zio.c revision 10685:931790026ac6
165793Smsmith/* 265793Smsmith * CDDL HEADER START 381082Sscottl * 465793Smsmith * The contents of this file are subject to the terms of the 581082Sscottl * Common Development and Distribution License (the "License"). 665793Smsmith * You may not use this file except in compliance with the License. 765793Smsmith * 865793Smsmith * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 965793Smsmith * or http://www.opensolaris.org/os/licensing. 1065793Smsmith * See the License for the specific language governing permissions 1165793Smsmith * and limitations under the License. 1265793Smsmith * 1365793Smsmith * When distributing Covered Code, include this CDDL HEADER in each 1465793Smsmith * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 1565793Smsmith * If applicable, add the following below this CDDL HEADER, with the 1665793Smsmith * fields enclosed by brackets "[]" replaced with your own identifying 1765793Smsmith * information: Portions Copyright [yyyy] [name of copyright owner] 1865793Smsmith * 1965793Smsmith * CDDL HEADER END 2065793Smsmith */ 2165793Smsmith/* 2265793Smsmith * Copyright 2009 Sun Microsystems, Inc. All rights reserved. 2365793Smsmith * Use is subject to license terms. 2465793Smsmith */ 2565793Smsmith 2665793Smsmith#include <sys/zfs_context.h> 2765793Smsmith#include <sys/fm/fs/zfs.h> 2865793Smsmith#include <sys/spa.h> 2965793Smsmith#include <sys/txg.h> 30119418Sobrien#include <sys/spa_impl.h> 31119418Sobrien#include <sys/vdev_impl.h> 32119418Sobrien#include <sys/zio_impl.h> 3381151Sscottl#include <sys/zio_compress.h> 3481151Sscottl#include <sys/zio_checksum.h> 3565793Smsmith 3665793Smsmith/* 3765793Smsmith * ========================================================================== 38129879Sphk * I/O priority table 3965793Smsmith * ========================================================================== 4065793Smsmith */ 4165793Smsmithuint8_t zio_priority_table[ZIO_PRIORITY_TABLE_SIZE] = { 4265793Smsmith 0, /* ZIO_PRIORITY_NOW */ 4365793Smsmith 0, /* ZIO_PRIORITY_SYNC_READ */ 4482527Sscottl 0, /* ZIO_PRIORITY_SYNC_WRITE */ 4582527Sscottl 6, /* ZIO_PRIORITY_ASYNC_READ */ 4682527Sscottl 4, /* ZIO_PRIORITY_ASYNC_WRITE */ 4782527Sscottl 4, /* ZIO_PRIORITY_FREE */ 4865793Smsmith 0, /* ZIO_PRIORITY_CACHE_FILL */ 4965793Smsmith 0, /* ZIO_PRIORITY_LOG_WRITE */ 5065793Smsmith 10, /* ZIO_PRIORITY_RESILVER */ 5165793Smsmith 20, /* ZIO_PRIORITY_SCRUB */ 52138635Sscottl}; 5365793Smsmith 5465793Smsmith/* 5565793Smsmith * ========================================================================== 5665793Smsmith * I/O type descriptions 5765793Smsmith * ========================================================================== 5865793Smsmith */ 5965793Smsmithchar *zio_type_name[ZIO_TYPES] = { 6065793Smsmith "null", "read", "write", "free", "claim", "ioctl" }; 6165793Smsmith 6265793Smsmith#define SYNC_PASS_DEFERRED_FREE 1 /* defer frees after this pass */ 6365793Smsmith#define SYNC_PASS_DONT_COMPRESS 4 /* don't compress after this pass */ 6465793Smsmith#define SYNC_PASS_REWRITE 1 /* rewrite new bps after this pass */ 65111525Sscottl 66111525Sscottl/* 67111525Sscottl * ========================================================================== 68111220Sphk * I/O kmem caches 6965793Smsmith * ========================================================================== 7089112Smsmith */ 7165793Smsmithkmem_cache_t *zio_cache; 7265793Smsmithkmem_cache_t *zio_link_cache; 7383114Sscottlkmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; 7483114Sscottlkmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; 7583114Sscottl 7683114Sscottl#ifdef _KERNEL 7765793Smsmithextern vmem_t *zio_alloc_arena; 7865793Smsmith#endif 7965793Smsmith 8083114Sscottl/* 8183114Sscottl * An allocating zio is one that either currently has the DVA allocate 8283114Sscottl * stage set or will have it later in its lifetime. 8365793Smsmith */ 8465793Smsmith#define IO_IS_ALLOCATING(zio) \ 8565793Smsmith ((zio)->io_orig_pipeline & (1U << ZIO_STAGE_DVA_ALLOCATE)) 8665793Smsmith 8783114Sscottlvoid 8865793Smsmithzio_init(void) 8965793Smsmith{ 90206534Semaste size_t c; 9165793Smsmith vmem_t *data_alloc_arena = NULL; 9265793Smsmith 9365793Smsmith#ifdef _KERNEL 94111525Sscottl data_alloc_arena = zio_alloc_arena; 9565793Smsmith#endif 9683114Sscottl zio_cache = kmem_cache_create("zio_cache", 9765793Smsmith sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0); 98177567Semaste zio_link_cache = kmem_cache_create("zio_link_cache", 9983114Sscottl sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0); 100111525Sscottl 10165793Smsmith /* 102109088Sscottl * For small buffers, we want a cache for each multiple of 103109088Sscottl * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache 10483114Sscottl * for each quarter-power of 2. For large buffers, we want 105109088Sscottl * a cache for each multiple of PAGESIZE. 10665793Smsmith */ 10783114Sscottl for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { 108109088Sscottl size_t size = (c + 1) << SPA_MINBLOCKSHIFT; 109212773Semaste size_t p2 = size; 110212773Semaste size_t align = 0; 111212773Semaste 11283114Sscottl while (p2 & (p2 - 1)) 113109088Sscottl p2 &= p2 - 1; 11465793Smsmith 11583114Sscottl if (size <= 4 * SPA_MINBLOCKSIZE) { 11683114Sscottl align = SPA_MINBLOCKSIZE; 11765793Smsmith } else if (P2PHASE(size, PAGESIZE) == 0) { 11865793Smsmith align = PAGESIZE; 11983114Sscottl } else if (P2PHASE(size, p2 >> 2) == 0) { 12065793Smsmith align = p2 >> 2; 12165793Smsmith } 12265793Smsmith 123111525Sscottl if (align != 0) { 12465793Smsmith char name[36]; 12583114Sscottl (void) sprintf(name, "zio_buf_%lu", (ulong_t)size); 12665793Smsmith zio_buf_cache[c] = kmem_cache_create(name, size, 127177567Semaste align, NULL, NULL, NULL, NULL, NULL, KMC_NODEBUG); 12883114Sscottl 129111525Sscottl (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size); 13065793Smsmith zio_data_buf_cache[c] = kmem_cache_create(name, size, 13183114Sscottl align, NULL, NULL, NULL, NULL, data_alloc_arena, 13283114Sscottl KMC_NODEBUG); 13365793Smsmith } 13483114Sscottl } 13583114Sscottl 13665793Smsmith while (--c != 0) { 13765793Smsmith ASSERT(zio_buf_cache[c] != NULL); 13883114Sscottl if (zio_buf_cache[c - 1] == NULL) 13965793Smsmith zio_buf_cache[c - 1] = zio_buf_cache[c]; 14065793Smsmith 14165793Smsmith ASSERT(zio_data_buf_cache[c] != NULL); 14265793Smsmith if (zio_data_buf_cache[c - 1] == NULL) 14365793Smsmith zio_data_buf_cache[c - 1] = zio_data_buf_cache[c]; 14483114Sscottl } 14565793Smsmith 146111525Sscottl zio_inject_init(); 147177567Semaste} 14865793Smsmith 14983114Sscottlvoid 15083114Sscottlzio_fini(void) 15183114Sscottl{ 15283114Sscottl size_t c; 15383114Sscottl kmem_cache_t *last_cache = NULL; 15483114Sscottl kmem_cache_t *last_data_cache = NULL; 15583114Sscottl 15682527Sscottl for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { 15783114Sscottl if (zio_buf_cache[c] != last_cache) { 15883114Sscottl last_cache = zio_buf_cache[c]; 15983114Sscottl kmem_cache_destroy(zio_buf_cache[c]); 16083114Sscottl } 16183114Sscottl zio_buf_cache[c] = NULL; 16283114Sscottl 16365793Smsmith if (zio_data_buf_cache[c] != last_data_cache) { 16483114Sscottl last_data_cache = zio_data_buf_cache[c]; 16583114Sscottl kmem_cache_destroy(zio_data_buf_cache[c]); 16683114Sscottl } 167133540Sscottl zio_data_buf_cache[c] = NULL; 16883114Sscottl } 169133540Sscottl 170111532Sscottl kmem_cache_destroy(zio_link_cache); 17183114Sscottl kmem_cache_destroy(zio_cache); 17265793Smsmith 17365793Smsmith zio_inject_fini(); 17483114Sscottl} 17595350Sscottl 17695350Sscottl/* 17795350Sscottl * ========================================================================== 17895350Sscottl * Allocate and free I/O buffers 17995350Sscottl * ========================================================================== 18095350Sscottl */ 18195350Sscottl 18295350Sscottl/* 18395350Sscottl * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a 18495350Sscottl * crashdump if the kernel panics, so use it judiciously. Obviously, it's 18595350Sscottl * useful to inspect ZFS metadata, but if possible, we should avoid keeping 18695350Sscottl * excess / transient data in-core during a crashdump. 18795350Sscottl */ 18895350Sscottlvoid * 18995350Sscottlzio_buf_alloc(size_t size) 19095350Sscottl{ 19195350Sscottl size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 192116553Sscottl 193116553Sscottl ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 19495350Sscottl 19595350Sscottl return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE)); 19695350Sscottl} 19795350Sscottl 19895350Sscottl/* 19995350Sscottl * Use zio_data_buf_alloc to allocate data. The data will not appear in a 20095350Sscottl * crashdump if the kernel panics. This exists so that we will limit the amount 20195350Sscottl * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount 202177899Semaste * of kernel heap dumped to disk when the kernel panics) 203177899Semaste */ 204177899Semastevoid * 205177899Semastezio_data_buf_alloc(size_t size) 206177899Semaste{ 207177899Semaste size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 208177899Semaste 209177899Semaste ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 210177899Semaste 211177899Semaste return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE)); 212177899Semaste} 213177899Semaste 214177899Semastevoid 215177899Semastezio_buf_free(void *buf, size_t size) 216177899Semaste{ 217177899Semaste size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 218177899Semaste 219177899Semaste ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 220177899Semaste 221177899Semaste kmem_cache_free(zio_buf_cache[c], buf); 222177899Semaste} 223177899Semaste 224177899Semastevoid 225177899Semastezio_data_buf_free(void *buf, size_t size) 226177899Semaste{ 22782527Sscottl size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 22882527Sscottl 229195614Sjkim ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 23082527Sscottl 23182527Sscottl kmem_cache_free(zio_data_buf_cache[c], buf); 232111220Sphk} 23382527Sscottl 23483114Sscottl/* 23583114Sscottl * ========================================================================== 23695350Sscottl * Push and pop I/O transform buffers 237195614Sjkim * ========================================================================== 23895350Sscottl */ 23995350Sscottlstatic void 24095350Sscottlzio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize, 241111220Sphk zio_transform_func_t *transform) 242177899Semaste{ 243177899Semaste zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP); 24482527Sscottl 245111220Sphk zt->zt_orig_data = zio->io_data; 246111525Sscottl zt->zt_orig_size = zio->io_size; 24782527Sscottl zt->zt_bufsize = bufsize; 24883114Sscottl zt->zt_transform = transform; 24995350Sscottl 25082527Sscottl zt->zt_next = zio->io_transform_stack; 25183114Sscottl zio->io_transform_stack = zt; 25282527Sscottl 25395350Sscottl zio->io_data = data; 25495350Sscottl zio->io_size = size; 25595350Sscottl} 256212773Semaste 257212773Semastestatic void 25895350Sscottlzio_pop_transforms(zio_t *zio) 25995350Sscottl{ 26095350Sscottl zio_transform_t *zt; 26182527Sscottl 262130006Sscottl while ((zt = zio->io_transform_stack) != NULL) { 263130006Sscottl if (zt->zt_transform != NULL) 26482527Sscottl zt->zt_transform(zio, 26595350Sscottl zt->zt_orig_data, zt->zt_orig_size); 266195614Sjkim 267195614Sjkim zio_buf_free(zio->io_data, zt->zt_bufsize); 268177899Semaste 269177899Semaste zio->io_data = zt->zt_orig_data; 270177899Semaste zio->io_size = zt->zt_orig_size; 271177899Semaste zio->io_transform_stack = zt->zt_next; 272177899Semaste 273177899Semaste kmem_free(zt, sizeof (zio_transform_t)); 274177899Semaste } 275177899Semaste} 276177899Semaste 277177899Semaste/* 278177899Semaste * ========================================================================== 279177899Semaste * I/O transform callbacks for subblocks and decompression 280177899Semaste * ========================================================================== 281177899Semaste */ 282177899Semastestatic void 283177899Semastezio_subblock(zio_t *zio, void *data, uint64_t size) 284177899Semaste{ 285177899Semaste ASSERT(zio->io_size > size); 286177899Semaste 287177899Semaste if (zio->io_type == ZIO_TYPE_READ) 288177899Semaste bcopy(zio->io_data, data, size); 289177899Semaste} 290177899Semaste 291177899Semastestatic void 292116553Sscottlzio_decompress(zio_t *zio, void *data, uint64_t size) 293116553Sscottl{ 294116553Sscottl if (zio->io_error == 0 && 295116553Sscottl zio_decompress_data(BP_GET_COMPRESS(zio->io_bp), 296116553Sscottl zio->io_data, zio->io_size, data, size) != 0) 297116553Sscottl zio->io_error = EIO; 298116553Sscottl} 299116553Sscottl 300177899Semaste/* 301145811Sscottl * ========================================================================== 302116553Sscottl * I/O parent/child relationships and pipeline interlocks 30395350Sscottl * ========================================================================== 30495350Sscottl */ 30582527Sscottl/* 30695350Sscottl * NOTE - Callers to zio_walk_parents() and zio_walk_children must 307177899Semaste * continue calling these functions until they return NULL. 30883114Sscottl * Otherwise, the next caller will pick up the list walk in 309177899Semaste * some indeterminate state. (Otherwise every caller would 310212773Semaste * have to pass in a cookie to keep the state represented by 311212773Semaste * io_walk_link, which gets annoying.) 312212773Semaste */ 31395350Sscottlzio_t * 31483114Sscottlzio_walk_parents(zio_t *cio) 315116553Sscottl{ 316145811Sscottl zio_link_t *zl = cio->io_walk_link; 317145811Sscottl list_t *pl = &cio->io_parent_list; 318145811Sscottl 319145811Sscottl zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl); 320145811Sscottl cio->io_walk_link = zl; 32195350Sscottl 32295350Sscottl if (zl == NULL) 323132771Skan return (NULL); 32483114Sscottl 32582527Sscottl ASSERT(zl->zl_child == cio); 32683114Sscottl return (zl->zl_parent); 32782527Sscottl} 32882527Sscottl 32983114Sscottlzio_t * 33065793Smsmithzio_walk_children(zio_t *pio) 33165793Smsmith{ 33265793Smsmith zio_link_t *zl = pio->io_walk_link; 33370393Smsmith list_t *cl = &pio->io_child_list; 33465793Smsmith 33583114Sscottl zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl); 33665793Smsmith pio->io_walk_link = zl; 337111525Sscottl 338177567Semaste if (zl == NULL) 33983114Sscottl return (NULL); 340111691Sscottl 341103675Sphk ASSERT(zl->zl_parent == pio); 342111691Sscottl return (zl->zl_child); 34383114Sscottl} 34465793Smsmith 34565793Smsmithzio_t * 34683114Sscottlzio_unique_parent(zio_t *cio) 34765793Smsmith{ 34865793Smsmith zio_t *pio = zio_walk_parents(cio); 34965793Smsmith 35065793Smsmith VERIFY(zio_walk_parents(cio) == NULL); 35165793Smsmith return (pio); 35265793Smsmith} 353177567Semaste 35465793Smsmithvoid 35583114Sscottlzio_add_child(zio_t *pio, zio_t *cio) 35665793Smsmith{ 35765793Smsmith zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP); 35883114Sscottl 35965793Smsmith /* 36065793Smsmith * Logical I/Os can have logical, gang, or vdev children. 36165793Smsmith * Gang I/Os can have gang or vdev children. 36265793Smsmith * Vdev I/Os can only have vdev children. 36365793Smsmith * The following ASSERT captures all of these constraints. 36483114Sscottl */ 36583114Sscottl ASSERT(cio->io_child_type <= pio->io_child_type); 36683114Sscottl 367177567Semaste zl->zl_parent = pio; 36865793Smsmith zl->zl_child = cio; 36983114Sscottl 37083114Sscottl mutex_enter(&cio->io_lock); 37183114Sscottl mutex_enter(&pio->io_lock); 37283114Sscottl 37383114Sscottl ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0); 37465793Smsmith 37583114Sscottl for (int w = 0; w < ZIO_WAIT_TYPES; w++) 37683114Sscottl pio->io_children[cio->io_child_type][w] += !cio->io_state[w]; 37783114Sscottl 37883114Sscottl list_insert_head(&pio->io_child_list, zl); 37983114Sscottl list_insert_head(&cio->io_parent_list, zl); 380177619Semaste 381177619Semaste mutex_exit(&pio->io_lock); 382177619Semaste mutex_exit(&cio->io_lock); 38383114Sscottl} 38483114Sscottl 38583114Sscottlstatic void 38683114Sscottlzio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl) 38783114Sscottl{ 38883114Sscottl ASSERT(zl->zl_parent == pio); 38983114Sscottl ASSERT(zl->zl_child == cio); 39083114Sscottl 39183114Sscottl mutex_enter(&cio->io_lock); 39283114Sscottl mutex_enter(&pio->io_lock); 39383114Sscottl 39465793Smsmith list_remove(&pio->io_child_list, zl); 395177619Semaste list_remove(&cio->io_parent_list, zl); 396177619Semaste 397177619Semaste mutex_exit(&pio->io_lock); 39865793Smsmith mutex_exit(&cio->io_lock); 39983114Sscottl 40083114Sscottl kmem_cache_free(zio_link_cache, zl); 401125975Sphk} 402125975Sphk 403125975Sphkstatic boolean_t 404195614Sjkimzio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait) 405125975Sphk{ 406125975Sphk uint64_t *countp = &zio->io_children[child][wait]; 407125975Sphk boolean_t waiting = B_FALSE; 408125975Sphk 409125975Sphk mutex_enter(&zio->io_lock); 410125975Sphk ASSERT(zio->io_stall == NULL); 411125975Sphk if (*countp != 0) { 412125975Sphk zio->io_stage--; 413125975Sphk zio->io_stall = countp; 414125975Sphk waiting = B_TRUE; 41581082Sscottl } 41683114Sscottl mutex_exit(&zio->io_lock); 41765793Smsmith 41865793Smsmith return (waiting); 41983114Sscottl} 42065793Smsmith 42165793Smsmithstatic void 42265793Smsmithzio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait) 42365793Smsmith{ 42465793Smsmith uint64_t *countp = &pio->io_children[zio->io_child_type][wait]; 42583114Sscottl int *errorp = &pio->io_child_error[zio->io_child_type]; 42665793Smsmith 42783114Sscottl mutex_enter(&pio->io_lock); 428177567Semaste if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE)) 42965793Smsmith *errorp = zio_worst_error(*errorp, zio->io_error); 43083114Sscottl pio->io_reexecute |= zio->io_reexecute; 43183114Sscottl ASSERT3U(*countp, >, 0); 43283114Sscottl if (--*countp == 0 && pio->io_stall == countp) { 433125975Sphk pio->io_stall = NULL; 43465793Smsmith mutex_exit(&pio->io_lock); 43583114Sscottl zio_execute(pio); 43665793Smsmith } else { 437 mutex_exit(&pio->io_lock); 438 } 439} 440 441static void 442zio_inherit_child_errors(zio_t *zio, enum zio_child c) 443{ 444 if (zio->io_child_error[c] != 0 && zio->io_error == 0) 445 zio->io_error = zio->io_child_error[c]; 446} 447 448/* 449 * ========================================================================== 450 * Create the various types of I/O (read, write, free, etc) 451 * ========================================================================== 452 */ 453static zio_t * 454zio_create(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, 455 void *data, uint64_t size, zio_done_func_t *done, void *private, 456 zio_type_t type, int priority, int flags, vdev_t *vd, uint64_t offset, 457 const zbookmark_t *zb, uint8_t stage, uint32_t pipeline) 458{ 459 zio_t *zio; 460 461 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE); 462 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0); 463 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0); 464 465 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER)); 466 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER)); 467 ASSERT(vd || stage == ZIO_STAGE_OPEN); 468 469 zio = kmem_cache_alloc(zio_cache, KM_SLEEP); 470 bzero(zio, sizeof (zio_t)); 471 472 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL); 473 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL); 474 475 list_create(&zio->io_parent_list, sizeof (zio_link_t), 476 offsetof(zio_link_t, zl_parent_node)); 477 list_create(&zio->io_child_list, sizeof (zio_link_t), 478 offsetof(zio_link_t, zl_child_node)); 479 480 if (vd != NULL) 481 zio->io_child_type = ZIO_CHILD_VDEV; 482 else if (flags & ZIO_FLAG_GANG_CHILD) 483 zio->io_child_type = ZIO_CHILD_GANG; 484 else 485 zio->io_child_type = ZIO_CHILD_LOGICAL; 486 487 if (bp != NULL) { 488 zio->io_bp = bp; 489 zio->io_bp_copy = *bp; 490 zio->io_bp_orig = *bp; 491 if (type != ZIO_TYPE_WRITE) 492 zio->io_bp = &zio->io_bp_copy; /* so caller can free */ 493 if (zio->io_child_type == ZIO_CHILD_LOGICAL) 494 zio->io_logical = zio; 495 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp)) 496 pipeline |= ZIO_GANG_STAGES; 497 } 498 499 zio->io_spa = spa; 500 zio->io_txg = txg; 501 zio->io_data = data; 502 zio->io_size = size; 503 zio->io_done = done; 504 zio->io_private = private; 505 zio->io_type = type; 506 zio->io_priority = priority; 507 zio->io_vd = vd; 508 zio->io_offset = offset; 509 zio->io_orig_flags = zio->io_flags = flags; 510 zio->io_orig_stage = zio->io_stage = stage; 511 zio->io_orig_pipeline = zio->io_pipeline = pipeline; 512 513 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY); 514 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE); 515 516 if (zb != NULL) 517 zio->io_bookmark = *zb; 518 519 if (pio != NULL) { 520 if (zio->io_logical == NULL) 521 zio->io_logical = pio->io_logical; 522 if (zio->io_child_type == ZIO_CHILD_GANG) 523 zio->io_gang_leader = pio->io_gang_leader; 524 zio_add_child(pio, zio); 525 } 526 527 return (zio); 528} 529 530static void 531zio_destroy(zio_t *zio) 532{ 533 list_destroy(&zio->io_parent_list); 534 list_destroy(&zio->io_child_list); 535 mutex_destroy(&zio->io_lock); 536 cv_destroy(&zio->io_cv); 537 kmem_cache_free(zio_cache, zio); 538} 539 540zio_t * 541zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done, 542 void *private, int flags) 543{ 544 zio_t *zio; 545 546 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private, 547 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, 548 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE); 549 550 return (zio); 551} 552 553zio_t * 554zio_root(spa_t *spa, zio_done_func_t *done, void *private, int flags) 555{ 556 return (zio_null(NULL, spa, NULL, done, private, flags)); 557} 558 559zio_t * 560zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, 561 void *data, uint64_t size, zio_done_func_t *done, void *private, 562 int priority, int flags, const zbookmark_t *zb) 563{ 564 zio_t *zio; 565 566 zio = zio_create(pio, spa, bp->blk_birth, (blkptr_t *)bp, 567 data, size, done, private, 568 ZIO_TYPE_READ, priority, flags, NULL, 0, zb, 569 ZIO_STAGE_OPEN, ZIO_READ_PIPELINE); 570 571 return (zio); 572} 573 574void 575zio_skip_write(zio_t *zio) 576{ 577 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 578 ASSERT(zio->io_stage == ZIO_STAGE_READY); 579 ASSERT(!BP_IS_GANG(zio->io_bp)); 580 581 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; 582} 583 584zio_t * 585zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, 586 void *data, uint64_t size, zio_prop_t *zp, 587 zio_done_func_t *ready, zio_done_func_t *done, void *private, 588 int priority, int flags, const zbookmark_t *zb) 589{ 590 zio_t *zio; 591 592 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF && 593 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS && 594 zp->zp_compress >= ZIO_COMPRESS_OFF && 595 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS && 596 zp->zp_type < DMU_OT_NUMTYPES && 597 zp->zp_level < 32 && 598 zp->zp_ndvas > 0 && 599 zp->zp_ndvas <= spa_max_replication(spa)); 600 ASSERT(ready != NULL); 601 602 zio = zio_create(pio, spa, txg, bp, data, size, done, private, 603 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, 604 ZIO_STAGE_OPEN, ZIO_WRITE_PIPELINE); 605 606 zio->io_ready = ready; 607 zio->io_prop = *zp; 608 609 return (zio); 610} 611 612zio_t * 613zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data, 614 uint64_t size, zio_done_func_t *done, void *private, int priority, 615 int flags, zbookmark_t *zb) 616{ 617 zio_t *zio; 618 619 zio = zio_create(pio, spa, txg, bp, data, size, done, private, 620 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, 621 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE); 622 623 return (zio); 624} 625 626zio_t * 627zio_free(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, 628 zio_done_func_t *done, void *private, int flags) 629{ 630 zio_t *zio; 631 632 ASSERT(!BP_IS_HOLE(bp)); 633 634 if (bp->blk_fill == BLK_FILL_ALREADY_FREED) 635 return (zio_null(pio, spa, NULL, NULL, NULL, flags)); 636 637 if (txg == spa->spa_syncing_txg && 638 spa_sync_pass(spa) > SYNC_PASS_DEFERRED_FREE) { 639 bplist_enqueue_deferred(&spa->spa_sync_bplist, bp); 640 return (zio_null(pio, spa, NULL, NULL, NULL, flags)); 641 } 642 643 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), 644 done, private, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags, 645 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE); 646 647 return (zio); 648} 649 650zio_t * 651zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, 652 zio_done_func_t *done, void *private, int flags) 653{ 654 zio_t *zio; 655 656 /* 657 * A claim is an allocation of a specific block. Claims are needed 658 * to support immediate writes in the intent log. The issue is that 659 * immediate writes contain committed data, but in a txg that was 660 * *not* committed. Upon opening the pool after an unclean shutdown, 661 * the intent log claims all blocks that contain immediate write data 662 * so that the SPA knows they're in use. 663 * 664 * All claims *must* be resolved in the first txg -- before the SPA 665 * starts allocating blocks -- so that nothing is allocated twice. 666 */ 667 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa)); 668 ASSERT3U(spa_first_txg(spa), <=, txg); 669 670 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), 671 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags, 672 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE); 673 674 return (zio); 675} 676 677zio_t * 678zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, 679 zio_done_func_t *done, void *private, int priority, int flags) 680{ 681 zio_t *zio; 682 int c; 683 684 if (vd->vdev_children == 0) { 685 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private, 686 ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL, 687 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE); 688 689 zio->io_cmd = cmd; 690 } else { 691 zio = zio_null(pio, spa, NULL, NULL, NULL, flags); 692 693 for (c = 0; c < vd->vdev_children; c++) 694 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd, 695 done, private, priority, flags)); 696 } 697 698 return (zio); 699} 700 701zio_t * 702zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 703 void *data, int checksum, zio_done_func_t *done, void *private, 704 int priority, int flags, boolean_t labels) 705{ 706 zio_t *zio; 707 708 ASSERT(vd->vdev_children == 0); 709 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || 710 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); 711 ASSERT3U(offset + size, <=, vd->vdev_psize); 712 713 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private, 714 ZIO_TYPE_READ, priority, flags, vd, offset, NULL, 715 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE); 716 717 zio->io_prop.zp_checksum = checksum; 718 719 return (zio); 720} 721 722zio_t * 723zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 724 void *data, int checksum, zio_done_func_t *done, void *private, 725 int priority, int flags, boolean_t labels) 726{ 727 zio_t *zio; 728 729 ASSERT(vd->vdev_children == 0); 730 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || 731 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); 732 ASSERT3U(offset + size, <=, vd->vdev_psize); 733 734 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private, 735 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL, 736 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE); 737 738 zio->io_prop.zp_checksum = checksum; 739 740 if (zio_checksum_table[checksum].ci_zbt) { 741 /* 742 * zbt checksums are necessarily destructive -- they modify 743 * the end of the write buffer to hold the verifier/checksum. 744 * Therefore, we must make a local copy in case the data is 745 * being written to multiple places in parallel. 746 */ 747 void *wbuf = zio_buf_alloc(size); 748 bcopy(data, wbuf, size); 749 zio_push_transform(zio, wbuf, size, size, NULL); 750 } 751 752 return (zio); 753} 754 755/* 756 * Create a child I/O to do some work for us. 757 */ 758zio_t * 759zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset, 760 void *data, uint64_t size, int type, int priority, int flags, 761 zio_done_func_t *done, void *private) 762{ 763 uint32_t pipeline = ZIO_VDEV_CHILD_PIPELINE; 764 zio_t *zio; 765 766 ASSERT(vd->vdev_parent == 767 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev)); 768 769 if (type == ZIO_TYPE_READ && bp != NULL) { 770 /* 771 * If we have the bp, then the child should perform the 772 * checksum and the parent need not. This pushes error 773 * detection as close to the leaves as possible and 774 * eliminates redundant checksums in the interior nodes. 775 */ 776 pipeline |= 1U << ZIO_STAGE_CHECKSUM_VERIFY; 777 pio->io_pipeline &= ~(1U << ZIO_STAGE_CHECKSUM_VERIFY); 778 } 779 780 if (vd->vdev_children == 0) 781 offset += VDEV_LABEL_START_SIZE; 782 783 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, 784 done, private, type, priority, 785 (pio->io_flags & ZIO_FLAG_VDEV_INHERIT) | 786 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | flags, 787 vd, offset, &pio->io_bookmark, 788 ZIO_STAGE_VDEV_IO_START - 1, pipeline); 789 790 return (zio); 791} 792 793zio_t * 794zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size, 795 int type, int priority, int flags, zio_done_func_t *done, void *private) 796{ 797 zio_t *zio; 798 799 ASSERT(vd->vdev_ops->vdev_op_leaf); 800 801 zio = zio_create(NULL, vd->vdev_spa, 0, NULL, 802 data, size, done, private, type, priority, 803 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY, 804 vd, offset, NULL, 805 ZIO_STAGE_VDEV_IO_START - 1, ZIO_VDEV_CHILD_PIPELINE); 806 807 return (zio); 808} 809 810void 811zio_flush(zio_t *zio, vdev_t *vd) 812{ 813 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 814 NULL, NULL, ZIO_PRIORITY_NOW, 815 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY)); 816} 817 818/* 819 * ========================================================================== 820 * Prepare to read and write logical blocks 821 * ========================================================================== 822 */ 823 824static int 825zio_read_bp_init(zio_t *zio) 826{ 827 blkptr_t *bp = zio->io_bp; 828 829 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF && 830 zio->io_child_type == ZIO_CHILD_LOGICAL && 831 !(zio->io_flags & ZIO_FLAG_RAW)) { 832 uint64_t csize = BP_GET_PSIZE(bp); 833 void *cbuf = zio_buf_alloc(csize); 834 835 zio_push_transform(zio, cbuf, csize, csize, zio_decompress); 836 } 837 838 if (!dmu_ot[BP_GET_TYPE(bp)].ot_metadata && BP_GET_LEVEL(bp) == 0) 839 zio->io_flags |= ZIO_FLAG_DONT_CACHE; 840 841 return (ZIO_PIPELINE_CONTINUE); 842} 843 844static int 845zio_write_bp_init(zio_t *zio) 846{ 847 zio_prop_t *zp = &zio->io_prop; 848 int compress = zp->zp_compress; 849 blkptr_t *bp = zio->io_bp; 850 void *cbuf; 851 uint64_t lsize = zio->io_size; 852 uint64_t csize = lsize; 853 uint64_t cbufsize = 0; 854 int pass = 1; 855 856 /* 857 * If our children haven't all reached the ready stage, 858 * wait for them and then repeat this pipeline stage. 859 */ 860 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) || 861 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY)) 862 return (ZIO_PIPELINE_STOP); 863 864 if (!IO_IS_ALLOCATING(zio)) 865 return (ZIO_PIPELINE_CONTINUE); 866 867 ASSERT(compress != ZIO_COMPRESS_INHERIT); 868 869 if (bp->blk_birth == zio->io_txg) { 870 /* 871 * We're rewriting an existing block, which means we're 872 * working on behalf of spa_sync(). For spa_sync() to 873 * converge, it must eventually be the case that we don't 874 * have to allocate new blocks. But compression changes 875 * the blocksize, which forces a reallocate, and makes 876 * convergence take longer. Therefore, after the first 877 * few passes, stop compressing to ensure convergence. 878 */ 879 pass = spa_sync_pass(zio->io_spa); 880 881 if (pass > SYNC_PASS_DONT_COMPRESS) 882 compress = ZIO_COMPRESS_OFF; 883 884 /* Make sure someone doesn't change their mind on overwrites */ 885 ASSERT(MIN(zp->zp_ndvas + BP_IS_GANG(bp), 886 spa_max_replication(zio->io_spa)) == BP_GET_NDVAS(bp)); 887 } 888 889 if (compress != ZIO_COMPRESS_OFF) { 890 if (!zio_compress_data(compress, zio->io_data, zio->io_size, 891 &cbuf, &csize, &cbufsize)) { 892 compress = ZIO_COMPRESS_OFF; 893 } else if (csize != 0) { 894 zio_push_transform(zio, cbuf, csize, cbufsize, NULL); 895 } 896 } 897 898 /* 899 * The final pass of spa_sync() must be all rewrites, but the first 900 * few passes offer a trade-off: allocating blocks defers convergence, 901 * but newly allocated blocks are sequential, so they can be written 902 * to disk faster. Therefore, we allow the first few passes of 903 * spa_sync() to allocate new blocks, but force rewrites after that. 904 * There should only be a handful of blocks after pass 1 in any case. 905 */ 906 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == csize && 907 pass > SYNC_PASS_REWRITE) { 908 ASSERT(csize != 0); 909 uint32_t gang_stages = zio->io_pipeline & ZIO_GANG_STAGES; 910 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages; 911 zio->io_flags |= ZIO_FLAG_IO_REWRITE; 912 } else { 913 BP_ZERO(bp); 914 zio->io_pipeline = ZIO_WRITE_PIPELINE; 915 } 916 917 if (csize == 0) { 918 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 919 } else { 920 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER); 921 BP_SET_LSIZE(bp, lsize); 922 BP_SET_PSIZE(bp, csize); 923 BP_SET_COMPRESS(bp, compress); 924 BP_SET_CHECKSUM(bp, zp->zp_checksum); 925 BP_SET_TYPE(bp, zp->zp_type); 926 BP_SET_LEVEL(bp, zp->zp_level); 927 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER); 928 } 929 930 return (ZIO_PIPELINE_CONTINUE); 931} 932 933/* 934 * ========================================================================== 935 * Execute the I/O pipeline 936 * ========================================================================== 937 */ 938 939static void 940zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q) 941{ 942 zio_type_t t = zio->io_type; 943 944 /* 945 * If we're a config writer or a probe, the normal issue and 946 * interrupt threads may all be blocked waiting for the config lock. 947 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL. 948 */ 949 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE)) 950 t = ZIO_TYPE_NULL; 951 952 /* 953 * A similar issue exists for the L2ARC write thread until L2ARC 2.0. 954 */ 955 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux) 956 t = ZIO_TYPE_NULL; 957 958 (void) taskq_dispatch(zio->io_spa->spa_zio_taskq[t][q], 959 (task_func_t *)zio_execute, zio, TQ_SLEEP); 960} 961 962static boolean_t 963zio_taskq_member(zio_t *zio, enum zio_taskq_type q) 964{ 965 kthread_t *executor = zio->io_executor; 966 spa_t *spa = zio->io_spa; 967 968 for (zio_type_t t = 0; t < ZIO_TYPES; t++) 969 if (taskq_member(spa->spa_zio_taskq[t][q], executor)) 970 return (B_TRUE); 971 972 return (B_FALSE); 973} 974 975static int 976zio_issue_async(zio_t *zio) 977{ 978 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE); 979 980 return (ZIO_PIPELINE_STOP); 981} 982 983void 984zio_interrupt(zio_t *zio) 985{ 986 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT); 987} 988 989/* 990 * Execute the I/O pipeline until one of the following occurs: 991 * (1) the I/O completes; (2) the pipeline stalls waiting for 992 * dependent child I/Os; (3) the I/O issues, so we're waiting 993 * for an I/O completion interrupt; (4) the I/O is delegated by 994 * vdev-level caching or aggregation; (5) the I/O is deferred 995 * due to vdev-level queueing; (6) the I/O is handed off to 996 * another thread. In all cases, the pipeline stops whenever 997 * there's no CPU work; it never burns a thread in cv_wait(). 998 * 999 * There's no locking on io_stage because there's no legitimate way 1000 * for multiple threads to be attempting to process the same I/O. 1001 */ 1002static zio_pipe_stage_t *zio_pipeline[ZIO_STAGES]; 1003 1004void 1005zio_execute(zio_t *zio) 1006{ 1007 zio->io_executor = curthread; 1008 1009 while (zio->io_stage < ZIO_STAGE_DONE) { 1010 uint32_t pipeline = zio->io_pipeline; 1011 zio_stage_t stage = zio->io_stage; 1012 int rv; 1013 1014 ASSERT(!MUTEX_HELD(&zio->io_lock)); 1015 1016 while (((1U << ++stage) & pipeline) == 0) 1017 continue; 1018 1019 ASSERT(stage <= ZIO_STAGE_DONE); 1020 ASSERT(zio->io_stall == NULL); 1021 1022 /* 1023 * If we are in interrupt context and this pipeline stage 1024 * will grab a config lock that is held across I/O, 1025 * issue async to avoid deadlock. 1026 */ 1027 if (((1U << stage) & ZIO_CONFIG_LOCK_BLOCKING_STAGES) && 1028 zio->io_vd == NULL && 1029 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) { 1030 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE); 1031 return; 1032 } 1033 1034 zio->io_stage = stage; 1035 rv = zio_pipeline[stage](zio); 1036 1037 if (rv == ZIO_PIPELINE_STOP) 1038 return; 1039 1040 ASSERT(rv == ZIO_PIPELINE_CONTINUE); 1041 } 1042} 1043 1044/* 1045 * ========================================================================== 1046 * Initiate I/O, either sync or async 1047 * ========================================================================== 1048 */ 1049int 1050zio_wait(zio_t *zio) 1051{ 1052 int error; 1053 1054 ASSERT(zio->io_stage == ZIO_STAGE_OPEN); 1055 ASSERT(zio->io_executor == NULL); 1056 1057 zio->io_waiter = curthread; 1058 1059 zio_execute(zio); 1060 1061 mutex_enter(&zio->io_lock); 1062 while (zio->io_executor != NULL) 1063 cv_wait(&zio->io_cv, &zio->io_lock); 1064 mutex_exit(&zio->io_lock); 1065 1066 error = zio->io_error; 1067 zio_destroy(zio); 1068 1069 return (error); 1070} 1071 1072void 1073zio_nowait(zio_t *zio) 1074{ 1075 ASSERT(zio->io_executor == NULL); 1076 1077 if (zio->io_child_type == ZIO_CHILD_LOGICAL && 1078 zio_unique_parent(zio) == NULL) { 1079 /* 1080 * This is a logical async I/O with no parent to wait for it. 1081 * We add it to the spa_async_root_zio "Godfather" I/O which 1082 * will ensure they complete prior to unloading the pool. 1083 */ 1084 spa_t *spa = zio->io_spa; 1085 1086 zio_add_child(spa->spa_async_zio_root, zio); 1087 } 1088 1089 zio_execute(zio); 1090} 1091 1092/* 1093 * ========================================================================== 1094 * Reexecute or suspend/resume failed I/O 1095 * ========================================================================== 1096 */ 1097 1098static void 1099zio_reexecute(zio_t *pio) 1100{ 1101 zio_t *cio, *cio_next; 1102 1103 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL); 1104 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN); 1105 ASSERT(pio->io_gang_leader == NULL); 1106 ASSERT(pio->io_gang_tree == NULL); 1107 1108 pio->io_flags = pio->io_orig_flags; 1109 pio->io_stage = pio->io_orig_stage; 1110 pio->io_pipeline = pio->io_orig_pipeline; 1111 pio->io_reexecute = 0; 1112 pio->io_error = 0; 1113 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 1114 pio->io_state[w] = 0; 1115 for (int c = 0; c < ZIO_CHILD_TYPES; c++) 1116 pio->io_child_error[c] = 0; 1117 1118 if (IO_IS_ALLOCATING(pio)) { 1119 /* 1120 * Remember the failed bp so that the io_ready() callback 1121 * can update its accounting upon reexecution. The block 1122 * was already freed in zio_done(); we indicate this with 1123 * a fill count of -1 so that zio_free() knows to skip it. 1124 */ 1125 blkptr_t *bp = pio->io_bp; 1126 ASSERT(bp->blk_birth == 0 || bp->blk_birth == pio->io_txg); 1127 bp->blk_fill = BLK_FILL_ALREADY_FREED; 1128 pio->io_bp_orig = *bp; 1129 BP_ZERO(bp); 1130 } 1131 1132 /* 1133 * As we reexecute pio's children, new children could be created. 1134 * New children go to the head of pio's io_child_list, however, 1135 * so we will (correctly) not reexecute them. The key is that 1136 * the remainder of pio's io_child_list, from 'cio_next' onward, 1137 * cannot be affected by any side effects of reexecuting 'cio'. 1138 */ 1139 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) { 1140 cio_next = zio_walk_children(pio); 1141 mutex_enter(&pio->io_lock); 1142 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 1143 pio->io_children[cio->io_child_type][w]++; 1144 mutex_exit(&pio->io_lock); 1145 zio_reexecute(cio); 1146 } 1147 1148 /* 1149 * Now that all children have been reexecuted, execute the parent. 1150 * We don't reexecute "The Godfather" I/O here as it's the 1151 * responsibility of the caller to wait on him. 1152 */ 1153 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) 1154 zio_execute(pio); 1155} 1156 1157void 1158zio_suspend(spa_t *spa, zio_t *zio) 1159{ 1160 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC) 1161 fm_panic("Pool '%s' has encountered an uncorrectable I/O " 1162 "failure and the failure mode property for this pool " 1163 "is set to panic.", spa_name(spa)); 1164 1165 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0); 1166 1167 mutex_enter(&spa->spa_suspend_lock); 1168 1169 if (spa->spa_suspend_zio_root == NULL) 1170 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL, 1171 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | 1172 ZIO_FLAG_GODFATHER); 1173 1174 spa->spa_suspended = B_TRUE; 1175 1176 if (zio != NULL) { 1177 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); 1178 ASSERT(zio != spa->spa_suspend_zio_root); 1179 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1180 ASSERT(zio_unique_parent(zio) == NULL); 1181 ASSERT(zio->io_stage == ZIO_STAGE_DONE); 1182 zio_add_child(spa->spa_suspend_zio_root, zio); 1183 } 1184 1185 mutex_exit(&spa->spa_suspend_lock); 1186} 1187 1188int 1189zio_resume(spa_t *spa) 1190{ 1191 zio_t *pio; 1192 1193 /* 1194 * Reexecute all previously suspended i/o. 1195 */ 1196 mutex_enter(&spa->spa_suspend_lock); 1197 spa->spa_suspended = B_FALSE; 1198 cv_broadcast(&spa->spa_suspend_cv); 1199 pio = spa->spa_suspend_zio_root; 1200 spa->spa_suspend_zio_root = NULL; 1201 mutex_exit(&spa->spa_suspend_lock); 1202 1203 if (pio == NULL) 1204 return (0); 1205 1206 zio_reexecute(pio); 1207 return (zio_wait(pio)); 1208} 1209 1210void 1211zio_resume_wait(spa_t *spa) 1212{ 1213 mutex_enter(&spa->spa_suspend_lock); 1214 while (spa_suspended(spa)) 1215 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock); 1216 mutex_exit(&spa->spa_suspend_lock); 1217} 1218 1219/* 1220 * ========================================================================== 1221 * Gang blocks. 1222 * 1223 * A gang block is a collection of small blocks that looks to the DMU 1224 * like one large block. When zio_dva_allocate() cannot find a block 1225 * of the requested size, due to either severe fragmentation or the pool 1226 * being nearly full, it calls zio_write_gang_block() to construct the 1227 * block from smaller fragments. 1228 * 1229 * A gang block consists of a gang header (zio_gbh_phys_t) and up to 1230 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like 1231 * an indirect block: it's an array of block pointers. It consumes 1232 * only one sector and hence is allocatable regardless of fragmentation. 1233 * The gang header's bps point to its gang members, which hold the data. 1234 * 1235 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg> 1236 * as the verifier to ensure uniqueness of the SHA256 checksum. 1237 * Critically, the gang block bp's blk_cksum is the checksum of the data, 1238 * not the gang header. This ensures that data block signatures (needed for 1239 * deduplication) are independent of how the block is physically stored. 1240 * 1241 * Gang blocks can be nested: a gang member may itself be a gang block. 1242 * Thus every gang block is a tree in which root and all interior nodes are 1243 * gang headers, and the leaves are normal blocks that contain user data. 1244 * The root of the gang tree is called the gang leader. 1245 * 1246 * To perform any operation (read, rewrite, free, claim) on a gang block, 1247 * zio_gang_assemble() first assembles the gang tree (minus data leaves) 1248 * in the io_gang_tree field of the original logical i/o by recursively 1249 * reading the gang leader and all gang headers below it. This yields 1250 * an in-core tree containing the contents of every gang header and the 1251 * bps for every constituent of the gang block. 1252 * 1253 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree 1254 * and invokes a callback on each bp. To free a gang block, zio_gang_issue() 1255 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp. 1256 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim(). 1257 * zio_read_gang() is a wrapper around zio_read() that omits reading gang 1258 * headers, since we already have those in io_gang_tree. zio_rewrite_gang() 1259 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite() 1260 * of the gang header plus zio_checksum_compute() of the data to update the 1261 * gang header's blk_cksum as described above. 1262 * 1263 * The two-phase assemble/issue model solves the problem of partial failure -- 1264 * what if you'd freed part of a gang block but then couldn't read the 1265 * gang header for another part? Assembling the entire gang tree first 1266 * ensures that all the necessary gang header I/O has succeeded before 1267 * starting the actual work of free, claim, or write. Once the gang tree 1268 * is assembled, free and claim are in-memory operations that cannot fail. 1269 * 1270 * In the event that a gang write fails, zio_dva_unallocate() walks the 1271 * gang tree to immediately free (i.e. insert back into the space map) 1272 * everything we've allocated. This ensures that we don't get ENOSPC 1273 * errors during repeated suspend/resume cycles due to a flaky device. 1274 * 1275 * Gang rewrites only happen during sync-to-convergence. If we can't assemble 1276 * the gang tree, we won't modify the block, so we can safely defer the free 1277 * (knowing that the block is still intact). If we *can* assemble the gang 1278 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free 1279 * each constituent bp and we can allocate a new block on the next sync pass. 1280 * 1281 * In all cases, the gang tree allows complete recovery from partial failure. 1282 * ========================================================================== 1283 */ 1284 1285static zio_t * 1286zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1287{ 1288 if (gn != NULL) 1289 return (pio); 1290 1291 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp), 1292 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), 1293 &pio->io_bookmark)); 1294} 1295 1296zio_t * 1297zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1298{ 1299 zio_t *zio; 1300 1301 if (gn != NULL) { 1302 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, 1303 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority, 1304 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 1305 /* 1306 * As we rewrite each gang header, the pipeline will compute 1307 * a new gang block header checksum for it; but no one will 1308 * compute a new data checksum, so we do that here. The one 1309 * exception is the gang leader: the pipeline already computed 1310 * its data checksum because that stage precedes gang assembly. 1311 * (Presently, nothing actually uses interior data checksums; 1312 * this is just good hygiene.) 1313 */ 1314 if (gn != pio->io_gang_leader->io_gang_tree) { 1315 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp), 1316 data, BP_GET_PSIZE(bp)); 1317 } 1318 } else { 1319 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, 1320 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority, 1321 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 1322 } 1323 1324 return (zio); 1325} 1326 1327/* ARGSUSED */ 1328zio_t * 1329zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1330{ 1331 return (zio_free(pio, pio->io_spa, pio->io_txg, bp, 1332 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio))); 1333} 1334 1335/* ARGSUSED */ 1336zio_t * 1337zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1338{ 1339 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp, 1340 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio))); 1341} 1342 1343static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = { 1344 NULL, 1345 zio_read_gang, 1346 zio_rewrite_gang, 1347 zio_free_gang, 1348 zio_claim_gang, 1349 NULL 1350}; 1351 1352static void zio_gang_tree_assemble_done(zio_t *zio); 1353 1354static zio_gang_node_t * 1355zio_gang_node_alloc(zio_gang_node_t **gnpp) 1356{ 1357 zio_gang_node_t *gn; 1358 1359 ASSERT(*gnpp == NULL); 1360 1361 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP); 1362 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE); 1363 *gnpp = gn; 1364 1365 return (gn); 1366} 1367 1368static void 1369zio_gang_node_free(zio_gang_node_t **gnpp) 1370{ 1371 zio_gang_node_t *gn = *gnpp; 1372 1373 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) 1374 ASSERT(gn->gn_child[g] == NULL); 1375 1376 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE); 1377 kmem_free(gn, sizeof (*gn)); 1378 *gnpp = NULL; 1379} 1380 1381static void 1382zio_gang_tree_free(zio_gang_node_t **gnpp) 1383{ 1384 zio_gang_node_t *gn = *gnpp; 1385 1386 if (gn == NULL) 1387 return; 1388 1389 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) 1390 zio_gang_tree_free(&gn->gn_child[g]); 1391 1392 zio_gang_node_free(gnpp); 1393} 1394 1395static void 1396zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp) 1397{ 1398 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp); 1399 1400 ASSERT(gio->io_gang_leader == gio); 1401 ASSERT(BP_IS_GANG(bp)); 1402 1403 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh, 1404 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn, 1405 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark)); 1406} 1407 1408static void 1409zio_gang_tree_assemble_done(zio_t *zio) 1410{ 1411 zio_t *gio = zio->io_gang_leader; 1412 zio_gang_node_t *gn = zio->io_private; 1413 blkptr_t *bp = zio->io_bp; 1414 1415 ASSERT(gio == zio_unique_parent(zio)); 1416 ASSERT(zio_walk_children(zio) == NULL); 1417 1418 if (zio->io_error) 1419 return; 1420 1421 if (BP_SHOULD_BYTESWAP(bp)) 1422 byteswap_uint64_array(zio->io_data, zio->io_size); 1423 1424 ASSERT(zio->io_data == gn->gn_gbh); 1425 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE); 1426 ASSERT(gn->gn_gbh->zg_tail.zbt_magic == ZBT_MAGIC); 1427 1428 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 1429 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; 1430 if (!BP_IS_GANG(gbp)) 1431 continue; 1432 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]); 1433 } 1434} 1435 1436static void 1437zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data) 1438{ 1439 zio_t *gio = pio->io_gang_leader; 1440 zio_t *zio; 1441 1442 ASSERT(BP_IS_GANG(bp) == !!gn); 1443 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp)); 1444 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree); 1445 1446 /* 1447 * If you're a gang header, your data is in gn->gn_gbh. 1448 * If you're a gang member, your data is in 'data' and gn == NULL. 1449 */ 1450 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data); 1451 1452 if (gn != NULL) { 1453 ASSERT(gn->gn_gbh->zg_tail.zbt_magic == ZBT_MAGIC); 1454 1455 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 1456 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; 1457 if (BP_IS_HOLE(gbp)) 1458 continue; 1459 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data); 1460 data = (char *)data + BP_GET_PSIZE(gbp); 1461 } 1462 } 1463 1464 if (gn == gio->io_gang_tree) 1465 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data); 1466 1467 if (zio != pio) 1468 zio_nowait(zio); 1469} 1470 1471static int 1472zio_gang_assemble(zio_t *zio) 1473{ 1474 blkptr_t *bp = zio->io_bp; 1475 1476 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL); 1477 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 1478 1479 zio->io_gang_leader = zio; 1480 1481 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree); 1482 1483 return (ZIO_PIPELINE_CONTINUE); 1484} 1485 1486static int 1487zio_gang_issue(zio_t *zio) 1488{ 1489 blkptr_t *bp = zio->io_bp; 1490 1491 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE)) 1492 return (ZIO_PIPELINE_STOP); 1493 1494 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio); 1495 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 1496 1497 if (zio->io_child_error[ZIO_CHILD_GANG] == 0) 1498 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data); 1499 else 1500 zio_gang_tree_free(&zio->io_gang_tree); 1501 1502 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1503 1504 return (ZIO_PIPELINE_CONTINUE); 1505} 1506 1507static void 1508zio_write_gang_member_ready(zio_t *zio) 1509{ 1510 zio_t *pio = zio_unique_parent(zio); 1511 zio_t *gio = zio->io_gang_leader; 1512 dva_t *cdva = zio->io_bp->blk_dva; 1513 dva_t *pdva = pio->io_bp->blk_dva; 1514 uint64_t asize; 1515 1516 if (BP_IS_HOLE(zio->io_bp)) 1517 return; 1518 1519 ASSERT(BP_IS_HOLE(&zio->io_bp_orig)); 1520 1521 ASSERT(zio->io_child_type == ZIO_CHILD_GANG); 1522 ASSERT3U(zio->io_prop.zp_ndvas, ==, gio->io_prop.zp_ndvas); 1523 ASSERT3U(zio->io_prop.zp_ndvas, <=, BP_GET_NDVAS(zio->io_bp)); 1524 ASSERT3U(pio->io_prop.zp_ndvas, <=, BP_GET_NDVAS(pio->io_bp)); 1525 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp)); 1526 1527 mutex_enter(&pio->io_lock); 1528 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) { 1529 ASSERT(DVA_GET_GANG(&pdva[d])); 1530 asize = DVA_GET_ASIZE(&pdva[d]); 1531 asize += DVA_GET_ASIZE(&cdva[d]); 1532 DVA_SET_ASIZE(&pdva[d], asize); 1533 } 1534 mutex_exit(&pio->io_lock); 1535} 1536 1537static int 1538zio_write_gang_block(zio_t *pio) 1539{ 1540 spa_t *spa = pio->io_spa; 1541 blkptr_t *bp = pio->io_bp; 1542 zio_t *gio = pio->io_gang_leader; 1543 zio_t *zio; 1544 zio_gang_node_t *gn, **gnpp; 1545 zio_gbh_phys_t *gbh; 1546 uint64_t txg = pio->io_txg; 1547 uint64_t resid = pio->io_size; 1548 uint64_t lsize; 1549 int ndvas = gio->io_prop.zp_ndvas; 1550 int gbh_ndvas = MIN(ndvas + 1, spa_max_replication(spa)); 1551 zio_prop_t zp; 1552 int error; 1553 1554 error = metaslab_alloc(spa, spa->spa_normal_class, SPA_GANGBLOCKSIZE, 1555 bp, gbh_ndvas, txg, pio == gio ? NULL : gio->io_bp, 1556 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER); 1557 if (error) { 1558 pio->io_error = error; 1559 return (ZIO_PIPELINE_CONTINUE); 1560 } 1561 1562 if (pio == gio) { 1563 gnpp = &gio->io_gang_tree; 1564 } else { 1565 gnpp = pio->io_private; 1566 ASSERT(pio->io_ready == zio_write_gang_member_ready); 1567 } 1568 1569 gn = zio_gang_node_alloc(gnpp); 1570 gbh = gn->gn_gbh; 1571 bzero(gbh, SPA_GANGBLOCKSIZE); 1572 1573 /* 1574 * Create the gang header. 1575 */ 1576 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL, 1577 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 1578 1579 /* 1580 * Create and nowait the gang children. 1581 */ 1582 for (int g = 0; resid != 0; resid -= lsize, g++) { 1583 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g), 1584 SPA_MINBLOCKSIZE); 1585 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid); 1586 1587 zp.zp_checksum = gio->io_prop.zp_checksum; 1588 zp.zp_compress = ZIO_COMPRESS_OFF; 1589 zp.zp_type = DMU_OT_NONE; 1590 zp.zp_level = 0; 1591 zp.zp_ndvas = gio->io_prop.zp_ndvas; 1592 1593 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g], 1594 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp, 1595 zio_write_gang_member_ready, NULL, &gn->gn_child[g], 1596 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), 1597 &pio->io_bookmark)); 1598 } 1599 1600 /* 1601 * Set pio's pipeline to just wait for zio to finish. 1602 */ 1603 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1604 1605 zio_nowait(zio); 1606 1607 return (ZIO_PIPELINE_CONTINUE); 1608} 1609 1610/* 1611 * ========================================================================== 1612 * Allocate and free blocks 1613 * ========================================================================== 1614 */ 1615 1616static int 1617zio_dva_allocate(zio_t *zio) 1618{ 1619 spa_t *spa = zio->io_spa; 1620 metaslab_class_t *mc = spa->spa_normal_class; 1621 blkptr_t *bp = zio->io_bp; 1622 int error; 1623 1624 if (zio->io_gang_leader == NULL) { 1625 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 1626 zio->io_gang_leader = zio; 1627 } 1628 1629 ASSERT(BP_IS_HOLE(bp)); 1630 ASSERT3U(BP_GET_NDVAS(bp), ==, 0); 1631 ASSERT3U(zio->io_prop.zp_ndvas, >, 0); 1632 ASSERT3U(zio->io_prop.zp_ndvas, <=, spa_max_replication(spa)); 1633 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp)); 1634 1635 error = metaslab_alloc(spa, mc, zio->io_size, bp, 1636 zio->io_prop.zp_ndvas, zio->io_txg, NULL, 0); 1637 1638 if (error) { 1639 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE) 1640 return (zio_write_gang_block(zio)); 1641 zio->io_error = error; 1642 } 1643 1644 return (ZIO_PIPELINE_CONTINUE); 1645} 1646 1647static int 1648zio_dva_free(zio_t *zio) 1649{ 1650 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE); 1651 1652 return (ZIO_PIPELINE_CONTINUE); 1653} 1654 1655static int 1656zio_dva_claim(zio_t *zio) 1657{ 1658 int error; 1659 1660 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg); 1661 if (error) 1662 zio->io_error = error; 1663 1664 return (ZIO_PIPELINE_CONTINUE); 1665} 1666 1667/* 1668 * Undo an allocation. This is used by zio_done() when an I/O fails 1669 * and we want to give back the block we just allocated. 1670 * This handles both normal blocks and gang blocks. 1671 */ 1672static void 1673zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp) 1674{ 1675 spa_t *spa = zio->io_spa; 1676 boolean_t now = !(zio->io_flags & ZIO_FLAG_IO_REWRITE); 1677 1678 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp)); 1679 1680 if (zio->io_bp == bp && !now) { 1681 /* 1682 * This is a rewrite for sync-to-convergence. 1683 * We can't do a metaslab_free(NOW) because bp wasn't allocated 1684 * during this sync pass, which means that metaslab_sync() 1685 * already committed the allocation. 1686 */ 1687 ASSERT(DVA_EQUAL(BP_IDENTITY(bp), 1688 BP_IDENTITY(&zio->io_bp_orig))); 1689 ASSERT(spa_sync_pass(spa) > 1); 1690 1691 if (BP_IS_GANG(bp) && gn == NULL) { 1692 /* 1693 * This is a gang leader whose gang header(s) we 1694 * couldn't read now, so defer the free until later. 1695 * The block should still be intact because without 1696 * the headers, we'd never even start the rewrite. 1697 */ 1698 bplist_enqueue_deferred(&spa->spa_sync_bplist, bp); 1699 return; 1700 } 1701 } 1702 1703 if (!BP_IS_HOLE(bp)) 1704 metaslab_free(spa, bp, bp->blk_birth, now); 1705 1706 if (gn != NULL) { 1707 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 1708 zio_dva_unallocate(zio, gn->gn_child[g], 1709 &gn->gn_gbh->zg_blkptr[g]); 1710 } 1711 } 1712} 1713 1714/* 1715 * Try to allocate an intent log block. Return 0 on success, errno on failure. 1716 */ 1717int 1718zio_alloc_blk(spa_t *spa, uint64_t size, blkptr_t *new_bp, blkptr_t *old_bp, 1719 uint64_t txg, boolean_t bypass_slog) 1720{ 1721 int error = 1; 1722 1723 if (!bypass_slog) 1724 error = metaslab_alloc(spa, spa->spa_log_class, size, 1725 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID); 1726 1727 if (error) 1728 error = metaslab_alloc(spa, spa->spa_normal_class, size, 1729 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID); 1730 1731 if (error == 0) { 1732 BP_SET_LSIZE(new_bp, size); 1733 BP_SET_PSIZE(new_bp, size); 1734 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF); 1735 BP_SET_CHECKSUM(new_bp, ZIO_CHECKSUM_ZILOG); 1736 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); 1737 BP_SET_LEVEL(new_bp, 0); 1738 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER); 1739 } 1740 1741 return (error); 1742} 1743 1744/* 1745 * Free an intent log block. We know it can't be a gang block, so there's 1746 * nothing to do except metaslab_free() it. 1747 */ 1748void 1749zio_free_blk(spa_t *spa, blkptr_t *bp, uint64_t txg) 1750{ 1751 ASSERT(!BP_IS_GANG(bp)); 1752 1753 metaslab_free(spa, bp, txg, B_FALSE); 1754} 1755 1756/* 1757 * ========================================================================== 1758 * Read and write to physical devices 1759 * ========================================================================== 1760 */ 1761static int 1762zio_vdev_io_start(zio_t *zio) 1763{ 1764 vdev_t *vd = zio->io_vd; 1765 uint64_t align; 1766 spa_t *spa = zio->io_spa; 1767 1768 ASSERT(zio->io_error == 0); 1769 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0); 1770 1771 if (vd == NULL) { 1772 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) 1773 spa_config_enter(spa, SCL_ZIO, zio, RW_READER); 1774 1775 /* 1776 * The mirror_ops handle multiple DVAs in a single BP. 1777 */ 1778 return (vdev_mirror_ops.vdev_op_io_start(zio)); 1779 } 1780 1781 align = 1ULL << vd->vdev_top->vdev_ashift; 1782 1783 if (P2PHASE(zio->io_size, align) != 0) { 1784 uint64_t asize = P2ROUNDUP(zio->io_size, align); 1785 char *abuf = zio_buf_alloc(asize); 1786 ASSERT(vd == vd->vdev_top); 1787 if (zio->io_type == ZIO_TYPE_WRITE) { 1788 bcopy(zio->io_data, abuf, zio->io_size); 1789 bzero(abuf + zio->io_size, asize - zio->io_size); 1790 } 1791 zio_push_transform(zio, abuf, asize, asize, zio_subblock); 1792 } 1793 1794 ASSERT(P2PHASE(zio->io_offset, align) == 0); 1795 ASSERT(P2PHASE(zio->io_size, align) == 0); 1796 ASSERT(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa)); 1797 1798 /* 1799 * If this is a repair I/O, and there's no self-healing involved -- 1800 * that is, we're just resilvering what we expect to resilver -- 1801 * then don't do the I/O unless zio's txg is actually in vd's DTL. 1802 * This prevents spurious resilvering with nested replication. 1803 * For example, given a mirror of mirrors, (A+B)+(C+D), if only 1804 * A is out of date, we'll read from C+D, then use the data to 1805 * resilver A+B -- but we don't actually want to resilver B, just A. 1806 * The top-level mirror has no way to know this, so instead we just 1807 * discard unnecessary repairs as we work our way down the vdev tree. 1808 * The same logic applies to any form of nested replication: 1809 * ditto + mirror, RAID-Z + replacing, etc. This covers them all. 1810 */ 1811 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) && 1812 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) && 1813 zio->io_txg != 0 && /* not a delegated i/o */ 1814 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) { 1815 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 1816 zio_vdev_io_bypass(zio); 1817 return (ZIO_PIPELINE_CONTINUE); 1818 } 1819 1820 if (vd->vdev_ops->vdev_op_leaf && 1821 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) { 1822 1823 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0) 1824 return (ZIO_PIPELINE_CONTINUE); 1825 1826 if ((zio = vdev_queue_io(zio)) == NULL) 1827 return (ZIO_PIPELINE_STOP); 1828 1829 if (!vdev_accessible(vd, zio)) { 1830 zio->io_error = ENXIO; 1831 zio_interrupt(zio); 1832 return (ZIO_PIPELINE_STOP); 1833 } 1834 } 1835 1836 return (vd->vdev_ops->vdev_op_io_start(zio)); 1837} 1838 1839static int 1840zio_vdev_io_done(zio_t *zio) 1841{ 1842 vdev_t *vd = zio->io_vd; 1843 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops; 1844 boolean_t unexpected_error = B_FALSE; 1845 1846 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE)) 1847 return (ZIO_PIPELINE_STOP); 1848 1849 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE); 1850 1851 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) { 1852 1853 vdev_queue_io_done(zio); 1854 1855 if (zio->io_type == ZIO_TYPE_WRITE) 1856 vdev_cache_write(zio); 1857 1858 if (zio_injection_enabled && zio->io_error == 0) 1859 zio->io_error = zio_handle_device_injection(vd, 1860 zio, EIO); 1861 1862 if (zio_injection_enabled && zio->io_error == 0) 1863 zio->io_error = zio_handle_label_injection(zio, EIO); 1864 1865 if (zio->io_error) { 1866 if (!vdev_accessible(vd, zio)) { 1867 zio->io_error = ENXIO; 1868 } else { 1869 unexpected_error = B_TRUE; 1870 } 1871 } 1872 } 1873 1874 ops->vdev_op_io_done(zio); 1875 1876 if (unexpected_error) 1877 VERIFY(vdev_probe(vd, zio) == NULL); 1878 1879 return (ZIO_PIPELINE_CONTINUE); 1880} 1881 1882/* 1883 * For non-raidz ZIOs, we can just copy aside the bad data read from the 1884 * disk, and use that to finish the checksum ereport later. 1885 */ 1886static void 1887zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr, 1888 const void *good_buf) 1889{ 1890 /* no processing needed */ 1891 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE); 1892} 1893 1894/*ARGSUSED*/ 1895void 1896zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored) 1897{ 1898 void *buf = zio_buf_alloc(zio->io_size); 1899 1900 bcopy(zio->io_data, buf, zio->io_size); 1901 1902 zcr->zcr_cbinfo = zio->io_size; 1903 zcr->zcr_cbdata = buf; 1904 zcr->zcr_finish = zio_vsd_default_cksum_finish; 1905 zcr->zcr_free = zio_buf_free; 1906} 1907 1908static int 1909zio_vdev_io_assess(zio_t *zio) 1910{ 1911 vdev_t *vd = zio->io_vd; 1912 1913 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE)) 1914 return (ZIO_PIPELINE_STOP); 1915 1916 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) 1917 spa_config_exit(zio->io_spa, SCL_ZIO, zio); 1918 1919 if (zio->io_vsd != NULL) { 1920 zio->io_vsd_ops->vsd_free(zio); 1921 zio->io_vsd = NULL; 1922 } 1923 1924 if (zio_injection_enabled && zio->io_error == 0) 1925 zio->io_error = zio_handle_fault_injection(zio, EIO); 1926 1927 /* 1928 * If the I/O failed, determine whether we should attempt to retry it. 1929 */ 1930 if (zio->io_error && vd == NULL && 1931 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) { 1932 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */ 1933 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */ 1934 zio->io_error = 0; 1935 zio->io_flags |= ZIO_FLAG_IO_RETRY | 1936 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE; 1937 zio->io_stage = ZIO_STAGE_VDEV_IO_START - 1; 1938 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE); 1939 return (ZIO_PIPELINE_STOP); 1940 } 1941 1942 /* 1943 * If we got an error on a leaf device, convert it to ENXIO 1944 * if the device is not accessible at all. 1945 */ 1946 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf && 1947 !vdev_accessible(vd, zio)) 1948 zio->io_error = ENXIO; 1949 1950 /* 1951 * If we can't write to an interior vdev (mirror or RAID-Z), 1952 * set vdev_cant_write so that we stop trying to allocate from it. 1953 */ 1954 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE && 1955 vd != NULL && !vd->vdev_ops->vdev_op_leaf) 1956 vd->vdev_cant_write = B_TRUE; 1957 1958 if (zio->io_error) 1959 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1960 1961 return (ZIO_PIPELINE_CONTINUE); 1962} 1963 1964void 1965zio_vdev_io_reissue(zio_t *zio) 1966{ 1967 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); 1968 ASSERT(zio->io_error == 0); 1969 1970 zio->io_stage--; 1971} 1972 1973void 1974zio_vdev_io_redone(zio_t *zio) 1975{ 1976 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE); 1977 1978 zio->io_stage--; 1979} 1980 1981void 1982zio_vdev_io_bypass(zio_t *zio) 1983{ 1984 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); 1985 ASSERT(zio->io_error == 0); 1986 1987 zio->io_flags |= ZIO_FLAG_IO_BYPASS; 1988 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS - 1; 1989} 1990 1991/* 1992 * ========================================================================== 1993 * Generate and verify checksums 1994 * ========================================================================== 1995 */ 1996static int 1997zio_checksum_generate(zio_t *zio) 1998{ 1999 blkptr_t *bp = zio->io_bp; 2000 enum zio_checksum checksum; 2001 2002 if (bp == NULL) { 2003 /* 2004 * This is zio_write_phys(). 2005 * We're either generating a label checksum, or none at all. 2006 */ 2007 checksum = zio->io_prop.zp_checksum; 2008 2009 if (checksum == ZIO_CHECKSUM_OFF) 2010 return (ZIO_PIPELINE_CONTINUE); 2011 2012 ASSERT(checksum == ZIO_CHECKSUM_LABEL); 2013 } else { 2014 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) { 2015 ASSERT(!IO_IS_ALLOCATING(zio)); 2016 checksum = ZIO_CHECKSUM_GANG_HEADER; 2017 } else { 2018 checksum = BP_GET_CHECKSUM(bp); 2019 } 2020 } 2021 2022 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size); 2023 2024 return (ZIO_PIPELINE_CONTINUE); 2025} 2026 2027static int 2028zio_checksum_verify(zio_t *zio) 2029{ 2030 zio_bad_cksum_t info; 2031 2032 blkptr_t *bp = zio->io_bp; 2033 int error; 2034 2035 if (bp == NULL) { 2036 /* 2037 * This is zio_read_phys(). 2038 * We're either verifying a label checksum, or nothing at all. 2039 */ 2040 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF) 2041 return (ZIO_PIPELINE_CONTINUE); 2042 2043 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL); 2044 } 2045 2046 if ((error = zio_checksum_error(zio, &info)) != 0) { 2047 zio->io_error = error; 2048 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) { 2049 zfs_ereport_start_checksum(zio->io_spa, 2050 zio->io_vd, zio, zio->io_offset, 2051 zio->io_size, NULL, &info); 2052 } 2053 } 2054 2055 return (ZIO_PIPELINE_CONTINUE); 2056} 2057 2058/* 2059 * Called by RAID-Z to ensure we don't compute the checksum twice. 2060 */ 2061void 2062zio_checksum_verified(zio_t *zio) 2063{ 2064 zio->io_pipeline &= ~(1U << ZIO_STAGE_CHECKSUM_VERIFY); 2065} 2066 2067/* 2068 * ========================================================================== 2069 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other. 2070 * An error of 0 indictes success. ENXIO indicates whole-device failure, 2071 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO 2072 * indicate errors that are specific to one I/O, and most likely permanent. 2073 * Any other error is presumed to be worse because we weren't expecting it. 2074 * ========================================================================== 2075 */ 2076int 2077zio_worst_error(int e1, int e2) 2078{ 2079 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO }; 2080 int r1, r2; 2081 2082 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++) 2083 if (e1 == zio_error_rank[r1]) 2084 break; 2085 2086 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++) 2087 if (e2 == zio_error_rank[r2]) 2088 break; 2089 2090 return (r1 > r2 ? e1 : e2); 2091} 2092 2093/* 2094 * ========================================================================== 2095 * I/O completion 2096 * ========================================================================== 2097 */ 2098static int 2099zio_ready(zio_t *zio) 2100{ 2101 blkptr_t *bp = zio->io_bp; 2102 zio_t *pio, *pio_next; 2103 2104 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY)) 2105 return (ZIO_PIPELINE_STOP); 2106 2107 if (zio->io_ready) { 2108 ASSERT(IO_IS_ALLOCATING(zio)); 2109 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp)); 2110 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0); 2111 2112 zio->io_ready(zio); 2113 } 2114 2115 if (bp != NULL && bp != &zio->io_bp_copy) 2116 zio->io_bp_copy = *bp; 2117 2118 if (zio->io_error) 2119 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2120 2121 mutex_enter(&zio->io_lock); 2122 zio->io_state[ZIO_WAIT_READY] = 1; 2123 pio = zio_walk_parents(zio); 2124 mutex_exit(&zio->io_lock); 2125 2126 /* 2127 * As we notify zio's parents, new parents could be added. 2128 * New parents go to the head of zio's io_parent_list, however, 2129 * so we will (correctly) not notify them. The remainder of zio's 2130 * io_parent_list, from 'pio_next' onward, cannot change because 2131 * all parents must wait for us to be done before they can be done. 2132 */ 2133 for (; pio != NULL; pio = pio_next) { 2134 pio_next = zio_walk_parents(zio); 2135 zio_notify_parent(pio, zio, ZIO_WAIT_READY); 2136 } 2137 2138 return (ZIO_PIPELINE_CONTINUE); 2139} 2140 2141static int 2142zio_done(zio_t *zio) 2143{ 2144 spa_t *spa = zio->io_spa; 2145 zio_t *lio = zio->io_logical; 2146 blkptr_t *bp = zio->io_bp; 2147 vdev_t *vd = zio->io_vd; 2148 uint64_t psize = zio->io_size; 2149 zio_t *pio, *pio_next; 2150 2151 /* 2152 * If our children haven't all completed, 2153 * wait for them and then repeat this pipeline stage. 2154 */ 2155 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) || 2156 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) || 2157 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE)) 2158 return (ZIO_PIPELINE_STOP); 2159 2160 for (int c = 0; c < ZIO_CHILD_TYPES; c++) 2161 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 2162 ASSERT(zio->io_children[c][w] == 0); 2163 2164 if (bp != NULL) { 2165 ASSERT(bp->blk_pad[0] == 0); 2166 ASSERT(bp->blk_pad[1] == 0); 2167 ASSERT(bp->blk_pad[2] == 0); 2168 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 || 2169 (bp == zio_unique_parent(zio)->io_bp)); 2170 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) && 2171 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) { 2172 ASSERT(!BP_SHOULD_BYTESWAP(bp)); 2173 ASSERT3U(zio->io_prop.zp_ndvas, <=, BP_GET_NDVAS(bp)); 2174 ASSERT(BP_COUNT_GANG(bp) == 0 || 2175 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp))); 2176 } 2177 } 2178 2179 /* 2180 * If there were child vdev or gang errors, they apply to us now. 2181 */ 2182 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV); 2183 zio_inherit_child_errors(zio, ZIO_CHILD_GANG); 2184 2185 zio_pop_transforms(zio); /* note: may set zio->io_error */ 2186 2187 vdev_stat_update(zio, psize); 2188 2189 if (zio->io_error) { 2190 /* 2191 * If this I/O is attached to a particular vdev, 2192 * generate an error message describing the I/O failure 2193 * at the block level. We ignore these errors if the 2194 * device is currently unavailable. 2195 */ 2196 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd)) 2197 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0); 2198 2199 if ((zio->io_error == EIO || !(zio->io_flags & 2200 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) && 2201 zio == lio) { 2202 /* 2203 * For logical I/O requests, tell the SPA to log the 2204 * error and generate a logical data ereport. 2205 */ 2206 spa_log_error(spa, zio); 2207 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio, 2208 0, 0); 2209 } 2210 } 2211 2212 if (zio->io_error && zio == lio) { 2213 /* 2214 * Determine whether zio should be reexecuted. This will 2215 * propagate all the way to the root via zio_notify_parent(). 2216 */ 2217 ASSERT(vd == NULL && bp != NULL); 2218 2219 if (IO_IS_ALLOCATING(zio)) 2220 if (zio->io_error != ENOSPC) 2221 zio->io_reexecute |= ZIO_REEXECUTE_NOW; 2222 else 2223 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 2224 2225 if ((zio->io_type == ZIO_TYPE_READ || 2226 zio->io_type == ZIO_TYPE_FREE) && 2227 zio->io_error == ENXIO && 2228 spa->spa_load_state == SPA_LOAD_NONE && 2229 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE) 2230 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 2231 2232 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute) 2233 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 2234 2235 /* 2236 * Here is a possibly good place to attempt to do 2237 * either combinatorial reconstruction or error correction 2238 * based on checksums. It also might be a good place 2239 * to send out preliminary ereports before we suspend 2240 * processing. 2241 */ 2242 } 2243 2244 /* 2245 * If there were logical child errors, they apply to us now. 2246 * We defer this until now to avoid conflating logical child 2247 * errors with errors that happened to the zio itself when 2248 * updating vdev stats and reporting FMA events above. 2249 */ 2250 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL); 2251 2252 if ((zio->io_error || zio->io_reexecute) && IO_IS_ALLOCATING(zio) && 2253 zio->io_child_type == ZIO_CHILD_LOGICAL) { 2254 ASSERT(zio->io_child_type != ZIO_CHILD_GANG); 2255 zio_dva_unallocate(zio, zio->io_gang_tree, bp); 2256 } 2257 2258 zio_gang_tree_free(&zio->io_gang_tree); 2259 2260 /* 2261 * Godfather I/Os should never suspend. 2262 */ 2263 if ((zio->io_flags & ZIO_FLAG_GODFATHER) && 2264 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) 2265 zio->io_reexecute = 0; 2266 2267 if (zio->io_reexecute) { 2268 /* 2269 * This is a logical I/O that wants to reexecute. 2270 * 2271 * Reexecute is top-down. When an i/o fails, if it's not 2272 * the root, it simply notifies its parent and sticks around. 2273 * The parent, seeing that it still has children in zio_done(), 2274 * does the same. This percolates all the way up to the root. 2275 * The root i/o will reexecute or suspend the entire tree. 2276 * 2277 * This approach ensures that zio_reexecute() honors 2278 * all the original i/o dependency relationships, e.g. 2279 * parents not executing until children are ready. 2280 */ 2281 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2282 2283 zio->io_gang_leader = NULL; 2284 2285 mutex_enter(&zio->io_lock); 2286 zio->io_state[ZIO_WAIT_DONE] = 1; 2287 mutex_exit(&zio->io_lock); 2288 2289 /* 2290 * "The Godfather" I/O monitors its children but is 2291 * not a true parent to them. It will track them through 2292 * the pipeline but severs its ties whenever they get into 2293 * trouble (e.g. suspended). This allows "The Godfather" 2294 * I/O to return status without blocking. 2295 */ 2296 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) { 2297 zio_link_t *zl = zio->io_walk_link; 2298 pio_next = zio_walk_parents(zio); 2299 2300 if ((pio->io_flags & ZIO_FLAG_GODFATHER) && 2301 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) { 2302 zio_remove_child(pio, zio, zl); 2303 zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 2304 } 2305 } 2306 2307 if ((pio = zio_unique_parent(zio)) != NULL) { 2308 /* 2309 * We're not a root i/o, so there's nothing to do 2310 * but notify our parent. Don't propagate errors 2311 * upward since we haven't permanently failed yet. 2312 */ 2313 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); 2314 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE; 2315 zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 2316 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) { 2317 /* 2318 * We'd fail again if we reexecuted now, so suspend 2319 * until conditions improve (e.g. device comes online). 2320 */ 2321 zio_suspend(spa, zio); 2322 } else { 2323 /* 2324 * Reexecution is potentially a huge amount of work. 2325 * Hand it off to the otherwise-unused claim taskq. 2326 */ 2327 (void) taskq_dispatch( 2328 spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE], 2329 (task_func_t *)zio_reexecute, zio, TQ_SLEEP); 2330 } 2331 return (ZIO_PIPELINE_STOP); 2332 } 2333 2334 ASSERT(zio_walk_children(zio) == NULL); 2335 ASSERT(zio->io_reexecute == 0); 2336 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL)); 2337 2338 /* Report any checksum errors, since the IO is complete */ 2339 while (zio->io_cksum_report != NULL) { 2340 zio_cksum_report_t *rpt = zio->io_cksum_report; 2341 2342 zio->io_cksum_report = rpt->zcr_next; 2343 rpt->zcr_next = NULL; 2344 2345 /* only pass in our data buffer if we've succeeded. */ 2346 rpt->zcr_finish(rpt, 2347 (zio->io_error == 0) ? zio->io_data : NULL); 2348 2349 zfs_ereport_free_checksum(rpt); 2350 } 2351 2352 /* 2353 * It is the responsibility of the done callback to ensure that this 2354 * particular zio is no longer discoverable for adoption, and as 2355 * such, cannot acquire any new parents. 2356 */ 2357 if (zio->io_done) 2358 zio->io_done(zio); 2359 2360 mutex_enter(&zio->io_lock); 2361 zio->io_state[ZIO_WAIT_DONE] = 1; 2362 mutex_exit(&zio->io_lock); 2363 2364 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) { 2365 zio_link_t *zl = zio->io_walk_link; 2366 pio_next = zio_walk_parents(zio); 2367 zio_remove_child(pio, zio, zl); 2368 zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 2369 } 2370 2371 if (zio->io_waiter != NULL) { 2372 mutex_enter(&zio->io_lock); 2373 zio->io_executor = NULL; 2374 cv_broadcast(&zio->io_cv); 2375 mutex_exit(&zio->io_lock); 2376 } else { 2377 zio_destroy(zio); 2378 } 2379 2380 return (ZIO_PIPELINE_STOP); 2381} 2382 2383/* 2384 * ========================================================================== 2385 * I/O pipeline definition 2386 * ========================================================================== 2387 */ 2388static zio_pipe_stage_t *zio_pipeline[ZIO_STAGES] = { 2389 NULL, 2390 zio_issue_async, 2391 zio_read_bp_init, 2392 zio_write_bp_init, 2393 zio_checksum_generate, 2394 zio_gang_assemble, 2395 zio_gang_issue, 2396 zio_dva_allocate, 2397 zio_dva_free, 2398 zio_dva_claim, 2399 zio_ready, 2400 zio_vdev_io_start, 2401 zio_vdev_io_done, 2402 zio_vdev_io_assess, 2403 zio_checksum_verify, 2404 zio_done 2405}; 2406