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 (c) 2012, 2018 by Delphix. All rights reserved. 24 */ 25 26 #include <sys/zfs_context.h> 27 #include <sys/dmu_objset.h> 28 #include <sys/dmu_traverse.h> 29 #include <sys/dsl_dataset.h> 30 #include <sys/dsl_dir.h> 31 #include <sys/dsl_pool.h> 32 #include <sys/dnode.h> 33 #include <sys/spa.h> 34 #include <sys/spa_impl.h> 35 #include <sys/zio.h> 36 #include <sys/dmu_impl.h> 37 #include <sys/sa.h> 38 #include <sys/sa_impl.h> 39 #include <sys/callb.h> 40 #include <sys/zfeature.h> 41 42 int32_t zfs_pd_bytes_max = 50 * 1024 * 1024; /* 50MB */ 43 int32_t send_holes_without_birth_time = 1; 44 45 typedef struct prefetch_data { 46 kmutex_t pd_mtx; 47 kcondvar_t pd_cv; 48 int32_t pd_bytes_fetched; 49 int pd_flags; 50 boolean_t pd_cancel; 51 boolean_t pd_exited; 52 zbookmark_phys_t pd_resume; 53 } prefetch_data_t; 54 55 typedef struct traverse_data { 56 spa_t *td_spa; 57 uint64_t td_objset; 58 blkptr_t *td_rootbp; 59 uint64_t td_min_txg; 60 zbookmark_phys_t *td_resume; 61 int td_flags; 62 prefetch_data_t *td_pfd; 63 boolean_t td_paused; 64 uint64_t td_hole_birth_enabled_txg; 65 blkptr_cb_t *td_func; 66 void *td_arg; 67 boolean_t td_realloc_possible; 68 } traverse_data_t; 69 70 static int traverse_dnode(traverse_data_t *td, const blkptr_t *bp, 71 const dnode_phys_t *dnp, uint64_t objset, uint64_t object); 72 static void prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *, 73 uint64_t objset, uint64_t object); 74 75 static int 76 traverse_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) 77 { 78 traverse_data_t *td = arg; 79 zbookmark_phys_t zb; 80 81 if (BP_IS_HOLE(bp)) 82 return (0); 83 84 if (claim_txg == 0 && bp->blk_birth >= spa_min_claim_txg(td->td_spa)) 85 return (-1); 86 87 SET_BOOKMARK(&zb, td->td_objset, ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, 88 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]); 89 90 (void) td->td_func(td->td_spa, zilog, bp, &zb, NULL, td->td_arg); 91 92 return (0); 93 } 94 95 static int 96 traverse_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg) 97 { 98 traverse_data_t *td = arg; 99 100 if (lrc->lrc_txtype == TX_WRITE) { 101 lr_write_t *lr = (lr_write_t *)lrc; 102 blkptr_t *bp = &lr->lr_blkptr; 103 zbookmark_phys_t zb; 104 105 if (BP_IS_HOLE(bp)) 106 return (0); 107 108 if (claim_txg == 0 || bp->blk_birth < claim_txg) 109 return (0); 110 111 SET_BOOKMARK(&zb, td->td_objset, lr->lr_foid, 112 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp)); 113 114 (void) td->td_func(td->td_spa, zilog, bp, &zb, NULL, 115 td->td_arg); 116 } 117 return (0); 118 } 119 120 static void 121 traverse_zil(traverse_data_t *td, zil_header_t *zh) 122 { 123 uint64_t claim_txg = zh->zh_claim_txg; 124 125 /* 126 * We only want to visit blocks that have been claimed but not yet 127 * replayed; plus blocks that are already stable in read-only mode. 128 */ 129 if (claim_txg == 0 && spa_writeable(td->td_spa)) 130 return; 131 132 zilog_t *zilog = zil_alloc(spa_get_dsl(td->td_spa)->dp_meta_objset, zh); 133 (void) zil_parse(zilog, traverse_zil_block, traverse_zil_record, td, 134 claim_txg, !(td->td_flags & TRAVERSE_NO_DECRYPT)); 135 zil_free(zilog); 136 } 137 138 typedef enum resume_skip { 139 RESUME_SKIP_ALL, 140 RESUME_SKIP_NONE, 141 RESUME_SKIP_CHILDREN 142 } resume_skip_t; 143 144 /* 145 * Returns RESUME_SKIP_ALL if td indicates that we are resuming a traversal and 146 * the block indicated by zb does not need to be visited at all. Returns 147 * RESUME_SKIP_CHILDREN if we are resuming a post traversal and we reach the 148 * resume point. This indicates that this block should be visited but not its 149 * children (since they must have been visited in a previous traversal). 150 * Otherwise returns RESUME_SKIP_NONE. 151 */ 152 static resume_skip_t 153 resume_skip_check(traverse_data_t *td, const dnode_phys_t *dnp, 154 const zbookmark_phys_t *zb) 155 { 156 if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume)) { 157 /* 158 * If we already visited this bp & everything below, 159 * don't bother doing it again. 160 */ 161 if (zbookmark_subtree_completed(dnp, zb, td->td_resume)) 162 return (RESUME_SKIP_ALL); 163 164 /* 165 * If we found the block we're trying to resume from, zero 166 * the bookmark out to indicate that we have resumed. 167 */ 168 if (bcmp(zb, td->td_resume, sizeof (*zb)) == 0) { 169 bzero(td->td_resume, sizeof (*zb)); 170 if (td->td_flags & TRAVERSE_POST) 171 return (RESUME_SKIP_CHILDREN); 172 } 173 } 174 return (RESUME_SKIP_NONE); 175 } 176 177 static void 178 traverse_prefetch_metadata(traverse_data_t *td, 179 const blkptr_t *bp, const zbookmark_phys_t *zb) 180 { 181 arc_flags_t flags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH; 182 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE; 183 184 if (!(td->td_flags & TRAVERSE_PREFETCH_METADATA)) 185 return; 186 /* 187 * If we are in the process of resuming, don't prefetch, because 188 * some children will not be needed (and in fact may have already 189 * been freed). 190 */ 191 if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume)) 192 return; 193 if (BP_IS_HOLE(bp) || bp->blk_birth <= td->td_min_txg) 194 return; 195 if (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE) 196 return; 197 ASSERT(!BP_IS_REDACTED(bp)); 198 199 if ((td->td_flags & TRAVERSE_NO_DECRYPT) && BP_IS_PROTECTED(bp)) 200 zio_flags |= ZIO_FLAG_RAW; 201 202 (void) arc_read(NULL, td->td_spa, bp, NULL, NULL, 203 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb); 204 } 205 206 static boolean_t 207 prefetch_needed(prefetch_data_t *pfd, const blkptr_t *bp) 208 { 209 ASSERT(pfd->pd_flags & TRAVERSE_PREFETCH_DATA); 210 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp) || 211 BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG || BP_IS_REDACTED(bp)) 212 return (B_FALSE); 213 return (B_TRUE); 214 } 215 216 static int 217 traverse_visitbp(traverse_data_t *td, const dnode_phys_t *dnp, 218 const blkptr_t *bp, const zbookmark_phys_t *zb) 219 { 220 int err = 0; 221 arc_buf_t *buf = NULL; 222 prefetch_data_t *pd = td->td_pfd; 223 224 switch (resume_skip_check(td, dnp, zb)) { 225 case RESUME_SKIP_ALL: 226 return (0); 227 case RESUME_SKIP_CHILDREN: 228 goto post; 229 case RESUME_SKIP_NONE: 230 break; 231 default: 232 ASSERT(0); 233 } 234 235 if (bp->blk_birth == 0) { 236 /* 237 * Since this block has a birth time of 0 it must be one of 238 * two things: a hole created before the 239 * SPA_FEATURE_HOLE_BIRTH feature was enabled, or a hole 240 * which has always been a hole in an object. 241 * 242 * If a file is written sparsely, then the unwritten parts of 243 * the file were "always holes" -- that is, they have been 244 * holes since this object was allocated. However, we (and 245 * our callers) can not necessarily tell when an object was 246 * allocated. Therefore, if it's possible that this object 247 * was freed and then its object number reused, we need to 248 * visit all the holes with birth==0. 249 * 250 * If it isn't possible that the object number was reused, 251 * then if SPA_FEATURE_HOLE_BIRTH was enabled before we wrote 252 * all the blocks we will visit as part of this traversal, 253 * then this hole must have always existed, so we can skip 254 * it. We visit blocks born after (exclusive) td_min_txg. 255 * 256 * Note that the meta-dnode cannot be reallocated. 257 */ 258 if (!send_holes_without_birth_time && 259 (!td->td_realloc_possible || 260 zb->zb_object == DMU_META_DNODE_OBJECT) && 261 td->td_hole_birth_enabled_txg <= td->td_min_txg) 262 return (0); 263 } else if (bp->blk_birth <= td->td_min_txg) { 264 return (0); 265 } 266 267 if (pd != NULL && !pd->pd_exited && prefetch_needed(pd, bp)) { 268 uint64_t size = BP_GET_LSIZE(bp); 269 mutex_enter(&pd->pd_mtx); 270 ASSERT(pd->pd_bytes_fetched >= 0); 271 while (pd->pd_bytes_fetched < size && !pd->pd_exited) 272 cv_wait_sig(&pd->pd_cv, &pd->pd_mtx); 273 pd->pd_bytes_fetched -= size; 274 cv_broadcast(&pd->pd_cv); 275 mutex_exit(&pd->pd_mtx); 276 } 277 278 if (BP_IS_HOLE(bp) || BP_IS_REDACTED(bp)) { 279 err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg); 280 if (err != 0) 281 goto post; 282 return (0); 283 } 284 285 if (td->td_flags & TRAVERSE_PRE) { 286 err = td->td_func(td->td_spa, NULL, bp, zb, dnp, 287 td->td_arg); 288 if (err == TRAVERSE_VISIT_NO_CHILDREN) 289 return (0); 290 if (err != 0) 291 goto post; 292 } 293 294 if (BP_GET_LEVEL(bp) > 0) { 295 uint32_t flags = ARC_FLAG_WAIT; 296 int32_t i; 297 int32_t epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT; 298 zbookmark_phys_t *czb; 299 300 ASSERT(!BP_IS_PROTECTED(bp)); 301 302 err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf, 303 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb); 304 if (err != 0) 305 goto post; 306 307 czb = kmem_alloc(sizeof (zbookmark_phys_t), KM_SLEEP); 308 309 for (i = 0; i < epb; i++) { 310 SET_BOOKMARK(czb, zb->zb_objset, zb->zb_object, 311 zb->zb_level - 1, 312 zb->zb_blkid * epb + i); 313 traverse_prefetch_metadata(td, 314 &((blkptr_t *)buf->b_data)[i], czb); 315 } 316 317 /* recursively visitbp() blocks below this */ 318 for (i = 0; i < epb; i++) { 319 SET_BOOKMARK(czb, zb->zb_objset, zb->zb_object, 320 zb->zb_level - 1, 321 zb->zb_blkid * epb + i); 322 err = traverse_visitbp(td, dnp, 323 &((blkptr_t *)buf->b_data)[i], czb); 324 if (err != 0) 325 break; 326 } 327 328 kmem_free(czb, sizeof (zbookmark_phys_t)); 329 330 } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) { 331 uint32_t flags = ARC_FLAG_WAIT; 332 uint32_t zio_flags = ZIO_FLAG_CANFAIL; 333 int32_t i; 334 int32_t epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT; 335 dnode_phys_t *child_dnp; 336 337 /* 338 * dnode blocks might have their bonus buffers encrypted, so 339 * we must be careful to honor TRAVERSE_NO_DECRYPT 340 */ 341 if ((td->td_flags & TRAVERSE_NO_DECRYPT) && BP_IS_PROTECTED(bp)) 342 zio_flags |= ZIO_FLAG_RAW; 343 344 err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf, 345 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb); 346 if (err != 0) 347 goto post; 348 349 child_dnp = buf->b_data; 350 351 for (i = 0; i < epb; i += child_dnp[i].dn_extra_slots + 1) { 352 prefetch_dnode_metadata(td, &child_dnp[i], 353 zb->zb_objset, zb->zb_blkid * epb + i); 354 } 355 356 /* recursively visitbp() blocks below this */ 357 for (i = 0; i < epb; i += child_dnp[i].dn_extra_slots + 1) { 358 err = traverse_dnode(td, bp, &child_dnp[i], 359 zb->zb_objset, zb->zb_blkid * epb + i); 360 if (err != 0) 361 break; 362 } 363 } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) { 364 uint32_t zio_flags = ZIO_FLAG_CANFAIL; 365 arc_flags_t flags = ARC_FLAG_WAIT; 366 objset_phys_t *osp; 367 368 if ((td->td_flags & TRAVERSE_NO_DECRYPT) && BP_IS_PROTECTED(bp)) 369 zio_flags |= ZIO_FLAG_RAW; 370 371 err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf, 372 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb); 373 if (err != 0) 374 goto post; 375 376 osp = buf->b_data; 377 prefetch_dnode_metadata(td, &osp->os_meta_dnode, zb->zb_objset, 378 DMU_META_DNODE_OBJECT); 379 /* 380 * See the block comment above for the goal of this variable. 381 * If the maxblkid of the meta-dnode is 0, then we know that 382 * we've never had more than DNODES_PER_BLOCK objects in the 383 * dataset, which means we can't have reused any object ids. 384 */ 385 if (osp->os_meta_dnode.dn_maxblkid == 0) 386 td->td_realloc_possible = B_FALSE; 387 388 if (OBJSET_BUF_HAS_USERUSED(buf)) { 389 if (OBJSET_BUF_HAS_PROJECTUSED(buf)) 390 prefetch_dnode_metadata(td, 391 &osp->os_projectused_dnode, 392 zb->zb_objset, DMU_PROJECTUSED_OBJECT); 393 prefetch_dnode_metadata(td, &osp->os_groupused_dnode, 394 zb->zb_objset, DMU_GROUPUSED_OBJECT); 395 prefetch_dnode_metadata(td, &osp->os_userused_dnode, 396 zb->zb_objset, DMU_USERUSED_OBJECT); 397 } 398 399 err = traverse_dnode(td, bp, &osp->os_meta_dnode, zb->zb_objset, 400 DMU_META_DNODE_OBJECT); 401 if (err == 0 && OBJSET_BUF_HAS_USERUSED(buf)) { 402 if (OBJSET_BUF_HAS_PROJECTUSED(buf)) 403 err = traverse_dnode(td, bp, 404 &osp->os_projectused_dnode, zb->zb_objset, 405 DMU_PROJECTUSED_OBJECT); 406 if (err == 0) 407 err = traverse_dnode(td, bp, 408 &osp->os_groupused_dnode, zb->zb_objset, 409 DMU_GROUPUSED_OBJECT); 410 if (err == 0) 411 err = traverse_dnode(td, bp, 412 &osp->os_userused_dnode, zb->zb_objset, 413 DMU_USERUSED_OBJECT); 414 } 415 } 416 417 if (buf) 418 arc_buf_destroy(buf, &buf); 419 420 post: 421 if (err == 0 && (td->td_flags & TRAVERSE_POST)) 422 err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg); 423 424 if ((td->td_flags & TRAVERSE_HARD) && (err == EIO || err == ECKSUM)) { 425 /* 426 * Ignore this disk error as requested by the HARD flag, 427 * and continue traversal. 428 */ 429 err = 0; 430 } 431 432 /* 433 * If we are stopping here, set td_resume. 434 */ 435 if (td->td_resume != NULL && err != 0 && !td->td_paused) { 436 td->td_resume->zb_objset = zb->zb_objset; 437 td->td_resume->zb_object = zb->zb_object; 438 td->td_resume->zb_level = 0; 439 /* 440 * If we have stopped on an indirect block (e.g. due to 441 * i/o error), we have not visited anything below it. 442 * Set the bookmark to the first level-0 block that we need 443 * to visit. This way, the resuming code does not need to 444 * deal with resuming from indirect blocks. 445 * 446 * Note, if zb_level <= 0, dnp may be NULL, so we don't want 447 * to dereference it. 448 */ 449 td->td_resume->zb_blkid = zb->zb_blkid; 450 if (zb->zb_level > 0) { 451 td->td_resume->zb_blkid <<= zb->zb_level * 452 (dnp->dn_indblkshift - SPA_BLKPTRSHIFT); 453 } 454 td->td_paused = B_TRUE; 455 } 456 457 return (err); 458 } 459 460 static void 461 prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *dnp, 462 uint64_t objset, uint64_t object) 463 { 464 int j; 465 zbookmark_phys_t czb; 466 467 for (j = 0; j < dnp->dn_nblkptr; j++) { 468 SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j); 469 traverse_prefetch_metadata(td, &dnp->dn_blkptr[j], &czb); 470 } 471 472 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) { 473 SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID); 474 traverse_prefetch_metadata(td, DN_SPILL_BLKPTR(dnp), &czb); 475 } 476 } 477 478 static int 479 traverse_dnode(traverse_data_t *td, const blkptr_t *bp, const dnode_phys_t *dnp, 480 uint64_t objset, uint64_t object) 481 { 482 int j, err = 0; 483 zbookmark_phys_t czb; 484 485 if (object != DMU_META_DNODE_OBJECT && td->td_resume != NULL && 486 object < td->td_resume->zb_object) 487 return (0); 488 489 if (td->td_flags & TRAVERSE_PRE) { 490 SET_BOOKMARK(&czb, objset, object, ZB_DNODE_LEVEL, 491 ZB_DNODE_BLKID); 492 err = td->td_func(td->td_spa, NULL, bp, &czb, dnp, 493 td->td_arg); 494 if (err == TRAVERSE_VISIT_NO_CHILDREN) 495 return (0); 496 if (err != 0) 497 return (err); 498 } 499 500 for (j = 0; j < dnp->dn_nblkptr; j++) { 501 SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j); 502 err = traverse_visitbp(td, dnp, &dnp->dn_blkptr[j], &czb); 503 if (err != 0) 504 break; 505 } 506 507 if (err == 0 && (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) { 508 SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID); 509 err = traverse_visitbp(td, dnp, DN_SPILL_BLKPTR(dnp), &czb); 510 } 511 512 if (err == 0 && (td->td_flags & TRAVERSE_POST)) { 513 SET_BOOKMARK(&czb, objset, object, ZB_DNODE_LEVEL, 514 ZB_DNODE_BLKID); 515 err = td->td_func(td->td_spa, NULL, bp, &czb, dnp, 516 td->td_arg); 517 if (err == TRAVERSE_VISIT_NO_CHILDREN) 518 return (0); 519 if (err != 0) 520 return (err); 521 } 522 return (err); 523 } 524 525 /* ARGSUSED */ 526 static int 527 traverse_prefetcher(spa_t *spa, zilog_t *zilog, const blkptr_t *bp, 528 const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg) 529 { 530 prefetch_data_t *pfd = arg; 531 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE; 532 arc_flags_t aflags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH | 533 ARC_FLAG_PRESCIENT_PREFETCH; 534 535 ASSERT(pfd->pd_bytes_fetched >= 0); 536 if (zb->zb_level == ZB_DNODE_LEVEL) 537 return (0); 538 if (pfd->pd_cancel) 539 return (SET_ERROR(EINTR)); 540 541 if (!prefetch_needed(pfd, bp)) 542 return (0); 543 544 mutex_enter(&pfd->pd_mtx); 545 while (!pfd->pd_cancel && pfd->pd_bytes_fetched >= zfs_pd_bytes_max) 546 cv_wait_sig(&pfd->pd_cv, &pfd->pd_mtx); 547 pfd->pd_bytes_fetched += BP_GET_LSIZE(bp); 548 cv_broadcast(&pfd->pd_cv); 549 mutex_exit(&pfd->pd_mtx); 550 551 if ((pfd->pd_flags & TRAVERSE_NO_DECRYPT) && BP_IS_PROTECTED(bp)) 552 zio_flags |= ZIO_FLAG_RAW; 553 554 (void) arc_read(NULL, spa, bp, NULL, NULL, ZIO_PRIORITY_ASYNC_READ, 555 zio_flags, &aflags, zb); 556 557 return (0); 558 } 559 560 static void 561 traverse_prefetch_thread(void *arg) 562 { 563 traverse_data_t *td_main = arg; 564 traverse_data_t td = *td_main; 565 zbookmark_phys_t czb; 566 fstrans_cookie_t cookie = spl_fstrans_mark(); 567 568 td.td_func = traverse_prefetcher; 569 td.td_arg = td_main->td_pfd; 570 td.td_pfd = NULL; 571 td.td_resume = &td_main->td_pfd->pd_resume; 572 573 SET_BOOKMARK(&czb, td.td_objset, 574 ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID); 575 (void) traverse_visitbp(&td, NULL, td.td_rootbp, &czb); 576 577 mutex_enter(&td_main->td_pfd->pd_mtx); 578 td_main->td_pfd->pd_exited = B_TRUE; 579 cv_broadcast(&td_main->td_pfd->pd_cv); 580 mutex_exit(&td_main->td_pfd->pd_mtx); 581 spl_fstrans_unmark(cookie); 582 } 583 584 /* 585 * NB: dataset must not be changing on-disk (eg, is a snapshot or we are 586 * in syncing context). 587 */ 588 static int 589 traverse_impl(spa_t *spa, dsl_dataset_t *ds, uint64_t objset, blkptr_t *rootbp, 590 uint64_t txg_start, zbookmark_phys_t *resume, int flags, 591 blkptr_cb_t func, void *arg) 592 { 593 traverse_data_t *td; 594 prefetch_data_t *pd; 595 zbookmark_phys_t *czb; 596 int err; 597 598 ASSERT(ds == NULL || objset == ds->ds_object); 599 ASSERT(!(flags & TRAVERSE_PRE) || !(flags & TRAVERSE_POST)); 600 601 td = kmem_alloc(sizeof (traverse_data_t), KM_SLEEP); 602 pd = kmem_zalloc(sizeof (prefetch_data_t), KM_SLEEP); 603 czb = kmem_alloc(sizeof (zbookmark_phys_t), KM_SLEEP); 604 605 td->td_spa = spa; 606 td->td_objset = objset; 607 td->td_rootbp = rootbp; 608 td->td_min_txg = txg_start; 609 td->td_resume = resume; 610 td->td_func = func; 611 td->td_arg = arg; 612 td->td_pfd = pd; 613 td->td_flags = flags; 614 td->td_paused = B_FALSE; 615 td->td_realloc_possible = (txg_start == 0 ? B_FALSE : B_TRUE); 616 617 if (spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) { 618 VERIFY(spa_feature_enabled_txg(spa, 619 SPA_FEATURE_HOLE_BIRTH, &td->td_hole_birth_enabled_txg)); 620 } else { 621 td->td_hole_birth_enabled_txg = UINT64_MAX; 622 } 623 624 pd->pd_flags = flags; 625 if (resume != NULL) 626 pd->pd_resume = *resume; 627 mutex_init(&pd->pd_mtx, NULL, MUTEX_DEFAULT, NULL); 628 cv_init(&pd->pd_cv, NULL, CV_DEFAULT, NULL); 629 630 SET_BOOKMARK(czb, td->td_objset, 631 ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID); 632 633 /* See comment on ZIL traversal in dsl_scan_visitds. */ 634 if (ds != NULL && !ds->ds_is_snapshot && !BP_IS_HOLE(rootbp)) { 635 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 636 uint32_t flags = ARC_FLAG_WAIT; 637 objset_phys_t *osp; 638 arc_buf_t *buf; 639 ASSERT(!BP_IS_REDACTED(rootbp)); 640 641 if ((td->td_flags & TRAVERSE_NO_DECRYPT) && 642 BP_IS_PROTECTED(rootbp)) 643 zio_flags |= ZIO_FLAG_RAW; 644 645 err = arc_read(NULL, td->td_spa, rootbp, arc_getbuf_func, 646 &buf, ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, czb); 647 if (err != 0) { 648 /* 649 * If both TRAVERSE_HARD and TRAVERSE_PRE are set, 650 * continue to visitbp so that td_func can be called 651 * in pre stage, and err will reset to zero. 652 */ 653 if (!(td->td_flags & TRAVERSE_HARD) || 654 !(td->td_flags & TRAVERSE_PRE)) 655 goto out; 656 } else { 657 osp = buf->b_data; 658 traverse_zil(td, &osp->os_zil_header); 659 arc_buf_destroy(buf, &buf); 660 } 661 } 662 663 if (!(flags & TRAVERSE_PREFETCH_DATA) || 664 taskq_dispatch(spa->spa_prefetch_taskq, traverse_prefetch_thread, 665 td, TQ_NOQUEUE) == TASKQID_INVALID) 666 pd->pd_exited = B_TRUE; 667 668 err = traverse_visitbp(td, NULL, rootbp, czb); 669 670 mutex_enter(&pd->pd_mtx); 671 pd->pd_cancel = B_TRUE; 672 cv_broadcast(&pd->pd_cv); 673 while (!pd->pd_exited) 674 cv_wait_sig(&pd->pd_cv, &pd->pd_mtx); 675 mutex_exit(&pd->pd_mtx); 676 out: 677 mutex_destroy(&pd->pd_mtx); 678 cv_destroy(&pd->pd_cv); 679 680 kmem_free(czb, sizeof (zbookmark_phys_t)); 681 kmem_free(pd, sizeof (struct prefetch_data)); 682 kmem_free(td, sizeof (struct traverse_data)); 683 684 return (err); 685 } 686 687 /* 688 * NB: dataset must not be changing on-disk (eg, is a snapshot or we are 689 * in syncing context). 690 */ 691 int 692 traverse_dataset_resume(dsl_dataset_t *ds, uint64_t txg_start, 693 zbookmark_phys_t *resume, 694 int flags, blkptr_cb_t func, void *arg) 695 { 696 return (traverse_impl(ds->ds_dir->dd_pool->dp_spa, ds, ds->ds_object, 697 &dsl_dataset_phys(ds)->ds_bp, txg_start, resume, flags, func, arg)); 698 } 699 700 int 701 traverse_dataset(dsl_dataset_t *ds, uint64_t txg_start, 702 int flags, blkptr_cb_t func, void *arg) 703 { 704 return (traverse_dataset_resume(ds, txg_start, NULL, flags, func, arg)); 705 } 706 707 int 708 traverse_dataset_destroyed(spa_t *spa, blkptr_t *blkptr, 709 uint64_t txg_start, zbookmark_phys_t *resume, int flags, 710 blkptr_cb_t func, void *arg) 711 { 712 return (traverse_impl(spa, NULL, ZB_DESTROYED_OBJSET, 713 blkptr, txg_start, resume, flags, func, arg)); 714 } 715 716 /* 717 * NB: pool must not be changing on-disk (eg, from zdb or sync context). 718 */ 719 int 720 traverse_pool(spa_t *spa, uint64_t txg_start, int flags, 721 blkptr_cb_t func, void *arg) 722 { 723 int err; 724 dsl_pool_t *dp = spa_get_dsl(spa); 725 objset_t *mos = dp->dp_meta_objset; 726 boolean_t hard = (flags & TRAVERSE_HARD); 727 728 /* visit the MOS */ 729 err = traverse_impl(spa, NULL, 0, spa_get_rootblkptr(spa), 730 txg_start, NULL, flags, func, arg); 731 if (err != 0) 732 return (err); 733 734 /* visit each dataset */ 735 for (uint64_t obj = 1; err == 0; 736 err = dmu_object_next(mos, &obj, B_FALSE, txg_start)) { 737 dmu_object_info_t doi; 738 739 err = dmu_object_info(mos, obj, &doi); 740 if (err != 0) { 741 if (hard) 742 continue; 743 break; 744 } 745 746 if (doi.doi_bonus_type == DMU_OT_DSL_DATASET) { 747 dsl_dataset_t *ds; 748 uint64_t txg = txg_start; 749 750 dsl_pool_config_enter(dp, FTAG); 751 err = dsl_dataset_hold_obj(dp, obj, FTAG, &ds); 752 dsl_pool_config_exit(dp, FTAG); 753 if (err != 0) { 754 if (hard) 755 continue; 756 break; 757 } 758 if (dsl_dataset_phys(ds)->ds_prev_snap_txg > txg) 759 txg = dsl_dataset_phys(ds)->ds_prev_snap_txg; 760 err = traverse_dataset(ds, txg, flags, func, arg); 761 dsl_dataset_rele(ds, FTAG); 762 if (err != 0) 763 break; 764 } 765 } 766 if (err == ESRCH) 767 err = 0; 768 return (err); 769 } 770 771 EXPORT_SYMBOL(traverse_dataset); 772 EXPORT_SYMBOL(traverse_pool); 773 774 /* BEGIN CSTYLED */ 775 ZFS_MODULE_PARAM(zfs, zfs_, pd_bytes_max, INT, ZMOD_RW, 776 "Max number of bytes to prefetch"); 777 778 #if defined(_KERNEL) 779 module_param_named(ignore_hole_birth, send_holes_without_birth_time, int, 0644); 780 MODULE_PARM_DESC(ignore_hole_birth, 781 "Alias for send_holes_without_birth_time"); 782 #endif 783 784 ZFS_MODULE_PARAM(zfs, , send_holes_without_birth_time, INT, ZMOD_RW, 785 "Ignore hole_birth txg for zfs send"); 786 /* END CSTYLED */ 787