zio.c revision 297078
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, 2015 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#define ZIO_PIPELINE_CONTINUE 0x100 97#define ZIO_PIPELINE_STOP 0x101 98 99#define BP_SPANB(indblkshift, level) \ 100 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT))) 101#define COMPARE_META_LEVEL 0x80000000ul 102/* 103 * The following actions directly effect the spa's sync-to-convergence logic. 104 * The values below define the sync pass when we start performing the action. 105 * Care should be taken when changing these values as they directly impact 106 * spa_sync() performance. Tuning these values may introduce subtle performance 107 * pathologies and should only be done in the context of performance analysis. 108 * These tunables will eventually be removed and replaced with #defines once 109 * enough analysis has been done to determine optimal values. 110 * 111 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that 112 * regular blocks are not deferred. 113 */ 114int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */ 115TUNABLE_INT("vfs.zfs.sync_pass_deferred_free", &zfs_sync_pass_deferred_free); 116SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_deferred_free, CTLFLAG_RDTUN, 117 &zfs_sync_pass_deferred_free, 0, "defer frees starting in this pass"); 118int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */ 119TUNABLE_INT("vfs.zfs.sync_pass_dont_compress", &zfs_sync_pass_dont_compress); 120SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_dont_compress, CTLFLAG_RDTUN, 121 &zfs_sync_pass_dont_compress, 0, "don't compress starting in this pass"); 122int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */ 123TUNABLE_INT("vfs.zfs.sync_pass_rewrite", &zfs_sync_pass_rewrite); 124SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_rewrite, CTLFLAG_RDTUN, 125 &zfs_sync_pass_rewrite, 0, "rewrite new bps starting in this pass"); 126 127/* 128 * An allocating zio is one that either currently has the DVA allocate 129 * stage set or will have it later in its lifetime. 130 */ 131#define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE) 132 133boolean_t zio_requeue_io_start_cut_in_line = B_TRUE; 134 135#ifdef ZFS_DEBUG 136int zio_buf_debug_limit = 16384; 137#else 138int zio_buf_debug_limit = 0; 139#endif 140 141void 142zio_init(void) 143{ 144 size_t c; 145 zio_cache = kmem_cache_create("zio_cache", 146 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0); 147 zio_link_cache = kmem_cache_create("zio_link_cache", 148 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0); 149 if (!zio_use_uma) 150 goto out; 151 152 /* 153 * For small buffers, we want a cache for each multiple of 154 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache 155 * for each quarter-power of 2. 156 */ 157 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { 158 size_t size = (c + 1) << SPA_MINBLOCKSHIFT; 159 size_t p2 = size; 160 size_t align = 0; 161 size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0; 162 163 while (!ISP2(p2)) 164 p2 &= p2 - 1; 165 166#ifdef illumos 167#ifndef _KERNEL 168 /* 169 * If we are using watchpoints, put each buffer on its own page, 170 * to eliminate the performance overhead of trapping to the 171 * kernel when modifying a non-watched buffer that shares the 172 * page with a watched buffer. 173 */ 174 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE)) 175 continue; 176#endif 177#endif /* illumos */ 178 if (size <= 4 * SPA_MINBLOCKSIZE) { 179 align = SPA_MINBLOCKSIZE; 180 } else if (IS_P2ALIGNED(size, p2 >> 2)) { 181 align = MIN(p2 >> 2, PAGESIZE); 182 } 183 184 if (align != 0) { 185 char name[36]; 186 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size); 187 zio_buf_cache[c] = kmem_cache_create(name, size, 188 align, NULL, NULL, NULL, NULL, NULL, cflags); 189 190 /* 191 * Since zio_data bufs do not appear in crash dumps, we 192 * pass KMC_NOTOUCH so that no allocator metadata is 193 * stored with the buffers. 194 */ 195 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size); 196 zio_data_buf_cache[c] = kmem_cache_create(name, size, 197 align, NULL, NULL, NULL, NULL, NULL, 198 cflags | KMC_NOTOUCH | KMC_NODEBUG); 199 } 200 } 201 202 while (--c != 0) { 203 ASSERT(zio_buf_cache[c] != NULL); 204 if (zio_buf_cache[c - 1] == NULL) 205 zio_buf_cache[c - 1] = zio_buf_cache[c]; 206 207 ASSERT(zio_data_buf_cache[c] != NULL); 208 if (zio_data_buf_cache[c - 1] == NULL) 209 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c]; 210 } 211out: 212 213 zio_inject_init(); 214 215 zio_trim_ksp = kstat_create("zfs", 0, "zio_trim", "misc", 216 KSTAT_TYPE_NAMED, 217 sizeof(zio_trim_stats) / sizeof(kstat_named_t), 218 KSTAT_FLAG_VIRTUAL); 219 220 if (zio_trim_ksp != NULL) { 221 zio_trim_ksp->ks_data = &zio_trim_stats; 222 kstat_install(zio_trim_ksp); 223 } 224} 225 226void 227zio_fini(void) 228{ 229 size_t c; 230 kmem_cache_t *last_cache = NULL; 231 kmem_cache_t *last_data_cache = NULL; 232 233 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { 234 if (zio_buf_cache[c] != last_cache) { 235 last_cache = zio_buf_cache[c]; 236 kmem_cache_destroy(zio_buf_cache[c]); 237 } 238 zio_buf_cache[c] = NULL; 239 240 if (zio_data_buf_cache[c] != last_data_cache) { 241 last_data_cache = zio_data_buf_cache[c]; 242 kmem_cache_destroy(zio_data_buf_cache[c]); 243 } 244 zio_data_buf_cache[c] = NULL; 245 } 246 247 kmem_cache_destroy(zio_link_cache); 248 kmem_cache_destroy(zio_cache); 249 250 zio_inject_fini(); 251 252 if (zio_trim_ksp != NULL) { 253 kstat_delete(zio_trim_ksp); 254 zio_trim_ksp = NULL; 255 } 256} 257 258/* 259 * ========================================================================== 260 * Allocate and free I/O buffers 261 * ========================================================================== 262 */ 263 264/* 265 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a 266 * crashdump if the kernel panics, so use it judiciously. Obviously, it's 267 * useful to inspect ZFS metadata, but if possible, we should avoid keeping 268 * excess / transient data in-core during a crashdump. 269 */ 270void * 271zio_buf_alloc(size_t size) 272{ 273 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 274 int flags = zio_exclude_metadata ? KM_NODEBUG : 0; 275 276 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 277 278 if (zio_use_uma) 279 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE)); 280 else 281 return (kmem_alloc(size, KM_SLEEP|flags)); 282} 283 284/* 285 * Use zio_data_buf_alloc to allocate data. The data will not appear in a 286 * crashdump if the kernel panics. This exists so that we will limit the amount 287 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount 288 * of kernel heap dumped to disk when the kernel panics) 289 */ 290void * 291zio_data_buf_alloc(size_t size) 292{ 293 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 294 295 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 296 297 if (zio_use_uma) 298 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE)); 299 else 300 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG)); 301} 302 303void 304zio_buf_free(void *buf, size_t size) 305{ 306 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 307 308 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 309 310 if (zio_use_uma) 311 kmem_cache_free(zio_buf_cache[c], buf); 312 else 313 kmem_free(buf, size); 314} 315 316void 317zio_data_buf_free(void *buf, size_t size) 318{ 319 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 320 321 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 322 323 if (zio_use_uma) 324 kmem_cache_free(zio_data_buf_cache[c], buf); 325 else 326 kmem_free(buf, size); 327} 328 329/* 330 * ========================================================================== 331 * Push and pop I/O transform buffers 332 * ========================================================================== 333 */ 334static void 335zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize, 336 zio_transform_func_t *transform) 337{ 338 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP); 339 340 zt->zt_orig_data = zio->io_data; 341 zt->zt_orig_size = zio->io_size; 342 zt->zt_bufsize = bufsize; 343 zt->zt_transform = transform; 344 345 zt->zt_next = zio->io_transform_stack; 346 zio->io_transform_stack = zt; 347 348 zio->io_data = data; 349 zio->io_size = size; 350} 351 352static void 353zio_pop_transforms(zio_t *zio) 354{ 355 zio_transform_t *zt; 356 357 while ((zt = zio->io_transform_stack) != NULL) { 358 if (zt->zt_transform != NULL) 359 zt->zt_transform(zio, 360 zt->zt_orig_data, zt->zt_orig_size); 361 362 if (zt->zt_bufsize != 0) 363 zio_buf_free(zio->io_data, zt->zt_bufsize); 364 365 zio->io_data = zt->zt_orig_data; 366 zio->io_size = zt->zt_orig_size; 367 zio->io_transform_stack = zt->zt_next; 368 369 kmem_free(zt, sizeof (zio_transform_t)); 370 } 371} 372 373/* 374 * ========================================================================== 375 * I/O transform callbacks for subblocks and decompression 376 * ========================================================================== 377 */ 378static void 379zio_subblock(zio_t *zio, void *data, uint64_t size) 380{ 381 ASSERT(zio->io_size > size); 382 383 if (zio->io_type == ZIO_TYPE_READ) 384 bcopy(zio->io_data, data, size); 385} 386 387static void 388zio_decompress(zio_t *zio, void *data, uint64_t size) 389{ 390 if (zio->io_error == 0 && 391 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp), 392 zio->io_data, data, zio->io_size, size) != 0) 393 zio->io_error = SET_ERROR(EIO); 394} 395 396/* 397 * ========================================================================== 398 * I/O parent/child relationships and pipeline interlocks 399 * ========================================================================== 400 */ 401/* 402 * NOTE - Callers to zio_walk_parents() and zio_walk_children must 403 * continue calling these functions until they return NULL. 404 * Otherwise, the next caller will pick up the list walk in 405 * some indeterminate state. (Otherwise every caller would 406 * have to pass in a cookie to keep the state represented by 407 * io_walk_link, which gets annoying.) 408 */ 409zio_t * 410zio_walk_parents(zio_t *cio) 411{ 412 zio_link_t *zl = cio->io_walk_link; 413 list_t *pl = &cio->io_parent_list; 414 415 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl); 416 cio->io_walk_link = zl; 417 418 if (zl == NULL) 419 return (NULL); 420 421 ASSERT(zl->zl_child == cio); 422 return (zl->zl_parent); 423} 424 425zio_t * 426zio_walk_children(zio_t *pio) 427{ 428 zio_link_t *zl = pio->io_walk_link; 429 list_t *cl = &pio->io_child_list; 430 431 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl); 432 pio->io_walk_link = zl; 433 434 if (zl == NULL) 435 return (NULL); 436 437 ASSERT(zl->zl_parent == pio); 438 return (zl->zl_child); 439} 440 441zio_t * 442zio_unique_parent(zio_t *cio) 443{ 444 zio_t *pio = zio_walk_parents(cio); 445 446 VERIFY(zio_walk_parents(cio) == NULL); 447 return (pio); 448} 449 450void 451zio_add_child(zio_t *pio, zio_t *cio) 452{ 453 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP); 454 455 /* 456 * Logical I/Os can have logical, gang, or vdev children. 457 * Gang I/Os can have gang or vdev children. 458 * Vdev I/Os can only have vdev children. 459 * The following ASSERT captures all of these constraints. 460 */ 461 ASSERT(cio->io_child_type <= pio->io_child_type); 462 463 zl->zl_parent = pio; 464 zl->zl_child = cio; 465 466 mutex_enter(&cio->io_lock); 467 mutex_enter(&pio->io_lock); 468 469 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0); 470 471 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 472 pio->io_children[cio->io_child_type][w] += !cio->io_state[w]; 473 474 list_insert_head(&pio->io_child_list, zl); 475 list_insert_head(&cio->io_parent_list, zl); 476 477 pio->io_child_count++; 478 cio->io_parent_count++; 479 480 mutex_exit(&pio->io_lock); 481 mutex_exit(&cio->io_lock); 482} 483 484static void 485zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl) 486{ 487 ASSERT(zl->zl_parent == pio); 488 ASSERT(zl->zl_child == cio); 489 490 mutex_enter(&cio->io_lock); 491 mutex_enter(&pio->io_lock); 492 493 list_remove(&pio->io_child_list, zl); 494 list_remove(&cio->io_parent_list, zl); 495 496 pio->io_child_count--; 497 cio->io_parent_count--; 498 499 mutex_exit(&pio->io_lock); 500 mutex_exit(&cio->io_lock); 501 502 kmem_cache_free(zio_link_cache, zl); 503} 504 505static boolean_t 506zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait) 507{ 508 uint64_t *countp = &zio->io_children[child][wait]; 509 boolean_t waiting = B_FALSE; 510 511 mutex_enter(&zio->io_lock); 512 ASSERT(zio->io_stall == NULL); 513 if (*countp != 0) { 514 zio->io_stage >>= 1; 515 zio->io_stall = countp; 516 waiting = B_TRUE; 517 } 518 mutex_exit(&zio->io_lock); 519 520 return (waiting); 521} 522 523static void 524zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait) 525{ 526 uint64_t *countp = &pio->io_children[zio->io_child_type][wait]; 527 int *errorp = &pio->io_child_error[zio->io_child_type]; 528 529 mutex_enter(&pio->io_lock); 530 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE)) 531 *errorp = zio_worst_error(*errorp, zio->io_error); 532 pio->io_reexecute |= zio->io_reexecute; 533 ASSERT3U(*countp, >, 0); 534 535 (*countp)--; 536 537 if (*countp == 0 && pio->io_stall == countp) { 538 pio->io_stall = NULL; 539 mutex_exit(&pio->io_lock); 540 zio_execute(pio); 541 } else { 542 mutex_exit(&pio->io_lock); 543 } 544} 545 546static void 547zio_inherit_child_errors(zio_t *zio, enum zio_child c) 548{ 549 if (zio->io_child_error[c] != 0 && zio->io_error == 0) 550 zio->io_error = zio->io_child_error[c]; 551} 552 553/* 554 * ========================================================================== 555 * Create the various types of I/O (read, write, free, etc) 556 * ========================================================================== 557 */ 558static zio_t * 559zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 560 void *data, uint64_t size, zio_done_func_t *done, void *private, 561 zio_type_t type, zio_priority_t priority, enum zio_flag flags, 562 vdev_t *vd, uint64_t offset, const zbookmark_phys_t *zb, 563 enum zio_stage stage, enum zio_stage pipeline) 564{ 565 zio_t *zio; 566 567 ASSERT3U(type == ZIO_TYPE_FREE || size, <=, SPA_MAXBLOCKSIZE); 568 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0); 569 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0); 570 571 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER)); 572 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER)); 573 ASSERT(vd || stage == ZIO_STAGE_OPEN); 574 575 zio = kmem_cache_alloc(zio_cache, KM_SLEEP); 576 bzero(zio, sizeof (zio_t)); 577 578 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL); 579 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL); 580 581 list_create(&zio->io_parent_list, sizeof (zio_link_t), 582 offsetof(zio_link_t, zl_parent_node)); 583 list_create(&zio->io_child_list, sizeof (zio_link_t), 584 offsetof(zio_link_t, zl_child_node)); 585 586 if (vd != NULL) 587 zio->io_child_type = ZIO_CHILD_VDEV; 588 else if (flags & ZIO_FLAG_GANG_CHILD) 589 zio->io_child_type = ZIO_CHILD_GANG; 590 else if (flags & ZIO_FLAG_DDT_CHILD) 591 zio->io_child_type = ZIO_CHILD_DDT; 592 else 593 zio->io_child_type = ZIO_CHILD_LOGICAL; 594 595 if (bp != NULL) { 596 zio->io_bp = (blkptr_t *)bp; 597 zio->io_bp_copy = *bp; 598 zio->io_bp_orig = *bp; 599 if (type != ZIO_TYPE_WRITE || 600 zio->io_child_type == ZIO_CHILD_DDT) 601 zio->io_bp = &zio->io_bp_copy; /* so caller can free */ 602 if (zio->io_child_type == ZIO_CHILD_LOGICAL) 603 zio->io_logical = zio; 604 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp)) 605 pipeline |= ZIO_GANG_STAGES; 606 } 607 608 zio->io_spa = spa; 609 zio->io_txg = txg; 610 zio->io_done = done; 611 zio->io_private = private; 612 zio->io_type = type; 613 zio->io_priority = priority; 614 zio->io_vd = vd; 615 zio->io_offset = offset; 616 zio->io_orig_data = zio->io_data = data; 617 zio->io_orig_size = zio->io_size = size; 618 zio->io_orig_flags = zio->io_flags = flags; 619 zio->io_orig_stage = zio->io_stage = stage; 620 zio->io_orig_pipeline = zio->io_pipeline = pipeline; 621 622 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY); 623 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE); 624 625 if (zb != NULL) 626 zio->io_bookmark = *zb; 627 628 if (pio != NULL) { 629 if (zio->io_logical == NULL) 630 zio->io_logical = pio->io_logical; 631 if (zio->io_child_type == ZIO_CHILD_GANG) 632 zio->io_gang_leader = pio->io_gang_leader; 633 zio_add_child(pio, zio); 634 } 635 636 return (zio); 637} 638 639static void 640zio_destroy(zio_t *zio) 641{ 642 list_destroy(&zio->io_parent_list); 643 list_destroy(&zio->io_child_list); 644 mutex_destroy(&zio->io_lock); 645 cv_destroy(&zio->io_cv); 646 kmem_cache_free(zio_cache, zio); 647} 648 649zio_t * 650zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done, 651 void *private, enum zio_flag flags) 652{ 653 zio_t *zio; 654 655 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private, 656 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, 657 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE); 658 659 return (zio); 660} 661 662zio_t * 663zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags) 664{ 665 return (zio_null(NULL, spa, NULL, done, private, flags)); 666} 667 668void 669zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp) 670{ 671 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) { 672 zfs_panic_recover("blkptr at %p has invalid TYPE %llu", 673 bp, (longlong_t)BP_GET_TYPE(bp)); 674 } 675 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS || 676 BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) { 677 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu", 678 bp, (longlong_t)BP_GET_CHECKSUM(bp)); 679 } 680 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS || 681 BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) { 682 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu", 683 bp, (longlong_t)BP_GET_COMPRESS(bp)); 684 } 685 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) { 686 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu", 687 bp, (longlong_t)BP_GET_LSIZE(bp)); 688 } 689 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) { 690 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu", 691 bp, (longlong_t)BP_GET_PSIZE(bp)); 692 } 693 694 if (BP_IS_EMBEDDED(bp)) { 695 if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) { 696 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu", 697 bp, (longlong_t)BPE_GET_ETYPE(bp)); 698 } 699 } 700 701 /* 702 * Pool-specific checks. 703 * 704 * Note: it would be nice to verify that the blk_birth and 705 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze() 706 * allows the birth time of log blocks (and dmu_sync()-ed blocks 707 * that are in the log) to be arbitrarily large. 708 */ 709 for (int i = 0; i < BP_GET_NDVAS(bp); i++) { 710 uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]); 711 if (vdevid >= spa->spa_root_vdev->vdev_children) { 712 zfs_panic_recover("blkptr at %p DVA %u has invalid " 713 "VDEV %llu", 714 bp, i, (longlong_t)vdevid); 715 continue; 716 } 717 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid]; 718 if (vd == NULL) { 719 zfs_panic_recover("blkptr at %p DVA %u has invalid " 720 "VDEV %llu", 721 bp, i, (longlong_t)vdevid); 722 continue; 723 } 724 if (vd->vdev_ops == &vdev_hole_ops) { 725 zfs_panic_recover("blkptr at %p DVA %u has hole " 726 "VDEV %llu", 727 bp, i, (longlong_t)vdevid); 728 continue; 729 } 730 if (vd->vdev_ops == &vdev_missing_ops) { 731 /* 732 * "missing" vdevs are valid during import, but we 733 * don't have their detailed info (e.g. asize), so 734 * we can't perform any more checks on them. 735 */ 736 continue; 737 } 738 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]); 739 uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]); 740 if (BP_IS_GANG(bp)) 741 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE); 742 if (offset + asize > vd->vdev_asize) { 743 zfs_panic_recover("blkptr at %p DVA %u has invalid " 744 "OFFSET %llu", 745 bp, i, (longlong_t)offset); 746 } 747 } 748} 749 750zio_t * 751zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, 752 void *data, uint64_t size, zio_done_func_t *done, void *private, 753 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb) 754{ 755 zio_t *zio; 756 757 zfs_blkptr_verify(spa, bp); 758 759 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp, 760 data, size, done, private, 761 ZIO_TYPE_READ, priority, flags, NULL, 0, zb, 762 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? 763 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE); 764 765 return (zio); 766} 767 768zio_t * 769zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, 770 void *data, uint64_t size, const zio_prop_t *zp, 771 zio_done_func_t *ready, zio_done_func_t *physdone, zio_done_func_t *done, 772 void *private, 773 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb) 774{ 775 zio_t *zio; 776 777 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF && 778 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS && 779 zp->zp_compress >= ZIO_COMPRESS_OFF && 780 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS && 781 DMU_OT_IS_VALID(zp->zp_type) && 782 zp->zp_level < 32 && 783 zp->zp_copies > 0 && 784 zp->zp_copies <= spa_max_replication(spa)); 785 786 zio = zio_create(pio, spa, txg, bp, data, size, done, private, 787 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, 788 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? 789 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE); 790 791 zio->io_ready = ready; 792 zio->io_physdone = physdone; 793 zio->io_prop = *zp; 794 795 /* 796 * Data can be NULL if we are going to call zio_write_override() to 797 * provide the already-allocated BP. But we may need the data to 798 * verify a dedup hit (if requested). In this case, don't try to 799 * dedup (just take the already-allocated BP verbatim). 800 */ 801 if (data == NULL && zio->io_prop.zp_dedup_verify) { 802 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE; 803 } 804 805 return (zio); 806} 807 808zio_t * 809zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data, 810 uint64_t size, zio_done_func_t *done, void *private, 811 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb) 812{ 813 zio_t *zio; 814 815 zio = zio_create(pio, spa, txg, bp, data, size, done, private, 816 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, 817 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE); 818 819 return (zio); 820} 821 822void 823zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite) 824{ 825 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 826 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 827 ASSERT(zio->io_stage == ZIO_STAGE_OPEN); 828 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa)); 829 830 /* 831 * We must reset the io_prop to match the values that existed 832 * when the bp was first written by dmu_sync() keeping in mind 833 * that nopwrite and dedup are mutually exclusive. 834 */ 835 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup; 836 zio->io_prop.zp_nopwrite = nopwrite; 837 zio->io_prop.zp_copies = copies; 838 zio->io_bp_override = bp; 839} 840 841void 842zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp) 843{ 844 845 /* 846 * The check for EMBEDDED is a performance optimization. We 847 * process the free here (by ignoring it) rather than 848 * putting it on the list and then processing it in zio_free_sync(). 849 */ 850 if (BP_IS_EMBEDDED(bp)) 851 return; 852 metaslab_check_free(spa, bp); 853 854 /* 855 * Frees that are for the currently-syncing txg, are not going to be 856 * deferred, and which will not need to do a read (i.e. not GANG or 857 * DEDUP), can be processed immediately. Otherwise, put them on the 858 * in-memory list for later processing. 859 */ 860 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) || 861 txg != spa->spa_syncing_txg || 862 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) { 863 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp); 864 } else { 865 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 866 BP_GET_PSIZE(bp), 0))); 867 } 868} 869 870zio_t * 871zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 872 uint64_t size, enum zio_flag flags) 873{ 874 zio_t *zio; 875 enum zio_stage stage = ZIO_FREE_PIPELINE; 876 877 ASSERT(!BP_IS_HOLE(bp)); 878 ASSERT(spa_syncing_txg(spa) == txg); 879 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free); 880 881 if (BP_IS_EMBEDDED(bp)) 882 return (zio_null(pio, spa, NULL, NULL, NULL, 0)); 883 884 metaslab_check_free(spa, bp); 885 arc_freed(spa, bp); 886 887 if (zfs_trim_enabled) 888 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START | 889 ZIO_STAGE_VDEV_IO_ASSESS; 890 /* 891 * GANG and DEDUP blocks can induce a read (for the gang block header, 892 * or the DDT), so issue them asynchronously so that this thread is 893 * not tied up. 894 */ 895 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp)) 896 stage |= ZIO_STAGE_ISSUE_ASYNC; 897 898 flags |= ZIO_FLAG_DONT_QUEUE; 899 900 zio = zio_create(pio, spa, txg, bp, NULL, size, 901 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags, 902 NULL, 0, NULL, ZIO_STAGE_OPEN, stage); 903 904 return (zio); 905} 906 907zio_t * 908zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 909 zio_done_func_t *done, void *private, enum zio_flag flags) 910{ 911 zio_t *zio; 912 913 dprintf_bp(bp, "claiming in txg %llu", txg); 914 915 if (BP_IS_EMBEDDED(bp)) 916 return (zio_null(pio, spa, NULL, NULL, NULL, 0)); 917 918 /* 919 * A claim is an allocation of a specific block. Claims are needed 920 * to support immediate writes in the intent log. The issue is that 921 * immediate writes contain committed data, but in a txg that was 922 * *not* committed. Upon opening the pool after an unclean shutdown, 923 * the intent log claims all blocks that contain immediate write data 924 * so that the SPA knows they're in use. 925 * 926 * All claims *must* be resolved in the first txg -- before the SPA 927 * starts allocating blocks -- so that nothing is allocated twice. 928 * If txg == 0 we just verify that the block is claimable. 929 */ 930 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa)); 931 ASSERT(txg == spa_first_txg(spa) || txg == 0); 932 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */ 933 934 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), 935 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags, 936 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE); 937 938 return (zio); 939} 940 941zio_t * 942zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset, 943 uint64_t size, zio_done_func_t *done, void *private, 944 zio_priority_t priority, enum zio_flag flags) 945{ 946 zio_t *zio; 947 int c; 948 949 if (vd->vdev_children == 0) { 950 zio = zio_create(pio, spa, 0, NULL, NULL, size, done, private, 951 ZIO_TYPE_IOCTL, priority, flags, vd, offset, NULL, 952 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE); 953 954 zio->io_cmd = cmd; 955 } else { 956 zio = zio_null(pio, spa, NULL, NULL, NULL, flags); 957 958 for (c = 0; c < vd->vdev_children; c++) 959 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd, 960 offset, size, done, private, priority, flags)); 961 } 962 963 return (zio); 964} 965 966zio_t * 967zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 968 void *data, int checksum, zio_done_func_t *done, void *private, 969 zio_priority_t priority, enum zio_flag flags, boolean_t labels) 970{ 971 zio_t *zio; 972 973 ASSERT(vd->vdev_children == 0); 974 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || 975 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); 976 ASSERT3U(offset + size, <=, vd->vdev_psize); 977 978 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private, 979 ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset, 980 NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE); 981 982 zio->io_prop.zp_checksum = checksum; 983 984 return (zio); 985} 986 987zio_t * 988zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 989 void *data, int checksum, zio_done_func_t *done, void *private, 990 zio_priority_t priority, enum zio_flag flags, boolean_t labels) 991{ 992 zio_t *zio; 993 994 ASSERT(vd->vdev_children == 0); 995 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || 996 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); 997 ASSERT3U(offset + size, <=, vd->vdev_psize); 998 999 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private, 1000 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset, 1001 NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE); 1002 1003 zio->io_prop.zp_checksum = checksum; 1004 1005 if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) { 1006 /* 1007 * zec checksums are necessarily destructive -- they modify 1008 * the end of the write buffer to hold the verifier/checksum. 1009 * Therefore, we must make a local copy in case the data is 1010 * being written to multiple places in parallel. 1011 */ 1012 void *wbuf = zio_buf_alloc(size); 1013 bcopy(data, wbuf, size); 1014 zio_push_transform(zio, wbuf, size, size, NULL); 1015 } 1016 1017 return (zio); 1018} 1019 1020/* 1021 * Create a child I/O to do some work for us. 1022 */ 1023zio_t * 1024zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset, 1025 void *data, uint64_t size, int type, zio_priority_t priority, 1026 enum zio_flag flags, zio_done_func_t *done, void *private) 1027{ 1028 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE; 1029 zio_t *zio; 1030 1031 ASSERT(vd->vdev_parent == 1032 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev)); 1033 1034 if (type == ZIO_TYPE_READ && bp != NULL) { 1035 /* 1036 * If we have the bp, then the child should perform the 1037 * checksum and the parent need not. This pushes error 1038 * detection as close to the leaves as possible and 1039 * eliminates redundant checksums in the interior nodes. 1040 */ 1041 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY; 1042 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; 1043 } 1044 1045 /* Not all IO types require vdev io done stage e.g. free */ 1046 if (!(pio->io_pipeline & ZIO_STAGE_VDEV_IO_DONE)) 1047 pipeline &= ~ZIO_STAGE_VDEV_IO_DONE; 1048 1049 if (vd->vdev_children == 0) 1050 offset += VDEV_LABEL_START_SIZE; 1051 1052 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE; 1053 1054 /* 1055 * If we've decided to do a repair, the write is not speculative -- 1056 * even if the original read was. 1057 */ 1058 if (flags & ZIO_FLAG_IO_REPAIR) 1059 flags &= ~ZIO_FLAG_SPECULATIVE; 1060 1061 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, 1062 done, private, type, priority, flags, vd, offset, &pio->io_bookmark, 1063 ZIO_STAGE_VDEV_IO_START >> 1, pipeline); 1064 1065 zio->io_physdone = pio->io_physdone; 1066 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL) 1067 zio->io_logical->io_phys_children++; 1068 1069 return (zio); 1070} 1071 1072zio_t * 1073zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size, 1074 int type, zio_priority_t priority, enum zio_flag flags, 1075 zio_done_func_t *done, void *private) 1076{ 1077 zio_t *zio; 1078 1079 ASSERT(vd->vdev_ops->vdev_op_leaf); 1080 1081 zio = zio_create(NULL, vd->vdev_spa, 0, NULL, 1082 data, size, done, private, type, priority, 1083 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED, 1084 vd, offset, NULL, 1085 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE); 1086 1087 return (zio); 1088} 1089 1090void 1091zio_flush(zio_t *zio, vdev_t *vd) 1092{ 1093 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0, 1094 NULL, NULL, ZIO_PRIORITY_NOW, 1095 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY)); 1096} 1097 1098zio_t * 1099zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size) 1100{ 1101 1102 ASSERT(vd->vdev_ops->vdev_op_leaf); 1103 1104 return (zio_create(zio, spa, 0, NULL, NULL, size, NULL, NULL, 1105 ZIO_TYPE_FREE, ZIO_PRIORITY_TRIM, ZIO_FLAG_DONT_AGGREGATE | 1106 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY, 1107 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PHYS_PIPELINE)); 1108} 1109 1110void 1111zio_shrink(zio_t *zio, uint64_t size) 1112{ 1113 ASSERT(zio->io_executor == NULL); 1114 ASSERT(zio->io_orig_size == zio->io_size); 1115 ASSERT(size <= zio->io_size); 1116 1117 /* 1118 * We don't shrink for raidz because of problems with the 1119 * reconstruction when reading back less than the block size. 1120 * Note, BP_IS_RAIDZ() assumes no compression. 1121 */ 1122 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 1123 if (!BP_IS_RAIDZ(zio->io_bp)) 1124 zio->io_orig_size = zio->io_size = size; 1125} 1126 1127/* 1128 * ========================================================================== 1129 * Prepare to read and write logical blocks 1130 * ========================================================================== 1131 */ 1132 1133static int 1134zio_read_bp_init(zio_t *zio) 1135{ 1136 blkptr_t *bp = zio->io_bp; 1137 1138 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF && 1139 zio->io_child_type == ZIO_CHILD_LOGICAL && 1140 !(zio->io_flags & ZIO_FLAG_RAW)) { 1141 uint64_t psize = 1142 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp); 1143 void *cbuf = zio_buf_alloc(psize); 1144 1145 zio_push_transform(zio, cbuf, psize, psize, zio_decompress); 1146 } 1147 1148 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) { 1149 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1150 decode_embedded_bp_compressed(bp, zio->io_data); 1151 } else { 1152 ASSERT(!BP_IS_EMBEDDED(bp)); 1153 } 1154 1155 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0) 1156 zio->io_flags |= ZIO_FLAG_DONT_CACHE; 1157 1158 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP) 1159 zio->io_flags |= ZIO_FLAG_DONT_CACHE; 1160 1161 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL) 1162 zio->io_pipeline = ZIO_DDT_READ_PIPELINE; 1163 1164 return (ZIO_PIPELINE_CONTINUE); 1165} 1166 1167static int 1168zio_write_bp_init(zio_t *zio) 1169{ 1170 spa_t *spa = zio->io_spa; 1171 zio_prop_t *zp = &zio->io_prop; 1172 enum zio_compress compress = zp->zp_compress; 1173 blkptr_t *bp = zio->io_bp; 1174 uint64_t lsize = zio->io_size; 1175 uint64_t psize = lsize; 1176 int pass = 1; 1177 1178 /* 1179 * If our children haven't all reached the ready stage, 1180 * wait for them and then repeat this pipeline stage. 1181 */ 1182 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) || 1183 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY)) 1184 return (ZIO_PIPELINE_STOP); 1185 1186 if (!IO_IS_ALLOCATING(zio)) 1187 return (ZIO_PIPELINE_CONTINUE); 1188 1189 ASSERT(zio->io_child_type != ZIO_CHILD_DDT); 1190 1191 if (zio->io_bp_override) { 1192 ASSERT(bp->blk_birth != zio->io_txg); 1193 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0); 1194 1195 *bp = *zio->io_bp_override; 1196 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1197 1198 if (BP_IS_EMBEDDED(bp)) 1199 return (ZIO_PIPELINE_CONTINUE); 1200 1201 /* 1202 * If we've been overridden and nopwrite is set then 1203 * set the flag accordingly to indicate that a nopwrite 1204 * has already occurred. 1205 */ 1206 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) { 1207 ASSERT(!zp->zp_dedup); 1208 zio->io_flags |= ZIO_FLAG_NOPWRITE; 1209 return (ZIO_PIPELINE_CONTINUE); 1210 } 1211 1212 ASSERT(!zp->zp_nopwrite); 1213 1214 if (BP_IS_HOLE(bp) || !zp->zp_dedup) 1215 return (ZIO_PIPELINE_CONTINUE); 1216 1217 ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags & 1218 ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify); 1219 1220 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) { 1221 BP_SET_DEDUP(bp, 1); 1222 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE; 1223 return (ZIO_PIPELINE_CONTINUE); 1224 } 1225 } 1226 1227 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) { 1228 /* 1229 * We're rewriting an existing block, which means we're 1230 * working on behalf of spa_sync(). For spa_sync() to 1231 * converge, it must eventually be the case that we don't 1232 * have to allocate new blocks. But compression changes 1233 * the blocksize, which forces a reallocate, and makes 1234 * convergence take longer. Therefore, after the first 1235 * few passes, stop compressing to ensure convergence. 1236 */ 1237 pass = spa_sync_pass(spa); 1238 1239 ASSERT(zio->io_txg == spa_syncing_txg(spa)); 1240 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1241 ASSERT(!BP_GET_DEDUP(bp)); 1242 1243 if (pass >= zfs_sync_pass_dont_compress) 1244 compress = ZIO_COMPRESS_OFF; 1245 1246 /* Make sure someone doesn't change their mind on overwrites */ 1247 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp), 1248 spa_max_replication(spa)) == BP_GET_NDVAS(bp)); 1249 } 1250 1251 if (compress != ZIO_COMPRESS_OFF) { 1252 void *cbuf = zio_buf_alloc(lsize); 1253 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize); 1254 if (psize == 0 || psize == lsize) { 1255 compress = ZIO_COMPRESS_OFF; 1256 zio_buf_free(cbuf, lsize); 1257 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE && 1258 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) && 1259 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) { 1260 encode_embedded_bp_compressed(bp, 1261 cbuf, compress, lsize, psize); 1262 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA); 1263 BP_SET_TYPE(bp, zio->io_prop.zp_type); 1264 BP_SET_LEVEL(bp, zio->io_prop.zp_level); 1265 zio_buf_free(cbuf, lsize); 1266 bp->blk_birth = zio->io_txg; 1267 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1268 ASSERT(spa_feature_is_active(spa, 1269 SPA_FEATURE_EMBEDDED_DATA)); 1270 return (ZIO_PIPELINE_CONTINUE); 1271 } else { 1272 /* 1273 * Round up compressed size up to the ashift 1274 * of the smallest-ashift device, and zero the tail. 1275 * This ensures that the compressed size of the BP 1276 * (and thus compressratio property) are correct, 1277 * in that we charge for the padding used to fill out 1278 * the last sector. 1279 */ 1280 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT); 1281 size_t rounded = (size_t)P2ROUNDUP(psize, 1282 1ULL << spa->spa_min_ashift); 1283 if (rounded >= lsize) { 1284 compress = ZIO_COMPRESS_OFF; 1285 zio_buf_free(cbuf, lsize); 1286 psize = lsize; 1287 } else { 1288 bzero((char *)cbuf + psize, rounded - psize); 1289 psize = rounded; 1290 zio_push_transform(zio, cbuf, 1291 psize, lsize, NULL); 1292 } 1293 } 1294 } 1295 1296 /* 1297 * The final pass of spa_sync() must be all rewrites, but the first 1298 * few passes offer a trade-off: allocating blocks defers convergence, 1299 * but newly allocated blocks are sequential, so they can be written 1300 * to disk faster. Therefore, we allow the first few passes of 1301 * spa_sync() to allocate new blocks, but force rewrites after that. 1302 * There should only be a handful of blocks after pass 1 in any case. 1303 */ 1304 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg && 1305 BP_GET_PSIZE(bp) == psize && 1306 pass >= zfs_sync_pass_rewrite) { 1307 ASSERT(psize != 0); 1308 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES; 1309 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages; 1310 zio->io_flags |= ZIO_FLAG_IO_REWRITE; 1311 } else { 1312 BP_ZERO(bp); 1313 zio->io_pipeline = ZIO_WRITE_PIPELINE; 1314 } 1315 1316 if (psize == 0) { 1317 if (zio->io_bp_orig.blk_birth != 0 && 1318 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) { 1319 BP_SET_LSIZE(bp, lsize); 1320 BP_SET_TYPE(bp, zp->zp_type); 1321 BP_SET_LEVEL(bp, zp->zp_level); 1322 BP_SET_BIRTH(bp, zio->io_txg, 0); 1323 } 1324 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1325 } else { 1326 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER); 1327 BP_SET_LSIZE(bp, lsize); 1328 BP_SET_TYPE(bp, zp->zp_type); 1329 BP_SET_LEVEL(bp, zp->zp_level); 1330 BP_SET_PSIZE(bp, psize); 1331 BP_SET_COMPRESS(bp, compress); 1332 BP_SET_CHECKSUM(bp, zp->zp_checksum); 1333 BP_SET_DEDUP(bp, zp->zp_dedup); 1334 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER); 1335 if (zp->zp_dedup) { 1336 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1337 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 1338 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE; 1339 } 1340 if (zp->zp_nopwrite) { 1341 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1342 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 1343 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE; 1344 } 1345 } 1346 1347 return (ZIO_PIPELINE_CONTINUE); 1348} 1349 1350static int 1351zio_free_bp_init(zio_t *zio) 1352{ 1353 blkptr_t *bp = zio->io_bp; 1354 1355 if (zio->io_child_type == ZIO_CHILD_LOGICAL) { 1356 if (BP_GET_DEDUP(bp)) 1357 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE; 1358 } 1359 1360 return (ZIO_PIPELINE_CONTINUE); 1361} 1362 1363/* 1364 * ========================================================================== 1365 * Execute the I/O pipeline 1366 * ========================================================================== 1367 */ 1368 1369static void 1370zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline) 1371{ 1372 spa_t *spa = zio->io_spa; 1373 zio_type_t t = zio->io_type; 1374 int flags = (cutinline ? TQ_FRONT : 0); 1375 1376 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT); 1377 1378 /* 1379 * If we're a config writer or a probe, the normal issue and 1380 * interrupt threads may all be blocked waiting for the config lock. 1381 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL. 1382 */ 1383 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE)) 1384 t = ZIO_TYPE_NULL; 1385 1386 /* 1387 * A similar issue exists for the L2ARC write thread until L2ARC 2.0. 1388 */ 1389 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux) 1390 t = ZIO_TYPE_NULL; 1391 1392 /* 1393 * If this is a high priority I/O, then use the high priority taskq if 1394 * available. 1395 */ 1396 if (zio->io_priority == ZIO_PRIORITY_NOW && 1397 spa->spa_zio_taskq[t][q + 1].stqs_count != 0) 1398 q++; 1399 1400 ASSERT3U(q, <, ZIO_TASKQ_TYPES); 1401 1402 /* 1403 * NB: We are assuming that the zio can only be dispatched 1404 * to a single taskq at a time. It would be a grievous error 1405 * to dispatch the zio to another taskq at the same time. 1406 */ 1407#if defined(illumos) || !defined(_KERNEL) 1408 ASSERT(zio->io_tqent.tqent_next == NULL); 1409#else 1410 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0); 1411#endif 1412 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio, 1413 flags, &zio->io_tqent); 1414} 1415 1416static boolean_t 1417zio_taskq_member(zio_t *zio, zio_taskq_type_t q) 1418{ 1419 kthread_t *executor = zio->io_executor; 1420 spa_t *spa = zio->io_spa; 1421 1422 for (zio_type_t t = 0; t < ZIO_TYPES; t++) { 1423 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q]; 1424 uint_t i; 1425 for (i = 0; i < tqs->stqs_count; i++) { 1426 if (taskq_member(tqs->stqs_taskq[i], executor)) 1427 return (B_TRUE); 1428 } 1429 } 1430 1431 return (B_FALSE); 1432} 1433 1434static int 1435zio_issue_async(zio_t *zio) 1436{ 1437 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); 1438 1439 return (ZIO_PIPELINE_STOP); 1440} 1441 1442void 1443zio_interrupt(zio_t *zio) 1444{ 1445 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE); 1446} 1447 1448/* 1449 * Execute the I/O pipeline until one of the following occurs: 1450 * 1451 * (1) the I/O completes 1452 * (2) the pipeline stalls waiting for dependent child I/Os 1453 * (3) the I/O issues, so we're waiting for an I/O completion interrupt 1454 * (4) the I/O is delegated by vdev-level caching or aggregation 1455 * (5) the I/O is deferred due to vdev-level queueing 1456 * (6) the I/O is handed off to another thread. 1457 * 1458 * In all cases, the pipeline stops whenever there's no CPU work; it never 1459 * burns a thread in cv_wait(). 1460 * 1461 * There's no locking on io_stage because there's no legitimate way 1462 * for multiple threads to be attempting to process the same I/O. 1463 */ 1464static zio_pipe_stage_t *zio_pipeline[]; 1465 1466void 1467zio_execute(zio_t *zio) 1468{ 1469 zio->io_executor = curthread; 1470 1471 while (zio->io_stage < ZIO_STAGE_DONE) { 1472 enum zio_stage pipeline = zio->io_pipeline; 1473 enum zio_stage stage = zio->io_stage; 1474 int rv; 1475 1476 ASSERT(!MUTEX_HELD(&zio->io_lock)); 1477 ASSERT(ISP2(stage)); 1478 ASSERT(zio->io_stall == NULL); 1479 1480 do { 1481 stage <<= 1; 1482 } while ((stage & pipeline) == 0); 1483 1484 ASSERT(stage <= ZIO_STAGE_DONE); 1485 1486 /* 1487 * If we are in interrupt context and this pipeline stage 1488 * will grab a config lock that is held across I/O, 1489 * or may wait for an I/O that needs an interrupt thread 1490 * to complete, issue async to avoid deadlock. 1491 * 1492 * For VDEV_IO_START, we cut in line so that the io will 1493 * be sent to disk promptly. 1494 */ 1495 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL && 1496 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) { 1497 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? 1498 zio_requeue_io_start_cut_in_line : B_FALSE; 1499 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); 1500 return; 1501 } 1502 1503 zio->io_stage = stage; 1504 rv = zio_pipeline[highbit64(stage) - 1](zio); 1505 1506 if (rv == ZIO_PIPELINE_STOP) 1507 return; 1508 1509 ASSERT(rv == ZIO_PIPELINE_CONTINUE); 1510 } 1511} 1512 1513/* 1514 * ========================================================================== 1515 * Initiate I/O, either sync or async 1516 * ========================================================================== 1517 */ 1518int 1519zio_wait(zio_t *zio) 1520{ 1521 int error; 1522 1523 ASSERT(zio->io_stage == ZIO_STAGE_OPEN); 1524 ASSERT(zio->io_executor == NULL); 1525 1526 zio->io_waiter = curthread; 1527 1528 zio_execute(zio); 1529 1530 mutex_enter(&zio->io_lock); 1531 while (zio->io_executor != NULL) 1532 cv_wait(&zio->io_cv, &zio->io_lock); 1533 mutex_exit(&zio->io_lock); 1534 1535 error = zio->io_error; 1536 zio_destroy(zio); 1537 1538 return (error); 1539} 1540 1541void 1542zio_nowait(zio_t *zio) 1543{ 1544 ASSERT(zio->io_executor == NULL); 1545 1546 if (zio->io_child_type == ZIO_CHILD_LOGICAL && 1547 zio_unique_parent(zio) == NULL) { 1548 /* 1549 * This is a logical async I/O with no parent to wait for it. 1550 * We add it to the spa_async_root_zio "Godfather" I/O which 1551 * will ensure they complete prior to unloading the pool. 1552 */ 1553 spa_t *spa = zio->io_spa; 1554 1555 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio); 1556 } 1557 1558 zio_execute(zio); 1559} 1560 1561/* 1562 * ========================================================================== 1563 * Reexecute or suspend/resume failed I/O 1564 * ========================================================================== 1565 */ 1566 1567static void 1568zio_reexecute(zio_t *pio) 1569{ 1570 zio_t *cio, *cio_next; 1571 1572 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL); 1573 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN); 1574 ASSERT(pio->io_gang_leader == NULL); 1575 ASSERT(pio->io_gang_tree == NULL); 1576 1577 pio->io_flags = pio->io_orig_flags; 1578 pio->io_stage = pio->io_orig_stage; 1579 pio->io_pipeline = pio->io_orig_pipeline; 1580 pio->io_reexecute = 0; 1581 pio->io_flags |= ZIO_FLAG_REEXECUTED; 1582 pio->io_error = 0; 1583 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 1584 pio->io_state[w] = 0; 1585 for (int c = 0; c < ZIO_CHILD_TYPES; c++) 1586 pio->io_child_error[c] = 0; 1587 1588 if (IO_IS_ALLOCATING(pio)) 1589 BP_ZERO(pio->io_bp); 1590 1591 /* 1592 * As we reexecute pio's children, new children could be created. 1593 * New children go to the head of pio's io_child_list, however, 1594 * so we will (correctly) not reexecute them. The key is that 1595 * the remainder of pio's io_child_list, from 'cio_next' onward, 1596 * cannot be affected by any side effects of reexecuting 'cio'. 1597 */ 1598 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) { 1599 cio_next = zio_walk_children(pio); 1600 mutex_enter(&pio->io_lock); 1601 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 1602 pio->io_children[cio->io_child_type][w]++; 1603 mutex_exit(&pio->io_lock); 1604 zio_reexecute(cio); 1605 } 1606 1607 /* 1608 * Now that all children have been reexecuted, execute the parent. 1609 * We don't reexecute "The Godfather" I/O here as it's the 1610 * responsibility of the caller to wait on him. 1611 */ 1612 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) 1613 zio_execute(pio); 1614} 1615 1616void 1617zio_suspend(spa_t *spa, zio_t *zio) 1618{ 1619 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC) 1620 fm_panic("Pool '%s' has encountered an uncorrectable I/O " 1621 "failure and the failure mode property for this pool " 1622 "is set to panic.", spa_name(spa)); 1623 1624 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0); 1625 1626 mutex_enter(&spa->spa_suspend_lock); 1627 1628 if (spa->spa_suspend_zio_root == NULL) 1629 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL, 1630 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | 1631 ZIO_FLAG_GODFATHER); 1632 1633 spa->spa_suspended = B_TRUE; 1634 1635 if (zio != NULL) { 1636 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); 1637 ASSERT(zio != spa->spa_suspend_zio_root); 1638 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1639 ASSERT(zio_unique_parent(zio) == NULL); 1640 ASSERT(zio->io_stage == ZIO_STAGE_DONE); 1641 zio_add_child(spa->spa_suspend_zio_root, zio); 1642 } 1643 1644 mutex_exit(&spa->spa_suspend_lock); 1645} 1646 1647int 1648zio_resume(spa_t *spa) 1649{ 1650 zio_t *pio; 1651 1652 /* 1653 * Reexecute all previously suspended i/o. 1654 */ 1655 mutex_enter(&spa->spa_suspend_lock); 1656 spa->spa_suspended = B_FALSE; 1657 cv_broadcast(&spa->spa_suspend_cv); 1658 pio = spa->spa_suspend_zio_root; 1659 spa->spa_suspend_zio_root = NULL; 1660 mutex_exit(&spa->spa_suspend_lock); 1661 1662 if (pio == NULL) 1663 return (0); 1664 1665 zio_reexecute(pio); 1666 return (zio_wait(pio)); 1667} 1668 1669void 1670zio_resume_wait(spa_t *spa) 1671{ 1672 mutex_enter(&spa->spa_suspend_lock); 1673 while (spa_suspended(spa)) 1674 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock); 1675 mutex_exit(&spa->spa_suspend_lock); 1676} 1677 1678/* 1679 * ========================================================================== 1680 * Gang blocks. 1681 * 1682 * A gang block is a collection of small blocks that looks to the DMU 1683 * like one large block. When zio_dva_allocate() cannot find a block 1684 * of the requested size, due to either severe fragmentation or the pool 1685 * being nearly full, it calls zio_write_gang_block() to construct the 1686 * block from smaller fragments. 1687 * 1688 * A gang block consists of a gang header (zio_gbh_phys_t) and up to 1689 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like 1690 * an indirect block: it's an array of block pointers. It consumes 1691 * only one sector and hence is allocatable regardless of fragmentation. 1692 * The gang header's bps point to its gang members, which hold the data. 1693 * 1694 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg> 1695 * as the verifier to ensure uniqueness of the SHA256 checksum. 1696 * Critically, the gang block bp's blk_cksum is the checksum of the data, 1697 * not the gang header. This ensures that data block signatures (needed for 1698 * deduplication) are independent of how the block is physically stored. 1699 * 1700 * Gang blocks can be nested: a gang member may itself be a gang block. 1701 * Thus every gang block is a tree in which root and all interior nodes are 1702 * gang headers, and the leaves are normal blocks that contain user data. 1703 * The root of the gang tree is called the gang leader. 1704 * 1705 * To perform any operation (read, rewrite, free, claim) on a gang block, 1706 * zio_gang_assemble() first assembles the gang tree (minus data leaves) 1707 * in the io_gang_tree field of the original logical i/o by recursively 1708 * reading the gang leader and all gang headers below it. This yields 1709 * an in-core tree containing the contents of every gang header and the 1710 * bps for every constituent of the gang block. 1711 * 1712 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree 1713 * and invokes a callback on each bp. To free a gang block, zio_gang_issue() 1714 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp. 1715 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim(). 1716 * zio_read_gang() is a wrapper around zio_read() that omits reading gang 1717 * headers, since we already have those in io_gang_tree. zio_rewrite_gang() 1718 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite() 1719 * of the gang header plus zio_checksum_compute() of the data to update the 1720 * gang header's blk_cksum as described above. 1721 * 1722 * The two-phase assemble/issue model solves the problem of partial failure -- 1723 * what if you'd freed part of a gang block but then couldn't read the 1724 * gang header for another part? Assembling the entire gang tree first 1725 * ensures that all the necessary gang header I/O has succeeded before 1726 * starting the actual work of free, claim, or write. Once the gang tree 1727 * is assembled, free and claim are in-memory operations that cannot fail. 1728 * 1729 * In the event that a gang write fails, zio_dva_unallocate() walks the 1730 * gang tree to immediately free (i.e. insert back into the space map) 1731 * everything we've allocated. This ensures that we don't get ENOSPC 1732 * errors during repeated suspend/resume cycles due to a flaky device. 1733 * 1734 * Gang rewrites only happen during sync-to-convergence. If we can't assemble 1735 * the gang tree, we won't modify the block, so we can safely defer the free 1736 * (knowing that the block is still intact). If we *can* assemble the gang 1737 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free 1738 * each constituent bp and we can allocate a new block on the next sync pass. 1739 * 1740 * In all cases, the gang tree allows complete recovery from partial failure. 1741 * ========================================================================== 1742 */ 1743 1744static zio_t * 1745zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1746{ 1747 if (gn != NULL) 1748 return (pio); 1749 1750 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp), 1751 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), 1752 &pio->io_bookmark)); 1753} 1754 1755zio_t * 1756zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1757{ 1758 zio_t *zio; 1759 1760 if (gn != NULL) { 1761 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, 1762 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority, 1763 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 1764 /* 1765 * As we rewrite each gang header, the pipeline will compute 1766 * a new gang block header checksum for it; but no one will 1767 * compute a new data checksum, so we do that here. The one 1768 * exception is the gang leader: the pipeline already computed 1769 * its data checksum because that stage precedes gang assembly. 1770 * (Presently, nothing actually uses interior data checksums; 1771 * this is just good hygiene.) 1772 */ 1773 if (gn != pio->io_gang_leader->io_gang_tree) { 1774 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp), 1775 data, BP_GET_PSIZE(bp)); 1776 } 1777 /* 1778 * If we are here to damage data for testing purposes, 1779 * leave the GBH alone so that we can detect the damage. 1780 */ 1781 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE) 1782 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; 1783 } else { 1784 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, 1785 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority, 1786 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 1787 } 1788 1789 return (zio); 1790} 1791 1792/* ARGSUSED */ 1793zio_t * 1794zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1795{ 1796 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp, 1797 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp), 1798 ZIO_GANG_CHILD_FLAGS(pio))); 1799} 1800 1801/* ARGSUSED */ 1802zio_t * 1803zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1804{ 1805 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp, 1806 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio))); 1807} 1808 1809static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = { 1810 NULL, 1811 zio_read_gang, 1812 zio_rewrite_gang, 1813 zio_free_gang, 1814 zio_claim_gang, 1815 NULL 1816}; 1817 1818static void zio_gang_tree_assemble_done(zio_t *zio); 1819 1820static zio_gang_node_t * 1821zio_gang_node_alloc(zio_gang_node_t **gnpp) 1822{ 1823 zio_gang_node_t *gn; 1824 1825 ASSERT(*gnpp == NULL); 1826 1827 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP); 1828 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE); 1829 *gnpp = gn; 1830 1831 return (gn); 1832} 1833 1834static void 1835zio_gang_node_free(zio_gang_node_t **gnpp) 1836{ 1837 zio_gang_node_t *gn = *gnpp; 1838 1839 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) 1840 ASSERT(gn->gn_child[g] == NULL); 1841 1842 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE); 1843 kmem_free(gn, sizeof (*gn)); 1844 *gnpp = NULL; 1845} 1846 1847static void 1848zio_gang_tree_free(zio_gang_node_t **gnpp) 1849{ 1850 zio_gang_node_t *gn = *gnpp; 1851 1852 if (gn == NULL) 1853 return; 1854 1855 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) 1856 zio_gang_tree_free(&gn->gn_child[g]); 1857 1858 zio_gang_node_free(gnpp); 1859} 1860 1861static void 1862zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp) 1863{ 1864 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp); 1865 1866 ASSERT(gio->io_gang_leader == gio); 1867 ASSERT(BP_IS_GANG(bp)); 1868 1869 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh, 1870 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn, 1871 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark)); 1872} 1873 1874static void 1875zio_gang_tree_assemble_done(zio_t *zio) 1876{ 1877 zio_t *gio = zio->io_gang_leader; 1878 zio_gang_node_t *gn = zio->io_private; 1879 blkptr_t *bp = zio->io_bp; 1880 1881 ASSERT(gio == zio_unique_parent(zio)); 1882 ASSERT(zio->io_child_count == 0); 1883 1884 if (zio->io_error) 1885 return; 1886 1887 if (BP_SHOULD_BYTESWAP(bp)) 1888 byteswap_uint64_array(zio->io_data, zio->io_size); 1889 1890 ASSERT(zio->io_data == gn->gn_gbh); 1891 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE); 1892 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); 1893 1894 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 1895 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; 1896 if (!BP_IS_GANG(gbp)) 1897 continue; 1898 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]); 1899 } 1900} 1901 1902static void 1903zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data) 1904{ 1905 zio_t *gio = pio->io_gang_leader; 1906 zio_t *zio; 1907 1908 ASSERT(BP_IS_GANG(bp) == !!gn); 1909 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp)); 1910 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree); 1911 1912 /* 1913 * If you're a gang header, your data is in gn->gn_gbh. 1914 * If you're a gang member, your data is in 'data' and gn == NULL. 1915 */ 1916 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data); 1917 1918 if (gn != NULL) { 1919 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); 1920 1921 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 1922 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; 1923 if (BP_IS_HOLE(gbp)) 1924 continue; 1925 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data); 1926 data = (char *)data + BP_GET_PSIZE(gbp); 1927 } 1928 } 1929 1930 if (gn == gio->io_gang_tree && gio->io_data != NULL) 1931 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data); 1932 1933 if (zio != pio) 1934 zio_nowait(zio); 1935} 1936 1937static int 1938zio_gang_assemble(zio_t *zio) 1939{ 1940 blkptr_t *bp = zio->io_bp; 1941 1942 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL); 1943 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 1944 1945 zio->io_gang_leader = zio; 1946 1947 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree); 1948 1949 return (ZIO_PIPELINE_CONTINUE); 1950} 1951 1952static int 1953zio_gang_issue(zio_t *zio) 1954{ 1955 blkptr_t *bp = zio->io_bp; 1956 1957 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE)) 1958 return (ZIO_PIPELINE_STOP); 1959 1960 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio); 1961 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 1962 1963 if (zio->io_child_error[ZIO_CHILD_GANG] == 0) 1964 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data); 1965 else 1966 zio_gang_tree_free(&zio->io_gang_tree); 1967 1968 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1969 1970 return (ZIO_PIPELINE_CONTINUE); 1971} 1972 1973static void 1974zio_write_gang_member_ready(zio_t *zio) 1975{ 1976 zio_t *pio = zio_unique_parent(zio); 1977 zio_t *gio = zio->io_gang_leader; 1978 dva_t *cdva = zio->io_bp->blk_dva; 1979 dva_t *pdva = pio->io_bp->blk_dva; 1980 uint64_t asize; 1981 1982 if (BP_IS_HOLE(zio->io_bp)) 1983 return; 1984 1985 ASSERT(BP_IS_HOLE(&zio->io_bp_orig)); 1986 1987 ASSERT(zio->io_child_type == ZIO_CHILD_GANG); 1988 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies); 1989 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp)); 1990 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp)); 1991 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp)); 1992 1993 mutex_enter(&pio->io_lock); 1994 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) { 1995 ASSERT(DVA_GET_GANG(&pdva[d])); 1996 asize = DVA_GET_ASIZE(&pdva[d]); 1997 asize += DVA_GET_ASIZE(&cdva[d]); 1998 DVA_SET_ASIZE(&pdva[d], asize); 1999 } 2000 mutex_exit(&pio->io_lock); 2001} 2002 2003static int 2004zio_write_gang_block(zio_t *pio) 2005{ 2006 spa_t *spa = pio->io_spa; 2007 blkptr_t *bp = pio->io_bp; 2008 zio_t *gio = pio->io_gang_leader; 2009 zio_t *zio; 2010 zio_gang_node_t *gn, **gnpp; 2011 zio_gbh_phys_t *gbh; 2012 uint64_t txg = pio->io_txg; 2013 uint64_t resid = pio->io_size; 2014 uint64_t lsize; 2015 int copies = gio->io_prop.zp_copies; 2016 int gbh_copies = MIN(copies + 1, spa_max_replication(spa)); 2017 zio_prop_t zp; 2018 int error; 2019 2020 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE, 2021 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, 2022 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER); 2023 if (error) { 2024 pio->io_error = error; 2025 return (ZIO_PIPELINE_CONTINUE); 2026 } 2027 2028 if (pio == gio) { 2029 gnpp = &gio->io_gang_tree; 2030 } else { 2031 gnpp = pio->io_private; 2032 ASSERT(pio->io_ready == zio_write_gang_member_ready); 2033 } 2034 2035 gn = zio_gang_node_alloc(gnpp); 2036 gbh = gn->gn_gbh; 2037 bzero(gbh, SPA_GANGBLOCKSIZE); 2038 2039 /* 2040 * Create the gang header. 2041 */ 2042 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL, 2043 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 2044 2045 /* 2046 * Create and nowait the gang children. 2047 */ 2048 for (int g = 0; resid != 0; resid -= lsize, g++) { 2049 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g), 2050 SPA_MINBLOCKSIZE); 2051 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid); 2052 2053 zp.zp_checksum = gio->io_prop.zp_checksum; 2054 zp.zp_compress = ZIO_COMPRESS_OFF; 2055 zp.zp_type = DMU_OT_NONE; 2056 zp.zp_level = 0; 2057 zp.zp_copies = gio->io_prop.zp_copies; 2058 zp.zp_dedup = B_FALSE; 2059 zp.zp_dedup_verify = B_FALSE; 2060 zp.zp_nopwrite = B_FALSE; 2061 2062 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g], 2063 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp, 2064 zio_write_gang_member_ready, NULL, NULL, &gn->gn_child[g], 2065 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), 2066 &pio->io_bookmark)); 2067 } 2068 2069 /* 2070 * Set pio's pipeline to just wait for zio to finish. 2071 */ 2072 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2073 2074 zio_nowait(zio); 2075 2076 return (ZIO_PIPELINE_CONTINUE); 2077} 2078 2079/* 2080 * The zio_nop_write stage in the pipeline determines if allocating a 2081 * new bp is necessary. The nopwrite feature can handle writes in 2082 * either syncing or open context (i.e. zil writes) and as a result is 2083 * mutually exclusive with dedup. 2084 * 2085 * By leveraging a cryptographically secure checksum, such as SHA256, we 2086 * can compare the checksums of the new data and the old to determine if 2087 * allocating a new block is required. Note that our requirements for 2088 * cryptographic strength are fairly weak: there can't be any accidental 2089 * hash collisions, but we don't need to be secure against intentional 2090 * (malicious) collisions. To trigger a nopwrite, you have to be able 2091 * to write the file to begin with, and triggering an incorrect (hash 2092 * collision) nopwrite is no worse than simply writing to the file. 2093 * That said, there are no known attacks against the checksum algorithms 2094 * used for nopwrite, assuming that the salt and the checksums 2095 * themselves remain secret. 2096 */ 2097static int 2098zio_nop_write(zio_t *zio) 2099{ 2100 blkptr_t *bp = zio->io_bp; 2101 blkptr_t *bp_orig = &zio->io_bp_orig; 2102 zio_prop_t *zp = &zio->io_prop; 2103 2104 ASSERT(BP_GET_LEVEL(bp) == 0); 2105 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 2106 ASSERT(zp->zp_nopwrite); 2107 ASSERT(!zp->zp_dedup); 2108 ASSERT(zio->io_bp_override == NULL); 2109 ASSERT(IO_IS_ALLOCATING(zio)); 2110 2111 /* 2112 * Check to see if the original bp and the new bp have matching 2113 * characteristics (i.e. same checksum, compression algorithms, etc). 2114 * If they don't then just continue with the pipeline which will 2115 * allocate a new bp. 2116 */ 2117 if (BP_IS_HOLE(bp_orig) || 2118 !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags & 2119 ZCHECKSUM_FLAG_NOPWRITE) || 2120 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) || 2121 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) || 2122 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) || 2123 zp->zp_copies != BP_GET_NDVAS(bp_orig)) 2124 return (ZIO_PIPELINE_CONTINUE); 2125 2126 /* 2127 * If the checksums match then reset the pipeline so that we 2128 * avoid allocating a new bp and issuing any I/O. 2129 */ 2130 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) { 2131 ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags & 2132 ZCHECKSUM_FLAG_NOPWRITE); 2133 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig)); 2134 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig)); 2135 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF); 2136 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop, 2137 sizeof (uint64_t)) == 0); 2138 2139 *bp = *bp_orig; 2140 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2141 zio->io_flags |= ZIO_FLAG_NOPWRITE; 2142 } 2143 2144 return (ZIO_PIPELINE_CONTINUE); 2145} 2146 2147/* 2148 * ========================================================================== 2149 * Dedup 2150 * ========================================================================== 2151 */ 2152static void 2153zio_ddt_child_read_done(zio_t *zio) 2154{ 2155 blkptr_t *bp = zio->io_bp; 2156 ddt_entry_t *dde = zio->io_private; 2157 ddt_phys_t *ddp; 2158 zio_t *pio = zio_unique_parent(zio); 2159 2160 mutex_enter(&pio->io_lock); 2161 ddp = ddt_phys_select(dde, bp); 2162 if (zio->io_error == 0) 2163 ddt_phys_clear(ddp); /* this ddp doesn't need repair */ 2164 if (zio->io_error == 0 && dde->dde_repair_data == NULL) 2165 dde->dde_repair_data = zio->io_data; 2166 else 2167 zio_buf_free(zio->io_data, zio->io_size); 2168 mutex_exit(&pio->io_lock); 2169} 2170 2171static int 2172zio_ddt_read_start(zio_t *zio) 2173{ 2174 blkptr_t *bp = zio->io_bp; 2175 2176 ASSERT(BP_GET_DEDUP(bp)); 2177 ASSERT(BP_GET_PSIZE(bp) == zio->io_size); 2178 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2179 2180 if (zio->io_child_error[ZIO_CHILD_DDT]) { 2181 ddt_t *ddt = ddt_select(zio->io_spa, bp); 2182 ddt_entry_t *dde = ddt_repair_start(ddt, bp); 2183 ddt_phys_t *ddp = dde->dde_phys; 2184 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp); 2185 blkptr_t blk; 2186 2187 ASSERT(zio->io_vsd == NULL); 2188 zio->io_vsd = dde; 2189 2190 if (ddp_self == NULL) 2191 return (ZIO_PIPELINE_CONTINUE); 2192 2193 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) { 2194 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self) 2195 continue; 2196 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp, 2197 &blk); 2198 zio_nowait(zio_read(zio, zio->io_spa, &blk, 2199 zio_buf_alloc(zio->io_size), zio->io_size, 2200 zio_ddt_child_read_done, dde, zio->io_priority, 2201 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE, 2202 &zio->io_bookmark)); 2203 } 2204 return (ZIO_PIPELINE_CONTINUE); 2205 } 2206 2207 zio_nowait(zio_read(zio, zio->io_spa, bp, 2208 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority, 2209 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark)); 2210 2211 return (ZIO_PIPELINE_CONTINUE); 2212} 2213 2214static int 2215zio_ddt_read_done(zio_t *zio) 2216{ 2217 blkptr_t *bp = zio->io_bp; 2218 2219 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE)) 2220 return (ZIO_PIPELINE_STOP); 2221 2222 ASSERT(BP_GET_DEDUP(bp)); 2223 ASSERT(BP_GET_PSIZE(bp) == zio->io_size); 2224 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2225 2226 if (zio->io_child_error[ZIO_CHILD_DDT]) { 2227 ddt_t *ddt = ddt_select(zio->io_spa, bp); 2228 ddt_entry_t *dde = zio->io_vsd; 2229 if (ddt == NULL) { 2230 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE); 2231 return (ZIO_PIPELINE_CONTINUE); 2232 } 2233 if (dde == NULL) { 2234 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1; 2235 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); 2236 return (ZIO_PIPELINE_STOP); 2237 } 2238 if (dde->dde_repair_data != NULL) { 2239 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size); 2240 zio->io_child_error[ZIO_CHILD_DDT] = 0; 2241 } 2242 ddt_repair_done(ddt, dde); 2243 zio->io_vsd = NULL; 2244 } 2245 2246 ASSERT(zio->io_vsd == NULL); 2247 2248 return (ZIO_PIPELINE_CONTINUE); 2249} 2250 2251static boolean_t 2252zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde) 2253{ 2254 spa_t *spa = zio->io_spa; 2255 2256 /* 2257 * Note: we compare the original data, not the transformed data, 2258 * because when zio->io_bp is an override bp, we will not have 2259 * pushed the I/O transforms. That's an important optimization 2260 * because otherwise we'd compress/encrypt all dmu_sync() data twice. 2261 */ 2262 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { 2263 zio_t *lio = dde->dde_lead_zio[p]; 2264 2265 if (lio != NULL) { 2266 return (lio->io_orig_size != zio->io_orig_size || 2267 bcmp(zio->io_orig_data, lio->io_orig_data, 2268 zio->io_orig_size) != 0); 2269 } 2270 } 2271 2272 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { 2273 ddt_phys_t *ddp = &dde->dde_phys[p]; 2274 2275 if (ddp->ddp_phys_birth != 0) { 2276 arc_buf_t *abuf = NULL; 2277 arc_flags_t aflags = ARC_FLAG_WAIT; 2278 blkptr_t blk = *zio->io_bp; 2279 int error; 2280 2281 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth); 2282 2283 ddt_exit(ddt); 2284 2285 error = arc_read(NULL, spa, &blk, 2286 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ, 2287 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2288 &aflags, &zio->io_bookmark); 2289 2290 if (error == 0) { 2291 if (arc_buf_size(abuf) != zio->io_orig_size || 2292 bcmp(abuf->b_data, zio->io_orig_data, 2293 zio->io_orig_size) != 0) 2294 error = SET_ERROR(EEXIST); 2295 VERIFY(arc_buf_remove_ref(abuf, &abuf)); 2296 } 2297 2298 ddt_enter(ddt); 2299 return (error != 0); 2300 } 2301 } 2302 2303 return (B_FALSE); 2304} 2305 2306static void 2307zio_ddt_child_write_ready(zio_t *zio) 2308{ 2309 int p = zio->io_prop.zp_copies; 2310 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); 2311 ddt_entry_t *dde = zio->io_private; 2312 ddt_phys_t *ddp = &dde->dde_phys[p]; 2313 zio_t *pio; 2314 2315 if (zio->io_error) 2316 return; 2317 2318 ddt_enter(ddt); 2319 2320 ASSERT(dde->dde_lead_zio[p] == zio); 2321 2322 ddt_phys_fill(ddp, zio->io_bp); 2323 2324 while ((pio = zio_walk_parents(zio)) != NULL) 2325 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg); 2326 2327 ddt_exit(ddt); 2328} 2329 2330static void 2331zio_ddt_child_write_done(zio_t *zio) 2332{ 2333 int p = zio->io_prop.zp_copies; 2334 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); 2335 ddt_entry_t *dde = zio->io_private; 2336 ddt_phys_t *ddp = &dde->dde_phys[p]; 2337 2338 ddt_enter(ddt); 2339 2340 ASSERT(ddp->ddp_refcnt == 0); 2341 ASSERT(dde->dde_lead_zio[p] == zio); 2342 dde->dde_lead_zio[p] = NULL; 2343 2344 if (zio->io_error == 0) { 2345 while (zio_walk_parents(zio) != NULL) 2346 ddt_phys_addref(ddp); 2347 } else { 2348 ddt_phys_clear(ddp); 2349 } 2350 2351 ddt_exit(ddt); 2352} 2353 2354static void 2355zio_ddt_ditto_write_done(zio_t *zio) 2356{ 2357 int p = DDT_PHYS_DITTO; 2358 zio_prop_t *zp = &zio->io_prop; 2359 blkptr_t *bp = zio->io_bp; 2360 ddt_t *ddt = ddt_select(zio->io_spa, bp); 2361 ddt_entry_t *dde = zio->io_private; 2362 ddt_phys_t *ddp = &dde->dde_phys[p]; 2363 ddt_key_t *ddk = &dde->dde_key; 2364 2365 ddt_enter(ddt); 2366 2367 ASSERT(ddp->ddp_refcnt == 0); 2368 ASSERT(dde->dde_lead_zio[p] == zio); 2369 dde->dde_lead_zio[p] = NULL; 2370 2371 if (zio->io_error == 0) { 2372 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum)); 2373 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP); 2374 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp)); 2375 if (ddp->ddp_phys_birth != 0) 2376 ddt_phys_free(ddt, ddk, ddp, zio->io_txg); 2377 ddt_phys_fill(ddp, bp); 2378 } 2379 2380 ddt_exit(ddt); 2381} 2382 2383static int 2384zio_ddt_write(zio_t *zio) 2385{ 2386 spa_t *spa = zio->io_spa; 2387 blkptr_t *bp = zio->io_bp; 2388 uint64_t txg = zio->io_txg; 2389 zio_prop_t *zp = &zio->io_prop; 2390 int p = zp->zp_copies; 2391 int ditto_copies; 2392 zio_t *cio = NULL; 2393 zio_t *dio = NULL; 2394 ddt_t *ddt = ddt_select(spa, bp); 2395 ddt_entry_t *dde; 2396 ddt_phys_t *ddp; 2397 2398 ASSERT(BP_GET_DEDUP(bp)); 2399 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum); 2400 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override); 2401 2402 ddt_enter(ddt); 2403 dde = ddt_lookup(ddt, bp, B_TRUE); 2404 ddp = &dde->dde_phys[p]; 2405 2406 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) { 2407 /* 2408 * If we're using a weak checksum, upgrade to a strong checksum 2409 * and try again. If we're already using a strong checksum, 2410 * we can't resolve it, so just convert to an ordinary write. 2411 * (And automatically e-mail a paper to Nature?) 2412 */ 2413 if (!(zio_checksum_table[zp->zp_checksum].ci_flags & 2414 ZCHECKSUM_FLAG_DEDUP)) { 2415 zp->zp_checksum = spa_dedup_checksum(spa); 2416 zio_pop_transforms(zio); 2417 zio->io_stage = ZIO_STAGE_OPEN; 2418 BP_ZERO(bp); 2419 } else { 2420 zp->zp_dedup = B_FALSE; 2421 } 2422 zio->io_pipeline = ZIO_WRITE_PIPELINE; 2423 ddt_exit(ddt); 2424 return (ZIO_PIPELINE_CONTINUE); 2425 } 2426 2427 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp); 2428 ASSERT(ditto_copies < SPA_DVAS_PER_BP); 2429 2430 if (ditto_copies > ddt_ditto_copies_present(dde) && 2431 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) { 2432 zio_prop_t czp = *zp; 2433 2434 czp.zp_copies = ditto_copies; 2435 2436 /* 2437 * If we arrived here with an override bp, we won't have run 2438 * the transform stack, so we won't have the data we need to 2439 * generate a child i/o. So, toss the override bp and restart. 2440 * This is safe, because using the override bp is just an 2441 * optimization; and it's rare, so the cost doesn't matter. 2442 */ 2443 if (zio->io_bp_override) { 2444 zio_pop_transforms(zio); 2445 zio->io_stage = ZIO_STAGE_OPEN; 2446 zio->io_pipeline = ZIO_WRITE_PIPELINE; 2447 zio->io_bp_override = NULL; 2448 BP_ZERO(bp); 2449 ddt_exit(ddt); 2450 return (ZIO_PIPELINE_CONTINUE); 2451 } 2452 2453 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data, 2454 zio->io_orig_size, &czp, NULL, NULL, 2455 zio_ddt_ditto_write_done, dde, zio->io_priority, 2456 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); 2457 2458 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL); 2459 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio; 2460 } 2461 2462 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) { 2463 if (ddp->ddp_phys_birth != 0) 2464 ddt_bp_fill(ddp, bp, txg); 2465 if (dde->dde_lead_zio[p] != NULL) 2466 zio_add_child(zio, dde->dde_lead_zio[p]); 2467 else 2468 ddt_phys_addref(ddp); 2469 } else if (zio->io_bp_override) { 2470 ASSERT(bp->blk_birth == txg); 2471 ASSERT(BP_EQUAL(bp, zio->io_bp_override)); 2472 ddt_phys_fill(ddp, bp); 2473 ddt_phys_addref(ddp); 2474 } else { 2475 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data, 2476 zio->io_orig_size, zp, zio_ddt_child_write_ready, NULL, 2477 zio_ddt_child_write_done, dde, zio->io_priority, 2478 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); 2479 2480 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL); 2481 dde->dde_lead_zio[p] = cio; 2482 } 2483 2484 ddt_exit(ddt); 2485 2486 if (cio) 2487 zio_nowait(cio); 2488 if (dio) 2489 zio_nowait(dio); 2490 2491 return (ZIO_PIPELINE_CONTINUE); 2492} 2493 2494ddt_entry_t *freedde; /* for debugging */ 2495 2496static int 2497zio_ddt_free(zio_t *zio) 2498{ 2499 spa_t *spa = zio->io_spa; 2500 blkptr_t *bp = zio->io_bp; 2501 ddt_t *ddt = ddt_select(spa, bp); 2502 ddt_entry_t *dde; 2503 ddt_phys_t *ddp; 2504 2505 ASSERT(BP_GET_DEDUP(bp)); 2506 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2507 2508 ddt_enter(ddt); 2509 freedde = dde = ddt_lookup(ddt, bp, B_TRUE); 2510 ddp = ddt_phys_select(dde, bp); 2511 ddt_phys_decref(ddp); 2512 ddt_exit(ddt); 2513 2514 return (ZIO_PIPELINE_CONTINUE); 2515} 2516 2517/* 2518 * ========================================================================== 2519 * Allocate and free blocks 2520 * ========================================================================== 2521 */ 2522static int 2523zio_dva_allocate(zio_t *zio) 2524{ 2525 spa_t *spa = zio->io_spa; 2526 metaslab_class_t *mc = spa_normal_class(spa); 2527 blkptr_t *bp = zio->io_bp; 2528 int error; 2529 int flags = 0; 2530 2531 if (zio->io_gang_leader == NULL) { 2532 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 2533 zio->io_gang_leader = zio; 2534 } 2535 2536 ASSERT(BP_IS_HOLE(bp)); 2537 ASSERT0(BP_GET_NDVAS(bp)); 2538 ASSERT3U(zio->io_prop.zp_copies, >, 0); 2539 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa)); 2540 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp)); 2541 2542 /* 2543 * The dump device does not support gang blocks so allocation on 2544 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid 2545 * the "fast" gang feature. 2546 */ 2547 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0; 2548 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ? 2549 METASLAB_GANG_CHILD : 0; 2550 error = metaslab_alloc(spa, mc, zio->io_size, bp, 2551 zio->io_prop.zp_copies, zio->io_txg, NULL, flags); 2552 2553 if (error) { 2554 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, " 2555 "size %llu, error %d", spa_name(spa), zio, zio->io_size, 2556 error); 2557 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE) 2558 return (zio_write_gang_block(zio)); 2559 zio->io_error = error; 2560 } 2561 2562 return (ZIO_PIPELINE_CONTINUE); 2563} 2564 2565static int 2566zio_dva_free(zio_t *zio) 2567{ 2568 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE); 2569 2570 return (ZIO_PIPELINE_CONTINUE); 2571} 2572 2573static int 2574zio_dva_claim(zio_t *zio) 2575{ 2576 int error; 2577 2578 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg); 2579 if (error) 2580 zio->io_error = error; 2581 2582 return (ZIO_PIPELINE_CONTINUE); 2583} 2584 2585/* 2586 * Undo an allocation. This is used by zio_done() when an I/O fails 2587 * and we want to give back the block we just allocated. 2588 * This handles both normal blocks and gang blocks. 2589 */ 2590static void 2591zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp) 2592{ 2593 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp)); 2594 ASSERT(zio->io_bp_override == NULL); 2595 2596 if (!BP_IS_HOLE(bp)) 2597 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE); 2598 2599 if (gn != NULL) { 2600 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 2601 zio_dva_unallocate(zio, gn->gn_child[g], 2602 &gn->gn_gbh->zg_blkptr[g]); 2603 } 2604 } 2605} 2606 2607/* 2608 * Try to allocate an intent log block. Return 0 on success, errno on failure. 2609 */ 2610int 2611zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp, 2612 uint64_t size, boolean_t use_slog) 2613{ 2614 int error = 1; 2615 2616 ASSERT(txg > spa_syncing_txg(spa)); 2617 2618 /* 2619 * ZIL blocks are always contiguous (i.e. not gang blocks) so we 2620 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang" 2621 * when allocating them. 2622 */ 2623 if (use_slog) { 2624 error = metaslab_alloc(spa, spa_log_class(spa), size, 2625 new_bp, 1, txg, old_bp, 2626 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID); 2627 } 2628 2629 if (error) { 2630 error = metaslab_alloc(spa, spa_normal_class(spa), size, 2631 new_bp, 1, txg, old_bp, 2632 METASLAB_HINTBP_AVOID); 2633 } 2634 2635 if (error == 0) { 2636 BP_SET_LSIZE(new_bp, size); 2637 BP_SET_PSIZE(new_bp, size); 2638 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF); 2639 BP_SET_CHECKSUM(new_bp, 2640 spa_version(spa) >= SPA_VERSION_SLIM_ZIL 2641 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG); 2642 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); 2643 BP_SET_LEVEL(new_bp, 0); 2644 BP_SET_DEDUP(new_bp, 0); 2645 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER); 2646 } 2647 2648 return (error); 2649} 2650 2651/* 2652 * Free an intent log block. 2653 */ 2654void 2655zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp) 2656{ 2657 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG); 2658 ASSERT(!BP_IS_GANG(bp)); 2659 2660 zio_free(spa, txg, bp); 2661} 2662 2663/* 2664 * ========================================================================== 2665 * Read, write and delete to physical devices 2666 * ========================================================================== 2667 */ 2668 2669 2670/* 2671 * Issue an I/O to the underlying vdev. Typically the issue pipeline 2672 * stops after this stage and will resume upon I/O completion. 2673 * However, there are instances where the vdev layer may need to 2674 * continue the pipeline when an I/O was not issued. Since the I/O 2675 * that was sent to the vdev layer might be different than the one 2676 * currently active in the pipeline (see vdev_queue_io()), we explicitly 2677 * force the underlying vdev layers to call either zio_execute() or 2678 * zio_interrupt() to ensure that the pipeline continues with the correct I/O. 2679 */ 2680static int 2681zio_vdev_io_start(zio_t *zio) 2682{ 2683 vdev_t *vd = zio->io_vd; 2684 uint64_t align; 2685 spa_t *spa = zio->io_spa; 2686 int ret; 2687 2688 ASSERT(zio->io_error == 0); 2689 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0); 2690 2691 if (vd == NULL) { 2692 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) 2693 spa_config_enter(spa, SCL_ZIO, zio, RW_READER); 2694 2695 /* 2696 * The mirror_ops handle multiple DVAs in a single BP. 2697 */ 2698 vdev_mirror_ops.vdev_op_io_start(zio); 2699 return (ZIO_PIPELINE_STOP); 2700 } 2701 2702 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE && 2703 zio->io_priority == ZIO_PRIORITY_NOW) { 2704 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg); 2705 return (ZIO_PIPELINE_CONTINUE); 2706 } 2707 2708 /* 2709 * We keep track of time-sensitive I/Os so that the scan thread 2710 * can quickly react to certain workloads. In particular, we care 2711 * about non-scrubbing, top-level reads and writes with the following 2712 * characteristics: 2713 * - synchronous writes of user data to non-slog devices 2714 * - any reads of user data 2715 * When these conditions are met, adjust the timestamp of spa_last_io 2716 * which allows the scan thread to adjust its workload accordingly. 2717 */ 2718 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL && 2719 vd == vd->vdev_top && !vd->vdev_islog && 2720 zio->io_bookmark.zb_objset != DMU_META_OBJSET && 2721 zio->io_txg != spa_syncing_txg(spa)) { 2722 uint64_t old = spa->spa_last_io; 2723 uint64_t new = ddi_get_lbolt64(); 2724 if (old != new) 2725 (void) atomic_cas_64(&spa->spa_last_io, old, new); 2726 } 2727 2728 align = 1ULL << vd->vdev_top->vdev_ashift; 2729 2730 if ((!(zio->io_flags & ZIO_FLAG_PHYSICAL) || 2731 (vd->vdev_top->vdev_physical_ashift > SPA_MINBLOCKSHIFT)) && 2732 P2PHASE(zio->io_size, align) != 0) { 2733 /* Transform logical writes to be a full physical block size. */ 2734 uint64_t asize = P2ROUNDUP(zio->io_size, align); 2735 char *abuf = NULL; 2736 if (zio->io_type == ZIO_TYPE_READ || 2737 zio->io_type == ZIO_TYPE_WRITE) 2738 abuf = zio_buf_alloc(asize); 2739 ASSERT(vd == vd->vdev_top); 2740 if (zio->io_type == ZIO_TYPE_WRITE) { 2741 bcopy(zio->io_data, abuf, zio->io_size); 2742 bzero(abuf + zio->io_size, asize - zio->io_size); 2743 } 2744 zio_push_transform(zio, abuf, asize, abuf ? asize : 0, 2745 zio_subblock); 2746 } 2747 2748 /* 2749 * If this is not a physical io, make sure that it is properly aligned 2750 * before proceeding. 2751 */ 2752 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) { 2753 ASSERT0(P2PHASE(zio->io_offset, align)); 2754 ASSERT0(P2PHASE(zio->io_size, align)); 2755 } else { 2756 /* 2757 * For physical writes, we allow 512b aligned writes and assume 2758 * the device will perform a read-modify-write as necessary. 2759 */ 2760 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE)); 2761 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE)); 2762 } 2763 2764 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa)); 2765 2766 /* 2767 * If this is a repair I/O, and there's no self-healing involved -- 2768 * that is, we're just resilvering what we expect to resilver -- 2769 * then don't do the I/O unless zio's txg is actually in vd's DTL. 2770 * This prevents spurious resilvering with nested replication. 2771 * For example, given a mirror of mirrors, (A+B)+(C+D), if only 2772 * A is out of date, we'll read from C+D, then use the data to 2773 * resilver A+B -- but we don't actually want to resilver B, just A. 2774 * The top-level mirror has no way to know this, so instead we just 2775 * discard unnecessary repairs as we work our way down the vdev tree. 2776 * The same logic applies to any form of nested replication: 2777 * ditto + mirror, RAID-Z + replacing, etc. This covers them all. 2778 */ 2779 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) && 2780 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) && 2781 zio->io_txg != 0 && /* not a delegated i/o */ 2782 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) { 2783 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 2784 zio_vdev_io_bypass(zio); 2785 return (ZIO_PIPELINE_CONTINUE); 2786 } 2787 2788 if (vd->vdev_ops->vdev_op_leaf) { 2789 switch (zio->io_type) { 2790 case ZIO_TYPE_READ: 2791 if (vdev_cache_read(zio)) 2792 return (ZIO_PIPELINE_CONTINUE); 2793 /* FALLTHROUGH */ 2794 case ZIO_TYPE_WRITE: 2795 case ZIO_TYPE_FREE: 2796 if ((zio = vdev_queue_io(zio)) == NULL) 2797 return (ZIO_PIPELINE_STOP); 2798 2799 if (!vdev_accessible(vd, zio)) { 2800 zio->io_error = SET_ERROR(ENXIO); 2801 zio_interrupt(zio); 2802 return (ZIO_PIPELINE_STOP); 2803 } 2804 break; 2805 } 2806 /* 2807 * Note that we ignore repair writes for TRIM because they can 2808 * conflict with normal writes. This isn't an issue because, by 2809 * definition, we only repair blocks that aren't freed. 2810 */ 2811 if (zio->io_type == ZIO_TYPE_WRITE && 2812 !(zio->io_flags & ZIO_FLAG_IO_REPAIR) && 2813 !trim_map_write_start(zio)) 2814 return (ZIO_PIPELINE_STOP); 2815 } 2816 2817 vd->vdev_ops->vdev_op_io_start(zio); 2818 return (ZIO_PIPELINE_STOP); 2819} 2820 2821static int 2822zio_vdev_io_done(zio_t *zio) 2823{ 2824 vdev_t *vd = zio->io_vd; 2825 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops; 2826 boolean_t unexpected_error = B_FALSE; 2827 2828 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE)) 2829 return (ZIO_PIPELINE_STOP); 2830 2831 ASSERT(zio->io_type == ZIO_TYPE_READ || 2832 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE); 2833 2834 if (vd != NULL && vd->vdev_ops->vdev_op_leaf && 2835 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE || 2836 zio->io_type == ZIO_TYPE_FREE)) { 2837 2838 if (zio->io_type == ZIO_TYPE_WRITE && 2839 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) 2840 trim_map_write_done(zio); 2841 2842 vdev_queue_io_done(zio); 2843 2844 if (zio->io_type == ZIO_TYPE_WRITE) 2845 vdev_cache_write(zio); 2846 2847 if (zio_injection_enabled && zio->io_error == 0) 2848 zio->io_error = zio_handle_device_injection(vd, 2849 zio, EIO); 2850 2851 if (zio_injection_enabled && zio->io_error == 0) 2852 zio->io_error = zio_handle_label_injection(zio, EIO); 2853 2854 if (zio->io_error) { 2855 if (zio->io_error == ENOTSUP && 2856 zio->io_type == ZIO_TYPE_FREE) { 2857 /* Not all devices support TRIM. */ 2858 } else if (!vdev_accessible(vd, zio)) { 2859 zio->io_error = SET_ERROR(ENXIO); 2860 } else { 2861 unexpected_error = B_TRUE; 2862 } 2863 } 2864 } 2865 2866 ops->vdev_op_io_done(zio); 2867 2868 if (unexpected_error) 2869 VERIFY(vdev_probe(vd, zio) == NULL); 2870 2871 return (ZIO_PIPELINE_CONTINUE); 2872} 2873 2874/* 2875 * For non-raidz ZIOs, we can just copy aside the bad data read from the 2876 * disk, and use that to finish the checksum ereport later. 2877 */ 2878static void 2879zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr, 2880 const void *good_buf) 2881{ 2882 /* no processing needed */ 2883 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE); 2884} 2885 2886/*ARGSUSED*/ 2887void 2888zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored) 2889{ 2890 void *buf = zio_buf_alloc(zio->io_size); 2891 2892 bcopy(zio->io_data, buf, zio->io_size); 2893 2894 zcr->zcr_cbinfo = zio->io_size; 2895 zcr->zcr_cbdata = buf; 2896 zcr->zcr_finish = zio_vsd_default_cksum_finish; 2897 zcr->zcr_free = zio_buf_free; 2898} 2899 2900static int 2901zio_vdev_io_assess(zio_t *zio) 2902{ 2903 vdev_t *vd = zio->io_vd; 2904 2905 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE)) 2906 return (ZIO_PIPELINE_STOP); 2907 2908 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) 2909 spa_config_exit(zio->io_spa, SCL_ZIO, zio); 2910 2911 if (zio->io_vsd != NULL) { 2912 zio->io_vsd_ops->vsd_free(zio); 2913 zio->io_vsd = NULL; 2914 } 2915 2916 if (zio_injection_enabled && zio->io_error == 0) 2917 zio->io_error = zio_handle_fault_injection(zio, EIO); 2918 2919 if (zio->io_type == ZIO_TYPE_FREE && 2920 zio->io_priority != ZIO_PRIORITY_NOW) { 2921 switch (zio->io_error) { 2922 case 0: 2923 ZIO_TRIM_STAT_INCR(bytes, zio->io_size); 2924 ZIO_TRIM_STAT_BUMP(success); 2925 break; 2926 case EOPNOTSUPP: 2927 ZIO_TRIM_STAT_BUMP(unsupported); 2928 break; 2929 default: 2930 ZIO_TRIM_STAT_BUMP(failed); 2931 break; 2932 } 2933 } 2934 2935 /* 2936 * If the I/O failed, determine whether we should attempt to retry it. 2937 * 2938 * On retry, we cut in line in the issue queue, since we don't want 2939 * compression/checksumming/etc. work to prevent our (cheap) IO reissue. 2940 */ 2941 if (zio->io_error && vd == NULL && 2942 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) { 2943 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */ 2944 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */ 2945 zio->io_error = 0; 2946 zio->io_flags |= ZIO_FLAG_IO_RETRY | 2947 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE; 2948 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1; 2949 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, 2950 zio_requeue_io_start_cut_in_line); 2951 return (ZIO_PIPELINE_STOP); 2952 } 2953 2954 /* 2955 * If we got an error on a leaf device, convert it to ENXIO 2956 * if the device is not accessible at all. 2957 */ 2958 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf && 2959 !vdev_accessible(vd, zio)) 2960 zio->io_error = SET_ERROR(ENXIO); 2961 2962 /* 2963 * If we can't write to an interior vdev (mirror or RAID-Z), 2964 * set vdev_cant_write so that we stop trying to allocate from it. 2965 */ 2966 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE && 2967 vd != NULL && !vd->vdev_ops->vdev_op_leaf) { 2968 vd->vdev_cant_write = B_TRUE; 2969 } 2970 2971 if (zio->io_error) 2972 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2973 2974 if (vd != NULL && vd->vdev_ops->vdev_op_leaf && 2975 zio->io_physdone != NULL) { 2976 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED)); 2977 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV); 2978 zio->io_physdone(zio->io_logical); 2979 } 2980 2981 return (ZIO_PIPELINE_CONTINUE); 2982} 2983 2984void 2985zio_vdev_io_reissue(zio_t *zio) 2986{ 2987 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); 2988 ASSERT(zio->io_error == 0); 2989 2990 zio->io_stage >>= 1; 2991} 2992 2993void 2994zio_vdev_io_redone(zio_t *zio) 2995{ 2996 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE); 2997 2998 zio->io_stage >>= 1; 2999} 3000 3001void 3002zio_vdev_io_bypass(zio_t *zio) 3003{ 3004 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); 3005 ASSERT(zio->io_error == 0); 3006 3007 zio->io_flags |= ZIO_FLAG_IO_BYPASS; 3008 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1; 3009} 3010 3011/* 3012 * ========================================================================== 3013 * Generate and verify checksums 3014 * ========================================================================== 3015 */ 3016static int 3017zio_checksum_generate(zio_t *zio) 3018{ 3019 blkptr_t *bp = zio->io_bp; 3020 enum zio_checksum checksum; 3021 3022 if (bp == NULL) { 3023 /* 3024 * This is zio_write_phys(). 3025 * We're either generating a label checksum, or none at all. 3026 */ 3027 checksum = zio->io_prop.zp_checksum; 3028 3029 if (checksum == ZIO_CHECKSUM_OFF) 3030 return (ZIO_PIPELINE_CONTINUE); 3031 3032 ASSERT(checksum == ZIO_CHECKSUM_LABEL); 3033 } else { 3034 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) { 3035 ASSERT(!IO_IS_ALLOCATING(zio)); 3036 checksum = ZIO_CHECKSUM_GANG_HEADER; 3037 } else { 3038 checksum = BP_GET_CHECKSUM(bp); 3039 } 3040 } 3041 3042 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size); 3043 3044 return (ZIO_PIPELINE_CONTINUE); 3045} 3046 3047static int 3048zio_checksum_verify(zio_t *zio) 3049{ 3050 zio_bad_cksum_t info; 3051 blkptr_t *bp = zio->io_bp; 3052 int error; 3053 3054 ASSERT(zio->io_vd != NULL); 3055 3056 if (bp == NULL) { 3057 /* 3058 * This is zio_read_phys(). 3059 * We're either verifying a label checksum, or nothing at all. 3060 */ 3061 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF) 3062 return (ZIO_PIPELINE_CONTINUE); 3063 3064 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL); 3065 } 3066 3067 if ((error = zio_checksum_error(zio, &info)) != 0) { 3068 zio->io_error = error; 3069 if (error == ECKSUM && 3070 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) { 3071 zfs_ereport_start_checksum(zio->io_spa, 3072 zio->io_vd, zio, zio->io_offset, 3073 zio->io_size, NULL, &info); 3074 } 3075 } 3076 3077 return (ZIO_PIPELINE_CONTINUE); 3078} 3079 3080/* 3081 * Called by RAID-Z to ensure we don't compute the checksum twice. 3082 */ 3083void 3084zio_checksum_verified(zio_t *zio) 3085{ 3086 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; 3087} 3088 3089/* 3090 * ========================================================================== 3091 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other. 3092 * An error of 0 indicates success. ENXIO indicates whole-device failure, 3093 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO 3094 * indicate errors that are specific to one I/O, and most likely permanent. 3095 * Any other error is presumed to be worse because we weren't expecting it. 3096 * ========================================================================== 3097 */ 3098int 3099zio_worst_error(int e1, int e2) 3100{ 3101 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO }; 3102 int r1, r2; 3103 3104 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++) 3105 if (e1 == zio_error_rank[r1]) 3106 break; 3107 3108 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++) 3109 if (e2 == zio_error_rank[r2]) 3110 break; 3111 3112 return (r1 > r2 ? e1 : e2); 3113} 3114 3115/* 3116 * ========================================================================== 3117 * I/O completion 3118 * ========================================================================== 3119 */ 3120static int 3121zio_ready(zio_t *zio) 3122{ 3123 blkptr_t *bp = zio->io_bp; 3124 zio_t *pio, *pio_next; 3125 3126 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) || 3127 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY)) 3128 return (ZIO_PIPELINE_STOP); 3129 3130 if (zio->io_ready) { 3131 ASSERT(IO_IS_ALLOCATING(zio)); 3132 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) || 3133 (zio->io_flags & ZIO_FLAG_NOPWRITE)); 3134 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0); 3135 3136 zio->io_ready(zio); 3137 } 3138 3139 if (bp != NULL && bp != &zio->io_bp_copy) 3140 zio->io_bp_copy = *bp; 3141 3142 if (zio->io_error) 3143 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 3144 3145 mutex_enter(&zio->io_lock); 3146 zio->io_state[ZIO_WAIT_READY] = 1; 3147 pio = zio_walk_parents(zio); 3148 mutex_exit(&zio->io_lock); 3149 3150 /* 3151 * As we notify zio's parents, new parents could be added. 3152 * New parents go to the head of zio's io_parent_list, however, 3153 * so we will (correctly) not notify them. The remainder of zio's 3154 * io_parent_list, from 'pio_next' onward, cannot change because 3155 * all parents must wait for us to be done before they can be done. 3156 */ 3157 for (; pio != NULL; pio = pio_next) { 3158 pio_next = zio_walk_parents(zio); 3159 zio_notify_parent(pio, zio, ZIO_WAIT_READY); 3160 } 3161 3162 if (zio->io_flags & ZIO_FLAG_NODATA) { 3163 if (BP_IS_GANG(bp)) { 3164 zio->io_flags &= ~ZIO_FLAG_NODATA; 3165 } else { 3166 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE); 3167 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; 3168 } 3169 } 3170 3171 if (zio_injection_enabled && 3172 zio->io_spa->spa_syncing_txg == zio->io_txg) 3173 zio_handle_ignored_writes(zio); 3174 3175 return (ZIO_PIPELINE_CONTINUE); 3176} 3177 3178static int 3179zio_done(zio_t *zio) 3180{ 3181 spa_t *spa = zio->io_spa; 3182 zio_t *lio = zio->io_logical; 3183 blkptr_t *bp = zio->io_bp; 3184 vdev_t *vd = zio->io_vd; 3185 uint64_t psize = zio->io_size; 3186 zio_t *pio, *pio_next; 3187 3188 /* 3189 * If our children haven't all completed, 3190 * wait for them and then repeat this pipeline stage. 3191 */ 3192 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) || 3193 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) || 3194 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) || 3195 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE)) 3196 return (ZIO_PIPELINE_STOP); 3197 3198 for (int c = 0; c < ZIO_CHILD_TYPES; c++) 3199 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 3200 ASSERT(zio->io_children[c][w] == 0); 3201 3202 if (bp != NULL && !BP_IS_EMBEDDED(bp)) { 3203 ASSERT(bp->blk_pad[0] == 0); 3204 ASSERT(bp->blk_pad[1] == 0); 3205 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 || 3206 (bp == zio_unique_parent(zio)->io_bp)); 3207 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) && 3208 zio->io_bp_override == NULL && 3209 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) { 3210 ASSERT(!BP_SHOULD_BYTESWAP(bp)); 3211 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp)); 3212 ASSERT(BP_COUNT_GANG(bp) == 0 || 3213 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp))); 3214 } 3215 if (zio->io_flags & ZIO_FLAG_NOPWRITE) 3216 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig)); 3217 } 3218 3219 /* 3220 * If there were child vdev/gang/ddt errors, they apply to us now. 3221 */ 3222 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV); 3223 zio_inherit_child_errors(zio, ZIO_CHILD_GANG); 3224 zio_inherit_child_errors(zio, ZIO_CHILD_DDT); 3225 3226 /* 3227 * If the I/O on the transformed data was successful, generate any 3228 * checksum reports now while we still have the transformed data. 3229 */ 3230 if (zio->io_error == 0) { 3231 while (zio->io_cksum_report != NULL) { 3232 zio_cksum_report_t *zcr = zio->io_cksum_report; 3233 uint64_t align = zcr->zcr_align; 3234 uint64_t asize = P2ROUNDUP(psize, align); 3235 char *abuf = zio->io_data; 3236 3237 if (asize != psize) { 3238 abuf = zio_buf_alloc(asize); 3239 bcopy(zio->io_data, abuf, psize); 3240 bzero(abuf + psize, asize - psize); 3241 } 3242 3243 zio->io_cksum_report = zcr->zcr_next; 3244 zcr->zcr_next = NULL; 3245 zcr->zcr_finish(zcr, abuf); 3246 zfs_ereport_free_checksum(zcr); 3247 3248 if (asize != psize) 3249 zio_buf_free(abuf, asize); 3250 } 3251 } 3252 3253 zio_pop_transforms(zio); /* note: may set zio->io_error */ 3254 3255 vdev_stat_update(zio, psize); 3256 3257 if (zio->io_error) { 3258 /* 3259 * If this I/O is attached to a particular vdev, 3260 * generate an error message describing the I/O failure 3261 * at the block level. We ignore these errors if the 3262 * device is currently unavailable. 3263 */ 3264 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd)) 3265 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0); 3266 3267 if ((zio->io_error == EIO || !(zio->io_flags & 3268 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) && 3269 zio == lio) { 3270 /* 3271 * For logical I/O requests, tell the SPA to log the 3272 * error and generate a logical data ereport. 3273 */ 3274 spa_log_error(spa, zio); 3275 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio, 3276 0, 0); 3277 } 3278 } 3279 3280 if (zio->io_error && zio == lio) { 3281 /* 3282 * Determine whether zio should be reexecuted. This will 3283 * propagate all the way to the root via zio_notify_parent(). 3284 */ 3285 ASSERT(vd == NULL && bp != NULL); 3286 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 3287 3288 if (IO_IS_ALLOCATING(zio) && 3289 !(zio->io_flags & ZIO_FLAG_CANFAIL)) { 3290 if (zio->io_error != ENOSPC) 3291 zio->io_reexecute |= ZIO_REEXECUTE_NOW; 3292 else 3293 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 3294 } 3295 3296 if ((zio->io_type == ZIO_TYPE_READ || 3297 zio->io_type == ZIO_TYPE_FREE) && 3298 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && 3299 zio->io_error == ENXIO && 3300 spa_load_state(spa) == SPA_LOAD_NONE && 3301 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE) 3302 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 3303 3304 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute) 3305 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 3306 3307 /* 3308 * Here is a possibly good place to attempt to do 3309 * either combinatorial reconstruction or error correction 3310 * based on checksums. It also might be a good place 3311 * to send out preliminary ereports before we suspend 3312 * processing. 3313 */ 3314 } 3315 3316 /* 3317 * If there were logical child errors, they apply to us now. 3318 * We defer this until now to avoid conflating logical child 3319 * errors with errors that happened to the zio itself when 3320 * updating vdev stats and reporting FMA events above. 3321 */ 3322 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL); 3323 3324 if ((zio->io_error || zio->io_reexecute) && 3325 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio && 3326 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE))) 3327 zio_dva_unallocate(zio, zio->io_gang_tree, bp); 3328 3329 zio_gang_tree_free(&zio->io_gang_tree); 3330 3331 /* 3332 * Godfather I/Os should never suspend. 3333 */ 3334 if ((zio->io_flags & ZIO_FLAG_GODFATHER) && 3335 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) 3336 zio->io_reexecute = 0; 3337 3338 if (zio->io_reexecute) { 3339 /* 3340 * This is a logical I/O that wants to reexecute. 3341 * 3342 * Reexecute is top-down. When an i/o fails, if it's not 3343 * the root, it simply notifies its parent and sticks around. 3344 * The parent, seeing that it still has children in zio_done(), 3345 * does the same. This percolates all the way up to the root. 3346 * The root i/o will reexecute or suspend the entire tree. 3347 * 3348 * This approach ensures that zio_reexecute() honors 3349 * all the original i/o dependency relationships, e.g. 3350 * parents not executing until children are ready. 3351 */ 3352 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 3353 3354 zio->io_gang_leader = NULL; 3355 3356 mutex_enter(&zio->io_lock); 3357 zio->io_state[ZIO_WAIT_DONE] = 1; 3358 mutex_exit(&zio->io_lock); 3359 3360 /* 3361 * "The Godfather" I/O monitors its children but is 3362 * not a true parent to them. It will track them through 3363 * the pipeline but severs its ties whenever they get into 3364 * trouble (e.g. suspended). This allows "The Godfather" 3365 * I/O to return status without blocking. 3366 */ 3367 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) { 3368 zio_link_t *zl = zio->io_walk_link; 3369 pio_next = zio_walk_parents(zio); 3370 3371 if ((pio->io_flags & ZIO_FLAG_GODFATHER) && 3372 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) { 3373 zio_remove_child(pio, zio, zl); 3374 zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 3375 } 3376 } 3377 3378 if ((pio = zio_unique_parent(zio)) != NULL) { 3379 /* 3380 * We're not a root i/o, so there's nothing to do 3381 * but notify our parent. Don't propagate errors 3382 * upward since we haven't permanently failed yet. 3383 */ 3384 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); 3385 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE; 3386 zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 3387 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) { 3388 /* 3389 * We'd fail again if we reexecuted now, so suspend 3390 * until conditions improve (e.g. device comes online). 3391 */ 3392 zio_suspend(spa, zio); 3393 } else { 3394 /* 3395 * Reexecution is potentially a huge amount of work. 3396 * Hand it off to the otherwise-unused claim taskq. 3397 */ 3398#if defined(illumos) || !defined(_KERNEL) 3399 ASSERT(zio->io_tqent.tqent_next == NULL); 3400#else 3401 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0); 3402#endif 3403 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM, 3404 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio, 3405 0, &zio->io_tqent); 3406 } 3407 return (ZIO_PIPELINE_STOP); 3408 } 3409 3410 ASSERT(zio->io_child_count == 0); 3411 ASSERT(zio->io_reexecute == 0); 3412 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL)); 3413 3414 /* 3415 * Report any checksum errors, since the I/O is complete. 3416 */ 3417 while (zio->io_cksum_report != NULL) { 3418 zio_cksum_report_t *zcr = zio->io_cksum_report; 3419 zio->io_cksum_report = zcr->zcr_next; 3420 zcr->zcr_next = NULL; 3421 zcr->zcr_finish(zcr, NULL); 3422 zfs_ereport_free_checksum(zcr); 3423 } 3424 3425 /* 3426 * It is the responsibility of the done callback to ensure that this 3427 * particular zio is no longer discoverable for adoption, and as 3428 * such, cannot acquire any new parents. 3429 */ 3430 if (zio->io_done) 3431 zio->io_done(zio); 3432 3433 mutex_enter(&zio->io_lock); 3434 zio->io_state[ZIO_WAIT_DONE] = 1; 3435 mutex_exit(&zio->io_lock); 3436 3437 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) { 3438 zio_link_t *zl = zio->io_walk_link; 3439 pio_next = zio_walk_parents(zio); 3440 zio_remove_child(pio, zio, zl); 3441 zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 3442 } 3443 3444 if (zio->io_waiter != NULL) { 3445 mutex_enter(&zio->io_lock); 3446 zio->io_executor = NULL; 3447 cv_broadcast(&zio->io_cv); 3448 mutex_exit(&zio->io_lock); 3449 } else { 3450 zio_destroy(zio); 3451 } 3452 3453 return (ZIO_PIPELINE_STOP); 3454} 3455 3456/* 3457 * ========================================================================== 3458 * I/O pipeline definition 3459 * ========================================================================== 3460 */ 3461static zio_pipe_stage_t *zio_pipeline[] = { 3462 NULL, 3463 zio_read_bp_init, 3464 zio_free_bp_init, 3465 zio_issue_async, 3466 zio_write_bp_init, 3467 zio_checksum_generate, 3468 zio_nop_write, 3469 zio_ddt_read_start, 3470 zio_ddt_read_done, 3471 zio_ddt_write, 3472 zio_ddt_free, 3473 zio_gang_assemble, 3474 zio_gang_issue, 3475 zio_dva_allocate, 3476 zio_dva_free, 3477 zio_dva_claim, 3478 zio_ready, 3479 zio_vdev_io_start, 3480 zio_vdev_io_done, 3481 zio_vdev_io_assess, 3482 zio_checksum_verify, 3483 zio_done 3484}; 3485 3486 3487 3488 3489/* 3490 * Compare two zbookmark_phys_t's to see which we would reach first in a 3491 * pre-order traversal of the object tree. 3492 * 3493 * This is simple in every case aside from the meta-dnode object. For all other 3494 * objects, we traverse them in order (object 1 before object 2, and so on). 3495 * However, all of these objects are traversed while traversing object 0, since 3496 * the data it points to is the list of objects. Thus, we need to convert to a 3497 * canonical representation so we can compare meta-dnode bookmarks to 3498 * non-meta-dnode bookmarks. 3499 * 3500 * We do this by calculating "equivalents" for each field of the zbookmark. 3501 * zbookmarks outside of the meta-dnode use their own object and level, and 3502 * calculate the level 0 equivalent (the first L0 blkid that is contained in the 3503 * blocks this bookmark refers to) by multiplying their blkid by their span 3504 * (the number of L0 blocks contained within one block at their level). 3505 * zbookmarks inside the meta-dnode calculate their object equivalent 3506 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use 3507 * level + 1<<31 (any value larger than a level could ever be) for their level. 3508 * This causes them to always compare before a bookmark in their object 3509 * equivalent, compare appropriately to bookmarks in other objects, and to 3510 * compare appropriately to other bookmarks in the meta-dnode. 3511 */ 3512int 3513zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2, 3514 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2) 3515{ 3516 /* 3517 * These variables represent the "equivalent" values for the zbookmark, 3518 * after converting zbookmarks inside the meta dnode to their 3519 * normal-object equivalents. 3520 */ 3521 uint64_t zb1obj, zb2obj; 3522 uint64_t zb1L0, zb2L0; 3523 uint64_t zb1level, zb2level; 3524 3525 if (zb1->zb_object == zb2->zb_object && 3526 zb1->zb_level == zb2->zb_level && 3527 zb1->zb_blkid == zb2->zb_blkid) 3528 return (0); 3529 3530 /* 3531 * BP_SPANB calculates the span in blocks. 3532 */ 3533 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level); 3534 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level); 3535 3536 if (zb1->zb_object == DMU_META_DNODE_OBJECT) { 3537 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT)); 3538 zb1L0 = 0; 3539 zb1level = zb1->zb_level + COMPARE_META_LEVEL; 3540 } else { 3541 zb1obj = zb1->zb_object; 3542 zb1level = zb1->zb_level; 3543 } 3544 3545 if (zb2->zb_object == DMU_META_DNODE_OBJECT) { 3546 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT)); 3547 zb2L0 = 0; 3548 zb2level = zb2->zb_level + COMPARE_META_LEVEL; 3549 } else { 3550 zb2obj = zb2->zb_object; 3551 zb2level = zb2->zb_level; 3552 } 3553 3554 /* Now that we have a canonical representation, do the comparison. */ 3555 if (zb1obj != zb2obj) 3556 return (zb1obj < zb2obj ? -1 : 1); 3557 else if (zb1L0 != zb2L0) 3558 return (zb1L0 < zb2L0 ? -1 : 1); 3559 else if (zb1level != zb2level) 3560 return (zb1level > zb2level ? -1 : 1); 3561 /* 3562 * This can (theoretically) happen if the bookmarks have the same object 3563 * and level, but different blkids, if the block sizes are not the same. 3564 * There is presently no way to change the indirect block sizes 3565 */ 3566 return (0); 3567} 3568 3569/* 3570 * This function checks the following: given that last_block is the place that 3571 * our traversal stopped last time, does that guarantee that we've visited 3572 * every node under subtree_root? Therefore, we can't just use the raw output 3573 * of zbookmark_compare. We have to pass in a modified version of 3574 * subtree_root; by incrementing the block id, and then checking whether 3575 * last_block is before or equal to that, we can tell whether or not having 3576 * visited last_block implies that all of subtree_root's children have been 3577 * visited. 3578 */ 3579boolean_t 3580zbookmark_subtree_completed(const dnode_phys_t *dnp, 3581 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block) 3582{ 3583 zbookmark_phys_t mod_zb = *subtree_root; 3584 mod_zb.zb_blkid++; 3585 ASSERT(last_block->zb_level == 0); 3586 3587 /* The objset_phys_t isn't before anything. */ 3588 if (dnp == NULL) 3589 return (B_FALSE); 3590 3591 /* 3592 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the 3593 * data block size in sectors, because that variable is only used if 3594 * the bookmark refers to a block in the meta-dnode. Since we don't 3595 * know without examining it what object it refers to, and there's no 3596 * harm in passing in this value in other cases, we always pass it in. 3597 * 3598 * We pass in 0 for the indirect block size shift because zb2 must be 3599 * level 0. The indirect block size is only used to calculate the span 3600 * of the bookmark, but since the bookmark must be level 0, the span is 3601 * always 1, so the math works out. 3602 * 3603 * If you make changes to how the zbookmark_compare code works, be sure 3604 * to make sure that this code still works afterwards. 3605 */ 3606 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift, 3607 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb, 3608 last_block) <= 0); 3609} 3610