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