abuf, 679 ARC_BUFC_DATA); 680 uint64_t *ptr; 681 for (ptr = abuf->b_data; 682 (char *)ptr < (char *)abuf->b_data + blksz; 683 ptr++) 684 *ptr = 0x2f5baddb10cULL; 685 } else { 686 return (SET_ERROR(EIO)); 687 } 688 } 689 690 offset = zb->zb_blkid * blksz; 691 692 if (!(dsa->dsa_featureflags & 693 DMU_BACKUP_FEATURE_LARGE_BLOCKS) && 694 blksz > SPA_OLD_MAXBLOCKSIZE) { 695 char *buf = abuf->b_data; 696 while (blksz > 0 && err == 0) { 697 int n = MIN(blksz, SPA_OLD_MAXBLOCKSIZE); 698 err = dump_write(dsa, type, zb->zb_object, 699 offset, n, NULL, buf); 700 offset += n; 701 buf += n; 702 blksz -= n; 703 } 704 } else { 705 err = dump_write(dsa, type, zb->zb_object, 706 offset, blksz, bp, abuf->b_data); 707 } 708 arc_buf_destroy(abuf, &abuf); 709 } 710 711 ASSERT(err == 0 || err == EINTR); 712 return (err); 713} 714 715/* 716 * Pop the new data off the queue, and free the old data. 717 */ 718static struct send_block_record * 719get_next_record(bqueue_t *bq, struct send_block_record *data) 720{ 721 struct send_block_record *tmp = bqueue_dequeue(bq); 722 kmem_free(data, sizeof (*data)); 723 return (tmp); 724} 725 726/* 727 * Actually do the bulk of the work in a zfs send. 728 * 729 * Note: Releases dp using the specified tag. 730 */ 731static int 732dmu_send_impl(void *tag, dsl_pool_t *dp, dsl_dataset_t *to_ds, 733 zfs_bookmark_phys_t *ancestor_zb, 734 boolean_t is_clone, boolean_t embedok, boolean_t large_block_ok, int outfd, 735 uint64_t resumeobj, uint64_t resumeoff, 736#ifdef illumos 737 vnode_t *vp, offset_t *off) 738#else 739 struct file *fp, offset_t *off) 740#endif 741{ 742 objset_t *os; 743 dmu_replay_record_t *drr; 744 dmu_sendarg_t *dsp; 745 int err; 746 uint64_t fromtxg = 0; 747 uint64_t featureflags = 0; 748 struct send_thread_arg to_arg = { 0 }; 749 750 err = dmu_objset_from_ds(to_ds, &os); 751 if (err != 0) { 752 dsl_pool_rele(dp, tag); 753 return (err); 754 } 755 756 drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP); 757 drr->drr_type = DRR_BEGIN; 758 drr->drr_u.drr_begin.drr_magic = DMU_BACKUP_MAGIC; 759 DMU_SET_STREAM_HDRTYPE(drr->drr_u.drr_begin.drr_versioninfo, 760 DMU_SUBSTREAM); 761 762#ifdef _KERNEL 763 if (dmu_objset_type(os) == DMU_OST_ZFS) { 764 uint64_t version; 765 if (zfs_get_zplprop(os, ZFS_PROP_VERSION, &version) != 0) { 766 kmem_free(drr, sizeof (dmu_replay_record_t)); 767 dsl_pool_rele(dp, tag); 768 return (SET_ERROR(EINVAL)); 769 } 770 if (version >= ZPL_VERSION_SA) { 771 featureflags |= DMU_BACKUP_FEATURE_SA_SPILL; 772 } 773 } 774#endif 775 776 if (large_block_ok && to_ds->ds_feature_inuse[SPA_FEATURE_LARGE_BLOCKS]) 777 featureflags |= DMU_BACKUP_FEATURE_LARGE_BLOCKS; 778 if (embedok && 779 spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA)) { 780 featureflags |= DMU_BACKUP_FEATURE_EMBED_DATA; 781 if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS)) 782 featureflags |= DMU_BACKUP_FEATURE_EMBED_DATA_LZ4; 783 } 784 785 if (resumeobj != 0 || resumeoff != 0) { 786 featureflags |= DMU_BACKUP_FEATURE_RESUMING; 787 } 788 789 DMU_SET_FEATUREFLAGS(drr->drr_u.drr_begin.drr_versioninfo, 790 featureflags); 791 792 drr->drr_u.drr_begin.drr_creation_time = 793 dsl_dataset_phys(to_ds)->ds_creation_time; 794 drr->drr_u.drr_begin.drr_type = dmu_objset_type(os); 795 if (is_clone) 796 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CLONE; 797 drr->drr_u.drr_begin.drr_toguid = dsl_dataset_phys(to_ds)->ds_guid; 798 if (dsl_dataset_phys(to_ds)->ds_flags & DS_FLAG_CI_DATASET) 799 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CI_DATA; 800 if (zfs_send_set_freerecords_bit) 801 drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_FREERECORDS; 802 803 if (ancestor_zb != NULL) { 804 drr->drr_u.drr_begin.drr_fromguid = 805 ancestor_zb->zbm_guid; 806 fromtxg = ancestor_zb->zbm_creation_txg; 807 } 808 dsl_dataset_name(to_ds, drr->drr_u.drr_begin.drr_toname); 809 if (!to_ds->ds_is_snapshot) { 810 (void) strlcat(drr->drr_u.drr_begin.drr_toname, "@--head--", 811 sizeof (drr->drr_u.drr_begin.drr_toname)); 812 } 813 814 dsp = kmem_zalloc(sizeof (dmu_sendarg_t), KM_SLEEP); 815 816 dsp->dsa_drr = drr; 817 dsp->dsa_outfd = outfd; 818 dsp->dsa_proc = curproc; 819 dsp->dsa_td = curthread; 820 dsp->dsa_fp = fp; 821 dsp->dsa_os = os; 822 dsp->dsa_off = off; 823 dsp->dsa_toguid = dsl_dataset_phys(to_ds)->ds_guid; 824 dsp->dsa_pending_op = PENDING_NONE; 825 dsp->dsa_featureflags = featureflags; 826 dsp->dsa_resume_object = resumeobj; 827 dsp->dsa_resume_offset = resumeoff; 828 829 mutex_enter(&to_ds->ds_sendstream_lock); 830 list_insert_head(&to_ds->ds_sendstreams, dsp); 831 mutex_exit(&to_ds->ds_sendstream_lock); 832 833 dsl_dataset_long_hold(to_ds, FTAG); 834 dsl_pool_rele(dp, tag); 835 836 void *payload = NULL; 837 size_t payload_len = 0; 838 if (resumeobj != 0 || resumeoff != 0) { 839 dmu_object_info_t to_doi; 840 err = dmu_object_info(os, resumeobj, &to_doi); 841 if (err != 0) 842 goto out; 843 SET_BOOKMARK(&to_arg.resume, to_ds->ds_object, resumeobj, 0, 844 resumeoff / to_doi.doi_data_block_size); 845 846 nvlist_t *nvl = fnvlist_alloc(); 847 fnvlist_add_uint64(nvl, "resume_object", resumeobj); 848 fnvlist_add_uint64(nvl, "resume_offset", resumeoff); 849 payload = fnvlist_pack(nvl, &payload_len); 850 drr->drr_payloadlen = payload_len; 851 fnvlist_free(nvl); 852 } 853 854 err = dump_record(dsp, payload, payload_len); 855 fnvlist_pack_free(payload, payload_len); 856 if (err != 0) { 857 err = dsp->dsa_err; 858 goto out; 859 } 860 861 err = bqueue_init(&to_arg.q, zfs_send_queue_length, 862 offsetof(struct send_block_record, ln)); 863 to_arg.error_code = 0; 864 to_arg.cancel = B_FALSE; 865 to_arg.ds = to_ds; 866 to_arg.fromtxg = fromtxg; 867 to_arg.flags = TRAVERSE_PRE | TRAVERSE_PREFETCH; 868 (void) thread_create(NULL, 0, send_traverse_thread, &to_arg, 0, &p0, 869 TS_RUN, minclsyspri); 870 871 struct send_block_record *to_data; 872 to_data = bqueue_dequeue(&to_arg.q); 873 874 while (!to_data->eos_marker && err == 0) { 875 err = do_dump(dsp, to_data); 876 to_data = get_next_record(&to_arg.q, to_data); 877 if (issig(JUSTLOOKING) && issig(FORREAL)) 878 err = EINTR; 879 } 880 881 if (err != 0) { 882 to_arg.cancel = B_TRUE; 883 while (!to_data->eos_marker) { 884 to_data = get_next_record(&to_arg.q, to_data); 885 } 886 } 887 kmem_free(to_data, sizeof (*to_data)); 888 889 bqueue_destroy(&to_arg.q); 890 891 if (err == 0 && to_arg.error_code != 0) 892 err = to_arg.error_code; 893 894 if (err != 0) 895 goto out; 896 897 if (dsp->dsa_pending_op != PENDING_NONE) 898 if (dump_record(dsp, NULL, 0) != 0) 899 err = SET_ERROR(EINTR); 900 901 if (err != 0) { 902 if (err == EINTR && dsp->dsa_err != 0) 903 err = dsp->dsa_err; 904 goto out; 905 } 906 907 bzero(drr, sizeof (dmu_replay_record_t)); 908 drr->drr_type = DRR_END; 909 drr->drr_u.drr_end.drr_checksum = dsp->dsa_zc; 910 drr->drr_u.drr_end.drr_toguid = dsp->dsa_toguid; 911 912 if (dump_record(dsp, NULL, 0) != 0) 913 err = dsp->dsa_err; 914 915out: 916 mutex_enter(&to_ds->ds_sendstream_lock); 917 list_remove(&to_ds->ds_sendstreams, dsp); 918 mutex_exit(&to_ds->ds_sendstream_lock); 919 920 VERIFY(err != 0 || (dsp->dsa_sent_begin && dsp->dsa_sent_end)); 921 922 kmem_free(drr, sizeof (dmu_replay_record_t)); 923 kmem_free(dsp, sizeof (dmu_sendarg_t)); 924 925 dsl_dataset_long_rele(to_ds, FTAG); 926 927 return (err); 928} 929 930int 931dmu_send_obj(const char *pool, uint64_t tosnap, uint64_t fromsnap, 932 boolean_t embedok, boolean_t large_block_ok, 933#ifdef illumos 934 int outfd, vnode_t *vp, offset_t *off) 935#else 936 int outfd, struct file *fp, offset_t *off) 937#endif 938{ 939 dsl_pool_t *dp; 940 dsl_dataset_t *ds; 941 dsl_dataset_t *fromds = NULL; 942 int err; 943 944 err = dsl_pool_hold(pool, FTAG, &dp); 945 if (err != 0) 946 return (err); 947 948 err = dsl_dataset_hold_obj(dp, tosnap, FTAG, &ds); 949 if (err != 0) { 950 dsl_pool_rele(dp, FTAG); 951 return (err); 952 } 953 954 if (fromsnap != 0) { 955 zfs_bookmark_phys_t zb; 956 boolean_t is_clone; 957 958 err = dsl_dataset_hold_obj(dp, fromsnap, FTAG, &fromds); 959 if (err != 0) { 960 dsl_dataset_rele(ds, FTAG); 961 dsl_pool_rele(dp, FTAG); 962 return (err); 963 } 964 if (!dsl_dataset_is_before(ds, fromds, 0)) 965 err = SET_ERROR(EXDEV); 966 zb.zbm_creation_time = 967 dsl_dataset_phys(fromds)->ds_creation_time; 968 zb.zbm_creation_txg = dsl_dataset_phys(fromds)->ds_creation_txg; 969 zb.zbm_guid = dsl_dataset_phys(fromds)->ds_guid; 970 is_clone = (fromds->ds_dir != ds->ds_dir); 971 dsl_dataset_rele(fromds, FTAG); 972 err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone, 973 embedok, large_block_ok, outfd, 0, 0, fp, off); 974 } else { 975 err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE, 976 embedok, large_block_ok, outfd, 0, 0, fp, off); 977 } 978 dsl_dataset_rele(ds, FTAG); 979 return (err); 980} 981 982int 983dmu_send(const char *tosnap, const char *fromsnap, boolean_t embedok, 984 boolean_t large_block_ok, int outfd, uint64_t resumeobj, uint64_t resumeoff, 985#ifdef illumos 986 vnode_t *vp, offset_t *off) 987#else 988 struct file *fp, offset_t *off) 989#endif 990{ 991 dsl_pool_t *dp; 992 dsl_dataset_t *ds; 993 int err; 994 boolean_t owned = B_FALSE; 995 996 if (fromsnap != NULL && strpbrk(fromsnap, "@#") == NULL) 997 return (SET_ERROR(EINVAL)); 998 999 err = dsl_pool_hold(tosnap, FTAG, &dp); 1000 if (err != 0) 1001 return (err); 1002 1003 if (strchr(tosnap, '@') == NULL && spa_writeable(dp->dp_spa)) { 1004 /* 1005 * We are sending a filesystem or volume. Ensure 1006 * that it doesn't change by owning the dataset. 1007 */ 1008 err = dsl_dataset_own(dp, tosnap, FTAG, &ds); 1009 owned = B_TRUE; 1010 } else { 1011 err = dsl_dataset_hold(dp, tosnap, FTAG, &ds); 1012 } 1013 if (err != 0) { 1014 dsl_pool_rele(dp, FTAG); 1015 return (err); 1016 } 1017 1018 if (fromsnap != NULL) { 1019 zfs_bookmark_phys_t zb; 1020 boolean_t is_clone = B_FALSE; 1021 int fsnamelen = strchr(tosnap, '@') - tosnap; 1022 1023 /* 1024 * If the fromsnap is in a different filesystem, then 1025 * mark the send stream as a clone. 1026 */ 1027 if (strncmp(tosnap, fromsnap, fsnamelen) != 0 || 1028 (fromsnap[fsnamelen] != '@' && 1029 fromsnap[fsnamelen] != '#')) { 1030 is_clone = B_TRUE; 1031 } 1032 1033 if (strchr(fromsnap, '@')) { 1034 dsl_dataset_t *fromds; 1035 err = dsl_dataset_hold(dp, fromsnap, FTAG, &fromds); 1036 if (err == 0) { 1037 if (!dsl_dataset_is_before(ds, fromds, 0)) 1038 err = SET_ERROR(EXDEV); 1039 zb.zbm_creation_time = 1040 dsl_dataset_phys(fromds)->ds_creation_time; 1041 zb.zbm_creation_txg = 1042 dsl_dataset_phys(fromds)->ds_creation_txg; 1043 zb.zbm_guid = dsl_dataset_phys(fromds)->ds_guid; 1044 is_clone = (ds->ds_dir != fromds->ds_dir); 1045 dsl_dataset_rele(fromds, FTAG); 1046 } 1047 } else { 1048 err = dsl_bookmark_lookup(dp, fromsnap, ds, &zb); 1049 } 1050 if (err != 0) { 1051 dsl_dataset_rele(ds, FTAG); 1052 dsl_pool_rele(dp, FTAG); 1053 return (err); 1054 } 1055 err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone, 1056 embedok, large_block_ok, 1057 outfd, resumeobj, resumeoff, fp, off); 1058 } else { 1059 err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE, 1060 embedok, large_block_ok, 1061 outfd, resumeobj, resumeoff, fp, off); 1062 } 1063 if (owned) 1064 dsl_dataset_disown(ds, FTAG); 1065 else 1066 dsl_dataset_rele(ds, FTAG); 1067 return (err); 1068} 1069 1070static int 1071dmu_adjust_send_estimate_for_indirects(dsl_dataset_t *ds, uint64_t size, 1072 uint64_t *sizep) 1073{ 1074 int err; 1075 /* 1076 * Assume that space (both on-disk and in-stream) is dominated by 1077 * data. We will adjust for indirect blocks and the copies property, 1078 * but ignore per-object space used (eg, dnodes and DRR_OBJECT records). 1079 */ 1080 1081 /* 1082 * Subtract out approximate space used by indirect blocks. 1083 * Assume most space is used by data blocks (non-indirect, non-dnode). 1084 * Assume all blocks are recordsize. Assume ditto blocks and 1085 * internal fragmentation counter out compression. 1086 * 1087 * Therefore, space used by indirect blocks is sizeof(blkptr_t) per 1088 * block, which we observe in practice. 1089 */ 1090 uint64_t recordsize; 1091 err = dsl_prop_get_int_ds(ds, "recordsize", &recordsize); 1092 if (err != 0) 1093 return (err); 1094 size -= size / recordsize * sizeof (blkptr_t); 1095 1096 /* Add in the space for the record associated with each block. */ 1097 size += size / recordsize * sizeof (dmu_replay_record_t); 1098 1099 *sizep = size; 1100 1101 return (0); 1102} 1103 1104int 1105dmu_send_estimate(dsl_dataset_t *ds, dsl_dataset_t *fromds, uint64_t *sizep) 1106{ 1107 dsl_pool_t *dp = ds->ds_dir->dd_pool; 1108 int err; 1109 uint64_t size; 1110 1111 ASSERT(dsl_pool_config_held(dp)); 1112 1113 /* tosnap must be a snapshot */ 1114 if (!ds->ds_is_snapshot) 1115 return (SET_ERROR(EINVAL)); 1116 1117 /* fromsnap, if provided, must be a snapshot */ 1118 if (fromds != NULL && !fromds->ds_is_snapshot) 1119 return (SET_ERROR(EINVAL)); 1120 1121 /* 1122 * fromsnap must be an earlier snapshot from the same fs as tosnap, 1123 * or the origin's fs. 1124 */ 1125 if (fromds != NULL && !dsl_dataset_is_before(ds, fromds, 0)) 1126 return (SET_ERROR(EXDEV)); 1127 1128 /* Get uncompressed size estimate of changed data. */ 1129 if (fromds == NULL) { 1130 size = dsl_dataset_phys(ds)->ds_uncompressed_bytes; 1131 } else { 1132 uint64_t used, comp; 1133 err = dsl_dataset_space_written(fromds, ds, 1134 &used, &comp, &size); 1135 if (err != 0) 1136 return (err); 1137 } 1138 1139 err = dmu_adjust_send_estimate_for_indirects(ds, size, sizep); 1140 return (err); 1141} 1142 1143/* 1144 * Simple callback used to traverse the blocks of a snapshot and sum their 1145 * uncompressed size 1146 */ 1147/* ARGSUSED */ 1148static int 1149dmu_calculate_send_traversal(spa_t *spa, zilog_t *zilog, const blkptr_t *bp, 1150 const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg) 1151{ 1152 uint64_t *spaceptr = arg; 1153 if (bp != NULL && !BP_IS_HOLE(bp)) { 1154 *spaceptr += BP_GET_UCSIZE(bp); 1155 } 1156 return (0); 1157} 1158 1159/* 1160 * Given a desination snapshot and a TXG, calculate the approximate size of a 1161 * send stream sent from that TXG. from_txg may be zero, indicating that the 1162 * whole snapshot will be sent. 1163 */ 1164int 1165dmu_send_estimate_from_txg(dsl_dataset_t *ds, uint64_t from_txg, 1166 uint64_t *sizep) 1167{ 1168 dsl_pool_t *dp = ds->ds_dir->dd_pool; 1169 int err; 1170 uint64_t size = 0; 1171 1172 ASSERT(dsl_pool_config_held(dp)); 1173 1174 /* tosnap must be a snapshot */ 1175 if (!dsl_dataset_is_snapshot(ds)) 1176 return (SET_ERROR(EINVAL)); 1177 1178 /* verify that from_txg is before the provided snapshot was taken */ 1179 if (from_txg >= dsl_dataset_phys(ds)->ds_creation_txg) { 1180 return (SET_ERROR(EXDEV)); 1181 } 1182 1183 /* 1184 * traverse the blocks of the snapshot with birth times after 1185 * from_txg, summing their uncompressed size 1186 */ 1187 err = traverse_dataset(ds, from_txg, TRAVERSE_POST, 1188 dmu_calculate_send_traversal, &size); 1189 if (err) 1190 return (err); 1191 1192 err = dmu_adjust_send_estimate_for_indirects(ds, size, sizep); 1193 return (err); 1194} 1195 1196typedef struct dmu_recv_begin_arg { 1197 const char *drba_origin; 1198 dmu_recv_cookie_t *drba_cookie; 1199 cred_t *drba_cred; 1200 uint64_t drba_snapobj; 1201} dmu_recv_begin_arg_t; 1202 1203static int 1204recv_begin_check_existing_impl(dmu_recv_begin_arg_t *drba, dsl_dataset_t *ds, 1205 uint64_t fromguid) 1206{ 1207 uint64_t val; 1208 int error; 1209 dsl_pool_t *dp = ds->ds_dir->dd_pool; 1210 1211 /* temporary clone name must not exist */ 1212 error = zap_lookup(dp->dp_meta_objset, 1213 dsl_dir_phys(ds->ds_dir)->dd_child_dir_zapobj, recv_clone_name, 1214 8, 1, &val); 1215 if (error != ENOENT) 1216 return (error == 0 ? EBUSY : error); 1217 1218 /* new snapshot name must not exist */ 1219 error = zap_lookup(dp->dp_meta_objset, 1220 dsl_dataset_phys(ds)->ds_snapnames_zapobj, 1221 drba->drba_cookie->drc_tosnap, 8, 1, &val); 1222 if (error != ENOENT) 1223 return (error == 0 ? EEXIST : error); 1224 1225 /* 1226 * Check snapshot limit before receiving. We'll recheck again at the 1227 * end, but might as well abort before receiving if we're already over 1228 * the limit. 1229 * 1230 * Note that we do not check the file system limit with 1231 * dsl_dir_fscount_check because the temporary %clones don't count 1232 * against that limit. 1233 */ 1234 error = dsl_fs_ss_limit_check(ds->ds_dir, 1, ZFS_PROP_SNAPSHOT_LIMIT, 1235 NULL, drba->drba_cred); 1236 if (error != 0) 1237 return (error); 1238 1239 if (fromguid != 0) { 1240 dsl_dataset_t *snap; 1241 uint64_t obj = dsl_dataset_phys(ds)->ds_prev_snap_obj; 1242 1243 /* Find snapshot in this dir that matches fromguid. */ 1244 while (obj != 0) { 1245 error = dsl_dataset_hold_obj(dp, obj, FTAG, 1246 &snap); 1247 if (error != 0) 1248 return (SET_ERROR(ENODEV)); 1249 if (snap->ds_dir != ds->ds_dir) { 1250 dsl_dataset_rele(snap, FTAG); 1251 return (SET_ERROR(ENODEV)); 1252 } 1253 if (dsl_dataset_phys(snap)->ds_guid == fromguid) 1254 break; 1255 obj = dsl_dataset_phys(snap)->ds_prev_snap_obj; 1256 dsl_dataset_rele(snap, FTAG); 1257 } 1258 if (obj == 0) 1259 return (SET_ERROR(ENODEV)); 1260 1261 if (drba->drba_cookie->drc_force) { 1262 drba->drba_snapobj = obj; 1263 } else { 1264 /* 1265 * If we are not forcing, there must be no 1266 * changes since fromsnap. 1267 */ 1268 if (dsl_dataset_modified_since_snap(ds, snap)) { 1269 dsl_dataset_rele(snap, FTAG); 1270 return (SET_ERROR(ETXTBSY)); 1271 } 1272 drba->drba_snapobj = ds->ds_prev->ds_object; 1273 } 1274 1275 dsl_dataset_rele(snap, FTAG); 1276 } else { 1277 /* if full, then must be forced */ 1278 if (!drba->drba_cookie->drc_force) 1279 return (SET_ERROR(EEXIST)); 1280 /* start from $ORIGIN@$ORIGIN, if supported */ 1281 drba->drba_snapobj = dp->dp_origin_snap != NULL ? 1282 dp->dp_origin_snap->ds_object : 0; 1283 } 1284 1285 return (0); 1286 1287} 1288 1289static int 1290dmu_recv_begin_check(void *arg, dmu_tx_t *tx) 1291{ 1292 dmu_recv_begin_arg_t *drba = arg; 1293 dsl_pool_t *dp = dmu_tx_pool(tx); 1294 struct drr_begin *drrb = drba->drba_cookie->drc_drrb; 1295 uint64_t fromguid = drrb->drr_fromguid; 1296 int flags = drrb->drr_flags; 1297 int error; 1298 uint64_t featureflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo); 1299 dsl_dataset_t *ds; 1300 const char *tofs = drba->drba_cookie->drc_tofs; 1301 1302 /* already checked */ 1303 ASSERT3U(drrb->drr_magic, ==, DMU_BACKUP_MAGIC); 1304 ASSERT(!(featureflags & DMU_BACKUP_FEATURE_RESUMING)); 1305 1306 if (DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) == 1307 DMU_COMPOUNDSTREAM || 1308 drrb->drr_type >= DMU_OST_NUMTYPES || 1309 ((flags & DRR_FLAG_CLONE) && drba->drba_origin == NULL)) 1310 return (SET_ERROR(EINVAL)); 1311 1312 /* Verify pool version supports SA if SA_SPILL feature set */ 1313 if ((featureflags & DMU_BACKUP_FEATURE_SA_SPILL) && 1314 spa_version(dp->dp_spa) < SPA_VERSION_SA) 1315 return (SET_ERROR(ENOTSUP)); 1316 1317 if (drba->drba_cookie->drc_resumable && 1318 !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EXTENSIBLE_DATASET)) 1319 return (SET_ERROR(ENOTSUP)); 1320 1321 /* 1322 * The receiving code doesn't know how to translate a WRITE_EMBEDDED 1323 * record to a plan WRITE record, so the pool must have the 1324 * EMBEDDED_DATA feature enabled if the stream has WRITE_EMBEDDED 1325 * records. Same with WRITE_EMBEDDED records that use LZ4 compression. 1326 */ 1327 if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA) && 1328 !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA)) 1329 return (SET_ERROR(ENOTSUP)); 1330 if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA_LZ4) && 1331 !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS)) 1332 return (SET_ERROR(ENOTSUP)); 1333 1334 /* 1335 * The receiving code doesn't know how to translate large blocks 1336 * to smaller ones, so the pool must have the LARGE_BLOCKS 1337 * feature enabled if the stream has LARGE_BLOCKS. 1338 */ 1339 if ((featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) && 1340 !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LARGE_BLOCKS)) 1341 return (SET_ERROR(ENOTSUP)); 1342 1343 error = dsl_dataset_hold(dp, tofs, FTAG, &ds); 1344 if (error == 0) { 1345 /* target fs already exists; recv into temp clone */ 1346 1347 /* Can't recv a clone into an existing fs */ 1348 if (flags & DRR_FLAG_CLONE || drba->drba_origin) { 1349 dsl_dataset_rele(ds, FTAG); 1350 return (SET_ERROR(EINVAL)); 1351 } 1352 1353 error = recv_begin_check_existing_impl(drba, ds, fromguid); 1354 dsl_dataset_rele(ds, FTAG); 1355 } else if (error == ENOENT) { 1356 /* target fs does not exist; must be a full backup or clone */ 1357 char buf[ZFS_MAX_DATASET_NAME_LEN]; 1358 1359 /* 1360 * If it's a non-clone incremental, we are missing the 1361 * target fs, so fail the recv. 1362 */ 1363 if (fromguid != 0 && !(flags & DRR_FLAG_CLONE || 1364 drba->drba_origin)) 1365 return (SET_ERROR(ENOENT)); 1366 1367 /* 1368 * If we're receiving a full send as a clone, and it doesn't 1369 * contain all the necessary free records and freeobject 1370 * records, reject it. 1371 */ 1372 if (fromguid == 0 && drba->drba_origin && 1373 !(flags & DRR_FLAG_FREERECORDS)) 1374 return (SET_ERROR(EINVAL)); 1375 1376 /* Open the parent of tofs */ 1377 ASSERT3U(strlen(tofs), <, sizeof (buf)); 1378 (void) strlcpy(buf, tofs, strrchr(tofs, '/') - tofs + 1); 1379 error = dsl_dataset_hold(dp, buf, FTAG, &ds); 1380 if (error != 0) 1381 return (error); 1382 1383 /* 1384 * Check filesystem and snapshot limits before receiving. We'll 1385 * recheck snapshot limits again at the end (we create the 1386 * filesystems and increment those counts during begin_sync). 1387 */ 1388 error = dsl_fs_ss_limit_check(ds->ds_dir, 1, 1389 ZFS_PROP_FILESYSTEM_LIMIT, NULL, drba->drba_cred); 1390 if (error != 0) { 1391 dsl_dataset_rele(ds, FTAG); 1392 return (error); 1393 } 1394 1395 error = dsl_fs_ss_limit_check(ds->ds_dir, 1, 1396 ZFS_PROP_SNAPSHOT_LIMIT, NULL, drba->drba_cred); 1397 if (error != 0) { 1398 dsl_dataset_rele(ds, FTAG); 1399 return (error); 1400 } 1401 1402 if (drba->drba_origin != NULL) { 1403 dsl_dataset_t *origin; 1404 error = dsl_dataset_hold(dp, drba->drba_origin, 1405 FTAG, &origin); 1406 if (error != 0) { 1407 dsl_dataset_rele(ds, FTAG); 1408 return (error); 1409 } 1410 if (!origin->ds_is_snapshot) { 1411 dsl_dataset_rele(origin, FTAG); 1412 dsl_dataset_rele(ds, FTAG); 1413 return (SET_ERROR(EINVAL)); 1414 } 1415 if (dsl_dataset_phys(origin)->ds_guid != fromguid && 1416 fromguid != 0) { 1417 dsl_dataset_rele(origin, FTAG); 1418 dsl_dataset_rele(ds, FTAG); 1419 return (SET_ERROR(ENODEV)); 1420 } 1421 dsl_dataset_rele(origin, FTAG); 1422 } 1423 dsl_dataset_rele(ds, FTAG); 1424 error = 0; 1425 } 1426 return (error); 1427} 1428 1429static void 1430dmu_recv_begin_sync(void *arg, dmu_tx_t *tx) 1431{ 1432 dmu_recv_begin_arg_t *drba = arg; 1433 dsl_pool_t *dp = dmu_tx_pool(tx); 1434 objset_t *mos = dp->dp_meta_objset; 1435 struct drr_begin *drrb = drba->drba_cookie->drc_drrb; 1436 const char *tofs = drba->drba_cookie->drc_tofs; 1437 dsl_dataset_t *ds, *newds; 1438 uint64_t dsobj; 1439 int error; 1440 uint64_t crflags = 0; 1441 1442 if (drrb->drr_flags & DRR_FLAG_CI_DATA) 1443 crflags |= DS_FLAG_CI_DATASET; 1444 1445 error = dsl_dataset_hold(dp, tofs, FTAG, &ds); 1446 if (error == 0) { 1447 /* create temporary clone */ 1448 dsl_dataset_t *snap = NULL; 1449 if (drba->drba_snapobj != 0) { 1450 VERIFY0(dsl_dataset_hold_obj(dp, 1451 drba->drba_snapobj, FTAG, &snap)); 1452 } 1453 dsobj = dsl_dataset_create_sync(ds->ds_dir, recv_clone_name, 1454 snap, crflags, drba->drba_cred, tx); 1455 if (drba->drba_snapobj != 0) 1456 dsl_dataset_rele(snap, FTAG); 1457 dsl_dataset_rele(ds, FTAG); 1458 } else { 1459 dsl_dir_t *dd; 1460 const char *tail; 1461 dsl_dataset_t *origin = NULL; 1462 1463 VERIFY0(dsl_dir_hold(dp, tofs, FTAG, &dd, &tail)); 1464 1465 if (drba->drba_origin != NULL) { 1466 VERIFY0(dsl_dataset_hold(dp, drba->drba_origin, 1467 FTAG, &origin)); 1468 } 1469 1470 /* Create new dataset. */ 1471 dsobj = dsl_dataset_create_sync(dd, 1472 strrchr(tofs, '/') + 1, 1473 origin, crflags, drba->drba_cred, tx); 1474 if (origin != NULL) 1475 dsl_dataset_rele(origin, FTAG); 1476 dsl_dir_rele(dd, FTAG); 1477 drba->drba_cookie->drc_newfs = B_TRUE; 1478 } 1479 VERIFY0(dsl_dataset_own_obj(dp, dsobj, dmu_recv_tag, &newds)); 1480 1481 if (drba->drba_cookie->drc_resumable) { 1482 dsl_dataset_zapify(newds, tx); 1483 if (drrb->drr_fromguid != 0) { 1484 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_FROMGUID, 1485 8, 1, &drrb->drr_fromguid, tx)); 1486 } 1487 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_TOGUID, 1488 8, 1, &drrb->drr_toguid, tx)); 1489 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_TONAME, 1490 1, strlen(drrb->drr_toname) + 1, drrb->drr_toname, tx)); 1491 uint64_t one = 1; 1492 uint64_t zero = 0; 1493 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_OBJECT, 1494 8, 1, &one, tx)); 1495 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_OFFSET, 1496 8, 1, &zero, tx)); 1497 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_BYTES, 1498 8, 1, &zero, tx)); 1499 if (DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo) & 1500 DMU_BACKUP_FEATURE_EMBED_DATA) { 1501 VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_EMBEDOK, 1502 8, 1, &one, tx)); 1503 } 1504 } 1505 1506 dmu_buf_will_dirty(newds->ds_dbuf, tx); 1507 dsl_dataset_phys(newds)->ds_flags |= DS_FLAG_INCONSISTENT; 1508 1509 /* 1510 * If we actually created a non-clone, we need to create the 1511 * objset in our new dataset. 1512 */ 1513 if (BP_IS_HOLE(dsl_dataset_get_blkptr(newds))) { 1514 (void) dmu_objset_create_impl(dp->dp_spa, 1515 newds, dsl_dataset_get_blkptr(newds), drrb->drr_type, tx); 1516 } 1517 1518 drba->drba_cookie->drc_ds = newds; 1519 1520 spa_history_log_internal_ds(newds, "receive", tx, ""); 1521} 1522 1523static int 1524dmu_recv_resume_begin_check(void *arg, dmu_tx_t *tx) 1525{ 1526 dmu_recv_begin_arg_t *drba = arg; 1527 dsl_pool_t *dp = dmu_tx_pool(tx); 1528 struct drr_begin *drrb = drba->drba_cookie->drc_drrb; 1529 int error; 1530 uint64_t featureflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo); 1531 dsl_dataset_t *ds; 1532 const char *tofs = drba->drba_cookie->drc_tofs; 1533 1534 /* already checked */ 1535 ASSERT3U(drrb->drr_magic, ==, DMU_BACKUP_MAGIC); 1536 ASSERT(featureflags & DMU_BACKUP_FEATURE_RESUMING); 1537 1538 if (DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) == 1539 DMU_COMPOUNDSTREAM || 1540 drrb->drr_type >= DMU_OST_NUMTYPES) 1541 return (SET_ERROR(EINVAL)); 1542 1543 /* Verify pool version supports SA if SA_SPILL feature set */ 1544 if ((featureflags & DMU_BACKUP_FEATURE_SA_SPILL) && 1545 spa_version(dp->dp_spa) < SPA_VERSION_SA) 1546 return (SET_ERROR(ENOTSUP)); 1547 1548 /* 1549 * The receiving code doesn't know how to translate a WRITE_EMBEDDED 1550 * record to a plain WRITE record, so the pool must have the 1551 * EMBEDDED_DATA feature enabled if the stream has WRITE_EMBEDDED 1552 * records. Same with WRITE_EMBEDDED records that use LZ4 compression. 1553 */ 1554 if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA) && 1555 !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA)) 1556 return (SET_ERROR(ENOTSUP)); 1557 if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA_LZ4) && 1558 !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS)) 1559 return (SET_ERROR(ENOTSUP)); 1560 1561 /* 6 extra bytes for /%recv */ 1562 char recvname[ZFS_MAX_DATASET_NAME_LEN + 6]; 1563 1564 (void) snprintf(recvname, sizeof (recvname), "%s/%s", 1565 tofs, recv_clone_name); 1566 1567 if (dsl_dataset_hold(dp, recvname, FTAG, &ds) != 0) { 1568 /* %recv does not exist; continue in tofs */ 1569 error = dsl_dataset_hold(dp, tofs, FTAG, &ds); 1570 if (error != 0) 1571 return (error); 1572 } 1573 1574 /* check that ds is marked inconsistent */ 1575 if (!DS_IS_INCONSISTENT(ds)) { 1576 dsl_dataset_rele(ds, FTAG); 1577 return (SET_ERROR(EINVAL)); 1578 } 1579 1580 /* check that there is resuming data, and that the toguid matches */ 1581 if (!dsl_dataset_is_zapified(ds)) { 1582 dsl_dataset_rele(ds, FTAG); 1583 return (SET_ERROR(EINVAL)); 1584 } 1585 uint64_t val; 1586 error = zap_lookup(dp->dp_meta_objset, ds->ds_object, 1587 DS_FIELD_RESUME_TOGUID, sizeof (val), 1, &val); 1588 if (error != 0 || drrb->drr_toguid != val) { 1589 dsl_dataset_rele(ds, FTAG); 1590 return (SET_ERROR(EINVAL)); 1591 } 1592 1593 /* 1594 * Check if the receive is still running. If so, it will be owned. 1595 * Note that nothing else can own the dataset (e.g. after the receive 1596 * fails) because it will be marked inconsistent. 1597 */ 1598 if (dsl_dataset_has_owner(ds)) { 1599 dsl_dataset_rele(ds, FTAG); 1600 return (SET_ERROR(EBUSY)); 1601 } 1602 1603 /* There should not be any snapshots of this fs yet. */ 1604 if (ds->ds_prev != NULL && ds->ds_prev->ds_dir == ds->ds_dir) { 1605 dsl_dataset_rele(ds, FTAG); 1606 return (SET_ERROR(EINVAL)); 1607 } 1608 1609 /* 1610 * Note: resume point will be checked when we process the first WRITE 1611 * record. 1612 */ 1613 1614 /* check that the origin matches */ 1615 val = 0; 1616 (void) zap_lookup(dp->dp_meta_objset, ds->ds_object, 1617 DS_FIELD_RESUME_FROMGUID, sizeof (val), 1, &val); 1618 if (drrb->drr_fromguid != val) { 1619 dsl_dataset_rele(ds, FTAG); 1620 return (SET_ERROR(EINVAL)); 1621 } 1622 1623 dsl_dataset_rele(ds, FTAG); 1624 return (0); 1625} 1626 1627static void 1628dmu_recv_resume_begin_sync(void *arg, dmu_tx_t *tx) 1629{ 1630 dmu_recv_begin_arg_t *drba = arg; 1631 dsl_pool_t *dp = dmu_tx_pool(tx); 1632 const char *tofs = drba->drba_cookie->drc_tofs; 1633 dsl_dataset_t *ds; 1634 uint64_t dsobj; 1635 /* 6 extra bytes for /%recv */ 1636 char recvname[ZFS_MAX_DATASET_NAME_LEN + 6]; 1637 1638 (void) snprintf(recvname, sizeof (recvname), "%s/%s", 1639 tofs, recv_clone_name); 1640 1641 if (dsl_dataset_hold(dp, recvname, FTAG, &ds) != 0) { 1642 /* %recv does not exist; continue in tofs */ 1643 VERIFY0(dsl_dataset_hold(dp, tofs, FTAG, &ds)); 1644 drba->drba_cookie->drc_newfs = B_TRUE; 1645 } 1646 1647 /* clear the inconsistent flag so that we can own it */ 1648 ASSERT(DS_IS_INCONSISTENT(ds)); 1649 dmu_buf_will_dirty(ds->ds_dbuf, tx); 1650 dsl_dataset_phys(ds)->ds_flags &= ~DS_FLAG_INCONSISTENT; 1651 dsobj = ds->ds_object; 1652 dsl_dataset_rele(ds, FTAG); 1653 1654 VERIFY0(dsl_dataset_own_obj(dp, dsobj, dmu_recv_tag, &ds)); 1655 1656 dmu_buf_will_dirty(ds->ds_dbuf, tx); 1657 dsl_dataset_phys(ds)->ds_flags |= DS_FLAG_INCONSISTENT; 1658 1659 ASSERT(!BP_IS_HOLE(dsl_dataset_get_blkptr(ds))); 1660 1661 drba->drba_cookie->drc_ds = ds; 1662 1663 spa_history_log_internal_ds(ds, "resume receive", tx, ""); 1664} 1665 1666/* 1667 * NB: callers *MUST* call dmu_recv_stream() if dmu_recv_begin() 1668 * succeeds; otherwise we will leak the holds on the datasets. 1669 */ 1670int 1671dmu_recv_begin(char *tofs, char *tosnap, dmu_replay_record_t *drr_begin, 1672 boolean_t force, boolean_t resumable, char *origin, dmu_recv_cookie_t *drc) 1673{ 1674 dmu_recv_begin_arg_t drba = { 0 }; 1675 1676 bzero(drc, sizeof (dmu_recv_cookie_t)); 1677 drc->drc_drr_begin = drr_begin; 1678 drc->drc_drrb = &drr_begin->drr_u.drr_begin; 1679 drc->drc_tosnap = tosnap; 1680 drc->drc_tofs = tofs; 1681 drc->drc_force = force; 1682 drc->drc_resumable = resumable; 1683 drc->drc_cred = CRED(); 1684 1685 if (drc->drc_drrb->drr_magic == BSWAP_64(DMU_BACKUP_MAGIC)) { 1686 drc->drc_byteswap = B_TRUE; 1687 fletcher_4_incremental_byteswap(drr_begin, 1688 sizeof (dmu_replay_record_t), &drc->drc_cksum); 1689 byteswap_record(drr_begin); 1690 } else if (drc->drc_drrb->drr_magic == DMU_BACKUP_MAGIC) { 1691 fletcher_4_incremental_native(drr_begin, 1692 sizeof (dmu_replay_record_t), &drc->drc_cksum); 1693 } else { 1694 return (SET_ERROR(EINVAL)); 1695 } 1696 1697 drba.drba_origin = origin; 1698 drba.drba_cookie = drc; 1699 drba.drba_cred = CRED(); 1700 1701 if (DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo) & 1702 DMU_BACKUP_FEATURE_RESUMING) { 1703 return (dsl_sync_task(tofs, 1704 dmu_recv_resume_begin_check, dmu_recv_resume_begin_sync, 1705 &drba, 5, ZFS_SPACE_CHECK_NORMAL)); 1706 } else { 1707 return (dsl_sync_task(tofs, 1708 dmu_recv_begin_check, dmu_recv_begin_sync, 1709 &drba, 5, ZFS_SPACE_CHECK_NORMAL)); 1710 } 1711} 1712 1713struct receive_record_arg { 1714 dmu_replay_record_t header; 1715 void *payload; /* Pointer to a buffer containing the payload */ 1716 /* 1717 * If the record is a write, pointer to the arc_buf_t containing the 1718 * payload. 1719 */ 1720 arc_buf_t *write_buf; 1721 int payload_size; 1722 uint64_t bytes_read; /* bytes read from stream when record created */ 1723 boolean_t eos_marker; /* Marks the end of the stream */ 1724 bqueue_node_t node; 1725}; 1726 1727struct receive_writer_arg { 1728 objset_t *os; 1729 boolean_t byteswap; 1730 bqueue_t q; 1731 1732 /* 1733 * These three args are used to signal to the main thread that we're 1734 * done. 1735 */ 1736 kmutex_t mutex; 1737 kcondvar_t cv; 1738 boolean_t done; 1739 1740 int err; 1741 /* A map from guid to dataset to help handle dedup'd streams. */ 1742 avl_tree_t *guid_to_ds_map; 1743 boolean_t resumable; 1744 uint64_t last_object, last_offset; 1745 uint64_t bytes_read; /* bytes read when current record created */ 1746}; 1747 1748struct objlist { 1749 list_t list; /* List of struct receive_objnode. */ 1750 /* 1751 * Last object looked up. Used to assert that objects are being looked 1752 * up in ascending order. 1753 */ 1754 uint64_t last_lookup; 1755}; 1756 1757struct receive_objnode { 1758 list_node_t node; 1759 uint64_t object; 1760}; 1761 1762struct receive_arg { 1763 objset_t *os; 1764 kthread_t *td; 1765 struct file *fp; 1766 uint64_t voff; /* The current offset in the stream */ 1767 uint64_t bytes_read; 1768 /* 1769 * A record that has had its payload read in, but hasn't yet been handed 1770 * off to the worker thread. 1771 */ 1772 struct receive_record_arg *rrd; 1773 /* A record that has had its header read in, but not its payload. */ 1774 struct receive_record_arg *next_rrd; 1775 zio_cksum_t cksum; 1776 zio_cksum_t prev_cksum; 1777 int err; 1778 boolean_t byteswap; 1779 /* Sorted list of objects not to issue prefetches for. */ 1780 struct objlist ignore_objlist; 1781}; 1782 1783typedef struct guid_map_entry { 1784 uint64_t guid; 1785 dsl_dataset_t *gme_ds; 1786 avl_node_t avlnode; 1787} guid_map_entry_t; 1788 1789static int 1790guid_compare(const void *arg1, const void *arg2) 1791{ 1792 const guid_map_entry_t *gmep1 = arg1; 1793 const guid_map_entry_t *gmep2 = arg2; 1794 1795 if (gmep1->guid < gmep2->guid) 1796 return (-1); 1797 else if (gmep1->guid > gmep2->guid) 1798 return (1); 1799 return (0); 1800} 1801 1802static void 1803free_guid_map_onexit(void *arg) 1804{ 1805 avl_tree_t *ca = arg; 1806 void *cookie = NULL; 1807 guid_map_entry_t *gmep; 1808 1809 while ((gmep = avl_destroy_nodes(ca, &cookie)) != NULL) { 1810 dsl_dataset_long_rele(gmep->gme_ds, gmep); 1811 dsl_dataset_rele(gmep->gme_ds, gmep); 1812 kmem_free(gmep, sizeof (guid_map_entry_t)); 1813 } 1814 avl_destroy(ca); 1815 kmem_free(ca, sizeof (avl_tree_t)); 1816} 1817 1818static int 1819restore_bytes(struct receive_arg *ra, void *buf, int len, off_t off, ssize_t *resid) 1820{ 1821 struct uio auio; 1822 struct iovec aiov; 1823 int error; 1824 1825 aiov.iov_base = buf; 1826 aiov.iov_len = len; 1827 auio.uio_iov = &aiov; 1828 auio.uio_iovcnt = 1; 1829 auio.uio_resid = len; 1830 auio.uio_segflg = UIO_SYSSPACE; 1831 auio.uio_rw = UIO_READ; 1832 auio.uio_offset = off; 1833 auio.uio_td = ra->td; 1834#ifdef _KERNEL 1835 error = fo_read(ra->fp, &auio, ra->td->td_ucred, FOF_OFFSET, ra->td); 1836#else 1837 fprintf(stderr, "%s: returning EOPNOTSUPP\n", __func__); 1838 error = EOPNOTSUPP; 1839#endif 1840 *resid = auio.uio_resid; 1841 return (error); 1842} 1843 1844static int 1845receive_read(struct receive_arg *ra, int len, void *buf) 1846{ 1847 int done = 0; 1848 1849 /* 1850 * The code doesn't rely on this (lengths being multiples of 8). See 1851 * comment in dump_bytes. 1852 */ 1853 ASSERT0(len % 8); 1854 1855 while (done < len) { 1856 ssize_t resid; 1857 1858 ra->err = restore_bytes(ra, buf + done, 1859 len - done, ra->voff, &resid); 1860 1861 if (resid == len - done) { 1862 /* 1863 * Note: ECKSUM indicates that the receive 1864 * was interrupted and can potentially be resumed. 1865 */ 1866 ra->err = SET_ERROR(ECKSUM); 1867 } 1868 ra->voff += len - done - resid; 1869 done = len - resid; 1870 if (ra->err != 0) 1871 return (ra->err); 1872 } 1873 1874 ra->bytes_read += len; 1875 1876 ASSERT3U(done, ==, len); 1877 return (0); 1878} 1879 1880static void 1881byteswap_record(dmu_replay_record_t *drr) 1882{ 1883#define DO64(X) (drr->drr_u.X = BSWAP_64(drr->drr_u.X)) 1884#define DO32(X) (drr->drr_u.X = BSWAP_32(drr->drr_u.X)) 1885 drr->drr_type = BSWAP_32(drr->drr_type); 1886 drr->drr_payloadlen = BSWAP_32(drr->drr_payloadlen); 1887 1888 switch (drr->drr_type) { 1889 case DRR_BEGIN: 1890 DO64(drr_begin.drr_magic); 1891 DO64(drr_begin.drr_versioninfo); 1892 DO64(drr_begin.drr_creation_time); 1893 DO32(drr_begin.drr_type); 1894 DO32(drr_begin.drr_flags); 1895 DO64(drr_begin.drr_toguid); 1896 DO64(drr_begin.drr_fromguid); 1897 break; 1898 case DRR_OBJECT: 1899 DO64(drr_object.drr_object); 1900 DO32(drr_object.drr_type); 1901 DO32(drr_object.drr_bonustype); 1902 DO32(drr_object.drr_blksz); 1903 DO32(drr_object.drr_bonuslen); 1904 DO64(drr_object.drr_toguid); 1905 break; 1906 case DRR_FREEOBJECTS: 1907 DO64(drr_freeobjects.drr_firstobj); 1908 DO64(drr_freeobjects.drr_numobjs); 1909 DO64(drr_freeobjects.drr_toguid); 1910 break; 1911 case DRR_WRITE: 1912 DO64(drr_write.drr_object); 1913 DO32(drr_write.drr_type); 1914 DO64(drr_write.drr_offset); 1915 DO64(drr_write.drr_length); 1916 DO64(drr_write.drr_toguid); 1917 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_write.drr_key.ddk_cksum); 1918 DO64(drr_write.drr_key.ddk_prop); 1919 break; 1920 case DRR_WRITE_BYREF: 1921 DO64(drr_write_byref.drr_object); 1922 DO64(drr_write_byref.drr_offset); 1923 DO64(drr_write_byref.drr_length); 1924 DO64(drr_write_byref.drr_toguid); 1925 DO64(drr_write_byref.drr_refguid); 1926 DO64(drr_write_byref.drr_refobject); 1927 DO64(drr_write_byref.drr_refoffset); 1928 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_write_byref. 1929 drr_key.ddk_cksum); 1930 DO64(drr_write_byref.drr_key.ddk_prop); 1931 break; 1932 case DRR_WRITE_EMBEDDED: 1933 DO64(drr_write_embedded.drr_object); 1934 DO64(drr_write_embedded.drr_offset); 1935 DO64(drr_write_embedded.drr_length); 1936 DO64(drr_write_embedded.drr_toguid); 1937 DO32(drr_write_embedded.drr_lsize); 1938 DO32(drr_write_embedded.drr_psize); 1939 break; 1940 case DRR_FREE: 1941 DO64(drr_free.drr_object); 1942 DO64(drr_free.drr_offset); 1943 DO64(drr_free.drr_length); 1944 DO64(drr_free.drr_toguid); 1945 break; 1946 case DRR_SPILL: 1947 DO64(drr_spill.drr_object); 1948 DO64(drr_spill.drr_length); 1949 DO64(drr_spill.drr_toguid); 1950 break; 1951 case DRR_END: 1952 DO64(drr_end.drr_toguid); 1953 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_end.drr_checksum); 1954 break; 1955 } 1956 1957 if (drr->drr_type != DRR_BEGIN) { 1958 ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_checksum.drr_checksum); 1959 } 1960 1961#undef DO64 1962#undef DO32 1963} 1964 1965static inline uint8_t 1966deduce_nblkptr(dmu_object_type_t bonus_type, uint64_t bonus_size) 1967{ 1968 if (bonus_type == DMU_OT_SA) { 1969 return (1); 1970 } else { 1971 return (1 + 1972 ((DN_MAX_BONUSLEN - bonus_size) >> SPA_BLKPTRSHIFT)); 1973 } 1974} 1975 1976static void 1977save_resume_state(struct receive_writer_arg *rwa, 1978 uint64_t object, uint64_t offset, dmu_tx_t *tx) 1979{ 1980 int txgoff = dmu_tx_get_txg(tx) & TXG_MASK; 1981 1982 if (!rwa->resumable) 1983 return; 1984 1985 /* 1986 * We use ds_resume_bytes[] != 0 to indicate that we need to 1987 * update this on disk, so it must not be 0. 1988 */ 1989 ASSERT(rwa->bytes_read != 0); 1990 1991 /* 1992 * We only resume from write records, which have a valid 1993 * (non-meta-dnode) object number. 1994 */ 1995 ASSERT(object != 0); 1996 1997 /* 1998 * For resuming to work correctly, we must receive records in order, 1999 * sorted by object,offset. This is checked by the callers, but 2000 * assert it here for good measure. 2001 */ 2002 ASSERT3U(object, >=, rwa->os->os_dsl_dataset->ds_resume_object[txgoff]); 2003 ASSERT(object != rwa->os->os_dsl_dataset->ds_resume_object[txgoff] || 2004 offset >= rwa->os->os_dsl_dataset->ds_resume_offset[txgoff]); 2005 ASSERT3U(rwa->bytes_read, >=, 2006 rwa->os->os_dsl_dataset->ds_resume_bytes[txgoff]); 2007 2008 rwa->os->os_dsl_dataset->ds_resume_object[txgoff] = object; 2009 rwa->os->os_dsl_dataset->ds_resume_offset[txgoff] = offset; 2010 rwa->os->os_dsl_dataset->ds_resume_bytes[txgoff] = rwa->bytes_read; 2011} 2012 2013static int 2014receive_object(struct receive_writer_arg *rwa, struct drr_object *drro, 2015 void *data) 2016{ 2017 dmu_object_info_t doi; 2018 dmu_tx_t *tx; 2019 uint64_t object; 2020 int err; 2021 2022 if (drro->drr_type == DMU_OT_NONE || 2023 !DMU_OT_IS_VALID(drro->drr_type) || 2024 !DMU_OT_IS_VALID(drro->drr_bonustype) || 2025 drro->drr_checksumtype >= ZIO_CHECKSUM_FUNCTIONS || 2026 drro->drr_compress >= ZIO_COMPRESS_FUNCTIONS || 2027 P2PHASE(drro->drr_blksz, SPA_MINBLOCKSIZE) || 2028 drro->drr_blksz < SPA_MINBLOCKSIZE || 2029 drro->drr_blksz > spa_maxblocksize(dmu_objset_spa(rwa->os)) || 2030 drro->drr_bonuslen > DN_MAX_BONUSLEN) { 2031 return (SET_ERROR(EINVAL)); 2032 } 2033 2034 err = dmu_object_info(rwa->os, drro->drr_object, &doi); 2035 2036 if (err != 0 && err != ENOENT) 2037 return (SET_ERROR(EINVAL)); 2038 object = err == 0 ? drro->drr_object : DMU_NEW_OBJECT; 2039 2040 /* 2041 * If we are losing blkptrs or changing the block size this must 2042 * be a new file instance. We must clear out the previous file 2043 * contents before we can change this type of metadata in the dnode. 2044 */ 2045 if (err == 0) { 2046 int nblkptr; 2047 2048 nblkptr = deduce_nblkptr(drro->drr_bonustype, 2049 drro->drr_bonuslen); 2050 2051 if (drro->drr_blksz != doi.doi_data_block_size || 2052 nblkptr < doi.doi_nblkptr) { 2053 err = dmu_free_long_range(rwa->os, drro->drr_object, 2054 0, DMU_OBJECT_END); 2055 if (err != 0) 2056 return (SET_ERROR(EINVAL)); 2057 } 2058 } 2059 2060 tx = dmu_tx_create(rwa->os); 2061 dmu_tx_hold_bonus(tx, object); 2062 err = dmu_tx_assign(tx, TXG_WAIT); 2063 if (err != 0) { 2064 dmu_tx_abort(tx); 2065 return (err); 2066 } 2067 2068 if (object == DMU_NEW_OBJECT) { 2069 /* currently free, want to be allocated */ 2070 err = dmu_object_claim(rwa->os, drro->drr_object, 2071 drro->drr_type, drro->drr_blksz, 2072 drro->drr_bonustype, drro->drr_bonuslen, tx); 2073 } else if (drro->drr_type != doi.doi_type || 2074 drro->drr_blksz != doi.doi_data_block_size || 2075 drro->drr_bonustype != doi.doi_bonus_type || 2076 drro->drr_bonuslen != doi.doi_bonus_size) { 2077 /* currently allocated, but with different properties */ 2078 err = dmu_object_reclaim(rwa->os, drro->drr_object, 2079 drro->drr_type, drro->drr_blksz, 2080 drro->drr_bonustype, drro->drr_bonuslen, tx); 2081 } 2082 if (err != 0) { 2083 dmu_tx_commit(tx); 2084 return (SET_ERROR(EINVAL)); 2085 } 2086 2087 dmu_object_set_checksum(rwa->os, drro->drr_object, 2088 drro->drr_checksumtype, tx); 2089 dmu_object_set_compress(rwa->os, drro->drr_object, 2090 drro->drr_compress, tx); 2091 2092 if (data != NULL) { 2093 dmu_buf_t *db; 2094 2095 VERIFY0(dmu_bonus_hold(rwa->os, drro->drr_object, FTAG, &db)); 2096 dmu_buf_will_dirty(db, tx); 2097 2098 ASSERT3U(db->db_size, >=, drro->drr_bonuslen); 2099 bcopy(data, db->db_data, drro->drr_bonuslen); 2100 if (rwa->byteswap) { 2101 dmu_object_byteswap_t byteswap = 2102 DMU_OT_BYTESWAP(drro->drr_bonustype); 2103 dmu_ot_byteswap[byteswap].ob_func(db->db_data, 2104 drro->drr_bonuslen); 2105 } 2106 dmu_buf_rele(db, FTAG); 2107 } 2108 dmu_tx_commit(tx); 2109 2110 return (0); 2111} 2112 2113/* ARGSUSED */ 2114static int 2115receive_freeobjects(struct receive_writer_arg *rwa, 2116 struct drr_freeobjects *drrfo) 2117{ 2118 uint64_t obj; 2119 int next_err = 0; 2120 2121 if (drrfo->drr_firstobj + drrfo->drr_numobjs < drrfo->drr_firstobj) 2122 return (SET_ERROR(EINVAL)); 2123 2124 for (obj = drrfo->drr_firstobj; 2125 obj < drrfo->drr_firstobj + drrfo->drr_numobjs && next_err == 0; 2126 next_err = dmu_object_next(rwa->os, &obj, FALSE, 0)) { 2127 int err; 2128 2129 if (dmu_object_info(rwa->os, obj, NULL) != 0) 2130 continue; 2131 2132 err = dmu_free_long_object(rwa->os, obj); 2133 if (err != 0) 2134 return (err); 2135 } 2136 if (next_err != ESRCH) 2137 return (next_err); 2138 return (0); 2139} 2140 2141static int 2142receive_write(struct receive_writer_arg *rwa, struct drr_write *drrw, 2143 arc_buf_t *abuf) 2144{ 2145 dmu_tx_t *tx; 2146 int err; 2147 2148 if (drrw->drr_offset + drrw->drr_length < drrw->drr_offset || 2149 !DMU_OT_IS_VALID(drrw->drr_type)) 2150 return (SET_ERROR(EINVAL)); 2151 2152 /* 2153 * For resuming to work, records must be in increasing order 2154 * by (object, offset). 2155 */ 2156 if (drrw->drr_object < rwa->last_object || 2157 (drrw->drr_object == rwa->last_object && 2158 drrw->drr_offset < rwa->last_offset)) { 2159 return (SET_ERROR(EINVAL)); 2160 } 2161 rwa->last_object = drrw->drr_object; 2162 rwa->last_offset = drrw->drr_offset; 2163 2164 if (dmu_object_info(rwa->os, drrw->drr_object, NULL) != 0) 2165 return (SET_ERROR(EINVAL)); 2166 2167 tx = dmu_tx_create(rwa->os); 2168 2169 dmu_tx_hold_write(tx, drrw->drr_object, 2170 drrw->drr_offset, drrw->drr_length); 2171 err = dmu_tx_assign(tx, TXG_WAIT); 2172 if (err != 0) { 2173 dmu_tx_abort(tx); 2174 return (err); 2175 } 2176 if (rwa->byteswap) { 2177 dmu_object_byteswap_t byteswap = 2178 DMU_OT_BYTESWAP(drrw->drr_type); 2179 dmu_ot_byteswap[byteswap].ob_func(abuf->b_data, 2180 drrw->drr_length); 2181 } 2182 2183 dmu_buf_t *bonus; 2184 if (dmu_bonus_hold(rwa->os, drrw->drr_object, FTAG, &bonus) != 0) 2185 return (SET_ERROR(EINVAL)); 2186 dmu_assign_arcbuf(bonus, drrw->drr_offset, abuf, tx); 2187 2188 /* 2189 * Note: If the receive fails, we want the resume stream to start 2190 * with the same record that we last successfully received (as opposed 2191 * to the next record), so that we can verify that we are 2192 * resuming from the correct location. 2193 */ 2194 save_resume_state(rwa, drrw->drr_object, drrw->drr_offset, tx); 2195 dmu_tx_commit(tx); 2196 dmu_buf_rele(bonus, FTAG); 2197 2198 return (0); 2199} 2200 2201/* 2202 * Handle a DRR_WRITE_BYREF record. This record is used in dedup'ed 2203 * streams to refer to a copy of the data that is already on the 2204 * system because it came in earlier in the stream. This function 2205 * finds the earlier copy of the data, and uses that copy instead of 2206 * data from the stream to fulfill this write. 2207 */ 2208static int 2209receive_write_byref(struct receive_writer_arg *rwa, 2210 struct drr_write_byref *drrwbr) 2211{ 2212 dmu_tx_t *tx; 2213 int err; 2214 guid_map_entry_t gmesrch; 2215 guid_map_entry_t *gmep; 2216 avl_index_t where; 2217 objset_t *ref_os = NULL; 2218 dmu_buf_t *dbp; 2219 2220 if (drrwbr->drr_offset + drrwbr->drr_length < drrwbr->drr_offset) 2221 return (SET_ERROR(EINVAL)); 2222 2223 /* 2224 * If the GUID of the referenced dataset is different from the 2225 * GUID of the target dataset, find the referenced dataset. 2226 */ 2227 if (drrwbr->drr_toguid != drrwbr->drr_refguid) { 2228 gmesrch.guid = drrwbr->drr_refguid; 2229 if ((gmep = avl_find(rwa->guid_to_ds_map, &gmesrch, 2230 &where)) == NULL) { 2231 return (SET_ERROR(EINVAL)); 2232 } 2233 if (dmu_objset_from_ds(gmep->gme_ds, &ref_os)) 2234 return (SET_ERROR(EINVAL)); 2235 } else { 2236 ref_os = rwa->os; 2237 } 2238 2239 err = dmu_buf_hold(ref_os, drrwbr->drr_refobject, 2240 drrwbr->drr_refoffset, FTAG, &dbp, DMU_READ_PREFETCH); 2241 if (err != 0) 2242 return (err); 2243 2244 tx = dmu_tx_create(rwa->os); 2245 2246 dmu_tx_hold_write(tx, drrwbr->drr_object, 2247 drrwbr->drr_offset, drrwbr->drr_length); 2248 err = dmu_tx_assign(tx, TXG_WAIT); 2249 if (err != 0) { 2250 dmu_tx_abort(tx); 2251 return (err); 2252 } 2253 dmu_write(rwa->os, drrwbr->drr_object, 2254 drrwbr->drr_offset, drrwbr->drr_length, dbp->db_data, tx); 2255 dmu_buf_rele(dbp, FTAG); 2256 2257 /* See comment in restore_write. */ 2258 save_resume_state(rwa, drrwbr->drr_object, drrwbr->drr_offset, tx); 2259 dmu_tx_commit(tx); 2260 return (0); 2261} 2262 2263static int 2264receive_write_embedded(struct receive_writer_arg *rwa, 2265 struct drr_write_embedded *drrwe, void *data) 2266{ 2267 dmu_tx_t *tx; 2268 int err; 2269 2270 if (drrwe->drr_offset + drrwe->drr_length < drrwe->drr_offset) 2271 return (EINVAL); 2272 2273 if (drrwe->drr_psize > BPE_PAYLOAD_SIZE) 2274 return (EINVAL); 2275 2276 if (drrwe->drr_etype >= NUM_BP_EMBEDDED_TYPES) 2277 return (EINVAL); 2278 if (drrwe->drr_compression >= ZIO_COMPRESS_FUNCTIONS) 2279 return (EINVAL); 2280 2281 tx = dmu_tx_create(rwa->os); 2282 2283 dmu_tx_hold_write(tx, drrwe->drr_object, 2284 drrwe->drr_offset, drrwe->drr_length); 2285 err = dmu_tx_assign(tx, TXG_WAIT); 2286 if (err != 0) { 2287 dmu_tx_abort(tx); 2288 return (err); 2289 } 2290 2291 dmu_write_embedded(rwa->os, drrwe->drr_object, 2292 drrwe->drr_offset, data, drrwe->drr_etype, 2293 drrwe->drr_compression, drrwe->drr_lsize, drrwe->drr_psize, 2294 rwa->byteswap ^ ZFS_HOST_BYTEORDER, tx); 2295 2296 /* See comment in restore_write. */ 2297 save_resume_state(rwa, drrwe->drr_object, drrwe->drr_offset, tx); 2298 dmu_tx_commit(tx); 2299 return (0); 2300} 2301 2302static int 2303receive_spill(struct receive_writer_arg *rwa, struct drr_spill *drrs, 2304 void *data) 2305{ 2306 dmu_tx_t *tx; 2307 dmu_buf_t *db, *db_spill; 2308 int err; 2309 2310 if (drrs->drr_length < SPA_MINBLOCKSIZE || 2311 drrs->drr_length > spa_maxblocksize(dmu_objset_spa(rwa->os))) 2312 return (SET_ERROR(EINVAL)); 2313 2314 if (dmu_object_info(rwa->os, drrs->drr_object, NULL) != 0) 2315 return (SET_ERROR(EINVAL)); 2316 2317 VERIFY0(dmu_bonus_hold(rwa->os, drrs->drr_object, FTAG, &db)); 2318 if ((err = dmu_spill_hold_by_bonus(db, FTAG, &db_spill)) != 0) { 2319 dmu_buf_rele(db, FTAG); 2320 return (err); 2321 } 2322 2323 tx = dmu_tx_create(rwa->os); 2324 2325 dmu_tx_hold_spill(tx, db->db_object); 2326 2327 err = dmu_tx_assign(tx, TXG_WAIT); 2328 if (err != 0) { 2329 dmu_buf_rele(db, FTAG); 2330 dmu_buf_rele(db_spill, FTAG); 2331 dmu_tx_abort(tx); 2332 return (err); 2333 } 2334 dmu_buf_will_dirty(db_spill, tx); 2335 2336 if (db_spill->db_size < drrs->drr_length) 2337 VERIFY(0 == dbuf_spill_set_blksz(db_spill, 2338 drrs->drr_length, tx)); 2339 bcopy(data, db_spill->db_data, drrs->drr_length); 2340 2341 dmu_buf_rele(db, FTAG); 2342 dmu_buf_rele(db_spill, FTAG); 2343 2344 dmu_tx_commit(tx); 2345 return (0); 2346} 2347 2348/* ARGSUSED */ 2349static int 2350receive_free(struct receive_writer_arg *rwa, struct drr_free *drrf) 2351{ 2352 int err; 2353 2354 if (drrf->drr_length != -1ULL && 2355 drrf->drr_offset + drrf->drr_length < drrf->drr_offset) 2356 return (SET_ERROR(EINVAL)); 2357 2358 if (dmu_object_info(rwa->os, drrf->drr_object, NULL) != 0) 2359 return (SET_ERROR(EINVAL)); 2360 2361 err = dmu_free_long_range(rwa->os, drrf->drr_object, 2362 drrf->drr_offset, drrf->drr_length); 2363 2364 return (err); 2365} 2366 2367/* used to destroy the drc_ds on error */ 2368static void 2369dmu_recv_cleanup_ds(dmu_recv_cookie_t *drc) 2370{ 2371 if (drc->drc_resumable) { 2372 /* wait for our resume state to be written to disk */ 2373 txg_wait_synced(drc->drc_ds->ds_dir->dd_pool, 0); 2374 dsl_dataset_disown(drc->drc_ds, dmu_recv_tag); 2375 } else { 2376 char name[ZFS_MAX_DATASET_NAME_LEN]; 2377 dsl_dataset_name(drc->drc_ds, name); 2378 dsl_dataset_disown(drc->drc_ds, dmu_recv_tag); 2379 (void) dsl_destroy_head(name); 2380 } 2381} 2382 2383static void 2384receive_cksum(struct receive_arg *ra, int len, void *buf) 2385{ 2386 if (ra->byteswap) { 2387 fletcher_4_incremental_byteswap(buf, len, &ra->cksum); 2388 } else { 2389 fletcher_4_incremental_native(buf, len, &ra->cksum); 2390 } 2391} 2392 2393/* 2394 * Read the payload into a buffer of size len, and update the current record's 2395 * payload field. 2396 * Allocate ra->next_rrd and read the next record's header into 2397 * ra->next_rrd->header. 2398 * Verify checksum of payload and next record. 2399 */ 2400static int 2401receive_read_payload_and_next_header(struct receive_arg *ra, int len, void *buf) 2402{ 2403 int err; 2404 2405 if (len != 0) { 2406 ASSERT3U(len, <=, SPA_MAXBLOCKSIZE); 2407 err = receive_read(ra, len, buf); 2408 if (err != 0) 2409 return (err); 2410 receive_cksum(ra, len, buf); 2411 2412 /* note: rrd is NULL when reading the begin record's payload */ 2413 if (ra->rrd != NULL) { 2414 ra->rrd->payload = buf; 2415 ra->rrd->payload_size = len; 2416 ra->rrd->bytes_read = ra->bytes_read; 2417 } 2418 } 2419 2420 ra->prev_cksum = ra->cksum; 2421 2422 ra->next_rrd = kmem_zalloc(sizeof (*ra->next_rrd), KM_SLEEP); 2423 err = receive_read(ra, sizeof (ra->next_rrd->header), 2424 &ra->next_rrd->header); 2425 ra->next_rrd->bytes_read = ra->bytes_read; 2426 if (err != 0) { 2427 kmem_free(ra->next_rrd, sizeof (*ra->next_rrd)); 2428 ra->next_rrd = NULL; 2429 return (err); 2430 } 2431 if (ra->next_rrd->header.drr_type == DRR_BEGIN) { 2432 kmem_free(ra->next_rrd, sizeof (*ra->next_rrd)); 2433 ra->next_rrd = NULL; 2434 return (SET_ERROR(EINVAL)); 2435 } 2436 2437 /* 2438 * Note: checksum is of everything up to but not including the 2439 * checksum itself. 2440 */ 2441 ASSERT3U(offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum), 2442 ==, sizeof (dmu_replay_record_t) - sizeof (zio_cksum_t)); 2443 receive_cksum(ra, 2444 offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum), 2445 &ra->next_rrd->header); 2446 2447 zio_cksum_t cksum_orig = 2448 ra->next_rrd->header.drr_u.drr_checksum.drr_checksum; 2449 zio_cksum_t *cksump = 2450 &ra->next_rrd->header.drr_u.drr_checksum.drr_checksum; 2451 2452 if (ra->byteswap) 2453 byteswap_record(&ra->next_rrd->header); 2454 2455 if ((!ZIO_CHECKSUM_IS_ZERO(cksump)) && 2456 !ZIO_CHECKSUM_EQUAL(ra->cksum, *cksump)) { 2457 kmem_free(ra->next_rrd, sizeof (*ra->next_rrd)); 2458 ra->next_rrd = NULL; 2459 return (SET_ERROR(ECKSUM)); 2460 } 2461 2462 receive_cksum(ra, sizeof (cksum_orig), &cksum_orig); 2463 2464 return (0); 2465} 2466 2467static void 2468objlist_create(struct objlist *list) 2469{ 2470 list_create(&list->list, sizeof (struct receive_objnode), 2471 offsetof(struct receive_objnode, node)); 2472 list->last_lookup = 0; 2473} 2474 2475static void 2476objlist_destroy(struct objlist *list) 2477{ 2478 for (struct receive_objnode *n = list_remove_head(&list->list); 2479 n != NULL; n = list_remove_head(&list->list)) { 2480 kmem_free(n, sizeof (*n)); 2481 } 2482 list_destroy(&list->list); 2483} 2484 2485/* 2486 * This function looks through the objlist to see if the specified object number 2487 * is contained in the objlist. In the process, it will remove all object 2488 * numbers in the list that are smaller than the specified object number. Thus, 2489 * any lookup of an object number smaller than a previously looked up object 2490 * number will always return false; therefore, all lookups should be done in 2491 * ascending order. 2492 */ 2493static boolean_t 2494objlist_exists(struct objlist *list, uint64_t object) 2495{ 2496 struct receive_objnode *node = list_head(&list->list); 2497 ASSERT3U(object, >=, list->last_lookup); 2498 list->last_lookup = object; 2499 while (node != NULL && node->object < object) { 2500 VERIFY3P(node, ==, list_remove_head(&list->list)); 2501 kmem_free(node, sizeof (*node)); 2502 node = list_head(&list->list); 2503 } 2504 return (node != NULL && node->object == object); 2505} 2506 2507/* 2508 * The objlist is a list of object numbers stored in ascending order. However, 2509 * the insertion of new object numbers does not seek out the correct location to 2510 * store a new object number; instead, it appends it to the list for simplicity. 2511 * Thus, any users must take care to only insert new object numbers in ascending 2512 * order. 2513 */ 2514static void 2515objlist_insert(struct objlist *list, uint64_t object) 2516{ 2517 struct receive_objnode *node = kmem_zalloc(sizeof (*node), KM_SLEEP); 2518 node->object = object; 2519#ifdef ZFS_DEBUG 2520 struct receive_objnode *last_object = list_tail(&list->list); 2521 uint64_t last_objnum = (last_object != NULL ? last_object->object : 0); 2522 ASSERT3U(node->object, >, last_objnum); 2523#endif 2524 list_insert_tail(&list->list, node); 2525} 2526 2527/* 2528 * Issue the prefetch reads for any necessary indirect blocks. 2529 * 2530 * We use the object ignore list to tell us whether or not to issue prefetches 2531 * for a given object. We do this for both correctness (in case the blocksize 2532 * of an object has changed) and performance (if the object doesn't exist, don't 2533 * needlessly try to issue prefetches). We also trim the list as we go through 2534 * the stream to prevent it from growing to an unbounded size. 2535 * 2536 * The object numbers within will always be in sorted order, and any write 2537 * records we see will also be in sorted order, but they're not sorted with 2538 * respect to each other (i.e. we can get several object records before 2539 * receiving each object's write records). As a result, once we've reached a 2540 * given object number, we can safely remove any reference to lower object 2541 * numbers in the ignore list. In practice, we receive up to 32 object records 2542 * before receiving write records, so the list can have up to 32 nodes in it. 2543 */ 2544/* ARGSUSED */ 2545static void 2546receive_read_prefetch(struct receive_arg *ra, 2547 uint64_t object, uint64_t offset, uint64_t length) 2548{ 2549 if (!objlist_exists(&ra->ignore_objlist, object)) { 2550 dmu_prefetch(ra->os, object, 1, offset, length, 2551 ZIO_PRIORITY_SYNC_READ); 2552 } 2553} 2554 2555/* 2556 * Read records off the stream, issuing any necessary prefetches. 2557 */ 2558static int 2559receive_read_record(struct receive_arg *ra) 2560{ 2561 int err; 2562 2563 switch (ra->rrd->header.drr_type) { 2564 case DRR_OBJECT: 2565 { 2566 struct drr_object *drro = &ra->rrd->header.drr_u.drr_object; 2567 uint32_t size = P2ROUNDUP(drro->drr_bonuslen, 8); 2568 void *buf = kmem_zalloc(size, KM_SLEEP); 2569 dmu_object_info_t doi; 2570 err = receive_read_payload_and_next_header(ra, size, buf); 2571 if (err != 0) { 2572 kmem_free(buf, size); 2573 return (err); 2574 } 2575 err = dmu_object_info(ra->os, drro->drr_object, &doi); 2576 /* 2577 * See receive_read_prefetch for an explanation why we're 2578 * storing this object in the ignore_obj_list. 2579 */ 2580 if (err == ENOENT || 2581 (err == 0 && doi.doi_data_block_size != drro->drr_blksz)) { 2582 objlist_insert(&ra->ignore_objlist, drro->drr_object); 2583 err = 0; 2584 } 2585 return (err); 2586 } 2587 case DRR_FREEOBJECTS: 2588 { 2589 err = receive_read_payload_and_next_header(ra, 0, NULL); 2590 return (err); 2591 } 2592 case DRR_WRITE: 2593 { 2594 struct drr_write *drrw = &ra->rrd->header.drr_u.drr_write; 2595 arc_buf_t *abuf = arc_loan_buf(dmu_objset_spa(ra->os), 2596 drrw->drr_length); 2597 2598 err = receive_read_payload_and_next_header(ra, 2599 drrw->drr_length, abuf->b_data); 2600 if (err != 0) { 2601 dmu_return_arcbuf(abuf); 2602 return (err); 2603 } 2604 ra->rrd->write_buf = abuf; 2605 receive_read_prefetch(ra, drrw->drr_object, drrw->drr_offset, 2606 drrw->drr_length); 2607 return (err); 2608 } 2609 case DRR_WRITE_BYREF: 2610 { 2611 struct drr_write_byref *drrwb = 2612 &ra->rrd->header.drr_u.drr_write_byref; 2613 err = receive_read_payload_and_next_header(ra, 0, NULL); 2614 receive_read_prefetch(ra, drrwb->drr_object, drrwb->drr_offset, 2615 drrwb->drr_length); 2616 return (err); 2617 } 2618 case DRR_WRITE_EMBEDDED: 2619 { 2620 struct drr_write_embedded *drrwe = 2621 &ra->rrd->header.drr_u.drr_write_embedded; 2622 uint32_t size = P2ROUNDUP(drrwe->drr_psize, 8); 2623 void *buf = kmem_zalloc(size, KM_SLEEP); 2624 2625 err = receive_read_payload_and_next_header(ra, size, buf); 2626 if (err != 0) { 2627 kmem_free(buf, size); 2628 return (err); 2629 } 2630 2631 receive_read_prefetch(ra, drrwe->drr_object, drrwe->drr_offset, 2632 drrwe->drr_length); 2633 return (err); 2634 } 2635 case DRR_FREE: 2636 { 2637 /* 2638 * It might be beneficial to prefetch indirect blocks here, but 2639 * we don't really have the data to decide for sure. 2640 */ 2641 err = receive_read_payload_and_next_header(ra, 0, NULL); 2642 return (err); 2643 } 2644 case DRR_END: 2645 { 2646 struct drr_end *drre = &ra->rrd->header.drr_u.drr_end; 2647 if (!ZIO_CHECKSUM_EQUAL(ra->prev_cksum, drre->drr_checksum)) 2648 return (SET_ERROR(ECKSUM)); 2649 return (0); 2650 } 2651 case DRR_SPILL: 2652 { 2653 struct drr_spill *drrs = &ra->rrd->header.drr_u.drr_spill; 2654 void *buf = kmem_zalloc(drrs->drr_length, KM_SLEEP); 2655 err = receive_read_payload_and_next_header(ra, drrs->drr_length, 2656 buf); 2657 if (err != 0) 2658 kmem_free(buf, drrs->drr_length); 2659 return (err); 2660 } 2661 default: 2662 return (SET_ERROR(EINVAL)); 2663 } 2664} 2665 2666/* 2667 * Commit the records to the pool. 2668 */ 2669static int 2670receive_process_record(struct receive_writer_arg *rwa, 2671 struct receive_record_arg *rrd) 2672{ 2673 int err; 2674 2675 /* Processing in order, therefore bytes_read should be increasing. */ 2676 ASSERT3U(rrd->bytes_read, >=, rwa->bytes_read); 2677 rwa->bytes_read = rrd->bytes_read; 2678 2679 switch (rrd->header.drr_type) { 2680 case DRR_OBJECT: 2681 { 2682 struct drr_object *drro = &rrd->header.drr_u.drr_object; 2683 err = receive_object(rwa, drro, rrd->payload); 2684 kmem_free(rrd->payload, rrd->payload_size); 2685 rrd->payload = NULL; 2686 return (err); 2687 } 2688 case DRR_FREEOBJECTS: 2689 { 2690 struct drr_freeobjects *drrfo = 2691 &rrd->header.drr_u.drr_freeobjects; 2692 return (receive_freeobjects(rwa, drrfo)); 2693 } 2694 case DRR_WRITE: 2695 { 2696 struct drr_write *drrw = &rrd->header.drr_u.drr_write; 2697 err = receive_write(rwa, drrw, rrd->write_buf); 2698 /* if receive_write() is successful, it consumes the arc_buf */ 2699 if (err != 0) 2700 dmu_return_arcbuf(rrd->write_buf); 2701 rrd->write_buf = NULL; 2702 rrd->payload = NULL; 2703 return (err); 2704 } 2705 case DRR_WRITE_BYREF: 2706 { 2707 struct drr_write_byref *drrwbr = 2708 &rrd->header.drr_u.drr_write_byref; 2709 return (receive_write_byref(rwa, drrwbr)); 2710 } 2711 case DRR_WRITE_EMBEDDED: 2712 { 2713 struct drr_write_embedded *drrwe = 2714 &rrd->header.drr_u.drr_write_embedded; 2715 err = receive_write_embedded(rwa, drrwe, rrd->payload); 2716 kmem_free(rrd->payload, rrd->payload_size); 2717 rrd->payload = NULL; 2718 return (err); 2719 } 2720 case DRR_FREE: 2721 { 2722 struct drr_free *drrf = &rrd->header.drr_u.drr_free; 2723 return (receive_free(rwa, drrf)); 2724 } 2725 case DRR_SPILL: 2726 { 2727 struct drr_spill *drrs = &rrd->header.drr_u.drr_spill; 2728 err = receive_spill(rwa, drrs, rrd->payload); 2729 kmem_free(rrd->payload, rrd->payload_size); 2730 rrd->payload = NULL; 2731 return (err); 2732 } 2733 default: 2734 return (SET_ERROR(EINVAL)); 2735 } 2736} 2737 2738/* 2739 * dmu_recv_stream's worker thread; pull records off the queue, and then call 2740 * receive_process_record When we're done, signal the main thread and exit. 2741 */ 2742static void 2743receive_writer_thread(void *arg) 2744{ 2745 struct receive_writer_arg *rwa = arg; 2746 struct receive_record_arg *rrd; 2747 for (rrd = bqueue_dequeue(&rwa->q); !rrd->eos_marker; 2748 rrd = bqueue_dequeue(&rwa->q)) { 2749 /* 2750 * If there's an error, the main thread will stop putting things 2751 * on the queue, but we need to clear everything in it before we 2752 * can exit. 2753 */ 2754 if (rwa->err == 0) { 2755 rwa->err = receive_process_record(rwa, rrd); 2756 } else if (rrd->write_buf != NULL) { 2757 dmu_return_arcbuf(rrd->write_buf); 2758 rrd->write_buf = NULL; 2759 rrd->payload = NULL; 2760 } else if (rrd->payload != NULL) { 2761 kmem_free(rrd->payload, rrd->payload_size); 2762 rrd->payload = NULL; 2763 } 2764 kmem_free(rrd, sizeof (*rrd)); 2765 } 2766 kmem_free(rrd, sizeof (*rrd)); 2767 mutex_enter(&rwa->mutex); 2768 rwa->done = B_TRUE; 2769 cv_signal(&rwa->cv); 2770 mutex_exit(&rwa->mutex); 2771 thread_exit(); 2772} 2773 2774static int 2775resume_check(struct receive_arg *ra, nvlist_t *begin_nvl) 2776{ 2777 uint64_t val; 2778 objset_t *mos = dmu_objset_pool(ra->os)->dp_meta_objset; 2779 uint64_t dsobj = dmu_objset_id(ra->os); 2780 uint64_t resume_obj, resume_off; 2781 2782 if (nvlist_lookup_uint64(begin_nvl, 2783 "resume_object", &resume_obj) != 0 || 2784 nvlist_lookup_uint64(begin_nvl, 2785 "resume_offset", &resume_off) != 0) { 2786 return (SET_ERROR(EINVAL)); 2787 } 2788 VERIFY0(zap_lookup(mos, dsobj, 2789 DS_FIELD_RESUME_OBJECT, sizeof (val), 1, &val)); 2790 if (resume_obj != val) 2791 return (SET_ERROR(EINVAL)); 2792 VERIFY0(zap_lookup(mos, dsobj, 2793 DS_FIELD_RESUME_OFFSET, sizeof (val), 1, &val)); 2794 if (resume_off != val) 2795 return (SET_ERROR(EINVAL)); 2796 2797 return (0); 2798} 2799 2800/* 2801 * Read in the stream's records, one by one, and apply them to the pool. There 2802 * are two threads involved; the thread that calls this function will spin up a 2803 * worker thread, read the records off the stream one by one, and issue 2804 * prefetches for any necessary indirect blocks. It will then push the records 2805 * onto an internal blocking queue. The worker thread will pull the records off 2806 * the queue, and actually write the data into the DMU. This way, the worker 2807 * thread doesn't have to wait for reads to complete, since everything it needs 2808 * (the indirect blocks) will be prefetched. 2809 * 2810 * NB: callers *must* call dmu_recv_end() if this succeeds. 2811 */ 2812int 2813dmu_recv_stream(dmu_recv_cookie_t *drc, struct file *fp, offset_t *voffp, 2814 int cleanup_fd, uint64_t *action_handlep) 2815{ 2816 int err = 0; 2817 struct receive_arg ra = { 0 }; 2818 struct receive_writer_arg rwa = { 0 }; 2819 int featureflags; 2820 nvlist_t *begin_nvl = NULL; 2821 2822 ra.byteswap = drc->drc_byteswap; 2823 ra.cksum = drc->drc_cksum; 2824 ra.td = curthread; 2825 ra.fp = fp; 2826 ra.voff = *voffp; 2827 2828 if (dsl_dataset_is_zapified(drc->drc_ds)) { 2829 (void) zap_lookup(drc->drc_ds->ds_dir->dd_pool->dp_meta_objset, 2830 drc->drc_ds->ds_object, DS_FIELD_RESUME_BYTES, 2831 sizeof (ra.bytes_read), 1, &ra.bytes_read); 2832 } 2833 2834 objlist_create(&ra.ignore_objlist); 2835 2836 /* these were verified in dmu_recv_begin */ 2837 ASSERT3U(DMU_GET_STREAM_HDRTYPE(drc->drc_drrb->drr_versioninfo), ==, 2838 DMU_SUBSTREAM); 2839 ASSERT3U(drc->drc_drrb->drr_type, <, DMU_OST_NUMTYPES); 2840 2841 /* 2842 * Open the objset we are modifying. 2843 */ 2844 VERIFY0(dmu_objset_from_ds(drc->drc_ds, &ra.os)); 2845 2846 ASSERT(dsl_dataset_phys(drc->drc_ds)->ds_flags & DS_FLAG_INCONSISTENT); 2847 2848 featureflags = DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo); 2849 2850 /* if this stream is dedup'ed, set up the avl tree for guid mapping */ 2851 if (featureflags & DMU_BACKUP_FEATURE_DEDUP) { 2852 minor_t minor; 2853 2854 if (cleanup_fd == -1) { 2855 ra.err = SET_ERROR(EBADF); 2856 goto out; 2857 } 2858 ra.err = zfs_onexit_fd_hold(cleanup_fd, &minor); 2859 if (ra.err != 0) { 2860 cleanup_fd = -1; 2861 goto out; 2862 } 2863 2864 if (*action_handlep == 0) { 2865 rwa.guid_to_ds_map = 2866 kmem_alloc(sizeof (avl_tree_t), KM_SLEEP); 2867 avl_create(rwa.guid_to_ds_map, guid_compare, 2868 sizeof (guid_map_entry_t), 2869 offsetof(guid_map_entry_t, avlnode)); 2870 err = zfs_onexit_add_cb(minor, 2871 free_guid_map_onexit, rwa.guid_to_ds_map, 2872 action_handlep); 2873 if (ra.err != 0) 2874 goto out; 2875 } else { 2876 err = zfs_onexit_cb_data(minor, *action_handlep, 2877 (void **)&rwa.guid_to_ds_map); 2878 if (ra.err != 0) 2879 goto out; 2880 } 2881 2882 drc->drc_guid_to_ds_map = rwa.guid_to_ds_map; 2883 } 2884 2885 uint32_t payloadlen = drc->drc_drr_begin->drr_payloadlen; 2886 void *payload = NULL; 2887 if (payloadlen != 0) 2888 payload = kmem_alloc(payloadlen, KM_SLEEP); 2889 2890 err = receive_read_payload_and_next_header(&ra, payloadlen, payload); 2891 if (err != 0) { 2892 if (payloadlen != 0) 2893 kmem_free(payload, payloadlen); 2894 goto out; 2895 } 2896 if (payloadlen != 0) { 2897 err = nvlist_unpack(payload, payloadlen, &begin_nvl, KM_SLEEP); 2898 kmem_free(payload, payloadlen); 2899 if (err != 0) 2900 goto out; 2901 } 2902 2903 if (featureflags & DMU_BACKUP_FEATURE_RESUMING) { 2904 err = resume_check(&ra, begin_nvl); 2905 if (err != 0) 2906 goto out; 2907 } 2908 2909 (void) bqueue_init(&rwa.q, zfs_recv_queue_length, 2910 offsetof(struct receive_record_arg, node)); 2911 cv_init(&rwa.cv, NULL, CV_DEFAULT, NULL); 2912 mutex_init(&rwa.mutex, NULL, MUTEX_DEFAULT, NULL); 2913 rwa.os = ra.os; 2914 rwa.byteswap = drc->drc_byteswap; 2915 rwa.resumable = drc->drc_resumable; 2916 2917 (void) thread_create(NULL, 0, receive_writer_thread, &rwa, 0, &p0, 2918 TS_RUN, minclsyspri); 2919 /* 2920 * We're reading rwa.err without locks, which is safe since we are the 2921 * only reader, and the worker thread is the only writer. It's ok if we 2922 * miss a write for an iteration or two of the loop, since the writer 2923 * thread will keep freeing records we send it until we send it an eos 2924 * marker. 2925 * 2926 * We can leave this loop in 3 ways: First, if rwa.err is 2927 * non-zero. In that case, the writer thread will free the rrd we just 2928 * pushed. Second, if we're interrupted; in that case, either it's the 2929 * first loop and ra.rrd was never allocated, or it's later, and ra.rrd 2930 * has been handed off to the writer thread who will free it. Finally, 2931 * if receive_read_record fails or we're at the end of the stream, then 2932 * we free ra.rrd and exit. 2933 */ 2934 while (rwa.err == 0) { 2935 if (issig(JUSTLOOKING) && issig(FORREAL)) { 2936 err = SET_ERROR(EINTR); 2937 break; 2938 } 2939 2940 ASSERT3P(ra.rrd, ==, NULL); 2941 ra.rrd = ra.next_rrd; 2942 ra.next_rrd = NULL; 2943 /* Allocates and loads header into ra.next_rrd */ 2944 err = receive_read_record(&ra); 2945 2946 if (ra.rrd->header.drr_type == DRR_END || err != 0) { 2947 kmem_free(ra.rrd, sizeof (*ra.rrd)); 2948 ra.rrd = NULL; 2949 break; 2950 } 2951 2952 bqueue_enqueue(&rwa.q, ra.rrd, 2953 sizeof (struct receive_record_arg) + ra.rrd->payload_size); 2954 ra.rrd = NULL; 2955 } 2956 if (ra.next_rrd == NULL) 2957 ra.next_rrd = kmem_zalloc(sizeof (*ra.next_rrd), KM_SLEEP); 2958 ra.next_rrd->eos_marker = B_TRUE; 2959 bqueue_enqueue(&rwa.q, ra.next_rrd, 1); 2960 2961 mutex_enter(&rwa.mutex); 2962 while (!rwa.done) { 2963 cv_wait(&rwa.cv, &rwa.mutex); 2964 } 2965 mutex_exit(&rwa.mutex); 2966 2967 cv_destroy(&rwa.cv); 2968 mutex_destroy(&rwa.mutex); 2969 bqueue_destroy(&rwa.q); 2970 if (err == 0) 2971 err = rwa.err; 2972 2973out: 2974 nvlist_free(begin_nvl); 2975 if ((featureflags & DMU_BACKUP_FEATURE_DEDUP) && (cleanup_fd != -1)) 2976 zfs_onexit_fd_rele(cleanup_fd); 2977 2978 if (err != 0) { 2979 /* 2980 * Clean up references. If receive is not resumable, 2981 * destroy what we created, so we don't leave it in 2982 * the inconsistent state. 2983 */ 2984 dmu_recv_cleanup_ds(drc); 2985 } 2986 2987 *voffp = ra.voff; 2988 objlist_destroy(&ra.ignore_objlist); 2989 return (err); 2990} 2991 2992static int 2993dmu_recv_end_check(void *arg, dmu_tx_t *tx) 2994{ 2995 dmu_recv_cookie_t *drc = arg; 2996 dsl_pool_t *dp = dmu_tx_pool(tx); 2997 int error; 2998 2999 ASSERT3P(drc->drc_ds->ds_owner, ==, dmu_recv_tag); 3000 3001 if (!drc->drc_newfs) { 3002 dsl_dataset_t *origin_head; 3003 3004 error = dsl_dataset_hold(dp, drc->drc_tofs, FTAG, &origin_head); 3005 if (error != 0) 3006 return (error); 3007 if (drc->drc_force) { 3008 /* 3009 * We will destroy any snapshots in tofs (i.e. before 3010 * origin_head) that are after the origin (which is 3011 * the snap before drc_ds, because drc_ds can not 3012 * have any snaps of its own). 3013 */ 3014 uint64_t obj; 3015 3016 obj = dsl_dataset_phys(origin_head)->ds_prev_snap_obj; 3017 while (obj != 3018 dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj) { 3019 dsl_dataset_t *snap; 3020 error = dsl_dataset_hold_obj(dp, obj, FTAG, 3021 &snap); 3022 if (error != 0) 3023 break; 3024 if (snap->ds_dir != origin_head->ds_dir) 3025 error = SET_ERROR(EINVAL); 3026 if (error == 0) { 3027 error = dsl_destroy_snapshot_check_impl( 3028 snap, B_FALSE); 3029 } 3030 obj = dsl_dataset_phys(snap)->ds_prev_snap_obj; 3031 dsl_dataset_rele(snap, FTAG); 3032 if (error != 0) 3033 break; 3034 } 3035 if (error != 0) { 3036 dsl_dataset_rele(origin_head, FTAG); 3037 return (error); 3038 } 3039 } 3040 error = dsl_dataset_clone_swap_check_impl(drc->drc_ds, 3041 origin_head, drc->drc_force, drc->drc_owner, tx); 3042 if (error != 0) { 3043 dsl_dataset_rele(origin_head, FTAG); 3044 return (error); 3045 } 3046 error = dsl_dataset_snapshot_check_impl(origin_head, 3047 drc->drc_tosnap, tx, B_TRUE, 1, drc->drc_cred); 3048 dsl_dataset_rele(origin_head, FTAG); 3049 if (error != 0) 3050 return (error); 3051 3052 error = dsl_destroy_head_check_impl(drc->drc_ds, 1); 3053 } else { 3054 error = dsl_dataset_snapshot_check_impl(drc->drc_ds, 3055 drc->drc_tosnap, tx, B_TRUE, 1, drc->drc_cred); 3056 } 3057 return (error); 3058} 3059 3060static void 3061dmu_recv_end_sync(void *arg, dmu_tx_t *tx) 3062{ 3063 dmu_recv_cookie_t *drc = arg; 3064 dsl_pool_t *dp = dmu_tx_pool(tx); 3065 3066 spa_history_log_internal_ds(drc->drc_ds, "finish receiving", 3067 tx, "snap=%s", drc->drc_tosnap); 3068 3069 if (!drc->drc_newfs) { 3070 dsl_dataset_t *origin_head; 3071 3072 VERIFY0(dsl_dataset_hold(dp, drc->drc_tofs, FTAG, 3073 &origin_head)); 3074 3075 if (drc->drc_force) { 3076 /* 3077 * Destroy any snapshots of drc_tofs (origin_head) 3078 * after the origin (the snap before drc_ds). 3079 */ 3080 uint64_t obj; 3081 3082 obj = dsl_dataset_phys(origin_head)->ds_prev_snap_obj; 3083 while (obj != 3084 dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj) { 3085 dsl_dataset_t *snap; 3086 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG, 3087 &snap)); 3088 ASSERT3P(snap->ds_dir, ==, origin_head->ds_dir); 3089 obj = dsl_dataset_phys(snap)->ds_prev_snap_obj; 3090 dsl_destroy_snapshot_sync_impl(snap, 3091 B_FALSE, tx); 3092 dsl_dataset_rele(snap, FTAG); 3093 } 3094 } 3095 VERIFY3P(drc->drc_ds->ds_prev, ==, 3096 origin_head->ds_prev); 3097 3098 dsl_dataset_clone_swap_sync_impl(drc->drc_ds, 3099 origin_head, tx); 3100 dsl_dataset_snapshot_sync_impl(origin_head, 3101 drc->drc_tosnap, tx); 3102 3103 /* set snapshot's creation time and guid */ 3104 dmu_buf_will_dirty(origin_head->ds_prev->ds_dbuf, tx); 3105 dsl_dataset_phys(origin_head->ds_prev)->ds_creation_time = 3106 drc->drc_drrb->drr_creation_time; 3107 dsl_dataset_phys(origin_head->ds_prev)->ds_guid = 3108 drc->drc_drrb->drr_toguid; 3109 dsl_dataset_phys(origin_head->ds_prev)->ds_flags &= 3110 ~DS_FLAG_INCONSISTENT; 3111 3112 dmu_buf_will_dirty(origin_head->ds_dbuf, tx); 3113 dsl_dataset_phys(origin_head)->ds_flags &= 3114 ~DS_FLAG_INCONSISTENT; 3115 3116 dsl_dataset_rele(origin_head, FTAG); 3117 dsl_destroy_head_sync_impl(drc->drc_ds, tx); 3118 3119 if (drc->drc_owner != NULL) 3120 VERIFY3P(origin_head->ds_owner, ==, drc->drc_owner); 3121 } else { 3122 dsl_dataset_t *ds = drc->drc_ds; 3123 3124 dsl_dataset_snapshot_sync_impl(ds, drc->drc_tosnap, tx); 3125 3126 /* set snapshot's creation time and guid */ 3127 dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx); 3128 dsl_dataset_phys(ds->ds_prev)->ds_creation_time = 3129 drc->drc_drrb->drr_creation_time; 3130 dsl_dataset_phys(ds->ds_prev)->ds_guid = 3131 drc->drc_drrb->drr_toguid; 3132 dsl_dataset_phys(ds->ds_prev)->ds_flags &= 3133 ~DS_FLAG_INCONSISTENT; 3134 3135 dmu_buf_will_dirty(ds->ds_dbuf, tx); 3136 dsl_dataset_phys(ds)->ds_flags &= ~DS_FLAG_INCONSISTENT; 3137 if (dsl_dataset_has_resume_receive_state(ds)) { 3138 (void) zap_remove(dp->dp_meta_objset, ds->ds_object, 3139 DS_FIELD_RESUME_FROMGUID, tx); 3140 (void) zap_remove(dp->dp_meta_objset, ds->ds_object, 3141 DS_FIELD_RESUME_OBJECT, tx); 3142 (void) zap_remove(dp->dp_meta_objset, ds->ds_object, 3143 DS_FIELD_RESUME_OFFSET, tx); 3144 (void) zap_remove(dp->dp_meta_objset, ds->ds_object, 3145 DS_FIELD_RESUME_BYTES, tx); 3146 (void) zap_remove(dp->dp_meta_objset, ds->ds_object, 3147 DS_FIELD_RESUME_TOGUID, tx); 3148 (void) zap_remove(dp->dp_meta_objset, ds->ds_object, 3149 DS_FIELD_RESUME_TONAME, tx); 3150 } 3151 } 3152 drc->drc_newsnapobj = dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj; 3153 /* 3154 * Release the hold from dmu_recv_begin. This must be done before 3155 * we return to open context, so that when we free the dataset's dnode, 3156 * we can evict its bonus buffer. 3157 */ 3158 dsl_dataset_disown(drc->drc_ds, dmu_recv_tag); 3159 drc->drc_ds = NULL; 3160} 3161 3162static int 3163add_ds_to_guidmap(const char *name, avl_tree_t *guid_map, uint64_t snapobj) 3164{ 3165 dsl_pool_t *dp; 3166 dsl_dataset_t *snapds; 3167 guid_map_entry_t *gmep; 3168 int err; 3169 3170 ASSERT(guid_map != NULL); 3171 3172 err = dsl_pool_hold(name, FTAG, &dp); 3173 if (err != 0) 3174 return (err); 3175 gmep = kmem_alloc(sizeof (*gmep), KM_SLEEP); 3176 err = dsl_dataset_hold_obj(dp, snapobj, gmep, &snapds); 3177 if (err == 0) { 3178 gmep->guid = dsl_dataset_phys(snapds)->ds_guid; 3179 gmep->gme_ds = snapds; 3180 avl_add(guid_map, gmep); 3181 dsl_dataset_long_hold(snapds, gmep); 3182 } else 3183 kmem_free(gmep, sizeof (*gmep)); 3184 3185 dsl_pool_rele(dp, FTAG); 3186 return (err); 3187} 3188 3189static int dmu_recv_end_modified_blocks = 3; 3190 3191static int 3192dmu_recv_existing_end(dmu_recv_cookie_t *drc) 3193{ 3194 int error; 3195 3196#ifdef _KERNEL 3197 /* 3198 * We will be destroying the ds; make sure its origin is unmounted if 3199 * necessary. 3200 */ 3201 char name[ZFS_MAX_DATASET_NAME_LEN]; 3202 dsl_dataset_name(drc->drc_ds, name); 3203 zfs_destroy_unmount_origin(name); 3204#endif 3205 3206 error = dsl_sync_task(drc->drc_tofs, 3207 dmu_recv_end_check, dmu_recv_end_sync, drc, 3208 dmu_recv_end_modified_blocks, ZFS_SPACE_CHECK_NORMAL); 3209 3210 if (error != 0) 3211 dmu_recv_cleanup_ds(drc); 3212 return (error); 3213} 3214 3215static int 3216dmu_recv_new_end(dmu_recv_cookie_t *drc) 3217{ 3218 int error; 3219 3220 error = dsl_sync_task(drc->drc_tofs, 3221 dmu_recv_end_check, dmu_recv_end_sync, drc, 3222 dmu_recv_end_modified_blocks, ZFS_SPACE_CHECK_NORMAL); 3223 3224 if (error != 0) { 3225 dmu_recv_cleanup_ds(drc); 3226 } else if (drc->drc_guid_to_ds_map != NULL) { 3227 (void) add_ds_to_guidmap(drc->drc_tofs, 3228 drc->drc_guid_to_ds_map, 3229 drc->drc_newsnapobj); 3230 } 3231 return (error); 3232} 3233 3234int 3235dmu_recv_end(dmu_recv_cookie_t *drc, void *owner) 3236{ 3237 drc->drc_owner = owner; 3238 3239 if (drc->drc_newfs) 3240 return (dmu_recv_new_end(drc)); 3241 else 3242 return (dmu_recv_existing_end(drc)); 3243} 3244 3245/* 3246 * Return TRUE if this objset is currently being received into. 3247 */ 3248boolean_t 3249dmu_objset_is_receiving(objset_t *os) 3250{ 3251 return (os->os_dsl_dataset != NULL && 3252 os->os_dsl_dataset->ds_owner == dmu_recv_tag); 3253} 3254