zio.c revision 273348
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/zfs_context.h> 28#include <sys/fm/fs/zfs.h> 29#include <sys/spa.h> 30#include <sys/txg.h> 31#include <sys/spa_impl.h> 32#include <sys/vdev_impl.h> 33#include <sys/zio_impl.h> 34#include <sys/zio_compress.h> 35#include <sys/zio_checksum.h> 36#include <sys/dmu_objset.h> 37#include <sys/arc.h> 38#include <sys/ddt.h> 39#include <sys/trim_map.h> 40#include <sys/blkptr.h> 41#include <sys/zfeature.h> 42 43SYSCTL_DECL(_vfs_zfs); 44SYSCTL_NODE(_vfs_zfs, OID_AUTO, zio, CTLFLAG_RW, 0, "ZFS ZIO"); 45#if defined(__amd64__) 46static int zio_use_uma = 1; 47#else 48static int zio_use_uma = 0; 49#endif 50TUNABLE_INT("vfs.zfs.zio.use_uma", &zio_use_uma); 51SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, use_uma, CTLFLAG_RDTUN, &zio_use_uma, 0, 52 "Use uma(9) for ZIO allocations"); 53static int zio_exclude_metadata = 0; 54TUNABLE_INT("vfs.zfs.zio.exclude_metadata", &zio_exclude_metadata); 55SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, exclude_metadata, CTLFLAG_RDTUN, &zio_exclude_metadata, 0, 56 "Exclude metadata buffers from dumps as well"); 57 58zio_trim_stats_t zio_trim_stats = { 59 { "bytes", KSTAT_DATA_UINT64, 60 "Number of bytes successfully TRIMmed" }, 61 { "success", KSTAT_DATA_UINT64, 62 "Number of successful TRIM requests" }, 63 { "unsupported", KSTAT_DATA_UINT64, 64 "Number of TRIM requests that failed because TRIM is not supported" }, 65 { "failed", KSTAT_DATA_UINT64, 66 "Number of TRIM requests that failed for reasons other than not supported" }, 67}; 68 69static kstat_t *zio_trim_ksp; 70 71/* 72 * ========================================================================== 73 * I/O type descriptions 74 * ========================================================================== 75 */ 76const char *zio_type_name[ZIO_TYPES] = { 77 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim", 78 "zio_ioctl" 79}; 80 81/* 82 * ========================================================================== 83 * I/O kmem caches 84 * ========================================================================== 85 */ 86kmem_cache_t *zio_cache; 87kmem_cache_t *zio_link_cache; 88kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; 89kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; 90 91#ifdef _KERNEL 92extern vmem_t *zio_alloc_arena; 93#endif 94 95/* 96 * The following actions directly effect the spa's sync-to-convergence logic. 97 * The values below define the sync pass when we start performing the action. 98 * Care should be taken when changing these values as they directly impact 99 * spa_sync() performance. Tuning these values may introduce subtle performance 100 * pathologies and should only be done in the context of performance analysis. 101 * These tunables will eventually be removed and replaced with #defines once 102 * enough analysis has been done to determine optimal values. 103 * 104 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that 105 * regular blocks are not deferred. 106 */ 107int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */ 108TUNABLE_INT("vfs.zfs.sync_pass_deferred_free", &zfs_sync_pass_deferred_free); 109SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_deferred_free, CTLFLAG_RDTUN, 110 &zfs_sync_pass_deferred_free, 0, "defer frees starting in this pass"); 111int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */ 112TUNABLE_INT("vfs.zfs.sync_pass_dont_compress", &zfs_sync_pass_dont_compress); 113SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_dont_compress, CTLFLAG_RDTUN, 114 &zfs_sync_pass_dont_compress, 0, "don't compress starting in this pass"); 115int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */ 116TUNABLE_INT("vfs.zfs.sync_pass_rewrite", &zfs_sync_pass_rewrite); 117SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_rewrite, CTLFLAG_RDTUN, 118 &zfs_sync_pass_rewrite, 0, "rewrite new bps starting in this pass"); 119 120/* 121 * An allocating zio is one that either currently has the DVA allocate 122 * stage set or will have it later in its lifetime. 123 */ 124#define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE) 125 126boolean_t zio_requeue_io_start_cut_in_line = B_TRUE; 127 128#ifdef ZFS_DEBUG 129int zio_buf_debug_limit = 16384; 130#else 131int zio_buf_debug_limit = 0; 132#endif 133 134void 135zio_init(void) 136{ 137 size_t c; 138 zio_cache = kmem_cache_create("zio_cache", 139 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0); 140 zio_link_cache = kmem_cache_create("zio_link_cache", 141 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0); 142 if (!zio_use_uma) 143 goto out; 144 145 /* 146 * For small buffers, we want a cache for each multiple of 147 * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache 148 * for each quarter-power of 2. For large buffers, we want 149 * a cache for each multiple of PAGESIZE. 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 (p2 & (p2 - 1)) 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, PAGESIZE)) { 175 align = PAGESIZE; 176 } else if (IS_P2ALIGNED(size, p2 >> 2)) { 177 align = p2 >> 2; 178 } 179 180 if (align != 0) { 181 char name[36]; 182 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size); 183 zio_buf_cache[c] = kmem_cache_create(name, size, 184 align, NULL, NULL, NULL, NULL, NULL, cflags); 185 186 /* 187 * Since zio_data bufs do not appear in crash dumps, we 188 * pass KMC_NOTOUCH so that no allocator metadata is 189 * stored with the buffers. 190 */ 191 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size); 192 zio_data_buf_cache[c] = kmem_cache_create(name, size, 193 align, NULL, NULL, NULL, NULL, NULL, 194 cflags | KMC_NOTOUCH | KMC_NODEBUG); 195 } 196 } 197 198 while (--c != 0) { 199 ASSERT(zio_buf_cache[c] != NULL); 200 if (zio_buf_cache[c - 1] == NULL) 201 zio_buf_cache[c - 1] = zio_buf_cache[c]; 202 203 ASSERT(zio_data_buf_cache[c] != NULL); 204 if (zio_data_buf_cache[c - 1] == NULL) 205 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c]; 206 } 207out: 208 209 zio_inject_init(); 210 211 zio_trim_ksp = kstat_create("zfs", 0, "zio_trim", "misc", 212 KSTAT_TYPE_NAMED, 213 sizeof(zio_trim_stats) / sizeof(kstat_named_t), 214 KSTAT_FLAG_VIRTUAL); 215 216 if (zio_trim_ksp != NULL) { 217 zio_trim_ksp->ks_data = &zio_trim_stats; 218 kstat_install(zio_trim_ksp); 219 } 220} 221 222void 223zio_fini(void) 224{ 225 size_t c; 226 kmem_cache_t *last_cache = NULL; 227 kmem_cache_t *last_data_cache = NULL; 228 229 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { 230 if (zio_buf_cache[c] != last_cache) { 231 last_cache = zio_buf_cache[c]; 232 kmem_cache_destroy(zio_buf_cache[c]); 233 } 234 zio_buf_cache[c] = NULL; 235 236 if (zio_data_buf_cache[c] != last_data_cache) { 237 last_data_cache = zio_data_buf_cache[c]; 238 kmem_cache_destroy(zio_data_buf_cache[c]); 239 } 240 zio_data_buf_cache[c] = NULL; 241 } 242 243 kmem_cache_destroy(zio_link_cache); 244 kmem_cache_destroy(zio_cache); 245 246 zio_inject_fini(); 247 248 if (zio_trim_ksp != NULL) { 249 kstat_delete(zio_trim_ksp); 250 zio_trim_ksp = NULL; 251 } 252} 253 254/* 255 * ========================================================================== 256 * Allocate and free I/O buffers 257 * ========================================================================== 258 */ 259 260/* 261 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a 262 * crashdump if the kernel panics, so use it judiciously. Obviously, it's 263 * useful to inspect ZFS metadata, but if possible, we should avoid keeping 264 * excess / transient data in-core during a crashdump. 265 */ 266void * 267zio_buf_alloc(size_t size) 268{ 269 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 270 int flags = zio_exclude_metadata ? KM_NODEBUG : 0; 271 272 ASSERT3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 273 274 if (zio_use_uma) 275 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE)); 276 else 277 return (kmem_alloc(size, KM_SLEEP|flags)); 278} 279 280/* 281 * Use zio_data_buf_alloc to allocate data. The data will not appear in a 282 * crashdump if the kernel panics. This exists so that we will limit the amount 283 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount 284 * of kernel heap dumped to disk when the kernel panics) 285 */ 286void * 287zio_data_buf_alloc(size_t size) 288{ 289 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 290 291 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 292 293 if (zio_use_uma) 294 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE)); 295 else 296 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG)); 297} 298 299void 300zio_buf_free(void *buf, size_t size) 301{ 302 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 303 304 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 305 306 if (zio_use_uma) 307 kmem_cache_free(zio_buf_cache[c], buf); 308 else 309 kmem_free(buf, size); 310} 311 312void 313zio_data_buf_free(void *buf, size_t size) 314{ 315 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 316 317 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 318 319 if (zio_use_uma) 320 kmem_cache_free(zio_data_buf_cache[c], buf); 321 else 322 kmem_free(buf, size); 323} 324 325/* 326 * ========================================================================== 327 * Push and pop I/O transform buffers 328 * ========================================================================== 329 */ 330static void 331zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize, 332 zio_transform_func_t *transform) 333{ 334 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP); 335 336 zt->zt_orig_data = zio->io_data; 337 zt->zt_orig_size = zio->io_size; 338 zt->zt_bufsize = bufsize; 339 zt->zt_transform = transform; 340 341 zt->zt_next = zio->io_transform_stack; 342 zio->io_transform_stack = zt; 343 344 zio->io_data = data; 345 zio->io_size = size; 346} 347 348static void 349zio_pop_transforms(zio_t *zio) 350{ 351 zio_transform_t *zt; 352 353 while ((zt = zio->io_transform_stack) != NULL) { 354 if (zt->zt_transform != NULL) 355 zt->zt_transform(zio, 356 zt->zt_orig_data, zt->zt_orig_size); 357 358 if (zt->zt_bufsize != 0) 359 zio_buf_free(zio->io_data, zt->zt_bufsize); 360 361 zio->io_data = zt->zt_orig_data; 362 zio->io_size = zt->zt_orig_size; 363 zio->io_transform_stack = zt->zt_next; 364 365 kmem_free(zt, sizeof (zio_transform_t)); 366 } 367} 368 369/* 370 * ========================================================================== 371 * I/O transform callbacks for subblocks and decompression 372 * ========================================================================== 373 */ 374static void 375zio_subblock(zio_t *zio, void *data, uint64_t size) 376{ 377 ASSERT(zio->io_size > size); 378 379 if (zio->io_type == ZIO_TYPE_READ) 380 bcopy(zio->io_data, data, size); 381} 382 383static void 384zio_decompress(zio_t *zio, void *data, uint64_t size) 385{ 386 if (zio->io_error == 0 && 387 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp), 388 zio->io_data, data, zio->io_size, size) != 0) 389 zio->io_error = SET_ERROR(EIO); 390} 391 392/* 393 * ========================================================================== 394 * I/O parent/child relationships and pipeline interlocks 395 * ========================================================================== 396 */ 397/* 398 * NOTE - Callers to zio_walk_parents() and zio_walk_children must 399 * continue calling these functions until they return NULL. 400 * Otherwise, the next caller will pick up the list walk in 401 * some indeterminate state. (Otherwise every caller would 402 * have to pass in a cookie to keep the state represented by 403 * io_walk_link, which gets annoying.) 404 */ 405zio_t * 406zio_walk_parents(zio_t *cio) 407{ 408 zio_link_t *zl = cio->io_walk_link; 409 list_t *pl = &cio->io_parent_list; 410 411 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl); 412 cio->io_walk_link = zl; 413 414 if (zl == NULL) 415 return (NULL); 416 417 ASSERT(zl->zl_child == cio); 418 return (zl->zl_parent); 419} 420 421zio_t * 422zio_walk_children(zio_t *pio) 423{ 424 zio_link_t *zl = pio->io_walk_link; 425 list_t *cl = &pio->io_child_list; 426 427 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl); 428 pio->io_walk_link = zl; 429 430 if (zl == NULL) 431 return (NULL); 432 433 ASSERT(zl->zl_parent == pio); 434 return (zl->zl_child); 435} 436 437zio_t * 438zio_unique_parent(zio_t *cio) 439{ 440 zio_t *pio = zio_walk_parents(cio); 441 442 VERIFY(zio_walk_parents(cio) == NULL); 443 return (pio); 444} 445 446void 447zio_add_child(zio_t *pio, zio_t *cio) 448{ 449 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP); 450 451 /* 452 * Logical I/Os can have logical, gang, or vdev children. 453 * Gang I/Os can have gang or vdev children. 454 * Vdev I/Os can only have vdev children. 455 * The following ASSERT captures all of these constraints. 456 */ 457 ASSERT(cio->io_child_type <= pio->io_child_type); 458 459 zl->zl_parent = pio; 460 zl->zl_child = cio; 461 462 mutex_enter(&cio->io_lock); 463 mutex_enter(&pio->io_lock); 464 465 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0); 466 467 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 468 pio->io_children[cio->io_child_type][w] += !cio->io_state[w]; 469 470 list_insert_head(&pio->io_child_list, zl); 471 list_insert_head(&cio->io_parent_list, zl); 472 473 pio->io_child_count++; 474 cio->io_parent_count++; 475 476 mutex_exit(&pio->io_lock); 477 mutex_exit(&cio->io_lock); 478} 479 480static void 481zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl) 482{ 483 ASSERT(zl->zl_parent == pio); 484 ASSERT(zl->zl_child == cio); 485 486 mutex_enter(&cio->io_lock); 487 mutex_enter(&pio->io_lock); 488 489 list_remove(&pio->io_child_list, zl); 490 list_remove(&cio->io_parent_list, zl); 491 492 pio->io_child_count--; 493 cio->io_parent_count--; 494 495 mutex_exit(&pio->io_lock); 496 mutex_exit(&cio->io_lock); 497 498 kmem_cache_free(zio_link_cache, zl); 499} 500 501static boolean_t 502zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait) 503{ 504 uint64_t *countp = &zio->io_children[child][wait]; 505 boolean_t waiting = B_FALSE; 506 507 mutex_enter(&zio->io_lock); 508 ASSERT(zio->io_stall == NULL); 509 if (*countp != 0) { 510 zio->io_stage >>= 1; 511 zio->io_stall = countp; 512 waiting = B_TRUE; 513 } 514 mutex_exit(&zio->io_lock); 515 516 return (waiting); 517} 518 519static void 520zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait) 521{ 522 uint64_t *countp = &pio->io_children[zio->io_child_type][wait]; 523 int *errorp = &pio->io_child_error[zio->io_child_type]; 524 525 mutex_enter(&pio->io_lock); 526 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE)) 527 *errorp = zio_worst_error(*errorp, zio->io_error); 528 pio->io_reexecute |= zio->io_reexecute; 529 ASSERT3U(*countp, >, 0); 530 531 (*countp)--; 532 533 if (*countp == 0 && pio->io_stall == countp) { 534 pio->io_stall = NULL; 535 mutex_exit(&pio->io_lock); 536 zio_execute(pio); 537 } else { 538 mutex_exit(&pio->io_lock); 539 } 540} 541 542static void 543zio_inherit_child_errors(zio_t *zio, enum zio_child c) 544{ 545 if (zio->io_child_error[c] != 0 && zio->io_error == 0) 546 zio->io_error = zio->io_child_error[c]; 547} 548 549/* 550 * ========================================================================== 551 * Create the various types of I/O (read, write, free, etc) 552 * ========================================================================== 553 */ 554static zio_t * 555zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 556 void *data, uint64_t size, zio_done_func_t *done, void *private, 557 zio_type_t type, zio_priority_t priority, enum zio_flag flags, 558 vdev_t *vd, uint64_t offset, const zbookmark_phys_t *zb, 559 enum zio_stage stage, enum zio_stage pipeline) 560{ 561 zio_t *zio; 562 563 ASSERT3U(type == ZIO_TYPE_FREE || size, <=, SPA_MAXBLOCKSIZE); 564 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0); 565 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0); 566 567 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER)); 568 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER)); 569 ASSERT(vd || stage == ZIO_STAGE_OPEN); 570 571 zio = kmem_cache_alloc(zio_cache, KM_SLEEP); 572 bzero(zio, sizeof (zio_t)); 573 574 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL); 575 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL); 576 577 list_create(&zio->io_parent_list, sizeof (zio_link_t), 578 offsetof(zio_link_t, zl_parent_node)); 579 list_create(&zio->io_child_list, sizeof (zio_link_t), 580 offsetof(zio_link_t, zl_child_node)); 581 582 if (vd != NULL) 583 zio->io_child_type = ZIO_CHILD_VDEV; 584 else if (flags & ZIO_FLAG_GANG_CHILD) 585 zio->io_child_type = ZIO_CHILD_GANG; 586 else if (flags & ZIO_FLAG_DDT_CHILD) 587 zio->io_child_type = ZIO_CHILD_DDT; 588 else 589 zio->io_child_type = ZIO_CHILD_LOGICAL; 590 591 if (bp != NULL) { 592 zio->io_bp = (blkptr_t *)bp; 593 zio->io_bp_copy = *bp; 594 zio->io_bp_orig = *bp; 595 if (type != ZIO_TYPE_WRITE || 596 zio->io_child_type == ZIO_CHILD_DDT) 597 zio->io_bp = &zio->io_bp_copy; /* so caller can free */ 598 if (zio->io_child_type == ZIO_CHILD_LOGICAL) 599 zio->io_logical = zio; 600 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp)) 601 pipeline |= ZIO_GANG_STAGES; 602 } 603 604 zio->io_spa = spa; 605 zio->io_txg = txg; 606 zio->io_done = done; 607 zio->io_private = private; 608 zio->io_type = type; 609 zio->io_priority = priority; 610 zio->io_vd = vd; 611 zio->io_offset = offset; 612 zio->io_orig_data = zio->io_data = data; 613 zio->io_orig_size = zio->io_size = size; 614 zio->io_orig_flags = zio->io_flags = flags; 615 zio->io_orig_stage = zio->io_stage = stage; 616 zio->io_orig_pipeline = zio->io_pipeline = pipeline; 617 618 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY); 619 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE); 620 621 if (zb != NULL) 622 zio->io_bookmark = *zb; 623 624 if (pio != NULL) { 625 if (zio->io_logical == NULL) 626 zio->io_logical = pio->io_logical; 627 if (zio->io_child_type == ZIO_CHILD_GANG) 628 zio->io_gang_leader = pio->io_gang_leader; 629 zio_add_child(pio, zio); 630 } 631 632 return (zio); 633} 634 635static void 636zio_destroy(zio_t *zio) 637{ 638 list_destroy(&zio->io_parent_list); 639 list_destroy(&zio->io_child_list); 640 mutex_destroy(&zio->io_lock); 641 cv_destroy(&zio->io_cv); 642 kmem_cache_free(zio_cache, zio); 643} 644 645zio_t * 646zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done, 647 void *private, enum zio_flag flags) 648{ 649 zio_t *zio; 650 651 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private, 652 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, 653 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE); 654 655 return (zio); 656} 657 658zio_t * 659zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags) 660{ 661 return (zio_null(NULL, spa, NULL, done, private, flags)); 662} 663 664zio_t * 665zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, 666 void *data, uint64_t size, zio_done_func_t *done, void *private, 667 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb) 668{ 669 zio_t *zio; 670 671 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp, 672 data, size, done, private, 673 ZIO_TYPE_READ, priority, flags, NULL, 0, zb, 674 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? 675 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE); 676 677 return (zio); 678} 679 680zio_t * 681zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, 682 void *data, uint64_t size, const zio_prop_t *zp, 683 zio_done_func_t *ready, zio_done_func_t *physdone, zio_done_func_t *done, 684 void *private, 685 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb) 686{ 687 zio_t *zio; 688 689 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF && 690 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS && 691 zp->zp_compress >= ZIO_COMPRESS_OFF && 692 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS && 693 DMU_OT_IS_VALID(zp->zp_type) && 694 zp->zp_level < 32 && 695 zp->zp_copies > 0 && 696 zp->zp_copies <= spa_max_replication(spa)); 697 698 zio = zio_create(pio, spa, txg, bp, data, size, done, private, 699 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, 700 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? 701 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE); 702 703 zio->io_ready = ready; 704 zio->io_physdone = physdone; 705 zio->io_prop = *zp; 706 707 /* 708 * Data can be NULL if we are going to call zio_write_override() to 709 * provide the already-allocated BP. But we may need the data to 710 * verify a dedup hit (if requested). In this case, don't try to 711 * dedup (just take the already-allocated BP verbatim). 712 */ 713 if (data == NULL && zio->io_prop.zp_dedup_verify) { 714 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE; 715 } 716 717 return (zio); 718} 719 720zio_t * 721zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data, 722 uint64_t size, zio_done_func_t *done, void *private, 723 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb) 724{ 725 zio_t *zio; 726 727 zio = zio_create(pio, spa, txg, bp, data, size, done, private, 728 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, 729 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE); 730 731 return (zio); 732} 733 734void 735zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite) 736{ 737 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 738 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 739 ASSERT(zio->io_stage == ZIO_STAGE_OPEN); 740 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa)); 741 742 /* 743 * We must reset the io_prop to match the values that existed 744 * when the bp was first written by dmu_sync() keeping in mind 745 * that nopwrite and dedup are mutually exclusive. 746 */ 747 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup; 748 zio->io_prop.zp_nopwrite = nopwrite; 749 zio->io_prop.zp_copies = copies; 750 zio->io_bp_override = bp; 751} 752 753void 754zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp) 755{ 756 757 /* 758 * The check for EMBEDDED is a performance optimization. We 759 * process the free here (by ignoring it) rather than 760 * putting it on the list and then processing it in zio_free_sync(). 761 */ 762 if (BP_IS_EMBEDDED(bp)) 763 return; 764 metaslab_check_free(spa, bp); 765 766 /* 767 * Frees that are for the currently-syncing txg, are not going to be 768 * deferred, and which will not need to do a read (i.e. not GANG or 769 * DEDUP), can be processed immediately. Otherwise, put them on the 770 * in-memory list for later processing. 771 */ 772 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) || 773 txg != spa->spa_syncing_txg || 774 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) { 775 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp); 776 } else { 777 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 778 BP_GET_PSIZE(bp), 0))); 779 } 780} 781 782zio_t * 783zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 784 uint64_t size, enum zio_flag flags) 785{ 786 zio_t *zio; 787 enum zio_stage stage = ZIO_FREE_PIPELINE; 788 789 ASSERT(!BP_IS_HOLE(bp)); 790 ASSERT(spa_syncing_txg(spa) == txg); 791 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free); 792 793 if (BP_IS_EMBEDDED(bp)) 794 return (zio_null(pio, spa, NULL, NULL, NULL, 0)); 795 796 metaslab_check_free(spa, bp); 797 arc_freed(spa, bp); 798 799 if (zfs_trim_enabled) 800 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START | 801 ZIO_STAGE_VDEV_IO_ASSESS; 802 /* 803 * GANG and DEDUP blocks can induce a read (for the gang block header, 804 * or the DDT), so issue them asynchronously so that this thread is 805 * not tied up. 806 */ 807 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp)) 808 stage |= ZIO_STAGE_ISSUE_ASYNC; 809 810 flags |= ZIO_FLAG_DONT_QUEUE; 811 812 zio = zio_create(pio, spa, txg, bp, NULL, size, 813 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags, 814 NULL, 0, NULL, ZIO_STAGE_OPEN, stage); 815 816 return (zio); 817} 818 819zio_t * 820zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 821 zio_done_func_t *done, void *private, enum zio_flag flags) 822{ 823 zio_t *zio; 824 825 dprintf_bp(bp, "claiming in txg %llu", txg); 826 827 if (BP_IS_EMBEDDED(bp)) 828 return (zio_null(pio, spa, NULL, NULL, NULL, 0)); 829 830 /* 831 * A claim is an allocation of a specific block. Claims are needed 832 * to support immediate writes in the intent log. The issue is that 833 * immediate writes contain committed data, but in a txg that was 834 * *not* committed. Upon opening the pool after an unclean shutdown, 835 * the intent log claims all blocks that contain immediate write data 836 * so that the SPA knows they're in use. 837 * 838 * All claims *must* be resolved in the first txg -- before the SPA 839 * starts allocating blocks -- so that nothing is allocated twice. 840 * If txg == 0 we just verify that the block is claimable. 841 */ 842 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa)); 843 ASSERT(txg == spa_first_txg(spa) || txg == 0); 844 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */ 845 846 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), 847 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags, 848 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE); 849 850 return (zio); 851} 852 853zio_t * 854zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset, 855 uint64_t size, zio_done_func_t *done, void *private, 856 zio_priority_t priority, enum zio_flag flags) 857{ 858 zio_t *zio; 859 int c; 860 861 if (vd->vdev_children == 0) { 862 zio = zio_create(pio, spa, 0, NULL, NULL, size, done, private, 863 ZIO_TYPE_IOCTL, priority, flags, vd, offset, NULL, 864 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE); 865 866 zio->io_cmd = cmd; 867 } else { 868 zio = zio_null(pio, spa, NULL, NULL, NULL, flags); 869 870 for (c = 0; c < vd->vdev_children; c++) 871 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd, 872 offset, size, done, private, priority, flags)); 873 } 874 875 return (zio); 876} 877 878zio_t * 879zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 880 void *data, int checksum, zio_done_func_t *done, void *private, 881 zio_priority_t priority, enum zio_flag flags, boolean_t labels) 882{ 883 zio_t *zio; 884 885 ASSERT(vd->vdev_children == 0); 886 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || 887 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); 888 ASSERT3U(offset + size, <=, vd->vdev_psize); 889 890 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private, 891 ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset, 892 NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE); 893 894 zio->io_prop.zp_checksum = checksum; 895 896 return (zio); 897} 898 899zio_t * 900zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 901 void *data, int checksum, zio_done_func_t *done, void *private, 902 zio_priority_t priority, enum zio_flag flags, boolean_t labels) 903{ 904 zio_t *zio; 905 906 ASSERT(vd->vdev_children == 0); 907 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || 908 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); 909 ASSERT3U(offset + size, <=, vd->vdev_psize); 910 911 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private, 912 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset, 913 NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE); 914 915 zio->io_prop.zp_checksum = checksum; 916 917 if (zio_checksum_table[checksum].ci_eck) { 918 /* 919 * zec checksums are necessarily destructive -- they modify 920 * the end of the write buffer to hold the verifier/checksum. 921 * Therefore, we must make a local copy in case the data is 922 * being written to multiple places in parallel. 923 */ 924 void *wbuf = zio_buf_alloc(size); 925 bcopy(data, wbuf, size); 926 zio_push_transform(zio, wbuf, size, size, NULL); 927 } 928 929 return (zio); 930} 931 932/* 933 * Create a child I/O to do some work for us. 934 */ 935zio_t * 936zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset, 937 void *data, uint64_t size, int type, zio_priority_t priority, 938 enum zio_flag flags, zio_done_func_t *done, void *private) 939{ 940 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE; 941 zio_t *zio; 942 943 ASSERT(vd->vdev_parent == 944 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev)); 945 946 if (type == ZIO_TYPE_READ && bp != NULL) { 947 /* 948 * If we have the bp, then the child should perform the 949 * checksum and the parent need not. This pushes error 950 * detection as close to the leaves as possible and 951 * eliminates redundant checksums in the interior nodes. 952 */ 953 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY; 954 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; 955 } 956 957 /* Not all IO types require vdev io done stage e.g. free */ 958 if (!(pio->io_pipeline & ZIO_STAGE_VDEV_IO_DONE)) 959 pipeline &= ~ZIO_STAGE_VDEV_IO_DONE; 960 961 if (vd->vdev_children == 0) 962 offset += VDEV_LABEL_START_SIZE; 963 964 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE; 965 966 /* 967 * If we've decided to do a repair, the write is not speculative -- 968 * even if the original read was. 969 */ 970 if (flags & ZIO_FLAG_IO_REPAIR) 971 flags &= ~ZIO_FLAG_SPECULATIVE; 972 973 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, 974 done, private, type, priority, flags, vd, offset, &pio->io_bookmark, 975 ZIO_STAGE_VDEV_IO_START >> 1, pipeline); 976 977 zio->io_physdone = pio->io_physdone; 978 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL) 979 zio->io_logical->io_phys_children++; 980 981 return (zio); 982} 983 984zio_t * 985zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size, 986 int type, zio_priority_t priority, enum zio_flag flags, 987 zio_done_func_t *done, void *private) 988{ 989 zio_t *zio; 990 991 ASSERT(vd->vdev_ops->vdev_op_leaf); 992 993 zio = zio_create(NULL, vd->vdev_spa, 0, NULL, 994 data, size, done, private, type, priority, 995 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED, 996 vd, offset, NULL, 997 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE); 998 999 return (zio); 1000} 1001 1002void 1003zio_flush(zio_t *zio, vdev_t *vd) 1004{ 1005 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0, 1006 NULL, NULL, ZIO_PRIORITY_NOW, 1007 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY)); 1008} 1009 1010zio_t * 1011zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size) 1012{ 1013 1014 ASSERT(vd->vdev_ops->vdev_op_leaf); 1015 1016 return (zio_create(zio, spa, 0, NULL, NULL, size, NULL, NULL, 1017 ZIO_TYPE_FREE, ZIO_PRIORITY_TRIM, ZIO_FLAG_DONT_AGGREGATE | 1018 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY, 1019 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PHYS_PIPELINE)); 1020} 1021 1022void 1023zio_shrink(zio_t *zio, uint64_t size) 1024{ 1025 ASSERT(zio->io_executor == NULL); 1026 ASSERT(zio->io_orig_size == zio->io_size); 1027 ASSERT(size <= zio->io_size); 1028 1029 /* 1030 * We don't shrink for raidz because of problems with the 1031 * reconstruction when reading back less than the block size. 1032 * Note, BP_IS_RAIDZ() assumes no compression. 1033 */ 1034 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 1035 if (!BP_IS_RAIDZ(zio->io_bp)) 1036 zio->io_orig_size = zio->io_size = size; 1037} 1038 1039/* 1040 * ========================================================================== 1041 * Prepare to read and write logical blocks 1042 * ========================================================================== 1043 */ 1044 1045static int 1046zio_read_bp_init(zio_t *zio) 1047{ 1048 blkptr_t *bp = zio->io_bp; 1049 1050 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF && 1051 zio->io_child_type == ZIO_CHILD_LOGICAL && 1052 !(zio->io_flags & ZIO_FLAG_RAW)) { 1053 uint64_t psize = 1054 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp); 1055 void *cbuf = zio_buf_alloc(psize); 1056 1057 zio_push_transform(zio, cbuf, psize, psize, zio_decompress); 1058 } 1059 1060 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) { 1061 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1062 decode_embedded_bp_compressed(bp, zio->io_data); 1063 } else { 1064 ASSERT(!BP_IS_EMBEDDED(bp)); 1065 } 1066 1067 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0) 1068 zio->io_flags |= ZIO_FLAG_DONT_CACHE; 1069 1070 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP) 1071 zio->io_flags |= ZIO_FLAG_DONT_CACHE; 1072 1073 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL) 1074 zio->io_pipeline = ZIO_DDT_READ_PIPELINE; 1075 1076 return (ZIO_PIPELINE_CONTINUE); 1077} 1078 1079static int 1080zio_write_bp_init(zio_t *zio) 1081{ 1082 spa_t *spa = zio->io_spa; 1083 zio_prop_t *zp = &zio->io_prop; 1084 enum zio_compress compress = zp->zp_compress; 1085 blkptr_t *bp = zio->io_bp; 1086 uint64_t lsize = zio->io_size; 1087 uint64_t psize = lsize; 1088 int pass = 1; 1089 1090 /* 1091 * If our children haven't all reached the ready stage, 1092 * wait for them and then repeat this pipeline stage. 1093 */ 1094 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) || 1095 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY)) 1096 return (ZIO_PIPELINE_STOP); 1097 1098 if (!IO_IS_ALLOCATING(zio)) 1099 return (ZIO_PIPELINE_CONTINUE); 1100 1101 ASSERT(zio->io_child_type != ZIO_CHILD_DDT); 1102 1103 if (zio->io_bp_override) { 1104 ASSERT(bp->blk_birth != zio->io_txg); 1105 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0); 1106 1107 *bp = *zio->io_bp_override; 1108 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1109 1110 if (BP_IS_EMBEDDED(bp)) 1111 return (ZIO_PIPELINE_CONTINUE); 1112 1113 /* 1114 * If we've been overridden and nopwrite is set then 1115 * set the flag accordingly to indicate that a nopwrite 1116 * has already occurred. 1117 */ 1118 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) { 1119 ASSERT(!zp->zp_dedup); 1120 zio->io_flags |= ZIO_FLAG_NOPWRITE; 1121 return (ZIO_PIPELINE_CONTINUE); 1122 } 1123 1124 ASSERT(!zp->zp_nopwrite); 1125 1126 if (BP_IS_HOLE(bp) || !zp->zp_dedup) 1127 return (ZIO_PIPELINE_CONTINUE); 1128 1129 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup || 1130 zp->zp_dedup_verify); 1131 1132 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) { 1133 BP_SET_DEDUP(bp, 1); 1134 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE; 1135 return (ZIO_PIPELINE_CONTINUE); 1136 } 1137 zio->io_bp_override = NULL; 1138 BP_ZERO(bp); 1139 } 1140 1141 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) { 1142 /* 1143 * We're rewriting an existing block, which means we're 1144 * working on behalf of spa_sync(). For spa_sync() to 1145 * converge, it must eventually be the case that we don't 1146 * have to allocate new blocks. But compression changes 1147 * the blocksize, which forces a reallocate, and makes 1148 * convergence take longer. Therefore, after the first 1149 * few passes, stop compressing to ensure convergence. 1150 */ 1151 pass = spa_sync_pass(spa); 1152 1153 ASSERT(zio->io_txg == spa_syncing_txg(spa)); 1154 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1155 ASSERT(!BP_GET_DEDUP(bp)); 1156 1157 if (pass >= zfs_sync_pass_dont_compress) 1158 compress = ZIO_COMPRESS_OFF; 1159 1160 /* Make sure someone doesn't change their mind on overwrites */ 1161 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp), 1162 spa_max_replication(spa)) == BP_GET_NDVAS(bp)); 1163 } 1164 1165 if (compress != ZIO_COMPRESS_OFF) { 1166 void *cbuf = zio_buf_alloc(lsize); 1167 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize); 1168 if (psize == 0 || psize == lsize) { 1169 compress = ZIO_COMPRESS_OFF; 1170 zio_buf_free(cbuf, lsize); 1171 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE && 1172 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) && 1173 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) { 1174 encode_embedded_bp_compressed(bp, 1175 cbuf, compress, lsize, psize); 1176 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA); 1177 BP_SET_TYPE(bp, zio->io_prop.zp_type); 1178 BP_SET_LEVEL(bp, zio->io_prop.zp_level); 1179 zio_buf_free(cbuf, lsize); 1180 bp->blk_birth = zio->io_txg; 1181 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1182 ASSERT(spa_feature_is_active(spa, 1183 SPA_FEATURE_EMBEDDED_DATA)); 1184 return (ZIO_PIPELINE_CONTINUE); 1185 } else { 1186 /* 1187 * Round up compressed size to MINBLOCKSIZE and 1188 * zero the tail. 1189 */ 1190 size_t rounded = 1191 P2ROUNDUP(psize, (size_t)SPA_MINBLOCKSIZE); 1192 if (rounded > psize) { 1193 bzero((char *)cbuf + psize, rounded - psize); 1194 psize = rounded; 1195 } 1196 if (psize == lsize) { 1197 compress = ZIO_COMPRESS_OFF; 1198 zio_buf_free(cbuf, lsize); 1199 } else { 1200 zio_push_transform(zio, cbuf, 1201 psize, lsize, NULL); 1202 } 1203 } 1204 } 1205 1206 /* 1207 * The final pass of spa_sync() must be all rewrites, but the first 1208 * few passes offer a trade-off: allocating blocks defers convergence, 1209 * but newly allocated blocks are sequential, so they can be written 1210 * to disk faster. Therefore, we allow the first few passes of 1211 * spa_sync() to allocate new blocks, but force rewrites after that. 1212 * There should only be a handful of blocks after pass 1 in any case. 1213 */ 1214 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg && 1215 BP_GET_PSIZE(bp) == psize && 1216 pass >= zfs_sync_pass_rewrite) { 1217 ASSERT(psize != 0); 1218 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES; 1219 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages; 1220 zio->io_flags |= ZIO_FLAG_IO_REWRITE; 1221 } else { 1222 BP_ZERO(bp); 1223 zio->io_pipeline = ZIO_WRITE_PIPELINE; 1224 } 1225 1226 if (psize == 0) { 1227 if (zio->io_bp_orig.blk_birth != 0 && 1228 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) { 1229 BP_SET_LSIZE(bp, lsize); 1230 BP_SET_TYPE(bp, zp->zp_type); 1231 BP_SET_LEVEL(bp, zp->zp_level); 1232 BP_SET_BIRTH(bp, zio->io_txg, 0); 1233 } 1234 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1235 } else { 1236 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER); 1237 BP_SET_LSIZE(bp, lsize); 1238 BP_SET_TYPE(bp, zp->zp_type); 1239 BP_SET_LEVEL(bp, zp->zp_level); 1240 BP_SET_PSIZE(bp, psize); 1241 BP_SET_COMPRESS(bp, compress); 1242 BP_SET_CHECKSUM(bp, zp->zp_checksum); 1243 BP_SET_DEDUP(bp, zp->zp_dedup); 1244 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER); 1245 if (zp->zp_dedup) { 1246 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1247 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 1248 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE; 1249 } 1250 if (zp->zp_nopwrite) { 1251 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1252 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 1253 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE; 1254 } 1255 } 1256 1257 return (ZIO_PIPELINE_CONTINUE); 1258} 1259 1260static int 1261zio_free_bp_init(zio_t *zio) 1262{ 1263 blkptr_t *bp = zio->io_bp; 1264 1265 if (zio->io_child_type == ZIO_CHILD_LOGICAL) { 1266 if (BP_GET_DEDUP(bp)) 1267 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE; 1268 } 1269 1270 return (ZIO_PIPELINE_CONTINUE); 1271} 1272 1273/* 1274 * ========================================================================== 1275 * Execute the I/O pipeline 1276 * ========================================================================== 1277 */ 1278 1279static void 1280zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline) 1281{ 1282 spa_t *spa = zio->io_spa; 1283 zio_type_t t = zio->io_type; 1284 int flags = (cutinline ? TQ_FRONT : 0); 1285 1286 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT); 1287 1288 /* 1289 * If we're a config writer or a probe, the normal issue and 1290 * interrupt threads may all be blocked waiting for the config lock. 1291 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL. 1292 */ 1293 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE)) 1294 t = ZIO_TYPE_NULL; 1295 1296 /* 1297 * A similar issue exists for the L2ARC write thread until L2ARC 2.0. 1298 */ 1299 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux) 1300 t = ZIO_TYPE_NULL; 1301 1302 /* 1303 * If this is a high priority I/O, then use the high priority taskq if 1304 * available. 1305 */ 1306 if (zio->io_priority == ZIO_PRIORITY_NOW && 1307 spa->spa_zio_taskq[t][q + 1].stqs_count != 0) 1308 q++; 1309 1310 ASSERT3U(q, <, ZIO_TASKQ_TYPES); 1311 1312 /* 1313 * NB: We are assuming that the zio can only be dispatched 1314 * to a single taskq at a time. It would be a grievous error 1315 * to dispatch the zio to another taskq at the same time. 1316 */ 1317#if defined(illumos) || !defined(_KERNEL) 1318 ASSERT(zio->io_tqent.tqent_next == NULL); 1319#else 1320 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0); 1321#endif 1322 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio, 1323 flags, &zio->io_tqent); 1324} 1325 1326static boolean_t 1327zio_taskq_member(zio_t *zio, zio_taskq_type_t q) 1328{ 1329 kthread_t *executor = zio->io_executor; 1330 spa_t *spa = zio->io_spa; 1331 1332 for (zio_type_t t = 0; t < ZIO_TYPES; t++) { 1333 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q]; 1334 uint_t i; 1335 for (i = 0; i < tqs->stqs_count; i++) { 1336 if (taskq_member(tqs->stqs_taskq[i], executor)) 1337 return (B_TRUE); 1338 } 1339 } 1340 1341 return (B_FALSE); 1342} 1343 1344static int 1345zio_issue_async(zio_t *zio) 1346{ 1347 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); 1348 1349 return (ZIO_PIPELINE_STOP); 1350} 1351 1352void 1353zio_interrupt(zio_t *zio) 1354{ 1355 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE); 1356} 1357 1358/* 1359 * Execute the I/O pipeline until one of the following occurs: 1360 * 1361 * (1) the I/O completes 1362 * (2) the pipeline stalls waiting for dependent child I/Os 1363 * (3) the I/O issues, so we're waiting for an I/O completion interrupt 1364 * (4) the I/O is delegated by vdev-level caching or aggregation 1365 * (5) the I/O is deferred due to vdev-level queueing 1366 * (6) the I/O is handed off to another thread. 1367 * 1368 * In all cases, the pipeline stops whenever there's no CPU work; it never 1369 * burns a thread in cv_wait(). 1370 * 1371 * There's no locking on io_stage because there's no legitimate way 1372 * for multiple threads to be attempting to process the same I/O. 1373 */ 1374static zio_pipe_stage_t *zio_pipeline[]; 1375 1376void 1377zio_execute(zio_t *zio) 1378{ 1379 zio->io_executor = curthread; 1380 1381 while (zio->io_stage < ZIO_STAGE_DONE) { 1382 enum zio_stage pipeline = zio->io_pipeline; 1383 enum zio_stage stage = zio->io_stage; 1384 int rv; 1385 1386 ASSERT(!MUTEX_HELD(&zio->io_lock)); 1387 ASSERT(ISP2(stage)); 1388 ASSERT(zio->io_stall == NULL); 1389 1390 do { 1391 stage <<= 1; 1392 } while ((stage & pipeline) == 0); 1393 1394 ASSERT(stage <= ZIO_STAGE_DONE); 1395 1396 /* 1397 * If we are in interrupt context and this pipeline stage 1398 * will grab a config lock that is held across I/O, 1399 * or may wait for an I/O that needs an interrupt thread 1400 * to complete, issue async to avoid deadlock. 1401 * 1402 * For VDEV_IO_START, we cut in line so that the io will 1403 * be sent to disk promptly. 1404 */ 1405 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL && 1406 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) { 1407 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? 1408 zio_requeue_io_start_cut_in_line : B_FALSE; 1409 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); 1410 return; 1411 } 1412 1413 zio->io_stage = stage; 1414 rv = zio_pipeline[highbit64(stage) - 1](zio); 1415 1416 if (rv == ZIO_PIPELINE_STOP) 1417 return; 1418 1419 ASSERT(rv == ZIO_PIPELINE_CONTINUE); 1420 } 1421} 1422 1423/* 1424 * ========================================================================== 1425 * Initiate I/O, either sync or async 1426 * ========================================================================== 1427 */ 1428int 1429zio_wait(zio_t *zio) 1430{ 1431 int error; 1432 1433 ASSERT(zio->io_stage == ZIO_STAGE_OPEN); 1434 ASSERT(zio->io_executor == NULL); 1435 1436 zio->io_waiter = curthread; 1437 1438 zio_execute(zio); 1439 1440 mutex_enter(&zio->io_lock); 1441 while (zio->io_executor != NULL) 1442 cv_wait(&zio->io_cv, &zio->io_lock); 1443 mutex_exit(&zio->io_lock); 1444 1445 error = zio->io_error; 1446 zio_destroy(zio); 1447 1448 return (error); 1449} 1450 1451void 1452zio_nowait(zio_t *zio) 1453{ 1454 ASSERT(zio->io_executor == NULL); 1455 1456 if (zio->io_child_type == ZIO_CHILD_LOGICAL && 1457 zio_unique_parent(zio) == NULL) { 1458 /* 1459 * This is a logical async I/O with no parent to wait for it. 1460 * We add it to the spa_async_root_zio "Godfather" I/O which 1461 * will ensure they complete prior to unloading the pool. 1462 */ 1463 spa_t *spa = zio->io_spa; 1464 1465 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio); 1466 } 1467 1468 zio_execute(zio); 1469} 1470 1471/* 1472 * ========================================================================== 1473 * Reexecute or suspend/resume failed I/O 1474 * ========================================================================== 1475 */ 1476 1477static void 1478zio_reexecute(zio_t *pio) 1479{ 1480 zio_t *cio, *cio_next; 1481 1482 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL); 1483 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN); 1484 ASSERT(pio->io_gang_leader == NULL); 1485 ASSERT(pio->io_gang_tree == NULL); 1486 1487 pio->io_flags = pio->io_orig_flags; 1488 pio->io_stage = pio->io_orig_stage; 1489 pio->io_pipeline = pio->io_orig_pipeline; 1490 pio->io_reexecute = 0; 1491 pio->io_flags |= ZIO_FLAG_REEXECUTED; 1492 pio->io_error = 0; 1493 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 1494 pio->io_state[w] = 0; 1495 for (int c = 0; c < ZIO_CHILD_TYPES; c++) 1496 pio->io_child_error[c] = 0; 1497 1498 if (IO_IS_ALLOCATING(pio)) 1499 BP_ZERO(pio->io_bp); 1500 1501 /* 1502 * As we reexecute pio's children, new children could be created. 1503 * New children go to the head of pio's io_child_list, however, 1504 * so we will (correctly) not reexecute them. The key is that 1505 * the remainder of pio's io_child_list, from 'cio_next' onward, 1506 * cannot be affected by any side effects of reexecuting 'cio'. 1507 */ 1508 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) { 1509 cio_next = zio_walk_children(pio); 1510 mutex_enter(&pio->io_lock); 1511 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 1512 pio->io_children[cio->io_child_type][w]++; 1513 mutex_exit(&pio->io_lock); 1514 zio_reexecute(cio); 1515 } 1516 1517 /* 1518 * Now that all children have been reexecuted, execute the parent. 1519 * We don't reexecute "The Godfather" I/O here as it's the 1520 * responsibility of the caller to wait on him. 1521 */ 1522 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) 1523 zio_execute(pio); 1524} 1525 1526void 1527zio_suspend(spa_t *spa, zio_t *zio) 1528{ 1529 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC) 1530 fm_panic("Pool '%s' has encountered an uncorrectable I/O " 1531 "failure and the failure mode property for this pool " 1532 "is set to panic.", spa_name(spa)); 1533 1534 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0); 1535 1536 mutex_enter(&spa->spa_suspend_lock); 1537 1538 if (spa->spa_suspend_zio_root == NULL) 1539 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL, 1540 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | 1541 ZIO_FLAG_GODFATHER); 1542 1543 spa->spa_suspended = B_TRUE; 1544 1545 if (zio != NULL) { 1546 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); 1547 ASSERT(zio != spa->spa_suspend_zio_root); 1548 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1549 ASSERT(zio_unique_parent(zio) == NULL); 1550 ASSERT(zio->io_stage == ZIO_STAGE_DONE); 1551 zio_add_child(spa->spa_suspend_zio_root, zio); 1552 } 1553 1554 mutex_exit(&spa->spa_suspend_lock); 1555} 1556 1557int 1558zio_resume(spa_t *spa) 1559{ 1560 zio_t *pio; 1561 1562 /* 1563 * Reexecute all previously suspended i/o. 1564 */ 1565 mutex_enter(&spa->spa_suspend_lock); 1566 spa->spa_suspended = B_FALSE; 1567 cv_broadcast(&spa->spa_suspend_cv); 1568 pio = spa->spa_suspend_zio_root; 1569 spa->spa_suspend_zio_root = NULL; 1570 mutex_exit(&spa->spa_suspend_lock); 1571 1572 if (pio == NULL) 1573 return (0); 1574 1575 zio_reexecute(pio); 1576 return (zio_wait(pio)); 1577} 1578 1579void 1580zio_resume_wait(spa_t *spa) 1581{ 1582 mutex_enter(&spa->spa_suspend_lock); 1583 while (spa_suspended(spa)) 1584 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock); 1585 mutex_exit(&spa->spa_suspend_lock); 1586} 1587 1588/* 1589 * ========================================================================== 1590 * Gang blocks. 1591 * 1592 * A gang block is a collection of small blocks that looks to the DMU 1593 * like one large block. When zio_dva_allocate() cannot find a block 1594 * of the requested size, due to either severe fragmentation or the pool 1595 * being nearly full, it calls zio_write_gang_block() to construct the 1596 * block from smaller fragments. 1597 * 1598 * A gang block consists of a gang header (zio_gbh_phys_t) and up to 1599 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like 1600 * an indirect block: it's an array of block pointers. It consumes 1601 * only one sector and hence is allocatable regardless of fragmentation. 1602 * The gang header's bps point to its gang members, which hold the data. 1603 * 1604 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg> 1605 * as the verifier to ensure uniqueness of the SHA256 checksum. 1606 * Critically, the gang block bp's blk_cksum is the checksum of the data, 1607 * not the gang header. This ensures that data block signatures (needed for 1608 * deduplication) are independent of how the block is physically stored. 1609 * 1610 * Gang blocks can be nested: a gang member may itself be a gang block. 1611 * Thus every gang block is a tree in which root and all interior nodes are 1612 * gang headers, and the leaves are normal blocks that contain user data. 1613 * The root of the gang tree is called the gang leader. 1614 * 1615 * To perform any operation (read, rewrite, free, claim) on a gang block, 1616 * zio_gang_assemble() first assembles the gang tree (minus data leaves) 1617 * in the io_gang_tree field of the original logical i/o by recursively 1618 * reading the gang leader and all gang headers below it. This yields 1619 * an in-core tree containing the contents of every gang header and the 1620 * bps for every constituent of the gang block. 1621 * 1622 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree 1623 * and invokes a callback on each bp. To free a gang block, zio_gang_issue() 1624 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp. 1625 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim(). 1626 * zio_read_gang() is a wrapper around zio_read() that omits reading gang 1627 * headers, since we already have those in io_gang_tree. zio_rewrite_gang() 1628 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite() 1629 * of the gang header plus zio_checksum_compute() of the data to update the 1630 * gang header's blk_cksum as described above. 1631 * 1632 * The two-phase assemble/issue model solves the problem of partial failure -- 1633 * what if you'd freed part of a gang block but then couldn't read the 1634 * gang header for another part? Assembling the entire gang tree first 1635 * ensures that all the necessary gang header I/O has succeeded before 1636 * starting the actual work of free, claim, or write. Once the gang tree 1637 * is assembled, free and claim are in-memory operations that cannot fail. 1638 * 1639 * In the event that a gang write fails, zio_dva_unallocate() walks the 1640 * gang tree to immediately free (i.e. insert back into the space map) 1641 * everything we've allocated. This ensures that we don't get ENOSPC 1642 * errors during repeated suspend/resume cycles due to a flaky device. 1643 * 1644 * Gang rewrites only happen during sync-to-convergence. If we can't assemble 1645 * the gang tree, we won't modify the block, so we can safely defer the free 1646 * (knowing that the block is still intact). If we *can* assemble the gang 1647 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free 1648 * each constituent bp and we can allocate a new block on the next sync pass. 1649 * 1650 * In all cases, the gang tree allows complete recovery from partial failure. 1651 * ========================================================================== 1652 */ 1653 1654static zio_t * 1655zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1656{ 1657 if (gn != NULL) 1658 return (pio); 1659 1660 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp), 1661 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), 1662 &pio->io_bookmark)); 1663} 1664 1665zio_t * 1666zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1667{ 1668 zio_t *zio; 1669 1670 if (gn != NULL) { 1671 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, 1672 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority, 1673 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 1674 /* 1675 * As we rewrite each gang header, the pipeline will compute 1676 * a new gang block header checksum for it; but no one will 1677 * compute a new data checksum, so we do that here. The one 1678 * exception is the gang leader: the pipeline already computed 1679 * its data checksum because that stage precedes gang assembly. 1680 * (Presently, nothing actually uses interior data checksums; 1681 * this is just good hygiene.) 1682 */ 1683 if (gn != pio->io_gang_leader->io_gang_tree) { 1684 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp), 1685 data, BP_GET_PSIZE(bp)); 1686 } 1687 /* 1688 * If we are here to damage data for testing purposes, 1689 * leave the GBH alone so that we can detect the damage. 1690 */ 1691 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE) 1692 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; 1693 } else { 1694 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, 1695 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority, 1696 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 1697 } 1698 1699 return (zio); 1700} 1701 1702/* ARGSUSED */ 1703zio_t * 1704zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1705{ 1706 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp, 1707 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp), 1708 ZIO_GANG_CHILD_FLAGS(pio))); 1709} 1710 1711/* ARGSUSED */ 1712zio_t * 1713zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1714{ 1715 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp, 1716 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio))); 1717} 1718 1719static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = { 1720 NULL, 1721 zio_read_gang, 1722 zio_rewrite_gang, 1723 zio_free_gang, 1724 zio_claim_gang, 1725 NULL 1726}; 1727 1728static void zio_gang_tree_assemble_done(zio_t *zio); 1729 1730static zio_gang_node_t * 1731zio_gang_node_alloc(zio_gang_node_t **gnpp) 1732{ 1733 zio_gang_node_t *gn; 1734 1735 ASSERT(*gnpp == NULL); 1736 1737 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP); 1738 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE); 1739 *gnpp = gn; 1740 1741 return (gn); 1742} 1743 1744static void 1745zio_gang_node_free(zio_gang_node_t **gnpp) 1746{ 1747 zio_gang_node_t *gn = *gnpp; 1748 1749 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) 1750 ASSERT(gn->gn_child[g] == NULL); 1751 1752 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE); 1753 kmem_free(gn, sizeof (*gn)); 1754 *gnpp = NULL; 1755} 1756 1757static void 1758zio_gang_tree_free(zio_gang_node_t **gnpp) 1759{ 1760 zio_gang_node_t *gn = *gnpp; 1761 1762 if (gn == NULL) 1763 return; 1764 1765 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) 1766 zio_gang_tree_free(&gn->gn_child[g]); 1767 1768 zio_gang_node_free(gnpp); 1769} 1770 1771static void 1772zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp) 1773{ 1774 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp); 1775 1776 ASSERT(gio->io_gang_leader == gio); 1777 ASSERT(BP_IS_GANG(bp)); 1778 1779 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh, 1780 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn, 1781 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark)); 1782} 1783 1784static void 1785zio_gang_tree_assemble_done(zio_t *zio) 1786{ 1787 zio_t *gio = zio->io_gang_leader; 1788 zio_gang_node_t *gn = zio->io_private; 1789 blkptr_t *bp = zio->io_bp; 1790 1791 ASSERT(gio == zio_unique_parent(zio)); 1792 ASSERT(zio->io_child_count == 0); 1793 1794 if (zio->io_error) 1795 return; 1796 1797 if (BP_SHOULD_BYTESWAP(bp)) 1798 byteswap_uint64_array(zio->io_data, zio->io_size); 1799 1800 ASSERT(zio->io_data == gn->gn_gbh); 1801 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE); 1802 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); 1803 1804 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 1805 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; 1806 if (!BP_IS_GANG(gbp)) 1807 continue; 1808 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]); 1809 } 1810} 1811 1812static void 1813zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data) 1814{ 1815 zio_t *gio = pio->io_gang_leader; 1816 zio_t *zio; 1817 1818 ASSERT(BP_IS_GANG(bp) == !!gn); 1819 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp)); 1820 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree); 1821 1822 /* 1823 * If you're a gang header, your data is in gn->gn_gbh. 1824 * If you're a gang member, your data is in 'data' and gn == NULL. 1825 */ 1826 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data); 1827 1828 if (gn != NULL) { 1829 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); 1830 1831 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 1832 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; 1833 if (BP_IS_HOLE(gbp)) 1834 continue; 1835 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data); 1836 data = (char *)data + BP_GET_PSIZE(gbp); 1837 } 1838 } 1839 1840 if (gn == gio->io_gang_tree && gio->io_data != NULL) 1841 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data); 1842 1843 if (zio != pio) 1844 zio_nowait(zio); 1845} 1846 1847static int 1848zio_gang_assemble(zio_t *zio) 1849{ 1850 blkptr_t *bp = zio->io_bp; 1851 1852 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL); 1853 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 1854 1855 zio->io_gang_leader = zio; 1856 1857 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree); 1858 1859 return (ZIO_PIPELINE_CONTINUE); 1860} 1861 1862static int 1863zio_gang_issue(zio_t *zio) 1864{ 1865 blkptr_t *bp = zio->io_bp; 1866 1867 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE)) 1868 return (ZIO_PIPELINE_STOP); 1869 1870 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio); 1871 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 1872 1873 if (zio->io_child_error[ZIO_CHILD_GANG] == 0) 1874 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data); 1875 else 1876 zio_gang_tree_free(&zio->io_gang_tree); 1877 1878 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1879 1880 return (ZIO_PIPELINE_CONTINUE); 1881} 1882 1883static void 1884zio_write_gang_member_ready(zio_t *zio) 1885{ 1886 zio_t *pio = zio_unique_parent(zio); 1887 zio_t *gio = zio->io_gang_leader; 1888 dva_t *cdva = zio->io_bp->blk_dva; 1889 dva_t *pdva = pio->io_bp->blk_dva; 1890 uint64_t asize; 1891 1892 if (BP_IS_HOLE(zio->io_bp)) 1893 return; 1894 1895 ASSERT(BP_IS_HOLE(&zio->io_bp_orig)); 1896 1897 ASSERT(zio->io_child_type == ZIO_CHILD_GANG); 1898 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies); 1899 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp)); 1900 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp)); 1901 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp)); 1902 1903 mutex_enter(&pio->io_lock); 1904 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) { 1905 ASSERT(DVA_GET_GANG(&pdva[d])); 1906 asize = DVA_GET_ASIZE(&pdva[d]); 1907 asize += DVA_GET_ASIZE(&cdva[d]); 1908 DVA_SET_ASIZE(&pdva[d], asize); 1909 } 1910 mutex_exit(&pio->io_lock); 1911} 1912 1913static int 1914zio_write_gang_block(zio_t *pio) 1915{ 1916 spa_t *spa = pio->io_spa; 1917 blkptr_t *bp = pio->io_bp; 1918 zio_t *gio = pio->io_gang_leader; 1919 zio_t *zio; 1920 zio_gang_node_t *gn, **gnpp; 1921 zio_gbh_phys_t *gbh; 1922 uint64_t txg = pio->io_txg; 1923 uint64_t resid = pio->io_size; 1924 uint64_t lsize; 1925 int copies = gio->io_prop.zp_copies; 1926 int gbh_copies = MIN(copies + 1, spa_max_replication(spa)); 1927 zio_prop_t zp; 1928 int error; 1929 1930 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE, 1931 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, 1932 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER); 1933 if (error) { 1934 pio->io_error = error; 1935 return (ZIO_PIPELINE_CONTINUE); 1936 } 1937 1938 if (pio == gio) { 1939 gnpp = &gio->io_gang_tree; 1940 } else { 1941 gnpp = pio->io_private; 1942 ASSERT(pio->io_ready == zio_write_gang_member_ready); 1943 } 1944 1945 gn = zio_gang_node_alloc(gnpp); 1946 gbh = gn->gn_gbh; 1947 bzero(gbh, SPA_GANGBLOCKSIZE); 1948 1949 /* 1950 * Create the gang header. 1951 */ 1952 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL, 1953 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 1954 1955 /* 1956 * Create and nowait the gang children. 1957 */ 1958 for (int g = 0; resid != 0; resid -= lsize, g++) { 1959 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g), 1960 SPA_MINBLOCKSIZE); 1961 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid); 1962 1963 zp.zp_checksum = gio->io_prop.zp_checksum; 1964 zp.zp_compress = ZIO_COMPRESS_OFF; 1965 zp.zp_type = DMU_OT_NONE; 1966 zp.zp_level = 0; 1967 zp.zp_copies = gio->io_prop.zp_copies; 1968 zp.zp_dedup = B_FALSE; 1969 zp.zp_dedup_verify = B_FALSE; 1970 zp.zp_nopwrite = B_FALSE; 1971 1972 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g], 1973 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp, 1974 zio_write_gang_member_ready, NULL, NULL, &gn->gn_child[g], 1975 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), 1976 &pio->io_bookmark)); 1977 } 1978 1979 /* 1980 * Set pio's pipeline to just wait for zio to finish. 1981 */ 1982 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1983 1984 zio_nowait(zio); 1985 1986 return (ZIO_PIPELINE_CONTINUE); 1987} 1988 1989/* 1990 * The zio_nop_write stage in the pipeline determines if allocating 1991 * a new bp is necessary. By leveraging a cryptographically secure checksum, 1992 * such as SHA256, we can compare the checksums of the new data and the old 1993 * to determine if allocating a new block is required. The nopwrite 1994 * feature can handle writes in either syncing or open context (i.e. zil 1995 * writes) and as a result is mutually exclusive with dedup. 1996 */ 1997static int 1998zio_nop_write(zio_t *zio) 1999{ 2000 blkptr_t *bp = zio->io_bp; 2001 blkptr_t *bp_orig = &zio->io_bp_orig; 2002 zio_prop_t *zp = &zio->io_prop; 2003 2004 ASSERT(BP_GET_LEVEL(bp) == 0); 2005 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 2006 ASSERT(zp->zp_nopwrite); 2007 ASSERT(!zp->zp_dedup); 2008 ASSERT(zio->io_bp_override == NULL); 2009 ASSERT(IO_IS_ALLOCATING(zio)); 2010 2011 /* 2012 * Check to see if the original bp and the new bp have matching 2013 * characteristics (i.e. same checksum, compression algorithms, etc). 2014 * If they don't then just continue with the pipeline which will 2015 * allocate a new bp. 2016 */ 2017 if (BP_IS_HOLE(bp_orig) || 2018 !zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup || 2019 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) || 2020 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) || 2021 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) || 2022 zp->zp_copies != BP_GET_NDVAS(bp_orig)) 2023 return (ZIO_PIPELINE_CONTINUE); 2024 2025 /* 2026 * If the checksums match then reset the pipeline so that we 2027 * avoid allocating a new bp and issuing any I/O. 2028 */ 2029 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) { 2030 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup); 2031 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig)); 2032 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig)); 2033 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF); 2034 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop, 2035 sizeof (uint64_t)) == 0); 2036 2037 *bp = *bp_orig; 2038 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2039 zio->io_flags |= ZIO_FLAG_NOPWRITE; 2040 } 2041 2042 return (ZIO_PIPELINE_CONTINUE); 2043} 2044 2045/* 2046 * ========================================================================== 2047 * Dedup 2048 * ========================================================================== 2049 */ 2050static void 2051zio_ddt_child_read_done(zio_t *zio) 2052{ 2053 blkptr_t *bp = zio->io_bp; 2054 ddt_entry_t *dde = zio->io_private; 2055 ddt_phys_t *ddp; 2056 zio_t *pio = zio_unique_parent(zio); 2057 2058 mutex_enter(&pio->io_lock); 2059 ddp = ddt_phys_select(dde, bp); 2060 if (zio->io_error == 0) 2061 ddt_phys_clear(ddp); /* this ddp doesn't need repair */ 2062 if (zio->io_error == 0 && dde->dde_repair_data == NULL) 2063 dde->dde_repair_data = zio->io_data; 2064 else 2065 zio_buf_free(zio->io_data, zio->io_size); 2066 mutex_exit(&pio->io_lock); 2067} 2068 2069static int 2070zio_ddt_read_start(zio_t *zio) 2071{ 2072 blkptr_t *bp = zio->io_bp; 2073 2074 ASSERT(BP_GET_DEDUP(bp)); 2075 ASSERT(BP_GET_PSIZE(bp) == zio->io_size); 2076 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2077 2078 if (zio->io_child_error[ZIO_CHILD_DDT]) { 2079 ddt_t *ddt = ddt_select(zio->io_spa, bp); 2080 ddt_entry_t *dde = ddt_repair_start(ddt, bp); 2081 ddt_phys_t *ddp = dde->dde_phys; 2082 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp); 2083 blkptr_t blk; 2084 2085 ASSERT(zio->io_vsd == NULL); 2086 zio->io_vsd = dde; 2087 2088 if (ddp_self == NULL) 2089 return (ZIO_PIPELINE_CONTINUE); 2090 2091 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) { 2092 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self) 2093 continue; 2094 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp, 2095 &blk); 2096 zio_nowait(zio_read(zio, zio->io_spa, &blk, 2097 zio_buf_alloc(zio->io_size), zio->io_size, 2098 zio_ddt_child_read_done, dde, zio->io_priority, 2099 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE, 2100 &zio->io_bookmark)); 2101 } 2102 return (ZIO_PIPELINE_CONTINUE); 2103 } 2104 2105 zio_nowait(zio_read(zio, zio->io_spa, bp, 2106 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority, 2107 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark)); 2108 2109 return (ZIO_PIPELINE_CONTINUE); 2110} 2111 2112static int 2113zio_ddt_read_done(zio_t *zio) 2114{ 2115 blkptr_t *bp = zio->io_bp; 2116 2117 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE)) 2118 return (ZIO_PIPELINE_STOP); 2119 2120 ASSERT(BP_GET_DEDUP(bp)); 2121 ASSERT(BP_GET_PSIZE(bp) == zio->io_size); 2122 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2123 2124 if (zio->io_child_error[ZIO_CHILD_DDT]) { 2125 ddt_t *ddt = ddt_select(zio->io_spa, bp); 2126 ddt_entry_t *dde = zio->io_vsd; 2127 if (ddt == NULL) { 2128 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE); 2129 return (ZIO_PIPELINE_CONTINUE); 2130 } 2131 if (dde == NULL) { 2132 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1; 2133 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); 2134 return (ZIO_PIPELINE_STOP); 2135 } 2136 if (dde->dde_repair_data != NULL) { 2137 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size); 2138 zio->io_child_error[ZIO_CHILD_DDT] = 0; 2139 } 2140 ddt_repair_done(ddt, dde); 2141 zio->io_vsd = NULL; 2142 } 2143 2144 ASSERT(zio->io_vsd == NULL); 2145 2146 return (ZIO_PIPELINE_CONTINUE); 2147} 2148 2149static boolean_t 2150zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde) 2151{ 2152 spa_t *spa = zio->io_spa; 2153 2154 /* 2155 * Note: we compare the original data, not the transformed data, 2156 * because when zio->io_bp is an override bp, we will not have 2157 * pushed the I/O transforms. That's an important optimization 2158 * because otherwise we'd compress/encrypt all dmu_sync() data twice. 2159 */ 2160 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { 2161 zio_t *lio = dde->dde_lead_zio[p]; 2162 2163 if (lio != NULL) { 2164 return (lio->io_orig_size != zio->io_orig_size || 2165 bcmp(zio->io_orig_data, lio->io_orig_data, 2166 zio->io_orig_size) != 0); 2167 } 2168 } 2169 2170 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { 2171 ddt_phys_t *ddp = &dde->dde_phys[p]; 2172 2173 if (ddp->ddp_phys_birth != 0) { 2174 arc_buf_t *abuf = NULL; 2175 uint32_t aflags = ARC_WAIT; 2176 blkptr_t blk = *zio->io_bp; 2177 int error; 2178 2179 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth); 2180 2181 ddt_exit(ddt); 2182 2183 error = arc_read(NULL, spa, &blk, 2184 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ, 2185 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2186 &aflags, &zio->io_bookmark); 2187 2188 if (error == 0) { 2189 if (arc_buf_size(abuf) != zio->io_orig_size || 2190 bcmp(abuf->b_data, zio->io_orig_data, 2191 zio->io_orig_size) != 0) 2192 error = SET_ERROR(EEXIST); 2193 VERIFY(arc_buf_remove_ref(abuf, &abuf)); 2194 } 2195 2196 ddt_enter(ddt); 2197 return (error != 0); 2198 } 2199 } 2200 2201 return (B_FALSE); 2202} 2203 2204static void 2205zio_ddt_child_write_ready(zio_t *zio) 2206{ 2207 int p = zio->io_prop.zp_copies; 2208 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); 2209 ddt_entry_t *dde = zio->io_private; 2210 ddt_phys_t *ddp = &dde->dde_phys[p]; 2211 zio_t *pio; 2212 2213 if (zio->io_error) 2214 return; 2215 2216 ddt_enter(ddt); 2217 2218 ASSERT(dde->dde_lead_zio[p] == zio); 2219 2220 ddt_phys_fill(ddp, zio->io_bp); 2221 2222 while ((pio = zio_walk_parents(zio)) != NULL) 2223 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg); 2224 2225 ddt_exit(ddt); 2226} 2227 2228static void 2229zio_ddt_child_write_done(zio_t *zio) 2230{ 2231 int p = zio->io_prop.zp_copies; 2232 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); 2233 ddt_entry_t *dde = zio->io_private; 2234 ddt_phys_t *ddp = &dde->dde_phys[p]; 2235 2236 ddt_enter(ddt); 2237 2238 ASSERT(ddp->ddp_refcnt == 0); 2239 ASSERT(dde->dde_lead_zio[p] == zio); 2240 dde->dde_lead_zio[p] = NULL; 2241 2242 if (zio->io_error == 0) { 2243 while (zio_walk_parents(zio) != NULL) 2244 ddt_phys_addref(ddp); 2245 } else { 2246 ddt_phys_clear(ddp); 2247 } 2248 2249 ddt_exit(ddt); 2250} 2251 2252static void 2253zio_ddt_ditto_write_done(zio_t *zio) 2254{ 2255 int p = DDT_PHYS_DITTO; 2256 zio_prop_t *zp = &zio->io_prop; 2257 blkptr_t *bp = zio->io_bp; 2258 ddt_t *ddt = ddt_select(zio->io_spa, bp); 2259 ddt_entry_t *dde = zio->io_private; 2260 ddt_phys_t *ddp = &dde->dde_phys[p]; 2261 ddt_key_t *ddk = &dde->dde_key; 2262 2263 ddt_enter(ddt); 2264 2265 ASSERT(ddp->ddp_refcnt == 0); 2266 ASSERT(dde->dde_lead_zio[p] == zio); 2267 dde->dde_lead_zio[p] = NULL; 2268 2269 if (zio->io_error == 0) { 2270 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum)); 2271 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP); 2272 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp)); 2273 if (ddp->ddp_phys_birth != 0) 2274 ddt_phys_free(ddt, ddk, ddp, zio->io_txg); 2275 ddt_phys_fill(ddp, bp); 2276 } 2277 2278 ddt_exit(ddt); 2279} 2280 2281static int 2282zio_ddt_write(zio_t *zio) 2283{ 2284 spa_t *spa = zio->io_spa; 2285 blkptr_t *bp = zio->io_bp; 2286 uint64_t txg = zio->io_txg; 2287 zio_prop_t *zp = &zio->io_prop; 2288 int p = zp->zp_copies; 2289 int ditto_copies; 2290 zio_t *cio = NULL; 2291 zio_t *dio = NULL; 2292 ddt_t *ddt = ddt_select(spa, bp); 2293 ddt_entry_t *dde; 2294 ddt_phys_t *ddp; 2295 2296 ASSERT(BP_GET_DEDUP(bp)); 2297 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum); 2298 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override); 2299 2300 ddt_enter(ddt); 2301 dde = ddt_lookup(ddt, bp, B_TRUE); 2302 ddp = &dde->dde_phys[p]; 2303 2304 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) { 2305 /* 2306 * If we're using a weak checksum, upgrade to a strong checksum 2307 * and try again. If we're already using a strong checksum, 2308 * we can't resolve it, so just convert to an ordinary write. 2309 * (And automatically e-mail a paper to Nature?) 2310 */ 2311 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) { 2312 zp->zp_checksum = spa_dedup_checksum(spa); 2313 zio_pop_transforms(zio); 2314 zio->io_stage = ZIO_STAGE_OPEN; 2315 BP_ZERO(bp); 2316 } else { 2317 zp->zp_dedup = B_FALSE; 2318 } 2319 zio->io_pipeline = ZIO_WRITE_PIPELINE; 2320 ddt_exit(ddt); 2321 return (ZIO_PIPELINE_CONTINUE); 2322 } 2323 2324 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp); 2325 ASSERT(ditto_copies < SPA_DVAS_PER_BP); 2326 2327 if (ditto_copies > ddt_ditto_copies_present(dde) && 2328 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) { 2329 zio_prop_t czp = *zp; 2330 2331 czp.zp_copies = ditto_copies; 2332 2333 /* 2334 * If we arrived here with an override bp, we won't have run 2335 * the transform stack, so we won't have the data we need to 2336 * generate a child i/o. So, toss the override bp and restart. 2337 * This is safe, because using the override bp is just an 2338 * optimization; and it's rare, so the cost doesn't matter. 2339 */ 2340 if (zio->io_bp_override) { 2341 zio_pop_transforms(zio); 2342 zio->io_stage = ZIO_STAGE_OPEN; 2343 zio->io_pipeline = ZIO_WRITE_PIPELINE; 2344 zio->io_bp_override = NULL; 2345 BP_ZERO(bp); 2346 ddt_exit(ddt); 2347 return (ZIO_PIPELINE_CONTINUE); 2348 } 2349 2350 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data, 2351 zio->io_orig_size, &czp, NULL, NULL, 2352 zio_ddt_ditto_write_done, dde, zio->io_priority, 2353 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); 2354 2355 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL); 2356 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio; 2357 } 2358 2359 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) { 2360 if (ddp->ddp_phys_birth != 0) 2361 ddt_bp_fill(ddp, bp, txg); 2362 if (dde->dde_lead_zio[p] != NULL) 2363 zio_add_child(zio, dde->dde_lead_zio[p]); 2364 else 2365 ddt_phys_addref(ddp); 2366 } else if (zio->io_bp_override) { 2367 ASSERT(bp->blk_birth == txg); 2368 ASSERT(BP_EQUAL(bp, zio->io_bp_override)); 2369 ddt_phys_fill(ddp, bp); 2370 ddt_phys_addref(ddp); 2371 } else { 2372 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data, 2373 zio->io_orig_size, zp, zio_ddt_child_write_ready, NULL, 2374 zio_ddt_child_write_done, dde, zio->io_priority, 2375 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); 2376 2377 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL); 2378 dde->dde_lead_zio[p] = cio; 2379 } 2380 2381 ddt_exit(ddt); 2382 2383 if (cio) 2384 zio_nowait(cio); 2385 if (dio) 2386 zio_nowait(dio); 2387 2388 return (ZIO_PIPELINE_CONTINUE); 2389} 2390 2391ddt_entry_t *freedde; /* for debugging */ 2392 2393static int 2394zio_ddt_free(zio_t *zio) 2395{ 2396 spa_t *spa = zio->io_spa; 2397 blkptr_t *bp = zio->io_bp; 2398 ddt_t *ddt = ddt_select(spa, bp); 2399 ddt_entry_t *dde; 2400 ddt_phys_t *ddp; 2401 2402 ASSERT(BP_GET_DEDUP(bp)); 2403 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2404 2405 ddt_enter(ddt); 2406 freedde = dde = ddt_lookup(ddt, bp, B_TRUE); 2407 ddp = ddt_phys_select(dde, bp); 2408 ddt_phys_decref(ddp); 2409 ddt_exit(ddt); 2410 2411 return (ZIO_PIPELINE_CONTINUE); 2412} 2413 2414/* 2415 * ========================================================================== 2416 * Allocate and free blocks 2417 * ========================================================================== 2418 */ 2419static int 2420zio_dva_allocate(zio_t *zio) 2421{ 2422 spa_t *spa = zio->io_spa; 2423 metaslab_class_t *mc = spa_normal_class(spa); 2424 blkptr_t *bp = zio->io_bp; 2425 int error; 2426 int flags = 0; 2427 2428 if (zio->io_gang_leader == NULL) { 2429 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 2430 zio->io_gang_leader = zio; 2431 } 2432 2433 ASSERT(BP_IS_HOLE(bp)); 2434 ASSERT0(BP_GET_NDVAS(bp)); 2435 ASSERT3U(zio->io_prop.zp_copies, >, 0); 2436 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa)); 2437 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp)); 2438 2439 /* 2440 * The dump device does not support gang blocks so allocation on 2441 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid 2442 * the "fast" gang feature. 2443 */ 2444 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0; 2445 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ? 2446 METASLAB_GANG_CHILD : 0; 2447 error = metaslab_alloc(spa, mc, zio->io_size, bp, 2448 zio->io_prop.zp_copies, zio->io_txg, NULL, flags); 2449 2450 if (error) { 2451 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, " 2452 "size %llu, error %d", spa_name(spa), zio, zio->io_size, 2453 error); 2454 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE) 2455 return (zio_write_gang_block(zio)); 2456 zio->io_error = error; 2457 } 2458 2459 return (ZIO_PIPELINE_CONTINUE); 2460} 2461 2462static int 2463zio_dva_free(zio_t *zio) 2464{ 2465 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE); 2466 2467 return (ZIO_PIPELINE_CONTINUE); 2468} 2469 2470static int 2471zio_dva_claim(zio_t *zio) 2472{ 2473 int error; 2474 2475 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg); 2476 if (error) 2477 zio->io_error = error; 2478 2479 return (ZIO_PIPELINE_CONTINUE); 2480} 2481 2482/* 2483 * Undo an allocation. This is used by zio_done() when an I/O fails 2484 * and we want to give back the block we just allocated. 2485 * This handles both normal blocks and gang blocks. 2486 */ 2487static void 2488zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp) 2489{ 2490 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp)); 2491 ASSERT(zio->io_bp_override == NULL); 2492 2493 if (!BP_IS_HOLE(bp)) 2494 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE); 2495 2496 if (gn != NULL) { 2497 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 2498 zio_dva_unallocate(zio, gn->gn_child[g], 2499 &gn->gn_gbh->zg_blkptr[g]); 2500 } 2501 } 2502} 2503 2504/* 2505 * Try to allocate an intent log block. Return 0 on success, errno on failure. 2506 */ 2507int 2508zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp, 2509 uint64_t size, boolean_t use_slog) 2510{ 2511 int error = 1; 2512 2513 ASSERT(txg > spa_syncing_txg(spa)); 2514 2515 /* 2516 * ZIL blocks are always contiguous (i.e. not gang blocks) so we 2517 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang" 2518 * when allocating them. 2519 */ 2520 if (use_slog) { 2521 error = metaslab_alloc(spa, spa_log_class(spa), size, 2522 new_bp, 1, txg, old_bp, 2523 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID); 2524 } 2525 2526 if (error) { 2527 error = metaslab_alloc(spa, spa_normal_class(spa), size, 2528 new_bp, 1, txg, old_bp, 2529 METASLAB_HINTBP_AVOID); 2530 } 2531 2532 if (error == 0) { 2533 BP_SET_LSIZE(new_bp, size); 2534 BP_SET_PSIZE(new_bp, size); 2535 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF); 2536 BP_SET_CHECKSUM(new_bp, 2537 spa_version(spa) >= SPA_VERSION_SLIM_ZIL 2538 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG); 2539 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); 2540 BP_SET_LEVEL(new_bp, 0); 2541 BP_SET_DEDUP(new_bp, 0); 2542 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER); 2543 } 2544 2545 return (error); 2546} 2547 2548/* 2549 * Free an intent log block. 2550 */ 2551void 2552zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp) 2553{ 2554 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG); 2555 ASSERT(!BP_IS_GANG(bp)); 2556 2557 zio_free(spa, txg, bp); 2558} 2559 2560/* 2561 * ========================================================================== 2562 * Read, write and delete to physical devices 2563 * ========================================================================== 2564 */ 2565static int 2566zio_vdev_io_start(zio_t *zio) 2567{ 2568 vdev_t *vd = zio->io_vd; 2569 uint64_t align; 2570 spa_t *spa = zio->io_spa; 2571 int ret; 2572 2573 ASSERT(zio->io_error == 0); 2574 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0); 2575 2576 if (vd == NULL) { 2577 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) 2578 spa_config_enter(spa, SCL_ZIO, zio, RW_READER); 2579 2580 /* 2581 * The mirror_ops handle multiple DVAs in a single BP. 2582 */ 2583 return (vdev_mirror_ops.vdev_op_io_start(zio)); 2584 } 2585 2586 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE && 2587 zio->io_priority == ZIO_PRIORITY_NOW) { 2588 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg); 2589 return (ZIO_PIPELINE_CONTINUE); 2590 } 2591 2592 /* 2593 * We keep track of time-sensitive I/Os so that the scan thread 2594 * can quickly react to certain workloads. In particular, we care 2595 * about non-scrubbing, top-level reads and writes with the following 2596 * characteristics: 2597 * - synchronous writes of user data to non-slog devices 2598 * - any reads of user data 2599 * When these conditions are met, adjust the timestamp of spa_last_io 2600 * which allows the scan thread to adjust its workload accordingly. 2601 */ 2602 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL && 2603 vd == vd->vdev_top && !vd->vdev_islog && 2604 zio->io_bookmark.zb_objset != DMU_META_OBJSET && 2605 zio->io_txg != spa_syncing_txg(spa)) { 2606 uint64_t old = spa->spa_last_io; 2607 uint64_t new = ddi_get_lbolt64(); 2608 if (old != new) 2609 (void) atomic_cas_64(&spa->spa_last_io, old, new); 2610 } 2611 2612 align = 1ULL << vd->vdev_top->vdev_ashift; 2613 2614 if ((!(zio->io_flags & ZIO_FLAG_PHYSICAL) || 2615 (vd->vdev_top->vdev_physical_ashift > SPA_MINBLOCKSHIFT)) && 2616 P2PHASE(zio->io_size, align) != 0) { 2617 /* Transform logical writes to be a full physical block size. */ 2618 uint64_t asize = P2ROUNDUP(zio->io_size, align); 2619 char *abuf = NULL; 2620 if (zio->io_type == ZIO_TYPE_READ || 2621 zio->io_type == ZIO_TYPE_WRITE) 2622 abuf = zio_buf_alloc(asize); 2623 ASSERT(vd == vd->vdev_top); 2624 if (zio->io_type == ZIO_TYPE_WRITE) { 2625 bcopy(zio->io_data, abuf, zio->io_size); 2626 bzero(abuf + zio->io_size, asize - zio->io_size); 2627 } 2628 zio_push_transform(zio, abuf, asize, abuf ? asize : 0, 2629 zio_subblock); 2630 } 2631 2632 /* 2633 * If this is not a physical io, make sure that it is properly aligned 2634 * before proceeding. 2635 */ 2636 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) { 2637 ASSERT0(P2PHASE(zio->io_offset, align)); 2638 ASSERT0(P2PHASE(zio->io_size, align)); 2639 } else { 2640 /* 2641 * For physical writes, we allow 512b aligned writes and assume 2642 * the device will perform a read-modify-write as necessary. 2643 */ 2644 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE)); 2645 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE)); 2646 } 2647 2648 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa)); 2649 2650 /* 2651 * If this is a repair I/O, and there's no self-healing involved -- 2652 * that is, we're just resilvering what we expect to resilver -- 2653 * then don't do the I/O unless zio's txg is actually in vd's DTL. 2654 * This prevents spurious resilvering with nested replication. 2655 * For example, given a mirror of mirrors, (A+B)+(C+D), if only 2656 * A is out of date, we'll read from C+D, then use the data to 2657 * resilver A+B -- but we don't actually want to resilver B, just A. 2658 * The top-level mirror has no way to know this, so instead we just 2659 * discard unnecessary repairs as we work our way down the vdev tree. 2660 * The same logic applies to any form of nested replication: 2661 * ditto + mirror, RAID-Z + replacing, etc. This covers them all. 2662 */ 2663 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) && 2664 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) && 2665 zio->io_txg != 0 && /* not a delegated i/o */ 2666 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) { 2667 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 2668 zio_vdev_io_bypass(zio); 2669 return (ZIO_PIPELINE_CONTINUE); 2670 } 2671 2672 if (vd->vdev_ops->vdev_op_leaf) { 2673 switch (zio->io_type) { 2674 case ZIO_TYPE_READ: 2675 if (vdev_cache_read(zio)) 2676 return (ZIO_PIPELINE_CONTINUE); 2677 /* FALLTHROUGH */ 2678 case ZIO_TYPE_WRITE: 2679 case ZIO_TYPE_FREE: 2680 if ((zio = vdev_queue_io(zio)) == NULL) 2681 return (ZIO_PIPELINE_STOP); 2682 2683 if (!vdev_accessible(vd, zio)) { 2684 zio->io_error = SET_ERROR(ENXIO); 2685 zio_interrupt(zio); 2686 return (ZIO_PIPELINE_STOP); 2687 } 2688 break; 2689 } 2690 /* 2691 * Note that we ignore repair writes for TRIM because they can 2692 * conflict with normal writes. This isn't an issue because, by 2693 * definition, we only repair blocks that aren't freed. 2694 */ 2695 if (zio->io_type == ZIO_TYPE_WRITE && 2696 !(zio->io_flags & ZIO_FLAG_IO_REPAIR) && 2697 !trim_map_write_start(zio)) 2698 return (ZIO_PIPELINE_STOP); 2699 } 2700 2701 ret = vd->vdev_ops->vdev_op_io_start(zio); 2702 ASSERT(ret == ZIO_PIPELINE_STOP); 2703 2704 return (ret); 2705} 2706 2707static int 2708zio_vdev_io_done(zio_t *zio) 2709{ 2710 vdev_t *vd = zio->io_vd; 2711 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops; 2712 boolean_t unexpected_error = B_FALSE; 2713 2714 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE)) 2715 return (ZIO_PIPELINE_STOP); 2716 2717 ASSERT(zio->io_type == ZIO_TYPE_READ || 2718 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE); 2719 2720 if (vd != NULL && vd->vdev_ops->vdev_op_leaf && 2721 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE || 2722 zio->io_type == ZIO_TYPE_FREE)) { 2723 2724 if (zio->io_type == ZIO_TYPE_WRITE && 2725 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) 2726 trim_map_write_done(zio); 2727 2728 vdev_queue_io_done(zio); 2729 2730 if (zio->io_type == ZIO_TYPE_WRITE) 2731 vdev_cache_write(zio); 2732 2733 if (zio_injection_enabled && zio->io_error == 0) 2734 zio->io_error = zio_handle_device_injection(vd, 2735 zio, EIO); 2736 2737 if (zio_injection_enabled && zio->io_error == 0) 2738 zio->io_error = zio_handle_label_injection(zio, EIO); 2739 2740 if (zio->io_error) { 2741 if (zio->io_error == ENOTSUP && 2742 zio->io_type == ZIO_TYPE_FREE) { 2743 /* Not all devices support TRIM. */ 2744 } else if (!vdev_accessible(vd, zio)) { 2745 zio->io_error = SET_ERROR(ENXIO); 2746 } else { 2747 unexpected_error = B_TRUE; 2748 } 2749 } 2750 } 2751 2752 ops->vdev_op_io_done(zio); 2753 2754 if (unexpected_error) 2755 VERIFY(vdev_probe(vd, zio) == NULL); 2756 2757 return (ZIO_PIPELINE_CONTINUE); 2758} 2759 2760/* 2761 * For non-raidz ZIOs, we can just copy aside the bad data read from the 2762 * disk, and use that to finish the checksum ereport later. 2763 */ 2764static void 2765zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr, 2766 const void *good_buf) 2767{ 2768 /* no processing needed */ 2769 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE); 2770} 2771 2772/*ARGSUSED*/ 2773void 2774zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored) 2775{ 2776 void *buf = zio_buf_alloc(zio->io_size); 2777 2778 bcopy(zio->io_data, buf, zio->io_size); 2779 2780 zcr->zcr_cbinfo = zio->io_size; 2781 zcr->zcr_cbdata = buf; 2782 zcr->zcr_finish = zio_vsd_default_cksum_finish; 2783 zcr->zcr_free = zio_buf_free; 2784} 2785 2786static int 2787zio_vdev_io_assess(zio_t *zio) 2788{ 2789 vdev_t *vd = zio->io_vd; 2790 2791 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE)) 2792 return (ZIO_PIPELINE_STOP); 2793 2794 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) 2795 spa_config_exit(zio->io_spa, SCL_ZIO, zio); 2796 2797 if (zio->io_vsd != NULL) { 2798 zio->io_vsd_ops->vsd_free(zio); 2799 zio->io_vsd = NULL; 2800 } 2801 2802 if (zio_injection_enabled && zio->io_error == 0) 2803 zio->io_error = zio_handle_fault_injection(zio, EIO); 2804 2805 if (zio->io_type == ZIO_TYPE_FREE && 2806 zio->io_priority != ZIO_PRIORITY_NOW) { 2807 switch (zio->io_error) { 2808 case 0: 2809 ZIO_TRIM_STAT_INCR(bytes, zio->io_size); 2810 ZIO_TRIM_STAT_BUMP(success); 2811 break; 2812 case EOPNOTSUPP: 2813 ZIO_TRIM_STAT_BUMP(unsupported); 2814 break; 2815 default: 2816 ZIO_TRIM_STAT_BUMP(failed); 2817 break; 2818 } 2819 } 2820 2821 /* 2822 * If the I/O failed, determine whether we should attempt to retry it. 2823 * 2824 * On retry, we cut in line in the issue queue, since we don't want 2825 * compression/checksumming/etc. work to prevent our (cheap) IO reissue. 2826 */ 2827 if (zio->io_error && vd == NULL && 2828 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) { 2829 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */ 2830 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */ 2831 zio->io_error = 0; 2832 zio->io_flags |= ZIO_FLAG_IO_RETRY | 2833 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE; 2834 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1; 2835 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, 2836 zio_requeue_io_start_cut_in_line); 2837 return (ZIO_PIPELINE_STOP); 2838 } 2839 2840 /* 2841 * If we got an error on a leaf device, convert it to ENXIO 2842 * if the device is not accessible at all. 2843 */ 2844 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf && 2845 !vdev_accessible(vd, zio)) 2846 zio->io_error = SET_ERROR(ENXIO); 2847 2848 /* 2849 * If we can't write to an interior vdev (mirror or RAID-Z), 2850 * set vdev_cant_write so that we stop trying to allocate from it. 2851 */ 2852 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE && 2853 vd != NULL && !vd->vdev_ops->vdev_op_leaf) { 2854 vd->vdev_cant_write = B_TRUE; 2855 } 2856 2857 if (zio->io_error) 2858 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2859 2860 if (vd != NULL && vd->vdev_ops->vdev_op_leaf && 2861 zio->io_physdone != NULL) { 2862 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED)); 2863 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV); 2864 zio->io_physdone(zio->io_logical); 2865 } 2866 2867 return (ZIO_PIPELINE_CONTINUE); 2868} 2869 2870void 2871zio_vdev_io_reissue(zio_t *zio) 2872{ 2873 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); 2874 ASSERT(zio->io_error == 0); 2875 2876 zio->io_stage >>= 1; 2877} 2878 2879void 2880zio_vdev_io_redone(zio_t *zio) 2881{ 2882 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE); 2883 2884 zio->io_stage >>= 1; 2885} 2886 2887void 2888zio_vdev_io_bypass(zio_t *zio) 2889{ 2890 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); 2891 ASSERT(zio->io_error == 0); 2892 2893 zio->io_flags |= ZIO_FLAG_IO_BYPASS; 2894 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1; 2895} 2896 2897/* 2898 * ========================================================================== 2899 * Generate and verify checksums 2900 * ========================================================================== 2901 */ 2902static int 2903zio_checksum_generate(zio_t *zio) 2904{ 2905 blkptr_t *bp = zio->io_bp; 2906 enum zio_checksum checksum; 2907 2908 if (bp == NULL) { 2909 /* 2910 * This is zio_write_phys(). 2911 * We're either generating a label checksum, or none at all. 2912 */ 2913 checksum = zio->io_prop.zp_checksum; 2914 2915 if (checksum == ZIO_CHECKSUM_OFF) 2916 return (ZIO_PIPELINE_CONTINUE); 2917 2918 ASSERT(checksum == ZIO_CHECKSUM_LABEL); 2919 } else { 2920 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) { 2921 ASSERT(!IO_IS_ALLOCATING(zio)); 2922 checksum = ZIO_CHECKSUM_GANG_HEADER; 2923 } else { 2924 checksum = BP_GET_CHECKSUM(bp); 2925 } 2926 } 2927 2928 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size); 2929 2930 return (ZIO_PIPELINE_CONTINUE); 2931} 2932 2933static int 2934zio_checksum_verify(zio_t *zio) 2935{ 2936 zio_bad_cksum_t info; 2937 blkptr_t *bp = zio->io_bp; 2938 int error; 2939 2940 ASSERT(zio->io_vd != NULL); 2941 2942 if (bp == NULL) { 2943 /* 2944 * This is zio_read_phys(). 2945 * We're either verifying a label checksum, or nothing at all. 2946 */ 2947 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF) 2948 return (ZIO_PIPELINE_CONTINUE); 2949 2950 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL); 2951 } 2952 2953 if ((error = zio_checksum_error(zio, &info)) != 0) { 2954 zio->io_error = error; 2955 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) { 2956 zfs_ereport_start_checksum(zio->io_spa, 2957 zio->io_vd, zio, zio->io_offset, 2958 zio->io_size, NULL, &info); 2959 } 2960 } 2961 2962 return (ZIO_PIPELINE_CONTINUE); 2963} 2964 2965/* 2966 * Called by RAID-Z to ensure we don't compute the checksum twice. 2967 */ 2968void 2969zio_checksum_verified(zio_t *zio) 2970{ 2971 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; 2972} 2973 2974/* 2975 * ========================================================================== 2976 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other. 2977 * An error of 0 indicates success. ENXIO indicates whole-device failure, 2978 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO 2979 * indicate errors that are specific to one I/O, and most likely permanent. 2980 * Any other error is presumed to be worse because we weren't expecting it. 2981 * ========================================================================== 2982 */ 2983int 2984zio_worst_error(int e1, int e2) 2985{ 2986 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO }; 2987 int r1, r2; 2988 2989 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++) 2990 if (e1 == zio_error_rank[r1]) 2991 break; 2992 2993 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++) 2994 if (e2 == zio_error_rank[r2]) 2995 break; 2996 2997 return (r1 > r2 ? e1 : e2); 2998} 2999 3000/* 3001 * ========================================================================== 3002 * I/O completion 3003 * ========================================================================== 3004 */ 3005static int 3006zio_ready(zio_t *zio) 3007{ 3008 blkptr_t *bp = zio->io_bp; 3009 zio_t *pio, *pio_next; 3010 3011 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) || 3012 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY)) 3013 return (ZIO_PIPELINE_STOP); 3014 3015 if (zio->io_ready) { 3016 ASSERT(IO_IS_ALLOCATING(zio)); 3017 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) || 3018 (zio->io_flags & ZIO_FLAG_NOPWRITE)); 3019 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0); 3020 3021 zio->io_ready(zio); 3022 } 3023 3024 if (bp != NULL && bp != &zio->io_bp_copy) 3025 zio->io_bp_copy = *bp; 3026 3027 if (zio->io_error) 3028 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 3029 3030 mutex_enter(&zio->io_lock); 3031 zio->io_state[ZIO_WAIT_READY] = 1; 3032 pio = zio_walk_parents(zio); 3033 mutex_exit(&zio->io_lock); 3034 3035 /* 3036 * As we notify zio's parents, new parents could be added. 3037 * New parents go to the head of zio's io_parent_list, however, 3038 * so we will (correctly) not notify them. The remainder of zio's 3039 * io_parent_list, from 'pio_next' onward, cannot change because 3040 * all parents must wait for us to be done before they can be done. 3041 */ 3042 for (; pio != NULL; pio = pio_next) { 3043 pio_next = zio_walk_parents(zio); 3044 zio_notify_parent(pio, zio, ZIO_WAIT_READY); 3045 } 3046 3047 if (zio->io_flags & ZIO_FLAG_NODATA) { 3048 if (BP_IS_GANG(bp)) { 3049 zio->io_flags &= ~ZIO_FLAG_NODATA; 3050 } else { 3051 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE); 3052 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; 3053 } 3054 } 3055 3056 if (zio_injection_enabled && 3057 zio->io_spa->spa_syncing_txg == zio->io_txg) 3058 zio_handle_ignored_writes(zio); 3059 3060 return (ZIO_PIPELINE_CONTINUE); 3061} 3062 3063static int 3064zio_done(zio_t *zio) 3065{ 3066 spa_t *spa = zio->io_spa; 3067 zio_t *lio = zio->io_logical; 3068 blkptr_t *bp = zio->io_bp; 3069 vdev_t *vd = zio->io_vd; 3070 uint64_t psize = zio->io_size; 3071 zio_t *pio, *pio_next; 3072 3073 /* 3074 * If our children haven't all completed, 3075 * wait for them and then repeat this pipeline stage. 3076 */ 3077 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) || 3078 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) || 3079 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) || 3080 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE)) 3081 return (ZIO_PIPELINE_STOP); 3082 3083 for (int c = 0; c < ZIO_CHILD_TYPES; c++) 3084 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 3085 ASSERT(zio->io_children[c][w] == 0); 3086 3087 if (bp != NULL && !BP_IS_EMBEDDED(bp)) { 3088 ASSERT(bp->blk_pad[0] == 0); 3089 ASSERT(bp->blk_pad[1] == 0); 3090 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 || 3091 (bp == zio_unique_parent(zio)->io_bp)); 3092 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) && 3093 zio->io_bp_override == NULL && 3094 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) { 3095 ASSERT(!BP_SHOULD_BYTESWAP(bp)); 3096 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp)); 3097 ASSERT(BP_COUNT_GANG(bp) == 0 || 3098 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp))); 3099 } 3100 if (zio->io_flags & ZIO_FLAG_NOPWRITE) 3101 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig)); 3102 } 3103 3104 /* 3105 * If there were child vdev/gang/ddt errors, they apply to us now. 3106 */ 3107 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV); 3108 zio_inherit_child_errors(zio, ZIO_CHILD_GANG); 3109 zio_inherit_child_errors(zio, ZIO_CHILD_DDT); 3110 3111 /* 3112 * If the I/O on the transformed data was successful, generate any 3113 * checksum reports now while we still have the transformed data. 3114 */ 3115 if (zio->io_error == 0) { 3116 while (zio->io_cksum_report != NULL) { 3117 zio_cksum_report_t *zcr = zio->io_cksum_report; 3118 uint64_t align = zcr->zcr_align; 3119 uint64_t asize = P2ROUNDUP(psize, align); 3120 char *abuf = zio->io_data; 3121 3122 if (asize != psize) { 3123 abuf = zio_buf_alloc(asize); 3124 bcopy(zio->io_data, abuf, psize); 3125 bzero(abuf + psize, asize - psize); 3126 } 3127 3128 zio->io_cksum_report = zcr->zcr_next; 3129 zcr->zcr_next = NULL; 3130 zcr->zcr_finish(zcr, abuf); 3131 zfs_ereport_free_checksum(zcr); 3132 3133 if (asize != psize) 3134 zio_buf_free(abuf, asize); 3135 } 3136 } 3137 3138 zio_pop_transforms(zio); /* note: may set zio->io_error */ 3139 3140 vdev_stat_update(zio, psize); 3141 3142 if (zio->io_error) { 3143 /* 3144 * If this I/O is attached to a particular vdev, 3145 * generate an error message describing the I/O failure 3146 * at the block level. We ignore these errors if the 3147 * device is currently unavailable. 3148 */ 3149 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd)) 3150 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0); 3151 3152 if ((zio->io_error == EIO || !(zio->io_flags & 3153 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) && 3154 zio == lio) { 3155 /* 3156 * For logical I/O requests, tell the SPA to log the 3157 * error and generate a logical data ereport. 3158 */ 3159 spa_log_error(spa, zio); 3160 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio, 3161 0, 0); 3162 } 3163 } 3164 3165 if (zio->io_error && zio == lio) { 3166 /* 3167 * Determine whether zio should be reexecuted. This will 3168 * propagate all the way to the root via zio_notify_parent(). 3169 */ 3170 ASSERT(vd == NULL && bp != NULL); 3171 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 3172 3173 if (IO_IS_ALLOCATING(zio) && 3174 !(zio->io_flags & ZIO_FLAG_CANFAIL)) { 3175 if (zio->io_error != ENOSPC) 3176 zio->io_reexecute |= ZIO_REEXECUTE_NOW; 3177 else 3178 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 3179 } 3180 3181 if ((zio->io_type == ZIO_TYPE_READ || 3182 zio->io_type == ZIO_TYPE_FREE) && 3183 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && 3184 zio->io_error == ENXIO && 3185 spa_load_state(spa) == SPA_LOAD_NONE && 3186 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE) 3187 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 3188 3189 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute) 3190 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 3191 3192 /* 3193 * Here is a possibly good place to attempt to do 3194 * either combinatorial reconstruction or error correction 3195 * based on checksums. It also might be a good place 3196 * to send out preliminary ereports before we suspend 3197 * processing. 3198 */ 3199 } 3200 3201 /* 3202 * If there were logical child errors, they apply to us now. 3203 * We defer this until now to avoid conflating logical child 3204 * errors with errors that happened to the zio itself when 3205 * updating vdev stats and reporting FMA events above. 3206 */ 3207 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL); 3208 3209 if ((zio->io_error || zio->io_reexecute) && 3210 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio && 3211 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE))) 3212 zio_dva_unallocate(zio, zio->io_gang_tree, bp); 3213 3214 zio_gang_tree_free(&zio->io_gang_tree); 3215 3216 /* 3217 * Godfather I/Os should never suspend. 3218 */ 3219 if ((zio->io_flags & ZIO_FLAG_GODFATHER) && 3220 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) 3221 zio->io_reexecute = 0; 3222 3223 if (zio->io_reexecute) { 3224 /* 3225 * This is a logical I/O that wants to reexecute. 3226 * 3227 * Reexecute is top-down. When an i/o fails, if it's not 3228 * the root, it simply notifies its parent and sticks around. 3229 * The parent, seeing that it still has children in zio_done(), 3230 * does the same. This percolates all the way up to the root. 3231 * The root i/o will reexecute or suspend the entire tree. 3232 * 3233 * This approach ensures that zio_reexecute() honors 3234 * all the original i/o dependency relationships, e.g. 3235 * parents not executing until children are ready. 3236 */ 3237 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 3238 3239 zio->io_gang_leader = NULL; 3240 3241 mutex_enter(&zio->io_lock); 3242 zio->io_state[ZIO_WAIT_DONE] = 1; 3243 mutex_exit(&zio->io_lock); 3244 3245 /* 3246 * "The Godfather" I/O monitors its children but is 3247 * not a true parent to them. It will track them through 3248 * the pipeline but severs its ties whenever they get into 3249 * trouble (e.g. suspended). This allows "The Godfather" 3250 * I/O to return status without blocking. 3251 */ 3252 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) { 3253 zio_link_t *zl = zio->io_walk_link; 3254 pio_next = zio_walk_parents(zio); 3255 3256 if ((pio->io_flags & ZIO_FLAG_GODFATHER) && 3257 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) { 3258 zio_remove_child(pio, zio, zl); 3259 zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 3260 } 3261 } 3262 3263 if ((pio = zio_unique_parent(zio)) != NULL) { 3264 /* 3265 * We're not a root i/o, so there's nothing to do 3266 * but notify our parent. Don't propagate errors 3267 * upward since we haven't permanently failed yet. 3268 */ 3269 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); 3270 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE; 3271 zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 3272 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) { 3273 /* 3274 * We'd fail again if we reexecuted now, so suspend 3275 * until conditions improve (e.g. device comes online). 3276 */ 3277 zio_suspend(spa, zio); 3278 } else { 3279 /* 3280 * Reexecution is potentially a huge amount of work. 3281 * Hand it off to the otherwise-unused claim taskq. 3282 */ 3283#if defined(illumos) || !defined(_KERNEL) 3284 ASSERT(zio->io_tqent.tqent_next == NULL); 3285#else 3286 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0); 3287#endif 3288 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM, 3289 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio, 3290 0, &zio->io_tqent); 3291 } 3292 return (ZIO_PIPELINE_STOP); 3293 } 3294 3295 ASSERT(zio->io_child_count == 0); 3296 ASSERT(zio->io_reexecute == 0); 3297 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL)); 3298 3299 /* 3300 * Report any checksum errors, since the I/O is complete. 3301 */ 3302 while (zio->io_cksum_report != NULL) { 3303 zio_cksum_report_t *zcr = zio->io_cksum_report; 3304 zio->io_cksum_report = zcr->zcr_next; 3305 zcr->zcr_next = NULL; 3306 zcr->zcr_finish(zcr, NULL); 3307 zfs_ereport_free_checksum(zcr); 3308 } 3309 3310 /* 3311 * It is the responsibility of the done callback to ensure that this 3312 * particular zio is no longer discoverable for adoption, and as 3313 * such, cannot acquire any new parents. 3314 */ 3315 if (zio->io_done) 3316 zio->io_done(zio); 3317 3318 mutex_enter(&zio->io_lock); 3319 zio->io_state[ZIO_WAIT_DONE] = 1; 3320 mutex_exit(&zio->io_lock); 3321 3322 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) { 3323 zio_link_t *zl = zio->io_walk_link; 3324 pio_next = zio_walk_parents(zio); 3325 zio_remove_child(pio, zio, zl); 3326 zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 3327 } 3328 3329 if (zio->io_waiter != NULL) { 3330 mutex_enter(&zio->io_lock); 3331 zio->io_executor = NULL; 3332 cv_broadcast(&zio->io_cv); 3333 mutex_exit(&zio->io_lock); 3334 } else { 3335 zio_destroy(zio); 3336 } 3337 3338 return (ZIO_PIPELINE_STOP); 3339} 3340 3341/* 3342 * ========================================================================== 3343 * I/O pipeline definition 3344 * ========================================================================== 3345 */ 3346static zio_pipe_stage_t *zio_pipeline[] = { 3347 NULL, 3348 zio_read_bp_init, 3349 zio_free_bp_init, 3350 zio_issue_async, 3351 zio_write_bp_init, 3352 zio_checksum_generate, 3353 zio_nop_write, 3354 zio_ddt_read_start, 3355 zio_ddt_read_done, 3356 zio_ddt_write, 3357 zio_ddt_free, 3358 zio_gang_assemble, 3359 zio_gang_issue, 3360 zio_dva_allocate, 3361 zio_dva_free, 3362 zio_dva_claim, 3363 zio_ready, 3364 zio_vdev_io_start, 3365 zio_vdev_io_done, 3366 zio_vdev_io_assess, 3367 zio_checksum_verify, 3368 zio_done 3369}; 3370 3371/* dnp is the dnode for zb1->zb_object */ 3372boolean_t 3373zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_phys_t *zb1, 3374 const zbookmark_phys_t *zb2) 3375{ 3376 uint64_t zb1nextL0, zb2thisobj; 3377 3378 ASSERT(zb1->zb_objset == zb2->zb_objset); 3379 ASSERT(zb2->zb_level == 0); 3380 3381 /* The objset_phys_t isn't before anything. */ 3382 if (dnp == NULL) 3383 return (B_FALSE); 3384 3385 zb1nextL0 = (zb1->zb_blkid + 1) << 3386 ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT)); 3387 3388 zb2thisobj = zb2->zb_object ? zb2->zb_object : 3389 zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT); 3390 3391 if (zb1->zb_object == DMU_META_DNODE_OBJECT) { 3392 uint64_t nextobj = zb1nextL0 * 3393 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT; 3394 return (nextobj <= zb2thisobj); 3395 } 3396 3397 if (zb1->zb_object < zb2thisobj) 3398 return (B_TRUE); 3399 if (zb1->zb_object > zb2thisobj) 3400 return (B_FALSE); 3401 if (zb2->zb_object == DMU_META_DNODE_OBJECT) 3402 return (B_FALSE); 3403 return (zb1nextL0 <= zb2->zb_blkid); 3404} 3405