zio.c revision 277586
1/* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21/* 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23 * Copyright (c) 2011, 2014 by Delphix. All rights reserved. 24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved. 25 */ 26 27#include <sys/sysmacros.h> 28#include <sys/zfs_context.h> 29#include <sys/fm/fs/zfs.h> 30#include <sys/spa.h> 31#include <sys/txg.h> 32#include <sys/spa_impl.h> 33#include <sys/vdev_impl.h> 34#include <sys/zio_impl.h> 35#include <sys/zio_compress.h> 36#include <sys/zio_checksum.h> 37#include <sys/dmu_objset.h> 38#include <sys/arc.h> 39#include <sys/ddt.h> 40#include <sys/trim_map.h> 41#include <sys/blkptr.h> 42#include <sys/zfeature.h> 43 44SYSCTL_DECL(_vfs_zfs); 45SYSCTL_NODE(_vfs_zfs, OID_AUTO, zio, CTLFLAG_RW, 0, "ZFS ZIO"); 46#if defined(__amd64__) 47static int zio_use_uma = 1; 48#else 49static int zio_use_uma = 0; 50#endif 51TUNABLE_INT("vfs.zfs.zio.use_uma", &zio_use_uma); 52SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, use_uma, CTLFLAG_RDTUN, &zio_use_uma, 0, 53 "Use uma(9) for ZIO allocations"); 54static int zio_exclude_metadata = 0; 55TUNABLE_INT("vfs.zfs.zio.exclude_metadata", &zio_exclude_metadata); 56SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, exclude_metadata, CTLFLAG_RDTUN, &zio_exclude_metadata, 0, 57 "Exclude metadata buffers from dumps as well"); 58 59zio_trim_stats_t zio_trim_stats = { 60 { "bytes", KSTAT_DATA_UINT64, 61 "Number of bytes successfully TRIMmed" }, 62 { "success", KSTAT_DATA_UINT64, 63 "Number of successful TRIM requests" }, 64 { "unsupported", KSTAT_DATA_UINT64, 65 "Number of TRIM requests that failed because TRIM is not supported" }, 66 { "failed", KSTAT_DATA_UINT64, 67 "Number of TRIM requests that failed for reasons other than not supported" }, 68}; 69 70static kstat_t *zio_trim_ksp; 71 72/* 73 * ========================================================================== 74 * I/O type descriptions 75 * ========================================================================== 76 */ 77const char *zio_type_name[ZIO_TYPES] = { 78 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim", 79 "zio_ioctl" 80}; 81 82/* 83 * ========================================================================== 84 * I/O kmem caches 85 * ========================================================================== 86 */ 87kmem_cache_t *zio_cache; 88kmem_cache_t *zio_link_cache; 89kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; 90kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; 91 92#ifdef _KERNEL 93extern vmem_t *zio_alloc_arena; 94#endif 95 96/* 97 * The following actions directly effect the spa's sync-to-convergence logic. 98 * The values below define the sync pass when we start performing the action. 99 * Care should be taken when changing these values as they directly impact 100 * spa_sync() performance. Tuning these values may introduce subtle performance 101 * pathologies and should only be done in the context of performance analysis. 102 * These tunables will eventually be removed and replaced with #defines once 103 * enough analysis has been done to determine optimal values. 104 * 105 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that 106 * regular blocks are not deferred. 107 */ 108int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */ 109TUNABLE_INT("vfs.zfs.sync_pass_deferred_free", &zfs_sync_pass_deferred_free); 110SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_deferred_free, CTLFLAG_RDTUN, 111 &zfs_sync_pass_deferred_free, 0, "defer frees starting in this pass"); 112int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */ 113TUNABLE_INT("vfs.zfs.sync_pass_dont_compress", &zfs_sync_pass_dont_compress); 114SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_dont_compress, CTLFLAG_RDTUN, 115 &zfs_sync_pass_dont_compress, 0, "don't compress starting in this pass"); 116int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */ 117TUNABLE_INT("vfs.zfs.sync_pass_rewrite", &zfs_sync_pass_rewrite); 118SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_rewrite, CTLFLAG_RDTUN, 119 &zfs_sync_pass_rewrite, 0, "rewrite new bps starting in this pass"); 120 121/* 122 * An allocating zio is one that either currently has the DVA allocate 123 * stage set or will have it later in its lifetime. 124 */ 125#define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE) 126 127boolean_t zio_requeue_io_start_cut_in_line = B_TRUE; 128 129#ifdef ZFS_DEBUG 130int zio_buf_debug_limit = 16384; 131#else 132int zio_buf_debug_limit = 0; 133#endif 134 135void 136zio_init(void) 137{ 138 size_t c; 139 zio_cache = kmem_cache_create("zio_cache", 140 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0); 141 zio_link_cache = kmem_cache_create("zio_link_cache", 142 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0); 143 if (!zio_use_uma) 144 goto out; 145 146 /* 147 * For small buffers, we want a cache for each multiple of 148 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache 149 * for each quarter-power of 2. 150 */ 151 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { 152 size_t size = (c + 1) << SPA_MINBLOCKSHIFT; 153 size_t p2 = size; 154 size_t align = 0; 155 size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0; 156 157 while (!ISP2(p2)) 158 p2 &= p2 - 1; 159 160#ifdef illumos 161#ifndef _KERNEL 162 /* 163 * If we are using watchpoints, put each buffer on its own page, 164 * to eliminate the performance overhead of trapping to the 165 * kernel when modifying a non-watched buffer that shares the 166 * page with a watched buffer. 167 */ 168 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE)) 169 continue; 170#endif 171#endif /* illumos */ 172 if (size <= 4 * SPA_MINBLOCKSIZE) { 173 align = SPA_MINBLOCKSIZE; 174 } else if (IS_P2ALIGNED(size, p2 >> 2)) { 175 align = MIN(p2 >> 2, PAGESIZE); 176 } 177 178 if (align != 0) { 179 char name[36]; 180 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size); 181 zio_buf_cache[c] = kmem_cache_create(name, size, 182 align, NULL, NULL, NULL, NULL, NULL, cflags); 183 184 /* 185 * Since zio_data bufs do not appear in crash dumps, we 186 * pass KMC_NOTOUCH so that no allocator metadata is 187 * stored with the buffers. 188 */ 189 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size); 190 zio_data_buf_cache[c] = kmem_cache_create(name, size, 191 align, NULL, NULL, NULL, NULL, NULL, 192 cflags | KMC_NOTOUCH | KMC_NODEBUG); 193 } 194 } 195 196 while (--c != 0) { 197 ASSERT(zio_buf_cache[c] != NULL); 198 if (zio_buf_cache[c - 1] == NULL) 199 zio_buf_cache[c - 1] = zio_buf_cache[c]; 200 201 ASSERT(zio_data_buf_cache[c] != NULL); 202 if (zio_data_buf_cache[c - 1] == NULL) 203 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c]; 204 } 205out: 206 207 zio_inject_init(); 208 209 zio_trim_ksp = kstat_create("zfs", 0, "zio_trim", "misc", 210 KSTAT_TYPE_NAMED, 211 sizeof(zio_trim_stats) / sizeof(kstat_named_t), 212 KSTAT_FLAG_VIRTUAL); 213 214 if (zio_trim_ksp != NULL) { 215 zio_trim_ksp->ks_data = &zio_trim_stats; 216 kstat_install(zio_trim_ksp); 217 } 218} 219 220void 221zio_fini(void) 222{ 223 size_t c; 224 kmem_cache_t *last_cache = NULL; 225 kmem_cache_t *last_data_cache = NULL; 226 227 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { 228 if (zio_buf_cache[c] != last_cache) { 229 last_cache = zio_buf_cache[c]; 230 kmem_cache_destroy(zio_buf_cache[c]); 231 } 232 zio_buf_cache[c] = NULL; 233 234 if (zio_data_buf_cache[c] != last_data_cache) { 235 last_data_cache = zio_data_buf_cache[c]; 236 kmem_cache_destroy(zio_data_buf_cache[c]); 237 } 238 zio_data_buf_cache[c] = NULL; 239 } 240 241 kmem_cache_destroy(zio_link_cache); 242 kmem_cache_destroy(zio_cache); 243 244 zio_inject_fini(); 245 246 if (zio_trim_ksp != NULL) { 247 kstat_delete(zio_trim_ksp); 248 zio_trim_ksp = NULL; 249 } 250} 251 252/* 253 * ========================================================================== 254 * Allocate and free I/O buffers 255 * ========================================================================== 256 */ 257 258/* 259 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a 260 * crashdump if the kernel panics, so use it judiciously. Obviously, it's 261 * useful to inspect ZFS metadata, but if possible, we should avoid keeping 262 * excess / transient data in-core during a crashdump. 263 */ 264void * 265zio_buf_alloc(size_t size) 266{ 267 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 268 int flags = zio_exclude_metadata ? KM_NODEBUG : 0; 269 270 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 271 272 if (zio_use_uma) 273 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE)); 274 else 275 return (kmem_alloc(size, KM_SLEEP|flags)); 276} 277 278/* 279 * Use zio_data_buf_alloc to allocate data. The data will not appear in a 280 * crashdump if the kernel panics. This exists so that we will limit the amount 281 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount 282 * of kernel heap dumped to disk when the kernel panics) 283 */ 284void * 285zio_data_buf_alloc(size_t size) 286{ 287 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 288 289 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 290 291 if (zio_use_uma) 292 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE)); 293 else 294 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG)); 295} 296 297void 298zio_buf_free(void *buf, size_t size) 299{ 300 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 301 302 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 303 304 if (zio_use_uma) 305 kmem_cache_free(zio_buf_cache[c], buf); 306 else 307 kmem_free(buf, size); 308} 309 310void 311zio_data_buf_free(void *buf, size_t size) 312{ 313 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 314 315 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 316 317 if (zio_use_uma) 318 kmem_cache_free(zio_data_buf_cache[c], buf); 319 else 320 kmem_free(buf, size); 321} 322 323/* 324 * ========================================================================== 325 * Push and pop I/O transform buffers 326 * ========================================================================== 327 */ 328static void 329zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize, 330 zio_transform_func_t *transform) 331{ 332 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP); 333 334 zt->zt_orig_data = zio->io_data; 335 zt->zt_orig_size = zio->io_size; 336 zt->zt_bufsize = bufsize; 337 zt->zt_transform = transform; 338 339 zt->zt_next = zio->io_transform_stack; 340 zio->io_transform_stack = zt; 341 342 zio->io_data = data; 343 zio->io_size = size; 344} 345 346static void 347zio_pop_transforms(zio_t *zio) 348{ 349 zio_transform_t *zt; 350 351 while ((zt = zio->io_transform_stack) != NULL) { 352 if (zt->zt_transform != NULL) 353 zt->zt_transform(zio, 354 zt->zt_orig_data, zt->zt_orig_size); 355 356 if (zt->zt_bufsize != 0) 357 zio_buf_free(zio->io_data, zt->zt_bufsize); 358 359 zio->io_data = zt->zt_orig_data; 360 zio->io_size = zt->zt_orig_size; 361 zio->io_transform_stack = zt->zt_next; 362 363 kmem_free(zt, sizeof (zio_transform_t)); 364 } 365} 366 367/* 368 * ========================================================================== 369 * I/O transform callbacks for subblocks and decompression 370 * ========================================================================== 371 */ 372static void 373zio_subblock(zio_t *zio, void *data, uint64_t size) 374{ 375 ASSERT(zio->io_size > size); 376 377 if (zio->io_type == ZIO_TYPE_READ) 378 bcopy(zio->io_data, data, size); 379} 380 381static void 382zio_decompress(zio_t *zio, void *data, uint64_t size) 383{ 384 if (zio->io_error == 0 && 385 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp), 386 zio->io_data, data, zio->io_size, size) != 0) 387 zio->io_error = SET_ERROR(EIO); 388} 389 390/* 391 * ========================================================================== 392 * I/O parent/child relationships and pipeline interlocks 393 * ========================================================================== 394 */ 395/* 396 * NOTE - Callers to zio_walk_parents() and zio_walk_children must 397 * continue calling these functions until they return NULL. 398 * Otherwise, the next caller will pick up the list walk in 399 * some indeterminate state. (Otherwise every caller would 400 * have to pass in a cookie to keep the state represented by 401 * io_walk_link, which gets annoying.) 402 */ 403zio_t * 404zio_walk_parents(zio_t *cio) 405{ 406 zio_link_t *zl = cio->io_walk_link; 407 list_t *pl = &cio->io_parent_list; 408 409 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl); 410 cio->io_walk_link = zl; 411 412 if (zl == NULL) 413 return (NULL); 414 415 ASSERT(zl->zl_child == cio); 416 return (zl->zl_parent); 417} 418 419zio_t * 420zio_walk_children(zio_t *pio) 421{ 422 zio_link_t *zl = pio->io_walk_link; 423 list_t *cl = &pio->io_child_list; 424 425 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl); 426 pio->io_walk_link = zl; 427 428 if (zl == NULL) 429 return (NULL); 430 431 ASSERT(zl->zl_parent == pio); 432 return (zl->zl_child); 433} 434 435zio_t * 436zio_unique_parent(zio_t *cio) 437{ 438 zio_t *pio = zio_walk_parents(cio); 439 440 VERIFY(zio_walk_parents(cio) == NULL); 441 return (pio); 442} 443 444void 445zio_add_child(zio_t *pio, zio_t *cio) 446{ 447 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP); 448 449 /* 450 * Logical I/Os can have logical, gang, or vdev children. 451 * Gang I/Os can have gang or vdev children. 452 * Vdev I/Os can only have vdev children. 453 * The following ASSERT captures all of these constraints. 454 */ 455 ASSERT(cio->io_child_type <= pio->io_child_type); 456 457 zl->zl_parent = pio; 458 zl->zl_child = cio; 459 460 mutex_enter(&cio->io_lock); 461 mutex_enter(&pio->io_lock); 462 463 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0); 464 465 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 466 pio->io_children[cio->io_child_type][w] += !cio->io_state[w]; 467 468 list_insert_head(&pio->io_child_list, zl); 469 list_insert_head(&cio->io_parent_list, zl); 470 471 pio->io_child_count++; 472 cio->io_parent_count++; 473 474 mutex_exit(&pio->io_lock); 475 mutex_exit(&cio->io_lock); 476} 477 478static void 479zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl) 480{ 481 ASSERT(zl->zl_parent == pio); 482 ASSERT(zl->zl_child == cio); 483 484 mutex_enter(&cio->io_lock); 485 mutex_enter(&pio->io_lock); 486 487 list_remove(&pio->io_child_list, zl); 488 list_remove(&cio->io_parent_list, zl); 489 490 pio->io_child_count--; 491 cio->io_parent_count--; 492 493 mutex_exit(&pio->io_lock); 494 mutex_exit(&cio->io_lock); 495 496 kmem_cache_free(zio_link_cache, zl); 497} 498 499static boolean_t 500zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait) 501{ 502 uint64_t *countp = &zio->io_children[child][wait]; 503 boolean_t waiting = B_FALSE; 504 505 mutex_enter(&zio->io_lock); 506 ASSERT(zio->io_stall == NULL); 507 if (*countp != 0) { 508 zio->io_stage >>= 1; 509 zio->io_stall = countp; 510 waiting = B_TRUE; 511 } 512 mutex_exit(&zio->io_lock); 513 514 return (waiting); 515} 516 517static void 518zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait) 519{ 520 uint64_t *countp = &pio->io_children[zio->io_child_type][wait]; 521 int *errorp = &pio->io_child_error[zio->io_child_type]; 522 523 mutex_enter(&pio->io_lock); 524 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE)) 525 *errorp = zio_worst_error(*errorp, zio->io_error); 526 pio->io_reexecute |= zio->io_reexecute; 527 ASSERT3U(*countp, >, 0); 528 529 (*countp)--; 530 531 if (*countp == 0 && pio->io_stall == countp) { 532 pio->io_stall = NULL; 533 mutex_exit(&pio->io_lock); 534 zio_execute(pio); 535 } else { 536 mutex_exit(&pio->io_lock); 537 } 538} 539 540static void 541zio_inherit_child_errors(zio_t *zio, enum zio_child c) 542{ 543 if (zio->io_child_error[c] != 0 && zio->io_error == 0) 544 zio->io_error = zio->io_child_error[c]; 545} 546 547/* 548 * ========================================================================== 549 * Create the various types of I/O (read, write, free, etc) 550 * ========================================================================== 551 */ 552static zio_t * 553zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 554 void *data, uint64_t size, zio_done_func_t *done, void *private, 555 zio_type_t type, zio_priority_t priority, enum zio_flag flags, 556 vdev_t *vd, uint64_t offset, const zbookmark_phys_t *zb, 557 enum zio_stage stage, enum zio_stage pipeline) 558{ 559 zio_t *zio; 560 561 ASSERT3U(type == ZIO_TYPE_FREE || size, <=, SPA_MAXBLOCKSIZE); 562 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0); 563 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0); 564 565 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER)); 566 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER)); 567 ASSERT(vd || stage == ZIO_STAGE_OPEN); 568 569 zio = kmem_cache_alloc(zio_cache, KM_SLEEP); 570 bzero(zio, sizeof (zio_t)); 571 572 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL); 573 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL); 574 575 list_create(&zio->io_parent_list, sizeof (zio_link_t), 576 offsetof(zio_link_t, zl_parent_node)); 577 list_create(&zio->io_child_list, sizeof (zio_link_t), 578 offsetof(zio_link_t, zl_child_node)); 579 580 if (vd != NULL) 581 zio->io_child_type = ZIO_CHILD_VDEV; 582 else if (flags & ZIO_FLAG_GANG_CHILD) 583 zio->io_child_type = ZIO_CHILD_GANG; 584 else if (flags & ZIO_FLAG_DDT_CHILD) 585 zio->io_child_type = ZIO_CHILD_DDT; 586 else 587 zio->io_child_type = ZIO_CHILD_LOGICAL; 588 589 if (bp != NULL) { 590 zio->io_bp = (blkptr_t *)bp; 591 zio->io_bp_copy = *bp; 592 zio->io_bp_orig = *bp; 593 if (type != ZIO_TYPE_WRITE || 594 zio->io_child_type == ZIO_CHILD_DDT) 595 zio->io_bp = &zio->io_bp_copy; /* so caller can free */ 596 if (zio->io_child_type == ZIO_CHILD_LOGICAL) 597 zio->io_logical = zio; 598 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp)) 599 pipeline |= ZIO_GANG_STAGES; 600 } 601 602 zio->io_spa = spa; 603 zio->io_txg = txg; 604 zio->io_done = done; 605 zio->io_private = private; 606 zio->io_type = type; 607 zio->io_priority = priority; 608 zio->io_vd = vd; 609 zio->io_offset = offset; 610 zio->io_orig_data = zio->io_data = data; 611 zio->io_orig_size = zio->io_size = size; 612 zio->io_orig_flags = zio->io_flags = flags; 613 zio->io_orig_stage = zio->io_stage = stage; 614 zio->io_orig_pipeline = zio->io_pipeline = pipeline; 615 616 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY); 617 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE); 618 619 if (zb != NULL) 620 zio->io_bookmark = *zb; 621 622 if (pio != NULL) { 623 if (zio->io_logical == NULL) 624 zio->io_logical = pio->io_logical; 625 if (zio->io_child_type == ZIO_CHILD_GANG) 626 zio->io_gang_leader = pio->io_gang_leader; 627 zio_add_child(pio, zio); 628 } 629 630 return (zio); 631} 632 633static void 634zio_destroy(zio_t *zio) 635{ 636 list_destroy(&zio->io_parent_list); 637 list_destroy(&zio->io_child_list); 638 mutex_destroy(&zio->io_lock); 639 cv_destroy(&zio->io_cv); 640 kmem_cache_free(zio_cache, zio); 641} 642 643zio_t * 644zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done, 645 void *private, enum zio_flag flags) 646{ 647 zio_t *zio; 648 649 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private, 650 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, 651 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE); 652 653 return (zio); 654} 655 656zio_t * 657zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags) 658{ 659 return (zio_null(NULL, spa, NULL, done, private, flags)); 660} 661 662void 663zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp) 664{ 665 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) { 666 zfs_panic_recover("blkptr at %p has invalid TYPE %llu", 667 bp, (longlong_t)BP_GET_TYPE(bp)); 668 } 669 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS || 670 BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) { 671 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu", 672 bp, (longlong_t)BP_GET_CHECKSUM(bp)); 673 } 674 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS || 675 BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) { 676 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu", 677 bp, (longlong_t)BP_GET_COMPRESS(bp)); 678 } 679 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) { 680 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu", 681 bp, (longlong_t)BP_GET_LSIZE(bp)); 682 } 683 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) { 684 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu", 685 bp, (longlong_t)BP_GET_PSIZE(bp)); 686 } 687 688 if (BP_IS_EMBEDDED(bp)) { 689 if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) { 690 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu", 691 bp, (longlong_t)BPE_GET_ETYPE(bp)); 692 } 693 } 694 695 /* 696 * Pool-specific checks. 697 * 698 * Note: it would be nice to verify that the blk_birth and 699 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze() 700 * allows the birth time of log blocks (and dmu_sync()-ed blocks 701 * that are in the log) to be arbitrarily large. 702 */ 703 for (int i = 0; i < BP_GET_NDVAS(bp); i++) { 704 uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]); 705 if (vdevid >= spa->spa_root_vdev->vdev_children) { 706 zfs_panic_recover("blkptr at %p DVA %u has invalid " 707 "VDEV %llu", 708 bp, i, (longlong_t)vdevid); 709 } 710 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid]; 711 if (vd == NULL) { 712 zfs_panic_recover("blkptr at %p DVA %u has invalid " 713 "VDEV %llu", 714 bp, i, (longlong_t)vdevid); 715 } 716 if (vd->vdev_ops == &vdev_hole_ops) { 717 zfs_panic_recover("blkptr at %p DVA %u has hole " 718 "VDEV %llu", 719 bp, i, (longlong_t)vdevid); 720 721 } 722 if (vd->vdev_ops == &vdev_missing_ops) { 723 /* 724 * "missing" vdevs are valid during import, but we 725 * don't have their detailed info (e.g. asize), so 726 * we can't perform any more checks on them. 727 */ 728 continue; 729 } 730 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]); 731 uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]); 732 if (BP_IS_GANG(bp)) 733 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE); 734 if (offset + asize > vd->vdev_asize) { 735 zfs_panic_recover("blkptr at %p DVA %u has invalid " 736 "OFFSET %llu", 737 bp, i, (longlong_t)offset); 738 } 739 } 740} 741 742zio_t * 743zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, 744 void *data, uint64_t size, zio_done_func_t *done, void *private, 745 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb) 746{ 747 zio_t *zio; 748 749 zfs_blkptr_verify(spa, bp); 750 751 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp, 752 data, size, done, private, 753 ZIO_TYPE_READ, priority, flags, NULL, 0, zb, 754 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? 755 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE); 756 757 return (zio); 758} 759 760zio_t * 761zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, 762 void *data, uint64_t size, const zio_prop_t *zp, 763 zio_done_func_t *ready, zio_done_func_t *physdone, zio_done_func_t *done, 764 void *private, 765 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb) 766{ 767 zio_t *zio; 768 769 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF && 770 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS && 771 zp->zp_compress >= ZIO_COMPRESS_OFF && 772 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS && 773 DMU_OT_IS_VALID(zp->zp_type) && 774 zp->zp_level < 32 && 775 zp->zp_copies > 0 && 776 zp->zp_copies <= spa_max_replication(spa)); 777 778 zio = zio_create(pio, spa, txg, bp, data, size, done, private, 779 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, 780 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? 781 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE); 782 783 zio->io_ready = ready; 784 zio->io_physdone = physdone; 785 zio->io_prop = *zp; 786 787 /* 788 * Data can be NULL if we are going to call zio_write_override() to 789 * provide the already-allocated BP. But we may need the data to 790 * verify a dedup hit (if requested). In this case, don't try to 791 * dedup (just take the already-allocated BP verbatim). 792 */ 793 if (data == NULL && zio->io_prop.zp_dedup_verify) { 794 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE; 795 } 796 797 return (zio); 798} 799 800zio_t * 801zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data, 802 uint64_t size, zio_done_func_t *done, void *private, 803 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb) 804{ 805 zio_t *zio; 806 807 zio = zio_create(pio, spa, txg, bp, data, size, done, private, 808 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, 809 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE); 810 811 return (zio); 812} 813 814void 815zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite) 816{ 817 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 818 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 819 ASSERT(zio->io_stage == ZIO_STAGE_OPEN); 820 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa)); 821 822 /* 823 * We must reset the io_prop to match the values that existed 824 * when the bp was first written by dmu_sync() keeping in mind 825 * that nopwrite and dedup are mutually exclusive. 826 */ 827 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup; 828 zio->io_prop.zp_nopwrite = nopwrite; 829 zio->io_prop.zp_copies = copies; 830 zio->io_bp_override = bp; 831} 832 833void 834zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp) 835{ 836 837 /* 838 * The check for EMBEDDED is a performance optimization. We 839 * process the free here (by ignoring it) rather than 840 * putting it on the list and then processing it in zio_free_sync(). 841 */ 842 if (BP_IS_EMBEDDED(bp)) 843 return; 844 metaslab_check_free(spa, bp); 845 846 /* 847 * Frees that are for the currently-syncing txg, are not going to be 848 * deferred, and which will not need to do a read (i.e. not GANG or 849 * DEDUP), can be processed immediately. Otherwise, put them on the 850 * in-memory list for later processing. 851 */ 852 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) || 853 txg != spa->spa_syncing_txg || 854 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) { 855 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp); 856 } else { 857 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 858 BP_GET_PSIZE(bp), 0))); 859 } 860} 861 862zio_t * 863zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 864 uint64_t size, enum zio_flag flags) 865{ 866 zio_t *zio; 867 enum zio_stage stage = ZIO_FREE_PIPELINE; 868 869 ASSERT(!BP_IS_HOLE(bp)); 870 ASSERT(spa_syncing_txg(spa) == txg); 871 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free); 872 873 if (BP_IS_EMBEDDED(bp)) 874 return (zio_null(pio, spa, NULL, NULL, NULL, 0)); 875 876 metaslab_check_free(spa, bp); 877 arc_freed(spa, bp); 878 879 if (zfs_trim_enabled) 880 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START | 881 ZIO_STAGE_VDEV_IO_ASSESS; 882 /* 883 * GANG and DEDUP blocks can induce a read (for the gang block header, 884 * or the DDT), so issue them asynchronously so that this thread is 885 * not tied up. 886 */ 887 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp)) 888 stage |= ZIO_STAGE_ISSUE_ASYNC; 889 890 flags |= ZIO_FLAG_DONT_QUEUE; 891 892 zio = zio_create(pio, spa, txg, bp, NULL, size, 893 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags, 894 NULL, 0, NULL, ZIO_STAGE_OPEN, stage); 895 896 return (zio); 897} 898 899zio_t * 900zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 901 zio_done_func_t *done, void *private, enum zio_flag flags) 902{ 903 zio_t *zio; 904 905 dprintf_bp(bp, "claiming in txg %llu", txg); 906 907 if (BP_IS_EMBEDDED(bp)) 908 return (zio_null(pio, spa, NULL, NULL, NULL, 0)); 909 910 /* 911 * A claim is an allocation of a specific block. Claims are needed 912 * to support immediate writes in the intent log. The issue is that 913 * immediate writes contain committed data, but in a txg that was 914 * *not* committed. Upon opening the pool after an unclean shutdown, 915 * the intent log claims all blocks that contain immediate write data 916 * so that the SPA knows they're in use. 917 * 918 * All claims *must* be resolved in the first txg -- before the SPA 919 * starts allocating blocks -- so that nothing is allocated twice. 920 * If txg == 0 we just verify that the block is claimable. 921 */ 922 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa)); 923 ASSERT(txg == spa_first_txg(spa) || txg == 0); 924 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */ 925 926 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), 927 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags, 928 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE); 929 930 return (zio); 931} 932 933zio_t * 934zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset, 935 uint64_t size, zio_done_func_t *done, void *private, 936 zio_priority_t priority, enum zio_flag flags) 937{ 938 zio_t *zio; 939 int c; 940 941 if (vd->vdev_children == 0) { 942 zio = zio_create(pio, spa, 0, NULL, NULL, size, done, private, 943 ZIO_TYPE_IOCTL, priority, flags, vd, offset, NULL, 944 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE); 945 946 zio->io_cmd = cmd; 947 } else { 948 zio = zio_null(pio, spa, NULL, NULL, NULL, flags); 949 950 for (c = 0; c < vd->vdev_children; c++) 951 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd, 952 offset, size, done, private, priority, flags)); 953 } 954 955 return (zio); 956} 957 958zio_t * 959zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 960 void *data, int checksum, zio_done_func_t *done, void *private, 961 zio_priority_t priority, enum zio_flag flags, boolean_t labels) 962{ 963 zio_t *zio; 964 965 ASSERT(vd->vdev_children == 0); 966 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || 967 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); 968 ASSERT3U(offset + size, <=, vd->vdev_psize); 969 970 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private, 971 ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset, 972 NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE); 973 974 zio->io_prop.zp_checksum = checksum; 975 976 return (zio); 977} 978 979zio_t * 980zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 981 void *data, int checksum, zio_done_func_t *done, void *private, 982 zio_priority_t priority, enum zio_flag flags, boolean_t labels) 983{ 984 zio_t *zio; 985 986 ASSERT(vd->vdev_children == 0); 987 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || 988 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); 989 ASSERT3U(offset + size, <=, vd->vdev_psize); 990 991 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private, 992 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset, 993 NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE); 994 995 zio->io_prop.zp_checksum = checksum; 996 997 if (zio_checksum_table[checksum].ci_eck) { 998 /* 999 * zec checksums are necessarily destructive -- they modify 1000 * the end of the write buffer to hold the verifier/checksum. 1001 * Therefore, we must make a local copy in case the data is 1002 * being written to multiple places in parallel. 1003 */ 1004 void *wbuf = zio_buf_alloc(size); 1005 bcopy(data, wbuf, size); 1006 zio_push_transform(zio, wbuf, size, size, NULL); 1007 } 1008 1009 return (zio); 1010} 1011 1012/* 1013 * Create a child I/O to do some work for us. 1014 */ 1015zio_t * 1016zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset, 1017 void *data, uint64_t size, int type, zio_priority_t priority, 1018 enum zio_flag flags, zio_done_func_t *done, void *private) 1019{ 1020 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE; 1021 zio_t *zio; 1022 1023 ASSERT(vd->vdev_parent == 1024 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev)); 1025 1026 if (type == ZIO_TYPE_READ && bp != NULL) { 1027 /* 1028 * If we have the bp, then the child should perform the 1029 * checksum and the parent need not. This pushes error 1030 * detection as close to the leaves as possible and 1031 * eliminates redundant checksums in the interior nodes. 1032 */ 1033 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY; 1034 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; 1035 } 1036 1037 /* Not all IO types require vdev io done stage e.g. free */ 1038 if (!(pio->io_pipeline & ZIO_STAGE_VDEV_IO_DONE)) 1039 pipeline &= ~ZIO_STAGE_VDEV_IO_DONE; 1040 1041 if (vd->vdev_children == 0) 1042 offset += VDEV_LABEL_START_SIZE; 1043 1044 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE; 1045 1046 /* 1047 * If we've decided to do a repair, the write is not speculative -- 1048 * even if the original read was. 1049 */ 1050 if (flags & ZIO_FLAG_IO_REPAIR) 1051 flags &= ~ZIO_FLAG_SPECULATIVE; 1052 1053 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, 1054 done, private, type, priority, flags, vd, offset, &pio->io_bookmark, 1055 ZIO_STAGE_VDEV_IO_START >> 1, pipeline); 1056 1057 zio->io_physdone = pio->io_physdone; 1058 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL) 1059 zio->io_logical->io_phys_children++; 1060 1061 return (zio); 1062} 1063 1064zio_t * 1065zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size, 1066 int type, zio_priority_t priority, enum zio_flag flags, 1067 zio_done_func_t *done, void *private) 1068{ 1069 zio_t *zio; 1070 1071 ASSERT(vd->vdev_ops->vdev_op_leaf); 1072 1073 zio = zio_create(NULL, vd->vdev_spa, 0, NULL, 1074 data, size, done, private, type, priority, 1075 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED, 1076 vd, offset, NULL, 1077 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE); 1078 1079 return (zio); 1080} 1081 1082void 1083zio_flush(zio_t *zio, vdev_t *vd) 1084{ 1085 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0, 1086 NULL, NULL, ZIO_PRIORITY_NOW, 1087 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY)); 1088} 1089 1090zio_t * 1091zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size) 1092{ 1093 1094 ASSERT(vd->vdev_ops->vdev_op_leaf); 1095 1096 return (zio_create(zio, spa, 0, NULL, NULL, size, NULL, NULL, 1097 ZIO_TYPE_FREE, ZIO_PRIORITY_TRIM, ZIO_FLAG_DONT_AGGREGATE | 1098 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY, 1099 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PHYS_PIPELINE)); 1100} 1101 1102void 1103zio_shrink(zio_t *zio, uint64_t size) 1104{ 1105 ASSERT(zio->io_executor == NULL); 1106 ASSERT(zio->io_orig_size == zio->io_size); 1107 ASSERT(size <= zio->io_size); 1108 1109 /* 1110 * We don't shrink for raidz because of problems with the 1111 * reconstruction when reading back less than the block size. 1112 * Note, BP_IS_RAIDZ() assumes no compression. 1113 */ 1114 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 1115 if (!BP_IS_RAIDZ(zio->io_bp)) 1116 zio->io_orig_size = zio->io_size = size; 1117} 1118 1119/* 1120 * ========================================================================== 1121 * Prepare to read and write logical blocks 1122 * ========================================================================== 1123 */ 1124 1125static int 1126zio_read_bp_init(zio_t *zio) 1127{ 1128 blkptr_t *bp = zio->io_bp; 1129 1130 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF && 1131 zio->io_child_type == ZIO_CHILD_LOGICAL && 1132 !(zio->io_flags & ZIO_FLAG_RAW)) { 1133 uint64_t psize = 1134 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp); 1135 void *cbuf = zio_buf_alloc(psize); 1136 1137 zio_push_transform(zio, cbuf, psize, psize, zio_decompress); 1138 } 1139 1140 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) { 1141 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1142 decode_embedded_bp_compressed(bp, zio->io_data); 1143 } else { 1144 ASSERT(!BP_IS_EMBEDDED(bp)); 1145 } 1146 1147 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0) 1148 zio->io_flags |= ZIO_FLAG_DONT_CACHE; 1149 1150 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP) 1151 zio->io_flags |= ZIO_FLAG_DONT_CACHE; 1152 1153 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL) 1154 zio->io_pipeline = ZIO_DDT_READ_PIPELINE; 1155 1156 return (ZIO_PIPELINE_CONTINUE); 1157} 1158 1159static int 1160zio_write_bp_init(zio_t *zio) 1161{ 1162 spa_t *spa = zio->io_spa; 1163 zio_prop_t *zp = &zio->io_prop; 1164 enum zio_compress compress = zp->zp_compress; 1165 blkptr_t *bp = zio->io_bp; 1166 uint64_t lsize = zio->io_size; 1167 uint64_t psize = lsize; 1168 int pass = 1; 1169 1170 /* 1171 * If our children haven't all reached the ready stage, 1172 * wait for them and then repeat this pipeline stage. 1173 */ 1174 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) || 1175 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY)) 1176 return (ZIO_PIPELINE_STOP); 1177 1178 if (!IO_IS_ALLOCATING(zio)) 1179 return (ZIO_PIPELINE_CONTINUE); 1180 1181 ASSERT(zio->io_child_type != ZIO_CHILD_DDT); 1182 1183 if (zio->io_bp_override) { 1184 ASSERT(bp->blk_birth != zio->io_txg); 1185 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0); 1186 1187 *bp = *zio->io_bp_override; 1188 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1189 1190 if (BP_IS_EMBEDDED(bp)) 1191 return (ZIO_PIPELINE_CONTINUE); 1192 1193 /* 1194 * If we've been overridden and nopwrite is set then 1195 * set the flag accordingly to indicate that a nopwrite 1196 * has already occurred. 1197 */ 1198 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) { 1199 ASSERT(!zp->zp_dedup); 1200 zio->io_flags |= ZIO_FLAG_NOPWRITE; 1201 return (ZIO_PIPELINE_CONTINUE); 1202 } 1203 1204 ASSERT(!zp->zp_nopwrite); 1205 1206 if (BP_IS_HOLE(bp) || !zp->zp_dedup) 1207 return (ZIO_PIPELINE_CONTINUE); 1208 1209 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup || 1210 zp->zp_dedup_verify); 1211 1212 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) { 1213 BP_SET_DEDUP(bp, 1); 1214 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE; 1215 return (ZIO_PIPELINE_CONTINUE); 1216 } 1217 zio->io_bp_override = NULL; 1218 BP_ZERO(bp); 1219 } 1220 1221 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) { 1222 /* 1223 * We're rewriting an existing block, which means we're 1224 * working on behalf of spa_sync(). For spa_sync() to 1225 * converge, it must eventually be the case that we don't 1226 * have to allocate new blocks. But compression changes 1227 * the blocksize, which forces a reallocate, and makes 1228 * convergence take longer. Therefore, after the first 1229 * few passes, stop compressing to ensure convergence. 1230 */ 1231 pass = spa_sync_pass(spa); 1232 1233 ASSERT(zio->io_txg == spa_syncing_txg(spa)); 1234 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1235 ASSERT(!BP_GET_DEDUP(bp)); 1236 1237 if (pass >= zfs_sync_pass_dont_compress) 1238 compress = ZIO_COMPRESS_OFF; 1239 1240 /* Make sure someone doesn't change their mind on overwrites */ 1241 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp), 1242 spa_max_replication(spa)) == BP_GET_NDVAS(bp)); 1243 } 1244 1245 if (compress != ZIO_COMPRESS_OFF) { 1246 void *cbuf = zio_buf_alloc(lsize); 1247 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize); 1248 if (psize == 0 || psize == lsize) { 1249 compress = ZIO_COMPRESS_OFF; 1250 zio_buf_free(cbuf, lsize); 1251 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE && 1252 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) && 1253 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) { 1254 encode_embedded_bp_compressed(bp, 1255 cbuf, compress, lsize, psize); 1256 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA); 1257 BP_SET_TYPE(bp, zio->io_prop.zp_type); 1258 BP_SET_LEVEL(bp, zio->io_prop.zp_level); 1259 zio_buf_free(cbuf, lsize); 1260 bp->blk_birth = zio->io_txg; 1261 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1262 ASSERT(spa_feature_is_active(spa, 1263 SPA_FEATURE_EMBEDDED_DATA)); 1264 return (ZIO_PIPELINE_CONTINUE); 1265 } else { 1266 /* 1267 * Round up compressed size to MINBLOCKSIZE and 1268 * zero the tail. 1269 */ 1270 size_t rounded = 1271 P2ROUNDUP(psize, (size_t)SPA_MINBLOCKSIZE); 1272 if (rounded > psize) { 1273 bzero((char *)cbuf + psize, rounded - psize); 1274 psize = rounded; 1275 } 1276 if (psize == lsize) { 1277 compress = ZIO_COMPRESS_OFF; 1278 zio_buf_free(cbuf, lsize); 1279 } else { 1280 zio_push_transform(zio, cbuf, 1281 psize, lsize, NULL); 1282 } 1283 } 1284 } 1285 1286 /* 1287 * The final pass of spa_sync() must be all rewrites, but the first 1288 * few passes offer a trade-off: allocating blocks defers convergence, 1289 * but newly allocated blocks are sequential, so they can be written 1290 * to disk faster. Therefore, we allow the first few passes of 1291 * spa_sync() to allocate new blocks, but force rewrites after that. 1292 * There should only be a handful of blocks after pass 1 in any case. 1293 */ 1294 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg && 1295 BP_GET_PSIZE(bp) == psize && 1296 pass >= zfs_sync_pass_rewrite) { 1297 ASSERT(psize != 0); 1298 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES; 1299 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages; 1300 zio->io_flags |= ZIO_FLAG_IO_REWRITE; 1301 } else { 1302 BP_ZERO(bp); 1303 zio->io_pipeline = ZIO_WRITE_PIPELINE; 1304 } 1305 1306 if (psize == 0) { 1307 if (zio->io_bp_orig.blk_birth != 0 && 1308 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) { 1309 BP_SET_LSIZE(bp, lsize); 1310 BP_SET_TYPE(bp, zp->zp_type); 1311 BP_SET_LEVEL(bp, zp->zp_level); 1312 BP_SET_BIRTH(bp, zio->io_txg, 0); 1313 } 1314 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1315 } else { 1316 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER); 1317 BP_SET_LSIZE(bp, lsize); 1318 BP_SET_TYPE(bp, zp->zp_type); 1319 BP_SET_LEVEL(bp, zp->zp_level); 1320 BP_SET_PSIZE(bp, psize); 1321 BP_SET_COMPRESS(bp, compress); 1322 BP_SET_CHECKSUM(bp, zp->zp_checksum); 1323 BP_SET_DEDUP(bp, zp->zp_dedup); 1324 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER); 1325 if (zp->zp_dedup) { 1326 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1327 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 1328 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE; 1329 } 1330 if (zp->zp_nopwrite) { 1331 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1332 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 1333 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE; 1334 } 1335 } 1336 1337 return (ZIO_PIPELINE_CONTINUE); 1338} 1339 1340static int 1341zio_free_bp_init(zio_t *zio) 1342{ 1343 blkptr_t *bp = zio->io_bp; 1344 1345 if (zio->io_child_type == ZIO_CHILD_LOGICAL) { 1346 if (BP_GET_DEDUP(bp)) 1347 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE; 1348 } 1349 1350 return (ZIO_PIPELINE_CONTINUE); 1351} 1352 1353/* 1354 * ========================================================================== 1355 * Execute the I/O pipeline 1356 * ========================================================================== 1357 */ 1358 1359static void 1360zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline) 1361{ 1362 spa_t *spa = zio->io_spa; 1363 zio_type_t t = zio->io_type; 1364 int flags = (cutinline ? TQ_FRONT : 0); 1365 1366 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT); 1367 1368 /* 1369 * If we're a config writer or a probe, the normal issue and 1370 * interrupt threads may all be blocked waiting for the config lock. 1371 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL. 1372 */ 1373 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE)) 1374 t = ZIO_TYPE_NULL; 1375 1376 /* 1377 * A similar issue exists for the L2ARC write thread until L2ARC 2.0. 1378 */ 1379 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux) 1380 t = ZIO_TYPE_NULL; 1381 1382 /* 1383 * If this is a high priority I/O, then use the high priority taskq if 1384 * available. 1385 */ 1386 if (zio->io_priority == ZIO_PRIORITY_NOW && 1387 spa->spa_zio_taskq[t][q + 1].stqs_count != 0) 1388 q++; 1389 1390 ASSERT3U(q, <, ZIO_TASKQ_TYPES); 1391 1392 /* 1393 * NB: We are assuming that the zio can only be dispatched 1394 * to a single taskq at a time. It would be a grievous error 1395 * to dispatch the zio to another taskq at the same time. 1396 */ 1397#if defined(illumos) || !defined(_KERNEL) 1398 ASSERT(zio->io_tqent.tqent_next == NULL); 1399#else 1400 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0); 1401#endif 1402 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio, 1403 flags, &zio->io_tqent); 1404} 1405 1406static boolean_t 1407zio_taskq_member(zio_t *zio, zio_taskq_type_t q) 1408{ 1409 kthread_t *executor = zio->io_executor; 1410 spa_t *spa = zio->io_spa; 1411 1412 for (zio_type_t t = 0; t < ZIO_TYPES; t++) { 1413 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q]; 1414 uint_t i; 1415 for (i = 0; i < tqs->stqs_count; i++) { 1416 if (taskq_member(tqs->stqs_taskq[i], executor)) 1417 return (B_TRUE); 1418 } 1419 } 1420 1421 return (B_FALSE); 1422} 1423 1424static int 1425zio_issue_async(zio_t *zio) 1426{ 1427 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); 1428 1429 return (ZIO_PIPELINE_STOP); 1430} 1431 1432void 1433zio_interrupt(zio_t *zio) 1434{ 1435 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE); 1436} 1437 1438/* 1439 * Execute the I/O pipeline until one of the following occurs: 1440 * 1441 * (1) the I/O completes 1442 * (2) the pipeline stalls waiting for dependent child I/Os 1443 * (3) the I/O issues, so we're waiting for an I/O completion interrupt 1444 * (4) the I/O is delegated by vdev-level caching or aggregation 1445 * (5) the I/O is deferred due to vdev-level queueing 1446 * (6) the I/O is handed off to another thread. 1447 * 1448 * In all cases, the pipeline stops whenever there's no CPU work; it never 1449 * burns a thread in cv_wait(). 1450 * 1451 * There's no locking on io_stage because there's no legitimate way 1452 * for multiple threads to be attempting to process the same I/O. 1453 */ 1454static zio_pipe_stage_t *zio_pipeline[]; 1455 1456void 1457zio_execute(zio_t *zio) 1458{ 1459 zio->io_executor = curthread; 1460 1461 while (zio->io_stage < ZIO_STAGE_DONE) { 1462 enum zio_stage pipeline = zio->io_pipeline; 1463 enum zio_stage stage = zio->io_stage; 1464 int rv; 1465 1466 ASSERT(!MUTEX_HELD(&zio->io_lock)); 1467 ASSERT(ISP2(stage)); 1468 ASSERT(zio->io_stall == NULL); 1469 1470 do { 1471 stage <<= 1; 1472 } while ((stage & pipeline) == 0); 1473 1474 ASSERT(stage <= ZIO_STAGE_DONE); 1475 1476 /* 1477 * If we are in interrupt context and this pipeline stage 1478 * will grab a config lock that is held across I/O, 1479 * or may wait for an I/O that needs an interrupt thread 1480 * to complete, issue async to avoid deadlock. 1481 * 1482 * For VDEV_IO_START, we cut in line so that the io will 1483 * be sent to disk promptly. 1484 */ 1485 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL && 1486 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) { 1487 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? 1488 zio_requeue_io_start_cut_in_line : B_FALSE; 1489 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); 1490 return; 1491 } 1492 1493 zio->io_stage = stage; 1494 rv = zio_pipeline[highbit64(stage) - 1](zio); 1495 1496 if (rv == ZIO_PIPELINE_STOP) 1497 return; 1498 1499 ASSERT(rv == ZIO_PIPELINE_CONTINUE); 1500 } 1501} 1502 1503/* 1504 * ========================================================================== 1505 * Initiate I/O, either sync or async 1506 * ========================================================================== 1507 */ 1508int 1509zio_wait(zio_t *zio) 1510{ 1511 int error; 1512 1513 ASSERT(zio->io_stage == ZIO_STAGE_OPEN); 1514 ASSERT(zio->io_executor == NULL); 1515 1516 zio->io_waiter = curthread; 1517 1518 zio_execute(zio); 1519 1520 mutex_enter(&zio->io_lock); 1521 while (zio->io_executor != NULL) 1522 cv_wait(&zio->io_cv, &zio->io_lock); 1523 mutex_exit(&zio->io_lock); 1524 1525 error = zio->io_error; 1526 zio_destroy(zio); 1527 1528 return (error); 1529} 1530 1531void 1532zio_nowait(zio_t *zio) 1533{ 1534 ASSERT(zio->io_executor == NULL); 1535 1536 if (zio->io_child_type == ZIO_CHILD_LOGICAL && 1537 zio_unique_parent(zio) == NULL) { 1538 /* 1539 * This is a logical async I/O with no parent to wait for it. 1540 * We add it to the spa_async_root_zio "Godfather" I/O which 1541 * will ensure they complete prior to unloading the pool. 1542 */ 1543 spa_t *spa = zio->io_spa; 1544 1545 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio); 1546 } 1547 1548 zio_execute(zio); 1549} 1550 1551/* 1552 * ========================================================================== 1553 * Reexecute or suspend/resume failed I/O 1554 * ========================================================================== 1555 */ 1556 1557static void 1558zio_reexecute(zio_t *pio) 1559{ 1560 zio_t *cio, *cio_next; 1561 1562 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL); 1563 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN); 1564 ASSERT(pio->io_gang_leader == NULL); 1565 ASSERT(pio->io_gang_tree == NULL); 1566 1567 pio->io_flags = pio->io_orig_flags; 1568 pio->io_stage = pio->io_orig_stage; 1569 pio->io_pipeline = pio->io_orig_pipeline; 1570 pio->io_reexecute = 0; 1571 pio->io_flags |= ZIO_FLAG_REEXECUTED; 1572 pio->io_error = 0; 1573 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 1574 pio->io_state[w] = 0; 1575 for (int c = 0; c < ZIO_CHILD_TYPES; c++) 1576 pio->io_child_error[c] = 0; 1577 1578 if (IO_IS_ALLOCATING(pio)) 1579 BP_ZERO(pio->io_bp); 1580 1581 /* 1582 * As we reexecute pio's children, new children could be created. 1583 * New children go to the head of pio's io_child_list, however, 1584 * so we will (correctly) not reexecute them. The key is that 1585 * the remainder of pio's io_child_list, from 'cio_next' onward, 1586 * cannot be affected by any side effects of reexecuting 'cio'. 1587 */ 1588 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) { 1589 cio_next = zio_walk_children(pio); 1590 mutex_enter(&pio->io_lock); 1591 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 1592 pio->io_children[cio->io_child_type][w]++; 1593 mutex_exit(&pio->io_lock); 1594 zio_reexecute(cio); 1595 } 1596 1597 /* 1598 * Now that all children have been reexecuted, execute the parent. 1599 * We don't reexecute "The Godfather" I/O here as it's the 1600 * responsibility of the caller to wait on him. 1601 */ 1602 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) 1603 zio_execute(pio); 1604} 1605 1606void 1607zio_suspend(spa_t *spa, zio_t *zio) 1608{ 1609 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC) 1610 fm_panic("Pool '%s' has encountered an uncorrectable I/O " 1611 "failure and the failure mode property for this pool " 1612 "is set to panic.", spa_name(spa)); 1613 1614 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0); 1615 1616 mutex_enter(&spa->spa_suspend_lock); 1617 1618 if (spa->spa_suspend_zio_root == NULL) 1619 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL, 1620 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | 1621 ZIO_FLAG_GODFATHER); 1622 1623 spa->spa_suspended = B_TRUE; 1624 1625 if (zio != NULL) { 1626 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); 1627 ASSERT(zio != spa->spa_suspend_zio_root); 1628 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1629 ASSERT(zio_unique_parent(zio) == NULL); 1630 ASSERT(zio->io_stage == ZIO_STAGE_DONE); 1631 zio_add_child(spa->spa_suspend_zio_root, zio); 1632 } 1633 1634 mutex_exit(&spa->spa_suspend_lock); 1635} 1636 1637int 1638zio_resume(spa_t *spa) 1639{ 1640 zio_t *pio; 1641 1642 /* 1643 * Reexecute all previously suspended i/o. 1644 */ 1645 mutex_enter(&spa->spa_suspend_lock); 1646 spa->spa_suspended = B_FALSE; 1647 cv_broadcast(&spa->spa_suspend_cv); 1648 pio = spa->spa_suspend_zio_root; 1649 spa->spa_suspend_zio_root = NULL; 1650 mutex_exit(&spa->spa_suspend_lock); 1651 1652 if (pio == NULL) 1653 return (0); 1654 1655 zio_reexecute(pio); 1656 return (zio_wait(pio)); 1657} 1658 1659void 1660zio_resume_wait(spa_t *spa) 1661{ 1662 mutex_enter(&spa->spa_suspend_lock); 1663 while (spa_suspended(spa)) 1664 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock); 1665 mutex_exit(&spa->spa_suspend_lock); 1666} 1667 1668/* 1669 * ========================================================================== 1670 * Gang blocks. 1671 * 1672 * A gang block is a collection of small blocks that looks to the DMU 1673 * like one large block. When zio_dva_allocate() cannot find a block 1674 * of the requested size, due to either severe fragmentation or the pool 1675 * being nearly full, it calls zio_write_gang_block() to construct the 1676 * block from smaller fragments. 1677 * 1678 * A gang block consists of a gang header (zio_gbh_phys_t) and up to 1679 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like 1680 * an indirect block: it's an array of block pointers. It consumes 1681 * only one sector and hence is allocatable regardless of fragmentation. 1682 * The gang header's bps point to its gang members, which hold the data. 1683 * 1684 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg> 1685 * as the verifier to ensure uniqueness of the SHA256 checksum. 1686 * Critically, the gang block bp's blk_cksum is the checksum of the data, 1687 * not the gang header. This ensures that data block signatures (needed for 1688 * deduplication) are independent of how the block is physically stored. 1689 * 1690 * Gang blocks can be nested: a gang member may itself be a gang block. 1691 * Thus every gang block is a tree in which root and all interior nodes are 1692 * gang headers, and the leaves are normal blocks that contain user data. 1693 * The root of the gang tree is called the gang leader. 1694 * 1695 * To perform any operation (read, rewrite, free, claim) on a gang block, 1696 * zio_gang_assemble() first assembles the gang tree (minus data leaves) 1697 * in the io_gang_tree field of the original logical i/o by recursively 1698 * reading the gang leader and all gang headers below it. This yields 1699 * an in-core tree containing the contents of every gang header and the 1700 * bps for every constituent of the gang block. 1701 * 1702 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree 1703 * and invokes a callback on each bp. To free a gang block, zio_gang_issue() 1704 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp. 1705 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim(). 1706 * zio_read_gang() is a wrapper around zio_read() that omits reading gang 1707 * headers, since we already have those in io_gang_tree. zio_rewrite_gang() 1708 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite() 1709 * of the gang header plus zio_checksum_compute() of the data to update the 1710 * gang header's blk_cksum as described above. 1711 * 1712 * The two-phase assemble/issue model solves the problem of partial failure -- 1713 * what if you'd freed part of a gang block but then couldn't read the 1714 * gang header for another part? Assembling the entire gang tree first 1715 * ensures that all the necessary gang header I/O has succeeded before 1716 * starting the actual work of free, claim, or write. Once the gang tree 1717 * is assembled, free and claim are in-memory operations that cannot fail. 1718 * 1719 * In the event that a gang write fails, zio_dva_unallocate() walks the 1720 * gang tree to immediately free (i.e. insert back into the space map) 1721 * everything we've allocated. This ensures that we don't get ENOSPC 1722 * errors during repeated suspend/resume cycles due to a flaky device. 1723 * 1724 * Gang rewrites only happen during sync-to-convergence. If we can't assemble 1725 * the gang tree, we won't modify the block, so we can safely defer the free 1726 * (knowing that the block is still intact). If we *can* assemble the gang 1727 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free 1728 * each constituent bp and we can allocate a new block on the next sync pass. 1729 * 1730 * In all cases, the gang tree allows complete recovery from partial failure. 1731 * ========================================================================== 1732 */ 1733 1734static zio_t * 1735zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1736{ 1737 if (gn != NULL) 1738 return (pio); 1739 1740 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp), 1741 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), 1742 &pio->io_bookmark)); 1743} 1744 1745zio_t * 1746zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1747{ 1748 zio_t *zio; 1749 1750 if (gn != NULL) { 1751 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, 1752 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority, 1753 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 1754 /* 1755 * As we rewrite each gang header, the pipeline will compute 1756 * a new gang block header checksum for it; but no one will 1757 * compute a new data checksum, so we do that here. The one 1758 * exception is the gang leader: the pipeline already computed 1759 * its data checksum because that stage precedes gang assembly. 1760 * (Presently, nothing actually uses interior data checksums; 1761 * this is just good hygiene.) 1762 */ 1763 if (gn != pio->io_gang_leader->io_gang_tree) { 1764 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp), 1765 data, BP_GET_PSIZE(bp)); 1766 } 1767 /* 1768 * If we are here to damage data for testing purposes, 1769 * leave the GBH alone so that we can detect the damage. 1770 */ 1771 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE) 1772 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; 1773 } else { 1774 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, 1775 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority, 1776 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 1777 } 1778 1779 return (zio); 1780} 1781 1782/* ARGSUSED */ 1783zio_t * 1784zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1785{ 1786 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp, 1787 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp), 1788 ZIO_GANG_CHILD_FLAGS(pio))); 1789} 1790 1791/* ARGSUSED */ 1792zio_t * 1793zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1794{ 1795 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp, 1796 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio))); 1797} 1798 1799static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = { 1800 NULL, 1801 zio_read_gang, 1802 zio_rewrite_gang, 1803 zio_free_gang, 1804 zio_claim_gang, 1805 NULL 1806}; 1807 1808static void zio_gang_tree_assemble_done(zio_t *zio); 1809 1810static zio_gang_node_t * 1811zio_gang_node_alloc(zio_gang_node_t **gnpp) 1812{ 1813 zio_gang_node_t *gn; 1814 1815 ASSERT(*gnpp == NULL); 1816 1817 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP); 1818 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE); 1819 *gnpp = gn; 1820 1821 return (gn); 1822} 1823 1824static void 1825zio_gang_node_free(zio_gang_node_t **gnpp) 1826{ 1827 zio_gang_node_t *gn = *gnpp; 1828 1829 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) 1830 ASSERT(gn->gn_child[g] == NULL); 1831 1832 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE); 1833 kmem_free(gn, sizeof (*gn)); 1834 *gnpp = NULL; 1835} 1836 1837static void 1838zio_gang_tree_free(zio_gang_node_t **gnpp) 1839{ 1840 zio_gang_node_t *gn = *gnpp; 1841 1842 if (gn == NULL) 1843 return; 1844 1845 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) 1846 zio_gang_tree_free(&gn->gn_child[g]); 1847 1848 zio_gang_node_free(gnpp); 1849} 1850 1851static void 1852zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp) 1853{ 1854 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp); 1855 1856 ASSERT(gio->io_gang_leader == gio); 1857 ASSERT(BP_IS_GANG(bp)); 1858 1859 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh, 1860 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn, 1861 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark)); 1862} 1863 1864static void 1865zio_gang_tree_assemble_done(zio_t *zio) 1866{ 1867 zio_t *gio = zio->io_gang_leader; 1868 zio_gang_node_t *gn = zio->io_private; 1869 blkptr_t *bp = zio->io_bp; 1870 1871 ASSERT(gio == zio_unique_parent(zio)); 1872 ASSERT(zio->io_child_count == 0); 1873 1874 if (zio->io_error) 1875 return; 1876 1877 if (BP_SHOULD_BYTESWAP(bp)) 1878 byteswap_uint64_array(zio->io_data, zio->io_size); 1879 1880 ASSERT(zio->io_data == gn->gn_gbh); 1881 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE); 1882 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); 1883 1884 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 1885 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; 1886 if (!BP_IS_GANG(gbp)) 1887 continue; 1888 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]); 1889 } 1890} 1891 1892static void 1893zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data) 1894{ 1895 zio_t *gio = pio->io_gang_leader; 1896 zio_t *zio; 1897 1898 ASSERT(BP_IS_GANG(bp) == !!gn); 1899 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp)); 1900 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree); 1901 1902 /* 1903 * If you're a gang header, your data is in gn->gn_gbh. 1904 * If you're a gang member, your data is in 'data' and gn == NULL. 1905 */ 1906 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data); 1907 1908 if (gn != NULL) { 1909 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); 1910 1911 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 1912 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; 1913 if (BP_IS_HOLE(gbp)) 1914 continue; 1915 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data); 1916 data = (char *)data + BP_GET_PSIZE(gbp); 1917 } 1918 } 1919 1920 if (gn == gio->io_gang_tree && gio->io_data != NULL) 1921 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data); 1922 1923 if (zio != pio) 1924 zio_nowait(zio); 1925} 1926 1927static int 1928zio_gang_assemble(zio_t *zio) 1929{ 1930 blkptr_t *bp = zio->io_bp; 1931 1932 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL); 1933 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 1934 1935 zio->io_gang_leader = zio; 1936 1937 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree); 1938 1939 return (ZIO_PIPELINE_CONTINUE); 1940} 1941 1942static int 1943zio_gang_issue(zio_t *zio) 1944{ 1945 blkptr_t *bp = zio->io_bp; 1946 1947 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE)) 1948 return (ZIO_PIPELINE_STOP); 1949 1950 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio); 1951 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 1952 1953 if (zio->io_child_error[ZIO_CHILD_GANG] == 0) 1954 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data); 1955 else 1956 zio_gang_tree_free(&zio->io_gang_tree); 1957 1958 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1959 1960 return (ZIO_PIPELINE_CONTINUE); 1961} 1962 1963static void 1964zio_write_gang_member_ready(zio_t *zio) 1965{ 1966 zio_t *pio = zio_unique_parent(zio); 1967 zio_t *gio = zio->io_gang_leader; 1968 dva_t *cdva = zio->io_bp->blk_dva; 1969 dva_t *pdva = pio->io_bp->blk_dva; 1970 uint64_t asize; 1971 1972 if (BP_IS_HOLE(zio->io_bp)) 1973 return; 1974 1975 ASSERT(BP_IS_HOLE(&zio->io_bp_orig)); 1976 1977 ASSERT(zio->io_child_type == ZIO_CHILD_GANG); 1978 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies); 1979 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp)); 1980 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp)); 1981 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp)); 1982 1983 mutex_enter(&pio->io_lock); 1984 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) { 1985 ASSERT(DVA_GET_GANG(&pdva[d])); 1986 asize = DVA_GET_ASIZE(&pdva[d]); 1987 asize += DVA_GET_ASIZE(&cdva[d]); 1988 DVA_SET_ASIZE(&pdva[d], asize); 1989 } 1990 mutex_exit(&pio->io_lock); 1991} 1992 1993static int 1994zio_write_gang_block(zio_t *pio) 1995{ 1996 spa_t *spa = pio->io_spa; 1997 blkptr_t *bp = pio->io_bp; 1998 zio_t *gio = pio->io_gang_leader; 1999 zio_t *zio; 2000 zio_gang_node_t *gn, **gnpp; 2001 zio_gbh_phys_t *gbh; 2002 uint64_t txg = pio->io_txg; 2003 uint64_t resid = pio->io_size; 2004 uint64_t lsize; 2005 int copies = gio->io_prop.zp_copies; 2006 int gbh_copies = MIN(copies + 1, spa_max_replication(spa)); 2007 zio_prop_t zp; 2008 int error; 2009 2010 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE, 2011 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, 2012 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER); 2013 if (error) { 2014 pio->io_error = error; 2015 return (ZIO_PIPELINE_CONTINUE); 2016 } 2017 2018 if (pio == gio) { 2019 gnpp = &gio->io_gang_tree; 2020 } else { 2021 gnpp = pio->io_private; 2022 ASSERT(pio->io_ready == zio_write_gang_member_ready); 2023 } 2024 2025 gn = zio_gang_node_alloc(gnpp); 2026 gbh = gn->gn_gbh; 2027 bzero(gbh, SPA_GANGBLOCKSIZE); 2028 2029 /* 2030 * Create the gang header. 2031 */ 2032 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL, 2033 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 2034 2035 /* 2036 * Create and nowait the gang children. 2037 */ 2038 for (int g = 0; resid != 0; resid -= lsize, g++) { 2039 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g), 2040 SPA_MINBLOCKSIZE); 2041 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid); 2042 2043 zp.zp_checksum = gio->io_prop.zp_checksum; 2044 zp.zp_compress = ZIO_COMPRESS_OFF; 2045 zp.zp_type = DMU_OT_NONE; 2046 zp.zp_level = 0; 2047 zp.zp_copies = gio->io_prop.zp_copies; 2048 zp.zp_dedup = B_FALSE; 2049 zp.zp_dedup_verify = B_FALSE; 2050 zp.zp_nopwrite = B_FALSE; 2051 2052 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g], 2053 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp, 2054 zio_write_gang_member_ready, NULL, NULL, &gn->gn_child[g], 2055 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), 2056 &pio->io_bookmark)); 2057 } 2058 2059 /* 2060 * Set pio's pipeline to just wait for zio to finish. 2061 */ 2062 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2063 2064 zio_nowait(zio); 2065 2066 return (ZIO_PIPELINE_CONTINUE); 2067} 2068 2069/* 2070 * The zio_nop_write stage in the pipeline determines if allocating 2071 * a new bp is necessary. By leveraging a cryptographically secure checksum, 2072 * such as SHA256, we can compare the checksums of the new data and the old 2073 * to determine if allocating a new block is required. The nopwrite 2074 * feature can handle writes in either syncing or open context (i.e. zil 2075 * writes) and as a result is mutually exclusive with dedup. 2076 */ 2077static int 2078zio_nop_write(zio_t *zio) 2079{ 2080 blkptr_t *bp = zio->io_bp; 2081 blkptr_t *bp_orig = &zio->io_bp_orig; 2082 zio_prop_t *zp = &zio->io_prop; 2083 2084 ASSERT(BP_GET_LEVEL(bp) == 0); 2085 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 2086 ASSERT(zp->zp_nopwrite); 2087 ASSERT(!zp->zp_dedup); 2088 ASSERT(zio->io_bp_override == NULL); 2089 ASSERT(IO_IS_ALLOCATING(zio)); 2090 2091 /* 2092 * Check to see if the original bp and the new bp have matching 2093 * characteristics (i.e. same checksum, compression algorithms, etc). 2094 * If they don't then just continue with the pipeline which will 2095 * allocate a new bp. 2096 */ 2097 if (BP_IS_HOLE(bp_orig) || 2098 !zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup || 2099 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) || 2100 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) || 2101 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) || 2102 zp->zp_copies != BP_GET_NDVAS(bp_orig)) 2103 return (ZIO_PIPELINE_CONTINUE); 2104 2105 /* 2106 * If the checksums match then reset the pipeline so that we 2107 * avoid allocating a new bp and issuing any I/O. 2108 */ 2109 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) { 2110 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup); 2111 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig)); 2112 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig)); 2113 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF); 2114 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop, 2115 sizeof (uint64_t)) == 0); 2116 2117 *bp = *bp_orig; 2118 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2119 zio->io_flags |= ZIO_FLAG_NOPWRITE; 2120 } 2121 2122 return (ZIO_PIPELINE_CONTINUE); 2123} 2124 2125/* 2126 * ========================================================================== 2127 * Dedup 2128 * ========================================================================== 2129 */ 2130static void 2131zio_ddt_child_read_done(zio_t *zio) 2132{ 2133 blkptr_t *bp = zio->io_bp; 2134 ddt_entry_t *dde = zio->io_private; 2135 ddt_phys_t *ddp; 2136 zio_t *pio = zio_unique_parent(zio); 2137 2138 mutex_enter(&pio->io_lock); 2139 ddp = ddt_phys_select(dde, bp); 2140 if (zio->io_error == 0) 2141 ddt_phys_clear(ddp); /* this ddp doesn't need repair */ 2142 if (zio->io_error == 0 && dde->dde_repair_data == NULL) 2143 dde->dde_repair_data = zio->io_data; 2144 else 2145 zio_buf_free(zio->io_data, zio->io_size); 2146 mutex_exit(&pio->io_lock); 2147} 2148 2149static int 2150zio_ddt_read_start(zio_t *zio) 2151{ 2152 blkptr_t *bp = zio->io_bp; 2153 2154 ASSERT(BP_GET_DEDUP(bp)); 2155 ASSERT(BP_GET_PSIZE(bp) == zio->io_size); 2156 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2157 2158 if (zio->io_child_error[ZIO_CHILD_DDT]) { 2159 ddt_t *ddt = ddt_select(zio->io_spa, bp); 2160 ddt_entry_t *dde = ddt_repair_start(ddt, bp); 2161 ddt_phys_t *ddp = dde->dde_phys; 2162 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp); 2163 blkptr_t blk; 2164 2165 ASSERT(zio->io_vsd == NULL); 2166 zio->io_vsd = dde; 2167 2168 if (ddp_self == NULL) 2169 return (ZIO_PIPELINE_CONTINUE); 2170 2171 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) { 2172 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self) 2173 continue; 2174 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp, 2175 &blk); 2176 zio_nowait(zio_read(zio, zio->io_spa, &blk, 2177 zio_buf_alloc(zio->io_size), zio->io_size, 2178 zio_ddt_child_read_done, dde, zio->io_priority, 2179 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE, 2180 &zio->io_bookmark)); 2181 } 2182 return (ZIO_PIPELINE_CONTINUE); 2183 } 2184 2185 zio_nowait(zio_read(zio, zio->io_spa, bp, 2186 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority, 2187 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark)); 2188 2189 return (ZIO_PIPELINE_CONTINUE); 2190} 2191 2192static int 2193zio_ddt_read_done(zio_t *zio) 2194{ 2195 blkptr_t *bp = zio->io_bp; 2196 2197 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE)) 2198 return (ZIO_PIPELINE_STOP); 2199 2200 ASSERT(BP_GET_DEDUP(bp)); 2201 ASSERT(BP_GET_PSIZE(bp) == zio->io_size); 2202 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2203 2204 if (zio->io_child_error[ZIO_CHILD_DDT]) { 2205 ddt_t *ddt = ddt_select(zio->io_spa, bp); 2206 ddt_entry_t *dde = zio->io_vsd; 2207 if (ddt == NULL) { 2208 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE); 2209 return (ZIO_PIPELINE_CONTINUE); 2210 } 2211 if (dde == NULL) { 2212 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1; 2213 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); 2214 return (ZIO_PIPELINE_STOP); 2215 } 2216 if (dde->dde_repair_data != NULL) { 2217 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size); 2218 zio->io_child_error[ZIO_CHILD_DDT] = 0; 2219 } 2220 ddt_repair_done(ddt, dde); 2221 zio->io_vsd = NULL; 2222 } 2223 2224 ASSERT(zio->io_vsd == NULL); 2225 2226 return (ZIO_PIPELINE_CONTINUE); 2227} 2228 2229static boolean_t 2230zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde) 2231{ 2232 spa_t *spa = zio->io_spa; 2233 2234 /* 2235 * Note: we compare the original data, not the transformed data, 2236 * because when zio->io_bp is an override bp, we will not have 2237 * pushed the I/O transforms. That's an important optimization 2238 * because otherwise we'd compress/encrypt all dmu_sync() data twice. 2239 */ 2240 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { 2241 zio_t *lio = dde->dde_lead_zio[p]; 2242 2243 if (lio != NULL) { 2244 return (lio->io_orig_size != zio->io_orig_size || 2245 bcmp(zio->io_orig_data, lio->io_orig_data, 2246 zio->io_orig_size) != 0); 2247 } 2248 } 2249 2250 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { 2251 ddt_phys_t *ddp = &dde->dde_phys[p]; 2252 2253 if (ddp->ddp_phys_birth != 0) { 2254 arc_buf_t *abuf = NULL; 2255 arc_flags_t aflags = ARC_FLAG_WAIT; 2256 blkptr_t blk = *zio->io_bp; 2257 int error; 2258 2259 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth); 2260 2261 ddt_exit(ddt); 2262 2263 error = arc_read(NULL, spa, &blk, 2264 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ, 2265 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2266 &aflags, &zio->io_bookmark); 2267 2268 if (error == 0) { 2269 if (arc_buf_size(abuf) != zio->io_orig_size || 2270 bcmp(abuf->b_data, zio->io_orig_data, 2271 zio->io_orig_size) != 0) 2272 error = SET_ERROR(EEXIST); 2273 VERIFY(arc_buf_remove_ref(abuf, &abuf)); 2274 } 2275 2276 ddt_enter(ddt); 2277 return (error != 0); 2278 } 2279 } 2280 2281 return (B_FALSE); 2282} 2283 2284static void 2285zio_ddt_child_write_ready(zio_t *zio) 2286{ 2287 int p = zio->io_prop.zp_copies; 2288 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); 2289 ddt_entry_t *dde = zio->io_private; 2290 ddt_phys_t *ddp = &dde->dde_phys[p]; 2291 zio_t *pio; 2292 2293 if (zio->io_error) 2294 return; 2295 2296 ddt_enter(ddt); 2297 2298 ASSERT(dde->dde_lead_zio[p] == zio); 2299 2300 ddt_phys_fill(ddp, zio->io_bp); 2301 2302 while ((pio = zio_walk_parents(zio)) != NULL) 2303 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg); 2304 2305 ddt_exit(ddt); 2306} 2307 2308static void 2309zio_ddt_child_write_done(zio_t *zio) 2310{ 2311 int p = zio->io_prop.zp_copies; 2312 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); 2313 ddt_entry_t *dde = zio->io_private; 2314 ddt_phys_t *ddp = &dde->dde_phys[p]; 2315 2316 ddt_enter(ddt); 2317 2318 ASSERT(ddp->ddp_refcnt == 0); 2319 ASSERT(dde->dde_lead_zio[p] == zio); 2320 dde->dde_lead_zio[p] = NULL; 2321 2322 if (zio->io_error == 0) { 2323 while (zio_walk_parents(zio) != NULL) 2324 ddt_phys_addref(ddp); 2325 } else { 2326 ddt_phys_clear(ddp); 2327 } 2328 2329 ddt_exit(ddt); 2330} 2331 2332static void 2333zio_ddt_ditto_write_done(zio_t *zio) 2334{ 2335 int p = DDT_PHYS_DITTO; 2336 zio_prop_t *zp = &zio->io_prop; 2337 blkptr_t *bp = zio->io_bp; 2338 ddt_t *ddt = ddt_select(zio->io_spa, bp); 2339 ddt_entry_t *dde = zio->io_private; 2340 ddt_phys_t *ddp = &dde->dde_phys[p]; 2341 ddt_key_t *ddk = &dde->dde_key; 2342 2343 ddt_enter(ddt); 2344 2345 ASSERT(ddp->ddp_refcnt == 0); 2346 ASSERT(dde->dde_lead_zio[p] == zio); 2347 dde->dde_lead_zio[p] = NULL; 2348 2349 if (zio->io_error == 0) { 2350 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum)); 2351 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP); 2352 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp)); 2353 if (ddp->ddp_phys_birth != 0) 2354 ddt_phys_free(ddt, ddk, ddp, zio->io_txg); 2355 ddt_phys_fill(ddp, bp); 2356 } 2357 2358 ddt_exit(ddt); 2359} 2360 2361static int 2362zio_ddt_write(zio_t *zio) 2363{ 2364 spa_t *spa = zio->io_spa; 2365 blkptr_t *bp = zio->io_bp; 2366 uint64_t txg = zio->io_txg; 2367 zio_prop_t *zp = &zio->io_prop; 2368 int p = zp->zp_copies; 2369 int ditto_copies; 2370 zio_t *cio = NULL; 2371 zio_t *dio = NULL; 2372 ddt_t *ddt = ddt_select(spa, bp); 2373 ddt_entry_t *dde; 2374 ddt_phys_t *ddp; 2375 2376 ASSERT(BP_GET_DEDUP(bp)); 2377 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum); 2378 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override); 2379 2380 ddt_enter(ddt); 2381 dde = ddt_lookup(ddt, bp, B_TRUE); 2382 ddp = &dde->dde_phys[p]; 2383 2384 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) { 2385 /* 2386 * If we're using a weak checksum, upgrade to a strong checksum 2387 * and try again. If we're already using a strong checksum, 2388 * we can't resolve it, so just convert to an ordinary write. 2389 * (And automatically e-mail a paper to Nature?) 2390 */ 2391 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) { 2392 zp->zp_checksum = spa_dedup_checksum(spa); 2393 zio_pop_transforms(zio); 2394 zio->io_stage = ZIO_STAGE_OPEN; 2395 BP_ZERO(bp); 2396 } else { 2397 zp->zp_dedup = B_FALSE; 2398 } 2399 zio->io_pipeline = ZIO_WRITE_PIPELINE; 2400 ddt_exit(ddt); 2401 return (ZIO_PIPELINE_CONTINUE); 2402 } 2403 2404 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp); 2405 ASSERT(ditto_copies < SPA_DVAS_PER_BP); 2406 2407 if (ditto_copies > ddt_ditto_copies_present(dde) && 2408 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) { 2409 zio_prop_t czp = *zp; 2410 2411 czp.zp_copies = ditto_copies; 2412 2413 /* 2414 * If we arrived here with an override bp, we won't have run 2415 * the transform stack, so we won't have the data we need to 2416 * generate a child i/o. So, toss the override bp and restart. 2417 * This is safe, because using the override bp is just an 2418 * optimization; and it's rare, so the cost doesn't matter. 2419 */ 2420 if (zio->io_bp_override) { 2421 zio_pop_transforms(zio); 2422 zio->io_stage = ZIO_STAGE_OPEN; 2423 zio->io_pipeline = ZIO_WRITE_PIPELINE; 2424 zio->io_bp_override = NULL; 2425 BP_ZERO(bp); 2426 ddt_exit(ddt); 2427 return (ZIO_PIPELINE_CONTINUE); 2428 } 2429 2430 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data, 2431 zio->io_orig_size, &czp, NULL, NULL, 2432 zio_ddt_ditto_write_done, dde, zio->io_priority, 2433 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); 2434 2435 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL); 2436 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio; 2437 } 2438 2439 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) { 2440 if (ddp->ddp_phys_birth != 0) 2441 ddt_bp_fill(ddp, bp, txg); 2442 if (dde->dde_lead_zio[p] != NULL) 2443 zio_add_child(zio, dde->dde_lead_zio[p]); 2444 else 2445 ddt_phys_addref(ddp); 2446 } else if (zio->io_bp_override) { 2447 ASSERT(bp->blk_birth == txg); 2448 ASSERT(BP_EQUAL(bp, zio->io_bp_override)); 2449 ddt_phys_fill(ddp, bp); 2450 ddt_phys_addref(ddp); 2451 } else { 2452 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data, 2453 zio->io_orig_size, zp, zio_ddt_child_write_ready, NULL, 2454 zio_ddt_child_write_done, dde, zio->io_priority, 2455 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); 2456 2457 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL); 2458 dde->dde_lead_zio[p] = cio; 2459 } 2460 2461 ddt_exit(ddt); 2462 2463 if (cio) 2464 zio_nowait(cio); 2465 if (dio) 2466 zio_nowait(dio); 2467 2468 return (ZIO_PIPELINE_CONTINUE); 2469} 2470 2471ddt_entry_t *freedde; /* for debugging */ 2472 2473static int 2474zio_ddt_free(zio_t *zio) 2475{ 2476 spa_t *spa = zio->io_spa; 2477 blkptr_t *bp = zio->io_bp; 2478 ddt_t *ddt = ddt_select(spa, bp); 2479 ddt_entry_t *dde; 2480 ddt_phys_t *ddp; 2481 2482 ASSERT(BP_GET_DEDUP(bp)); 2483 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2484 2485 ddt_enter(ddt); 2486 freedde = dde = ddt_lookup(ddt, bp, B_TRUE); 2487 ddp = ddt_phys_select(dde, bp); 2488 ddt_phys_decref(ddp); 2489 ddt_exit(ddt); 2490 2491 return (ZIO_PIPELINE_CONTINUE); 2492} 2493 2494/* 2495 * ========================================================================== 2496 * Allocate and free blocks 2497 * ========================================================================== 2498 */ 2499static int 2500zio_dva_allocate(zio_t *zio) 2501{ 2502 spa_t *spa = zio->io_spa; 2503 metaslab_class_t *mc = spa_normal_class(spa); 2504 blkptr_t *bp = zio->io_bp; 2505 int error; 2506 int flags = 0; 2507 2508 if (zio->io_gang_leader == NULL) { 2509 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 2510 zio->io_gang_leader = zio; 2511 } 2512 2513 ASSERT(BP_IS_HOLE(bp)); 2514 ASSERT0(BP_GET_NDVAS(bp)); 2515 ASSERT3U(zio->io_prop.zp_copies, >, 0); 2516 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa)); 2517 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp)); 2518 2519 /* 2520 * The dump device does not support gang blocks so allocation on 2521 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid 2522 * the "fast" gang feature. 2523 */ 2524 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0; 2525 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ? 2526 METASLAB_GANG_CHILD : 0; 2527 error = metaslab_alloc(spa, mc, zio->io_size, bp, 2528 zio->io_prop.zp_copies, zio->io_txg, NULL, flags); 2529 2530 if (error) { 2531 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, " 2532 "size %llu, error %d", spa_name(spa), zio, zio->io_size, 2533 error); 2534 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE) 2535 return (zio_write_gang_block(zio)); 2536 zio->io_error = error; 2537 } 2538 2539 return (ZIO_PIPELINE_CONTINUE); 2540} 2541 2542static int 2543zio_dva_free(zio_t *zio) 2544{ 2545 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE); 2546 2547 return (ZIO_PIPELINE_CONTINUE); 2548} 2549 2550static int 2551zio_dva_claim(zio_t *zio) 2552{ 2553 int error; 2554 2555 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg); 2556 if (error) 2557 zio->io_error = error; 2558 2559 return (ZIO_PIPELINE_CONTINUE); 2560} 2561 2562/* 2563 * Undo an allocation. This is used by zio_done() when an I/O fails 2564 * and we want to give back the block we just allocated. 2565 * This handles both normal blocks and gang blocks. 2566 */ 2567static void 2568zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp) 2569{ 2570 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp)); 2571 ASSERT(zio->io_bp_override == NULL); 2572 2573 if (!BP_IS_HOLE(bp)) 2574 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE); 2575 2576 if (gn != NULL) { 2577 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 2578 zio_dva_unallocate(zio, gn->gn_child[g], 2579 &gn->gn_gbh->zg_blkptr[g]); 2580 } 2581 } 2582} 2583 2584/* 2585 * Try to allocate an intent log block. Return 0 on success, errno on failure. 2586 */ 2587int 2588zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp, 2589 uint64_t size, boolean_t use_slog) 2590{ 2591 int error = 1; 2592 2593 ASSERT(txg > spa_syncing_txg(spa)); 2594 2595 /* 2596 * ZIL blocks are always contiguous (i.e. not gang blocks) so we 2597 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang" 2598 * when allocating them. 2599 */ 2600 if (use_slog) { 2601 error = metaslab_alloc(spa, spa_log_class(spa), size, 2602 new_bp, 1, txg, old_bp, 2603 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID); 2604 } 2605 2606 if (error) { 2607 error = metaslab_alloc(spa, spa_normal_class(spa), size, 2608 new_bp, 1, txg, old_bp, 2609 METASLAB_HINTBP_AVOID); 2610 } 2611 2612 if (error == 0) { 2613 BP_SET_LSIZE(new_bp, size); 2614 BP_SET_PSIZE(new_bp, size); 2615 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF); 2616 BP_SET_CHECKSUM(new_bp, 2617 spa_version(spa) >= SPA_VERSION_SLIM_ZIL 2618 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG); 2619 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); 2620 BP_SET_LEVEL(new_bp, 0); 2621 BP_SET_DEDUP(new_bp, 0); 2622 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER); 2623 } 2624 2625 return (error); 2626} 2627 2628/* 2629 * Free an intent log block. 2630 */ 2631void 2632zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp) 2633{ 2634 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG); 2635 ASSERT(!BP_IS_GANG(bp)); 2636 2637 zio_free(spa, txg, bp); 2638} 2639 2640/* 2641 * ========================================================================== 2642 * Read, write and delete to physical devices 2643 * ========================================================================== 2644 */ 2645static int 2646zio_vdev_io_start(zio_t *zio) 2647{ 2648 vdev_t *vd = zio->io_vd; 2649 uint64_t align; 2650 spa_t *spa = zio->io_spa; 2651 int ret; 2652 2653 ASSERT(zio->io_error == 0); 2654 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0); 2655 2656 if (vd == NULL) { 2657 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) 2658 spa_config_enter(spa, SCL_ZIO, zio, RW_READER); 2659 2660 /* 2661 * The mirror_ops handle multiple DVAs in a single BP. 2662 */ 2663 return (vdev_mirror_ops.vdev_op_io_start(zio)); 2664 } 2665 2666 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE && 2667 zio->io_priority == ZIO_PRIORITY_NOW) { 2668 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg); 2669 return (ZIO_PIPELINE_CONTINUE); 2670 } 2671 2672 /* 2673 * We keep track of time-sensitive I/Os so that the scan thread 2674 * can quickly react to certain workloads. In particular, we care 2675 * about non-scrubbing, top-level reads and writes with the following 2676 * characteristics: 2677 * - synchronous writes of user data to non-slog devices 2678 * - any reads of user data 2679 * When these conditions are met, adjust the timestamp of spa_last_io 2680 * which allows the scan thread to adjust its workload accordingly. 2681 */ 2682 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL && 2683 vd == vd->vdev_top && !vd->vdev_islog && 2684 zio->io_bookmark.zb_objset != DMU_META_OBJSET && 2685 zio->io_txg != spa_syncing_txg(spa)) { 2686 uint64_t old = spa->spa_last_io; 2687 uint64_t new = ddi_get_lbolt64(); 2688 if (old != new) 2689 (void) atomic_cas_64(&spa->spa_last_io, old, new); 2690 } 2691 2692 align = 1ULL << vd->vdev_top->vdev_ashift; 2693 2694 if ((!(zio->io_flags & ZIO_FLAG_PHYSICAL) || 2695 (vd->vdev_top->vdev_physical_ashift > SPA_MINBLOCKSHIFT)) && 2696 P2PHASE(zio->io_size, align) != 0) { 2697 /* Transform logical writes to be a full physical block size. */ 2698 uint64_t asize = P2ROUNDUP(zio->io_size, align); 2699 char *abuf = NULL; 2700 if (zio->io_type == ZIO_TYPE_READ || 2701 zio->io_type == ZIO_TYPE_WRITE) 2702 abuf = zio_buf_alloc(asize); 2703 ASSERT(vd == vd->vdev_top); 2704 if (zio->io_type == ZIO_TYPE_WRITE) { 2705 bcopy(zio->io_data, abuf, zio->io_size); 2706 bzero(abuf + zio->io_size, asize - zio->io_size); 2707 } 2708 zio_push_transform(zio, abuf, asize, abuf ? asize : 0, 2709 zio_subblock); 2710 } 2711 2712 /* 2713 * If this is not a physical io, make sure that it is properly aligned 2714 * before proceeding. 2715 */ 2716 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) { 2717 ASSERT0(P2PHASE(zio->io_offset, align)); 2718 ASSERT0(P2PHASE(zio->io_size, align)); 2719 } else { 2720 /* 2721 * For physical writes, we allow 512b aligned writes and assume 2722 * the device will perform a read-modify-write as necessary. 2723 */ 2724 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE)); 2725 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE)); 2726 } 2727 2728 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa)); 2729 2730 /* 2731 * If this is a repair I/O, and there's no self-healing involved -- 2732 * that is, we're just resilvering what we expect to resilver -- 2733 * then don't do the I/O unless zio's txg is actually in vd's DTL. 2734 * This prevents spurious resilvering with nested replication. 2735 * For example, given a mirror of mirrors, (A+B)+(C+D), if only 2736 * A is out of date, we'll read from C+D, then use the data to 2737 * resilver A+B -- but we don't actually want to resilver B, just A. 2738 * The top-level mirror has no way to know this, so instead we just 2739 * discard unnecessary repairs as we work our way down the vdev tree. 2740 * The same logic applies to any form of nested replication: 2741 * ditto + mirror, RAID-Z + replacing, etc. This covers them all. 2742 */ 2743 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) && 2744 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) && 2745 zio->io_txg != 0 && /* not a delegated i/o */ 2746 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) { 2747 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 2748 zio_vdev_io_bypass(zio); 2749 return (ZIO_PIPELINE_CONTINUE); 2750 } 2751 2752 if (vd->vdev_ops->vdev_op_leaf) { 2753 switch (zio->io_type) { 2754 case ZIO_TYPE_READ: 2755 if (vdev_cache_read(zio)) 2756 return (ZIO_PIPELINE_CONTINUE); 2757 /* FALLTHROUGH */ 2758 case ZIO_TYPE_WRITE: 2759 case ZIO_TYPE_FREE: 2760 if ((zio = vdev_queue_io(zio)) == NULL) 2761 return (ZIO_PIPELINE_STOP); 2762 2763 if (!vdev_accessible(vd, zio)) { 2764 zio->io_error = SET_ERROR(ENXIO); 2765 zio_interrupt(zio); 2766 return (ZIO_PIPELINE_STOP); 2767 } 2768 break; 2769 } 2770 /* 2771 * Note that we ignore repair writes for TRIM because they can 2772 * conflict with normal writes. This isn't an issue because, by 2773 * definition, we only repair blocks that aren't freed. 2774 */ 2775 if (zio->io_type == ZIO_TYPE_WRITE && 2776 !(zio->io_flags & ZIO_FLAG_IO_REPAIR) && 2777 !trim_map_write_start(zio)) 2778 return (ZIO_PIPELINE_STOP); 2779 } 2780 2781 ret = vd->vdev_ops->vdev_op_io_start(zio); 2782 ASSERT(ret == ZIO_PIPELINE_STOP); 2783 2784 return (ret); 2785} 2786 2787static int 2788zio_vdev_io_done(zio_t *zio) 2789{ 2790 vdev_t *vd = zio->io_vd; 2791 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops; 2792 boolean_t unexpected_error = B_FALSE; 2793 2794 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE)) 2795 return (ZIO_PIPELINE_STOP); 2796 2797 ASSERT(zio->io_type == ZIO_TYPE_READ || 2798 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE); 2799 2800 if (vd != NULL && vd->vdev_ops->vdev_op_leaf && 2801 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE || 2802 zio->io_type == ZIO_TYPE_FREE)) { 2803 2804 if (zio->io_type == ZIO_TYPE_WRITE && 2805 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) 2806 trim_map_write_done(zio); 2807 2808 vdev_queue_io_done(zio); 2809 2810 if (zio->io_type == ZIO_TYPE_WRITE) 2811 vdev_cache_write(zio); 2812 2813 if (zio_injection_enabled && zio->io_error == 0) 2814 zio->io_error = zio_handle_device_injection(vd, 2815 zio, EIO); 2816 2817 if (zio_injection_enabled && zio->io_error == 0) 2818 zio->io_error = zio_handle_label_injection(zio, EIO); 2819 2820 if (zio->io_error) { 2821 if (zio->io_error == ENOTSUP && 2822 zio->io_type == ZIO_TYPE_FREE) { 2823 /* Not all devices support TRIM. */ 2824 } else if (!vdev_accessible(vd, zio)) { 2825 zio->io_error = SET_ERROR(ENXIO); 2826 } else { 2827 unexpected_error = B_TRUE; 2828 } 2829 } 2830 } 2831 2832 ops->vdev_op_io_done(zio); 2833 2834 if (unexpected_error) 2835 VERIFY(vdev_probe(vd, zio) == NULL); 2836 2837 return (ZIO_PIPELINE_CONTINUE); 2838} 2839 2840/* 2841 * For non-raidz ZIOs, we can just copy aside the bad data read from the 2842 * disk, and use that to finish the checksum ereport later. 2843 */ 2844static void 2845zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr, 2846 const void *good_buf) 2847{ 2848 /* no processing needed */ 2849 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE); 2850} 2851 2852/*ARGSUSED*/ 2853void 2854zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored) 2855{ 2856 void *buf = zio_buf_alloc(zio->io_size); 2857 2858 bcopy(zio->io_data, buf, zio->io_size); 2859 2860 zcr->zcr_cbinfo = zio->io_size; 2861 zcr->zcr_cbdata = buf; 2862 zcr->zcr_finish = zio_vsd_default_cksum_finish; 2863 zcr->zcr_free = zio_buf_free; 2864} 2865 2866static int 2867zio_vdev_io_assess(zio_t *zio) 2868{ 2869 vdev_t *vd = zio->io_vd; 2870 2871 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE)) 2872 return (ZIO_PIPELINE_STOP); 2873 2874 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) 2875 spa_config_exit(zio->io_spa, SCL_ZIO, zio); 2876 2877 if (zio->io_vsd != NULL) { 2878 zio->io_vsd_ops->vsd_free(zio); 2879 zio->io_vsd = NULL; 2880 } 2881 2882 if (zio_injection_enabled && zio->io_error == 0) 2883 zio->io_error = zio_handle_fault_injection(zio, EIO); 2884 2885 if (zio->io_type == ZIO_TYPE_FREE && 2886 zio->io_priority != ZIO_PRIORITY_NOW) { 2887 switch (zio->io_error) { 2888 case 0: 2889 ZIO_TRIM_STAT_INCR(bytes, zio->io_size); 2890 ZIO_TRIM_STAT_BUMP(success); 2891 break; 2892 case EOPNOTSUPP: 2893 ZIO_TRIM_STAT_BUMP(unsupported); 2894 break; 2895 default: 2896 ZIO_TRIM_STAT_BUMP(failed); 2897 break; 2898 } 2899 } 2900 2901 /* 2902 * If the I/O failed, determine whether we should attempt to retry it. 2903 * 2904 * On retry, we cut in line in the issue queue, since we don't want 2905 * compression/checksumming/etc. work to prevent our (cheap) IO reissue. 2906 */ 2907 if (zio->io_error && vd == NULL && 2908 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) { 2909 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */ 2910 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */ 2911 zio->io_error = 0; 2912 zio->io_flags |= ZIO_FLAG_IO_RETRY | 2913 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE; 2914 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1; 2915 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, 2916 zio_requeue_io_start_cut_in_line); 2917 return (ZIO_PIPELINE_STOP); 2918 } 2919 2920 /* 2921 * If we got an error on a leaf device, convert it to ENXIO 2922 * if the device is not accessible at all. 2923 */ 2924 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf && 2925 !vdev_accessible(vd, zio)) 2926 zio->io_error = SET_ERROR(ENXIO); 2927 2928 /* 2929 * If we can't write to an interior vdev (mirror or RAID-Z), 2930 * set vdev_cant_write so that we stop trying to allocate from it. 2931 */ 2932 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE && 2933 vd != NULL && !vd->vdev_ops->vdev_op_leaf) { 2934 vd->vdev_cant_write = B_TRUE; 2935 } 2936 2937 if (zio->io_error) 2938 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2939 2940 if (vd != NULL && vd->vdev_ops->vdev_op_leaf && 2941 zio->io_physdone != NULL) { 2942 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED)); 2943 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV); 2944 zio->io_physdone(zio->io_logical); 2945 } 2946 2947 return (ZIO_PIPELINE_CONTINUE); 2948} 2949 2950void 2951zio_vdev_io_reissue(zio_t *zio) 2952{ 2953 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); 2954 ASSERT(zio->io_error == 0); 2955 2956 zio->io_stage >>= 1; 2957} 2958 2959void 2960zio_vdev_io_redone(zio_t *zio) 2961{ 2962 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE); 2963 2964 zio->io_stage >>= 1; 2965} 2966 2967void 2968zio_vdev_io_bypass(zio_t *zio) 2969{ 2970 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); 2971 ASSERT(zio->io_error == 0); 2972 2973 zio->io_flags |= ZIO_FLAG_IO_BYPASS; 2974 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1; 2975} 2976 2977/* 2978 * ========================================================================== 2979 * Generate and verify checksums 2980 * ========================================================================== 2981 */ 2982static int 2983zio_checksum_generate(zio_t *zio) 2984{ 2985 blkptr_t *bp = zio->io_bp; 2986 enum zio_checksum checksum; 2987 2988 if (bp == NULL) { 2989 /* 2990 * This is zio_write_phys(). 2991 * We're either generating a label checksum, or none at all. 2992 */ 2993 checksum = zio->io_prop.zp_checksum; 2994 2995 if (checksum == ZIO_CHECKSUM_OFF) 2996 return (ZIO_PIPELINE_CONTINUE); 2997 2998 ASSERT(checksum == ZIO_CHECKSUM_LABEL); 2999 } else { 3000 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) { 3001 ASSERT(!IO_IS_ALLOCATING(zio)); 3002 checksum = ZIO_CHECKSUM_GANG_HEADER; 3003 } else { 3004 checksum = BP_GET_CHECKSUM(bp); 3005 } 3006 } 3007 3008 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size); 3009 3010 return (ZIO_PIPELINE_CONTINUE); 3011} 3012 3013static int 3014zio_checksum_verify(zio_t *zio) 3015{ 3016 zio_bad_cksum_t info; 3017 blkptr_t *bp = zio->io_bp; 3018 int error; 3019 3020 ASSERT(zio->io_vd != NULL); 3021 3022 if (bp == NULL) { 3023 /* 3024 * This is zio_read_phys(). 3025 * We're either verifying a label checksum, or nothing at all. 3026 */ 3027 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF) 3028 return (ZIO_PIPELINE_CONTINUE); 3029 3030 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL); 3031 } 3032 3033 if ((error = zio_checksum_error(zio, &info)) != 0) { 3034 zio->io_error = error; 3035 if (error == ECKSUM && 3036 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) { 3037 zfs_ereport_start_checksum(zio->io_spa, 3038 zio->io_vd, zio, zio->io_offset, 3039 zio->io_size, NULL, &info); 3040 } 3041 } 3042 3043 return (ZIO_PIPELINE_CONTINUE); 3044} 3045 3046/* 3047 * Called by RAID-Z to ensure we don't compute the checksum twice. 3048 */ 3049void 3050zio_checksum_verified(zio_t *zio) 3051{ 3052 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; 3053} 3054 3055/* 3056 * ========================================================================== 3057 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other. 3058 * An error of 0 indicates success. ENXIO indicates whole-device failure, 3059 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO 3060 * indicate errors that are specific to one I/O, and most likely permanent. 3061 * Any other error is presumed to be worse because we weren't expecting it. 3062 * ========================================================================== 3063 */ 3064int 3065zio_worst_error(int e1, int e2) 3066{ 3067 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO }; 3068 int r1, r2; 3069 3070 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++) 3071 if (e1 == zio_error_rank[r1]) 3072 break; 3073 3074 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++) 3075 if (e2 == zio_error_rank[r2]) 3076 break; 3077 3078 return (r1 > r2 ? e1 : e2); 3079} 3080 3081/* 3082 * ========================================================================== 3083 * I/O completion 3084 * ========================================================================== 3085 */ 3086static int 3087zio_ready(zio_t *zio) 3088{ 3089 blkptr_t *bp = zio->io_bp; 3090 zio_t *pio, *pio_next; 3091 3092 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) || 3093 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY)) 3094 return (ZIO_PIPELINE_STOP); 3095 3096 if (zio->io_ready) { 3097 ASSERT(IO_IS_ALLOCATING(zio)); 3098 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) || 3099 (zio->io_flags & ZIO_FLAG_NOPWRITE)); 3100 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0); 3101 3102 zio->io_ready(zio); 3103 } 3104 3105 if (bp != NULL && bp != &zio->io_bp_copy) 3106 zio->io_bp_copy = *bp; 3107 3108 if (zio->io_error) 3109 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 3110 3111 mutex_enter(&zio->io_lock); 3112 zio->io_state[ZIO_WAIT_READY] = 1; 3113 pio = zio_walk_parents(zio); 3114 mutex_exit(&zio->io_lock); 3115 3116 /* 3117 * As we notify zio's parents, new parents could be added. 3118 * New parents go to the head of zio's io_parent_list, however, 3119 * so we will (correctly) not notify them. The remainder of zio's 3120 * io_parent_list, from 'pio_next' onward, cannot change because 3121 * all parents must wait for us to be done before they can be done. 3122 */ 3123 for (; pio != NULL; pio = pio_next) { 3124 pio_next = zio_walk_parents(zio); 3125 zio_notify_parent(pio, zio, ZIO_WAIT_READY); 3126 } 3127 3128 if (zio->io_flags & ZIO_FLAG_NODATA) { 3129 if (BP_IS_GANG(bp)) { 3130 zio->io_flags &= ~ZIO_FLAG_NODATA; 3131 } else { 3132 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE); 3133 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; 3134 } 3135 } 3136 3137 if (zio_injection_enabled && 3138 zio->io_spa->spa_syncing_txg == zio->io_txg) 3139 zio_handle_ignored_writes(zio); 3140 3141 return (ZIO_PIPELINE_CONTINUE); 3142} 3143 3144static int 3145zio_done(zio_t *zio) 3146{ 3147 spa_t *spa = zio->io_spa; 3148 zio_t *lio = zio->io_logical; 3149 blkptr_t *bp = zio->io_bp; 3150 vdev_t *vd = zio->io_vd; 3151 uint64_t psize = zio->io_size; 3152 zio_t *pio, *pio_next; 3153 3154 /* 3155 * If our children haven't all completed, 3156 * wait for them and then repeat this pipeline stage. 3157 */ 3158 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) || 3159 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) || 3160 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) || 3161 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE)) 3162 return (ZIO_PIPELINE_STOP); 3163 3164 for (int c = 0; c < ZIO_CHILD_TYPES; c++) 3165 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 3166 ASSERT(zio->io_children[c][w] == 0); 3167 3168 if (bp != NULL && !BP_IS_EMBEDDED(bp)) { 3169 ASSERT(bp->blk_pad[0] == 0); 3170 ASSERT(bp->blk_pad[1] == 0); 3171 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 || 3172 (bp == zio_unique_parent(zio)->io_bp)); 3173 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) && 3174 zio->io_bp_override == NULL && 3175 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) { 3176 ASSERT(!BP_SHOULD_BYTESWAP(bp)); 3177 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp)); 3178 ASSERT(BP_COUNT_GANG(bp) == 0 || 3179 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp))); 3180 } 3181 if (zio->io_flags & ZIO_FLAG_NOPWRITE) 3182 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig)); 3183 } 3184 3185 /* 3186 * If there were child vdev/gang/ddt errors, they apply to us now. 3187 */ 3188 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV); 3189 zio_inherit_child_errors(zio, ZIO_CHILD_GANG); 3190 zio_inherit_child_errors(zio, ZIO_CHILD_DDT); 3191 3192 /* 3193 * If the I/O on the transformed data was successful, generate any 3194 * checksum reports now while we still have the transformed data. 3195 */ 3196 if (zio->io_error == 0) { 3197 while (zio->io_cksum_report != NULL) { 3198 zio_cksum_report_t *zcr = zio->io_cksum_report; 3199 uint64_t align = zcr->zcr_align; 3200 uint64_t asize = P2ROUNDUP(psize, align); 3201 char *abuf = zio->io_data; 3202 3203 if (asize != psize) { 3204 abuf = zio_buf_alloc(asize); 3205 bcopy(zio->io_data, abuf, psize); 3206 bzero(abuf + psize, asize - psize); 3207 } 3208 3209 zio->io_cksum_report = zcr->zcr_next; 3210 zcr->zcr_next = NULL; 3211 zcr->zcr_finish(zcr, abuf); 3212 zfs_ereport_free_checksum(zcr); 3213 3214 if (asize != psize) 3215 zio_buf_free(abuf, asize); 3216 } 3217 } 3218 3219 zio_pop_transforms(zio); /* note: may set zio->io_error */ 3220 3221 vdev_stat_update(zio, psize); 3222 3223 if (zio->io_error) { 3224 /* 3225 * If this I/O is attached to a particular vdev, 3226 * generate an error message describing the I/O failure 3227 * at the block level. We ignore these errors if the 3228 * device is currently unavailable. 3229 */ 3230 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd)) 3231 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0); 3232 3233 if ((zio->io_error == EIO || !(zio->io_flags & 3234 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) && 3235 zio == lio) { 3236 /* 3237 * For logical I/O requests, tell the SPA to log the 3238 * error and generate a logical data ereport. 3239 */ 3240 spa_log_error(spa, zio); 3241 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio, 3242 0, 0); 3243 } 3244 } 3245 3246 if (zio->io_error && zio == lio) { 3247 /* 3248 * Determine whether zio should be reexecuted. This will 3249 * propagate all the way to the root via zio_notify_parent(). 3250 */ 3251 ASSERT(vd == NULL && bp != NULL); 3252 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 3253 3254 if (IO_IS_ALLOCATING(zio) && 3255 !(zio->io_flags & ZIO_FLAG_CANFAIL)) { 3256 if (zio->io_error != ENOSPC) 3257 zio->io_reexecute |= ZIO_REEXECUTE_NOW; 3258 else 3259 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 3260 } 3261 3262 if ((zio->io_type == ZIO_TYPE_READ || 3263 zio->io_type == ZIO_TYPE_FREE) && 3264 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && 3265 zio->io_error == ENXIO && 3266 spa_load_state(spa) == SPA_LOAD_NONE && 3267 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE) 3268 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 3269 3270 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute) 3271 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 3272 3273 /* 3274 * Here is a possibly good place to attempt to do 3275 * either combinatorial reconstruction or error correction 3276 * based on checksums. It also might be a good place 3277 * to send out preliminary ereports before we suspend 3278 * processing. 3279 */ 3280 } 3281 3282 /* 3283 * If there were logical child errors, they apply to us now. 3284 * We defer this until now to avoid conflating logical child 3285 * errors with errors that happened to the zio itself when 3286 * updating vdev stats and reporting FMA events above. 3287 */ 3288 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL); 3289 3290 if ((zio->io_error || zio->io_reexecute) && 3291 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio && 3292 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE))) 3293 zio_dva_unallocate(zio, zio->io_gang_tree, bp); 3294 3295 zio_gang_tree_free(&zio->io_gang_tree); 3296 3297 /* 3298 * Godfather I/Os should never suspend. 3299 */ 3300 if ((zio->io_flags & ZIO_FLAG_GODFATHER) && 3301 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) 3302 zio->io_reexecute = 0; 3303 3304 if (zio->io_reexecute) { 3305 /* 3306 * This is a logical I/O that wants to reexecute. 3307 * 3308 * Reexecute is top-down. When an i/o fails, if it's not 3309 * the root, it simply notifies its parent and sticks around. 3310 * The parent, seeing that it still has children in zio_done(), 3311 * does the same. This percolates all the way up to the root. 3312 * The root i/o will reexecute or suspend the entire tree. 3313 * 3314 * This approach ensures that zio_reexecute() honors 3315 * all the original i/o dependency relationships, e.g. 3316 * parents not executing until children are ready. 3317 */ 3318 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 3319 3320 zio->io_gang_leader = NULL; 3321 3322 mutex_enter(&zio->io_lock); 3323 zio->io_state[ZIO_WAIT_DONE] = 1; 3324 mutex_exit(&zio->io_lock); 3325 3326 /* 3327 * "The Godfather" I/O monitors its children but is 3328 * not a true parent to them. It will track them through 3329 * the pipeline but severs its ties whenever they get into 3330 * trouble (e.g. suspended). This allows "The Godfather" 3331 * I/O to return status without blocking. 3332 */ 3333 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) { 3334 zio_link_t *zl = zio->io_walk_link; 3335 pio_next = zio_walk_parents(zio); 3336 3337 if ((pio->io_flags & ZIO_FLAG_GODFATHER) && 3338 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) { 3339 zio_remove_child(pio, zio, zl); 3340 zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 3341 } 3342 } 3343 3344 if ((pio = zio_unique_parent(zio)) != NULL) { 3345 /* 3346 * We're not a root i/o, so there's nothing to do 3347 * but notify our parent. Don't propagate errors 3348 * upward since we haven't permanently failed yet. 3349 */ 3350 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); 3351 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE; 3352 zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 3353 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) { 3354 /* 3355 * We'd fail again if we reexecuted now, so suspend 3356 * until conditions improve (e.g. device comes online). 3357 */ 3358 zio_suspend(spa, zio); 3359 } else { 3360 /* 3361 * Reexecution is potentially a huge amount of work. 3362 * Hand it off to the otherwise-unused claim taskq. 3363 */ 3364#if defined(illumos) || !defined(_KERNEL) 3365 ASSERT(zio->io_tqent.tqent_next == NULL); 3366#else 3367 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0); 3368#endif 3369 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM, 3370 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio, 3371 0, &zio->io_tqent); 3372 } 3373 return (ZIO_PIPELINE_STOP); 3374 } 3375 3376 ASSERT(zio->io_child_count == 0); 3377 ASSERT(zio->io_reexecute == 0); 3378 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL)); 3379 3380 /* 3381 * Report any checksum errors, since the I/O is complete. 3382 */ 3383 while (zio->io_cksum_report != NULL) { 3384 zio_cksum_report_t *zcr = zio->io_cksum_report; 3385 zio->io_cksum_report = zcr->zcr_next; 3386 zcr->zcr_next = NULL; 3387 zcr->zcr_finish(zcr, NULL); 3388 zfs_ereport_free_checksum(zcr); 3389 } 3390 3391 /* 3392 * It is the responsibility of the done callback to ensure that this 3393 * particular zio is no longer discoverable for adoption, and as 3394 * such, cannot acquire any new parents. 3395 */ 3396 if (zio->io_done) 3397 zio->io_done(zio); 3398 3399 mutex_enter(&zio->io_lock); 3400 zio->io_state[ZIO_WAIT_DONE] = 1; 3401 mutex_exit(&zio->io_lock); 3402 3403 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) { 3404 zio_link_t *zl = zio->io_walk_link; 3405 pio_next = zio_walk_parents(zio); 3406 zio_remove_child(pio, zio, zl); 3407 zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 3408 } 3409 3410 if (zio->io_waiter != NULL) { 3411 mutex_enter(&zio->io_lock); 3412 zio->io_executor = NULL; 3413 cv_broadcast(&zio->io_cv); 3414 mutex_exit(&zio->io_lock); 3415 } else { 3416 zio_destroy(zio); 3417 } 3418 3419 return (ZIO_PIPELINE_STOP); 3420} 3421 3422/* 3423 * ========================================================================== 3424 * I/O pipeline definition 3425 * ========================================================================== 3426 */ 3427static zio_pipe_stage_t *zio_pipeline[] = { 3428 NULL, 3429 zio_read_bp_init, 3430 zio_free_bp_init, 3431 zio_issue_async, 3432 zio_write_bp_init, 3433 zio_checksum_generate, 3434 zio_nop_write, 3435 zio_ddt_read_start, 3436 zio_ddt_read_done, 3437 zio_ddt_write, 3438 zio_ddt_free, 3439 zio_gang_assemble, 3440 zio_gang_issue, 3441 zio_dva_allocate, 3442 zio_dva_free, 3443 zio_dva_claim, 3444 zio_ready, 3445 zio_vdev_io_start, 3446 zio_vdev_io_done, 3447 zio_vdev_io_assess, 3448 zio_checksum_verify, 3449 zio_done 3450}; 3451 3452/* dnp is the dnode for zb1->zb_object */ 3453boolean_t 3454zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_phys_t *zb1, 3455 const zbookmark_phys_t *zb2) 3456{ 3457 uint64_t zb1nextL0, zb2thisobj; 3458 3459 ASSERT(zb1->zb_objset == zb2->zb_objset); 3460 ASSERT(zb2->zb_level == 0); 3461 3462 /* The objset_phys_t isn't before anything. */ 3463 if (dnp == NULL) 3464 return (B_FALSE); 3465 3466 zb1nextL0 = (zb1->zb_blkid + 1) << 3467 ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT)); 3468 3469 zb2thisobj = zb2->zb_object ? zb2->zb_object : 3470 zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT); 3471 3472 if (zb1->zb_object == DMU_META_DNODE_OBJECT) { 3473 uint64_t nextobj = zb1nextL0 * 3474 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT; 3475 return (nextobj <= zb2thisobj); 3476 } 3477 3478 if (zb1->zb_object < zb2thisobj) 3479 return (B_TRUE); 3480 if (zb1->zb_object > zb2thisobj) 3481 return (B_FALSE); 3482 if (zb2->zb_object == DMU_META_DNODE_OBJECT) 3483 return (B_FALSE); 3484 return (zb1nextL0 <= zb2->zb_blkid); 3485} 3486