1 // SPDX-License-Identifier: CDDL-1.0 2 /* 3 * CDDL HEADER START 4 * 5 * The contents of this file are subject to the terms of the 6 * Common Development and Distribution License (the "License"). 7 * You may not use this file except in compliance with the License. 8 * 9 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 10 * or https://opensource.org/licenses/CDDL-1.0. 11 * See the License for the specific language governing permissions 12 * and limitations under the License. 13 * 14 * When distributing Covered Code, include this CDDL HEADER in each 15 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 16 * If applicable, add the following below this CDDL HEADER, with the 17 * fields enclosed by brackets "[]" replaced with your own identifying 18 * information: Portions Copyright [yyyy] [name of copyright owner] 19 * 20 * CDDL HEADER END 21 */ 22 /* 23 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 24 * Copyright (c) 2011, 2022 by Delphix. All rights reserved. 25 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved. 26 * Copyright (c) 2017, Intel Corporation. 27 * Copyright (c) 2019, 2023, 2024, 2025, Klara, Inc. 28 * Copyright (c) 2019, Allan Jude 29 * Copyright (c) 2021, Datto, Inc. 30 * Copyright (c) 2021, 2024 by George Melikov. All rights reserved. 31 */ 32 33 #include <sys/sysmacros.h> 34 #include <sys/zfs_context.h> 35 #include <sys/fm/fs/zfs.h> 36 #include <sys/spa.h> 37 #include <sys/txg.h> 38 #include <sys/spa_impl.h> 39 #include <sys/vdev_impl.h> 40 #include <sys/vdev_trim.h> 41 #include <sys/zio_impl.h> 42 #include <sys/zio_compress.h> 43 #include <sys/zio_checksum.h> 44 #include <sys/dmu_objset.h> 45 #include <sys/arc.h> 46 #include <sys/brt.h> 47 #include <sys/ddt.h> 48 #include <sys/blkptr.h> 49 #include <sys/zfeature.h> 50 #include <sys/dsl_scan.h> 51 #include <sys/metaslab_impl.h> 52 #include <sys/time.h> 53 #include <sys/trace_zfs.h> 54 #include <sys/abd.h> 55 #include <sys/dsl_crypt.h> 56 #include <cityhash.h> 57 58 /* 59 * ========================================================================== 60 * I/O type descriptions 61 * ========================================================================== 62 */ 63 const char *const zio_type_name[ZIO_TYPES] = { 64 /* 65 * Note: Linux kernel thread name length is limited 66 * so these names will differ from upstream open zfs. 67 */ 68 "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_flush", "z_trim" 69 }; 70 71 int zio_dva_throttle_enabled = B_TRUE; 72 static int zio_deadman_log_all = B_FALSE; 73 74 /* 75 * ========================================================================== 76 * I/O kmem caches 77 * ========================================================================== 78 */ 79 static kmem_cache_t *zio_cache; 80 static kmem_cache_t *zio_link_cache; 81 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; 82 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; 83 #if defined(ZFS_DEBUG) && !defined(_KERNEL) 84 static uint64_t zio_buf_cache_allocs[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; 85 static uint64_t zio_buf_cache_frees[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; 86 #endif 87 88 /* Mark IOs as "slow" if they take longer than 30 seconds */ 89 static uint_t zio_slow_io_ms = (30 * MILLISEC); 90 91 #define BP_SPANB(indblkshift, level) \ 92 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT))) 93 #define COMPARE_META_LEVEL 0x80000000ul 94 /* 95 * The following actions directly effect the spa's sync-to-convergence logic. 96 * The values below define the sync pass when we start performing the action. 97 * Care should be taken when changing these values as they directly impact 98 * spa_sync() performance. Tuning these values may introduce subtle performance 99 * pathologies and should only be done in the context of performance analysis. 100 * These tunables will eventually be removed and replaced with #defines once 101 * enough analysis has been done to determine optimal values. 102 * 103 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that 104 * regular blocks are not deferred. 105 * 106 * Starting in sync pass 8 (zfs_sync_pass_dont_compress), we disable 107 * compression (including of metadata). In practice, we don't have this 108 * many sync passes, so this has no effect. 109 * 110 * The original intent was that disabling compression would help the sync 111 * passes to converge. However, in practice disabling compression increases 112 * the average number of sync passes, because when we turn compression off, a 113 * lot of block's size will change and thus we have to re-allocate (not 114 * overwrite) them. It also increases the number of 128KB allocations (e.g. 115 * for indirect blocks and spacemaps) because these will not be compressed. 116 * The 128K allocations are especially detrimental to performance on highly 117 * fragmented systems, which may have very few free segments of this size, 118 * and may need to load new metaslabs to satisfy 128K allocations. 119 */ 120 121 /* defer frees starting in this pass */ 122 uint_t zfs_sync_pass_deferred_free = 2; 123 124 /* don't compress starting in this pass */ 125 static uint_t zfs_sync_pass_dont_compress = 8; 126 127 /* rewrite new bps starting in this pass */ 128 static uint_t zfs_sync_pass_rewrite = 2; 129 130 /* 131 * An allocating zio is one that either currently has the DVA allocate 132 * stage set or will have it later in its lifetime. 133 */ 134 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE) 135 136 /* 137 * Enable smaller cores by excluding metadata 138 * allocations as well. 139 */ 140 int zio_exclude_metadata = 0; 141 static int zio_requeue_io_start_cut_in_line = 1; 142 143 #ifdef ZFS_DEBUG 144 static const int zio_buf_debug_limit = 16384; 145 #else 146 static const int zio_buf_debug_limit = 0; 147 #endif 148 149 typedef struct zio_stats { 150 kstat_named_t ziostat_total_allocations; 151 kstat_named_t ziostat_alloc_class_fallbacks; 152 kstat_named_t ziostat_gang_writes; 153 kstat_named_t ziostat_gang_multilevel; 154 } zio_stats_t; 155 156 static zio_stats_t zio_stats = { 157 { "total_allocations", KSTAT_DATA_UINT64 }, 158 { "alloc_class_fallbacks", KSTAT_DATA_UINT64 }, 159 { "gang_writes", KSTAT_DATA_UINT64 }, 160 { "gang_multilevel", KSTAT_DATA_UINT64 }, 161 }; 162 163 struct { 164 wmsum_t ziostat_total_allocations; 165 wmsum_t ziostat_alloc_class_fallbacks; 166 wmsum_t ziostat_gang_writes; 167 wmsum_t ziostat_gang_multilevel; 168 } ziostat_sums; 169 170 #define ZIOSTAT_BUMP(stat) wmsum_add(&ziostat_sums.stat, 1); 171 172 static kstat_t *zio_ksp; 173 174 static inline void __zio_execute(zio_t *zio); 175 176 static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t); 177 178 static int 179 zio_kstats_update(kstat_t *ksp, int rw) 180 { 181 zio_stats_t *zs = ksp->ks_data; 182 if (rw == KSTAT_WRITE) 183 return (EACCES); 184 185 zs->ziostat_total_allocations.value.ui64 = 186 wmsum_value(&ziostat_sums.ziostat_total_allocations); 187 zs->ziostat_alloc_class_fallbacks.value.ui64 = 188 wmsum_value(&ziostat_sums.ziostat_alloc_class_fallbacks); 189 zs->ziostat_gang_writes.value.ui64 = 190 wmsum_value(&ziostat_sums.ziostat_gang_writes); 191 zs->ziostat_gang_multilevel.value.ui64 = 192 wmsum_value(&ziostat_sums.ziostat_gang_multilevel); 193 return (0); 194 } 195 196 void 197 zio_init(void) 198 { 199 size_t c; 200 201 zio_cache = kmem_cache_create("zio_cache", 202 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0); 203 zio_link_cache = kmem_cache_create("zio_link_cache", 204 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0); 205 206 wmsum_init(&ziostat_sums.ziostat_total_allocations, 0); 207 wmsum_init(&ziostat_sums.ziostat_alloc_class_fallbacks, 0); 208 wmsum_init(&ziostat_sums.ziostat_gang_writes, 0); 209 wmsum_init(&ziostat_sums.ziostat_gang_multilevel, 0); 210 zio_ksp = kstat_create("zfs", 0, "zio_stats", 211 "misc", KSTAT_TYPE_NAMED, sizeof (zio_stats) / 212 sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 213 if (zio_ksp != NULL) { 214 zio_ksp->ks_data = &zio_stats; 215 zio_ksp->ks_update = zio_kstats_update; 216 kstat_install(zio_ksp); 217 } 218 219 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { 220 size_t size = (c + 1) << SPA_MINBLOCKSHIFT; 221 size_t align, cflags, data_cflags; 222 char name[32]; 223 224 /* 225 * Create cache for each half-power of 2 size, starting from 226 * SPA_MINBLOCKSIZE. It should give us memory space efficiency 227 * of ~7/8, sufficient for transient allocations mostly using 228 * these caches. 229 */ 230 size_t p2 = size; 231 while (!ISP2(p2)) 232 p2 &= p2 - 1; 233 if (!IS_P2ALIGNED(size, p2 / 2)) 234 continue; 235 236 #ifndef _KERNEL 237 /* 238 * If we are using watchpoints, put each buffer on its own page, 239 * to eliminate the performance overhead of trapping to the 240 * kernel when modifying a non-watched buffer that shares the 241 * page with a watched buffer. 242 */ 243 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE)) 244 continue; 245 #endif 246 247 if (IS_P2ALIGNED(size, PAGESIZE)) 248 align = PAGESIZE; 249 else 250 align = 1 << (highbit64(size ^ (size - 1)) - 1); 251 252 cflags = (zio_exclude_metadata || size > zio_buf_debug_limit) ? 253 KMC_NODEBUG : 0; 254 data_cflags = KMC_NODEBUG; 255 if (abd_size_alloc_linear(size)) { 256 cflags |= KMC_RECLAIMABLE; 257 data_cflags |= KMC_RECLAIMABLE; 258 } 259 if (cflags == data_cflags) { 260 /* 261 * Resulting kmem caches would be identical. 262 * Save memory by creating only one. 263 */ 264 (void) snprintf(name, sizeof (name), 265 "zio_buf_comb_%lu", (ulong_t)size); 266 zio_buf_cache[c] = kmem_cache_create(name, size, align, 267 NULL, NULL, NULL, NULL, NULL, cflags); 268 zio_data_buf_cache[c] = zio_buf_cache[c]; 269 continue; 270 } 271 (void) snprintf(name, sizeof (name), "zio_buf_%lu", 272 (ulong_t)size); 273 zio_buf_cache[c] = kmem_cache_create(name, size, align, 274 NULL, NULL, NULL, NULL, NULL, cflags); 275 276 (void) snprintf(name, sizeof (name), "zio_data_buf_%lu", 277 (ulong_t)size); 278 zio_data_buf_cache[c] = kmem_cache_create(name, size, align, 279 NULL, NULL, NULL, NULL, NULL, data_cflags); 280 } 281 282 while (--c != 0) { 283 ASSERT(zio_buf_cache[c] != NULL); 284 if (zio_buf_cache[c - 1] == NULL) 285 zio_buf_cache[c - 1] = zio_buf_cache[c]; 286 287 ASSERT(zio_data_buf_cache[c] != NULL); 288 if (zio_data_buf_cache[c - 1] == NULL) 289 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c]; 290 } 291 292 zio_inject_init(); 293 294 lz4_init(); 295 } 296 297 void 298 zio_fini(void) 299 { 300 size_t n = SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; 301 302 #if defined(ZFS_DEBUG) && !defined(_KERNEL) 303 for (size_t i = 0; i < n; i++) { 304 if (zio_buf_cache_allocs[i] != zio_buf_cache_frees[i]) 305 (void) printf("zio_fini: [%d] %llu != %llu\n", 306 (int)((i + 1) << SPA_MINBLOCKSHIFT), 307 (long long unsigned)zio_buf_cache_allocs[i], 308 (long long unsigned)zio_buf_cache_frees[i]); 309 } 310 #endif 311 312 /* 313 * The same kmem cache can show up multiple times in both zio_buf_cache 314 * and zio_data_buf_cache. Do a wasteful but trivially correct scan to 315 * sort it out. 316 */ 317 for (size_t i = 0; i < n; i++) { 318 kmem_cache_t *cache = zio_buf_cache[i]; 319 if (cache == NULL) 320 continue; 321 for (size_t j = i; j < n; j++) { 322 if (cache == zio_buf_cache[j]) 323 zio_buf_cache[j] = NULL; 324 if (cache == zio_data_buf_cache[j]) 325 zio_data_buf_cache[j] = NULL; 326 } 327 kmem_cache_destroy(cache); 328 } 329 330 for (size_t i = 0; i < n; i++) { 331 kmem_cache_t *cache = zio_data_buf_cache[i]; 332 if (cache == NULL) 333 continue; 334 for (size_t j = i; j < n; j++) { 335 if (cache == zio_data_buf_cache[j]) 336 zio_data_buf_cache[j] = NULL; 337 } 338 kmem_cache_destroy(cache); 339 } 340 341 for (size_t i = 0; i < n; i++) { 342 VERIFY3P(zio_buf_cache[i], ==, NULL); 343 VERIFY3P(zio_data_buf_cache[i], ==, NULL); 344 } 345 346 if (zio_ksp != NULL) { 347 kstat_delete(zio_ksp); 348 zio_ksp = NULL; 349 } 350 351 wmsum_fini(&ziostat_sums.ziostat_total_allocations); 352 wmsum_fini(&ziostat_sums.ziostat_alloc_class_fallbacks); 353 wmsum_fini(&ziostat_sums.ziostat_gang_writes); 354 wmsum_fini(&ziostat_sums.ziostat_gang_multilevel); 355 356 kmem_cache_destroy(zio_link_cache); 357 kmem_cache_destroy(zio_cache); 358 359 zio_inject_fini(); 360 361 lz4_fini(); 362 } 363 364 /* 365 * ========================================================================== 366 * Allocate and free I/O buffers 367 * ========================================================================== 368 */ 369 370 #if defined(ZFS_DEBUG) && defined(_KERNEL) 371 #define ZFS_ZIO_BUF_CANARY 1 372 #endif 373 374 #ifdef ZFS_ZIO_BUF_CANARY 375 static const ulong_t zio_buf_canary = (ulong_t)0xdeadc0dedead210b; 376 377 /* 378 * Use empty space after the buffer to detect overflows. 379 * 380 * Since zio_init() creates kmem caches only for certain set of buffer sizes, 381 * allocations of different sizes may have some unused space after the data. 382 * Filling part of that space with a known pattern on allocation and checking 383 * it on free should allow us to detect some buffer overflows. 384 */ 385 static void 386 zio_buf_put_canary(ulong_t *p, size_t size, kmem_cache_t **cache, size_t c) 387 { 388 size_t off = P2ROUNDUP(size, sizeof (ulong_t)); 389 ulong_t *canary = p + off / sizeof (ulong_t); 390 size_t asize = (c + 1) << SPA_MINBLOCKSHIFT; 391 if (c + 1 < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT && 392 cache[c] == cache[c + 1]) 393 asize = (c + 2) << SPA_MINBLOCKSHIFT; 394 for (; off < asize; canary++, off += sizeof (ulong_t)) 395 *canary = zio_buf_canary; 396 } 397 398 static void 399 zio_buf_check_canary(ulong_t *p, size_t size, kmem_cache_t **cache, size_t c) 400 { 401 size_t off = P2ROUNDUP(size, sizeof (ulong_t)); 402 ulong_t *canary = p + off / sizeof (ulong_t); 403 size_t asize = (c + 1) << SPA_MINBLOCKSHIFT; 404 if (c + 1 < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT && 405 cache[c] == cache[c + 1]) 406 asize = (c + 2) << SPA_MINBLOCKSHIFT; 407 for (; off < asize; canary++, off += sizeof (ulong_t)) { 408 if (unlikely(*canary != zio_buf_canary)) { 409 PANIC("ZIO buffer overflow %p (%zu) + %zu %#lx != %#lx", 410 p, size, (canary - p) * sizeof (ulong_t), 411 *canary, zio_buf_canary); 412 } 413 } 414 } 415 #endif 416 417 /* 418 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a 419 * crashdump if the kernel panics, so use it judiciously. Obviously, it's 420 * useful to inspect ZFS metadata, but if possible, we should avoid keeping 421 * excess / transient data in-core during a crashdump. 422 */ 423 void * 424 zio_buf_alloc(size_t size) 425 { 426 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 427 428 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 429 #if defined(ZFS_DEBUG) && !defined(_KERNEL) 430 atomic_add_64(&zio_buf_cache_allocs[c], 1); 431 #endif 432 433 void *p = kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE); 434 #ifdef ZFS_ZIO_BUF_CANARY 435 zio_buf_put_canary(p, size, zio_buf_cache, c); 436 #endif 437 return (p); 438 } 439 440 /* 441 * Use zio_data_buf_alloc to allocate data. The data will not appear in a 442 * crashdump if the kernel panics. This exists so that we will limit the amount 443 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount 444 * of kernel heap dumped to disk when the kernel panics) 445 */ 446 void * 447 zio_data_buf_alloc(size_t size) 448 { 449 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 450 451 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 452 453 void *p = kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE); 454 #ifdef ZFS_ZIO_BUF_CANARY 455 zio_buf_put_canary(p, size, zio_data_buf_cache, c); 456 #endif 457 return (p); 458 } 459 460 void 461 zio_buf_free(void *buf, size_t size) 462 { 463 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 464 465 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 466 #if defined(ZFS_DEBUG) && !defined(_KERNEL) 467 atomic_add_64(&zio_buf_cache_frees[c], 1); 468 #endif 469 470 #ifdef ZFS_ZIO_BUF_CANARY 471 zio_buf_check_canary(buf, size, zio_buf_cache, c); 472 #endif 473 kmem_cache_free(zio_buf_cache[c], buf); 474 } 475 476 void 477 zio_data_buf_free(void *buf, size_t size) 478 { 479 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 480 481 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 482 483 #ifdef ZFS_ZIO_BUF_CANARY 484 zio_buf_check_canary(buf, size, zio_data_buf_cache, c); 485 #endif 486 kmem_cache_free(zio_data_buf_cache[c], buf); 487 } 488 489 static void 490 zio_abd_free(void *abd, size_t size) 491 { 492 (void) size; 493 abd_free((abd_t *)abd); 494 } 495 496 /* 497 * ========================================================================== 498 * Push and pop I/O transform buffers 499 * ========================================================================== 500 */ 501 void 502 zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize, 503 zio_transform_func_t *transform) 504 { 505 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP); 506 507 zt->zt_orig_abd = zio->io_abd; 508 zt->zt_orig_size = zio->io_size; 509 zt->zt_bufsize = bufsize; 510 zt->zt_transform = transform; 511 512 zt->zt_next = zio->io_transform_stack; 513 zio->io_transform_stack = zt; 514 515 zio->io_abd = data; 516 zio->io_size = size; 517 } 518 519 void 520 zio_pop_transforms(zio_t *zio) 521 { 522 zio_transform_t *zt; 523 524 while ((zt = zio->io_transform_stack) != NULL) { 525 if (zt->zt_transform != NULL) 526 zt->zt_transform(zio, 527 zt->zt_orig_abd, zt->zt_orig_size); 528 529 if (zt->zt_bufsize != 0) 530 abd_free(zio->io_abd); 531 532 zio->io_abd = zt->zt_orig_abd; 533 zio->io_size = zt->zt_orig_size; 534 zio->io_transform_stack = zt->zt_next; 535 536 kmem_free(zt, sizeof (zio_transform_t)); 537 } 538 } 539 540 /* 541 * ========================================================================== 542 * I/O transform callbacks for subblocks, decompression, and decryption 543 * ========================================================================== 544 */ 545 static void 546 zio_subblock(zio_t *zio, abd_t *data, uint64_t size) 547 { 548 ASSERT(zio->io_size > size); 549 550 if (zio->io_type == ZIO_TYPE_READ) 551 abd_copy(data, zio->io_abd, size); 552 } 553 554 static void 555 zio_decompress(zio_t *zio, abd_t *data, uint64_t size) 556 { 557 if (zio->io_error == 0) { 558 int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp), 559 zio->io_abd, data, zio->io_size, size, 560 &zio->io_prop.zp_complevel); 561 562 if (zio_injection_enabled && ret == 0) 563 ret = zio_handle_fault_injection(zio, EINVAL); 564 565 if (ret != 0) 566 zio->io_error = SET_ERROR(EIO); 567 } 568 } 569 570 static void 571 zio_decrypt(zio_t *zio, abd_t *data, uint64_t size) 572 { 573 int ret; 574 void *tmp; 575 blkptr_t *bp = zio->io_bp; 576 spa_t *spa = zio->io_spa; 577 uint64_t dsobj = zio->io_bookmark.zb_objset; 578 uint64_t lsize = BP_GET_LSIZE(bp); 579 dmu_object_type_t ot = BP_GET_TYPE(bp); 580 uint8_t salt[ZIO_DATA_SALT_LEN]; 581 uint8_t iv[ZIO_DATA_IV_LEN]; 582 uint8_t mac[ZIO_DATA_MAC_LEN]; 583 boolean_t no_crypt = B_FALSE; 584 585 ASSERT(BP_USES_CRYPT(bp)); 586 ASSERT3U(size, !=, 0); 587 588 if (zio->io_error != 0) 589 return; 590 591 /* 592 * Verify the cksum of MACs stored in an indirect bp. It will always 593 * be possible to verify this since it does not require an encryption 594 * key. 595 */ 596 if (BP_HAS_INDIRECT_MAC_CKSUM(bp)) { 597 zio_crypt_decode_mac_bp(bp, mac); 598 599 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF) { 600 /* 601 * We haven't decompressed the data yet, but 602 * zio_crypt_do_indirect_mac_checksum() requires 603 * decompressed data to be able to parse out the MACs 604 * from the indirect block. We decompress it now and 605 * throw away the result after we are finished. 606 */ 607 abd_t *abd = abd_alloc_linear(lsize, B_TRUE); 608 ret = zio_decompress_data(BP_GET_COMPRESS(bp), 609 zio->io_abd, abd, zio->io_size, lsize, 610 &zio->io_prop.zp_complevel); 611 if (ret != 0) { 612 abd_free(abd); 613 ret = SET_ERROR(EIO); 614 goto error; 615 } 616 ret = zio_crypt_do_indirect_mac_checksum_abd(B_FALSE, 617 abd, lsize, BP_SHOULD_BYTESWAP(bp), mac); 618 abd_free(abd); 619 } else { 620 ret = zio_crypt_do_indirect_mac_checksum_abd(B_FALSE, 621 zio->io_abd, size, BP_SHOULD_BYTESWAP(bp), mac); 622 } 623 abd_copy(data, zio->io_abd, size); 624 625 if (zio_injection_enabled && ot != DMU_OT_DNODE && ret == 0) { 626 ret = zio_handle_decrypt_injection(spa, 627 &zio->io_bookmark, ot, ECKSUM); 628 } 629 if (ret != 0) 630 goto error; 631 632 return; 633 } 634 635 /* 636 * If this is an authenticated block, just check the MAC. It would be 637 * nice to separate this out into its own flag, but when this was done, 638 * we had run out of bits in what is now zio_flag_t. Future cleanup 639 * could make this a flag bit. 640 */ 641 if (BP_IS_AUTHENTICATED(bp)) { 642 if (ot == DMU_OT_OBJSET) { 643 ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa, 644 dsobj, zio->io_abd, size, BP_SHOULD_BYTESWAP(bp)); 645 } else { 646 zio_crypt_decode_mac_bp(bp, mac); 647 ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj, 648 zio->io_abd, size, mac); 649 if (zio_injection_enabled && ret == 0) { 650 ret = zio_handle_decrypt_injection(spa, 651 &zio->io_bookmark, ot, ECKSUM); 652 } 653 } 654 abd_copy(data, zio->io_abd, size); 655 656 if (ret != 0) 657 goto error; 658 659 return; 660 } 661 662 zio_crypt_decode_params_bp(bp, salt, iv); 663 664 if (ot == DMU_OT_INTENT_LOG) { 665 tmp = abd_borrow_buf_copy(zio->io_abd, sizeof (zil_chain_t)); 666 zio_crypt_decode_mac_zil(tmp, mac); 667 abd_return_buf(zio->io_abd, tmp, sizeof (zil_chain_t)); 668 } else { 669 zio_crypt_decode_mac_bp(bp, mac); 670 } 671 672 ret = spa_do_crypt_abd(B_FALSE, spa, &zio->io_bookmark, BP_GET_TYPE(bp), 673 BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp), salt, iv, mac, size, data, 674 zio->io_abd, &no_crypt); 675 if (no_crypt) 676 abd_copy(data, zio->io_abd, size); 677 678 if (ret != 0) 679 goto error; 680 681 return; 682 683 error: 684 /* assert that the key was found unless this was speculative */ 685 ASSERT(ret != EACCES || (zio->io_flags & ZIO_FLAG_SPECULATIVE)); 686 687 /* 688 * If there was a decryption / authentication error return EIO as 689 * the io_error. If this was not a speculative zio, create an ereport. 690 */ 691 if (ret == ECKSUM) { 692 zio->io_error = SET_ERROR(EIO); 693 if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) { 694 spa_log_error(spa, &zio->io_bookmark, 695 BP_GET_LOGICAL_BIRTH(zio->io_bp)); 696 (void) zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION, 697 spa, NULL, &zio->io_bookmark, zio, 0); 698 } 699 } else { 700 zio->io_error = ret; 701 } 702 } 703 704 /* 705 * ========================================================================== 706 * I/O parent/child relationships and pipeline interlocks 707 * ========================================================================== 708 */ 709 zio_t * 710 zio_walk_parents(zio_t *cio, zio_link_t **zl) 711 { 712 list_t *pl = &cio->io_parent_list; 713 714 *zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl); 715 if (*zl == NULL) 716 return (NULL); 717 718 ASSERT((*zl)->zl_child == cio); 719 return ((*zl)->zl_parent); 720 } 721 722 zio_t * 723 zio_walk_children(zio_t *pio, zio_link_t **zl) 724 { 725 list_t *cl = &pio->io_child_list; 726 727 ASSERT(MUTEX_HELD(&pio->io_lock)); 728 729 *zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl); 730 if (*zl == NULL) 731 return (NULL); 732 733 ASSERT((*zl)->zl_parent == pio); 734 return ((*zl)->zl_child); 735 } 736 737 zio_t * 738 zio_unique_parent(zio_t *cio) 739 { 740 zio_link_t *zl = NULL; 741 zio_t *pio = zio_walk_parents(cio, &zl); 742 743 VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL); 744 return (pio); 745 } 746 747 void 748 zio_add_child(zio_t *pio, zio_t *cio) 749 { 750 /* 751 * Logical I/Os can have logical, gang, or vdev children. 752 * Gang I/Os can have gang or vdev children. 753 * Vdev I/Os can only have vdev children. 754 * The following ASSERT captures all of these constraints. 755 */ 756 ASSERT3S(cio->io_child_type, <=, pio->io_child_type); 757 758 /* Parent should not have READY stage if child doesn't have it. */ 759 IMPLY((cio->io_pipeline & ZIO_STAGE_READY) == 0 && 760 (cio->io_child_type != ZIO_CHILD_VDEV), 761 (pio->io_pipeline & ZIO_STAGE_READY) == 0); 762 763 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP); 764 zl->zl_parent = pio; 765 zl->zl_child = cio; 766 767 mutex_enter(&pio->io_lock); 768 mutex_enter(&cio->io_lock); 769 770 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0); 771 772 uint64_t *countp = pio->io_children[cio->io_child_type]; 773 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 774 countp[w] += !cio->io_state[w]; 775 776 list_insert_head(&pio->io_child_list, zl); 777 list_insert_head(&cio->io_parent_list, zl); 778 779 mutex_exit(&cio->io_lock); 780 mutex_exit(&pio->io_lock); 781 } 782 783 void 784 zio_add_child_first(zio_t *pio, zio_t *cio) 785 { 786 /* 787 * Logical I/Os can have logical, gang, or vdev children. 788 * Gang I/Os can have gang or vdev children. 789 * Vdev I/Os can only have vdev children. 790 * The following ASSERT captures all of these constraints. 791 */ 792 ASSERT3S(cio->io_child_type, <=, pio->io_child_type); 793 794 /* Parent should not have READY stage if child doesn't have it. */ 795 IMPLY((cio->io_pipeline & ZIO_STAGE_READY) == 0 && 796 (cio->io_child_type != ZIO_CHILD_VDEV), 797 (pio->io_pipeline & ZIO_STAGE_READY) == 0); 798 799 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP); 800 zl->zl_parent = pio; 801 zl->zl_child = cio; 802 803 ASSERT(list_is_empty(&cio->io_parent_list)); 804 list_insert_head(&cio->io_parent_list, zl); 805 806 mutex_enter(&pio->io_lock); 807 808 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0); 809 810 uint64_t *countp = pio->io_children[cio->io_child_type]; 811 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 812 countp[w] += !cio->io_state[w]; 813 814 list_insert_head(&pio->io_child_list, zl); 815 816 mutex_exit(&pio->io_lock); 817 } 818 819 static void 820 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl) 821 { 822 ASSERT(zl->zl_parent == pio); 823 ASSERT(zl->zl_child == cio); 824 825 mutex_enter(&pio->io_lock); 826 mutex_enter(&cio->io_lock); 827 828 list_remove(&pio->io_child_list, zl); 829 list_remove(&cio->io_parent_list, zl); 830 831 mutex_exit(&cio->io_lock); 832 mutex_exit(&pio->io_lock); 833 kmem_cache_free(zio_link_cache, zl); 834 } 835 836 static boolean_t 837 zio_wait_for_children(zio_t *zio, uint8_t childbits, enum zio_wait_type wait) 838 { 839 boolean_t waiting = B_FALSE; 840 841 mutex_enter(&zio->io_lock); 842 ASSERT(zio->io_stall == NULL); 843 for (int c = 0; c < ZIO_CHILD_TYPES; c++) { 844 if (!(ZIO_CHILD_BIT_IS_SET(childbits, c))) 845 continue; 846 847 uint64_t *countp = &zio->io_children[c][wait]; 848 if (*countp != 0) { 849 zio->io_stage >>= 1; 850 ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN); 851 zio->io_stall = countp; 852 waiting = B_TRUE; 853 break; 854 } 855 } 856 mutex_exit(&zio->io_lock); 857 return (waiting); 858 } 859 860 __attribute__((always_inline)) 861 static inline void 862 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait, 863 zio_t **next_to_executep) 864 { 865 uint64_t *countp = &pio->io_children[zio->io_child_type][wait]; 866 int *errorp = &pio->io_child_error[zio->io_child_type]; 867 868 mutex_enter(&pio->io_lock); 869 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE)) 870 *errorp = zio_worst_error(*errorp, zio->io_error); 871 pio->io_reexecute |= zio->io_reexecute; 872 ASSERT3U(*countp, >, 0); 873 874 /* 875 * Propogate the Direct I/O checksum verify failure to the parent. 876 */ 877 if (zio->io_flags & ZIO_FLAG_DIO_CHKSUM_ERR) 878 pio->io_flags |= ZIO_FLAG_DIO_CHKSUM_ERR; 879 880 (*countp)--; 881 882 if (*countp == 0 && pio->io_stall == countp) { 883 zio_taskq_type_t type = 884 pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE : 885 ZIO_TASKQ_INTERRUPT; 886 pio->io_stall = NULL; 887 mutex_exit(&pio->io_lock); 888 889 /* 890 * If we can tell the caller to execute this parent next, do 891 * so. We do this if the parent's zio type matches the child's 892 * type, or if it's a zio_null() with no done callback, and so 893 * has no actual work to do. Otherwise dispatch the parent zio 894 * in its own taskq. 895 * 896 * Having the caller execute the parent when possible reduces 897 * locking on the zio taskq's, reduces context switch 898 * overhead, and has no recursion penalty. Note that one 899 * read from disk typically causes at least 3 zio's: a 900 * zio_null(), the logical zio_read(), and then a physical 901 * zio. When the physical ZIO completes, we are able to call 902 * zio_done() on all 3 of these zio's from one invocation of 903 * zio_execute() by returning the parent back to 904 * zio_execute(). Since the parent isn't executed until this 905 * thread returns back to zio_execute(), the caller should do 906 * so promptly. 907 * 908 * In other cases, dispatching the parent prevents 909 * overflowing the stack when we have deeply nested 910 * parent-child relationships, as we do with the "mega zio" 911 * of writes for spa_sync(), and the chain of ZIL blocks. 912 */ 913 if (next_to_executep != NULL && *next_to_executep == NULL && 914 (pio->io_type == zio->io_type || 915 (pio->io_type == ZIO_TYPE_NULL && !pio->io_done))) { 916 *next_to_executep = pio; 917 } else { 918 zio_taskq_dispatch(pio, type, B_FALSE); 919 } 920 } else { 921 mutex_exit(&pio->io_lock); 922 } 923 } 924 925 static void 926 zio_inherit_child_errors(zio_t *zio, enum zio_child c) 927 { 928 if (zio->io_child_error[c] != 0 && zio->io_error == 0) 929 zio->io_error = zio->io_child_error[c]; 930 } 931 932 int 933 zio_bookmark_compare(const void *x1, const void *x2) 934 { 935 const zio_t *z1 = x1; 936 const zio_t *z2 = x2; 937 938 if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset) 939 return (-1); 940 if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset) 941 return (1); 942 943 if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object) 944 return (-1); 945 if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object) 946 return (1); 947 948 if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level) 949 return (-1); 950 if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level) 951 return (1); 952 953 if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid) 954 return (-1); 955 if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid) 956 return (1); 957 958 if (z1 < z2) 959 return (-1); 960 if (z1 > z2) 961 return (1); 962 963 return (0); 964 } 965 966 /* 967 * ========================================================================== 968 * Create the various types of I/O (read, write, free, etc) 969 * ========================================================================== 970 */ 971 static zio_t * 972 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 973 abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done, 974 void *private, zio_type_t type, zio_priority_t priority, 975 zio_flag_t flags, vdev_t *vd, uint64_t offset, 976 const zbookmark_phys_t *zb, enum zio_stage stage, 977 enum zio_stage pipeline) 978 { 979 zio_t *zio; 980 981 IMPLY(type != ZIO_TYPE_TRIM, psize <= SPA_MAXBLOCKSIZE); 982 ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0); 983 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0); 984 985 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER)); 986 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER)); 987 ASSERT(vd || stage == ZIO_STAGE_OPEN); 988 989 IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW_COMPRESS) != 0); 990 991 zio = kmem_cache_alloc(zio_cache, KM_SLEEP); 992 memset(zio, 0, sizeof (zio_t)); 993 994 mutex_init(&zio->io_lock, NULL, MUTEX_NOLOCKDEP, NULL); 995 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL); 996 997 list_create(&zio->io_parent_list, sizeof (zio_link_t), 998 offsetof(zio_link_t, zl_parent_node)); 999 list_create(&zio->io_child_list, sizeof (zio_link_t), 1000 offsetof(zio_link_t, zl_child_node)); 1001 metaslab_trace_init(&zio->io_alloc_list); 1002 1003 if (vd != NULL) 1004 zio->io_child_type = ZIO_CHILD_VDEV; 1005 else if (flags & ZIO_FLAG_GANG_CHILD) 1006 zio->io_child_type = ZIO_CHILD_GANG; 1007 else if (flags & ZIO_FLAG_DDT_CHILD) 1008 zio->io_child_type = ZIO_CHILD_DDT; 1009 else 1010 zio->io_child_type = ZIO_CHILD_LOGICAL; 1011 1012 if (bp != NULL) { 1013 if (type != ZIO_TYPE_WRITE || 1014 zio->io_child_type == ZIO_CHILD_DDT) { 1015 zio->io_bp_copy = *bp; 1016 zio->io_bp = &zio->io_bp_copy; /* so caller can free */ 1017 } else { 1018 zio->io_bp = (blkptr_t *)bp; 1019 } 1020 zio->io_bp_orig = *bp; 1021 if (zio->io_child_type == ZIO_CHILD_LOGICAL) 1022 zio->io_logical = zio; 1023 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp)) 1024 pipeline |= ZIO_GANG_STAGES; 1025 if (flags & ZIO_FLAG_PREALLOCATED) { 1026 BP_ZERO_DVAS(zio->io_bp); 1027 BP_SET_BIRTH(zio->io_bp, 0, 0); 1028 } 1029 } 1030 1031 zio->io_spa = spa; 1032 zio->io_txg = txg; 1033 zio->io_done = done; 1034 zio->io_private = private; 1035 zio->io_type = type; 1036 zio->io_priority = priority; 1037 zio->io_vd = vd; 1038 zio->io_offset = offset; 1039 zio->io_orig_abd = zio->io_abd = data; 1040 zio->io_orig_size = zio->io_size = psize; 1041 zio->io_lsize = lsize; 1042 zio->io_orig_flags = zio->io_flags = flags; 1043 zio->io_orig_stage = zio->io_stage = stage; 1044 zio->io_orig_pipeline = zio->io_pipeline = pipeline; 1045 zio->io_pipeline_trace = ZIO_STAGE_OPEN; 1046 zio->io_allocator = ZIO_ALLOCATOR_NONE; 1047 1048 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY) || 1049 (pipeline & ZIO_STAGE_READY) == 0; 1050 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE); 1051 1052 if (zb != NULL) 1053 zio->io_bookmark = *zb; 1054 1055 if (pio != NULL) { 1056 zio->io_metaslab_class = pio->io_metaslab_class; 1057 if (zio->io_logical == NULL) 1058 zio->io_logical = pio->io_logical; 1059 if (zio->io_child_type == ZIO_CHILD_GANG) 1060 zio->io_gang_leader = pio->io_gang_leader; 1061 zio_add_child_first(pio, zio); 1062 } 1063 1064 taskq_init_ent(&zio->io_tqent); 1065 1066 return (zio); 1067 } 1068 1069 void 1070 zio_destroy(zio_t *zio) 1071 { 1072 metaslab_trace_fini(&zio->io_alloc_list); 1073 list_destroy(&zio->io_parent_list); 1074 list_destroy(&zio->io_child_list); 1075 mutex_destroy(&zio->io_lock); 1076 cv_destroy(&zio->io_cv); 1077 kmem_cache_free(zio_cache, zio); 1078 } 1079 1080 /* 1081 * ZIO intended to be between others. Provides synchronization at READY 1082 * and DONE pipeline stages and calls the respective callbacks. 1083 */ 1084 zio_t * 1085 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done, 1086 void *private, zio_flag_t flags) 1087 { 1088 zio_t *zio; 1089 1090 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private, 1091 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, 1092 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE); 1093 1094 return (zio); 1095 } 1096 1097 /* 1098 * ZIO intended to be a root of a tree. Unlike null ZIO does not have a 1099 * READY pipeline stage (is ready on creation), so it should not be used 1100 * as child of any ZIO that may need waiting for grandchildren READY stage 1101 * (any other ZIO type). 1102 */ 1103 zio_t * 1104 zio_root(spa_t *spa, zio_done_func_t *done, void *private, zio_flag_t flags) 1105 { 1106 zio_t *zio; 1107 1108 zio = zio_create(NULL, spa, 0, NULL, NULL, 0, 0, done, private, 1109 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, NULL, 0, NULL, 1110 ZIO_STAGE_OPEN, ZIO_ROOT_PIPELINE); 1111 1112 return (zio); 1113 } 1114 1115 static int 1116 zfs_blkptr_verify_log(spa_t *spa, const blkptr_t *bp, 1117 enum blk_verify_flag blk_verify, const char *fmt, ...) 1118 { 1119 va_list adx; 1120 char buf[256]; 1121 1122 va_start(adx, fmt); 1123 (void) vsnprintf(buf, sizeof (buf), fmt, adx); 1124 va_end(adx); 1125 1126 zfs_dbgmsg("bad blkptr at %px: " 1127 "DVA[0]=%#llx/%#llx " 1128 "DVA[1]=%#llx/%#llx " 1129 "DVA[2]=%#llx/%#llx " 1130 "prop=%#llx " 1131 "pad=%#llx,%#llx " 1132 "phys_birth=%#llx " 1133 "birth=%#llx " 1134 "fill=%#llx " 1135 "cksum=%#llx/%#llx/%#llx/%#llx", 1136 bp, 1137 (long long)bp->blk_dva[0].dva_word[0], 1138 (long long)bp->blk_dva[0].dva_word[1], 1139 (long long)bp->blk_dva[1].dva_word[0], 1140 (long long)bp->blk_dva[1].dva_word[1], 1141 (long long)bp->blk_dva[2].dva_word[0], 1142 (long long)bp->blk_dva[2].dva_word[1], 1143 (long long)bp->blk_prop, 1144 (long long)bp->blk_pad[0], 1145 (long long)bp->blk_pad[1], 1146 (long long)BP_GET_PHYSICAL_BIRTH(bp), 1147 (long long)BP_GET_LOGICAL_BIRTH(bp), 1148 (long long)bp->blk_fill, 1149 (long long)bp->blk_cksum.zc_word[0], 1150 (long long)bp->blk_cksum.zc_word[1], 1151 (long long)bp->blk_cksum.zc_word[2], 1152 (long long)bp->blk_cksum.zc_word[3]); 1153 switch (blk_verify) { 1154 case BLK_VERIFY_HALT: 1155 zfs_panic_recover("%s: %s", spa_name(spa), buf); 1156 break; 1157 case BLK_VERIFY_LOG: 1158 zfs_dbgmsg("%s: %s", spa_name(spa), buf); 1159 break; 1160 case BLK_VERIFY_ONLY: 1161 break; 1162 } 1163 1164 return (1); 1165 } 1166 1167 /* 1168 * Verify the block pointer fields contain reasonable values. This means 1169 * it only contains known object types, checksum/compression identifiers, 1170 * block sizes within the maximum allowed limits, valid DVAs, etc. 1171 * 1172 * If everything checks out 0 is returned. The zfs_blkptr_verify 1173 * argument controls the behavior when an invalid field is detected. 1174 * 1175 * Values for blk_verify_flag: 1176 * BLK_VERIFY_ONLY: evaluate the block 1177 * BLK_VERIFY_LOG: evaluate the block and log problems 1178 * BLK_VERIFY_HALT: call zfs_panic_recover on error 1179 * 1180 * Values for blk_config_flag: 1181 * BLK_CONFIG_HELD: caller holds SCL_VDEV for writer 1182 * BLK_CONFIG_NEEDED: caller holds no config lock, SCL_VDEV will be 1183 * obtained for reader 1184 * BLK_CONFIG_SKIP: skip checks which require SCL_VDEV, for better 1185 * performance 1186 */ 1187 int 1188 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp, 1189 enum blk_config_flag blk_config, enum blk_verify_flag blk_verify) 1190 { 1191 int errors = 0; 1192 1193 if (unlikely(!DMU_OT_IS_VALID(BP_GET_TYPE(bp)))) { 1194 errors += zfs_blkptr_verify_log(spa, bp, blk_verify, 1195 "blkptr at %px has invalid TYPE %llu", 1196 bp, (longlong_t)BP_GET_TYPE(bp)); 1197 } 1198 if (unlikely(BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS)) { 1199 errors += zfs_blkptr_verify_log(spa, bp, blk_verify, 1200 "blkptr at %px has invalid COMPRESS %llu", 1201 bp, (longlong_t)BP_GET_COMPRESS(bp)); 1202 } 1203 if (unlikely(BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE)) { 1204 errors += zfs_blkptr_verify_log(spa, bp, blk_verify, 1205 "blkptr at %px has invalid LSIZE %llu", 1206 bp, (longlong_t)BP_GET_LSIZE(bp)); 1207 } 1208 if (BP_IS_EMBEDDED(bp)) { 1209 if (unlikely(BPE_GET_ETYPE(bp) >= NUM_BP_EMBEDDED_TYPES)) { 1210 errors += zfs_blkptr_verify_log(spa, bp, blk_verify, 1211 "blkptr at %px has invalid ETYPE %llu", 1212 bp, (longlong_t)BPE_GET_ETYPE(bp)); 1213 } 1214 if (unlikely(BPE_GET_PSIZE(bp) > BPE_PAYLOAD_SIZE)) { 1215 errors += zfs_blkptr_verify_log(spa, bp, blk_verify, 1216 "blkptr at %px has invalid PSIZE %llu", 1217 bp, (longlong_t)BPE_GET_PSIZE(bp)); 1218 } 1219 return (errors ? ECKSUM : 0); 1220 } else if (BP_IS_HOLE(bp)) { 1221 /* 1222 * Holes are allowed (expected, even) to have no DVAs, no 1223 * checksum, and no psize. 1224 */ 1225 return (errors ? ECKSUM : 0); 1226 } else if (unlikely(!DVA_IS_VALID(&bp->blk_dva[0]))) { 1227 /* Non-hole, non-embedded BPs _must_ have at least one DVA */ 1228 errors += zfs_blkptr_verify_log(spa, bp, blk_verify, 1229 "blkptr at %px has no valid DVAs", bp); 1230 } 1231 if (unlikely(BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS)) { 1232 errors += zfs_blkptr_verify_log(spa, bp, blk_verify, 1233 "blkptr at %px has invalid CHECKSUM %llu", 1234 bp, (longlong_t)BP_GET_CHECKSUM(bp)); 1235 } 1236 if (unlikely(BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE)) { 1237 errors += zfs_blkptr_verify_log(spa, bp, blk_verify, 1238 "blkptr at %px has invalid PSIZE %llu", 1239 bp, (longlong_t)BP_GET_PSIZE(bp)); 1240 } 1241 1242 /* 1243 * Do not verify individual DVAs if the config is not trusted. This 1244 * will be done once the zio is executed in vdev_mirror_map_alloc. 1245 */ 1246 if (unlikely(!spa->spa_trust_config)) 1247 return (errors ? ECKSUM : 0); 1248 1249 switch (blk_config) { 1250 case BLK_CONFIG_HELD: 1251 ASSERT(spa_config_held(spa, SCL_VDEV, RW_WRITER)); 1252 break; 1253 case BLK_CONFIG_NEEDED: 1254 spa_config_enter(spa, SCL_VDEV, bp, RW_READER); 1255 break; 1256 case BLK_CONFIG_NEEDED_TRY: 1257 if (!spa_config_tryenter(spa, SCL_VDEV, bp, RW_READER)) 1258 return (EBUSY); 1259 break; 1260 case BLK_CONFIG_SKIP: 1261 return (errors ? ECKSUM : 0); 1262 default: 1263 panic("invalid blk_config %u", blk_config); 1264 } 1265 1266 /* 1267 * Pool-specific checks. 1268 * 1269 * Note: it would be nice to verify that the logical birth 1270 * and physical birth are not too large. However, 1271 * spa_freeze() allows the birth time of log blocks (and 1272 * dmu_sync()-ed blocks that are in the log) to be arbitrarily 1273 * large. 1274 */ 1275 for (int i = 0; i < BP_GET_NDVAS(bp); i++) { 1276 const dva_t *dva = &bp->blk_dva[i]; 1277 uint64_t vdevid = DVA_GET_VDEV(dva); 1278 1279 if (unlikely(vdevid >= spa->spa_root_vdev->vdev_children)) { 1280 errors += zfs_blkptr_verify_log(spa, bp, blk_verify, 1281 "blkptr at %px DVA %u has invalid VDEV %llu", 1282 bp, i, (longlong_t)vdevid); 1283 continue; 1284 } 1285 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid]; 1286 if (unlikely(vd == NULL)) { 1287 errors += zfs_blkptr_verify_log(spa, bp, blk_verify, 1288 "blkptr at %px DVA %u has invalid VDEV %llu", 1289 bp, i, (longlong_t)vdevid); 1290 continue; 1291 } 1292 if (unlikely(vd->vdev_ops == &vdev_hole_ops)) { 1293 errors += zfs_blkptr_verify_log(spa, bp, blk_verify, 1294 "blkptr at %px DVA %u has hole VDEV %llu", 1295 bp, i, (longlong_t)vdevid); 1296 continue; 1297 } 1298 if (vd->vdev_ops == &vdev_missing_ops) { 1299 /* 1300 * "missing" vdevs are valid during import, but we 1301 * don't have their detailed info (e.g. asize), so 1302 * we can't perform any more checks on them. 1303 */ 1304 continue; 1305 } 1306 uint64_t offset = DVA_GET_OFFSET(dva); 1307 uint64_t asize = DVA_GET_ASIZE(dva); 1308 if (DVA_GET_GANG(dva)) 1309 asize = vdev_gang_header_asize(vd); 1310 if (unlikely(offset + asize > vd->vdev_asize)) { 1311 errors += zfs_blkptr_verify_log(spa, bp, blk_verify, 1312 "blkptr at %px DVA %u has invalid OFFSET %llu", 1313 bp, i, (longlong_t)offset); 1314 } 1315 } 1316 if (blk_config == BLK_CONFIG_NEEDED || blk_config == 1317 BLK_CONFIG_NEEDED_TRY) 1318 spa_config_exit(spa, SCL_VDEV, bp); 1319 1320 return (errors ? ECKSUM : 0); 1321 } 1322 1323 boolean_t 1324 zfs_dva_valid(spa_t *spa, const dva_t *dva, const blkptr_t *bp) 1325 { 1326 (void) bp; 1327 uint64_t vdevid = DVA_GET_VDEV(dva); 1328 1329 if (vdevid >= spa->spa_root_vdev->vdev_children) 1330 return (B_FALSE); 1331 1332 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid]; 1333 if (vd == NULL) 1334 return (B_FALSE); 1335 1336 if (vd->vdev_ops == &vdev_hole_ops) 1337 return (B_FALSE); 1338 1339 if (vd->vdev_ops == &vdev_missing_ops) { 1340 return (B_FALSE); 1341 } 1342 1343 uint64_t offset = DVA_GET_OFFSET(dva); 1344 uint64_t asize = DVA_GET_ASIZE(dva); 1345 1346 if (DVA_GET_GANG(dva)) 1347 asize = vdev_gang_header_asize(vd); 1348 if (offset + asize > vd->vdev_asize) 1349 return (B_FALSE); 1350 1351 return (B_TRUE); 1352 } 1353 1354 zio_t * 1355 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, 1356 abd_t *data, uint64_t size, zio_done_func_t *done, void *private, 1357 zio_priority_t priority, zio_flag_t flags, const zbookmark_phys_t *zb) 1358 { 1359 zio_t *zio; 1360 1361 zio = zio_create(pio, spa, BP_GET_BIRTH(bp), bp, 1362 data, size, size, done, private, 1363 ZIO_TYPE_READ, priority, flags, NULL, 0, zb, 1364 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? 1365 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE); 1366 1367 return (zio); 1368 } 1369 1370 zio_t * 1371 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, 1372 abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp, 1373 zio_done_func_t *ready, zio_done_func_t *children_ready, 1374 zio_done_func_t *done, void *private, zio_priority_t priority, 1375 zio_flag_t flags, const zbookmark_phys_t *zb) 1376 { 1377 zio_t *zio; 1378 enum zio_stage pipeline = zp->zp_direct_write == B_TRUE ? 1379 ZIO_DIRECT_WRITE_PIPELINE : (flags & ZIO_FLAG_DDT_CHILD) ? 1380 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE; 1381 1382 1383 zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private, 1384 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, 1385 ZIO_STAGE_OPEN, pipeline); 1386 1387 zio->io_ready = ready; 1388 zio->io_children_ready = children_ready; 1389 zio->io_prop = *zp; 1390 1391 /* 1392 * Data can be NULL if we are going to call zio_write_override() to 1393 * provide the already-allocated BP. But we may need the data to 1394 * verify a dedup hit (if requested). In this case, don't try to 1395 * dedup (just take the already-allocated BP verbatim). Encrypted 1396 * dedup blocks need data as well so we also disable dedup in this 1397 * case. 1398 */ 1399 if (data == NULL && 1400 (zio->io_prop.zp_dedup_verify || zio->io_prop.zp_encrypt)) { 1401 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE; 1402 } 1403 1404 return (zio); 1405 } 1406 1407 zio_t * 1408 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data, 1409 uint64_t size, zio_done_func_t *done, void *private, 1410 zio_priority_t priority, zio_flag_t flags, zbookmark_phys_t *zb) 1411 { 1412 zio_t *zio; 1413 1414 zio = zio_create(pio, spa, txg, bp, data, size, size, done, private, 1415 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb, 1416 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE); 1417 1418 return (zio); 1419 } 1420 1421 void 1422 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, int gang_copies, 1423 boolean_t nopwrite, boolean_t brtwrite) 1424 { 1425 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 1426 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1427 ASSERT(zio->io_stage == ZIO_STAGE_OPEN); 1428 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa)); 1429 ASSERT(!brtwrite || !nopwrite); 1430 1431 /* 1432 * We must reset the io_prop to match the values that existed 1433 * when the bp was first written by dmu_sync() keeping in mind 1434 * that nopwrite and dedup are mutually exclusive. 1435 */ 1436 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup; 1437 zio->io_prop.zp_nopwrite = nopwrite; 1438 zio->io_prop.zp_brtwrite = brtwrite; 1439 zio->io_prop.zp_copies = copies; 1440 zio->io_prop.zp_gang_copies = gang_copies; 1441 zio->io_bp_override = bp; 1442 } 1443 1444 void 1445 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp) 1446 { 1447 1448 (void) zfs_blkptr_verify(spa, bp, BLK_CONFIG_NEEDED, BLK_VERIFY_HALT); 1449 1450 /* 1451 * The check for EMBEDDED is a performance optimization. We 1452 * process the free here (by ignoring it) rather than 1453 * putting it on the list and then processing it in zio_free_sync(). 1454 */ 1455 if (BP_IS_EMBEDDED(bp)) 1456 return; 1457 1458 /* 1459 * Frees that are for the currently-syncing txg, are not going to be 1460 * deferred, and which will not need to do a read (i.e. not GANG or 1461 * DEDUP), can be processed immediately. Otherwise, put them on the 1462 * in-memory list for later processing. 1463 * 1464 * Note that we only defer frees after zfs_sync_pass_deferred_free 1465 * when the log space map feature is disabled. [see relevant comment 1466 * in spa_sync_iterate_to_convergence()] 1467 */ 1468 if (BP_IS_GANG(bp) || 1469 BP_GET_DEDUP(bp) || 1470 txg != spa->spa_syncing_txg || 1471 (spa_sync_pass(spa) >= zfs_sync_pass_deferred_free && 1472 !spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP)) || 1473 brt_maybe_exists(spa, bp)) { 1474 metaslab_check_free(spa, bp); 1475 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp); 1476 } else { 1477 VERIFY3P(zio_free_sync(NULL, spa, txg, bp, 0), ==, NULL); 1478 } 1479 } 1480 1481 /* 1482 * To improve performance, this function may return NULL if we were able 1483 * to do the free immediately. This avoids the cost of creating a zio 1484 * (and linking it to the parent, etc). 1485 */ 1486 zio_t * 1487 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 1488 zio_flag_t flags) 1489 { 1490 ASSERT(!BP_IS_HOLE(bp)); 1491 ASSERT(spa_syncing_txg(spa) == txg); 1492 1493 if (BP_IS_EMBEDDED(bp)) 1494 return (NULL); 1495 1496 metaslab_check_free(spa, bp); 1497 arc_freed(spa, bp); 1498 dsl_scan_freed(spa, bp); 1499 1500 if (BP_IS_GANG(bp) || 1501 BP_GET_DEDUP(bp) || 1502 brt_maybe_exists(spa, bp)) { 1503 /* 1504 * GANG, DEDUP and BRT blocks can induce a read (for the gang 1505 * block header, the DDT or the BRT), so issue them 1506 * asynchronously so that this thread is not tied up. 1507 */ 1508 enum zio_stage stage = 1509 ZIO_FREE_PIPELINE | ZIO_STAGE_ISSUE_ASYNC; 1510 1511 return (zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), 1512 BP_GET_PSIZE(bp), NULL, NULL, 1513 ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, 1514 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage)); 1515 } else { 1516 metaslab_free(spa, bp, txg, B_FALSE); 1517 return (NULL); 1518 } 1519 } 1520 1521 zio_t * 1522 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 1523 zio_done_func_t *done, void *private, zio_flag_t flags) 1524 { 1525 zio_t *zio; 1526 1527 (void) zfs_blkptr_verify(spa, bp, (flags & ZIO_FLAG_CONFIG_WRITER) ? 1528 BLK_CONFIG_HELD : BLK_CONFIG_NEEDED, BLK_VERIFY_HALT); 1529 1530 if (BP_IS_EMBEDDED(bp)) 1531 return (zio_null(pio, spa, NULL, NULL, NULL, 0)); 1532 1533 /* 1534 * A claim is an allocation of a specific block. Claims are needed 1535 * to support immediate writes in the intent log. The issue is that 1536 * immediate writes contain committed data, but in a txg that was 1537 * *not* committed. Upon opening the pool after an unclean shutdown, 1538 * the intent log claims all blocks that contain immediate write data 1539 * so that the SPA knows they're in use. 1540 * 1541 * All claims *must* be resolved in the first txg -- before the SPA 1542 * starts allocating blocks -- so that nothing is allocated twice. 1543 * If txg == 0 we just verify that the block is claimable. 1544 */ 1545 ASSERT3U(BP_GET_LOGICAL_BIRTH(&spa->spa_uberblock.ub_rootbp), <, 1546 spa_min_claim_txg(spa)); 1547 ASSERT(txg == spa_min_claim_txg(spa) || txg == 0); 1548 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(8) */ 1549 1550 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), 1551 BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, 1552 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE); 1553 ASSERT0(zio->io_queued_timestamp); 1554 1555 return (zio); 1556 } 1557 1558 zio_t * 1559 zio_trim(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 1560 zio_done_func_t *done, void *private, zio_priority_t priority, 1561 zio_flag_t flags, enum trim_flag trim_flags) 1562 { 1563 zio_t *zio; 1564 1565 ASSERT0(vd->vdev_children); 1566 ASSERT0(P2PHASE(offset, 1ULL << vd->vdev_ashift)); 1567 ASSERT0(P2PHASE(size, 1ULL << vd->vdev_ashift)); 1568 ASSERT3U(size, !=, 0); 1569 1570 zio = zio_create(pio, vd->vdev_spa, 0, NULL, NULL, size, size, done, 1571 private, ZIO_TYPE_TRIM, priority, flags | ZIO_FLAG_PHYSICAL, 1572 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_TRIM_PIPELINE); 1573 zio->io_trim_flags = trim_flags; 1574 1575 return (zio); 1576 } 1577 1578 zio_t * 1579 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 1580 abd_t *data, int checksum, zio_done_func_t *done, void *private, 1581 zio_priority_t priority, zio_flag_t flags, boolean_t labels) 1582 { 1583 zio_t *zio; 1584 1585 ASSERT(vd->vdev_children == 0); 1586 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || 1587 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); 1588 ASSERT3U(offset + size, <=, vd->vdev_psize); 1589 1590 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done, 1591 private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, 1592 offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE); 1593 1594 zio->io_prop.zp_checksum = checksum; 1595 1596 return (zio); 1597 } 1598 1599 zio_t * 1600 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 1601 abd_t *data, int checksum, zio_done_func_t *done, void *private, 1602 zio_priority_t priority, zio_flag_t flags, boolean_t labels) 1603 { 1604 zio_t *zio; 1605 1606 ASSERT(vd->vdev_children == 0); 1607 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || 1608 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); 1609 ASSERT3U(offset + size, <=, vd->vdev_psize); 1610 1611 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done, 1612 private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, 1613 offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE); 1614 1615 zio->io_prop.zp_checksum = checksum; 1616 1617 if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) { 1618 /* 1619 * zec checksums are necessarily destructive -- they modify 1620 * the end of the write buffer to hold the verifier/checksum. 1621 * Therefore, we must make a local copy in case the data is 1622 * being written to multiple places in parallel. 1623 */ 1624 abd_t *wbuf = abd_alloc_sametype(data, size); 1625 abd_copy(wbuf, data, size); 1626 1627 zio_push_transform(zio, wbuf, size, size, NULL); 1628 } 1629 1630 return (zio); 1631 } 1632 1633 /* 1634 * Create a child I/O to do some work for us. 1635 */ 1636 zio_t * 1637 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset, 1638 abd_t *data, uint64_t size, int type, zio_priority_t priority, 1639 zio_flag_t flags, zio_done_func_t *done, void *private) 1640 { 1641 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE; 1642 zio_t *zio; 1643 1644 /* 1645 * vdev child I/Os do not propagate their error to the parent. 1646 * Therefore, for correct operation the caller *must* check for 1647 * and handle the error in the child i/o's done callback. 1648 * The only exceptions are i/os that we don't care about 1649 * (OPTIONAL or REPAIR). 1650 */ 1651 ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) || 1652 done != NULL); 1653 1654 if (type == ZIO_TYPE_READ && bp != NULL) { 1655 /* 1656 * If we have the bp, then the child should perform the 1657 * checksum and the parent need not. This pushes error 1658 * detection as close to the leaves as possible and 1659 * eliminates redundant checksums in the interior nodes. 1660 */ 1661 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY; 1662 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; 1663 /* 1664 * We never allow the mirror VDEV to attempt reading from any 1665 * additional data copies after the first Direct I/O checksum 1666 * verify failure. This is to avoid bad data being written out 1667 * through the mirror during self healing. See comment in 1668 * vdev_mirror_io_done() for more details. 1669 */ 1670 ASSERT0(pio->io_flags & ZIO_FLAG_DIO_CHKSUM_ERR); 1671 } else if (type == ZIO_TYPE_WRITE && 1672 pio->io_prop.zp_direct_write == B_TRUE) { 1673 /* 1674 * By default we only will verify checksums for Direct I/O 1675 * writes for Linux. FreeBSD is able to place user pages under 1676 * write protection before issuing them to the ZIO pipeline. 1677 * 1678 * Checksum validation errors will only be reported through 1679 * the top-level VDEV, which is set by this child ZIO. 1680 */ 1681 ASSERT3P(bp, !=, NULL); 1682 ASSERT3U(pio->io_child_type, ==, ZIO_CHILD_LOGICAL); 1683 pipeline |= ZIO_STAGE_DIO_CHECKSUM_VERIFY; 1684 } 1685 1686 if (vd->vdev_ops->vdev_op_leaf) { 1687 ASSERT0(vd->vdev_children); 1688 offset += VDEV_LABEL_START_SIZE; 1689 } 1690 1691 flags |= ZIO_VDEV_CHILD_FLAGS(pio); 1692 1693 /* 1694 * If we've decided to do a repair, the write is not speculative -- 1695 * even if the original read was. 1696 */ 1697 if (flags & ZIO_FLAG_IO_REPAIR) 1698 flags &= ~ZIO_FLAG_SPECULATIVE; 1699 1700 /* 1701 * If we're creating a child I/O that is not associated with a 1702 * top-level vdev, then the child zio is not an allocating I/O. 1703 * If this is a retried I/O then we ignore it since we will 1704 * have already processed the original allocating I/O. 1705 */ 1706 if (flags & ZIO_FLAG_IO_ALLOCATING && 1707 (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) { 1708 ASSERT(pio->io_metaslab_class != NULL); 1709 ASSERT(pio->io_metaslab_class->mc_alloc_throttle_enabled); 1710 ASSERT(type == ZIO_TYPE_WRITE); 1711 ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE); 1712 ASSERT(!(flags & ZIO_FLAG_IO_REPAIR)); 1713 ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) || 1714 pio->io_child_type == ZIO_CHILD_GANG); 1715 1716 flags &= ~ZIO_FLAG_IO_ALLOCATING; 1717 } 1718 1719 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size, 1720 done, private, type, priority, flags, vd, offset, &pio->io_bookmark, 1721 ZIO_STAGE_VDEV_IO_START >> 1, pipeline); 1722 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV); 1723 1724 return (zio); 1725 } 1726 1727 zio_t * 1728 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size, 1729 zio_type_t type, zio_priority_t priority, zio_flag_t flags, 1730 zio_done_func_t *done, void *private) 1731 { 1732 zio_t *zio; 1733 1734 ASSERT(vd->vdev_ops->vdev_op_leaf); 1735 1736 zio = zio_create(NULL, vd->vdev_spa, 0, NULL, 1737 data, size, size, done, private, type, priority, 1738 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED, 1739 vd, offset, NULL, 1740 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE); 1741 1742 return (zio); 1743 } 1744 1745 1746 /* 1747 * Send a flush command to the given vdev. Unlike most zio creation functions, 1748 * the flush zios are issued immediately. You can wait on pio to pause until 1749 * the flushes complete. 1750 */ 1751 void 1752 zio_flush(zio_t *pio, vdev_t *vd) 1753 { 1754 const zio_flag_t flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | 1755 ZIO_FLAG_DONT_RETRY; 1756 1757 if (vd->vdev_nowritecache) 1758 return; 1759 1760 if (vd->vdev_children == 0) { 1761 zio_nowait(zio_create(pio, vd->vdev_spa, 0, NULL, NULL, 0, 0, 1762 NULL, NULL, ZIO_TYPE_FLUSH, ZIO_PRIORITY_NOW, flags, vd, 0, 1763 NULL, ZIO_STAGE_OPEN, ZIO_FLUSH_PIPELINE)); 1764 } else { 1765 for (uint64_t c = 0; c < vd->vdev_children; c++) 1766 zio_flush(pio, vd->vdev_child[c]); 1767 } 1768 } 1769 1770 void 1771 zio_shrink(zio_t *zio, uint64_t size) 1772 { 1773 ASSERT3P(zio->io_executor, ==, NULL); 1774 ASSERT3U(zio->io_orig_size, ==, zio->io_size); 1775 ASSERT3U(size, <=, zio->io_size); 1776 1777 /* 1778 * We don't shrink for raidz because of problems with the 1779 * reconstruction when reading back less than the block size. 1780 * Note, BP_IS_RAIDZ() assumes no compression. 1781 */ 1782 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 1783 if (!BP_IS_RAIDZ(zio->io_bp)) { 1784 /* we are not doing a raw write */ 1785 ASSERT3U(zio->io_size, ==, zio->io_lsize); 1786 zio->io_orig_size = zio->io_size = zio->io_lsize = size; 1787 } 1788 } 1789 1790 /* 1791 * Round provided allocation size up to a value that can be allocated 1792 * by at least some vdev(s) in the pool with minimum or no additional 1793 * padding and without extra space usage on others 1794 */ 1795 static uint64_t 1796 zio_roundup_alloc_size(spa_t *spa, uint64_t size) 1797 { 1798 if (size > spa->spa_min_alloc) 1799 return (roundup(size, spa->spa_gcd_alloc)); 1800 return (spa->spa_min_alloc); 1801 } 1802 1803 size_t 1804 zio_get_compression_max_size(enum zio_compress compress, uint64_t gcd_alloc, 1805 uint64_t min_alloc, size_t s_len) 1806 { 1807 size_t d_len; 1808 1809 /* minimum 12.5% must be saved (legacy value, may be changed later) */ 1810 d_len = s_len - (s_len >> 3); 1811 1812 /* ZLE can't use exactly d_len bytes, it needs more, so ignore it */ 1813 if (compress == ZIO_COMPRESS_ZLE) 1814 return (d_len); 1815 1816 d_len = d_len - d_len % gcd_alloc; 1817 1818 if (d_len < min_alloc) 1819 return (BPE_PAYLOAD_SIZE); 1820 return (d_len); 1821 } 1822 1823 /* 1824 * ========================================================================== 1825 * Prepare to read and write logical blocks 1826 * ========================================================================== 1827 */ 1828 1829 static zio_t * 1830 zio_read_bp_init(zio_t *zio) 1831 { 1832 blkptr_t *bp = zio->io_bp; 1833 uint64_t psize = 1834 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp); 1835 1836 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy); 1837 1838 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF && 1839 zio->io_child_type == ZIO_CHILD_LOGICAL && 1840 !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) { 1841 zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize), 1842 psize, psize, zio_decompress); 1843 } 1844 1845 if (((BP_IS_PROTECTED(bp) && !(zio->io_flags & ZIO_FLAG_RAW_ENCRYPT)) || 1846 BP_HAS_INDIRECT_MAC_CKSUM(bp)) && 1847 zio->io_child_type == ZIO_CHILD_LOGICAL) { 1848 zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize), 1849 psize, psize, zio_decrypt); 1850 } 1851 1852 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) { 1853 int psize = BPE_GET_PSIZE(bp); 1854 void *data = abd_borrow_buf(zio->io_abd, psize); 1855 1856 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1857 decode_embedded_bp_compressed(bp, data); 1858 abd_return_buf_copy(zio->io_abd, data, psize); 1859 } else { 1860 ASSERT(!BP_IS_EMBEDDED(bp)); 1861 } 1862 1863 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL) 1864 zio->io_pipeline = ZIO_DDT_READ_PIPELINE; 1865 1866 return (zio); 1867 } 1868 1869 static zio_t * 1870 zio_write_bp_init(zio_t *zio) 1871 { 1872 if (!IO_IS_ALLOCATING(zio)) 1873 return (zio); 1874 1875 ASSERT(zio->io_child_type != ZIO_CHILD_DDT); 1876 1877 if (zio->io_bp_override) { 1878 blkptr_t *bp = zio->io_bp; 1879 zio_prop_t *zp = &zio->io_prop; 1880 1881 ASSERT(BP_GET_LOGICAL_BIRTH(bp) != zio->io_txg); 1882 1883 *bp = *zio->io_bp_override; 1884 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1885 1886 if (zp->zp_brtwrite) 1887 return (zio); 1888 1889 ASSERT(!BP_GET_DEDUP(zio->io_bp_override)); 1890 1891 if (BP_IS_EMBEDDED(bp)) 1892 return (zio); 1893 1894 /* 1895 * If we've been overridden and nopwrite is set then 1896 * set the flag accordingly to indicate that a nopwrite 1897 * has already occurred. 1898 */ 1899 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) { 1900 ASSERT(!zp->zp_dedup); 1901 ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum); 1902 zio->io_flags |= ZIO_FLAG_NOPWRITE; 1903 return (zio); 1904 } 1905 1906 ASSERT(!zp->zp_nopwrite); 1907 1908 if (BP_IS_HOLE(bp) || !zp->zp_dedup) 1909 return (zio); 1910 1911 ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags & 1912 ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify); 1913 1914 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum && 1915 !zp->zp_encrypt) { 1916 BP_SET_DEDUP(bp, 1); 1917 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE; 1918 return (zio); 1919 } 1920 1921 /* 1922 * We were unable to handle this as an override bp, treat 1923 * it as a regular write I/O. 1924 */ 1925 zio->io_bp_override = NULL; 1926 *bp = zio->io_bp_orig; 1927 zio->io_pipeline = zio->io_orig_pipeline; 1928 } 1929 1930 return (zio); 1931 } 1932 1933 static zio_t * 1934 zio_write_compress(zio_t *zio) 1935 { 1936 spa_t *spa = zio->io_spa; 1937 zio_prop_t *zp = &zio->io_prop; 1938 enum zio_compress compress = zp->zp_compress; 1939 blkptr_t *bp = zio->io_bp; 1940 uint64_t lsize = zio->io_lsize; 1941 uint64_t psize = zio->io_size; 1942 uint32_t pass = 1; 1943 1944 /* 1945 * If our children haven't all reached the ready stage, 1946 * wait for them and then repeat this pipeline stage. 1947 */ 1948 if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT | 1949 ZIO_CHILD_GANG_BIT, ZIO_WAIT_READY)) { 1950 return (NULL); 1951 } 1952 1953 if (!IO_IS_ALLOCATING(zio)) 1954 return (zio); 1955 1956 if (zio->io_children_ready != NULL) { 1957 /* 1958 * Now that all our children are ready, run the callback 1959 * associated with this zio in case it wants to modify the 1960 * data to be written. 1961 */ 1962 ASSERT3U(zp->zp_level, >, 0); 1963 zio->io_children_ready(zio); 1964 } 1965 1966 ASSERT(zio->io_child_type != ZIO_CHILD_DDT); 1967 ASSERT(zio->io_bp_override == NULL); 1968 1969 if (!BP_IS_HOLE(bp) && BP_GET_LOGICAL_BIRTH(bp) == zio->io_txg) { 1970 /* 1971 * We're rewriting an existing block, which means we're 1972 * working on behalf of spa_sync(). For spa_sync() to 1973 * converge, it must eventually be the case that we don't 1974 * have to allocate new blocks. But compression changes 1975 * the blocksize, which forces a reallocate, and makes 1976 * convergence take longer. Therefore, after the first 1977 * few passes, stop compressing to ensure convergence. 1978 */ 1979 pass = spa_sync_pass(spa); 1980 1981 ASSERT(zio->io_txg == spa_syncing_txg(spa)); 1982 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1983 ASSERT(!BP_GET_DEDUP(bp)); 1984 1985 if (pass >= zfs_sync_pass_dont_compress) 1986 compress = ZIO_COMPRESS_OFF; 1987 1988 /* Make sure someone doesn't change their mind on overwrites */ 1989 ASSERT(BP_IS_EMBEDDED(bp) || BP_IS_GANG(bp) || 1990 MIN(zp->zp_copies, spa_max_replication(spa)) 1991 == BP_GET_NDVAS(bp)); 1992 } 1993 1994 /* If it's a compressed write that is not raw, compress the buffer. */ 1995 if (compress != ZIO_COMPRESS_OFF && 1996 !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) { 1997 abd_t *cabd = NULL; 1998 if (abd_cmp_zero(zio->io_abd, lsize) == 0) 1999 psize = 0; 2000 else if (compress == ZIO_COMPRESS_EMPTY) 2001 psize = lsize; 2002 else 2003 psize = zio_compress_data(compress, zio->io_abd, &cabd, 2004 lsize, 2005 zio_get_compression_max_size(compress, 2006 spa->spa_gcd_alloc, spa->spa_min_alloc, lsize), 2007 zp->zp_complevel); 2008 if (psize == 0) { 2009 compress = ZIO_COMPRESS_OFF; 2010 } else if (psize >= lsize) { 2011 compress = ZIO_COMPRESS_OFF; 2012 if (cabd != NULL) 2013 abd_free(cabd); 2014 } else if (psize <= BPE_PAYLOAD_SIZE && !zp->zp_encrypt && 2015 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) && 2016 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) { 2017 void *cbuf = abd_borrow_buf_copy(cabd, lsize); 2018 encode_embedded_bp_compressed(bp, 2019 cbuf, compress, lsize, psize); 2020 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA); 2021 BP_SET_TYPE(bp, zio->io_prop.zp_type); 2022 BP_SET_LEVEL(bp, zio->io_prop.zp_level); 2023 abd_return_buf(cabd, cbuf, lsize); 2024 abd_free(cabd); 2025 BP_SET_LOGICAL_BIRTH(bp, zio->io_txg); 2026 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2027 ASSERT(spa_feature_is_active(spa, 2028 SPA_FEATURE_EMBEDDED_DATA)); 2029 return (zio); 2030 } else { 2031 /* 2032 * Round compressed size up to the minimum allocation 2033 * size of the smallest-ashift device, and zero the 2034 * tail. This ensures that the compressed size of the 2035 * BP (and thus compressratio property) are correct, 2036 * in that we charge for the padding used to fill out 2037 * the last sector. 2038 */ 2039 size_t rounded = (size_t)zio_roundup_alloc_size(spa, 2040 psize); 2041 if (rounded >= lsize) { 2042 compress = ZIO_COMPRESS_OFF; 2043 abd_free(cabd); 2044 psize = lsize; 2045 } else { 2046 abd_zero_off(cabd, psize, rounded - psize); 2047 psize = rounded; 2048 zio_push_transform(zio, cabd, 2049 psize, lsize, NULL); 2050 } 2051 } 2052 2053 /* 2054 * We were unable to handle this as an override bp, treat 2055 * it as a regular write I/O. 2056 */ 2057 zio->io_bp_override = NULL; 2058 *bp = zio->io_bp_orig; 2059 zio->io_pipeline = zio->io_orig_pipeline; 2060 2061 } else if ((zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) != 0 && 2062 zp->zp_type == DMU_OT_DNODE) { 2063 /* 2064 * The DMU actually relies on the zio layer's compression 2065 * to free metadnode blocks that have had all contained 2066 * dnodes freed. As a result, even when doing a raw 2067 * receive, we must check whether the block can be compressed 2068 * to a hole. 2069 */ 2070 if (abd_cmp_zero(zio->io_abd, lsize) == 0) { 2071 psize = 0; 2072 compress = ZIO_COMPRESS_OFF; 2073 } else { 2074 psize = lsize; 2075 } 2076 } else if (zio->io_flags & ZIO_FLAG_RAW_COMPRESS && 2077 !(zio->io_flags & ZIO_FLAG_RAW_ENCRYPT)) { 2078 /* 2079 * If we are raw receiving an encrypted dataset we should not 2080 * take this codepath because it will change the on-disk block 2081 * and decryption will fail. 2082 */ 2083 size_t rounded = MIN((size_t)zio_roundup_alloc_size(spa, psize), 2084 lsize); 2085 2086 if (rounded != psize) { 2087 abd_t *cdata = abd_alloc_linear(rounded, B_TRUE); 2088 abd_zero_off(cdata, psize, rounded - psize); 2089 abd_copy_off(cdata, zio->io_abd, 0, 0, psize); 2090 psize = rounded; 2091 zio_push_transform(zio, cdata, 2092 psize, rounded, NULL); 2093 } 2094 } else { 2095 ASSERT3U(psize, !=, 0); 2096 } 2097 2098 /* 2099 * The final pass of spa_sync() must be all rewrites, but the first 2100 * few passes offer a trade-off: allocating blocks defers convergence, 2101 * but newly allocated blocks are sequential, so they can be written 2102 * to disk faster. Therefore, we allow the first few passes of 2103 * spa_sync() to allocate new blocks, but force rewrites after that. 2104 * There should only be a handful of blocks after pass 1 in any case. 2105 */ 2106 if (!BP_IS_HOLE(bp) && BP_GET_LOGICAL_BIRTH(bp) == zio->io_txg && 2107 BP_GET_PSIZE(bp) == psize && 2108 pass >= zfs_sync_pass_rewrite) { 2109 VERIFY3U(psize, !=, 0); 2110 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES; 2111 2112 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages; 2113 zio->io_flags |= ZIO_FLAG_IO_REWRITE; 2114 } else { 2115 BP_ZERO(bp); 2116 zio->io_pipeline = ZIO_WRITE_PIPELINE; 2117 } 2118 2119 if (psize == 0) { 2120 if (BP_GET_LOGICAL_BIRTH(&zio->io_bp_orig) != 0 && 2121 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) { 2122 BP_SET_LSIZE(bp, lsize); 2123 BP_SET_TYPE(bp, zp->zp_type); 2124 BP_SET_LEVEL(bp, zp->zp_level); 2125 BP_SET_BIRTH(bp, zio->io_txg, 0); 2126 } 2127 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2128 } else { 2129 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER); 2130 BP_SET_LSIZE(bp, lsize); 2131 BP_SET_TYPE(bp, zp->zp_type); 2132 BP_SET_LEVEL(bp, zp->zp_level); 2133 BP_SET_PSIZE(bp, psize); 2134 BP_SET_COMPRESS(bp, compress); 2135 BP_SET_CHECKSUM(bp, zp->zp_checksum); 2136 BP_SET_DEDUP(bp, zp->zp_dedup); 2137 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER); 2138 if (zp->zp_dedup) { 2139 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2140 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 2141 ASSERT(!zp->zp_encrypt || 2142 DMU_OT_IS_ENCRYPTED(zp->zp_type)); 2143 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE; 2144 } 2145 if (zp->zp_nopwrite) { 2146 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2147 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 2148 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE; 2149 } 2150 } 2151 return (zio); 2152 } 2153 2154 static zio_t * 2155 zio_free_bp_init(zio_t *zio) 2156 { 2157 blkptr_t *bp = zio->io_bp; 2158 2159 if (zio->io_child_type == ZIO_CHILD_LOGICAL) { 2160 if (BP_GET_DEDUP(bp)) 2161 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE; 2162 } 2163 2164 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy); 2165 2166 return (zio); 2167 } 2168 2169 /* 2170 * ========================================================================== 2171 * Execute the I/O pipeline 2172 * ========================================================================== 2173 */ 2174 2175 static void 2176 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline) 2177 { 2178 spa_t *spa = zio->io_spa; 2179 zio_type_t t = zio->io_type; 2180 2181 /* 2182 * If we're a config writer or a probe, the normal issue and 2183 * interrupt threads may all be blocked waiting for the config lock. 2184 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL. 2185 */ 2186 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE)) 2187 t = ZIO_TYPE_NULL; 2188 2189 /* 2190 * A similar issue exists for the L2ARC write thread until L2ARC 2.0. 2191 */ 2192 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux) 2193 t = ZIO_TYPE_NULL; 2194 2195 /* 2196 * If this is a high priority I/O, then use the high priority taskq if 2197 * available or cut the line otherwise. 2198 */ 2199 if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE) { 2200 if (spa->spa_zio_taskq[t][q + 1].stqs_count != 0) 2201 q++; 2202 else 2203 cutinline = B_TRUE; 2204 } 2205 2206 ASSERT3U(q, <, ZIO_TASKQ_TYPES); 2207 2208 spa_taskq_dispatch(spa, t, q, zio_execute, zio, cutinline); 2209 } 2210 2211 static boolean_t 2212 zio_taskq_member(zio_t *zio, zio_taskq_type_t q) 2213 { 2214 spa_t *spa = zio->io_spa; 2215 2216 taskq_t *tq = taskq_of_curthread(); 2217 2218 for (zio_type_t t = 0; t < ZIO_TYPES; t++) { 2219 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q]; 2220 uint_t i; 2221 for (i = 0; i < tqs->stqs_count; i++) { 2222 if (tqs->stqs_taskq[i] == tq) 2223 return (B_TRUE); 2224 } 2225 } 2226 2227 return (B_FALSE); 2228 } 2229 2230 static zio_t * 2231 zio_issue_async(zio_t *zio) 2232 { 2233 ASSERT((zio->io_type != ZIO_TYPE_WRITE) || ZIO_HAS_ALLOCATOR(zio)); 2234 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); 2235 return (NULL); 2236 } 2237 2238 void 2239 zio_interrupt(void *zio) 2240 { 2241 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE); 2242 } 2243 2244 void 2245 zio_delay_interrupt(zio_t *zio) 2246 { 2247 /* 2248 * The timeout_generic() function isn't defined in userspace, so 2249 * rather than trying to implement the function, the zio delay 2250 * functionality has been disabled for userspace builds. 2251 */ 2252 2253 #ifdef _KERNEL 2254 /* 2255 * If io_target_timestamp is zero, then no delay has been registered 2256 * for this IO, thus jump to the end of this function and "skip" the 2257 * delay; issuing it directly to the zio layer. 2258 */ 2259 if (zio->io_target_timestamp != 0) { 2260 hrtime_t now = gethrtime(); 2261 2262 if (now >= zio->io_target_timestamp) { 2263 /* 2264 * This IO has already taken longer than the target 2265 * delay to complete, so we don't want to delay it 2266 * any longer; we "miss" the delay and issue it 2267 * directly to the zio layer. This is likely due to 2268 * the target latency being set to a value less than 2269 * the underlying hardware can satisfy (e.g. delay 2270 * set to 1ms, but the disks take 10ms to complete an 2271 * IO request). 2272 */ 2273 2274 DTRACE_PROBE2(zio__delay__miss, zio_t *, zio, 2275 hrtime_t, now); 2276 2277 zio_interrupt(zio); 2278 } else { 2279 taskqid_t tid; 2280 hrtime_t diff = zio->io_target_timestamp - now; 2281 int ticks = MAX(1, NSEC_TO_TICK(diff)); 2282 clock_t expire_at_tick = ddi_get_lbolt() + ticks; 2283 2284 DTRACE_PROBE3(zio__delay__hit, zio_t *, zio, 2285 hrtime_t, now, hrtime_t, diff); 2286 2287 tid = taskq_dispatch_delay(system_taskq, zio_interrupt, 2288 zio, TQ_NOSLEEP, expire_at_tick); 2289 if (tid == TASKQID_INVALID) { 2290 /* 2291 * Couldn't allocate a task. Just finish the 2292 * zio without a delay. 2293 */ 2294 zio_interrupt(zio); 2295 } 2296 } 2297 return; 2298 } 2299 #endif 2300 DTRACE_PROBE1(zio__delay__skip, zio_t *, zio); 2301 zio_interrupt(zio); 2302 } 2303 2304 static void 2305 zio_deadman_impl(zio_t *pio, int ziodepth) 2306 { 2307 zio_t *cio, *cio_next; 2308 zio_link_t *zl = NULL; 2309 vdev_t *vd = pio->io_vd; 2310 2311 if (zio_deadman_log_all || (vd != NULL && vd->vdev_ops->vdev_op_leaf)) { 2312 vdev_queue_t *vq = vd ? &vd->vdev_queue : NULL; 2313 zbookmark_phys_t *zb = &pio->io_bookmark; 2314 uint64_t delta = gethrtime() - pio->io_timestamp; 2315 uint64_t failmode = spa_get_deadman_failmode(pio->io_spa); 2316 2317 zfs_dbgmsg("slow zio[%d]: zio=%px timestamp=%llu " 2318 "delta=%llu queued=%llu io=%llu " 2319 "path=%s " 2320 "last=%llu type=%d " 2321 "priority=%d flags=0x%llx stage=0x%x " 2322 "pipeline=0x%x pipeline-trace=0x%x " 2323 "objset=%llu object=%llu " 2324 "level=%llu blkid=%llu " 2325 "offset=%llu size=%llu " 2326 "error=%d", 2327 ziodepth, pio, pio->io_timestamp, 2328 (u_longlong_t)delta, pio->io_delta, pio->io_delay, 2329 vd ? vd->vdev_path : "NULL", 2330 vq ? vq->vq_io_complete_ts : 0, pio->io_type, 2331 pio->io_priority, (u_longlong_t)pio->io_flags, 2332 pio->io_stage, pio->io_pipeline, pio->io_pipeline_trace, 2333 (u_longlong_t)zb->zb_objset, (u_longlong_t)zb->zb_object, 2334 (u_longlong_t)zb->zb_level, (u_longlong_t)zb->zb_blkid, 2335 (u_longlong_t)pio->io_offset, (u_longlong_t)pio->io_size, 2336 pio->io_error); 2337 (void) zfs_ereport_post(FM_EREPORT_ZFS_DEADMAN, 2338 pio->io_spa, vd, zb, pio, 0); 2339 2340 if (failmode == ZIO_FAILURE_MODE_CONTINUE && 2341 taskq_empty_ent(&pio->io_tqent)) { 2342 zio_interrupt(pio); 2343 } 2344 } 2345 2346 mutex_enter(&pio->io_lock); 2347 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) { 2348 cio_next = zio_walk_children(pio, &zl); 2349 zio_deadman_impl(cio, ziodepth + 1); 2350 } 2351 mutex_exit(&pio->io_lock); 2352 } 2353 2354 /* 2355 * Log the critical information describing this zio and all of its children 2356 * using the zfs_dbgmsg() interface then post deadman event for the ZED. 2357 */ 2358 void 2359 zio_deadman(zio_t *pio, const char *tag) 2360 { 2361 spa_t *spa = pio->io_spa; 2362 char *name = spa_name(spa); 2363 2364 if (!zfs_deadman_enabled || spa_suspended(spa)) 2365 return; 2366 2367 zio_deadman_impl(pio, 0); 2368 2369 switch (spa_get_deadman_failmode(spa)) { 2370 case ZIO_FAILURE_MODE_WAIT: 2371 zfs_dbgmsg("%s waiting for hung I/O to pool '%s'", tag, name); 2372 break; 2373 2374 case ZIO_FAILURE_MODE_CONTINUE: 2375 zfs_dbgmsg("%s restarting hung I/O for pool '%s'", tag, name); 2376 break; 2377 2378 case ZIO_FAILURE_MODE_PANIC: 2379 fm_panic("%s determined I/O to pool '%s' is hung.", tag, name); 2380 break; 2381 } 2382 } 2383 2384 /* 2385 * Execute the I/O pipeline until one of the following occurs: 2386 * (1) the I/O completes; (2) the pipeline stalls waiting for 2387 * dependent child I/Os; (3) the I/O issues, so we're waiting 2388 * for an I/O completion interrupt; (4) the I/O is delegated by 2389 * vdev-level caching or aggregation; (5) the I/O is deferred 2390 * due to vdev-level queueing; (6) the I/O is handed off to 2391 * another thread. In all cases, the pipeline stops whenever 2392 * there's no CPU work; it never burns a thread in cv_wait_io(). 2393 * 2394 * There's no locking on io_stage because there's no legitimate way 2395 * for multiple threads to be attempting to process the same I/O. 2396 */ 2397 static zio_pipe_stage_t *zio_pipeline[]; 2398 2399 /* 2400 * zio_execute() is a wrapper around the static function 2401 * __zio_execute() so that we can force __zio_execute() to be 2402 * inlined. This reduces stack overhead which is important 2403 * because __zio_execute() is called recursively in several zio 2404 * code paths. zio_execute() itself cannot be inlined because 2405 * it is externally visible. 2406 */ 2407 void 2408 zio_execute(void *zio) 2409 { 2410 fstrans_cookie_t cookie; 2411 2412 cookie = spl_fstrans_mark(); 2413 __zio_execute(zio); 2414 spl_fstrans_unmark(cookie); 2415 } 2416 2417 /* 2418 * Used to determine if in the current context the stack is sized large 2419 * enough to allow zio_execute() to be called recursively. A minimum 2420 * stack size of 16K is required to avoid needing to re-dispatch the zio. 2421 */ 2422 static boolean_t 2423 zio_execute_stack_check(zio_t *zio) 2424 { 2425 #if !defined(HAVE_LARGE_STACKS) 2426 dsl_pool_t *dp = spa_get_dsl(zio->io_spa); 2427 2428 /* Executing in txg_sync_thread() context. */ 2429 if (dp && curthread == dp->dp_tx.tx_sync_thread) 2430 return (B_TRUE); 2431 2432 /* Pool initialization outside of zio_taskq context. */ 2433 if (dp && spa_is_initializing(dp->dp_spa) && 2434 !zio_taskq_member(zio, ZIO_TASKQ_ISSUE) && 2435 !zio_taskq_member(zio, ZIO_TASKQ_ISSUE_HIGH)) 2436 return (B_TRUE); 2437 #else 2438 (void) zio; 2439 #endif /* HAVE_LARGE_STACKS */ 2440 2441 return (B_FALSE); 2442 } 2443 2444 __attribute__((always_inline)) 2445 static inline void 2446 __zio_execute(zio_t *zio) 2447 { 2448 ASSERT3U(zio->io_queued_timestamp, >, 0); 2449 2450 while (zio->io_stage < ZIO_STAGE_DONE) { 2451 enum zio_stage pipeline = zio->io_pipeline; 2452 enum zio_stage stage = zio->io_stage; 2453 2454 zio->io_executor = curthread; 2455 2456 ASSERT(!MUTEX_HELD(&zio->io_lock)); 2457 ASSERT(ISP2(stage)); 2458 ASSERT(zio->io_stall == NULL); 2459 2460 do { 2461 stage <<= 1; 2462 } while ((stage & pipeline) == 0); 2463 2464 ASSERT(stage <= ZIO_STAGE_DONE); 2465 2466 /* 2467 * If we are in interrupt context and this pipeline stage 2468 * will grab a config lock that is held across I/O, 2469 * or may wait for an I/O that needs an interrupt thread 2470 * to complete, issue async to avoid deadlock. 2471 * 2472 * For VDEV_IO_START, we cut in line so that the io will 2473 * be sent to disk promptly. 2474 */ 2475 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL && 2476 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) { 2477 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? 2478 zio_requeue_io_start_cut_in_line : B_FALSE; 2479 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); 2480 return; 2481 } 2482 2483 /* 2484 * If the current context doesn't have large enough stacks 2485 * the zio must be issued asynchronously to prevent overflow. 2486 */ 2487 if (zio_execute_stack_check(zio)) { 2488 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? 2489 zio_requeue_io_start_cut_in_line : B_FALSE; 2490 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); 2491 return; 2492 } 2493 2494 zio->io_stage = stage; 2495 zio->io_pipeline_trace |= zio->io_stage; 2496 2497 /* 2498 * The zio pipeline stage returns the next zio to execute 2499 * (typically the same as this one), or NULL if we should 2500 * stop. 2501 */ 2502 zio = zio_pipeline[highbit64(stage) - 1](zio); 2503 2504 if (zio == NULL) 2505 return; 2506 } 2507 } 2508 2509 2510 /* 2511 * ========================================================================== 2512 * Initiate I/O, either sync or async 2513 * ========================================================================== 2514 */ 2515 int 2516 zio_wait(zio_t *zio) 2517 { 2518 /* 2519 * Some routines, like zio_free_sync(), may return a NULL zio 2520 * to avoid the performance overhead of creating and then destroying 2521 * an unneeded zio. For the callers' simplicity, we accept a NULL 2522 * zio and ignore it. 2523 */ 2524 if (zio == NULL) 2525 return (0); 2526 2527 long timeout = MSEC_TO_TICK(zfs_deadman_ziotime_ms); 2528 int error; 2529 2530 ASSERT3S(zio->io_stage, ==, ZIO_STAGE_OPEN); 2531 ASSERT3P(zio->io_executor, ==, NULL); 2532 2533 zio->io_waiter = curthread; 2534 ASSERT0(zio->io_queued_timestamp); 2535 zio->io_queued_timestamp = gethrtime(); 2536 2537 if (zio->io_type == ZIO_TYPE_WRITE) { 2538 spa_select_allocator(zio); 2539 } 2540 __zio_execute(zio); 2541 2542 mutex_enter(&zio->io_lock); 2543 while (zio->io_executor != NULL) { 2544 error = cv_timedwait_io(&zio->io_cv, &zio->io_lock, 2545 ddi_get_lbolt() + timeout); 2546 2547 if (zfs_deadman_enabled && error == -1 && 2548 gethrtime() - zio->io_queued_timestamp > 2549 spa_deadman_ziotime(zio->io_spa)) { 2550 mutex_exit(&zio->io_lock); 2551 timeout = MSEC_TO_TICK(zfs_deadman_checktime_ms); 2552 zio_deadman(zio, FTAG); 2553 mutex_enter(&zio->io_lock); 2554 } 2555 } 2556 mutex_exit(&zio->io_lock); 2557 2558 error = zio->io_error; 2559 zio_destroy(zio); 2560 2561 return (error); 2562 } 2563 2564 void 2565 zio_nowait(zio_t *zio) 2566 { 2567 /* 2568 * See comment in zio_wait(). 2569 */ 2570 if (zio == NULL) 2571 return; 2572 2573 ASSERT3P(zio->io_executor, ==, NULL); 2574 2575 if (zio->io_child_type == ZIO_CHILD_LOGICAL && 2576 list_is_empty(&zio->io_parent_list)) { 2577 zio_t *pio; 2578 2579 /* 2580 * This is a logical async I/O with no parent to wait for it. 2581 * We add it to the spa_async_root_zio "Godfather" I/O which 2582 * will ensure they complete prior to unloading the pool. 2583 */ 2584 spa_t *spa = zio->io_spa; 2585 pio = spa->spa_async_zio_root[CPU_SEQID_UNSTABLE]; 2586 2587 zio_add_child(pio, zio); 2588 } 2589 2590 ASSERT0(zio->io_queued_timestamp); 2591 zio->io_queued_timestamp = gethrtime(); 2592 if (zio->io_type == ZIO_TYPE_WRITE) { 2593 spa_select_allocator(zio); 2594 } 2595 __zio_execute(zio); 2596 } 2597 2598 /* 2599 * ========================================================================== 2600 * Reexecute, cancel, or suspend/resume failed I/O 2601 * ========================================================================== 2602 */ 2603 2604 static void 2605 zio_reexecute(void *arg) 2606 { 2607 zio_t *pio = arg; 2608 zio_t *cio, *cio_next, *gio; 2609 2610 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL); 2611 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN); 2612 ASSERT(pio->io_gang_leader == NULL); 2613 ASSERT(pio->io_gang_tree == NULL); 2614 2615 mutex_enter(&pio->io_lock); 2616 pio->io_flags = pio->io_orig_flags; 2617 pio->io_stage = pio->io_orig_stage; 2618 pio->io_pipeline = pio->io_orig_pipeline; 2619 pio->io_reexecute = 0; 2620 pio->io_flags |= ZIO_FLAG_REEXECUTED; 2621 pio->io_pipeline_trace = 0; 2622 pio->io_error = 0; 2623 pio->io_state[ZIO_WAIT_READY] = (pio->io_stage >= ZIO_STAGE_READY) || 2624 (pio->io_pipeline & ZIO_STAGE_READY) == 0; 2625 pio->io_state[ZIO_WAIT_DONE] = (pio->io_stage >= ZIO_STAGE_DONE); 2626 2627 /* 2628 * It's possible for a failed ZIO to be a descendant of more than one 2629 * ZIO tree. When reexecuting it, we have to be sure to add its wait 2630 * states to all parent wait counts. 2631 * 2632 * Those parents, in turn, may have other children that are currently 2633 * active, usually because they've already been reexecuted after 2634 * resuming. Those children may be executing and may call 2635 * zio_notify_parent() at the same time as we're updating our parent's 2636 * counts. To avoid races while updating the counts, we take 2637 * gio->io_lock before each update. 2638 */ 2639 zio_link_t *zl = NULL; 2640 while ((gio = zio_walk_parents(pio, &zl)) != NULL) { 2641 mutex_enter(&gio->io_lock); 2642 for (int w = 0; w < ZIO_WAIT_TYPES; w++) { 2643 gio->io_children[pio->io_child_type][w] += 2644 !pio->io_state[w]; 2645 } 2646 mutex_exit(&gio->io_lock); 2647 } 2648 2649 for (int c = 0; c < ZIO_CHILD_TYPES; c++) 2650 pio->io_child_error[c] = 0; 2651 2652 if (IO_IS_ALLOCATING(pio)) 2653 BP_ZERO(pio->io_bp); 2654 2655 /* 2656 * As we reexecute pio's children, new children could be created. 2657 * New children go to the head of pio's io_child_list, however, 2658 * so we will (correctly) not reexecute them. The key is that 2659 * the remainder of pio's io_child_list, from 'cio_next' onward, 2660 * cannot be affected by any side effects of reexecuting 'cio'. 2661 */ 2662 zl = NULL; 2663 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) { 2664 cio_next = zio_walk_children(pio, &zl); 2665 mutex_exit(&pio->io_lock); 2666 zio_reexecute(cio); 2667 mutex_enter(&pio->io_lock); 2668 } 2669 mutex_exit(&pio->io_lock); 2670 2671 /* 2672 * Now that all children have been reexecuted, execute the parent. 2673 * We don't reexecute "The Godfather" I/O here as it's the 2674 * responsibility of the caller to wait on it. 2675 */ 2676 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) { 2677 pio->io_queued_timestamp = gethrtime(); 2678 __zio_execute(pio); 2679 } 2680 } 2681 2682 void 2683 zio_suspend(spa_t *spa, zio_t *zio, zio_suspend_reason_t reason) 2684 { 2685 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC) 2686 fm_panic("Pool '%s' has encountered an uncorrectable I/O " 2687 "failure and the failure mode property for this pool " 2688 "is set to panic.", spa_name(spa)); 2689 2690 if (reason != ZIO_SUSPEND_MMP) { 2691 cmn_err(CE_WARN, "Pool '%s' has encountered an uncorrectable " 2692 "I/O failure and has been suspended.", spa_name(spa)); 2693 } 2694 2695 (void) zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, 2696 NULL, NULL, 0); 2697 2698 mutex_enter(&spa->spa_suspend_lock); 2699 2700 if (spa->spa_suspend_zio_root == NULL) 2701 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL, 2702 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | 2703 ZIO_FLAG_GODFATHER); 2704 2705 spa->spa_suspended = reason; 2706 2707 if (zio != NULL) { 2708 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); 2709 ASSERT(zio != spa->spa_suspend_zio_root); 2710 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2711 ASSERT(zio_unique_parent(zio) == NULL); 2712 ASSERT(zio->io_stage == ZIO_STAGE_DONE); 2713 zio_add_child(spa->spa_suspend_zio_root, zio); 2714 } 2715 2716 mutex_exit(&spa->spa_suspend_lock); 2717 } 2718 2719 int 2720 zio_resume(spa_t *spa) 2721 { 2722 zio_t *pio; 2723 2724 /* 2725 * Reexecute all previously suspended i/o. 2726 */ 2727 mutex_enter(&spa->spa_suspend_lock); 2728 if (spa->spa_suspended != ZIO_SUSPEND_NONE) 2729 cmn_err(CE_WARN, "Pool '%s' was suspended and is being " 2730 "resumed. Failed I/O will be retried.", 2731 spa_name(spa)); 2732 spa->spa_suspended = ZIO_SUSPEND_NONE; 2733 cv_broadcast(&spa->spa_suspend_cv); 2734 pio = spa->spa_suspend_zio_root; 2735 spa->spa_suspend_zio_root = NULL; 2736 mutex_exit(&spa->spa_suspend_lock); 2737 2738 if (pio == NULL) 2739 return (0); 2740 2741 zio_reexecute(pio); 2742 return (zio_wait(pio)); 2743 } 2744 2745 void 2746 zio_resume_wait(spa_t *spa) 2747 { 2748 mutex_enter(&spa->spa_suspend_lock); 2749 while (spa_suspended(spa)) 2750 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock); 2751 mutex_exit(&spa->spa_suspend_lock); 2752 } 2753 2754 /* 2755 * ========================================================================== 2756 * Gang blocks. 2757 * 2758 * A gang block is a collection of small blocks that looks to the DMU 2759 * like one large block. When zio_dva_allocate() cannot find a block 2760 * of the requested size, due to either severe fragmentation or the pool 2761 * being nearly full, it calls zio_write_gang_block() to construct the 2762 * block from smaller fragments. 2763 * 2764 * A gang block consists of a gang header (zio_gbh_phys_t) and up to 2765 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like 2766 * an indirect block: it's an array of block pointers. It consumes 2767 * only one sector and hence is allocatable regardless of fragmentation. 2768 * The gang header's bps point to its gang members, which hold the data. 2769 * 2770 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg> 2771 * as the verifier to ensure uniqueness of the SHA256 checksum. 2772 * Critically, the gang block bp's blk_cksum is the checksum of the data, 2773 * not the gang header. This ensures that data block signatures (needed for 2774 * deduplication) are independent of how the block is physically stored. 2775 * 2776 * Gang blocks can be nested: a gang member may itself be a gang block. 2777 * Thus every gang block is a tree in which root and all interior nodes are 2778 * gang headers, and the leaves are normal blocks that contain user data. 2779 * The root of the gang tree is called the gang leader. 2780 * 2781 * To perform any operation (read, rewrite, free, claim) on a gang block, 2782 * zio_gang_assemble() first assembles the gang tree (minus data leaves) 2783 * in the io_gang_tree field of the original logical i/o by recursively 2784 * reading the gang leader and all gang headers below it. This yields 2785 * an in-core tree containing the contents of every gang header and the 2786 * bps for every constituent of the gang block. 2787 * 2788 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree 2789 * and invokes a callback on each bp. To free a gang block, zio_gang_issue() 2790 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp. 2791 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim(). 2792 * zio_read_gang() is a wrapper around zio_read() that omits reading gang 2793 * headers, since we already have those in io_gang_tree. zio_rewrite_gang() 2794 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite() 2795 * of the gang header plus zio_checksum_compute() of the data to update the 2796 * gang header's blk_cksum as described above. 2797 * 2798 * The two-phase assemble/issue model solves the problem of partial failure -- 2799 * what if you'd freed part of a gang block but then couldn't read the 2800 * gang header for another part? Assembling the entire gang tree first 2801 * ensures that all the necessary gang header I/O has succeeded before 2802 * starting the actual work of free, claim, or write. Once the gang tree 2803 * is assembled, free and claim are in-memory operations that cannot fail. 2804 * 2805 * In the event that a gang write fails, zio_dva_unallocate() walks the 2806 * gang tree to immediately free (i.e. insert back into the space map) 2807 * everything we've allocated. This ensures that we don't get ENOSPC 2808 * errors during repeated suspend/resume cycles due to a flaky device. 2809 * 2810 * Gang rewrites only happen during sync-to-convergence. If we can't assemble 2811 * the gang tree, we won't modify the block, so we can safely defer the free 2812 * (knowing that the block is still intact). If we *can* assemble the gang 2813 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free 2814 * each constituent bp and we can allocate a new block on the next sync pass. 2815 * 2816 * In all cases, the gang tree allows complete recovery from partial failure. 2817 * ========================================================================== 2818 */ 2819 2820 static void 2821 zio_gang_issue_func_done(zio_t *zio) 2822 { 2823 abd_free(zio->io_abd); 2824 } 2825 2826 static zio_t * 2827 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, 2828 uint64_t offset) 2829 { 2830 if (gn != NULL) 2831 return (pio); 2832 2833 return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset), 2834 BP_GET_PSIZE(bp), zio_gang_issue_func_done, 2835 NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), 2836 &pio->io_bookmark)); 2837 } 2838 2839 static zio_t * 2840 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, 2841 uint64_t offset) 2842 { 2843 zio_t *zio; 2844 2845 if (gn != NULL) { 2846 abd_t *gbh_abd = 2847 abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE); 2848 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, 2849 gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL, 2850 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), 2851 &pio->io_bookmark); 2852 /* 2853 * As we rewrite each gang header, the pipeline will compute 2854 * a new gang block header checksum for it; but no one will 2855 * compute a new data checksum, so we do that here. The one 2856 * exception is the gang leader: the pipeline already computed 2857 * its data checksum because that stage precedes gang assembly. 2858 * (Presently, nothing actually uses interior data checksums; 2859 * this is just good hygiene.) 2860 */ 2861 if (gn != pio->io_gang_leader->io_gang_tree) { 2862 abd_t *buf = abd_get_offset(data, offset); 2863 2864 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp), 2865 buf, BP_GET_PSIZE(bp)); 2866 2867 abd_free(buf); 2868 } 2869 /* 2870 * If we are here to damage data for testing purposes, 2871 * leave the GBH alone so that we can detect the damage. 2872 */ 2873 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE) 2874 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; 2875 } else { 2876 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, 2877 abd_get_offset(data, offset), BP_GET_PSIZE(bp), 2878 zio_gang_issue_func_done, NULL, pio->io_priority, 2879 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 2880 } 2881 2882 return (zio); 2883 } 2884 2885 static zio_t * 2886 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, 2887 uint64_t offset) 2888 { 2889 (void) gn, (void) data, (void) offset; 2890 2891 zio_t *zio = zio_free_sync(pio, pio->io_spa, pio->io_txg, bp, 2892 ZIO_GANG_CHILD_FLAGS(pio)); 2893 if (zio == NULL) { 2894 zio = zio_null(pio, pio->io_spa, 2895 NULL, NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)); 2896 } 2897 return (zio); 2898 } 2899 2900 static zio_t * 2901 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, 2902 uint64_t offset) 2903 { 2904 (void) gn, (void) data, (void) offset; 2905 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp, 2906 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio))); 2907 } 2908 2909 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = { 2910 NULL, 2911 zio_read_gang, 2912 zio_rewrite_gang, 2913 zio_free_gang, 2914 zio_claim_gang, 2915 NULL 2916 }; 2917 2918 static void zio_gang_tree_assemble_done(zio_t *zio); 2919 2920 static zio_gang_node_t * 2921 zio_gang_node_alloc(zio_gang_node_t **gnpp) 2922 { 2923 zio_gang_node_t *gn; 2924 2925 ASSERT(*gnpp == NULL); 2926 2927 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP); 2928 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE); 2929 *gnpp = gn; 2930 2931 return (gn); 2932 } 2933 2934 static void 2935 zio_gang_node_free(zio_gang_node_t **gnpp) 2936 { 2937 zio_gang_node_t *gn = *gnpp; 2938 2939 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) 2940 ASSERT(gn->gn_child[g] == NULL); 2941 2942 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE); 2943 kmem_free(gn, sizeof (*gn)); 2944 *gnpp = NULL; 2945 } 2946 2947 static void 2948 zio_gang_tree_free(zio_gang_node_t **gnpp) 2949 { 2950 zio_gang_node_t *gn = *gnpp; 2951 2952 if (gn == NULL) 2953 return; 2954 2955 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) 2956 zio_gang_tree_free(&gn->gn_child[g]); 2957 2958 zio_gang_node_free(gnpp); 2959 } 2960 2961 static void 2962 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp) 2963 { 2964 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp); 2965 abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE); 2966 2967 ASSERT(gio->io_gang_leader == gio); 2968 ASSERT(BP_IS_GANG(bp)); 2969 2970 zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE, 2971 zio_gang_tree_assemble_done, gn, gio->io_priority, 2972 ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark)); 2973 } 2974 2975 static void 2976 zio_gang_tree_assemble_done(zio_t *zio) 2977 { 2978 zio_t *gio = zio->io_gang_leader; 2979 zio_gang_node_t *gn = zio->io_private; 2980 blkptr_t *bp = zio->io_bp; 2981 2982 ASSERT(gio == zio_unique_parent(zio)); 2983 ASSERT(list_is_empty(&zio->io_child_list)); 2984 2985 if (zio->io_error) 2986 return; 2987 2988 /* this ABD was created from a linear buf in zio_gang_tree_assemble */ 2989 if (BP_SHOULD_BYTESWAP(bp)) 2990 byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size); 2991 2992 ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh); 2993 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE); 2994 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); 2995 2996 abd_free(zio->io_abd); 2997 2998 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 2999 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; 3000 if (!BP_IS_GANG(gbp)) 3001 continue; 3002 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]); 3003 } 3004 } 3005 3006 static void 3007 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data, 3008 uint64_t offset) 3009 { 3010 zio_t *gio = pio->io_gang_leader; 3011 zio_t *zio; 3012 3013 ASSERT(BP_IS_GANG(bp) == !!gn); 3014 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp)); 3015 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree); 3016 3017 /* 3018 * If you're a gang header, your data is in gn->gn_gbh. 3019 * If you're a gang member, your data is in 'data' and gn == NULL. 3020 */ 3021 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset); 3022 3023 if (gn != NULL) { 3024 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); 3025 3026 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 3027 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; 3028 if (BP_IS_HOLE(gbp)) 3029 continue; 3030 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data, 3031 offset); 3032 offset += BP_GET_PSIZE(gbp); 3033 } 3034 } 3035 3036 if (gn == gio->io_gang_tree) 3037 ASSERT3U(gio->io_size, ==, offset); 3038 3039 if (zio != pio) 3040 zio_nowait(zio); 3041 } 3042 3043 static zio_t * 3044 zio_gang_assemble(zio_t *zio) 3045 { 3046 blkptr_t *bp = zio->io_bp; 3047 3048 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL); 3049 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 3050 3051 zio->io_gang_leader = zio; 3052 3053 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree); 3054 3055 return (zio); 3056 } 3057 3058 static zio_t * 3059 zio_gang_issue(zio_t *zio) 3060 { 3061 blkptr_t *bp = zio->io_bp; 3062 3063 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT, ZIO_WAIT_DONE)) { 3064 return (NULL); 3065 } 3066 3067 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio); 3068 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 3069 3070 if (zio->io_child_error[ZIO_CHILD_GANG] == 0) 3071 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd, 3072 0); 3073 else 3074 zio_gang_tree_free(&zio->io_gang_tree); 3075 3076 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 3077 3078 return (zio); 3079 } 3080 3081 static void 3082 zio_gang_inherit_allocator(zio_t *pio, zio_t *cio) 3083 { 3084 cio->io_allocator = pio->io_allocator; 3085 } 3086 3087 static void 3088 zio_write_gang_member_ready(zio_t *zio) 3089 { 3090 zio_t *pio = zio_unique_parent(zio); 3091 dva_t *cdva = zio->io_bp->blk_dva; 3092 dva_t *pdva = pio->io_bp->blk_dva; 3093 uint64_t asize; 3094 zio_t *gio __maybe_unused = zio->io_gang_leader; 3095 3096 if (BP_IS_HOLE(zio->io_bp)) 3097 return; 3098 3099 /* 3100 * If we're getting direct-invoked from zio_write_gang_block(), 3101 * the bp_orig will be set. 3102 */ 3103 ASSERT(BP_IS_HOLE(&zio->io_bp_orig) || 3104 zio->io_flags & ZIO_FLAG_PREALLOCATED); 3105 3106 ASSERT(zio->io_child_type == ZIO_CHILD_GANG); 3107 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies); 3108 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp)); 3109 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp)); 3110 VERIFY3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp)); 3111 3112 mutex_enter(&pio->io_lock); 3113 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) { 3114 ASSERT(DVA_GET_GANG(&pdva[d])); 3115 asize = DVA_GET_ASIZE(&pdva[d]); 3116 asize += DVA_GET_ASIZE(&cdva[d]); 3117 DVA_SET_ASIZE(&pdva[d], asize); 3118 } 3119 mutex_exit(&pio->io_lock); 3120 } 3121 3122 static void 3123 zio_write_gang_done(zio_t *zio) 3124 { 3125 /* 3126 * The io_abd field will be NULL for a zio with no data. The io_flags 3127 * will initially have the ZIO_FLAG_NODATA bit flag set, but we can't 3128 * check for it here as it is cleared in zio_ready. 3129 */ 3130 if (zio->io_abd != NULL) 3131 abd_free(zio->io_abd); 3132 } 3133 3134 static zio_t * 3135 zio_write_gang_block(zio_t *pio, metaslab_class_t *mc) 3136 { 3137 spa_t *spa = pio->io_spa; 3138 blkptr_t *bp = pio->io_bp; 3139 zio_t *gio = pio->io_gang_leader; 3140 zio_t *zio; 3141 zio_gang_node_t *gn, **gnpp; 3142 zio_gbh_phys_t *gbh; 3143 abd_t *gbh_abd; 3144 uint64_t txg = pio->io_txg; 3145 uint64_t resid = pio->io_size; 3146 zio_prop_t zp; 3147 int error; 3148 boolean_t has_data = !(pio->io_flags & ZIO_FLAG_NODATA); 3149 3150 /* 3151 * Store multiple copies of the GBH, so that we can still traverse 3152 * all the data (e.g. to free or scrub) even if a block is damaged. 3153 * This value respects the redundant_metadata property. 3154 */ 3155 int gbh_copies = gio->io_prop.zp_gang_copies; 3156 if (gbh_copies == 0) { 3157 /* 3158 * This should only happen in the case where we're filling in 3159 * DDT entries for a parent that wants more copies than the DDT 3160 * has. In that case, we cannot gang without creating a mixed 3161 * blkptr, which is illegal. 3162 */ 3163 ASSERT3U(gio->io_child_type, ==, ZIO_CHILD_DDT); 3164 pio->io_error = EAGAIN; 3165 return (pio); 3166 } 3167 ASSERT3S(gbh_copies, >, 0); 3168 ASSERT3S(gbh_copies, <=, SPA_DVAS_PER_BP); 3169 3170 ASSERT(ZIO_HAS_ALLOCATOR(pio)); 3171 int flags = METASLAB_GANG_HEADER; 3172 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { 3173 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); 3174 ASSERT(has_data); 3175 3176 flags |= METASLAB_ASYNC_ALLOC; 3177 } 3178 3179 error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE, 3180 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags, 3181 &pio->io_alloc_list, pio->io_allocator, pio); 3182 if (error) { 3183 pio->io_error = error; 3184 return (pio); 3185 } 3186 3187 if (pio == gio) { 3188 gnpp = &gio->io_gang_tree; 3189 } else { 3190 gnpp = pio->io_private; 3191 ASSERT(pio->io_ready == zio_write_gang_member_ready); 3192 } 3193 3194 gn = zio_gang_node_alloc(gnpp); 3195 gbh = gn->gn_gbh; 3196 memset(gbh, 0, SPA_GANGBLOCKSIZE); 3197 gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE); 3198 3199 /* 3200 * Create the gang header. 3201 */ 3202 zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE, 3203 zio_write_gang_done, NULL, pio->io_priority, 3204 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 3205 3206 zio_gang_inherit_allocator(pio, zio); 3207 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { 3208 boolean_t more; 3209 VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies, 3210 zio, B_TRUE, &more)); 3211 } 3212 3213 /* 3214 * Create and nowait the gang children. First, we try to do 3215 * opportunistic allocations. If that fails to generate enough 3216 * space, we fall back to normal zio_write calls for nested gang. 3217 */ 3218 for (int g = 0; resid != 0; g++) { 3219 flags &= METASLAB_ASYNC_ALLOC; 3220 flags |= METASLAB_GANG_CHILD; 3221 zp.zp_checksum = gio->io_prop.zp_checksum; 3222 zp.zp_compress = ZIO_COMPRESS_OFF; 3223 zp.zp_complevel = gio->io_prop.zp_complevel; 3224 zp.zp_type = zp.zp_storage_type = DMU_OT_NONE; 3225 zp.zp_level = 0; 3226 zp.zp_copies = gio->io_prop.zp_copies; 3227 zp.zp_gang_copies = gio->io_prop.zp_gang_copies; 3228 zp.zp_dedup = B_FALSE; 3229 zp.zp_dedup_verify = B_FALSE; 3230 zp.zp_nopwrite = B_FALSE; 3231 zp.zp_encrypt = gio->io_prop.zp_encrypt; 3232 zp.zp_byteorder = gio->io_prop.zp_byteorder; 3233 zp.zp_direct_write = B_FALSE; 3234 memset(zp.zp_salt, 0, ZIO_DATA_SALT_LEN); 3235 memset(zp.zp_iv, 0, ZIO_DATA_IV_LEN); 3236 memset(zp.zp_mac, 0, ZIO_DATA_MAC_LEN); 3237 3238 uint64_t min_size = zio_roundup_alloc_size(spa, 3239 resid / (SPA_GBH_NBLKPTRS - g)); 3240 min_size = MIN(min_size, resid); 3241 bp = &gbh->zg_blkptr[g]; 3242 3243 zio_alloc_list_t cio_list; 3244 metaslab_trace_init(&cio_list); 3245 uint64_t allocated_size = UINT64_MAX; 3246 error = metaslab_alloc_range(spa, mc, min_size, resid, 3247 bp, gio->io_prop.zp_copies, txg, NULL, 3248 flags, &cio_list, zio->io_allocator, NULL, &allocated_size); 3249 3250 boolean_t allocated = error == 0; 3251 3252 uint64_t psize = allocated ? MIN(resid, allocated_size) : 3253 min_size; 3254 3255 zio_t *cio = zio_write(zio, spa, txg, bp, has_data ? 3256 abd_get_offset(pio->io_abd, pio->io_size - resid) : NULL, 3257 psize, psize, &zp, zio_write_gang_member_ready, NULL, 3258 zio_write_gang_done, &gn->gn_child[g], pio->io_priority, 3259 ZIO_GANG_CHILD_FLAGS(pio) | 3260 (allocated ? ZIO_FLAG_PREALLOCATED : 0), &pio->io_bookmark); 3261 3262 resid -= psize; 3263 zio_gang_inherit_allocator(zio, cio); 3264 if (allocated) { 3265 metaslab_trace_move(&cio_list, &cio->io_alloc_list); 3266 metaslab_group_alloc_increment_all(spa, 3267 &cio->io_bp_orig, zio->io_allocator, flags, psize, 3268 cio); 3269 } 3270 /* 3271 * We do not reserve for the child writes, since we already 3272 * reserved for the parent. Unreserve though will be called 3273 * for individual children. We can do this since sum of all 3274 * child's physical sizes is equal to parent's physical size. 3275 * It would not work for potentially bigger allocation sizes. 3276 */ 3277 3278 zio_nowait(cio); 3279 } 3280 3281 /* 3282 * Set pio's pipeline to just wait for zio to finish. 3283 */ 3284 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 3285 3286 zio_nowait(zio); 3287 3288 return (pio); 3289 } 3290 3291 /* 3292 * The zio_nop_write stage in the pipeline determines if allocating a 3293 * new bp is necessary. The nopwrite feature can handle writes in 3294 * either syncing or open context (i.e. zil writes) and as a result is 3295 * mutually exclusive with dedup. 3296 * 3297 * By leveraging a cryptographically secure checksum, such as SHA256, we 3298 * can compare the checksums of the new data and the old to determine if 3299 * allocating a new block is required. Note that our requirements for 3300 * cryptographic strength are fairly weak: there can't be any accidental 3301 * hash collisions, but we don't need to be secure against intentional 3302 * (malicious) collisions. To trigger a nopwrite, you have to be able 3303 * to write the file to begin with, and triggering an incorrect (hash 3304 * collision) nopwrite is no worse than simply writing to the file. 3305 * That said, there are no known attacks against the checksum algorithms 3306 * used for nopwrite, assuming that the salt and the checksums 3307 * themselves remain secret. 3308 */ 3309 static zio_t * 3310 zio_nop_write(zio_t *zio) 3311 { 3312 blkptr_t *bp = zio->io_bp; 3313 blkptr_t *bp_orig = &zio->io_bp_orig; 3314 zio_prop_t *zp = &zio->io_prop; 3315 3316 ASSERT(BP_IS_HOLE(bp)); 3317 ASSERT(BP_GET_LEVEL(bp) == 0); 3318 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 3319 ASSERT(zp->zp_nopwrite); 3320 ASSERT(!zp->zp_dedup); 3321 ASSERT(zio->io_bp_override == NULL); 3322 ASSERT(IO_IS_ALLOCATING(zio)); 3323 3324 /* 3325 * Check to see if the original bp and the new bp have matching 3326 * characteristics (i.e. same checksum, compression algorithms, etc). 3327 * If they don't then just continue with the pipeline which will 3328 * allocate a new bp. 3329 */ 3330 if (BP_IS_HOLE(bp_orig) || 3331 !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags & 3332 ZCHECKSUM_FLAG_NOPWRITE) || 3333 BP_IS_ENCRYPTED(bp) || BP_IS_ENCRYPTED(bp_orig) || 3334 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) || 3335 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) || 3336 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) || 3337 zp->zp_copies != BP_GET_NDVAS(bp_orig)) 3338 return (zio); 3339 3340 /* 3341 * If the checksums match then reset the pipeline so that we 3342 * avoid allocating a new bp and issuing any I/O. 3343 */ 3344 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) { 3345 ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags & 3346 ZCHECKSUM_FLAG_NOPWRITE); 3347 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig)); 3348 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig)); 3349 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF); 3350 ASSERT3U(bp->blk_prop, ==, bp_orig->blk_prop); 3351 3352 /* 3353 * If we're overwriting a block that is currently on an 3354 * indirect vdev, then ignore the nopwrite request and 3355 * allow a new block to be allocated on a concrete vdev. 3356 */ 3357 spa_config_enter(zio->io_spa, SCL_VDEV, FTAG, RW_READER); 3358 for (int d = 0; d < BP_GET_NDVAS(bp_orig); d++) { 3359 vdev_t *tvd = vdev_lookup_top(zio->io_spa, 3360 DVA_GET_VDEV(&bp_orig->blk_dva[d])); 3361 if (tvd->vdev_ops == &vdev_indirect_ops) { 3362 spa_config_exit(zio->io_spa, SCL_VDEV, FTAG); 3363 return (zio); 3364 } 3365 } 3366 spa_config_exit(zio->io_spa, SCL_VDEV, FTAG); 3367 3368 *bp = *bp_orig; 3369 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 3370 zio->io_flags |= ZIO_FLAG_NOPWRITE; 3371 } 3372 3373 return (zio); 3374 } 3375 3376 /* 3377 * ========================================================================== 3378 * Block Reference Table 3379 * ========================================================================== 3380 */ 3381 static zio_t * 3382 zio_brt_free(zio_t *zio) 3383 { 3384 blkptr_t *bp; 3385 3386 bp = zio->io_bp; 3387 3388 if (BP_GET_LEVEL(bp) > 0 || 3389 BP_IS_METADATA(bp) || 3390 !brt_maybe_exists(zio->io_spa, bp)) { 3391 return (zio); 3392 } 3393 3394 if (!brt_entry_decref(zio->io_spa, bp)) { 3395 /* 3396 * This isn't the last reference, so we cannot free 3397 * the data yet. 3398 */ 3399 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 3400 } 3401 3402 return (zio); 3403 } 3404 3405 /* 3406 * ========================================================================== 3407 * Dedup 3408 * ========================================================================== 3409 */ 3410 static void 3411 zio_ddt_child_read_done(zio_t *zio) 3412 { 3413 blkptr_t *bp = zio->io_bp; 3414 ddt_t *ddt; 3415 ddt_entry_t *dde = zio->io_private; 3416 zio_t *pio = zio_unique_parent(zio); 3417 3418 mutex_enter(&pio->io_lock); 3419 ddt = ddt_select(zio->io_spa, bp); 3420 3421 if (zio->io_error == 0) { 3422 ddt_phys_variant_t v = ddt_phys_select(ddt, dde, bp); 3423 /* this phys variant doesn't need repair */ 3424 ddt_phys_clear(dde->dde_phys, v); 3425 } 3426 3427 if (zio->io_error == 0 && dde->dde_io->dde_repair_abd == NULL) 3428 dde->dde_io->dde_repair_abd = zio->io_abd; 3429 else 3430 abd_free(zio->io_abd); 3431 mutex_exit(&pio->io_lock); 3432 } 3433 3434 static zio_t * 3435 zio_ddt_read_start(zio_t *zio) 3436 { 3437 blkptr_t *bp = zio->io_bp; 3438 3439 ASSERT(BP_GET_DEDUP(bp)); 3440 ASSERT(BP_GET_PSIZE(bp) == zio->io_size); 3441 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 3442 3443 if (zio->io_child_error[ZIO_CHILD_DDT]) { 3444 ddt_t *ddt = ddt_select(zio->io_spa, bp); 3445 ddt_entry_t *dde = ddt_repair_start(ddt, bp); 3446 ddt_phys_variant_t v_self = ddt_phys_select(ddt, dde, bp); 3447 ddt_univ_phys_t *ddp = dde->dde_phys; 3448 blkptr_t blk; 3449 3450 ASSERT(zio->io_vsd == NULL); 3451 zio->io_vsd = dde; 3452 3453 if (v_self == DDT_PHYS_NONE) 3454 return (zio); 3455 3456 /* issue I/O for the other copies */ 3457 for (int p = 0; p < DDT_NPHYS(ddt); p++) { 3458 ddt_phys_variant_t v = DDT_PHYS_VARIANT(ddt, p); 3459 3460 if (ddt_phys_birth(ddp, v) == 0 || v == v_self) 3461 continue; 3462 3463 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, 3464 ddp, v, &blk); 3465 zio_nowait(zio_read(zio, zio->io_spa, &blk, 3466 abd_alloc_for_io(zio->io_size, B_TRUE), 3467 zio->io_size, zio_ddt_child_read_done, dde, 3468 zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) | 3469 ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark)); 3470 } 3471 return (zio); 3472 } 3473 3474 zio_nowait(zio_read(zio, zio->io_spa, bp, 3475 zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority, 3476 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark)); 3477 3478 return (zio); 3479 } 3480 3481 static zio_t * 3482 zio_ddt_read_done(zio_t *zio) 3483 { 3484 blkptr_t *bp = zio->io_bp; 3485 3486 if (zio_wait_for_children(zio, ZIO_CHILD_DDT_BIT, ZIO_WAIT_DONE)) { 3487 return (NULL); 3488 } 3489 3490 ASSERT(BP_GET_DEDUP(bp)); 3491 ASSERT(BP_GET_PSIZE(bp) == zio->io_size); 3492 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 3493 3494 if (zio->io_child_error[ZIO_CHILD_DDT]) { 3495 ddt_t *ddt = ddt_select(zio->io_spa, bp); 3496 ddt_entry_t *dde = zio->io_vsd; 3497 if (ddt == NULL) { 3498 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE); 3499 return (zio); 3500 } 3501 if (dde == NULL) { 3502 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1; 3503 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); 3504 return (NULL); 3505 } 3506 if (dde->dde_io->dde_repair_abd != NULL) { 3507 abd_copy(zio->io_abd, dde->dde_io->dde_repair_abd, 3508 zio->io_size); 3509 zio->io_child_error[ZIO_CHILD_DDT] = 0; 3510 } 3511 ddt_repair_done(ddt, dde); 3512 zio->io_vsd = NULL; 3513 } 3514 3515 ASSERT(zio->io_vsd == NULL); 3516 3517 return (zio); 3518 } 3519 3520 static boolean_t 3521 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde) 3522 { 3523 spa_t *spa = zio->io_spa; 3524 boolean_t do_raw = !!(zio->io_flags & ZIO_FLAG_RAW); 3525 3526 ASSERT(!(zio->io_bp_override && do_raw)); 3527 3528 /* 3529 * Note: we compare the original data, not the transformed data, 3530 * because when zio->io_bp is an override bp, we will not have 3531 * pushed the I/O transforms. That's an important optimization 3532 * because otherwise we'd compress/encrypt all dmu_sync() data twice. 3533 * However, we should never get a raw, override zio so in these 3534 * cases we can compare the io_abd directly. This is useful because 3535 * it allows us to do dedup verification even if we don't have access 3536 * to the original data (for instance, if the encryption keys aren't 3537 * loaded). 3538 */ 3539 3540 for (int p = 0; p < DDT_NPHYS(ddt); p++) { 3541 if (DDT_PHYS_IS_DITTO(ddt, p)) 3542 continue; 3543 3544 if (dde->dde_io == NULL) 3545 continue; 3546 3547 zio_t *lio = dde->dde_io->dde_lead_zio[p]; 3548 if (lio == NULL) 3549 continue; 3550 3551 if (do_raw) 3552 return (lio->io_size != zio->io_size || 3553 abd_cmp(zio->io_abd, lio->io_abd) != 0); 3554 3555 return (lio->io_orig_size != zio->io_orig_size || 3556 abd_cmp(zio->io_orig_abd, lio->io_orig_abd) != 0); 3557 } 3558 3559 for (int p = 0; p < DDT_NPHYS(ddt); p++) { 3560 ddt_phys_variant_t v = DDT_PHYS_VARIANT(ddt, p); 3561 uint64_t phys_birth = ddt_phys_birth(dde->dde_phys, v); 3562 3563 if (phys_birth != 0 && do_raw) { 3564 blkptr_t blk = *zio->io_bp; 3565 uint64_t psize; 3566 abd_t *tmpabd; 3567 int error; 3568 3569 ddt_bp_fill(dde->dde_phys, v, &blk, phys_birth); 3570 psize = BP_GET_PSIZE(&blk); 3571 3572 if (psize != zio->io_size) 3573 return (B_TRUE); 3574 3575 ddt_exit(ddt); 3576 3577 tmpabd = abd_alloc_for_io(psize, B_TRUE); 3578 3579 error = zio_wait(zio_read(NULL, spa, &blk, tmpabd, 3580 psize, NULL, NULL, ZIO_PRIORITY_SYNC_READ, 3581 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | 3582 ZIO_FLAG_RAW, &zio->io_bookmark)); 3583 3584 if (error == 0) { 3585 if (abd_cmp(tmpabd, zio->io_abd) != 0) 3586 error = SET_ERROR(ENOENT); 3587 } 3588 3589 abd_free(tmpabd); 3590 ddt_enter(ddt); 3591 return (error != 0); 3592 } else if (phys_birth != 0) { 3593 arc_buf_t *abuf = NULL; 3594 arc_flags_t aflags = ARC_FLAG_WAIT; 3595 blkptr_t blk = *zio->io_bp; 3596 int error; 3597 3598 ddt_bp_fill(dde->dde_phys, v, &blk, phys_birth); 3599 3600 if (BP_GET_LSIZE(&blk) != zio->io_orig_size) 3601 return (B_TRUE); 3602 3603 ddt_exit(ddt); 3604 3605 error = arc_read(NULL, spa, &blk, 3606 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ, 3607 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 3608 &aflags, &zio->io_bookmark); 3609 3610 if (error == 0) { 3611 if (abd_cmp_buf(zio->io_orig_abd, abuf->b_data, 3612 zio->io_orig_size) != 0) 3613 error = SET_ERROR(ENOENT); 3614 arc_buf_destroy(abuf, &abuf); 3615 } 3616 3617 ddt_enter(ddt); 3618 return (error != 0); 3619 } 3620 } 3621 3622 return (B_FALSE); 3623 } 3624 3625 static void 3626 zio_ddt_child_write_done(zio_t *zio) 3627 { 3628 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); 3629 ddt_entry_t *dde = zio->io_private; 3630 3631 zio_link_t *zl = NULL; 3632 ASSERT3P(zio_walk_parents(zio, &zl), !=, NULL); 3633 3634 int p = DDT_PHYS_FOR_COPIES(ddt, zio->io_prop.zp_copies); 3635 ddt_phys_variant_t v = DDT_PHYS_VARIANT(ddt, p); 3636 ddt_univ_phys_t *ddp = dde->dde_phys; 3637 3638 ddt_enter(ddt); 3639 3640 /* we're the lead, so once we're done there's no one else outstanding */ 3641 if (dde->dde_io->dde_lead_zio[p] == zio) 3642 dde->dde_io->dde_lead_zio[p] = NULL; 3643 3644 ddt_univ_phys_t *orig = &dde->dde_io->dde_orig_phys; 3645 3646 if (zio->io_error != 0) { 3647 /* 3648 * The write failed, so we're about to abort the entire IO 3649 * chain. We need to revert the entry back to what it was at 3650 * the last time it was successfully extended. 3651 */ 3652 ddt_phys_unextend(ddp, orig, v); 3653 ddt_phys_clear(orig, v); 3654 3655 ddt_exit(ddt); 3656 return; 3657 } 3658 3659 /* 3660 * Add references for all dedup writes that were waiting on the 3661 * physical one, skipping any other physical writes that are waiting. 3662 */ 3663 zio_t *pio; 3664 zl = NULL; 3665 while ((pio = zio_walk_parents(zio, &zl)) != NULL) { 3666 if (!(pio->io_flags & ZIO_FLAG_DDT_CHILD)) 3667 ddt_phys_addref(ddp, v); 3668 } 3669 3670 /* 3671 * We've successfully added new DVAs to the entry. Clear the saved 3672 * state or, if there's still outstanding IO, remember it so we can 3673 * revert to a known good state if that IO fails. 3674 */ 3675 if (dde->dde_io->dde_lead_zio[p] == NULL) 3676 ddt_phys_clear(orig, v); 3677 else 3678 ddt_phys_copy(orig, ddp, v); 3679 3680 ddt_exit(ddt); 3681 } 3682 3683 static void 3684 zio_ddt_child_write_ready(zio_t *zio) 3685 { 3686 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); 3687 ddt_entry_t *dde = zio->io_private; 3688 3689 zio_link_t *zl = NULL; 3690 ASSERT3P(zio_walk_parents(zio, &zl), !=, NULL); 3691 3692 int p = DDT_PHYS_FOR_COPIES(ddt, zio->io_prop.zp_copies); 3693 ddt_phys_variant_t v = DDT_PHYS_VARIANT(ddt, p); 3694 3695 if (ddt_phys_is_gang(dde->dde_phys, v)) { 3696 for (int i = 0; i < BP_GET_NDVAS(zio->io_bp); i++) { 3697 dva_t *d = &zio->io_bp->blk_dva[i]; 3698 metaslab_group_alloc_decrement(zio->io_spa, 3699 DVA_GET_VDEV(d), zio->io_allocator, 3700 METASLAB_ASYNC_ALLOC, zio->io_size, zio); 3701 } 3702 zio->io_error = EAGAIN; 3703 } 3704 3705 if (zio->io_error != 0) 3706 return; 3707 3708 ddt_enter(ddt); 3709 3710 ddt_phys_extend(dde->dde_phys, v, zio->io_bp); 3711 3712 zio_t *pio; 3713 zl = NULL; 3714 while ((pio = zio_walk_parents(zio, &zl)) != NULL) { 3715 if (!(pio->io_flags & ZIO_FLAG_DDT_CHILD)) 3716 ddt_bp_fill(dde->dde_phys, v, pio->io_bp, zio->io_txg); 3717 } 3718 3719 ddt_exit(ddt); 3720 } 3721 3722 static zio_t * 3723 zio_ddt_write(zio_t *zio) 3724 { 3725 spa_t *spa = zio->io_spa; 3726 blkptr_t *bp = zio->io_bp; 3727 uint64_t txg = zio->io_txg; 3728 zio_prop_t *zp = &zio->io_prop; 3729 ddt_t *ddt = ddt_select(spa, bp); 3730 ddt_entry_t *dde; 3731 3732 ASSERT(BP_GET_DEDUP(bp)); 3733 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum); 3734 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override); 3735 ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW))); 3736 /* 3737 * Deduplication will not take place for Direct I/O writes. The 3738 * ddt_tree will be emptied in syncing context. Direct I/O writes take 3739 * place in the open-context. Direct I/O write can not attempt to 3740 * modify the ddt_tree while issuing out a write. 3741 */ 3742 ASSERT3B(zio->io_prop.zp_direct_write, ==, B_FALSE); 3743 3744 ddt_enter(ddt); 3745 /* 3746 * Search DDT for matching entry. Skip DVAs verification here, since 3747 * they can go only from override, and once we get here the override 3748 * pointer can't have "D" flag to be confused with pruned DDT entries. 3749 */ 3750 IMPLY(zio->io_bp_override, !BP_GET_DEDUP(zio->io_bp_override)); 3751 dde = ddt_lookup(ddt, bp, B_FALSE); 3752 if (dde == NULL) { 3753 /* DDT size is over its quota so no new entries */ 3754 zp->zp_dedup = B_FALSE; 3755 BP_SET_DEDUP(bp, B_FALSE); 3756 if (zio->io_bp_override == NULL) 3757 zio->io_pipeline = ZIO_WRITE_PIPELINE; 3758 ddt_exit(ddt); 3759 return (zio); 3760 } 3761 3762 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) { 3763 /* 3764 * If we're using a weak checksum, upgrade to a strong checksum 3765 * and try again. If we're already using a strong checksum, 3766 * we can't resolve it, so just convert to an ordinary write. 3767 * (And automatically e-mail a paper to Nature?) 3768 */ 3769 if (!(zio_checksum_table[zp->zp_checksum].ci_flags & 3770 ZCHECKSUM_FLAG_DEDUP)) { 3771 zp->zp_checksum = spa_dedup_checksum(spa); 3772 zio_pop_transforms(zio); 3773 zio->io_stage = ZIO_STAGE_OPEN; 3774 BP_ZERO(bp); 3775 } else { 3776 zp->zp_dedup = B_FALSE; 3777 BP_SET_DEDUP(bp, B_FALSE); 3778 } 3779 ASSERT(!BP_GET_DEDUP(bp)); 3780 zio->io_pipeline = ZIO_WRITE_PIPELINE; 3781 ddt_exit(ddt); 3782 return (zio); 3783 } 3784 3785 int p = DDT_PHYS_FOR_COPIES(ddt, zp->zp_copies); 3786 ddt_phys_variant_t v = DDT_PHYS_VARIANT(ddt, p); 3787 ddt_univ_phys_t *ddp = dde->dde_phys; 3788 3789 /* 3790 * In the common cases, at this point we have a regular BP with no 3791 * allocated DVAs, and the corresponding DDT entry for its checksum. 3792 * Our goal is to fill the BP with enough DVAs to satisfy its copies= 3793 * requirement. 3794 * 3795 * One of three things needs to happen to fulfill this: 3796 * 3797 * - if the DDT entry has enough DVAs to satisfy the BP, we just copy 3798 * them out of the entry and return; 3799 * 3800 * - if the DDT entry has no DVAs (ie its brand new), then we have to 3801 * issue the write as normal so that DVAs can be allocated and the 3802 * data land on disk. We then copy the DVAs into the DDT entry on 3803 * return. 3804 * 3805 * - if the DDT entry has some DVAs, but too few, we have to issue the 3806 * write, adjusted to have allocate fewer copies. When it returns, we 3807 * add the new DVAs to the DDT entry, and update the BP to have the 3808 * full amount it originally requested. 3809 * 3810 * In all cases, if there's already a writing IO in flight, we need to 3811 * defer the action until after the write is done. If our action is to 3812 * write, we need to adjust our request for additional DVAs to match 3813 * what will be in the DDT entry after it completes. In this way every 3814 * IO can be guaranteed to recieve enough DVAs simply by joining the 3815 * end of the chain and letting the sequence play out. 3816 */ 3817 3818 /* 3819 * Number of DVAs in the DDT entry. If the BP is encrypted we ignore 3820 * the third one as normal. 3821 */ 3822 int have_dvas = ddt_phys_dva_count(ddp, v, BP_IS_ENCRYPTED(bp)); 3823 IMPLY(have_dvas == 0, ddt_phys_birth(ddp, v) == 0); 3824 boolean_t is_ganged = ddt_phys_is_gang(ddp, v); 3825 3826 /* Number of DVAs requested by the IO. */ 3827 uint8_t need_dvas = zp->zp_copies; 3828 /* Number of DVAs in outstanding writes for this dde. */ 3829 uint8_t parent_dvas = 0; 3830 3831 /* 3832 * What we do next depends on whether or not there's IO outstanding that 3833 * will update this entry. 3834 */ 3835 if (dde->dde_io == NULL || dde->dde_io->dde_lead_zio[p] == NULL) { 3836 /* 3837 * No IO outstanding, so we only need to worry about ourselves. 3838 */ 3839 3840 /* 3841 * Override BPs bring their own DVAs and their own problems. 3842 */ 3843 if (zio->io_bp_override) { 3844 /* 3845 * For a brand-new entry, all the work has been done 3846 * for us, and we can just fill it out from the provided 3847 * block and leave. 3848 */ 3849 if (have_dvas == 0) { 3850 ASSERT(BP_GET_LOGICAL_BIRTH(bp) == txg); 3851 ASSERT(BP_EQUAL(bp, zio->io_bp_override)); 3852 ddt_phys_extend(ddp, v, bp); 3853 ddt_phys_addref(ddp, v); 3854 ddt_exit(ddt); 3855 return (zio); 3856 } 3857 3858 /* 3859 * If we already have this entry, then we want to treat 3860 * it like a regular write. To do this we just wipe 3861 * them out and proceed like a regular write. 3862 * 3863 * Even if there are some DVAs in the entry, we still 3864 * have to clear them out. We can't use them to fill 3865 * out the dedup entry, as they are all referenced 3866 * together by a bp already on disk, and will be freed 3867 * as a group. 3868 */ 3869 BP_ZERO_DVAS(bp); 3870 BP_SET_BIRTH(bp, 0, 0); 3871 } 3872 3873 /* 3874 * If there are enough DVAs in the entry to service our request, 3875 * then we can just use them as-is. 3876 */ 3877 if (have_dvas >= need_dvas) { 3878 ddt_bp_fill(ddp, v, bp, txg); 3879 ddt_phys_addref(ddp, v); 3880 ddt_exit(ddt); 3881 return (zio); 3882 } 3883 3884 /* 3885 * Otherwise, we have to issue IO to fill the entry up to the 3886 * amount we need. 3887 */ 3888 need_dvas -= have_dvas; 3889 } else { 3890 /* 3891 * There's a write in-flight. If there's already enough DVAs on 3892 * the entry, then either there were already enough to start 3893 * with, or the in-flight IO is between READY and DONE, and so 3894 * has extended the entry with new DVAs. Either way, we don't 3895 * need to do anything, we can just slot in behind it. 3896 */ 3897 3898 if (zio->io_bp_override) { 3899 /* 3900 * If there's a write out, then we're soon going to 3901 * have our own copies of this block, so clear out the 3902 * override block and treat it as a regular dedup 3903 * write. See comment above. 3904 */ 3905 BP_ZERO_DVAS(bp); 3906 BP_SET_BIRTH(bp, 0, 0); 3907 } 3908 3909 if (have_dvas >= need_dvas) { 3910 /* 3911 * A minor point: there might already be enough 3912 * committed DVAs in the entry to service our request, 3913 * but we don't know which are completed and which are 3914 * allocated but not yet written. In this case, should 3915 * the IO for the new DVAs fail, we will be on the end 3916 * of the IO chain and will also recieve an error, even 3917 * though our request could have been serviced. 3918 * 3919 * This is an extremely rare case, as it requires the 3920 * original block to be copied with a request for a 3921 * larger number of DVAs, then copied again requesting 3922 * the same (or already fulfilled) number of DVAs while 3923 * the first request is active, and then that first 3924 * request errors. In return, the logic required to 3925 * catch and handle it is complex. For now, I'm just 3926 * not going to bother with it. 3927 */ 3928 3929 /* 3930 * We always fill the bp here as we may have arrived 3931 * after the in-flight write has passed READY, and so 3932 * missed out. 3933 */ 3934 ddt_bp_fill(ddp, v, bp, txg); 3935 zio_add_child(zio, dde->dde_io->dde_lead_zio[p]); 3936 ddt_exit(ddt); 3937 return (zio); 3938 } 3939 3940 /* 3941 * There's not enough in the entry yet, so we need to look at 3942 * the write in-flight and see how many DVAs it will have once 3943 * it completes. 3944 * 3945 * The in-flight write has potentially had its copies request 3946 * reduced (if we're filling out an existing entry), so we need 3947 * to reach in and get the original write to find out what it is 3948 * expecting. 3949 * 3950 * Note that the parent of the lead zio will always have the 3951 * highest zp_copies of any zio in the chain, because ones that 3952 * can be serviced without additional IO are always added to 3953 * the back of the chain. 3954 */ 3955 zio_link_t *zl = NULL; 3956 zio_t *pio = 3957 zio_walk_parents(dde->dde_io->dde_lead_zio[p], &zl); 3958 ASSERT(pio); 3959 parent_dvas = pio->io_prop.zp_copies; 3960 3961 if (parent_dvas >= need_dvas) { 3962 zio_add_child(zio, dde->dde_io->dde_lead_zio[p]); 3963 ddt_exit(ddt); 3964 return (zio); 3965 } 3966 3967 /* 3968 * Still not enough, so we will need to issue to get the 3969 * shortfall. 3970 */ 3971 need_dvas -= parent_dvas; 3972 } 3973 3974 if (is_ganged) { 3975 zp->zp_dedup = B_FALSE; 3976 BP_SET_DEDUP(bp, B_FALSE); 3977 zio->io_pipeline = ZIO_WRITE_PIPELINE; 3978 ddt_exit(ddt); 3979 return (zio); 3980 } 3981 3982 /* 3983 * We need to write. We will create a new write with the copies 3984 * property adjusted to match the number of DVAs we need to need to 3985 * grow the DDT entry by to satisfy the request. 3986 */ 3987 zio_prop_t czp = *zp; 3988 if (have_dvas > 0 || parent_dvas > 0) { 3989 czp.zp_copies = need_dvas; 3990 czp.zp_gang_copies = 0; 3991 } else { 3992 ASSERT3U(czp.zp_copies, ==, need_dvas); 3993 } 3994 3995 zio_t *cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd, 3996 zio->io_orig_size, zio->io_orig_size, &czp, 3997 zio_ddt_child_write_ready, NULL, 3998 zio_ddt_child_write_done, dde, zio->io_priority, 3999 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); 4000 4001 zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL); 4002 4003 /* 4004 * We are the new lead zio, because our parent has the highest 4005 * zp_copies that has been requested for this entry so far. 4006 */ 4007 ddt_alloc_entry_io(dde); 4008 if (dde->dde_io->dde_lead_zio[p] == NULL) { 4009 /* 4010 * First time out, take a copy of the stable entry to revert 4011 * to if there's an error (see zio_ddt_child_write_done()) 4012 */ 4013 ddt_phys_copy(&dde->dde_io->dde_orig_phys, dde->dde_phys, v); 4014 } else { 4015 /* 4016 * Make the existing chain our child, because it cannot 4017 * complete until we have. 4018 */ 4019 zio_add_child(cio, dde->dde_io->dde_lead_zio[p]); 4020 } 4021 dde->dde_io->dde_lead_zio[p] = cio; 4022 4023 ddt_exit(ddt); 4024 4025 zio_nowait(cio); 4026 4027 return (zio); 4028 } 4029 4030 static ddt_entry_t *freedde; /* for debugging */ 4031 4032 static zio_t * 4033 zio_ddt_free(zio_t *zio) 4034 { 4035 spa_t *spa = zio->io_spa; 4036 blkptr_t *bp = zio->io_bp; 4037 ddt_t *ddt = ddt_select(spa, bp); 4038 ddt_entry_t *dde = NULL; 4039 4040 ASSERT(BP_GET_DEDUP(bp)); 4041 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 4042 4043 ddt_enter(ddt); 4044 freedde = dde = ddt_lookup(ddt, bp, B_TRUE); 4045 if (dde) { 4046 ddt_phys_variant_t v = ddt_phys_select(ddt, dde, bp); 4047 if (v != DDT_PHYS_NONE) 4048 ddt_phys_decref(dde->dde_phys, v); 4049 } 4050 ddt_exit(ddt); 4051 4052 /* 4053 * When no entry was found, it must have been pruned, 4054 * so we can free it now instead of decrementing the 4055 * refcount in the DDT. 4056 */ 4057 if (!dde) { 4058 BP_SET_DEDUP(bp, 0); 4059 zio->io_pipeline |= ZIO_STAGE_DVA_FREE; 4060 } 4061 4062 return (zio); 4063 } 4064 4065 /* 4066 * ========================================================================== 4067 * Allocate and free blocks 4068 * ========================================================================== 4069 */ 4070 4071 static zio_t * 4072 zio_io_to_allocate(metaslab_class_allocator_t *mca, boolean_t *more) 4073 { 4074 zio_t *zio; 4075 4076 ASSERT(MUTEX_HELD(&mca->mca_lock)); 4077 4078 zio = avl_first(&mca->mca_tree); 4079 if (zio == NULL) { 4080 *more = B_FALSE; 4081 return (NULL); 4082 } 4083 4084 ASSERT(IO_IS_ALLOCATING(zio)); 4085 ASSERT(ZIO_HAS_ALLOCATOR(zio)); 4086 4087 /* 4088 * Try to place a reservation for this zio. If we're unable to 4089 * reserve then we throttle. 4090 */ 4091 if (!metaslab_class_throttle_reserve(zio->io_metaslab_class, 4092 zio->io_prop.zp_copies, zio, B_FALSE, more)) { 4093 return (NULL); 4094 } 4095 4096 avl_remove(&mca->mca_tree, zio); 4097 ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE); 4098 4099 if (avl_is_empty(&mca->mca_tree)) 4100 *more = B_FALSE; 4101 return (zio); 4102 } 4103 4104 static zio_t * 4105 zio_dva_throttle(zio_t *zio) 4106 { 4107 spa_t *spa = zio->io_spa; 4108 zio_t *nio; 4109 metaslab_class_t *mc; 4110 boolean_t more; 4111 4112 /* 4113 * If not already chosen, choose an appropriate allocation class. 4114 */ 4115 mc = zio->io_metaslab_class; 4116 if (mc == NULL) 4117 mc = spa_preferred_class(spa, zio); 4118 4119 if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE || 4120 !mc->mc_alloc_throttle_enabled || 4121 zio->io_child_type == ZIO_CHILD_GANG || 4122 zio->io_flags & ZIO_FLAG_NODATA) { 4123 return (zio); 4124 } 4125 4126 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 4127 ASSERT(ZIO_HAS_ALLOCATOR(zio)); 4128 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 4129 ASSERT3U(zio->io_queued_timestamp, >, 0); 4130 ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE); 4131 4132 zio->io_metaslab_class = mc; 4133 metaslab_class_allocator_t *mca = &mc->mc_allocator[zio->io_allocator]; 4134 mutex_enter(&mca->mca_lock); 4135 avl_add(&mca->mca_tree, zio); 4136 nio = zio_io_to_allocate(mca, &more); 4137 mutex_exit(&mca->mca_lock); 4138 return (nio); 4139 } 4140 4141 static void 4142 zio_allocate_dispatch(metaslab_class_t *mc, int allocator) 4143 { 4144 metaslab_class_allocator_t *mca = &mc->mc_allocator[allocator]; 4145 zio_t *zio; 4146 boolean_t more; 4147 4148 do { 4149 mutex_enter(&mca->mca_lock); 4150 zio = zio_io_to_allocate(mca, &more); 4151 mutex_exit(&mca->mca_lock); 4152 if (zio == NULL) 4153 return; 4154 4155 ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE); 4156 ASSERT0(zio->io_error); 4157 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE); 4158 } while (more); 4159 } 4160 4161 static zio_t * 4162 zio_dva_allocate(zio_t *zio) 4163 { 4164 spa_t *spa = zio->io_spa; 4165 metaslab_class_t *mc; 4166 blkptr_t *bp = zio->io_bp; 4167 int error; 4168 int flags = 0; 4169 4170 if (zio->io_gang_leader == NULL) { 4171 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 4172 zio->io_gang_leader = zio; 4173 } 4174 if (zio->io_flags & ZIO_FLAG_PREALLOCATED) { 4175 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_GANG); 4176 memcpy(zio->io_bp->blk_dva, zio->io_bp_orig.blk_dva, 4177 3 * sizeof (dva_t)); 4178 BP_SET_BIRTH(zio->io_bp, BP_GET_LOGICAL_BIRTH(&zio->io_bp_orig), 4179 BP_GET_PHYSICAL_BIRTH(&zio->io_bp_orig)); 4180 return (zio); 4181 } 4182 4183 ASSERT(BP_IS_HOLE(bp)); 4184 ASSERT0(BP_GET_NDVAS(bp)); 4185 ASSERT3U(zio->io_prop.zp_copies, >, 0); 4186 4187 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa)); 4188 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp)); 4189 4190 if (zio->io_flags & ZIO_FLAG_GANG_CHILD) 4191 flags |= METASLAB_GANG_CHILD; 4192 if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE) 4193 flags |= METASLAB_ASYNC_ALLOC; 4194 4195 /* 4196 * If not already chosen, choose an appropriate allocation class. 4197 */ 4198 mc = zio->io_metaslab_class; 4199 if (mc == NULL) { 4200 mc = spa_preferred_class(spa, zio); 4201 zio->io_metaslab_class = mc; 4202 } 4203 ZIOSTAT_BUMP(ziostat_total_allocations); 4204 4205 again: 4206 /* 4207 * Try allocating the block in the usual metaslab class. 4208 * If that's full, allocate it in the normal class. 4209 * If that's full, allocate as a gang block, 4210 * and if all are full, the allocation fails (which shouldn't happen). 4211 * 4212 * Note that we do not fall back on embedded slog (ZIL) space, to 4213 * preserve unfragmented slog space, which is critical for decent 4214 * sync write performance. If a log allocation fails, we will fall 4215 * back to spa_sync() which is abysmal for performance. 4216 */ 4217 ASSERT(ZIO_HAS_ALLOCATOR(zio)); 4218 error = metaslab_alloc(spa, mc, zio->io_size, bp, 4219 zio->io_prop.zp_copies, zio->io_txg, NULL, flags, 4220 &zio->io_alloc_list, zio->io_allocator, zio); 4221 4222 /* 4223 * Fallback to normal class when an alloc class is full 4224 */ 4225 if (error == ENOSPC && mc != spa_normal_class(spa)) { 4226 /* 4227 * When the dedup or special class is spilling into the normal 4228 * class, there can still be significant space available due 4229 * to deferred frees that are in-flight. We track the txg when 4230 * this occurred and back off adding new DDT entries for a few 4231 * txgs to allow the free blocks to be processed. 4232 */ 4233 if ((mc == spa_dedup_class(spa) || (spa_special_has_ddt(spa) && 4234 mc == spa_special_class(spa))) && 4235 spa->spa_dedup_class_full_txg != zio->io_txg) { 4236 spa->spa_dedup_class_full_txg = zio->io_txg; 4237 zfs_dbgmsg("%s[%d]: %s class spilling, req size %d, " 4238 "%llu allocated of %llu", 4239 spa_name(spa), (int)zio->io_txg, 4240 mc == spa_dedup_class(spa) ? "dedup" : "special", 4241 (int)zio->io_size, 4242 (u_longlong_t)metaslab_class_get_alloc(mc), 4243 (u_longlong_t)metaslab_class_get_space(mc)); 4244 } 4245 4246 /* 4247 * If we are holding old class reservation, drop it. 4248 * Dispatch the next ZIO(s) there if some are waiting. 4249 */ 4250 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) { 4251 if (metaslab_class_throttle_unreserve(mc, 4252 zio->io_prop.zp_copies, zio)) { 4253 zio_allocate_dispatch(zio->io_metaslab_class, 4254 zio->io_allocator); 4255 } 4256 zio->io_flags &= ~ZIO_FLAG_IO_ALLOCATING; 4257 } 4258 4259 if (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC) { 4260 zfs_dbgmsg("%s: metaslab allocation failure, " 4261 "trying normal class: zio %px, size %llu, error %d", 4262 spa_name(spa), zio, (u_longlong_t)zio->io_size, 4263 error); 4264 } 4265 zio->io_metaslab_class = mc = spa_normal_class(spa); 4266 ZIOSTAT_BUMP(ziostat_alloc_class_fallbacks); 4267 4268 /* 4269 * If normal class uses throttling, return to that pipeline 4270 * stage. Otherwise just do another allocation attempt. 4271 */ 4272 if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE && 4273 mc->mc_alloc_throttle_enabled && 4274 zio->io_child_type != ZIO_CHILD_GANG && 4275 !(zio->io_flags & ZIO_FLAG_NODATA)) { 4276 zio->io_stage = ZIO_STAGE_DVA_THROTTLE >> 1; 4277 return (zio); 4278 } 4279 goto again; 4280 } 4281 4282 if (error == ENOSPC && zio->io_size > spa->spa_min_alloc) { 4283 if (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC) { 4284 zfs_dbgmsg("%s: metaslab allocation failure, " 4285 "trying ganging: zio %px, size %llu, error %d", 4286 spa_name(spa), zio, (u_longlong_t)zio->io_size, 4287 error); 4288 } 4289 ZIOSTAT_BUMP(ziostat_gang_writes); 4290 if (flags & METASLAB_GANG_CHILD) 4291 ZIOSTAT_BUMP(ziostat_gang_multilevel); 4292 return (zio_write_gang_block(zio, mc)); 4293 } 4294 if (error != 0) { 4295 if (error != ENOSPC || 4296 (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC)) { 4297 zfs_dbgmsg("%s: metaslab allocation failure: zio %px, " 4298 "size %llu, error %d", 4299 spa_name(spa), zio, (u_longlong_t)zio->io_size, 4300 error); 4301 } 4302 zio->io_error = error; 4303 } 4304 4305 return (zio); 4306 } 4307 4308 static zio_t * 4309 zio_dva_free(zio_t *zio) 4310 { 4311 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE); 4312 4313 return (zio); 4314 } 4315 4316 static zio_t * 4317 zio_dva_claim(zio_t *zio) 4318 { 4319 int error; 4320 4321 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg); 4322 if (error) 4323 zio->io_error = error; 4324 4325 return (zio); 4326 } 4327 4328 /* 4329 * Undo an allocation. This is used by zio_done() when an I/O fails 4330 * and we want to give back the block we just allocated. 4331 * This handles both normal blocks and gang blocks. 4332 */ 4333 static void 4334 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp) 4335 { 4336 ASSERT(BP_GET_LOGICAL_BIRTH(bp) == zio->io_txg || BP_IS_HOLE(bp)); 4337 ASSERT(zio->io_bp_override == NULL); 4338 4339 if (!BP_IS_HOLE(bp)) { 4340 metaslab_free(zio->io_spa, bp, BP_GET_LOGICAL_BIRTH(bp), 4341 B_TRUE); 4342 } 4343 4344 if (gn != NULL) { 4345 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 4346 zio_dva_unallocate(zio, gn->gn_child[g], 4347 &gn->gn_gbh->zg_blkptr[g]); 4348 } 4349 } 4350 } 4351 4352 /* 4353 * Try to allocate an intent log block. Return 0 on success, errno on failure. 4354 */ 4355 int 4356 zio_alloc_zil(spa_t *spa, objset_t *os, uint64_t txg, blkptr_t *new_bp, 4357 uint64_t size, boolean_t *slog) 4358 { 4359 int error = 1; 4360 zio_alloc_list_t io_alloc_list; 4361 4362 ASSERT(txg > spa_syncing_txg(spa)); 4363 4364 metaslab_trace_init(&io_alloc_list); 4365 4366 /* 4367 * Block pointer fields are useful to metaslabs for stats and debugging. 4368 * Fill in the obvious ones before calling into metaslab_alloc(). 4369 */ 4370 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); 4371 BP_SET_PSIZE(new_bp, size); 4372 BP_SET_LEVEL(new_bp, 0); 4373 4374 /* 4375 * When allocating a zil block, we don't have information about 4376 * the final destination of the block except the objset it's part 4377 * of, so we just hash the objset ID to pick the allocator to get 4378 * some parallelism. 4379 */ 4380 int flags = METASLAB_ZIL; 4381 int allocator = (uint_t)cityhash1(os->os_dsl_dataset->ds_object) 4382 % spa->spa_alloc_count; 4383 ZIOSTAT_BUMP(ziostat_total_allocations); 4384 error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1, 4385 txg, NULL, flags, &io_alloc_list, allocator, NULL); 4386 *slog = (error == 0); 4387 if (error != 0) { 4388 error = metaslab_alloc(spa, spa_embedded_log_class(spa), size, 4389 new_bp, 1, txg, NULL, flags, &io_alloc_list, allocator, 4390 NULL); 4391 } 4392 if (error != 0) { 4393 ZIOSTAT_BUMP(ziostat_alloc_class_fallbacks); 4394 error = metaslab_alloc(spa, spa_normal_class(spa), size, 4395 new_bp, 1, txg, NULL, flags, &io_alloc_list, allocator, 4396 NULL); 4397 } 4398 metaslab_trace_fini(&io_alloc_list); 4399 4400 if (error == 0) { 4401 BP_SET_LSIZE(new_bp, size); 4402 BP_SET_PSIZE(new_bp, size); 4403 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF); 4404 BP_SET_CHECKSUM(new_bp, 4405 spa_version(spa) >= SPA_VERSION_SLIM_ZIL 4406 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG); 4407 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); 4408 BP_SET_LEVEL(new_bp, 0); 4409 BP_SET_DEDUP(new_bp, 0); 4410 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER); 4411 4412 /* 4413 * encrypted blocks will require an IV and salt. We generate 4414 * these now since we will not be rewriting the bp at 4415 * rewrite time. 4416 */ 4417 if (os->os_encrypted) { 4418 uint8_t iv[ZIO_DATA_IV_LEN]; 4419 uint8_t salt[ZIO_DATA_SALT_LEN]; 4420 4421 BP_SET_CRYPT(new_bp, B_TRUE); 4422 VERIFY0(spa_crypt_get_salt(spa, 4423 dmu_objset_id(os), salt)); 4424 VERIFY0(zio_crypt_generate_iv(iv)); 4425 4426 zio_crypt_encode_params_bp(new_bp, salt, iv); 4427 } 4428 } else { 4429 zfs_dbgmsg("%s: zil block allocation failure: " 4430 "size %llu, error %d", spa_name(spa), (u_longlong_t)size, 4431 error); 4432 } 4433 4434 return (error); 4435 } 4436 4437 /* 4438 * ========================================================================== 4439 * Read and write to physical devices 4440 * ========================================================================== 4441 */ 4442 4443 /* 4444 * Issue an I/O to the underlying vdev. Typically the issue pipeline 4445 * stops after this stage and will resume upon I/O completion. 4446 * However, there are instances where the vdev layer may need to 4447 * continue the pipeline when an I/O was not issued. Since the I/O 4448 * that was sent to the vdev layer might be different than the one 4449 * currently active in the pipeline (see vdev_queue_io()), we explicitly 4450 * force the underlying vdev layers to call either zio_execute() or 4451 * zio_interrupt() to ensure that the pipeline continues with the correct I/O. 4452 */ 4453 static zio_t * 4454 zio_vdev_io_start(zio_t *zio) 4455 { 4456 vdev_t *vd = zio->io_vd; 4457 uint64_t align; 4458 spa_t *spa = zio->io_spa; 4459 4460 zio->io_delay = 0; 4461 4462 ASSERT(zio->io_error == 0); 4463 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0); 4464 4465 if (vd == NULL) { 4466 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) 4467 spa_config_enter(spa, SCL_ZIO, zio, RW_READER); 4468 4469 /* 4470 * The mirror_ops handle multiple DVAs in a single BP. 4471 */ 4472 vdev_mirror_ops.vdev_op_io_start(zio); 4473 return (NULL); 4474 } 4475 4476 ASSERT3P(zio->io_logical, !=, zio); 4477 if (zio->io_type == ZIO_TYPE_WRITE) { 4478 ASSERT(spa->spa_trust_config); 4479 4480 /* 4481 * Note: the code can handle other kinds of writes, 4482 * but we don't expect them. 4483 */ 4484 if (zio->io_vd->vdev_noalloc) { 4485 ASSERT(zio->io_flags & 4486 (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL | 4487 ZIO_FLAG_RESILVER | ZIO_FLAG_INDUCE_DAMAGE)); 4488 } 4489 } 4490 4491 align = 1ULL << vd->vdev_top->vdev_ashift; 4492 4493 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) && 4494 P2PHASE(zio->io_size, align) != 0) { 4495 /* Transform logical writes to be a full physical block size. */ 4496 uint64_t asize = P2ROUNDUP(zio->io_size, align); 4497 abd_t *abuf = abd_alloc_sametype(zio->io_abd, asize); 4498 ASSERT(vd == vd->vdev_top); 4499 if (zio->io_type == ZIO_TYPE_WRITE) { 4500 abd_copy(abuf, zio->io_abd, zio->io_size); 4501 abd_zero_off(abuf, zio->io_size, asize - zio->io_size); 4502 } 4503 zio_push_transform(zio, abuf, asize, asize, zio_subblock); 4504 } 4505 4506 /* 4507 * If this is not a physical io, make sure that it is properly aligned 4508 * before proceeding. 4509 */ 4510 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) { 4511 ASSERT0(P2PHASE(zio->io_offset, align)); 4512 ASSERT0(P2PHASE(zio->io_size, align)); 4513 } else { 4514 /* 4515 * For physical writes, we allow 512b aligned writes and assume 4516 * the device will perform a read-modify-write as necessary. 4517 */ 4518 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE)); 4519 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE)); 4520 } 4521 4522 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa)); 4523 4524 /* 4525 * If this is a repair I/O, and there's no self-healing involved -- 4526 * that is, we're just resilvering what we expect to resilver -- 4527 * then don't do the I/O unless zio's txg is actually in vd's DTL. 4528 * This prevents spurious resilvering. 4529 * 4530 * There are a few ways that we can end up creating these spurious 4531 * resilver i/os: 4532 * 4533 * 1. A resilver i/o will be issued if any DVA in the BP has a 4534 * dirty DTL. The mirror code will issue resilver writes to 4535 * each DVA, including the one(s) that are not on vdevs with dirty 4536 * DTLs. 4537 * 4538 * 2. With nested replication, which happens when we have a 4539 * "replacing" or "spare" vdev that's a child of a mirror or raidz. 4540 * For example, given mirror(replacing(A+B), C), it's likely that 4541 * only A is out of date (it's the new device). In this case, we'll 4542 * read from C, then use the data to resilver A+B -- but we don't 4543 * actually want to resilver B, just A. The top-level mirror has no 4544 * way to know this, so instead we just discard unnecessary repairs 4545 * as we work our way down the vdev tree. 4546 * 4547 * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc. 4548 * The same logic applies to any form of nested replication: ditto 4549 * + mirror, RAID-Z + replacing, etc. 4550 * 4551 * However, indirect vdevs point off to other vdevs which may have 4552 * DTL's, so we never bypass them. The child i/os on concrete vdevs 4553 * will be properly bypassed instead. 4554 * 4555 * Leaf DTL_PARTIAL can be empty when a legitimate write comes from 4556 * a dRAID spare vdev. For example, when a dRAID spare is first 4557 * used, its spare blocks need to be written to but the leaf vdev's 4558 * of such blocks can have empty DTL_PARTIAL. 4559 * 4560 * There seemed no clean way to allow such writes while bypassing 4561 * spurious ones. At this point, just avoid all bypassing for dRAID 4562 * for correctness. 4563 */ 4564 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) && 4565 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) && 4566 zio->io_txg != 0 && /* not a delegated i/o */ 4567 vd->vdev_ops != &vdev_indirect_ops && 4568 vd->vdev_top->vdev_ops != &vdev_draid_ops && 4569 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) { 4570 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 4571 zio_vdev_io_bypass(zio); 4572 return (zio); 4573 } 4574 4575 /* 4576 * Select the next best leaf I/O to process. Distributed spares are 4577 * excluded since they dispatch the I/O directly to a leaf vdev after 4578 * applying the dRAID mapping. 4579 */ 4580 if (vd->vdev_ops->vdev_op_leaf && 4581 vd->vdev_ops != &vdev_draid_spare_ops && 4582 (zio->io_type == ZIO_TYPE_READ || 4583 zio->io_type == ZIO_TYPE_WRITE || 4584 zio->io_type == ZIO_TYPE_TRIM)) { 4585 4586 if ((zio = vdev_queue_io(zio)) == NULL) 4587 return (NULL); 4588 4589 if (!vdev_accessible(vd, zio)) { 4590 zio->io_error = SET_ERROR(ENXIO); 4591 zio_interrupt(zio); 4592 return (NULL); 4593 } 4594 zio->io_delay = gethrtime(); 4595 4596 if (zio_handle_device_injection(vd, zio, ENOSYS) != 0) { 4597 /* 4598 * "no-op" injections return success, but do no actual 4599 * work. Just return it. 4600 */ 4601 zio_delay_interrupt(zio); 4602 return (NULL); 4603 } 4604 } 4605 4606 vd->vdev_ops->vdev_op_io_start(zio); 4607 return (NULL); 4608 } 4609 4610 static zio_t * 4611 zio_vdev_io_done(zio_t *zio) 4612 { 4613 vdev_t *vd = zio->io_vd; 4614 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops; 4615 boolean_t unexpected_error = B_FALSE; 4616 4617 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) { 4618 return (NULL); 4619 } 4620 4621 ASSERT(zio->io_type == ZIO_TYPE_READ || 4622 zio->io_type == ZIO_TYPE_WRITE || 4623 zio->io_type == ZIO_TYPE_FLUSH || 4624 zio->io_type == ZIO_TYPE_TRIM); 4625 4626 if (zio->io_delay) 4627 zio->io_delay = gethrtime() - zio->io_delay; 4628 4629 if (vd != NULL && vd->vdev_ops->vdev_op_leaf && 4630 vd->vdev_ops != &vdev_draid_spare_ops) { 4631 if (zio->io_type != ZIO_TYPE_FLUSH) 4632 vdev_queue_io_done(zio); 4633 4634 if (zio_injection_enabled && zio->io_error == 0) 4635 zio->io_error = zio_handle_device_injections(vd, zio, 4636 EIO, EILSEQ); 4637 4638 if (zio_injection_enabled && zio->io_error == 0) 4639 zio->io_error = zio_handle_label_injection(zio, EIO); 4640 4641 if (zio->io_error && zio->io_type != ZIO_TYPE_FLUSH && 4642 zio->io_type != ZIO_TYPE_TRIM) { 4643 if (!vdev_accessible(vd, zio)) { 4644 zio->io_error = SET_ERROR(ENXIO); 4645 } else { 4646 unexpected_error = B_TRUE; 4647 } 4648 } 4649 } 4650 4651 ops->vdev_op_io_done(zio); 4652 4653 if (unexpected_error && vd->vdev_remove_wanted == B_FALSE) 4654 VERIFY(vdev_probe(vd, zio) == NULL); 4655 4656 return (zio); 4657 } 4658 4659 /* 4660 * This function is used to change the priority of an existing zio that is 4661 * currently in-flight. This is used by the arc to upgrade priority in the 4662 * event that a demand read is made for a block that is currently queued 4663 * as a scrub or async read IO. Otherwise, the high priority read request 4664 * would end up having to wait for the lower priority IO. 4665 */ 4666 void 4667 zio_change_priority(zio_t *pio, zio_priority_t priority) 4668 { 4669 zio_t *cio, *cio_next; 4670 zio_link_t *zl = NULL; 4671 4672 ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 4673 4674 if (pio->io_vd != NULL && pio->io_vd->vdev_ops->vdev_op_leaf) { 4675 vdev_queue_change_io_priority(pio, priority); 4676 } else { 4677 pio->io_priority = priority; 4678 } 4679 4680 mutex_enter(&pio->io_lock); 4681 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) { 4682 cio_next = zio_walk_children(pio, &zl); 4683 zio_change_priority(cio, priority); 4684 } 4685 mutex_exit(&pio->io_lock); 4686 } 4687 4688 /* 4689 * For non-raidz ZIOs, we can just copy aside the bad data read from the 4690 * disk, and use that to finish the checksum ereport later. 4691 */ 4692 static void 4693 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr, 4694 const abd_t *good_buf) 4695 { 4696 /* no processing needed */ 4697 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE); 4698 } 4699 4700 void 4701 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr) 4702 { 4703 void *abd = abd_alloc_sametype(zio->io_abd, zio->io_size); 4704 4705 abd_copy(abd, zio->io_abd, zio->io_size); 4706 4707 zcr->zcr_cbinfo = zio->io_size; 4708 zcr->zcr_cbdata = abd; 4709 zcr->zcr_finish = zio_vsd_default_cksum_finish; 4710 zcr->zcr_free = zio_abd_free; 4711 } 4712 4713 static zio_t * 4714 zio_vdev_io_assess(zio_t *zio) 4715 { 4716 vdev_t *vd = zio->io_vd; 4717 4718 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) { 4719 return (NULL); 4720 } 4721 4722 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) 4723 spa_config_exit(zio->io_spa, SCL_ZIO, zio); 4724 4725 if (zio->io_vsd != NULL) { 4726 zio->io_vsd_ops->vsd_free(zio); 4727 zio->io_vsd = NULL; 4728 } 4729 4730 /* 4731 * If a Direct I/O operation has a checksum verify error then this I/O 4732 * should not attempt to be issued again. 4733 */ 4734 if (zio->io_flags & ZIO_FLAG_DIO_CHKSUM_ERR) { 4735 if (zio->io_type == ZIO_TYPE_WRITE) { 4736 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_LOGICAL); 4737 ASSERT3U(zio->io_error, ==, EIO); 4738 } 4739 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 4740 return (zio); 4741 } 4742 4743 if (zio_injection_enabled && zio->io_error == 0) 4744 zio->io_error = zio_handle_fault_injection(zio, EIO); 4745 4746 /* 4747 * If the I/O failed, determine whether we should attempt to retry it. 4748 * 4749 * On retry, we cut in line in the issue queue, since we don't want 4750 * compression/checksumming/etc. work to prevent our (cheap) IO reissue. 4751 */ 4752 if (zio->io_error && vd == NULL && 4753 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) { 4754 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */ 4755 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */ 4756 zio->io_error = 0; 4757 zio->io_flags |= ZIO_FLAG_IO_RETRY | ZIO_FLAG_DONT_AGGREGATE; 4758 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1; 4759 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, 4760 zio_requeue_io_start_cut_in_line); 4761 return (NULL); 4762 } 4763 4764 /* 4765 * If we got an error on a leaf device, convert it to ENXIO 4766 * if the device is not accessible at all. 4767 */ 4768 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf && 4769 !vdev_accessible(vd, zio)) 4770 zio->io_error = SET_ERROR(ENXIO); 4771 4772 /* 4773 * If we can't write to an interior vdev (mirror or RAID-Z), 4774 * set vdev_cant_write so that we stop trying to allocate from it. 4775 */ 4776 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE && 4777 vd != NULL && !vd->vdev_ops->vdev_op_leaf) { 4778 vdev_dbgmsg(vd, "zio_vdev_io_assess(zio=%px) setting " 4779 "cant_write=TRUE due to write failure with ENXIO", 4780 zio); 4781 vd->vdev_cant_write = B_TRUE; 4782 } 4783 4784 /* 4785 * If a cache flush returns ENOTSUP we know that no future 4786 * attempts will ever succeed. In this case we set a persistent 4787 * boolean flag so that we don't bother with it in the future, and 4788 * then we act like the flush succeeded. 4789 */ 4790 if (zio->io_error == ENOTSUP && zio->io_type == ZIO_TYPE_FLUSH && 4791 vd != NULL) { 4792 vd->vdev_nowritecache = B_TRUE; 4793 zio->io_error = 0; 4794 } 4795 4796 if (zio->io_error) 4797 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 4798 4799 return (zio); 4800 } 4801 4802 void 4803 zio_vdev_io_reissue(zio_t *zio) 4804 { 4805 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); 4806 ASSERT(zio->io_error == 0); 4807 4808 zio->io_stage >>= 1; 4809 } 4810 4811 void 4812 zio_vdev_io_redone(zio_t *zio) 4813 { 4814 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE); 4815 4816 zio->io_stage >>= 1; 4817 } 4818 4819 void 4820 zio_vdev_io_bypass(zio_t *zio) 4821 { 4822 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); 4823 ASSERT(zio->io_error == 0); 4824 4825 zio->io_flags |= ZIO_FLAG_IO_BYPASS; 4826 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1; 4827 } 4828 4829 /* 4830 * ========================================================================== 4831 * Encrypt and store encryption parameters 4832 * ========================================================================== 4833 */ 4834 4835 4836 /* 4837 * This function is used for ZIO_STAGE_ENCRYPT. It is responsible for 4838 * managing the storage of encryption parameters and passing them to the 4839 * lower-level encryption functions. 4840 */ 4841 static zio_t * 4842 zio_encrypt(zio_t *zio) 4843 { 4844 zio_prop_t *zp = &zio->io_prop; 4845 spa_t *spa = zio->io_spa; 4846 blkptr_t *bp = zio->io_bp; 4847 uint64_t psize = BP_GET_PSIZE(bp); 4848 uint64_t dsobj = zio->io_bookmark.zb_objset; 4849 dmu_object_type_t ot = BP_GET_TYPE(bp); 4850 void *enc_buf = NULL; 4851 abd_t *eabd = NULL; 4852 uint8_t salt[ZIO_DATA_SALT_LEN]; 4853 uint8_t iv[ZIO_DATA_IV_LEN]; 4854 uint8_t mac[ZIO_DATA_MAC_LEN]; 4855 boolean_t no_crypt = B_FALSE; 4856 4857 /* the root zio already encrypted the data */ 4858 if (zio->io_child_type == ZIO_CHILD_GANG) 4859 return (zio); 4860 4861 /* only ZIL blocks are re-encrypted on rewrite */ 4862 if (!IO_IS_ALLOCATING(zio) && ot != DMU_OT_INTENT_LOG) 4863 return (zio); 4864 4865 if (!(zp->zp_encrypt || BP_IS_ENCRYPTED(bp))) { 4866 BP_SET_CRYPT(bp, B_FALSE); 4867 return (zio); 4868 } 4869 4870 /* if we are doing raw encryption set the provided encryption params */ 4871 if (zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) { 4872 ASSERT0(BP_GET_LEVEL(bp)); 4873 BP_SET_CRYPT(bp, B_TRUE); 4874 BP_SET_BYTEORDER(bp, zp->zp_byteorder); 4875 if (ot != DMU_OT_OBJSET) 4876 zio_crypt_encode_mac_bp(bp, zp->zp_mac); 4877 4878 /* dnode blocks must be written out in the provided byteorder */ 4879 if (zp->zp_byteorder != ZFS_HOST_BYTEORDER && 4880 ot == DMU_OT_DNODE) { 4881 void *bswap_buf = zio_buf_alloc(psize); 4882 abd_t *babd = abd_get_from_buf(bswap_buf, psize); 4883 4884 ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF); 4885 abd_copy_to_buf(bswap_buf, zio->io_abd, psize); 4886 dmu_ot_byteswap[DMU_OT_BYTESWAP(ot)].ob_func(bswap_buf, 4887 psize); 4888 4889 abd_take_ownership_of_buf(babd, B_TRUE); 4890 zio_push_transform(zio, babd, psize, psize, NULL); 4891 } 4892 4893 if (DMU_OT_IS_ENCRYPTED(ot)) 4894 zio_crypt_encode_params_bp(bp, zp->zp_salt, zp->zp_iv); 4895 return (zio); 4896 } 4897 4898 /* indirect blocks only maintain a cksum of the lower level MACs */ 4899 if (BP_GET_LEVEL(bp) > 0) { 4900 BP_SET_CRYPT(bp, B_TRUE); 4901 VERIFY0(zio_crypt_do_indirect_mac_checksum_abd(B_TRUE, 4902 zio->io_orig_abd, BP_GET_LSIZE(bp), BP_SHOULD_BYTESWAP(bp), 4903 mac)); 4904 zio_crypt_encode_mac_bp(bp, mac); 4905 return (zio); 4906 } 4907 4908 /* 4909 * Objset blocks are a special case since they have 2 256-bit MACs 4910 * embedded within them. 4911 */ 4912 if (ot == DMU_OT_OBJSET) { 4913 ASSERT0(DMU_OT_IS_ENCRYPTED(ot)); 4914 ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF); 4915 BP_SET_CRYPT(bp, B_TRUE); 4916 VERIFY0(spa_do_crypt_objset_mac_abd(B_TRUE, spa, dsobj, 4917 zio->io_abd, psize, BP_SHOULD_BYTESWAP(bp))); 4918 return (zio); 4919 } 4920 4921 /* unencrypted object types are only authenticated with a MAC */ 4922 if (!DMU_OT_IS_ENCRYPTED(ot)) { 4923 BP_SET_CRYPT(bp, B_TRUE); 4924 VERIFY0(spa_do_crypt_mac_abd(B_TRUE, spa, dsobj, 4925 zio->io_abd, psize, mac)); 4926 zio_crypt_encode_mac_bp(bp, mac); 4927 return (zio); 4928 } 4929 4930 /* 4931 * Later passes of sync-to-convergence may decide to rewrite data 4932 * in place to avoid more disk reallocations. This presents a problem 4933 * for encryption because this constitutes rewriting the new data with 4934 * the same encryption key and IV. However, this only applies to blocks 4935 * in the MOS (particularly the spacemaps) and we do not encrypt the 4936 * MOS. We assert that the zio is allocating or an intent log write 4937 * to enforce this. 4938 */ 4939 ASSERT(IO_IS_ALLOCATING(zio) || ot == DMU_OT_INTENT_LOG); 4940 ASSERT(BP_GET_LEVEL(bp) == 0 || ot == DMU_OT_INTENT_LOG); 4941 ASSERT(spa_feature_is_active(spa, SPA_FEATURE_ENCRYPTION)); 4942 ASSERT3U(psize, !=, 0); 4943 4944 enc_buf = zio_buf_alloc(psize); 4945 eabd = abd_get_from_buf(enc_buf, psize); 4946 abd_take_ownership_of_buf(eabd, B_TRUE); 4947 4948 /* 4949 * For an explanation of what encryption parameters are stored 4950 * where, see the block comment in zio_crypt.c. 4951 */ 4952 if (ot == DMU_OT_INTENT_LOG) { 4953 zio_crypt_decode_params_bp(bp, salt, iv); 4954 } else { 4955 BP_SET_CRYPT(bp, B_TRUE); 4956 } 4957 4958 /* Perform the encryption. This should not fail */ 4959 VERIFY0(spa_do_crypt_abd(B_TRUE, spa, &zio->io_bookmark, 4960 BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp), 4961 salt, iv, mac, psize, zio->io_abd, eabd, &no_crypt)); 4962 4963 /* encode encryption metadata into the bp */ 4964 if (ot == DMU_OT_INTENT_LOG) { 4965 /* 4966 * ZIL blocks store the MAC in the embedded checksum, so the 4967 * transform must always be applied. 4968 */ 4969 zio_crypt_encode_mac_zil(enc_buf, mac); 4970 zio_push_transform(zio, eabd, psize, psize, NULL); 4971 } else { 4972 BP_SET_CRYPT(bp, B_TRUE); 4973 zio_crypt_encode_params_bp(bp, salt, iv); 4974 zio_crypt_encode_mac_bp(bp, mac); 4975 4976 if (no_crypt) { 4977 ASSERT3U(ot, ==, DMU_OT_DNODE); 4978 abd_free(eabd); 4979 } else { 4980 zio_push_transform(zio, eabd, psize, psize, NULL); 4981 } 4982 } 4983 4984 return (zio); 4985 } 4986 4987 /* 4988 * ========================================================================== 4989 * Generate and verify checksums 4990 * ========================================================================== 4991 */ 4992 static zio_t * 4993 zio_checksum_generate(zio_t *zio) 4994 { 4995 blkptr_t *bp = zio->io_bp; 4996 enum zio_checksum checksum; 4997 4998 if (bp == NULL) { 4999 /* 5000 * This is zio_write_phys(). 5001 * We're either generating a label checksum, or none at all. 5002 */ 5003 checksum = zio->io_prop.zp_checksum; 5004 5005 if (checksum == ZIO_CHECKSUM_OFF) 5006 return (zio); 5007 5008 ASSERT(checksum == ZIO_CHECKSUM_LABEL); 5009 } else { 5010 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) { 5011 ASSERT(!IO_IS_ALLOCATING(zio)); 5012 checksum = ZIO_CHECKSUM_GANG_HEADER; 5013 } else { 5014 checksum = BP_GET_CHECKSUM(bp); 5015 } 5016 } 5017 5018 zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size); 5019 5020 return (zio); 5021 } 5022 5023 static zio_t * 5024 zio_checksum_verify(zio_t *zio) 5025 { 5026 zio_bad_cksum_t info; 5027 blkptr_t *bp = zio->io_bp; 5028 int error; 5029 5030 ASSERT(zio->io_vd != NULL); 5031 5032 if (bp == NULL) { 5033 /* 5034 * This is zio_read_phys(). 5035 * We're either verifying a label checksum, or nothing at all. 5036 */ 5037 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF) 5038 return (zio); 5039 5040 ASSERT3U(zio->io_prop.zp_checksum, ==, ZIO_CHECKSUM_LABEL); 5041 } 5042 5043 ASSERT0(zio->io_flags & ZIO_FLAG_DIO_CHKSUM_ERR); 5044 IMPLY(zio->io_flags & ZIO_FLAG_DIO_READ, 5045 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)); 5046 5047 if ((error = zio_checksum_error(zio, &info)) != 0) { 5048 zio->io_error = error; 5049 if (error == ECKSUM && 5050 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) { 5051 if (zio->io_flags & ZIO_FLAG_DIO_READ) { 5052 zio->io_flags |= ZIO_FLAG_DIO_CHKSUM_ERR; 5053 zio_t *pio = zio_unique_parent(zio); 5054 /* 5055 * Any Direct I/O read that has a checksum 5056 * error must be treated as suspicous as the 5057 * contents of the buffer could be getting 5058 * manipulated while the I/O is taking place. 5059 * 5060 * The checksum verify error will only be 5061 * reported here for disk and file VDEV's and 5062 * will be reported on those that the failure 5063 * occurred on. Other types of VDEV's report the 5064 * verify failure in their own code paths. 5065 */ 5066 if (pio->io_child_type == ZIO_CHILD_LOGICAL) { 5067 zio_dio_chksum_verify_error_report(zio); 5068 } 5069 } else { 5070 mutex_enter(&zio->io_vd->vdev_stat_lock); 5071 zio->io_vd->vdev_stat.vs_checksum_errors++; 5072 mutex_exit(&zio->io_vd->vdev_stat_lock); 5073 (void) zfs_ereport_start_checksum(zio->io_spa, 5074 zio->io_vd, &zio->io_bookmark, zio, 5075 zio->io_offset, zio->io_size, &info); 5076 } 5077 } 5078 } 5079 5080 return (zio); 5081 } 5082 5083 static zio_t * 5084 zio_dio_checksum_verify(zio_t *zio) 5085 { 5086 zio_t *pio = zio_unique_parent(zio); 5087 int error; 5088 5089 ASSERT3P(zio->io_vd, !=, NULL); 5090 ASSERT3P(zio->io_bp, !=, NULL); 5091 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV); 5092 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE); 5093 ASSERT3B(pio->io_prop.zp_direct_write, ==, B_TRUE); 5094 ASSERT3U(pio->io_child_type, ==, ZIO_CHILD_LOGICAL); 5095 5096 if (zfs_vdev_direct_write_verify == 0 || zio->io_error != 0) 5097 goto out; 5098 5099 if ((error = zio_checksum_error(zio, NULL)) != 0) { 5100 zio->io_error = error; 5101 if (error == ECKSUM) { 5102 zio->io_flags |= ZIO_FLAG_DIO_CHKSUM_ERR; 5103 zio_dio_chksum_verify_error_report(zio); 5104 } 5105 } 5106 5107 out: 5108 return (zio); 5109 } 5110 5111 5112 /* 5113 * Called by RAID-Z to ensure we don't compute the checksum twice. 5114 */ 5115 void 5116 zio_checksum_verified(zio_t *zio) 5117 { 5118 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; 5119 } 5120 5121 /* 5122 * Report Direct I/O checksum verify error and create ZED event. 5123 */ 5124 void 5125 zio_dio_chksum_verify_error_report(zio_t *zio) 5126 { 5127 ASSERT(zio->io_flags & ZIO_FLAG_DIO_CHKSUM_ERR); 5128 5129 if (zio->io_child_type == ZIO_CHILD_LOGICAL) 5130 return; 5131 5132 mutex_enter(&zio->io_vd->vdev_stat_lock); 5133 zio->io_vd->vdev_stat.vs_dio_verify_errors++; 5134 mutex_exit(&zio->io_vd->vdev_stat_lock); 5135 if (zio->io_type == ZIO_TYPE_WRITE) { 5136 /* 5137 * Convert checksum error for writes into EIO. 5138 */ 5139 zio->io_error = SET_ERROR(EIO); 5140 /* 5141 * Report dio_verify_wr ZED event. 5142 */ 5143 (void) zfs_ereport_post(FM_EREPORT_ZFS_DIO_VERIFY_WR, 5144 zio->io_spa, zio->io_vd, &zio->io_bookmark, zio, 0); 5145 } else { 5146 /* 5147 * Report dio_verify_rd ZED event. 5148 */ 5149 (void) zfs_ereport_post(FM_EREPORT_ZFS_DIO_VERIFY_RD, 5150 zio->io_spa, zio->io_vd, &zio->io_bookmark, zio, 0); 5151 } 5152 } 5153 5154 /* 5155 * ========================================================================== 5156 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other. 5157 * An error of 0 indicates success. ENXIO indicates whole-device failure, 5158 * which may be transient (e.g. unplugged) or permanent. ECKSUM and EIO 5159 * indicate errors that are specific to one I/O, and most likely permanent. 5160 * Any other error is presumed to be worse because we weren't expecting it. 5161 * ========================================================================== 5162 */ 5163 int 5164 zio_worst_error(int e1, int e2) 5165 { 5166 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO }; 5167 int r1, r2; 5168 5169 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++) 5170 if (e1 == zio_error_rank[r1]) 5171 break; 5172 5173 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++) 5174 if (e2 == zio_error_rank[r2]) 5175 break; 5176 5177 return (r1 > r2 ? e1 : e2); 5178 } 5179 5180 /* 5181 * ========================================================================== 5182 * I/O completion 5183 * ========================================================================== 5184 */ 5185 static zio_t * 5186 zio_ready(zio_t *zio) 5187 { 5188 blkptr_t *bp = zio->io_bp; 5189 zio_t *pio, *pio_next; 5190 zio_link_t *zl = NULL; 5191 5192 if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT | 5193 ZIO_CHILD_GANG_BIT | ZIO_CHILD_DDT_BIT, ZIO_WAIT_READY)) { 5194 return (NULL); 5195 } 5196 5197 if (zio->io_ready) { 5198 ASSERT(IO_IS_ALLOCATING(zio)); 5199 ASSERT(BP_GET_LOGICAL_BIRTH(bp) == zio->io_txg || 5200 BP_IS_HOLE(bp) || (zio->io_flags & ZIO_FLAG_NOPWRITE)); 5201 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0); 5202 5203 zio->io_ready(zio); 5204 } 5205 5206 #ifdef ZFS_DEBUG 5207 if (bp != NULL && bp != &zio->io_bp_copy) 5208 zio->io_bp_copy = *bp; 5209 #endif 5210 5211 if (zio->io_error != 0) { 5212 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 5213 5214 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) { 5215 ASSERT(IO_IS_ALLOCATING(zio)); 5216 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); 5217 ASSERT(zio->io_metaslab_class != NULL); 5218 ASSERT(ZIO_HAS_ALLOCATOR(zio)); 5219 5220 /* 5221 * We were unable to allocate anything, unreserve and 5222 * issue the next I/O to allocate. 5223 */ 5224 if (metaslab_class_throttle_unreserve( 5225 zio->io_metaslab_class, zio->io_prop.zp_copies, 5226 zio)) { 5227 zio_allocate_dispatch(zio->io_metaslab_class, 5228 zio->io_allocator); 5229 } 5230 } 5231 } 5232 5233 mutex_enter(&zio->io_lock); 5234 zio->io_state[ZIO_WAIT_READY] = 1; 5235 pio = zio_walk_parents(zio, &zl); 5236 mutex_exit(&zio->io_lock); 5237 5238 /* 5239 * As we notify zio's parents, new parents could be added. 5240 * New parents go to the head of zio's io_parent_list, however, 5241 * so we will (correctly) not notify them. The remainder of zio's 5242 * io_parent_list, from 'pio_next' onward, cannot change because 5243 * all parents must wait for us to be done before they can be done. 5244 */ 5245 for (; pio != NULL; pio = pio_next) { 5246 pio_next = zio_walk_parents(zio, &zl); 5247 zio_notify_parent(pio, zio, ZIO_WAIT_READY, NULL); 5248 } 5249 5250 if (zio->io_flags & ZIO_FLAG_NODATA) { 5251 if (bp != NULL && BP_IS_GANG(bp)) { 5252 zio->io_flags &= ~ZIO_FLAG_NODATA; 5253 } else { 5254 ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE); 5255 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; 5256 } 5257 } 5258 5259 if (zio_injection_enabled && 5260 zio->io_spa->spa_syncing_txg == zio->io_txg) 5261 zio_handle_ignored_writes(zio); 5262 5263 return (zio); 5264 } 5265 5266 /* 5267 * Update the allocation throttle accounting. 5268 */ 5269 static void 5270 zio_dva_throttle_done(zio_t *zio) 5271 { 5272 zio_t *pio = zio_unique_parent(zio); 5273 vdev_t *vd = zio->io_vd; 5274 int flags = METASLAB_ASYNC_ALLOC; 5275 const void *tag = pio; 5276 5277 ASSERT3P(zio->io_bp, !=, NULL); 5278 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE); 5279 ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE); 5280 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV); 5281 ASSERT(vd != NULL); 5282 ASSERT3P(vd, ==, vd->vdev_top); 5283 ASSERT(zio_injection_enabled || !(zio->io_flags & ZIO_FLAG_IO_RETRY)); 5284 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR)); 5285 ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING); 5286 5287 /* 5288 * Parents of gang children can have two flavors -- ones that allocated 5289 * the gang header (will have ZIO_FLAG_IO_REWRITE set) and ones that 5290 * allocated the constituent blocks. The first use their parent as tag. 5291 */ 5292 if (pio->io_child_type == ZIO_CHILD_GANG && 5293 (pio->io_flags & ZIO_FLAG_IO_REWRITE)) 5294 tag = zio_unique_parent(pio); 5295 5296 ASSERT(IO_IS_ALLOCATING(pio) || (pio->io_child_type == ZIO_CHILD_GANG && 5297 (pio->io_flags & ZIO_FLAG_IO_REWRITE))); 5298 ASSERT(ZIO_HAS_ALLOCATOR(pio)); 5299 ASSERT3P(zio, !=, zio->io_logical); 5300 ASSERT(zio->io_logical != NULL); 5301 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR)); 5302 ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE); 5303 ASSERT(zio->io_metaslab_class != NULL); 5304 ASSERT(zio->io_metaslab_class->mc_alloc_throttle_enabled); 5305 5306 metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, 5307 pio->io_allocator, flags, pio->io_size, tag); 5308 5309 if (metaslab_class_throttle_unreserve(zio->io_metaslab_class, 1, pio)) { 5310 zio_allocate_dispatch(zio->io_metaslab_class, 5311 pio->io_allocator); 5312 } 5313 } 5314 5315 static zio_t * 5316 zio_done(zio_t *zio) 5317 { 5318 /* 5319 * Always attempt to keep stack usage minimal here since 5320 * we can be called recursively up to 19 levels deep. 5321 */ 5322 const uint64_t psize = zio->io_size; 5323 zio_t *pio, *pio_next; 5324 zio_link_t *zl = NULL; 5325 5326 /* 5327 * If our children haven't all completed, 5328 * wait for them and then repeat this pipeline stage. 5329 */ 5330 if (zio_wait_for_children(zio, ZIO_CHILD_ALL_BITS, ZIO_WAIT_DONE)) { 5331 return (NULL); 5332 } 5333 5334 /* 5335 * If the allocation throttle is enabled, then update the accounting. 5336 * We only track child I/Os that are part of an allocating async 5337 * write. We must do this since the allocation is performed 5338 * by the logical I/O but the actual write is done by child I/Os. 5339 */ 5340 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING && 5341 zio->io_child_type == ZIO_CHILD_VDEV) 5342 zio_dva_throttle_done(zio); 5343 5344 for (int c = 0; c < ZIO_CHILD_TYPES; c++) 5345 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 5346 ASSERT(zio->io_children[c][w] == 0); 5347 5348 if (zio->io_bp != NULL && !BP_IS_EMBEDDED(zio->io_bp)) { 5349 ASSERT(zio->io_bp->blk_pad[0] == 0); 5350 ASSERT(zio->io_bp->blk_pad[1] == 0); 5351 ASSERT(memcmp(zio->io_bp, &zio->io_bp_copy, 5352 sizeof (blkptr_t)) == 0 || 5353 (zio->io_bp == zio_unique_parent(zio)->io_bp)); 5354 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) && 5355 zio->io_bp_override == NULL && 5356 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) { 5357 ASSERT3U(zio->io_prop.zp_copies, <=, 5358 BP_GET_NDVAS(zio->io_bp)); 5359 ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 || 5360 (BP_COUNT_GANG(zio->io_bp) == 5361 BP_GET_NDVAS(zio->io_bp))); 5362 } 5363 if (zio->io_flags & ZIO_FLAG_NOPWRITE) 5364 VERIFY(BP_EQUAL(zio->io_bp, &zio->io_bp_orig)); 5365 } 5366 5367 /* 5368 * If there were child vdev/gang/ddt errors, they apply to us now. 5369 */ 5370 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV); 5371 zio_inherit_child_errors(zio, ZIO_CHILD_GANG); 5372 zio_inherit_child_errors(zio, ZIO_CHILD_DDT); 5373 5374 /* 5375 * If the I/O on the transformed data was successful, generate any 5376 * checksum reports now while we still have the transformed data. 5377 */ 5378 if (zio->io_error == 0) { 5379 while (zio->io_cksum_report != NULL) { 5380 zio_cksum_report_t *zcr = zio->io_cksum_report; 5381 uint64_t align = zcr->zcr_align; 5382 uint64_t asize = P2ROUNDUP(psize, align); 5383 abd_t *adata = zio->io_abd; 5384 5385 if (adata != NULL && asize != psize) { 5386 adata = abd_alloc(asize, B_TRUE); 5387 abd_copy(adata, zio->io_abd, psize); 5388 abd_zero_off(adata, psize, asize - psize); 5389 } 5390 5391 zio->io_cksum_report = zcr->zcr_next; 5392 zcr->zcr_next = NULL; 5393 zcr->zcr_finish(zcr, adata); 5394 zfs_ereport_free_checksum(zcr); 5395 5396 if (adata != NULL && asize != psize) 5397 abd_free(adata); 5398 } 5399 } 5400 5401 zio_pop_transforms(zio); /* note: may set zio->io_error */ 5402 5403 vdev_stat_update(zio, psize); 5404 5405 /* 5406 * If this I/O is attached to a particular vdev is slow, exceeding 5407 * 30 seconds to complete, post an error described the I/O delay. 5408 * We ignore these errors if the device is currently unavailable. 5409 */ 5410 if (zio->io_delay >= MSEC2NSEC(zio_slow_io_ms)) { 5411 if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd)) { 5412 /* 5413 * We want to only increment our slow IO counters if 5414 * the IO is valid (i.e. not if the drive is removed). 5415 * 5416 * zfs_ereport_post() will also do these checks, but 5417 * it can also ratelimit and have other failures, so we 5418 * need to increment the slow_io counters independent 5419 * of it. 5420 */ 5421 if (zfs_ereport_is_valid(FM_EREPORT_ZFS_DELAY, 5422 zio->io_spa, zio->io_vd, zio)) { 5423 mutex_enter(&zio->io_vd->vdev_stat_lock); 5424 zio->io_vd->vdev_stat.vs_slow_ios++; 5425 mutex_exit(&zio->io_vd->vdev_stat_lock); 5426 5427 (void) zfs_ereport_post(FM_EREPORT_ZFS_DELAY, 5428 zio->io_spa, zio->io_vd, &zio->io_bookmark, 5429 zio, 0); 5430 } 5431 } 5432 } 5433 5434 if (zio->io_error) { 5435 /* 5436 * If this I/O is attached to a particular vdev, 5437 * generate an error message describing the I/O failure 5438 * at the block level. We ignore these errors if the 5439 * device is currently unavailable. 5440 */ 5441 if (zio->io_error != ECKSUM && zio->io_vd != NULL && 5442 !vdev_is_dead(zio->io_vd) && 5443 !(zio->io_flags & ZIO_FLAG_DIO_CHKSUM_ERR)) { 5444 int ret = zfs_ereport_post(FM_EREPORT_ZFS_IO, 5445 zio->io_spa, zio->io_vd, &zio->io_bookmark, zio, 0); 5446 if (ret != EALREADY) { 5447 mutex_enter(&zio->io_vd->vdev_stat_lock); 5448 if (zio->io_type == ZIO_TYPE_READ) 5449 zio->io_vd->vdev_stat.vs_read_errors++; 5450 else if (zio->io_type == ZIO_TYPE_WRITE) 5451 zio->io_vd->vdev_stat.vs_write_errors++; 5452 mutex_exit(&zio->io_vd->vdev_stat_lock); 5453 } 5454 } 5455 5456 if ((zio->io_error == EIO || !(zio->io_flags & 5457 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) && 5458 !(zio->io_flags & ZIO_FLAG_DIO_CHKSUM_ERR) && 5459 zio == zio->io_logical) { 5460 /* 5461 * For logical I/O requests, tell the SPA to log the 5462 * error and generate a logical data ereport. 5463 */ 5464 spa_log_error(zio->io_spa, &zio->io_bookmark, 5465 BP_GET_LOGICAL_BIRTH(zio->io_bp)); 5466 (void) zfs_ereport_post(FM_EREPORT_ZFS_DATA, 5467 zio->io_spa, NULL, &zio->io_bookmark, zio, 0); 5468 } 5469 } 5470 5471 if (zio->io_error && zio == zio->io_logical) { 5472 5473 /* 5474 * A DDT child tried to create a mixed gang/non-gang BP. We're 5475 * going to have to just retry as a non-dedup IO. 5476 */ 5477 if (zio->io_error == EAGAIN && IO_IS_ALLOCATING(zio) && 5478 zio->io_prop.zp_dedup) { 5479 zio->io_reexecute |= ZIO_REEXECUTE_NOW; 5480 zio->io_prop.zp_dedup = B_FALSE; 5481 } 5482 /* 5483 * Determine whether zio should be reexecuted. This will 5484 * propagate all the way to the root via zio_notify_parent(). 5485 */ 5486 ASSERT(zio->io_vd == NULL && zio->io_bp != NULL); 5487 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 5488 5489 if (IO_IS_ALLOCATING(zio) && 5490 !(zio->io_flags & ZIO_FLAG_CANFAIL) && 5491 !(zio->io_flags & ZIO_FLAG_DIO_CHKSUM_ERR)) { 5492 if (zio->io_error != ENOSPC) 5493 zio->io_reexecute |= ZIO_REEXECUTE_NOW; 5494 else 5495 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 5496 } 5497 5498 if ((zio->io_type == ZIO_TYPE_READ || 5499 zio->io_type == ZIO_TYPE_FREE) && 5500 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && 5501 zio->io_error == ENXIO && 5502 spa_load_state(zio->io_spa) == SPA_LOAD_NONE && 5503 spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE) 5504 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 5505 5506 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute) 5507 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 5508 5509 /* 5510 * Here is a possibly good place to attempt to do 5511 * either combinatorial reconstruction or error correction 5512 * based on checksums. It also might be a good place 5513 * to send out preliminary ereports before we suspend 5514 * processing. 5515 */ 5516 } 5517 5518 /* 5519 * If there were logical child errors, they apply to us now. 5520 * We defer this until now to avoid conflating logical child 5521 * errors with errors that happened to the zio itself when 5522 * updating vdev stats and reporting FMA events above. 5523 */ 5524 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL); 5525 5526 if ((zio->io_error || zio->io_reexecute) && 5527 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio && 5528 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE))) 5529 zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp); 5530 5531 zio_gang_tree_free(&zio->io_gang_tree); 5532 5533 /* 5534 * Godfather I/Os should never suspend. 5535 */ 5536 if ((zio->io_flags & ZIO_FLAG_GODFATHER) && 5537 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) 5538 zio->io_reexecute &= ~ZIO_REEXECUTE_SUSPEND; 5539 5540 if (zio->io_reexecute) { 5541 /* 5542 * A Direct I/O operation that has a checksum verify error 5543 * should not attempt to reexecute. Instead, the error should 5544 * just be propagated back. 5545 */ 5546 ASSERT(!(zio->io_flags & ZIO_FLAG_DIO_CHKSUM_ERR)); 5547 5548 /* 5549 * This is a logical I/O that wants to reexecute. 5550 * 5551 * Reexecute is top-down. When an i/o fails, if it's not 5552 * the root, it simply notifies its parent and sticks around. 5553 * The parent, seeing that it still has children in zio_done(), 5554 * does the same. This percolates all the way up to the root. 5555 * The root i/o will reexecute or suspend the entire tree. 5556 * 5557 * This approach ensures that zio_reexecute() honors 5558 * all the original i/o dependency relationships, e.g. 5559 * parents not executing until children are ready. 5560 */ 5561 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 5562 5563 zio->io_gang_leader = NULL; 5564 5565 mutex_enter(&zio->io_lock); 5566 zio->io_state[ZIO_WAIT_DONE] = 1; 5567 mutex_exit(&zio->io_lock); 5568 5569 /* 5570 * "The Godfather" I/O monitors its children but is 5571 * not a true parent to them. It will track them through 5572 * the pipeline but severs its ties whenever they get into 5573 * trouble (e.g. suspended). This allows "The Godfather" 5574 * I/O to return status without blocking. 5575 */ 5576 zl = NULL; 5577 for (pio = zio_walk_parents(zio, &zl); pio != NULL; 5578 pio = pio_next) { 5579 zio_link_t *remove_zl = zl; 5580 pio_next = zio_walk_parents(zio, &zl); 5581 5582 if ((pio->io_flags & ZIO_FLAG_GODFATHER) && 5583 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) { 5584 zio_remove_child(pio, zio, remove_zl); 5585 /* 5586 * This is a rare code path, so we don't 5587 * bother with "next_to_execute". 5588 */ 5589 zio_notify_parent(pio, zio, ZIO_WAIT_DONE, 5590 NULL); 5591 } 5592 } 5593 5594 if ((pio = zio_unique_parent(zio)) != NULL) { 5595 /* 5596 * We're not a root i/o, so there's nothing to do 5597 * but notify our parent. Don't propagate errors 5598 * upward since we haven't permanently failed yet. 5599 */ 5600 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); 5601 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE; 5602 /* 5603 * This is a rare code path, so we don't bother with 5604 * "next_to_execute". 5605 */ 5606 zio_notify_parent(pio, zio, ZIO_WAIT_DONE, NULL); 5607 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) { 5608 /* 5609 * We'd fail again if we reexecuted now, so suspend 5610 * until conditions improve (e.g. device comes online). 5611 */ 5612 zio_suspend(zio->io_spa, zio, ZIO_SUSPEND_IOERR); 5613 } else { 5614 /* 5615 * Reexecution is potentially a huge amount of work. 5616 * Hand it off to the otherwise-unused claim taskq. 5617 */ 5618 spa_taskq_dispatch(zio->io_spa, 5619 ZIO_TYPE_CLAIM, ZIO_TASKQ_ISSUE, 5620 zio_reexecute, zio, B_FALSE); 5621 } 5622 return (NULL); 5623 } 5624 5625 ASSERT(list_is_empty(&zio->io_child_list)); 5626 ASSERT(zio->io_reexecute == 0); 5627 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL)); 5628 5629 /* 5630 * Report any checksum errors, since the I/O is complete. 5631 */ 5632 while (zio->io_cksum_report != NULL) { 5633 zio_cksum_report_t *zcr = zio->io_cksum_report; 5634 zio->io_cksum_report = zcr->zcr_next; 5635 zcr->zcr_next = NULL; 5636 zcr->zcr_finish(zcr, NULL); 5637 zfs_ereport_free_checksum(zcr); 5638 } 5639 5640 /* 5641 * It is the responsibility of the done callback to ensure that this 5642 * particular zio is no longer discoverable for adoption, and as 5643 * such, cannot acquire any new parents. 5644 */ 5645 if (zio->io_done) 5646 zio->io_done(zio); 5647 5648 mutex_enter(&zio->io_lock); 5649 zio->io_state[ZIO_WAIT_DONE] = 1; 5650 mutex_exit(&zio->io_lock); 5651 5652 /* 5653 * We are done executing this zio. We may want to execute a parent 5654 * next. See the comment in zio_notify_parent(). 5655 */ 5656 zio_t *next_to_execute = NULL; 5657 zl = NULL; 5658 for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) { 5659 zio_link_t *remove_zl = zl; 5660 pio_next = zio_walk_parents(zio, &zl); 5661 zio_remove_child(pio, zio, remove_zl); 5662 zio_notify_parent(pio, zio, ZIO_WAIT_DONE, &next_to_execute); 5663 } 5664 5665 if (zio->io_waiter != NULL) { 5666 mutex_enter(&zio->io_lock); 5667 zio->io_executor = NULL; 5668 cv_broadcast(&zio->io_cv); 5669 mutex_exit(&zio->io_lock); 5670 } else { 5671 zio_destroy(zio); 5672 } 5673 5674 return (next_to_execute); 5675 } 5676 5677 /* 5678 * ========================================================================== 5679 * I/O pipeline definition 5680 * ========================================================================== 5681 */ 5682 static zio_pipe_stage_t *zio_pipeline[] = { 5683 NULL, 5684 zio_read_bp_init, 5685 zio_write_bp_init, 5686 zio_free_bp_init, 5687 zio_issue_async, 5688 zio_write_compress, 5689 zio_encrypt, 5690 zio_checksum_generate, 5691 zio_nop_write, 5692 zio_brt_free, 5693 zio_ddt_read_start, 5694 zio_ddt_read_done, 5695 zio_ddt_write, 5696 zio_ddt_free, 5697 zio_gang_assemble, 5698 zio_gang_issue, 5699 zio_dva_throttle, 5700 zio_dva_allocate, 5701 zio_dva_free, 5702 zio_dva_claim, 5703 zio_ready, 5704 zio_vdev_io_start, 5705 zio_vdev_io_done, 5706 zio_vdev_io_assess, 5707 zio_checksum_verify, 5708 zio_dio_checksum_verify, 5709 zio_done 5710 }; 5711 5712 5713 5714 5715 /* 5716 * Compare two zbookmark_phys_t's to see which we would reach first in a 5717 * pre-order traversal of the object tree. 5718 * 5719 * This is simple in every case aside from the meta-dnode object. For all other 5720 * objects, we traverse them in order (object 1 before object 2, and so on). 5721 * However, all of these objects are traversed while traversing object 0, since 5722 * the data it points to is the list of objects. Thus, we need to convert to a 5723 * canonical representation so we can compare meta-dnode bookmarks to 5724 * non-meta-dnode bookmarks. 5725 * 5726 * We do this by calculating "equivalents" for each field of the zbookmark. 5727 * zbookmarks outside of the meta-dnode use their own object and level, and 5728 * calculate the level 0 equivalent (the first L0 blkid that is contained in the 5729 * blocks this bookmark refers to) by multiplying their blkid by their span 5730 * (the number of L0 blocks contained within one block at their level). 5731 * zbookmarks inside the meta-dnode calculate their object equivalent 5732 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use 5733 * level + 1<<31 (any value larger than a level could ever be) for their level. 5734 * This causes them to always compare before a bookmark in their object 5735 * equivalent, compare appropriately to bookmarks in other objects, and to 5736 * compare appropriately to other bookmarks in the meta-dnode. 5737 */ 5738 int 5739 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2, 5740 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2) 5741 { 5742 /* 5743 * These variables represent the "equivalent" values for the zbookmark, 5744 * after converting zbookmarks inside the meta dnode to their 5745 * normal-object equivalents. 5746 */ 5747 uint64_t zb1obj, zb2obj; 5748 uint64_t zb1L0, zb2L0; 5749 uint64_t zb1level, zb2level; 5750 5751 if (zb1->zb_object == zb2->zb_object && 5752 zb1->zb_level == zb2->zb_level && 5753 zb1->zb_blkid == zb2->zb_blkid) 5754 return (0); 5755 5756 IMPLY(zb1->zb_level > 0, ibs1 >= SPA_MINBLOCKSHIFT); 5757 IMPLY(zb2->zb_level > 0, ibs2 >= SPA_MINBLOCKSHIFT); 5758 5759 /* 5760 * BP_SPANB calculates the span in blocks. 5761 */ 5762 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level); 5763 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level); 5764 5765 if (zb1->zb_object == DMU_META_DNODE_OBJECT) { 5766 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT)); 5767 zb1L0 = 0; 5768 zb1level = zb1->zb_level + COMPARE_META_LEVEL; 5769 } else { 5770 zb1obj = zb1->zb_object; 5771 zb1level = zb1->zb_level; 5772 } 5773 5774 if (zb2->zb_object == DMU_META_DNODE_OBJECT) { 5775 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT)); 5776 zb2L0 = 0; 5777 zb2level = zb2->zb_level + COMPARE_META_LEVEL; 5778 } else { 5779 zb2obj = zb2->zb_object; 5780 zb2level = zb2->zb_level; 5781 } 5782 5783 /* Now that we have a canonical representation, do the comparison. */ 5784 if (zb1obj != zb2obj) 5785 return (zb1obj < zb2obj ? -1 : 1); 5786 else if (zb1L0 != zb2L0) 5787 return (zb1L0 < zb2L0 ? -1 : 1); 5788 else if (zb1level != zb2level) 5789 return (zb1level > zb2level ? -1 : 1); 5790 /* 5791 * This can (theoretically) happen if the bookmarks have the same object 5792 * and level, but different blkids, if the block sizes are not the same. 5793 * There is presently no way to change the indirect block sizes 5794 */ 5795 return (0); 5796 } 5797 5798 /* 5799 * This function checks the following: given that last_block is the place that 5800 * our traversal stopped last time, does that guarantee that we've visited 5801 * every node under subtree_root? Therefore, we can't just use the raw output 5802 * of zbookmark_compare. We have to pass in a modified version of 5803 * subtree_root; by incrementing the block id, and then checking whether 5804 * last_block is before or equal to that, we can tell whether or not having 5805 * visited last_block implies that all of subtree_root's children have been 5806 * visited. 5807 */ 5808 boolean_t 5809 zbookmark_subtree_completed(const dnode_phys_t *dnp, 5810 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block) 5811 { 5812 zbookmark_phys_t mod_zb = *subtree_root; 5813 mod_zb.zb_blkid++; 5814 ASSERT0(last_block->zb_level); 5815 5816 /* The objset_phys_t isn't before anything. */ 5817 if (dnp == NULL) 5818 return (B_FALSE); 5819 5820 /* 5821 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the 5822 * data block size in sectors, because that variable is only used if 5823 * the bookmark refers to a block in the meta-dnode. Since we don't 5824 * know without examining it what object it refers to, and there's no 5825 * harm in passing in this value in other cases, we always pass it in. 5826 * 5827 * We pass in 0 for the indirect block size shift because zb2 must be 5828 * level 0. The indirect block size is only used to calculate the span 5829 * of the bookmark, but since the bookmark must be level 0, the span is 5830 * always 1, so the math works out. 5831 * 5832 * If you make changes to how the zbookmark_compare code works, be sure 5833 * to make sure that this code still works afterwards. 5834 */ 5835 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift, 5836 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb, 5837 last_block) <= 0); 5838 } 5839 5840 /* 5841 * This function is similar to zbookmark_subtree_completed(), but returns true 5842 * if subtree_root is equal or ahead of last_block, i.e. still to be done. 5843 */ 5844 boolean_t 5845 zbookmark_subtree_tbd(const dnode_phys_t *dnp, 5846 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block) 5847 { 5848 ASSERT0(last_block->zb_level); 5849 if (dnp == NULL) 5850 return (B_FALSE); 5851 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift, 5852 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, subtree_root, 5853 last_block) >= 0); 5854 } 5855 5856 EXPORT_SYMBOL(zio_type_name); 5857 EXPORT_SYMBOL(zio_buf_alloc); 5858 EXPORT_SYMBOL(zio_data_buf_alloc); 5859 EXPORT_SYMBOL(zio_buf_free); 5860 EXPORT_SYMBOL(zio_data_buf_free); 5861 5862 ZFS_MODULE_PARAM(zfs_zio, zio_, slow_io_ms, INT, ZMOD_RW, 5863 "Max I/O completion time (milliseconds) before marking it as slow"); 5864 5865 ZFS_MODULE_PARAM(zfs_zio, zio_, requeue_io_start_cut_in_line, INT, ZMOD_RW, 5866 "Prioritize requeued I/O"); 5867 5868 ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_deferred_free, UINT, ZMOD_RW, 5869 "Defer frees starting in this pass"); 5870 5871 ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_dont_compress, UINT, ZMOD_RW, 5872 "Don't compress starting in this pass"); 5873 5874 ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_rewrite, UINT, ZMOD_RW, 5875 "Rewrite new bps starting in this pass"); 5876 5877 ZFS_MODULE_PARAM(zfs_zio, zio_, dva_throttle_enabled, INT, ZMOD_RW, 5878 "Throttle block allocations in the ZIO pipeline"); 5879 5880 ZFS_MODULE_PARAM(zfs_zio, zio_, deadman_log_all, INT, ZMOD_RW, 5881 "Log all slow ZIOs, not just those with vdevs"); 5882