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 2010 Sun Microsystems, Inc. All rights reserved. 23 * Use is subject to license terms. 24 */ 25 26 #include <sys/zfs_context.h> 27 #include <sys/fm/fs/zfs.h> 28 #include <sys/spa.h> 29 #include <sys/txg.h> 30 #include <sys/spa_impl.h> 31 #include <sys/vdev_impl.h> 32 #include <sys/zio_impl.h> 33 #include <sys/zio_compress.h> 34 #include <sys/zio_checksum.h> 35 #include <sys/dmu_objset.h> 36 #include <sys/arc.h> 37 #include <sys/ddt.h> 38 39 /* 40 * ========================================================================== 41 * I/O priority table 42 * ========================================================================== 43 */ 44 uint8_t zio_priority_table[ZIO_PRIORITY_TABLE_SIZE] = { 45 0, /* ZIO_PRIORITY_NOW */ 46 0, /* ZIO_PRIORITY_SYNC_READ */ 47 0, /* ZIO_PRIORITY_SYNC_WRITE */ 48 0, /* ZIO_PRIORITY_LOG_WRITE */ 49 1, /* ZIO_PRIORITY_CACHE_FILL */ 50 1, /* ZIO_PRIORITY_AGG */ 51 4, /* ZIO_PRIORITY_FREE */ 52 4, /* ZIO_PRIORITY_ASYNC_WRITE */ 53 6, /* ZIO_PRIORITY_ASYNC_READ */ 54 10, /* ZIO_PRIORITY_RESILVER */ 55 20, /* ZIO_PRIORITY_SCRUB */ 56 }; 57 58 /* 59 * ========================================================================== 60 * I/O type descriptions 61 * ========================================================================== 62 */ 63 char *zio_type_name[ZIO_TYPES] = { 64 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim", 65 "zio_ioctl" 66 }; 67 68 /* 69 * ========================================================================== 70 * I/O kmem caches 71 * ========================================================================== 72 */ 73 kmem_cache_t *zio_cache; 74 kmem_cache_t *zio_link_cache; 75 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; 76 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; 77 78 #if defined(_KERNEL) && !defined(__NetBSD__) 79 extern vmem_t *zio_alloc_arena; 80 #endif 81 82 /* 83 * An allocating zio is one that either currently has the DVA allocate 84 * stage set or will have it later in its lifetime. 85 */ 86 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE) 87 88 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE; 89 90 #ifdef ZFS_DEBUG 91 int zio_buf_debug_limit = 16384; 92 #else 93 int zio_buf_debug_limit = 0; 94 #endif 95 96 void 97 zio_init(void) 98 { 99 size_t c; 100 vmem_t *data_alloc_arena = NULL; 101 102 #if defined(_KERNEL) && !defined(__NetBSD__) 103 data_alloc_arena = zio_alloc_arena; 104 #endif 105 zio_cache = kmem_cache_create("zio_cache", 106 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0); 107 zio_link_cache = kmem_cache_create("zio_link_cache", 108 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0); 109 110 #ifndef __NetBSD__ 111 /* 112 * For small buffers, we want a cache for each multiple of 113 * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache 114 * for each quarter-power of 2. For large buffers, we want 115 * a cache for each multiple of PAGESIZE. 116 */ 117 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { 118 size_t size = (c + 1) << SPA_MINBLOCKSHIFT; 119 size_t p2 = size; 120 size_t align = 0; 121 122 while (p2 & (p2 - 1)) 123 p2 &= p2 - 1; 124 125 if (size <= 4 * SPA_MINBLOCKSIZE) { 126 align = SPA_MINBLOCKSIZE; 127 } else if (P2PHASE(size, PAGESIZE) == 0) { 128 align = PAGESIZE; 129 } else if (P2PHASE(size, p2 >> 2) == 0) { 130 align = p2 >> 2; 131 } 132 133 if (align != 0) { 134 char name[36]; 135 (void) snprintf(name, sizeof(name), "zio_buf_%lu", 136 (ulong_t)size); 137 zio_buf_cache[c] = kmem_cache_create(name, size, 138 align, NULL, NULL, NULL, NULL, NULL, 139 size > zio_buf_debug_limit ? KMC_NODEBUG : 0); 140 141 (void) snprintf(name, sizeof(name), "zio_data_buf_%lu", 142 (ulong_t)size); 143 zio_data_buf_cache[c] = kmem_cache_create(name, size, 144 align, NULL, NULL, NULL, NULL, data_alloc_arena, 145 size > zio_buf_debug_limit ? KMC_NODEBUG : 0); 146 } 147 } 148 149 while (--c != 0) { 150 ASSERT(zio_buf_cache[c] != NULL); 151 if (zio_buf_cache[c - 1] == NULL) 152 zio_buf_cache[c - 1] = zio_buf_cache[c]; 153 154 ASSERT(zio_data_buf_cache[c] != NULL); 155 if (zio_data_buf_cache[c - 1] == NULL) 156 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c]; 157 } 158 #endif /* __NetBSD__ */ 159 zio_inject_init(); 160 } 161 162 void 163 zio_fini(void) 164 { 165 size_t c; 166 kmem_cache_t *last_cache = NULL; 167 kmem_cache_t *last_data_cache = NULL; 168 169 #ifndef __NetBSD__ 170 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { 171 if (zio_buf_cache[c] != last_cache) { 172 last_cache = zio_buf_cache[c]; 173 kmem_cache_destroy(zio_buf_cache[c]); 174 } 175 zio_buf_cache[c] = NULL; 176 177 if (zio_data_buf_cache[c] != last_data_cache) { 178 last_data_cache = zio_data_buf_cache[c]; 179 kmem_cache_destroy(zio_data_buf_cache[c]); 180 } 181 zio_data_buf_cache[c] = NULL; 182 } 183 #endif /* __NetBSD__ */ 184 185 kmem_cache_destroy(zio_link_cache); 186 kmem_cache_destroy(zio_cache); 187 188 zio_inject_fini(); 189 } 190 191 /* 192 * ========================================================================== 193 * Allocate and free I/O buffers 194 * ========================================================================== 195 */ 196 197 /* 198 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a 199 * crashdump if the kernel panics, so use it judiciously. Obviously, it's 200 * useful to inspect ZFS metadata, but if possible, we should avoid keeping 201 * excess / transient data in-core during a crashdump. 202 */ 203 void * 204 zio_buf_alloc(size_t size) 205 { 206 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 207 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 208 #ifdef __NetBSD__ 209 return (kmem_alloc(size, KM_SLEEP)); 210 #else 211 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE)); 212 #endif 213 } 214 215 /* 216 * Use zio_data_buf_alloc to allocate data. The data will not appear in a 217 * crashdump if the kernel panics. This exists so that we will limit the amount 218 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount 219 * of kernel heap dumped to disk when the kernel panics) 220 */ 221 void * 222 zio_data_buf_alloc(size_t size) 223 { 224 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 225 226 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 227 #ifdef __NetBSD__ 228 return (kmem_alloc(size, KM_SLEEP)); 229 #else 230 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE)); 231 #endif 232 } 233 234 void 235 zio_buf_free(void *buf, size_t size) 236 { 237 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 238 239 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 240 241 #ifdef __NetBSD__ 242 kmem_free(buf, size); 243 #else 244 kmem_cache_free(zio_buf_cache[c], buf); 245 #endif 246 } 247 248 void 249 zio_data_buf_free(void *buf, size_t size) 250 { 251 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 252 253 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 254 255 #ifdef __NetBSD__ 256 kmem_free(buf, size); 257 #else 258 kmem_cache_free(zio_data_buf_cache[c], buf); 259 #endif 260 } 261 262 /* 263 * ========================================================================== 264 * Push and pop I/O transform buffers 265 * ========================================================================== 266 */ 267 static void 268 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize, 269 zio_transform_func_t *transform) 270 { 271 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP); 272 273 zt->zt_orig_data = zio->io_data; 274 zt->zt_orig_size = zio->io_size; 275 zt->zt_bufsize = bufsize; 276 zt->zt_transform = transform; 277 278 zt->zt_next = zio->io_transform_stack; 279 zio->io_transform_stack = zt; 280 281 zio->io_data = data; 282 zio->io_size = size; 283 } 284 285 static void 286 zio_pop_transforms(zio_t *zio) 287 { 288 zio_transform_t *zt; 289 290 while ((zt = zio->io_transform_stack) != NULL) { 291 if (zt->zt_transform != NULL) 292 zt->zt_transform(zio, 293 zt->zt_orig_data, zt->zt_orig_size); 294 295 if (zt->zt_bufsize != 0) 296 zio_buf_free(zio->io_data, zt->zt_bufsize); 297 298 zio->io_data = zt->zt_orig_data; 299 zio->io_size = zt->zt_orig_size; 300 zio->io_transform_stack = zt->zt_next; 301 302 kmem_free(zt, sizeof (zio_transform_t)); 303 } 304 } 305 306 /* 307 * ========================================================================== 308 * I/O transform callbacks for subblocks and decompression 309 * ========================================================================== 310 */ 311 static void 312 zio_subblock(zio_t *zio, void *data, uint64_t size) 313 { 314 ASSERT(zio->io_size > size); 315 316 if (zio->io_type == ZIO_TYPE_READ) 317 bcopy(zio->io_data, data, size); 318 } 319 320 static void 321 zio_decompress(zio_t *zio, void *data, uint64_t size) 322 { 323 if (zio->io_error == 0 && 324 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp), 325 zio->io_data, data, zio->io_size, size) != 0) 326 zio->io_error = EIO; 327 } 328 329 /* 330 * ========================================================================== 331 * I/O parent/child relationships and pipeline interlocks 332 * ========================================================================== 333 */ 334 /* 335 * NOTE - Callers to zio_walk_parents() and zio_walk_children must 336 * continue calling these functions until they return NULL. 337 * Otherwise, the next caller will pick up the list walk in 338 * some indeterminate state. (Otherwise every caller would 339 * have to pass in a cookie to keep the state represented by 340 * io_walk_link, which gets annoying.) 341 */ 342 zio_t * 343 zio_walk_parents(zio_t *cio) 344 { 345 zio_link_t *zl = cio->io_walk_link; 346 list_t *pl = &cio->io_parent_list; 347 348 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl); 349 cio->io_walk_link = zl; 350 351 if (zl == NULL) 352 return (NULL); 353 354 ASSERT(zl->zl_child == cio); 355 return (zl->zl_parent); 356 } 357 358 zio_t * 359 zio_walk_children(zio_t *pio) 360 { 361 zio_link_t *zl = pio->io_walk_link; 362 list_t *cl = &pio->io_child_list; 363 364 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl); 365 pio->io_walk_link = zl; 366 367 if (zl == NULL) 368 return (NULL); 369 370 ASSERT(zl->zl_parent == pio); 371 return (zl->zl_child); 372 } 373 374 zio_t * 375 zio_unique_parent(zio_t *cio) 376 { 377 zio_t *pio = zio_walk_parents(cio); 378 379 VERIFY(zio_walk_parents(cio) == NULL); 380 return (pio); 381 } 382 383 void 384 zio_add_child(zio_t *pio, zio_t *cio) 385 { 386 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP); 387 388 /* 389 * Logical I/Os can have logical, gang, or vdev children. 390 * Gang I/Os can have gang or vdev children. 391 * Vdev I/Os can only have vdev children. 392 * The following ASSERT captures all of these constraints. 393 */ 394 ASSERT(cio->io_child_type <= pio->io_child_type); 395 396 zl->zl_parent = pio; 397 zl->zl_child = cio; 398 399 mutex_enter(&cio->io_lock); 400 mutex_enter(&pio->io_lock); 401 402 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0); 403 404 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 405 pio->io_children[cio->io_child_type][w] += !cio->io_state[w]; 406 407 list_insert_head(&pio->io_child_list, zl); 408 list_insert_head(&cio->io_parent_list, zl); 409 410 pio->io_child_count++; 411 cio->io_parent_count++; 412 413 mutex_exit(&pio->io_lock); 414 mutex_exit(&cio->io_lock); 415 } 416 417 static void 418 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl) 419 { 420 ASSERT(zl->zl_parent == pio); 421 ASSERT(zl->zl_child == cio); 422 423 mutex_enter(&cio->io_lock); 424 mutex_enter(&pio->io_lock); 425 426 list_remove(&pio->io_child_list, zl); 427 list_remove(&cio->io_parent_list, zl); 428 429 pio->io_child_count--; 430 cio->io_parent_count--; 431 432 mutex_exit(&pio->io_lock); 433 mutex_exit(&cio->io_lock); 434 435 kmem_cache_free(zio_link_cache, zl); 436 } 437 438 static boolean_t 439 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait) 440 { 441 uint64_t *countp = &zio->io_children[child][wait]; 442 boolean_t waiting = B_FALSE; 443 444 mutex_enter(&zio->io_lock); 445 ASSERT(zio->io_stall == NULL); 446 if (*countp != 0) { 447 zio->io_stage >>= 1; 448 zio->io_stall = countp; 449 waiting = B_TRUE; 450 } 451 mutex_exit(&zio->io_lock); 452 453 return (waiting); 454 } 455 456 static void 457 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait) 458 { 459 uint64_t *countp = &pio->io_children[zio->io_child_type][wait]; 460 int *errorp = &pio->io_child_error[zio->io_child_type]; 461 462 mutex_enter(&pio->io_lock); 463 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE)) 464 *errorp = zio_worst_error(*errorp, zio->io_error); 465 pio->io_reexecute |= zio->io_reexecute; 466 ASSERT3U(*countp, >, 0); 467 if (--*countp == 0 && pio->io_stall == countp) { 468 pio->io_stall = NULL; 469 mutex_exit(&pio->io_lock); 470 zio_execute(pio); 471 } else { 472 mutex_exit(&pio->io_lock); 473 } 474 } 475 476 static void 477 zio_inherit_child_errors(zio_t *zio, enum zio_child c) 478 { 479 if (zio->io_child_error[c] != 0 && zio->io_error == 0) 480 zio->io_error = zio->io_child_error[c]; 481 } 482 483 /* 484 * ========================================================================== 485 * Create the various types of I/O (read, write, free, etc) 486 * ========================================================================== 487 */ 488 static zio_t * 489 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 490 void *data, uint64_t size, zio_done_func_t *done, void *private, 491 zio_type_t type, int priority, enum zio_flag flags, 492 vdev_t *vd, uint64_t offset, const zbookmark_t *zb, 493 enum zio_stage stage, enum zio_stage pipeline) 494 { 495 zio_t *zio; 496 497 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE); 498 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0); 499 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0); 500 501 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER)); 502 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER)); 503 ASSERT(vd || stage == ZIO_STAGE_OPEN); 504 505 zio = kmem_cache_alloc(zio_cache, KM_SLEEP); 506 bzero(zio, sizeof (zio_t)); 507 508 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL); 509 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL); 510 511 list_create(&zio->io_parent_list, sizeof (zio_link_t), 512 offsetof(zio_link_t, zl_parent_node)); 513 list_create(&zio->io_child_list, sizeof (zio_link_t), 514 offsetof(zio_link_t, zl_child_node)); 515 516 if (vd != NULL) 517 zio->io_child_type = ZIO_CHILD_VDEV; 518 else if (flags & ZIO_FLAG_GANG_CHILD) 519 zio->io_child_type = ZIO_CHILD_GANG; 520 else if (flags & ZIO_FLAG_DDT_CHILD) 521 zio->io_child_type = ZIO_CHILD_DDT; 522 else 523 zio->io_child_type = ZIO_CHILD_LOGICAL; 524 525 if (bp != NULL) { 526 zio->io_bp = (blkptr_t *)bp; 527 zio->io_bp_copy = *bp; 528 zio->io_bp_orig = *bp; 529 if (type != ZIO_TYPE_WRITE || 530 zio->io_child_type == ZIO_CHILD_DDT) 531 zio->io_bp = &zio->io_bp_copy; /* so caller can free */ 532 if (zio->io_child_type == ZIO_CHILD_LOGICAL) 533 zio->io_logical = zio; 534 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp)) 535 pipeline |= ZIO_GANG_STAGES; 536 } 537 538 zio->io_spa = spa; 539 zio->io_txg = txg; 540 zio->io_done = done; 541 zio->io_private = private; 542 zio->io_type = type; 543 zio->io_priority = priority; 544 zio->io_vd = vd; 545 zio->io_offset = offset; 546 zio->io_orig_data = zio->io_data = data; 547 zio->io_orig_size = zio->io_size = size; 548 zio->io_orig_flags = zio->io_flags = flags; 549 zio->io_orig_stage = zio->io_stage = stage; 550 zio->io_orig_pipeline = zio->io_pipeline = pipeline; 551 552 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY); 553 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE); 554 555 if (zb != NULL) 556 zio->io_bookmark = *zb; 557 558 if (pio != NULL) { 559 if (zio->io_logical == NULL) 560 zio->io_logical = pio->io_logical; 561 if (zio->io_child_type == ZIO_CHILD_GANG) 562 zio->io_gang_leader = pio->io_gang_leader; 563 zio_add_child(pio, zio); 564 } 565 566 return (zio); 567 } 568 569 static void 570 zio_destroy(zio_t *zio) 571 { 572 list_destroy(&zio->io_parent_list); 573 list_destroy(&zio->io_child_list); 574 mutex_destroy(&zio->io_lock); 575 cv_destroy(&zio->io_cv); 576 kmem_cache_free(zio_cache, zio); 577 } 578 579 zio_t * 580 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done, 581 void *private, enum zio_flag flags) 582 { 583 zio_t *zio; 584 585 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private, 586 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, 587 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE); 588 589 return (zio); 590 } 591 592 zio_t * 593 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags) 594 { 595 return (zio_null(NULL, spa, NULL, done, private, flags)); 596 } 597 598 zio_t * 599 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, 600 void *data, uint64_t size, zio_done_func_t *done, void *private, 601 int priority, enum zio_flag flags, const zbookmark_t *zb) 602 { 603 zio_t *zio; 604 605 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp, 606 data, size, done, private, 607 ZIO_TYPE_READ, priority, flags, NULL, 0, zb, 608 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? 609 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE); 610 611 return (zio); 612 } 613 614 zio_t * 615 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, 616 void *data, uint64_t size, const zio_prop_t *zp, 617 zio_done_func_t *ready, zio_done_func_t *done, void *private, 618 int priority, enum zio_flag flags, const zbookmark_t *zb) 619 { 620 zio_t *zio; 621 622 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF && 623 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS && 624 zp->zp_compress >= ZIO_COMPRESS_OFF && 625 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS && 626 zp->zp_type < DMU_OT_NUMTYPES && 627 zp->zp_level < 32 && 628 zp->zp_copies > 0 && 629 zp->zp_copies <= spa_max_replication(spa) && 630 zp->zp_dedup <= 1 && 631 zp->zp_dedup_verify <= 1); 632 633 zio = zio_create(pio, spa, txg, bp, data, size, done, private, 634 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, 635 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? 636 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE); 637 638 zio->io_ready = ready; 639 zio->io_prop = *zp; 640 641 return (zio); 642 } 643 644 zio_t * 645 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data, 646 uint64_t size, zio_done_func_t *done, void *private, int priority, 647 enum zio_flag flags, zbookmark_t *zb) 648 { 649 zio_t *zio; 650 651 zio = zio_create(pio, spa, txg, bp, data, size, done, private, 652 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, 653 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE); 654 655 return (zio); 656 } 657 658 void 659 zio_write_override(zio_t *zio, blkptr_t *bp, int copies) 660 { 661 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 662 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 663 ASSERT(zio->io_stage == ZIO_STAGE_OPEN); 664 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa)); 665 666 zio->io_prop.zp_copies = copies; 667 zio->io_bp_override = bp; 668 } 669 670 void 671 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp) 672 { 673 bplist_enqueue_deferred(&spa->spa_free_bplist[txg & TXG_MASK], bp); 674 } 675 676 zio_t * 677 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 678 enum zio_flag flags) 679 { 680 zio_t *zio; 681 682 ASSERT(!BP_IS_HOLE(bp)); 683 ASSERT(spa_syncing_txg(spa) == txg); 684 ASSERT(spa_sync_pass(spa) <= SYNC_PASS_DEFERRED_FREE); 685 686 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), 687 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags, 688 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE); 689 690 return (zio); 691 } 692 693 zio_t * 694 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 695 zio_done_func_t *done, void *private, enum zio_flag flags) 696 { 697 zio_t *zio; 698 699 /* 700 * A claim is an allocation of a specific block. Claims are needed 701 * to support immediate writes in the intent log. The issue is that 702 * immediate writes contain committed data, but in a txg that was 703 * *not* committed. Upon opening the pool after an unclean shutdown, 704 * the intent log claims all blocks that contain immediate write data 705 * so that the SPA knows they're in use. 706 * 707 * All claims *must* be resolved in the first txg -- before the SPA 708 * starts allocating blocks -- so that nothing is allocated twice. 709 * If txg == 0 we just verify that the block is claimable. 710 */ 711 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa)); 712 ASSERT(txg == spa_first_txg(spa) || txg == 0); 713 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */ 714 715 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), 716 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags, 717 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE); 718 719 return (zio); 720 } 721 722 zio_t * 723 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, 724 zio_done_func_t *done, void *private, int priority, enum zio_flag flags) 725 { 726 zio_t *zio; 727 int c; 728 729 if (vd->vdev_children == 0) { 730 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private, 731 ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL, 732 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE); 733 734 zio->io_cmd = cmd; 735 } else { 736 zio = zio_null(pio, spa, NULL, NULL, NULL, flags); 737 738 for (c = 0; c < vd->vdev_children; c++) 739 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd, 740 done, private, priority, flags)); 741 } 742 743 return (zio); 744 } 745 746 zio_t * 747 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 748 void *data, int checksum, zio_done_func_t *done, void *private, 749 int priority, enum zio_flag flags, boolean_t labels) 750 { 751 zio_t *zio; 752 753 ASSERT(vd->vdev_children == 0); 754 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || 755 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); 756 ASSERT3U(offset + size, <=, vd->vdev_psize); 757 758 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private, 759 ZIO_TYPE_READ, priority, flags, vd, offset, NULL, 760 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE); 761 762 zio->io_prop.zp_checksum = checksum; 763 764 return (zio); 765 } 766 767 zio_t * 768 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 769 void *data, int checksum, zio_done_func_t *done, void *private, 770 int priority, enum zio_flag flags, boolean_t labels) 771 { 772 zio_t *zio; 773 774 ASSERT(vd->vdev_children == 0); 775 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || 776 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); 777 ASSERT3U(offset + size, <=, vd->vdev_psize); 778 779 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private, 780 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL, 781 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE); 782 783 zio->io_prop.zp_checksum = checksum; 784 785 if (zio_checksum_table[checksum].ci_eck) { 786 /* 787 * zec checksums are necessarily destructive -- they modify 788 * the end of the write buffer to hold the verifier/checksum. 789 * Therefore, we must make a local copy in case the data is 790 * being written to multiple places in parallel. 791 */ 792 void *wbuf = zio_buf_alloc(size); 793 bcopy(data, wbuf, size); 794 zio_push_transform(zio, wbuf, size, size, NULL); 795 } 796 797 return (zio); 798 } 799 800 /* 801 * Create a child I/O to do some work for us. 802 */ 803 zio_t * 804 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset, 805 void *data, uint64_t size, int type, int priority, enum zio_flag flags, 806 zio_done_func_t *done, void *private) 807 { 808 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE; 809 zio_t *zio; 810 811 ASSERT(vd->vdev_parent == 812 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev)); 813 814 if (type == ZIO_TYPE_READ && bp != NULL) { 815 /* 816 * If we have the bp, then the child should perform the 817 * checksum and the parent need not. This pushes error 818 * detection as close to the leaves as possible and 819 * eliminates redundant checksums in the interior nodes. 820 */ 821 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY; 822 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; 823 } 824 825 if (vd->vdev_children == 0) 826 offset += VDEV_LABEL_START_SIZE; 827 828 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE; 829 830 /* 831 * If we've decided to do a repair, the write is not speculative -- 832 * even if the original read was. 833 */ 834 if (flags & ZIO_FLAG_IO_REPAIR) 835 flags &= ~ZIO_FLAG_SPECULATIVE; 836 837 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, 838 done, private, type, priority, flags, vd, offset, &pio->io_bookmark, 839 ZIO_STAGE_VDEV_IO_START >> 1, pipeline); 840 841 return (zio); 842 } 843 844 zio_t * 845 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size, 846 int type, int priority, enum zio_flag flags, 847 zio_done_func_t *done, void *private) 848 { 849 zio_t *zio; 850 851 ASSERT(vd->vdev_ops->vdev_op_leaf); 852 853 zio = zio_create(NULL, vd->vdev_spa, 0, NULL, 854 data, size, done, private, type, priority, 855 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY, 856 vd, offset, NULL, 857 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE); 858 859 return (zio); 860 } 861 862 void 863 zio_flush(zio_t *zio, vdev_t *vd) 864 { 865 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 866 NULL, NULL, ZIO_PRIORITY_NOW, 867 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY)); 868 } 869 870 void 871 zio_shrink(zio_t *zio, uint64_t size) 872 { 873 ASSERT(zio->io_executor == NULL); 874 ASSERT(zio->io_orig_size == zio->io_size); 875 ASSERT(size <= zio->io_size); 876 877 /* 878 * We don't shrink for raidz because of problems with the 879 * reconstruction when reading back less than the block size. 880 * Note, BP_IS_RAIDZ() assumes no compression. 881 */ 882 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 883 if (!BP_IS_RAIDZ(zio->io_bp)) 884 zio->io_orig_size = zio->io_size = size; 885 } 886 887 /* 888 * ========================================================================== 889 * Prepare to read and write logical blocks 890 * ========================================================================== 891 */ 892 893 static int 894 zio_read_bp_init(zio_t *zio) 895 { 896 blkptr_t *bp = zio->io_bp; 897 898 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF && 899 zio->io_child_type == ZIO_CHILD_LOGICAL && 900 !(zio->io_flags & ZIO_FLAG_RAW)) { 901 uint64_t psize = BP_GET_PSIZE(bp); 902 void *cbuf = zio_buf_alloc(psize); 903 904 zio_push_transform(zio, cbuf, psize, psize, zio_decompress); 905 } 906 907 if (!dmu_ot[BP_GET_TYPE(bp)].ot_metadata && BP_GET_LEVEL(bp) == 0) 908 zio->io_flags |= ZIO_FLAG_DONT_CACHE; 909 910 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP) 911 zio->io_flags |= ZIO_FLAG_DONT_CACHE; 912 913 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL) 914 zio->io_pipeline = ZIO_DDT_READ_PIPELINE; 915 916 return (ZIO_PIPELINE_CONTINUE); 917 } 918 919 static int 920 zio_write_bp_init(zio_t *zio) 921 { 922 spa_t *spa = zio->io_spa; 923 zio_prop_t *zp = &zio->io_prop; 924 enum zio_compress compress = zp->zp_compress; 925 blkptr_t *bp = zio->io_bp; 926 uint64_t lsize = zio->io_size; 927 uint64_t psize = lsize; 928 int pass = 1; 929 930 /* 931 * If our children haven't all reached the ready stage, 932 * wait for them and then repeat this pipeline stage. 933 */ 934 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) || 935 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY)) 936 return (ZIO_PIPELINE_STOP); 937 938 if (!IO_IS_ALLOCATING(zio)) 939 return (ZIO_PIPELINE_CONTINUE); 940 941 ASSERT(zio->io_child_type != ZIO_CHILD_DDT); 942 943 if (zio->io_bp_override) { 944 ASSERT(bp->blk_birth != zio->io_txg); 945 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0); 946 947 *bp = *zio->io_bp_override; 948 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 949 950 if (BP_IS_HOLE(bp) || !zp->zp_dedup) 951 return (ZIO_PIPELINE_CONTINUE); 952 953 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup || 954 zp->zp_dedup_verify); 955 956 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) { 957 BP_SET_DEDUP(bp, 1); 958 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE; 959 return (ZIO_PIPELINE_CONTINUE); 960 } 961 zio->io_bp_override = NULL; 962 BP_ZERO(bp); 963 } 964 965 if (bp->blk_birth == zio->io_txg) { 966 /* 967 * We're rewriting an existing block, which means we're 968 * working on behalf of spa_sync(). For spa_sync() to 969 * converge, it must eventually be the case that we don't 970 * have to allocate new blocks. But compression changes 971 * the blocksize, which forces a reallocate, and makes 972 * convergence take longer. Therefore, after the first 973 * few passes, stop compressing to ensure convergence. 974 */ 975 pass = spa_sync_pass(spa); 976 977 ASSERT(zio->io_txg == spa_syncing_txg(spa)); 978 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 979 ASSERT(!BP_GET_DEDUP(bp)); 980 981 if (pass > SYNC_PASS_DONT_COMPRESS) 982 compress = ZIO_COMPRESS_OFF; 983 984 /* Make sure someone doesn't change their mind on overwrites */ 985 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp), 986 spa_max_replication(spa)) == BP_GET_NDVAS(bp)); 987 } 988 989 if (compress != ZIO_COMPRESS_OFF) { 990 void *cbuf = zio_buf_alloc(lsize); 991 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize); 992 if (psize == 0 || psize == lsize) { 993 compress = ZIO_COMPRESS_OFF; 994 zio_buf_free(cbuf, lsize); 995 } else { 996 ASSERT(psize < lsize); 997 zio_push_transform(zio, cbuf, psize, lsize, NULL); 998 } 999 } 1000 1001 /* 1002 * The final pass of spa_sync() must be all rewrites, but the first 1003 * few passes offer a trade-off: allocating blocks defers convergence, 1004 * but newly allocated blocks are sequential, so they can be written 1005 * to disk faster. Therefore, we allow the first few passes of 1006 * spa_sync() to allocate new blocks, but force rewrites after that. 1007 * There should only be a handful of blocks after pass 1 in any case. 1008 */ 1009 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize && 1010 pass > SYNC_PASS_REWRITE) { 1011 ASSERT(psize != 0); 1012 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES; 1013 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages; 1014 zio->io_flags |= ZIO_FLAG_IO_REWRITE; 1015 } else { 1016 BP_ZERO(bp); 1017 zio->io_pipeline = ZIO_WRITE_PIPELINE; 1018 } 1019 1020 if (psize == 0) { 1021 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1022 } else { 1023 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER); 1024 BP_SET_LSIZE(bp, lsize); 1025 BP_SET_PSIZE(bp, psize); 1026 BP_SET_COMPRESS(bp, compress); 1027 BP_SET_CHECKSUM(bp, zp->zp_checksum); 1028 BP_SET_TYPE(bp, zp->zp_type); 1029 BP_SET_LEVEL(bp, zp->zp_level); 1030 BP_SET_DEDUP(bp, zp->zp_dedup); 1031 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER); 1032 if (zp->zp_dedup) { 1033 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1034 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 1035 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE; 1036 } 1037 } 1038 1039 return (ZIO_PIPELINE_CONTINUE); 1040 } 1041 1042 static int 1043 zio_free_bp_init(zio_t *zio) 1044 { 1045 blkptr_t *bp = zio->io_bp; 1046 1047 if (zio->io_child_type == ZIO_CHILD_LOGICAL) { 1048 if (BP_GET_DEDUP(bp)) 1049 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE; 1050 else 1051 arc_free(zio->io_spa, bp); 1052 } 1053 1054 return (ZIO_PIPELINE_CONTINUE); 1055 } 1056 1057 /* 1058 * ========================================================================== 1059 * Execute the I/O pipeline 1060 * ========================================================================== 1061 */ 1062 1063 static void 1064 zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q, boolean_t cutinline) 1065 { 1066 spa_t *spa = zio->io_spa; 1067 zio_type_t t = zio->io_type; 1068 int flags = TQ_SLEEP | (cutinline ? TQ_FRONT : 0); 1069 1070 /* 1071 * If we're a config writer or a probe, the normal issue and 1072 * interrupt threads may all be blocked waiting for the config lock. 1073 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL. 1074 */ 1075 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE)) 1076 t = ZIO_TYPE_NULL; 1077 1078 /* 1079 * A similar issue exists for the L2ARC write thread until L2ARC 2.0. 1080 */ 1081 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux) 1082 t = ZIO_TYPE_NULL; 1083 1084 /* 1085 * If this is a high priority I/O, then use the high priority taskq. 1086 */ 1087 if (zio->io_priority == ZIO_PRIORITY_NOW && 1088 spa->spa_zio_taskq[t][q + 1] != NULL) 1089 q++; 1090 1091 ASSERT3U(q, <, ZIO_TASKQ_TYPES); 1092 (void) taskq_dispatch(spa->spa_zio_taskq[t][q], 1093 (task_func_t *)zio_execute, zio, flags); 1094 } 1095 1096 static boolean_t 1097 zio_taskq_member(zio_t *zio, enum zio_taskq_type q) 1098 { 1099 kthread_t *executor = zio->io_executor; 1100 spa_t *spa = zio->io_spa; 1101 1102 for (zio_type_t t = 0; t < ZIO_TYPES; t++) 1103 if (taskq_member(spa->spa_zio_taskq[t][q], executor)) 1104 return (B_TRUE); 1105 1106 return (B_FALSE); 1107 } 1108 1109 static int 1110 zio_issue_async(zio_t *zio) 1111 { 1112 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); 1113 1114 return (ZIO_PIPELINE_STOP); 1115 } 1116 1117 void 1118 zio_interrupt(zio_t *zio) 1119 { 1120 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE); 1121 } 1122 1123 /* 1124 * Execute the I/O pipeline until one of the following occurs: 1125 * (1) the I/O completes; (2) the pipeline stalls waiting for 1126 * dependent child I/Os; (3) the I/O issues, so we're waiting 1127 * for an I/O completion interrupt; (4) the I/O is delegated by 1128 * vdev-level caching or aggregation; (5) the I/O is deferred 1129 * due to vdev-level queueing; (6) the I/O is handed off to 1130 * another thread. In all cases, the pipeline stops whenever 1131 * there's no CPU work; it never burns a thread in cv_wait(). 1132 * 1133 * There's no locking on io_stage because there's no legitimate way 1134 * for multiple threads to be attempting to process the same I/O. 1135 */ 1136 static zio_pipe_stage_t *zio_pipeline[]; 1137 1138 void 1139 zio_execute(zio_t *zio) 1140 { 1141 zio->io_executor = curthread; 1142 1143 while (zio->io_stage < ZIO_STAGE_DONE) { 1144 enum zio_stage pipeline = zio->io_pipeline; 1145 enum zio_stage stage = zio->io_stage; 1146 int rv; 1147 1148 ASSERT(!MUTEX_HELD(&zio->io_lock)); 1149 ASSERT(ISP2(stage)); 1150 ASSERT(zio->io_stall == NULL); 1151 1152 do { 1153 stage <<= 1; 1154 } while ((stage & pipeline) == 0); 1155 1156 ASSERT(stage <= ZIO_STAGE_DONE); 1157 1158 /* 1159 * If we are in interrupt context and this pipeline stage 1160 * will grab a config lock that is held across I/O, 1161 * or may wait for an I/O that needs an interrupt thread 1162 * to complete, issue async to avoid deadlock. 1163 * 1164 * For VDEV_IO_START, we cut in line so that the io will 1165 * be sent to disk promptly. 1166 */ 1167 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL && 1168 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) { 1169 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? 1170 zio_requeue_io_start_cut_in_line : B_FALSE; 1171 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); 1172 return; 1173 } 1174 1175 zio->io_stage = stage; 1176 rv = zio_pipeline[highbit(stage) - 1](zio); 1177 1178 if (rv == ZIO_PIPELINE_STOP) 1179 return; 1180 1181 ASSERT(rv == ZIO_PIPELINE_CONTINUE); 1182 } 1183 } 1184 1185 /* 1186 * ========================================================================== 1187 * Initiate I/O, either sync or async 1188 * ========================================================================== 1189 */ 1190 int 1191 zio_wait(zio_t *zio) 1192 { 1193 int error; 1194 1195 ASSERT(zio->io_stage == ZIO_STAGE_OPEN); 1196 ASSERT(zio->io_executor == NULL); 1197 1198 zio->io_waiter = curthread; 1199 1200 zio_execute(zio); 1201 1202 mutex_enter(&zio->io_lock); 1203 while (zio->io_executor != NULL) 1204 cv_wait(&zio->io_cv, &zio->io_lock); 1205 mutex_exit(&zio->io_lock); 1206 1207 error = zio->io_error; 1208 zio_destroy(zio); 1209 1210 return (error); 1211 } 1212 1213 void 1214 zio_nowait(zio_t *zio) 1215 { 1216 ASSERT(zio->io_executor == NULL); 1217 1218 if (zio->io_child_type == ZIO_CHILD_LOGICAL && 1219 zio_unique_parent(zio) == NULL) { 1220 /* 1221 * This is a logical async I/O with no parent to wait for it. 1222 * We add it to the spa_async_root_zio "Godfather" I/O which 1223 * will ensure they complete prior to unloading the pool. 1224 */ 1225 spa_t *spa = zio->io_spa; 1226 1227 zio_add_child(spa->spa_async_zio_root, zio); 1228 } 1229 1230 zio_execute(zio); 1231 } 1232 1233 /* 1234 * ========================================================================== 1235 * Reexecute or suspend/resume failed I/O 1236 * ========================================================================== 1237 */ 1238 1239 static void 1240 zio_reexecute(zio_t *pio) 1241 { 1242 zio_t *cio, *cio_next; 1243 1244 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL); 1245 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN); 1246 ASSERT(pio->io_gang_leader == NULL); 1247 ASSERT(pio->io_gang_tree == NULL); 1248 1249 pio->io_flags = pio->io_orig_flags; 1250 pio->io_stage = pio->io_orig_stage; 1251 pio->io_pipeline = pio->io_orig_pipeline; 1252 pio->io_reexecute = 0; 1253 pio->io_error = 0; 1254 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 1255 pio->io_state[w] = 0; 1256 for (int c = 0; c < ZIO_CHILD_TYPES; c++) 1257 pio->io_child_error[c] = 0; 1258 1259 if (IO_IS_ALLOCATING(pio)) 1260 BP_ZERO(pio->io_bp); 1261 1262 /* 1263 * As we reexecute pio's children, new children could be created. 1264 * New children go to the head of pio's io_child_list, however, 1265 * so we will (correctly) not reexecute them. The key is that 1266 * the remainder of pio's io_child_list, from 'cio_next' onward, 1267 * cannot be affected by any side effects of reexecuting 'cio'. 1268 */ 1269 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) { 1270 cio_next = zio_walk_children(pio); 1271 mutex_enter(&pio->io_lock); 1272 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 1273 pio->io_children[cio->io_child_type][w]++; 1274 mutex_exit(&pio->io_lock); 1275 zio_reexecute(cio); 1276 } 1277 1278 /* 1279 * Now that all children have been reexecuted, execute the parent. 1280 * We don't reexecute "The Godfather" I/O here as it's the 1281 * responsibility of the caller to wait on him. 1282 */ 1283 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) 1284 zio_execute(pio); 1285 } 1286 1287 void 1288 zio_suspend(spa_t *spa, zio_t *zio) 1289 { 1290 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC) 1291 fm_panic("Pool '%s' has encountered an uncorrectable I/O " 1292 "failure and the failure mode property for this pool " 1293 "is set to panic.", spa_name(spa)); 1294 1295 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0); 1296 1297 mutex_enter(&spa->spa_suspend_lock); 1298 1299 if (spa->spa_suspend_zio_root == NULL) 1300 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL, 1301 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | 1302 ZIO_FLAG_GODFATHER); 1303 1304 spa->spa_suspended = B_TRUE; 1305 1306 if (zio != NULL) { 1307 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); 1308 ASSERT(zio != spa->spa_suspend_zio_root); 1309 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1310 ASSERT(zio_unique_parent(zio) == NULL); 1311 ASSERT(zio->io_stage == ZIO_STAGE_DONE); 1312 zio_add_child(spa->spa_suspend_zio_root, zio); 1313 } 1314 1315 mutex_exit(&spa->spa_suspend_lock); 1316 } 1317 1318 int 1319 zio_resume(spa_t *spa) 1320 { 1321 zio_t *pio; 1322 1323 /* 1324 * Reexecute all previously suspended i/o. 1325 */ 1326 mutex_enter(&spa->spa_suspend_lock); 1327 spa->spa_suspended = B_FALSE; 1328 cv_broadcast(&spa->spa_suspend_cv); 1329 pio = spa->spa_suspend_zio_root; 1330 spa->spa_suspend_zio_root = NULL; 1331 mutex_exit(&spa->spa_suspend_lock); 1332 1333 if (pio == NULL) 1334 return (0); 1335 1336 zio_reexecute(pio); 1337 return (zio_wait(pio)); 1338 } 1339 1340 void 1341 zio_resume_wait(spa_t *spa) 1342 { 1343 mutex_enter(&spa->spa_suspend_lock); 1344 while (spa_suspended(spa)) 1345 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock); 1346 mutex_exit(&spa->spa_suspend_lock); 1347 } 1348 1349 /* 1350 * ========================================================================== 1351 * Gang blocks. 1352 * 1353 * A gang block is a collection of small blocks that looks to the DMU 1354 * like one large block. When zio_dva_allocate() cannot find a block 1355 * of the requested size, due to either severe fragmentation or the pool 1356 * being nearly full, it calls zio_write_gang_block() to construct the 1357 * block from smaller fragments. 1358 * 1359 * A gang block consists of a gang header (zio_gbh_phys_t) and up to 1360 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like 1361 * an indirect block: it's an array of block pointers. It consumes 1362 * only one sector and hence is allocatable regardless of fragmentation. 1363 * The gang header's bps point to its gang members, which hold the data. 1364 * 1365 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg> 1366 * as the verifier to ensure uniqueness of the SHA256 checksum. 1367 * Critically, the gang block bp's blk_cksum is the checksum of the data, 1368 * not the gang header. This ensures that data block signatures (needed for 1369 * deduplication) are independent of how the block is physically stored. 1370 * 1371 * Gang blocks can be nested: a gang member may itself be a gang block. 1372 * Thus every gang block is a tree in which root and all interior nodes are 1373 * gang headers, and the leaves are normal blocks that contain user data. 1374 * The root of the gang tree is called the gang leader. 1375 * 1376 * To perform any operation (read, rewrite, free, claim) on a gang block, 1377 * zio_gang_assemble() first assembles the gang tree (minus data leaves) 1378 * in the io_gang_tree field of the original logical i/o by recursively 1379 * reading the gang leader and all gang headers below it. This yields 1380 * an in-core tree containing the contents of every gang header and the 1381 * bps for every constituent of the gang block. 1382 * 1383 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree 1384 * and invokes a callback on each bp. To free a gang block, zio_gang_issue() 1385 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp. 1386 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim(). 1387 * zio_read_gang() is a wrapper around zio_read() that omits reading gang 1388 * headers, since we already have those in io_gang_tree. zio_rewrite_gang() 1389 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite() 1390 * of the gang header plus zio_checksum_compute() of the data to update the 1391 * gang header's blk_cksum as described above. 1392 * 1393 * The two-phase assemble/issue model solves the problem of partial failure -- 1394 * what if you'd freed part of a gang block but then couldn't read the 1395 * gang header for another part? Assembling the entire gang tree first 1396 * ensures that all the necessary gang header I/O has succeeded before 1397 * starting the actual work of free, claim, or write. Once the gang tree 1398 * is assembled, free and claim are in-memory operations that cannot fail. 1399 * 1400 * In the event that a gang write fails, zio_dva_unallocate() walks the 1401 * gang tree to immediately free (i.e. insert back into the space map) 1402 * everything we've allocated. This ensures that we don't get ENOSPC 1403 * errors during repeated suspend/resume cycles due to a flaky device. 1404 * 1405 * Gang rewrites only happen during sync-to-convergence. If we can't assemble 1406 * the gang tree, we won't modify the block, so we can safely defer the free 1407 * (knowing that the block is still intact). If we *can* assemble the gang 1408 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free 1409 * each constituent bp and we can allocate a new block on the next sync pass. 1410 * 1411 * In all cases, the gang tree allows complete recovery from partial failure. 1412 * ========================================================================== 1413 */ 1414 1415 static zio_t * 1416 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1417 { 1418 if (gn != NULL) 1419 return (pio); 1420 1421 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp), 1422 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), 1423 &pio->io_bookmark)); 1424 } 1425 1426 zio_t * 1427 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1428 { 1429 zio_t *zio; 1430 1431 if (gn != NULL) { 1432 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, 1433 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority, 1434 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 1435 /* 1436 * As we rewrite each gang header, the pipeline will compute 1437 * a new gang block header checksum for it; but no one will 1438 * compute a new data checksum, so we do that here. The one 1439 * exception is the gang leader: the pipeline already computed 1440 * its data checksum because that stage precedes gang assembly. 1441 * (Presently, nothing actually uses interior data checksums; 1442 * this is just good hygiene.) 1443 */ 1444 if (gn != pio->io_gang_leader->io_gang_tree) { 1445 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp), 1446 data, BP_GET_PSIZE(bp)); 1447 } 1448 /* 1449 * If we are here to damage data for testing purposes, 1450 * leave the GBH alone so that we can detect the damage. 1451 */ 1452 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE) 1453 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; 1454 } else { 1455 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, 1456 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority, 1457 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 1458 } 1459 1460 return (zio); 1461 } 1462 1463 /* ARGSUSED */ 1464 zio_t * 1465 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1466 { 1467 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp, 1468 ZIO_GANG_CHILD_FLAGS(pio))); 1469 } 1470 1471 /* ARGSUSED */ 1472 zio_t * 1473 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1474 { 1475 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp, 1476 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio))); 1477 } 1478 1479 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = { 1480 NULL, 1481 zio_read_gang, 1482 zio_rewrite_gang, 1483 zio_free_gang, 1484 zio_claim_gang, 1485 NULL 1486 }; 1487 1488 static void zio_gang_tree_assemble_done(zio_t *zio); 1489 1490 static zio_gang_node_t * 1491 zio_gang_node_alloc(zio_gang_node_t **gnpp) 1492 { 1493 zio_gang_node_t *gn; 1494 1495 ASSERT(*gnpp == NULL); 1496 1497 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP); 1498 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE); 1499 *gnpp = gn; 1500 1501 return (gn); 1502 } 1503 1504 static void 1505 zio_gang_node_free(zio_gang_node_t **gnpp) 1506 { 1507 zio_gang_node_t *gn = *gnpp; 1508 1509 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) 1510 ASSERT(gn->gn_child[g] == NULL); 1511 1512 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE); 1513 kmem_free(gn, sizeof (*gn)); 1514 *gnpp = NULL; 1515 } 1516 1517 static void 1518 zio_gang_tree_free(zio_gang_node_t **gnpp) 1519 { 1520 zio_gang_node_t *gn = *gnpp; 1521 1522 if (gn == NULL) 1523 return; 1524 1525 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) 1526 zio_gang_tree_free(&gn->gn_child[g]); 1527 1528 zio_gang_node_free(gnpp); 1529 } 1530 1531 static void 1532 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp) 1533 { 1534 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp); 1535 1536 ASSERT(gio->io_gang_leader == gio); 1537 ASSERT(BP_IS_GANG(bp)); 1538 1539 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh, 1540 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn, 1541 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark)); 1542 } 1543 1544 static void 1545 zio_gang_tree_assemble_done(zio_t *zio) 1546 { 1547 zio_t *gio = zio->io_gang_leader; 1548 zio_gang_node_t *gn = zio->io_private; 1549 blkptr_t *bp = zio->io_bp; 1550 1551 ASSERT(gio == zio_unique_parent(zio)); 1552 ASSERT(zio->io_child_count == 0); 1553 1554 if (zio->io_error) 1555 return; 1556 1557 if (BP_SHOULD_BYTESWAP(bp)) 1558 byteswap_uint64_array(zio->io_data, zio->io_size); 1559 1560 ASSERT(zio->io_data == gn->gn_gbh); 1561 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE); 1562 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); 1563 1564 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 1565 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; 1566 if (!BP_IS_GANG(gbp)) 1567 continue; 1568 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]); 1569 } 1570 } 1571 1572 static void 1573 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data) 1574 { 1575 zio_t *gio = pio->io_gang_leader; 1576 zio_t *zio; 1577 1578 ASSERT(BP_IS_GANG(bp) == !!gn); 1579 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp)); 1580 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree); 1581 1582 /* 1583 * If you're a gang header, your data is in gn->gn_gbh. 1584 * If you're a gang member, your data is in 'data' and gn == NULL. 1585 */ 1586 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data); 1587 1588 if (gn != NULL) { 1589 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); 1590 1591 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 1592 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; 1593 if (BP_IS_HOLE(gbp)) 1594 continue; 1595 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data); 1596 data = (char *)data + BP_GET_PSIZE(gbp); 1597 } 1598 } 1599 1600 if (gn == gio->io_gang_tree) 1601 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data); 1602 1603 if (zio != pio) 1604 zio_nowait(zio); 1605 } 1606 1607 static int 1608 zio_gang_assemble(zio_t *zio) 1609 { 1610 blkptr_t *bp = zio->io_bp; 1611 1612 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL); 1613 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 1614 1615 zio->io_gang_leader = zio; 1616 1617 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree); 1618 1619 return (ZIO_PIPELINE_CONTINUE); 1620 } 1621 1622 static int 1623 zio_gang_issue(zio_t *zio) 1624 { 1625 blkptr_t *bp = zio->io_bp; 1626 1627 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE)) 1628 return (ZIO_PIPELINE_STOP); 1629 1630 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio); 1631 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 1632 1633 if (zio->io_child_error[ZIO_CHILD_GANG] == 0) 1634 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data); 1635 else 1636 zio_gang_tree_free(&zio->io_gang_tree); 1637 1638 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1639 1640 return (ZIO_PIPELINE_CONTINUE); 1641 } 1642 1643 static void 1644 zio_write_gang_member_ready(zio_t *zio) 1645 { 1646 zio_t *pio = zio_unique_parent(zio); 1647 zio_t *gio = zio->io_gang_leader; 1648 dva_t *cdva = zio->io_bp->blk_dva; 1649 dva_t *pdva = pio->io_bp->blk_dva; 1650 uint64_t asize; 1651 1652 if (BP_IS_HOLE(zio->io_bp)) 1653 return; 1654 1655 ASSERT(BP_IS_HOLE(&zio->io_bp_orig)); 1656 1657 ASSERT(zio->io_child_type == ZIO_CHILD_GANG); 1658 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies); 1659 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp)); 1660 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp)); 1661 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp)); 1662 1663 mutex_enter(&pio->io_lock); 1664 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) { 1665 ASSERT(DVA_GET_GANG(&pdva[d])); 1666 asize = DVA_GET_ASIZE(&pdva[d]); 1667 asize += DVA_GET_ASIZE(&cdva[d]); 1668 DVA_SET_ASIZE(&pdva[d], asize); 1669 } 1670 mutex_exit(&pio->io_lock); 1671 } 1672 1673 static int 1674 zio_write_gang_block(zio_t *pio) 1675 { 1676 spa_t *spa = pio->io_spa; 1677 blkptr_t *bp = pio->io_bp; 1678 zio_t *gio = pio->io_gang_leader; 1679 zio_t *zio; 1680 zio_gang_node_t *gn, **gnpp; 1681 zio_gbh_phys_t *gbh; 1682 uint64_t txg = pio->io_txg; 1683 uint64_t resid = pio->io_size; 1684 uint64_t lsize; 1685 int copies = gio->io_prop.zp_copies; 1686 int gbh_copies = MIN(copies + 1, spa_max_replication(spa)); 1687 zio_prop_t zp; 1688 int error; 1689 1690 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE, 1691 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, 1692 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER); 1693 if (error) { 1694 pio->io_error = error; 1695 return (ZIO_PIPELINE_CONTINUE); 1696 } 1697 1698 if (pio == gio) { 1699 gnpp = &gio->io_gang_tree; 1700 } else { 1701 gnpp = pio->io_private; 1702 ASSERT(pio->io_ready == zio_write_gang_member_ready); 1703 } 1704 1705 gn = zio_gang_node_alloc(gnpp); 1706 gbh = gn->gn_gbh; 1707 bzero(gbh, SPA_GANGBLOCKSIZE); 1708 1709 /* 1710 * Create the gang header. 1711 */ 1712 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL, 1713 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 1714 1715 /* 1716 * Create and nowait the gang children. 1717 */ 1718 for (int g = 0; resid != 0; resid -= lsize, g++) { 1719 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g), 1720 SPA_MINBLOCKSIZE); 1721 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid); 1722 1723 zp.zp_checksum = gio->io_prop.zp_checksum; 1724 zp.zp_compress = ZIO_COMPRESS_OFF; 1725 zp.zp_type = DMU_OT_NONE; 1726 zp.zp_level = 0; 1727 zp.zp_copies = gio->io_prop.zp_copies; 1728 zp.zp_dedup = 0; 1729 zp.zp_dedup_verify = 0; 1730 1731 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g], 1732 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp, 1733 zio_write_gang_member_ready, NULL, &gn->gn_child[g], 1734 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), 1735 &pio->io_bookmark)); 1736 } 1737 1738 /* 1739 * Set pio's pipeline to just wait for zio to finish. 1740 */ 1741 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1742 1743 zio_nowait(zio); 1744 1745 return (ZIO_PIPELINE_CONTINUE); 1746 } 1747 1748 /* 1749 * ========================================================================== 1750 * Dedup 1751 * ========================================================================== 1752 */ 1753 static void 1754 zio_ddt_child_read_done(zio_t *zio) 1755 { 1756 blkptr_t *bp = zio->io_bp; 1757 ddt_entry_t *dde = zio->io_private; 1758 ddt_phys_t *ddp; 1759 zio_t *pio = zio_unique_parent(zio); 1760 1761 mutex_enter(&pio->io_lock); 1762 ddp = ddt_phys_select(dde, bp); 1763 if (zio->io_error == 0) 1764 ddt_phys_clear(ddp); /* this ddp doesn't need repair */ 1765 if (zio->io_error == 0 && dde->dde_repair_data == NULL) 1766 dde->dde_repair_data = zio->io_data; 1767 else 1768 zio_buf_free(zio->io_data, zio->io_size); 1769 mutex_exit(&pio->io_lock); 1770 } 1771 1772 static int 1773 zio_ddt_read_start(zio_t *zio) 1774 { 1775 blkptr_t *bp = zio->io_bp; 1776 1777 ASSERT(BP_GET_DEDUP(bp)); 1778 ASSERT(BP_GET_PSIZE(bp) == zio->io_size); 1779 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1780 1781 if (zio->io_child_error[ZIO_CHILD_DDT]) { 1782 ddt_t *ddt = ddt_select(zio->io_spa, bp); 1783 ddt_entry_t *dde = ddt_repair_start(ddt, bp); 1784 ddt_phys_t *ddp = dde->dde_phys; 1785 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp); 1786 blkptr_t blk; 1787 1788 ASSERT(zio->io_vsd == NULL); 1789 zio->io_vsd = dde; 1790 1791 if (ddp_self == NULL) 1792 return (ZIO_PIPELINE_CONTINUE); 1793 1794 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) { 1795 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self) 1796 continue; 1797 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp, 1798 &blk); 1799 zio_nowait(zio_read(zio, zio->io_spa, &blk, 1800 zio_buf_alloc(zio->io_size), zio->io_size, 1801 zio_ddt_child_read_done, dde, zio->io_priority, 1802 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE, 1803 &zio->io_bookmark)); 1804 } 1805 return (ZIO_PIPELINE_CONTINUE); 1806 } 1807 1808 zio_nowait(zio_read(zio, zio->io_spa, bp, 1809 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority, 1810 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark)); 1811 1812 return (ZIO_PIPELINE_CONTINUE); 1813 } 1814 1815 static int 1816 zio_ddt_read_done(zio_t *zio) 1817 { 1818 blkptr_t *bp = zio->io_bp; 1819 1820 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE)) 1821 return (ZIO_PIPELINE_STOP); 1822 1823 ASSERT(BP_GET_DEDUP(bp)); 1824 ASSERT(BP_GET_PSIZE(bp) == zio->io_size); 1825 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1826 1827 if (zio->io_child_error[ZIO_CHILD_DDT]) { 1828 ddt_t *ddt = ddt_select(zio->io_spa, bp); 1829 ddt_entry_t *dde = zio->io_vsd; 1830 if (ddt == NULL) { 1831 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE); 1832 return (ZIO_PIPELINE_CONTINUE); 1833 } 1834 if (dde == NULL) { 1835 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1; 1836 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); 1837 return (ZIO_PIPELINE_STOP); 1838 } 1839 if (dde->dde_repair_data != NULL) { 1840 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size); 1841 zio->io_child_error[ZIO_CHILD_DDT] = 0; 1842 } 1843 ddt_repair_done(ddt, dde); 1844 zio->io_vsd = NULL; 1845 } 1846 1847 ASSERT(zio->io_vsd == NULL); 1848 1849 return (ZIO_PIPELINE_CONTINUE); 1850 } 1851 1852 static boolean_t 1853 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde) 1854 { 1855 spa_t *spa = zio->io_spa; 1856 1857 /* 1858 * Note: we compare the original data, not the transformed data, 1859 * because when zio->io_bp is an override bp, we will not have 1860 * pushed the I/O transforms. That's an important optimization 1861 * because otherwise we'd compress/encrypt all dmu_sync() data twice. 1862 */ 1863 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { 1864 zio_t *lio = dde->dde_lead_zio[p]; 1865 1866 if (lio != NULL) { 1867 return (lio->io_orig_size != zio->io_orig_size || 1868 bcmp(zio->io_orig_data, lio->io_orig_data, 1869 zio->io_orig_size) != 0); 1870 } 1871 } 1872 1873 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { 1874 ddt_phys_t *ddp = &dde->dde_phys[p]; 1875 1876 if (ddp->ddp_phys_birth != 0) { 1877 arc_buf_t *abuf = NULL; 1878 uint32_t aflags = ARC_WAIT; 1879 blkptr_t blk = *zio->io_bp; 1880 int error; 1881 1882 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth); 1883 1884 ddt_exit(ddt); 1885 1886 error = arc_read_nolock(NULL, spa, &blk, 1887 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ, 1888 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 1889 &aflags, &zio->io_bookmark); 1890 1891 if (error == 0) { 1892 if (arc_buf_size(abuf) != zio->io_orig_size || 1893 bcmp(abuf->b_data, zio->io_orig_data, 1894 zio->io_orig_size) != 0) 1895 error = EEXIST; 1896 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1); 1897 } 1898 1899 ddt_enter(ddt); 1900 return (error != 0); 1901 } 1902 } 1903 1904 return (B_FALSE); 1905 } 1906 1907 static void 1908 zio_ddt_child_write_ready(zio_t *zio) 1909 { 1910 int p = zio->io_prop.zp_copies; 1911 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); 1912 ddt_entry_t *dde = zio->io_private; 1913 ddt_phys_t *ddp = &dde->dde_phys[p]; 1914 zio_t *pio; 1915 1916 if (zio->io_error) 1917 return; 1918 1919 ddt_enter(ddt); 1920 1921 ASSERT(dde->dde_lead_zio[p] == zio); 1922 1923 ddt_phys_fill(ddp, zio->io_bp); 1924 1925 while ((pio = zio_walk_parents(zio)) != NULL) 1926 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg); 1927 1928 ddt_exit(ddt); 1929 } 1930 1931 static void 1932 zio_ddt_child_write_done(zio_t *zio) 1933 { 1934 int p = zio->io_prop.zp_copies; 1935 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); 1936 ddt_entry_t *dde = zio->io_private; 1937 ddt_phys_t *ddp = &dde->dde_phys[p]; 1938 1939 ddt_enter(ddt); 1940 1941 ASSERT(ddp->ddp_refcnt == 0); 1942 ASSERT(dde->dde_lead_zio[p] == zio); 1943 dde->dde_lead_zio[p] = NULL; 1944 1945 if (zio->io_error == 0) { 1946 while (zio_walk_parents(zio) != NULL) 1947 ddt_phys_addref(ddp); 1948 } else { 1949 ddt_phys_clear(ddp); 1950 } 1951 1952 ddt_exit(ddt); 1953 } 1954 1955 static void 1956 zio_ddt_ditto_write_done(zio_t *zio) 1957 { 1958 int p = DDT_PHYS_DITTO; 1959 zio_prop_t *zp = &zio->io_prop; 1960 blkptr_t *bp = zio->io_bp; 1961 ddt_t *ddt = ddt_select(zio->io_spa, bp); 1962 ddt_entry_t *dde = zio->io_private; 1963 ddt_phys_t *ddp = &dde->dde_phys[p]; 1964 ddt_key_t *ddk = &dde->dde_key; 1965 1966 ddt_enter(ddt); 1967 1968 ASSERT(ddp->ddp_refcnt == 0); 1969 ASSERT(dde->dde_lead_zio[p] == zio); 1970 dde->dde_lead_zio[p] = NULL; 1971 1972 if (zio->io_error == 0) { 1973 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum)); 1974 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP); 1975 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp)); 1976 if (ddp->ddp_phys_birth != 0) 1977 ddt_phys_free(ddt, ddk, ddp, zio->io_txg); 1978 ddt_phys_fill(ddp, bp); 1979 } 1980 1981 ddt_exit(ddt); 1982 } 1983 1984 static int 1985 zio_ddt_write(zio_t *zio) 1986 { 1987 spa_t *spa = zio->io_spa; 1988 blkptr_t *bp = zio->io_bp; 1989 uint64_t txg = zio->io_txg; 1990 zio_prop_t *zp = &zio->io_prop; 1991 int p = zp->zp_copies; 1992 int ditto_copies; 1993 zio_t *cio = NULL; 1994 zio_t *dio = NULL; 1995 ddt_t *ddt = ddt_select(spa, bp); 1996 ddt_entry_t *dde; 1997 ddt_phys_t *ddp; 1998 1999 ASSERT(BP_GET_DEDUP(bp)); 2000 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum); 2001 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override); 2002 2003 ddt_enter(ddt); 2004 dde = ddt_lookup(ddt, bp, B_TRUE); 2005 ddp = &dde->dde_phys[p]; 2006 2007 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) { 2008 /* 2009 * If we're using a weak checksum, upgrade to a strong checksum 2010 * and try again. If we're already using a strong checksum, 2011 * we can't resolve it, so just convert to an ordinary write. 2012 * (And automatically e-mail a paper to Nature?) 2013 */ 2014 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) { 2015 zp->zp_checksum = spa_dedup_checksum(spa); 2016 zio_pop_transforms(zio); 2017 zio->io_stage = ZIO_STAGE_OPEN; 2018 BP_ZERO(bp); 2019 } else { 2020 zp->zp_dedup = 0; 2021 } 2022 zio->io_pipeline = ZIO_WRITE_PIPELINE; 2023 ddt_exit(ddt); 2024 return (ZIO_PIPELINE_CONTINUE); 2025 } 2026 2027 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp); 2028 ASSERT(ditto_copies < SPA_DVAS_PER_BP); 2029 2030 if (ditto_copies > ddt_ditto_copies_present(dde) && 2031 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) { 2032 zio_prop_t czp = *zp; 2033 2034 czp.zp_copies = ditto_copies; 2035 2036 /* 2037 * If we arrived here with an override bp, we won't have run 2038 * the transform stack, so we won't have the data we need to 2039 * generate a child i/o. So, toss the override bp and restart. 2040 * This is safe, because using the override bp is just an 2041 * optimization; and it's rare, so the cost doesn't matter. 2042 */ 2043 if (zio->io_bp_override) { 2044 zio_pop_transforms(zio); 2045 zio->io_stage = ZIO_STAGE_OPEN; 2046 zio->io_pipeline = ZIO_WRITE_PIPELINE; 2047 zio->io_bp_override = NULL; 2048 BP_ZERO(bp); 2049 ddt_exit(ddt); 2050 return (ZIO_PIPELINE_CONTINUE); 2051 } 2052 2053 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data, 2054 zio->io_orig_size, &czp, NULL, 2055 zio_ddt_ditto_write_done, dde, zio->io_priority, 2056 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); 2057 2058 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL); 2059 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio; 2060 } 2061 2062 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) { 2063 if (ddp->ddp_phys_birth != 0) 2064 ddt_bp_fill(ddp, bp, txg); 2065 if (dde->dde_lead_zio[p] != NULL) 2066 zio_add_child(zio, dde->dde_lead_zio[p]); 2067 else 2068 ddt_phys_addref(ddp); 2069 } else if (zio->io_bp_override) { 2070 ASSERT(bp->blk_birth == txg); 2071 ASSERT(BP_EQUAL(bp, zio->io_bp_override)); 2072 ddt_phys_fill(ddp, bp); 2073 ddt_phys_addref(ddp); 2074 } else { 2075 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data, 2076 zio->io_orig_size, zp, zio_ddt_child_write_ready, 2077 zio_ddt_child_write_done, dde, zio->io_priority, 2078 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); 2079 2080 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL); 2081 dde->dde_lead_zio[p] = cio; 2082 } 2083 2084 ddt_exit(ddt); 2085 2086 if (cio) 2087 zio_nowait(cio); 2088 if (dio) 2089 zio_nowait(dio); 2090 2091 return (ZIO_PIPELINE_CONTINUE); 2092 } 2093 2094 static int 2095 zio_ddt_free(zio_t *zio) 2096 { 2097 spa_t *spa = zio->io_spa; 2098 blkptr_t *bp = zio->io_bp; 2099 ddt_t *ddt = ddt_select(spa, bp); 2100 ddt_entry_t *dde; 2101 ddt_phys_t *ddp; 2102 2103 ASSERT(BP_GET_DEDUP(bp)); 2104 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2105 2106 ddt_enter(ddt); 2107 dde = ddt_lookup(ddt, bp, B_TRUE); 2108 ddp = ddt_phys_select(dde, bp); 2109 ddt_phys_decref(ddp); 2110 ddt_exit(ddt); 2111 2112 return (ZIO_PIPELINE_CONTINUE); 2113 } 2114 2115 /* 2116 * ========================================================================== 2117 * Allocate and free blocks 2118 * ========================================================================== 2119 */ 2120 static int 2121 zio_dva_allocate(zio_t *zio) 2122 { 2123 spa_t *spa = zio->io_spa; 2124 metaslab_class_t *mc = spa_normal_class(spa); 2125 blkptr_t *bp = zio->io_bp; 2126 int error; 2127 2128 if (zio->io_gang_leader == NULL) { 2129 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 2130 zio->io_gang_leader = zio; 2131 } 2132 2133 ASSERT(BP_IS_HOLE(bp)); 2134 ASSERT3U(BP_GET_NDVAS(bp), ==, 0); 2135 ASSERT3U(zio->io_prop.zp_copies, >, 0); 2136 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa)); 2137 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp)); 2138 2139 error = metaslab_alloc(spa, mc, zio->io_size, bp, 2140 zio->io_prop.zp_copies, zio->io_txg, NULL, 0); 2141 2142 if (error) { 2143 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE) 2144 return (zio_write_gang_block(zio)); 2145 zio->io_error = error; 2146 } 2147 2148 return (ZIO_PIPELINE_CONTINUE); 2149 } 2150 2151 static int 2152 zio_dva_free(zio_t *zio) 2153 { 2154 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE); 2155 2156 return (ZIO_PIPELINE_CONTINUE); 2157 } 2158 2159 static int 2160 zio_dva_claim(zio_t *zio) 2161 { 2162 int error; 2163 2164 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg); 2165 if (error) 2166 zio->io_error = error; 2167 2168 return (ZIO_PIPELINE_CONTINUE); 2169 } 2170 2171 /* 2172 * Undo an allocation. This is used by zio_done() when an I/O fails 2173 * and we want to give back the block we just allocated. 2174 * This handles both normal blocks and gang blocks. 2175 */ 2176 static void 2177 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp) 2178 { 2179 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp)); 2180 ASSERT(zio->io_bp_override == NULL); 2181 2182 if (!BP_IS_HOLE(bp)) 2183 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE); 2184 2185 if (gn != NULL) { 2186 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 2187 zio_dva_unallocate(zio, gn->gn_child[g], 2188 &gn->gn_gbh->zg_blkptr[g]); 2189 } 2190 } 2191 } 2192 2193 /* 2194 * Try to allocate an intent log block. Return 0 on success, errno on failure. 2195 */ 2196 int 2197 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp, 2198 uint64_t size, boolean_t use_slog) 2199 { 2200 int error = 1; 2201 2202 ASSERT(txg > spa_syncing_txg(spa)); 2203 2204 if (use_slog) 2205 error = metaslab_alloc(spa, spa_log_class(spa), size, 2206 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID); 2207 2208 if (error) 2209 error = metaslab_alloc(spa, spa_normal_class(spa), size, 2210 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID); 2211 2212 if (error == 0) { 2213 BP_SET_LSIZE(new_bp, size); 2214 BP_SET_PSIZE(new_bp, size); 2215 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF); 2216 BP_SET_CHECKSUM(new_bp, 2217 spa_version(spa) >= SPA_VERSION_SLIM_ZIL 2218 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG); 2219 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); 2220 BP_SET_LEVEL(new_bp, 0); 2221 BP_SET_DEDUP(new_bp, 0); 2222 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER); 2223 } 2224 2225 return (error); 2226 } 2227 2228 /* 2229 * Free an intent log block. 2230 */ 2231 void 2232 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp) 2233 { 2234 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG); 2235 ASSERT(!BP_IS_GANG(bp)); 2236 2237 zio_free(spa, txg, bp); 2238 } 2239 2240 /* 2241 * ========================================================================== 2242 * Read and write to physical devices 2243 * ========================================================================== 2244 */ 2245 static int 2246 zio_vdev_io_start(zio_t *zio) 2247 { 2248 vdev_t *vd = zio->io_vd; 2249 uint64_t align; 2250 spa_t *spa = zio->io_spa; 2251 2252 ASSERT(zio->io_error == 0); 2253 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0); 2254 2255 if (vd == NULL) { 2256 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) 2257 spa_config_enter(spa, SCL_ZIO, zio, RW_READER); 2258 2259 /* 2260 * The mirror_ops handle multiple DVAs in a single BP. 2261 */ 2262 return (vdev_mirror_ops.vdev_op_io_start(zio)); 2263 } 2264 2265 align = 1ULL << vd->vdev_top->vdev_ashift; 2266 2267 if (P2PHASE(zio->io_size, align) != 0) { 2268 uint64_t asize = P2ROUNDUP(zio->io_size, align); 2269 char *abuf = zio_buf_alloc(asize); 2270 ASSERT(vd == vd->vdev_top); 2271 if (zio->io_type == ZIO_TYPE_WRITE) { 2272 bcopy(zio->io_data, abuf, zio->io_size); 2273 bzero(abuf + zio->io_size, asize - zio->io_size); 2274 } 2275 zio_push_transform(zio, abuf, asize, asize, zio_subblock); 2276 } 2277 2278 ASSERT(P2PHASE(zio->io_offset, align) == 0); 2279 ASSERT(P2PHASE(zio->io_size, align) == 0); 2280 ASSERT(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa)); 2281 2282 /* 2283 * If this is a repair I/O, and there's no self-healing involved -- 2284 * that is, we're just resilvering what we expect to resilver -- 2285 * then don't do the I/O unless zio's txg is actually in vd's DTL. 2286 * This prevents spurious resilvering with nested replication. 2287 * For example, given a mirror of mirrors, (A+B)+(C+D), if only 2288 * A is out of date, we'll read from C+D, then use the data to 2289 * resilver A+B -- but we don't actually want to resilver B, just A. 2290 * The top-level mirror has no way to know this, so instead we just 2291 * discard unnecessary repairs as we work our way down the vdev tree. 2292 * The same logic applies to any form of nested replication: 2293 * ditto + mirror, RAID-Z + replacing, etc. This covers them all. 2294 */ 2295 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) && 2296 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) && 2297 zio->io_txg != 0 && /* not a delegated i/o */ 2298 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) { 2299 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 2300 zio_vdev_io_bypass(zio); 2301 return (ZIO_PIPELINE_CONTINUE); 2302 } 2303 2304 if (vd->vdev_ops->vdev_op_leaf && 2305 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) { 2306 2307 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0) 2308 return (ZIO_PIPELINE_CONTINUE); 2309 2310 if ((zio = vdev_queue_io(zio)) == NULL) 2311 return (ZIO_PIPELINE_STOP); 2312 2313 if (!vdev_accessible(vd, zio)) { 2314 zio->io_error = ENXIO; 2315 zio_interrupt(zio); 2316 return (ZIO_PIPELINE_STOP); 2317 } 2318 } 2319 2320 return (vd->vdev_ops->vdev_op_io_start(zio)); 2321 } 2322 2323 static int 2324 zio_vdev_io_done(zio_t *zio) 2325 { 2326 vdev_t *vd = zio->io_vd; 2327 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops; 2328 boolean_t unexpected_error = B_FALSE; 2329 2330 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE)) 2331 return (ZIO_PIPELINE_STOP); 2332 2333 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE); 2334 2335 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) { 2336 2337 vdev_queue_io_done(zio); 2338 2339 if (zio->io_type == ZIO_TYPE_WRITE) 2340 vdev_cache_write(zio); 2341 2342 if (zio_injection_enabled && zio->io_error == 0) 2343 zio->io_error = zio_handle_device_injection(vd, 2344 zio, EIO); 2345 2346 if (zio_injection_enabled && zio->io_error == 0) 2347 zio->io_error = zio_handle_label_injection(zio, EIO); 2348 2349 if (zio->io_error) { 2350 if (!vdev_accessible(vd, zio)) { 2351 zio->io_error = ENXIO; 2352 } else { 2353 unexpected_error = B_TRUE; 2354 } 2355 } 2356 } 2357 2358 ops->vdev_op_io_done(zio); 2359 2360 if (unexpected_error) 2361 VERIFY(vdev_probe(vd, zio) == NULL); 2362 2363 return (ZIO_PIPELINE_CONTINUE); 2364 } 2365 2366 /* 2367 * For non-raidz ZIOs, we can just copy aside the bad data read from the 2368 * disk, and use that to finish the checksum ereport later. 2369 */ 2370 static void 2371 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr, 2372 const void *good_buf) 2373 { 2374 /* no processing needed */ 2375 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE); 2376 } 2377 2378 /*ARGSUSED*/ 2379 void 2380 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored) 2381 { 2382 void *buf = zio_buf_alloc(zio->io_size); 2383 2384 bcopy(zio->io_data, buf, zio->io_size); 2385 2386 zcr->zcr_cbinfo = zio->io_size; 2387 zcr->zcr_cbdata = buf; 2388 zcr->zcr_finish = zio_vsd_default_cksum_finish; 2389 zcr->zcr_free = zio_buf_free; 2390 } 2391 2392 static int 2393 zio_vdev_io_assess(zio_t *zio) 2394 { 2395 vdev_t *vd = zio->io_vd; 2396 2397 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE)) 2398 return (ZIO_PIPELINE_STOP); 2399 2400 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) 2401 spa_config_exit(zio->io_spa, SCL_ZIO, zio); 2402 2403 if (zio->io_vsd != NULL) { 2404 zio->io_vsd_ops->vsd_free(zio); 2405 zio->io_vsd = NULL; 2406 } 2407 2408 if (zio_injection_enabled && zio->io_error == 0) 2409 zio->io_error = zio_handle_fault_injection(zio, EIO); 2410 2411 /* 2412 * If the I/O failed, determine whether we should attempt to retry it. 2413 * 2414 * On retry, we cut in line in the issue queue, since we don't want 2415 * compression/checksumming/etc. work to prevent our (cheap) IO reissue. 2416 */ 2417 if (zio->io_error && vd == NULL && 2418 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) { 2419 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */ 2420 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */ 2421 zio->io_error = 0; 2422 zio->io_flags |= ZIO_FLAG_IO_RETRY | 2423 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE; 2424 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1; 2425 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, 2426 zio_requeue_io_start_cut_in_line); 2427 return (ZIO_PIPELINE_STOP); 2428 } 2429 2430 /* 2431 * If we got an error on a leaf device, convert it to ENXIO 2432 * if the device is not accessible at all. 2433 */ 2434 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf && 2435 !vdev_accessible(vd, zio)) 2436 zio->io_error = ENXIO; 2437 2438 /* 2439 * If we can't write to an interior vdev (mirror or RAID-Z), 2440 * set vdev_cant_write so that we stop trying to allocate from it. 2441 */ 2442 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE && 2443 vd != NULL && !vd->vdev_ops->vdev_op_leaf) 2444 vd->vdev_cant_write = B_TRUE; 2445 2446 if (zio->io_error) 2447 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2448 2449 return (ZIO_PIPELINE_CONTINUE); 2450 } 2451 2452 void 2453 zio_vdev_io_reissue(zio_t *zio) 2454 { 2455 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); 2456 ASSERT(zio->io_error == 0); 2457 2458 zio->io_stage >>= 1; 2459 } 2460 2461 void 2462 zio_vdev_io_redone(zio_t *zio) 2463 { 2464 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE); 2465 2466 zio->io_stage >>= 1; 2467 } 2468 2469 void 2470 zio_vdev_io_bypass(zio_t *zio) 2471 { 2472 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); 2473 ASSERT(zio->io_error == 0); 2474 2475 zio->io_flags |= ZIO_FLAG_IO_BYPASS; 2476 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1; 2477 } 2478 2479 /* 2480 * ========================================================================== 2481 * Generate and verify checksums 2482 * ========================================================================== 2483 */ 2484 static int 2485 zio_checksum_generate(zio_t *zio) 2486 { 2487 blkptr_t *bp = zio->io_bp; 2488 enum zio_checksum checksum; 2489 2490 if (bp == NULL) { 2491 /* 2492 * This is zio_write_phys(). 2493 * We're either generating a label checksum, or none at all. 2494 */ 2495 checksum = zio->io_prop.zp_checksum; 2496 2497 if (checksum == ZIO_CHECKSUM_OFF) 2498 return (ZIO_PIPELINE_CONTINUE); 2499 2500 ASSERT(checksum == ZIO_CHECKSUM_LABEL); 2501 } else { 2502 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) { 2503 ASSERT(!IO_IS_ALLOCATING(zio)); 2504 checksum = ZIO_CHECKSUM_GANG_HEADER; 2505 } else { 2506 checksum = BP_GET_CHECKSUM(bp); 2507 } 2508 } 2509 2510 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size); 2511 2512 return (ZIO_PIPELINE_CONTINUE); 2513 } 2514 2515 static int 2516 zio_checksum_verify(zio_t *zio) 2517 { 2518 zio_bad_cksum_t info; 2519 blkptr_t *bp = zio->io_bp; 2520 int error; 2521 2522 ASSERT(zio->io_vd != NULL); 2523 2524 if (bp == NULL) { 2525 /* 2526 * This is zio_read_phys(). 2527 * We're either verifying a label checksum, or nothing at all. 2528 */ 2529 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF) 2530 return (ZIO_PIPELINE_CONTINUE); 2531 2532 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL); 2533 } 2534 2535 if ((error = zio_checksum_error(zio, &info)) != 0) { 2536 zio->io_error = error; 2537 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) { 2538 zfs_ereport_start_checksum(zio->io_spa, 2539 zio->io_vd, zio, zio->io_offset, 2540 zio->io_size, NULL, &info); 2541 } 2542 } 2543 2544 return (ZIO_PIPELINE_CONTINUE); 2545 } 2546 2547 /* 2548 * Called by RAID-Z to ensure we don't compute the checksum twice. 2549 */ 2550 void 2551 zio_checksum_verified(zio_t *zio) 2552 { 2553 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; 2554 } 2555 2556 /* 2557 * ========================================================================== 2558 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other. 2559 * An error of 0 indictes success. ENXIO indicates whole-device failure, 2560 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO 2561 * indicate errors that are specific to one I/O, and most likely permanent. 2562 * Any other error is presumed to be worse because we weren't expecting it. 2563 * ========================================================================== 2564 */ 2565 int 2566 zio_worst_error(int e1, int e2) 2567 { 2568 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO }; 2569 int r1, r2; 2570 2571 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++) 2572 if (e1 == zio_error_rank[r1]) 2573 break; 2574 2575 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++) 2576 if (e2 == zio_error_rank[r2]) 2577 break; 2578 2579 return (r1 > r2 ? e1 : e2); 2580 } 2581 2582 /* 2583 * ========================================================================== 2584 * I/O completion 2585 * ========================================================================== 2586 */ 2587 static int 2588 zio_ready(zio_t *zio) 2589 { 2590 blkptr_t *bp = zio->io_bp; 2591 zio_t *pio, *pio_next; 2592 2593 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) || 2594 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY)) 2595 return (ZIO_PIPELINE_STOP); 2596 2597 if (zio->io_ready) { 2598 ASSERT(IO_IS_ALLOCATING(zio)); 2599 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp)); 2600 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0); 2601 2602 zio->io_ready(zio); 2603 } 2604 2605 if (bp != NULL && bp != &zio->io_bp_copy) 2606 zio->io_bp_copy = *bp; 2607 2608 if (zio->io_error) 2609 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2610 2611 mutex_enter(&zio->io_lock); 2612 zio->io_state[ZIO_WAIT_READY] = 1; 2613 pio = zio_walk_parents(zio); 2614 mutex_exit(&zio->io_lock); 2615 2616 /* 2617 * As we notify zio's parents, new parents could be added. 2618 * New parents go to the head of zio's io_parent_list, however, 2619 * so we will (correctly) not notify them. The remainder of zio's 2620 * io_parent_list, from 'pio_next' onward, cannot change because 2621 * all parents must wait for us to be done before they can be done. 2622 */ 2623 for (; pio != NULL; pio = pio_next) { 2624 pio_next = zio_walk_parents(zio); 2625 zio_notify_parent(pio, zio, ZIO_WAIT_READY); 2626 } 2627 2628 if (zio->io_flags & ZIO_FLAG_NODATA) { 2629 if (BP_IS_GANG(bp)) { 2630 zio->io_flags &= ~ZIO_FLAG_NODATA; 2631 } else { 2632 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE); 2633 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; 2634 } 2635 } 2636 2637 if (zio_injection_enabled && 2638 zio->io_spa->spa_syncing_txg == zio->io_txg) 2639 zio_handle_ignored_writes(zio); 2640 2641 return (ZIO_PIPELINE_CONTINUE); 2642 } 2643 2644 static int 2645 zio_done(zio_t *zio) 2646 { 2647 spa_t *spa = zio->io_spa; 2648 zio_t *lio = zio->io_logical; 2649 blkptr_t *bp = zio->io_bp; 2650 vdev_t *vd = zio->io_vd; 2651 uint64_t psize = zio->io_size; 2652 zio_t *pio, *pio_next; 2653 2654 /* 2655 * If our children haven't all completed, 2656 * wait for them and then repeat this pipeline stage. 2657 */ 2658 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) || 2659 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) || 2660 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) || 2661 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE)) 2662 return (ZIO_PIPELINE_STOP); 2663 2664 for (int c = 0; c < ZIO_CHILD_TYPES; c++) 2665 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 2666 ASSERT(zio->io_children[c][w] == 0); 2667 2668 if (bp != NULL) { 2669 ASSERT(bp->blk_pad[0] == 0); 2670 ASSERT(bp->blk_pad[1] == 0); 2671 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 || 2672 (bp == zio_unique_parent(zio)->io_bp)); 2673 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) && 2674 zio->io_bp_override == NULL && 2675 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) { 2676 ASSERT(!BP_SHOULD_BYTESWAP(bp)); 2677 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp)); 2678 ASSERT(BP_COUNT_GANG(bp) == 0 || 2679 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp))); 2680 } 2681 } 2682 2683 /* 2684 * If there were child vdev/gang/ddt errors, they apply to us now. 2685 */ 2686 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV); 2687 zio_inherit_child_errors(zio, ZIO_CHILD_GANG); 2688 zio_inherit_child_errors(zio, ZIO_CHILD_DDT); 2689 2690 /* 2691 * If the I/O on the transformed data was successful, generate any 2692 * checksum reports now while we still have the transformed data. 2693 */ 2694 if (zio->io_error == 0) { 2695 while (zio->io_cksum_report != NULL) { 2696 zio_cksum_report_t *zcr = zio->io_cksum_report; 2697 uint64_t align = zcr->zcr_align; 2698 uint64_t asize = P2ROUNDUP(psize, align); 2699 char *abuf = zio->io_data; 2700 2701 if (asize != psize) { 2702 abuf = zio_buf_alloc(asize); 2703 bcopy(zio->io_data, abuf, psize); 2704 bzero(abuf + psize, asize - psize); 2705 } 2706 2707 zio->io_cksum_report = zcr->zcr_next; 2708 zcr->zcr_next = NULL; 2709 zcr->zcr_finish(zcr, abuf); 2710 zfs_ereport_free_checksum(zcr); 2711 2712 if (asize != psize) 2713 zio_buf_free(abuf, asize); 2714 } 2715 } 2716 2717 zio_pop_transforms(zio); /* note: may set zio->io_error */ 2718 2719 vdev_stat_update(zio, psize); 2720 2721 if (zio->io_error) { 2722 /* 2723 * If this I/O is attached to a particular vdev, 2724 * generate an error message describing the I/O failure 2725 * at the block level. We ignore these errors if the 2726 * device is currently unavailable. 2727 */ 2728 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd)) 2729 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0); 2730 2731 if ((zio->io_error == EIO || !(zio->io_flags & 2732 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) && 2733 zio == lio) { 2734 /* 2735 * For logical I/O requests, tell the SPA to log the 2736 * error and generate a logical data ereport. 2737 */ 2738 spa_log_error(spa, zio); 2739 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio, 2740 0, 0); 2741 } 2742 } 2743 2744 if (zio->io_error && zio == lio) { 2745 /* 2746 * Determine whether zio should be reexecuted. This will 2747 * propagate all the way to the root via zio_notify_parent(). 2748 */ 2749 ASSERT(vd == NULL && bp != NULL); 2750 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2751 2752 if (IO_IS_ALLOCATING(zio) && 2753 !(zio->io_flags & ZIO_FLAG_CANFAIL)) { 2754 if (zio->io_error != ENOSPC) 2755 zio->io_reexecute |= ZIO_REEXECUTE_NOW; 2756 else 2757 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 2758 } 2759 2760 if ((zio->io_type == ZIO_TYPE_READ || 2761 zio->io_type == ZIO_TYPE_FREE) && 2762 zio->io_error == ENXIO && 2763 spa_load_state(spa) == SPA_LOAD_NONE && 2764 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE) 2765 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 2766 2767 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute) 2768 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 2769 2770 /* 2771 * Here is a possibly good place to attempt to do 2772 * either combinatorial reconstruction or error correction 2773 * based on checksums. It also might be a good place 2774 * to send out preliminary ereports before we suspend 2775 * processing. 2776 */ 2777 } 2778 2779 /* 2780 * If there were logical child errors, they apply to us now. 2781 * We defer this until now to avoid conflating logical child 2782 * errors with errors that happened to the zio itself when 2783 * updating vdev stats and reporting FMA events above. 2784 */ 2785 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL); 2786 2787 if ((zio->io_error || zio->io_reexecute) && 2788 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio && 2789 !(zio->io_flags & ZIO_FLAG_IO_REWRITE)) 2790 zio_dva_unallocate(zio, zio->io_gang_tree, bp); 2791 2792 zio_gang_tree_free(&zio->io_gang_tree); 2793 2794 /* 2795 * Godfather I/Os should never suspend. 2796 */ 2797 if ((zio->io_flags & ZIO_FLAG_GODFATHER) && 2798 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) 2799 zio->io_reexecute = 0; 2800 2801 if (zio->io_reexecute) { 2802 /* 2803 * This is a logical I/O that wants to reexecute. 2804 * 2805 * Reexecute is top-down. When an i/o fails, if it's not 2806 * the root, it simply notifies its parent and sticks around. 2807 * The parent, seeing that it still has children in zio_done(), 2808 * does the same. This percolates all the way up to the root. 2809 * The root i/o will reexecute or suspend the entire tree. 2810 * 2811 * This approach ensures that zio_reexecute() honors 2812 * all the original i/o dependency relationships, e.g. 2813 * parents not executing until children are ready. 2814 */ 2815 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2816 2817 zio->io_gang_leader = NULL; 2818 2819 mutex_enter(&zio->io_lock); 2820 zio->io_state[ZIO_WAIT_DONE] = 1; 2821 mutex_exit(&zio->io_lock); 2822 2823 /* 2824 * "The Godfather" I/O monitors its children but is 2825 * not a true parent to them. It will track them through 2826 * the pipeline but severs its ties whenever they get into 2827 * trouble (e.g. suspended). This allows "The Godfather" 2828 * I/O to return status without blocking. 2829 */ 2830 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) { 2831 zio_link_t *zl = zio->io_walk_link; 2832 pio_next = zio_walk_parents(zio); 2833 2834 if ((pio->io_flags & ZIO_FLAG_GODFATHER) && 2835 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) { 2836 zio_remove_child(pio, zio, zl); 2837 zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 2838 } 2839 } 2840 2841 if ((pio = zio_unique_parent(zio)) != NULL) { 2842 /* 2843 * We're not a root i/o, so there's nothing to do 2844 * but notify our parent. Don't propagate errors 2845 * upward since we haven't permanently failed yet. 2846 */ 2847 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); 2848 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE; 2849 zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 2850 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) { 2851 /* 2852 * We'd fail again if we reexecuted now, so suspend 2853 * until conditions improve (e.g. device comes online). 2854 */ 2855 zio_suspend(spa, zio); 2856 } else { 2857 /* 2858 * Reexecution is potentially a huge amount of work. 2859 * Hand it off to the otherwise-unused claim taskq. 2860 */ 2861 (void) taskq_dispatch( 2862 spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE], 2863 (task_func_t *)zio_reexecute, zio, TQ_SLEEP); 2864 } 2865 return (ZIO_PIPELINE_STOP); 2866 } 2867 2868 ASSERT(zio->io_child_count == 0); 2869 ASSERT(zio->io_reexecute == 0); 2870 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL)); 2871 2872 /* 2873 * Report any checksum errors, since the I/O is complete. 2874 */ 2875 while (zio->io_cksum_report != NULL) { 2876 zio_cksum_report_t *zcr = zio->io_cksum_report; 2877 zio->io_cksum_report = zcr->zcr_next; 2878 zcr->zcr_next = NULL; 2879 zcr->zcr_finish(zcr, NULL); 2880 zfs_ereport_free_checksum(zcr); 2881 } 2882 2883 /* 2884 * It is the responsibility of the done callback to ensure that this 2885 * particular zio is no longer discoverable for adoption, and as 2886 * such, cannot acquire any new parents. 2887 */ 2888 if (zio->io_done) 2889 zio->io_done(zio); 2890 2891 mutex_enter(&zio->io_lock); 2892 zio->io_state[ZIO_WAIT_DONE] = 1; 2893 mutex_exit(&zio->io_lock); 2894 2895 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) { 2896 zio_link_t *zl = zio->io_walk_link; 2897 pio_next = zio_walk_parents(zio); 2898 zio_remove_child(pio, zio, zl); 2899 zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 2900 } 2901 2902 if (zio->io_waiter != NULL) { 2903 mutex_enter(&zio->io_lock); 2904 zio->io_executor = NULL; 2905 cv_broadcast(&zio->io_cv); 2906 mutex_exit(&zio->io_lock); 2907 } else { 2908 zio_destroy(zio); 2909 } 2910 2911 return (ZIO_PIPELINE_STOP); 2912 } 2913 2914 /* 2915 * ========================================================================== 2916 * I/O pipeline definition 2917 * ========================================================================== 2918 */ 2919 static zio_pipe_stage_t *zio_pipeline[] = { 2920 NULL, 2921 zio_read_bp_init, 2922 zio_free_bp_init, 2923 zio_issue_async, 2924 zio_write_bp_init, 2925 zio_checksum_generate, 2926 zio_ddt_read_start, 2927 zio_ddt_read_done, 2928 zio_ddt_write, 2929 zio_ddt_free, 2930 zio_gang_assemble, 2931 zio_gang_issue, 2932 zio_dva_allocate, 2933 zio_dva_free, 2934 zio_dva_claim, 2935 zio_ready, 2936 zio_vdev_io_start, 2937 zio_vdev_io_done, 2938 zio_vdev_io_assess, 2939 zio_checksum_verify, 2940 zio_done 2941 }; 2942