zio.c revision 307279
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, 2016 by Delphix. All rights reserved. 24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved. 25 * Copyright (c) 2014 Integros [integros.com] 26 */ 27 28#include <sys/sysmacros.h> 29#include <sys/zfs_context.h> 30#include <sys/fm/fs/zfs.h> 31#include <sys/spa.h> 32#include <sys/txg.h> 33#include <sys/spa_impl.h> 34#include <sys/vdev_impl.h> 35#include <sys/zio_impl.h> 36#include <sys/zio_compress.h> 37#include <sys/zio_checksum.h> 38#include <sys/dmu_objset.h> 39#include <sys/arc.h> 40#include <sys/ddt.h> 41#include <sys/trim_map.h> 42#include <sys/blkptr.h> 43#include <sys/zfeature.h> 44#include <sys/metaslab_impl.h> 45 46SYSCTL_DECL(_vfs_zfs); 47SYSCTL_NODE(_vfs_zfs, OID_AUTO, zio, CTLFLAG_RW, 0, "ZFS ZIO"); 48#if defined(__amd64__) 49static int zio_use_uma = 1; 50#else 51static int zio_use_uma = 0; 52#endif 53TUNABLE_INT("vfs.zfs.zio.use_uma", &zio_use_uma); 54SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, use_uma, CTLFLAG_RDTUN, &zio_use_uma, 0, 55 "Use uma(9) for ZIO allocations"); 56static int zio_exclude_metadata = 0; 57TUNABLE_INT("vfs.zfs.zio.exclude_metadata", &zio_exclude_metadata); 58SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, exclude_metadata, CTLFLAG_RDTUN, &zio_exclude_metadata, 0, 59 "Exclude metadata buffers from dumps as well"); 60 61zio_trim_stats_t zio_trim_stats = { 62 { "bytes", KSTAT_DATA_UINT64, 63 "Number of bytes successfully TRIMmed" }, 64 { "success", KSTAT_DATA_UINT64, 65 "Number of successful TRIM requests" }, 66 { "unsupported", KSTAT_DATA_UINT64, 67 "Number of TRIM requests that failed because TRIM is not supported" }, 68 { "failed", KSTAT_DATA_UINT64, 69 "Number of TRIM requests that failed for reasons other than not supported" }, 70}; 71 72static kstat_t *zio_trim_ksp; 73 74/* 75 * ========================================================================== 76 * I/O type descriptions 77 * ========================================================================== 78 */ 79const char *zio_type_name[ZIO_TYPES] = { 80 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim", 81 "zio_ioctl" 82}; 83 84boolean_t zio_dva_throttle_enabled = B_TRUE; 85SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, dva_throttle_enabled, CTLFLAG_RDTUN, 86 &zio_dva_throttle_enabled, 0, ""); 87 88/* 89 * ========================================================================== 90 * I/O kmem caches 91 * ========================================================================== 92 */ 93kmem_cache_t *zio_cache; 94kmem_cache_t *zio_link_cache; 95kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; 96kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; 97 98#ifdef _KERNEL 99extern vmem_t *zio_alloc_arena; 100#endif 101 102#define ZIO_PIPELINE_CONTINUE 0x100 103#define ZIO_PIPELINE_STOP 0x101 104 105#define BP_SPANB(indblkshift, level) \ 106 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT))) 107#define COMPARE_META_LEVEL 0x80000000ul 108/* 109 * The following actions directly effect the spa's sync-to-convergence logic. 110 * The values below define the sync pass when we start performing the action. 111 * Care should be taken when changing these values as they directly impact 112 * spa_sync() performance. Tuning these values may introduce subtle performance 113 * pathologies and should only be done in the context of performance analysis. 114 * These tunables will eventually be removed and replaced with #defines once 115 * enough analysis has been done to determine optimal values. 116 * 117 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that 118 * regular blocks are not deferred. 119 */ 120int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */ 121TUNABLE_INT("vfs.zfs.sync_pass_deferred_free", &zfs_sync_pass_deferred_free); 122SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_deferred_free, CTLFLAG_RDTUN, 123 &zfs_sync_pass_deferred_free, 0, "defer frees starting in this pass"); 124int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */ 125TUNABLE_INT("vfs.zfs.sync_pass_dont_compress", &zfs_sync_pass_dont_compress); 126SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_dont_compress, CTLFLAG_RDTUN, 127 &zfs_sync_pass_dont_compress, 0, "don't compress starting in this pass"); 128int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */ 129TUNABLE_INT("vfs.zfs.sync_pass_rewrite", &zfs_sync_pass_rewrite); 130SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_rewrite, CTLFLAG_RDTUN, 131 &zfs_sync_pass_rewrite, 0, "rewrite new bps starting in this pass"); 132 133/* 134 * An allocating zio is one that either currently has the DVA allocate 135 * stage set or will have it later in its lifetime. 136 */ 137#define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE) 138 139boolean_t zio_requeue_io_start_cut_in_line = B_TRUE; 140 141#ifdef illumos 142#ifdef ZFS_DEBUG 143int zio_buf_debug_limit = 16384; 144#else 145int zio_buf_debug_limit = 0; 146#endif 147#endif 148 149static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t); 150 151void 152zio_init(void) 153{ 154 size_t c; 155 zio_cache = kmem_cache_create("zio_cache", 156 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0); 157 zio_link_cache = kmem_cache_create("zio_link_cache", 158 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0); 159 if (!zio_use_uma) 160 goto out; 161 162 /* 163 * For small buffers, we want a cache for each multiple of 164 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache 165 * for each quarter-power of 2. 166 */ 167 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { 168 size_t size = (c + 1) << SPA_MINBLOCKSHIFT; 169 size_t p2 = size; 170 size_t align = 0; 171 int cflags = zio_exclude_metadata ? KMC_NODEBUG : 0; 172 173 while (!ISP2(p2)) 174 p2 &= p2 - 1; 175 176#ifdef illumos 177#ifndef _KERNEL 178 /* 179 * If we are using watchpoints, put each buffer on its own page, 180 * to eliminate the performance overhead of trapping to the 181 * kernel when modifying a non-watched buffer that shares the 182 * page with a watched buffer. 183 */ 184 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE)) 185 continue; 186#endif 187#endif /* illumos */ 188 if (size <= 4 * SPA_MINBLOCKSIZE) { 189 align = SPA_MINBLOCKSIZE; 190 } else if (IS_P2ALIGNED(size, p2 >> 2)) { 191 align = MIN(p2 >> 2, PAGESIZE); 192 } 193 194 if (align != 0) { 195 char name[36]; 196 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size); 197 zio_buf_cache[c] = kmem_cache_create(name, size, 198 align, NULL, NULL, NULL, NULL, NULL, cflags); 199 200 /* 201 * Since zio_data bufs do not appear in crash dumps, we 202 * pass KMC_NOTOUCH so that no allocator metadata is 203 * stored with the buffers. 204 */ 205 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size); 206 zio_data_buf_cache[c] = kmem_cache_create(name, size, 207 align, NULL, NULL, NULL, NULL, NULL, 208 cflags | KMC_NOTOUCH | KMC_NODEBUG); 209 } 210 } 211 212 while (--c != 0) { 213 ASSERT(zio_buf_cache[c] != NULL); 214 if (zio_buf_cache[c - 1] == NULL) 215 zio_buf_cache[c - 1] = zio_buf_cache[c]; 216 217 ASSERT(zio_data_buf_cache[c] != NULL); 218 if (zio_data_buf_cache[c - 1] == NULL) 219 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c]; 220 } 221out: 222 223 zio_inject_init(); 224 225 zio_trim_ksp = kstat_create("zfs", 0, "zio_trim", "misc", 226 KSTAT_TYPE_NAMED, 227 sizeof(zio_trim_stats) / sizeof(kstat_named_t), 228 KSTAT_FLAG_VIRTUAL); 229 230 if (zio_trim_ksp != NULL) { 231 zio_trim_ksp->ks_data = &zio_trim_stats; 232 kstat_install(zio_trim_ksp); 233 } 234} 235 236void 237zio_fini(void) 238{ 239 size_t c; 240 kmem_cache_t *last_cache = NULL; 241 kmem_cache_t *last_data_cache = NULL; 242 243 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { 244 if (zio_buf_cache[c] != last_cache) { 245 last_cache = zio_buf_cache[c]; 246 kmem_cache_destroy(zio_buf_cache[c]); 247 } 248 zio_buf_cache[c] = NULL; 249 250 if (zio_data_buf_cache[c] != last_data_cache) { 251 last_data_cache = zio_data_buf_cache[c]; 252 kmem_cache_destroy(zio_data_buf_cache[c]); 253 } 254 zio_data_buf_cache[c] = NULL; 255 } 256 257 kmem_cache_destroy(zio_link_cache); 258 kmem_cache_destroy(zio_cache); 259 260 zio_inject_fini(); 261 262 if (zio_trim_ksp != NULL) { 263 kstat_delete(zio_trim_ksp); 264 zio_trim_ksp = NULL; 265 } 266} 267 268/* 269 * ========================================================================== 270 * Allocate and free I/O buffers 271 * ========================================================================== 272 */ 273 274/* 275 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a 276 * crashdump if the kernel panics, so use it judiciously. Obviously, it's 277 * useful to inspect ZFS metadata, but if possible, we should avoid keeping 278 * excess / transient data in-core during a crashdump. 279 */ 280void * 281zio_buf_alloc(size_t size) 282{ 283 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 284 int flags = zio_exclude_metadata ? KM_NODEBUG : 0; 285 286 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 287 288 if (zio_use_uma) 289 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE)); 290 else 291 return (kmem_alloc(size, KM_SLEEP|flags)); 292} 293 294/* 295 * Use zio_data_buf_alloc to allocate data. The data will not appear in a 296 * crashdump if the kernel panics. This exists so that we will limit the amount 297 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount 298 * of kernel heap dumped to disk when the kernel panics) 299 */ 300void * 301zio_data_buf_alloc(size_t size) 302{ 303 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 304 305 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 306 307 if (zio_use_uma) 308 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE)); 309 else 310 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG)); 311} 312 313void 314zio_buf_free(void *buf, size_t size) 315{ 316 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 317 318 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 319 320 if (zio_use_uma) 321 kmem_cache_free(zio_buf_cache[c], buf); 322 else 323 kmem_free(buf, size); 324} 325 326void 327zio_data_buf_free(void *buf, size_t size) 328{ 329 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 330 331 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 332 333 if (zio_use_uma) 334 kmem_cache_free(zio_data_buf_cache[c], buf); 335 else 336 kmem_free(buf, size); 337} 338 339/* 340 * ========================================================================== 341 * Push and pop I/O transform buffers 342 * ========================================================================== 343 */ 344void 345zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize, 346 zio_transform_func_t *transform) 347{ 348 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP); 349 350 zt->zt_orig_data = zio->io_data; 351 zt->zt_orig_size = zio->io_size; 352 zt->zt_bufsize = bufsize; 353 zt->zt_transform = transform; 354 355 zt->zt_next = zio->io_transform_stack; 356 zio->io_transform_stack = zt; 357 358 zio->io_data = data; 359 zio->io_size = size; 360} 361 362void 363zio_pop_transforms(zio_t *zio) 364{ 365 zio_transform_t *zt; 366 367 while ((zt = zio->io_transform_stack) != NULL) { 368 if (zt->zt_transform != NULL) 369 zt->zt_transform(zio, 370 zt->zt_orig_data, zt->zt_orig_size); 371 372 if (zt->zt_bufsize != 0) 373 zio_buf_free(zio->io_data, zt->zt_bufsize); 374 375 zio->io_data = zt->zt_orig_data; 376 zio->io_size = zt->zt_orig_size; 377 zio->io_transform_stack = zt->zt_next; 378 379 kmem_free(zt, sizeof (zio_transform_t)); 380 } 381} 382 383/* 384 * ========================================================================== 385 * I/O transform callbacks for subblocks and decompression 386 * ========================================================================== 387 */ 388static void 389zio_subblock(zio_t *zio, void *data, uint64_t size) 390{ 391 ASSERT(zio->io_size > size); 392 393 if (zio->io_type == ZIO_TYPE_READ) 394 bcopy(zio->io_data, data, size); 395} 396 397static void 398zio_decompress(zio_t *zio, void *data, uint64_t size) 399{ 400 if (zio->io_error == 0 && 401 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp), 402 zio->io_data, data, zio->io_size, size) != 0) 403 zio->io_error = SET_ERROR(EIO); 404} 405 406/* 407 * ========================================================================== 408 * I/O parent/child relationships and pipeline interlocks 409 * ========================================================================== 410 */ 411zio_t * 412zio_walk_parents(zio_t *cio, zio_link_t **zl) 413{ 414 list_t *pl = &cio->io_parent_list; 415 416 *zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl); 417 if (*zl == NULL) 418 return (NULL); 419 420 ASSERT((*zl)->zl_child == cio); 421 return ((*zl)->zl_parent); 422} 423 424zio_t * 425zio_walk_children(zio_t *pio, zio_link_t **zl) 426{ 427 list_t *cl = &pio->io_child_list; 428 429 *zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl); 430 if (*zl == NULL) 431 return (NULL); 432 433 ASSERT((*zl)->zl_parent == pio); 434 return ((*zl)->zl_child); 435} 436 437zio_t * 438zio_unique_parent(zio_t *cio) 439{ 440 zio_link_t *zl = NULL; 441 zio_t *pio = zio_walk_parents(cio, &zl); 442 443 VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL); 444 return (pio); 445} 446 447void 448zio_add_child(zio_t *pio, zio_t *cio) 449{ 450 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP); 451 452 /* 453 * Logical I/Os can have logical, gang, or vdev children. 454 * Gang I/Os can have gang or vdev children. 455 * Vdev I/Os can only have vdev children. 456 * The following ASSERT captures all of these constraints. 457 */ 458 ASSERT(cio->io_child_type <= pio->io_child_type); 459 460 zl->zl_parent = pio; 461 zl->zl_child = cio; 462 463 mutex_enter(&cio->io_lock); 464 mutex_enter(&pio->io_lock); 465 466 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0); 467 468 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 469 pio->io_children[cio->io_child_type][w] += !cio->io_state[w]; 470 471 list_insert_head(&pio->io_child_list, zl); 472 list_insert_head(&cio->io_parent_list, zl); 473 474 pio->io_child_count++; 475 cio->io_parent_count++; 476 477 mutex_exit(&pio->io_lock); 478 mutex_exit(&cio->io_lock); 479} 480 481static void 482zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl) 483{ 484 ASSERT(zl->zl_parent == pio); 485 ASSERT(zl->zl_child == cio); 486 487 mutex_enter(&cio->io_lock); 488 mutex_enter(&pio->io_lock); 489 490 list_remove(&pio->io_child_list, zl); 491 list_remove(&cio->io_parent_list, zl); 492 493 pio->io_child_count--; 494 cio->io_parent_count--; 495 496 mutex_exit(&pio->io_lock); 497 mutex_exit(&cio->io_lock); 498 499 kmem_cache_free(zio_link_cache, zl); 500} 501 502static boolean_t 503zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait) 504{ 505 uint64_t *countp = &zio->io_children[child][wait]; 506 boolean_t waiting = B_FALSE; 507 508 mutex_enter(&zio->io_lock); 509 ASSERT(zio->io_stall == NULL); 510 if (*countp != 0) { 511 zio->io_stage >>= 1; 512 ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN); 513 zio->io_stall = countp; 514 waiting = B_TRUE; 515 } 516 mutex_exit(&zio->io_lock); 517 518 return (waiting); 519} 520 521static void 522zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait) 523{ 524 uint64_t *countp = &pio->io_children[zio->io_child_type][wait]; 525 int *errorp = &pio->io_child_error[zio->io_child_type]; 526 527 mutex_enter(&pio->io_lock); 528 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE)) 529 *errorp = zio_worst_error(*errorp, zio->io_error); 530 pio->io_reexecute |= zio->io_reexecute; 531 ASSERT3U(*countp, >, 0); 532 533 (*countp)--; 534 535 if (*countp == 0 && pio->io_stall == countp) { 536 zio_taskq_type_t type = 537 pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE : 538 ZIO_TASKQ_INTERRUPT; 539 pio->io_stall = NULL; 540 mutex_exit(&pio->io_lock); 541 /* 542 * Dispatch the parent zio in its own taskq so that 543 * the child can continue to make progress. This also 544 * prevents overflowing the stack when we have deeply nested 545 * parent-child relationships. 546 */ 547 zio_taskq_dispatch(pio, type, B_FALSE); 548 } else { 549 mutex_exit(&pio->io_lock); 550 } 551} 552 553static void 554zio_inherit_child_errors(zio_t *zio, enum zio_child c) 555{ 556 if (zio->io_child_error[c] != 0 && zio->io_error == 0) 557 zio->io_error = zio->io_child_error[c]; 558} 559 560int 561zio_timestamp_compare(const void *x1, const void *x2) 562{ 563 const zio_t *z1 = x1; 564 const zio_t *z2 = x2; 565 566 if (z1->io_queued_timestamp < z2->io_queued_timestamp) 567 return (-1); 568 if (z1->io_queued_timestamp > z2->io_queued_timestamp) 569 return (1); 570 571 if (z1->io_offset < z2->io_offset) 572 return (-1); 573 if (z1->io_offset > z2->io_offset) 574 return (1); 575 576 if (z1 < z2) 577 return (-1); 578 if (z1 > z2) 579 return (1); 580 581 return (0); 582} 583 584/* 585 * ========================================================================== 586 * Create the various types of I/O (read, write, free, etc) 587 * ========================================================================== 588 */ 589static zio_t * 590zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 591 void *data, uint64_t size, zio_done_func_t *done, void *private, 592 zio_type_t type, zio_priority_t priority, enum zio_flag flags, 593 vdev_t *vd, uint64_t offset, const zbookmark_phys_t *zb, 594 enum zio_stage stage, enum zio_stage pipeline) 595{ 596 zio_t *zio; 597 598 ASSERT3U(type == ZIO_TYPE_FREE || size, <=, SPA_MAXBLOCKSIZE); 599 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0); 600 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0); 601 602 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER)); 603 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER)); 604 ASSERT(vd || stage == ZIO_STAGE_OPEN); 605 606 zio = kmem_cache_alloc(zio_cache, KM_SLEEP); 607 bzero(zio, sizeof (zio_t)); 608 609 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL); 610 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL); 611 612 list_create(&zio->io_parent_list, sizeof (zio_link_t), 613 offsetof(zio_link_t, zl_parent_node)); 614 list_create(&zio->io_child_list, sizeof (zio_link_t), 615 offsetof(zio_link_t, zl_child_node)); 616 617 if (vd != NULL) 618 zio->io_child_type = ZIO_CHILD_VDEV; 619 else if (flags & ZIO_FLAG_GANG_CHILD) 620 zio->io_child_type = ZIO_CHILD_GANG; 621 else if (flags & ZIO_FLAG_DDT_CHILD) 622 zio->io_child_type = ZIO_CHILD_DDT; 623 else 624 zio->io_child_type = ZIO_CHILD_LOGICAL; 625 626 if (bp != NULL) { 627 zio->io_bp = (blkptr_t *)bp; 628 zio->io_bp_copy = *bp; 629 zio->io_bp_orig = *bp; 630 if (type != ZIO_TYPE_WRITE || 631 zio->io_child_type == ZIO_CHILD_DDT) 632 zio->io_bp = &zio->io_bp_copy; /* so caller can free */ 633 if (zio->io_child_type == ZIO_CHILD_LOGICAL) 634 zio->io_logical = zio; 635 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp)) 636 pipeline |= ZIO_GANG_STAGES; 637 } 638 639 zio->io_spa = spa; 640 zio->io_txg = txg; 641 zio->io_done = done; 642 zio->io_private = private; 643 zio->io_type = type; 644 zio->io_priority = priority; 645 zio->io_vd = vd; 646 zio->io_offset = offset; 647 zio->io_orig_data = zio->io_data = data; 648 zio->io_orig_size = zio->io_size = size; 649 zio->io_orig_flags = zio->io_flags = flags; 650 zio->io_orig_stage = zio->io_stage = stage; 651 zio->io_orig_pipeline = zio->io_pipeline = pipeline; 652 zio->io_pipeline_trace = ZIO_STAGE_OPEN; 653 654 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY); 655 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE); 656 657 if (zb != NULL) 658 zio->io_bookmark = *zb; 659 660 if (pio != NULL) { 661 if (zio->io_logical == NULL) 662 zio->io_logical = pio->io_logical; 663 if (zio->io_child_type == ZIO_CHILD_GANG) 664 zio->io_gang_leader = pio->io_gang_leader; 665 zio_add_child(pio, zio); 666 } 667 668 return (zio); 669} 670 671static void 672zio_destroy(zio_t *zio) 673{ 674 list_destroy(&zio->io_parent_list); 675 list_destroy(&zio->io_child_list); 676 mutex_destroy(&zio->io_lock); 677 cv_destroy(&zio->io_cv); 678 kmem_cache_free(zio_cache, zio); 679} 680 681zio_t * 682zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done, 683 void *private, enum zio_flag flags) 684{ 685 zio_t *zio; 686 687 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private, 688 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, 689 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE); 690 691 return (zio); 692} 693 694zio_t * 695zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags) 696{ 697 return (zio_null(NULL, spa, NULL, done, private, flags)); 698} 699 700void 701zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp) 702{ 703 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) { 704 zfs_panic_recover("blkptr at %p has invalid TYPE %llu", 705 bp, (longlong_t)BP_GET_TYPE(bp)); 706 } 707 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS || 708 BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) { 709 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu", 710 bp, (longlong_t)BP_GET_CHECKSUM(bp)); 711 } 712 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS || 713 BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) { 714 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu", 715 bp, (longlong_t)BP_GET_COMPRESS(bp)); 716 } 717 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) { 718 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu", 719 bp, (longlong_t)BP_GET_LSIZE(bp)); 720 } 721 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) { 722 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu", 723 bp, (longlong_t)BP_GET_PSIZE(bp)); 724 } 725 726 if (BP_IS_EMBEDDED(bp)) { 727 if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) { 728 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu", 729 bp, (longlong_t)BPE_GET_ETYPE(bp)); 730 } 731 } 732 733 /* 734 * Pool-specific checks. 735 * 736 * Note: it would be nice to verify that the blk_birth and 737 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze() 738 * allows the birth time of log blocks (and dmu_sync()-ed blocks 739 * that are in the log) to be arbitrarily large. 740 */ 741 for (int i = 0; i < BP_GET_NDVAS(bp); i++) { 742 uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]); 743 if (vdevid >= spa->spa_root_vdev->vdev_children) { 744 zfs_panic_recover("blkptr at %p DVA %u has invalid " 745 "VDEV %llu", 746 bp, i, (longlong_t)vdevid); 747 continue; 748 } 749 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid]; 750 if (vd == NULL) { 751 zfs_panic_recover("blkptr at %p DVA %u has invalid " 752 "VDEV %llu", 753 bp, i, (longlong_t)vdevid); 754 continue; 755 } 756 if (vd->vdev_ops == &vdev_hole_ops) { 757 zfs_panic_recover("blkptr at %p DVA %u has hole " 758 "VDEV %llu", 759 bp, i, (longlong_t)vdevid); 760 continue; 761 } 762 if (vd->vdev_ops == &vdev_missing_ops) { 763 /* 764 * "missing" vdevs are valid during import, but we 765 * don't have their detailed info (e.g. asize), so 766 * we can't perform any more checks on them. 767 */ 768 continue; 769 } 770 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]); 771 uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]); 772 if (BP_IS_GANG(bp)) 773 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE); 774 if (offset + asize > vd->vdev_asize) { 775 zfs_panic_recover("blkptr at %p DVA %u has invalid " 776 "OFFSET %llu", 777 bp, i, (longlong_t)offset); 778 } 779 } 780} 781 782zio_t * 783zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, 784 void *data, uint64_t size, zio_done_func_t *done, void *private, 785 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb) 786{ 787 zio_t *zio; 788 789 zfs_blkptr_verify(spa, bp); 790 791 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp, 792 data, size, done, private, 793 ZIO_TYPE_READ, priority, flags, NULL, 0, zb, 794 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? 795 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE); 796 797 return (zio); 798} 799 800zio_t * 801zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, 802 void *data, uint64_t size, const zio_prop_t *zp, 803 zio_done_func_t *ready, zio_done_func_t *children_ready, 804 zio_done_func_t *physdone, zio_done_func_t *done, 805 void *private, zio_priority_t priority, enum zio_flag flags, 806 const zbookmark_phys_t *zb) 807{ 808 zio_t *zio; 809 810 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF && 811 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS && 812 zp->zp_compress >= ZIO_COMPRESS_OFF && 813 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS && 814 DMU_OT_IS_VALID(zp->zp_type) && 815 zp->zp_level < 32 && 816 zp->zp_copies > 0 && 817 zp->zp_copies <= spa_max_replication(spa)); 818 819 zio = zio_create(pio, spa, txg, bp, data, size, done, private, 820 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, 821 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? 822 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE); 823 824 zio->io_ready = ready; 825 zio->io_children_ready = children_ready; 826 zio->io_physdone = physdone; 827 zio->io_prop = *zp; 828 829 /* 830 * Data can be NULL if we are going to call zio_write_override() to 831 * provide the already-allocated BP. But we may need the data to 832 * verify a dedup hit (if requested). In this case, don't try to 833 * dedup (just take the already-allocated BP verbatim). 834 */ 835 if (data == NULL && zio->io_prop.zp_dedup_verify) { 836 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE; 837 } 838 839 return (zio); 840} 841 842zio_t * 843zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data, 844 uint64_t size, zio_done_func_t *done, void *private, 845 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb) 846{ 847 zio_t *zio; 848 849 zio = zio_create(pio, spa, txg, bp, data, size, done, private, 850 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb, 851 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE); 852 853 return (zio); 854} 855 856void 857zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite) 858{ 859 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 860 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 861 ASSERT(zio->io_stage == ZIO_STAGE_OPEN); 862 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa)); 863 864 /* 865 * We must reset the io_prop to match the values that existed 866 * when the bp was first written by dmu_sync() keeping in mind 867 * that nopwrite and dedup are mutually exclusive. 868 */ 869 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup; 870 zio->io_prop.zp_nopwrite = nopwrite; 871 zio->io_prop.zp_copies = copies; 872 zio->io_bp_override = bp; 873} 874 875void 876zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp) 877{ 878 879 /* 880 * The check for EMBEDDED is a performance optimization. We 881 * process the free here (by ignoring it) rather than 882 * putting it on the list and then processing it in zio_free_sync(). 883 */ 884 if (BP_IS_EMBEDDED(bp)) 885 return; 886 metaslab_check_free(spa, bp); 887 888 /* 889 * Frees that are for the currently-syncing txg, are not going to be 890 * deferred, and which will not need to do a read (i.e. not GANG or 891 * DEDUP), can be processed immediately. Otherwise, put them on the 892 * in-memory list for later processing. 893 */ 894 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) || 895 txg != spa->spa_syncing_txg || 896 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) { 897 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp); 898 } else { 899 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 900 BP_GET_PSIZE(bp), 0))); 901 } 902} 903 904zio_t * 905zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 906 uint64_t size, enum zio_flag flags) 907{ 908 zio_t *zio; 909 enum zio_stage stage = ZIO_FREE_PIPELINE; 910 911 ASSERT(!BP_IS_HOLE(bp)); 912 ASSERT(spa_syncing_txg(spa) == txg); 913 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free); 914 915 if (BP_IS_EMBEDDED(bp)) 916 return (zio_null(pio, spa, NULL, NULL, NULL, 0)); 917 918 metaslab_check_free(spa, bp); 919 arc_freed(spa, bp); 920 921 if (zfs_trim_enabled) 922 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START | 923 ZIO_STAGE_VDEV_IO_ASSESS; 924 /* 925 * GANG and DEDUP blocks can induce a read (for the gang block header, 926 * or the DDT), so issue them asynchronously so that this thread is 927 * not tied up. 928 */ 929 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp)) 930 stage |= ZIO_STAGE_ISSUE_ASYNC; 931 932 flags |= ZIO_FLAG_DONT_QUEUE; 933 934 zio = zio_create(pio, spa, txg, bp, NULL, size, 935 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags, 936 NULL, 0, NULL, ZIO_STAGE_OPEN, stage); 937 938 return (zio); 939} 940 941zio_t * 942zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 943 zio_done_func_t *done, void *private, enum zio_flag flags) 944{ 945 zio_t *zio; 946 947 dprintf_bp(bp, "claiming in txg %llu", txg); 948 949 if (BP_IS_EMBEDDED(bp)) 950 return (zio_null(pio, spa, NULL, NULL, NULL, 0)); 951 952 /* 953 * A claim is an allocation of a specific block. Claims are needed 954 * to support immediate writes in the intent log. The issue is that 955 * immediate writes contain committed data, but in a txg that was 956 * *not* committed. Upon opening the pool after an unclean shutdown, 957 * the intent log claims all blocks that contain immediate write data 958 * so that the SPA knows they're in use. 959 * 960 * All claims *must* be resolved in the first txg -- before the SPA 961 * starts allocating blocks -- so that nothing is allocated twice. 962 * If txg == 0 we just verify that the block is claimable. 963 */ 964 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa)); 965 ASSERT(txg == spa_first_txg(spa) || txg == 0); 966 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */ 967 968 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), 969 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags, 970 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE); 971 ASSERT0(zio->io_queued_timestamp); 972 973 return (zio); 974} 975 976zio_t * 977zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset, 978 uint64_t size, zio_done_func_t *done, void *private, 979 zio_priority_t priority, enum zio_flag flags) 980{ 981 zio_t *zio; 982 int c; 983 984 if (vd->vdev_children == 0) { 985 zio = zio_create(pio, spa, 0, NULL, NULL, size, done, private, 986 ZIO_TYPE_IOCTL, priority, flags, vd, offset, NULL, 987 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE); 988 989 zio->io_cmd = cmd; 990 } else { 991 zio = zio_null(pio, spa, NULL, NULL, NULL, flags); 992 993 for (c = 0; c < vd->vdev_children; c++) 994 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd, 995 offset, size, done, private, priority, flags)); 996 } 997 998 return (zio); 999} 1000 1001zio_t * 1002zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 1003 void *data, int checksum, zio_done_func_t *done, void *private, 1004 zio_priority_t priority, enum zio_flag flags, boolean_t labels) 1005{ 1006 zio_t *zio; 1007 1008 ASSERT(vd->vdev_children == 0); 1009 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || 1010 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); 1011 ASSERT3U(offset + size, <=, vd->vdev_psize); 1012 1013 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private, 1014 ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset, 1015 NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE); 1016 1017 zio->io_prop.zp_checksum = checksum; 1018 1019 return (zio); 1020} 1021 1022zio_t * 1023zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 1024 void *data, int checksum, zio_done_func_t *done, void *private, 1025 zio_priority_t priority, enum zio_flag flags, boolean_t labels) 1026{ 1027 zio_t *zio; 1028 1029 ASSERT(vd->vdev_children == 0); 1030 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || 1031 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); 1032 ASSERT3U(offset + size, <=, vd->vdev_psize); 1033 1034 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private, 1035 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset, 1036 NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE); 1037 1038 zio->io_prop.zp_checksum = checksum; 1039 1040 if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) { 1041 /* 1042 * zec checksums are necessarily destructive -- they modify 1043 * the end of the write buffer to hold the verifier/checksum. 1044 * Therefore, we must make a local copy in case the data is 1045 * being written to multiple places in parallel. 1046 */ 1047 void *wbuf = zio_buf_alloc(size); 1048 bcopy(data, wbuf, size); 1049 zio_push_transform(zio, wbuf, size, size, NULL); 1050 } 1051 1052 return (zio); 1053} 1054 1055/* 1056 * Create a child I/O to do some work for us. 1057 */ 1058zio_t * 1059zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset, 1060 void *data, uint64_t size, int type, zio_priority_t priority, 1061 enum zio_flag flags, zio_done_func_t *done, void *private) 1062{ 1063 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE; 1064 zio_t *zio; 1065 1066 ASSERT(vd->vdev_parent == 1067 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev)); 1068 1069 if (type == ZIO_TYPE_READ && bp != NULL) { 1070 /* 1071 * If we have the bp, then the child should perform the 1072 * checksum and the parent need not. This pushes error 1073 * detection as close to the leaves as possible and 1074 * eliminates redundant checksums in the interior nodes. 1075 */ 1076 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY; 1077 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; 1078 } 1079 1080 /* Not all IO types require vdev io done stage e.g. free */ 1081 if (!(pio->io_pipeline & ZIO_STAGE_VDEV_IO_DONE)) 1082 pipeline &= ~ZIO_STAGE_VDEV_IO_DONE; 1083 1084 if (vd->vdev_children == 0) 1085 offset += VDEV_LABEL_START_SIZE; 1086 1087 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE; 1088 1089 /* 1090 * If we've decided to do a repair, the write is not speculative -- 1091 * even if the original read was. 1092 */ 1093 if (flags & ZIO_FLAG_IO_REPAIR) 1094 flags &= ~ZIO_FLAG_SPECULATIVE; 1095 1096 /* 1097 * If we're creating a child I/O that is not associated with a 1098 * top-level vdev, then the child zio is not an allocating I/O. 1099 * If this is a retried I/O then we ignore it since we will 1100 * have already processed the original allocating I/O. 1101 */ 1102 if (flags & ZIO_FLAG_IO_ALLOCATING && 1103 (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) { 1104 metaslab_class_t *mc = spa_normal_class(pio->io_spa); 1105 1106 ASSERT(mc->mc_alloc_throttle_enabled); 1107 ASSERT(type == ZIO_TYPE_WRITE); 1108 ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE); 1109 ASSERT(!(flags & ZIO_FLAG_IO_REPAIR)); 1110 ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) || 1111 pio->io_child_type == ZIO_CHILD_GANG); 1112 1113 flags &= ~ZIO_FLAG_IO_ALLOCATING; 1114 } 1115 1116 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, 1117 done, private, type, priority, flags, vd, offset, &pio->io_bookmark, 1118 ZIO_STAGE_VDEV_IO_START >> 1, pipeline); 1119 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV); 1120 1121 zio->io_physdone = pio->io_physdone; 1122 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL) 1123 zio->io_logical->io_phys_children++; 1124 1125 return (zio); 1126} 1127 1128zio_t * 1129zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size, 1130 int type, zio_priority_t priority, enum zio_flag flags, 1131 zio_done_func_t *done, void *private) 1132{ 1133 zio_t *zio; 1134 1135 ASSERT(vd->vdev_ops->vdev_op_leaf); 1136 1137 zio = zio_create(NULL, vd->vdev_spa, 0, NULL, 1138 data, size, done, private, type, priority, 1139 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED, 1140 vd, offset, NULL, 1141 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE); 1142 1143 return (zio); 1144} 1145 1146void 1147zio_flush(zio_t *zio, vdev_t *vd) 1148{ 1149 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0, 1150 NULL, NULL, ZIO_PRIORITY_NOW, 1151 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY)); 1152} 1153 1154zio_t * 1155zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size) 1156{ 1157 1158 ASSERT(vd->vdev_ops->vdev_op_leaf); 1159 1160 return (zio_create(zio, spa, 0, NULL, NULL, size, NULL, NULL, 1161 ZIO_TYPE_FREE, ZIO_PRIORITY_TRIM, ZIO_FLAG_DONT_AGGREGATE | 1162 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY, 1163 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PHYS_PIPELINE)); 1164} 1165 1166void 1167zio_shrink(zio_t *zio, uint64_t size) 1168{ 1169 ASSERT(zio->io_executor == NULL); 1170 ASSERT(zio->io_orig_size == zio->io_size); 1171 ASSERT(size <= zio->io_size); 1172 1173 /* 1174 * We don't shrink for raidz because of problems with the 1175 * reconstruction when reading back less than the block size. 1176 * Note, BP_IS_RAIDZ() assumes no compression. 1177 */ 1178 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 1179 if (!BP_IS_RAIDZ(zio->io_bp)) 1180 zio->io_orig_size = zio->io_size = size; 1181} 1182 1183/* 1184 * ========================================================================== 1185 * Prepare to read and write logical blocks 1186 * ========================================================================== 1187 */ 1188 1189static int 1190zio_read_bp_init(zio_t *zio) 1191{ 1192 blkptr_t *bp = zio->io_bp; 1193 1194 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF && 1195 zio->io_child_type == ZIO_CHILD_LOGICAL && 1196 !(zio->io_flags & ZIO_FLAG_RAW)) { 1197 uint64_t psize = 1198 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp); 1199 void *cbuf = zio_buf_alloc(psize); 1200 1201 zio_push_transform(zio, cbuf, psize, psize, zio_decompress); 1202 } 1203 1204 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) { 1205 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1206 decode_embedded_bp_compressed(bp, zio->io_data); 1207 } else { 1208 ASSERT(!BP_IS_EMBEDDED(bp)); 1209 } 1210 1211 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0) 1212 zio->io_flags |= ZIO_FLAG_DONT_CACHE; 1213 1214 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP) 1215 zio->io_flags |= ZIO_FLAG_DONT_CACHE; 1216 1217 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL) 1218 zio->io_pipeline = ZIO_DDT_READ_PIPELINE; 1219 1220 return (ZIO_PIPELINE_CONTINUE); 1221} 1222 1223static int 1224zio_write_bp_init(zio_t *zio) 1225{ 1226 if (!IO_IS_ALLOCATING(zio)) 1227 return (ZIO_PIPELINE_CONTINUE); 1228 1229 ASSERT(zio->io_child_type != ZIO_CHILD_DDT); 1230 1231 if (zio->io_bp_override) { 1232 blkptr_t *bp = zio->io_bp; 1233 zio_prop_t *zp = &zio->io_prop; 1234 1235 ASSERT(bp->blk_birth != zio->io_txg); 1236 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0); 1237 1238 *bp = *zio->io_bp_override; 1239 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1240 1241 if (BP_IS_EMBEDDED(bp)) 1242 return (ZIO_PIPELINE_CONTINUE); 1243 1244 /* 1245 * If we've been overridden and nopwrite is set then 1246 * set the flag accordingly to indicate that a nopwrite 1247 * has already occurred. 1248 */ 1249 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) { 1250 ASSERT(!zp->zp_dedup); 1251 ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum); 1252 zio->io_flags |= ZIO_FLAG_NOPWRITE; 1253 return (ZIO_PIPELINE_CONTINUE); 1254 } 1255 1256 ASSERT(!zp->zp_nopwrite); 1257 1258 if (BP_IS_HOLE(bp) || !zp->zp_dedup) 1259 return (ZIO_PIPELINE_CONTINUE); 1260 1261 ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags & 1262 ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify); 1263 1264 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) { 1265 BP_SET_DEDUP(bp, 1); 1266 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE; 1267 return (ZIO_PIPELINE_CONTINUE); 1268 } 1269 1270 /* 1271 * We were unable to handle this as an override bp, treat 1272 * it as a regular write I/O. 1273 */ 1274 zio->io_bp_override = NULL; 1275 *bp = zio->io_bp_orig; 1276 zio->io_pipeline = zio->io_orig_pipeline; 1277 } 1278 1279 return (ZIO_PIPELINE_CONTINUE); 1280} 1281 1282static int 1283zio_write_compress(zio_t *zio) 1284{ 1285 spa_t *spa = zio->io_spa; 1286 zio_prop_t *zp = &zio->io_prop; 1287 enum zio_compress compress = zp->zp_compress; 1288 blkptr_t *bp = zio->io_bp; 1289 uint64_t lsize = zio->io_size; 1290 uint64_t psize = lsize; 1291 int pass = 1; 1292 1293 /* 1294 * If our children haven't all reached the ready stage, 1295 * wait for them and then repeat this pipeline stage. 1296 */ 1297 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) || 1298 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY)) 1299 return (ZIO_PIPELINE_STOP); 1300 1301 if (!IO_IS_ALLOCATING(zio)) 1302 return (ZIO_PIPELINE_CONTINUE); 1303 1304 if (zio->io_children_ready != NULL) { 1305 /* 1306 * Now that all our children are ready, run the callback 1307 * associated with this zio in case it wants to modify the 1308 * data to be written. 1309 */ 1310 ASSERT3U(zp->zp_level, >, 0); 1311 zio->io_children_ready(zio); 1312 } 1313 1314 ASSERT(zio->io_child_type != ZIO_CHILD_DDT); 1315 ASSERT(zio->io_bp_override == NULL); 1316 1317 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) { 1318 /* 1319 * We're rewriting an existing block, which means we're 1320 * working on behalf of spa_sync(). For spa_sync() to 1321 * converge, it must eventually be the case that we don't 1322 * have to allocate new blocks. But compression changes 1323 * the blocksize, which forces a reallocate, and makes 1324 * convergence take longer. Therefore, after the first 1325 * few passes, stop compressing to ensure convergence. 1326 */ 1327 pass = spa_sync_pass(spa); 1328 1329 ASSERT(zio->io_txg == spa_syncing_txg(spa)); 1330 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1331 ASSERT(!BP_GET_DEDUP(bp)); 1332 1333 if (pass >= zfs_sync_pass_dont_compress) 1334 compress = ZIO_COMPRESS_OFF; 1335 1336 /* Make sure someone doesn't change their mind on overwrites */ 1337 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp), 1338 spa_max_replication(spa)) == BP_GET_NDVAS(bp)); 1339 } 1340 1341 if (compress != ZIO_COMPRESS_OFF) { 1342 void *cbuf = zio_buf_alloc(lsize); 1343 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize); 1344 if (psize == 0 || psize == lsize) { 1345 compress = ZIO_COMPRESS_OFF; 1346 zio_buf_free(cbuf, lsize); 1347 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE && 1348 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) && 1349 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) { 1350 encode_embedded_bp_compressed(bp, 1351 cbuf, compress, lsize, psize); 1352 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA); 1353 BP_SET_TYPE(bp, zio->io_prop.zp_type); 1354 BP_SET_LEVEL(bp, zio->io_prop.zp_level); 1355 zio_buf_free(cbuf, lsize); 1356 bp->blk_birth = zio->io_txg; 1357 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1358 ASSERT(spa_feature_is_active(spa, 1359 SPA_FEATURE_EMBEDDED_DATA)); 1360 return (ZIO_PIPELINE_CONTINUE); 1361 } else { 1362 /* 1363 * Round up compressed size up to the ashift 1364 * of the smallest-ashift device, and zero the tail. 1365 * This ensures that the compressed size of the BP 1366 * (and thus compressratio property) are correct, 1367 * in that we charge for the padding used to fill out 1368 * the last sector. 1369 */ 1370 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT); 1371 size_t rounded = (size_t)P2ROUNDUP(psize, 1372 1ULL << spa->spa_min_ashift); 1373 if (rounded >= lsize) { 1374 compress = ZIO_COMPRESS_OFF; 1375 zio_buf_free(cbuf, lsize); 1376 psize = lsize; 1377 } else { 1378 bzero((char *)cbuf + psize, rounded - psize); 1379 psize = rounded; 1380 zio_push_transform(zio, cbuf, 1381 psize, lsize, NULL); 1382 } 1383 } 1384 1385 /* 1386 * We were unable to handle this as an override bp, treat 1387 * it as a regular write I/O. 1388 */ 1389 zio->io_bp_override = NULL; 1390 *bp = zio->io_bp_orig; 1391 zio->io_pipeline = zio->io_orig_pipeline; 1392 } 1393 1394 /* 1395 * The final pass of spa_sync() must be all rewrites, but the first 1396 * few passes offer a trade-off: allocating blocks defers convergence, 1397 * but newly allocated blocks are sequential, so they can be written 1398 * to disk faster. Therefore, we allow the first few passes of 1399 * spa_sync() to allocate new blocks, but force rewrites after that. 1400 * There should only be a handful of blocks after pass 1 in any case. 1401 */ 1402 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg && 1403 BP_GET_PSIZE(bp) == psize && 1404 pass >= zfs_sync_pass_rewrite) { 1405 ASSERT(psize != 0); 1406 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES; 1407 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages; 1408 zio->io_flags |= ZIO_FLAG_IO_REWRITE; 1409 } else { 1410 BP_ZERO(bp); 1411 zio->io_pipeline = ZIO_WRITE_PIPELINE; 1412 } 1413 1414 if (psize == 0) { 1415 if (zio->io_bp_orig.blk_birth != 0 && 1416 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) { 1417 BP_SET_LSIZE(bp, lsize); 1418 BP_SET_TYPE(bp, zp->zp_type); 1419 BP_SET_LEVEL(bp, zp->zp_level); 1420 BP_SET_BIRTH(bp, zio->io_txg, 0); 1421 } 1422 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1423 } else { 1424 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER); 1425 BP_SET_LSIZE(bp, lsize); 1426 BP_SET_TYPE(bp, zp->zp_type); 1427 BP_SET_LEVEL(bp, zp->zp_level); 1428 BP_SET_PSIZE(bp, psize); 1429 BP_SET_COMPRESS(bp, compress); 1430 BP_SET_CHECKSUM(bp, zp->zp_checksum); 1431 BP_SET_DEDUP(bp, zp->zp_dedup); 1432 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER); 1433 if (zp->zp_dedup) { 1434 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1435 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 1436 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE; 1437 } 1438 if (zp->zp_nopwrite) { 1439 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1440 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 1441 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE; 1442 } 1443 } 1444 return (ZIO_PIPELINE_CONTINUE); 1445} 1446 1447static int 1448zio_free_bp_init(zio_t *zio) 1449{ 1450 blkptr_t *bp = zio->io_bp; 1451 1452 if (zio->io_child_type == ZIO_CHILD_LOGICAL) { 1453 if (BP_GET_DEDUP(bp)) 1454 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE; 1455 } 1456 1457 return (ZIO_PIPELINE_CONTINUE); 1458} 1459 1460/* 1461 * ========================================================================== 1462 * Execute the I/O pipeline 1463 * ========================================================================== 1464 */ 1465 1466static void 1467zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline) 1468{ 1469 spa_t *spa = zio->io_spa; 1470 zio_type_t t = zio->io_type; 1471 int flags = (cutinline ? TQ_FRONT : 0); 1472 1473 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT); 1474 1475 /* 1476 * If we're a config writer or a probe, the normal issue and 1477 * interrupt threads may all be blocked waiting for the config lock. 1478 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL. 1479 */ 1480 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE)) 1481 t = ZIO_TYPE_NULL; 1482 1483 /* 1484 * A similar issue exists for the L2ARC write thread until L2ARC 2.0. 1485 */ 1486 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux) 1487 t = ZIO_TYPE_NULL; 1488 1489 /* 1490 * If this is a high priority I/O, then use the high priority taskq if 1491 * available. 1492 */ 1493 if (zio->io_priority == ZIO_PRIORITY_NOW && 1494 spa->spa_zio_taskq[t][q + 1].stqs_count != 0) 1495 q++; 1496 1497 ASSERT3U(q, <, ZIO_TASKQ_TYPES); 1498 1499 /* 1500 * NB: We are assuming that the zio can only be dispatched 1501 * to a single taskq at a time. It would be a grievous error 1502 * to dispatch the zio to another taskq at the same time. 1503 */ 1504#if defined(illumos) || !defined(_KERNEL) 1505 ASSERT(zio->io_tqent.tqent_next == NULL); 1506#else 1507 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0); 1508#endif 1509 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio, 1510 flags, &zio->io_tqent); 1511} 1512 1513static boolean_t 1514zio_taskq_member(zio_t *zio, zio_taskq_type_t q) 1515{ 1516 kthread_t *executor = zio->io_executor; 1517 spa_t *spa = zio->io_spa; 1518 1519 for (zio_type_t t = 0; t < ZIO_TYPES; t++) { 1520 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q]; 1521 uint_t i; 1522 for (i = 0; i < tqs->stqs_count; i++) { 1523 if (taskq_member(tqs->stqs_taskq[i], executor)) 1524 return (B_TRUE); 1525 } 1526 } 1527 1528 return (B_FALSE); 1529} 1530 1531static int 1532zio_issue_async(zio_t *zio) 1533{ 1534 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); 1535 1536 return (ZIO_PIPELINE_STOP); 1537} 1538 1539void 1540zio_interrupt(zio_t *zio) 1541{ 1542 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE); 1543} 1544 1545void 1546zio_delay_interrupt(zio_t *zio) 1547{ 1548 /* 1549 * The timeout_generic() function isn't defined in userspace, so 1550 * rather than trying to implement the function, the zio delay 1551 * functionality has been disabled for userspace builds. 1552 */ 1553 1554#ifdef _KERNEL 1555 /* 1556 * If io_target_timestamp is zero, then no delay has been registered 1557 * for this IO, thus jump to the end of this function and "skip" the 1558 * delay; issuing it directly to the zio layer. 1559 */ 1560 if (zio->io_target_timestamp != 0) { 1561 hrtime_t now = gethrtime(); 1562 1563 if (now >= zio->io_target_timestamp) { 1564 /* 1565 * This IO has already taken longer than the target 1566 * delay to complete, so we don't want to delay it 1567 * any longer; we "miss" the delay and issue it 1568 * directly to the zio layer. This is likely due to 1569 * the target latency being set to a value less than 1570 * the underlying hardware can satisfy (e.g. delay 1571 * set to 1ms, but the disks take 10ms to complete an 1572 * IO request). 1573 */ 1574 1575 DTRACE_PROBE2(zio__delay__miss, zio_t *, zio, 1576 hrtime_t, now); 1577 1578 zio_interrupt(zio); 1579 } else { 1580 hrtime_t diff = zio->io_target_timestamp - now; 1581 1582 DTRACE_PROBE3(zio__delay__hit, zio_t *, zio, 1583 hrtime_t, now, hrtime_t, diff); 1584 1585 (void) timeout_generic(CALLOUT_NORMAL, 1586 (void (*)(void *))zio_interrupt, zio, diff, 1, 0); 1587 } 1588 1589 return; 1590 } 1591#endif 1592 1593 DTRACE_PROBE1(zio__delay__skip, zio_t *, zio); 1594 zio_interrupt(zio); 1595} 1596 1597/* 1598 * Execute the I/O pipeline until one of the following occurs: 1599 * 1600 * (1) the I/O completes 1601 * (2) the pipeline stalls waiting for dependent child I/Os 1602 * (3) the I/O issues, so we're waiting for an I/O completion interrupt 1603 * (4) the I/O is delegated by vdev-level caching or aggregation 1604 * (5) the I/O is deferred due to vdev-level queueing 1605 * (6) the I/O is handed off to another thread. 1606 * 1607 * In all cases, the pipeline stops whenever there's no CPU work; it never 1608 * burns a thread in cv_wait(). 1609 * 1610 * There's no locking on io_stage because there's no legitimate way 1611 * for multiple threads to be attempting to process the same I/O. 1612 */ 1613static zio_pipe_stage_t *zio_pipeline[]; 1614 1615void 1616zio_execute(zio_t *zio) 1617{ 1618 zio->io_executor = curthread; 1619 1620 ASSERT3U(zio->io_queued_timestamp, >, 0); 1621 1622 while (zio->io_stage < ZIO_STAGE_DONE) { 1623 enum zio_stage pipeline = zio->io_pipeline; 1624 enum zio_stage stage = zio->io_stage; 1625 int rv; 1626 1627 ASSERT(!MUTEX_HELD(&zio->io_lock)); 1628 ASSERT(ISP2(stage)); 1629 ASSERT(zio->io_stall == NULL); 1630 1631 do { 1632 stage <<= 1; 1633 } while ((stage & pipeline) == 0); 1634 1635 ASSERT(stage <= ZIO_STAGE_DONE); 1636 1637 /* 1638 * If we are in interrupt context and this pipeline stage 1639 * will grab a config lock that is held across I/O, 1640 * or may wait for an I/O that needs an interrupt thread 1641 * to complete, issue async to avoid deadlock. 1642 * 1643 * For VDEV_IO_START, we cut in line so that the io will 1644 * be sent to disk promptly. 1645 */ 1646 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL && 1647 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) { 1648 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? 1649 zio_requeue_io_start_cut_in_line : B_FALSE; 1650 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); 1651 return; 1652 } 1653 1654 zio->io_stage = stage; 1655 zio->io_pipeline_trace |= zio->io_stage; 1656 rv = zio_pipeline[highbit64(stage) - 1](zio); 1657 1658 if (rv == ZIO_PIPELINE_STOP) 1659 return; 1660 1661 ASSERT(rv == ZIO_PIPELINE_CONTINUE); 1662 } 1663} 1664 1665/* 1666 * ========================================================================== 1667 * Initiate I/O, either sync or async 1668 * ========================================================================== 1669 */ 1670int 1671zio_wait(zio_t *zio) 1672{ 1673 int error; 1674 1675 ASSERT(zio->io_stage == ZIO_STAGE_OPEN); 1676 ASSERT(zio->io_executor == NULL); 1677 1678 zio->io_waiter = curthread; 1679 ASSERT0(zio->io_queued_timestamp); 1680 zio->io_queued_timestamp = gethrtime(); 1681 1682 zio_execute(zio); 1683 1684 mutex_enter(&zio->io_lock); 1685 while (zio->io_executor != NULL) 1686 cv_wait(&zio->io_cv, &zio->io_lock); 1687 mutex_exit(&zio->io_lock); 1688 1689 error = zio->io_error; 1690 zio_destroy(zio); 1691 1692 return (error); 1693} 1694 1695void 1696zio_nowait(zio_t *zio) 1697{ 1698 ASSERT(zio->io_executor == NULL); 1699 1700 if (zio->io_child_type == ZIO_CHILD_LOGICAL && 1701 zio_unique_parent(zio) == NULL) { 1702 /* 1703 * This is a logical async I/O with no parent to wait for it. 1704 * We add it to the spa_async_root_zio "Godfather" I/O which 1705 * will ensure they complete prior to unloading the pool. 1706 */ 1707 spa_t *spa = zio->io_spa; 1708 1709 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio); 1710 } 1711 1712 ASSERT0(zio->io_queued_timestamp); 1713 zio->io_queued_timestamp = gethrtime(); 1714 zio_execute(zio); 1715} 1716 1717/* 1718 * ========================================================================== 1719 * Reexecute or suspend/resume failed I/O 1720 * ========================================================================== 1721 */ 1722 1723static void 1724zio_reexecute(zio_t *pio) 1725{ 1726 zio_t *cio, *cio_next; 1727 1728 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL); 1729 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN); 1730 ASSERT(pio->io_gang_leader == NULL); 1731 ASSERT(pio->io_gang_tree == NULL); 1732 1733 pio->io_flags = pio->io_orig_flags; 1734 pio->io_stage = pio->io_orig_stage; 1735 pio->io_pipeline = pio->io_orig_pipeline; 1736 pio->io_reexecute = 0; 1737 pio->io_flags |= ZIO_FLAG_REEXECUTED; 1738 pio->io_pipeline_trace = 0; 1739 pio->io_error = 0; 1740 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 1741 pio->io_state[w] = 0; 1742 for (int c = 0; c < ZIO_CHILD_TYPES; c++) 1743 pio->io_child_error[c] = 0; 1744 1745 if (IO_IS_ALLOCATING(pio)) 1746 BP_ZERO(pio->io_bp); 1747 1748 /* 1749 * As we reexecute pio's children, new children could be created. 1750 * New children go to the head of pio's io_child_list, however, 1751 * so we will (correctly) not reexecute them. The key is that 1752 * the remainder of pio's io_child_list, from 'cio_next' onward, 1753 * cannot be affected by any side effects of reexecuting 'cio'. 1754 */ 1755 zio_link_t *zl = NULL; 1756 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) { 1757 cio_next = zio_walk_children(pio, &zl); 1758 mutex_enter(&pio->io_lock); 1759 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 1760 pio->io_children[cio->io_child_type][w]++; 1761 mutex_exit(&pio->io_lock); 1762 zio_reexecute(cio); 1763 } 1764 1765 /* 1766 * Now that all children have been reexecuted, execute the parent. 1767 * We don't reexecute "The Godfather" I/O here as it's the 1768 * responsibility of the caller to wait on him. 1769 */ 1770 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) { 1771 pio->io_queued_timestamp = gethrtime(); 1772 zio_execute(pio); 1773 } 1774} 1775 1776void 1777zio_suspend(spa_t *spa, zio_t *zio) 1778{ 1779 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC) 1780 fm_panic("Pool '%s' has encountered an uncorrectable I/O " 1781 "failure and the failure mode property for this pool " 1782 "is set to panic.", spa_name(spa)); 1783 1784 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0); 1785 1786 mutex_enter(&spa->spa_suspend_lock); 1787 1788 if (spa->spa_suspend_zio_root == NULL) 1789 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL, 1790 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | 1791 ZIO_FLAG_GODFATHER); 1792 1793 spa->spa_suspended = B_TRUE; 1794 1795 if (zio != NULL) { 1796 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); 1797 ASSERT(zio != spa->spa_suspend_zio_root); 1798 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1799 ASSERT(zio_unique_parent(zio) == NULL); 1800 ASSERT(zio->io_stage == ZIO_STAGE_DONE); 1801 zio_add_child(spa->spa_suspend_zio_root, zio); 1802 } 1803 1804 mutex_exit(&spa->spa_suspend_lock); 1805} 1806 1807int 1808zio_resume(spa_t *spa) 1809{ 1810 zio_t *pio; 1811 1812 /* 1813 * Reexecute all previously suspended i/o. 1814 */ 1815 mutex_enter(&spa->spa_suspend_lock); 1816 spa->spa_suspended = B_FALSE; 1817 cv_broadcast(&spa->spa_suspend_cv); 1818 pio = spa->spa_suspend_zio_root; 1819 spa->spa_suspend_zio_root = NULL; 1820 mutex_exit(&spa->spa_suspend_lock); 1821 1822 if (pio == NULL) 1823 return (0); 1824 1825 zio_reexecute(pio); 1826 return (zio_wait(pio)); 1827} 1828 1829void 1830zio_resume_wait(spa_t *spa) 1831{ 1832 mutex_enter(&spa->spa_suspend_lock); 1833 while (spa_suspended(spa)) 1834 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock); 1835 mutex_exit(&spa->spa_suspend_lock); 1836} 1837 1838/* 1839 * ========================================================================== 1840 * Gang blocks. 1841 * 1842 * A gang block is a collection of small blocks that looks to the DMU 1843 * like one large block. When zio_dva_allocate() cannot find a block 1844 * of the requested size, due to either severe fragmentation or the pool 1845 * being nearly full, it calls zio_write_gang_block() to construct the 1846 * block from smaller fragments. 1847 * 1848 * A gang block consists of a gang header (zio_gbh_phys_t) and up to 1849 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like 1850 * an indirect block: it's an array of block pointers. It consumes 1851 * only one sector and hence is allocatable regardless of fragmentation. 1852 * The gang header's bps point to its gang members, which hold the data. 1853 * 1854 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg> 1855 * as the verifier to ensure uniqueness of the SHA256 checksum. 1856 * Critically, the gang block bp's blk_cksum is the checksum of the data, 1857 * not the gang header. This ensures that data block signatures (needed for 1858 * deduplication) are independent of how the block is physically stored. 1859 * 1860 * Gang blocks can be nested: a gang member may itself be a gang block. 1861 * Thus every gang block is a tree in which root and all interior nodes are 1862 * gang headers, and the leaves are normal blocks that contain user data. 1863 * The root of the gang tree is called the gang leader. 1864 * 1865 * To perform any operation (read, rewrite, free, claim) on a gang block, 1866 * zio_gang_assemble() first assembles the gang tree (minus data leaves) 1867 * in the io_gang_tree field of the original logical i/o by recursively 1868 * reading the gang leader and all gang headers below it. This yields 1869 * an in-core tree containing the contents of every gang header and the 1870 * bps for every constituent of the gang block. 1871 * 1872 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree 1873 * and invokes a callback on each bp. To free a gang block, zio_gang_issue() 1874 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp. 1875 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim(). 1876 * zio_read_gang() is a wrapper around zio_read() that omits reading gang 1877 * headers, since we already have those in io_gang_tree. zio_rewrite_gang() 1878 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite() 1879 * of the gang header plus zio_checksum_compute() of the data to update the 1880 * gang header's blk_cksum as described above. 1881 * 1882 * The two-phase assemble/issue model solves the problem of partial failure -- 1883 * what if you'd freed part of a gang block but then couldn't read the 1884 * gang header for another part? Assembling the entire gang tree first 1885 * ensures that all the necessary gang header I/O has succeeded before 1886 * starting the actual work of free, claim, or write. Once the gang tree 1887 * is assembled, free and claim are in-memory operations that cannot fail. 1888 * 1889 * In the event that a gang write fails, zio_dva_unallocate() walks the 1890 * gang tree to immediately free (i.e. insert back into the space map) 1891 * everything we've allocated. This ensures that we don't get ENOSPC 1892 * errors during repeated suspend/resume cycles due to a flaky device. 1893 * 1894 * Gang rewrites only happen during sync-to-convergence. If we can't assemble 1895 * the gang tree, we won't modify the block, so we can safely defer the free 1896 * (knowing that the block is still intact). If we *can* assemble the gang 1897 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free 1898 * each constituent bp and we can allocate a new block on the next sync pass. 1899 * 1900 * In all cases, the gang tree allows complete recovery from partial failure. 1901 * ========================================================================== 1902 */ 1903 1904static zio_t * 1905zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1906{ 1907 if (gn != NULL) 1908 return (pio); 1909 1910 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp), 1911 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), 1912 &pio->io_bookmark)); 1913} 1914 1915zio_t * 1916zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1917{ 1918 zio_t *zio; 1919 1920 if (gn != NULL) { 1921 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, 1922 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority, 1923 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 1924 /* 1925 * As we rewrite each gang header, the pipeline will compute 1926 * a new gang block header checksum for it; but no one will 1927 * compute a new data checksum, so we do that here. The one 1928 * exception is the gang leader: the pipeline already computed 1929 * its data checksum because that stage precedes gang assembly. 1930 * (Presently, nothing actually uses interior data checksums; 1931 * this is just good hygiene.) 1932 */ 1933 if (gn != pio->io_gang_leader->io_gang_tree) { 1934 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp), 1935 data, BP_GET_PSIZE(bp)); 1936 } 1937 /* 1938 * If we are here to damage data for testing purposes, 1939 * leave the GBH alone so that we can detect the damage. 1940 */ 1941 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE) 1942 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; 1943 } else { 1944 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, 1945 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority, 1946 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 1947 } 1948 1949 return (zio); 1950} 1951 1952/* ARGSUSED */ 1953zio_t * 1954zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1955{ 1956 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp, 1957 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp), 1958 ZIO_GANG_CHILD_FLAGS(pio))); 1959} 1960 1961/* ARGSUSED */ 1962zio_t * 1963zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1964{ 1965 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp, 1966 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio))); 1967} 1968 1969static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = { 1970 NULL, 1971 zio_read_gang, 1972 zio_rewrite_gang, 1973 zio_free_gang, 1974 zio_claim_gang, 1975 NULL 1976}; 1977 1978static void zio_gang_tree_assemble_done(zio_t *zio); 1979 1980static zio_gang_node_t * 1981zio_gang_node_alloc(zio_gang_node_t **gnpp) 1982{ 1983 zio_gang_node_t *gn; 1984 1985 ASSERT(*gnpp == NULL); 1986 1987 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP); 1988 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE); 1989 *gnpp = gn; 1990 1991 return (gn); 1992} 1993 1994static void 1995zio_gang_node_free(zio_gang_node_t **gnpp) 1996{ 1997 zio_gang_node_t *gn = *gnpp; 1998 1999 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) 2000 ASSERT(gn->gn_child[g] == NULL); 2001 2002 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE); 2003 kmem_free(gn, sizeof (*gn)); 2004 *gnpp = NULL; 2005} 2006 2007static void 2008zio_gang_tree_free(zio_gang_node_t **gnpp) 2009{ 2010 zio_gang_node_t *gn = *gnpp; 2011 2012 if (gn == NULL) 2013 return; 2014 2015 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) 2016 zio_gang_tree_free(&gn->gn_child[g]); 2017 2018 zio_gang_node_free(gnpp); 2019} 2020 2021static void 2022zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp) 2023{ 2024 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp); 2025 2026 ASSERT(gio->io_gang_leader == gio); 2027 ASSERT(BP_IS_GANG(bp)); 2028 2029 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh, 2030 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn, 2031 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark)); 2032} 2033 2034static void 2035zio_gang_tree_assemble_done(zio_t *zio) 2036{ 2037 zio_t *gio = zio->io_gang_leader; 2038 zio_gang_node_t *gn = zio->io_private; 2039 blkptr_t *bp = zio->io_bp; 2040 2041 ASSERT(gio == zio_unique_parent(zio)); 2042 ASSERT(zio->io_child_count == 0); 2043 2044 if (zio->io_error) 2045 return; 2046 2047 if (BP_SHOULD_BYTESWAP(bp)) 2048 byteswap_uint64_array(zio->io_data, zio->io_size); 2049 2050 ASSERT(zio->io_data == gn->gn_gbh); 2051 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE); 2052 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); 2053 2054 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 2055 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; 2056 if (!BP_IS_GANG(gbp)) 2057 continue; 2058 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]); 2059 } 2060} 2061 2062static void 2063zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data) 2064{ 2065 zio_t *gio = pio->io_gang_leader; 2066 zio_t *zio; 2067 2068 ASSERT(BP_IS_GANG(bp) == !!gn); 2069 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp)); 2070 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree); 2071 2072 /* 2073 * If you're a gang header, your data is in gn->gn_gbh. 2074 * If you're a gang member, your data is in 'data' and gn == NULL. 2075 */ 2076 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data); 2077 2078 if (gn != NULL) { 2079 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); 2080 2081 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 2082 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; 2083 if (BP_IS_HOLE(gbp)) 2084 continue; 2085 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data); 2086 data = (char *)data + BP_GET_PSIZE(gbp); 2087 } 2088 } 2089 2090 if (gn == gio->io_gang_tree && gio->io_data != NULL) 2091 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data); 2092 2093 if (zio != pio) 2094 zio_nowait(zio); 2095} 2096 2097static int 2098zio_gang_assemble(zio_t *zio) 2099{ 2100 blkptr_t *bp = zio->io_bp; 2101 2102 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL); 2103 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 2104 2105 zio->io_gang_leader = zio; 2106 2107 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree); 2108 2109 return (ZIO_PIPELINE_CONTINUE); 2110} 2111 2112static int 2113zio_gang_issue(zio_t *zio) 2114{ 2115 blkptr_t *bp = zio->io_bp; 2116 2117 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE)) 2118 return (ZIO_PIPELINE_STOP); 2119 2120 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio); 2121 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 2122 2123 if (zio->io_child_error[ZIO_CHILD_GANG] == 0) 2124 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data); 2125 else 2126 zio_gang_tree_free(&zio->io_gang_tree); 2127 2128 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2129 2130 return (ZIO_PIPELINE_CONTINUE); 2131} 2132 2133static void 2134zio_write_gang_member_ready(zio_t *zio) 2135{ 2136 zio_t *pio = zio_unique_parent(zio); 2137 zio_t *gio = zio->io_gang_leader; 2138 dva_t *cdva = zio->io_bp->blk_dva; 2139 dva_t *pdva = pio->io_bp->blk_dva; 2140 uint64_t asize; 2141 2142 if (BP_IS_HOLE(zio->io_bp)) 2143 return; 2144 2145 ASSERT(BP_IS_HOLE(&zio->io_bp_orig)); 2146 2147 ASSERT(zio->io_child_type == ZIO_CHILD_GANG); 2148 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies); 2149 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp)); 2150 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp)); 2151 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp)); 2152 2153 mutex_enter(&pio->io_lock); 2154 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) { 2155 ASSERT(DVA_GET_GANG(&pdva[d])); 2156 asize = DVA_GET_ASIZE(&pdva[d]); 2157 asize += DVA_GET_ASIZE(&cdva[d]); 2158 DVA_SET_ASIZE(&pdva[d], asize); 2159 } 2160 mutex_exit(&pio->io_lock); 2161} 2162 2163static int 2164zio_write_gang_block(zio_t *pio) 2165{ 2166 spa_t *spa = pio->io_spa; 2167 metaslab_class_t *mc = spa_normal_class(spa); 2168 blkptr_t *bp = pio->io_bp; 2169 zio_t *gio = pio->io_gang_leader; 2170 zio_t *zio; 2171 zio_gang_node_t *gn, **gnpp; 2172 zio_gbh_phys_t *gbh; 2173 uint64_t txg = pio->io_txg; 2174 uint64_t resid = pio->io_size; 2175 uint64_t lsize; 2176 int copies = gio->io_prop.zp_copies; 2177 int gbh_copies = MIN(copies + 1, spa_max_replication(spa)); 2178 zio_prop_t zp; 2179 int error; 2180 2181 int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER; 2182 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { 2183 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); 2184 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA)); 2185 2186 flags |= METASLAB_ASYNC_ALLOC; 2187 VERIFY(refcount_held(&mc->mc_alloc_slots, pio)); 2188 2189 /* 2190 * The logical zio has already placed a reservation for 2191 * 'copies' allocation slots but gang blocks may require 2192 * additional copies. These additional copies 2193 * (i.e. gbh_copies - copies) are guaranteed to succeed 2194 * since metaslab_class_throttle_reserve() always allows 2195 * additional reservations for gang blocks. 2196 */ 2197 VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies, 2198 pio, flags)); 2199 } 2200 2201 error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE, 2202 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags, pio); 2203 if (error) { 2204 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { 2205 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); 2206 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA)); 2207 2208 /* 2209 * If we failed to allocate the gang block header then 2210 * we remove any additional allocation reservations that 2211 * we placed here. The original reservation will 2212 * be removed when the logical I/O goes to the ready 2213 * stage. 2214 */ 2215 metaslab_class_throttle_unreserve(mc, 2216 gbh_copies - copies, pio); 2217 } 2218 pio->io_error = error; 2219 return (ZIO_PIPELINE_CONTINUE); 2220 } 2221 2222 if (pio == gio) { 2223 gnpp = &gio->io_gang_tree; 2224 } else { 2225 gnpp = pio->io_private; 2226 ASSERT(pio->io_ready == zio_write_gang_member_ready); 2227 } 2228 2229 gn = zio_gang_node_alloc(gnpp); 2230 gbh = gn->gn_gbh; 2231 bzero(gbh, SPA_GANGBLOCKSIZE); 2232 2233 /* 2234 * Create the gang header. 2235 */ 2236 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL, 2237 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 2238 2239 /* 2240 * Create and nowait the gang children. 2241 */ 2242 for (int g = 0; resid != 0; resid -= lsize, g++) { 2243 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g), 2244 SPA_MINBLOCKSIZE); 2245 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid); 2246 2247 zp.zp_checksum = gio->io_prop.zp_checksum; 2248 zp.zp_compress = ZIO_COMPRESS_OFF; 2249 zp.zp_type = DMU_OT_NONE; 2250 zp.zp_level = 0; 2251 zp.zp_copies = gio->io_prop.zp_copies; 2252 zp.zp_dedup = B_FALSE; 2253 zp.zp_dedup_verify = B_FALSE; 2254 zp.zp_nopwrite = B_FALSE; 2255 2256 zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g], 2257 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp, 2258 zio_write_gang_member_ready, NULL, NULL, NULL, 2259 &gn->gn_child[g], pio->io_priority, 2260 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 2261 2262 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { 2263 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); 2264 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA)); 2265 2266 /* 2267 * Gang children won't throttle but we should 2268 * account for their work, so reserve an allocation 2269 * slot for them here. 2270 */ 2271 VERIFY(metaslab_class_throttle_reserve(mc, 2272 zp.zp_copies, cio, flags)); 2273 } 2274 zio_nowait(cio); 2275 } 2276 2277 /* 2278 * Set pio's pipeline to just wait for zio to finish. 2279 */ 2280 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2281 2282 zio_nowait(zio); 2283 2284 return (ZIO_PIPELINE_CONTINUE); 2285} 2286 2287/* 2288 * The zio_nop_write stage in the pipeline determines if allocating a 2289 * new bp is necessary. The nopwrite feature can handle writes in 2290 * either syncing or open context (i.e. zil writes) and as a result is 2291 * mutually exclusive with dedup. 2292 * 2293 * By leveraging a cryptographically secure checksum, such as SHA256, we 2294 * can compare the checksums of the new data and the old to determine if 2295 * allocating a new block is required. Note that our requirements for 2296 * cryptographic strength are fairly weak: there can't be any accidental 2297 * hash collisions, but we don't need to be secure against intentional 2298 * (malicious) collisions. To trigger a nopwrite, you have to be able 2299 * to write the file to begin with, and triggering an incorrect (hash 2300 * collision) nopwrite is no worse than simply writing to the file. 2301 * That said, there are no known attacks against the checksum algorithms 2302 * used for nopwrite, assuming that the salt and the checksums 2303 * themselves remain secret. 2304 */ 2305static int 2306zio_nop_write(zio_t *zio) 2307{ 2308 blkptr_t *bp = zio->io_bp; 2309 blkptr_t *bp_orig = &zio->io_bp_orig; 2310 zio_prop_t *zp = &zio->io_prop; 2311 2312 ASSERT(BP_GET_LEVEL(bp) == 0); 2313 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 2314 ASSERT(zp->zp_nopwrite); 2315 ASSERT(!zp->zp_dedup); 2316 ASSERT(zio->io_bp_override == NULL); 2317 ASSERT(IO_IS_ALLOCATING(zio)); 2318 2319 /* 2320 * Check to see if the original bp and the new bp have matching 2321 * characteristics (i.e. same checksum, compression algorithms, etc). 2322 * If they don't then just continue with the pipeline which will 2323 * allocate a new bp. 2324 */ 2325 if (BP_IS_HOLE(bp_orig) || 2326 !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags & 2327 ZCHECKSUM_FLAG_NOPWRITE) || 2328 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) || 2329 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) || 2330 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) || 2331 zp->zp_copies != BP_GET_NDVAS(bp_orig)) 2332 return (ZIO_PIPELINE_CONTINUE); 2333 2334 /* 2335 * If the checksums match then reset the pipeline so that we 2336 * avoid allocating a new bp and issuing any I/O. 2337 */ 2338 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) { 2339 ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags & 2340 ZCHECKSUM_FLAG_NOPWRITE); 2341 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig)); 2342 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig)); 2343 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF); 2344 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop, 2345 sizeof (uint64_t)) == 0); 2346 2347 *bp = *bp_orig; 2348 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2349 zio->io_flags |= ZIO_FLAG_NOPWRITE; 2350 } 2351 2352 return (ZIO_PIPELINE_CONTINUE); 2353} 2354 2355/* 2356 * ========================================================================== 2357 * Dedup 2358 * ========================================================================== 2359 */ 2360static void 2361zio_ddt_child_read_done(zio_t *zio) 2362{ 2363 blkptr_t *bp = zio->io_bp; 2364 ddt_entry_t *dde = zio->io_private; 2365 ddt_phys_t *ddp; 2366 zio_t *pio = zio_unique_parent(zio); 2367 2368 mutex_enter(&pio->io_lock); 2369 ddp = ddt_phys_select(dde, bp); 2370 if (zio->io_error == 0) 2371 ddt_phys_clear(ddp); /* this ddp doesn't need repair */ 2372 if (zio->io_error == 0 && dde->dde_repair_data == NULL) 2373 dde->dde_repair_data = zio->io_data; 2374 else 2375 zio_buf_free(zio->io_data, zio->io_size); 2376 mutex_exit(&pio->io_lock); 2377} 2378 2379static int 2380zio_ddt_read_start(zio_t *zio) 2381{ 2382 blkptr_t *bp = zio->io_bp; 2383 2384 ASSERT(BP_GET_DEDUP(bp)); 2385 ASSERT(BP_GET_PSIZE(bp) == zio->io_size); 2386 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2387 2388 if (zio->io_child_error[ZIO_CHILD_DDT]) { 2389 ddt_t *ddt = ddt_select(zio->io_spa, bp); 2390 ddt_entry_t *dde = ddt_repair_start(ddt, bp); 2391 ddt_phys_t *ddp = dde->dde_phys; 2392 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp); 2393 blkptr_t blk; 2394 2395 ASSERT(zio->io_vsd == NULL); 2396 zio->io_vsd = dde; 2397 2398 if (ddp_self == NULL) 2399 return (ZIO_PIPELINE_CONTINUE); 2400 2401 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) { 2402 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self) 2403 continue; 2404 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp, 2405 &blk); 2406 zio_nowait(zio_read(zio, zio->io_spa, &blk, 2407 zio_buf_alloc(zio->io_size), zio->io_size, 2408 zio_ddt_child_read_done, dde, zio->io_priority, 2409 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE, 2410 &zio->io_bookmark)); 2411 } 2412 return (ZIO_PIPELINE_CONTINUE); 2413 } 2414 2415 zio_nowait(zio_read(zio, zio->io_spa, bp, 2416 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority, 2417 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark)); 2418 2419 return (ZIO_PIPELINE_CONTINUE); 2420} 2421 2422static int 2423zio_ddt_read_done(zio_t *zio) 2424{ 2425 blkptr_t *bp = zio->io_bp; 2426 2427 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE)) 2428 return (ZIO_PIPELINE_STOP); 2429 2430 ASSERT(BP_GET_DEDUP(bp)); 2431 ASSERT(BP_GET_PSIZE(bp) == zio->io_size); 2432 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2433 2434 if (zio->io_child_error[ZIO_CHILD_DDT]) { 2435 ddt_t *ddt = ddt_select(zio->io_spa, bp); 2436 ddt_entry_t *dde = zio->io_vsd; 2437 if (ddt == NULL) { 2438 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE); 2439 return (ZIO_PIPELINE_CONTINUE); 2440 } 2441 if (dde == NULL) { 2442 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1; 2443 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); 2444 return (ZIO_PIPELINE_STOP); 2445 } 2446 if (dde->dde_repair_data != NULL) { 2447 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size); 2448 zio->io_child_error[ZIO_CHILD_DDT] = 0; 2449 } 2450 ddt_repair_done(ddt, dde); 2451 zio->io_vsd = NULL; 2452 } 2453 2454 ASSERT(zio->io_vsd == NULL); 2455 2456 return (ZIO_PIPELINE_CONTINUE); 2457} 2458 2459static boolean_t 2460zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde) 2461{ 2462 spa_t *spa = zio->io_spa; 2463 2464 /* 2465 * Note: we compare the original data, not the transformed data, 2466 * because when zio->io_bp is an override bp, we will not have 2467 * pushed the I/O transforms. That's an important optimization 2468 * because otherwise we'd compress/encrypt all dmu_sync() data twice. 2469 */ 2470 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { 2471 zio_t *lio = dde->dde_lead_zio[p]; 2472 2473 if (lio != NULL) { 2474 return (lio->io_orig_size != zio->io_orig_size || 2475 bcmp(zio->io_orig_data, lio->io_orig_data, 2476 zio->io_orig_size) != 0); 2477 } 2478 } 2479 2480 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { 2481 ddt_phys_t *ddp = &dde->dde_phys[p]; 2482 2483 if (ddp->ddp_phys_birth != 0) { 2484 arc_buf_t *abuf = NULL; 2485 arc_flags_t aflags = ARC_FLAG_WAIT; 2486 blkptr_t blk = *zio->io_bp; 2487 int error; 2488 2489 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth); 2490 2491 ddt_exit(ddt); 2492 2493 error = arc_read(NULL, spa, &blk, 2494 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ, 2495 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2496 &aflags, &zio->io_bookmark); 2497 2498 if (error == 0) { 2499 if (arc_buf_size(abuf) != zio->io_orig_size || 2500 bcmp(abuf->b_data, zio->io_orig_data, 2501 zio->io_orig_size) != 0) 2502 error = SET_ERROR(EEXIST); 2503 arc_buf_destroy(abuf, &abuf); 2504 } 2505 2506 ddt_enter(ddt); 2507 return (error != 0); 2508 } 2509 } 2510 2511 return (B_FALSE); 2512} 2513 2514static void 2515zio_ddt_child_write_ready(zio_t *zio) 2516{ 2517 int p = zio->io_prop.zp_copies; 2518 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); 2519 ddt_entry_t *dde = zio->io_private; 2520 ddt_phys_t *ddp = &dde->dde_phys[p]; 2521 zio_t *pio; 2522 2523 if (zio->io_error) 2524 return; 2525 2526 ddt_enter(ddt); 2527 2528 ASSERT(dde->dde_lead_zio[p] == zio); 2529 2530 ddt_phys_fill(ddp, zio->io_bp); 2531 2532 zio_link_t *zl = NULL; 2533 while ((pio = zio_walk_parents(zio, &zl)) != NULL) 2534 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg); 2535 2536 ddt_exit(ddt); 2537} 2538 2539static void 2540zio_ddt_child_write_done(zio_t *zio) 2541{ 2542 int p = zio->io_prop.zp_copies; 2543 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); 2544 ddt_entry_t *dde = zio->io_private; 2545 ddt_phys_t *ddp = &dde->dde_phys[p]; 2546 2547 ddt_enter(ddt); 2548 2549 ASSERT(ddp->ddp_refcnt == 0); 2550 ASSERT(dde->dde_lead_zio[p] == zio); 2551 dde->dde_lead_zio[p] = NULL; 2552 2553 if (zio->io_error == 0) { 2554 zio_link_t *zl = NULL; 2555 while (zio_walk_parents(zio, &zl) != NULL) 2556 ddt_phys_addref(ddp); 2557 } else { 2558 ddt_phys_clear(ddp); 2559 } 2560 2561 ddt_exit(ddt); 2562} 2563 2564static void 2565zio_ddt_ditto_write_done(zio_t *zio) 2566{ 2567 int p = DDT_PHYS_DITTO; 2568 zio_prop_t *zp = &zio->io_prop; 2569 blkptr_t *bp = zio->io_bp; 2570 ddt_t *ddt = ddt_select(zio->io_spa, bp); 2571 ddt_entry_t *dde = zio->io_private; 2572 ddt_phys_t *ddp = &dde->dde_phys[p]; 2573 ddt_key_t *ddk = &dde->dde_key; 2574 2575 ddt_enter(ddt); 2576 2577 ASSERT(ddp->ddp_refcnt == 0); 2578 ASSERT(dde->dde_lead_zio[p] == zio); 2579 dde->dde_lead_zio[p] = NULL; 2580 2581 if (zio->io_error == 0) { 2582 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum)); 2583 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP); 2584 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp)); 2585 if (ddp->ddp_phys_birth != 0) 2586 ddt_phys_free(ddt, ddk, ddp, zio->io_txg); 2587 ddt_phys_fill(ddp, bp); 2588 } 2589 2590 ddt_exit(ddt); 2591} 2592 2593static int 2594zio_ddt_write(zio_t *zio) 2595{ 2596 spa_t *spa = zio->io_spa; 2597 blkptr_t *bp = zio->io_bp; 2598 uint64_t txg = zio->io_txg; 2599 zio_prop_t *zp = &zio->io_prop; 2600 int p = zp->zp_copies; 2601 int ditto_copies; 2602 zio_t *cio = NULL; 2603 zio_t *dio = NULL; 2604 ddt_t *ddt = ddt_select(spa, bp); 2605 ddt_entry_t *dde; 2606 ddt_phys_t *ddp; 2607 2608 ASSERT(BP_GET_DEDUP(bp)); 2609 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum); 2610 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override); 2611 2612 ddt_enter(ddt); 2613 dde = ddt_lookup(ddt, bp, B_TRUE); 2614 ddp = &dde->dde_phys[p]; 2615 2616 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) { 2617 /* 2618 * If we're using a weak checksum, upgrade to a strong checksum 2619 * and try again. If we're already using a strong checksum, 2620 * we can't resolve it, so just convert to an ordinary write. 2621 * (And automatically e-mail a paper to Nature?) 2622 */ 2623 if (!(zio_checksum_table[zp->zp_checksum].ci_flags & 2624 ZCHECKSUM_FLAG_DEDUP)) { 2625 zp->zp_checksum = spa_dedup_checksum(spa); 2626 zio_pop_transforms(zio); 2627 zio->io_stage = ZIO_STAGE_OPEN; 2628 BP_ZERO(bp); 2629 } else { 2630 zp->zp_dedup = B_FALSE; 2631 } 2632 zio->io_pipeline = ZIO_WRITE_PIPELINE; 2633 ddt_exit(ddt); 2634 return (ZIO_PIPELINE_CONTINUE); 2635 } 2636 2637 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp); 2638 ASSERT(ditto_copies < SPA_DVAS_PER_BP); 2639 2640 if (ditto_copies > ddt_ditto_copies_present(dde) && 2641 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) { 2642 zio_prop_t czp = *zp; 2643 2644 czp.zp_copies = ditto_copies; 2645 2646 /* 2647 * If we arrived here with an override bp, we won't have run 2648 * the transform stack, so we won't have the data we need to 2649 * generate a child i/o. So, toss the override bp and restart. 2650 * This is safe, because using the override bp is just an 2651 * optimization; and it's rare, so the cost doesn't matter. 2652 */ 2653 if (zio->io_bp_override) { 2654 zio_pop_transforms(zio); 2655 zio->io_stage = ZIO_STAGE_OPEN; 2656 zio->io_pipeline = ZIO_WRITE_PIPELINE; 2657 zio->io_bp_override = NULL; 2658 BP_ZERO(bp); 2659 ddt_exit(ddt); 2660 return (ZIO_PIPELINE_CONTINUE); 2661 } 2662 2663 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data, 2664 zio->io_orig_size, &czp, NULL, NULL, 2665 NULL, zio_ddt_ditto_write_done, dde, zio->io_priority, 2666 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); 2667 2668 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL); 2669 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio; 2670 } 2671 2672 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) { 2673 if (ddp->ddp_phys_birth != 0) 2674 ddt_bp_fill(ddp, bp, txg); 2675 if (dde->dde_lead_zio[p] != NULL) 2676 zio_add_child(zio, dde->dde_lead_zio[p]); 2677 else 2678 ddt_phys_addref(ddp); 2679 } else if (zio->io_bp_override) { 2680 ASSERT(bp->blk_birth == txg); 2681 ASSERT(BP_EQUAL(bp, zio->io_bp_override)); 2682 ddt_phys_fill(ddp, bp); 2683 ddt_phys_addref(ddp); 2684 } else { 2685 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data, 2686 zio->io_orig_size, zp, 2687 zio_ddt_child_write_ready, NULL, NULL, 2688 zio_ddt_child_write_done, dde, zio->io_priority, 2689 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); 2690 2691 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL); 2692 dde->dde_lead_zio[p] = cio; 2693 } 2694 2695 ddt_exit(ddt); 2696 2697 if (cio) 2698 zio_nowait(cio); 2699 if (dio) 2700 zio_nowait(dio); 2701 2702 return (ZIO_PIPELINE_CONTINUE); 2703} 2704 2705ddt_entry_t *freedde; /* for debugging */ 2706 2707static int 2708zio_ddt_free(zio_t *zio) 2709{ 2710 spa_t *spa = zio->io_spa; 2711 blkptr_t *bp = zio->io_bp; 2712 ddt_t *ddt = ddt_select(spa, bp); 2713 ddt_entry_t *dde; 2714 ddt_phys_t *ddp; 2715 2716 ASSERT(BP_GET_DEDUP(bp)); 2717 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2718 2719 ddt_enter(ddt); 2720 freedde = dde = ddt_lookup(ddt, bp, B_TRUE); 2721 ddp = ddt_phys_select(dde, bp); 2722 ddt_phys_decref(ddp); 2723 ddt_exit(ddt); 2724 2725 return (ZIO_PIPELINE_CONTINUE); 2726} 2727 2728/* 2729 * ========================================================================== 2730 * Allocate and free blocks 2731 * ========================================================================== 2732 */ 2733 2734static zio_t * 2735zio_io_to_allocate(spa_t *spa) 2736{ 2737 zio_t *zio; 2738 2739 ASSERT(MUTEX_HELD(&spa->spa_alloc_lock)); 2740 2741 zio = avl_first(&spa->spa_alloc_tree); 2742 if (zio == NULL) 2743 return (NULL); 2744 2745 ASSERT(IO_IS_ALLOCATING(zio)); 2746 2747 /* 2748 * Try to place a reservation for this zio. If we're unable to 2749 * reserve then we throttle. 2750 */ 2751 if (!metaslab_class_throttle_reserve(spa_normal_class(spa), 2752 zio->io_prop.zp_copies, zio, 0)) { 2753 return (NULL); 2754 } 2755 2756 avl_remove(&spa->spa_alloc_tree, zio); 2757 ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE); 2758 2759 return (zio); 2760} 2761 2762static int 2763zio_dva_throttle(zio_t *zio) 2764{ 2765 spa_t *spa = zio->io_spa; 2766 zio_t *nio; 2767 2768 if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE || 2769 !spa_normal_class(zio->io_spa)->mc_alloc_throttle_enabled || 2770 zio->io_child_type == ZIO_CHILD_GANG || 2771 zio->io_flags & ZIO_FLAG_NODATA) { 2772 return (ZIO_PIPELINE_CONTINUE); 2773 } 2774 2775 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 2776 2777 ASSERT3U(zio->io_queued_timestamp, >, 0); 2778 ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE); 2779 2780 mutex_enter(&spa->spa_alloc_lock); 2781 2782 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 2783 avl_add(&spa->spa_alloc_tree, zio); 2784 2785 nio = zio_io_to_allocate(zio->io_spa); 2786 mutex_exit(&spa->spa_alloc_lock); 2787 2788 if (nio == zio) 2789 return (ZIO_PIPELINE_CONTINUE); 2790 2791 if (nio != NULL) { 2792 ASSERT3U(nio->io_queued_timestamp, <=, 2793 zio->io_queued_timestamp); 2794 ASSERT(nio->io_stage == ZIO_STAGE_DVA_THROTTLE); 2795 /* 2796 * We are passing control to a new zio so make sure that 2797 * it is processed by a different thread. We do this to 2798 * avoid stack overflows that can occur when parents are 2799 * throttled and children are making progress. We allow 2800 * it to go to the head of the taskq since it's already 2801 * been waiting. 2802 */ 2803 zio_taskq_dispatch(nio, ZIO_TASKQ_ISSUE, B_TRUE); 2804 } 2805 return (ZIO_PIPELINE_STOP); 2806} 2807 2808void 2809zio_allocate_dispatch(spa_t *spa) 2810{ 2811 zio_t *zio; 2812 2813 mutex_enter(&spa->spa_alloc_lock); 2814 zio = zio_io_to_allocate(spa); 2815 mutex_exit(&spa->spa_alloc_lock); 2816 if (zio == NULL) 2817 return; 2818 2819 ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE); 2820 ASSERT0(zio->io_error); 2821 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE); 2822} 2823 2824static int 2825zio_dva_allocate(zio_t *zio) 2826{ 2827 spa_t *spa = zio->io_spa; 2828 metaslab_class_t *mc = spa_normal_class(spa); 2829 blkptr_t *bp = zio->io_bp; 2830 int error; 2831 int flags = 0; 2832 2833 if (zio->io_gang_leader == NULL) { 2834 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 2835 zio->io_gang_leader = zio; 2836 } 2837 2838 ASSERT(BP_IS_HOLE(bp)); 2839 ASSERT0(BP_GET_NDVAS(bp)); 2840 ASSERT3U(zio->io_prop.zp_copies, >, 0); 2841 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa)); 2842 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp)); 2843 2844 if (zio->io_flags & ZIO_FLAG_NODATA) { 2845 flags |= METASLAB_DONT_THROTTLE; 2846 } 2847 if (zio->io_flags & ZIO_FLAG_GANG_CHILD) { 2848 flags |= METASLAB_GANG_CHILD; 2849 } 2850 if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE) { 2851 flags |= METASLAB_ASYNC_ALLOC; 2852 } 2853 2854 error = metaslab_alloc(spa, mc, zio->io_size, bp, 2855 zio->io_prop.zp_copies, zio->io_txg, NULL, flags, zio); 2856 2857 if (error != 0) { 2858 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, " 2859 "size %llu, error %d", spa_name(spa), zio, zio->io_size, 2860 error); 2861 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE) 2862 return (zio_write_gang_block(zio)); 2863 zio->io_error = error; 2864 } 2865 2866 return (ZIO_PIPELINE_CONTINUE); 2867} 2868 2869static int 2870zio_dva_free(zio_t *zio) 2871{ 2872 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE); 2873 2874 return (ZIO_PIPELINE_CONTINUE); 2875} 2876 2877static int 2878zio_dva_claim(zio_t *zio) 2879{ 2880 int error; 2881 2882 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg); 2883 if (error) 2884 zio->io_error = error; 2885 2886 return (ZIO_PIPELINE_CONTINUE); 2887} 2888 2889/* 2890 * Undo an allocation. This is used by zio_done() when an I/O fails 2891 * and we want to give back the block we just allocated. 2892 * This handles both normal blocks and gang blocks. 2893 */ 2894static void 2895zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp) 2896{ 2897 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp)); 2898 ASSERT(zio->io_bp_override == NULL); 2899 2900 if (!BP_IS_HOLE(bp)) 2901 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE); 2902 2903 if (gn != NULL) { 2904 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 2905 zio_dva_unallocate(zio, gn->gn_child[g], 2906 &gn->gn_gbh->zg_blkptr[g]); 2907 } 2908 } 2909} 2910 2911/* 2912 * Try to allocate an intent log block. Return 0 on success, errno on failure. 2913 */ 2914int 2915zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp, 2916 uint64_t size, boolean_t use_slog) 2917{ 2918 int error = 1; 2919 2920 ASSERT(txg > spa_syncing_txg(spa)); 2921 2922 if (use_slog) { 2923 error = metaslab_alloc(spa, spa_log_class(spa), size, 2924 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID, NULL); 2925 } 2926 2927 if (error) { 2928 error = metaslab_alloc(spa, spa_normal_class(spa), size, 2929 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID, NULL); 2930 } 2931 2932 if (error == 0) { 2933 BP_SET_LSIZE(new_bp, size); 2934 BP_SET_PSIZE(new_bp, size); 2935 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF); 2936 BP_SET_CHECKSUM(new_bp, 2937 spa_version(spa) >= SPA_VERSION_SLIM_ZIL 2938 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG); 2939 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); 2940 BP_SET_LEVEL(new_bp, 0); 2941 BP_SET_DEDUP(new_bp, 0); 2942 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER); 2943 } 2944 2945 return (error); 2946} 2947 2948/* 2949 * Free an intent log block. 2950 */ 2951void 2952zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp) 2953{ 2954 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG); 2955 ASSERT(!BP_IS_GANG(bp)); 2956 2957 zio_free(spa, txg, bp); 2958} 2959 2960/* 2961 * ========================================================================== 2962 * Read, write and delete to physical devices 2963 * ========================================================================== 2964 */ 2965 2966 2967/* 2968 * Issue an I/O to the underlying vdev. Typically the issue pipeline 2969 * stops after this stage and will resume upon I/O completion. 2970 * However, there are instances where the vdev layer may need to 2971 * continue the pipeline when an I/O was not issued. Since the I/O 2972 * that was sent to the vdev layer might be different than the one 2973 * currently active in the pipeline (see vdev_queue_io()), we explicitly 2974 * force the underlying vdev layers to call either zio_execute() or 2975 * zio_interrupt() to ensure that the pipeline continues with the correct I/O. 2976 */ 2977static int 2978zio_vdev_io_start(zio_t *zio) 2979{ 2980 vdev_t *vd = zio->io_vd; 2981 uint64_t align; 2982 spa_t *spa = zio->io_spa; 2983 int ret; 2984 2985 ASSERT(zio->io_error == 0); 2986 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0); 2987 2988 if (vd == NULL) { 2989 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) 2990 spa_config_enter(spa, SCL_ZIO, zio, RW_READER); 2991 2992 /* 2993 * The mirror_ops handle multiple DVAs in a single BP. 2994 */ 2995 vdev_mirror_ops.vdev_op_io_start(zio); 2996 return (ZIO_PIPELINE_STOP); 2997 } 2998 2999 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE && 3000 zio->io_priority == ZIO_PRIORITY_NOW) { 3001 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg); 3002 return (ZIO_PIPELINE_CONTINUE); 3003 } 3004 3005 ASSERT3P(zio->io_logical, !=, zio); 3006 3007 /* 3008 * We keep track of time-sensitive I/Os so that the scan thread 3009 * can quickly react to certain workloads. In particular, we care 3010 * about non-scrubbing, top-level reads and writes with the following 3011 * characteristics: 3012 * - synchronous writes of user data to non-slog devices 3013 * - any reads of user data 3014 * When these conditions are met, adjust the timestamp of spa_last_io 3015 * which allows the scan thread to adjust its workload accordingly. 3016 */ 3017 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL && 3018 vd == vd->vdev_top && !vd->vdev_islog && 3019 zio->io_bookmark.zb_objset != DMU_META_OBJSET && 3020 zio->io_txg != spa_syncing_txg(spa)) { 3021 uint64_t old = spa->spa_last_io; 3022 uint64_t new = ddi_get_lbolt64(); 3023 if (old != new) 3024 (void) atomic_cas_64(&spa->spa_last_io, old, new); 3025 } 3026 3027 align = 1ULL << vd->vdev_top->vdev_ashift; 3028 3029 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) && 3030 P2PHASE(zio->io_size, align) != 0) { 3031 /* Transform logical writes to be a full physical block size. */ 3032 uint64_t asize = P2ROUNDUP(zio->io_size, align); 3033 char *abuf = NULL; 3034 if (zio->io_type == ZIO_TYPE_READ || 3035 zio->io_type == ZIO_TYPE_WRITE) 3036 abuf = zio_buf_alloc(asize); 3037 ASSERT(vd == vd->vdev_top); 3038 if (zio->io_type == ZIO_TYPE_WRITE) { 3039 bcopy(zio->io_data, abuf, zio->io_size); 3040 bzero(abuf + zio->io_size, asize - zio->io_size); 3041 } 3042 zio_push_transform(zio, abuf, asize, abuf ? asize : 0, 3043 zio_subblock); 3044 } 3045 3046 /* 3047 * If this is not a physical io, make sure that it is properly aligned 3048 * before proceeding. 3049 */ 3050 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) { 3051 ASSERT0(P2PHASE(zio->io_offset, align)); 3052 ASSERT0(P2PHASE(zio->io_size, align)); 3053 } else { 3054 /* 3055 * For the physical io we allow alignment 3056 * to a logical block size. 3057 */ 3058 uint64_t log_align = 3059 1ULL << vd->vdev_top->vdev_logical_ashift; 3060 ASSERT0(P2PHASE(zio->io_offset, log_align)); 3061 ASSERT0(P2PHASE(zio->io_size, log_align)); 3062 } 3063 3064 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa)); 3065 3066 /* 3067 * If this is a repair I/O, and there's no self-healing involved -- 3068 * that is, we're just resilvering what we expect to resilver -- 3069 * then don't do the I/O unless zio's txg is actually in vd's DTL. 3070 * This prevents spurious resilvering with nested replication. 3071 * For example, given a mirror of mirrors, (A+B)+(C+D), if only 3072 * A is out of date, we'll read from C+D, then use the data to 3073 * resilver A+B -- but we don't actually want to resilver B, just A. 3074 * The top-level mirror has no way to know this, so instead we just 3075 * discard unnecessary repairs as we work our way down the vdev tree. 3076 * The same logic applies to any form of nested replication: 3077 * ditto + mirror, RAID-Z + replacing, etc. This covers them all. 3078 */ 3079 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) && 3080 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) && 3081 zio->io_txg != 0 && /* not a delegated i/o */ 3082 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) { 3083 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 3084 zio_vdev_io_bypass(zio); 3085 return (ZIO_PIPELINE_CONTINUE); 3086 } 3087 3088 if (vd->vdev_ops->vdev_op_leaf) { 3089 switch (zio->io_type) { 3090 case ZIO_TYPE_READ: 3091 if (vdev_cache_read(zio)) 3092 return (ZIO_PIPELINE_CONTINUE); 3093 /* FALLTHROUGH */ 3094 case ZIO_TYPE_WRITE: 3095 case ZIO_TYPE_FREE: 3096 if ((zio = vdev_queue_io(zio)) == NULL) 3097 return (ZIO_PIPELINE_STOP); 3098 3099 if (!vdev_accessible(vd, zio)) { 3100 zio->io_error = SET_ERROR(ENXIO); 3101 zio_interrupt(zio); 3102 return (ZIO_PIPELINE_STOP); 3103 } 3104 break; 3105 } 3106 /* 3107 * Note that we ignore repair writes for TRIM because they can 3108 * conflict with normal writes. This isn't an issue because, by 3109 * definition, we only repair blocks that aren't freed. 3110 */ 3111 if (zio->io_type == ZIO_TYPE_WRITE && 3112 !(zio->io_flags & ZIO_FLAG_IO_REPAIR) && 3113 !trim_map_write_start(zio)) 3114 return (ZIO_PIPELINE_STOP); 3115 } 3116 3117 vd->vdev_ops->vdev_op_io_start(zio); 3118 return (ZIO_PIPELINE_STOP); 3119} 3120 3121static int 3122zio_vdev_io_done(zio_t *zio) 3123{ 3124 vdev_t *vd = zio->io_vd; 3125 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops; 3126 boolean_t unexpected_error = B_FALSE; 3127 3128 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE)) 3129 return (ZIO_PIPELINE_STOP); 3130 3131 ASSERT(zio->io_type == ZIO_TYPE_READ || 3132 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE); 3133 3134 if (vd != NULL && vd->vdev_ops->vdev_op_leaf && 3135 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE || 3136 zio->io_type == ZIO_TYPE_FREE)) { 3137 3138 if (zio->io_type == ZIO_TYPE_WRITE && 3139 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) 3140 trim_map_write_done(zio); 3141 3142 vdev_queue_io_done(zio); 3143 3144 if (zio->io_type == ZIO_TYPE_WRITE) 3145 vdev_cache_write(zio); 3146 3147 if (zio_injection_enabled && zio->io_error == 0) 3148 zio->io_error = zio_handle_device_injection(vd, 3149 zio, EIO); 3150 3151 if (zio_injection_enabled && zio->io_error == 0) 3152 zio->io_error = zio_handle_label_injection(zio, EIO); 3153 3154 if (zio->io_error) { 3155 if (zio->io_error == ENOTSUP && 3156 zio->io_type == ZIO_TYPE_FREE) { 3157 /* Not all devices support TRIM. */ 3158 } else if (!vdev_accessible(vd, zio)) { 3159 zio->io_error = SET_ERROR(ENXIO); 3160 } else { 3161 unexpected_error = B_TRUE; 3162 } 3163 } 3164 } 3165 3166 ops->vdev_op_io_done(zio); 3167 3168 if (unexpected_error) 3169 VERIFY(vdev_probe(vd, zio) == NULL); 3170 3171 return (ZIO_PIPELINE_CONTINUE); 3172} 3173 3174/* 3175 * For non-raidz ZIOs, we can just copy aside the bad data read from the 3176 * disk, and use that to finish the checksum ereport later. 3177 */ 3178static void 3179zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr, 3180 const void *good_buf) 3181{ 3182 /* no processing needed */ 3183 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE); 3184} 3185 3186/*ARGSUSED*/ 3187void 3188zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored) 3189{ 3190 void *buf = zio_buf_alloc(zio->io_size); 3191 3192 bcopy(zio->io_data, buf, zio->io_size); 3193 3194 zcr->zcr_cbinfo = zio->io_size; 3195 zcr->zcr_cbdata = buf; 3196 zcr->zcr_finish = zio_vsd_default_cksum_finish; 3197 zcr->zcr_free = zio_buf_free; 3198} 3199 3200static int 3201zio_vdev_io_assess(zio_t *zio) 3202{ 3203 vdev_t *vd = zio->io_vd; 3204 3205 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE)) 3206 return (ZIO_PIPELINE_STOP); 3207 3208 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) 3209 spa_config_exit(zio->io_spa, SCL_ZIO, zio); 3210 3211 if (zio->io_vsd != NULL) { 3212 zio->io_vsd_ops->vsd_free(zio); 3213 zio->io_vsd = NULL; 3214 } 3215 3216 if (zio_injection_enabled && zio->io_error == 0) 3217 zio->io_error = zio_handle_fault_injection(zio, EIO); 3218 3219 if (zio->io_type == ZIO_TYPE_FREE && 3220 zio->io_priority != ZIO_PRIORITY_NOW) { 3221 switch (zio->io_error) { 3222 case 0: 3223 ZIO_TRIM_STAT_INCR(bytes, zio->io_size); 3224 ZIO_TRIM_STAT_BUMP(success); 3225 break; 3226 case EOPNOTSUPP: 3227 ZIO_TRIM_STAT_BUMP(unsupported); 3228 break; 3229 default: 3230 ZIO_TRIM_STAT_BUMP(failed); 3231 break; 3232 } 3233 } 3234 3235 /* 3236 * If the I/O failed, determine whether we should attempt to retry it. 3237 * 3238 * On retry, we cut in line in the issue queue, since we don't want 3239 * compression/checksumming/etc. work to prevent our (cheap) IO reissue. 3240 */ 3241 if (zio->io_error && vd == NULL && 3242 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) { 3243 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */ 3244 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */ 3245 zio->io_error = 0; 3246 zio->io_flags |= ZIO_FLAG_IO_RETRY | 3247 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE; 3248 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1; 3249 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, 3250 zio_requeue_io_start_cut_in_line); 3251 return (ZIO_PIPELINE_STOP); 3252 } 3253 3254 /* 3255 * If we got an error on a leaf device, convert it to ENXIO 3256 * if the device is not accessible at all. 3257 */ 3258 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf && 3259 !vdev_accessible(vd, zio)) 3260 zio->io_error = SET_ERROR(ENXIO); 3261 3262 /* 3263 * If we can't write to an interior vdev (mirror or RAID-Z), 3264 * set vdev_cant_write so that we stop trying to allocate from it. 3265 */ 3266 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE && 3267 vd != NULL && !vd->vdev_ops->vdev_op_leaf) { 3268 vd->vdev_cant_write = B_TRUE; 3269 } 3270 3271 if (zio->io_error) 3272 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 3273 3274 if (vd != NULL && vd->vdev_ops->vdev_op_leaf && 3275 zio->io_physdone != NULL) { 3276 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED)); 3277 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV); 3278 zio->io_physdone(zio->io_logical); 3279 } 3280 3281 return (ZIO_PIPELINE_CONTINUE); 3282} 3283 3284void 3285zio_vdev_io_reissue(zio_t *zio) 3286{ 3287 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); 3288 ASSERT(zio->io_error == 0); 3289 3290 zio->io_stage >>= 1; 3291} 3292 3293void 3294zio_vdev_io_redone(zio_t *zio) 3295{ 3296 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE); 3297 3298 zio->io_stage >>= 1; 3299} 3300 3301void 3302zio_vdev_io_bypass(zio_t *zio) 3303{ 3304 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); 3305 ASSERT(zio->io_error == 0); 3306 3307 zio->io_flags |= ZIO_FLAG_IO_BYPASS; 3308 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1; 3309} 3310 3311/* 3312 * ========================================================================== 3313 * Generate and verify checksums 3314 * ========================================================================== 3315 */ 3316static int 3317zio_checksum_generate(zio_t *zio) 3318{ 3319 blkptr_t *bp = zio->io_bp; 3320 enum zio_checksum checksum; 3321 3322 if (bp == NULL) { 3323 /* 3324 * This is zio_write_phys(). 3325 * We're either generating a label checksum, or none at all. 3326 */ 3327 checksum = zio->io_prop.zp_checksum; 3328 3329 if (checksum == ZIO_CHECKSUM_OFF) 3330 return (ZIO_PIPELINE_CONTINUE); 3331 3332 ASSERT(checksum == ZIO_CHECKSUM_LABEL); 3333 } else { 3334 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) { 3335 ASSERT(!IO_IS_ALLOCATING(zio)); 3336 checksum = ZIO_CHECKSUM_GANG_HEADER; 3337 } else { 3338 checksum = BP_GET_CHECKSUM(bp); 3339 } 3340 } 3341 3342 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size); 3343 3344 return (ZIO_PIPELINE_CONTINUE); 3345} 3346 3347static int 3348zio_checksum_verify(zio_t *zio) 3349{ 3350 zio_bad_cksum_t info; 3351 blkptr_t *bp = zio->io_bp; 3352 int error; 3353 3354 ASSERT(zio->io_vd != NULL); 3355 3356 if (bp == NULL) { 3357 /* 3358 * This is zio_read_phys(). 3359 * We're either verifying a label checksum, or nothing at all. 3360 */ 3361 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF) 3362 return (ZIO_PIPELINE_CONTINUE); 3363 3364 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL); 3365 } 3366 3367 if ((error = zio_checksum_error(zio, &info)) != 0) { 3368 zio->io_error = error; 3369 if (error == ECKSUM && 3370 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) { 3371 zfs_ereport_start_checksum(zio->io_spa, 3372 zio->io_vd, zio, zio->io_offset, 3373 zio->io_size, NULL, &info); 3374 } 3375 } 3376 3377 return (ZIO_PIPELINE_CONTINUE); 3378} 3379 3380/* 3381 * Called by RAID-Z to ensure we don't compute the checksum twice. 3382 */ 3383void 3384zio_checksum_verified(zio_t *zio) 3385{ 3386 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; 3387} 3388 3389/* 3390 * ========================================================================== 3391 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other. 3392 * An error of 0 indicates success. ENXIO indicates whole-device failure, 3393 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO 3394 * indicate errors that are specific to one I/O, and most likely permanent. 3395 * Any other error is presumed to be worse because we weren't expecting it. 3396 * ========================================================================== 3397 */ 3398int 3399zio_worst_error(int e1, int e2) 3400{ 3401 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO }; 3402 int r1, r2; 3403 3404 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++) 3405 if (e1 == zio_error_rank[r1]) 3406 break; 3407 3408 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++) 3409 if (e2 == zio_error_rank[r2]) 3410 break; 3411 3412 return (r1 > r2 ? e1 : e2); 3413} 3414 3415/* 3416 * ========================================================================== 3417 * I/O completion 3418 * ========================================================================== 3419 */ 3420static int 3421zio_ready(zio_t *zio) 3422{ 3423 blkptr_t *bp = zio->io_bp; 3424 zio_t *pio, *pio_next; 3425 zio_link_t *zl = NULL; 3426 3427 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) || 3428 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY)) 3429 return (ZIO_PIPELINE_STOP); 3430 3431 if (zio->io_ready) { 3432 ASSERT(IO_IS_ALLOCATING(zio)); 3433 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) || 3434 (zio->io_flags & ZIO_FLAG_NOPWRITE)); 3435 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0); 3436 3437 zio->io_ready(zio); 3438 } 3439 3440 if (bp != NULL && bp != &zio->io_bp_copy) 3441 zio->io_bp_copy = *bp; 3442 3443 if (zio->io_error != 0) { 3444 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 3445 3446 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) { 3447 ASSERT(IO_IS_ALLOCATING(zio)); 3448 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); 3449 /* 3450 * We were unable to allocate anything, unreserve and 3451 * issue the next I/O to allocate. 3452 */ 3453 metaslab_class_throttle_unreserve( 3454 spa_normal_class(zio->io_spa), 3455 zio->io_prop.zp_copies, zio); 3456 zio_allocate_dispatch(zio->io_spa); 3457 } 3458 } 3459 3460 mutex_enter(&zio->io_lock); 3461 zio->io_state[ZIO_WAIT_READY] = 1; 3462 pio = zio_walk_parents(zio, &zl); 3463 mutex_exit(&zio->io_lock); 3464 3465 /* 3466 * As we notify zio's parents, new parents could be added. 3467 * New parents go to the head of zio's io_parent_list, however, 3468 * so we will (correctly) not notify them. The remainder of zio's 3469 * io_parent_list, from 'pio_next' onward, cannot change because 3470 * all parents must wait for us to be done before they can be done. 3471 */ 3472 for (; pio != NULL; pio = pio_next) { 3473 pio_next = zio_walk_parents(zio, &zl); 3474 zio_notify_parent(pio, zio, ZIO_WAIT_READY); 3475 } 3476 3477 if (zio->io_flags & ZIO_FLAG_NODATA) { 3478 if (BP_IS_GANG(bp)) { 3479 zio->io_flags &= ~ZIO_FLAG_NODATA; 3480 } else { 3481 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE); 3482 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; 3483 } 3484 } 3485 3486 if (zio_injection_enabled && 3487 zio->io_spa->spa_syncing_txg == zio->io_txg) 3488 zio_handle_ignored_writes(zio); 3489 3490 return (ZIO_PIPELINE_CONTINUE); 3491} 3492 3493/* 3494 * Update the allocation throttle accounting. 3495 */ 3496static void 3497zio_dva_throttle_done(zio_t *zio) 3498{ 3499 zio_t *lio = zio->io_logical; 3500 zio_t *pio = zio_unique_parent(zio); 3501 vdev_t *vd = zio->io_vd; 3502 int flags = METASLAB_ASYNC_ALLOC; 3503 3504 ASSERT3P(zio->io_bp, !=, NULL); 3505 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE); 3506 ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE); 3507 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV); 3508 ASSERT(vd != NULL); 3509 ASSERT3P(vd, ==, vd->vdev_top); 3510 ASSERT(!(zio->io_flags & (ZIO_FLAG_IO_REPAIR | ZIO_FLAG_IO_RETRY))); 3511 ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING); 3512 ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE)); 3513 ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA)); 3514 3515 /* 3516 * Parents of gang children can have two flavors -- ones that 3517 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set) 3518 * and ones that allocated the constituent blocks. The allocation 3519 * throttle needs to know the allocating parent zio so we must find 3520 * it here. 3521 */ 3522 if (pio->io_child_type == ZIO_CHILD_GANG) { 3523 /* 3524 * If our parent is a rewrite gang child then our grandparent 3525 * would have been the one that performed the allocation. 3526 */ 3527 if (pio->io_flags & ZIO_FLAG_IO_REWRITE) 3528 pio = zio_unique_parent(pio); 3529 flags |= METASLAB_GANG_CHILD; 3530 } 3531 3532 ASSERT(IO_IS_ALLOCATING(pio)); 3533 ASSERT3P(zio, !=, zio->io_logical); 3534 ASSERT(zio->io_logical != NULL); 3535 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR)); 3536 ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE); 3537 3538 mutex_enter(&pio->io_lock); 3539 metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags); 3540 mutex_exit(&pio->io_lock); 3541 3542 metaslab_class_throttle_unreserve(spa_normal_class(zio->io_spa), 3543 1, pio); 3544 3545 /* 3546 * Call into the pipeline to see if there is more work that 3547 * needs to be done. If there is work to be done it will be 3548 * dispatched to another taskq thread. 3549 */ 3550 zio_allocate_dispatch(zio->io_spa); 3551} 3552 3553static int 3554zio_done(zio_t *zio) 3555{ 3556 spa_t *spa = zio->io_spa; 3557 zio_t *lio = zio->io_logical; 3558 blkptr_t *bp = zio->io_bp; 3559 vdev_t *vd = zio->io_vd; 3560 uint64_t psize = zio->io_size; 3561 zio_t *pio, *pio_next; 3562 metaslab_class_t *mc = spa_normal_class(spa); 3563 zio_link_t *zl = NULL; 3564 3565 /* 3566 * If our children haven't all completed, 3567 * wait for them and then repeat this pipeline stage. 3568 */ 3569 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) || 3570 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) || 3571 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) || 3572 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE)) 3573 return (ZIO_PIPELINE_STOP); 3574 3575 /* 3576 * If the allocation throttle is enabled, then update the accounting. 3577 * We only track child I/Os that are part of an allocating async 3578 * write. We must do this since the allocation is performed 3579 * by the logical I/O but the actual write is done by child I/Os. 3580 */ 3581 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING && 3582 zio->io_child_type == ZIO_CHILD_VDEV) { 3583 ASSERT(mc->mc_alloc_throttle_enabled); 3584 zio_dva_throttle_done(zio); 3585 } 3586 3587 /* 3588 * If the allocation throttle is enabled, verify that 3589 * we have decremented the refcounts for every I/O that was throttled. 3590 */ 3591 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) { 3592 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 3593 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); 3594 ASSERT(bp != NULL); 3595 metaslab_group_alloc_verify(spa, zio->io_bp, zio); 3596 VERIFY(refcount_not_held(&mc->mc_alloc_slots, zio)); 3597 } 3598 3599 for (int c = 0; c < ZIO_CHILD_TYPES; c++) 3600 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 3601 ASSERT(zio->io_children[c][w] == 0); 3602 3603 if (bp != NULL && !BP_IS_EMBEDDED(bp)) { 3604 ASSERT(bp->blk_pad[0] == 0); 3605 ASSERT(bp->blk_pad[1] == 0); 3606 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 || 3607 (bp == zio_unique_parent(zio)->io_bp)); 3608 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) && 3609 zio->io_bp_override == NULL && 3610 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) { 3611 ASSERT(!BP_SHOULD_BYTESWAP(bp)); 3612 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp)); 3613 ASSERT(BP_COUNT_GANG(bp) == 0 || 3614 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp))); 3615 } 3616 if (zio->io_flags & ZIO_FLAG_NOPWRITE) 3617 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig)); 3618 } 3619 3620 /* 3621 * If there were child vdev/gang/ddt errors, they apply to us now. 3622 */ 3623 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV); 3624 zio_inherit_child_errors(zio, ZIO_CHILD_GANG); 3625 zio_inherit_child_errors(zio, ZIO_CHILD_DDT); 3626 3627 /* 3628 * If the I/O on the transformed data was successful, generate any 3629 * checksum reports now while we still have the transformed data. 3630 */ 3631 if (zio->io_error == 0) { 3632 while (zio->io_cksum_report != NULL) { 3633 zio_cksum_report_t *zcr = zio->io_cksum_report; 3634 uint64_t align = zcr->zcr_align; 3635 uint64_t asize = P2ROUNDUP(psize, align); 3636 char *abuf = zio->io_data; 3637 3638 if (asize != psize) { 3639 abuf = zio_buf_alloc(asize); 3640 bcopy(zio->io_data, abuf, psize); 3641 bzero(abuf + psize, asize - psize); 3642 } 3643 3644 zio->io_cksum_report = zcr->zcr_next; 3645 zcr->zcr_next = NULL; 3646 zcr->zcr_finish(zcr, abuf); 3647 zfs_ereport_free_checksum(zcr); 3648 3649 if (asize != psize) 3650 zio_buf_free(abuf, asize); 3651 } 3652 } 3653 3654 zio_pop_transforms(zio); /* note: may set zio->io_error */ 3655 3656 vdev_stat_update(zio, psize); 3657 3658 if (zio->io_error) { 3659 /* 3660 * If this I/O is attached to a particular vdev, 3661 * generate an error message describing the I/O failure 3662 * at the block level. We ignore these errors if the 3663 * device is currently unavailable. 3664 */ 3665 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd)) 3666 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0); 3667 3668 if ((zio->io_error == EIO || !(zio->io_flags & 3669 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) && 3670 zio == lio) { 3671 /* 3672 * For logical I/O requests, tell the SPA to log the 3673 * error and generate a logical data ereport. 3674 */ 3675 spa_log_error(spa, zio); 3676 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio, 3677 0, 0); 3678 } 3679 } 3680 3681 if (zio->io_error && zio == lio) { 3682 /* 3683 * Determine whether zio should be reexecuted. This will 3684 * propagate all the way to the root via zio_notify_parent(). 3685 */ 3686 ASSERT(vd == NULL && bp != NULL); 3687 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 3688 3689 if (IO_IS_ALLOCATING(zio) && 3690 !(zio->io_flags & ZIO_FLAG_CANFAIL)) { 3691 if (zio->io_error != ENOSPC) 3692 zio->io_reexecute |= ZIO_REEXECUTE_NOW; 3693 else 3694 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 3695 } 3696 3697 if ((zio->io_type == ZIO_TYPE_READ || 3698 zio->io_type == ZIO_TYPE_FREE) && 3699 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && 3700 zio->io_error == ENXIO && 3701 spa_load_state(spa) == SPA_LOAD_NONE && 3702 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE) 3703 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 3704 3705 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute) 3706 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 3707 3708 /* 3709 * Here is a possibly good place to attempt to do 3710 * either combinatorial reconstruction or error correction 3711 * based on checksums. It also might be a good place 3712 * to send out preliminary ereports before we suspend 3713 * processing. 3714 */ 3715 } 3716 3717 /* 3718 * If there were logical child errors, they apply to us now. 3719 * We defer this until now to avoid conflating logical child 3720 * errors with errors that happened to the zio itself when 3721 * updating vdev stats and reporting FMA events above. 3722 */ 3723 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL); 3724 3725 if ((zio->io_error || zio->io_reexecute) && 3726 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio && 3727 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE))) 3728 zio_dva_unallocate(zio, zio->io_gang_tree, bp); 3729 3730 zio_gang_tree_free(&zio->io_gang_tree); 3731 3732 /* 3733 * Godfather I/Os should never suspend. 3734 */ 3735 if ((zio->io_flags & ZIO_FLAG_GODFATHER) && 3736 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) 3737 zio->io_reexecute = 0; 3738 3739 if (zio->io_reexecute) { 3740 /* 3741 * This is a logical I/O that wants to reexecute. 3742 * 3743 * Reexecute is top-down. When an i/o fails, if it's not 3744 * the root, it simply notifies its parent and sticks around. 3745 * The parent, seeing that it still has children in zio_done(), 3746 * does the same. This percolates all the way up to the root. 3747 * The root i/o will reexecute or suspend the entire tree. 3748 * 3749 * This approach ensures that zio_reexecute() honors 3750 * all the original i/o dependency relationships, e.g. 3751 * parents not executing until children are ready. 3752 */ 3753 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 3754 3755 zio->io_gang_leader = NULL; 3756 3757 mutex_enter(&zio->io_lock); 3758 zio->io_state[ZIO_WAIT_DONE] = 1; 3759 mutex_exit(&zio->io_lock); 3760 3761 /* 3762 * "The Godfather" I/O monitors its children but is 3763 * not a true parent to them. It will track them through 3764 * the pipeline but severs its ties whenever they get into 3765 * trouble (e.g. suspended). This allows "The Godfather" 3766 * I/O to return status without blocking. 3767 */ 3768 zl = NULL; 3769 for (pio = zio_walk_parents(zio, &zl); pio != NULL; 3770 pio = pio_next) { 3771 zio_link_t *remove_zl = zl; 3772 pio_next = zio_walk_parents(zio, &zl); 3773 3774 if ((pio->io_flags & ZIO_FLAG_GODFATHER) && 3775 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) { 3776 zio_remove_child(pio, zio, remove_zl); 3777 zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 3778 } 3779 } 3780 3781 if ((pio = zio_unique_parent(zio)) != NULL) { 3782 /* 3783 * We're not a root i/o, so there's nothing to do 3784 * but notify our parent. Don't propagate errors 3785 * upward since we haven't permanently failed yet. 3786 */ 3787 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); 3788 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE; 3789 zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 3790 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) { 3791 /* 3792 * We'd fail again if we reexecuted now, so suspend 3793 * until conditions improve (e.g. device comes online). 3794 */ 3795 zio_suspend(spa, zio); 3796 } else { 3797 /* 3798 * Reexecution is potentially a huge amount of work. 3799 * Hand it off to the otherwise-unused claim taskq. 3800 */ 3801#if defined(illumos) || !defined(_KERNEL) 3802 ASSERT(zio->io_tqent.tqent_next == NULL); 3803#else 3804 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0); 3805#endif 3806 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM, 3807 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio, 3808 0, &zio->io_tqent); 3809 } 3810 return (ZIO_PIPELINE_STOP); 3811 } 3812 3813 ASSERT(zio->io_child_count == 0); 3814 ASSERT(zio->io_reexecute == 0); 3815 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL)); 3816 3817 /* 3818 * Report any checksum errors, since the I/O is complete. 3819 */ 3820 while (zio->io_cksum_report != NULL) { 3821 zio_cksum_report_t *zcr = zio->io_cksum_report; 3822 zio->io_cksum_report = zcr->zcr_next; 3823 zcr->zcr_next = NULL; 3824 zcr->zcr_finish(zcr, NULL); 3825 zfs_ereport_free_checksum(zcr); 3826 } 3827 3828 /* 3829 * It is the responsibility of the done callback to ensure that this 3830 * particular zio is no longer discoverable for adoption, and as 3831 * such, cannot acquire any new parents. 3832 */ 3833 if (zio->io_done) 3834 zio->io_done(zio); 3835 3836 mutex_enter(&zio->io_lock); 3837 zio->io_state[ZIO_WAIT_DONE] = 1; 3838 mutex_exit(&zio->io_lock); 3839 3840 zl = NULL; 3841 for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) { 3842 zio_link_t *remove_zl = zl; 3843 pio_next = zio_walk_parents(zio, &zl); 3844 zio_remove_child(pio, zio, remove_zl); 3845 zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 3846 } 3847 3848 if (zio->io_waiter != NULL) { 3849 mutex_enter(&zio->io_lock); 3850 zio->io_executor = NULL; 3851 cv_broadcast(&zio->io_cv); 3852 mutex_exit(&zio->io_lock); 3853 } else { 3854 zio_destroy(zio); 3855 } 3856 3857 return (ZIO_PIPELINE_STOP); 3858} 3859 3860/* 3861 * ========================================================================== 3862 * I/O pipeline definition 3863 * ========================================================================== 3864 */ 3865static zio_pipe_stage_t *zio_pipeline[] = { 3866 NULL, 3867 zio_read_bp_init, 3868 zio_write_bp_init, 3869 zio_free_bp_init, 3870 zio_issue_async, 3871 zio_write_compress, 3872 zio_checksum_generate, 3873 zio_nop_write, 3874 zio_ddt_read_start, 3875 zio_ddt_read_done, 3876 zio_ddt_write, 3877 zio_ddt_free, 3878 zio_gang_assemble, 3879 zio_gang_issue, 3880 zio_dva_throttle, 3881 zio_dva_allocate, 3882 zio_dva_free, 3883 zio_dva_claim, 3884 zio_ready, 3885 zio_vdev_io_start, 3886 zio_vdev_io_done, 3887 zio_vdev_io_assess, 3888 zio_checksum_verify, 3889 zio_done 3890}; 3891 3892 3893 3894 3895/* 3896 * Compare two zbookmark_phys_t's to see which we would reach first in a 3897 * pre-order traversal of the object tree. 3898 * 3899 * This is simple in every case aside from the meta-dnode object. For all other 3900 * objects, we traverse them in order (object 1 before object 2, and so on). 3901 * However, all of these objects are traversed while traversing object 0, since 3902 * the data it points to is the list of objects. Thus, we need to convert to a 3903 * canonical representation so we can compare meta-dnode bookmarks to 3904 * non-meta-dnode bookmarks. 3905 * 3906 * We do this by calculating "equivalents" for each field of the zbookmark. 3907 * zbookmarks outside of the meta-dnode use their own object and level, and 3908 * calculate the level 0 equivalent (the first L0 blkid that is contained in the 3909 * blocks this bookmark refers to) by multiplying their blkid by their span 3910 * (the number of L0 blocks contained within one block at their level). 3911 * zbookmarks inside the meta-dnode calculate their object equivalent 3912 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use 3913 * level + 1<<31 (any value larger than a level could ever be) for their level. 3914 * This causes them to always compare before a bookmark in their object 3915 * equivalent, compare appropriately to bookmarks in other objects, and to 3916 * compare appropriately to other bookmarks in the meta-dnode. 3917 */ 3918int 3919zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2, 3920 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2) 3921{ 3922 /* 3923 * These variables represent the "equivalent" values for the zbookmark, 3924 * after converting zbookmarks inside the meta dnode to their 3925 * normal-object equivalents. 3926 */ 3927 uint64_t zb1obj, zb2obj; 3928 uint64_t zb1L0, zb2L0; 3929 uint64_t zb1level, zb2level; 3930 3931 if (zb1->zb_object == zb2->zb_object && 3932 zb1->zb_level == zb2->zb_level && 3933 zb1->zb_blkid == zb2->zb_blkid) 3934 return (0); 3935 3936 /* 3937 * BP_SPANB calculates the span in blocks. 3938 */ 3939 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level); 3940 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level); 3941 3942 if (zb1->zb_object == DMU_META_DNODE_OBJECT) { 3943 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT)); 3944 zb1L0 = 0; 3945 zb1level = zb1->zb_level + COMPARE_META_LEVEL; 3946 } else { 3947 zb1obj = zb1->zb_object; 3948 zb1level = zb1->zb_level; 3949 } 3950 3951 if (zb2->zb_object == DMU_META_DNODE_OBJECT) { 3952 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT)); 3953 zb2L0 = 0; 3954 zb2level = zb2->zb_level + COMPARE_META_LEVEL; 3955 } else { 3956 zb2obj = zb2->zb_object; 3957 zb2level = zb2->zb_level; 3958 } 3959 3960 /* Now that we have a canonical representation, do the comparison. */ 3961 if (zb1obj != zb2obj) 3962 return (zb1obj < zb2obj ? -1 : 1); 3963 else if (zb1L0 != zb2L0) 3964 return (zb1L0 < zb2L0 ? -1 : 1); 3965 else if (zb1level != zb2level) 3966 return (zb1level > zb2level ? -1 : 1); 3967 /* 3968 * This can (theoretically) happen if the bookmarks have the same object 3969 * and level, but different blkids, if the block sizes are not the same. 3970 * There is presently no way to change the indirect block sizes 3971 */ 3972 return (0); 3973} 3974 3975/* 3976 * This function checks the following: given that last_block is the place that 3977 * our traversal stopped last time, does that guarantee that we've visited 3978 * every node under subtree_root? Therefore, we can't just use the raw output 3979 * of zbookmark_compare. We have to pass in a modified version of 3980 * subtree_root; by incrementing the block id, and then checking whether 3981 * last_block is before or equal to that, we can tell whether or not having 3982 * visited last_block implies that all of subtree_root's children have been 3983 * visited. 3984 */ 3985boolean_t 3986zbookmark_subtree_completed(const dnode_phys_t *dnp, 3987 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block) 3988{ 3989 zbookmark_phys_t mod_zb = *subtree_root; 3990 mod_zb.zb_blkid++; 3991 ASSERT(last_block->zb_level == 0); 3992 3993 /* The objset_phys_t isn't before anything. */ 3994 if (dnp == NULL) 3995 return (B_FALSE); 3996 3997 /* 3998 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the 3999 * data block size in sectors, because that variable is only used if 4000 * the bookmark refers to a block in the meta-dnode. Since we don't 4001 * know without examining it what object it refers to, and there's no 4002 * harm in passing in this value in other cases, we always pass it in. 4003 * 4004 * We pass in 0 for the indirect block size shift because zb2 must be 4005 * level 0. The indirect block size is only used to calculate the span 4006 * of the bookmark, but since the bookmark must be level 0, the span is 4007 * always 1, so the math works out. 4008 * 4009 * If you make changes to how the zbookmark_compare code works, be sure 4010 * to make sure that this code still works afterwards. 4011 */ 4012 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift, 4013 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb, 4014 last_block) <= 0); 4015} 4016