1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21 /* 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23 * Copyright (c) 2011, 2015 by Delphix. All rights reserved. 24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved. 25 */ 26 27 #include <sys/sysmacros.h> 28 #include <sys/zfs_context.h> 29 #include <sys/fm/fs/zfs.h> 30 #include <sys/spa.h> 31 #include <sys/txg.h> 32 #include <sys/spa_impl.h> 33 #include <sys/vdev_impl.h> 34 #include <sys/zio_impl.h> 35 #include <sys/zio_compress.h> 36 #include <sys/zio_checksum.h> 37 #include <sys/dmu_objset.h> 38 #include <sys/arc.h> 39 #include <sys/ddt.h> 40 #include <sys/blkptr.h> 41 #include <sys/zfeature.h> 42 43 /* 44 * ========================================================================== 45 * I/O type descriptions 46 * ========================================================================== 47 */ 48 const char *zio_type_name[ZIO_TYPES] = { 49 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim", 50 "zio_ioctl" 51 }; 52 53 /* 54 * ========================================================================== 55 * I/O kmem caches 56 * ========================================================================== 57 */ 58 kmem_cache_t *zio_cache; 59 kmem_cache_t *zio_link_cache; 60 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; 61 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; 62 63 #ifdef _KERNEL 64 extern vmem_t *zio_alloc_arena; 65 #endif 66 67 #define ZIO_PIPELINE_CONTINUE 0x100 68 #define ZIO_PIPELINE_STOP 0x101 69 70 #define BP_SPANB(indblkshift, level) \ 71 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT))) 72 #define COMPARE_META_LEVEL 0x80000000ul 73 /* 74 * The following actions directly effect the spa's sync-to-convergence logic. 75 * The values below define the sync pass when we start performing the action. 76 * Care should be taken when changing these values as they directly impact 77 * spa_sync() performance. Tuning these values may introduce subtle performance 78 * pathologies and should only be done in the context of performance analysis. 79 * These tunables will eventually be removed and replaced with #defines once 80 * enough analysis has been done to determine optimal values. 81 * 82 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that 83 * regular blocks are not deferred. 84 */ 85 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */ 86 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */ 87 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */ 88 89 /* 90 * An allocating zio is one that either currently has the DVA allocate 91 * stage set or will have it later in its lifetime. 92 */ 93 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE) 94 95 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE; 96 97 #ifdef ZFS_DEBUG 98 int zio_buf_debug_limit = 16384; 99 #else 100 int zio_buf_debug_limit = 0; 101 #endif 102 103 void 104 zio_init(void) 105 { 106 size_t c; 107 vmem_t *data_alloc_arena = NULL; 108 109 #ifdef _KERNEL 110 data_alloc_arena = zio_alloc_arena; 111 #endif 112 zio_cache = kmem_cache_create("zio_cache", 113 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0); 114 zio_link_cache = kmem_cache_create("zio_link_cache", 115 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0); 116 117 /* 118 * For small buffers, we want a cache for each multiple of 119 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache 120 * for each quarter-power of 2. 121 */ 122 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { 123 size_t size = (c + 1) << SPA_MINBLOCKSHIFT; 124 size_t p2 = size; 125 size_t align = 0; 126 size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0; 127 128 while (!ISP2(p2)) 129 p2 &= p2 - 1; 130 131 #ifndef _KERNEL 132 /* 133 * If we are using watchpoints, put each buffer on its own page, 134 * to eliminate the performance overhead of trapping to the 135 * kernel when modifying a non-watched buffer that shares the 136 * page with a watched buffer. 137 */ 138 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE)) 139 continue; 140 #endif 141 if (size <= 4 * SPA_MINBLOCKSIZE) { 142 align = SPA_MINBLOCKSIZE; 143 } else if (IS_P2ALIGNED(size, p2 >> 2)) { 144 align = MIN(p2 >> 2, PAGESIZE); 145 } 146 147 if (align != 0) { 148 char name[36]; 149 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size); 150 zio_buf_cache[c] = kmem_cache_create(name, size, 151 align, NULL, NULL, NULL, NULL, NULL, cflags); 152 153 /* 154 * Since zio_data bufs do not appear in crash dumps, we 155 * pass KMC_NOTOUCH so that no allocator metadata is 156 * stored with the buffers. 157 */ 158 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size); 159 zio_data_buf_cache[c] = kmem_cache_create(name, size, 160 align, NULL, NULL, NULL, NULL, data_alloc_arena, 161 cflags | KMC_NOTOUCH); 162 } 163 } 164 165 while (--c != 0) { 166 ASSERT(zio_buf_cache[c] != NULL); 167 if (zio_buf_cache[c - 1] == NULL) 168 zio_buf_cache[c - 1] = zio_buf_cache[c]; 169 170 ASSERT(zio_data_buf_cache[c] != NULL); 171 if (zio_data_buf_cache[c - 1] == NULL) 172 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c]; 173 } 174 175 zio_inject_init(); 176 } 177 178 void 179 zio_fini(void) 180 { 181 size_t c; 182 kmem_cache_t *last_cache = NULL; 183 kmem_cache_t *last_data_cache = NULL; 184 185 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { 186 if (zio_buf_cache[c] != last_cache) { 187 last_cache = zio_buf_cache[c]; 188 kmem_cache_destroy(zio_buf_cache[c]); 189 } 190 zio_buf_cache[c] = NULL; 191 192 if (zio_data_buf_cache[c] != last_data_cache) { 193 last_data_cache = zio_data_buf_cache[c]; 194 kmem_cache_destroy(zio_data_buf_cache[c]); 195 } 196 zio_data_buf_cache[c] = NULL; 197 } 198 199 kmem_cache_destroy(zio_link_cache); 200 kmem_cache_destroy(zio_cache); 201 202 zio_inject_fini(); 203 } 204 205 /* 206 * ========================================================================== 207 * Allocate and free I/O buffers 208 * ========================================================================== 209 */ 210 211 /* 212 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a 213 * crashdump if the kernel panics, so use it judiciously. Obviously, it's 214 * useful to inspect ZFS metadata, but if possible, we should avoid keeping 215 * excess / transient data in-core during a crashdump. 216 */ 217 void * 218 zio_buf_alloc(size_t size) 219 { 220 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 221 222 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 223 224 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE)); 225 } 226 227 /* 228 * Use zio_data_buf_alloc to allocate data. The data will not appear in a 229 * crashdump if the kernel panics. This exists so that we will limit the amount 230 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount 231 * of kernel heap dumped to disk when the kernel panics) 232 */ 233 void * 234 zio_data_buf_alloc(size_t size) 235 { 236 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 237 238 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 239 240 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE)); 241 } 242 243 void 244 zio_buf_free(void *buf, size_t size) 245 { 246 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 247 248 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 249 250 kmem_cache_free(zio_buf_cache[c], buf); 251 } 252 253 void 254 zio_data_buf_free(void *buf, size_t size) 255 { 256 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 257 258 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 259 260 kmem_cache_free(zio_data_buf_cache[c], buf); 261 } 262 263 /* 264 * ========================================================================== 265 * Push and pop I/O transform buffers 266 * ========================================================================== 267 */ 268 static void 269 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize, 270 zio_transform_func_t *transform) 271 { 272 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP); 273 274 zt->zt_orig_data = zio->io_data; 275 zt->zt_orig_size = zio->io_size; 276 zt->zt_bufsize = bufsize; 277 zt->zt_transform = transform; 278 279 zt->zt_next = zio->io_transform_stack; 280 zio->io_transform_stack = zt; 281 282 zio->io_data = data; 283 zio->io_size = size; 284 } 285 286 static void 287 zio_pop_transforms(zio_t *zio) 288 { 289 zio_transform_t *zt; 290 291 while ((zt = zio->io_transform_stack) != NULL) { 292 if (zt->zt_transform != NULL) 293 zt->zt_transform(zio, 294 zt->zt_orig_data, zt->zt_orig_size); 295 296 if (zt->zt_bufsize != 0) 297 zio_buf_free(zio->io_data, zt->zt_bufsize); 298 299 zio->io_data = zt->zt_orig_data; 300 zio->io_size = zt->zt_orig_size; 301 zio->io_transform_stack = zt->zt_next; 302 303 kmem_free(zt, sizeof (zio_transform_t)); 304 } 305 } 306 307 /* 308 * ========================================================================== 309 * I/O transform callbacks for subblocks and decompression 310 * ========================================================================== 311 */ 312 static void 313 zio_subblock(zio_t *zio, void *data, uint64_t size) 314 { 315 ASSERT(zio->io_size > size); 316 317 if (zio->io_type == ZIO_TYPE_READ) 318 bcopy(zio->io_data, data, size); 319 } 320 321 static void 322 zio_decompress(zio_t *zio, void *data, uint64_t size) 323 { 324 if (zio->io_error == 0 && 325 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp), 326 zio->io_data, data, zio->io_size, size) != 0) 327 zio->io_error = SET_ERROR(EIO); 328 } 329 330 /* 331 * ========================================================================== 332 * I/O parent/child relationships and pipeline interlocks 333 * ========================================================================== 334 */ 335 /* 336 * NOTE - Callers to zio_walk_parents() and zio_walk_children must 337 * continue calling these functions until they return NULL. 338 * Otherwise, the next caller will pick up the list walk in 339 * some indeterminate state. (Otherwise every caller would 340 * have to pass in a cookie to keep the state represented by 341 * io_walk_link, which gets annoying.) 342 */ 343 zio_t * 344 zio_walk_parents(zio_t *cio) 345 { 346 zio_link_t *zl = cio->io_walk_link; 347 list_t *pl = &cio->io_parent_list; 348 349 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl); 350 cio->io_walk_link = zl; 351 352 if (zl == NULL) 353 return (NULL); 354 355 ASSERT(zl->zl_child == cio); 356 return (zl->zl_parent); 357 } 358 359 zio_t * 360 zio_walk_children(zio_t *pio) 361 { 362 zio_link_t *zl = pio->io_walk_link; 363 list_t *cl = &pio->io_child_list; 364 365 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl); 366 pio->io_walk_link = zl; 367 368 if (zl == NULL) 369 return (NULL); 370 371 ASSERT(zl->zl_parent == pio); 372 return (zl->zl_child); 373 } 374 375 zio_t * 376 zio_unique_parent(zio_t *cio) 377 { 378 zio_t *pio = zio_walk_parents(cio); 379 380 VERIFY(zio_walk_parents(cio) == NULL); 381 return (pio); 382 } 383 384 void 385 zio_add_child(zio_t *pio, zio_t *cio) 386 { 387 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP); 388 389 /* 390 * Logical I/Os can have logical, gang, or vdev children. 391 * Gang I/Os can have gang or vdev children. 392 * Vdev I/Os can only have vdev children. 393 * The following ASSERT captures all of these constraints. 394 */ 395 ASSERT(cio->io_child_type <= pio->io_child_type); 396 397 zl->zl_parent = pio; 398 zl->zl_child = cio; 399 400 mutex_enter(&cio->io_lock); 401 mutex_enter(&pio->io_lock); 402 403 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0); 404 405 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 406 pio->io_children[cio->io_child_type][w] += !cio->io_state[w]; 407 408 list_insert_head(&pio->io_child_list, zl); 409 list_insert_head(&cio->io_parent_list, zl); 410 411 pio->io_child_count++; 412 cio->io_parent_count++; 413 414 mutex_exit(&pio->io_lock); 415 mutex_exit(&cio->io_lock); 416 } 417 418 static void 419 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl) 420 { 421 ASSERT(zl->zl_parent == pio); 422 ASSERT(zl->zl_child == cio); 423 424 mutex_enter(&cio->io_lock); 425 mutex_enter(&pio->io_lock); 426 427 list_remove(&pio->io_child_list, zl); 428 list_remove(&cio->io_parent_list, zl); 429 430 pio->io_child_count--; 431 cio->io_parent_count--; 432 433 mutex_exit(&pio->io_lock); 434 mutex_exit(&cio->io_lock); 435 436 kmem_cache_free(zio_link_cache, zl); 437 } 438 439 static boolean_t 440 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait) 441 { 442 uint64_t *countp = &zio->io_children[child][wait]; 443 boolean_t waiting = B_FALSE; 444 445 mutex_enter(&zio->io_lock); 446 ASSERT(zio->io_stall == NULL); 447 if (*countp != 0) { 448 zio->io_stage >>= 1; 449 zio->io_stall = countp; 450 waiting = B_TRUE; 451 } 452 mutex_exit(&zio->io_lock); 453 454 return (waiting); 455 } 456 457 static void 458 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait) 459 { 460 uint64_t *countp = &pio->io_children[zio->io_child_type][wait]; 461 int *errorp = &pio->io_child_error[zio->io_child_type]; 462 463 mutex_enter(&pio->io_lock); 464 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE)) 465 *errorp = zio_worst_error(*errorp, zio->io_error); 466 pio->io_reexecute |= zio->io_reexecute; 467 ASSERT3U(*countp, >, 0); 468 469 (*countp)--; 470 471 if (*countp == 0 && pio->io_stall == countp) { 472 pio->io_stall = NULL; 473 mutex_exit(&pio->io_lock); 474 zio_execute(pio); 475 } else { 476 mutex_exit(&pio->io_lock); 477 } 478 } 479 480 static void 481 zio_inherit_child_errors(zio_t *zio, enum zio_child c) 482 { 483 if (zio->io_child_error[c] != 0 && zio->io_error == 0) 484 zio->io_error = zio->io_child_error[c]; 485 } 486 487 /* 488 * ========================================================================== 489 * Create the various types of I/O (read, write, free, etc) 490 * ========================================================================== 491 */ 492 static zio_t * 493 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 494 void *data, uint64_t size, zio_done_func_t *done, void *private, 495 zio_type_t type, zio_priority_t priority, enum zio_flag flags, 496 vdev_t *vd, uint64_t offset, const zbookmark_phys_t *zb, 497 enum zio_stage stage, enum zio_stage pipeline) 498 { 499 zio_t *zio; 500 501 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE); 502 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0); 503 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0); 504 505 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER)); 506 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER)); 507 ASSERT(vd || stage == ZIO_STAGE_OPEN); 508 509 zio = kmem_cache_alloc(zio_cache, KM_SLEEP); 510 bzero(zio, sizeof (zio_t)); 511 512 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL); 513 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL); 514 515 list_create(&zio->io_parent_list, sizeof (zio_link_t), 516 offsetof(zio_link_t, zl_parent_node)); 517 list_create(&zio->io_child_list, sizeof (zio_link_t), 518 offsetof(zio_link_t, zl_child_node)); 519 520 if (vd != NULL) 521 zio->io_child_type = ZIO_CHILD_VDEV; 522 else if (flags & ZIO_FLAG_GANG_CHILD) 523 zio->io_child_type = ZIO_CHILD_GANG; 524 else if (flags & ZIO_FLAG_DDT_CHILD) 525 zio->io_child_type = ZIO_CHILD_DDT; 526 else 527 zio->io_child_type = ZIO_CHILD_LOGICAL; 528 529 if (bp != NULL) { 530 zio->io_bp = (blkptr_t *)bp; 531 zio->io_bp_copy = *bp; 532 zio->io_bp_orig = *bp; 533 if (type != ZIO_TYPE_WRITE || 534 zio->io_child_type == ZIO_CHILD_DDT) 535 zio->io_bp = &zio->io_bp_copy; /* so caller can free */ 536 if (zio->io_child_type == ZIO_CHILD_LOGICAL) 537 zio->io_logical = zio; 538 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp)) 539 pipeline |= ZIO_GANG_STAGES; 540 } 541 542 zio->io_spa = spa; 543 zio->io_txg = txg; 544 zio->io_done = done; 545 zio->io_private = private; 546 zio->io_type = type; 547 zio->io_priority = priority; 548 zio->io_vd = vd; 549 zio->io_offset = offset; 550 zio->io_orig_data = zio->io_data = data; 551 zio->io_orig_size = zio->io_size = size; 552 zio->io_orig_flags = zio->io_flags = flags; 553 zio->io_orig_stage = zio->io_stage = stage; 554 zio->io_orig_pipeline = zio->io_pipeline = pipeline; 555 556 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY); 557 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE); 558 559 if (zb != NULL) 560 zio->io_bookmark = *zb; 561 562 if (pio != NULL) { 563 if (zio->io_logical == NULL) 564 zio->io_logical = pio->io_logical; 565 if (zio->io_child_type == ZIO_CHILD_GANG) 566 zio->io_gang_leader = pio->io_gang_leader; 567 zio_add_child(pio, zio); 568 } 569 570 return (zio); 571 } 572 573 static void 574 zio_destroy(zio_t *zio) 575 { 576 list_destroy(&zio->io_parent_list); 577 list_destroy(&zio->io_child_list); 578 mutex_destroy(&zio->io_lock); 579 cv_destroy(&zio->io_cv); 580 kmem_cache_free(zio_cache, zio); 581 } 582 583 zio_t * 584 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done, 585 void *private, enum zio_flag flags) 586 { 587 zio_t *zio; 588 589 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private, 590 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, 591 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE); 592 593 return (zio); 594 } 595 596 zio_t * 597 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags) 598 { 599 return (zio_null(NULL, spa, NULL, done, private, flags)); 600 } 601 602 void 603 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp) 604 { 605 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) { 606 zfs_panic_recover("blkptr at %p has invalid TYPE %llu", 607 bp, (longlong_t)BP_GET_TYPE(bp)); 608 } 609 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS || 610 BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) { 611 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu", 612 bp, (longlong_t)BP_GET_CHECKSUM(bp)); 613 } 614 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS || 615 BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) { 616 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu", 617 bp, (longlong_t)BP_GET_COMPRESS(bp)); 618 } 619 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) { 620 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu", 621 bp, (longlong_t)BP_GET_LSIZE(bp)); 622 } 623 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) { 624 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu", 625 bp, (longlong_t)BP_GET_PSIZE(bp)); 626 } 627 628 if (BP_IS_EMBEDDED(bp)) { 629 if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) { 630 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu", 631 bp, (longlong_t)BPE_GET_ETYPE(bp)); 632 } 633 } 634 635 /* 636 * Pool-specific checks. 637 * 638 * Note: it would be nice to verify that the blk_birth and 639 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze() 640 * allows the birth time of log blocks (and dmu_sync()-ed blocks 641 * that are in the log) to be arbitrarily large. 642 */ 643 for (int i = 0; i < BP_GET_NDVAS(bp); i++) { 644 uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]); 645 if (vdevid >= spa->spa_root_vdev->vdev_children) { 646 zfs_panic_recover("blkptr at %p DVA %u has invalid " 647 "VDEV %llu", 648 bp, i, (longlong_t)vdevid); 649 continue; 650 } 651 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid]; 652 if (vd == NULL) { 653 zfs_panic_recover("blkptr at %p DVA %u has invalid " 654 "VDEV %llu", 655 bp, i, (longlong_t)vdevid); 656 continue; 657 } 658 if (vd->vdev_ops == &vdev_hole_ops) { 659 zfs_panic_recover("blkptr at %p DVA %u has hole " 660 "VDEV %llu", 661 bp, i, (longlong_t)vdevid); 662 continue; 663 } 664 if (vd->vdev_ops == &vdev_missing_ops) { 665 /* 666 * "missing" vdevs are valid during import, but we 667 * don't have their detailed info (e.g. asize), so 668 * we can't perform any more checks on them. 669 */ 670 continue; 671 } 672 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]); 673 uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]); 674 if (BP_IS_GANG(bp)) 675 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE); 676 if (offset + asize > vd->vdev_asize) { 677 zfs_panic_recover("blkptr at %p DVA %u has invalid " 678 "OFFSET %llu", 679 bp, i, (longlong_t)offset); 680 } 681 } 682 } 683 684 zio_t * 685 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, 686 void *data, uint64_t size, zio_done_func_t *done, void *private, 687 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb) 688 { 689 zio_t *zio; 690 691 zfs_blkptr_verify(spa, bp); 692 693 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp, 694 data, size, done, private, 695 ZIO_TYPE_READ, priority, flags, NULL, 0, zb, 696 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? 697 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE); 698 699 return (zio); 700 } 701 702 zio_t * 703 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, 704 void *data, uint64_t size, const zio_prop_t *zp, 705 zio_done_func_t *ready, zio_done_func_t *physdone, zio_done_func_t *done, 706 void *private, 707 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb) 708 { 709 zio_t *zio; 710 711 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF && 712 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS && 713 zp->zp_compress >= ZIO_COMPRESS_OFF && 714 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS && 715 DMU_OT_IS_VALID(zp->zp_type) && 716 zp->zp_level < 32 && 717 zp->zp_copies > 0 && 718 zp->zp_copies <= spa_max_replication(spa)); 719 720 zio = zio_create(pio, spa, txg, bp, data, size, done, private, 721 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, 722 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? 723 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE); 724 725 zio->io_ready = ready; 726 zio->io_physdone = physdone; 727 zio->io_prop = *zp; 728 729 /* 730 * Data can be NULL if we are going to call zio_write_override() to 731 * provide the already-allocated BP. But we may need the data to 732 * verify a dedup hit (if requested). In this case, don't try to 733 * dedup (just take the already-allocated BP verbatim). 734 */ 735 if (data == NULL && zio->io_prop.zp_dedup_verify) { 736 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE; 737 } 738 739 return (zio); 740 } 741 742 zio_t * 743 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data, 744 uint64_t size, zio_done_func_t *done, void *private, 745 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb) 746 { 747 zio_t *zio; 748 749 zio = zio_create(pio, spa, txg, bp, data, size, done, private, 750 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, 751 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE); 752 753 return (zio); 754 } 755 756 void 757 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite) 758 { 759 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 760 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 761 ASSERT(zio->io_stage == ZIO_STAGE_OPEN); 762 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa)); 763 764 /* 765 * We must reset the io_prop to match the values that existed 766 * when the bp was first written by dmu_sync() keeping in mind 767 * that nopwrite and dedup are mutually exclusive. 768 */ 769 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup; 770 zio->io_prop.zp_nopwrite = nopwrite; 771 zio->io_prop.zp_copies = copies; 772 zio->io_bp_override = bp; 773 } 774 775 void 776 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp) 777 { 778 779 /* 780 * The check for EMBEDDED is a performance optimization. We 781 * process the free here (by ignoring it) rather than 782 * putting it on the list and then processing it in zio_free_sync(). 783 */ 784 if (BP_IS_EMBEDDED(bp)) 785 return; 786 metaslab_check_free(spa, bp); 787 788 /* 789 * Frees that are for the currently-syncing txg, are not going to be 790 * deferred, and which will not need to do a read (i.e. not GANG or 791 * DEDUP), can be processed immediately. Otherwise, put them on the 792 * in-memory list for later processing. 793 */ 794 if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp) || 795 txg != spa->spa_syncing_txg || 796 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) { 797 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp); 798 } else { 799 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 0))); 800 } 801 } 802 803 zio_t * 804 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 805 enum zio_flag flags) 806 { 807 zio_t *zio; 808 enum zio_stage stage = ZIO_FREE_PIPELINE; 809 810 ASSERT(!BP_IS_HOLE(bp)); 811 ASSERT(spa_syncing_txg(spa) == txg); 812 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free); 813 814 if (BP_IS_EMBEDDED(bp)) 815 return (zio_null(pio, spa, NULL, NULL, NULL, 0)); 816 817 metaslab_check_free(spa, bp); 818 arc_freed(spa, bp); 819 820 /* 821 * GANG and DEDUP blocks can induce a read (for the gang block header, 822 * or the DDT), so issue them asynchronously so that this thread is 823 * not tied up. 824 */ 825 if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp)) 826 stage |= ZIO_STAGE_ISSUE_ASYNC; 827 828 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), 829 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags, 830 NULL, 0, NULL, ZIO_STAGE_OPEN, stage); 831 832 return (zio); 833 } 834 835 zio_t * 836 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 837 zio_done_func_t *done, void *private, enum zio_flag flags) 838 { 839 zio_t *zio; 840 841 dprintf_bp(bp, "claiming in txg %llu", txg); 842 843 if (BP_IS_EMBEDDED(bp)) 844 return (zio_null(pio, spa, NULL, NULL, NULL, 0)); 845 846 /* 847 * A claim is an allocation of a specific block. Claims are needed 848 * to support immediate writes in the intent log. The issue is that 849 * immediate writes contain committed data, but in a txg that was 850 * *not* committed. Upon opening the pool after an unclean shutdown, 851 * the intent log claims all blocks that contain immediate write data 852 * so that the SPA knows they're in use. 853 * 854 * All claims *must* be resolved in the first txg -- before the SPA 855 * starts allocating blocks -- so that nothing is allocated twice. 856 * If txg == 0 we just verify that the block is claimable. 857 */ 858 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa)); 859 ASSERT(txg == spa_first_txg(spa) || txg == 0); 860 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */ 861 862 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), 863 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags, 864 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE); 865 866 return (zio); 867 } 868 869 zio_t * 870 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, 871 zio_done_func_t *done, void *private, enum zio_flag flags) 872 { 873 zio_t *zio; 874 int c; 875 876 if (vd->vdev_children == 0) { 877 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private, 878 ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, 879 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE); 880 881 zio->io_cmd = cmd; 882 } else { 883 zio = zio_null(pio, spa, NULL, NULL, NULL, flags); 884 885 for (c = 0; c < vd->vdev_children; c++) 886 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd, 887 done, private, flags)); 888 } 889 890 return (zio); 891 } 892 893 zio_t * 894 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 895 void *data, int checksum, zio_done_func_t *done, void *private, 896 zio_priority_t priority, enum zio_flag flags, boolean_t labels) 897 { 898 zio_t *zio; 899 900 ASSERT(vd->vdev_children == 0); 901 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || 902 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); 903 ASSERT3U(offset + size, <=, vd->vdev_psize); 904 905 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private, 906 ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset, 907 NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE); 908 909 zio->io_prop.zp_checksum = checksum; 910 911 return (zio); 912 } 913 914 zio_t * 915 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 916 void *data, int checksum, zio_done_func_t *done, void *private, 917 zio_priority_t priority, enum zio_flag flags, boolean_t labels) 918 { 919 zio_t *zio; 920 921 ASSERT(vd->vdev_children == 0); 922 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || 923 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); 924 ASSERT3U(offset + size, <=, vd->vdev_psize); 925 926 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private, 927 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset, 928 NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE); 929 930 zio->io_prop.zp_checksum = checksum; 931 932 if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) { 933 /* 934 * zec checksums are necessarily destructive -- they modify 935 * the end of the write buffer to hold the verifier/checksum. 936 * Therefore, we must make a local copy in case the data is 937 * being written to multiple places in parallel. 938 */ 939 void *wbuf = zio_buf_alloc(size); 940 bcopy(data, wbuf, size); 941 zio_push_transform(zio, wbuf, size, size, NULL); 942 } 943 944 return (zio); 945 } 946 947 /* 948 * Create a child I/O to do some work for us. 949 */ 950 zio_t * 951 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset, 952 void *data, uint64_t size, int type, zio_priority_t priority, 953 enum zio_flag flags, zio_done_func_t *done, void *private) 954 { 955 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE; 956 zio_t *zio; 957 958 ASSERT(vd->vdev_parent == 959 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev)); 960 961 if (type == ZIO_TYPE_READ && bp != NULL) { 962 /* 963 * If we have the bp, then the child should perform the 964 * checksum and the parent need not. This pushes error 965 * detection as close to the leaves as possible and 966 * eliminates redundant checksums in the interior nodes. 967 */ 968 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY; 969 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; 970 } 971 972 if (vd->vdev_children == 0) 973 offset += VDEV_LABEL_START_SIZE; 974 975 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE; 976 977 /* 978 * If we've decided to do a repair, the write is not speculative -- 979 * even if the original read was. 980 */ 981 if (flags & ZIO_FLAG_IO_REPAIR) 982 flags &= ~ZIO_FLAG_SPECULATIVE; 983 984 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, 985 done, private, type, priority, flags, vd, offset, &pio->io_bookmark, 986 ZIO_STAGE_VDEV_IO_START >> 1, pipeline); 987 988 zio->io_physdone = pio->io_physdone; 989 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL) 990 zio->io_logical->io_phys_children++; 991 992 return (zio); 993 } 994 995 zio_t * 996 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size, 997 int type, zio_priority_t priority, enum zio_flag flags, 998 zio_done_func_t *done, void *private) 999 { 1000 zio_t *zio; 1001 1002 ASSERT(vd->vdev_ops->vdev_op_leaf); 1003 1004 zio = zio_create(NULL, vd->vdev_spa, 0, NULL, 1005 data, size, done, private, type, priority, 1006 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED, 1007 vd, offset, NULL, 1008 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE); 1009 1010 return (zio); 1011 } 1012 1013 void 1014 zio_flush(zio_t *zio, vdev_t *vd) 1015 { 1016 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 1017 NULL, NULL, 1018 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY)); 1019 } 1020 1021 void 1022 zio_shrink(zio_t *zio, uint64_t size) 1023 { 1024 ASSERT(zio->io_executor == NULL); 1025 ASSERT(zio->io_orig_size == zio->io_size); 1026 ASSERT(size <= zio->io_size); 1027 1028 /* 1029 * We don't shrink for raidz because of problems with the 1030 * reconstruction when reading back less than the block size. 1031 * Note, BP_IS_RAIDZ() assumes no compression. 1032 */ 1033 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 1034 if (!BP_IS_RAIDZ(zio->io_bp)) 1035 zio->io_orig_size = zio->io_size = size; 1036 } 1037 1038 /* 1039 * ========================================================================== 1040 * Prepare to read and write logical blocks 1041 * ========================================================================== 1042 */ 1043 1044 static int 1045 zio_read_bp_init(zio_t *zio) 1046 { 1047 blkptr_t *bp = zio->io_bp; 1048 1049 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF && 1050 zio->io_child_type == ZIO_CHILD_LOGICAL && 1051 !(zio->io_flags & ZIO_FLAG_RAW)) { 1052 uint64_t psize = 1053 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp); 1054 void *cbuf = zio_buf_alloc(psize); 1055 1056 zio_push_transform(zio, cbuf, psize, psize, zio_decompress); 1057 } 1058 1059 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) { 1060 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1061 decode_embedded_bp_compressed(bp, zio->io_data); 1062 } else { 1063 ASSERT(!BP_IS_EMBEDDED(bp)); 1064 } 1065 1066 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0) 1067 zio->io_flags |= ZIO_FLAG_DONT_CACHE; 1068 1069 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP) 1070 zio->io_flags |= ZIO_FLAG_DONT_CACHE; 1071 1072 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL) 1073 zio->io_pipeline = ZIO_DDT_READ_PIPELINE; 1074 1075 return (ZIO_PIPELINE_CONTINUE); 1076 } 1077 1078 static int 1079 zio_write_bp_init(zio_t *zio) 1080 { 1081 spa_t *spa = zio->io_spa; 1082 zio_prop_t *zp = &zio->io_prop; 1083 enum zio_compress compress = zp->zp_compress; 1084 blkptr_t *bp = zio->io_bp; 1085 uint64_t lsize = zio->io_size; 1086 uint64_t psize = lsize; 1087 int pass = 1; 1088 1089 /* 1090 * If our children haven't all reached the ready stage, 1091 * wait for them and then repeat this pipeline stage. 1092 */ 1093 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) || 1094 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY)) 1095 return (ZIO_PIPELINE_STOP); 1096 1097 if (!IO_IS_ALLOCATING(zio)) 1098 return (ZIO_PIPELINE_CONTINUE); 1099 1100 ASSERT(zio->io_child_type != ZIO_CHILD_DDT); 1101 1102 if (zio->io_bp_override) { 1103 ASSERT(bp->blk_birth != zio->io_txg); 1104 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0); 1105 1106 *bp = *zio->io_bp_override; 1107 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1108 1109 if (BP_IS_EMBEDDED(bp)) 1110 return (ZIO_PIPELINE_CONTINUE); 1111 1112 /* 1113 * If we've been overridden and nopwrite is set then 1114 * set the flag accordingly to indicate that a nopwrite 1115 * has already occurred. 1116 */ 1117 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) { 1118 ASSERT(!zp->zp_dedup); 1119 zio->io_flags |= ZIO_FLAG_NOPWRITE; 1120 return (ZIO_PIPELINE_CONTINUE); 1121 } 1122 1123 ASSERT(!zp->zp_nopwrite); 1124 1125 if (BP_IS_HOLE(bp) || !zp->zp_dedup) 1126 return (ZIO_PIPELINE_CONTINUE); 1127 1128 ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags & 1129 ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify); 1130 1131 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) { 1132 BP_SET_DEDUP(bp, 1); 1133 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE; 1134 return (ZIO_PIPELINE_CONTINUE); 1135 } 1136 zio->io_bp_override = NULL; 1137 BP_ZERO(bp); 1138 } 1139 1140 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) { 1141 /* 1142 * We're rewriting an existing block, which means we're 1143 * working on behalf of spa_sync(). For spa_sync() to 1144 * converge, it must eventually be the case that we don't 1145 * have to allocate new blocks. But compression changes 1146 * the blocksize, which forces a reallocate, and makes 1147 * convergence take longer. Therefore, after the first 1148 * few passes, stop compressing to ensure convergence. 1149 */ 1150 pass = spa_sync_pass(spa); 1151 1152 ASSERT(zio->io_txg == spa_syncing_txg(spa)); 1153 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1154 ASSERT(!BP_GET_DEDUP(bp)); 1155 1156 if (pass >= zfs_sync_pass_dont_compress) 1157 compress = ZIO_COMPRESS_OFF; 1158 1159 /* Make sure someone doesn't change their mind on overwrites */ 1160 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp), 1161 spa_max_replication(spa)) == BP_GET_NDVAS(bp)); 1162 } 1163 1164 if (compress != ZIO_COMPRESS_OFF) { 1165 void *cbuf = zio_buf_alloc(lsize); 1166 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize); 1167 if (psize == 0 || psize == lsize) { 1168 compress = ZIO_COMPRESS_OFF; 1169 zio_buf_free(cbuf, lsize); 1170 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE && 1171 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) && 1172 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) { 1173 encode_embedded_bp_compressed(bp, 1174 cbuf, compress, lsize, psize); 1175 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA); 1176 BP_SET_TYPE(bp, zio->io_prop.zp_type); 1177 BP_SET_LEVEL(bp, zio->io_prop.zp_level); 1178 zio_buf_free(cbuf, lsize); 1179 bp->blk_birth = zio->io_txg; 1180 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1181 ASSERT(spa_feature_is_active(spa, 1182 SPA_FEATURE_EMBEDDED_DATA)); 1183 return (ZIO_PIPELINE_CONTINUE); 1184 } else { 1185 /* 1186 * Round up compressed size up to the ashift 1187 * of the smallest-ashift device, and zero the tail. 1188 * This ensures that the compressed size of the BP 1189 * (and thus compressratio property) are correct, 1190 * in that we charge for the padding used to fill out 1191 * the last sector. 1192 */ 1193 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT); 1194 size_t rounded = (size_t)P2ROUNDUP(psize, 1195 1ULL << spa->spa_min_ashift); 1196 if (rounded >= lsize) { 1197 compress = ZIO_COMPRESS_OFF; 1198 zio_buf_free(cbuf, lsize); 1199 psize = lsize; 1200 } else { 1201 bzero((char *)cbuf + psize, rounded - psize); 1202 psize = rounded; 1203 zio_push_transform(zio, cbuf, 1204 psize, lsize, NULL); 1205 } 1206 } 1207 } 1208 1209 /* 1210 * The final pass of spa_sync() must be all rewrites, but the first 1211 * few passes offer a trade-off: allocating blocks defers convergence, 1212 * but newly allocated blocks are sequential, so they can be written 1213 * to disk faster. Therefore, we allow the first few passes of 1214 * spa_sync() to allocate new blocks, but force rewrites after that. 1215 * There should only be a handful of blocks after pass 1 in any case. 1216 */ 1217 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg && 1218 BP_GET_PSIZE(bp) == psize && 1219 pass >= zfs_sync_pass_rewrite) { 1220 ASSERT(psize != 0); 1221 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES; 1222 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages; 1223 zio->io_flags |= ZIO_FLAG_IO_REWRITE; 1224 } else { 1225 BP_ZERO(bp); 1226 zio->io_pipeline = ZIO_WRITE_PIPELINE; 1227 } 1228 1229 if (psize == 0) { 1230 if (zio->io_bp_orig.blk_birth != 0 && 1231 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) { 1232 BP_SET_LSIZE(bp, lsize); 1233 BP_SET_TYPE(bp, zp->zp_type); 1234 BP_SET_LEVEL(bp, zp->zp_level); 1235 BP_SET_BIRTH(bp, zio->io_txg, 0); 1236 } 1237 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1238 } else { 1239 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER); 1240 BP_SET_LSIZE(bp, lsize); 1241 BP_SET_TYPE(bp, zp->zp_type); 1242 BP_SET_LEVEL(bp, zp->zp_level); 1243 BP_SET_PSIZE(bp, psize); 1244 BP_SET_COMPRESS(bp, compress); 1245 BP_SET_CHECKSUM(bp, zp->zp_checksum); 1246 BP_SET_DEDUP(bp, zp->zp_dedup); 1247 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER); 1248 if (zp->zp_dedup) { 1249 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1250 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 1251 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE; 1252 } 1253 if (zp->zp_nopwrite) { 1254 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1255 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 1256 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE; 1257 } 1258 } 1259 1260 return (ZIO_PIPELINE_CONTINUE); 1261 } 1262 1263 static int 1264 zio_free_bp_init(zio_t *zio) 1265 { 1266 blkptr_t *bp = zio->io_bp; 1267 1268 if (zio->io_child_type == ZIO_CHILD_LOGICAL) { 1269 if (BP_GET_DEDUP(bp)) 1270 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE; 1271 } 1272 1273 return (ZIO_PIPELINE_CONTINUE); 1274 } 1275 1276 /* 1277 * ========================================================================== 1278 * Execute the I/O pipeline 1279 * ========================================================================== 1280 */ 1281 1282 static void 1283 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline) 1284 { 1285 spa_t *spa = zio->io_spa; 1286 zio_type_t t = zio->io_type; 1287 int flags = (cutinline ? TQ_FRONT : 0); 1288 1289 /* 1290 * If we're a config writer or a probe, the normal issue and 1291 * interrupt threads may all be blocked waiting for the config lock. 1292 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL. 1293 */ 1294 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE)) 1295 t = ZIO_TYPE_NULL; 1296 1297 /* 1298 * A similar issue exists for the L2ARC write thread until L2ARC 2.0. 1299 */ 1300 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux) 1301 t = ZIO_TYPE_NULL; 1302 1303 /* 1304 * If this is a high priority I/O, then use the high priority taskq if 1305 * available. 1306 */ 1307 if (zio->io_priority == ZIO_PRIORITY_NOW && 1308 spa->spa_zio_taskq[t][q + 1].stqs_count != 0) 1309 q++; 1310 1311 ASSERT3U(q, <, ZIO_TASKQ_TYPES); 1312 1313 /* 1314 * NB: We are assuming that the zio can only be dispatched 1315 * to a single taskq at a time. It would be a grievous error 1316 * to dispatch the zio to another taskq at the same time. 1317 */ 1318 ASSERT(zio->io_tqent.tqent_next == NULL); 1319 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio, 1320 flags, &zio->io_tqent); 1321 } 1322 1323 static boolean_t 1324 zio_taskq_member(zio_t *zio, zio_taskq_type_t q) 1325 { 1326 kthread_t *executor = zio->io_executor; 1327 spa_t *spa = zio->io_spa; 1328 1329 for (zio_type_t t = 0; t < ZIO_TYPES; t++) { 1330 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q]; 1331 uint_t i; 1332 for (i = 0; i < tqs->stqs_count; i++) { 1333 if (taskq_member(tqs->stqs_taskq[i], executor)) 1334 return (B_TRUE); 1335 } 1336 } 1337 1338 return (B_FALSE); 1339 } 1340 1341 static int 1342 zio_issue_async(zio_t *zio) 1343 { 1344 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); 1345 1346 return (ZIO_PIPELINE_STOP); 1347 } 1348 1349 void 1350 zio_interrupt(zio_t *zio) 1351 { 1352 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE); 1353 } 1354 1355 void 1356 zio_delay_interrupt(zio_t *zio) 1357 { 1358 /* 1359 * The timeout_generic() function isn't defined in userspace, so 1360 * rather than trying to implement the function, the zio delay 1361 * functionality has been disabled for userspace builds. 1362 */ 1363 1364 #ifdef _KERNEL 1365 /* 1366 * If io_target_timestamp is zero, then no delay has been registered 1367 * for this IO, thus jump to the end of this function and "skip" the 1368 * delay; issuing it directly to the zio layer. 1369 */ 1370 if (zio->io_target_timestamp != 0) { 1371 hrtime_t now = gethrtime(); 1372 1373 if (now >= zio->io_target_timestamp) { 1374 /* 1375 * This IO has already taken longer than the target 1376 * delay to complete, so we don't want to delay it 1377 * any longer; we "miss" the delay and issue it 1378 * directly to the zio layer. This is likely due to 1379 * the target latency being set to a value less than 1380 * the underlying hardware can satisfy (e.g. delay 1381 * set to 1ms, but the disks take 10ms to complete an 1382 * IO request). 1383 */ 1384 1385 DTRACE_PROBE2(zio__delay__miss, zio_t *, zio, 1386 hrtime_t, now); 1387 1388 zio_interrupt(zio); 1389 } else { 1390 hrtime_t diff = zio->io_target_timestamp - now; 1391 1392 DTRACE_PROBE3(zio__delay__hit, zio_t *, zio, 1393 hrtime_t, now, hrtime_t, diff); 1394 1395 (void) timeout_generic(CALLOUT_NORMAL, 1396 (void (*)(void *))zio_interrupt, zio, diff, 1, 0); 1397 } 1398 1399 return; 1400 } 1401 #endif 1402 1403 DTRACE_PROBE1(zio__delay__skip, zio_t *, zio); 1404 zio_interrupt(zio); 1405 } 1406 1407 /* 1408 * Execute the I/O pipeline until one of the following occurs: 1409 * 1410 * (1) the I/O completes 1411 * (2) the pipeline stalls waiting for dependent child I/Os 1412 * (3) the I/O issues, so we're waiting for an I/O completion interrupt 1413 * (4) the I/O is delegated by vdev-level caching or aggregation 1414 * (5) the I/O is deferred due to vdev-level queueing 1415 * (6) the I/O is handed off to another thread. 1416 * 1417 * In all cases, the pipeline stops whenever there's no CPU work; it never 1418 * burns a thread in cv_wait(). 1419 * 1420 * There's no locking on io_stage because there's no legitimate way 1421 * for multiple threads to be attempting to process the same I/O. 1422 */ 1423 static zio_pipe_stage_t *zio_pipeline[]; 1424 1425 void 1426 zio_execute(zio_t *zio) 1427 { 1428 zio->io_executor = curthread; 1429 1430 while (zio->io_stage < ZIO_STAGE_DONE) { 1431 enum zio_stage pipeline = zio->io_pipeline; 1432 enum zio_stage stage = zio->io_stage; 1433 int rv; 1434 1435 ASSERT(!MUTEX_HELD(&zio->io_lock)); 1436 ASSERT(ISP2(stage)); 1437 ASSERT(zio->io_stall == NULL); 1438 1439 do { 1440 stage <<= 1; 1441 } while ((stage & pipeline) == 0); 1442 1443 ASSERT(stage <= ZIO_STAGE_DONE); 1444 1445 /* 1446 * If we are in interrupt context and this pipeline stage 1447 * will grab a config lock that is held across I/O, 1448 * or may wait for an I/O that needs an interrupt thread 1449 * to complete, issue async to avoid deadlock. 1450 * 1451 * For VDEV_IO_START, we cut in line so that the io will 1452 * be sent to disk promptly. 1453 */ 1454 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL && 1455 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) { 1456 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? 1457 zio_requeue_io_start_cut_in_line : B_FALSE; 1458 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); 1459 return; 1460 } 1461 1462 zio->io_stage = stage; 1463 rv = zio_pipeline[highbit64(stage) - 1](zio); 1464 1465 if (rv == ZIO_PIPELINE_STOP) 1466 return; 1467 1468 ASSERT(rv == ZIO_PIPELINE_CONTINUE); 1469 } 1470 } 1471 1472 /* 1473 * ========================================================================== 1474 * Initiate I/O, either sync or async 1475 * ========================================================================== 1476 */ 1477 int 1478 zio_wait(zio_t *zio) 1479 { 1480 int error; 1481 1482 ASSERT(zio->io_stage == ZIO_STAGE_OPEN); 1483 ASSERT(zio->io_executor == NULL); 1484 1485 zio->io_waiter = curthread; 1486 1487 zio_execute(zio); 1488 1489 mutex_enter(&zio->io_lock); 1490 while (zio->io_executor != NULL) 1491 cv_wait(&zio->io_cv, &zio->io_lock); 1492 mutex_exit(&zio->io_lock); 1493 1494 error = zio->io_error; 1495 zio_destroy(zio); 1496 1497 return (error); 1498 } 1499 1500 void 1501 zio_nowait(zio_t *zio) 1502 { 1503 ASSERT(zio->io_executor == NULL); 1504 1505 if (zio->io_child_type == ZIO_CHILD_LOGICAL && 1506 zio_unique_parent(zio) == NULL) { 1507 /* 1508 * This is a logical async I/O with no parent to wait for it. 1509 * We add it to the spa_async_root_zio "Godfather" I/O which 1510 * will ensure they complete prior to unloading the pool. 1511 */ 1512 spa_t *spa = zio->io_spa; 1513 1514 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio); 1515 } 1516 1517 zio_execute(zio); 1518 } 1519 1520 /* 1521 * ========================================================================== 1522 * Reexecute or suspend/resume failed I/O 1523 * ========================================================================== 1524 */ 1525 1526 static void 1527 zio_reexecute(zio_t *pio) 1528 { 1529 zio_t *cio, *cio_next; 1530 1531 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL); 1532 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN); 1533 ASSERT(pio->io_gang_leader == NULL); 1534 ASSERT(pio->io_gang_tree == NULL); 1535 1536 pio->io_flags = pio->io_orig_flags; 1537 pio->io_stage = pio->io_orig_stage; 1538 pio->io_pipeline = pio->io_orig_pipeline; 1539 pio->io_reexecute = 0; 1540 pio->io_flags |= ZIO_FLAG_REEXECUTED; 1541 pio->io_error = 0; 1542 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 1543 pio->io_state[w] = 0; 1544 for (int c = 0; c < ZIO_CHILD_TYPES; c++) 1545 pio->io_child_error[c] = 0; 1546 1547 if (IO_IS_ALLOCATING(pio)) 1548 BP_ZERO(pio->io_bp); 1549 1550 /* 1551 * As we reexecute pio's children, new children could be created. 1552 * New children go to the head of pio's io_child_list, however, 1553 * so we will (correctly) not reexecute them. The key is that 1554 * the remainder of pio's io_child_list, from 'cio_next' onward, 1555 * cannot be affected by any side effects of reexecuting 'cio'. 1556 */ 1557 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) { 1558 cio_next = zio_walk_children(pio); 1559 mutex_enter(&pio->io_lock); 1560 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 1561 pio->io_children[cio->io_child_type][w]++; 1562 mutex_exit(&pio->io_lock); 1563 zio_reexecute(cio); 1564 } 1565 1566 /* 1567 * Now that all children have been reexecuted, execute the parent. 1568 * We don't reexecute "The Godfather" I/O here as it's the 1569 * responsibility of the caller to wait on him. 1570 */ 1571 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) 1572 zio_execute(pio); 1573 } 1574 1575 void 1576 zio_suspend(spa_t *spa, zio_t *zio) 1577 { 1578 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC) 1579 fm_panic("Pool '%s' has encountered an uncorrectable I/O " 1580 "failure and the failure mode property for this pool " 1581 "is set to panic.", spa_name(spa)); 1582 1583 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0); 1584 1585 mutex_enter(&spa->spa_suspend_lock); 1586 1587 if (spa->spa_suspend_zio_root == NULL) 1588 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL, 1589 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | 1590 ZIO_FLAG_GODFATHER); 1591 1592 spa->spa_suspended = B_TRUE; 1593 1594 if (zio != NULL) { 1595 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); 1596 ASSERT(zio != spa->spa_suspend_zio_root); 1597 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1598 ASSERT(zio_unique_parent(zio) == NULL); 1599 ASSERT(zio->io_stage == ZIO_STAGE_DONE); 1600 zio_add_child(spa->spa_suspend_zio_root, zio); 1601 } 1602 1603 mutex_exit(&spa->spa_suspend_lock); 1604 } 1605 1606 int 1607 zio_resume(spa_t *spa) 1608 { 1609 zio_t *pio; 1610 1611 /* 1612 * Reexecute all previously suspended i/o. 1613 */ 1614 mutex_enter(&spa->spa_suspend_lock); 1615 spa->spa_suspended = B_FALSE; 1616 cv_broadcast(&spa->spa_suspend_cv); 1617 pio = spa->spa_suspend_zio_root; 1618 spa->spa_suspend_zio_root = NULL; 1619 mutex_exit(&spa->spa_suspend_lock); 1620 1621 if (pio == NULL) 1622 return (0); 1623 1624 zio_reexecute(pio); 1625 return (zio_wait(pio)); 1626 } 1627 1628 void 1629 zio_resume_wait(spa_t *spa) 1630 { 1631 mutex_enter(&spa->spa_suspend_lock); 1632 while (spa_suspended(spa)) 1633 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock); 1634 mutex_exit(&spa->spa_suspend_lock); 1635 } 1636 1637 /* 1638 * ========================================================================== 1639 * Gang blocks. 1640 * 1641 * A gang block is a collection of small blocks that looks to the DMU 1642 * like one large block. When zio_dva_allocate() cannot find a block 1643 * of the requested size, due to either severe fragmentation or the pool 1644 * being nearly full, it calls zio_write_gang_block() to construct the 1645 * block from smaller fragments. 1646 * 1647 * A gang block consists of a gang header (zio_gbh_phys_t) and up to 1648 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like 1649 * an indirect block: it's an array of block pointers. It consumes 1650 * only one sector and hence is allocatable regardless of fragmentation. 1651 * The gang header's bps point to its gang members, which hold the data. 1652 * 1653 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg> 1654 * as the verifier to ensure uniqueness of the SHA256 checksum. 1655 * Critically, the gang block bp's blk_cksum is the checksum of the data, 1656 * not the gang header. This ensures that data block signatures (needed for 1657 * deduplication) are independent of how the block is physically stored. 1658 * 1659 * Gang blocks can be nested: a gang member may itself be a gang block. 1660 * Thus every gang block is a tree in which root and all interior nodes are 1661 * gang headers, and the leaves are normal blocks that contain user data. 1662 * The root of the gang tree is called the gang leader. 1663 * 1664 * To perform any operation (read, rewrite, free, claim) on a gang block, 1665 * zio_gang_assemble() first assembles the gang tree (minus data leaves) 1666 * in the io_gang_tree field of the original logical i/o by recursively 1667 * reading the gang leader and all gang headers below it. This yields 1668 * an in-core tree containing the contents of every gang header and the 1669 * bps for every constituent of the gang block. 1670 * 1671 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree 1672 * and invokes a callback on each bp. To free a gang block, zio_gang_issue() 1673 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp. 1674 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim(). 1675 * zio_read_gang() is a wrapper around zio_read() that omits reading gang 1676 * headers, since we already have those in io_gang_tree. zio_rewrite_gang() 1677 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite() 1678 * of the gang header plus zio_checksum_compute() of the data to update the 1679 * gang header's blk_cksum as described above. 1680 * 1681 * The two-phase assemble/issue model solves the problem of partial failure -- 1682 * what if you'd freed part of a gang block but then couldn't read the 1683 * gang header for another part? Assembling the entire gang tree first 1684 * ensures that all the necessary gang header I/O has succeeded before 1685 * starting the actual work of free, claim, or write. Once the gang tree 1686 * is assembled, free and claim are in-memory operations that cannot fail. 1687 * 1688 * In the event that a gang write fails, zio_dva_unallocate() walks the 1689 * gang tree to immediately free (i.e. insert back into the space map) 1690 * everything we've allocated. This ensures that we don't get ENOSPC 1691 * errors during repeated suspend/resume cycles due to a flaky device. 1692 * 1693 * Gang rewrites only happen during sync-to-convergence. If we can't assemble 1694 * the gang tree, we won't modify the block, so we can safely defer the free 1695 * (knowing that the block is still intact). If we *can* assemble the gang 1696 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free 1697 * each constituent bp and we can allocate a new block on the next sync pass. 1698 * 1699 * In all cases, the gang tree allows complete recovery from partial failure. 1700 * ========================================================================== 1701 */ 1702 1703 static zio_t * 1704 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1705 { 1706 if (gn != NULL) 1707 return (pio); 1708 1709 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp), 1710 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), 1711 &pio->io_bookmark)); 1712 } 1713 1714 zio_t * 1715 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1716 { 1717 zio_t *zio; 1718 1719 if (gn != NULL) { 1720 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, 1721 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority, 1722 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 1723 /* 1724 * As we rewrite each gang header, the pipeline will compute 1725 * a new gang block header checksum for it; but no one will 1726 * compute a new data checksum, so we do that here. The one 1727 * exception is the gang leader: the pipeline already computed 1728 * its data checksum because that stage precedes gang assembly. 1729 * (Presently, nothing actually uses interior data checksums; 1730 * this is just good hygiene.) 1731 */ 1732 if (gn != pio->io_gang_leader->io_gang_tree) { 1733 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp), 1734 data, BP_GET_PSIZE(bp)); 1735 } 1736 /* 1737 * If we are here to damage data for testing purposes, 1738 * leave the GBH alone so that we can detect the damage. 1739 */ 1740 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE) 1741 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; 1742 } else { 1743 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, 1744 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority, 1745 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 1746 } 1747 1748 return (zio); 1749 } 1750 1751 /* ARGSUSED */ 1752 zio_t * 1753 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1754 { 1755 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp, 1756 ZIO_GANG_CHILD_FLAGS(pio))); 1757 } 1758 1759 /* ARGSUSED */ 1760 zio_t * 1761 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1762 { 1763 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp, 1764 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio))); 1765 } 1766 1767 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = { 1768 NULL, 1769 zio_read_gang, 1770 zio_rewrite_gang, 1771 zio_free_gang, 1772 zio_claim_gang, 1773 NULL 1774 }; 1775 1776 static void zio_gang_tree_assemble_done(zio_t *zio); 1777 1778 static zio_gang_node_t * 1779 zio_gang_node_alloc(zio_gang_node_t **gnpp) 1780 { 1781 zio_gang_node_t *gn; 1782 1783 ASSERT(*gnpp == NULL); 1784 1785 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP); 1786 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE); 1787 *gnpp = gn; 1788 1789 return (gn); 1790 } 1791 1792 static void 1793 zio_gang_node_free(zio_gang_node_t **gnpp) 1794 { 1795 zio_gang_node_t *gn = *gnpp; 1796 1797 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) 1798 ASSERT(gn->gn_child[g] == NULL); 1799 1800 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE); 1801 kmem_free(gn, sizeof (*gn)); 1802 *gnpp = NULL; 1803 } 1804 1805 static void 1806 zio_gang_tree_free(zio_gang_node_t **gnpp) 1807 { 1808 zio_gang_node_t *gn = *gnpp; 1809 1810 if (gn == NULL) 1811 return; 1812 1813 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) 1814 zio_gang_tree_free(&gn->gn_child[g]); 1815 1816 zio_gang_node_free(gnpp); 1817 } 1818 1819 static void 1820 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp) 1821 { 1822 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp); 1823 1824 ASSERT(gio->io_gang_leader == gio); 1825 ASSERT(BP_IS_GANG(bp)); 1826 1827 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh, 1828 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn, 1829 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark)); 1830 } 1831 1832 static void 1833 zio_gang_tree_assemble_done(zio_t *zio) 1834 { 1835 zio_t *gio = zio->io_gang_leader; 1836 zio_gang_node_t *gn = zio->io_private; 1837 blkptr_t *bp = zio->io_bp; 1838 1839 ASSERT(gio == zio_unique_parent(zio)); 1840 ASSERT(zio->io_child_count == 0); 1841 1842 if (zio->io_error) 1843 return; 1844 1845 if (BP_SHOULD_BYTESWAP(bp)) 1846 byteswap_uint64_array(zio->io_data, zio->io_size); 1847 1848 ASSERT(zio->io_data == gn->gn_gbh); 1849 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE); 1850 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); 1851 1852 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 1853 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; 1854 if (!BP_IS_GANG(gbp)) 1855 continue; 1856 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]); 1857 } 1858 } 1859 1860 static void 1861 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data) 1862 { 1863 zio_t *gio = pio->io_gang_leader; 1864 zio_t *zio; 1865 1866 ASSERT(BP_IS_GANG(bp) == !!gn); 1867 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp)); 1868 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree); 1869 1870 /* 1871 * If you're a gang header, your data is in gn->gn_gbh. 1872 * If you're a gang member, your data is in 'data' and gn == NULL. 1873 */ 1874 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data); 1875 1876 if (gn != NULL) { 1877 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); 1878 1879 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 1880 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; 1881 if (BP_IS_HOLE(gbp)) 1882 continue; 1883 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data); 1884 data = (char *)data + BP_GET_PSIZE(gbp); 1885 } 1886 } 1887 1888 if (gn == gio->io_gang_tree) 1889 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data); 1890 1891 if (zio != pio) 1892 zio_nowait(zio); 1893 } 1894 1895 static int 1896 zio_gang_assemble(zio_t *zio) 1897 { 1898 blkptr_t *bp = zio->io_bp; 1899 1900 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL); 1901 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 1902 1903 zio->io_gang_leader = zio; 1904 1905 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree); 1906 1907 return (ZIO_PIPELINE_CONTINUE); 1908 } 1909 1910 static int 1911 zio_gang_issue(zio_t *zio) 1912 { 1913 blkptr_t *bp = zio->io_bp; 1914 1915 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE)) 1916 return (ZIO_PIPELINE_STOP); 1917 1918 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio); 1919 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 1920 1921 if (zio->io_child_error[ZIO_CHILD_GANG] == 0) 1922 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data); 1923 else 1924 zio_gang_tree_free(&zio->io_gang_tree); 1925 1926 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1927 1928 return (ZIO_PIPELINE_CONTINUE); 1929 } 1930 1931 static void 1932 zio_write_gang_member_ready(zio_t *zio) 1933 { 1934 zio_t *pio = zio_unique_parent(zio); 1935 zio_t *gio = zio->io_gang_leader; 1936 dva_t *cdva = zio->io_bp->blk_dva; 1937 dva_t *pdva = pio->io_bp->blk_dva; 1938 uint64_t asize; 1939 1940 if (BP_IS_HOLE(zio->io_bp)) 1941 return; 1942 1943 ASSERT(BP_IS_HOLE(&zio->io_bp_orig)); 1944 1945 ASSERT(zio->io_child_type == ZIO_CHILD_GANG); 1946 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies); 1947 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp)); 1948 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp)); 1949 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp)); 1950 1951 mutex_enter(&pio->io_lock); 1952 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) { 1953 ASSERT(DVA_GET_GANG(&pdva[d])); 1954 asize = DVA_GET_ASIZE(&pdva[d]); 1955 asize += DVA_GET_ASIZE(&cdva[d]); 1956 DVA_SET_ASIZE(&pdva[d], asize); 1957 } 1958 mutex_exit(&pio->io_lock); 1959 } 1960 1961 static int 1962 zio_write_gang_block(zio_t *pio) 1963 { 1964 spa_t *spa = pio->io_spa; 1965 blkptr_t *bp = pio->io_bp; 1966 zio_t *gio = pio->io_gang_leader; 1967 zio_t *zio; 1968 zio_gang_node_t *gn, **gnpp; 1969 zio_gbh_phys_t *gbh; 1970 uint64_t txg = pio->io_txg; 1971 uint64_t resid = pio->io_size; 1972 uint64_t lsize; 1973 int copies = gio->io_prop.zp_copies; 1974 int gbh_copies = MIN(copies + 1, spa_max_replication(spa)); 1975 zio_prop_t zp; 1976 int error; 1977 1978 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE, 1979 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, 1980 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER); 1981 if (error) { 1982 pio->io_error = error; 1983 return (ZIO_PIPELINE_CONTINUE); 1984 } 1985 1986 if (pio == gio) { 1987 gnpp = &gio->io_gang_tree; 1988 } else { 1989 gnpp = pio->io_private; 1990 ASSERT(pio->io_ready == zio_write_gang_member_ready); 1991 } 1992 1993 gn = zio_gang_node_alloc(gnpp); 1994 gbh = gn->gn_gbh; 1995 bzero(gbh, SPA_GANGBLOCKSIZE); 1996 1997 /* 1998 * Create the gang header. 1999 */ 2000 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL, 2001 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 2002 2003 /* 2004 * Create and nowait the gang children. 2005 */ 2006 for (int g = 0; resid != 0; resid -= lsize, g++) { 2007 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g), 2008 SPA_MINBLOCKSIZE); 2009 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid); 2010 2011 zp.zp_checksum = gio->io_prop.zp_checksum; 2012 zp.zp_compress = ZIO_COMPRESS_OFF; 2013 zp.zp_type = DMU_OT_NONE; 2014 zp.zp_level = 0; 2015 zp.zp_copies = gio->io_prop.zp_copies; 2016 zp.zp_dedup = B_FALSE; 2017 zp.zp_dedup_verify = B_FALSE; 2018 zp.zp_nopwrite = B_FALSE; 2019 2020 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g], 2021 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp, 2022 zio_write_gang_member_ready, NULL, NULL, &gn->gn_child[g], 2023 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), 2024 &pio->io_bookmark)); 2025 } 2026 2027 /* 2028 * Set pio's pipeline to just wait for zio to finish. 2029 */ 2030 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2031 2032 zio_nowait(zio); 2033 2034 return (ZIO_PIPELINE_CONTINUE); 2035 } 2036 2037 /* 2038 * The zio_nop_write stage in the pipeline determines if allocating a 2039 * new bp is necessary. The nopwrite feature can handle writes in 2040 * either syncing or open context (i.e. zil writes) and as a result is 2041 * mutually exclusive with dedup. 2042 * 2043 * By leveraging a cryptographically secure checksum, such as SHA256, we 2044 * can compare the checksums of the new data and the old to determine if 2045 * allocating a new block is required. Note that our requirements for 2046 * cryptographic strength are fairly weak: there can't be any accidental 2047 * hash collisions, but we don't need to be secure against intentional 2048 * (malicious) collisions. To trigger a nopwrite, you have to be able 2049 * to write the file to begin with, and triggering an incorrect (hash 2050 * collision) nopwrite is no worse than simply writing to the file. 2051 * That said, there are no known attacks against the checksum algorithms 2052 * used for nopwrite, assuming that the salt and the checksums 2053 * themselves remain secret. 2054 */ 2055 static int 2056 zio_nop_write(zio_t *zio) 2057 { 2058 blkptr_t *bp = zio->io_bp; 2059 blkptr_t *bp_orig = &zio->io_bp_orig; 2060 zio_prop_t *zp = &zio->io_prop; 2061 2062 ASSERT(BP_GET_LEVEL(bp) == 0); 2063 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 2064 ASSERT(zp->zp_nopwrite); 2065 ASSERT(!zp->zp_dedup); 2066 ASSERT(zio->io_bp_override == NULL); 2067 ASSERT(IO_IS_ALLOCATING(zio)); 2068 2069 /* 2070 * Check to see if the original bp and the new bp have matching 2071 * characteristics (i.e. same checksum, compression algorithms, etc). 2072 * If they don't then just continue with the pipeline which will 2073 * allocate a new bp. 2074 */ 2075 if (BP_IS_HOLE(bp_orig) || 2076 !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags & 2077 ZCHECKSUM_FLAG_NOPWRITE) || 2078 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) || 2079 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) || 2080 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) || 2081 zp->zp_copies != BP_GET_NDVAS(bp_orig)) 2082 return (ZIO_PIPELINE_CONTINUE); 2083 2084 /* 2085 * If the checksums match then reset the pipeline so that we 2086 * avoid allocating a new bp and issuing any I/O. 2087 */ 2088 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) { 2089 ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags & 2090 ZCHECKSUM_FLAG_NOPWRITE); 2091 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig)); 2092 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig)); 2093 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF); 2094 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop, 2095 sizeof (uint64_t)) == 0); 2096 2097 *bp = *bp_orig; 2098 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2099 zio->io_flags |= ZIO_FLAG_NOPWRITE; 2100 } 2101 2102 return (ZIO_PIPELINE_CONTINUE); 2103 } 2104 2105 /* 2106 * ========================================================================== 2107 * Dedup 2108 * ========================================================================== 2109 */ 2110 static void 2111 zio_ddt_child_read_done(zio_t *zio) 2112 { 2113 blkptr_t *bp = zio->io_bp; 2114 ddt_entry_t *dde = zio->io_private; 2115 ddt_phys_t *ddp; 2116 zio_t *pio = zio_unique_parent(zio); 2117 2118 mutex_enter(&pio->io_lock); 2119 ddp = ddt_phys_select(dde, bp); 2120 if (zio->io_error == 0) 2121 ddt_phys_clear(ddp); /* this ddp doesn't need repair */ 2122 if (zio->io_error == 0 && dde->dde_repair_data == NULL) 2123 dde->dde_repair_data = zio->io_data; 2124 else 2125 zio_buf_free(zio->io_data, zio->io_size); 2126 mutex_exit(&pio->io_lock); 2127 } 2128 2129 static int 2130 zio_ddt_read_start(zio_t *zio) 2131 { 2132 blkptr_t *bp = zio->io_bp; 2133 2134 ASSERT(BP_GET_DEDUP(bp)); 2135 ASSERT(BP_GET_PSIZE(bp) == zio->io_size); 2136 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2137 2138 if (zio->io_child_error[ZIO_CHILD_DDT]) { 2139 ddt_t *ddt = ddt_select(zio->io_spa, bp); 2140 ddt_entry_t *dde = ddt_repair_start(ddt, bp); 2141 ddt_phys_t *ddp = dde->dde_phys; 2142 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp); 2143 blkptr_t blk; 2144 2145 ASSERT(zio->io_vsd == NULL); 2146 zio->io_vsd = dde; 2147 2148 if (ddp_self == NULL) 2149 return (ZIO_PIPELINE_CONTINUE); 2150 2151 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) { 2152 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self) 2153 continue; 2154 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp, 2155 &blk); 2156 zio_nowait(zio_read(zio, zio->io_spa, &blk, 2157 zio_buf_alloc(zio->io_size), zio->io_size, 2158 zio_ddt_child_read_done, dde, zio->io_priority, 2159 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE, 2160 &zio->io_bookmark)); 2161 } 2162 return (ZIO_PIPELINE_CONTINUE); 2163 } 2164 2165 zio_nowait(zio_read(zio, zio->io_spa, bp, 2166 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority, 2167 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark)); 2168 2169 return (ZIO_PIPELINE_CONTINUE); 2170 } 2171 2172 static int 2173 zio_ddt_read_done(zio_t *zio) 2174 { 2175 blkptr_t *bp = zio->io_bp; 2176 2177 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE)) 2178 return (ZIO_PIPELINE_STOP); 2179 2180 ASSERT(BP_GET_DEDUP(bp)); 2181 ASSERT(BP_GET_PSIZE(bp) == zio->io_size); 2182 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2183 2184 if (zio->io_child_error[ZIO_CHILD_DDT]) { 2185 ddt_t *ddt = ddt_select(zio->io_spa, bp); 2186 ddt_entry_t *dde = zio->io_vsd; 2187 if (ddt == NULL) { 2188 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE); 2189 return (ZIO_PIPELINE_CONTINUE); 2190 } 2191 if (dde == NULL) { 2192 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1; 2193 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); 2194 return (ZIO_PIPELINE_STOP); 2195 } 2196 if (dde->dde_repair_data != NULL) { 2197 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size); 2198 zio->io_child_error[ZIO_CHILD_DDT] = 0; 2199 } 2200 ddt_repair_done(ddt, dde); 2201 zio->io_vsd = NULL; 2202 } 2203 2204 ASSERT(zio->io_vsd == NULL); 2205 2206 return (ZIO_PIPELINE_CONTINUE); 2207 } 2208 2209 static boolean_t 2210 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde) 2211 { 2212 spa_t *spa = zio->io_spa; 2213 2214 /* 2215 * Note: we compare the original data, not the transformed data, 2216 * because when zio->io_bp is an override bp, we will not have 2217 * pushed the I/O transforms. That's an important optimization 2218 * because otherwise we'd compress/encrypt all dmu_sync() data twice. 2219 */ 2220 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { 2221 zio_t *lio = dde->dde_lead_zio[p]; 2222 2223 if (lio != NULL) { 2224 return (lio->io_orig_size != zio->io_orig_size || 2225 bcmp(zio->io_orig_data, lio->io_orig_data, 2226 zio->io_orig_size) != 0); 2227 } 2228 } 2229 2230 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { 2231 ddt_phys_t *ddp = &dde->dde_phys[p]; 2232 2233 if (ddp->ddp_phys_birth != 0) { 2234 arc_buf_t *abuf = NULL; 2235 arc_flags_t aflags = ARC_FLAG_WAIT; 2236 blkptr_t blk = *zio->io_bp; 2237 int error; 2238 2239 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth); 2240 2241 ddt_exit(ddt); 2242 2243 error = arc_read(NULL, spa, &blk, 2244 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ, 2245 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2246 &aflags, &zio->io_bookmark); 2247 2248 if (error == 0) { 2249 if (arc_buf_size(abuf) != zio->io_orig_size || 2250 bcmp(abuf->b_data, zio->io_orig_data, 2251 zio->io_orig_size) != 0) 2252 error = SET_ERROR(EEXIST); 2253 VERIFY(arc_buf_remove_ref(abuf, &abuf)); 2254 } 2255 2256 ddt_enter(ddt); 2257 return (error != 0); 2258 } 2259 } 2260 2261 return (B_FALSE); 2262 } 2263 2264 static void 2265 zio_ddt_child_write_ready(zio_t *zio) 2266 { 2267 int p = zio->io_prop.zp_copies; 2268 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); 2269 ddt_entry_t *dde = zio->io_private; 2270 ddt_phys_t *ddp = &dde->dde_phys[p]; 2271 zio_t *pio; 2272 2273 if (zio->io_error) 2274 return; 2275 2276 ddt_enter(ddt); 2277 2278 ASSERT(dde->dde_lead_zio[p] == zio); 2279 2280 ddt_phys_fill(ddp, zio->io_bp); 2281 2282 while ((pio = zio_walk_parents(zio)) != NULL) 2283 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg); 2284 2285 ddt_exit(ddt); 2286 } 2287 2288 static void 2289 zio_ddt_child_write_done(zio_t *zio) 2290 { 2291 int p = zio->io_prop.zp_copies; 2292 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); 2293 ddt_entry_t *dde = zio->io_private; 2294 ddt_phys_t *ddp = &dde->dde_phys[p]; 2295 2296 ddt_enter(ddt); 2297 2298 ASSERT(ddp->ddp_refcnt == 0); 2299 ASSERT(dde->dde_lead_zio[p] == zio); 2300 dde->dde_lead_zio[p] = NULL; 2301 2302 if (zio->io_error == 0) { 2303 while (zio_walk_parents(zio) != NULL) 2304 ddt_phys_addref(ddp); 2305 } else { 2306 ddt_phys_clear(ddp); 2307 } 2308 2309 ddt_exit(ddt); 2310 } 2311 2312 static void 2313 zio_ddt_ditto_write_done(zio_t *zio) 2314 { 2315 int p = DDT_PHYS_DITTO; 2316 zio_prop_t *zp = &zio->io_prop; 2317 blkptr_t *bp = zio->io_bp; 2318 ddt_t *ddt = ddt_select(zio->io_spa, bp); 2319 ddt_entry_t *dde = zio->io_private; 2320 ddt_phys_t *ddp = &dde->dde_phys[p]; 2321 ddt_key_t *ddk = &dde->dde_key; 2322 2323 ddt_enter(ddt); 2324 2325 ASSERT(ddp->ddp_refcnt == 0); 2326 ASSERT(dde->dde_lead_zio[p] == zio); 2327 dde->dde_lead_zio[p] = NULL; 2328 2329 if (zio->io_error == 0) { 2330 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum)); 2331 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP); 2332 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp)); 2333 if (ddp->ddp_phys_birth != 0) 2334 ddt_phys_free(ddt, ddk, ddp, zio->io_txg); 2335 ddt_phys_fill(ddp, bp); 2336 } 2337 2338 ddt_exit(ddt); 2339 } 2340 2341 static int 2342 zio_ddt_write(zio_t *zio) 2343 { 2344 spa_t *spa = zio->io_spa; 2345 blkptr_t *bp = zio->io_bp; 2346 uint64_t txg = zio->io_txg; 2347 zio_prop_t *zp = &zio->io_prop; 2348 int p = zp->zp_copies; 2349 int ditto_copies; 2350 zio_t *cio = NULL; 2351 zio_t *dio = NULL; 2352 ddt_t *ddt = ddt_select(spa, bp); 2353 ddt_entry_t *dde; 2354 ddt_phys_t *ddp; 2355 2356 ASSERT(BP_GET_DEDUP(bp)); 2357 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum); 2358 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override); 2359 2360 ddt_enter(ddt); 2361 dde = ddt_lookup(ddt, bp, B_TRUE); 2362 ddp = &dde->dde_phys[p]; 2363 2364 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) { 2365 /* 2366 * If we're using a weak checksum, upgrade to a strong checksum 2367 * and try again. If we're already using a strong checksum, 2368 * we can't resolve it, so just convert to an ordinary write. 2369 * (And automatically e-mail a paper to Nature?) 2370 */ 2371 if (!(zio_checksum_table[zp->zp_checksum].ci_flags & 2372 ZCHECKSUM_FLAG_DEDUP)) { 2373 zp->zp_checksum = spa_dedup_checksum(spa); 2374 zio_pop_transforms(zio); 2375 zio->io_stage = ZIO_STAGE_OPEN; 2376 BP_ZERO(bp); 2377 } else { 2378 zp->zp_dedup = B_FALSE; 2379 } 2380 zio->io_pipeline = ZIO_WRITE_PIPELINE; 2381 ddt_exit(ddt); 2382 return (ZIO_PIPELINE_CONTINUE); 2383 } 2384 2385 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp); 2386 ASSERT(ditto_copies < SPA_DVAS_PER_BP); 2387 2388 if (ditto_copies > ddt_ditto_copies_present(dde) && 2389 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) { 2390 zio_prop_t czp = *zp; 2391 2392 czp.zp_copies = ditto_copies; 2393 2394 /* 2395 * If we arrived here with an override bp, we won't have run 2396 * the transform stack, so we won't have the data we need to 2397 * generate a child i/o. So, toss the override bp and restart. 2398 * This is safe, because using the override bp is just an 2399 * optimization; and it's rare, so the cost doesn't matter. 2400 */ 2401 if (zio->io_bp_override) { 2402 zio_pop_transforms(zio); 2403 zio->io_stage = ZIO_STAGE_OPEN; 2404 zio->io_pipeline = ZIO_WRITE_PIPELINE; 2405 zio->io_bp_override = NULL; 2406 BP_ZERO(bp); 2407 ddt_exit(ddt); 2408 return (ZIO_PIPELINE_CONTINUE); 2409 } 2410 2411 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data, 2412 zio->io_orig_size, &czp, NULL, NULL, 2413 zio_ddt_ditto_write_done, dde, zio->io_priority, 2414 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); 2415 2416 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL); 2417 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio; 2418 } 2419 2420 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) { 2421 if (ddp->ddp_phys_birth != 0) 2422 ddt_bp_fill(ddp, bp, txg); 2423 if (dde->dde_lead_zio[p] != NULL) 2424 zio_add_child(zio, dde->dde_lead_zio[p]); 2425 else 2426 ddt_phys_addref(ddp); 2427 } else if (zio->io_bp_override) { 2428 ASSERT(bp->blk_birth == txg); 2429 ASSERT(BP_EQUAL(bp, zio->io_bp_override)); 2430 ddt_phys_fill(ddp, bp); 2431 ddt_phys_addref(ddp); 2432 } else { 2433 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data, 2434 zio->io_orig_size, zp, zio_ddt_child_write_ready, NULL, 2435 zio_ddt_child_write_done, dde, zio->io_priority, 2436 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); 2437 2438 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL); 2439 dde->dde_lead_zio[p] = cio; 2440 } 2441 2442 ddt_exit(ddt); 2443 2444 if (cio) 2445 zio_nowait(cio); 2446 if (dio) 2447 zio_nowait(dio); 2448 2449 return (ZIO_PIPELINE_CONTINUE); 2450 } 2451 2452 ddt_entry_t *freedde; /* for debugging */ 2453 2454 static int 2455 zio_ddt_free(zio_t *zio) 2456 { 2457 spa_t *spa = zio->io_spa; 2458 blkptr_t *bp = zio->io_bp; 2459 ddt_t *ddt = ddt_select(spa, bp); 2460 ddt_entry_t *dde; 2461 ddt_phys_t *ddp; 2462 2463 ASSERT(BP_GET_DEDUP(bp)); 2464 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2465 2466 ddt_enter(ddt); 2467 freedde = dde = ddt_lookup(ddt, bp, B_TRUE); 2468 ddp = ddt_phys_select(dde, bp); 2469 ddt_phys_decref(ddp); 2470 ddt_exit(ddt); 2471 2472 return (ZIO_PIPELINE_CONTINUE); 2473 } 2474 2475 /* 2476 * ========================================================================== 2477 * Allocate and free blocks 2478 * ========================================================================== 2479 */ 2480 static int 2481 zio_dva_allocate(zio_t *zio) 2482 { 2483 spa_t *spa = zio->io_spa; 2484 metaslab_class_t *mc = spa_normal_class(spa); 2485 blkptr_t *bp = zio->io_bp; 2486 int error; 2487 int flags = 0; 2488 2489 if (zio->io_gang_leader == NULL) { 2490 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 2491 zio->io_gang_leader = zio; 2492 } 2493 2494 ASSERT(BP_IS_HOLE(bp)); 2495 ASSERT0(BP_GET_NDVAS(bp)); 2496 ASSERT3U(zio->io_prop.zp_copies, >, 0); 2497 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa)); 2498 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp)); 2499 2500 /* 2501 * The dump device does not support gang blocks so allocation on 2502 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid 2503 * the "fast" gang feature. 2504 */ 2505 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0; 2506 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ? 2507 METASLAB_GANG_CHILD : 0; 2508 error = metaslab_alloc(spa, mc, zio->io_size, bp, 2509 zio->io_prop.zp_copies, zio->io_txg, NULL, flags); 2510 2511 if (error) { 2512 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, " 2513 "size %llu, error %d", spa_name(spa), zio, zio->io_size, 2514 error); 2515 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE) 2516 return (zio_write_gang_block(zio)); 2517 zio->io_error = error; 2518 } 2519 2520 return (ZIO_PIPELINE_CONTINUE); 2521 } 2522 2523 static int 2524 zio_dva_free(zio_t *zio) 2525 { 2526 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE); 2527 2528 return (ZIO_PIPELINE_CONTINUE); 2529 } 2530 2531 static int 2532 zio_dva_claim(zio_t *zio) 2533 { 2534 int error; 2535 2536 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg); 2537 if (error) 2538 zio->io_error = error; 2539 2540 return (ZIO_PIPELINE_CONTINUE); 2541 } 2542 2543 /* 2544 * Undo an allocation. This is used by zio_done() when an I/O fails 2545 * and we want to give back the block we just allocated. 2546 * This handles both normal blocks and gang blocks. 2547 */ 2548 static void 2549 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp) 2550 { 2551 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp)); 2552 ASSERT(zio->io_bp_override == NULL); 2553 2554 if (!BP_IS_HOLE(bp)) 2555 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE); 2556 2557 if (gn != NULL) { 2558 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 2559 zio_dva_unallocate(zio, gn->gn_child[g], 2560 &gn->gn_gbh->zg_blkptr[g]); 2561 } 2562 } 2563 } 2564 2565 /* 2566 * Try to allocate an intent log block. Return 0 on success, errno on failure. 2567 */ 2568 int 2569 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp, 2570 uint64_t size, boolean_t use_slog) 2571 { 2572 int error = 1; 2573 2574 ASSERT(txg > spa_syncing_txg(spa)); 2575 2576 /* 2577 * ZIL blocks are always contiguous (i.e. not gang blocks) so we 2578 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang" 2579 * when allocating them. 2580 */ 2581 if (use_slog) { 2582 error = metaslab_alloc(spa, spa_log_class(spa), size, 2583 new_bp, 1, txg, old_bp, 2584 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID); 2585 } 2586 2587 if (error) { 2588 error = metaslab_alloc(spa, spa_normal_class(spa), size, 2589 new_bp, 1, txg, old_bp, 2590 METASLAB_HINTBP_AVOID); 2591 } 2592 2593 if (error == 0) { 2594 BP_SET_LSIZE(new_bp, size); 2595 BP_SET_PSIZE(new_bp, size); 2596 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF); 2597 BP_SET_CHECKSUM(new_bp, 2598 spa_version(spa) >= SPA_VERSION_SLIM_ZIL 2599 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG); 2600 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); 2601 BP_SET_LEVEL(new_bp, 0); 2602 BP_SET_DEDUP(new_bp, 0); 2603 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER); 2604 } 2605 2606 return (error); 2607 } 2608 2609 /* 2610 * Free an intent log block. 2611 */ 2612 void 2613 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp) 2614 { 2615 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG); 2616 ASSERT(!BP_IS_GANG(bp)); 2617 2618 zio_free(spa, txg, bp); 2619 } 2620 2621 /* 2622 * ========================================================================== 2623 * Read and write to physical devices 2624 * ========================================================================== 2625 */ 2626 2627 2628 /* 2629 * Issue an I/O to the underlying vdev. Typically the issue pipeline 2630 * stops after this stage and will resume upon I/O completion. 2631 * However, there are instances where the vdev layer may need to 2632 * continue the pipeline when an I/O was not issued. Since the I/O 2633 * that was sent to the vdev layer might be different than the one 2634 * currently active in the pipeline (see vdev_queue_io()), we explicitly 2635 * force the underlying vdev layers to call either zio_execute() or 2636 * zio_interrupt() to ensure that the pipeline continues with the correct I/O. 2637 */ 2638 static int 2639 zio_vdev_io_start(zio_t *zio) 2640 { 2641 vdev_t *vd = zio->io_vd; 2642 uint64_t align; 2643 spa_t *spa = zio->io_spa; 2644 2645 ASSERT(zio->io_error == 0); 2646 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0); 2647 2648 if (vd == NULL) { 2649 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) 2650 spa_config_enter(spa, SCL_ZIO, zio, RW_READER); 2651 2652 /* 2653 * The mirror_ops handle multiple DVAs in a single BP. 2654 */ 2655 vdev_mirror_ops.vdev_op_io_start(zio); 2656 return (ZIO_PIPELINE_STOP); 2657 } 2658 2659 /* 2660 * We keep track of time-sensitive I/Os so that the scan thread 2661 * can quickly react to certain workloads. In particular, we care 2662 * about non-scrubbing, top-level reads and writes with the following 2663 * characteristics: 2664 * - synchronous writes of user data to non-slog devices 2665 * - any reads of user data 2666 * When these conditions are met, adjust the timestamp of spa_last_io 2667 * which allows the scan thread to adjust its workload accordingly. 2668 */ 2669 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL && 2670 vd == vd->vdev_top && !vd->vdev_islog && 2671 zio->io_bookmark.zb_objset != DMU_META_OBJSET && 2672 zio->io_txg != spa_syncing_txg(spa)) { 2673 uint64_t old = spa->spa_last_io; 2674 uint64_t new = ddi_get_lbolt64(); 2675 if (old != new) 2676 (void) atomic_cas_64(&spa->spa_last_io, old, new); 2677 } 2678 2679 align = 1ULL << vd->vdev_top->vdev_ashift; 2680 2681 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) && 2682 P2PHASE(zio->io_size, align) != 0) { 2683 /* Transform logical writes to be a full physical block size. */ 2684 uint64_t asize = P2ROUNDUP(zio->io_size, align); 2685 char *abuf = zio_buf_alloc(asize); 2686 ASSERT(vd == vd->vdev_top); 2687 if (zio->io_type == ZIO_TYPE_WRITE) { 2688 bcopy(zio->io_data, abuf, zio->io_size); 2689 bzero(abuf + zio->io_size, asize - zio->io_size); 2690 } 2691 zio_push_transform(zio, abuf, asize, asize, zio_subblock); 2692 } 2693 2694 /* 2695 * If this is not a physical io, make sure that it is properly aligned 2696 * before proceeding. 2697 */ 2698 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) { 2699 ASSERT0(P2PHASE(zio->io_offset, align)); 2700 ASSERT0(P2PHASE(zio->io_size, align)); 2701 } else { 2702 /* 2703 * For physical writes, we allow 512b aligned writes and assume 2704 * the device will perform a read-modify-write as necessary. 2705 */ 2706 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE)); 2707 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE)); 2708 } 2709 2710 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa)); 2711 2712 /* 2713 * If this is a repair I/O, and there's no self-healing involved -- 2714 * that is, we're just resilvering what we expect to resilver -- 2715 * then don't do the I/O unless zio's txg is actually in vd's DTL. 2716 * This prevents spurious resilvering with nested replication. 2717 * For example, given a mirror of mirrors, (A+B)+(C+D), if only 2718 * A is out of date, we'll read from C+D, then use the data to 2719 * resilver A+B -- but we don't actually want to resilver B, just A. 2720 * The top-level mirror has no way to know this, so instead we just 2721 * discard unnecessary repairs as we work our way down the vdev tree. 2722 * The same logic applies to any form of nested replication: 2723 * ditto + mirror, RAID-Z + replacing, etc. This covers them all. 2724 */ 2725 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) && 2726 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) && 2727 zio->io_txg != 0 && /* not a delegated i/o */ 2728 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) { 2729 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 2730 zio_vdev_io_bypass(zio); 2731 return (ZIO_PIPELINE_CONTINUE); 2732 } 2733 2734 if (vd->vdev_ops->vdev_op_leaf && 2735 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) { 2736 2737 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio)) 2738 return (ZIO_PIPELINE_CONTINUE); 2739 2740 if ((zio = vdev_queue_io(zio)) == NULL) 2741 return (ZIO_PIPELINE_STOP); 2742 2743 if (!vdev_accessible(vd, zio)) { 2744 zio->io_error = SET_ERROR(ENXIO); 2745 zio_interrupt(zio); 2746 return (ZIO_PIPELINE_STOP); 2747 } 2748 } 2749 2750 vd->vdev_ops->vdev_op_io_start(zio); 2751 return (ZIO_PIPELINE_STOP); 2752 } 2753 2754 static int 2755 zio_vdev_io_done(zio_t *zio) 2756 { 2757 vdev_t *vd = zio->io_vd; 2758 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops; 2759 boolean_t unexpected_error = B_FALSE; 2760 2761 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE)) 2762 return (ZIO_PIPELINE_STOP); 2763 2764 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE); 2765 2766 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) { 2767 2768 vdev_queue_io_done(zio); 2769 2770 if (zio->io_type == ZIO_TYPE_WRITE) 2771 vdev_cache_write(zio); 2772 2773 if (zio_injection_enabled && zio->io_error == 0) 2774 zio->io_error = zio_handle_device_injection(vd, 2775 zio, EIO); 2776 2777 if (zio_injection_enabled && zio->io_error == 0) 2778 zio->io_error = zio_handle_label_injection(zio, EIO); 2779 2780 if (zio->io_error) { 2781 if (!vdev_accessible(vd, zio)) { 2782 zio->io_error = SET_ERROR(ENXIO); 2783 } else { 2784 unexpected_error = B_TRUE; 2785 } 2786 } 2787 } 2788 2789 ops->vdev_op_io_done(zio); 2790 2791 if (unexpected_error) 2792 VERIFY(vdev_probe(vd, zio) == NULL); 2793 2794 return (ZIO_PIPELINE_CONTINUE); 2795 } 2796 2797 /* 2798 * For non-raidz ZIOs, we can just copy aside the bad data read from the 2799 * disk, and use that to finish the checksum ereport later. 2800 */ 2801 static void 2802 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr, 2803 const void *good_buf) 2804 { 2805 /* no processing needed */ 2806 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE); 2807 } 2808 2809 /*ARGSUSED*/ 2810 void 2811 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored) 2812 { 2813 void *buf = zio_buf_alloc(zio->io_size); 2814 2815 bcopy(zio->io_data, buf, zio->io_size); 2816 2817 zcr->zcr_cbinfo = zio->io_size; 2818 zcr->zcr_cbdata = buf; 2819 zcr->zcr_finish = zio_vsd_default_cksum_finish; 2820 zcr->zcr_free = zio_buf_free; 2821 } 2822 2823 static int 2824 zio_vdev_io_assess(zio_t *zio) 2825 { 2826 vdev_t *vd = zio->io_vd; 2827 2828 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE)) 2829 return (ZIO_PIPELINE_STOP); 2830 2831 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) 2832 spa_config_exit(zio->io_spa, SCL_ZIO, zio); 2833 2834 if (zio->io_vsd != NULL) { 2835 zio->io_vsd_ops->vsd_free(zio); 2836 zio->io_vsd = NULL; 2837 } 2838 2839 if (zio_injection_enabled && zio->io_error == 0) 2840 zio->io_error = zio_handle_fault_injection(zio, EIO); 2841 2842 /* 2843 * If the I/O failed, determine whether we should attempt to retry it. 2844 * 2845 * On retry, we cut in line in the issue queue, since we don't want 2846 * compression/checksumming/etc. work to prevent our (cheap) IO reissue. 2847 */ 2848 if (zio->io_error && vd == NULL && 2849 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) { 2850 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */ 2851 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */ 2852 zio->io_error = 0; 2853 zio->io_flags |= ZIO_FLAG_IO_RETRY | 2854 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE; 2855 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1; 2856 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, 2857 zio_requeue_io_start_cut_in_line); 2858 return (ZIO_PIPELINE_STOP); 2859 } 2860 2861 /* 2862 * If we got an error on a leaf device, convert it to ENXIO 2863 * if the device is not accessible at all. 2864 */ 2865 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf && 2866 !vdev_accessible(vd, zio)) 2867 zio->io_error = SET_ERROR(ENXIO); 2868 2869 /* 2870 * If we can't write to an interior vdev (mirror or RAID-Z), 2871 * set vdev_cant_write so that we stop trying to allocate from it. 2872 */ 2873 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE && 2874 vd != NULL && !vd->vdev_ops->vdev_op_leaf) { 2875 vd->vdev_cant_write = B_TRUE; 2876 } 2877 2878 if (zio->io_error) 2879 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2880 2881 if (vd != NULL && vd->vdev_ops->vdev_op_leaf && 2882 zio->io_physdone != NULL) { 2883 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED)); 2884 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV); 2885 zio->io_physdone(zio->io_logical); 2886 } 2887 2888 return (ZIO_PIPELINE_CONTINUE); 2889 } 2890 2891 void 2892 zio_vdev_io_reissue(zio_t *zio) 2893 { 2894 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); 2895 ASSERT(zio->io_error == 0); 2896 2897 zio->io_stage >>= 1; 2898 } 2899 2900 void 2901 zio_vdev_io_redone(zio_t *zio) 2902 { 2903 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE); 2904 2905 zio->io_stage >>= 1; 2906 } 2907 2908 void 2909 zio_vdev_io_bypass(zio_t *zio) 2910 { 2911 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); 2912 ASSERT(zio->io_error == 0); 2913 2914 zio->io_flags |= ZIO_FLAG_IO_BYPASS; 2915 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1; 2916 } 2917 2918 /* 2919 * ========================================================================== 2920 * Generate and verify checksums 2921 * ========================================================================== 2922 */ 2923 static int 2924 zio_checksum_generate(zio_t *zio) 2925 { 2926 blkptr_t *bp = zio->io_bp; 2927 enum zio_checksum checksum; 2928 2929 if (bp == NULL) { 2930 /* 2931 * This is zio_write_phys(). 2932 * We're either generating a label checksum, or none at all. 2933 */ 2934 checksum = zio->io_prop.zp_checksum; 2935 2936 if (checksum == ZIO_CHECKSUM_OFF) 2937 return (ZIO_PIPELINE_CONTINUE); 2938 2939 ASSERT(checksum == ZIO_CHECKSUM_LABEL); 2940 } else { 2941 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) { 2942 ASSERT(!IO_IS_ALLOCATING(zio)); 2943 checksum = ZIO_CHECKSUM_GANG_HEADER; 2944 } else { 2945 checksum = BP_GET_CHECKSUM(bp); 2946 } 2947 } 2948 2949 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size); 2950 2951 return (ZIO_PIPELINE_CONTINUE); 2952 } 2953 2954 static int 2955 zio_checksum_verify(zio_t *zio) 2956 { 2957 zio_bad_cksum_t info; 2958 blkptr_t *bp = zio->io_bp; 2959 int error; 2960 2961 ASSERT(zio->io_vd != NULL); 2962 2963 if (bp == NULL) { 2964 /* 2965 * This is zio_read_phys(). 2966 * We're either verifying a label checksum, or nothing at all. 2967 */ 2968 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF) 2969 return (ZIO_PIPELINE_CONTINUE); 2970 2971 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL); 2972 } 2973 2974 if ((error = zio_checksum_error(zio, &info)) != 0) { 2975 zio->io_error = error; 2976 if (error == ECKSUM && 2977 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) { 2978 zfs_ereport_start_checksum(zio->io_spa, 2979 zio->io_vd, zio, zio->io_offset, 2980 zio->io_size, NULL, &info); 2981 } 2982 } 2983 2984 return (ZIO_PIPELINE_CONTINUE); 2985 } 2986 2987 /* 2988 * Called by RAID-Z to ensure we don't compute the checksum twice. 2989 */ 2990 void 2991 zio_checksum_verified(zio_t *zio) 2992 { 2993 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; 2994 } 2995 2996 /* 2997 * ========================================================================== 2998 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other. 2999 * An error of 0 indicates success. ENXIO indicates whole-device failure, 3000 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO 3001 * indicate errors that are specific to one I/O, and most likely permanent. 3002 * Any other error is presumed to be worse because we weren't expecting it. 3003 * ========================================================================== 3004 */ 3005 int 3006 zio_worst_error(int e1, int e2) 3007 { 3008 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO }; 3009 int r1, r2; 3010 3011 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++) 3012 if (e1 == zio_error_rank[r1]) 3013 break; 3014 3015 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++) 3016 if (e2 == zio_error_rank[r2]) 3017 break; 3018 3019 return (r1 > r2 ? e1 : e2); 3020 } 3021 3022 /* 3023 * ========================================================================== 3024 * I/O completion 3025 * ========================================================================== 3026 */ 3027 static int 3028 zio_ready(zio_t *zio) 3029 { 3030 blkptr_t *bp = zio->io_bp; 3031 zio_t *pio, *pio_next; 3032 3033 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) || 3034 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY)) 3035 return (ZIO_PIPELINE_STOP); 3036 3037 if (zio->io_ready) { 3038 ASSERT(IO_IS_ALLOCATING(zio)); 3039 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) || 3040 (zio->io_flags & ZIO_FLAG_NOPWRITE)); 3041 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0); 3042 3043 zio->io_ready(zio); 3044 } 3045 3046 if (bp != NULL && bp != &zio->io_bp_copy) 3047 zio->io_bp_copy = *bp; 3048 3049 if (zio->io_error) 3050 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 3051 3052 mutex_enter(&zio->io_lock); 3053 zio->io_state[ZIO_WAIT_READY] = 1; 3054 pio = zio_walk_parents(zio); 3055 mutex_exit(&zio->io_lock); 3056 3057 /* 3058 * As we notify zio's parents, new parents could be added. 3059 * New parents go to the head of zio's io_parent_list, however, 3060 * so we will (correctly) not notify them. The remainder of zio's 3061 * io_parent_list, from 'pio_next' onward, cannot change because 3062 * all parents must wait for us to be done before they can be done. 3063 */ 3064 for (; pio != NULL; pio = pio_next) { 3065 pio_next = zio_walk_parents(zio); 3066 zio_notify_parent(pio, zio, ZIO_WAIT_READY); 3067 } 3068 3069 if (zio->io_flags & ZIO_FLAG_NODATA) { 3070 if (BP_IS_GANG(bp)) { 3071 zio->io_flags &= ~ZIO_FLAG_NODATA; 3072 } else { 3073 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE); 3074 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; 3075 } 3076 } 3077 3078 if (zio_injection_enabled && 3079 zio->io_spa->spa_syncing_txg == zio->io_txg) 3080 zio_handle_ignored_writes(zio); 3081 3082 return (ZIO_PIPELINE_CONTINUE); 3083 } 3084 3085 static int 3086 zio_done(zio_t *zio) 3087 { 3088 spa_t *spa = zio->io_spa; 3089 zio_t *lio = zio->io_logical; 3090 blkptr_t *bp = zio->io_bp; 3091 vdev_t *vd = zio->io_vd; 3092 uint64_t psize = zio->io_size; 3093 zio_t *pio, *pio_next; 3094 3095 /* 3096 * If our children haven't all completed, 3097 * wait for them and then repeat this pipeline stage. 3098 */ 3099 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) || 3100 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) || 3101 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) || 3102 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE)) 3103 return (ZIO_PIPELINE_STOP); 3104 3105 for (int c = 0; c < ZIO_CHILD_TYPES; c++) 3106 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 3107 ASSERT(zio->io_children[c][w] == 0); 3108 3109 if (bp != NULL && !BP_IS_EMBEDDED(bp)) { 3110 ASSERT(bp->blk_pad[0] == 0); 3111 ASSERT(bp->blk_pad[1] == 0); 3112 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 || 3113 (bp == zio_unique_parent(zio)->io_bp)); 3114 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) && 3115 zio->io_bp_override == NULL && 3116 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) { 3117 ASSERT(!BP_SHOULD_BYTESWAP(bp)); 3118 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp)); 3119 ASSERT(BP_COUNT_GANG(bp) == 0 || 3120 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp))); 3121 } 3122 if (zio->io_flags & ZIO_FLAG_NOPWRITE) 3123 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig)); 3124 } 3125 3126 /* 3127 * If there were child vdev/gang/ddt errors, they apply to us now. 3128 */ 3129 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV); 3130 zio_inherit_child_errors(zio, ZIO_CHILD_GANG); 3131 zio_inherit_child_errors(zio, ZIO_CHILD_DDT); 3132 3133 /* 3134 * If the I/O on the transformed data was successful, generate any 3135 * checksum reports now while we still have the transformed data. 3136 */ 3137 if (zio->io_error == 0) { 3138 while (zio->io_cksum_report != NULL) { 3139 zio_cksum_report_t *zcr = zio->io_cksum_report; 3140 uint64_t align = zcr->zcr_align; 3141 uint64_t asize = P2ROUNDUP(psize, align); 3142 char *abuf = zio->io_data; 3143 3144 if (asize != psize) { 3145 abuf = zio_buf_alloc(asize); 3146 bcopy(zio->io_data, abuf, psize); 3147 bzero(abuf + psize, asize - psize); 3148 } 3149 3150 zio->io_cksum_report = zcr->zcr_next; 3151 zcr->zcr_next = NULL; 3152 zcr->zcr_finish(zcr, abuf); 3153 zfs_ereport_free_checksum(zcr); 3154 3155 if (asize != psize) 3156 zio_buf_free(abuf, asize); 3157 } 3158 } 3159 3160 zio_pop_transforms(zio); /* note: may set zio->io_error */ 3161 3162 vdev_stat_update(zio, psize); 3163 3164 if (zio->io_error) { 3165 /* 3166 * If this I/O is attached to a particular vdev, 3167 * generate an error message describing the I/O failure 3168 * at the block level. We ignore these errors if the 3169 * device is currently unavailable. 3170 */ 3171 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd)) 3172 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0); 3173 3174 if ((zio->io_error == EIO || !(zio->io_flags & 3175 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) && 3176 zio == lio) { 3177 /* 3178 * For logical I/O requests, tell the SPA to log the 3179 * error and generate a logical data ereport. 3180 */ 3181 spa_log_error(spa, zio); 3182 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio, 3183 0, 0); 3184 } 3185 } 3186 3187 if (zio->io_error && zio == lio) { 3188 /* 3189 * Determine whether zio should be reexecuted. This will 3190 * propagate all the way to the root via zio_notify_parent(). 3191 */ 3192 ASSERT(vd == NULL && bp != NULL); 3193 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 3194 3195 if (IO_IS_ALLOCATING(zio) && 3196 !(zio->io_flags & ZIO_FLAG_CANFAIL)) { 3197 if (zio->io_error != ENOSPC) 3198 zio->io_reexecute |= ZIO_REEXECUTE_NOW; 3199 else 3200 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 3201 } 3202 3203 if ((zio->io_type == ZIO_TYPE_READ || 3204 zio->io_type == ZIO_TYPE_FREE) && 3205 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && 3206 zio->io_error == ENXIO && 3207 spa_load_state(spa) == SPA_LOAD_NONE && 3208 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE) 3209 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 3210 3211 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute) 3212 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 3213 3214 /* 3215 * Here is a possibly good place to attempt to do 3216 * either combinatorial reconstruction or error correction 3217 * based on checksums. It also might be a good place 3218 * to send out preliminary ereports before we suspend 3219 * processing. 3220 */ 3221 } 3222 3223 /* 3224 * If there were logical child errors, they apply to us now. 3225 * We defer this until now to avoid conflating logical child 3226 * errors with errors that happened to the zio itself when 3227 * updating vdev stats and reporting FMA events above. 3228 */ 3229 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL); 3230 3231 if ((zio->io_error || zio->io_reexecute) && 3232 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio && 3233 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE))) 3234 zio_dva_unallocate(zio, zio->io_gang_tree, bp); 3235 3236 zio_gang_tree_free(&zio->io_gang_tree); 3237 3238 /* 3239 * Godfather I/Os should never suspend. 3240 */ 3241 if ((zio->io_flags & ZIO_FLAG_GODFATHER) && 3242 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) 3243 zio->io_reexecute = 0; 3244 3245 if (zio->io_reexecute) { 3246 /* 3247 * This is a logical I/O that wants to reexecute. 3248 * 3249 * Reexecute is top-down. When an i/o fails, if it's not 3250 * the root, it simply notifies its parent and sticks around. 3251 * The parent, seeing that it still has children in zio_done(), 3252 * does the same. This percolates all the way up to the root. 3253 * The root i/o will reexecute or suspend the entire tree. 3254 * 3255 * This approach ensures that zio_reexecute() honors 3256 * all the original i/o dependency relationships, e.g. 3257 * parents not executing until children are ready. 3258 */ 3259 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 3260 3261 zio->io_gang_leader = NULL; 3262 3263 mutex_enter(&zio->io_lock); 3264 zio->io_state[ZIO_WAIT_DONE] = 1; 3265 mutex_exit(&zio->io_lock); 3266 3267 /* 3268 * "The Godfather" I/O monitors its children but is 3269 * not a true parent to them. It will track them through 3270 * the pipeline but severs its ties whenever they get into 3271 * trouble (e.g. suspended). This allows "The Godfather" 3272 * I/O to return status without blocking. 3273 */ 3274 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) { 3275 zio_link_t *zl = zio->io_walk_link; 3276 pio_next = zio_walk_parents(zio); 3277 3278 if ((pio->io_flags & ZIO_FLAG_GODFATHER) && 3279 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) { 3280 zio_remove_child(pio, zio, zl); 3281 zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 3282 } 3283 } 3284 3285 if ((pio = zio_unique_parent(zio)) != NULL) { 3286 /* 3287 * We're not a root i/o, so there's nothing to do 3288 * but notify our parent. Don't propagate errors 3289 * upward since we haven't permanently failed yet. 3290 */ 3291 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); 3292 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE; 3293 zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 3294 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) { 3295 /* 3296 * We'd fail again if we reexecuted now, so suspend 3297 * until conditions improve (e.g. device comes online). 3298 */ 3299 zio_suspend(spa, zio); 3300 } else { 3301 /* 3302 * Reexecution is potentially a huge amount of work. 3303 * Hand it off to the otherwise-unused claim taskq. 3304 */ 3305 ASSERT(zio->io_tqent.tqent_next == NULL); 3306 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM, 3307 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio, 3308 0, &zio->io_tqent); 3309 } 3310 return (ZIO_PIPELINE_STOP); 3311 } 3312 3313 ASSERT(zio->io_child_count == 0); 3314 ASSERT(zio->io_reexecute == 0); 3315 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL)); 3316 3317 /* 3318 * Report any checksum errors, since the I/O is complete. 3319 */ 3320 while (zio->io_cksum_report != NULL) { 3321 zio_cksum_report_t *zcr = zio->io_cksum_report; 3322 zio->io_cksum_report = zcr->zcr_next; 3323 zcr->zcr_next = NULL; 3324 zcr->zcr_finish(zcr, NULL); 3325 zfs_ereport_free_checksum(zcr); 3326 } 3327 3328 /* 3329 * It is the responsibility of the done callback to ensure that this 3330 * particular zio is no longer discoverable for adoption, and as 3331 * such, cannot acquire any new parents. 3332 */ 3333 if (zio->io_done) 3334 zio->io_done(zio); 3335 3336 mutex_enter(&zio->io_lock); 3337 zio->io_state[ZIO_WAIT_DONE] = 1; 3338 mutex_exit(&zio->io_lock); 3339 3340 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) { 3341 zio_link_t *zl = zio->io_walk_link; 3342 pio_next = zio_walk_parents(zio); 3343 zio_remove_child(pio, zio, zl); 3344 zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 3345 } 3346 3347 if (zio->io_waiter != NULL) { 3348 mutex_enter(&zio->io_lock); 3349 zio->io_executor = NULL; 3350 cv_broadcast(&zio->io_cv); 3351 mutex_exit(&zio->io_lock); 3352 } else { 3353 zio_destroy(zio); 3354 } 3355 3356 return (ZIO_PIPELINE_STOP); 3357 } 3358 3359 /* 3360 * ========================================================================== 3361 * I/O pipeline definition 3362 * ========================================================================== 3363 */ 3364 static zio_pipe_stage_t *zio_pipeline[] = { 3365 NULL, 3366 zio_read_bp_init, 3367 zio_free_bp_init, 3368 zio_issue_async, 3369 zio_write_bp_init, 3370 zio_checksum_generate, 3371 zio_nop_write, 3372 zio_ddt_read_start, 3373 zio_ddt_read_done, 3374 zio_ddt_write, 3375 zio_ddt_free, 3376 zio_gang_assemble, 3377 zio_gang_issue, 3378 zio_dva_allocate, 3379 zio_dva_free, 3380 zio_dva_claim, 3381 zio_ready, 3382 zio_vdev_io_start, 3383 zio_vdev_io_done, 3384 zio_vdev_io_assess, 3385 zio_checksum_verify, 3386 zio_done 3387 }; 3388 3389 3390 3391 3392 /* 3393 * Compare two zbookmark_phys_t's to see which we would reach first in a 3394 * pre-order traversal of the object tree. 3395 * 3396 * This is simple in every case aside from the meta-dnode object. For all other 3397 * objects, we traverse them in order (object 1 before object 2, and so on). 3398 * However, all of these objects are traversed while traversing object 0, since 3399 * the data it points to is the list of objects. Thus, we need to convert to a 3400 * canonical representation so we can compare meta-dnode bookmarks to 3401 * non-meta-dnode bookmarks. 3402 * 3403 * We do this by calculating "equivalents" for each field of the zbookmark. 3404 * zbookmarks outside of the meta-dnode use their own object and level, and 3405 * calculate the level 0 equivalent (the first L0 blkid that is contained in the 3406 * blocks this bookmark refers to) by multiplying their blkid by their span 3407 * (the number of L0 blocks contained within one block at their level). 3408 * zbookmarks inside the meta-dnode calculate their object equivalent 3409 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use 3410 * level + 1<<31 (any value larger than a level could ever be) for their level. 3411 * This causes them to always compare before a bookmark in their object 3412 * equivalent, compare appropriately to bookmarks in other objects, and to 3413 * compare appropriately to other bookmarks in the meta-dnode. 3414 */ 3415 int 3416 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2, 3417 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2) 3418 { 3419 /* 3420 * These variables represent the "equivalent" values for the zbookmark, 3421 * after converting zbookmarks inside the meta dnode to their 3422 * normal-object equivalents. 3423 */ 3424 uint64_t zb1obj, zb2obj; 3425 uint64_t zb1L0, zb2L0; 3426 uint64_t zb1level, zb2level; 3427 3428 if (zb1->zb_object == zb2->zb_object && 3429 zb1->zb_level == zb2->zb_level && 3430 zb1->zb_blkid == zb2->zb_blkid) 3431 return (0); 3432 3433 /* 3434 * BP_SPANB calculates the span in blocks. 3435 */ 3436 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level); 3437 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level); 3438 3439 if (zb1->zb_object == DMU_META_DNODE_OBJECT) { 3440 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT)); 3441 zb1L0 = 0; 3442 zb1level = zb1->zb_level + COMPARE_META_LEVEL; 3443 } else { 3444 zb1obj = zb1->zb_object; 3445 zb1level = zb1->zb_level; 3446 } 3447 3448 if (zb2->zb_object == DMU_META_DNODE_OBJECT) { 3449 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT)); 3450 zb2L0 = 0; 3451 zb2level = zb2->zb_level + COMPARE_META_LEVEL; 3452 } else { 3453 zb2obj = zb2->zb_object; 3454 zb2level = zb2->zb_level; 3455 } 3456 3457 /* Now that we have a canonical representation, do the comparison. */ 3458 if (zb1obj != zb2obj) 3459 return (zb1obj < zb2obj ? -1 : 1); 3460 else if (zb1L0 != zb2L0) 3461 return (zb1L0 < zb2L0 ? -1 : 1); 3462 else if (zb1level != zb2level) 3463 return (zb1level > zb2level ? -1 : 1); 3464 /* 3465 * This can (theoretically) happen if the bookmarks have the same object 3466 * and level, but different blkids, if the block sizes are not the same. 3467 * There is presently no way to change the indirect block sizes 3468 */ 3469 return (0); 3470 } 3471 3472 /* 3473 * This function checks the following: given that last_block is the place that 3474 * our traversal stopped last time, does that guarantee that we've visited 3475 * every node under subtree_root? Therefore, we can't just use the raw output 3476 * of zbookmark_compare. We have to pass in a modified version of 3477 * subtree_root; by incrementing the block id, and then checking whether 3478 * last_block is before or equal to that, we can tell whether or not having 3479 * visited last_block implies that all of subtree_root's children have been 3480 * visited. 3481 */ 3482 boolean_t 3483 zbookmark_subtree_completed(const dnode_phys_t *dnp, 3484 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block) 3485 { 3486 zbookmark_phys_t mod_zb = *subtree_root; 3487 mod_zb.zb_blkid++; 3488 ASSERT(last_block->zb_level == 0); 3489 3490 /* The objset_phys_t isn't before anything. */ 3491 if (dnp == NULL) 3492 return (B_FALSE); 3493 3494 /* 3495 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the 3496 * data block size in sectors, because that variable is only used if 3497 * the bookmark refers to a block in the meta-dnode. Since we don't 3498 * know without examining it what object it refers to, and there's no 3499 * harm in passing in this value in other cases, we always pass it in. 3500 * 3501 * We pass in 0 for the indirect block size shift because zb2 must be 3502 * level 0. The indirect block size is only used to calculate the span 3503 * of the bookmark, but since the bookmark must be level 0, the span is 3504 * always 1, so the math works out. 3505 * 3506 * If you make changes to how the zbookmark_compare code works, be sure 3507 * to make sure that this code still works afterwards. 3508 */ 3509 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift, 3510 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb, 3511 last_block) <= 0); 3512 } 3513