155 p2 &= p2 - 1; 156 157#ifdef illumos 158#ifndef _KERNEL 159 /* 160 * If we are using watchpoints, put each buffer on its own page, 161 * to eliminate the performance overhead of trapping to the 162 * kernel when modifying a non-watched buffer that shares the 163 * page with a watched buffer. 164 */ 165 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE)) 166 continue; 167#endif 168#endif /* illumos */ 169 if (size <= 4 * SPA_MINBLOCKSIZE) { 170 align = SPA_MINBLOCKSIZE; 171 } else if (IS_P2ALIGNED(size, p2 >> 2)) { 172 align = MIN(p2 >> 2, PAGESIZE); 173 } 174 175 if (align != 0) { 176 char name[36]; 177 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size); 178 zio_buf_cache[c] = kmem_cache_create(name, size, 179 align, NULL, NULL, NULL, NULL, NULL, cflags); 180 181 /* 182 * Since zio_data bufs do not appear in crash dumps, we 183 * pass KMC_NOTOUCH so that no allocator metadata is 184 * stored with the buffers. 185 */ 186 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size); 187 zio_data_buf_cache[c] = kmem_cache_create(name, size, 188 align, NULL, NULL, NULL, NULL, NULL, 189 cflags | KMC_NOTOUCH | KMC_NODEBUG); 190 } 191 } 192 193 while (--c != 0) { 194 ASSERT(zio_buf_cache[c] != NULL); 195 if (zio_buf_cache[c - 1] == NULL) 196 zio_buf_cache[c - 1] = zio_buf_cache[c]; 197 198 ASSERT(zio_data_buf_cache[c] != NULL); 199 if (zio_data_buf_cache[c - 1] == NULL) 200 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c]; 201 } 202out: 203 204 zio_inject_init(); 205 206 zio_trim_ksp = kstat_create("zfs", 0, "zio_trim", "misc", 207 KSTAT_TYPE_NAMED, 208 sizeof(zio_trim_stats) / sizeof(kstat_named_t), 209 KSTAT_FLAG_VIRTUAL); 210 211 if (zio_trim_ksp != NULL) { 212 zio_trim_ksp->ks_data = &zio_trim_stats; 213 kstat_install(zio_trim_ksp); 214 } 215} 216 217void 218zio_fini(void) 219{ 220 size_t c; 221 kmem_cache_t *last_cache = NULL; 222 kmem_cache_t *last_data_cache = NULL; 223 224 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { 225 if (zio_buf_cache[c] != last_cache) { 226 last_cache = zio_buf_cache[c]; 227 kmem_cache_destroy(zio_buf_cache[c]); 228 } 229 zio_buf_cache[c] = NULL; 230 231 if (zio_data_buf_cache[c] != last_data_cache) { 232 last_data_cache = zio_data_buf_cache[c]; 233 kmem_cache_destroy(zio_data_buf_cache[c]); 234 } 235 zio_data_buf_cache[c] = NULL; 236 } 237 238 kmem_cache_destroy(zio_link_cache); 239 kmem_cache_destroy(zio_cache); 240 241 zio_inject_fini(); 242 243 if (zio_trim_ksp != NULL) { 244 kstat_delete(zio_trim_ksp); 245 zio_trim_ksp = NULL; 246 } 247} 248 249/* 250 * ========================================================================== 251 * Allocate and free I/O buffers 252 * ========================================================================== 253 */ 254 255/* 256 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a 257 * crashdump if the kernel panics, so use it judiciously. Obviously, it's 258 * useful to inspect ZFS metadata, but if possible, we should avoid keeping 259 * excess / transient data in-core during a crashdump. 260 */ 261void * 262zio_buf_alloc(size_t size) 263{ 264 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 265 int flags = zio_exclude_metadata ? KM_NODEBUG : 0; 266 267 ASSERT3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 268 269 if (zio_use_uma) 270 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE)); 271 else 272 return (kmem_alloc(size, KM_SLEEP|flags)); 273} 274 275/* 276 * Use zio_data_buf_alloc to allocate data. The data will not appear in a 277 * crashdump if the kernel panics. This exists so that we will limit the amount 278 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount 279 * of kernel heap dumped to disk when the kernel panics) 280 */ 281void * 282zio_data_buf_alloc(size_t size) 283{ 284 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 285 286 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 287 288 if (zio_use_uma) 289 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE)); 290 else 291 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG)); 292} 293 294void 295zio_buf_free(void *buf, size_t size) 296{ 297 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 298 299 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 300 301 if (zio_use_uma) 302 kmem_cache_free(zio_buf_cache[c], buf); 303 else 304 kmem_free(buf, size); 305} 306 307void 308zio_data_buf_free(void *buf, size_t size) 309{ 310 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 311 312 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 313 314 if (zio_use_uma) 315 kmem_cache_free(zio_data_buf_cache[c], buf); 316 else 317 kmem_free(buf, size); 318} 319 320/* 321 * ========================================================================== 322 * Push and pop I/O transform buffers 323 * ========================================================================== 324 */ 325static void 326zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize, 327 zio_transform_func_t *transform) 328{ 329 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP); 330 331 zt->zt_orig_data = zio->io_data; 332 zt->zt_orig_size = zio->io_size; 333 zt->zt_bufsize = bufsize; 334 zt->zt_transform = transform; 335 336 zt->zt_next = zio->io_transform_stack; 337 zio->io_transform_stack = zt; 338 339 zio->io_data = data; 340 zio->io_size = size; 341} 342 343static void 344zio_pop_transforms(zio_t *zio) 345{ 346 zio_transform_t *zt; 347 348 while ((zt = zio->io_transform_stack) != NULL) { 349 if (zt->zt_transform != NULL) 350 zt->zt_transform(zio, 351 zt->zt_orig_data, zt->zt_orig_size); 352 353 if (zt->zt_bufsize != 0) 354 zio_buf_free(zio->io_data, zt->zt_bufsize); 355 356 zio->io_data = zt->zt_orig_data; 357 zio->io_size = zt->zt_orig_size; 358 zio->io_transform_stack = zt->zt_next; 359 360 kmem_free(zt, sizeof (zio_transform_t)); 361 } 362} 363 364/* 365 * ========================================================================== 366 * I/O transform callbacks for subblocks and decompression 367 * ========================================================================== 368 */ 369static void 370zio_subblock(zio_t *zio, void *data, uint64_t size) 371{ 372 ASSERT(zio->io_size > size); 373 374 if (zio->io_type == ZIO_TYPE_READ) 375 bcopy(zio->io_data, data, size); 376} 377 378static void 379zio_decompress(zio_t *zio, void *data, uint64_t size) 380{ 381 if (zio->io_error == 0 && 382 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp), 383 zio->io_data, data, zio->io_size, size) != 0) 384 zio->io_error = SET_ERROR(EIO); 385} 386 387/* 388 * ========================================================================== 389 * I/O parent/child relationships and pipeline interlocks 390 * ========================================================================== 391 */ 392/* 393 * NOTE - Callers to zio_walk_parents() and zio_walk_children must 394 * continue calling these functions until they return NULL. 395 * Otherwise, the next caller will pick up the list walk in 396 * some indeterminate state. (Otherwise every caller would 397 * have to pass in a cookie to keep the state represented by 398 * io_walk_link, which gets annoying.) 399 */ 400zio_t * 401zio_walk_parents(zio_t *cio) 402{ 403 zio_link_t *zl = cio->io_walk_link; 404 list_t *pl = &cio->io_parent_list; 405 406 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl); 407 cio->io_walk_link = zl; 408 409 if (zl == NULL) 410 return (NULL); 411 412 ASSERT(zl->zl_child == cio); 413 return (zl->zl_parent); 414} 415 416zio_t * 417zio_walk_children(zio_t *pio) 418{ 419 zio_link_t *zl = pio->io_walk_link; 420 list_t *cl = &pio->io_child_list; 421 422 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl); 423 pio->io_walk_link = zl; 424 425 if (zl == NULL) 426 return (NULL); 427 428 ASSERT(zl->zl_parent == pio); 429 return (zl->zl_child); 430} 431 432zio_t * 433zio_unique_parent(zio_t *cio) 434{ 435 zio_t *pio = zio_walk_parents(cio); 436 437 VERIFY(zio_walk_parents(cio) == NULL); 438 return (pio); 439} 440 441void 442zio_add_child(zio_t *pio, zio_t *cio) 443{ 444 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP); 445 446 /* 447 * Logical I/Os can have logical, gang, or vdev children. 448 * Gang I/Os can have gang or vdev children. 449 * Vdev I/Os can only have vdev children. 450 * The following ASSERT captures all of these constraints. 451 */ 452 ASSERT(cio->io_child_type <= pio->io_child_type); 453 454 zl->zl_parent = pio; 455 zl->zl_child = cio; 456 457 mutex_enter(&cio->io_lock); 458 mutex_enter(&pio->io_lock); 459 460 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0); 461 462 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 463 pio->io_children[cio->io_child_type][w] += !cio->io_state[w]; 464 465 list_insert_head(&pio->io_child_list, zl); 466 list_insert_head(&cio->io_parent_list, zl); 467 468 pio->io_child_count++; 469 cio->io_parent_count++; 470 471 mutex_exit(&pio->io_lock); 472 mutex_exit(&cio->io_lock); 473} 474 475static void 476zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl) 477{ 478 ASSERT(zl->zl_parent == pio); 479 ASSERT(zl->zl_child == cio); 480 481 mutex_enter(&cio->io_lock); 482 mutex_enter(&pio->io_lock); 483 484 list_remove(&pio->io_child_list, zl); 485 list_remove(&cio->io_parent_list, zl); 486 487 pio->io_child_count--; 488 cio->io_parent_count--; 489 490 mutex_exit(&pio->io_lock); 491 mutex_exit(&cio->io_lock); 492 493 kmem_cache_free(zio_link_cache, zl); 494} 495 496static boolean_t 497zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait) 498{ 499 uint64_t *countp = &zio->io_children[child][wait]; 500 boolean_t waiting = B_FALSE; 501 502 mutex_enter(&zio->io_lock); 503 ASSERT(zio->io_stall == NULL); 504 if (*countp != 0) { 505 zio->io_stage >>= 1; 506 zio->io_stall = countp; 507 waiting = B_TRUE; 508 } 509 mutex_exit(&zio->io_lock); 510 511 return (waiting); 512} 513 514static void 515zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait) 516{ 517 uint64_t *countp = &pio->io_children[zio->io_child_type][wait]; 518 int *errorp = &pio->io_child_error[zio->io_child_type]; 519 520 mutex_enter(&pio->io_lock); 521 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE)) 522 *errorp = zio_worst_error(*errorp, zio->io_error); 523 pio->io_reexecute |= zio->io_reexecute; 524 ASSERT3U(*countp, >, 0); 525 526 (*countp)--; 527 528 if (*countp == 0 && pio->io_stall == countp) { 529 pio->io_stall = NULL; 530 mutex_exit(&pio->io_lock); 531 zio_execute(pio); 532 } else { 533 mutex_exit(&pio->io_lock); 534 } 535} 536 537static void 538zio_inherit_child_errors(zio_t *zio, enum zio_child c) 539{ 540 if (zio->io_child_error[c] != 0 && zio->io_error == 0) 541 zio->io_error = zio->io_child_error[c]; 542} 543 544/* 545 * ========================================================================== 546 * Create the various types of I/O (read, write, free, etc) 547 * ========================================================================== 548 */ 549static zio_t * 550zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 551 void *data, uint64_t size, zio_done_func_t *done, void *private, 552 zio_type_t type, zio_priority_t priority, enum zio_flag flags, 553 vdev_t *vd, uint64_t offset, const zbookmark_phys_t *zb, 554 enum zio_stage stage, enum zio_stage pipeline) 555{ 556 zio_t *zio; 557 558 ASSERT3U(type == ZIO_TYPE_FREE || size, <=, SPA_MAXBLOCKSIZE); 559 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0); 560 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0); 561 562 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER)); 563 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER)); 564 ASSERT(vd || stage == ZIO_STAGE_OPEN); 565 566 zio = kmem_cache_alloc(zio_cache, KM_SLEEP); 567 bzero(zio, sizeof (zio_t)); 568 569 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL); 570 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL); 571 572 list_create(&zio->io_parent_list, sizeof (zio_link_t), 573 offsetof(zio_link_t, zl_parent_node)); 574 list_create(&zio->io_child_list, sizeof (zio_link_t), 575 offsetof(zio_link_t, zl_child_node)); 576 577 if (vd != NULL) 578 zio->io_child_type = ZIO_CHILD_VDEV; 579 else if (flags & ZIO_FLAG_GANG_CHILD) 580 zio->io_child_type = ZIO_CHILD_GANG; 581 else if (flags & ZIO_FLAG_DDT_CHILD) 582 zio->io_child_type = ZIO_CHILD_DDT; 583 else 584 zio->io_child_type = ZIO_CHILD_LOGICAL; 585 586 if (bp != NULL) { 587 zio->io_bp = (blkptr_t *)bp; 588 zio->io_bp_copy = *bp; 589 zio->io_bp_orig = *bp; 590 if (type != ZIO_TYPE_WRITE || 591 zio->io_child_type == ZIO_CHILD_DDT) 592 zio->io_bp = &zio->io_bp_copy; /* so caller can free */ 593 if (zio->io_child_type == ZIO_CHILD_LOGICAL) 594 zio->io_logical = zio; 595 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp)) 596 pipeline |= ZIO_GANG_STAGES; 597 } 598 599 zio->io_spa = spa; 600 zio->io_txg = txg; 601 zio->io_done = done; 602 zio->io_private = private; 603 zio->io_type = type; 604 zio->io_priority = priority; 605 zio->io_vd = vd; 606 zio->io_offset = offset; 607 zio->io_orig_data = zio->io_data = data; 608 zio->io_orig_size = zio->io_size = size; 609 zio->io_orig_flags = zio->io_flags = flags; 610 zio->io_orig_stage = zio->io_stage = stage; 611 zio->io_orig_pipeline = zio->io_pipeline = pipeline; 612 613 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY); 614 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE); 615 616 if (zb != NULL) 617 zio->io_bookmark = *zb; 618 619 if (pio != NULL) { 620 if (zio->io_logical == NULL) 621 zio->io_logical = pio->io_logical; 622 if (zio->io_child_type == ZIO_CHILD_GANG) 623 zio->io_gang_leader = pio->io_gang_leader; 624 zio_add_child(pio, zio); 625 } 626 627 return (zio); 628} 629 630static void 631zio_destroy(zio_t *zio) 632{ 633 list_destroy(&zio->io_parent_list); 634 list_destroy(&zio->io_child_list); 635 mutex_destroy(&zio->io_lock); 636 cv_destroy(&zio->io_cv); 637 kmem_cache_free(zio_cache, zio); 638} 639 640zio_t * 641zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done, 642 void *private, enum zio_flag flags) 643{ 644 zio_t *zio; 645 646 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private, 647 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, 648 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE); 649 650 return (zio); 651} 652 653zio_t * 654zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags) 655{ 656 return (zio_null(NULL, spa, NULL, done, private, flags)); 657} 658 659zio_t * 660zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, 661 void *data, uint64_t size, zio_done_func_t *done, void *private, 662 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb) 663{ 664 zio_t *zio; 665 666 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp, 667 data, size, done, private, 668 ZIO_TYPE_READ, priority, flags, NULL, 0, zb, 669 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? 670 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE); 671 672 return (zio); 673} 674 675zio_t * 676zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, 677 void *data, uint64_t size, const zio_prop_t *zp, 678 zio_done_func_t *ready, zio_done_func_t *physdone, zio_done_func_t *done, 679 void *private, 680 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb) 681{ 682 zio_t *zio; 683 684 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF && 685 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS && 686 zp->zp_compress >= ZIO_COMPRESS_OFF && 687 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS && 688 DMU_OT_IS_VALID(zp->zp_type) && 689 zp->zp_level < 32 && 690 zp->zp_copies > 0 && 691 zp->zp_copies <= spa_max_replication(spa)); 692 693 zio = zio_create(pio, spa, txg, bp, data, size, done, private, 694 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, 695 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? 696 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE); 697 698 zio->io_ready = ready; 699 zio->io_physdone = physdone; 700 zio->io_prop = *zp; 701 702 /* 703 * Data can be NULL if we are going to call zio_write_override() to 704 * provide the already-allocated BP. But we may need the data to 705 * verify a dedup hit (if requested). In this case, don't try to 706 * dedup (just take the already-allocated BP verbatim). 707 */ 708 if (data == NULL && zio->io_prop.zp_dedup_verify) { 709 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE; 710 } 711 712 return (zio); 713} 714 715zio_t * 716zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data, 717 uint64_t size, zio_done_func_t *done, void *private, 718 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb) 719{ 720 zio_t *zio; 721 722 zio = zio_create(pio, spa, txg, bp, data, size, done, private, 723 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, 724 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE); 725 726 return (zio); 727} 728 729void 730zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite) 731{ 732 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 733 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 734 ASSERT(zio->io_stage == ZIO_STAGE_OPEN); 735 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa)); 736 737 /* 738 * We must reset the io_prop to match the values that existed 739 * when the bp was first written by dmu_sync() keeping in mind 740 * that nopwrite and dedup are mutually exclusive. 741 */ 742 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup; 743 zio->io_prop.zp_nopwrite = nopwrite; 744 zio->io_prop.zp_copies = copies; 745 zio->io_bp_override = bp; 746} 747 748void 749zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp) 750{ 751 752 /* 753 * The check for EMBEDDED is a performance optimization. We 754 * process the free here (by ignoring it) rather than 755 * putting it on the list and then processing it in zio_free_sync(). 756 */ 757 if (BP_IS_EMBEDDED(bp)) 758 return; 759 metaslab_check_free(spa, bp); 760 761 /* 762 * Frees that are for the currently-syncing txg, are not going to be 763 * deferred, and which will not need to do a read (i.e. not GANG or 764 * DEDUP), can be processed immediately. Otherwise, put them on the 765 * in-memory list for later processing. 766 */ 767 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) || 768 txg != spa->spa_syncing_txg || 769 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) { 770 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp); 771 } else { 772 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 773 BP_GET_PSIZE(bp), 0))); 774 } 775} 776 777zio_t * 778zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 779 uint64_t size, enum zio_flag flags) 780{ 781 zio_t *zio; 782 enum zio_stage stage = ZIO_FREE_PIPELINE; 783 784 ASSERT(!BP_IS_HOLE(bp)); 785 ASSERT(spa_syncing_txg(spa) == txg); 786 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free); 787 788 if (BP_IS_EMBEDDED(bp)) 789 return (zio_null(pio, spa, NULL, NULL, NULL, 0)); 790 791 metaslab_check_free(spa, bp); 792 arc_freed(spa, bp); 793 794 if (zfs_trim_enabled) 795 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START | 796 ZIO_STAGE_VDEV_IO_ASSESS; 797 /* 798 * GANG and DEDUP blocks can induce a read (for the gang block header, 799 * or the DDT), so issue them asynchronously so that this thread is 800 * not tied up. 801 */ 802 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp)) 803 stage |= ZIO_STAGE_ISSUE_ASYNC; 804 805 flags |= ZIO_FLAG_DONT_QUEUE; 806 807 zio = zio_create(pio, spa, txg, bp, NULL, size, 808 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags, 809 NULL, 0, NULL, ZIO_STAGE_OPEN, stage); 810 811 return (zio); 812} 813 814zio_t * 815zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 816 zio_done_func_t *done, void *private, enum zio_flag flags) 817{ 818 zio_t *zio; 819 820 dprintf_bp(bp, "claiming in txg %llu", txg); 821 822 if (BP_IS_EMBEDDED(bp)) 823 return (zio_null(pio, spa, NULL, NULL, NULL, 0)); 824 825 /* 826 * A claim is an allocation of a specific block. Claims are needed 827 * to support immediate writes in the intent log. The issue is that 828 * immediate writes contain committed data, but in a txg that was 829 * *not* committed. Upon opening the pool after an unclean shutdown, 830 * the intent log claims all blocks that contain immediate write data 831 * so that the SPA knows they're in use. 832 * 833 * All claims *must* be resolved in the first txg -- before the SPA 834 * starts allocating blocks -- so that nothing is allocated twice. 835 * If txg == 0 we just verify that the block is claimable. 836 */ 837 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa)); 838 ASSERT(txg == spa_first_txg(spa) || txg == 0); 839 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */ 840 841 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), 842 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags, 843 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE); 844 845 return (zio); 846} 847 848zio_t * 849zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset, 850 uint64_t size, zio_done_func_t *done, void *private, 851 zio_priority_t priority, enum zio_flag flags) 852{ 853 zio_t *zio; 854 int c; 855 856 if (vd->vdev_children == 0) { 857 zio = zio_create(pio, spa, 0, NULL, NULL, size, done, private, 858 ZIO_TYPE_IOCTL, priority, flags, vd, offset, NULL, 859 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE); 860 861 zio->io_cmd = cmd; 862 } else { 863 zio = zio_null(pio, spa, NULL, NULL, NULL, flags); 864 865 for (c = 0; c < vd->vdev_children; c++) 866 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd, 867 offset, size, done, private, priority, flags)); 868 } 869 870 return (zio); 871} 872 873zio_t * 874zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 875 void *data, int checksum, zio_done_func_t *done, void *private, 876 zio_priority_t priority, enum zio_flag flags, boolean_t labels) 877{ 878 zio_t *zio; 879 880 ASSERT(vd->vdev_children == 0); 881 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || 882 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); 883 ASSERT3U(offset + size, <=, vd->vdev_psize); 884 885 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private, 886 ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset, 887 NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE); 888 889 zio->io_prop.zp_checksum = checksum; 890 891 return (zio); 892} 893 894zio_t * 895zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 896 void *data, int checksum, zio_done_func_t *done, void *private, 897 zio_priority_t priority, enum zio_flag flags, boolean_t labels) 898{ 899 zio_t *zio; 900 901 ASSERT(vd->vdev_children == 0); 902 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || 903 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); 904 ASSERT3U(offset + size, <=, vd->vdev_psize); 905 906 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private, 907 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset, 908 NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE); 909 910 zio->io_prop.zp_checksum = checksum; 911 912 if (zio_checksum_table[checksum].ci_eck) { 913 /* 914 * zec checksums are necessarily destructive -- they modify 915 * the end of the write buffer to hold the verifier/checksum. 916 * Therefore, we must make a local copy in case the data is 917 * being written to multiple places in parallel. 918 */ 919 void *wbuf = zio_buf_alloc(size); 920 bcopy(data, wbuf, size); 921 zio_push_transform(zio, wbuf, size, size, NULL); 922 } 923 924 return (zio); 925} 926 927/* 928 * Create a child I/O to do some work for us. 929 */ 930zio_t * 931zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset, 932 void *data, uint64_t size, int type, zio_priority_t priority, 933 enum zio_flag flags, zio_done_func_t *done, void *private) 934{ 935 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE; 936 zio_t *zio; 937 938 ASSERT(vd->vdev_parent == 939 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev)); 940 941 if (type == ZIO_TYPE_READ && bp != NULL) { 942 /* 943 * If we have the bp, then the child should perform the 944 * checksum and the parent need not. This pushes error 945 * detection as close to the leaves as possible and 946 * eliminates redundant checksums in the interior nodes. 947 */ 948 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY; 949 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; 950 } 951 952 /* Not all IO types require vdev io done stage e.g. free */ 953 if (!(pio->io_pipeline & ZIO_STAGE_VDEV_IO_DONE)) 954 pipeline &= ~ZIO_STAGE_VDEV_IO_DONE; 955 956 if (vd->vdev_children == 0) 957 offset += VDEV_LABEL_START_SIZE; 958 959 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE; 960 961 /* 962 * If we've decided to do a repair, the write is not speculative -- 963 * even if the original read was. 964 */ 965 if (flags & ZIO_FLAG_IO_REPAIR) 966 flags &= ~ZIO_FLAG_SPECULATIVE; 967 968 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, 969 done, private, type, priority, flags, vd, offset, &pio->io_bookmark, 970 ZIO_STAGE_VDEV_IO_START >> 1, pipeline); 971 972 zio->io_physdone = pio->io_physdone; 973 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL) 974 zio->io_logical->io_phys_children++; 975 976 return (zio); 977} 978 979zio_t * 980zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size, 981 int type, zio_priority_t priority, enum zio_flag flags, 982 zio_done_func_t *done, void *private) 983{ 984 zio_t *zio; 985 986 ASSERT(vd->vdev_ops->vdev_op_leaf); 987 988 zio = zio_create(NULL, vd->vdev_spa, 0, NULL, 989 data, size, done, private, type, priority, 990 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED, 991 vd, offset, NULL, 992 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE); 993 994 return (zio); 995} 996 997void 998zio_flush(zio_t *zio, vdev_t *vd) 999{ 1000 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0, 1001 NULL, NULL, ZIO_PRIORITY_NOW, 1002 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY)); 1003} 1004 1005zio_t * 1006zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size) 1007{ 1008 1009 ASSERT(vd->vdev_ops->vdev_op_leaf); 1010 1011 return (zio_create(zio, spa, 0, NULL, NULL, size, NULL, NULL, 1012 ZIO_TYPE_FREE, ZIO_PRIORITY_TRIM, ZIO_FLAG_DONT_AGGREGATE | 1013 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY, 1014 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PHYS_PIPELINE)); 1015} 1016 1017void 1018zio_shrink(zio_t *zio, uint64_t size) 1019{ 1020 ASSERT(zio->io_executor == NULL); 1021 ASSERT(zio->io_orig_size == zio->io_size); 1022 ASSERT(size <= zio->io_size); 1023 1024 /* 1025 * We don't shrink for raidz because of problems with the 1026 * reconstruction when reading back less than the block size. 1027 * Note, BP_IS_RAIDZ() assumes no compression. 1028 */ 1029 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 1030 if (!BP_IS_RAIDZ(zio->io_bp)) 1031 zio->io_orig_size = zio->io_size = size; 1032} 1033 1034/* 1035 * ========================================================================== 1036 * Prepare to read and write logical blocks 1037 * ========================================================================== 1038 */ 1039 1040static int 1041zio_read_bp_init(zio_t *zio) 1042{ 1043 blkptr_t *bp = zio->io_bp; 1044 1045 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF && 1046 zio->io_child_type == ZIO_CHILD_LOGICAL && 1047 !(zio->io_flags & ZIO_FLAG_RAW)) { 1048 uint64_t psize = 1049 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp); 1050 void *cbuf = zio_buf_alloc(psize); 1051 1052 zio_push_transform(zio, cbuf, psize, psize, zio_decompress); 1053 } 1054 1055 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) { 1056 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1057 decode_embedded_bp_compressed(bp, zio->io_data); 1058 } else { 1059 ASSERT(!BP_IS_EMBEDDED(bp)); 1060 } 1061 1062 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0) 1063 zio->io_flags |= ZIO_FLAG_DONT_CACHE; 1064 1065 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP) 1066 zio->io_flags |= ZIO_FLAG_DONT_CACHE; 1067 1068 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL) 1069 zio->io_pipeline = ZIO_DDT_READ_PIPELINE; 1070 1071 return (ZIO_PIPELINE_CONTINUE); 1072} 1073 1074static int 1075zio_write_bp_init(zio_t *zio) 1076{ 1077 spa_t *spa = zio->io_spa; 1078 zio_prop_t *zp = &zio->io_prop; 1079 enum zio_compress compress = zp->zp_compress; 1080 blkptr_t *bp = zio->io_bp; 1081 uint64_t lsize = zio->io_size; 1082 uint64_t psize = lsize; 1083 int pass = 1; 1084 1085 /* 1086 * If our children haven't all reached the ready stage, 1087 * wait for them and then repeat this pipeline stage. 1088 */ 1089 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) || 1090 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY)) 1091 return (ZIO_PIPELINE_STOP); 1092 1093 if (!IO_IS_ALLOCATING(zio)) 1094 return (ZIO_PIPELINE_CONTINUE); 1095 1096 ASSERT(zio->io_child_type != ZIO_CHILD_DDT); 1097 1098 if (zio->io_bp_override) { 1099 ASSERT(bp->blk_birth != zio->io_txg); 1100 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0); 1101 1102 *bp = *zio->io_bp_override; 1103 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1104 1105 if (BP_IS_EMBEDDED(bp)) 1106 return (ZIO_PIPELINE_CONTINUE); 1107 1108 /* 1109 * If we've been overridden and nopwrite is set then 1110 * set the flag accordingly to indicate that a nopwrite 1111 * has already occurred. 1112 */ 1113 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) { 1114 ASSERT(!zp->zp_dedup); 1115 zio->io_flags |= ZIO_FLAG_NOPWRITE; 1116 return (ZIO_PIPELINE_CONTINUE); 1117 } 1118 1119 ASSERT(!zp->zp_nopwrite); 1120 1121 if (BP_IS_HOLE(bp) || !zp->zp_dedup) 1122 return (ZIO_PIPELINE_CONTINUE); 1123 1124 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup || 1125 zp->zp_dedup_verify); 1126 1127 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) { 1128 BP_SET_DEDUP(bp, 1); 1129 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE; 1130 return (ZIO_PIPELINE_CONTINUE); 1131 } 1132 zio->io_bp_override = NULL; 1133 BP_ZERO(bp); 1134 } 1135 1136 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) { 1137 /* 1138 * We're rewriting an existing block, which means we're 1139 * working on behalf of spa_sync(). For spa_sync() to 1140 * converge, it must eventually be the case that we don't 1141 * have to allocate new blocks. But compression changes 1142 * the blocksize, which forces a reallocate, and makes 1143 * convergence take longer. Therefore, after the first 1144 * few passes, stop compressing to ensure convergence. 1145 */ 1146 pass = spa_sync_pass(spa); 1147 1148 ASSERT(zio->io_txg == spa_syncing_txg(spa)); 1149 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1150 ASSERT(!BP_GET_DEDUP(bp)); 1151 1152 if (pass >= zfs_sync_pass_dont_compress) 1153 compress = ZIO_COMPRESS_OFF; 1154 1155 /* Make sure someone doesn't change their mind on overwrites */ 1156 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp), 1157 spa_max_replication(spa)) == BP_GET_NDVAS(bp)); 1158 } 1159 1160 if (compress != ZIO_COMPRESS_OFF) { 1161 void *cbuf = zio_buf_alloc(lsize); 1162 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize); 1163 if (psize == 0 || psize == lsize) { 1164 compress = ZIO_COMPRESS_OFF; 1165 zio_buf_free(cbuf, lsize); 1166 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE && 1167 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) && 1168 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) { 1169 encode_embedded_bp_compressed(bp, 1170 cbuf, compress, lsize, psize); 1171 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA); 1172 BP_SET_TYPE(bp, zio->io_prop.zp_type); 1173 BP_SET_LEVEL(bp, zio->io_prop.zp_level); 1174 zio_buf_free(cbuf, lsize); 1175 bp->blk_birth = zio->io_txg; 1176 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1177 ASSERT(spa_feature_is_active(spa, 1178 SPA_FEATURE_EMBEDDED_DATA)); 1179 return (ZIO_PIPELINE_CONTINUE); 1180 } else { 1181 /* 1182 * Round up compressed size to MINBLOCKSIZE and 1183 * zero the tail. 1184 */ 1185 size_t rounded = 1186 P2ROUNDUP(psize, (size_t)SPA_MINBLOCKSIZE); 1187 if (rounded > psize) { 1188 bzero((char *)cbuf + psize, rounded - psize); 1189 psize = rounded; 1190 } 1191 if (psize == lsize) { 1192 compress = ZIO_COMPRESS_OFF; 1193 zio_buf_free(cbuf, lsize); 1194 } else { 1195 zio_push_transform(zio, cbuf, 1196 psize, lsize, NULL); 1197 } 1198 } 1199 } 1200 1201 /* 1202 * The final pass of spa_sync() must be all rewrites, but the first 1203 * few passes offer a trade-off: allocating blocks defers convergence, 1204 * but newly allocated blocks are sequential, so they can be written 1205 * to disk faster. Therefore, we allow the first few passes of 1206 * spa_sync() to allocate new blocks, but force rewrites after that. 1207 * There should only be a handful of blocks after pass 1 in any case. 1208 */ 1209 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg && 1210 BP_GET_PSIZE(bp) == psize && 1211 pass >= zfs_sync_pass_rewrite) { 1212 ASSERT(psize != 0); 1213 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES; 1214 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages; 1215 zio->io_flags |= ZIO_FLAG_IO_REWRITE; 1216 } else { 1217 BP_ZERO(bp); 1218 zio->io_pipeline = ZIO_WRITE_PIPELINE; 1219 } 1220 1221 if (psize == 0) { 1222 if (zio->io_bp_orig.blk_birth != 0 && 1223 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) { 1224 BP_SET_LSIZE(bp, lsize); 1225 BP_SET_TYPE(bp, zp->zp_type); 1226 BP_SET_LEVEL(bp, zp->zp_level); 1227 BP_SET_BIRTH(bp, zio->io_txg, 0); 1228 } 1229 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1230 } else { 1231 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER); 1232 BP_SET_LSIZE(bp, lsize); 1233 BP_SET_TYPE(bp, zp->zp_type); 1234 BP_SET_LEVEL(bp, zp->zp_level); 1235 BP_SET_PSIZE(bp, psize); 1236 BP_SET_COMPRESS(bp, compress); 1237 BP_SET_CHECKSUM(bp, zp->zp_checksum); 1238 BP_SET_DEDUP(bp, zp->zp_dedup); 1239 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER); 1240 if (zp->zp_dedup) { 1241 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1242 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 1243 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE; 1244 } 1245 if (zp->zp_nopwrite) { 1246 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1247 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 1248 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE; 1249 } 1250 } 1251 1252 return (ZIO_PIPELINE_CONTINUE); 1253} 1254 1255static int 1256zio_free_bp_init(zio_t *zio) 1257{ 1258 blkptr_t *bp = zio->io_bp; 1259 1260 if (zio->io_child_type == ZIO_CHILD_LOGICAL) { 1261 if (BP_GET_DEDUP(bp)) 1262 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE; 1263 } 1264 1265 return (ZIO_PIPELINE_CONTINUE); 1266} 1267 1268/* 1269 * ========================================================================== 1270 * Execute the I/O pipeline 1271 * ========================================================================== 1272 */ 1273 1274static void 1275zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline) 1276{ 1277 spa_t *spa = zio->io_spa; 1278 zio_type_t t = zio->io_type; 1279 int flags = (cutinline ? TQ_FRONT : 0); 1280 1281 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT); 1282 1283 /* 1284 * If we're a config writer or a probe, the normal issue and 1285 * interrupt threads may all be blocked waiting for the config lock. 1286 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL. 1287 */ 1288 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE)) 1289 t = ZIO_TYPE_NULL; 1290 1291 /* 1292 * A similar issue exists for the L2ARC write thread until L2ARC 2.0. 1293 */ 1294 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux) 1295 t = ZIO_TYPE_NULL; 1296 1297 /* 1298 * If this is a high priority I/O, then use the high priority taskq if 1299 * available. 1300 */ 1301 if (zio->io_priority == ZIO_PRIORITY_NOW && 1302 spa->spa_zio_taskq[t][q + 1].stqs_count != 0) 1303 q++; 1304 1305 ASSERT3U(q, <, ZIO_TASKQ_TYPES); 1306 1307 /* 1308 * NB: We are assuming that the zio can only be dispatched 1309 * to a single taskq at a time. It would be a grievous error 1310 * to dispatch the zio to another taskq at the same time. 1311 */ 1312#if defined(illumos) || !defined(_KERNEL) 1313 ASSERT(zio->io_tqent.tqent_next == NULL); 1314#else 1315 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0); 1316#endif 1317 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio, 1318 flags, &zio->io_tqent); 1319} 1320 1321static boolean_t 1322zio_taskq_member(zio_t *zio, zio_taskq_type_t q) 1323{ 1324 kthread_t *executor = zio->io_executor; 1325 spa_t *spa = zio->io_spa; 1326 1327 for (zio_type_t t = 0; t < ZIO_TYPES; t++) { 1328 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q]; 1329 uint_t i; 1330 for (i = 0; i < tqs->stqs_count; i++) { 1331 if (taskq_member(tqs->stqs_taskq[i], executor)) 1332 return (B_TRUE); 1333 } 1334 } 1335 1336 return (B_FALSE); 1337} 1338 1339static int 1340zio_issue_async(zio_t *zio) 1341{ 1342 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); 1343 1344 return (ZIO_PIPELINE_STOP); 1345} 1346 1347void 1348zio_interrupt(zio_t *zio) 1349{ 1350 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE); 1351} 1352 1353/* 1354 * Execute the I/O pipeline until one of the following occurs: 1355 * 1356 * (1) the I/O completes 1357 * (2) the pipeline stalls waiting for dependent child I/Os 1358 * (3) the I/O issues, so we're waiting for an I/O completion interrupt 1359 * (4) the I/O is delegated by vdev-level caching or aggregation 1360 * (5) the I/O is deferred due to vdev-level queueing 1361 * (6) the I/O is handed off to another thread. 1362 * 1363 * In all cases, the pipeline stops whenever there's no CPU work; it never 1364 * burns a thread in cv_wait(). 1365 * 1366 * There's no locking on io_stage because there's no legitimate way 1367 * for multiple threads to be attempting to process the same I/O. 1368 */ 1369static zio_pipe_stage_t *zio_pipeline[]; 1370 1371void 1372zio_execute(zio_t *zio) 1373{ 1374 zio->io_executor = curthread; 1375 1376 while (zio->io_stage < ZIO_STAGE_DONE) { 1377 enum zio_stage pipeline = zio->io_pipeline; 1378 enum zio_stage stage = zio->io_stage; 1379 int rv; 1380 1381 ASSERT(!MUTEX_HELD(&zio->io_lock)); 1382 ASSERT(ISP2(stage)); 1383 ASSERT(zio->io_stall == NULL); 1384 1385 do { 1386 stage <<= 1; 1387 } while ((stage & pipeline) == 0); 1388 1389 ASSERT(stage <= ZIO_STAGE_DONE); 1390 1391 /* 1392 * If we are in interrupt context and this pipeline stage 1393 * will grab a config lock that is held across I/O, 1394 * or may wait for an I/O that needs an interrupt thread 1395 * to complete, issue async to avoid deadlock. 1396 * 1397 * For VDEV_IO_START, we cut in line so that the io will 1398 * be sent to disk promptly. 1399 */ 1400 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL && 1401 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) { 1402 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? 1403 zio_requeue_io_start_cut_in_line : B_FALSE; 1404 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); 1405 return; 1406 } 1407 1408 zio->io_stage = stage; 1409 rv = zio_pipeline[highbit64(stage) - 1](zio); 1410 1411 if (rv == ZIO_PIPELINE_STOP) 1412 return; 1413 1414 ASSERT(rv == ZIO_PIPELINE_CONTINUE); 1415 } 1416} 1417 1418/* 1419 * ========================================================================== 1420 * Initiate I/O, either sync or async 1421 * ========================================================================== 1422 */ 1423int 1424zio_wait(zio_t *zio) 1425{ 1426 int error; 1427 1428 ASSERT(zio->io_stage == ZIO_STAGE_OPEN); 1429 ASSERT(zio->io_executor == NULL); 1430 1431 zio->io_waiter = curthread; 1432 1433 zio_execute(zio); 1434 1435 mutex_enter(&zio->io_lock); 1436 while (zio->io_executor != NULL) 1437 cv_wait(&zio->io_cv, &zio->io_lock); 1438 mutex_exit(&zio->io_lock); 1439 1440 error = zio->io_error; 1441 zio_destroy(zio); 1442 1443 return (error); 1444} 1445 1446void 1447zio_nowait(zio_t *zio) 1448{ 1449 ASSERT(zio->io_executor == NULL); 1450 1451 if (zio->io_child_type == ZIO_CHILD_LOGICAL && 1452 zio_unique_parent(zio) == NULL) { 1453 /* 1454 * This is a logical async I/O with no parent to wait for it. 1455 * We add it to the spa_async_root_zio "Godfather" I/O which 1456 * will ensure they complete prior to unloading the pool. 1457 */ 1458 spa_t *spa = zio->io_spa; 1459 1460 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio); 1461 } 1462 1463 zio_execute(zio); 1464} 1465 1466/* 1467 * ========================================================================== 1468 * Reexecute or suspend/resume failed I/O 1469 * ========================================================================== 1470 */ 1471 1472static void 1473zio_reexecute(zio_t *pio) 1474{ 1475 zio_t *cio, *cio_next; 1476 1477 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL); 1478 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN); 1479 ASSERT(pio->io_gang_leader == NULL); 1480 ASSERT(pio->io_gang_tree == NULL); 1481 1482 pio->io_flags = pio->io_orig_flags; 1483 pio->io_stage = pio->io_orig_stage; 1484 pio->io_pipeline = pio->io_orig_pipeline; 1485 pio->io_reexecute = 0; 1486 pio->io_flags |= ZIO_FLAG_REEXECUTED; 1487 pio->io_error = 0; 1488 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 1489 pio->io_state[w] = 0; 1490 for (int c = 0; c < ZIO_CHILD_TYPES; c++) 1491 pio->io_child_error[c] = 0; 1492 1493 if (IO_IS_ALLOCATING(pio)) 1494 BP_ZERO(pio->io_bp); 1495 1496 /* 1497 * As we reexecute pio's children, new children could be created. 1498 * New children go to the head of pio's io_child_list, however, 1499 * so we will (correctly) not reexecute them. The key is that 1500 * the remainder of pio's io_child_list, from 'cio_next' onward, 1501 * cannot be affected by any side effects of reexecuting 'cio'. 1502 */ 1503 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) { 1504 cio_next = zio_walk_children(pio); 1505 mutex_enter(&pio->io_lock); 1506 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 1507 pio->io_children[cio->io_child_type][w]++; 1508 mutex_exit(&pio->io_lock); 1509 zio_reexecute(cio); 1510 } 1511 1512 /* 1513 * Now that all children have been reexecuted, execute the parent. 1514 * We don't reexecute "The Godfather" I/O here as it's the 1515 * responsibility of the caller to wait on him. 1516 */ 1517 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) 1518 zio_execute(pio); 1519} 1520 1521void 1522zio_suspend(spa_t *spa, zio_t *zio) 1523{ 1524 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC) 1525 fm_panic("Pool '%s' has encountered an uncorrectable I/O " 1526 "failure and the failure mode property for this pool " 1527 "is set to panic.", spa_name(spa)); 1528 1529 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0); 1530 1531 mutex_enter(&spa->spa_suspend_lock); 1532 1533 if (spa->spa_suspend_zio_root == NULL) 1534 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL, 1535 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | 1536 ZIO_FLAG_GODFATHER); 1537 1538 spa->spa_suspended = B_TRUE; 1539 1540 if (zio != NULL) { 1541 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); 1542 ASSERT(zio != spa->spa_suspend_zio_root); 1543 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1544 ASSERT(zio_unique_parent(zio) == NULL); 1545 ASSERT(zio->io_stage == ZIO_STAGE_DONE); 1546 zio_add_child(spa->spa_suspend_zio_root, zio); 1547 } 1548 1549 mutex_exit(&spa->spa_suspend_lock); 1550} 1551 1552int 1553zio_resume(spa_t *spa) 1554{ 1555 zio_t *pio; 1556 1557 /* 1558 * Reexecute all previously suspended i/o. 1559 */ 1560 mutex_enter(&spa->spa_suspend_lock); 1561 spa->spa_suspended = B_FALSE; 1562 cv_broadcast(&spa->spa_suspend_cv); 1563 pio = spa->spa_suspend_zio_root; 1564 spa->spa_suspend_zio_root = NULL; 1565 mutex_exit(&spa->spa_suspend_lock); 1566 1567 if (pio == NULL) 1568 return (0); 1569 1570 zio_reexecute(pio); 1571 return (zio_wait(pio)); 1572} 1573 1574void 1575zio_resume_wait(spa_t *spa) 1576{ 1577 mutex_enter(&spa->spa_suspend_lock); 1578 while (spa_suspended(spa)) 1579 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock); 1580 mutex_exit(&spa->spa_suspend_lock); 1581} 1582 1583/* 1584 * ========================================================================== 1585 * Gang blocks. 1586 * 1587 * A gang block is a collection of small blocks that looks to the DMU 1588 * like one large block. When zio_dva_allocate() cannot find a block 1589 * of the requested size, due to either severe fragmentation or the pool 1590 * being nearly full, it calls zio_write_gang_block() to construct the 1591 * block from smaller fragments. 1592 * 1593 * A gang block consists of a gang header (zio_gbh_phys_t) and up to 1594 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like 1595 * an indirect block: it's an array of block pointers. It consumes 1596 * only one sector and hence is allocatable regardless of fragmentation. 1597 * The gang header's bps point to its gang members, which hold the data. 1598 * 1599 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg> 1600 * as the verifier to ensure uniqueness of the SHA256 checksum. 1601 * Critically, the gang block bp's blk_cksum is the checksum of the data, 1602 * not the gang header. This ensures that data block signatures (needed for 1603 * deduplication) are independent of how the block is physically stored. 1604 * 1605 * Gang blocks can be nested: a gang member may itself be a gang block. 1606 * Thus every gang block is a tree in which root and all interior nodes are 1607 * gang headers, and the leaves are normal blocks that contain user data. 1608 * The root of the gang tree is called the gang leader. 1609 * 1610 * To perform any operation (read, rewrite, free, claim) on a gang block, 1611 * zio_gang_assemble() first assembles the gang tree (minus data leaves) 1612 * in the io_gang_tree field of the original logical i/o by recursively 1613 * reading the gang leader and all gang headers below it. This yields 1614 * an in-core tree containing the contents of every gang header and the 1615 * bps for every constituent of the gang block. 1616 * 1617 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree 1618 * and invokes a callback on each bp. To free a gang block, zio_gang_issue() 1619 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp. 1620 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim(). 1621 * zio_read_gang() is a wrapper around zio_read() that omits reading gang 1622 * headers, since we already have those in io_gang_tree. zio_rewrite_gang() 1623 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite() 1624 * of the gang header plus zio_checksum_compute() of the data to update the 1625 * gang header's blk_cksum as described above. 1626 * 1627 * The two-phase assemble/issue model solves the problem of partial failure -- 1628 * what if you'd freed part of a gang block but then couldn't read the 1629 * gang header for another part? Assembling the entire gang tree first 1630 * ensures that all the necessary gang header I/O has succeeded before 1631 * starting the actual work of free, claim, or write. Once the gang tree 1632 * is assembled, free and claim are in-memory operations that cannot fail. 1633 * 1634 * In the event that a gang write fails, zio_dva_unallocate() walks the 1635 * gang tree to immediately free (i.e. insert back into the space map) 1636 * everything we've allocated. This ensures that we don't get ENOSPC 1637 * errors during repeated suspend/resume cycles due to a flaky device. 1638 * 1639 * Gang rewrites only happen during sync-to-convergence. If we can't assemble 1640 * the gang tree, we won't modify the block, so we can safely defer the free 1641 * (knowing that the block is still intact). If we *can* assemble the gang 1642 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free 1643 * each constituent bp and we can allocate a new block on the next sync pass. 1644 * 1645 * In all cases, the gang tree allows complete recovery from partial failure. 1646 * ========================================================================== 1647 */ 1648 1649static zio_t * 1650zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1651{ 1652 if (gn != NULL) 1653 return (pio); 1654 1655 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp), 1656 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), 1657 &pio->io_bookmark)); 1658} 1659 1660zio_t * 1661zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1662{ 1663 zio_t *zio; 1664 1665 if (gn != NULL) { 1666 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, 1667 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority, 1668 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 1669 /* 1670 * As we rewrite each gang header, the pipeline will compute 1671 * a new gang block header checksum for it; but no one will 1672 * compute a new data checksum, so we do that here. The one 1673 * exception is the gang leader: the pipeline already computed 1674 * its data checksum because that stage precedes gang assembly. 1675 * (Presently, nothing actually uses interior data checksums; 1676 * this is just good hygiene.) 1677 */ 1678 if (gn != pio->io_gang_leader->io_gang_tree) { 1679 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp), 1680 data, BP_GET_PSIZE(bp)); 1681 } 1682 /* 1683 * If we are here to damage data for testing purposes, 1684 * leave the GBH alone so that we can detect the damage. 1685 */ 1686 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE) 1687 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; 1688 } else { 1689 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, 1690 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority, 1691 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 1692 } 1693 1694 return (zio); 1695} 1696 1697/* ARGSUSED */ 1698zio_t * 1699zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1700{ 1701 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp, 1702 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp), 1703 ZIO_GANG_CHILD_FLAGS(pio))); 1704} 1705 1706/* ARGSUSED */ 1707zio_t * 1708zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1709{ 1710 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp, 1711 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio))); 1712} 1713 1714static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = { 1715 NULL, 1716 zio_read_gang, 1717 zio_rewrite_gang, 1718 zio_free_gang, 1719 zio_claim_gang, 1720 NULL 1721}; 1722 1723static void zio_gang_tree_assemble_done(zio_t *zio); 1724 1725static zio_gang_node_t * 1726zio_gang_node_alloc(zio_gang_node_t **gnpp) 1727{ 1728 zio_gang_node_t *gn; 1729 1730 ASSERT(*gnpp == NULL); 1731 1732 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP); 1733 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE); 1734 *gnpp = gn; 1735 1736 return (gn); 1737} 1738 1739static void 1740zio_gang_node_free(zio_gang_node_t **gnpp) 1741{ 1742 zio_gang_node_t *gn = *gnpp; 1743 1744 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) 1745 ASSERT(gn->gn_child[g] == NULL); 1746 1747 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE); 1748 kmem_free(gn, sizeof (*gn)); 1749 *gnpp = NULL; 1750} 1751 1752static void 1753zio_gang_tree_free(zio_gang_node_t **gnpp) 1754{ 1755 zio_gang_node_t *gn = *gnpp; 1756 1757 if (gn == NULL) 1758 return; 1759 1760 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) 1761 zio_gang_tree_free(&gn->gn_child[g]); 1762 1763 zio_gang_node_free(gnpp); 1764} 1765 1766static void 1767zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp) 1768{ 1769 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp); 1770 1771 ASSERT(gio->io_gang_leader == gio); 1772 ASSERT(BP_IS_GANG(bp)); 1773 1774 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh, 1775 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn, 1776 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark)); 1777} 1778 1779static void 1780zio_gang_tree_assemble_done(zio_t *zio) 1781{ 1782 zio_t *gio = zio->io_gang_leader; 1783 zio_gang_node_t *gn = zio->io_private; 1784 blkptr_t *bp = zio->io_bp; 1785 1786 ASSERT(gio == zio_unique_parent(zio)); 1787 ASSERT(zio->io_child_count == 0); 1788 1789 if (zio->io_error) 1790 return; 1791 1792 if (BP_SHOULD_BYTESWAP(bp)) 1793 byteswap_uint64_array(zio->io_data, zio->io_size); 1794 1795 ASSERT(zio->io_data == gn->gn_gbh); 1796 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE); 1797 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); 1798 1799 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 1800 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; 1801 if (!BP_IS_GANG(gbp)) 1802 continue; 1803 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]); 1804 } 1805} 1806 1807static void 1808zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data) 1809{ 1810 zio_t *gio = pio->io_gang_leader; 1811 zio_t *zio; 1812 1813 ASSERT(BP_IS_GANG(bp) == !!gn); 1814 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp)); 1815 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree); 1816 1817 /* 1818 * If you're a gang header, your data is in gn->gn_gbh. 1819 * If you're a gang member, your data is in 'data' and gn == NULL. 1820 */ 1821 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data); 1822 1823 if (gn != NULL) { 1824 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); 1825 1826 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 1827 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; 1828 if (BP_IS_HOLE(gbp)) 1829 continue; 1830 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data); 1831 data = (char *)data + BP_GET_PSIZE(gbp); 1832 } 1833 } 1834 1835 if (gn == gio->io_gang_tree && gio->io_data != NULL) 1836 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data); 1837 1838 if (zio != pio) 1839 zio_nowait(zio); 1840} 1841 1842static int 1843zio_gang_assemble(zio_t *zio) 1844{ 1845 blkptr_t *bp = zio->io_bp; 1846 1847 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL); 1848 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 1849 1850 zio->io_gang_leader = zio; 1851 1852 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree); 1853 1854 return (ZIO_PIPELINE_CONTINUE); 1855} 1856 1857static int 1858zio_gang_issue(zio_t *zio) 1859{ 1860 blkptr_t *bp = zio->io_bp; 1861 1862 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE)) 1863 return (ZIO_PIPELINE_STOP); 1864 1865 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio); 1866 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 1867 1868 if (zio->io_child_error[ZIO_CHILD_GANG] == 0) 1869 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data); 1870 else 1871 zio_gang_tree_free(&zio->io_gang_tree); 1872 1873 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1874 1875 return (ZIO_PIPELINE_CONTINUE); 1876} 1877 1878static void 1879zio_write_gang_member_ready(zio_t *zio) 1880{ 1881 zio_t *pio = zio_unique_parent(zio); 1882 zio_t *gio = zio->io_gang_leader; 1883 dva_t *cdva = zio->io_bp->blk_dva; 1884 dva_t *pdva = pio->io_bp->blk_dva; 1885 uint64_t asize; 1886 1887 if (BP_IS_HOLE(zio->io_bp)) 1888 return; 1889 1890 ASSERT(BP_IS_HOLE(&zio->io_bp_orig)); 1891 1892 ASSERT(zio->io_child_type == ZIO_CHILD_GANG); 1893 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies); 1894 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp)); 1895 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp)); 1896 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp)); 1897 1898 mutex_enter(&pio->io_lock); 1899 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) { 1900 ASSERT(DVA_GET_GANG(&pdva[d])); 1901 asize = DVA_GET_ASIZE(&pdva[d]); 1902 asize += DVA_GET_ASIZE(&cdva[d]); 1903 DVA_SET_ASIZE(&pdva[d], asize); 1904 } 1905 mutex_exit(&pio->io_lock); 1906} 1907 1908static int 1909zio_write_gang_block(zio_t *pio) 1910{ 1911 spa_t *spa = pio->io_spa; 1912 blkptr_t *bp = pio->io_bp; 1913 zio_t *gio = pio->io_gang_leader; 1914 zio_t *zio; 1915 zio_gang_node_t *gn, **gnpp; 1916 zio_gbh_phys_t *gbh; 1917 uint64_t txg = pio->io_txg; 1918 uint64_t resid = pio->io_size; 1919 uint64_t lsize; 1920 int copies = gio->io_prop.zp_copies; 1921 int gbh_copies = MIN(copies + 1, spa_max_replication(spa)); 1922 zio_prop_t zp; 1923 int error; 1924 1925 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE, 1926 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, 1927 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER); 1928 if (error) { 1929 pio->io_error = error; 1930 return (ZIO_PIPELINE_CONTINUE); 1931 } 1932 1933 if (pio == gio) { 1934 gnpp = &gio->io_gang_tree; 1935 } else { 1936 gnpp = pio->io_private; 1937 ASSERT(pio->io_ready == zio_write_gang_member_ready); 1938 } 1939 1940 gn = zio_gang_node_alloc(gnpp); 1941 gbh = gn->gn_gbh; 1942 bzero(gbh, SPA_GANGBLOCKSIZE); 1943 1944 /* 1945 * Create the gang header. 1946 */ 1947 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL, 1948 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 1949 1950 /* 1951 * Create and nowait the gang children. 1952 */ 1953 for (int g = 0; resid != 0; resid -= lsize, g++) { 1954 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g), 1955 SPA_MINBLOCKSIZE); 1956 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid); 1957 1958 zp.zp_checksum = gio->io_prop.zp_checksum; 1959 zp.zp_compress = ZIO_COMPRESS_OFF; 1960 zp.zp_type = DMU_OT_NONE; 1961 zp.zp_level = 0; 1962 zp.zp_copies = gio->io_prop.zp_copies; 1963 zp.zp_dedup = B_FALSE; 1964 zp.zp_dedup_verify = B_FALSE; 1965 zp.zp_nopwrite = B_FALSE; 1966 1967 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g], 1968 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp, 1969 zio_write_gang_member_ready, NULL, NULL, &gn->gn_child[g], 1970 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), 1971 &pio->io_bookmark)); 1972 } 1973 1974 /* 1975 * Set pio's pipeline to just wait for zio to finish. 1976 */ 1977 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1978 1979 zio_nowait(zio); 1980 1981 return (ZIO_PIPELINE_CONTINUE); 1982} 1983 1984/* 1985 * The zio_nop_write stage in the pipeline determines if allocating 1986 * a new bp is necessary. By leveraging a cryptographically secure checksum, 1987 * such as SHA256, we can compare the checksums of the new data and the old 1988 * to determine if allocating a new block is required. The nopwrite 1989 * feature can handle writes in either syncing or open context (i.e. zil 1990 * writes) and as a result is mutually exclusive with dedup. 1991 */ 1992static int 1993zio_nop_write(zio_t *zio) 1994{ 1995 blkptr_t *bp = zio->io_bp; 1996 blkptr_t *bp_orig = &zio->io_bp_orig; 1997 zio_prop_t *zp = &zio->io_prop; 1998 1999 ASSERT(BP_GET_LEVEL(bp) == 0); 2000 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 2001 ASSERT(zp->zp_nopwrite); 2002 ASSERT(!zp->zp_dedup); 2003 ASSERT(zio->io_bp_override == NULL); 2004 ASSERT(IO_IS_ALLOCATING(zio)); 2005 2006 /* 2007 * Check to see if the original bp and the new bp have matching 2008 * characteristics (i.e. same checksum, compression algorithms, etc). 2009 * If they don't then just continue with the pipeline which will 2010 * allocate a new bp. 2011 */ 2012 if (BP_IS_HOLE(bp_orig) || 2013 !zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup || 2014 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) || 2015 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) || 2016 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) || 2017 zp->zp_copies != BP_GET_NDVAS(bp_orig)) 2018 return (ZIO_PIPELINE_CONTINUE); 2019 2020 /* 2021 * If the checksums match then reset the pipeline so that we 2022 * avoid allocating a new bp and issuing any I/O. 2023 */ 2024 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) { 2025 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup); 2026 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig)); 2027 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig)); 2028 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF); 2029 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop, 2030 sizeof (uint64_t)) == 0); 2031 2032 *bp = *bp_orig; 2033 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2034 zio->io_flags |= ZIO_FLAG_NOPWRITE; 2035 } 2036 2037 return (ZIO_PIPELINE_CONTINUE); 2038} 2039 2040/* 2041 * ========================================================================== 2042 * Dedup 2043 * ========================================================================== 2044 */ 2045static void 2046zio_ddt_child_read_done(zio_t *zio) 2047{ 2048 blkptr_t *bp = zio->io_bp; 2049 ddt_entry_t *dde = zio->io_private; 2050 ddt_phys_t *ddp; 2051 zio_t *pio = zio_unique_parent(zio); 2052 2053 mutex_enter(&pio->io_lock); 2054 ddp = ddt_phys_select(dde, bp); 2055 if (zio->io_error == 0) 2056 ddt_phys_clear(ddp); /* this ddp doesn't need repair */ 2057 if (zio->io_error == 0 && dde->dde_repair_data == NULL) 2058 dde->dde_repair_data = zio->io_data; 2059 else 2060 zio_buf_free(zio->io_data, zio->io_size); 2061 mutex_exit(&pio->io_lock); 2062} 2063 2064static int 2065zio_ddt_read_start(zio_t *zio) 2066{ 2067 blkptr_t *bp = zio->io_bp; 2068 2069 ASSERT(BP_GET_DEDUP(bp)); 2070 ASSERT(BP_GET_PSIZE(bp) == zio->io_size); 2071 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2072 2073 if (zio->io_child_error[ZIO_CHILD_DDT]) { 2074 ddt_t *ddt = ddt_select(zio->io_spa, bp); 2075 ddt_entry_t *dde = ddt_repair_start(ddt, bp); 2076 ddt_phys_t *ddp = dde->dde_phys; 2077 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp); 2078 blkptr_t blk; 2079 2080 ASSERT(zio->io_vsd == NULL); 2081 zio->io_vsd = dde; 2082 2083 if (ddp_self == NULL) 2084 return (ZIO_PIPELINE_CONTINUE); 2085 2086 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) { 2087 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self) 2088 continue; 2089 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp, 2090 &blk); 2091 zio_nowait(zio_read(zio, zio->io_spa, &blk, 2092 zio_buf_alloc(zio->io_size), zio->io_size, 2093 zio_ddt_child_read_done, dde, zio->io_priority, 2094 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE, 2095 &zio->io_bookmark)); 2096 } 2097 return (ZIO_PIPELINE_CONTINUE); 2098 } 2099 2100 zio_nowait(zio_read(zio, zio->io_spa, bp, 2101 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority, 2102 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark)); 2103 2104 return (ZIO_PIPELINE_CONTINUE); 2105} 2106 2107static int 2108zio_ddt_read_done(zio_t *zio) 2109{ 2110 blkptr_t *bp = zio->io_bp; 2111 2112 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE)) 2113 return (ZIO_PIPELINE_STOP); 2114 2115 ASSERT(BP_GET_DEDUP(bp)); 2116 ASSERT(BP_GET_PSIZE(bp) == zio->io_size); 2117 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2118 2119 if (zio->io_child_error[ZIO_CHILD_DDT]) { 2120 ddt_t *ddt = ddt_select(zio->io_spa, bp); 2121 ddt_entry_t *dde = zio->io_vsd; 2122 if (ddt == NULL) { 2123 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE); 2124 return (ZIO_PIPELINE_CONTINUE); 2125 } 2126 if (dde == NULL) { 2127 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1; 2128 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); 2129 return (ZIO_PIPELINE_STOP); 2130 } 2131 if (dde->dde_repair_data != NULL) { 2132 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size); 2133 zio->io_child_error[ZIO_CHILD_DDT] = 0; 2134 } 2135 ddt_repair_done(ddt, dde); 2136 zio->io_vsd = NULL; 2137 } 2138 2139 ASSERT(zio->io_vsd == NULL); 2140 2141 return (ZIO_PIPELINE_CONTINUE); 2142} 2143 2144static boolean_t 2145zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde) 2146{ 2147 spa_t *spa = zio->io_spa; 2148 2149 /* 2150 * Note: we compare the original data, not the transformed data, 2151 * because when zio->io_bp is an override bp, we will not have 2152 * pushed the I/O transforms. That's an important optimization 2153 * because otherwise we'd compress/encrypt all dmu_sync() data twice. 2154 */ 2155 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { 2156 zio_t *lio = dde->dde_lead_zio[p]; 2157 2158 if (lio != NULL) { 2159 return (lio->io_orig_size != zio->io_orig_size || 2160 bcmp(zio->io_orig_data, lio->io_orig_data, 2161 zio->io_orig_size) != 0); 2162 } 2163 } 2164 2165 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { 2166 ddt_phys_t *ddp = &dde->dde_phys[p]; 2167 2168 if (ddp->ddp_phys_birth != 0) { 2169 arc_buf_t *abuf = NULL; 2170 uint32_t aflags = ARC_WAIT; 2171 blkptr_t blk = *zio->io_bp; 2172 int error; 2173 2174 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth); 2175 2176 ddt_exit(ddt); 2177 2178 error = arc_read(NULL, spa, &blk, 2179 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ, 2180 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2181 &aflags, &zio->io_bookmark); 2182 2183 if (error == 0) { 2184 if (arc_buf_size(abuf) != zio->io_orig_size || 2185 bcmp(abuf->b_data, zio->io_orig_data, 2186 zio->io_orig_size) != 0) 2187 error = SET_ERROR(EEXIST); 2188 VERIFY(arc_buf_remove_ref(abuf, &abuf)); 2189 } 2190 2191 ddt_enter(ddt); 2192 return (error != 0); 2193 } 2194 } 2195 2196 return (B_FALSE); 2197} 2198 2199static void 2200zio_ddt_child_write_ready(zio_t *zio) 2201{ 2202 int p = zio->io_prop.zp_copies; 2203 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); 2204 ddt_entry_t *dde = zio->io_private; 2205 ddt_phys_t *ddp = &dde->dde_phys[p]; 2206 zio_t *pio; 2207 2208 if (zio->io_error) 2209 return; 2210 2211 ddt_enter(ddt); 2212 2213 ASSERT(dde->dde_lead_zio[p] == zio); 2214 2215 ddt_phys_fill(ddp, zio->io_bp); 2216 2217 while ((pio = zio_walk_parents(zio)) != NULL) 2218 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg); 2219 2220 ddt_exit(ddt); 2221} 2222 2223static void 2224zio_ddt_child_write_done(zio_t *zio) 2225{ 2226 int p = zio->io_prop.zp_copies; 2227 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); 2228 ddt_entry_t *dde = zio->io_private; 2229 ddt_phys_t *ddp = &dde->dde_phys[p]; 2230 2231 ddt_enter(ddt); 2232 2233 ASSERT(ddp->ddp_refcnt == 0); 2234 ASSERT(dde->dde_lead_zio[p] == zio); 2235 dde->dde_lead_zio[p] = NULL; 2236 2237 if (zio->io_error == 0) { 2238 while (zio_walk_parents(zio) != NULL) 2239 ddt_phys_addref(ddp); 2240 } else { 2241 ddt_phys_clear(ddp); 2242 } 2243 2244 ddt_exit(ddt); 2245} 2246 2247static void 2248zio_ddt_ditto_write_done(zio_t *zio) 2249{ 2250 int p = DDT_PHYS_DITTO; 2251 zio_prop_t *zp = &zio->io_prop; 2252 blkptr_t *bp = zio->io_bp; 2253 ddt_t *ddt = ddt_select(zio->io_spa, bp); 2254 ddt_entry_t *dde = zio->io_private; 2255 ddt_phys_t *ddp = &dde->dde_phys[p]; 2256 ddt_key_t *ddk = &dde->dde_key; 2257 2258 ddt_enter(ddt); 2259 2260 ASSERT(ddp->ddp_refcnt == 0); 2261 ASSERT(dde->dde_lead_zio[p] == zio); 2262 dde->dde_lead_zio[p] = NULL; 2263 2264 if (zio->io_error == 0) { 2265 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum)); 2266 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP); 2267 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp)); 2268 if (ddp->ddp_phys_birth != 0) 2269 ddt_phys_free(ddt, ddk, ddp, zio->io_txg); 2270 ddt_phys_fill(ddp, bp); 2271 } 2272 2273 ddt_exit(ddt); 2274} 2275 2276static int 2277zio_ddt_write(zio_t *zio) 2278{ 2279 spa_t *spa = zio->io_spa; 2280 blkptr_t *bp = zio->io_bp; 2281 uint64_t txg = zio->io_txg; 2282 zio_prop_t *zp = &zio->io_prop; 2283 int p = zp->zp_copies; 2284 int ditto_copies; 2285 zio_t *cio = NULL; 2286 zio_t *dio = NULL; 2287 ddt_t *ddt = ddt_select(spa, bp); 2288 ddt_entry_t *dde; 2289 ddt_phys_t *ddp; 2290 2291 ASSERT(BP_GET_DEDUP(bp)); 2292 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum); 2293 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override); 2294 2295 ddt_enter(ddt); 2296 dde = ddt_lookup(ddt, bp, B_TRUE); 2297 ddp = &dde->dde_phys[p]; 2298 2299 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) { 2300 /* 2301 * If we're using a weak checksum, upgrade to a strong checksum 2302 * and try again. If we're already using a strong checksum, 2303 * we can't resolve it, so just convert to an ordinary write. 2304 * (And automatically e-mail a paper to Nature?) 2305 */ 2306 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) { 2307 zp->zp_checksum = spa_dedup_checksum(spa); 2308 zio_pop_transforms(zio); 2309 zio->io_stage = ZIO_STAGE_OPEN; 2310 BP_ZERO(bp); 2311 } else { 2312 zp->zp_dedup = B_FALSE; 2313 } 2314 zio->io_pipeline = ZIO_WRITE_PIPELINE; 2315 ddt_exit(ddt); 2316 return (ZIO_PIPELINE_CONTINUE); 2317 } 2318 2319 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp); 2320 ASSERT(ditto_copies < SPA_DVAS_PER_BP); 2321 2322 if (ditto_copies > ddt_ditto_copies_present(dde) && 2323 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) { 2324 zio_prop_t czp = *zp; 2325 2326 czp.zp_copies = ditto_copies; 2327 2328 /* 2329 * If we arrived here with an override bp, we won't have run 2330 * the transform stack, so we won't have the data we need to 2331 * generate a child i/o. So, toss the override bp and restart. 2332 * This is safe, because using the override bp is just an 2333 * optimization; and it's rare, so the cost doesn't matter. 2334 */ 2335 if (zio->io_bp_override) { 2336 zio_pop_transforms(zio); 2337 zio->io_stage = ZIO_STAGE_OPEN; 2338 zio->io_pipeline = ZIO_WRITE_PIPELINE; 2339 zio->io_bp_override = NULL; 2340 BP_ZERO(bp); 2341 ddt_exit(ddt); 2342 return (ZIO_PIPELINE_CONTINUE); 2343 } 2344 2345 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data, 2346 zio->io_orig_size, &czp, NULL, NULL, 2347 zio_ddt_ditto_write_done, dde, zio->io_priority, 2348 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); 2349 2350 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL); 2351 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio; 2352 } 2353 2354 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) { 2355 if (ddp->ddp_phys_birth != 0) 2356 ddt_bp_fill(ddp, bp, txg); 2357 if (dde->dde_lead_zio[p] != NULL) 2358 zio_add_child(zio, dde->dde_lead_zio[p]); 2359 else 2360 ddt_phys_addref(ddp); 2361 } else if (zio->io_bp_override) { 2362 ASSERT(bp->blk_birth == txg); 2363 ASSERT(BP_EQUAL(bp, zio->io_bp_override)); 2364 ddt_phys_fill(ddp, bp); 2365 ddt_phys_addref(ddp); 2366 } else { 2367 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data, 2368 zio->io_orig_size, zp, zio_ddt_child_write_ready, NULL, 2369 zio_ddt_child_write_done, dde, zio->io_priority, 2370 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); 2371 2372 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL); 2373 dde->dde_lead_zio[p] = cio; 2374 } 2375 2376 ddt_exit(ddt); 2377 2378 if (cio) 2379 zio_nowait(cio); 2380 if (dio) 2381 zio_nowait(dio); 2382 2383 return (ZIO_PIPELINE_CONTINUE); 2384} 2385 2386ddt_entry_t *freedde; /* for debugging */ 2387 2388static int 2389zio_ddt_free(zio_t *zio) 2390{ 2391 spa_t *spa = zio->io_spa; 2392 blkptr_t *bp = zio->io_bp; 2393 ddt_t *ddt = ddt_select(spa, bp); 2394 ddt_entry_t *dde; 2395 ddt_phys_t *ddp; 2396 2397 ASSERT(BP_GET_DEDUP(bp)); 2398 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2399 2400 ddt_enter(ddt); 2401 freedde = dde = ddt_lookup(ddt, bp, B_TRUE); 2402 ddp = ddt_phys_select(dde, bp); 2403 ddt_phys_decref(ddp); 2404 ddt_exit(ddt); 2405 2406 return (ZIO_PIPELINE_CONTINUE); 2407} 2408 2409/* 2410 * ========================================================================== 2411 * Allocate and free blocks 2412 * ========================================================================== 2413 */ 2414static int 2415zio_dva_allocate(zio_t *zio) 2416{ 2417 spa_t *spa = zio->io_spa; 2418 metaslab_class_t *mc = spa_normal_class(spa); 2419 blkptr_t *bp = zio->io_bp; 2420 int error; 2421 int flags = 0; 2422 2423 if (zio->io_gang_leader == NULL) { 2424 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 2425 zio->io_gang_leader = zio; 2426 } 2427 2428 ASSERT(BP_IS_HOLE(bp)); 2429 ASSERT0(BP_GET_NDVAS(bp)); 2430 ASSERT3U(zio->io_prop.zp_copies, >, 0); 2431 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa)); 2432 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp)); 2433 2434 /* 2435 * The dump device does not support gang blocks so allocation on 2436 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid 2437 * the "fast" gang feature. 2438 */ 2439 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0; 2440 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ? 2441 METASLAB_GANG_CHILD : 0; 2442 error = metaslab_alloc(spa, mc, zio->io_size, bp, 2443 zio->io_prop.zp_copies, zio->io_txg, NULL, flags); 2444 2445 if (error) { 2446 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, " 2447 "size %llu, error %d", spa_name(spa), zio, zio->io_size, 2448 error); 2449 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE) 2450 return (zio_write_gang_block(zio)); 2451 zio->io_error = error; 2452 } 2453 2454 return (ZIO_PIPELINE_CONTINUE); 2455} 2456 2457static int 2458zio_dva_free(zio_t *zio) 2459{ 2460 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE); 2461 2462 return (ZIO_PIPELINE_CONTINUE); 2463} 2464 2465static int 2466zio_dva_claim(zio_t *zio) 2467{ 2468 int error; 2469 2470 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg); 2471 if (error) 2472 zio->io_error = error; 2473 2474 return (ZIO_PIPELINE_CONTINUE); 2475} 2476 2477/* 2478 * Undo an allocation. This is used by zio_done() when an I/O fails 2479 * and we want to give back the block we just allocated. 2480 * This handles both normal blocks and gang blocks. 2481 */ 2482static void 2483zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp) 2484{ 2485 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp)); 2486 ASSERT(zio->io_bp_override == NULL); 2487 2488 if (!BP_IS_HOLE(bp)) 2489 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE); 2490 2491 if (gn != NULL) { 2492 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 2493 zio_dva_unallocate(zio, gn->gn_child[g], 2494 &gn->gn_gbh->zg_blkptr[g]); 2495 } 2496 } 2497} 2498 2499/* 2500 * Try to allocate an intent log block. Return 0 on success, errno on failure. 2501 */ 2502int 2503zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp, 2504 uint64_t size, boolean_t use_slog) 2505{ 2506 int error = 1; 2507 2508 ASSERT(txg > spa_syncing_txg(spa)); 2509 2510 /* 2511 * ZIL blocks are always contiguous (i.e. not gang blocks) so we 2512 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang" 2513 * when allocating them. 2514 */ 2515 if (use_slog) { 2516 error = metaslab_alloc(spa, spa_log_class(spa), size, 2517 new_bp, 1, txg, old_bp, 2518 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID); 2519 } 2520 2521 if (error) { 2522 error = metaslab_alloc(spa, spa_normal_class(spa), size, 2523 new_bp, 1, txg, old_bp, 2524 METASLAB_HINTBP_AVOID); 2525 } 2526 2527 if (error == 0) { 2528 BP_SET_LSIZE(new_bp, size); 2529 BP_SET_PSIZE(new_bp, size); 2530 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF); 2531 BP_SET_CHECKSUM(new_bp, 2532 spa_version(spa) >= SPA_VERSION_SLIM_ZIL 2533 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG); 2534 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); 2535 BP_SET_LEVEL(new_bp, 0); 2536 BP_SET_DEDUP(new_bp, 0); 2537 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER); 2538 } 2539 2540 return (error); 2541} 2542 2543/* 2544 * Free an intent log block. 2545 */ 2546void 2547zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp) 2548{ 2549 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG); 2550 ASSERT(!BP_IS_GANG(bp)); 2551 2552 zio_free(spa, txg, bp); 2553} 2554 2555/* 2556 * ========================================================================== 2557 * Read, write and delete to physical devices 2558 * ========================================================================== 2559 */ 2560 2561 2562/* 2563 * Issue an I/O to the underlying vdev. Typically the issue pipeline 2564 * stops after this stage and will resume upon I/O completion. 2565 * However, there are instances where the vdev layer may need to 2566 * continue the pipeline when an I/O was not issued. Since the I/O 2567 * that was sent to the vdev layer might be different than the one 2568 * currently active in the pipeline (see vdev_queue_io()), we explicitly 2569 * force the underlying vdev layers to call either zio_execute() or 2570 * zio_interrupt() to ensure that the pipeline continues with the correct I/O. 2571 */ 2572static int 2573zio_vdev_io_start(zio_t *zio) 2574{ 2575 vdev_t *vd = zio->io_vd; 2576 uint64_t align; 2577 spa_t *spa = zio->io_spa; 2578 int ret; 2579 2580 ASSERT(zio->io_error == 0); 2581 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0); 2582 2583 if (vd == NULL) { 2584 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) 2585 spa_config_enter(spa, SCL_ZIO, zio, RW_READER); 2586 2587 /* 2588 * The mirror_ops handle multiple DVAs in a single BP. 2589 */ 2590 vdev_mirror_ops.vdev_op_io_start(zio); 2591 return (ZIO_PIPELINE_STOP); 2592 } 2593 2594 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE && 2595 zio->io_priority == ZIO_PRIORITY_NOW) { 2596 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg); 2597 return (ZIO_PIPELINE_CONTINUE); 2598 } 2599 2600 /* 2601 * We keep track of time-sensitive I/Os so that the scan thread 2602 * can quickly react to certain workloads. In particular, we care 2603 * about non-scrubbing, top-level reads and writes with the following 2604 * characteristics: 2605 * - synchronous writes of user data to non-slog devices 2606 * - any reads of user data 2607 * When these conditions are met, adjust the timestamp of spa_last_io 2608 * which allows the scan thread to adjust its workload accordingly. 2609 */ 2610 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL && 2611 vd == vd->vdev_top && !vd->vdev_islog && 2612 zio->io_bookmark.zb_objset != DMU_META_OBJSET && 2613 zio->io_txg != spa_syncing_txg(spa)) { 2614 uint64_t old = spa->spa_last_io; 2615 uint64_t new = ddi_get_lbolt64(); 2616 if (old != new) 2617 (void) atomic_cas_64(&spa->spa_last_io, old, new); 2618 } 2619 2620 align = 1ULL << vd->vdev_top->vdev_ashift; 2621 2622 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) && 2623 P2PHASE(zio->io_size, align) != 0) { 2624 /* Transform logical writes to be a full physical block size. */ 2625 uint64_t asize = P2ROUNDUP(zio->io_size, align); 2626 char *abuf = NULL; 2627 if (zio->io_type == ZIO_TYPE_READ || 2628 zio->io_type == ZIO_TYPE_WRITE) 2629 abuf = zio_buf_alloc(asize); 2630 ASSERT(vd == vd->vdev_top); 2631 if (zio->io_type == ZIO_TYPE_WRITE) { 2632 bcopy(zio->io_data, abuf, zio->io_size); 2633 bzero(abuf + zio->io_size, asize - zio->io_size); 2634 } 2635 zio_push_transform(zio, abuf, asize, abuf ? asize : 0, 2636 zio_subblock); 2637 } 2638 2639 /* 2640 * If this is not a physical io, make sure that it is properly aligned 2641 * before proceeding. 2642 */ 2643 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) { 2644 ASSERT0(P2PHASE(zio->io_offset, align)); 2645 ASSERT0(P2PHASE(zio->io_size, align)); 2646 } else { 2647 /* 2648 * For physical writes, we allow 512b aligned writes and assume 2649 * the device will perform a read-modify-write as necessary. 2650 */ 2651 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE)); 2652 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE)); 2653 } 2654 2655 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa)); 2656 2657 /* 2658 * If this is a repair I/O, and there's no self-healing involved -- 2659 * that is, we're just resilvering what we expect to resilver -- 2660 * then don't do the I/O unless zio's txg is actually in vd's DTL. 2661 * This prevents spurious resilvering with nested replication. 2662 * For example, given a mirror of mirrors, (A+B)+(C+D), if only 2663 * A is out of date, we'll read from C+D, then use the data to 2664 * resilver A+B -- but we don't actually want to resilver B, just A. 2665 * The top-level mirror has no way to know this, so instead we just 2666 * discard unnecessary repairs as we work our way down the vdev tree. 2667 * The same logic applies to any form of nested replication: 2668 * ditto + mirror, RAID-Z + replacing, etc. This covers them all. 2669 */ 2670 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) && 2671 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) && 2672 zio->io_txg != 0 && /* not a delegated i/o */ 2673 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) { 2674 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 2675 zio_vdev_io_bypass(zio); 2676 return (ZIO_PIPELINE_CONTINUE); 2677 } 2678 2679 if (vd->vdev_ops->vdev_op_leaf) { 2680 switch (zio->io_type) { 2681 case ZIO_TYPE_READ: 2682 if (vdev_cache_read(zio)) 2683 return (ZIO_PIPELINE_CONTINUE); 2684 /* FALLTHROUGH */ 2685 case ZIO_TYPE_WRITE: 2686 case ZIO_TYPE_FREE: 2687 if ((zio = vdev_queue_io(zio)) == NULL) 2688 return (ZIO_PIPELINE_STOP); 2689 2690 if (!vdev_accessible(vd, zio)) { 2691 zio->io_error = SET_ERROR(ENXIO); 2692 zio_interrupt(zio); 2693 return (ZIO_PIPELINE_STOP); 2694 } 2695 break; 2696 } 2697 /* 2698 * Note that we ignore repair writes for TRIM because they can 2699 * conflict with normal writes. This isn't an issue because, by 2700 * definition, we only repair blocks that aren't freed. 2701 */ 2702 if (zio->io_type == ZIO_TYPE_WRITE && 2703 !(zio->io_flags & ZIO_FLAG_IO_REPAIR) && 2704 !trim_map_write_start(zio)) 2705 return (ZIO_PIPELINE_STOP); 2706 } 2707 2708 vd->vdev_ops->vdev_op_io_start(zio); 2709 return (ZIO_PIPELINE_STOP); 2710} 2711 2712static int 2713zio_vdev_io_done(zio_t *zio) 2714{ 2715 vdev_t *vd = zio->io_vd; 2716 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops; 2717 boolean_t unexpected_error = B_FALSE; 2718 2719 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE)) 2720 return (ZIO_PIPELINE_STOP); 2721 2722 ASSERT(zio->io_type == ZIO_TYPE_READ || 2723 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE); 2724 2725 if (vd != NULL && vd->vdev_ops->vdev_op_leaf && 2726 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE || 2727 zio->io_type == ZIO_TYPE_FREE)) { 2728 2729 if (zio->io_type == ZIO_TYPE_WRITE && 2730 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) 2731 trim_map_write_done(zio); 2732 2733 vdev_queue_io_done(zio); 2734 2735 if (zio->io_type == ZIO_TYPE_WRITE) 2736 vdev_cache_write(zio); 2737 2738 if (zio_injection_enabled && zio->io_error == 0) 2739 zio->io_error = zio_handle_device_injection(vd, 2740 zio, EIO); 2741 2742 if (zio_injection_enabled && zio->io_error == 0) 2743 zio->io_error = zio_handle_label_injection(zio, EIO); 2744 2745 if (zio->io_error) { 2746 if (zio->io_error == ENOTSUP && 2747 zio->io_type == ZIO_TYPE_FREE) { 2748 /* Not all devices support TRIM. */ 2749 } else if (!vdev_accessible(vd, zio)) { 2750 zio->io_error = SET_ERROR(ENXIO); 2751 } else { 2752 unexpected_error = B_TRUE; 2753 } 2754 } 2755 } 2756 2757 ops->vdev_op_io_done(zio); 2758 2759 if (unexpected_error) 2760 VERIFY(vdev_probe(vd, zio) == NULL); 2761 2762 return (ZIO_PIPELINE_CONTINUE); 2763} 2764 2765/* 2766 * For non-raidz ZIOs, we can just copy aside the bad data read from the 2767 * disk, and use that to finish the checksum ereport later. 2768 */ 2769static void 2770zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr, 2771 const void *good_buf) 2772{ 2773 /* no processing needed */ 2774 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE); 2775} 2776 2777/*ARGSUSED*/ 2778void 2779zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored) 2780{ 2781 void *buf = zio_buf_alloc(zio->io_size); 2782 2783 bcopy(zio->io_data, buf, zio->io_size); 2784 2785 zcr->zcr_cbinfo = zio->io_size; 2786 zcr->zcr_cbdata = buf; 2787 zcr->zcr_finish = zio_vsd_default_cksum_finish; 2788 zcr->zcr_free = zio_buf_free; 2789} 2790 2791static int 2792zio_vdev_io_assess(zio_t *zio) 2793{ 2794 vdev_t *vd = zio->io_vd; 2795 2796 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE)) 2797 return (ZIO_PIPELINE_STOP); 2798 2799 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) 2800 spa_config_exit(zio->io_spa, SCL_ZIO, zio); 2801 2802 if (zio->io_vsd != NULL) { 2803 zio->io_vsd_ops->vsd_free(zio); 2804 zio->io_vsd = NULL; 2805 } 2806 2807 if (zio_injection_enabled && zio->io_error == 0) 2808 zio->io_error = zio_handle_fault_injection(zio, EIO); 2809 2810 if (zio->io_type == ZIO_TYPE_FREE && 2811 zio->io_priority != ZIO_PRIORITY_NOW) { 2812 switch (zio->io_error) { 2813 case 0: 2814 ZIO_TRIM_STAT_INCR(bytes, zio->io_size); 2815 ZIO_TRIM_STAT_BUMP(success); 2816 break; 2817 case EOPNOTSUPP: 2818 ZIO_TRIM_STAT_BUMP(unsupported); 2819 break; 2820 default: 2821 ZIO_TRIM_STAT_BUMP(failed); 2822 break; 2823 } 2824 } 2825 2826 /* 2827 * If the I/O failed, determine whether we should attempt to retry it. 2828 * 2829 * On retry, we cut in line in the issue queue, since we don't want 2830 * compression/checksumming/etc. work to prevent our (cheap) IO reissue. 2831 */ 2832 if (zio->io_error && vd == NULL && 2833 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) { 2834 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */ 2835 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */ 2836 zio->io_error = 0; 2837 zio->io_flags |= ZIO_FLAG_IO_RETRY | 2838 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE; 2839 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1; 2840 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, 2841 zio_requeue_io_start_cut_in_line); 2842 return (ZIO_PIPELINE_STOP); 2843 } 2844 2845 /* 2846 * If we got an error on a leaf device, convert it to ENXIO 2847 * if the device is not accessible at all. 2848 */ 2849 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf && 2850 !vdev_accessible(vd, zio)) 2851 zio->io_error = SET_ERROR(ENXIO); 2852 2853 /* 2854 * If we can't write to an interior vdev (mirror or RAID-Z), 2855 * set vdev_cant_write so that we stop trying to allocate from it. 2856 */ 2857 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE && 2858 vd != NULL && !vd->vdev_ops->vdev_op_leaf) { 2859 vd->vdev_cant_write = B_TRUE; 2860 } 2861 2862 if (zio->io_error) 2863 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2864 2865 if (vd != NULL && vd->vdev_ops->vdev_op_leaf && 2866 zio->io_physdone != NULL) { 2867 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED)); 2868 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV); 2869 zio->io_physdone(zio->io_logical); 2870 } 2871 2872 return (ZIO_PIPELINE_CONTINUE); 2873} 2874 2875void 2876zio_vdev_io_reissue(zio_t *zio) 2877{ 2878 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); 2879 ASSERT(zio->io_error == 0); 2880 2881 zio->io_stage >>= 1; 2882} 2883 2884void 2885zio_vdev_io_redone(zio_t *zio) 2886{ 2887 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE); 2888 2889 zio->io_stage >>= 1; 2890} 2891 2892void 2893zio_vdev_io_bypass(zio_t *zio) 2894{ 2895 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); 2896 ASSERT(zio->io_error == 0); 2897 2898 zio->io_flags |= ZIO_FLAG_IO_BYPASS; 2899 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1; 2900} 2901 2902/* 2903 * ========================================================================== 2904 * Generate and verify checksums 2905 * ========================================================================== 2906 */ 2907static int 2908zio_checksum_generate(zio_t *zio) 2909{ 2910 blkptr_t *bp = zio->io_bp; 2911 enum zio_checksum checksum; 2912 2913 if (bp == NULL) { 2914 /* 2915 * This is zio_write_phys(). 2916 * We're either generating a label checksum, or none at all. 2917 */ 2918 checksum = zio->io_prop.zp_checksum; 2919 2920 if (checksum == ZIO_CHECKSUM_OFF) 2921 return (ZIO_PIPELINE_CONTINUE); 2922 2923 ASSERT(checksum == ZIO_CHECKSUM_LABEL); 2924 } else { 2925 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) { 2926 ASSERT(!IO_IS_ALLOCATING(zio)); 2927 checksum = ZIO_CHECKSUM_GANG_HEADER; 2928 } else { 2929 checksum = BP_GET_CHECKSUM(bp); 2930 } 2931 } 2932 2933 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size); 2934 2935 return (ZIO_PIPELINE_CONTINUE); 2936} 2937 2938static int 2939zio_checksum_verify(zio_t *zio) 2940{ 2941 zio_bad_cksum_t info; 2942 blkptr_t *bp = zio->io_bp; 2943 int error; 2944 2945 ASSERT(zio->io_vd != NULL); 2946 2947 if (bp == NULL) { 2948 /* 2949 * This is zio_read_phys(). 2950 * We're either verifying a label checksum, or nothing at all. 2951 */ 2952 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF) 2953 return (ZIO_PIPELINE_CONTINUE); 2954 2955 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL); 2956 } 2957 2958 if ((error = zio_checksum_error(zio, &info)) != 0) { 2959 zio->io_error = error; 2960 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) { 2961 zfs_ereport_start_checksum(zio->io_spa, 2962 zio->io_vd, zio, zio->io_offset, 2963 zio->io_size, NULL, &info); 2964 } 2965 } 2966 2967 return (ZIO_PIPELINE_CONTINUE); 2968} 2969 2970/* 2971 * Called by RAID-Z to ensure we don't compute the checksum twice. 2972 */ 2973void 2974zio_checksum_verified(zio_t *zio) 2975{ 2976 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; 2977} 2978 2979/* 2980 * ========================================================================== 2981 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other. 2982 * An error of 0 indicates success. ENXIO indicates whole-device failure, 2983 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO 2984 * indicate errors that are specific to one I/O, and most likely permanent. 2985 * Any other error is presumed to be worse because we weren't expecting it. 2986 * ========================================================================== 2987 */ 2988int 2989zio_worst_error(int e1, int e2) 2990{ 2991 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO }; 2992 int r1, r2; 2993 2994 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++) 2995 if (e1 == zio_error_rank[r1]) 2996 break; 2997 2998 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++) 2999 if (e2 == zio_error_rank[r2]) 3000 break; 3001 3002 return (r1 > r2 ? e1 : e2); 3003} 3004 3005/* 3006 * ========================================================================== 3007 * I/O completion 3008 * ========================================================================== 3009 */ 3010static int 3011zio_ready(zio_t *zio) 3012{ 3013 blkptr_t *bp = zio->io_bp; 3014 zio_t *pio, *pio_next; 3015 3016 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) || 3017 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY)) 3018 return (ZIO_PIPELINE_STOP); 3019 3020 if (zio->io_ready) { 3021 ASSERT(IO_IS_ALLOCATING(zio)); 3022 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) || 3023 (zio->io_flags & ZIO_FLAG_NOPWRITE)); 3024 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0); 3025 3026 zio->io_ready(zio); 3027 } 3028 3029 if (bp != NULL && bp != &zio->io_bp_copy) 3030 zio->io_bp_copy = *bp; 3031 3032 if (zio->io_error) 3033 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 3034 3035 mutex_enter(&zio->io_lock); 3036 zio->io_state[ZIO_WAIT_READY] = 1; 3037 pio = zio_walk_parents(zio); 3038 mutex_exit(&zio->io_lock); 3039 3040 /* 3041 * As we notify zio's parents, new parents could be added. 3042 * New parents go to the head of zio's io_parent_list, however, 3043 * so we will (correctly) not notify them. The remainder of zio's 3044 * io_parent_list, from 'pio_next' onward, cannot change because 3045 * all parents must wait for us to be done before they can be done. 3046 */ 3047 for (; pio != NULL; pio = pio_next) { 3048 pio_next = zio_walk_parents(zio); 3049 zio_notify_parent(pio, zio, ZIO_WAIT_READY); 3050 } 3051 3052 if (zio->io_flags & ZIO_FLAG_NODATA) { 3053 if (BP_IS_GANG(bp)) { 3054 zio->io_flags &= ~ZIO_FLAG_NODATA; 3055 } else { 3056 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE); 3057 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; 3058 } 3059 } 3060 3061 if (zio_injection_enabled && 3062 zio->io_spa->spa_syncing_txg == zio->io_txg) 3063 zio_handle_ignored_writes(zio); 3064 3065 return (ZIO_PIPELINE_CONTINUE); 3066} 3067 3068static int 3069zio_done(zio_t *zio) 3070{ 3071 spa_t *spa = zio->io_spa; 3072 zio_t *lio = zio->io_logical; 3073 blkptr_t *bp = zio->io_bp; 3074 vdev_t *vd = zio->io_vd; 3075 uint64_t psize = zio->io_size; 3076 zio_t *pio, *pio_next; 3077 3078 /* 3079 * If our children haven't all completed, 3080 * wait for them and then repeat this pipeline stage. 3081 */ 3082 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) || 3083 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) || 3084 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) || 3085 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE)) 3086 return (ZIO_PIPELINE_STOP); 3087 3088 for (int c = 0; c < ZIO_CHILD_TYPES; c++) 3089 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 3090 ASSERT(zio->io_children[c][w] == 0); 3091 3092 if (bp != NULL && !BP_IS_EMBEDDED(bp)) { 3093 ASSERT(bp->blk_pad[0] == 0); 3094 ASSERT(bp->blk_pad[1] == 0); 3095 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 || 3096 (bp == zio_unique_parent(zio)->io_bp)); 3097 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) && 3098 zio->io_bp_override == NULL && 3099 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) { 3100 ASSERT(!BP_SHOULD_BYTESWAP(bp)); 3101 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp)); 3102 ASSERT(BP_COUNT_GANG(bp) == 0 || 3103 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp))); 3104 } 3105 if (zio->io_flags & ZIO_FLAG_NOPWRITE) 3106 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig)); 3107 } 3108 3109 /* 3110 * If there were child vdev/gang/ddt errors, they apply to us now. 3111 */ 3112 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV); 3113 zio_inherit_child_errors(zio, ZIO_CHILD_GANG); 3114 zio_inherit_child_errors(zio, ZIO_CHILD_DDT); 3115 3116 /* 3117 * If the I/O on the transformed data was successful, generate any 3118 * checksum reports now while we still have the transformed data. 3119 */ 3120 if (zio->io_error == 0) { 3121 while (zio->io_cksum_report != NULL) { 3122 zio_cksum_report_t *zcr = zio->io_cksum_report; 3123 uint64_t align = zcr->zcr_align; 3124 uint64_t asize = P2ROUNDUP(psize, align); 3125 char *abuf = zio->io_data; 3126 3127 if (asize != psize) { 3128 abuf = zio_buf_alloc(asize); 3129 bcopy(zio->io_data, abuf, psize); 3130 bzero(abuf + psize, asize - psize); 3131 } 3132 3133 zio->io_cksum_report = zcr->zcr_next; 3134 zcr->zcr_next = NULL; 3135 zcr->zcr_finish(zcr, abuf); 3136 zfs_ereport_free_checksum(zcr); 3137 3138 if (asize != psize) 3139 zio_buf_free(abuf, asize); 3140 } 3141 } 3142 3143 zio_pop_transforms(zio); /* note: may set zio->io_error */ 3144 3145 vdev_stat_update(zio, psize); 3146 3147 if (zio->io_error) { 3148 /* 3149 * If this I/O is attached to a particular vdev, 3150 * generate an error message describing the I/O failure 3151 * at the block level. We ignore these errors if the 3152 * device is currently unavailable. 3153 */ 3154 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd)) 3155 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0); 3156 3157 if ((zio->io_error == EIO || !(zio->io_flags & 3158 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) && 3159 zio == lio) { 3160 /* 3161 * For logical I/O requests, tell the SPA to log the 3162 * error and generate a logical data ereport. 3163 */ 3164 spa_log_error(spa, zio); 3165 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio, 3166 0, 0); 3167 } 3168 } 3169 3170 if (zio->io_error && zio == lio) { 3171 /* 3172 * Determine whether zio should be reexecuted. This will 3173 * propagate all the way to the root via zio_notify_parent(). 3174 */ 3175 ASSERT(vd == NULL && bp != NULL); 3176 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 3177 3178 if (IO_IS_ALLOCATING(zio) && 3179 !(zio->io_flags & ZIO_FLAG_CANFAIL)) { 3180 if (zio->io_error != ENOSPC) 3181 zio->io_reexecute |= ZIO_REEXECUTE_NOW; 3182 else 3183 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 3184 } 3185 3186 if ((zio->io_type == ZIO_TYPE_READ || 3187 zio->io_type == ZIO_TYPE_FREE) && 3188 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && 3189 zio->io_error == ENXIO && 3190 spa_load_state(spa) == SPA_LOAD_NONE && 3191 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE) 3192 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 3193 3194 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute) 3195 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 3196 3197 /* 3198 * Here is a possibly good place to attempt to do 3199 * either combinatorial reconstruction or error correction 3200 * based on checksums. It also might be a good place 3201 * to send out preliminary ereports before we suspend 3202 * processing. 3203 */ 3204 } 3205 3206 /* 3207 * If there were logical child errors, they apply to us now. 3208 * We defer this until now to avoid conflating logical child 3209 * errors with errors that happened to the zio itself when 3210 * updating vdev stats and reporting FMA events above. 3211 */ 3212 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL); 3213 3214 if ((zio->io_error || zio->io_reexecute) && 3215 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio && 3216 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE))) 3217 zio_dva_unallocate(zio, zio->io_gang_tree, bp); 3218 3219 zio_gang_tree_free(&zio->io_gang_tree); 3220 3221 /* 3222 * Godfather I/Os should never suspend. 3223 */ 3224 if ((zio->io_flags & ZIO_FLAG_GODFATHER) && 3225 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) 3226 zio->io_reexecute = 0; 3227 3228 if (zio->io_reexecute) { 3229 /* 3230 * This is a logical I/O that wants to reexecute. 3231 * 3232 * Reexecute is top-down. When an i/o fails, if it's not 3233 * the root, it simply notifies its parent and sticks around. 3234 * The parent, seeing that it still has children in zio_done(), 3235 * does the same. This percolates all the way up to the root. 3236 * The root i/o will reexecute or suspend the entire tree. 3237 * 3238 * This approach ensures that zio_reexecute() honors 3239 * all the original i/o dependency relationships, e.g. 3240 * parents not executing until children are ready. 3241 */ 3242 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 3243 3244 zio->io_gang_leader = NULL; 3245 3246 mutex_enter(&zio->io_lock); 3247 zio->io_state[ZIO_WAIT_DONE] = 1; 3248 mutex_exit(&zio->io_lock); 3249 3250 /* 3251 * "The Godfather" I/O monitors its children but is 3252 * not a true parent to them. It will track them through 3253 * the pipeline but severs its ties whenever they get into 3254 * trouble (e.g. suspended). This allows "The Godfather" 3255 * I/O to return status without blocking. 3256 */ 3257 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) { 3258 zio_link_t *zl = zio->io_walk_link; 3259 pio_next = zio_walk_parents(zio); 3260 3261 if ((pio->io_flags & ZIO_FLAG_GODFATHER) && 3262 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) { 3263 zio_remove_child(pio, zio, zl); 3264 zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 3265 } 3266 } 3267 3268 if ((pio = zio_unique_parent(zio)) != NULL) { 3269 /* 3270 * We're not a root i/o, so there's nothing to do 3271 * but notify our parent. Don't propagate errors 3272 * upward since we haven't permanently failed yet. 3273 */ 3274 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); 3275 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE; 3276 zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 3277 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) { 3278 /* 3279 * We'd fail again if we reexecuted now, so suspend 3280 * until conditions improve (e.g. device comes online). 3281 */ 3282 zio_suspend(spa, zio); 3283 } else { 3284 /* 3285 * Reexecution is potentially a huge amount of work. 3286 * Hand it off to the otherwise-unused claim taskq. 3287 */ 3288#if defined(illumos) || !defined(_KERNEL) 3289 ASSERT(zio->io_tqent.tqent_next == NULL); 3290#else 3291 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0); 3292#endif 3293 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM, 3294 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio, 3295 0, &zio->io_tqent); 3296 } 3297 return (ZIO_PIPELINE_STOP); 3298 } 3299 3300 ASSERT(zio->io_child_count == 0); 3301 ASSERT(zio->io_reexecute == 0); 3302 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL)); 3303 3304 /* 3305 * Report any checksum errors, since the I/O is complete. 3306 */ 3307 while (zio->io_cksum_report != NULL) { 3308 zio_cksum_report_t *zcr = zio->io_cksum_report; 3309 zio->io_cksum_report = zcr->zcr_next; 3310 zcr->zcr_next = NULL; 3311 zcr->zcr_finish(zcr, NULL); 3312 zfs_ereport_free_checksum(zcr); 3313 } 3314 3315 /* 3316 * It is the responsibility of the done callback to ensure that this 3317 * particular zio is no longer discoverable for adoption, and as 3318 * such, cannot acquire any new parents. 3319 */ 3320 if (zio->io_done) 3321 zio->io_done(zio); 3322 3323 mutex_enter(&zio->io_lock); 3324 zio->io_state[ZIO_WAIT_DONE] = 1; 3325 mutex_exit(&zio->io_lock); 3326 3327 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) { 3328 zio_link_t *zl = zio->io_walk_link; 3329 pio_next = zio_walk_parents(zio); 3330 zio_remove_child(pio, zio, zl); 3331 zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 3332 } 3333 3334 if (zio->io_waiter != NULL) { 3335 mutex_enter(&zio->io_lock); 3336 zio->io_executor = NULL; 3337 cv_broadcast(&zio->io_cv); 3338 mutex_exit(&zio->io_lock); 3339 } else { 3340 zio_destroy(zio); 3341 } 3342 3343 return (ZIO_PIPELINE_STOP); 3344} 3345 3346/* 3347 * ========================================================================== 3348 * I/O pipeline definition 3349 * ========================================================================== 3350 */ 3351static zio_pipe_stage_t *zio_pipeline[] = { 3352 NULL, 3353 zio_read_bp_init, 3354 zio_free_bp_init, 3355 zio_issue_async, 3356 zio_write_bp_init, 3357 zio_checksum_generate, 3358 zio_nop_write, 3359 zio_ddt_read_start, 3360 zio_ddt_read_done, 3361 zio_ddt_write, 3362 zio_ddt_free, 3363 zio_gang_assemble, 3364 zio_gang_issue, 3365 zio_dva_allocate, 3366 zio_dva_free, 3367 zio_dva_claim, 3368 zio_ready, 3369 zio_vdev_io_start, 3370 zio_vdev_io_done, 3371 zio_vdev_io_assess, 3372 zio_checksum_verify, 3373 zio_done 3374}; 3375 3376/* dnp is the dnode for zb1->zb_object */ 3377boolean_t 3378zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_phys_t *zb1, 3379 const zbookmark_phys_t *zb2) 3380{ 3381 uint64_t zb1nextL0, zb2thisobj; 3382 3383 ASSERT(zb1->zb_objset == zb2->zb_objset); 3384 ASSERT(zb2->zb_level == 0); 3385 3386 /* The objset_phys_t isn't before anything. */ 3387 if (dnp == NULL) 3388 return (B_FALSE); 3389 3390 zb1nextL0 = (zb1->zb_blkid + 1) << 3391 ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT)); 3392 3393 zb2thisobj = zb2->zb_object ? zb2->zb_object : 3394 zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT); 3395 3396 if (zb1->zb_object == DMU_META_DNODE_OBJECT) { 3397 uint64_t nextobj = zb1nextL0 * 3398 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT; 3399 return (nextobj <= zb2thisobj); 3400 } 3401 3402 if (zb1->zb_object < zb2thisobj) 3403 return (B_TRUE); 3404 if (zb1->zb_object > zb2thisobj) 3405 return (B_FALSE); 3406 if (zb2->zb_object == DMU_META_DNODE_OBJECT) 3407 return (B_FALSE); 3408 return (zb1nextL0 <= zb2->zb_blkid); 3409}
| 156 p2 &= p2 - 1; 157 158#ifdef illumos 159#ifndef _KERNEL 160 /* 161 * If we are using watchpoints, put each buffer on its own page, 162 * to eliminate the performance overhead of trapping to the 163 * kernel when modifying a non-watched buffer that shares the 164 * page with a watched buffer. 165 */ 166 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE)) 167 continue; 168#endif 169#endif /* illumos */ 170 if (size <= 4 * SPA_MINBLOCKSIZE) { 171 align = SPA_MINBLOCKSIZE; 172 } else if (IS_P2ALIGNED(size, p2 >> 2)) { 173 align = MIN(p2 >> 2, PAGESIZE); 174 } 175 176 if (align != 0) { 177 char name[36]; 178 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size); 179 zio_buf_cache[c] = kmem_cache_create(name, size, 180 align, NULL, NULL, NULL, NULL, NULL, cflags); 181 182 /* 183 * Since zio_data bufs do not appear in crash dumps, we 184 * pass KMC_NOTOUCH so that no allocator metadata is 185 * stored with the buffers. 186 */ 187 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size); 188 zio_data_buf_cache[c] = kmem_cache_create(name, size, 189 align, NULL, NULL, NULL, NULL, NULL, 190 cflags | KMC_NOTOUCH | KMC_NODEBUG); 191 } 192 } 193 194 while (--c != 0) { 195 ASSERT(zio_buf_cache[c] != NULL); 196 if (zio_buf_cache[c - 1] == NULL) 197 zio_buf_cache[c - 1] = zio_buf_cache[c]; 198 199 ASSERT(zio_data_buf_cache[c] != NULL); 200 if (zio_data_buf_cache[c - 1] == NULL) 201 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c]; 202 } 203out: 204 205 zio_inject_init(); 206 207 zio_trim_ksp = kstat_create("zfs", 0, "zio_trim", "misc", 208 KSTAT_TYPE_NAMED, 209 sizeof(zio_trim_stats) / sizeof(kstat_named_t), 210 KSTAT_FLAG_VIRTUAL); 211 212 if (zio_trim_ksp != NULL) { 213 zio_trim_ksp->ks_data = &zio_trim_stats; 214 kstat_install(zio_trim_ksp); 215 } 216} 217 218void 219zio_fini(void) 220{ 221 size_t c; 222 kmem_cache_t *last_cache = NULL; 223 kmem_cache_t *last_data_cache = NULL; 224 225 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { 226 if (zio_buf_cache[c] != last_cache) { 227 last_cache = zio_buf_cache[c]; 228 kmem_cache_destroy(zio_buf_cache[c]); 229 } 230 zio_buf_cache[c] = NULL; 231 232 if (zio_data_buf_cache[c] != last_data_cache) { 233 last_data_cache = zio_data_buf_cache[c]; 234 kmem_cache_destroy(zio_data_buf_cache[c]); 235 } 236 zio_data_buf_cache[c] = NULL; 237 } 238 239 kmem_cache_destroy(zio_link_cache); 240 kmem_cache_destroy(zio_cache); 241 242 zio_inject_fini(); 243 244 if (zio_trim_ksp != NULL) { 245 kstat_delete(zio_trim_ksp); 246 zio_trim_ksp = NULL; 247 } 248} 249 250/* 251 * ========================================================================== 252 * Allocate and free I/O buffers 253 * ========================================================================== 254 */ 255 256/* 257 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a 258 * crashdump if the kernel panics, so use it judiciously. Obviously, it's 259 * useful to inspect ZFS metadata, but if possible, we should avoid keeping 260 * excess / transient data in-core during a crashdump. 261 */ 262void * 263zio_buf_alloc(size_t size) 264{ 265 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 266 int flags = zio_exclude_metadata ? KM_NODEBUG : 0; 267 268 ASSERT3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 269 270 if (zio_use_uma) 271 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE)); 272 else 273 return (kmem_alloc(size, KM_SLEEP|flags)); 274} 275 276/* 277 * Use zio_data_buf_alloc to allocate data. The data will not appear in a 278 * crashdump if the kernel panics. This exists so that we will limit the amount 279 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount 280 * of kernel heap dumped to disk when the kernel panics) 281 */ 282void * 283zio_data_buf_alloc(size_t size) 284{ 285 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 286 287 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 288 289 if (zio_use_uma) 290 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE)); 291 else 292 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG)); 293} 294 295void 296zio_buf_free(void *buf, size_t size) 297{ 298 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 299 300 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 301 302 if (zio_use_uma) 303 kmem_cache_free(zio_buf_cache[c], buf); 304 else 305 kmem_free(buf, size); 306} 307 308void 309zio_data_buf_free(void *buf, size_t size) 310{ 311 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 312 313 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 314 315 if (zio_use_uma) 316 kmem_cache_free(zio_data_buf_cache[c], buf); 317 else 318 kmem_free(buf, size); 319} 320 321/* 322 * ========================================================================== 323 * Push and pop I/O transform buffers 324 * ========================================================================== 325 */ 326static void 327zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize, 328 zio_transform_func_t *transform) 329{ 330 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP); 331 332 zt->zt_orig_data = zio->io_data; 333 zt->zt_orig_size = zio->io_size; 334 zt->zt_bufsize = bufsize; 335 zt->zt_transform = transform; 336 337 zt->zt_next = zio->io_transform_stack; 338 zio->io_transform_stack = zt; 339 340 zio->io_data = data; 341 zio->io_size = size; 342} 343 344static void 345zio_pop_transforms(zio_t *zio) 346{ 347 zio_transform_t *zt; 348 349 while ((zt = zio->io_transform_stack) != NULL) { 350 if (zt->zt_transform != NULL) 351 zt->zt_transform(zio, 352 zt->zt_orig_data, zt->zt_orig_size); 353 354 if (zt->zt_bufsize != 0) 355 zio_buf_free(zio->io_data, zt->zt_bufsize); 356 357 zio->io_data = zt->zt_orig_data; 358 zio->io_size = zt->zt_orig_size; 359 zio->io_transform_stack = zt->zt_next; 360 361 kmem_free(zt, sizeof (zio_transform_t)); 362 } 363} 364 365/* 366 * ========================================================================== 367 * I/O transform callbacks for subblocks and decompression 368 * ========================================================================== 369 */ 370static void 371zio_subblock(zio_t *zio, void *data, uint64_t size) 372{ 373 ASSERT(zio->io_size > size); 374 375 if (zio->io_type == ZIO_TYPE_READ) 376 bcopy(zio->io_data, data, size); 377} 378 379static void 380zio_decompress(zio_t *zio, void *data, uint64_t size) 381{ 382 if (zio->io_error == 0 && 383 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp), 384 zio->io_data, data, zio->io_size, size) != 0) 385 zio->io_error = SET_ERROR(EIO); 386} 387 388/* 389 * ========================================================================== 390 * I/O parent/child relationships and pipeline interlocks 391 * ========================================================================== 392 */ 393/* 394 * NOTE - Callers to zio_walk_parents() and zio_walk_children must 395 * continue calling these functions until they return NULL. 396 * Otherwise, the next caller will pick up the list walk in 397 * some indeterminate state. (Otherwise every caller would 398 * have to pass in a cookie to keep the state represented by 399 * io_walk_link, which gets annoying.) 400 */ 401zio_t * 402zio_walk_parents(zio_t *cio) 403{ 404 zio_link_t *zl = cio->io_walk_link; 405 list_t *pl = &cio->io_parent_list; 406 407 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl); 408 cio->io_walk_link = zl; 409 410 if (zl == NULL) 411 return (NULL); 412 413 ASSERT(zl->zl_child == cio); 414 return (zl->zl_parent); 415} 416 417zio_t * 418zio_walk_children(zio_t *pio) 419{ 420 zio_link_t *zl = pio->io_walk_link; 421 list_t *cl = &pio->io_child_list; 422 423 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl); 424 pio->io_walk_link = zl; 425 426 if (zl == NULL) 427 return (NULL); 428 429 ASSERT(zl->zl_parent == pio); 430 return (zl->zl_child); 431} 432 433zio_t * 434zio_unique_parent(zio_t *cio) 435{ 436 zio_t *pio = zio_walk_parents(cio); 437 438 VERIFY(zio_walk_parents(cio) == NULL); 439 return (pio); 440} 441 442void 443zio_add_child(zio_t *pio, zio_t *cio) 444{ 445 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP); 446 447 /* 448 * Logical I/Os can have logical, gang, or vdev children. 449 * Gang I/Os can have gang or vdev children. 450 * Vdev I/Os can only have vdev children. 451 * The following ASSERT captures all of these constraints. 452 */ 453 ASSERT(cio->io_child_type <= pio->io_child_type); 454 455 zl->zl_parent = pio; 456 zl->zl_child = cio; 457 458 mutex_enter(&cio->io_lock); 459 mutex_enter(&pio->io_lock); 460 461 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0); 462 463 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 464 pio->io_children[cio->io_child_type][w] += !cio->io_state[w]; 465 466 list_insert_head(&pio->io_child_list, zl); 467 list_insert_head(&cio->io_parent_list, zl); 468 469 pio->io_child_count++; 470 cio->io_parent_count++; 471 472 mutex_exit(&pio->io_lock); 473 mutex_exit(&cio->io_lock); 474} 475 476static void 477zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl) 478{ 479 ASSERT(zl->zl_parent == pio); 480 ASSERT(zl->zl_child == cio); 481 482 mutex_enter(&cio->io_lock); 483 mutex_enter(&pio->io_lock); 484 485 list_remove(&pio->io_child_list, zl); 486 list_remove(&cio->io_parent_list, zl); 487 488 pio->io_child_count--; 489 cio->io_parent_count--; 490 491 mutex_exit(&pio->io_lock); 492 mutex_exit(&cio->io_lock); 493 494 kmem_cache_free(zio_link_cache, zl); 495} 496 497static boolean_t 498zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait) 499{ 500 uint64_t *countp = &zio->io_children[child][wait]; 501 boolean_t waiting = B_FALSE; 502 503 mutex_enter(&zio->io_lock); 504 ASSERT(zio->io_stall == NULL); 505 if (*countp != 0) { 506 zio->io_stage >>= 1; 507 zio->io_stall = countp; 508 waiting = B_TRUE; 509 } 510 mutex_exit(&zio->io_lock); 511 512 return (waiting); 513} 514 515static void 516zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait) 517{ 518 uint64_t *countp = &pio->io_children[zio->io_child_type][wait]; 519 int *errorp = &pio->io_child_error[zio->io_child_type]; 520 521 mutex_enter(&pio->io_lock); 522 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE)) 523 *errorp = zio_worst_error(*errorp, zio->io_error); 524 pio->io_reexecute |= zio->io_reexecute; 525 ASSERT3U(*countp, >, 0); 526 527 (*countp)--; 528 529 if (*countp == 0 && pio->io_stall == countp) { 530 pio->io_stall = NULL; 531 mutex_exit(&pio->io_lock); 532 zio_execute(pio); 533 } else { 534 mutex_exit(&pio->io_lock); 535 } 536} 537 538static void 539zio_inherit_child_errors(zio_t *zio, enum zio_child c) 540{ 541 if (zio->io_child_error[c] != 0 && zio->io_error == 0) 542 zio->io_error = zio->io_child_error[c]; 543} 544 545/* 546 * ========================================================================== 547 * Create the various types of I/O (read, write, free, etc) 548 * ========================================================================== 549 */ 550static zio_t * 551zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 552 void *data, uint64_t size, zio_done_func_t *done, void *private, 553 zio_type_t type, zio_priority_t priority, enum zio_flag flags, 554 vdev_t *vd, uint64_t offset, const zbookmark_phys_t *zb, 555 enum zio_stage stage, enum zio_stage pipeline) 556{ 557 zio_t *zio; 558 559 ASSERT3U(type == ZIO_TYPE_FREE || size, <=, SPA_MAXBLOCKSIZE); 560 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0); 561 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0); 562 563 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER)); 564 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER)); 565 ASSERT(vd || stage == ZIO_STAGE_OPEN); 566 567 zio = kmem_cache_alloc(zio_cache, KM_SLEEP); 568 bzero(zio, sizeof (zio_t)); 569 570 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL); 571 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL); 572 573 list_create(&zio->io_parent_list, sizeof (zio_link_t), 574 offsetof(zio_link_t, zl_parent_node)); 575 list_create(&zio->io_child_list, sizeof (zio_link_t), 576 offsetof(zio_link_t, zl_child_node)); 577 578 if (vd != NULL) 579 zio->io_child_type = ZIO_CHILD_VDEV; 580 else if (flags & ZIO_FLAG_GANG_CHILD) 581 zio->io_child_type = ZIO_CHILD_GANG; 582 else if (flags & ZIO_FLAG_DDT_CHILD) 583 zio->io_child_type = ZIO_CHILD_DDT; 584 else 585 zio->io_child_type = ZIO_CHILD_LOGICAL; 586 587 if (bp != NULL) { 588 zio->io_bp = (blkptr_t *)bp; 589 zio->io_bp_copy = *bp; 590 zio->io_bp_orig = *bp; 591 if (type != ZIO_TYPE_WRITE || 592 zio->io_child_type == ZIO_CHILD_DDT) 593 zio->io_bp = &zio->io_bp_copy; /* so caller can free */ 594 if (zio->io_child_type == ZIO_CHILD_LOGICAL) 595 zio->io_logical = zio; 596 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp)) 597 pipeline |= ZIO_GANG_STAGES; 598 } 599 600 zio->io_spa = spa; 601 zio->io_txg = txg; 602 zio->io_done = done; 603 zio->io_private = private; 604 zio->io_type = type; 605 zio->io_priority = priority; 606 zio->io_vd = vd; 607 zio->io_offset = offset; 608 zio->io_orig_data = zio->io_data = data; 609 zio->io_orig_size = zio->io_size = size; 610 zio->io_orig_flags = zio->io_flags = flags; 611 zio->io_orig_stage = zio->io_stage = stage; 612 zio->io_orig_pipeline = zio->io_pipeline = pipeline; 613 614 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY); 615 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE); 616 617 if (zb != NULL) 618 zio->io_bookmark = *zb; 619 620 if (pio != NULL) { 621 if (zio->io_logical == NULL) 622 zio->io_logical = pio->io_logical; 623 if (zio->io_child_type == ZIO_CHILD_GANG) 624 zio->io_gang_leader = pio->io_gang_leader; 625 zio_add_child(pio, zio); 626 } 627 628 return (zio); 629} 630 631static void 632zio_destroy(zio_t *zio) 633{ 634 list_destroy(&zio->io_parent_list); 635 list_destroy(&zio->io_child_list); 636 mutex_destroy(&zio->io_lock); 637 cv_destroy(&zio->io_cv); 638 kmem_cache_free(zio_cache, zio); 639} 640 641zio_t * 642zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done, 643 void *private, enum zio_flag flags) 644{ 645 zio_t *zio; 646 647 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private, 648 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, 649 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE); 650 651 return (zio); 652} 653 654zio_t * 655zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags) 656{ 657 return (zio_null(NULL, spa, NULL, done, private, flags)); 658} 659 660zio_t * 661zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, 662 void *data, uint64_t size, zio_done_func_t *done, void *private, 663 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb) 664{ 665 zio_t *zio; 666 667 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp, 668 data, size, done, private, 669 ZIO_TYPE_READ, priority, flags, NULL, 0, zb, 670 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? 671 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE); 672 673 return (zio); 674} 675 676zio_t * 677zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, 678 void *data, uint64_t size, const zio_prop_t *zp, 679 zio_done_func_t *ready, zio_done_func_t *physdone, zio_done_func_t *done, 680 void *private, 681 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb) 682{ 683 zio_t *zio; 684 685 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF && 686 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS && 687 zp->zp_compress >= ZIO_COMPRESS_OFF && 688 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS && 689 DMU_OT_IS_VALID(zp->zp_type) && 690 zp->zp_level < 32 && 691 zp->zp_copies > 0 && 692 zp->zp_copies <= spa_max_replication(spa)); 693 694 zio = zio_create(pio, spa, txg, bp, data, size, done, private, 695 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, 696 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? 697 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE); 698 699 zio->io_ready = ready; 700 zio->io_physdone = physdone; 701 zio->io_prop = *zp; 702 703 /* 704 * Data can be NULL if we are going to call zio_write_override() to 705 * provide the already-allocated BP. But we may need the data to 706 * verify a dedup hit (if requested). In this case, don't try to 707 * dedup (just take the already-allocated BP verbatim). 708 */ 709 if (data == NULL && zio->io_prop.zp_dedup_verify) { 710 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE; 711 } 712 713 return (zio); 714} 715 716zio_t * 717zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data, 718 uint64_t size, zio_done_func_t *done, void *private, 719 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb) 720{ 721 zio_t *zio; 722 723 zio = zio_create(pio, spa, txg, bp, data, size, done, private, 724 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, 725 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE); 726 727 return (zio); 728} 729 730void 731zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite) 732{ 733 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 734 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 735 ASSERT(zio->io_stage == ZIO_STAGE_OPEN); 736 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa)); 737 738 /* 739 * We must reset the io_prop to match the values that existed 740 * when the bp was first written by dmu_sync() keeping in mind 741 * that nopwrite and dedup are mutually exclusive. 742 */ 743 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup; 744 zio->io_prop.zp_nopwrite = nopwrite; 745 zio->io_prop.zp_copies = copies; 746 zio->io_bp_override = bp; 747} 748 749void 750zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp) 751{ 752 753 /* 754 * The check for EMBEDDED is a performance optimization. We 755 * process the free here (by ignoring it) rather than 756 * putting it on the list and then processing it in zio_free_sync(). 757 */ 758 if (BP_IS_EMBEDDED(bp)) 759 return; 760 metaslab_check_free(spa, bp); 761 762 /* 763 * Frees that are for the currently-syncing txg, are not going to be 764 * deferred, and which will not need to do a read (i.e. not GANG or 765 * DEDUP), can be processed immediately. Otherwise, put them on the 766 * in-memory list for later processing. 767 */ 768 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) || 769 txg != spa->spa_syncing_txg || 770 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) { 771 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp); 772 } else { 773 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 774 BP_GET_PSIZE(bp), 0))); 775 } 776} 777 778zio_t * 779zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 780 uint64_t size, enum zio_flag flags) 781{ 782 zio_t *zio; 783 enum zio_stage stage = ZIO_FREE_PIPELINE; 784 785 ASSERT(!BP_IS_HOLE(bp)); 786 ASSERT(spa_syncing_txg(spa) == txg); 787 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free); 788 789 if (BP_IS_EMBEDDED(bp)) 790 return (zio_null(pio, spa, NULL, NULL, NULL, 0)); 791 792 metaslab_check_free(spa, bp); 793 arc_freed(spa, bp); 794 795 if (zfs_trim_enabled) 796 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START | 797 ZIO_STAGE_VDEV_IO_ASSESS; 798 /* 799 * GANG and DEDUP blocks can induce a read (for the gang block header, 800 * or the DDT), so issue them asynchronously so that this thread is 801 * not tied up. 802 */ 803 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp)) 804 stage |= ZIO_STAGE_ISSUE_ASYNC; 805 806 flags |= ZIO_FLAG_DONT_QUEUE; 807 808 zio = zio_create(pio, spa, txg, bp, NULL, size, 809 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags, 810 NULL, 0, NULL, ZIO_STAGE_OPEN, stage); 811 812 return (zio); 813} 814 815zio_t * 816zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 817 zio_done_func_t *done, void *private, enum zio_flag flags) 818{ 819 zio_t *zio; 820 821 dprintf_bp(bp, "claiming in txg %llu", txg); 822 823 if (BP_IS_EMBEDDED(bp)) 824 return (zio_null(pio, spa, NULL, NULL, NULL, 0)); 825 826 /* 827 * A claim is an allocation of a specific block. Claims are needed 828 * to support immediate writes in the intent log. The issue is that 829 * immediate writes contain committed data, but in a txg that was 830 * *not* committed. Upon opening the pool after an unclean shutdown, 831 * the intent log claims all blocks that contain immediate write data 832 * so that the SPA knows they're in use. 833 * 834 * All claims *must* be resolved in the first txg -- before the SPA 835 * starts allocating blocks -- so that nothing is allocated twice. 836 * If txg == 0 we just verify that the block is claimable. 837 */ 838 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa)); 839 ASSERT(txg == spa_first_txg(spa) || txg == 0); 840 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */ 841 842 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), 843 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags, 844 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE); 845 846 return (zio); 847} 848 849zio_t * 850zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset, 851 uint64_t size, zio_done_func_t *done, void *private, 852 zio_priority_t priority, enum zio_flag flags) 853{ 854 zio_t *zio; 855 int c; 856 857 if (vd->vdev_children == 0) { 858 zio = zio_create(pio, spa, 0, NULL, NULL, size, done, private, 859 ZIO_TYPE_IOCTL, priority, flags, vd, offset, NULL, 860 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE); 861 862 zio->io_cmd = cmd; 863 } else { 864 zio = zio_null(pio, spa, NULL, NULL, NULL, flags); 865 866 for (c = 0; c < vd->vdev_children; c++) 867 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd, 868 offset, size, done, private, priority, flags)); 869 } 870 871 return (zio); 872} 873 874zio_t * 875zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 876 void *data, int checksum, zio_done_func_t *done, void *private, 877 zio_priority_t priority, enum zio_flag flags, boolean_t labels) 878{ 879 zio_t *zio; 880 881 ASSERT(vd->vdev_children == 0); 882 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || 883 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); 884 ASSERT3U(offset + size, <=, vd->vdev_psize); 885 886 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private, 887 ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset, 888 NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE); 889 890 zio->io_prop.zp_checksum = checksum; 891 892 return (zio); 893} 894 895zio_t * 896zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 897 void *data, int checksum, zio_done_func_t *done, void *private, 898 zio_priority_t priority, enum zio_flag flags, boolean_t labels) 899{ 900 zio_t *zio; 901 902 ASSERT(vd->vdev_children == 0); 903 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || 904 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); 905 ASSERT3U(offset + size, <=, vd->vdev_psize); 906 907 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private, 908 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset, 909 NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE); 910 911 zio->io_prop.zp_checksum = checksum; 912 913 if (zio_checksum_table[checksum].ci_eck) { 914 /* 915 * zec checksums are necessarily destructive -- they modify 916 * the end of the write buffer to hold the verifier/checksum. 917 * Therefore, we must make a local copy in case the data is 918 * being written to multiple places in parallel. 919 */ 920 void *wbuf = zio_buf_alloc(size); 921 bcopy(data, wbuf, size); 922 zio_push_transform(zio, wbuf, size, size, NULL); 923 } 924 925 return (zio); 926} 927 928/* 929 * Create a child I/O to do some work for us. 930 */ 931zio_t * 932zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset, 933 void *data, uint64_t size, int type, zio_priority_t priority, 934 enum zio_flag flags, zio_done_func_t *done, void *private) 935{ 936 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE; 937 zio_t *zio; 938 939 ASSERT(vd->vdev_parent == 940 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev)); 941 942 if (type == ZIO_TYPE_READ && bp != NULL) { 943 /* 944 * If we have the bp, then the child should perform the 945 * checksum and the parent need not. This pushes error 946 * detection as close to the leaves as possible and 947 * eliminates redundant checksums in the interior nodes. 948 */ 949 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY; 950 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; 951 } 952 953 /* Not all IO types require vdev io done stage e.g. free */ 954 if (!(pio->io_pipeline & ZIO_STAGE_VDEV_IO_DONE)) 955 pipeline &= ~ZIO_STAGE_VDEV_IO_DONE; 956 957 if (vd->vdev_children == 0) 958 offset += VDEV_LABEL_START_SIZE; 959 960 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE; 961 962 /* 963 * If we've decided to do a repair, the write is not speculative -- 964 * even if the original read was. 965 */ 966 if (flags & ZIO_FLAG_IO_REPAIR) 967 flags &= ~ZIO_FLAG_SPECULATIVE; 968 969 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, 970 done, private, type, priority, flags, vd, offset, &pio->io_bookmark, 971 ZIO_STAGE_VDEV_IO_START >> 1, pipeline); 972 973 zio->io_physdone = pio->io_physdone; 974 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL) 975 zio->io_logical->io_phys_children++; 976 977 return (zio); 978} 979 980zio_t * 981zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size, 982 int type, zio_priority_t priority, enum zio_flag flags, 983 zio_done_func_t *done, void *private) 984{ 985 zio_t *zio; 986 987 ASSERT(vd->vdev_ops->vdev_op_leaf); 988 989 zio = zio_create(NULL, vd->vdev_spa, 0, NULL, 990 data, size, done, private, type, priority, 991 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED, 992 vd, offset, NULL, 993 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE); 994 995 return (zio); 996} 997 998void 999zio_flush(zio_t *zio, vdev_t *vd) 1000{ 1001 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0, 1002 NULL, NULL, ZIO_PRIORITY_NOW, 1003 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY)); 1004} 1005 1006zio_t * 1007zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size) 1008{ 1009 1010 ASSERT(vd->vdev_ops->vdev_op_leaf); 1011 1012 return (zio_create(zio, spa, 0, NULL, NULL, size, NULL, NULL, 1013 ZIO_TYPE_FREE, ZIO_PRIORITY_TRIM, ZIO_FLAG_DONT_AGGREGATE | 1014 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY, 1015 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PHYS_PIPELINE)); 1016} 1017 1018void 1019zio_shrink(zio_t *zio, uint64_t size) 1020{ 1021 ASSERT(zio->io_executor == NULL); 1022 ASSERT(zio->io_orig_size == zio->io_size); 1023 ASSERT(size <= zio->io_size); 1024 1025 /* 1026 * We don't shrink for raidz because of problems with the 1027 * reconstruction when reading back less than the block size. 1028 * Note, BP_IS_RAIDZ() assumes no compression. 1029 */ 1030 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 1031 if (!BP_IS_RAIDZ(zio->io_bp)) 1032 zio->io_orig_size = zio->io_size = size; 1033} 1034 1035/* 1036 * ========================================================================== 1037 * Prepare to read and write logical blocks 1038 * ========================================================================== 1039 */ 1040 1041static int 1042zio_read_bp_init(zio_t *zio) 1043{ 1044 blkptr_t *bp = zio->io_bp; 1045 1046 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF && 1047 zio->io_child_type == ZIO_CHILD_LOGICAL && 1048 !(zio->io_flags & ZIO_FLAG_RAW)) { 1049 uint64_t psize = 1050 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp); 1051 void *cbuf = zio_buf_alloc(psize); 1052 1053 zio_push_transform(zio, cbuf, psize, psize, zio_decompress); 1054 } 1055 1056 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) { 1057 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1058 decode_embedded_bp_compressed(bp, zio->io_data); 1059 } else { 1060 ASSERT(!BP_IS_EMBEDDED(bp)); 1061 } 1062 1063 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0) 1064 zio->io_flags |= ZIO_FLAG_DONT_CACHE; 1065 1066 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP) 1067 zio->io_flags |= ZIO_FLAG_DONT_CACHE; 1068 1069 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL) 1070 zio->io_pipeline = ZIO_DDT_READ_PIPELINE; 1071 1072 return (ZIO_PIPELINE_CONTINUE); 1073} 1074 1075static int 1076zio_write_bp_init(zio_t *zio) 1077{ 1078 spa_t *spa = zio->io_spa; 1079 zio_prop_t *zp = &zio->io_prop; 1080 enum zio_compress compress = zp->zp_compress; 1081 blkptr_t *bp = zio->io_bp; 1082 uint64_t lsize = zio->io_size; 1083 uint64_t psize = lsize; 1084 int pass = 1; 1085 1086 /* 1087 * If our children haven't all reached the ready stage, 1088 * wait for them and then repeat this pipeline stage. 1089 */ 1090 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) || 1091 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY)) 1092 return (ZIO_PIPELINE_STOP); 1093 1094 if (!IO_IS_ALLOCATING(zio)) 1095 return (ZIO_PIPELINE_CONTINUE); 1096 1097 ASSERT(zio->io_child_type != ZIO_CHILD_DDT); 1098 1099 if (zio->io_bp_override) { 1100 ASSERT(bp->blk_birth != zio->io_txg); 1101 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0); 1102 1103 *bp = *zio->io_bp_override; 1104 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1105 1106 if (BP_IS_EMBEDDED(bp)) 1107 return (ZIO_PIPELINE_CONTINUE); 1108 1109 /* 1110 * If we've been overridden and nopwrite is set then 1111 * set the flag accordingly to indicate that a nopwrite 1112 * has already occurred. 1113 */ 1114 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) { 1115 ASSERT(!zp->zp_dedup); 1116 zio->io_flags |= ZIO_FLAG_NOPWRITE; 1117 return (ZIO_PIPELINE_CONTINUE); 1118 } 1119 1120 ASSERT(!zp->zp_nopwrite); 1121 1122 if (BP_IS_HOLE(bp) || !zp->zp_dedup) 1123 return (ZIO_PIPELINE_CONTINUE); 1124 1125 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup || 1126 zp->zp_dedup_verify); 1127 1128 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) { 1129 BP_SET_DEDUP(bp, 1); 1130 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE; 1131 return (ZIO_PIPELINE_CONTINUE); 1132 } 1133 zio->io_bp_override = NULL; 1134 BP_ZERO(bp); 1135 } 1136 1137 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) { 1138 /* 1139 * We're rewriting an existing block, which means we're 1140 * working on behalf of spa_sync(). For spa_sync() to 1141 * converge, it must eventually be the case that we don't 1142 * have to allocate new blocks. But compression changes 1143 * the blocksize, which forces a reallocate, and makes 1144 * convergence take longer. Therefore, after the first 1145 * few passes, stop compressing to ensure convergence. 1146 */ 1147 pass = spa_sync_pass(spa); 1148 1149 ASSERT(zio->io_txg == spa_syncing_txg(spa)); 1150 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1151 ASSERT(!BP_GET_DEDUP(bp)); 1152 1153 if (pass >= zfs_sync_pass_dont_compress) 1154 compress = ZIO_COMPRESS_OFF; 1155 1156 /* Make sure someone doesn't change their mind on overwrites */ 1157 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp), 1158 spa_max_replication(spa)) == BP_GET_NDVAS(bp)); 1159 } 1160 1161 if (compress != ZIO_COMPRESS_OFF) { 1162 void *cbuf = zio_buf_alloc(lsize); 1163 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize); 1164 if (psize == 0 || psize == lsize) { 1165 compress = ZIO_COMPRESS_OFF; 1166 zio_buf_free(cbuf, lsize); 1167 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE && 1168 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) && 1169 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) { 1170 encode_embedded_bp_compressed(bp, 1171 cbuf, compress, lsize, psize); 1172 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA); 1173 BP_SET_TYPE(bp, zio->io_prop.zp_type); 1174 BP_SET_LEVEL(bp, zio->io_prop.zp_level); 1175 zio_buf_free(cbuf, lsize); 1176 bp->blk_birth = zio->io_txg; 1177 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1178 ASSERT(spa_feature_is_active(spa, 1179 SPA_FEATURE_EMBEDDED_DATA)); 1180 return (ZIO_PIPELINE_CONTINUE); 1181 } else { 1182 /* 1183 * Round up compressed size to MINBLOCKSIZE and 1184 * zero the tail. 1185 */ 1186 size_t rounded = 1187 P2ROUNDUP(psize, (size_t)SPA_MINBLOCKSIZE); 1188 if (rounded > psize) { 1189 bzero((char *)cbuf + psize, rounded - psize); 1190 psize = rounded; 1191 } 1192 if (psize == lsize) { 1193 compress = ZIO_COMPRESS_OFF; 1194 zio_buf_free(cbuf, lsize); 1195 } else { 1196 zio_push_transform(zio, cbuf, 1197 psize, lsize, NULL); 1198 } 1199 } 1200 } 1201 1202 /* 1203 * The final pass of spa_sync() must be all rewrites, but the first 1204 * few passes offer a trade-off: allocating blocks defers convergence, 1205 * but newly allocated blocks are sequential, so they can be written 1206 * to disk faster. Therefore, we allow the first few passes of 1207 * spa_sync() to allocate new blocks, but force rewrites after that. 1208 * There should only be a handful of blocks after pass 1 in any case. 1209 */ 1210 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg && 1211 BP_GET_PSIZE(bp) == psize && 1212 pass >= zfs_sync_pass_rewrite) { 1213 ASSERT(psize != 0); 1214 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES; 1215 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages; 1216 zio->io_flags |= ZIO_FLAG_IO_REWRITE; 1217 } else { 1218 BP_ZERO(bp); 1219 zio->io_pipeline = ZIO_WRITE_PIPELINE; 1220 } 1221 1222 if (psize == 0) { 1223 if (zio->io_bp_orig.blk_birth != 0 && 1224 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) { 1225 BP_SET_LSIZE(bp, lsize); 1226 BP_SET_TYPE(bp, zp->zp_type); 1227 BP_SET_LEVEL(bp, zp->zp_level); 1228 BP_SET_BIRTH(bp, zio->io_txg, 0); 1229 } 1230 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1231 } else { 1232 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER); 1233 BP_SET_LSIZE(bp, lsize); 1234 BP_SET_TYPE(bp, zp->zp_type); 1235 BP_SET_LEVEL(bp, zp->zp_level); 1236 BP_SET_PSIZE(bp, psize); 1237 BP_SET_COMPRESS(bp, compress); 1238 BP_SET_CHECKSUM(bp, zp->zp_checksum); 1239 BP_SET_DEDUP(bp, zp->zp_dedup); 1240 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER); 1241 if (zp->zp_dedup) { 1242 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1243 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 1244 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE; 1245 } 1246 if (zp->zp_nopwrite) { 1247 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1248 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 1249 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE; 1250 } 1251 } 1252 1253 return (ZIO_PIPELINE_CONTINUE); 1254} 1255 1256static int 1257zio_free_bp_init(zio_t *zio) 1258{ 1259 blkptr_t *bp = zio->io_bp; 1260 1261 if (zio->io_child_type == ZIO_CHILD_LOGICAL) { 1262 if (BP_GET_DEDUP(bp)) 1263 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE; 1264 } 1265 1266 return (ZIO_PIPELINE_CONTINUE); 1267} 1268 1269/* 1270 * ========================================================================== 1271 * Execute the I/O pipeline 1272 * ========================================================================== 1273 */ 1274 1275static void 1276zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline) 1277{ 1278 spa_t *spa = zio->io_spa; 1279 zio_type_t t = zio->io_type; 1280 int flags = (cutinline ? TQ_FRONT : 0); 1281 1282 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT); 1283 1284 /* 1285 * If we're a config writer or a probe, the normal issue and 1286 * interrupt threads may all be blocked waiting for the config lock. 1287 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL. 1288 */ 1289 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE)) 1290 t = ZIO_TYPE_NULL; 1291 1292 /* 1293 * A similar issue exists for the L2ARC write thread until L2ARC 2.0. 1294 */ 1295 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux) 1296 t = ZIO_TYPE_NULL; 1297 1298 /* 1299 * If this is a high priority I/O, then use the high priority taskq if 1300 * available. 1301 */ 1302 if (zio->io_priority == ZIO_PRIORITY_NOW && 1303 spa->spa_zio_taskq[t][q + 1].stqs_count != 0) 1304 q++; 1305 1306 ASSERT3U(q, <, ZIO_TASKQ_TYPES); 1307 1308 /* 1309 * NB: We are assuming that the zio can only be dispatched 1310 * to a single taskq at a time. It would be a grievous error 1311 * to dispatch the zio to another taskq at the same time. 1312 */ 1313#if defined(illumos) || !defined(_KERNEL) 1314 ASSERT(zio->io_tqent.tqent_next == NULL); 1315#else 1316 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0); 1317#endif 1318 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio, 1319 flags, &zio->io_tqent); 1320} 1321 1322static boolean_t 1323zio_taskq_member(zio_t *zio, zio_taskq_type_t q) 1324{ 1325 kthread_t *executor = zio->io_executor; 1326 spa_t *spa = zio->io_spa; 1327 1328 for (zio_type_t t = 0; t < ZIO_TYPES; t++) { 1329 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q]; 1330 uint_t i; 1331 for (i = 0; i < tqs->stqs_count; i++) { 1332 if (taskq_member(tqs->stqs_taskq[i], executor)) 1333 return (B_TRUE); 1334 } 1335 } 1336 1337 return (B_FALSE); 1338} 1339 1340static int 1341zio_issue_async(zio_t *zio) 1342{ 1343 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); 1344 1345 return (ZIO_PIPELINE_STOP); 1346} 1347 1348void 1349zio_interrupt(zio_t *zio) 1350{ 1351 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE); 1352} 1353 1354/* 1355 * Execute the I/O pipeline until one of the following occurs: 1356 * 1357 * (1) the I/O completes 1358 * (2) the pipeline stalls waiting for dependent child I/Os 1359 * (3) the I/O issues, so we're waiting for an I/O completion interrupt 1360 * (4) the I/O is delegated by vdev-level caching or aggregation 1361 * (5) the I/O is deferred due to vdev-level queueing 1362 * (6) the I/O is handed off to another thread. 1363 * 1364 * In all cases, the pipeline stops whenever there's no CPU work; it never 1365 * burns a thread in cv_wait(). 1366 * 1367 * There's no locking on io_stage because there's no legitimate way 1368 * for multiple threads to be attempting to process the same I/O. 1369 */ 1370static zio_pipe_stage_t *zio_pipeline[]; 1371 1372void 1373zio_execute(zio_t *zio) 1374{ 1375 zio->io_executor = curthread; 1376 1377 while (zio->io_stage < ZIO_STAGE_DONE) { 1378 enum zio_stage pipeline = zio->io_pipeline; 1379 enum zio_stage stage = zio->io_stage; 1380 int rv; 1381 1382 ASSERT(!MUTEX_HELD(&zio->io_lock)); 1383 ASSERT(ISP2(stage)); 1384 ASSERT(zio->io_stall == NULL); 1385 1386 do { 1387 stage <<= 1; 1388 } while ((stage & pipeline) == 0); 1389 1390 ASSERT(stage <= ZIO_STAGE_DONE); 1391 1392 /* 1393 * If we are in interrupt context and this pipeline stage 1394 * will grab a config lock that is held across I/O, 1395 * or may wait for an I/O that needs an interrupt thread 1396 * to complete, issue async to avoid deadlock. 1397 * 1398 * For VDEV_IO_START, we cut in line so that the io will 1399 * be sent to disk promptly. 1400 */ 1401 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL && 1402 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) { 1403 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? 1404 zio_requeue_io_start_cut_in_line : B_FALSE; 1405 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); 1406 return; 1407 } 1408 1409 zio->io_stage = stage; 1410 rv = zio_pipeline[highbit64(stage) - 1](zio); 1411 1412 if (rv == ZIO_PIPELINE_STOP) 1413 return; 1414 1415 ASSERT(rv == ZIO_PIPELINE_CONTINUE); 1416 } 1417} 1418 1419/* 1420 * ========================================================================== 1421 * Initiate I/O, either sync or async 1422 * ========================================================================== 1423 */ 1424int 1425zio_wait(zio_t *zio) 1426{ 1427 int error; 1428 1429 ASSERT(zio->io_stage == ZIO_STAGE_OPEN); 1430 ASSERT(zio->io_executor == NULL); 1431 1432 zio->io_waiter = curthread; 1433 1434 zio_execute(zio); 1435 1436 mutex_enter(&zio->io_lock); 1437 while (zio->io_executor != NULL) 1438 cv_wait(&zio->io_cv, &zio->io_lock); 1439 mutex_exit(&zio->io_lock); 1440 1441 error = zio->io_error; 1442 zio_destroy(zio); 1443 1444 return (error); 1445} 1446 1447void 1448zio_nowait(zio_t *zio) 1449{ 1450 ASSERT(zio->io_executor == NULL); 1451 1452 if (zio->io_child_type == ZIO_CHILD_LOGICAL && 1453 zio_unique_parent(zio) == NULL) { 1454 /* 1455 * This is a logical async I/O with no parent to wait for it. 1456 * We add it to the spa_async_root_zio "Godfather" I/O which 1457 * will ensure they complete prior to unloading the pool. 1458 */ 1459 spa_t *spa = zio->io_spa; 1460 1461 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio); 1462 } 1463 1464 zio_execute(zio); 1465} 1466 1467/* 1468 * ========================================================================== 1469 * Reexecute or suspend/resume failed I/O 1470 * ========================================================================== 1471 */ 1472 1473static void 1474zio_reexecute(zio_t *pio) 1475{ 1476 zio_t *cio, *cio_next; 1477 1478 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL); 1479 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN); 1480 ASSERT(pio->io_gang_leader == NULL); 1481 ASSERT(pio->io_gang_tree == NULL); 1482 1483 pio->io_flags = pio->io_orig_flags; 1484 pio->io_stage = pio->io_orig_stage; 1485 pio->io_pipeline = pio->io_orig_pipeline; 1486 pio->io_reexecute = 0; 1487 pio->io_flags |= ZIO_FLAG_REEXECUTED; 1488 pio->io_error = 0; 1489 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 1490 pio->io_state[w] = 0; 1491 for (int c = 0; c < ZIO_CHILD_TYPES; c++) 1492 pio->io_child_error[c] = 0; 1493 1494 if (IO_IS_ALLOCATING(pio)) 1495 BP_ZERO(pio->io_bp); 1496 1497 /* 1498 * As we reexecute pio's children, new children could be created. 1499 * New children go to the head of pio's io_child_list, however, 1500 * so we will (correctly) not reexecute them. The key is that 1501 * the remainder of pio's io_child_list, from 'cio_next' onward, 1502 * cannot be affected by any side effects of reexecuting 'cio'. 1503 */ 1504 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) { 1505 cio_next = zio_walk_children(pio); 1506 mutex_enter(&pio->io_lock); 1507 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 1508 pio->io_children[cio->io_child_type][w]++; 1509 mutex_exit(&pio->io_lock); 1510 zio_reexecute(cio); 1511 } 1512 1513 /* 1514 * Now that all children have been reexecuted, execute the parent. 1515 * We don't reexecute "The Godfather" I/O here as it's the 1516 * responsibility of the caller to wait on him. 1517 */ 1518 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) 1519 zio_execute(pio); 1520} 1521 1522void 1523zio_suspend(spa_t *spa, zio_t *zio) 1524{ 1525 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC) 1526 fm_panic("Pool '%s' has encountered an uncorrectable I/O " 1527 "failure and the failure mode property for this pool " 1528 "is set to panic.", spa_name(spa)); 1529 1530 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0); 1531 1532 mutex_enter(&spa->spa_suspend_lock); 1533 1534 if (spa->spa_suspend_zio_root == NULL) 1535 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL, 1536 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | 1537 ZIO_FLAG_GODFATHER); 1538 1539 spa->spa_suspended = B_TRUE; 1540 1541 if (zio != NULL) { 1542 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); 1543 ASSERT(zio != spa->spa_suspend_zio_root); 1544 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1545 ASSERT(zio_unique_parent(zio) == NULL); 1546 ASSERT(zio->io_stage == ZIO_STAGE_DONE); 1547 zio_add_child(spa->spa_suspend_zio_root, zio); 1548 } 1549 1550 mutex_exit(&spa->spa_suspend_lock); 1551} 1552 1553int 1554zio_resume(spa_t *spa) 1555{ 1556 zio_t *pio; 1557 1558 /* 1559 * Reexecute all previously suspended i/o. 1560 */ 1561 mutex_enter(&spa->spa_suspend_lock); 1562 spa->spa_suspended = B_FALSE; 1563 cv_broadcast(&spa->spa_suspend_cv); 1564 pio = spa->spa_suspend_zio_root; 1565 spa->spa_suspend_zio_root = NULL; 1566 mutex_exit(&spa->spa_suspend_lock); 1567 1568 if (pio == NULL) 1569 return (0); 1570 1571 zio_reexecute(pio); 1572 return (zio_wait(pio)); 1573} 1574 1575void 1576zio_resume_wait(spa_t *spa) 1577{ 1578 mutex_enter(&spa->spa_suspend_lock); 1579 while (spa_suspended(spa)) 1580 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock); 1581 mutex_exit(&spa->spa_suspend_lock); 1582} 1583 1584/* 1585 * ========================================================================== 1586 * Gang blocks. 1587 * 1588 * A gang block is a collection of small blocks that looks to the DMU 1589 * like one large block. When zio_dva_allocate() cannot find a block 1590 * of the requested size, due to either severe fragmentation or the pool 1591 * being nearly full, it calls zio_write_gang_block() to construct the 1592 * block from smaller fragments. 1593 * 1594 * A gang block consists of a gang header (zio_gbh_phys_t) and up to 1595 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like 1596 * an indirect block: it's an array of block pointers. It consumes 1597 * only one sector and hence is allocatable regardless of fragmentation. 1598 * The gang header's bps point to its gang members, which hold the data. 1599 * 1600 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg> 1601 * as the verifier to ensure uniqueness of the SHA256 checksum. 1602 * Critically, the gang block bp's blk_cksum is the checksum of the data, 1603 * not the gang header. This ensures that data block signatures (needed for 1604 * deduplication) are independent of how the block is physically stored. 1605 * 1606 * Gang blocks can be nested: a gang member may itself be a gang block. 1607 * Thus every gang block is a tree in which root and all interior nodes are 1608 * gang headers, and the leaves are normal blocks that contain user data. 1609 * The root of the gang tree is called the gang leader. 1610 * 1611 * To perform any operation (read, rewrite, free, claim) on a gang block, 1612 * zio_gang_assemble() first assembles the gang tree (minus data leaves) 1613 * in the io_gang_tree field of the original logical i/o by recursively 1614 * reading the gang leader and all gang headers below it. This yields 1615 * an in-core tree containing the contents of every gang header and the 1616 * bps for every constituent of the gang block. 1617 * 1618 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree 1619 * and invokes a callback on each bp. To free a gang block, zio_gang_issue() 1620 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp. 1621 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim(). 1622 * zio_read_gang() is a wrapper around zio_read() that omits reading gang 1623 * headers, since we already have those in io_gang_tree. zio_rewrite_gang() 1624 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite() 1625 * of the gang header plus zio_checksum_compute() of the data to update the 1626 * gang header's blk_cksum as described above. 1627 * 1628 * The two-phase assemble/issue model solves the problem of partial failure -- 1629 * what if you'd freed part of a gang block but then couldn't read the 1630 * gang header for another part? Assembling the entire gang tree first 1631 * ensures that all the necessary gang header I/O has succeeded before 1632 * starting the actual work of free, claim, or write. Once the gang tree 1633 * is assembled, free and claim are in-memory operations that cannot fail. 1634 * 1635 * In the event that a gang write fails, zio_dva_unallocate() walks the 1636 * gang tree to immediately free (i.e. insert back into the space map) 1637 * everything we've allocated. This ensures that we don't get ENOSPC 1638 * errors during repeated suspend/resume cycles due to a flaky device. 1639 * 1640 * Gang rewrites only happen during sync-to-convergence. If we can't assemble 1641 * the gang tree, we won't modify the block, so we can safely defer the free 1642 * (knowing that the block is still intact). If we *can* assemble the gang 1643 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free 1644 * each constituent bp and we can allocate a new block on the next sync pass. 1645 * 1646 * In all cases, the gang tree allows complete recovery from partial failure. 1647 * ========================================================================== 1648 */ 1649 1650static zio_t * 1651zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1652{ 1653 if (gn != NULL) 1654 return (pio); 1655 1656 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp), 1657 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), 1658 &pio->io_bookmark)); 1659} 1660 1661zio_t * 1662zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1663{ 1664 zio_t *zio; 1665 1666 if (gn != NULL) { 1667 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, 1668 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority, 1669 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 1670 /* 1671 * As we rewrite each gang header, the pipeline will compute 1672 * a new gang block header checksum for it; but no one will 1673 * compute a new data checksum, so we do that here. The one 1674 * exception is the gang leader: the pipeline already computed 1675 * its data checksum because that stage precedes gang assembly. 1676 * (Presently, nothing actually uses interior data checksums; 1677 * this is just good hygiene.) 1678 */ 1679 if (gn != pio->io_gang_leader->io_gang_tree) { 1680 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp), 1681 data, BP_GET_PSIZE(bp)); 1682 } 1683 /* 1684 * If we are here to damage data for testing purposes, 1685 * leave the GBH alone so that we can detect the damage. 1686 */ 1687 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE) 1688 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; 1689 } else { 1690 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, 1691 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority, 1692 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 1693 } 1694 1695 return (zio); 1696} 1697 1698/* ARGSUSED */ 1699zio_t * 1700zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1701{ 1702 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp, 1703 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp), 1704 ZIO_GANG_CHILD_FLAGS(pio))); 1705} 1706 1707/* ARGSUSED */ 1708zio_t * 1709zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1710{ 1711 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp, 1712 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio))); 1713} 1714 1715static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = { 1716 NULL, 1717 zio_read_gang, 1718 zio_rewrite_gang, 1719 zio_free_gang, 1720 zio_claim_gang, 1721 NULL 1722}; 1723 1724static void zio_gang_tree_assemble_done(zio_t *zio); 1725 1726static zio_gang_node_t * 1727zio_gang_node_alloc(zio_gang_node_t **gnpp) 1728{ 1729 zio_gang_node_t *gn; 1730 1731 ASSERT(*gnpp == NULL); 1732 1733 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP); 1734 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE); 1735 *gnpp = gn; 1736 1737 return (gn); 1738} 1739 1740static void 1741zio_gang_node_free(zio_gang_node_t **gnpp) 1742{ 1743 zio_gang_node_t *gn = *gnpp; 1744 1745 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) 1746 ASSERT(gn->gn_child[g] == NULL); 1747 1748 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE); 1749 kmem_free(gn, sizeof (*gn)); 1750 *gnpp = NULL; 1751} 1752 1753static void 1754zio_gang_tree_free(zio_gang_node_t **gnpp) 1755{ 1756 zio_gang_node_t *gn = *gnpp; 1757 1758 if (gn == NULL) 1759 return; 1760 1761 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) 1762 zio_gang_tree_free(&gn->gn_child[g]); 1763 1764 zio_gang_node_free(gnpp); 1765} 1766 1767static void 1768zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp) 1769{ 1770 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp); 1771 1772 ASSERT(gio->io_gang_leader == gio); 1773 ASSERT(BP_IS_GANG(bp)); 1774 1775 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh, 1776 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn, 1777 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark)); 1778} 1779 1780static void 1781zio_gang_tree_assemble_done(zio_t *zio) 1782{ 1783 zio_t *gio = zio->io_gang_leader; 1784 zio_gang_node_t *gn = zio->io_private; 1785 blkptr_t *bp = zio->io_bp; 1786 1787 ASSERT(gio == zio_unique_parent(zio)); 1788 ASSERT(zio->io_child_count == 0); 1789 1790 if (zio->io_error) 1791 return; 1792 1793 if (BP_SHOULD_BYTESWAP(bp)) 1794 byteswap_uint64_array(zio->io_data, zio->io_size); 1795 1796 ASSERT(zio->io_data == gn->gn_gbh); 1797 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE); 1798 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); 1799 1800 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 1801 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; 1802 if (!BP_IS_GANG(gbp)) 1803 continue; 1804 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]); 1805 } 1806} 1807 1808static void 1809zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data) 1810{ 1811 zio_t *gio = pio->io_gang_leader; 1812 zio_t *zio; 1813 1814 ASSERT(BP_IS_GANG(bp) == !!gn); 1815 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp)); 1816 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree); 1817 1818 /* 1819 * If you're a gang header, your data is in gn->gn_gbh. 1820 * If you're a gang member, your data is in 'data' and gn == NULL. 1821 */ 1822 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data); 1823 1824 if (gn != NULL) { 1825 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); 1826 1827 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 1828 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; 1829 if (BP_IS_HOLE(gbp)) 1830 continue; 1831 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data); 1832 data = (char *)data + BP_GET_PSIZE(gbp); 1833 } 1834 } 1835 1836 if (gn == gio->io_gang_tree && gio->io_data != NULL) 1837 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data); 1838 1839 if (zio != pio) 1840 zio_nowait(zio); 1841} 1842 1843static int 1844zio_gang_assemble(zio_t *zio) 1845{ 1846 blkptr_t *bp = zio->io_bp; 1847 1848 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL); 1849 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 1850 1851 zio->io_gang_leader = zio; 1852 1853 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree); 1854 1855 return (ZIO_PIPELINE_CONTINUE); 1856} 1857 1858static int 1859zio_gang_issue(zio_t *zio) 1860{ 1861 blkptr_t *bp = zio->io_bp; 1862 1863 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE)) 1864 return (ZIO_PIPELINE_STOP); 1865 1866 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio); 1867 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 1868 1869 if (zio->io_child_error[ZIO_CHILD_GANG] == 0) 1870 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data); 1871 else 1872 zio_gang_tree_free(&zio->io_gang_tree); 1873 1874 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1875 1876 return (ZIO_PIPELINE_CONTINUE); 1877} 1878 1879static void 1880zio_write_gang_member_ready(zio_t *zio) 1881{ 1882 zio_t *pio = zio_unique_parent(zio); 1883 zio_t *gio = zio->io_gang_leader; 1884 dva_t *cdva = zio->io_bp->blk_dva; 1885 dva_t *pdva = pio->io_bp->blk_dva; 1886 uint64_t asize; 1887 1888 if (BP_IS_HOLE(zio->io_bp)) 1889 return; 1890 1891 ASSERT(BP_IS_HOLE(&zio->io_bp_orig)); 1892 1893 ASSERT(zio->io_child_type == ZIO_CHILD_GANG); 1894 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies); 1895 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp)); 1896 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp)); 1897 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp)); 1898 1899 mutex_enter(&pio->io_lock); 1900 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) { 1901 ASSERT(DVA_GET_GANG(&pdva[d])); 1902 asize = DVA_GET_ASIZE(&pdva[d]); 1903 asize += DVA_GET_ASIZE(&cdva[d]); 1904 DVA_SET_ASIZE(&pdva[d], asize); 1905 } 1906 mutex_exit(&pio->io_lock); 1907} 1908 1909static int 1910zio_write_gang_block(zio_t *pio) 1911{ 1912 spa_t *spa = pio->io_spa; 1913 blkptr_t *bp = pio->io_bp; 1914 zio_t *gio = pio->io_gang_leader; 1915 zio_t *zio; 1916 zio_gang_node_t *gn, **gnpp; 1917 zio_gbh_phys_t *gbh; 1918 uint64_t txg = pio->io_txg; 1919 uint64_t resid = pio->io_size; 1920 uint64_t lsize; 1921 int copies = gio->io_prop.zp_copies; 1922 int gbh_copies = MIN(copies + 1, spa_max_replication(spa)); 1923 zio_prop_t zp; 1924 int error; 1925 1926 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE, 1927 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, 1928 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER); 1929 if (error) { 1930 pio->io_error = error; 1931 return (ZIO_PIPELINE_CONTINUE); 1932 } 1933 1934 if (pio == gio) { 1935 gnpp = &gio->io_gang_tree; 1936 } else { 1937 gnpp = pio->io_private; 1938 ASSERT(pio->io_ready == zio_write_gang_member_ready); 1939 } 1940 1941 gn = zio_gang_node_alloc(gnpp); 1942 gbh = gn->gn_gbh; 1943 bzero(gbh, SPA_GANGBLOCKSIZE); 1944 1945 /* 1946 * Create the gang header. 1947 */ 1948 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL, 1949 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 1950 1951 /* 1952 * Create and nowait the gang children. 1953 */ 1954 for (int g = 0; resid != 0; resid -= lsize, g++) { 1955 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g), 1956 SPA_MINBLOCKSIZE); 1957 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid); 1958 1959 zp.zp_checksum = gio->io_prop.zp_checksum; 1960 zp.zp_compress = ZIO_COMPRESS_OFF; 1961 zp.zp_type = DMU_OT_NONE; 1962 zp.zp_level = 0; 1963 zp.zp_copies = gio->io_prop.zp_copies; 1964 zp.zp_dedup = B_FALSE; 1965 zp.zp_dedup_verify = B_FALSE; 1966 zp.zp_nopwrite = B_FALSE; 1967 1968 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g], 1969 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp, 1970 zio_write_gang_member_ready, NULL, NULL, &gn->gn_child[g], 1971 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), 1972 &pio->io_bookmark)); 1973 } 1974 1975 /* 1976 * Set pio's pipeline to just wait for zio to finish. 1977 */ 1978 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1979 1980 zio_nowait(zio); 1981 1982 return (ZIO_PIPELINE_CONTINUE); 1983} 1984 1985/* 1986 * The zio_nop_write stage in the pipeline determines if allocating 1987 * a new bp is necessary. By leveraging a cryptographically secure checksum, 1988 * such as SHA256, we can compare the checksums of the new data and the old 1989 * to determine if allocating a new block is required. The nopwrite 1990 * feature can handle writes in either syncing or open context (i.e. zil 1991 * writes) and as a result is mutually exclusive with dedup. 1992 */ 1993static int 1994zio_nop_write(zio_t *zio) 1995{ 1996 blkptr_t *bp = zio->io_bp; 1997 blkptr_t *bp_orig = &zio->io_bp_orig; 1998 zio_prop_t *zp = &zio->io_prop; 1999 2000 ASSERT(BP_GET_LEVEL(bp) == 0); 2001 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 2002 ASSERT(zp->zp_nopwrite); 2003 ASSERT(!zp->zp_dedup); 2004 ASSERT(zio->io_bp_override == NULL); 2005 ASSERT(IO_IS_ALLOCATING(zio)); 2006 2007 /* 2008 * Check to see if the original bp and the new bp have matching 2009 * characteristics (i.e. same checksum, compression algorithms, etc). 2010 * If they don't then just continue with the pipeline which will 2011 * allocate a new bp. 2012 */ 2013 if (BP_IS_HOLE(bp_orig) || 2014 !zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup || 2015 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) || 2016 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) || 2017 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) || 2018 zp->zp_copies != BP_GET_NDVAS(bp_orig)) 2019 return (ZIO_PIPELINE_CONTINUE); 2020 2021 /* 2022 * If the checksums match then reset the pipeline so that we 2023 * avoid allocating a new bp and issuing any I/O. 2024 */ 2025 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) { 2026 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup); 2027 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig)); 2028 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig)); 2029 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF); 2030 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop, 2031 sizeof (uint64_t)) == 0); 2032 2033 *bp = *bp_orig; 2034 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2035 zio->io_flags |= ZIO_FLAG_NOPWRITE; 2036 } 2037 2038 return (ZIO_PIPELINE_CONTINUE); 2039} 2040 2041/* 2042 * ========================================================================== 2043 * Dedup 2044 * ========================================================================== 2045 */ 2046static void 2047zio_ddt_child_read_done(zio_t *zio) 2048{ 2049 blkptr_t *bp = zio->io_bp; 2050 ddt_entry_t *dde = zio->io_private; 2051 ddt_phys_t *ddp; 2052 zio_t *pio = zio_unique_parent(zio); 2053 2054 mutex_enter(&pio->io_lock); 2055 ddp = ddt_phys_select(dde, bp); 2056 if (zio->io_error == 0) 2057 ddt_phys_clear(ddp); /* this ddp doesn't need repair */ 2058 if (zio->io_error == 0 && dde->dde_repair_data == NULL) 2059 dde->dde_repair_data = zio->io_data; 2060 else 2061 zio_buf_free(zio->io_data, zio->io_size); 2062 mutex_exit(&pio->io_lock); 2063} 2064 2065static int 2066zio_ddt_read_start(zio_t *zio) 2067{ 2068 blkptr_t *bp = zio->io_bp; 2069 2070 ASSERT(BP_GET_DEDUP(bp)); 2071 ASSERT(BP_GET_PSIZE(bp) == zio->io_size); 2072 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2073 2074 if (zio->io_child_error[ZIO_CHILD_DDT]) { 2075 ddt_t *ddt = ddt_select(zio->io_spa, bp); 2076 ddt_entry_t *dde = ddt_repair_start(ddt, bp); 2077 ddt_phys_t *ddp = dde->dde_phys; 2078 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp); 2079 blkptr_t blk; 2080 2081 ASSERT(zio->io_vsd == NULL); 2082 zio->io_vsd = dde; 2083 2084 if (ddp_self == NULL) 2085 return (ZIO_PIPELINE_CONTINUE); 2086 2087 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) { 2088 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self) 2089 continue; 2090 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp, 2091 &blk); 2092 zio_nowait(zio_read(zio, zio->io_spa, &blk, 2093 zio_buf_alloc(zio->io_size), zio->io_size, 2094 zio_ddt_child_read_done, dde, zio->io_priority, 2095 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE, 2096 &zio->io_bookmark)); 2097 } 2098 return (ZIO_PIPELINE_CONTINUE); 2099 } 2100 2101 zio_nowait(zio_read(zio, zio->io_spa, bp, 2102 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority, 2103 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark)); 2104 2105 return (ZIO_PIPELINE_CONTINUE); 2106} 2107 2108static int 2109zio_ddt_read_done(zio_t *zio) 2110{ 2111 blkptr_t *bp = zio->io_bp; 2112 2113 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE)) 2114 return (ZIO_PIPELINE_STOP); 2115 2116 ASSERT(BP_GET_DEDUP(bp)); 2117 ASSERT(BP_GET_PSIZE(bp) == zio->io_size); 2118 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2119 2120 if (zio->io_child_error[ZIO_CHILD_DDT]) { 2121 ddt_t *ddt = ddt_select(zio->io_spa, bp); 2122 ddt_entry_t *dde = zio->io_vsd; 2123 if (ddt == NULL) { 2124 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE); 2125 return (ZIO_PIPELINE_CONTINUE); 2126 } 2127 if (dde == NULL) { 2128 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1; 2129 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); 2130 return (ZIO_PIPELINE_STOP); 2131 } 2132 if (dde->dde_repair_data != NULL) { 2133 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size); 2134 zio->io_child_error[ZIO_CHILD_DDT] = 0; 2135 } 2136 ddt_repair_done(ddt, dde); 2137 zio->io_vsd = NULL; 2138 } 2139 2140 ASSERT(zio->io_vsd == NULL); 2141 2142 return (ZIO_PIPELINE_CONTINUE); 2143} 2144 2145static boolean_t 2146zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde) 2147{ 2148 spa_t *spa = zio->io_spa; 2149 2150 /* 2151 * Note: we compare the original data, not the transformed data, 2152 * because when zio->io_bp is an override bp, we will not have 2153 * pushed the I/O transforms. That's an important optimization 2154 * because otherwise we'd compress/encrypt all dmu_sync() data twice. 2155 */ 2156 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { 2157 zio_t *lio = dde->dde_lead_zio[p]; 2158 2159 if (lio != NULL) { 2160 return (lio->io_orig_size != zio->io_orig_size || 2161 bcmp(zio->io_orig_data, lio->io_orig_data, 2162 zio->io_orig_size) != 0); 2163 } 2164 } 2165 2166 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { 2167 ddt_phys_t *ddp = &dde->dde_phys[p]; 2168 2169 if (ddp->ddp_phys_birth != 0) { 2170 arc_buf_t *abuf = NULL; 2171 uint32_t aflags = ARC_WAIT; 2172 blkptr_t blk = *zio->io_bp; 2173 int error; 2174 2175 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth); 2176 2177 ddt_exit(ddt); 2178 2179 error = arc_read(NULL, spa, &blk, 2180 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ, 2181 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2182 &aflags, &zio->io_bookmark); 2183 2184 if (error == 0) { 2185 if (arc_buf_size(abuf) != zio->io_orig_size || 2186 bcmp(abuf->b_data, zio->io_orig_data, 2187 zio->io_orig_size) != 0) 2188 error = SET_ERROR(EEXIST); 2189 VERIFY(arc_buf_remove_ref(abuf, &abuf)); 2190 } 2191 2192 ddt_enter(ddt); 2193 return (error != 0); 2194 } 2195 } 2196 2197 return (B_FALSE); 2198} 2199 2200static void 2201zio_ddt_child_write_ready(zio_t *zio) 2202{ 2203 int p = zio->io_prop.zp_copies; 2204 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); 2205 ddt_entry_t *dde = zio->io_private; 2206 ddt_phys_t *ddp = &dde->dde_phys[p]; 2207 zio_t *pio; 2208 2209 if (zio->io_error) 2210 return; 2211 2212 ddt_enter(ddt); 2213 2214 ASSERT(dde->dde_lead_zio[p] == zio); 2215 2216 ddt_phys_fill(ddp, zio->io_bp); 2217 2218 while ((pio = zio_walk_parents(zio)) != NULL) 2219 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg); 2220 2221 ddt_exit(ddt); 2222} 2223 2224static void 2225zio_ddt_child_write_done(zio_t *zio) 2226{ 2227 int p = zio->io_prop.zp_copies; 2228 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); 2229 ddt_entry_t *dde = zio->io_private; 2230 ddt_phys_t *ddp = &dde->dde_phys[p]; 2231 2232 ddt_enter(ddt); 2233 2234 ASSERT(ddp->ddp_refcnt == 0); 2235 ASSERT(dde->dde_lead_zio[p] == zio); 2236 dde->dde_lead_zio[p] = NULL; 2237 2238 if (zio->io_error == 0) { 2239 while (zio_walk_parents(zio) != NULL) 2240 ddt_phys_addref(ddp); 2241 } else { 2242 ddt_phys_clear(ddp); 2243 } 2244 2245 ddt_exit(ddt); 2246} 2247 2248static void 2249zio_ddt_ditto_write_done(zio_t *zio) 2250{ 2251 int p = DDT_PHYS_DITTO; 2252 zio_prop_t *zp = &zio->io_prop; 2253 blkptr_t *bp = zio->io_bp; 2254 ddt_t *ddt = ddt_select(zio->io_spa, bp); 2255 ddt_entry_t *dde = zio->io_private; 2256 ddt_phys_t *ddp = &dde->dde_phys[p]; 2257 ddt_key_t *ddk = &dde->dde_key; 2258 2259 ddt_enter(ddt); 2260 2261 ASSERT(ddp->ddp_refcnt == 0); 2262 ASSERT(dde->dde_lead_zio[p] == zio); 2263 dde->dde_lead_zio[p] = NULL; 2264 2265 if (zio->io_error == 0) { 2266 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum)); 2267 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP); 2268 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp)); 2269 if (ddp->ddp_phys_birth != 0) 2270 ddt_phys_free(ddt, ddk, ddp, zio->io_txg); 2271 ddt_phys_fill(ddp, bp); 2272 } 2273 2274 ddt_exit(ddt); 2275} 2276 2277static int 2278zio_ddt_write(zio_t *zio) 2279{ 2280 spa_t *spa = zio->io_spa; 2281 blkptr_t *bp = zio->io_bp; 2282 uint64_t txg = zio->io_txg; 2283 zio_prop_t *zp = &zio->io_prop; 2284 int p = zp->zp_copies; 2285 int ditto_copies; 2286 zio_t *cio = NULL; 2287 zio_t *dio = NULL; 2288 ddt_t *ddt = ddt_select(spa, bp); 2289 ddt_entry_t *dde; 2290 ddt_phys_t *ddp; 2291 2292 ASSERT(BP_GET_DEDUP(bp)); 2293 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum); 2294 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override); 2295 2296 ddt_enter(ddt); 2297 dde = ddt_lookup(ddt, bp, B_TRUE); 2298 ddp = &dde->dde_phys[p]; 2299 2300 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) { 2301 /* 2302 * If we're using a weak checksum, upgrade to a strong checksum 2303 * and try again. If we're already using a strong checksum, 2304 * we can't resolve it, so just convert to an ordinary write. 2305 * (And automatically e-mail a paper to Nature?) 2306 */ 2307 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) { 2308 zp->zp_checksum = spa_dedup_checksum(spa); 2309 zio_pop_transforms(zio); 2310 zio->io_stage = ZIO_STAGE_OPEN; 2311 BP_ZERO(bp); 2312 } else { 2313 zp->zp_dedup = B_FALSE; 2314 } 2315 zio->io_pipeline = ZIO_WRITE_PIPELINE; 2316 ddt_exit(ddt); 2317 return (ZIO_PIPELINE_CONTINUE); 2318 } 2319 2320 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp); 2321 ASSERT(ditto_copies < SPA_DVAS_PER_BP); 2322 2323 if (ditto_copies > ddt_ditto_copies_present(dde) && 2324 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) { 2325 zio_prop_t czp = *zp; 2326 2327 czp.zp_copies = ditto_copies; 2328 2329 /* 2330 * If we arrived here with an override bp, we won't have run 2331 * the transform stack, so we won't have the data we need to 2332 * generate a child i/o. So, toss the override bp and restart. 2333 * This is safe, because using the override bp is just an 2334 * optimization; and it's rare, so the cost doesn't matter. 2335 */ 2336 if (zio->io_bp_override) { 2337 zio_pop_transforms(zio); 2338 zio->io_stage = ZIO_STAGE_OPEN; 2339 zio->io_pipeline = ZIO_WRITE_PIPELINE; 2340 zio->io_bp_override = NULL; 2341 BP_ZERO(bp); 2342 ddt_exit(ddt); 2343 return (ZIO_PIPELINE_CONTINUE); 2344 } 2345 2346 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data, 2347 zio->io_orig_size, &czp, NULL, NULL, 2348 zio_ddt_ditto_write_done, dde, zio->io_priority, 2349 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); 2350 2351 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL); 2352 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio; 2353 } 2354 2355 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) { 2356 if (ddp->ddp_phys_birth != 0) 2357 ddt_bp_fill(ddp, bp, txg); 2358 if (dde->dde_lead_zio[p] != NULL) 2359 zio_add_child(zio, dde->dde_lead_zio[p]); 2360 else 2361 ddt_phys_addref(ddp); 2362 } else if (zio->io_bp_override) { 2363 ASSERT(bp->blk_birth == txg); 2364 ASSERT(BP_EQUAL(bp, zio->io_bp_override)); 2365 ddt_phys_fill(ddp, bp); 2366 ddt_phys_addref(ddp); 2367 } else { 2368 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data, 2369 zio->io_orig_size, zp, zio_ddt_child_write_ready, NULL, 2370 zio_ddt_child_write_done, dde, zio->io_priority, 2371 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); 2372 2373 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL); 2374 dde->dde_lead_zio[p] = cio; 2375 } 2376 2377 ddt_exit(ddt); 2378 2379 if (cio) 2380 zio_nowait(cio); 2381 if (dio) 2382 zio_nowait(dio); 2383 2384 return (ZIO_PIPELINE_CONTINUE); 2385} 2386 2387ddt_entry_t *freedde; /* for debugging */ 2388 2389static int 2390zio_ddt_free(zio_t *zio) 2391{ 2392 spa_t *spa = zio->io_spa; 2393 blkptr_t *bp = zio->io_bp; 2394 ddt_t *ddt = ddt_select(spa, bp); 2395 ddt_entry_t *dde; 2396 ddt_phys_t *ddp; 2397 2398 ASSERT(BP_GET_DEDUP(bp)); 2399 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2400 2401 ddt_enter(ddt); 2402 freedde = dde = ddt_lookup(ddt, bp, B_TRUE); 2403 ddp = ddt_phys_select(dde, bp); 2404 ddt_phys_decref(ddp); 2405 ddt_exit(ddt); 2406 2407 return (ZIO_PIPELINE_CONTINUE); 2408} 2409 2410/* 2411 * ========================================================================== 2412 * Allocate and free blocks 2413 * ========================================================================== 2414 */ 2415static int 2416zio_dva_allocate(zio_t *zio) 2417{ 2418 spa_t *spa = zio->io_spa; 2419 metaslab_class_t *mc = spa_normal_class(spa); 2420 blkptr_t *bp = zio->io_bp; 2421 int error; 2422 int flags = 0; 2423 2424 if (zio->io_gang_leader == NULL) { 2425 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 2426 zio->io_gang_leader = zio; 2427 } 2428 2429 ASSERT(BP_IS_HOLE(bp)); 2430 ASSERT0(BP_GET_NDVAS(bp)); 2431 ASSERT3U(zio->io_prop.zp_copies, >, 0); 2432 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa)); 2433 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp)); 2434 2435 /* 2436 * The dump device does not support gang blocks so allocation on 2437 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid 2438 * the "fast" gang feature. 2439 */ 2440 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0; 2441 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ? 2442 METASLAB_GANG_CHILD : 0; 2443 error = metaslab_alloc(spa, mc, zio->io_size, bp, 2444 zio->io_prop.zp_copies, zio->io_txg, NULL, flags); 2445 2446 if (error) { 2447 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, " 2448 "size %llu, error %d", spa_name(spa), zio, zio->io_size, 2449 error); 2450 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE) 2451 return (zio_write_gang_block(zio)); 2452 zio->io_error = error; 2453 } 2454 2455 return (ZIO_PIPELINE_CONTINUE); 2456} 2457 2458static int 2459zio_dva_free(zio_t *zio) 2460{ 2461 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE); 2462 2463 return (ZIO_PIPELINE_CONTINUE); 2464} 2465 2466static int 2467zio_dva_claim(zio_t *zio) 2468{ 2469 int error; 2470 2471 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg); 2472 if (error) 2473 zio->io_error = error; 2474 2475 return (ZIO_PIPELINE_CONTINUE); 2476} 2477 2478/* 2479 * Undo an allocation. This is used by zio_done() when an I/O fails 2480 * and we want to give back the block we just allocated. 2481 * This handles both normal blocks and gang blocks. 2482 */ 2483static void 2484zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp) 2485{ 2486 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp)); 2487 ASSERT(zio->io_bp_override == NULL); 2488 2489 if (!BP_IS_HOLE(bp)) 2490 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE); 2491 2492 if (gn != NULL) { 2493 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 2494 zio_dva_unallocate(zio, gn->gn_child[g], 2495 &gn->gn_gbh->zg_blkptr[g]); 2496 } 2497 } 2498} 2499 2500/* 2501 * Try to allocate an intent log block. Return 0 on success, errno on failure. 2502 */ 2503int 2504zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp, 2505 uint64_t size, boolean_t use_slog) 2506{ 2507 int error = 1; 2508 2509 ASSERT(txg > spa_syncing_txg(spa)); 2510 2511 /* 2512 * ZIL blocks are always contiguous (i.e. not gang blocks) so we 2513 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang" 2514 * when allocating them. 2515 */ 2516 if (use_slog) { 2517 error = metaslab_alloc(spa, spa_log_class(spa), size, 2518 new_bp, 1, txg, old_bp, 2519 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID); 2520 } 2521 2522 if (error) { 2523 error = metaslab_alloc(spa, spa_normal_class(spa), size, 2524 new_bp, 1, txg, old_bp, 2525 METASLAB_HINTBP_AVOID); 2526 } 2527 2528 if (error == 0) { 2529 BP_SET_LSIZE(new_bp, size); 2530 BP_SET_PSIZE(new_bp, size); 2531 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF); 2532 BP_SET_CHECKSUM(new_bp, 2533 spa_version(spa) >= SPA_VERSION_SLIM_ZIL 2534 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG); 2535 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); 2536 BP_SET_LEVEL(new_bp, 0); 2537 BP_SET_DEDUP(new_bp, 0); 2538 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER); 2539 } 2540 2541 return (error); 2542} 2543 2544/* 2545 * Free an intent log block. 2546 */ 2547void 2548zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp) 2549{ 2550 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG); 2551 ASSERT(!BP_IS_GANG(bp)); 2552 2553 zio_free(spa, txg, bp); 2554} 2555 2556/* 2557 * ========================================================================== 2558 * Read, write and delete to physical devices 2559 * ========================================================================== 2560 */ 2561 2562 2563/* 2564 * Issue an I/O to the underlying vdev. Typically the issue pipeline 2565 * stops after this stage and will resume upon I/O completion. 2566 * However, there are instances where the vdev layer may need to 2567 * continue the pipeline when an I/O was not issued. Since the I/O 2568 * that was sent to the vdev layer might be different than the one 2569 * currently active in the pipeline (see vdev_queue_io()), we explicitly 2570 * force the underlying vdev layers to call either zio_execute() or 2571 * zio_interrupt() to ensure that the pipeline continues with the correct I/O. 2572 */ 2573static int 2574zio_vdev_io_start(zio_t *zio) 2575{ 2576 vdev_t *vd = zio->io_vd; 2577 uint64_t align; 2578 spa_t *spa = zio->io_spa; 2579 int ret; 2580 2581 ASSERT(zio->io_error == 0); 2582 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0); 2583 2584 if (vd == NULL) { 2585 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) 2586 spa_config_enter(spa, SCL_ZIO, zio, RW_READER); 2587 2588 /* 2589 * The mirror_ops handle multiple DVAs in a single BP. 2590 */ 2591 vdev_mirror_ops.vdev_op_io_start(zio); 2592 return (ZIO_PIPELINE_STOP); 2593 } 2594 2595 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE && 2596 zio->io_priority == ZIO_PRIORITY_NOW) { 2597 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg); 2598 return (ZIO_PIPELINE_CONTINUE); 2599 } 2600 2601 /* 2602 * We keep track of time-sensitive I/Os so that the scan thread 2603 * can quickly react to certain workloads. In particular, we care 2604 * about non-scrubbing, top-level reads and writes with the following 2605 * characteristics: 2606 * - synchronous writes of user data to non-slog devices 2607 * - any reads of user data 2608 * When these conditions are met, adjust the timestamp of spa_last_io 2609 * which allows the scan thread to adjust its workload accordingly. 2610 */ 2611 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL && 2612 vd == vd->vdev_top && !vd->vdev_islog && 2613 zio->io_bookmark.zb_objset != DMU_META_OBJSET && 2614 zio->io_txg != spa_syncing_txg(spa)) { 2615 uint64_t old = spa->spa_last_io; 2616 uint64_t new = ddi_get_lbolt64(); 2617 if (old != new) 2618 (void) atomic_cas_64(&spa->spa_last_io, old, new); 2619 } 2620 2621 align = 1ULL << vd->vdev_top->vdev_ashift; 2622 2623 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) && 2624 P2PHASE(zio->io_size, align) != 0) { 2625 /* Transform logical writes to be a full physical block size. */ 2626 uint64_t asize = P2ROUNDUP(zio->io_size, align); 2627 char *abuf = NULL; 2628 if (zio->io_type == ZIO_TYPE_READ || 2629 zio->io_type == ZIO_TYPE_WRITE) 2630 abuf = zio_buf_alloc(asize); 2631 ASSERT(vd == vd->vdev_top); 2632 if (zio->io_type == ZIO_TYPE_WRITE) { 2633 bcopy(zio->io_data, abuf, zio->io_size); 2634 bzero(abuf + zio->io_size, asize - zio->io_size); 2635 } 2636 zio_push_transform(zio, abuf, asize, abuf ? asize : 0, 2637 zio_subblock); 2638 } 2639 2640 /* 2641 * If this is not a physical io, make sure that it is properly aligned 2642 * before proceeding. 2643 */ 2644 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) { 2645 ASSERT0(P2PHASE(zio->io_offset, align)); 2646 ASSERT0(P2PHASE(zio->io_size, align)); 2647 } else { 2648 /* 2649 * For physical writes, we allow 512b aligned writes and assume 2650 * the device will perform a read-modify-write as necessary. 2651 */ 2652 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE)); 2653 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE)); 2654 } 2655 2656 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa)); 2657 2658 /* 2659 * If this is a repair I/O, and there's no self-healing involved -- 2660 * that is, we're just resilvering what we expect to resilver -- 2661 * then don't do the I/O unless zio's txg is actually in vd's DTL. 2662 * This prevents spurious resilvering with nested replication. 2663 * For example, given a mirror of mirrors, (A+B)+(C+D), if only 2664 * A is out of date, we'll read from C+D, then use the data to 2665 * resilver A+B -- but we don't actually want to resilver B, just A. 2666 * The top-level mirror has no way to know this, so instead we just 2667 * discard unnecessary repairs as we work our way down the vdev tree. 2668 * The same logic applies to any form of nested replication: 2669 * ditto + mirror, RAID-Z + replacing, etc. This covers them all. 2670 */ 2671 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) && 2672 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) && 2673 zio->io_txg != 0 && /* not a delegated i/o */ 2674 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) { 2675 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 2676 zio_vdev_io_bypass(zio); 2677 return (ZIO_PIPELINE_CONTINUE); 2678 } 2679 2680 if (vd->vdev_ops->vdev_op_leaf) { 2681 switch (zio->io_type) { 2682 case ZIO_TYPE_READ: 2683 if (vdev_cache_read(zio)) 2684 return (ZIO_PIPELINE_CONTINUE); 2685 /* FALLTHROUGH */ 2686 case ZIO_TYPE_WRITE: 2687 case ZIO_TYPE_FREE: 2688 if ((zio = vdev_queue_io(zio)) == NULL) 2689 return (ZIO_PIPELINE_STOP); 2690 2691 if (!vdev_accessible(vd, zio)) { 2692 zio->io_error = SET_ERROR(ENXIO); 2693 zio_interrupt(zio); 2694 return (ZIO_PIPELINE_STOP); 2695 } 2696 break; 2697 } 2698 /* 2699 * Note that we ignore repair writes for TRIM because they can 2700 * conflict with normal writes. This isn't an issue because, by 2701 * definition, we only repair blocks that aren't freed. 2702 */ 2703 if (zio->io_type == ZIO_TYPE_WRITE && 2704 !(zio->io_flags & ZIO_FLAG_IO_REPAIR) && 2705 !trim_map_write_start(zio)) 2706 return (ZIO_PIPELINE_STOP); 2707 } 2708 2709 vd->vdev_ops->vdev_op_io_start(zio); 2710 return (ZIO_PIPELINE_STOP); 2711} 2712 2713static int 2714zio_vdev_io_done(zio_t *zio) 2715{ 2716 vdev_t *vd = zio->io_vd; 2717 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops; 2718 boolean_t unexpected_error = B_FALSE; 2719 2720 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE)) 2721 return (ZIO_PIPELINE_STOP); 2722 2723 ASSERT(zio->io_type == ZIO_TYPE_READ || 2724 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE); 2725 2726 if (vd != NULL && vd->vdev_ops->vdev_op_leaf && 2727 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE || 2728 zio->io_type == ZIO_TYPE_FREE)) { 2729 2730 if (zio->io_type == ZIO_TYPE_WRITE && 2731 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) 2732 trim_map_write_done(zio); 2733 2734 vdev_queue_io_done(zio); 2735 2736 if (zio->io_type == ZIO_TYPE_WRITE) 2737 vdev_cache_write(zio); 2738 2739 if (zio_injection_enabled && zio->io_error == 0) 2740 zio->io_error = zio_handle_device_injection(vd, 2741 zio, EIO); 2742 2743 if (zio_injection_enabled && zio->io_error == 0) 2744 zio->io_error = zio_handle_label_injection(zio, EIO); 2745 2746 if (zio->io_error) { 2747 if (zio->io_error == ENOTSUP && 2748 zio->io_type == ZIO_TYPE_FREE) { 2749 /* Not all devices support TRIM. */ 2750 } else if (!vdev_accessible(vd, zio)) { 2751 zio->io_error = SET_ERROR(ENXIO); 2752 } else { 2753 unexpected_error = B_TRUE; 2754 } 2755 } 2756 } 2757 2758 ops->vdev_op_io_done(zio); 2759 2760 if (unexpected_error) 2761 VERIFY(vdev_probe(vd, zio) == NULL); 2762 2763 return (ZIO_PIPELINE_CONTINUE); 2764} 2765 2766/* 2767 * For non-raidz ZIOs, we can just copy aside the bad data read from the 2768 * disk, and use that to finish the checksum ereport later. 2769 */ 2770static void 2771zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr, 2772 const void *good_buf) 2773{ 2774 /* no processing needed */ 2775 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE); 2776} 2777 2778/*ARGSUSED*/ 2779void 2780zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored) 2781{ 2782 void *buf = zio_buf_alloc(zio->io_size); 2783 2784 bcopy(zio->io_data, buf, zio->io_size); 2785 2786 zcr->zcr_cbinfo = zio->io_size; 2787 zcr->zcr_cbdata = buf; 2788 zcr->zcr_finish = zio_vsd_default_cksum_finish; 2789 zcr->zcr_free = zio_buf_free; 2790} 2791 2792static int 2793zio_vdev_io_assess(zio_t *zio) 2794{ 2795 vdev_t *vd = zio->io_vd; 2796 2797 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE)) 2798 return (ZIO_PIPELINE_STOP); 2799 2800 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) 2801 spa_config_exit(zio->io_spa, SCL_ZIO, zio); 2802 2803 if (zio->io_vsd != NULL) { 2804 zio->io_vsd_ops->vsd_free(zio); 2805 zio->io_vsd = NULL; 2806 } 2807 2808 if (zio_injection_enabled && zio->io_error == 0) 2809 zio->io_error = zio_handle_fault_injection(zio, EIO); 2810 2811 if (zio->io_type == ZIO_TYPE_FREE && 2812 zio->io_priority != ZIO_PRIORITY_NOW) { 2813 switch (zio->io_error) { 2814 case 0: 2815 ZIO_TRIM_STAT_INCR(bytes, zio->io_size); 2816 ZIO_TRIM_STAT_BUMP(success); 2817 break; 2818 case EOPNOTSUPP: 2819 ZIO_TRIM_STAT_BUMP(unsupported); 2820 break; 2821 default: 2822 ZIO_TRIM_STAT_BUMP(failed); 2823 break; 2824 } 2825 } 2826 2827 /* 2828 * If the I/O failed, determine whether we should attempt to retry it. 2829 * 2830 * On retry, we cut in line in the issue queue, since we don't want 2831 * compression/checksumming/etc. work to prevent our (cheap) IO reissue. 2832 */ 2833 if (zio->io_error && vd == NULL && 2834 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) { 2835 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */ 2836 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */ 2837 zio->io_error = 0; 2838 zio->io_flags |= ZIO_FLAG_IO_RETRY | 2839 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE; 2840 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1; 2841 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, 2842 zio_requeue_io_start_cut_in_line); 2843 return (ZIO_PIPELINE_STOP); 2844 } 2845 2846 /* 2847 * If we got an error on a leaf device, convert it to ENXIO 2848 * if the device is not accessible at all. 2849 */ 2850 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf && 2851 !vdev_accessible(vd, zio)) 2852 zio->io_error = SET_ERROR(ENXIO); 2853 2854 /* 2855 * If we can't write to an interior vdev (mirror or RAID-Z), 2856 * set vdev_cant_write so that we stop trying to allocate from it. 2857 */ 2858 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE && 2859 vd != NULL && !vd->vdev_ops->vdev_op_leaf) { 2860 vd->vdev_cant_write = B_TRUE; 2861 } 2862 2863 if (zio->io_error) 2864 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2865 2866 if (vd != NULL && vd->vdev_ops->vdev_op_leaf && 2867 zio->io_physdone != NULL) { 2868 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED)); 2869 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV); 2870 zio->io_physdone(zio->io_logical); 2871 } 2872 2873 return (ZIO_PIPELINE_CONTINUE); 2874} 2875 2876void 2877zio_vdev_io_reissue(zio_t *zio) 2878{ 2879 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); 2880 ASSERT(zio->io_error == 0); 2881 2882 zio->io_stage >>= 1; 2883} 2884 2885void 2886zio_vdev_io_redone(zio_t *zio) 2887{ 2888 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE); 2889 2890 zio->io_stage >>= 1; 2891} 2892 2893void 2894zio_vdev_io_bypass(zio_t *zio) 2895{ 2896 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); 2897 ASSERT(zio->io_error == 0); 2898 2899 zio->io_flags |= ZIO_FLAG_IO_BYPASS; 2900 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1; 2901} 2902 2903/* 2904 * ========================================================================== 2905 * Generate and verify checksums 2906 * ========================================================================== 2907 */ 2908static int 2909zio_checksum_generate(zio_t *zio) 2910{ 2911 blkptr_t *bp = zio->io_bp; 2912 enum zio_checksum checksum; 2913 2914 if (bp == NULL) { 2915 /* 2916 * This is zio_write_phys(). 2917 * We're either generating a label checksum, or none at all. 2918 */ 2919 checksum = zio->io_prop.zp_checksum; 2920 2921 if (checksum == ZIO_CHECKSUM_OFF) 2922 return (ZIO_PIPELINE_CONTINUE); 2923 2924 ASSERT(checksum == ZIO_CHECKSUM_LABEL); 2925 } else { 2926 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) { 2927 ASSERT(!IO_IS_ALLOCATING(zio)); 2928 checksum = ZIO_CHECKSUM_GANG_HEADER; 2929 } else { 2930 checksum = BP_GET_CHECKSUM(bp); 2931 } 2932 } 2933 2934 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size); 2935 2936 return (ZIO_PIPELINE_CONTINUE); 2937} 2938 2939static int 2940zio_checksum_verify(zio_t *zio) 2941{ 2942 zio_bad_cksum_t info; 2943 blkptr_t *bp = zio->io_bp; 2944 int error; 2945 2946 ASSERT(zio->io_vd != NULL); 2947 2948 if (bp == NULL) { 2949 /* 2950 * This is zio_read_phys(). 2951 * We're either verifying a label checksum, or nothing at all. 2952 */ 2953 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF) 2954 return (ZIO_PIPELINE_CONTINUE); 2955 2956 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL); 2957 } 2958 2959 if ((error = zio_checksum_error(zio, &info)) != 0) { 2960 zio->io_error = error; 2961 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) { 2962 zfs_ereport_start_checksum(zio->io_spa, 2963 zio->io_vd, zio, zio->io_offset, 2964 zio->io_size, NULL, &info); 2965 } 2966 } 2967 2968 return (ZIO_PIPELINE_CONTINUE); 2969} 2970 2971/* 2972 * Called by RAID-Z to ensure we don't compute the checksum twice. 2973 */ 2974void 2975zio_checksum_verified(zio_t *zio) 2976{ 2977 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; 2978} 2979 2980/* 2981 * ========================================================================== 2982 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other. 2983 * An error of 0 indicates success. ENXIO indicates whole-device failure, 2984 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO 2985 * indicate errors that are specific to one I/O, and most likely permanent. 2986 * Any other error is presumed to be worse because we weren't expecting it. 2987 * ========================================================================== 2988 */ 2989int 2990zio_worst_error(int e1, int e2) 2991{ 2992 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO }; 2993 int r1, r2; 2994 2995 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++) 2996 if (e1 == zio_error_rank[r1]) 2997 break; 2998 2999 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++) 3000 if (e2 == zio_error_rank[r2]) 3001 break; 3002 3003 return (r1 > r2 ? e1 : e2); 3004} 3005 3006/* 3007 * ========================================================================== 3008 * I/O completion 3009 * ========================================================================== 3010 */ 3011static int 3012zio_ready(zio_t *zio) 3013{ 3014 blkptr_t *bp = zio->io_bp; 3015 zio_t *pio, *pio_next; 3016 3017 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) || 3018 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY)) 3019 return (ZIO_PIPELINE_STOP); 3020 3021 if (zio->io_ready) { 3022 ASSERT(IO_IS_ALLOCATING(zio)); 3023 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) || 3024 (zio->io_flags & ZIO_FLAG_NOPWRITE)); 3025 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0); 3026 3027 zio->io_ready(zio); 3028 } 3029 3030 if (bp != NULL && bp != &zio->io_bp_copy) 3031 zio->io_bp_copy = *bp; 3032 3033 if (zio->io_error) 3034 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 3035 3036 mutex_enter(&zio->io_lock); 3037 zio->io_state[ZIO_WAIT_READY] = 1; 3038 pio = zio_walk_parents(zio); 3039 mutex_exit(&zio->io_lock); 3040 3041 /* 3042 * As we notify zio's parents, new parents could be added. 3043 * New parents go to the head of zio's io_parent_list, however, 3044 * so we will (correctly) not notify them. The remainder of zio's 3045 * io_parent_list, from 'pio_next' onward, cannot change because 3046 * all parents must wait for us to be done before they can be done. 3047 */ 3048 for (; pio != NULL; pio = pio_next) { 3049 pio_next = zio_walk_parents(zio); 3050 zio_notify_parent(pio, zio, ZIO_WAIT_READY); 3051 } 3052 3053 if (zio->io_flags & ZIO_FLAG_NODATA) { 3054 if (BP_IS_GANG(bp)) { 3055 zio->io_flags &= ~ZIO_FLAG_NODATA; 3056 } else { 3057 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE); 3058 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; 3059 } 3060 } 3061 3062 if (zio_injection_enabled && 3063 zio->io_spa->spa_syncing_txg == zio->io_txg) 3064 zio_handle_ignored_writes(zio); 3065 3066 return (ZIO_PIPELINE_CONTINUE); 3067} 3068 3069static int 3070zio_done(zio_t *zio) 3071{ 3072 spa_t *spa = zio->io_spa; 3073 zio_t *lio = zio->io_logical; 3074 blkptr_t *bp = zio->io_bp; 3075 vdev_t *vd = zio->io_vd; 3076 uint64_t psize = zio->io_size; 3077 zio_t *pio, *pio_next; 3078 3079 /* 3080 * If our children haven't all completed, 3081 * wait for them and then repeat this pipeline stage. 3082 */ 3083 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) || 3084 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) || 3085 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) || 3086 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE)) 3087 return (ZIO_PIPELINE_STOP); 3088 3089 for (int c = 0; c < ZIO_CHILD_TYPES; c++) 3090 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 3091 ASSERT(zio->io_children[c][w] == 0); 3092 3093 if (bp != NULL && !BP_IS_EMBEDDED(bp)) { 3094 ASSERT(bp->blk_pad[0] == 0); 3095 ASSERT(bp->blk_pad[1] == 0); 3096 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 || 3097 (bp == zio_unique_parent(zio)->io_bp)); 3098 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) && 3099 zio->io_bp_override == NULL && 3100 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) { 3101 ASSERT(!BP_SHOULD_BYTESWAP(bp)); 3102 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp)); 3103 ASSERT(BP_COUNT_GANG(bp) == 0 || 3104 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp))); 3105 } 3106 if (zio->io_flags & ZIO_FLAG_NOPWRITE) 3107 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig)); 3108 } 3109 3110 /* 3111 * If there were child vdev/gang/ddt errors, they apply to us now. 3112 */ 3113 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV); 3114 zio_inherit_child_errors(zio, ZIO_CHILD_GANG); 3115 zio_inherit_child_errors(zio, ZIO_CHILD_DDT); 3116 3117 /* 3118 * If the I/O on the transformed data was successful, generate any 3119 * checksum reports now while we still have the transformed data. 3120 */ 3121 if (zio->io_error == 0) { 3122 while (zio->io_cksum_report != NULL) { 3123 zio_cksum_report_t *zcr = zio->io_cksum_report; 3124 uint64_t align = zcr->zcr_align; 3125 uint64_t asize = P2ROUNDUP(psize, align); 3126 char *abuf = zio->io_data; 3127 3128 if (asize != psize) { 3129 abuf = zio_buf_alloc(asize); 3130 bcopy(zio->io_data, abuf, psize); 3131 bzero(abuf + psize, asize - psize); 3132 } 3133 3134 zio->io_cksum_report = zcr->zcr_next; 3135 zcr->zcr_next = NULL; 3136 zcr->zcr_finish(zcr, abuf); 3137 zfs_ereport_free_checksum(zcr); 3138 3139 if (asize != psize) 3140 zio_buf_free(abuf, asize); 3141 } 3142 } 3143 3144 zio_pop_transforms(zio); /* note: may set zio->io_error */ 3145 3146 vdev_stat_update(zio, psize); 3147 3148 if (zio->io_error) { 3149 /* 3150 * If this I/O is attached to a particular vdev, 3151 * generate an error message describing the I/O failure 3152 * at the block level. We ignore these errors if the 3153 * device is currently unavailable. 3154 */ 3155 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd)) 3156 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0); 3157 3158 if ((zio->io_error == EIO || !(zio->io_flags & 3159 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) && 3160 zio == lio) { 3161 /* 3162 * For logical I/O requests, tell the SPA to log the 3163 * error and generate a logical data ereport. 3164 */ 3165 spa_log_error(spa, zio); 3166 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio, 3167 0, 0); 3168 } 3169 } 3170 3171 if (zio->io_error && zio == lio) { 3172 /* 3173 * Determine whether zio should be reexecuted. This will 3174 * propagate all the way to the root via zio_notify_parent(). 3175 */ 3176 ASSERT(vd == NULL && bp != NULL); 3177 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 3178 3179 if (IO_IS_ALLOCATING(zio) && 3180 !(zio->io_flags & ZIO_FLAG_CANFAIL)) { 3181 if (zio->io_error != ENOSPC) 3182 zio->io_reexecute |= ZIO_REEXECUTE_NOW; 3183 else 3184 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 3185 } 3186 3187 if ((zio->io_type == ZIO_TYPE_READ || 3188 zio->io_type == ZIO_TYPE_FREE) && 3189 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && 3190 zio->io_error == ENXIO && 3191 spa_load_state(spa) == SPA_LOAD_NONE && 3192 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE) 3193 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 3194 3195 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute) 3196 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 3197 3198 /* 3199 * Here is a possibly good place to attempt to do 3200 * either combinatorial reconstruction or error correction 3201 * based on checksums. It also might be a good place 3202 * to send out preliminary ereports before we suspend 3203 * processing. 3204 */ 3205 } 3206 3207 /* 3208 * If there were logical child errors, they apply to us now. 3209 * We defer this until now to avoid conflating logical child 3210 * errors with errors that happened to the zio itself when 3211 * updating vdev stats and reporting FMA events above. 3212 */ 3213 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL); 3214 3215 if ((zio->io_error || zio->io_reexecute) && 3216 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio && 3217 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE))) 3218 zio_dva_unallocate(zio, zio->io_gang_tree, bp); 3219 3220 zio_gang_tree_free(&zio->io_gang_tree); 3221 3222 /* 3223 * Godfather I/Os should never suspend. 3224 */ 3225 if ((zio->io_flags & ZIO_FLAG_GODFATHER) && 3226 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) 3227 zio->io_reexecute = 0; 3228 3229 if (zio->io_reexecute) { 3230 /* 3231 * This is a logical I/O that wants to reexecute. 3232 * 3233 * Reexecute is top-down. When an i/o fails, if it's not 3234 * the root, it simply notifies its parent and sticks around. 3235 * The parent, seeing that it still has children in zio_done(), 3236 * does the same. This percolates all the way up to the root. 3237 * The root i/o will reexecute or suspend the entire tree. 3238 * 3239 * This approach ensures that zio_reexecute() honors 3240 * all the original i/o dependency relationships, e.g. 3241 * parents not executing until children are ready. 3242 */ 3243 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 3244 3245 zio->io_gang_leader = NULL; 3246 3247 mutex_enter(&zio->io_lock); 3248 zio->io_state[ZIO_WAIT_DONE] = 1; 3249 mutex_exit(&zio->io_lock); 3250 3251 /* 3252 * "The Godfather" I/O monitors its children but is 3253 * not a true parent to them. It will track them through 3254 * the pipeline but severs its ties whenever they get into 3255 * trouble (e.g. suspended). This allows "The Godfather" 3256 * I/O to return status without blocking. 3257 */ 3258 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) { 3259 zio_link_t *zl = zio->io_walk_link; 3260 pio_next = zio_walk_parents(zio); 3261 3262 if ((pio->io_flags & ZIO_FLAG_GODFATHER) && 3263 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) { 3264 zio_remove_child(pio, zio, zl); 3265 zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 3266 } 3267 } 3268 3269 if ((pio = zio_unique_parent(zio)) != NULL) { 3270 /* 3271 * We're not a root i/o, so there's nothing to do 3272 * but notify our parent. Don't propagate errors 3273 * upward since we haven't permanently failed yet. 3274 */ 3275 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); 3276 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE; 3277 zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 3278 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) { 3279 /* 3280 * We'd fail again if we reexecuted now, so suspend 3281 * until conditions improve (e.g. device comes online). 3282 */ 3283 zio_suspend(spa, zio); 3284 } else { 3285 /* 3286 * Reexecution is potentially a huge amount of work. 3287 * Hand it off to the otherwise-unused claim taskq. 3288 */ 3289#if defined(illumos) || !defined(_KERNEL) 3290 ASSERT(zio->io_tqent.tqent_next == NULL); 3291#else 3292 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0); 3293#endif 3294 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM, 3295 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio, 3296 0, &zio->io_tqent); 3297 } 3298 return (ZIO_PIPELINE_STOP); 3299 } 3300 3301 ASSERT(zio->io_child_count == 0); 3302 ASSERT(zio->io_reexecute == 0); 3303 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL)); 3304 3305 /* 3306 * Report any checksum errors, since the I/O is complete. 3307 */ 3308 while (zio->io_cksum_report != NULL) { 3309 zio_cksum_report_t *zcr = zio->io_cksum_report; 3310 zio->io_cksum_report = zcr->zcr_next; 3311 zcr->zcr_next = NULL; 3312 zcr->zcr_finish(zcr, NULL); 3313 zfs_ereport_free_checksum(zcr); 3314 } 3315 3316 /* 3317 * It is the responsibility of the done callback to ensure that this 3318 * particular zio is no longer discoverable for adoption, and as 3319 * such, cannot acquire any new parents. 3320 */ 3321 if (zio->io_done) 3322 zio->io_done(zio); 3323 3324 mutex_enter(&zio->io_lock); 3325 zio->io_state[ZIO_WAIT_DONE] = 1; 3326 mutex_exit(&zio->io_lock); 3327 3328 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) { 3329 zio_link_t *zl = zio->io_walk_link; 3330 pio_next = zio_walk_parents(zio); 3331 zio_remove_child(pio, zio, zl); 3332 zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 3333 } 3334 3335 if (zio->io_waiter != NULL) { 3336 mutex_enter(&zio->io_lock); 3337 zio->io_executor = NULL; 3338 cv_broadcast(&zio->io_cv); 3339 mutex_exit(&zio->io_lock); 3340 } else { 3341 zio_destroy(zio); 3342 } 3343 3344 return (ZIO_PIPELINE_STOP); 3345} 3346 3347/* 3348 * ========================================================================== 3349 * I/O pipeline definition 3350 * ========================================================================== 3351 */ 3352static zio_pipe_stage_t *zio_pipeline[] = { 3353 NULL, 3354 zio_read_bp_init, 3355 zio_free_bp_init, 3356 zio_issue_async, 3357 zio_write_bp_init, 3358 zio_checksum_generate, 3359 zio_nop_write, 3360 zio_ddt_read_start, 3361 zio_ddt_read_done, 3362 zio_ddt_write, 3363 zio_ddt_free, 3364 zio_gang_assemble, 3365 zio_gang_issue, 3366 zio_dva_allocate, 3367 zio_dva_free, 3368 zio_dva_claim, 3369 zio_ready, 3370 zio_vdev_io_start, 3371 zio_vdev_io_done, 3372 zio_vdev_io_assess, 3373 zio_checksum_verify, 3374 zio_done 3375}; 3376 3377/* dnp is the dnode for zb1->zb_object */ 3378boolean_t 3379zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_phys_t *zb1, 3380 const zbookmark_phys_t *zb2) 3381{ 3382 uint64_t zb1nextL0, zb2thisobj; 3383 3384 ASSERT(zb1->zb_objset == zb2->zb_objset); 3385 ASSERT(zb2->zb_level == 0); 3386 3387 /* The objset_phys_t isn't before anything. */ 3388 if (dnp == NULL) 3389 return (B_FALSE); 3390 3391 zb1nextL0 = (zb1->zb_blkid + 1) << 3392 ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT)); 3393 3394 zb2thisobj = zb2->zb_object ? zb2->zb_object : 3395 zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT); 3396 3397 if (zb1->zb_object == DMU_META_DNODE_OBJECT) { 3398 uint64_t nextobj = zb1nextL0 * 3399 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT; 3400 return (nextobj <= zb2thisobj); 3401 } 3402 3403 if (zb1->zb_object < zb2thisobj) 3404 return (B_TRUE); 3405 if (zb1->zb_object > zb2thisobj) 3406 return (B_FALSE); 3407 if (zb2->zb_object == DMU_META_DNODE_OBJECT) 3408 return (B_FALSE); 3409 return (zb1nextL0 <= zb2->zb_blkid); 3410}
|