zio.c revision 277575
1/* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21/* 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23 * Copyright (c) 2011, 2014 by Delphix. All rights reserved. 24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved. 25 */ 26 27#include <sys/sysmacros.h> 28#include <sys/zfs_context.h> 29#include <sys/fm/fs/zfs.h> 30#include <sys/spa.h> 31#include <sys/txg.h> 32#include <sys/spa_impl.h> 33#include <sys/vdev_impl.h> 34#include <sys/zio_impl.h> 35#include <sys/zio_compress.h> 36#include <sys/zio_checksum.h> 37#include <sys/dmu_objset.h> 38#include <sys/arc.h> 39#include <sys/ddt.h> 40#include <sys/trim_map.h> 41#include <sys/blkptr.h> 42#include <sys/zfeature.h> 43 44SYSCTL_DECL(_vfs_zfs); 45SYSCTL_NODE(_vfs_zfs, OID_AUTO, zio, CTLFLAG_RW, 0, "ZFS ZIO"); 46#if defined(__amd64__) 47static int zio_use_uma = 1; 48#else 49static int zio_use_uma = 0; 50#endif 51TUNABLE_INT("vfs.zfs.zio.use_uma", &zio_use_uma); 52SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, use_uma, CTLFLAG_RDTUN, &zio_use_uma, 0, 53 "Use uma(9) for ZIO allocations"); 54static int zio_exclude_metadata = 0; 55TUNABLE_INT("vfs.zfs.zio.exclude_metadata", &zio_exclude_metadata); 56SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, exclude_metadata, CTLFLAG_RDTUN, &zio_exclude_metadata, 0, 57 "Exclude metadata buffers from dumps as well"); 58 59zio_trim_stats_t zio_trim_stats = { 60 { "bytes", KSTAT_DATA_UINT64, 61 "Number of bytes successfully TRIMmed" }, 62 { "success", KSTAT_DATA_UINT64, 63 "Number of successful TRIM requests" }, 64 { "unsupported", KSTAT_DATA_UINT64, 65 "Number of TRIM requests that failed because TRIM is not supported" }, 66 { "failed", KSTAT_DATA_UINT64, 67 "Number of TRIM requests that failed for reasons other than not supported" }, 68}; 69 70static kstat_t *zio_trim_ksp; 71 72/* 73 * ========================================================================== 74 * I/O type descriptions 75 * ========================================================================== 76 */ 77const char *zio_type_name[ZIO_TYPES] = { 78 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim", 79 "zio_ioctl" 80}; 81 82/* 83 * ========================================================================== 84 * I/O kmem caches 85 * ========================================================================== 86 */ 87kmem_cache_t *zio_cache; 88kmem_cache_t *zio_link_cache; 89kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; 90kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; 91 92#ifdef _KERNEL 93extern vmem_t *zio_alloc_arena; 94#endif 95 96/* 97 * The following actions directly effect the spa's sync-to-convergence logic. 98 * The values below define the sync pass when we start performing the action. 99 * Care should be taken when changing these values as they directly impact 100 * spa_sync() performance. Tuning these values may introduce subtle performance 101 * pathologies and should only be done in the context of performance analysis. 102 * These tunables will eventually be removed and replaced with #defines once 103 * enough analysis has been done to determine optimal values. 104 * 105 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that 106 * regular blocks are not deferred. 107 */ 108int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */ 109TUNABLE_INT("vfs.zfs.sync_pass_deferred_free", &zfs_sync_pass_deferred_free); 110SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_deferred_free, CTLFLAG_RDTUN, 111 &zfs_sync_pass_deferred_free, 0, "defer frees starting in this pass"); 112int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */ 113TUNABLE_INT("vfs.zfs.sync_pass_dont_compress", &zfs_sync_pass_dont_compress); 114SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_dont_compress, CTLFLAG_RDTUN, 115 &zfs_sync_pass_dont_compress, 0, "don't compress starting in this pass"); 116int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */ 117TUNABLE_INT("vfs.zfs.sync_pass_rewrite", &zfs_sync_pass_rewrite); 118SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_rewrite, CTLFLAG_RDTUN, 119 &zfs_sync_pass_rewrite, 0, "rewrite new bps starting in this pass"); 120 121/* 122 * An allocating zio is one that either currently has the DVA allocate 123 * stage set or will have it later in its lifetime. 124 */ 125#define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE) 126 127boolean_t zio_requeue_io_start_cut_in_line = B_TRUE; 128 129#ifdef ZFS_DEBUG 130int zio_buf_debug_limit = 16384; 131#else 132int zio_buf_debug_limit = 0; 133#endif 134 135void 136zio_init(void) 137{ 138 size_t c; 139 zio_cache = kmem_cache_create("zio_cache", 140 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0); 141 zio_link_cache = kmem_cache_create("zio_link_cache", 142 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0); 143 if (!zio_use_uma) 144 goto out; 145 146 /* 147 * For small buffers, we want a cache for each multiple of 148 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache 149 * for each quarter-power of 2. 150 */ 151 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { 152 size_t size = (c + 1) << SPA_MINBLOCKSHIFT; 153 size_t p2 = size; 154 size_t align = 0; 155 size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0; 156 157 while (!ISP2(p2)) 158 p2 &= p2 - 1; 159 160#ifdef illumos 161#ifndef _KERNEL 162 /* 163 * If we are using watchpoints, put each buffer on its own page, 164 * to eliminate the performance overhead of trapping to the 165 * kernel when modifying a non-watched buffer that shares the 166 * page with a watched buffer. 167 */ 168 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE)) 169 continue; 170#endif 171#endif /* illumos */ 172 if (size <= 4 * SPA_MINBLOCKSIZE) { 173 align = SPA_MINBLOCKSIZE; 174 } else if (IS_P2ALIGNED(size, p2 >> 2)) { 175 align = MIN(p2 >> 2, PAGESIZE); 176 } 177 178 if (align != 0) { 179 char name[36]; 180 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size); 181 zio_buf_cache[c] = kmem_cache_create(name, size, 182 align, NULL, NULL, NULL, NULL, NULL, cflags); 183 184 /* 185 * Since zio_data bufs do not appear in crash dumps, we 186 * pass KMC_NOTOUCH so that no allocator metadata is 187 * stored with the buffers. 188 */ 189 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size); 190 zio_data_buf_cache[c] = kmem_cache_create(name, size, 191 align, NULL, NULL, NULL, NULL, NULL, 192 cflags | KMC_NOTOUCH | KMC_NODEBUG); 193 } 194 } 195 196 while (--c != 0) { 197 ASSERT(zio_buf_cache[c] != NULL); 198 if (zio_buf_cache[c - 1] == NULL) 199 zio_buf_cache[c - 1] = zio_buf_cache[c]; 200 201 ASSERT(zio_data_buf_cache[c] != NULL); 202 if (zio_data_buf_cache[c - 1] == NULL) 203 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c]; 204 } 205out: 206 207 zio_inject_init(); 208 209 zio_trim_ksp = kstat_create("zfs", 0, "zio_trim", "misc", 210 KSTAT_TYPE_NAMED, 211 sizeof(zio_trim_stats) / sizeof(kstat_named_t), 212 KSTAT_FLAG_VIRTUAL); 213 214 if (zio_trim_ksp != NULL) { 215 zio_trim_ksp->ks_data = &zio_trim_stats; 216 kstat_install(zio_trim_ksp); 217 } 218} 219 220void 221zio_fini(void) 222{ 223 size_t c; 224 kmem_cache_t *last_cache = NULL; 225 kmem_cache_t *last_data_cache = NULL; 226 227 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { 228 if (zio_buf_cache[c] != last_cache) { 229 last_cache = zio_buf_cache[c]; 230 kmem_cache_destroy(zio_buf_cache[c]); 231 } 232 zio_buf_cache[c] = NULL; 233 234 if (zio_data_buf_cache[c] != last_data_cache) { 235 last_data_cache = zio_data_buf_cache[c]; 236 kmem_cache_destroy(zio_data_buf_cache[c]); 237 } 238 zio_data_buf_cache[c] = NULL; 239 } 240 241 kmem_cache_destroy(zio_link_cache); 242 kmem_cache_destroy(zio_cache); 243 244 zio_inject_fini(); 245 246 if (zio_trim_ksp != NULL) { 247 kstat_delete(zio_trim_ksp); 248 zio_trim_ksp = NULL; 249 } 250} 251 252/* 253 * ========================================================================== 254 * Allocate and free I/O buffers 255 * ========================================================================== 256 */ 257 258/* 259 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a 260 * crashdump if the kernel panics, so use it judiciously. Obviously, it's 261 * useful to inspect ZFS metadata, but if possible, we should avoid keeping 262 * excess / transient data in-core during a crashdump. 263 */ 264void * 265zio_buf_alloc(size_t size) 266{ 267 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 268 int flags = zio_exclude_metadata ? KM_NODEBUG : 0; 269 270 ASSERT3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 271 272 if (zio_use_uma) 273 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE)); 274 else 275 return (kmem_alloc(size, KM_SLEEP|flags)); 276} 277 278/* 279 * Use zio_data_buf_alloc to allocate data. The data will not appear in a 280 * crashdump if the kernel panics. This exists so that we will limit the amount 281 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount 282 * of kernel heap dumped to disk when the kernel panics) 283 */ 284void * 285zio_data_buf_alloc(size_t size) 286{ 287 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 288 289 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 290 291 if (zio_use_uma) 292 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE)); 293 else 294 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG)); 295} 296 297void 298zio_buf_free(void *buf, size_t size) 299{ 300 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 301 302 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 303 304 if (zio_use_uma) 305 kmem_cache_free(zio_buf_cache[c], buf); 306 else 307 kmem_free(buf, size); 308} 309 310void 311zio_data_buf_free(void *buf, size_t size) 312{ 313 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 314 315 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 316 317 if (zio_use_uma) 318 kmem_cache_free(zio_data_buf_cache[c], buf); 319 else 320 kmem_free(buf, size); 321} 322 323/* 324 * ========================================================================== 325 * Push and pop I/O transform buffers 326 * ========================================================================== 327 */ 328static void 329zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize, 330 zio_transform_func_t *transform) 331{ 332 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP); 333 334 zt->zt_orig_data = zio->io_data; 335 zt->zt_orig_size = zio->io_size; 336 zt->zt_bufsize = bufsize; 337 zt->zt_transform = transform; 338 339 zt->zt_next = zio->io_transform_stack; 340 zio->io_transform_stack = zt; 341 342 zio->io_data = data; 343 zio->io_size = size; 344} 345 346static void 347zio_pop_transforms(zio_t *zio) 348{ 349 zio_transform_t *zt; 350 351 while ((zt = zio->io_transform_stack) != NULL) { 352 if (zt->zt_transform != NULL) 353 zt->zt_transform(zio, 354 zt->zt_orig_data, zt->zt_orig_size); 355 356 if (zt->zt_bufsize != 0) 357 zio_buf_free(zio->io_data, zt->zt_bufsize); 358 359 zio->io_data = zt->zt_orig_data; 360 zio->io_size = zt->zt_orig_size; 361 zio->io_transform_stack = zt->zt_next; 362 363 kmem_free(zt, sizeof (zio_transform_t)); 364 } 365} 366 367/* 368 * ========================================================================== 369 * I/O transform callbacks for subblocks and decompression 370 * ========================================================================== 371 */ 372static void 373zio_subblock(zio_t *zio, void *data, uint64_t size) 374{ 375 ASSERT(zio->io_size > size); 376 377 if (zio->io_type == ZIO_TYPE_READ) 378 bcopy(zio->io_data, data, size); 379} 380 381static void 382zio_decompress(zio_t *zio, void *data, uint64_t size) 383{ 384 if (zio->io_error == 0 && 385 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp), 386 zio->io_data, data, zio->io_size, size) != 0) 387 zio->io_error = SET_ERROR(EIO); 388} 389 390/* 391 * ========================================================================== 392 * I/O parent/child relationships and pipeline interlocks 393 * ========================================================================== 394 */ 395/* 396 * NOTE - Callers to zio_walk_parents() and zio_walk_children must 397 * continue calling these functions until they return NULL. 398 * Otherwise, the next caller will pick up the list walk in 399 * some indeterminate state. (Otherwise every caller would 400 * have to pass in a cookie to keep the state represented by 401 * io_walk_link, which gets annoying.) 402 */ 403zio_t * 404zio_walk_parents(zio_t *cio) 405{ 406 zio_link_t *zl = cio->io_walk_link; 407 list_t *pl = &cio->io_parent_list; 408 409 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl); 410 cio->io_walk_link = zl; 411 412 if (zl == NULL) 413 return (NULL); 414 415 ASSERT(zl->zl_child == cio); 416 return (zl->zl_parent); 417} 418 419zio_t * 420zio_walk_children(zio_t *pio) 421{ 422 zio_link_t *zl = pio->io_walk_link; 423 list_t *cl = &pio->io_child_list; 424 425 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl); 426 pio->io_walk_link = zl; 427 428 if (zl == NULL) 429 return (NULL); 430 431 ASSERT(zl->zl_parent == pio); 432 return (zl->zl_child); 433} 434 435zio_t * 436zio_unique_parent(zio_t *cio) 437{ 438 zio_t *pio = zio_walk_parents(cio); 439 440 VERIFY(zio_walk_parents(cio) == NULL); 441 return (pio); 442} 443 444void 445zio_add_child(zio_t *pio, zio_t *cio) 446{ 447 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP); 448 449 /* 450 * Logical I/Os can have logical, gang, or vdev children. 451 * Gang I/Os can have gang or vdev children. 452 * Vdev I/Os can only have vdev children. 453 * The following ASSERT captures all of these constraints. 454 */ 455 ASSERT(cio->io_child_type <= pio->io_child_type); 456 457 zl->zl_parent = pio; 458 zl->zl_child = cio; 459 460 mutex_enter(&cio->io_lock); 461 mutex_enter(&pio->io_lock); 462 463 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0); 464 465 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 466 pio->io_children[cio->io_child_type][w] += !cio->io_state[w]; 467 468 list_insert_head(&pio->io_child_list, zl); 469 list_insert_head(&cio->io_parent_list, zl); 470 471 pio->io_child_count++; 472 cio->io_parent_count++; 473 474 mutex_exit(&pio->io_lock); 475 mutex_exit(&cio->io_lock); 476} 477 478static void 479zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl) 480{ 481 ASSERT(zl->zl_parent == pio); 482 ASSERT(zl->zl_child == cio); 483 484 mutex_enter(&cio->io_lock); 485 mutex_enter(&pio->io_lock); 486 487 list_remove(&pio->io_child_list, zl); 488 list_remove(&cio->io_parent_list, zl); 489 490 pio->io_child_count--; 491 cio->io_parent_count--; 492 493 mutex_exit(&pio->io_lock); 494 mutex_exit(&cio->io_lock); 495 496 kmem_cache_free(zio_link_cache, zl); 497} 498 499static boolean_t 500zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait) 501{ 502 uint64_t *countp = &zio->io_children[child][wait]; 503 boolean_t waiting = B_FALSE; 504 505 mutex_enter(&zio->io_lock); 506 ASSERT(zio->io_stall == NULL); 507 if (*countp != 0) { 508 zio->io_stage >>= 1; 509 zio->io_stall = countp; 510 waiting = B_TRUE; 511 } 512 mutex_exit(&zio->io_lock); 513 514 return (waiting); 515} 516 517static void 518zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait) 519{ 520 uint64_t *countp = &pio->io_children[zio->io_child_type][wait]; 521 int *errorp = &pio->io_child_error[zio->io_child_type]; 522 523 mutex_enter(&pio->io_lock); 524 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE)) 525 *errorp = zio_worst_error(*errorp, zio->io_error); 526 pio->io_reexecute |= zio->io_reexecute; 527 ASSERT3U(*countp, >, 0); 528 529 (*countp)--; 530 531 if (*countp == 0 && pio->io_stall == countp) { 532 pio->io_stall = NULL; 533 mutex_exit(&pio->io_lock); 534 zio_execute(pio); 535 } else { 536 mutex_exit(&pio->io_lock); 537 } 538} 539 540static void 541zio_inherit_child_errors(zio_t *zio, enum zio_child c) 542{ 543 if (zio->io_child_error[c] != 0 && zio->io_error == 0) 544 zio->io_error = zio->io_child_error[c]; 545} 546 547/* 548 * ========================================================================== 549 * Create the various types of I/O (read, write, free, etc) 550 * ========================================================================== 551 */ 552static zio_t * 553zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 554 void *data, uint64_t size, zio_done_func_t *done, void *private, 555 zio_type_t type, zio_priority_t priority, enum zio_flag flags, 556 vdev_t *vd, uint64_t offset, const zbookmark_phys_t *zb, 557 enum zio_stage stage, enum zio_stage pipeline) 558{ 559 zio_t *zio; 560 561 ASSERT3U(type == ZIO_TYPE_FREE || size, <=, SPA_MAXBLOCKSIZE); 562 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0); 563 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0); 564 565 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER)); 566 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER)); 567 ASSERT(vd || stage == ZIO_STAGE_OPEN); 568 569 zio = kmem_cache_alloc(zio_cache, KM_SLEEP); 570 bzero(zio, sizeof (zio_t)); 571 572 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL); 573 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL); 574 575 list_create(&zio->io_parent_list, sizeof (zio_link_t), 576 offsetof(zio_link_t, zl_parent_node)); 577 list_create(&zio->io_child_list, sizeof (zio_link_t), 578 offsetof(zio_link_t, zl_child_node)); 579 580 if (vd != NULL) 581 zio->io_child_type = ZIO_CHILD_VDEV; 582 else if (flags & ZIO_FLAG_GANG_CHILD) 583 zio->io_child_type = ZIO_CHILD_GANG; 584 else if (flags & ZIO_FLAG_DDT_CHILD) 585 zio->io_child_type = ZIO_CHILD_DDT; 586 else 587 zio->io_child_type = ZIO_CHILD_LOGICAL; 588 589 if (bp != NULL) { 590 zio->io_bp = (blkptr_t *)bp; 591 zio->io_bp_copy = *bp; 592 zio->io_bp_orig = *bp; 593 if (type != ZIO_TYPE_WRITE || 594 zio->io_child_type == ZIO_CHILD_DDT) 595 zio->io_bp = &zio->io_bp_copy; /* so caller can free */ 596 if (zio->io_child_type == ZIO_CHILD_LOGICAL) 597 zio->io_logical = zio; 598 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp)) 599 pipeline |= ZIO_GANG_STAGES; 600 } 601 602 zio->io_spa = spa; 603 zio->io_txg = txg; 604 zio->io_done = done; 605 zio->io_private = private; 606 zio->io_type = type; 607 zio->io_priority = priority; 608 zio->io_vd = vd; 609 zio->io_offset = offset; 610 zio->io_orig_data = zio->io_data = data; 611 zio->io_orig_size = zio->io_size = size; 612 zio->io_orig_flags = zio->io_flags = flags; 613 zio->io_orig_stage = zio->io_stage = stage; 614 zio->io_orig_pipeline = zio->io_pipeline = pipeline; 615 616 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY); 617 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE); 618 619 if (zb != NULL) 620 zio->io_bookmark = *zb; 621 622 if (pio != NULL) { 623 if (zio->io_logical == NULL) 624 zio->io_logical = pio->io_logical; 625 if (zio->io_child_type == ZIO_CHILD_GANG) 626 zio->io_gang_leader = pio->io_gang_leader; 627 zio_add_child(pio, zio); 628 } 629 630 return (zio); 631} 632 633static void 634zio_destroy(zio_t *zio) 635{ 636 list_destroy(&zio->io_parent_list); 637 list_destroy(&zio->io_child_list); 638 mutex_destroy(&zio->io_lock); 639 cv_destroy(&zio->io_cv); 640 kmem_cache_free(zio_cache, zio); 641} 642 643zio_t * 644zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done, 645 void *private, enum zio_flag flags) 646{ 647 zio_t *zio; 648 649 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private, 650 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, 651 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE); 652 653 return (zio); 654} 655 656zio_t * 657zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags) 658{ 659 return (zio_null(NULL, spa, NULL, done, private, flags)); 660} 661 662zio_t * 663zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, 664 void *data, uint64_t size, zio_done_func_t *done, void *private, 665 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb) 666{ 667 zio_t *zio; 668 669 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp, 670 data, size, done, private, 671 ZIO_TYPE_READ, priority, flags, NULL, 0, zb, 672 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? 673 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE); 674 675 return (zio); 676} 677 678zio_t * 679zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, 680 void *data, uint64_t size, const zio_prop_t *zp, 681 zio_done_func_t *ready, zio_done_func_t *physdone, zio_done_func_t *done, 682 void *private, 683 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb) 684{ 685 zio_t *zio; 686 687 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF && 688 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS && 689 zp->zp_compress >= ZIO_COMPRESS_OFF && 690 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS && 691 DMU_OT_IS_VALID(zp->zp_type) && 692 zp->zp_level < 32 && 693 zp->zp_copies > 0 && 694 zp->zp_copies <= spa_max_replication(spa)); 695 696 zio = zio_create(pio, spa, txg, bp, data, size, done, private, 697 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, 698 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? 699 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE); 700 701 zio->io_ready = ready; 702 zio->io_physdone = physdone; 703 zio->io_prop = *zp; 704 705 /* 706 * Data can be NULL if we are going to call zio_write_override() to 707 * provide the already-allocated BP. But we may need the data to 708 * verify a dedup hit (if requested). In this case, don't try to 709 * dedup (just take the already-allocated BP verbatim). 710 */ 711 if (data == NULL && zio->io_prop.zp_dedup_verify) { 712 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE; 713 } 714 715 return (zio); 716} 717 718zio_t * 719zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data, 720 uint64_t size, zio_done_func_t *done, void *private, 721 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb) 722{ 723 zio_t *zio; 724 725 zio = zio_create(pio, spa, txg, bp, data, size, done, private, 726 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, 727 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE); 728 729 return (zio); 730} 731 732void 733zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite) 734{ 735 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 736 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 737 ASSERT(zio->io_stage == ZIO_STAGE_OPEN); 738 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa)); 739 740 /* 741 * We must reset the io_prop to match the values that existed 742 * when the bp was first written by dmu_sync() keeping in mind 743 * that nopwrite and dedup are mutually exclusive. 744 */ 745 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup; 746 zio->io_prop.zp_nopwrite = nopwrite; 747 zio->io_prop.zp_copies = copies; 748 zio->io_bp_override = bp; 749} 750 751void 752zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp) 753{ 754 755 /* 756 * The check for EMBEDDED is a performance optimization. We 757 * process the free here (by ignoring it) rather than 758 * putting it on the list and then processing it in zio_free_sync(). 759 */ 760 if (BP_IS_EMBEDDED(bp)) 761 return; 762 metaslab_check_free(spa, bp); 763 764 /* 765 * Frees that are for the currently-syncing txg, are not going to be 766 * deferred, and which will not need to do a read (i.e. not GANG or 767 * DEDUP), can be processed immediately. Otherwise, put them on the 768 * in-memory list for later processing. 769 */ 770 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) || 771 txg != spa->spa_syncing_txg || 772 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) { 773 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp); 774 } else { 775 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 776 BP_GET_PSIZE(bp), 0))); 777 } 778} 779 780zio_t * 781zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 782 uint64_t size, enum zio_flag flags) 783{ 784 zio_t *zio; 785 enum zio_stage stage = ZIO_FREE_PIPELINE; 786 787 ASSERT(!BP_IS_HOLE(bp)); 788 ASSERT(spa_syncing_txg(spa) == txg); 789 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free); 790 791 if (BP_IS_EMBEDDED(bp)) 792 return (zio_null(pio, spa, NULL, NULL, NULL, 0)); 793 794 metaslab_check_free(spa, bp); 795 arc_freed(spa, bp); 796 797 if (zfs_trim_enabled) 798 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START | 799 ZIO_STAGE_VDEV_IO_ASSESS; 800 /* 801 * GANG and DEDUP blocks can induce a read (for the gang block header, 802 * or the DDT), so issue them asynchronously so that this thread is 803 * not tied up. 804 */ 805 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp)) 806 stage |= ZIO_STAGE_ISSUE_ASYNC; 807 808 flags |= ZIO_FLAG_DONT_QUEUE; 809 810 zio = zio_create(pio, spa, txg, bp, NULL, size, 811 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags, 812 NULL, 0, NULL, ZIO_STAGE_OPEN, stage); 813 814 return (zio); 815} 816 817zio_t * 818zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 819 zio_done_func_t *done, void *private, enum zio_flag flags) 820{ 821 zio_t *zio; 822 823 dprintf_bp(bp, "claiming in txg %llu", txg); 824 825 if (BP_IS_EMBEDDED(bp)) 826 return (zio_null(pio, spa, NULL, NULL, NULL, 0)); 827 828 /* 829 * A claim is an allocation of a specific block. Claims are needed 830 * to support immediate writes in the intent log. The issue is that 831 * immediate writes contain committed data, but in a txg that was 832 * *not* committed. Upon opening the pool after an unclean shutdown, 833 * the intent log claims all blocks that contain immediate write data 834 * so that the SPA knows they're in use. 835 * 836 * All claims *must* be resolved in the first txg -- before the SPA 837 * starts allocating blocks -- so that nothing is allocated twice. 838 * If txg == 0 we just verify that the block is claimable. 839 */ 840 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa)); 841 ASSERT(txg == spa_first_txg(spa) || txg == 0); 842 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */ 843 844 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), 845 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags, 846 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE); 847 848 return (zio); 849} 850 851zio_t * 852zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset, 853 uint64_t size, zio_done_func_t *done, void *private, 854 zio_priority_t priority, enum zio_flag flags) 855{ 856 zio_t *zio; 857 int c; 858 859 if (vd->vdev_children == 0) { 860 zio = zio_create(pio, spa, 0, NULL, NULL, size, done, private, 861 ZIO_TYPE_IOCTL, priority, flags, vd, offset, NULL, 862 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE); 863 864 zio->io_cmd = cmd; 865 } else { 866 zio = zio_null(pio, spa, NULL, NULL, NULL, flags); 867 868 for (c = 0; c < vd->vdev_children; c++) 869 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd, 870 offset, size, done, private, priority, flags)); 871 } 872 873 return (zio); 874} 875 876zio_t * 877zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 878 void *data, int checksum, zio_done_func_t *done, void *private, 879 zio_priority_t priority, enum zio_flag flags, boolean_t labels) 880{ 881 zio_t *zio; 882 883 ASSERT(vd->vdev_children == 0); 884 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || 885 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); 886 ASSERT3U(offset + size, <=, vd->vdev_psize); 887 888 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private, 889 ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset, 890 NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE); 891 892 zio->io_prop.zp_checksum = checksum; 893 894 return (zio); 895} 896 897zio_t * 898zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 899 void *data, int checksum, zio_done_func_t *done, void *private, 900 zio_priority_t priority, enum zio_flag flags, boolean_t labels) 901{ 902 zio_t *zio; 903 904 ASSERT(vd->vdev_children == 0); 905 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || 906 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); 907 ASSERT3U(offset + size, <=, vd->vdev_psize); 908 909 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private, 910 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset, 911 NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE); 912 913 zio->io_prop.zp_checksum = checksum; 914 915 if (zio_checksum_table[checksum].ci_eck) { 916 /* 917 * zec checksums are necessarily destructive -- they modify 918 * the end of the write buffer to hold the verifier/checksum. 919 * Therefore, we must make a local copy in case the data is 920 * being written to multiple places in parallel. 921 */ 922 void *wbuf = zio_buf_alloc(size); 923 bcopy(data, wbuf, size); 924 zio_push_transform(zio, wbuf, size, size, NULL); 925 } 926 927 return (zio); 928} 929 930/* 931 * Create a child I/O to do some work for us. 932 */ 933zio_t * 934zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset, 935 void *data, uint64_t size, int type, zio_priority_t priority, 936 enum zio_flag flags, zio_done_func_t *done, void *private) 937{ 938 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE; 939 zio_t *zio; 940 941 ASSERT(vd->vdev_parent == 942 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev)); 943 944 if (type == ZIO_TYPE_READ && bp != NULL) { 945 /* 946 * If we have the bp, then the child should perform the 947 * checksum and the parent need not. This pushes error 948 * detection as close to the leaves as possible and 949 * eliminates redundant checksums in the interior nodes. 950 */ 951 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY; 952 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; 953 } 954 955 /* Not all IO types require vdev io done stage e.g. free */ 956 if (!(pio->io_pipeline & ZIO_STAGE_VDEV_IO_DONE)) 957 pipeline &= ~ZIO_STAGE_VDEV_IO_DONE; 958 959 if (vd->vdev_children == 0) 960 offset += VDEV_LABEL_START_SIZE; 961 962 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE; 963 964 /* 965 * If we've decided to do a repair, the write is not speculative -- 966 * even if the original read was. 967 */ 968 if (flags & ZIO_FLAG_IO_REPAIR) 969 flags &= ~ZIO_FLAG_SPECULATIVE; 970 971 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, 972 done, private, type, priority, flags, vd, offset, &pio->io_bookmark, 973 ZIO_STAGE_VDEV_IO_START >> 1, pipeline); 974 975 zio->io_physdone = pio->io_physdone; 976 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL) 977 zio->io_logical->io_phys_children++; 978 979 return (zio); 980} 981 982zio_t * 983zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size, 984 int type, zio_priority_t priority, enum zio_flag flags, 985 zio_done_func_t *done, void *private) 986{ 987 zio_t *zio; 988 989 ASSERT(vd->vdev_ops->vdev_op_leaf); 990 991 zio = zio_create(NULL, vd->vdev_spa, 0, NULL, 992 data, size, done, private, type, priority, 993 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED, 994 vd, offset, NULL, 995 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE); 996 997 return (zio); 998} 999 1000void 1001zio_flush(zio_t *zio, vdev_t *vd) 1002{ 1003 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0, 1004 NULL, NULL, ZIO_PRIORITY_NOW, 1005 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY)); 1006} 1007 1008zio_t * 1009zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size) 1010{ 1011 1012 ASSERT(vd->vdev_ops->vdev_op_leaf); 1013 1014 return (zio_create(zio, spa, 0, NULL, NULL, size, NULL, NULL, 1015 ZIO_TYPE_FREE, ZIO_PRIORITY_TRIM, ZIO_FLAG_DONT_AGGREGATE | 1016 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY, 1017 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PHYS_PIPELINE)); 1018} 1019 1020void 1021zio_shrink(zio_t *zio, uint64_t size) 1022{ 1023 ASSERT(zio->io_executor == NULL); 1024 ASSERT(zio->io_orig_size == zio->io_size); 1025 ASSERT(size <= zio->io_size); 1026 1027 /* 1028 * We don't shrink for raidz because of problems with the 1029 * reconstruction when reading back less than the block size. 1030 * Note, BP_IS_RAIDZ() assumes no compression. 1031 */ 1032 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 1033 if (!BP_IS_RAIDZ(zio->io_bp)) 1034 zio->io_orig_size = zio->io_size = size; 1035} 1036 1037/* 1038 * ========================================================================== 1039 * Prepare to read and write logical blocks 1040 * ========================================================================== 1041 */ 1042 1043static int 1044zio_read_bp_init(zio_t *zio) 1045{ 1046 blkptr_t *bp = zio->io_bp; 1047 1048 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF && 1049 zio->io_child_type == ZIO_CHILD_LOGICAL && 1050 !(zio->io_flags & ZIO_FLAG_RAW)) { 1051 uint64_t psize = 1052 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp); 1053 void *cbuf = zio_buf_alloc(psize); 1054 1055 zio_push_transform(zio, cbuf, psize, psize, zio_decompress); 1056 } 1057 1058 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) { 1059 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1060 decode_embedded_bp_compressed(bp, zio->io_data); 1061 } else { 1062 ASSERT(!BP_IS_EMBEDDED(bp)); 1063 } 1064 1065 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0) 1066 zio->io_flags |= ZIO_FLAG_DONT_CACHE; 1067 1068 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP) 1069 zio->io_flags |= ZIO_FLAG_DONT_CACHE; 1070 1071 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL) 1072 zio->io_pipeline = ZIO_DDT_READ_PIPELINE; 1073 1074 return (ZIO_PIPELINE_CONTINUE); 1075} 1076 1077static int 1078zio_write_bp_init(zio_t *zio) 1079{ 1080 spa_t *spa = zio->io_spa; 1081 zio_prop_t *zp = &zio->io_prop; 1082 enum zio_compress compress = zp->zp_compress; 1083 blkptr_t *bp = zio->io_bp; 1084 uint64_t lsize = zio->io_size; 1085 uint64_t psize = lsize; 1086 int pass = 1; 1087 1088 /* 1089 * If our children haven't all reached the ready stage, 1090 * wait for them and then repeat this pipeline stage. 1091 */ 1092 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) || 1093 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY)) 1094 return (ZIO_PIPELINE_STOP); 1095 1096 if (!IO_IS_ALLOCATING(zio)) 1097 return (ZIO_PIPELINE_CONTINUE); 1098 1099 ASSERT(zio->io_child_type != ZIO_CHILD_DDT); 1100 1101 if (zio->io_bp_override) { 1102 ASSERT(bp->blk_birth != zio->io_txg); 1103 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0); 1104 1105 *bp = *zio->io_bp_override; 1106 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1107 1108 if (BP_IS_EMBEDDED(bp)) 1109 return (ZIO_PIPELINE_CONTINUE); 1110 1111 /* 1112 * If we've been overridden and nopwrite is set then 1113 * set the flag accordingly to indicate that a nopwrite 1114 * has already occurred. 1115 */ 1116 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) { 1117 ASSERT(!zp->zp_dedup); 1118 zio->io_flags |= ZIO_FLAG_NOPWRITE; 1119 return (ZIO_PIPELINE_CONTINUE); 1120 } 1121 1122 ASSERT(!zp->zp_nopwrite); 1123 1124 if (BP_IS_HOLE(bp) || !zp->zp_dedup) 1125 return (ZIO_PIPELINE_CONTINUE); 1126 1127 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup || 1128 zp->zp_dedup_verify); 1129 1130 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) { 1131 BP_SET_DEDUP(bp, 1); 1132 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE; 1133 return (ZIO_PIPELINE_CONTINUE); 1134 } 1135 zio->io_bp_override = NULL; 1136 BP_ZERO(bp); 1137 } 1138 1139 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) { 1140 /* 1141 * We're rewriting an existing block, which means we're 1142 * working on behalf of spa_sync(). For spa_sync() to 1143 * converge, it must eventually be the case that we don't 1144 * have to allocate new blocks. But compression changes 1145 * the blocksize, which forces a reallocate, and makes 1146 * convergence take longer. Therefore, after the first 1147 * few passes, stop compressing to ensure convergence. 1148 */ 1149 pass = spa_sync_pass(spa); 1150 1151 ASSERT(zio->io_txg == spa_syncing_txg(spa)); 1152 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1153 ASSERT(!BP_GET_DEDUP(bp)); 1154 1155 if (pass >= zfs_sync_pass_dont_compress) 1156 compress = ZIO_COMPRESS_OFF; 1157 1158 /* Make sure someone doesn't change their mind on overwrites */ 1159 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp), 1160 spa_max_replication(spa)) == BP_GET_NDVAS(bp)); 1161 } 1162 1163 if (compress != ZIO_COMPRESS_OFF) { 1164 void *cbuf = zio_buf_alloc(lsize); 1165 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize); 1166 if (psize == 0 || psize == lsize) { 1167 compress = ZIO_COMPRESS_OFF; 1168 zio_buf_free(cbuf, lsize); 1169 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE && 1170 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) && 1171 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) { 1172 encode_embedded_bp_compressed(bp, 1173 cbuf, compress, lsize, psize); 1174 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA); 1175 BP_SET_TYPE(bp, zio->io_prop.zp_type); 1176 BP_SET_LEVEL(bp, zio->io_prop.zp_level); 1177 zio_buf_free(cbuf, lsize); 1178 bp->blk_birth = zio->io_txg; 1179 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1180 ASSERT(spa_feature_is_active(spa, 1181 SPA_FEATURE_EMBEDDED_DATA)); 1182 return (ZIO_PIPELINE_CONTINUE); 1183 } else { 1184 /* 1185 * Round up compressed size to MINBLOCKSIZE and 1186 * zero the tail. 1187 */ 1188 size_t rounded = 1189 P2ROUNDUP(psize, (size_t)SPA_MINBLOCKSIZE); 1190 if (rounded > psize) { 1191 bzero((char *)cbuf + psize, rounded - psize); 1192 psize = rounded; 1193 } 1194 if (psize == lsize) { 1195 compress = ZIO_COMPRESS_OFF; 1196 zio_buf_free(cbuf, lsize); 1197 } else { 1198 zio_push_transform(zio, cbuf, 1199 psize, lsize, NULL); 1200 } 1201 } 1202 } 1203 1204 /* 1205 * The final pass of spa_sync() must be all rewrites, but the first 1206 * few passes offer a trade-off: allocating blocks defers convergence, 1207 * but newly allocated blocks are sequential, so they can be written 1208 * to disk faster. Therefore, we allow the first few passes of 1209 * spa_sync() to allocate new blocks, but force rewrites after that. 1210 * There should only be a handful of blocks after pass 1 in any case. 1211 */ 1212 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg && 1213 BP_GET_PSIZE(bp) == psize && 1214 pass >= zfs_sync_pass_rewrite) { 1215 ASSERT(psize != 0); 1216 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES; 1217 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages; 1218 zio->io_flags |= ZIO_FLAG_IO_REWRITE; 1219 } else { 1220 BP_ZERO(bp); 1221 zio->io_pipeline = ZIO_WRITE_PIPELINE; 1222 } 1223 1224 if (psize == 0) { 1225 if (zio->io_bp_orig.blk_birth != 0 && 1226 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) { 1227 BP_SET_LSIZE(bp, lsize); 1228 BP_SET_TYPE(bp, zp->zp_type); 1229 BP_SET_LEVEL(bp, zp->zp_level); 1230 BP_SET_BIRTH(bp, zio->io_txg, 0); 1231 } 1232 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1233 } else { 1234 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER); 1235 BP_SET_LSIZE(bp, lsize); 1236 BP_SET_TYPE(bp, zp->zp_type); 1237 BP_SET_LEVEL(bp, zp->zp_level); 1238 BP_SET_PSIZE(bp, psize); 1239 BP_SET_COMPRESS(bp, compress); 1240 BP_SET_CHECKSUM(bp, zp->zp_checksum); 1241 BP_SET_DEDUP(bp, zp->zp_dedup); 1242 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER); 1243 if (zp->zp_dedup) { 1244 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1245 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 1246 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE; 1247 } 1248 if (zp->zp_nopwrite) { 1249 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1250 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 1251 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE; 1252 } 1253 } 1254 1255 return (ZIO_PIPELINE_CONTINUE); 1256} 1257 1258static int 1259zio_free_bp_init(zio_t *zio) 1260{ 1261 blkptr_t *bp = zio->io_bp; 1262 1263 if (zio->io_child_type == ZIO_CHILD_LOGICAL) { 1264 if (BP_GET_DEDUP(bp)) 1265 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE; 1266 } 1267 1268 return (ZIO_PIPELINE_CONTINUE); 1269} 1270 1271/* 1272 * ========================================================================== 1273 * Execute the I/O pipeline 1274 * ========================================================================== 1275 */ 1276 1277static void 1278zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline) 1279{ 1280 spa_t *spa = zio->io_spa; 1281 zio_type_t t = zio->io_type; 1282 int flags = (cutinline ? TQ_FRONT : 0); 1283 1284 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT); 1285 1286 /* 1287 * If we're a config writer or a probe, the normal issue and 1288 * interrupt threads may all be blocked waiting for the config lock. 1289 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL. 1290 */ 1291 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE)) 1292 t = ZIO_TYPE_NULL; 1293 1294 /* 1295 * A similar issue exists for the L2ARC write thread until L2ARC 2.0. 1296 */ 1297 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux) 1298 t = ZIO_TYPE_NULL; 1299 1300 /* 1301 * If this is a high priority I/O, then use the high priority taskq if 1302 * available. 1303 */ 1304 if (zio->io_priority == ZIO_PRIORITY_NOW && 1305 spa->spa_zio_taskq[t][q + 1].stqs_count != 0) 1306 q++; 1307 1308 ASSERT3U(q, <, ZIO_TASKQ_TYPES); 1309 1310 /* 1311 * NB: We are assuming that the zio can only be dispatched 1312 * to a single taskq at a time. It would be a grievous error 1313 * to dispatch the zio to another taskq at the same time. 1314 */ 1315#if defined(illumos) || !defined(_KERNEL) 1316 ASSERT(zio->io_tqent.tqent_next == NULL); 1317#else 1318 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0); 1319#endif 1320 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio, 1321 flags, &zio->io_tqent); 1322} 1323 1324static boolean_t 1325zio_taskq_member(zio_t *zio, zio_taskq_type_t q) 1326{ 1327 kthread_t *executor = zio->io_executor; 1328 spa_t *spa = zio->io_spa; 1329 1330 for (zio_type_t t = 0; t < ZIO_TYPES; t++) { 1331 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q]; 1332 uint_t i; 1333 for (i = 0; i < tqs->stqs_count; i++) { 1334 if (taskq_member(tqs->stqs_taskq[i], executor)) 1335 return (B_TRUE); 1336 } 1337 } 1338 1339 return (B_FALSE); 1340} 1341 1342static int 1343zio_issue_async(zio_t *zio) 1344{ 1345 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); 1346 1347 return (ZIO_PIPELINE_STOP); 1348} 1349 1350void 1351zio_interrupt(zio_t *zio) 1352{ 1353 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE); 1354} 1355 1356/* 1357 * Execute the I/O pipeline until one of the following occurs: 1358 * 1359 * (1) the I/O completes 1360 * (2) the pipeline stalls waiting for dependent child I/Os 1361 * (3) the I/O issues, so we're waiting for an I/O completion interrupt 1362 * (4) the I/O is delegated by vdev-level caching or aggregation 1363 * (5) the I/O is deferred due to vdev-level queueing 1364 * (6) the I/O is handed off to another thread. 1365 * 1366 * In all cases, the pipeline stops whenever there's no CPU work; it never 1367 * burns a thread in cv_wait(). 1368 * 1369 * There's no locking on io_stage because there's no legitimate way 1370 * for multiple threads to be attempting to process the same I/O. 1371 */ 1372static zio_pipe_stage_t *zio_pipeline[]; 1373 1374void 1375zio_execute(zio_t *zio) 1376{ 1377 zio->io_executor = curthread; 1378 1379 while (zio->io_stage < ZIO_STAGE_DONE) { 1380 enum zio_stage pipeline = zio->io_pipeline; 1381 enum zio_stage stage = zio->io_stage; 1382 int rv; 1383 1384 ASSERT(!MUTEX_HELD(&zio->io_lock)); 1385 ASSERT(ISP2(stage)); 1386 ASSERT(zio->io_stall == NULL); 1387 1388 do { 1389 stage <<= 1; 1390 } while ((stage & pipeline) == 0); 1391 1392 ASSERT(stage <= ZIO_STAGE_DONE); 1393 1394 /* 1395 * If we are in interrupt context and this pipeline stage 1396 * will grab a config lock that is held across I/O, 1397 * or may wait for an I/O that needs an interrupt thread 1398 * to complete, issue async to avoid deadlock. 1399 * 1400 * For VDEV_IO_START, we cut in line so that the io will 1401 * be sent to disk promptly. 1402 */ 1403 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL && 1404 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) { 1405 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? 1406 zio_requeue_io_start_cut_in_line : B_FALSE; 1407 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); 1408 return; 1409 } 1410 1411 zio->io_stage = stage; 1412 rv = zio_pipeline[highbit64(stage) - 1](zio); 1413 1414 if (rv == ZIO_PIPELINE_STOP) 1415 return; 1416 1417 ASSERT(rv == ZIO_PIPELINE_CONTINUE); 1418 } 1419} 1420 1421/* 1422 * ========================================================================== 1423 * Initiate I/O, either sync or async 1424 * ========================================================================== 1425 */ 1426int 1427zio_wait(zio_t *zio) 1428{ 1429 int error; 1430 1431 ASSERT(zio->io_stage == ZIO_STAGE_OPEN); 1432 ASSERT(zio->io_executor == NULL); 1433 1434 zio->io_waiter = curthread; 1435 1436 zio_execute(zio); 1437 1438 mutex_enter(&zio->io_lock); 1439 while (zio->io_executor != NULL) 1440 cv_wait(&zio->io_cv, &zio->io_lock); 1441 mutex_exit(&zio->io_lock); 1442 1443 error = zio->io_error; 1444 zio_destroy(zio); 1445 1446 return (error); 1447} 1448 1449void 1450zio_nowait(zio_t *zio) 1451{ 1452 ASSERT(zio->io_executor == NULL); 1453 1454 if (zio->io_child_type == ZIO_CHILD_LOGICAL && 1455 zio_unique_parent(zio) == NULL) { 1456 /* 1457 * This is a logical async I/O with no parent to wait for it. 1458 * We add it to the spa_async_root_zio "Godfather" I/O which 1459 * will ensure they complete prior to unloading the pool. 1460 */ 1461 spa_t *spa = zio->io_spa; 1462 1463 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio); 1464 } 1465 1466 zio_execute(zio); 1467} 1468 1469/* 1470 * ========================================================================== 1471 * Reexecute or suspend/resume failed I/O 1472 * ========================================================================== 1473 */ 1474 1475static void 1476zio_reexecute(zio_t *pio) 1477{ 1478 zio_t *cio, *cio_next; 1479 1480 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL); 1481 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN); 1482 ASSERT(pio->io_gang_leader == NULL); 1483 ASSERT(pio->io_gang_tree == NULL); 1484 1485 pio->io_flags = pio->io_orig_flags; 1486 pio->io_stage = pio->io_orig_stage; 1487 pio->io_pipeline = pio->io_orig_pipeline; 1488 pio->io_reexecute = 0; 1489 pio->io_flags |= ZIO_FLAG_REEXECUTED; 1490 pio->io_error = 0; 1491 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 1492 pio->io_state[w] = 0; 1493 for (int c = 0; c < ZIO_CHILD_TYPES; c++) 1494 pio->io_child_error[c] = 0; 1495 1496 if (IO_IS_ALLOCATING(pio)) 1497 BP_ZERO(pio->io_bp); 1498 1499 /* 1500 * As we reexecute pio's children, new children could be created. 1501 * New children go to the head of pio's io_child_list, however, 1502 * so we will (correctly) not reexecute them. The key is that 1503 * the remainder of pio's io_child_list, from 'cio_next' onward, 1504 * cannot be affected by any side effects of reexecuting 'cio'. 1505 */ 1506 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) { 1507 cio_next = zio_walk_children(pio); 1508 mutex_enter(&pio->io_lock); 1509 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 1510 pio->io_children[cio->io_child_type][w]++; 1511 mutex_exit(&pio->io_lock); 1512 zio_reexecute(cio); 1513 } 1514 1515 /* 1516 * Now that all children have been reexecuted, execute the parent. 1517 * We don't reexecute "The Godfather" I/O here as it's the 1518 * responsibility of the caller to wait on him. 1519 */ 1520 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) 1521 zio_execute(pio); 1522} 1523 1524void 1525zio_suspend(spa_t *spa, zio_t *zio) 1526{ 1527 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC) 1528 fm_panic("Pool '%s' has encountered an uncorrectable I/O " 1529 "failure and the failure mode property for this pool " 1530 "is set to panic.", spa_name(spa)); 1531 1532 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0); 1533 1534 mutex_enter(&spa->spa_suspend_lock); 1535 1536 if (spa->spa_suspend_zio_root == NULL) 1537 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL, 1538 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | 1539 ZIO_FLAG_GODFATHER); 1540 1541 spa->spa_suspended = B_TRUE; 1542 1543 if (zio != NULL) { 1544 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); 1545 ASSERT(zio != spa->spa_suspend_zio_root); 1546 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1547 ASSERT(zio_unique_parent(zio) == NULL); 1548 ASSERT(zio->io_stage == ZIO_STAGE_DONE); 1549 zio_add_child(spa->spa_suspend_zio_root, zio); 1550 } 1551 1552 mutex_exit(&spa->spa_suspend_lock); 1553} 1554 1555int 1556zio_resume(spa_t *spa) 1557{ 1558 zio_t *pio; 1559 1560 /* 1561 * Reexecute all previously suspended i/o. 1562 */ 1563 mutex_enter(&spa->spa_suspend_lock); 1564 spa->spa_suspended = B_FALSE; 1565 cv_broadcast(&spa->spa_suspend_cv); 1566 pio = spa->spa_suspend_zio_root; 1567 spa->spa_suspend_zio_root = NULL; 1568 mutex_exit(&spa->spa_suspend_lock); 1569 1570 if (pio == NULL) 1571 return (0); 1572 1573 zio_reexecute(pio); 1574 return (zio_wait(pio)); 1575} 1576 1577void 1578zio_resume_wait(spa_t *spa) 1579{ 1580 mutex_enter(&spa->spa_suspend_lock); 1581 while (spa_suspended(spa)) 1582 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock); 1583 mutex_exit(&spa->spa_suspend_lock); 1584} 1585 1586/* 1587 * ========================================================================== 1588 * Gang blocks. 1589 * 1590 * A gang block is a collection of small blocks that looks to the DMU 1591 * like one large block. When zio_dva_allocate() cannot find a block 1592 * of the requested size, due to either severe fragmentation or the pool 1593 * being nearly full, it calls zio_write_gang_block() to construct the 1594 * block from smaller fragments. 1595 * 1596 * A gang block consists of a gang header (zio_gbh_phys_t) and up to 1597 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like 1598 * an indirect block: it's an array of block pointers. It consumes 1599 * only one sector and hence is allocatable regardless of fragmentation. 1600 * The gang header's bps point to its gang members, which hold the data. 1601 * 1602 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg> 1603 * as the verifier to ensure uniqueness of the SHA256 checksum. 1604 * Critically, the gang block bp's blk_cksum is the checksum of the data, 1605 * not the gang header. This ensures that data block signatures (needed for 1606 * deduplication) are independent of how the block is physically stored. 1607 * 1608 * Gang blocks can be nested: a gang member may itself be a gang block. 1609 * Thus every gang block is a tree in which root and all interior nodes are 1610 * gang headers, and the leaves are normal blocks that contain user data. 1611 * The root of the gang tree is called the gang leader. 1612 * 1613 * To perform any operation (read, rewrite, free, claim) on a gang block, 1614 * zio_gang_assemble() first assembles the gang tree (minus data leaves) 1615 * in the io_gang_tree field of the original logical i/o by recursively 1616 * reading the gang leader and all gang headers below it. This yields 1617 * an in-core tree containing the contents of every gang header and the 1618 * bps for every constituent of the gang block. 1619 * 1620 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree 1621 * and invokes a callback on each bp. To free a gang block, zio_gang_issue() 1622 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp. 1623 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim(). 1624 * zio_read_gang() is a wrapper around zio_read() that omits reading gang 1625 * headers, since we already have those in io_gang_tree. zio_rewrite_gang() 1626 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite() 1627 * of the gang header plus zio_checksum_compute() of the data to update the 1628 * gang header's blk_cksum as described above. 1629 * 1630 * The two-phase assemble/issue model solves the problem of partial failure -- 1631 * what if you'd freed part of a gang block but then couldn't read the 1632 * gang header for another part? Assembling the entire gang tree first 1633 * ensures that all the necessary gang header I/O has succeeded before 1634 * starting the actual work of free, claim, or write. Once the gang tree 1635 * is assembled, free and claim are in-memory operations that cannot fail. 1636 * 1637 * In the event that a gang write fails, zio_dva_unallocate() walks the 1638 * gang tree to immediately free (i.e. insert back into the space map) 1639 * everything we've allocated. This ensures that we don't get ENOSPC 1640 * errors during repeated suspend/resume cycles due to a flaky device. 1641 * 1642 * Gang rewrites only happen during sync-to-convergence. If we can't assemble 1643 * the gang tree, we won't modify the block, so we can safely defer the free 1644 * (knowing that the block is still intact). If we *can* assemble the gang 1645 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free 1646 * each constituent bp and we can allocate a new block on the next sync pass. 1647 * 1648 * In all cases, the gang tree allows complete recovery from partial failure. 1649 * ========================================================================== 1650 */ 1651 1652static zio_t * 1653zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1654{ 1655 if (gn != NULL) 1656 return (pio); 1657 1658 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp), 1659 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), 1660 &pio->io_bookmark)); 1661} 1662 1663zio_t * 1664zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1665{ 1666 zio_t *zio; 1667 1668 if (gn != NULL) { 1669 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, 1670 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority, 1671 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 1672 /* 1673 * As we rewrite each gang header, the pipeline will compute 1674 * a new gang block header checksum for it; but no one will 1675 * compute a new data checksum, so we do that here. The one 1676 * exception is the gang leader: the pipeline already computed 1677 * its data checksum because that stage precedes gang assembly. 1678 * (Presently, nothing actually uses interior data checksums; 1679 * this is just good hygiene.) 1680 */ 1681 if (gn != pio->io_gang_leader->io_gang_tree) { 1682 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp), 1683 data, BP_GET_PSIZE(bp)); 1684 } 1685 /* 1686 * If we are here to damage data for testing purposes, 1687 * leave the GBH alone so that we can detect the damage. 1688 */ 1689 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE) 1690 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; 1691 } else { 1692 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, 1693 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority, 1694 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 1695 } 1696 1697 return (zio); 1698} 1699 1700/* ARGSUSED */ 1701zio_t * 1702zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1703{ 1704 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp, 1705 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp), 1706 ZIO_GANG_CHILD_FLAGS(pio))); 1707} 1708 1709/* ARGSUSED */ 1710zio_t * 1711zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1712{ 1713 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp, 1714 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio))); 1715} 1716 1717static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = { 1718 NULL, 1719 zio_read_gang, 1720 zio_rewrite_gang, 1721 zio_free_gang, 1722 zio_claim_gang, 1723 NULL 1724}; 1725 1726static void zio_gang_tree_assemble_done(zio_t *zio); 1727 1728static zio_gang_node_t * 1729zio_gang_node_alloc(zio_gang_node_t **gnpp) 1730{ 1731 zio_gang_node_t *gn; 1732 1733 ASSERT(*gnpp == NULL); 1734 1735 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP); 1736 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE); 1737 *gnpp = gn; 1738 1739 return (gn); 1740} 1741 1742static void 1743zio_gang_node_free(zio_gang_node_t **gnpp) 1744{ 1745 zio_gang_node_t *gn = *gnpp; 1746 1747 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) 1748 ASSERT(gn->gn_child[g] == NULL); 1749 1750 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE); 1751 kmem_free(gn, sizeof (*gn)); 1752 *gnpp = NULL; 1753} 1754 1755static void 1756zio_gang_tree_free(zio_gang_node_t **gnpp) 1757{ 1758 zio_gang_node_t *gn = *gnpp; 1759 1760 if (gn == NULL) 1761 return; 1762 1763 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) 1764 zio_gang_tree_free(&gn->gn_child[g]); 1765 1766 zio_gang_node_free(gnpp); 1767} 1768 1769static void 1770zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp) 1771{ 1772 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp); 1773 1774 ASSERT(gio->io_gang_leader == gio); 1775 ASSERT(BP_IS_GANG(bp)); 1776 1777 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh, 1778 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn, 1779 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark)); 1780} 1781 1782static void 1783zio_gang_tree_assemble_done(zio_t *zio) 1784{ 1785 zio_t *gio = zio->io_gang_leader; 1786 zio_gang_node_t *gn = zio->io_private; 1787 blkptr_t *bp = zio->io_bp; 1788 1789 ASSERT(gio == zio_unique_parent(zio)); 1790 ASSERT(zio->io_child_count == 0); 1791 1792 if (zio->io_error) 1793 return; 1794 1795 if (BP_SHOULD_BYTESWAP(bp)) 1796 byteswap_uint64_array(zio->io_data, zio->io_size); 1797 1798 ASSERT(zio->io_data == gn->gn_gbh); 1799 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE); 1800 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); 1801 1802 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 1803 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; 1804 if (!BP_IS_GANG(gbp)) 1805 continue; 1806 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]); 1807 } 1808} 1809 1810static void 1811zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data) 1812{ 1813 zio_t *gio = pio->io_gang_leader; 1814 zio_t *zio; 1815 1816 ASSERT(BP_IS_GANG(bp) == !!gn); 1817 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp)); 1818 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree); 1819 1820 /* 1821 * If you're a gang header, your data is in gn->gn_gbh. 1822 * If you're a gang member, your data is in 'data' and gn == NULL. 1823 */ 1824 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data); 1825 1826 if (gn != NULL) { 1827 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); 1828 1829 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 1830 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; 1831 if (BP_IS_HOLE(gbp)) 1832 continue; 1833 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data); 1834 data = (char *)data + BP_GET_PSIZE(gbp); 1835 } 1836 } 1837 1838 if (gn == gio->io_gang_tree && gio->io_data != NULL) 1839 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data); 1840 1841 if (zio != pio) 1842 zio_nowait(zio); 1843} 1844 1845static int 1846zio_gang_assemble(zio_t *zio) 1847{ 1848 blkptr_t *bp = zio->io_bp; 1849 1850 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL); 1851 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 1852 1853 zio->io_gang_leader = zio; 1854 1855 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree); 1856 1857 return (ZIO_PIPELINE_CONTINUE); 1858} 1859 1860static int 1861zio_gang_issue(zio_t *zio) 1862{ 1863 blkptr_t *bp = zio->io_bp; 1864 1865 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE)) 1866 return (ZIO_PIPELINE_STOP); 1867 1868 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio); 1869 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 1870 1871 if (zio->io_child_error[ZIO_CHILD_GANG] == 0) 1872 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data); 1873 else 1874 zio_gang_tree_free(&zio->io_gang_tree); 1875 1876 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1877 1878 return (ZIO_PIPELINE_CONTINUE); 1879} 1880 1881static void 1882zio_write_gang_member_ready(zio_t *zio) 1883{ 1884 zio_t *pio = zio_unique_parent(zio); 1885 zio_t *gio = zio->io_gang_leader; 1886 dva_t *cdva = zio->io_bp->blk_dva; 1887 dva_t *pdva = pio->io_bp->blk_dva; 1888 uint64_t asize; 1889 1890 if (BP_IS_HOLE(zio->io_bp)) 1891 return; 1892 1893 ASSERT(BP_IS_HOLE(&zio->io_bp_orig)); 1894 1895 ASSERT(zio->io_child_type == ZIO_CHILD_GANG); 1896 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies); 1897 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp)); 1898 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp)); 1899 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp)); 1900 1901 mutex_enter(&pio->io_lock); 1902 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) { 1903 ASSERT(DVA_GET_GANG(&pdva[d])); 1904 asize = DVA_GET_ASIZE(&pdva[d]); 1905 asize += DVA_GET_ASIZE(&cdva[d]); 1906 DVA_SET_ASIZE(&pdva[d], asize); 1907 } 1908 mutex_exit(&pio->io_lock); 1909} 1910 1911static int 1912zio_write_gang_block(zio_t *pio) 1913{ 1914 spa_t *spa = pio->io_spa; 1915 blkptr_t *bp = pio->io_bp; 1916 zio_t *gio = pio->io_gang_leader; 1917 zio_t *zio; 1918 zio_gang_node_t *gn, **gnpp; 1919 zio_gbh_phys_t *gbh; 1920 uint64_t txg = pio->io_txg; 1921 uint64_t resid = pio->io_size; 1922 uint64_t lsize; 1923 int copies = gio->io_prop.zp_copies; 1924 int gbh_copies = MIN(copies + 1, spa_max_replication(spa)); 1925 zio_prop_t zp; 1926 int error; 1927 1928 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE, 1929 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, 1930 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER); 1931 if (error) { 1932 pio->io_error = error; 1933 return (ZIO_PIPELINE_CONTINUE); 1934 } 1935 1936 if (pio == gio) { 1937 gnpp = &gio->io_gang_tree; 1938 } else { 1939 gnpp = pio->io_private; 1940 ASSERT(pio->io_ready == zio_write_gang_member_ready); 1941 } 1942 1943 gn = zio_gang_node_alloc(gnpp); 1944 gbh = gn->gn_gbh; 1945 bzero(gbh, SPA_GANGBLOCKSIZE); 1946 1947 /* 1948 * Create the gang header. 1949 */ 1950 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL, 1951 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 1952 1953 /* 1954 * Create and nowait the gang children. 1955 */ 1956 for (int g = 0; resid != 0; resid -= lsize, g++) { 1957 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g), 1958 SPA_MINBLOCKSIZE); 1959 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid); 1960 1961 zp.zp_checksum = gio->io_prop.zp_checksum; 1962 zp.zp_compress = ZIO_COMPRESS_OFF; 1963 zp.zp_type = DMU_OT_NONE; 1964 zp.zp_level = 0; 1965 zp.zp_copies = gio->io_prop.zp_copies; 1966 zp.zp_dedup = B_FALSE; 1967 zp.zp_dedup_verify = B_FALSE; 1968 zp.zp_nopwrite = B_FALSE; 1969 1970 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g], 1971 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp, 1972 zio_write_gang_member_ready, NULL, NULL, &gn->gn_child[g], 1973 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), 1974 &pio->io_bookmark)); 1975 } 1976 1977 /* 1978 * Set pio's pipeline to just wait for zio to finish. 1979 */ 1980 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1981 1982 zio_nowait(zio); 1983 1984 return (ZIO_PIPELINE_CONTINUE); 1985} 1986 1987/* 1988 * The zio_nop_write stage in the pipeline determines if allocating 1989 * a new bp is necessary. By leveraging a cryptographically secure checksum, 1990 * such as SHA256, we can compare the checksums of the new data and the old 1991 * to determine if allocating a new block is required. The nopwrite 1992 * feature can handle writes in either syncing or open context (i.e. zil 1993 * writes) and as a result is mutually exclusive with dedup. 1994 */ 1995static int 1996zio_nop_write(zio_t *zio) 1997{ 1998 blkptr_t *bp = zio->io_bp; 1999 blkptr_t *bp_orig = &zio->io_bp_orig; 2000 zio_prop_t *zp = &zio->io_prop; 2001 2002 ASSERT(BP_GET_LEVEL(bp) == 0); 2003 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 2004 ASSERT(zp->zp_nopwrite); 2005 ASSERT(!zp->zp_dedup); 2006 ASSERT(zio->io_bp_override == NULL); 2007 ASSERT(IO_IS_ALLOCATING(zio)); 2008 2009 /* 2010 * Check to see if the original bp and the new bp have matching 2011 * characteristics (i.e. same checksum, compression algorithms, etc). 2012 * If they don't then just continue with the pipeline which will 2013 * allocate a new bp. 2014 */ 2015 if (BP_IS_HOLE(bp_orig) || 2016 !zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup || 2017 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) || 2018 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) || 2019 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) || 2020 zp->zp_copies != BP_GET_NDVAS(bp_orig)) 2021 return (ZIO_PIPELINE_CONTINUE); 2022 2023 /* 2024 * If the checksums match then reset the pipeline so that we 2025 * avoid allocating a new bp and issuing any I/O. 2026 */ 2027 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) { 2028 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup); 2029 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig)); 2030 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig)); 2031 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF); 2032 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop, 2033 sizeof (uint64_t)) == 0); 2034 2035 *bp = *bp_orig; 2036 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2037 zio->io_flags |= ZIO_FLAG_NOPWRITE; 2038 } 2039 2040 return (ZIO_PIPELINE_CONTINUE); 2041} 2042 2043/* 2044 * ========================================================================== 2045 * Dedup 2046 * ========================================================================== 2047 */ 2048static void 2049zio_ddt_child_read_done(zio_t *zio) 2050{ 2051 blkptr_t *bp = zio->io_bp; 2052 ddt_entry_t *dde = zio->io_private; 2053 ddt_phys_t *ddp; 2054 zio_t *pio = zio_unique_parent(zio); 2055 2056 mutex_enter(&pio->io_lock); 2057 ddp = ddt_phys_select(dde, bp); 2058 if (zio->io_error == 0) 2059 ddt_phys_clear(ddp); /* this ddp doesn't need repair */ 2060 if (zio->io_error == 0 && dde->dde_repair_data == NULL) 2061 dde->dde_repair_data = zio->io_data; 2062 else 2063 zio_buf_free(zio->io_data, zio->io_size); 2064 mutex_exit(&pio->io_lock); 2065} 2066 2067static int 2068zio_ddt_read_start(zio_t *zio) 2069{ 2070 blkptr_t *bp = zio->io_bp; 2071 2072 ASSERT(BP_GET_DEDUP(bp)); 2073 ASSERT(BP_GET_PSIZE(bp) == zio->io_size); 2074 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2075 2076 if (zio->io_child_error[ZIO_CHILD_DDT]) { 2077 ddt_t *ddt = ddt_select(zio->io_spa, bp); 2078 ddt_entry_t *dde = ddt_repair_start(ddt, bp); 2079 ddt_phys_t *ddp = dde->dde_phys; 2080 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp); 2081 blkptr_t blk; 2082 2083 ASSERT(zio->io_vsd == NULL); 2084 zio->io_vsd = dde; 2085 2086 if (ddp_self == NULL) 2087 return (ZIO_PIPELINE_CONTINUE); 2088 2089 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) { 2090 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self) 2091 continue; 2092 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp, 2093 &blk); 2094 zio_nowait(zio_read(zio, zio->io_spa, &blk, 2095 zio_buf_alloc(zio->io_size), zio->io_size, 2096 zio_ddt_child_read_done, dde, zio->io_priority, 2097 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE, 2098 &zio->io_bookmark)); 2099 } 2100 return (ZIO_PIPELINE_CONTINUE); 2101 } 2102 2103 zio_nowait(zio_read(zio, zio->io_spa, bp, 2104 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority, 2105 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark)); 2106 2107 return (ZIO_PIPELINE_CONTINUE); 2108} 2109 2110static int 2111zio_ddt_read_done(zio_t *zio) 2112{ 2113 blkptr_t *bp = zio->io_bp; 2114 2115 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE)) 2116 return (ZIO_PIPELINE_STOP); 2117 2118 ASSERT(BP_GET_DEDUP(bp)); 2119 ASSERT(BP_GET_PSIZE(bp) == zio->io_size); 2120 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2121 2122 if (zio->io_child_error[ZIO_CHILD_DDT]) { 2123 ddt_t *ddt = ddt_select(zio->io_spa, bp); 2124 ddt_entry_t *dde = zio->io_vsd; 2125 if (ddt == NULL) { 2126 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE); 2127 return (ZIO_PIPELINE_CONTINUE); 2128 } 2129 if (dde == NULL) { 2130 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1; 2131 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); 2132 return (ZIO_PIPELINE_STOP); 2133 } 2134 if (dde->dde_repair_data != NULL) { 2135 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size); 2136 zio->io_child_error[ZIO_CHILD_DDT] = 0; 2137 } 2138 ddt_repair_done(ddt, dde); 2139 zio->io_vsd = NULL; 2140 } 2141 2142 ASSERT(zio->io_vsd == NULL); 2143 2144 return (ZIO_PIPELINE_CONTINUE); 2145} 2146 2147static boolean_t 2148zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde) 2149{ 2150 spa_t *spa = zio->io_spa; 2151 2152 /* 2153 * Note: we compare the original data, not the transformed data, 2154 * because when zio->io_bp is an override bp, we will not have 2155 * pushed the I/O transforms. That's an important optimization 2156 * because otherwise we'd compress/encrypt all dmu_sync() data twice. 2157 */ 2158 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { 2159 zio_t *lio = dde->dde_lead_zio[p]; 2160 2161 if (lio != NULL) { 2162 return (lio->io_orig_size != zio->io_orig_size || 2163 bcmp(zio->io_orig_data, lio->io_orig_data, 2164 zio->io_orig_size) != 0); 2165 } 2166 } 2167 2168 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { 2169 ddt_phys_t *ddp = &dde->dde_phys[p]; 2170 2171 if (ddp->ddp_phys_birth != 0) { 2172 arc_buf_t *abuf = NULL; 2173 uint32_t aflags = ARC_WAIT; 2174 blkptr_t blk = *zio->io_bp; 2175 int error; 2176 2177 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth); 2178 2179 ddt_exit(ddt); 2180 2181 error = arc_read(NULL, spa, &blk, 2182 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ, 2183 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2184 &aflags, &zio->io_bookmark); 2185 2186 if (error == 0) { 2187 if (arc_buf_size(abuf) != zio->io_orig_size || 2188 bcmp(abuf->b_data, zio->io_orig_data, 2189 zio->io_orig_size) != 0) 2190 error = SET_ERROR(EEXIST); 2191 VERIFY(arc_buf_remove_ref(abuf, &abuf)); 2192 } 2193 2194 ddt_enter(ddt); 2195 return (error != 0); 2196 } 2197 } 2198 2199 return (B_FALSE); 2200} 2201 2202static void 2203zio_ddt_child_write_ready(zio_t *zio) 2204{ 2205 int p = zio->io_prop.zp_copies; 2206 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); 2207 ddt_entry_t *dde = zio->io_private; 2208 ddt_phys_t *ddp = &dde->dde_phys[p]; 2209 zio_t *pio; 2210 2211 if (zio->io_error) 2212 return; 2213 2214 ddt_enter(ddt); 2215 2216 ASSERT(dde->dde_lead_zio[p] == zio); 2217 2218 ddt_phys_fill(ddp, zio->io_bp); 2219 2220 while ((pio = zio_walk_parents(zio)) != NULL) 2221 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg); 2222 2223 ddt_exit(ddt); 2224} 2225 2226static void 2227zio_ddt_child_write_done(zio_t *zio) 2228{ 2229 int p = zio->io_prop.zp_copies; 2230 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); 2231 ddt_entry_t *dde = zio->io_private; 2232 ddt_phys_t *ddp = &dde->dde_phys[p]; 2233 2234 ddt_enter(ddt); 2235 2236 ASSERT(ddp->ddp_refcnt == 0); 2237 ASSERT(dde->dde_lead_zio[p] == zio); 2238 dde->dde_lead_zio[p] = NULL; 2239 2240 if (zio->io_error == 0) { 2241 while (zio_walk_parents(zio) != NULL) 2242 ddt_phys_addref(ddp); 2243 } else { 2244 ddt_phys_clear(ddp); 2245 } 2246 2247 ddt_exit(ddt); 2248} 2249 2250static void 2251zio_ddt_ditto_write_done(zio_t *zio) 2252{ 2253 int p = DDT_PHYS_DITTO; 2254 zio_prop_t *zp = &zio->io_prop; 2255 blkptr_t *bp = zio->io_bp; 2256 ddt_t *ddt = ddt_select(zio->io_spa, bp); 2257 ddt_entry_t *dde = zio->io_private; 2258 ddt_phys_t *ddp = &dde->dde_phys[p]; 2259 ddt_key_t *ddk = &dde->dde_key; 2260 2261 ddt_enter(ddt); 2262 2263 ASSERT(ddp->ddp_refcnt == 0); 2264 ASSERT(dde->dde_lead_zio[p] == zio); 2265 dde->dde_lead_zio[p] = NULL; 2266 2267 if (zio->io_error == 0) { 2268 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum)); 2269 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP); 2270 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp)); 2271 if (ddp->ddp_phys_birth != 0) 2272 ddt_phys_free(ddt, ddk, ddp, zio->io_txg); 2273 ddt_phys_fill(ddp, bp); 2274 } 2275 2276 ddt_exit(ddt); 2277} 2278 2279static int 2280zio_ddt_write(zio_t *zio) 2281{ 2282 spa_t *spa = zio->io_spa; 2283 blkptr_t *bp = zio->io_bp; 2284 uint64_t txg = zio->io_txg; 2285 zio_prop_t *zp = &zio->io_prop; 2286 int p = zp->zp_copies; 2287 int ditto_copies; 2288 zio_t *cio = NULL; 2289 zio_t *dio = NULL; 2290 ddt_t *ddt = ddt_select(spa, bp); 2291 ddt_entry_t *dde; 2292 ddt_phys_t *ddp; 2293 2294 ASSERT(BP_GET_DEDUP(bp)); 2295 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum); 2296 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override); 2297 2298 ddt_enter(ddt); 2299 dde = ddt_lookup(ddt, bp, B_TRUE); 2300 ddp = &dde->dde_phys[p]; 2301 2302 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) { 2303 /* 2304 * If we're using a weak checksum, upgrade to a strong checksum 2305 * and try again. If we're already using a strong checksum, 2306 * we can't resolve it, so just convert to an ordinary write. 2307 * (And automatically e-mail a paper to Nature?) 2308 */ 2309 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) { 2310 zp->zp_checksum = spa_dedup_checksum(spa); 2311 zio_pop_transforms(zio); 2312 zio->io_stage = ZIO_STAGE_OPEN; 2313 BP_ZERO(bp); 2314 } else { 2315 zp->zp_dedup = B_FALSE; 2316 } 2317 zio->io_pipeline = ZIO_WRITE_PIPELINE; 2318 ddt_exit(ddt); 2319 return (ZIO_PIPELINE_CONTINUE); 2320 } 2321 2322 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp); 2323 ASSERT(ditto_copies < SPA_DVAS_PER_BP); 2324 2325 if (ditto_copies > ddt_ditto_copies_present(dde) && 2326 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) { 2327 zio_prop_t czp = *zp; 2328 2329 czp.zp_copies = ditto_copies; 2330 2331 /* 2332 * If we arrived here with an override bp, we won't have run 2333 * the transform stack, so we won't have the data we need to 2334 * generate a child i/o. So, toss the override bp and restart. 2335 * This is safe, because using the override bp is just an 2336 * optimization; and it's rare, so the cost doesn't matter. 2337 */ 2338 if (zio->io_bp_override) { 2339 zio_pop_transforms(zio); 2340 zio->io_stage = ZIO_STAGE_OPEN; 2341 zio->io_pipeline = ZIO_WRITE_PIPELINE; 2342 zio->io_bp_override = NULL; 2343 BP_ZERO(bp); 2344 ddt_exit(ddt); 2345 return (ZIO_PIPELINE_CONTINUE); 2346 } 2347 2348 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data, 2349 zio->io_orig_size, &czp, NULL, NULL, 2350 zio_ddt_ditto_write_done, dde, zio->io_priority, 2351 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); 2352 2353 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL); 2354 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio; 2355 } 2356 2357 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) { 2358 if (ddp->ddp_phys_birth != 0) 2359 ddt_bp_fill(ddp, bp, txg); 2360 if (dde->dde_lead_zio[p] != NULL) 2361 zio_add_child(zio, dde->dde_lead_zio[p]); 2362 else 2363 ddt_phys_addref(ddp); 2364 } else if (zio->io_bp_override) { 2365 ASSERT(bp->blk_birth == txg); 2366 ASSERT(BP_EQUAL(bp, zio->io_bp_override)); 2367 ddt_phys_fill(ddp, bp); 2368 ddt_phys_addref(ddp); 2369 } else { 2370 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data, 2371 zio->io_orig_size, zp, zio_ddt_child_write_ready, NULL, 2372 zio_ddt_child_write_done, dde, zio->io_priority, 2373 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); 2374 2375 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL); 2376 dde->dde_lead_zio[p] = cio; 2377 } 2378 2379 ddt_exit(ddt); 2380 2381 if (cio) 2382 zio_nowait(cio); 2383 if (dio) 2384 zio_nowait(dio); 2385 2386 return (ZIO_PIPELINE_CONTINUE); 2387} 2388 2389ddt_entry_t *freedde; /* for debugging */ 2390 2391static int 2392zio_ddt_free(zio_t *zio) 2393{ 2394 spa_t *spa = zio->io_spa; 2395 blkptr_t *bp = zio->io_bp; 2396 ddt_t *ddt = ddt_select(spa, bp); 2397 ddt_entry_t *dde; 2398 ddt_phys_t *ddp; 2399 2400 ASSERT(BP_GET_DEDUP(bp)); 2401 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2402 2403 ddt_enter(ddt); 2404 freedde = dde = ddt_lookup(ddt, bp, B_TRUE); 2405 ddp = ddt_phys_select(dde, bp); 2406 ddt_phys_decref(ddp); 2407 ddt_exit(ddt); 2408 2409 return (ZIO_PIPELINE_CONTINUE); 2410} 2411 2412/* 2413 * ========================================================================== 2414 * Allocate and free blocks 2415 * ========================================================================== 2416 */ 2417static int 2418zio_dva_allocate(zio_t *zio) 2419{ 2420 spa_t *spa = zio->io_spa; 2421 metaslab_class_t *mc = spa_normal_class(spa); 2422 blkptr_t *bp = zio->io_bp; 2423 int error; 2424 int flags = 0; 2425 2426 if (zio->io_gang_leader == NULL) { 2427 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 2428 zio->io_gang_leader = zio; 2429 } 2430 2431 ASSERT(BP_IS_HOLE(bp)); 2432 ASSERT0(BP_GET_NDVAS(bp)); 2433 ASSERT3U(zio->io_prop.zp_copies, >, 0); 2434 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa)); 2435 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp)); 2436 2437 /* 2438 * The dump device does not support gang blocks so allocation on 2439 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid 2440 * the "fast" gang feature. 2441 */ 2442 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0; 2443 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ? 2444 METASLAB_GANG_CHILD : 0; 2445 error = metaslab_alloc(spa, mc, zio->io_size, bp, 2446 zio->io_prop.zp_copies, zio->io_txg, NULL, flags); 2447 2448 if (error) { 2449 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, " 2450 "size %llu, error %d", spa_name(spa), zio, zio->io_size, 2451 error); 2452 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE) 2453 return (zio_write_gang_block(zio)); 2454 zio->io_error = error; 2455 } 2456 2457 return (ZIO_PIPELINE_CONTINUE); 2458} 2459 2460static int 2461zio_dva_free(zio_t *zio) 2462{ 2463 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE); 2464 2465 return (ZIO_PIPELINE_CONTINUE); 2466} 2467 2468static int 2469zio_dva_claim(zio_t *zio) 2470{ 2471 int error; 2472 2473 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg); 2474 if (error) 2475 zio->io_error = error; 2476 2477 return (ZIO_PIPELINE_CONTINUE); 2478} 2479 2480/* 2481 * Undo an allocation. This is used by zio_done() when an I/O fails 2482 * and we want to give back the block we just allocated. 2483 * This handles both normal blocks and gang blocks. 2484 */ 2485static void 2486zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp) 2487{ 2488 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp)); 2489 ASSERT(zio->io_bp_override == NULL); 2490 2491 if (!BP_IS_HOLE(bp)) 2492 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE); 2493 2494 if (gn != NULL) { 2495 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 2496 zio_dva_unallocate(zio, gn->gn_child[g], 2497 &gn->gn_gbh->zg_blkptr[g]); 2498 } 2499 } 2500} 2501 2502/* 2503 * Try to allocate an intent log block. Return 0 on success, errno on failure. 2504 */ 2505int 2506zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp, 2507 uint64_t size, boolean_t use_slog) 2508{ 2509 int error = 1; 2510 2511 ASSERT(txg > spa_syncing_txg(spa)); 2512 2513 /* 2514 * ZIL blocks are always contiguous (i.e. not gang blocks) so we 2515 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang" 2516 * when allocating them. 2517 */ 2518 if (use_slog) { 2519 error = metaslab_alloc(spa, spa_log_class(spa), size, 2520 new_bp, 1, txg, old_bp, 2521 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID); 2522 } 2523 2524 if (error) { 2525 error = metaslab_alloc(spa, spa_normal_class(spa), size, 2526 new_bp, 1, txg, old_bp, 2527 METASLAB_HINTBP_AVOID); 2528 } 2529 2530 if (error == 0) { 2531 BP_SET_LSIZE(new_bp, size); 2532 BP_SET_PSIZE(new_bp, size); 2533 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF); 2534 BP_SET_CHECKSUM(new_bp, 2535 spa_version(spa) >= SPA_VERSION_SLIM_ZIL 2536 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG); 2537 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); 2538 BP_SET_LEVEL(new_bp, 0); 2539 BP_SET_DEDUP(new_bp, 0); 2540 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER); 2541 } 2542 2543 return (error); 2544} 2545 2546/* 2547 * Free an intent log block. 2548 */ 2549void 2550zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp) 2551{ 2552 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG); 2553 ASSERT(!BP_IS_GANG(bp)); 2554 2555 zio_free(spa, txg, bp); 2556} 2557 2558/* 2559 * ========================================================================== 2560 * Read, write and delete to physical devices 2561 * ========================================================================== 2562 */ 2563static int 2564zio_vdev_io_start(zio_t *zio) 2565{ 2566 vdev_t *vd = zio->io_vd; 2567 uint64_t align; 2568 spa_t *spa = zio->io_spa; 2569 int ret; 2570 2571 ASSERT(zio->io_error == 0); 2572 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0); 2573 2574 if (vd == NULL) { 2575 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) 2576 spa_config_enter(spa, SCL_ZIO, zio, RW_READER); 2577 2578 /* 2579 * The mirror_ops handle multiple DVAs in a single BP. 2580 */ 2581 return (vdev_mirror_ops.vdev_op_io_start(zio)); 2582 } 2583 2584 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE && 2585 zio->io_priority == ZIO_PRIORITY_NOW) { 2586 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg); 2587 return (ZIO_PIPELINE_CONTINUE); 2588 } 2589 2590 /* 2591 * We keep track of time-sensitive I/Os so that the scan thread 2592 * can quickly react to certain workloads. In particular, we care 2593 * about non-scrubbing, top-level reads and writes with the following 2594 * characteristics: 2595 * - synchronous writes of user data to non-slog devices 2596 * - any reads of user data 2597 * When these conditions are met, adjust the timestamp of spa_last_io 2598 * which allows the scan thread to adjust its workload accordingly. 2599 */ 2600 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL && 2601 vd == vd->vdev_top && !vd->vdev_islog && 2602 zio->io_bookmark.zb_objset != DMU_META_OBJSET && 2603 zio->io_txg != spa_syncing_txg(spa)) { 2604 uint64_t old = spa->spa_last_io; 2605 uint64_t new = ddi_get_lbolt64(); 2606 if (old != new) 2607 (void) atomic_cas_64(&spa->spa_last_io, old, new); 2608 } 2609 2610 align = 1ULL << vd->vdev_top->vdev_ashift; 2611 2612 if ((!(zio->io_flags & ZIO_FLAG_PHYSICAL) || 2613 (vd->vdev_top->vdev_physical_ashift > SPA_MINBLOCKSHIFT)) && 2614 P2PHASE(zio->io_size, align) != 0) { 2615 /* Transform logical writes to be a full physical block size. */ 2616 uint64_t asize = P2ROUNDUP(zio->io_size, align); 2617 char *abuf = NULL; 2618 if (zio->io_type == ZIO_TYPE_READ || 2619 zio->io_type == ZIO_TYPE_WRITE) 2620 abuf = zio_buf_alloc(asize); 2621 ASSERT(vd == vd->vdev_top); 2622 if (zio->io_type == ZIO_TYPE_WRITE) { 2623 bcopy(zio->io_data, abuf, zio->io_size); 2624 bzero(abuf + zio->io_size, asize - zio->io_size); 2625 } 2626 zio_push_transform(zio, abuf, asize, abuf ? asize : 0, 2627 zio_subblock); 2628 } 2629 2630 /* 2631 * If this is not a physical io, make sure that it is properly aligned 2632 * before proceeding. 2633 */ 2634 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) { 2635 ASSERT0(P2PHASE(zio->io_offset, align)); 2636 ASSERT0(P2PHASE(zio->io_size, align)); 2637 } else { 2638 /* 2639 * For physical writes, we allow 512b aligned writes and assume 2640 * the device will perform a read-modify-write as necessary. 2641 */ 2642 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE)); 2643 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE)); 2644 } 2645 2646 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa)); 2647 2648 /* 2649 * If this is a repair I/O, and there's no self-healing involved -- 2650 * that is, we're just resilvering what we expect to resilver -- 2651 * then don't do the I/O unless zio's txg is actually in vd's DTL. 2652 * This prevents spurious resilvering with nested replication. 2653 * For example, given a mirror of mirrors, (A+B)+(C+D), if only 2654 * A is out of date, we'll read from C+D, then use the data to 2655 * resilver A+B -- but we don't actually want to resilver B, just A. 2656 * The top-level mirror has no way to know this, so instead we just 2657 * discard unnecessary repairs as we work our way down the vdev tree. 2658 * The same logic applies to any form of nested replication: 2659 * ditto + mirror, RAID-Z + replacing, etc. This covers them all. 2660 */ 2661 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) && 2662 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) && 2663 zio->io_txg != 0 && /* not a delegated i/o */ 2664 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) { 2665 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 2666 zio_vdev_io_bypass(zio); 2667 return (ZIO_PIPELINE_CONTINUE); 2668 } 2669 2670 if (vd->vdev_ops->vdev_op_leaf) { 2671 switch (zio->io_type) { 2672 case ZIO_TYPE_READ: 2673 if (vdev_cache_read(zio)) 2674 return (ZIO_PIPELINE_CONTINUE); 2675 /* FALLTHROUGH */ 2676 case ZIO_TYPE_WRITE: 2677 case ZIO_TYPE_FREE: 2678 if ((zio = vdev_queue_io(zio)) == NULL) 2679 return (ZIO_PIPELINE_STOP); 2680 2681 if (!vdev_accessible(vd, zio)) { 2682 zio->io_error = SET_ERROR(ENXIO); 2683 zio_interrupt(zio); 2684 return (ZIO_PIPELINE_STOP); 2685 } 2686 break; 2687 } 2688 /* 2689 * Note that we ignore repair writes for TRIM because they can 2690 * conflict with normal writes. This isn't an issue because, by 2691 * definition, we only repair blocks that aren't freed. 2692 */ 2693 if (zio->io_type == ZIO_TYPE_WRITE && 2694 !(zio->io_flags & ZIO_FLAG_IO_REPAIR) && 2695 !trim_map_write_start(zio)) 2696 return (ZIO_PIPELINE_STOP); 2697 } 2698 2699 ret = vd->vdev_ops->vdev_op_io_start(zio); 2700 ASSERT(ret == ZIO_PIPELINE_STOP); 2701 2702 return (ret); 2703} 2704 2705static int 2706zio_vdev_io_done(zio_t *zio) 2707{ 2708 vdev_t *vd = zio->io_vd; 2709 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops; 2710 boolean_t unexpected_error = B_FALSE; 2711 2712 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE)) 2713 return (ZIO_PIPELINE_STOP); 2714 2715 ASSERT(zio->io_type == ZIO_TYPE_READ || 2716 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE); 2717 2718 if (vd != NULL && vd->vdev_ops->vdev_op_leaf && 2719 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE || 2720 zio->io_type == ZIO_TYPE_FREE)) { 2721 2722 if (zio->io_type == ZIO_TYPE_WRITE && 2723 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) 2724 trim_map_write_done(zio); 2725 2726 vdev_queue_io_done(zio); 2727 2728 if (zio->io_type == ZIO_TYPE_WRITE) 2729 vdev_cache_write(zio); 2730 2731 if (zio_injection_enabled && zio->io_error == 0) 2732 zio->io_error = zio_handle_device_injection(vd, 2733 zio, EIO); 2734 2735 if (zio_injection_enabled && zio->io_error == 0) 2736 zio->io_error = zio_handle_label_injection(zio, EIO); 2737 2738 if (zio->io_error) { 2739 if (zio->io_error == ENOTSUP && 2740 zio->io_type == ZIO_TYPE_FREE) { 2741 /* Not all devices support TRIM. */ 2742 } else if (!vdev_accessible(vd, zio)) { 2743 zio->io_error = SET_ERROR(ENXIO); 2744 } else { 2745 unexpected_error = B_TRUE; 2746 } 2747 } 2748 } 2749 2750 ops->vdev_op_io_done(zio); 2751 2752 if (unexpected_error) 2753 VERIFY(vdev_probe(vd, zio) == NULL); 2754 2755 return (ZIO_PIPELINE_CONTINUE); 2756} 2757 2758/* 2759 * For non-raidz ZIOs, we can just copy aside the bad data read from the 2760 * disk, and use that to finish the checksum ereport later. 2761 */ 2762static void 2763zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr, 2764 const void *good_buf) 2765{ 2766 /* no processing needed */ 2767 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE); 2768} 2769 2770/*ARGSUSED*/ 2771void 2772zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored) 2773{ 2774 void *buf = zio_buf_alloc(zio->io_size); 2775 2776 bcopy(zio->io_data, buf, zio->io_size); 2777 2778 zcr->zcr_cbinfo = zio->io_size; 2779 zcr->zcr_cbdata = buf; 2780 zcr->zcr_finish = zio_vsd_default_cksum_finish; 2781 zcr->zcr_free = zio_buf_free; 2782} 2783 2784static int 2785zio_vdev_io_assess(zio_t *zio) 2786{ 2787 vdev_t *vd = zio->io_vd; 2788 2789 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE)) 2790 return (ZIO_PIPELINE_STOP); 2791 2792 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) 2793 spa_config_exit(zio->io_spa, SCL_ZIO, zio); 2794 2795 if (zio->io_vsd != NULL) { 2796 zio->io_vsd_ops->vsd_free(zio); 2797 zio->io_vsd = NULL; 2798 } 2799 2800 if (zio_injection_enabled && zio->io_error == 0) 2801 zio->io_error = zio_handle_fault_injection(zio, EIO); 2802 2803 if (zio->io_type == ZIO_TYPE_FREE && 2804 zio->io_priority != ZIO_PRIORITY_NOW) { 2805 switch (zio->io_error) { 2806 case 0: 2807 ZIO_TRIM_STAT_INCR(bytes, zio->io_size); 2808 ZIO_TRIM_STAT_BUMP(success); 2809 break; 2810 case EOPNOTSUPP: 2811 ZIO_TRIM_STAT_BUMP(unsupported); 2812 break; 2813 default: 2814 ZIO_TRIM_STAT_BUMP(failed); 2815 break; 2816 } 2817 } 2818 2819 /* 2820 * If the I/O failed, determine whether we should attempt to retry it. 2821 * 2822 * On retry, we cut in line in the issue queue, since we don't want 2823 * compression/checksumming/etc. work to prevent our (cheap) IO reissue. 2824 */ 2825 if (zio->io_error && vd == NULL && 2826 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) { 2827 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */ 2828 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */ 2829 zio->io_error = 0; 2830 zio->io_flags |= ZIO_FLAG_IO_RETRY | 2831 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE; 2832 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1; 2833 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, 2834 zio_requeue_io_start_cut_in_line); 2835 return (ZIO_PIPELINE_STOP); 2836 } 2837 2838 /* 2839 * If we got an error on a leaf device, convert it to ENXIO 2840 * if the device is not accessible at all. 2841 */ 2842 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf && 2843 !vdev_accessible(vd, zio)) 2844 zio->io_error = SET_ERROR(ENXIO); 2845 2846 /* 2847 * If we can't write to an interior vdev (mirror or RAID-Z), 2848 * set vdev_cant_write so that we stop trying to allocate from it. 2849 */ 2850 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE && 2851 vd != NULL && !vd->vdev_ops->vdev_op_leaf) { 2852 vd->vdev_cant_write = B_TRUE; 2853 } 2854 2855 if (zio->io_error) 2856 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2857 2858 if (vd != NULL && vd->vdev_ops->vdev_op_leaf && 2859 zio->io_physdone != NULL) { 2860 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED)); 2861 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV); 2862 zio->io_physdone(zio->io_logical); 2863 } 2864 2865 return (ZIO_PIPELINE_CONTINUE); 2866} 2867 2868void 2869zio_vdev_io_reissue(zio_t *zio) 2870{ 2871 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); 2872 ASSERT(zio->io_error == 0); 2873 2874 zio->io_stage >>= 1; 2875} 2876 2877void 2878zio_vdev_io_redone(zio_t *zio) 2879{ 2880 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE); 2881 2882 zio->io_stage >>= 1; 2883} 2884 2885void 2886zio_vdev_io_bypass(zio_t *zio) 2887{ 2888 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); 2889 ASSERT(zio->io_error == 0); 2890 2891 zio->io_flags |= ZIO_FLAG_IO_BYPASS; 2892 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1; 2893} 2894 2895/* 2896 * ========================================================================== 2897 * Generate and verify checksums 2898 * ========================================================================== 2899 */ 2900static int 2901zio_checksum_generate(zio_t *zio) 2902{ 2903 blkptr_t *bp = zio->io_bp; 2904 enum zio_checksum checksum; 2905 2906 if (bp == NULL) { 2907 /* 2908 * This is zio_write_phys(). 2909 * We're either generating a label checksum, or none at all. 2910 */ 2911 checksum = zio->io_prop.zp_checksum; 2912 2913 if (checksum == ZIO_CHECKSUM_OFF) 2914 return (ZIO_PIPELINE_CONTINUE); 2915 2916 ASSERT(checksum == ZIO_CHECKSUM_LABEL); 2917 } else { 2918 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) { 2919 ASSERT(!IO_IS_ALLOCATING(zio)); 2920 checksum = ZIO_CHECKSUM_GANG_HEADER; 2921 } else { 2922 checksum = BP_GET_CHECKSUM(bp); 2923 } 2924 } 2925 2926 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size); 2927 2928 return (ZIO_PIPELINE_CONTINUE); 2929} 2930 2931static int 2932zio_checksum_verify(zio_t *zio) 2933{ 2934 zio_bad_cksum_t info; 2935 blkptr_t *bp = zio->io_bp; 2936 int error; 2937 2938 ASSERT(zio->io_vd != NULL); 2939 2940 if (bp == NULL) { 2941 /* 2942 * This is zio_read_phys(). 2943 * We're either verifying a label checksum, or nothing at all. 2944 */ 2945 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF) 2946 return (ZIO_PIPELINE_CONTINUE); 2947 2948 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL); 2949 } 2950 2951 if ((error = zio_checksum_error(zio, &info)) != 0) { 2952 zio->io_error = error; 2953 if (error == ECKSUM && 2954 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) { 2955 zfs_ereport_start_checksum(zio->io_spa, 2956 zio->io_vd, zio, zio->io_offset, 2957 zio->io_size, NULL, &info); 2958 } 2959 } 2960 2961 return (ZIO_PIPELINE_CONTINUE); 2962} 2963 2964/* 2965 * Called by RAID-Z to ensure we don't compute the checksum twice. 2966 */ 2967void 2968zio_checksum_verified(zio_t *zio) 2969{ 2970 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; 2971} 2972 2973/* 2974 * ========================================================================== 2975 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other. 2976 * An error of 0 indicates success. ENXIO indicates whole-device failure, 2977 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO 2978 * indicate errors that are specific to one I/O, and most likely permanent. 2979 * Any other error is presumed to be worse because we weren't expecting it. 2980 * ========================================================================== 2981 */ 2982int 2983zio_worst_error(int e1, int e2) 2984{ 2985 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO }; 2986 int r1, r2; 2987 2988 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++) 2989 if (e1 == zio_error_rank[r1]) 2990 break; 2991 2992 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++) 2993 if (e2 == zio_error_rank[r2]) 2994 break; 2995 2996 return (r1 > r2 ? e1 : e2); 2997} 2998 2999/* 3000 * ========================================================================== 3001 * I/O completion 3002 * ========================================================================== 3003 */ 3004static int 3005zio_ready(zio_t *zio) 3006{ 3007 blkptr_t *bp = zio->io_bp; 3008 zio_t *pio, *pio_next; 3009 3010 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) || 3011 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY)) 3012 return (ZIO_PIPELINE_STOP); 3013 3014 if (zio->io_ready) { 3015 ASSERT(IO_IS_ALLOCATING(zio)); 3016 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) || 3017 (zio->io_flags & ZIO_FLAG_NOPWRITE)); 3018 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0); 3019 3020 zio->io_ready(zio); 3021 } 3022 3023 if (bp != NULL && bp != &zio->io_bp_copy) 3024 zio->io_bp_copy = *bp; 3025 3026 if (zio->io_error) 3027 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 3028 3029 mutex_enter(&zio->io_lock); 3030 zio->io_state[ZIO_WAIT_READY] = 1; 3031 pio = zio_walk_parents(zio); 3032 mutex_exit(&zio->io_lock); 3033 3034 /* 3035 * As we notify zio's parents, new parents could be added. 3036 * New parents go to the head of zio's io_parent_list, however, 3037 * so we will (correctly) not notify them. The remainder of zio's 3038 * io_parent_list, from 'pio_next' onward, cannot change because 3039 * all parents must wait for us to be done before they can be done. 3040 */ 3041 for (; pio != NULL; pio = pio_next) { 3042 pio_next = zio_walk_parents(zio); 3043 zio_notify_parent(pio, zio, ZIO_WAIT_READY); 3044 } 3045 3046 if (zio->io_flags & ZIO_FLAG_NODATA) { 3047 if (BP_IS_GANG(bp)) { 3048 zio->io_flags &= ~ZIO_FLAG_NODATA; 3049 } else { 3050 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE); 3051 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; 3052 } 3053 } 3054 3055 if (zio_injection_enabled && 3056 zio->io_spa->spa_syncing_txg == zio->io_txg) 3057 zio_handle_ignored_writes(zio); 3058 3059 return (ZIO_PIPELINE_CONTINUE); 3060} 3061 3062static int 3063zio_done(zio_t *zio) 3064{ 3065 spa_t *spa = zio->io_spa; 3066 zio_t *lio = zio->io_logical; 3067 blkptr_t *bp = zio->io_bp; 3068 vdev_t *vd = zio->io_vd; 3069 uint64_t psize = zio->io_size; 3070 zio_t *pio, *pio_next; 3071 3072 /* 3073 * If our children haven't all completed, 3074 * wait for them and then repeat this pipeline stage. 3075 */ 3076 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) || 3077 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) || 3078 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) || 3079 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE)) 3080 return (ZIO_PIPELINE_STOP); 3081 3082 for (int c = 0; c < ZIO_CHILD_TYPES; c++) 3083 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 3084 ASSERT(zio->io_children[c][w] == 0); 3085 3086 if (bp != NULL && !BP_IS_EMBEDDED(bp)) { 3087 ASSERT(bp->blk_pad[0] == 0); 3088 ASSERT(bp->blk_pad[1] == 0); 3089 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 || 3090 (bp == zio_unique_parent(zio)->io_bp)); 3091 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) && 3092 zio->io_bp_override == NULL && 3093 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) { 3094 ASSERT(!BP_SHOULD_BYTESWAP(bp)); 3095 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp)); 3096 ASSERT(BP_COUNT_GANG(bp) == 0 || 3097 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp))); 3098 } 3099 if (zio->io_flags & ZIO_FLAG_NOPWRITE) 3100 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig)); 3101 } 3102 3103 /* 3104 * If there were child vdev/gang/ddt errors, they apply to us now. 3105 */ 3106 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV); 3107 zio_inherit_child_errors(zio, ZIO_CHILD_GANG); 3108 zio_inherit_child_errors(zio, ZIO_CHILD_DDT); 3109 3110 /* 3111 * If the I/O on the transformed data was successful, generate any 3112 * checksum reports now while we still have the transformed data. 3113 */ 3114 if (zio->io_error == 0) { 3115 while (zio->io_cksum_report != NULL) { 3116 zio_cksum_report_t *zcr = zio->io_cksum_report; 3117 uint64_t align = zcr->zcr_align; 3118 uint64_t asize = P2ROUNDUP(psize, align); 3119 char *abuf = zio->io_data; 3120 3121 if (asize != psize) { 3122 abuf = zio_buf_alloc(asize); 3123 bcopy(zio->io_data, abuf, psize); 3124 bzero(abuf + psize, asize - psize); 3125 } 3126 3127 zio->io_cksum_report = zcr->zcr_next; 3128 zcr->zcr_next = NULL; 3129 zcr->zcr_finish(zcr, abuf); 3130 zfs_ereport_free_checksum(zcr); 3131 3132 if (asize != psize) 3133 zio_buf_free(abuf, asize); 3134 } 3135 } 3136 3137 zio_pop_transforms(zio); /* note: may set zio->io_error */ 3138 3139 vdev_stat_update(zio, psize); 3140 3141 if (zio->io_error) { 3142 /* 3143 * If this I/O is attached to a particular vdev, 3144 * generate an error message describing the I/O failure 3145 * at the block level. We ignore these errors if the 3146 * device is currently unavailable. 3147 */ 3148 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd)) 3149 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0); 3150 3151 if ((zio->io_error == EIO || !(zio->io_flags & 3152 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) && 3153 zio == lio) { 3154 /* 3155 * For logical I/O requests, tell the SPA to log the 3156 * error and generate a logical data ereport. 3157 */ 3158 spa_log_error(spa, zio); 3159 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio, 3160 0, 0); 3161 } 3162 } 3163 3164 if (zio->io_error && zio == lio) { 3165 /* 3166 * Determine whether zio should be reexecuted. This will 3167 * propagate all the way to the root via zio_notify_parent(). 3168 */ 3169 ASSERT(vd == NULL && bp != NULL); 3170 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 3171 3172 if (IO_IS_ALLOCATING(zio) && 3173 !(zio->io_flags & ZIO_FLAG_CANFAIL)) { 3174 if (zio->io_error != ENOSPC) 3175 zio->io_reexecute |= ZIO_REEXECUTE_NOW; 3176 else 3177 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 3178 } 3179 3180 if ((zio->io_type == ZIO_TYPE_READ || 3181 zio->io_type == ZIO_TYPE_FREE) && 3182 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && 3183 zio->io_error == ENXIO && 3184 spa_load_state(spa) == SPA_LOAD_NONE && 3185 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE) 3186 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 3187 3188 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute) 3189 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 3190 3191 /* 3192 * Here is a possibly good place to attempt to do 3193 * either combinatorial reconstruction or error correction 3194 * based on checksums. It also might be a good place 3195 * to send out preliminary ereports before we suspend 3196 * processing. 3197 */ 3198 } 3199 3200 /* 3201 * If there were logical child errors, they apply to us now. 3202 * We defer this until now to avoid conflating logical child 3203 * errors with errors that happened to the zio itself when 3204 * updating vdev stats and reporting FMA events above. 3205 */ 3206 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL); 3207 3208 if ((zio->io_error || zio->io_reexecute) && 3209 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio && 3210 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE))) 3211 zio_dva_unallocate(zio, zio->io_gang_tree, bp); 3212 3213 zio_gang_tree_free(&zio->io_gang_tree); 3214 3215 /* 3216 * Godfather I/Os should never suspend. 3217 */ 3218 if ((zio->io_flags & ZIO_FLAG_GODFATHER) && 3219 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) 3220 zio->io_reexecute = 0; 3221 3222 if (zio->io_reexecute) { 3223 /* 3224 * This is a logical I/O that wants to reexecute. 3225 * 3226 * Reexecute is top-down. When an i/o fails, if it's not 3227 * the root, it simply notifies its parent and sticks around. 3228 * The parent, seeing that it still has children in zio_done(), 3229 * does the same. This percolates all the way up to the root. 3230 * The root i/o will reexecute or suspend the entire tree. 3231 * 3232 * This approach ensures that zio_reexecute() honors 3233 * all the original i/o dependency relationships, e.g. 3234 * parents not executing until children are ready. 3235 */ 3236 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 3237 3238 zio->io_gang_leader = NULL; 3239 3240 mutex_enter(&zio->io_lock); 3241 zio->io_state[ZIO_WAIT_DONE] = 1; 3242 mutex_exit(&zio->io_lock); 3243 3244 /* 3245 * "The Godfather" I/O monitors its children but is 3246 * not a true parent to them. It will track them through 3247 * the pipeline but severs its ties whenever they get into 3248 * trouble (e.g. suspended). This allows "The Godfather" 3249 * I/O to return status without blocking. 3250 */ 3251 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) { 3252 zio_link_t *zl = zio->io_walk_link; 3253 pio_next = zio_walk_parents(zio); 3254 3255 if ((pio->io_flags & ZIO_FLAG_GODFATHER) && 3256 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) { 3257 zio_remove_child(pio, zio, zl); 3258 zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 3259 } 3260 } 3261 3262 if ((pio = zio_unique_parent(zio)) != NULL) { 3263 /* 3264 * We're not a root i/o, so there's nothing to do 3265 * but notify our parent. Don't propagate errors 3266 * upward since we haven't permanently failed yet. 3267 */ 3268 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); 3269 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE; 3270 zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 3271 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) { 3272 /* 3273 * We'd fail again if we reexecuted now, so suspend 3274 * until conditions improve (e.g. device comes online). 3275 */ 3276 zio_suspend(spa, zio); 3277 } else { 3278 /* 3279 * Reexecution is potentially a huge amount of work. 3280 * Hand it off to the otherwise-unused claim taskq. 3281 */ 3282#if defined(illumos) || !defined(_KERNEL) 3283 ASSERT(zio->io_tqent.tqent_next == NULL); 3284#else 3285 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0); 3286#endif 3287 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM, 3288 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio, 3289 0, &zio->io_tqent); 3290 } 3291 return (ZIO_PIPELINE_STOP); 3292 } 3293 3294 ASSERT(zio->io_child_count == 0); 3295 ASSERT(zio->io_reexecute == 0); 3296 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL)); 3297 3298 /* 3299 * Report any checksum errors, since the I/O is complete. 3300 */ 3301 while (zio->io_cksum_report != NULL) { 3302 zio_cksum_report_t *zcr = zio->io_cksum_report; 3303 zio->io_cksum_report = zcr->zcr_next; 3304 zcr->zcr_next = NULL; 3305 zcr->zcr_finish(zcr, NULL); 3306 zfs_ereport_free_checksum(zcr); 3307 } 3308 3309 /* 3310 * It is the responsibility of the done callback to ensure that this 3311 * particular zio is no longer discoverable for adoption, and as 3312 * such, cannot acquire any new parents. 3313 */ 3314 if (zio->io_done) 3315 zio->io_done(zio); 3316 3317 mutex_enter(&zio->io_lock); 3318 zio->io_state[ZIO_WAIT_DONE] = 1; 3319 mutex_exit(&zio->io_lock); 3320 3321 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) { 3322 zio_link_t *zl = zio->io_walk_link; 3323 pio_next = zio_walk_parents(zio); 3324 zio_remove_child(pio, zio, zl); 3325 zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 3326 } 3327 3328 if (zio->io_waiter != NULL) { 3329 mutex_enter(&zio->io_lock); 3330 zio->io_executor = NULL; 3331 cv_broadcast(&zio->io_cv); 3332 mutex_exit(&zio->io_lock); 3333 } else { 3334 zio_destroy(zio); 3335 } 3336 3337 return (ZIO_PIPELINE_STOP); 3338} 3339 3340/* 3341 * ========================================================================== 3342 * I/O pipeline definition 3343 * ========================================================================== 3344 */ 3345static zio_pipe_stage_t *zio_pipeline[] = { 3346 NULL, 3347 zio_read_bp_init, 3348 zio_free_bp_init, 3349 zio_issue_async, 3350 zio_write_bp_init, 3351 zio_checksum_generate, 3352 zio_nop_write, 3353 zio_ddt_read_start, 3354 zio_ddt_read_done, 3355 zio_ddt_write, 3356 zio_ddt_free, 3357 zio_gang_assemble, 3358 zio_gang_issue, 3359 zio_dva_allocate, 3360 zio_dva_free, 3361 zio_dva_claim, 3362 zio_ready, 3363 zio_vdev_io_start, 3364 zio_vdev_io_done, 3365 zio_vdev_io_assess, 3366 zio_checksum_verify, 3367 zio_done 3368}; 3369 3370/* dnp is the dnode for zb1->zb_object */ 3371boolean_t 3372zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_phys_t *zb1, 3373 const zbookmark_phys_t *zb2) 3374{ 3375 uint64_t zb1nextL0, zb2thisobj; 3376 3377 ASSERT(zb1->zb_objset == zb2->zb_objset); 3378 ASSERT(zb2->zb_level == 0); 3379 3380 /* The objset_phys_t isn't before anything. */ 3381 if (dnp == NULL) 3382 return (B_FALSE); 3383 3384 zb1nextL0 = (zb1->zb_blkid + 1) << 3385 ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT)); 3386 3387 zb2thisobj = zb2->zb_object ? zb2->zb_object : 3388 zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT); 3389 3390 if (zb1->zb_object == DMU_META_DNODE_OBJECT) { 3391 uint64_t nextobj = zb1nextL0 * 3392 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT; 3393 return (nextobj <= zb2thisobj); 3394 } 3395 3396 if (zb1->zb_object < zb2thisobj) 3397 return (B_TRUE); 3398 if (zb1->zb_object > zb2thisobj) 3399 return (B_FALSE); 3400 if (zb2->zb_object == DMU_META_DNODE_OBJECT) 3401 return (B_FALSE); 3402 return (zb1nextL0 <= zb2->zb_blkid); 3403} 3404