1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or https://opensource.org/licenses/CDDL-1.0. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21 /* 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23 * Copyright (c) 2011, 2022 by Delphix. All rights reserved. 24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved. 25 * Copyright (c) 2017, Intel Corporation. 26 * Copyright (c) 2019, Klara Inc. 27 * Copyright (c) 2019, Allan Jude 28 * Copyright (c) 2021, Datto, Inc. 29 */ 30 31 #include <sys/sysmacros.h> 32 #include <sys/zfs_context.h> 33 #include <sys/fm/fs/zfs.h> 34 #include <sys/spa.h> 35 #include <sys/txg.h> 36 #include <sys/spa_impl.h> 37 #include <sys/vdev_impl.h> 38 #include <sys/vdev_trim.h> 39 #include <sys/zio_impl.h> 40 #include <sys/zio_compress.h> 41 #include <sys/zio_checksum.h> 42 #include <sys/dmu_objset.h> 43 #include <sys/arc.h> 44 #include <sys/ddt.h> 45 #include <sys/blkptr.h> 46 #include <sys/zfeature.h> 47 #include <sys/dsl_scan.h> 48 #include <sys/metaslab_impl.h> 49 #include <sys/time.h> 50 #include <sys/trace_zfs.h> 51 #include <sys/abd.h> 52 #include <sys/dsl_crypt.h> 53 #include <cityhash.h> 54 55 /* 56 * ========================================================================== 57 * I/O type descriptions 58 * ========================================================================== 59 */ 60 const char *const zio_type_name[ZIO_TYPES] = { 61 /* 62 * Note: Linux kernel thread name length is limited 63 * so these names will differ from upstream open zfs. 64 */ 65 "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl", "z_trim" 66 }; 67 68 int zio_dva_throttle_enabled = B_TRUE; 69 static int zio_deadman_log_all = B_FALSE; 70 71 /* 72 * ========================================================================== 73 * I/O kmem caches 74 * ========================================================================== 75 */ 76 static kmem_cache_t *zio_cache; 77 static kmem_cache_t *zio_link_cache; 78 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; 79 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; 80 #if defined(ZFS_DEBUG) && !defined(_KERNEL) 81 static uint64_t zio_buf_cache_allocs[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; 82 static uint64_t zio_buf_cache_frees[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; 83 #endif 84 85 /* Mark IOs as "slow" if they take longer than 30 seconds */ 86 static uint_t zio_slow_io_ms = (30 * MILLISEC); 87 88 #define BP_SPANB(indblkshift, level) \ 89 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT))) 90 #define COMPARE_META_LEVEL 0x80000000ul 91 /* 92 * The following actions directly effect the spa's sync-to-convergence logic. 93 * The values below define the sync pass when we start performing the action. 94 * Care should be taken when changing these values as they directly impact 95 * spa_sync() performance. Tuning these values may introduce subtle performance 96 * pathologies and should only be done in the context of performance analysis. 97 * These tunables will eventually be removed and replaced with #defines once 98 * enough analysis has been done to determine optimal values. 99 * 100 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that 101 * regular blocks are not deferred. 102 * 103 * Starting in sync pass 8 (zfs_sync_pass_dont_compress), we disable 104 * compression (including of metadata). In practice, we don't have this 105 * many sync passes, so this has no effect. 106 * 107 * The original intent was that disabling compression would help the sync 108 * passes to converge. However, in practice disabling compression increases 109 * the average number of sync passes, because when we turn compression off, a 110 * lot of block's size will change and thus we have to re-allocate (not 111 * overwrite) them. It also increases the number of 128KB allocations (e.g. 112 * for indirect blocks and spacemaps) because these will not be compressed. 113 * The 128K allocations are especially detrimental to performance on highly 114 * fragmented systems, which may have very few free segments of this size, 115 * and may need to load new metaslabs to satisfy 128K allocations. 116 */ 117 118 /* defer frees starting in this pass */ 119 uint_t zfs_sync_pass_deferred_free = 2; 120 121 /* don't compress starting in this pass */ 122 static uint_t zfs_sync_pass_dont_compress = 8; 123 124 /* rewrite new bps starting in this pass */ 125 static uint_t zfs_sync_pass_rewrite = 2; 126 127 /* 128 * An allocating zio is one that either currently has the DVA allocate 129 * stage set or will have it later in its lifetime. 130 */ 131 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE) 132 133 /* 134 * Enable smaller cores by excluding metadata 135 * allocations as well. 136 */ 137 int zio_exclude_metadata = 0; 138 static int zio_requeue_io_start_cut_in_line = 1; 139 140 #ifdef ZFS_DEBUG 141 static const int zio_buf_debug_limit = 16384; 142 #else 143 static const int zio_buf_debug_limit = 0; 144 #endif 145 146 static inline void __zio_execute(zio_t *zio); 147 148 static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t); 149 150 void 151 zio_init(void) 152 { 153 size_t c; 154 155 zio_cache = kmem_cache_create("zio_cache", 156 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0); 157 zio_link_cache = kmem_cache_create("zio_link_cache", 158 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0); 159 160 /* 161 * For small buffers, we want a cache for each multiple of 162 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache 163 * for each quarter-power of 2. 164 */ 165 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { 166 size_t size = (c + 1) << SPA_MINBLOCKSHIFT; 167 size_t p2 = size; 168 size_t align = 0; 169 size_t data_cflags, cflags; 170 171 data_cflags = KMC_NODEBUG; 172 cflags = (zio_exclude_metadata || size > zio_buf_debug_limit) ? 173 KMC_NODEBUG : 0; 174 175 while (!ISP2(p2)) 176 p2 &= p2 - 1; 177 178 #ifndef _KERNEL 179 /* 180 * If we are using watchpoints, put each buffer on its own page, 181 * to eliminate the performance overhead of trapping to the 182 * kernel when modifying a non-watched buffer that shares the 183 * page with a watched buffer. 184 */ 185 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE)) 186 continue; 187 /* 188 * Here's the problem - on 4K native devices in userland on 189 * Linux using O_DIRECT, buffers must be 4K aligned or I/O 190 * will fail with EINVAL, causing zdb (and others) to coredump. 191 * Since userland probably doesn't need optimized buffer caches, 192 * we just force 4K alignment on everything. 193 */ 194 align = 8 * SPA_MINBLOCKSIZE; 195 #else 196 if (size < PAGESIZE) { 197 align = SPA_MINBLOCKSIZE; 198 } else if (IS_P2ALIGNED(size, p2 >> 2)) { 199 align = PAGESIZE; 200 } 201 #endif 202 203 if (align != 0) { 204 char name[36]; 205 if (cflags == data_cflags) { 206 /* 207 * Resulting kmem caches would be identical. 208 * Save memory by creating only one. 209 */ 210 (void) snprintf(name, sizeof (name), 211 "zio_buf_comb_%lu", (ulong_t)size); 212 zio_buf_cache[c] = kmem_cache_create(name, 213 size, align, NULL, NULL, NULL, NULL, NULL, 214 cflags); 215 zio_data_buf_cache[c] = zio_buf_cache[c]; 216 continue; 217 } 218 (void) snprintf(name, sizeof (name), "zio_buf_%lu", 219 (ulong_t)size); 220 zio_buf_cache[c] = kmem_cache_create(name, size, 221 align, NULL, NULL, NULL, NULL, NULL, cflags); 222 223 (void) snprintf(name, sizeof (name), "zio_data_buf_%lu", 224 (ulong_t)size); 225 zio_data_buf_cache[c] = kmem_cache_create(name, size, 226 align, NULL, NULL, NULL, NULL, NULL, data_cflags); 227 } 228 } 229 230 while (--c != 0) { 231 ASSERT(zio_buf_cache[c] != NULL); 232 if (zio_buf_cache[c - 1] == NULL) 233 zio_buf_cache[c - 1] = zio_buf_cache[c]; 234 235 ASSERT(zio_data_buf_cache[c] != NULL); 236 if (zio_data_buf_cache[c - 1] == NULL) 237 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c]; 238 } 239 240 zio_inject_init(); 241 242 lz4_init(); 243 } 244 245 void 246 zio_fini(void) 247 { 248 size_t n = SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; 249 250 #if defined(ZFS_DEBUG) && !defined(_KERNEL) 251 for (size_t i = 0; i < n; i++) { 252 if (zio_buf_cache_allocs[i] != zio_buf_cache_frees[i]) 253 (void) printf("zio_fini: [%d] %llu != %llu\n", 254 (int)((i + 1) << SPA_MINBLOCKSHIFT), 255 (long long unsigned)zio_buf_cache_allocs[i], 256 (long long unsigned)zio_buf_cache_frees[i]); 257 } 258 #endif 259 260 /* 261 * The same kmem cache can show up multiple times in both zio_buf_cache 262 * and zio_data_buf_cache. Do a wasteful but trivially correct scan to 263 * sort it out. 264 */ 265 for (size_t i = 0; i < n; i++) { 266 kmem_cache_t *cache = zio_buf_cache[i]; 267 if (cache == NULL) 268 continue; 269 for (size_t j = i; j < n; j++) { 270 if (cache == zio_buf_cache[j]) 271 zio_buf_cache[j] = NULL; 272 if (cache == zio_data_buf_cache[j]) 273 zio_data_buf_cache[j] = NULL; 274 } 275 kmem_cache_destroy(cache); 276 } 277 278 for (size_t i = 0; i < n; i++) { 279 kmem_cache_t *cache = zio_data_buf_cache[i]; 280 if (cache == NULL) 281 continue; 282 for (size_t j = i; j < n; j++) { 283 if (cache == zio_data_buf_cache[j]) 284 zio_data_buf_cache[j] = NULL; 285 } 286 kmem_cache_destroy(cache); 287 } 288 289 for (size_t i = 0; i < n; i++) { 290 VERIFY3P(zio_buf_cache[i], ==, NULL); 291 VERIFY3P(zio_data_buf_cache[i], ==, NULL); 292 } 293 294 kmem_cache_destroy(zio_link_cache); 295 kmem_cache_destroy(zio_cache); 296 297 zio_inject_fini(); 298 299 lz4_fini(); 300 } 301 302 /* 303 * ========================================================================== 304 * Allocate and free I/O buffers 305 * ========================================================================== 306 */ 307 308 /* 309 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a 310 * crashdump if the kernel panics, so use it judiciously. Obviously, it's 311 * useful to inspect ZFS metadata, but if possible, we should avoid keeping 312 * excess / transient data in-core during a crashdump. 313 */ 314 void * 315 zio_buf_alloc(size_t size) 316 { 317 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 318 319 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 320 #if defined(ZFS_DEBUG) && !defined(_KERNEL) 321 atomic_add_64(&zio_buf_cache_allocs[c], 1); 322 #endif 323 324 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE)); 325 } 326 327 /* 328 * Use zio_data_buf_alloc to allocate data. The data will not appear in a 329 * crashdump if the kernel panics. This exists so that we will limit the amount 330 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount 331 * of kernel heap dumped to disk when the kernel panics) 332 */ 333 void * 334 zio_data_buf_alloc(size_t size) 335 { 336 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 337 338 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 339 340 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE)); 341 } 342 343 void 344 zio_buf_free(void *buf, size_t size) 345 { 346 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 347 348 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 349 #if defined(ZFS_DEBUG) && !defined(_KERNEL) 350 atomic_add_64(&zio_buf_cache_frees[c], 1); 351 #endif 352 353 kmem_cache_free(zio_buf_cache[c], buf); 354 } 355 356 void 357 zio_data_buf_free(void *buf, size_t size) 358 { 359 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 360 361 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 362 363 kmem_cache_free(zio_data_buf_cache[c], buf); 364 } 365 366 static void 367 zio_abd_free(void *abd, size_t size) 368 { 369 (void) size; 370 abd_free((abd_t *)abd); 371 } 372 373 /* 374 * ========================================================================== 375 * Push and pop I/O transform buffers 376 * ========================================================================== 377 */ 378 void 379 zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize, 380 zio_transform_func_t *transform) 381 { 382 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP); 383 384 zt->zt_orig_abd = zio->io_abd; 385 zt->zt_orig_size = zio->io_size; 386 zt->zt_bufsize = bufsize; 387 zt->zt_transform = transform; 388 389 zt->zt_next = zio->io_transform_stack; 390 zio->io_transform_stack = zt; 391 392 zio->io_abd = data; 393 zio->io_size = size; 394 } 395 396 void 397 zio_pop_transforms(zio_t *zio) 398 { 399 zio_transform_t *zt; 400 401 while ((zt = zio->io_transform_stack) != NULL) { 402 if (zt->zt_transform != NULL) 403 zt->zt_transform(zio, 404 zt->zt_orig_abd, zt->zt_orig_size); 405 406 if (zt->zt_bufsize != 0) 407 abd_free(zio->io_abd); 408 409 zio->io_abd = zt->zt_orig_abd; 410 zio->io_size = zt->zt_orig_size; 411 zio->io_transform_stack = zt->zt_next; 412 413 kmem_free(zt, sizeof (zio_transform_t)); 414 } 415 } 416 417 /* 418 * ========================================================================== 419 * I/O transform callbacks for subblocks, decompression, and decryption 420 * ========================================================================== 421 */ 422 static void 423 zio_subblock(zio_t *zio, abd_t *data, uint64_t size) 424 { 425 ASSERT(zio->io_size > size); 426 427 if (zio->io_type == ZIO_TYPE_READ) 428 abd_copy(data, zio->io_abd, size); 429 } 430 431 static void 432 zio_decompress(zio_t *zio, abd_t *data, uint64_t size) 433 { 434 if (zio->io_error == 0) { 435 void *tmp = abd_borrow_buf(data, size); 436 int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp), 437 zio->io_abd, tmp, zio->io_size, size, 438 &zio->io_prop.zp_complevel); 439 abd_return_buf_copy(data, tmp, size); 440 441 if (zio_injection_enabled && ret == 0) 442 ret = zio_handle_fault_injection(zio, EINVAL); 443 444 if (ret != 0) 445 zio->io_error = SET_ERROR(EIO); 446 } 447 } 448 449 static void 450 zio_decrypt(zio_t *zio, abd_t *data, uint64_t size) 451 { 452 int ret; 453 void *tmp; 454 blkptr_t *bp = zio->io_bp; 455 spa_t *spa = zio->io_spa; 456 uint64_t dsobj = zio->io_bookmark.zb_objset; 457 uint64_t lsize = BP_GET_LSIZE(bp); 458 dmu_object_type_t ot = BP_GET_TYPE(bp); 459 uint8_t salt[ZIO_DATA_SALT_LEN]; 460 uint8_t iv[ZIO_DATA_IV_LEN]; 461 uint8_t mac[ZIO_DATA_MAC_LEN]; 462 boolean_t no_crypt = B_FALSE; 463 464 ASSERT(BP_USES_CRYPT(bp)); 465 ASSERT3U(size, !=, 0); 466 467 if (zio->io_error != 0) 468 return; 469 470 /* 471 * Verify the cksum of MACs stored in an indirect bp. It will always 472 * be possible to verify this since it does not require an encryption 473 * key. 474 */ 475 if (BP_HAS_INDIRECT_MAC_CKSUM(bp)) { 476 zio_crypt_decode_mac_bp(bp, mac); 477 478 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF) { 479 /* 480 * We haven't decompressed the data yet, but 481 * zio_crypt_do_indirect_mac_checksum() requires 482 * decompressed data to be able to parse out the MACs 483 * from the indirect block. We decompress it now and 484 * throw away the result after we are finished. 485 */ 486 tmp = zio_buf_alloc(lsize); 487 ret = zio_decompress_data(BP_GET_COMPRESS(bp), 488 zio->io_abd, tmp, zio->io_size, lsize, 489 &zio->io_prop.zp_complevel); 490 if (ret != 0) { 491 ret = SET_ERROR(EIO); 492 goto error; 493 } 494 ret = zio_crypt_do_indirect_mac_checksum(B_FALSE, 495 tmp, lsize, BP_SHOULD_BYTESWAP(bp), mac); 496 zio_buf_free(tmp, lsize); 497 } else { 498 ret = zio_crypt_do_indirect_mac_checksum_abd(B_FALSE, 499 zio->io_abd, size, BP_SHOULD_BYTESWAP(bp), mac); 500 } 501 abd_copy(data, zio->io_abd, size); 502 503 if (zio_injection_enabled && ot != DMU_OT_DNODE && ret == 0) { 504 ret = zio_handle_decrypt_injection(spa, 505 &zio->io_bookmark, ot, ECKSUM); 506 } 507 if (ret != 0) 508 goto error; 509 510 return; 511 } 512 513 /* 514 * If this is an authenticated block, just check the MAC. It would be 515 * nice to separate this out into its own flag, but when this was done, 516 * we had run out of bits in what is now zio_flag_t. Future cleanup 517 * could make this a flag bit. 518 */ 519 if (BP_IS_AUTHENTICATED(bp)) { 520 if (ot == DMU_OT_OBJSET) { 521 ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa, 522 dsobj, zio->io_abd, size, BP_SHOULD_BYTESWAP(bp)); 523 } else { 524 zio_crypt_decode_mac_bp(bp, mac); 525 ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj, 526 zio->io_abd, size, mac); 527 if (zio_injection_enabled && ret == 0) { 528 ret = zio_handle_decrypt_injection(spa, 529 &zio->io_bookmark, ot, ECKSUM); 530 } 531 } 532 abd_copy(data, zio->io_abd, size); 533 534 if (ret != 0) 535 goto error; 536 537 return; 538 } 539 540 zio_crypt_decode_params_bp(bp, salt, iv); 541 542 if (ot == DMU_OT_INTENT_LOG) { 543 tmp = abd_borrow_buf_copy(zio->io_abd, sizeof (zil_chain_t)); 544 zio_crypt_decode_mac_zil(tmp, mac); 545 abd_return_buf(zio->io_abd, tmp, sizeof (zil_chain_t)); 546 } else { 547 zio_crypt_decode_mac_bp(bp, mac); 548 } 549 550 ret = spa_do_crypt_abd(B_FALSE, spa, &zio->io_bookmark, BP_GET_TYPE(bp), 551 BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp), salt, iv, mac, size, data, 552 zio->io_abd, &no_crypt); 553 if (no_crypt) 554 abd_copy(data, zio->io_abd, size); 555 556 if (ret != 0) 557 goto error; 558 559 return; 560 561 error: 562 /* assert that the key was found unless this was speculative */ 563 ASSERT(ret != EACCES || (zio->io_flags & ZIO_FLAG_SPECULATIVE)); 564 565 /* 566 * If there was a decryption / authentication error return EIO as 567 * the io_error. If this was not a speculative zio, create an ereport. 568 */ 569 if (ret == ECKSUM) { 570 zio->io_error = SET_ERROR(EIO); 571 if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) { 572 spa_log_error(spa, &zio->io_bookmark); 573 (void) zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION, 574 spa, NULL, &zio->io_bookmark, zio, 0); 575 } 576 } else { 577 zio->io_error = ret; 578 } 579 } 580 581 /* 582 * ========================================================================== 583 * I/O parent/child relationships and pipeline interlocks 584 * ========================================================================== 585 */ 586 zio_t * 587 zio_walk_parents(zio_t *cio, zio_link_t **zl) 588 { 589 list_t *pl = &cio->io_parent_list; 590 591 *zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl); 592 if (*zl == NULL) 593 return (NULL); 594 595 ASSERT((*zl)->zl_child == cio); 596 return ((*zl)->zl_parent); 597 } 598 599 zio_t * 600 zio_walk_children(zio_t *pio, zio_link_t **zl) 601 { 602 list_t *cl = &pio->io_child_list; 603 604 ASSERT(MUTEX_HELD(&pio->io_lock)); 605 606 *zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl); 607 if (*zl == NULL) 608 return (NULL); 609 610 ASSERT((*zl)->zl_parent == pio); 611 return ((*zl)->zl_child); 612 } 613 614 zio_t * 615 zio_unique_parent(zio_t *cio) 616 { 617 zio_link_t *zl = NULL; 618 zio_t *pio = zio_walk_parents(cio, &zl); 619 620 VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL); 621 return (pio); 622 } 623 624 void 625 zio_add_child(zio_t *pio, zio_t *cio) 626 { 627 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP); 628 629 /* 630 * Logical I/Os can have logical, gang, or vdev children. 631 * Gang I/Os can have gang or vdev children. 632 * Vdev I/Os can only have vdev children. 633 * The following ASSERT captures all of these constraints. 634 */ 635 ASSERT3S(cio->io_child_type, <=, pio->io_child_type); 636 637 zl->zl_parent = pio; 638 zl->zl_child = cio; 639 640 mutex_enter(&pio->io_lock); 641 mutex_enter(&cio->io_lock); 642 643 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0); 644 645 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 646 pio->io_children[cio->io_child_type][w] += !cio->io_state[w]; 647 648 list_insert_head(&pio->io_child_list, zl); 649 list_insert_head(&cio->io_parent_list, zl); 650 651 pio->io_child_count++; 652 cio->io_parent_count++; 653 654 mutex_exit(&cio->io_lock); 655 mutex_exit(&pio->io_lock); 656 } 657 658 static void 659 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl) 660 { 661 ASSERT(zl->zl_parent == pio); 662 ASSERT(zl->zl_child == cio); 663 664 mutex_enter(&pio->io_lock); 665 mutex_enter(&cio->io_lock); 666 667 list_remove(&pio->io_child_list, zl); 668 list_remove(&cio->io_parent_list, zl); 669 670 pio->io_child_count--; 671 cio->io_parent_count--; 672 673 mutex_exit(&cio->io_lock); 674 mutex_exit(&pio->io_lock); 675 kmem_cache_free(zio_link_cache, zl); 676 } 677 678 static boolean_t 679 zio_wait_for_children(zio_t *zio, uint8_t childbits, enum zio_wait_type wait) 680 { 681 boolean_t waiting = B_FALSE; 682 683 mutex_enter(&zio->io_lock); 684 ASSERT(zio->io_stall == NULL); 685 for (int c = 0; c < ZIO_CHILD_TYPES; c++) { 686 if (!(ZIO_CHILD_BIT_IS_SET(childbits, c))) 687 continue; 688 689 uint64_t *countp = &zio->io_children[c][wait]; 690 if (*countp != 0) { 691 zio->io_stage >>= 1; 692 ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN); 693 zio->io_stall = countp; 694 waiting = B_TRUE; 695 break; 696 } 697 } 698 mutex_exit(&zio->io_lock); 699 return (waiting); 700 } 701 702 __attribute__((always_inline)) 703 static inline void 704 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait, 705 zio_t **next_to_executep) 706 { 707 uint64_t *countp = &pio->io_children[zio->io_child_type][wait]; 708 int *errorp = &pio->io_child_error[zio->io_child_type]; 709 710 mutex_enter(&pio->io_lock); 711 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE)) 712 *errorp = zio_worst_error(*errorp, zio->io_error); 713 pio->io_reexecute |= zio->io_reexecute; 714 ASSERT3U(*countp, >, 0); 715 716 (*countp)--; 717 718 if (*countp == 0 && pio->io_stall == countp) { 719 zio_taskq_type_t type = 720 pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE : 721 ZIO_TASKQ_INTERRUPT; 722 pio->io_stall = NULL; 723 mutex_exit(&pio->io_lock); 724 725 /* 726 * If we can tell the caller to execute this parent next, do 727 * so. We only do this if the parent's zio type matches the 728 * child's type. Otherwise dispatch the parent zio in its 729 * own taskq. 730 * 731 * Having the caller execute the parent when possible reduces 732 * locking on the zio taskq's, reduces context switch 733 * overhead, and has no recursion penalty. Note that one 734 * read from disk typically causes at least 3 zio's: a 735 * zio_null(), the logical zio_read(), and then a physical 736 * zio. When the physical ZIO completes, we are able to call 737 * zio_done() on all 3 of these zio's from one invocation of 738 * zio_execute() by returning the parent back to 739 * zio_execute(). Since the parent isn't executed until this 740 * thread returns back to zio_execute(), the caller should do 741 * so promptly. 742 * 743 * In other cases, dispatching the parent prevents 744 * overflowing the stack when we have deeply nested 745 * parent-child relationships, as we do with the "mega zio" 746 * of writes for spa_sync(), and the chain of ZIL blocks. 747 */ 748 if (next_to_executep != NULL && *next_to_executep == NULL && 749 pio->io_type == zio->io_type) { 750 *next_to_executep = pio; 751 } else { 752 zio_taskq_dispatch(pio, type, B_FALSE); 753 } 754 } else { 755 mutex_exit(&pio->io_lock); 756 } 757 } 758 759 static void 760 zio_inherit_child_errors(zio_t *zio, enum zio_child c) 761 { 762 if (zio->io_child_error[c] != 0 && zio->io_error == 0) 763 zio->io_error = zio->io_child_error[c]; 764 } 765 766 int 767 zio_bookmark_compare(const void *x1, const void *x2) 768 { 769 const zio_t *z1 = x1; 770 const zio_t *z2 = x2; 771 772 if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset) 773 return (-1); 774 if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset) 775 return (1); 776 777 if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object) 778 return (-1); 779 if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object) 780 return (1); 781 782 if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level) 783 return (-1); 784 if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level) 785 return (1); 786 787 if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid) 788 return (-1); 789 if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid) 790 return (1); 791 792 if (z1 < z2) 793 return (-1); 794 if (z1 > z2) 795 return (1); 796 797 return (0); 798 } 799 800 /* 801 * ========================================================================== 802 * Create the various types of I/O (read, write, free, etc) 803 * ========================================================================== 804 */ 805 static zio_t * 806 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 807 abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done, 808 void *private, zio_type_t type, zio_priority_t priority, 809 zio_flag_t flags, vdev_t *vd, uint64_t offset, 810 const zbookmark_phys_t *zb, enum zio_stage stage, 811 enum zio_stage pipeline) 812 { 813 zio_t *zio; 814 815 IMPLY(type != ZIO_TYPE_TRIM, psize <= SPA_MAXBLOCKSIZE); 816 ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0); 817 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0); 818 819 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER)); 820 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER)); 821 ASSERT(vd || stage == ZIO_STAGE_OPEN); 822 823 IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW_COMPRESS) != 0); 824 825 zio = kmem_cache_alloc(zio_cache, KM_SLEEP); 826 memset(zio, 0, sizeof (zio_t)); 827 828 mutex_init(&zio->io_lock, NULL, MUTEX_NOLOCKDEP, NULL); 829 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL); 830 831 list_create(&zio->io_parent_list, sizeof (zio_link_t), 832 offsetof(zio_link_t, zl_parent_node)); 833 list_create(&zio->io_child_list, sizeof (zio_link_t), 834 offsetof(zio_link_t, zl_child_node)); 835 metaslab_trace_init(&zio->io_alloc_list); 836 837 if (vd != NULL) 838 zio->io_child_type = ZIO_CHILD_VDEV; 839 else if (flags & ZIO_FLAG_GANG_CHILD) 840 zio->io_child_type = ZIO_CHILD_GANG; 841 else if (flags & ZIO_FLAG_DDT_CHILD) 842 zio->io_child_type = ZIO_CHILD_DDT; 843 else 844 zio->io_child_type = ZIO_CHILD_LOGICAL; 845 846 if (bp != NULL) { 847 zio->io_bp = (blkptr_t *)bp; 848 zio->io_bp_copy = *bp; 849 zio->io_bp_orig = *bp; 850 if (type != ZIO_TYPE_WRITE || 851 zio->io_child_type == ZIO_CHILD_DDT) 852 zio->io_bp = &zio->io_bp_copy; /* so caller can free */ 853 if (zio->io_child_type == ZIO_CHILD_LOGICAL) 854 zio->io_logical = zio; 855 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp)) 856 pipeline |= ZIO_GANG_STAGES; 857 } 858 859 zio->io_spa = spa; 860 zio->io_txg = txg; 861 zio->io_done = done; 862 zio->io_private = private; 863 zio->io_type = type; 864 zio->io_priority = priority; 865 zio->io_vd = vd; 866 zio->io_offset = offset; 867 zio->io_orig_abd = zio->io_abd = data; 868 zio->io_orig_size = zio->io_size = psize; 869 zio->io_lsize = lsize; 870 zio->io_orig_flags = zio->io_flags = flags; 871 zio->io_orig_stage = zio->io_stage = stage; 872 zio->io_orig_pipeline = zio->io_pipeline = pipeline; 873 zio->io_pipeline_trace = ZIO_STAGE_OPEN; 874 875 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY); 876 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE); 877 878 if (zb != NULL) 879 zio->io_bookmark = *zb; 880 881 if (pio != NULL) { 882 zio->io_metaslab_class = pio->io_metaslab_class; 883 if (zio->io_logical == NULL) 884 zio->io_logical = pio->io_logical; 885 if (zio->io_child_type == ZIO_CHILD_GANG) 886 zio->io_gang_leader = pio->io_gang_leader; 887 zio_add_child(pio, zio); 888 } 889 890 taskq_init_ent(&zio->io_tqent); 891 892 return (zio); 893 } 894 895 void 896 zio_destroy(zio_t *zio) 897 { 898 metaslab_trace_fini(&zio->io_alloc_list); 899 list_destroy(&zio->io_parent_list); 900 list_destroy(&zio->io_child_list); 901 mutex_destroy(&zio->io_lock); 902 cv_destroy(&zio->io_cv); 903 kmem_cache_free(zio_cache, zio); 904 } 905 906 zio_t * 907 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done, 908 void *private, zio_flag_t flags) 909 { 910 zio_t *zio; 911 912 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private, 913 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, 914 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE); 915 916 return (zio); 917 } 918 919 zio_t * 920 zio_root(spa_t *spa, zio_done_func_t *done, void *private, zio_flag_t flags) 921 { 922 return (zio_null(NULL, spa, NULL, done, private, flags)); 923 } 924 925 static int 926 zfs_blkptr_verify_log(spa_t *spa, const blkptr_t *bp, 927 enum blk_verify_flag blk_verify, const char *fmt, ...) 928 { 929 va_list adx; 930 char buf[256]; 931 932 va_start(adx, fmt); 933 (void) vsnprintf(buf, sizeof (buf), fmt, adx); 934 va_end(adx); 935 936 switch (blk_verify) { 937 case BLK_VERIFY_HALT: 938 dprintf_bp(bp, "blkptr at %p dprintf_bp():", bp); 939 zfs_panic_recover("%s: %s", spa_name(spa), buf); 940 break; 941 case BLK_VERIFY_LOG: 942 zfs_dbgmsg("%s: %s", spa_name(spa), buf); 943 break; 944 case BLK_VERIFY_ONLY: 945 break; 946 } 947 948 return (1); 949 } 950 951 /* 952 * Verify the block pointer fields contain reasonable values. This means 953 * it only contains known object types, checksum/compression identifiers, 954 * block sizes within the maximum allowed limits, valid DVAs, etc. 955 * 956 * If everything checks out B_TRUE is returned. The zfs_blkptr_verify 957 * argument controls the behavior when an invalid field is detected. 958 * 959 * Modes for zfs_blkptr_verify: 960 * 1) BLK_VERIFY_ONLY (evaluate the block) 961 * 2) BLK_VERIFY_LOG (evaluate the block and log problems) 962 * 3) BLK_VERIFY_HALT (call zfs_panic_recover on error) 963 */ 964 boolean_t 965 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp, boolean_t config_held, 966 enum blk_verify_flag blk_verify) 967 { 968 int errors = 0; 969 970 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) { 971 errors += zfs_blkptr_verify_log(spa, bp, blk_verify, 972 "blkptr at %p has invalid TYPE %llu", 973 bp, (longlong_t)BP_GET_TYPE(bp)); 974 } 975 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS) { 976 errors += zfs_blkptr_verify_log(spa, bp, blk_verify, 977 "blkptr at %p has invalid CHECKSUM %llu", 978 bp, (longlong_t)BP_GET_CHECKSUM(bp)); 979 } 980 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS) { 981 errors += zfs_blkptr_verify_log(spa, bp, blk_verify, 982 "blkptr at %p has invalid COMPRESS %llu", 983 bp, (longlong_t)BP_GET_COMPRESS(bp)); 984 } 985 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) { 986 errors += zfs_blkptr_verify_log(spa, bp, blk_verify, 987 "blkptr at %p has invalid LSIZE %llu", 988 bp, (longlong_t)BP_GET_LSIZE(bp)); 989 } 990 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) { 991 errors += zfs_blkptr_verify_log(spa, bp, blk_verify, 992 "blkptr at %p has invalid PSIZE %llu", 993 bp, (longlong_t)BP_GET_PSIZE(bp)); 994 } 995 996 if (BP_IS_EMBEDDED(bp)) { 997 if (BPE_GET_ETYPE(bp) >= NUM_BP_EMBEDDED_TYPES) { 998 errors += zfs_blkptr_verify_log(spa, bp, blk_verify, 999 "blkptr at %p has invalid ETYPE %llu", 1000 bp, (longlong_t)BPE_GET_ETYPE(bp)); 1001 } 1002 } 1003 1004 /* 1005 * Do not verify individual DVAs if the config is not trusted. This 1006 * will be done once the zio is executed in vdev_mirror_map_alloc. 1007 */ 1008 if (!spa->spa_trust_config) 1009 return (errors == 0); 1010 1011 if (!config_held) 1012 spa_config_enter(spa, SCL_VDEV, bp, RW_READER); 1013 else 1014 ASSERT(spa_config_held(spa, SCL_VDEV, RW_WRITER)); 1015 /* 1016 * Pool-specific checks. 1017 * 1018 * Note: it would be nice to verify that the blk_birth and 1019 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze() 1020 * allows the birth time of log blocks (and dmu_sync()-ed blocks 1021 * that are in the log) to be arbitrarily large. 1022 */ 1023 for (int i = 0; i < BP_GET_NDVAS(bp); i++) { 1024 const dva_t *dva = &bp->blk_dva[i]; 1025 uint64_t vdevid = DVA_GET_VDEV(dva); 1026 1027 if (vdevid >= spa->spa_root_vdev->vdev_children) { 1028 errors += zfs_blkptr_verify_log(spa, bp, blk_verify, 1029 "blkptr at %p DVA %u has invalid VDEV %llu", 1030 bp, i, (longlong_t)vdevid); 1031 continue; 1032 } 1033 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid]; 1034 if (vd == NULL) { 1035 errors += zfs_blkptr_verify_log(spa, bp, blk_verify, 1036 "blkptr at %p DVA %u has invalid VDEV %llu", 1037 bp, i, (longlong_t)vdevid); 1038 continue; 1039 } 1040 if (vd->vdev_ops == &vdev_hole_ops) { 1041 errors += zfs_blkptr_verify_log(spa, bp, blk_verify, 1042 "blkptr at %p DVA %u has hole VDEV %llu", 1043 bp, i, (longlong_t)vdevid); 1044 continue; 1045 } 1046 if (vd->vdev_ops == &vdev_missing_ops) { 1047 /* 1048 * "missing" vdevs are valid during import, but we 1049 * don't have their detailed info (e.g. asize), so 1050 * we can't perform any more checks on them. 1051 */ 1052 continue; 1053 } 1054 uint64_t offset = DVA_GET_OFFSET(dva); 1055 uint64_t asize = DVA_GET_ASIZE(dva); 1056 if (DVA_GET_GANG(dva)) 1057 asize = vdev_gang_header_asize(vd); 1058 if (offset + asize > vd->vdev_asize) { 1059 errors += zfs_blkptr_verify_log(spa, bp, blk_verify, 1060 "blkptr at %p DVA %u has invalid OFFSET %llu", 1061 bp, i, (longlong_t)offset); 1062 } 1063 } 1064 if (errors > 0) 1065 dprintf_bp(bp, "blkptr at %p dprintf_bp():", bp); 1066 if (!config_held) 1067 spa_config_exit(spa, SCL_VDEV, bp); 1068 1069 return (errors == 0); 1070 } 1071 1072 boolean_t 1073 zfs_dva_valid(spa_t *spa, const dva_t *dva, const blkptr_t *bp) 1074 { 1075 (void) bp; 1076 uint64_t vdevid = DVA_GET_VDEV(dva); 1077 1078 if (vdevid >= spa->spa_root_vdev->vdev_children) 1079 return (B_FALSE); 1080 1081 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid]; 1082 if (vd == NULL) 1083 return (B_FALSE); 1084 1085 if (vd->vdev_ops == &vdev_hole_ops) 1086 return (B_FALSE); 1087 1088 if (vd->vdev_ops == &vdev_missing_ops) { 1089 return (B_FALSE); 1090 } 1091 1092 uint64_t offset = DVA_GET_OFFSET(dva); 1093 uint64_t asize = DVA_GET_ASIZE(dva); 1094 1095 if (DVA_GET_GANG(dva)) 1096 asize = vdev_gang_header_asize(vd); 1097 if (offset + asize > vd->vdev_asize) 1098 return (B_FALSE); 1099 1100 return (B_TRUE); 1101 } 1102 1103 zio_t * 1104 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, 1105 abd_t *data, uint64_t size, zio_done_func_t *done, void *private, 1106 zio_priority_t priority, zio_flag_t flags, const zbookmark_phys_t *zb) 1107 { 1108 zio_t *zio; 1109 1110 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp, 1111 data, size, size, done, private, 1112 ZIO_TYPE_READ, priority, flags, NULL, 0, zb, 1113 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? 1114 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE); 1115 1116 return (zio); 1117 } 1118 1119 zio_t * 1120 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, 1121 abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp, 1122 zio_done_func_t *ready, zio_done_func_t *children_ready, 1123 zio_done_func_t *physdone, zio_done_func_t *done, 1124 void *private, zio_priority_t priority, zio_flag_t flags, 1125 const zbookmark_phys_t *zb) 1126 { 1127 zio_t *zio; 1128 1129 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF && 1130 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS && 1131 zp->zp_compress >= ZIO_COMPRESS_OFF && 1132 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS && 1133 DMU_OT_IS_VALID(zp->zp_type) && 1134 zp->zp_level < 32 && 1135 zp->zp_copies > 0 && 1136 zp->zp_copies <= spa_max_replication(spa)); 1137 1138 zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private, 1139 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, 1140 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? 1141 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE); 1142 1143 zio->io_ready = ready; 1144 zio->io_children_ready = children_ready; 1145 zio->io_physdone = physdone; 1146 zio->io_prop = *zp; 1147 1148 /* 1149 * Data can be NULL if we are going to call zio_write_override() to 1150 * provide the already-allocated BP. But we may need the data to 1151 * verify a dedup hit (if requested). In this case, don't try to 1152 * dedup (just take the already-allocated BP verbatim). Encrypted 1153 * dedup blocks need data as well so we also disable dedup in this 1154 * case. 1155 */ 1156 if (data == NULL && 1157 (zio->io_prop.zp_dedup_verify || zio->io_prop.zp_encrypt)) { 1158 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE; 1159 } 1160 1161 return (zio); 1162 } 1163 1164 zio_t * 1165 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data, 1166 uint64_t size, zio_done_func_t *done, void *private, 1167 zio_priority_t priority, zio_flag_t flags, zbookmark_phys_t *zb) 1168 { 1169 zio_t *zio; 1170 1171 zio = zio_create(pio, spa, txg, bp, data, size, size, done, private, 1172 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb, 1173 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE); 1174 1175 return (zio); 1176 } 1177 1178 void 1179 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite) 1180 { 1181 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 1182 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1183 ASSERT(zio->io_stage == ZIO_STAGE_OPEN); 1184 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa)); 1185 1186 /* 1187 * We must reset the io_prop to match the values that existed 1188 * when the bp was first written by dmu_sync() keeping in mind 1189 * that nopwrite and dedup are mutually exclusive. 1190 */ 1191 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup; 1192 zio->io_prop.zp_nopwrite = nopwrite; 1193 zio->io_prop.zp_copies = copies; 1194 zio->io_bp_override = bp; 1195 } 1196 1197 void 1198 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp) 1199 { 1200 1201 (void) zfs_blkptr_verify(spa, bp, B_FALSE, BLK_VERIFY_HALT); 1202 1203 /* 1204 * The check for EMBEDDED is a performance optimization. We 1205 * process the free here (by ignoring it) rather than 1206 * putting it on the list and then processing it in zio_free_sync(). 1207 */ 1208 if (BP_IS_EMBEDDED(bp)) 1209 return; 1210 1211 /* 1212 * Frees that are for the currently-syncing txg, are not going to be 1213 * deferred, and which will not need to do a read (i.e. not GANG or 1214 * DEDUP), can be processed immediately. Otherwise, put them on the 1215 * in-memory list for later processing. 1216 * 1217 * Note that we only defer frees after zfs_sync_pass_deferred_free 1218 * when the log space map feature is disabled. [see relevant comment 1219 * in spa_sync_iterate_to_convergence()] 1220 */ 1221 if (BP_IS_GANG(bp) || 1222 BP_GET_DEDUP(bp) || 1223 txg != spa->spa_syncing_txg || 1224 (spa_sync_pass(spa) >= zfs_sync_pass_deferred_free && 1225 !spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP))) { 1226 metaslab_check_free(spa, bp); 1227 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp); 1228 } else { 1229 VERIFY3P(zio_free_sync(NULL, spa, txg, bp, 0), ==, NULL); 1230 } 1231 } 1232 1233 /* 1234 * To improve performance, this function may return NULL if we were able 1235 * to do the free immediately. This avoids the cost of creating a zio 1236 * (and linking it to the parent, etc). 1237 */ 1238 zio_t * 1239 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 1240 zio_flag_t flags) 1241 { 1242 ASSERT(!BP_IS_HOLE(bp)); 1243 ASSERT(spa_syncing_txg(spa) == txg); 1244 1245 if (BP_IS_EMBEDDED(bp)) 1246 return (NULL); 1247 1248 metaslab_check_free(spa, bp); 1249 arc_freed(spa, bp); 1250 dsl_scan_freed(spa, bp); 1251 1252 if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp)) { 1253 /* 1254 * GANG and DEDUP blocks can induce a read (for the gang block 1255 * header, or the DDT), so issue them asynchronously so that 1256 * this thread is not tied up. 1257 */ 1258 enum zio_stage stage = 1259 ZIO_FREE_PIPELINE | ZIO_STAGE_ISSUE_ASYNC; 1260 1261 return (zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), 1262 BP_GET_PSIZE(bp), NULL, NULL, 1263 ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, 1264 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage)); 1265 } else { 1266 metaslab_free(spa, bp, txg, B_FALSE); 1267 return (NULL); 1268 } 1269 } 1270 1271 zio_t * 1272 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 1273 zio_done_func_t *done, void *private, zio_flag_t flags) 1274 { 1275 zio_t *zio; 1276 1277 (void) zfs_blkptr_verify(spa, bp, flags & ZIO_FLAG_CONFIG_WRITER, 1278 BLK_VERIFY_HALT); 1279 1280 if (BP_IS_EMBEDDED(bp)) 1281 return (zio_null(pio, spa, NULL, NULL, NULL, 0)); 1282 1283 /* 1284 * A claim is an allocation of a specific block. Claims are needed 1285 * to support immediate writes in the intent log. The issue is that 1286 * immediate writes contain committed data, but in a txg that was 1287 * *not* committed. Upon opening the pool after an unclean shutdown, 1288 * the intent log claims all blocks that contain immediate write data 1289 * so that the SPA knows they're in use. 1290 * 1291 * All claims *must* be resolved in the first txg -- before the SPA 1292 * starts allocating blocks -- so that nothing is allocated twice. 1293 * If txg == 0 we just verify that the block is claimable. 1294 */ 1295 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, 1296 spa_min_claim_txg(spa)); 1297 ASSERT(txg == spa_min_claim_txg(spa) || txg == 0); 1298 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(8) */ 1299 1300 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), 1301 BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, 1302 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE); 1303 ASSERT0(zio->io_queued_timestamp); 1304 1305 return (zio); 1306 } 1307 1308 zio_t * 1309 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, 1310 zio_done_func_t *done, void *private, zio_flag_t flags) 1311 { 1312 zio_t *zio; 1313 int c; 1314 1315 if (vd->vdev_children == 0) { 1316 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private, 1317 ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, 1318 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE); 1319 1320 zio->io_cmd = cmd; 1321 } else { 1322 zio = zio_null(pio, spa, NULL, NULL, NULL, flags); 1323 1324 for (c = 0; c < vd->vdev_children; c++) 1325 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd, 1326 done, private, flags)); 1327 } 1328 1329 return (zio); 1330 } 1331 1332 zio_t * 1333 zio_trim(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 1334 zio_done_func_t *done, void *private, zio_priority_t priority, 1335 zio_flag_t flags, enum trim_flag trim_flags) 1336 { 1337 zio_t *zio; 1338 1339 ASSERT0(vd->vdev_children); 1340 ASSERT0(P2PHASE(offset, 1ULL << vd->vdev_ashift)); 1341 ASSERT0(P2PHASE(size, 1ULL << vd->vdev_ashift)); 1342 ASSERT3U(size, !=, 0); 1343 1344 zio = zio_create(pio, vd->vdev_spa, 0, NULL, NULL, size, size, done, 1345 private, ZIO_TYPE_TRIM, priority, flags | ZIO_FLAG_PHYSICAL, 1346 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_TRIM_PIPELINE); 1347 zio->io_trim_flags = trim_flags; 1348 1349 return (zio); 1350 } 1351 1352 zio_t * 1353 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 1354 abd_t *data, int checksum, zio_done_func_t *done, void *private, 1355 zio_priority_t priority, zio_flag_t flags, boolean_t labels) 1356 { 1357 zio_t *zio; 1358 1359 ASSERT(vd->vdev_children == 0); 1360 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || 1361 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); 1362 ASSERT3U(offset + size, <=, vd->vdev_psize); 1363 1364 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done, 1365 private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, 1366 offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE); 1367 1368 zio->io_prop.zp_checksum = checksum; 1369 1370 return (zio); 1371 } 1372 1373 zio_t * 1374 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 1375 abd_t *data, int checksum, zio_done_func_t *done, void *private, 1376 zio_priority_t priority, zio_flag_t flags, boolean_t labels) 1377 { 1378 zio_t *zio; 1379 1380 ASSERT(vd->vdev_children == 0); 1381 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || 1382 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); 1383 ASSERT3U(offset + size, <=, vd->vdev_psize); 1384 1385 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done, 1386 private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, 1387 offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE); 1388 1389 zio->io_prop.zp_checksum = checksum; 1390 1391 if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) { 1392 /* 1393 * zec checksums are necessarily destructive -- they modify 1394 * the end of the write buffer to hold the verifier/checksum. 1395 * Therefore, we must make a local copy in case the data is 1396 * being written to multiple places in parallel. 1397 */ 1398 abd_t *wbuf = abd_alloc_sametype(data, size); 1399 abd_copy(wbuf, data, size); 1400 1401 zio_push_transform(zio, wbuf, size, size, NULL); 1402 } 1403 1404 return (zio); 1405 } 1406 1407 /* 1408 * Create a child I/O to do some work for us. 1409 */ 1410 zio_t * 1411 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset, 1412 abd_t *data, uint64_t size, int type, zio_priority_t priority, 1413 zio_flag_t flags, zio_done_func_t *done, void *private) 1414 { 1415 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE; 1416 zio_t *zio; 1417 1418 /* 1419 * vdev child I/Os do not propagate their error to the parent. 1420 * Therefore, for correct operation the caller *must* check for 1421 * and handle the error in the child i/o's done callback. 1422 * The only exceptions are i/os that we don't care about 1423 * (OPTIONAL or REPAIR). 1424 */ 1425 ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) || 1426 done != NULL); 1427 1428 if (type == ZIO_TYPE_READ && bp != NULL) { 1429 /* 1430 * If we have the bp, then the child should perform the 1431 * checksum and the parent need not. This pushes error 1432 * detection as close to the leaves as possible and 1433 * eliminates redundant checksums in the interior nodes. 1434 */ 1435 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY; 1436 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; 1437 } 1438 1439 if (vd->vdev_ops->vdev_op_leaf) { 1440 ASSERT0(vd->vdev_children); 1441 offset += VDEV_LABEL_START_SIZE; 1442 } 1443 1444 flags |= ZIO_VDEV_CHILD_FLAGS(pio); 1445 1446 /* 1447 * If we've decided to do a repair, the write is not speculative -- 1448 * even if the original read was. 1449 */ 1450 if (flags & ZIO_FLAG_IO_REPAIR) 1451 flags &= ~ZIO_FLAG_SPECULATIVE; 1452 1453 /* 1454 * If we're creating a child I/O that is not associated with a 1455 * top-level vdev, then the child zio is not an allocating I/O. 1456 * If this is a retried I/O then we ignore it since we will 1457 * have already processed the original allocating I/O. 1458 */ 1459 if (flags & ZIO_FLAG_IO_ALLOCATING && 1460 (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) { 1461 ASSERT(pio->io_metaslab_class != NULL); 1462 ASSERT(pio->io_metaslab_class->mc_alloc_throttle_enabled); 1463 ASSERT(type == ZIO_TYPE_WRITE); 1464 ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE); 1465 ASSERT(!(flags & ZIO_FLAG_IO_REPAIR)); 1466 ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) || 1467 pio->io_child_type == ZIO_CHILD_GANG); 1468 1469 flags &= ~ZIO_FLAG_IO_ALLOCATING; 1470 } 1471 1472 1473 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size, 1474 done, private, type, priority, flags, vd, offset, &pio->io_bookmark, 1475 ZIO_STAGE_VDEV_IO_START >> 1, pipeline); 1476 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV); 1477 1478 zio->io_physdone = pio->io_physdone; 1479 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL) 1480 zio->io_logical->io_phys_children++; 1481 1482 return (zio); 1483 } 1484 1485 zio_t * 1486 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size, 1487 zio_type_t type, zio_priority_t priority, zio_flag_t flags, 1488 zio_done_func_t *done, void *private) 1489 { 1490 zio_t *zio; 1491 1492 ASSERT(vd->vdev_ops->vdev_op_leaf); 1493 1494 zio = zio_create(NULL, vd->vdev_spa, 0, NULL, 1495 data, size, size, done, private, type, priority, 1496 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED, 1497 vd, offset, NULL, 1498 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE); 1499 1500 return (zio); 1501 } 1502 1503 void 1504 zio_flush(zio_t *zio, vdev_t *vd) 1505 { 1506 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 1507 NULL, NULL, 1508 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY)); 1509 } 1510 1511 void 1512 zio_shrink(zio_t *zio, uint64_t size) 1513 { 1514 ASSERT3P(zio->io_executor, ==, NULL); 1515 ASSERT3U(zio->io_orig_size, ==, zio->io_size); 1516 ASSERT3U(size, <=, zio->io_size); 1517 1518 /* 1519 * We don't shrink for raidz because of problems with the 1520 * reconstruction when reading back less than the block size. 1521 * Note, BP_IS_RAIDZ() assumes no compression. 1522 */ 1523 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 1524 if (!BP_IS_RAIDZ(zio->io_bp)) { 1525 /* we are not doing a raw write */ 1526 ASSERT3U(zio->io_size, ==, zio->io_lsize); 1527 zio->io_orig_size = zio->io_size = zio->io_lsize = size; 1528 } 1529 } 1530 1531 /* 1532 * ========================================================================== 1533 * Prepare to read and write logical blocks 1534 * ========================================================================== 1535 */ 1536 1537 static zio_t * 1538 zio_read_bp_init(zio_t *zio) 1539 { 1540 blkptr_t *bp = zio->io_bp; 1541 uint64_t psize = 1542 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp); 1543 1544 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy); 1545 1546 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF && 1547 zio->io_child_type == ZIO_CHILD_LOGICAL && 1548 !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) { 1549 zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize), 1550 psize, psize, zio_decompress); 1551 } 1552 1553 if (((BP_IS_PROTECTED(bp) && !(zio->io_flags & ZIO_FLAG_RAW_ENCRYPT)) || 1554 BP_HAS_INDIRECT_MAC_CKSUM(bp)) && 1555 zio->io_child_type == ZIO_CHILD_LOGICAL) { 1556 zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize), 1557 psize, psize, zio_decrypt); 1558 } 1559 1560 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) { 1561 int psize = BPE_GET_PSIZE(bp); 1562 void *data = abd_borrow_buf(zio->io_abd, psize); 1563 1564 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1565 decode_embedded_bp_compressed(bp, data); 1566 abd_return_buf_copy(zio->io_abd, data, psize); 1567 } else { 1568 ASSERT(!BP_IS_EMBEDDED(bp)); 1569 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy); 1570 } 1571 1572 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0) 1573 zio->io_flags |= ZIO_FLAG_DONT_CACHE; 1574 1575 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP) 1576 zio->io_flags |= ZIO_FLAG_DONT_CACHE; 1577 1578 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL) 1579 zio->io_pipeline = ZIO_DDT_READ_PIPELINE; 1580 1581 return (zio); 1582 } 1583 1584 static zio_t * 1585 zio_write_bp_init(zio_t *zio) 1586 { 1587 if (!IO_IS_ALLOCATING(zio)) 1588 return (zio); 1589 1590 ASSERT(zio->io_child_type != ZIO_CHILD_DDT); 1591 1592 if (zio->io_bp_override) { 1593 blkptr_t *bp = zio->io_bp; 1594 zio_prop_t *zp = &zio->io_prop; 1595 1596 ASSERT(bp->blk_birth != zio->io_txg); 1597 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0); 1598 1599 *bp = *zio->io_bp_override; 1600 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1601 1602 if (BP_IS_EMBEDDED(bp)) 1603 return (zio); 1604 1605 /* 1606 * If we've been overridden and nopwrite is set then 1607 * set the flag accordingly to indicate that a nopwrite 1608 * has already occurred. 1609 */ 1610 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) { 1611 ASSERT(!zp->zp_dedup); 1612 ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum); 1613 zio->io_flags |= ZIO_FLAG_NOPWRITE; 1614 return (zio); 1615 } 1616 1617 ASSERT(!zp->zp_nopwrite); 1618 1619 if (BP_IS_HOLE(bp) || !zp->zp_dedup) 1620 return (zio); 1621 1622 ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags & 1623 ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify); 1624 1625 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum && 1626 !zp->zp_encrypt) { 1627 BP_SET_DEDUP(bp, 1); 1628 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE; 1629 return (zio); 1630 } 1631 1632 /* 1633 * We were unable to handle this as an override bp, treat 1634 * it as a regular write I/O. 1635 */ 1636 zio->io_bp_override = NULL; 1637 *bp = zio->io_bp_orig; 1638 zio->io_pipeline = zio->io_orig_pipeline; 1639 } 1640 1641 return (zio); 1642 } 1643 1644 static zio_t * 1645 zio_write_compress(zio_t *zio) 1646 { 1647 spa_t *spa = zio->io_spa; 1648 zio_prop_t *zp = &zio->io_prop; 1649 enum zio_compress compress = zp->zp_compress; 1650 blkptr_t *bp = zio->io_bp; 1651 uint64_t lsize = zio->io_lsize; 1652 uint64_t psize = zio->io_size; 1653 uint32_t pass = 1; 1654 1655 /* 1656 * If our children haven't all reached the ready stage, 1657 * wait for them and then repeat this pipeline stage. 1658 */ 1659 if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT | 1660 ZIO_CHILD_GANG_BIT, ZIO_WAIT_READY)) { 1661 return (NULL); 1662 } 1663 1664 if (!IO_IS_ALLOCATING(zio)) 1665 return (zio); 1666 1667 if (zio->io_children_ready != NULL) { 1668 /* 1669 * Now that all our children are ready, run the callback 1670 * associated with this zio in case it wants to modify the 1671 * data to be written. 1672 */ 1673 ASSERT3U(zp->zp_level, >, 0); 1674 zio->io_children_ready(zio); 1675 } 1676 1677 ASSERT(zio->io_child_type != ZIO_CHILD_DDT); 1678 ASSERT(zio->io_bp_override == NULL); 1679 1680 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) { 1681 /* 1682 * We're rewriting an existing block, which means we're 1683 * working on behalf of spa_sync(). For spa_sync() to 1684 * converge, it must eventually be the case that we don't 1685 * have to allocate new blocks. But compression changes 1686 * the blocksize, which forces a reallocate, and makes 1687 * convergence take longer. Therefore, after the first 1688 * few passes, stop compressing to ensure convergence. 1689 */ 1690 pass = spa_sync_pass(spa); 1691 1692 ASSERT(zio->io_txg == spa_syncing_txg(spa)); 1693 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1694 ASSERT(!BP_GET_DEDUP(bp)); 1695 1696 if (pass >= zfs_sync_pass_dont_compress) 1697 compress = ZIO_COMPRESS_OFF; 1698 1699 /* Make sure someone doesn't change their mind on overwrites */ 1700 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp), 1701 spa_max_replication(spa)) == BP_GET_NDVAS(bp)); 1702 } 1703 1704 /* If it's a compressed write that is not raw, compress the buffer. */ 1705 if (compress != ZIO_COMPRESS_OFF && 1706 !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) { 1707 void *cbuf = zio_buf_alloc(lsize); 1708 psize = zio_compress_data(compress, zio->io_abd, cbuf, lsize, 1709 zp->zp_complevel); 1710 if (psize == 0 || psize >= lsize) { 1711 compress = ZIO_COMPRESS_OFF; 1712 zio_buf_free(cbuf, lsize); 1713 } else if (!zp->zp_dedup && !zp->zp_encrypt && 1714 psize <= BPE_PAYLOAD_SIZE && 1715 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) && 1716 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) { 1717 encode_embedded_bp_compressed(bp, 1718 cbuf, compress, lsize, psize); 1719 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA); 1720 BP_SET_TYPE(bp, zio->io_prop.zp_type); 1721 BP_SET_LEVEL(bp, zio->io_prop.zp_level); 1722 zio_buf_free(cbuf, lsize); 1723 bp->blk_birth = zio->io_txg; 1724 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1725 ASSERT(spa_feature_is_active(spa, 1726 SPA_FEATURE_EMBEDDED_DATA)); 1727 return (zio); 1728 } else { 1729 /* 1730 * Round compressed size up to the minimum allocation 1731 * size of the smallest-ashift device, and zero the 1732 * tail. This ensures that the compressed size of the 1733 * BP (and thus compressratio property) are correct, 1734 * in that we charge for the padding used to fill out 1735 * the last sector. 1736 */ 1737 ASSERT3U(spa->spa_min_alloc, >=, SPA_MINBLOCKSHIFT); 1738 size_t rounded = (size_t)roundup(psize, 1739 spa->spa_min_alloc); 1740 if (rounded >= lsize) { 1741 compress = ZIO_COMPRESS_OFF; 1742 zio_buf_free(cbuf, lsize); 1743 psize = lsize; 1744 } else { 1745 abd_t *cdata = abd_get_from_buf(cbuf, lsize); 1746 abd_take_ownership_of_buf(cdata, B_TRUE); 1747 abd_zero_off(cdata, psize, rounded - psize); 1748 psize = rounded; 1749 zio_push_transform(zio, cdata, 1750 psize, lsize, NULL); 1751 } 1752 } 1753 1754 /* 1755 * We were unable to handle this as an override bp, treat 1756 * it as a regular write I/O. 1757 */ 1758 zio->io_bp_override = NULL; 1759 *bp = zio->io_bp_orig; 1760 zio->io_pipeline = zio->io_orig_pipeline; 1761 1762 } else if ((zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) != 0 && 1763 zp->zp_type == DMU_OT_DNODE) { 1764 /* 1765 * The DMU actually relies on the zio layer's compression 1766 * to free metadnode blocks that have had all contained 1767 * dnodes freed. As a result, even when doing a raw 1768 * receive, we must check whether the block can be compressed 1769 * to a hole. 1770 */ 1771 psize = zio_compress_data(ZIO_COMPRESS_EMPTY, 1772 zio->io_abd, NULL, lsize, zp->zp_complevel); 1773 if (psize == 0 || psize >= lsize) 1774 compress = ZIO_COMPRESS_OFF; 1775 } else if (zio->io_flags & ZIO_FLAG_RAW_COMPRESS && 1776 !(zio->io_flags & ZIO_FLAG_RAW_ENCRYPT)) { 1777 /* 1778 * If we are raw receiving an encrypted dataset we should not 1779 * take this codepath because it will change the on-disk block 1780 * and decryption will fail. 1781 */ 1782 size_t rounded = MIN((size_t)roundup(psize, 1783 spa->spa_min_alloc), lsize); 1784 1785 if (rounded != psize) { 1786 abd_t *cdata = abd_alloc_linear(rounded, B_TRUE); 1787 abd_zero_off(cdata, psize, rounded - psize); 1788 abd_copy_off(cdata, zio->io_abd, 0, 0, psize); 1789 psize = rounded; 1790 zio_push_transform(zio, cdata, 1791 psize, rounded, NULL); 1792 } 1793 } else { 1794 ASSERT3U(psize, !=, 0); 1795 } 1796 1797 /* 1798 * The final pass of spa_sync() must be all rewrites, but the first 1799 * few passes offer a trade-off: allocating blocks defers convergence, 1800 * but newly allocated blocks are sequential, so they can be written 1801 * to disk faster. Therefore, we allow the first few passes of 1802 * spa_sync() to allocate new blocks, but force rewrites after that. 1803 * There should only be a handful of blocks after pass 1 in any case. 1804 */ 1805 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg && 1806 BP_GET_PSIZE(bp) == psize && 1807 pass >= zfs_sync_pass_rewrite) { 1808 VERIFY3U(psize, !=, 0); 1809 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES; 1810 1811 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages; 1812 zio->io_flags |= ZIO_FLAG_IO_REWRITE; 1813 } else { 1814 BP_ZERO(bp); 1815 zio->io_pipeline = ZIO_WRITE_PIPELINE; 1816 } 1817 1818 if (psize == 0) { 1819 if (zio->io_bp_orig.blk_birth != 0 && 1820 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) { 1821 BP_SET_LSIZE(bp, lsize); 1822 BP_SET_TYPE(bp, zp->zp_type); 1823 BP_SET_LEVEL(bp, zp->zp_level); 1824 BP_SET_BIRTH(bp, zio->io_txg, 0); 1825 } 1826 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1827 } else { 1828 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER); 1829 BP_SET_LSIZE(bp, lsize); 1830 BP_SET_TYPE(bp, zp->zp_type); 1831 BP_SET_LEVEL(bp, zp->zp_level); 1832 BP_SET_PSIZE(bp, psize); 1833 BP_SET_COMPRESS(bp, compress); 1834 BP_SET_CHECKSUM(bp, zp->zp_checksum); 1835 BP_SET_DEDUP(bp, zp->zp_dedup); 1836 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER); 1837 if (zp->zp_dedup) { 1838 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1839 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 1840 ASSERT(!zp->zp_encrypt || 1841 DMU_OT_IS_ENCRYPTED(zp->zp_type)); 1842 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE; 1843 } 1844 if (zp->zp_nopwrite) { 1845 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1846 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 1847 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE; 1848 } 1849 } 1850 return (zio); 1851 } 1852 1853 static zio_t * 1854 zio_free_bp_init(zio_t *zio) 1855 { 1856 blkptr_t *bp = zio->io_bp; 1857 1858 if (zio->io_child_type == ZIO_CHILD_LOGICAL) { 1859 if (BP_GET_DEDUP(bp)) 1860 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE; 1861 } 1862 1863 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy); 1864 1865 return (zio); 1866 } 1867 1868 /* 1869 * ========================================================================== 1870 * Execute the I/O pipeline 1871 * ========================================================================== 1872 */ 1873 1874 static void 1875 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline) 1876 { 1877 spa_t *spa = zio->io_spa; 1878 zio_type_t t = zio->io_type; 1879 int flags = (cutinline ? TQ_FRONT : 0); 1880 1881 /* 1882 * If we're a config writer or a probe, the normal issue and 1883 * interrupt threads may all be blocked waiting for the config lock. 1884 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL. 1885 */ 1886 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE)) 1887 t = ZIO_TYPE_NULL; 1888 1889 /* 1890 * A similar issue exists for the L2ARC write thread until L2ARC 2.0. 1891 */ 1892 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux) 1893 t = ZIO_TYPE_NULL; 1894 1895 /* 1896 * If this is a high priority I/O, then use the high priority taskq if 1897 * available. 1898 */ 1899 if ((zio->io_priority == ZIO_PRIORITY_NOW || 1900 zio->io_priority == ZIO_PRIORITY_SYNC_WRITE) && 1901 spa->spa_zio_taskq[t][q + 1].stqs_count != 0) 1902 q++; 1903 1904 ASSERT3U(q, <, ZIO_TASKQ_TYPES); 1905 1906 /* 1907 * NB: We are assuming that the zio can only be dispatched 1908 * to a single taskq at a time. It would be a grievous error 1909 * to dispatch the zio to another taskq at the same time. 1910 */ 1911 ASSERT(taskq_empty_ent(&zio->io_tqent)); 1912 spa_taskq_dispatch_ent(spa, t, q, zio_execute, zio, flags, 1913 &zio->io_tqent); 1914 } 1915 1916 static boolean_t 1917 zio_taskq_member(zio_t *zio, zio_taskq_type_t q) 1918 { 1919 spa_t *spa = zio->io_spa; 1920 1921 taskq_t *tq = taskq_of_curthread(); 1922 1923 for (zio_type_t t = 0; t < ZIO_TYPES; t++) { 1924 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q]; 1925 uint_t i; 1926 for (i = 0; i < tqs->stqs_count; i++) { 1927 if (tqs->stqs_taskq[i] == tq) 1928 return (B_TRUE); 1929 } 1930 } 1931 1932 return (B_FALSE); 1933 } 1934 1935 static zio_t * 1936 zio_issue_async(zio_t *zio) 1937 { 1938 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); 1939 1940 return (NULL); 1941 } 1942 1943 void 1944 zio_interrupt(void *zio) 1945 { 1946 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE); 1947 } 1948 1949 void 1950 zio_delay_interrupt(zio_t *zio) 1951 { 1952 /* 1953 * The timeout_generic() function isn't defined in userspace, so 1954 * rather than trying to implement the function, the zio delay 1955 * functionality has been disabled for userspace builds. 1956 */ 1957 1958 #ifdef _KERNEL 1959 /* 1960 * If io_target_timestamp is zero, then no delay has been registered 1961 * for this IO, thus jump to the end of this function and "skip" the 1962 * delay; issuing it directly to the zio layer. 1963 */ 1964 if (zio->io_target_timestamp != 0) { 1965 hrtime_t now = gethrtime(); 1966 1967 if (now >= zio->io_target_timestamp) { 1968 /* 1969 * This IO has already taken longer than the target 1970 * delay to complete, so we don't want to delay it 1971 * any longer; we "miss" the delay and issue it 1972 * directly to the zio layer. This is likely due to 1973 * the target latency being set to a value less than 1974 * the underlying hardware can satisfy (e.g. delay 1975 * set to 1ms, but the disks take 10ms to complete an 1976 * IO request). 1977 */ 1978 1979 DTRACE_PROBE2(zio__delay__miss, zio_t *, zio, 1980 hrtime_t, now); 1981 1982 zio_interrupt(zio); 1983 } else { 1984 taskqid_t tid; 1985 hrtime_t diff = zio->io_target_timestamp - now; 1986 clock_t expire_at_tick = ddi_get_lbolt() + 1987 NSEC_TO_TICK(diff); 1988 1989 DTRACE_PROBE3(zio__delay__hit, zio_t *, zio, 1990 hrtime_t, now, hrtime_t, diff); 1991 1992 if (NSEC_TO_TICK(diff) == 0) { 1993 /* Our delay is less than a jiffy - just spin */ 1994 zfs_sleep_until(zio->io_target_timestamp); 1995 zio_interrupt(zio); 1996 } else { 1997 /* 1998 * Use taskq_dispatch_delay() in the place of 1999 * OpenZFS's timeout_generic(). 2000 */ 2001 tid = taskq_dispatch_delay(system_taskq, 2002 zio_interrupt, zio, TQ_NOSLEEP, 2003 expire_at_tick); 2004 if (tid == TASKQID_INVALID) { 2005 /* 2006 * Couldn't allocate a task. Just 2007 * finish the zio without a delay. 2008 */ 2009 zio_interrupt(zio); 2010 } 2011 } 2012 } 2013 return; 2014 } 2015 #endif 2016 DTRACE_PROBE1(zio__delay__skip, zio_t *, zio); 2017 zio_interrupt(zio); 2018 } 2019 2020 static void 2021 zio_deadman_impl(zio_t *pio, int ziodepth) 2022 { 2023 zio_t *cio, *cio_next; 2024 zio_link_t *zl = NULL; 2025 vdev_t *vd = pio->io_vd; 2026 2027 if (zio_deadman_log_all || (vd != NULL && vd->vdev_ops->vdev_op_leaf)) { 2028 vdev_queue_t *vq = vd ? &vd->vdev_queue : NULL; 2029 zbookmark_phys_t *zb = &pio->io_bookmark; 2030 uint64_t delta = gethrtime() - pio->io_timestamp; 2031 uint64_t failmode = spa_get_deadman_failmode(pio->io_spa); 2032 2033 zfs_dbgmsg("slow zio[%d]: zio=%px timestamp=%llu " 2034 "delta=%llu queued=%llu io=%llu " 2035 "path=%s " 2036 "last=%llu type=%d " 2037 "priority=%d flags=0x%llx stage=0x%x " 2038 "pipeline=0x%x pipeline-trace=0x%x " 2039 "objset=%llu object=%llu " 2040 "level=%llu blkid=%llu " 2041 "offset=%llu size=%llu " 2042 "error=%d", 2043 ziodepth, pio, pio->io_timestamp, 2044 (u_longlong_t)delta, pio->io_delta, pio->io_delay, 2045 vd ? vd->vdev_path : "NULL", 2046 vq ? vq->vq_io_complete_ts : 0, pio->io_type, 2047 pio->io_priority, (u_longlong_t)pio->io_flags, 2048 pio->io_stage, pio->io_pipeline, pio->io_pipeline_trace, 2049 (u_longlong_t)zb->zb_objset, (u_longlong_t)zb->zb_object, 2050 (u_longlong_t)zb->zb_level, (u_longlong_t)zb->zb_blkid, 2051 (u_longlong_t)pio->io_offset, (u_longlong_t)pio->io_size, 2052 pio->io_error); 2053 (void) zfs_ereport_post(FM_EREPORT_ZFS_DEADMAN, 2054 pio->io_spa, vd, zb, pio, 0); 2055 2056 if (failmode == ZIO_FAILURE_MODE_CONTINUE && 2057 taskq_empty_ent(&pio->io_tqent)) { 2058 zio_interrupt(pio); 2059 } 2060 } 2061 2062 mutex_enter(&pio->io_lock); 2063 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) { 2064 cio_next = zio_walk_children(pio, &zl); 2065 zio_deadman_impl(cio, ziodepth + 1); 2066 } 2067 mutex_exit(&pio->io_lock); 2068 } 2069 2070 /* 2071 * Log the critical information describing this zio and all of its children 2072 * using the zfs_dbgmsg() interface then post deadman event for the ZED. 2073 */ 2074 void 2075 zio_deadman(zio_t *pio, const char *tag) 2076 { 2077 spa_t *spa = pio->io_spa; 2078 char *name = spa_name(spa); 2079 2080 if (!zfs_deadman_enabled || spa_suspended(spa)) 2081 return; 2082 2083 zio_deadman_impl(pio, 0); 2084 2085 switch (spa_get_deadman_failmode(spa)) { 2086 case ZIO_FAILURE_MODE_WAIT: 2087 zfs_dbgmsg("%s waiting for hung I/O to pool '%s'", tag, name); 2088 break; 2089 2090 case ZIO_FAILURE_MODE_CONTINUE: 2091 zfs_dbgmsg("%s restarting hung I/O for pool '%s'", tag, name); 2092 break; 2093 2094 case ZIO_FAILURE_MODE_PANIC: 2095 fm_panic("%s determined I/O to pool '%s' is hung.", tag, name); 2096 break; 2097 } 2098 } 2099 2100 /* 2101 * Execute the I/O pipeline until one of the following occurs: 2102 * (1) the I/O completes; (2) the pipeline stalls waiting for 2103 * dependent child I/Os; (3) the I/O issues, so we're waiting 2104 * for an I/O completion interrupt; (4) the I/O is delegated by 2105 * vdev-level caching or aggregation; (5) the I/O is deferred 2106 * due to vdev-level queueing; (6) the I/O is handed off to 2107 * another thread. In all cases, the pipeline stops whenever 2108 * there's no CPU work; it never burns a thread in cv_wait_io(). 2109 * 2110 * There's no locking on io_stage because there's no legitimate way 2111 * for multiple threads to be attempting to process the same I/O. 2112 */ 2113 static zio_pipe_stage_t *zio_pipeline[]; 2114 2115 /* 2116 * zio_execute() is a wrapper around the static function 2117 * __zio_execute() so that we can force __zio_execute() to be 2118 * inlined. This reduces stack overhead which is important 2119 * because __zio_execute() is called recursively in several zio 2120 * code paths. zio_execute() itself cannot be inlined because 2121 * it is externally visible. 2122 */ 2123 void 2124 zio_execute(void *zio) 2125 { 2126 fstrans_cookie_t cookie; 2127 2128 cookie = spl_fstrans_mark(); 2129 __zio_execute(zio); 2130 spl_fstrans_unmark(cookie); 2131 } 2132 2133 /* 2134 * Used to determine if in the current context the stack is sized large 2135 * enough to allow zio_execute() to be called recursively. A minimum 2136 * stack size of 16K is required to avoid needing to re-dispatch the zio. 2137 */ 2138 static boolean_t 2139 zio_execute_stack_check(zio_t *zio) 2140 { 2141 #if !defined(HAVE_LARGE_STACKS) 2142 dsl_pool_t *dp = spa_get_dsl(zio->io_spa); 2143 2144 /* Executing in txg_sync_thread() context. */ 2145 if (dp && curthread == dp->dp_tx.tx_sync_thread) 2146 return (B_TRUE); 2147 2148 /* Pool initialization outside of zio_taskq context. */ 2149 if (dp && spa_is_initializing(dp->dp_spa) && 2150 !zio_taskq_member(zio, ZIO_TASKQ_ISSUE) && 2151 !zio_taskq_member(zio, ZIO_TASKQ_ISSUE_HIGH)) 2152 return (B_TRUE); 2153 #else 2154 (void) zio; 2155 #endif /* HAVE_LARGE_STACKS */ 2156 2157 return (B_FALSE); 2158 } 2159 2160 __attribute__((always_inline)) 2161 static inline void 2162 __zio_execute(zio_t *zio) 2163 { 2164 ASSERT3U(zio->io_queued_timestamp, >, 0); 2165 2166 while (zio->io_stage < ZIO_STAGE_DONE) { 2167 enum zio_stage pipeline = zio->io_pipeline; 2168 enum zio_stage stage = zio->io_stage; 2169 2170 zio->io_executor = curthread; 2171 2172 ASSERT(!MUTEX_HELD(&zio->io_lock)); 2173 ASSERT(ISP2(stage)); 2174 ASSERT(zio->io_stall == NULL); 2175 2176 do { 2177 stage <<= 1; 2178 } while ((stage & pipeline) == 0); 2179 2180 ASSERT(stage <= ZIO_STAGE_DONE); 2181 2182 /* 2183 * If we are in interrupt context and this pipeline stage 2184 * will grab a config lock that is held across I/O, 2185 * or may wait for an I/O that needs an interrupt thread 2186 * to complete, issue async to avoid deadlock. 2187 * 2188 * For VDEV_IO_START, we cut in line so that the io will 2189 * be sent to disk promptly. 2190 */ 2191 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL && 2192 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) { 2193 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? 2194 zio_requeue_io_start_cut_in_line : B_FALSE; 2195 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); 2196 return; 2197 } 2198 2199 /* 2200 * If the current context doesn't have large enough stacks 2201 * the zio must be issued asynchronously to prevent overflow. 2202 */ 2203 if (zio_execute_stack_check(zio)) { 2204 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? 2205 zio_requeue_io_start_cut_in_line : B_FALSE; 2206 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); 2207 return; 2208 } 2209 2210 zio->io_stage = stage; 2211 zio->io_pipeline_trace |= zio->io_stage; 2212 2213 /* 2214 * The zio pipeline stage returns the next zio to execute 2215 * (typically the same as this one), or NULL if we should 2216 * stop. 2217 */ 2218 zio = zio_pipeline[highbit64(stage) - 1](zio); 2219 2220 if (zio == NULL) 2221 return; 2222 } 2223 } 2224 2225 2226 /* 2227 * ========================================================================== 2228 * Initiate I/O, either sync or async 2229 * ========================================================================== 2230 */ 2231 int 2232 zio_wait(zio_t *zio) 2233 { 2234 /* 2235 * Some routines, like zio_free_sync(), may return a NULL zio 2236 * to avoid the performance overhead of creating and then destroying 2237 * an unneeded zio. For the callers' simplicity, we accept a NULL 2238 * zio and ignore it. 2239 */ 2240 if (zio == NULL) 2241 return (0); 2242 2243 long timeout = MSEC_TO_TICK(zfs_deadman_ziotime_ms); 2244 int error; 2245 2246 ASSERT3S(zio->io_stage, ==, ZIO_STAGE_OPEN); 2247 ASSERT3P(zio->io_executor, ==, NULL); 2248 2249 zio->io_waiter = curthread; 2250 ASSERT0(zio->io_queued_timestamp); 2251 zio->io_queued_timestamp = gethrtime(); 2252 2253 __zio_execute(zio); 2254 2255 mutex_enter(&zio->io_lock); 2256 while (zio->io_executor != NULL) { 2257 error = cv_timedwait_io(&zio->io_cv, &zio->io_lock, 2258 ddi_get_lbolt() + timeout); 2259 2260 if (zfs_deadman_enabled && error == -1 && 2261 gethrtime() - zio->io_queued_timestamp > 2262 spa_deadman_ziotime(zio->io_spa)) { 2263 mutex_exit(&zio->io_lock); 2264 timeout = MSEC_TO_TICK(zfs_deadman_checktime_ms); 2265 zio_deadman(zio, FTAG); 2266 mutex_enter(&zio->io_lock); 2267 } 2268 } 2269 mutex_exit(&zio->io_lock); 2270 2271 error = zio->io_error; 2272 zio_destroy(zio); 2273 2274 return (error); 2275 } 2276 2277 void 2278 zio_nowait(zio_t *zio) 2279 { 2280 /* 2281 * See comment in zio_wait(). 2282 */ 2283 if (zio == NULL) 2284 return; 2285 2286 ASSERT3P(zio->io_executor, ==, NULL); 2287 2288 if (zio->io_child_type == ZIO_CHILD_LOGICAL && 2289 zio_unique_parent(zio) == NULL) { 2290 zio_t *pio; 2291 2292 /* 2293 * This is a logical async I/O with no parent to wait for it. 2294 * We add it to the spa_async_root_zio "Godfather" I/O which 2295 * will ensure they complete prior to unloading the pool. 2296 */ 2297 spa_t *spa = zio->io_spa; 2298 pio = spa->spa_async_zio_root[CPU_SEQID_UNSTABLE]; 2299 2300 zio_add_child(pio, zio); 2301 } 2302 2303 ASSERT0(zio->io_queued_timestamp); 2304 zio->io_queued_timestamp = gethrtime(); 2305 __zio_execute(zio); 2306 } 2307 2308 /* 2309 * ========================================================================== 2310 * Reexecute, cancel, or suspend/resume failed I/O 2311 * ========================================================================== 2312 */ 2313 2314 static void 2315 zio_reexecute(void *arg) 2316 { 2317 zio_t *pio = arg; 2318 zio_t *cio, *cio_next; 2319 2320 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL); 2321 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN); 2322 ASSERT(pio->io_gang_leader == NULL); 2323 ASSERT(pio->io_gang_tree == NULL); 2324 2325 pio->io_flags = pio->io_orig_flags; 2326 pio->io_stage = pio->io_orig_stage; 2327 pio->io_pipeline = pio->io_orig_pipeline; 2328 pio->io_reexecute = 0; 2329 pio->io_flags |= ZIO_FLAG_REEXECUTED; 2330 pio->io_pipeline_trace = 0; 2331 pio->io_error = 0; 2332 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 2333 pio->io_state[w] = 0; 2334 for (int c = 0; c < ZIO_CHILD_TYPES; c++) 2335 pio->io_child_error[c] = 0; 2336 2337 if (IO_IS_ALLOCATING(pio)) 2338 BP_ZERO(pio->io_bp); 2339 2340 /* 2341 * As we reexecute pio's children, new children could be created. 2342 * New children go to the head of pio's io_child_list, however, 2343 * so we will (correctly) not reexecute them. The key is that 2344 * the remainder of pio's io_child_list, from 'cio_next' onward, 2345 * cannot be affected by any side effects of reexecuting 'cio'. 2346 */ 2347 zio_link_t *zl = NULL; 2348 mutex_enter(&pio->io_lock); 2349 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) { 2350 cio_next = zio_walk_children(pio, &zl); 2351 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 2352 pio->io_children[cio->io_child_type][w]++; 2353 mutex_exit(&pio->io_lock); 2354 zio_reexecute(cio); 2355 mutex_enter(&pio->io_lock); 2356 } 2357 mutex_exit(&pio->io_lock); 2358 2359 /* 2360 * Now that all children have been reexecuted, execute the parent. 2361 * We don't reexecute "The Godfather" I/O here as it's the 2362 * responsibility of the caller to wait on it. 2363 */ 2364 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) { 2365 pio->io_queued_timestamp = gethrtime(); 2366 __zio_execute(pio); 2367 } 2368 } 2369 2370 void 2371 zio_suspend(spa_t *spa, zio_t *zio, zio_suspend_reason_t reason) 2372 { 2373 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC) 2374 fm_panic("Pool '%s' has encountered an uncorrectable I/O " 2375 "failure and the failure mode property for this pool " 2376 "is set to panic.", spa_name(spa)); 2377 2378 cmn_err(CE_WARN, "Pool '%s' has encountered an uncorrectable I/O " 2379 "failure and has been suspended.\n", spa_name(spa)); 2380 2381 (void) zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, 2382 NULL, NULL, 0); 2383 2384 mutex_enter(&spa->spa_suspend_lock); 2385 2386 if (spa->spa_suspend_zio_root == NULL) 2387 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL, 2388 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | 2389 ZIO_FLAG_GODFATHER); 2390 2391 spa->spa_suspended = reason; 2392 2393 if (zio != NULL) { 2394 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); 2395 ASSERT(zio != spa->spa_suspend_zio_root); 2396 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2397 ASSERT(zio_unique_parent(zio) == NULL); 2398 ASSERT(zio->io_stage == ZIO_STAGE_DONE); 2399 zio_add_child(spa->spa_suspend_zio_root, zio); 2400 } 2401 2402 mutex_exit(&spa->spa_suspend_lock); 2403 } 2404 2405 int 2406 zio_resume(spa_t *spa) 2407 { 2408 zio_t *pio; 2409 2410 /* 2411 * Reexecute all previously suspended i/o. 2412 */ 2413 mutex_enter(&spa->spa_suspend_lock); 2414 spa->spa_suspended = ZIO_SUSPEND_NONE; 2415 cv_broadcast(&spa->spa_suspend_cv); 2416 pio = spa->spa_suspend_zio_root; 2417 spa->spa_suspend_zio_root = NULL; 2418 mutex_exit(&spa->spa_suspend_lock); 2419 2420 if (pio == NULL) 2421 return (0); 2422 2423 zio_reexecute(pio); 2424 return (zio_wait(pio)); 2425 } 2426 2427 void 2428 zio_resume_wait(spa_t *spa) 2429 { 2430 mutex_enter(&spa->spa_suspend_lock); 2431 while (spa_suspended(spa)) 2432 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock); 2433 mutex_exit(&spa->spa_suspend_lock); 2434 } 2435 2436 /* 2437 * ========================================================================== 2438 * Gang blocks. 2439 * 2440 * A gang block is a collection of small blocks that looks to the DMU 2441 * like one large block. When zio_dva_allocate() cannot find a block 2442 * of the requested size, due to either severe fragmentation or the pool 2443 * being nearly full, it calls zio_write_gang_block() to construct the 2444 * block from smaller fragments. 2445 * 2446 * A gang block consists of a gang header (zio_gbh_phys_t) and up to 2447 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like 2448 * an indirect block: it's an array of block pointers. It consumes 2449 * only one sector and hence is allocatable regardless of fragmentation. 2450 * The gang header's bps point to its gang members, which hold the data. 2451 * 2452 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg> 2453 * as the verifier to ensure uniqueness of the SHA256 checksum. 2454 * Critically, the gang block bp's blk_cksum is the checksum of the data, 2455 * not the gang header. This ensures that data block signatures (needed for 2456 * deduplication) are independent of how the block is physically stored. 2457 * 2458 * Gang blocks can be nested: a gang member may itself be a gang block. 2459 * Thus every gang block is a tree in which root and all interior nodes are 2460 * gang headers, and the leaves are normal blocks that contain user data. 2461 * The root of the gang tree is called the gang leader. 2462 * 2463 * To perform any operation (read, rewrite, free, claim) on a gang block, 2464 * zio_gang_assemble() first assembles the gang tree (minus data leaves) 2465 * in the io_gang_tree field of the original logical i/o by recursively 2466 * reading the gang leader and all gang headers below it. This yields 2467 * an in-core tree containing the contents of every gang header and the 2468 * bps for every constituent of the gang block. 2469 * 2470 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree 2471 * and invokes a callback on each bp. To free a gang block, zio_gang_issue() 2472 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp. 2473 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim(). 2474 * zio_read_gang() is a wrapper around zio_read() that omits reading gang 2475 * headers, since we already have those in io_gang_tree. zio_rewrite_gang() 2476 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite() 2477 * of the gang header plus zio_checksum_compute() of the data to update the 2478 * gang header's blk_cksum as described above. 2479 * 2480 * The two-phase assemble/issue model solves the problem of partial failure -- 2481 * what if you'd freed part of a gang block but then couldn't read the 2482 * gang header for another part? Assembling the entire gang tree first 2483 * ensures that all the necessary gang header I/O has succeeded before 2484 * starting the actual work of free, claim, or write. Once the gang tree 2485 * is assembled, free and claim are in-memory operations that cannot fail. 2486 * 2487 * In the event that a gang write fails, zio_dva_unallocate() walks the 2488 * gang tree to immediately free (i.e. insert back into the space map) 2489 * everything we've allocated. This ensures that we don't get ENOSPC 2490 * errors during repeated suspend/resume cycles due to a flaky device. 2491 * 2492 * Gang rewrites only happen during sync-to-convergence. If we can't assemble 2493 * the gang tree, we won't modify the block, so we can safely defer the free 2494 * (knowing that the block is still intact). If we *can* assemble the gang 2495 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free 2496 * each constituent bp and we can allocate a new block on the next sync pass. 2497 * 2498 * In all cases, the gang tree allows complete recovery from partial failure. 2499 * ========================================================================== 2500 */ 2501 2502 static void 2503 zio_gang_issue_func_done(zio_t *zio) 2504 { 2505 abd_free(zio->io_abd); 2506 } 2507 2508 static zio_t * 2509 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, 2510 uint64_t offset) 2511 { 2512 if (gn != NULL) 2513 return (pio); 2514 2515 return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset), 2516 BP_GET_PSIZE(bp), zio_gang_issue_func_done, 2517 NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), 2518 &pio->io_bookmark)); 2519 } 2520 2521 static zio_t * 2522 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, 2523 uint64_t offset) 2524 { 2525 zio_t *zio; 2526 2527 if (gn != NULL) { 2528 abd_t *gbh_abd = 2529 abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE); 2530 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, 2531 gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL, 2532 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), 2533 &pio->io_bookmark); 2534 /* 2535 * As we rewrite each gang header, the pipeline will compute 2536 * a new gang block header checksum for it; but no one will 2537 * compute a new data checksum, so we do that here. The one 2538 * exception is the gang leader: the pipeline already computed 2539 * its data checksum because that stage precedes gang assembly. 2540 * (Presently, nothing actually uses interior data checksums; 2541 * this is just good hygiene.) 2542 */ 2543 if (gn != pio->io_gang_leader->io_gang_tree) { 2544 abd_t *buf = abd_get_offset(data, offset); 2545 2546 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp), 2547 buf, BP_GET_PSIZE(bp)); 2548 2549 abd_free(buf); 2550 } 2551 /* 2552 * If we are here to damage data for testing purposes, 2553 * leave the GBH alone so that we can detect the damage. 2554 */ 2555 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE) 2556 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; 2557 } else { 2558 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, 2559 abd_get_offset(data, offset), BP_GET_PSIZE(bp), 2560 zio_gang_issue_func_done, NULL, pio->io_priority, 2561 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 2562 } 2563 2564 return (zio); 2565 } 2566 2567 static zio_t * 2568 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, 2569 uint64_t offset) 2570 { 2571 (void) gn, (void) data, (void) offset; 2572 2573 zio_t *zio = zio_free_sync(pio, pio->io_spa, pio->io_txg, bp, 2574 ZIO_GANG_CHILD_FLAGS(pio)); 2575 if (zio == NULL) { 2576 zio = zio_null(pio, pio->io_spa, 2577 NULL, NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)); 2578 } 2579 return (zio); 2580 } 2581 2582 static zio_t * 2583 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, 2584 uint64_t offset) 2585 { 2586 (void) gn, (void) data, (void) offset; 2587 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp, 2588 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio))); 2589 } 2590 2591 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = { 2592 NULL, 2593 zio_read_gang, 2594 zio_rewrite_gang, 2595 zio_free_gang, 2596 zio_claim_gang, 2597 NULL 2598 }; 2599 2600 static void zio_gang_tree_assemble_done(zio_t *zio); 2601 2602 static zio_gang_node_t * 2603 zio_gang_node_alloc(zio_gang_node_t **gnpp) 2604 { 2605 zio_gang_node_t *gn; 2606 2607 ASSERT(*gnpp == NULL); 2608 2609 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP); 2610 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE); 2611 *gnpp = gn; 2612 2613 return (gn); 2614 } 2615 2616 static void 2617 zio_gang_node_free(zio_gang_node_t **gnpp) 2618 { 2619 zio_gang_node_t *gn = *gnpp; 2620 2621 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) 2622 ASSERT(gn->gn_child[g] == NULL); 2623 2624 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE); 2625 kmem_free(gn, sizeof (*gn)); 2626 *gnpp = NULL; 2627 } 2628 2629 static void 2630 zio_gang_tree_free(zio_gang_node_t **gnpp) 2631 { 2632 zio_gang_node_t *gn = *gnpp; 2633 2634 if (gn == NULL) 2635 return; 2636 2637 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) 2638 zio_gang_tree_free(&gn->gn_child[g]); 2639 2640 zio_gang_node_free(gnpp); 2641 } 2642 2643 static void 2644 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp) 2645 { 2646 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp); 2647 abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE); 2648 2649 ASSERT(gio->io_gang_leader == gio); 2650 ASSERT(BP_IS_GANG(bp)); 2651 2652 zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE, 2653 zio_gang_tree_assemble_done, gn, gio->io_priority, 2654 ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark)); 2655 } 2656 2657 static void 2658 zio_gang_tree_assemble_done(zio_t *zio) 2659 { 2660 zio_t *gio = zio->io_gang_leader; 2661 zio_gang_node_t *gn = zio->io_private; 2662 blkptr_t *bp = zio->io_bp; 2663 2664 ASSERT(gio == zio_unique_parent(zio)); 2665 ASSERT(zio->io_child_count == 0); 2666 2667 if (zio->io_error) 2668 return; 2669 2670 /* this ABD was created from a linear buf in zio_gang_tree_assemble */ 2671 if (BP_SHOULD_BYTESWAP(bp)) 2672 byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size); 2673 2674 ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh); 2675 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE); 2676 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); 2677 2678 abd_free(zio->io_abd); 2679 2680 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 2681 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; 2682 if (!BP_IS_GANG(gbp)) 2683 continue; 2684 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]); 2685 } 2686 } 2687 2688 static void 2689 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data, 2690 uint64_t offset) 2691 { 2692 zio_t *gio = pio->io_gang_leader; 2693 zio_t *zio; 2694 2695 ASSERT(BP_IS_GANG(bp) == !!gn); 2696 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp)); 2697 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree); 2698 2699 /* 2700 * If you're a gang header, your data is in gn->gn_gbh. 2701 * If you're a gang member, your data is in 'data' and gn == NULL. 2702 */ 2703 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset); 2704 2705 if (gn != NULL) { 2706 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); 2707 2708 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 2709 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; 2710 if (BP_IS_HOLE(gbp)) 2711 continue; 2712 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data, 2713 offset); 2714 offset += BP_GET_PSIZE(gbp); 2715 } 2716 } 2717 2718 if (gn == gio->io_gang_tree) 2719 ASSERT3U(gio->io_size, ==, offset); 2720 2721 if (zio != pio) 2722 zio_nowait(zio); 2723 } 2724 2725 static zio_t * 2726 zio_gang_assemble(zio_t *zio) 2727 { 2728 blkptr_t *bp = zio->io_bp; 2729 2730 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL); 2731 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 2732 2733 zio->io_gang_leader = zio; 2734 2735 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree); 2736 2737 return (zio); 2738 } 2739 2740 static zio_t * 2741 zio_gang_issue(zio_t *zio) 2742 { 2743 blkptr_t *bp = zio->io_bp; 2744 2745 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT, ZIO_WAIT_DONE)) { 2746 return (NULL); 2747 } 2748 2749 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio); 2750 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 2751 2752 if (zio->io_child_error[ZIO_CHILD_GANG] == 0) 2753 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd, 2754 0); 2755 else 2756 zio_gang_tree_free(&zio->io_gang_tree); 2757 2758 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2759 2760 return (zio); 2761 } 2762 2763 static void 2764 zio_write_gang_member_ready(zio_t *zio) 2765 { 2766 zio_t *pio = zio_unique_parent(zio); 2767 dva_t *cdva = zio->io_bp->blk_dva; 2768 dva_t *pdva = pio->io_bp->blk_dva; 2769 uint64_t asize; 2770 zio_t *gio __maybe_unused = zio->io_gang_leader; 2771 2772 if (BP_IS_HOLE(zio->io_bp)) 2773 return; 2774 2775 ASSERT(BP_IS_HOLE(&zio->io_bp_orig)); 2776 2777 ASSERT(zio->io_child_type == ZIO_CHILD_GANG); 2778 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies); 2779 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp)); 2780 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp)); 2781 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp)); 2782 2783 mutex_enter(&pio->io_lock); 2784 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) { 2785 ASSERT(DVA_GET_GANG(&pdva[d])); 2786 asize = DVA_GET_ASIZE(&pdva[d]); 2787 asize += DVA_GET_ASIZE(&cdva[d]); 2788 DVA_SET_ASIZE(&pdva[d], asize); 2789 } 2790 mutex_exit(&pio->io_lock); 2791 } 2792 2793 static void 2794 zio_write_gang_done(zio_t *zio) 2795 { 2796 /* 2797 * The io_abd field will be NULL for a zio with no data. The io_flags 2798 * will initially have the ZIO_FLAG_NODATA bit flag set, but we can't 2799 * check for it here as it is cleared in zio_ready. 2800 */ 2801 if (zio->io_abd != NULL) 2802 abd_free(zio->io_abd); 2803 } 2804 2805 static zio_t * 2806 zio_write_gang_block(zio_t *pio, metaslab_class_t *mc) 2807 { 2808 spa_t *spa = pio->io_spa; 2809 blkptr_t *bp = pio->io_bp; 2810 zio_t *gio = pio->io_gang_leader; 2811 zio_t *zio; 2812 zio_gang_node_t *gn, **gnpp; 2813 zio_gbh_phys_t *gbh; 2814 abd_t *gbh_abd; 2815 uint64_t txg = pio->io_txg; 2816 uint64_t resid = pio->io_size; 2817 uint64_t lsize; 2818 int copies = gio->io_prop.zp_copies; 2819 int gbh_copies; 2820 zio_prop_t zp; 2821 int error; 2822 boolean_t has_data = !(pio->io_flags & ZIO_FLAG_NODATA); 2823 2824 /* 2825 * encrypted blocks need DVA[2] free so encrypted gang headers can't 2826 * have a third copy. 2827 */ 2828 gbh_copies = MIN(copies + 1, spa_max_replication(spa)); 2829 if (BP_IS_ENCRYPTED(bp) && gbh_copies >= SPA_DVAS_PER_BP) 2830 gbh_copies = SPA_DVAS_PER_BP - 1; 2831 2832 int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER; 2833 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { 2834 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); 2835 ASSERT(has_data); 2836 2837 flags |= METASLAB_ASYNC_ALLOC; 2838 VERIFY(zfs_refcount_held(&mc->mc_allocator[pio->io_allocator]. 2839 mca_alloc_slots, pio)); 2840 2841 /* 2842 * The logical zio has already placed a reservation for 2843 * 'copies' allocation slots but gang blocks may require 2844 * additional copies. These additional copies 2845 * (i.e. gbh_copies - copies) are guaranteed to succeed 2846 * since metaslab_class_throttle_reserve() always allows 2847 * additional reservations for gang blocks. 2848 */ 2849 VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies, 2850 pio->io_allocator, pio, flags)); 2851 } 2852 2853 error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE, 2854 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags, 2855 &pio->io_alloc_list, pio, pio->io_allocator); 2856 if (error) { 2857 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { 2858 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); 2859 ASSERT(has_data); 2860 2861 /* 2862 * If we failed to allocate the gang block header then 2863 * we remove any additional allocation reservations that 2864 * we placed here. The original reservation will 2865 * be removed when the logical I/O goes to the ready 2866 * stage. 2867 */ 2868 metaslab_class_throttle_unreserve(mc, 2869 gbh_copies - copies, pio->io_allocator, pio); 2870 } 2871 2872 pio->io_error = error; 2873 return (pio); 2874 } 2875 2876 if (pio == gio) { 2877 gnpp = &gio->io_gang_tree; 2878 } else { 2879 gnpp = pio->io_private; 2880 ASSERT(pio->io_ready == zio_write_gang_member_ready); 2881 } 2882 2883 gn = zio_gang_node_alloc(gnpp); 2884 gbh = gn->gn_gbh; 2885 memset(gbh, 0, SPA_GANGBLOCKSIZE); 2886 gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE); 2887 2888 /* 2889 * Create the gang header. 2890 */ 2891 zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE, 2892 zio_write_gang_done, NULL, pio->io_priority, 2893 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 2894 2895 /* 2896 * Create and nowait the gang children. 2897 */ 2898 for (int g = 0; resid != 0; resid -= lsize, g++) { 2899 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g), 2900 SPA_MINBLOCKSIZE); 2901 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid); 2902 2903 zp.zp_checksum = gio->io_prop.zp_checksum; 2904 zp.zp_compress = ZIO_COMPRESS_OFF; 2905 zp.zp_complevel = gio->io_prop.zp_complevel; 2906 zp.zp_type = DMU_OT_NONE; 2907 zp.zp_level = 0; 2908 zp.zp_copies = gio->io_prop.zp_copies; 2909 zp.zp_dedup = B_FALSE; 2910 zp.zp_dedup_verify = B_FALSE; 2911 zp.zp_nopwrite = B_FALSE; 2912 zp.zp_encrypt = gio->io_prop.zp_encrypt; 2913 zp.zp_byteorder = gio->io_prop.zp_byteorder; 2914 memset(zp.zp_salt, 0, ZIO_DATA_SALT_LEN); 2915 memset(zp.zp_iv, 0, ZIO_DATA_IV_LEN); 2916 memset(zp.zp_mac, 0, ZIO_DATA_MAC_LEN); 2917 2918 zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g], 2919 has_data ? abd_get_offset(pio->io_abd, pio->io_size - 2920 resid) : NULL, lsize, lsize, &zp, 2921 zio_write_gang_member_ready, NULL, NULL, 2922 zio_write_gang_done, &gn->gn_child[g], pio->io_priority, 2923 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 2924 2925 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { 2926 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); 2927 ASSERT(has_data); 2928 2929 /* 2930 * Gang children won't throttle but we should 2931 * account for their work, so reserve an allocation 2932 * slot for them here. 2933 */ 2934 VERIFY(metaslab_class_throttle_reserve(mc, 2935 zp.zp_copies, cio->io_allocator, cio, flags)); 2936 } 2937 zio_nowait(cio); 2938 } 2939 2940 /* 2941 * Set pio's pipeline to just wait for zio to finish. 2942 */ 2943 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2944 2945 /* 2946 * We didn't allocate this bp, so make sure it doesn't get unmarked. 2947 */ 2948 pio->io_flags &= ~ZIO_FLAG_FASTWRITE; 2949 2950 zio_nowait(zio); 2951 2952 return (pio); 2953 } 2954 2955 /* 2956 * The zio_nop_write stage in the pipeline determines if allocating a 2957 * new bp is necessary. The nopwrite feature can handle writes in 2958 * either syncing or open context (i.e. zil writes) and as a result is 2959 * mutually exclusive with dedup. 2960 * 2961 * By leveraging a cryptographically secure checksum, such as SHA256, we 2962 * can compare the checksums of the new data and the old to determine if 2963 * allocating a new block is required. Note that our requirements for 2964 * cryptographic strength are fairly weak: there can't be any accidental 2965 * hash collisions, but we don't need to be secure against intentional 2966 * (malicious) collisions. To trigger a nopwrite, you have to be able 2967 * to write the file to begin with, and triggering an incorrect (hash 2968 * collision) nopwrite is no worse than simply writing to the file. 2969 * That said, there are no known attacks against the checksum algorithms 2970 * used for nopwrite, assuming that the salt and the checksums 2971 * themselves remain secret. 2972 */ 2973 static zio_t * 2974 zio_nop_write(zio_t *zio) 2975 { 2976 blkptr_t *bp = zio->io_bp; 2977 blkptr_t *bp_orig = &zio->io_bp_orig; 2978 zio_prop_t *zp = &zio->io_prop; 2979 2980 ASSERT(BP_IS_HOLE(bp)); 2981 ASSERT(BP_GET_LEVEL(bp) == 0); 2982 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 2983 ASSERT(zp->zp_nopwrite); 2984 ASSERT(!zp->zp_dedup); 2985 ASSERT(zio->io_bp_override == NULL); 2986 ASSERT(IO_IS_ALLOCATING(zio)); 2987 2988 /* 2989 * Check to see if the original bp and the new bp have matching 2990 * characteristics (i.e. same checksum, compression algorithms, etc). 2991 * If they don't then just continue with the pipeline which will 2992 * allocate a new bp. 2993 */ 2994 if (BP_IS_HOLE(bp_orig) || 2995 !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags & 2996 ZCHECKSUM_FLAG_NOPWRITE) || 2997 BP_IS_ENCRYPTED(bp) || BP_IS_ENCRYPTED(bp_orig) || 2998 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) || 2999 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) || 3000 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) || 3001 zp->zp_copies != BP_GET_NDVAS(bp_orig)) 3002 return (zio); 3003 3004 /* 3005 * If the checksums match then reset the pipeline so that we 3006 * avoid allocating a new bp and issuing any I/O. 3007 */ 3008 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) { 3009 ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags & 3010 ZCHECKSUM_FLAG_NOPWRITE); 3011 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig)); 3012 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig)); 3013 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF); 3014 ASSERT3U(bp->blk_prop, ==, bp_orig->blk_prop); 3015 3016 /* 3017 * If we're overwriting a block that is currently on an 3018 * indirect vdev, then ignore the nopwrite request and 3019 * allow a new block to be allocated on a concrete vdev. 3020 */ 3021 spa_config_enter(zio->io_spa, SCL_VDEV, FTAG, RW_READER); 3022 for (int d = 0; d < BP_GET_NDVAS(bp_orig); d++) { 3023 vdev_t *tvd = vdev_lookup_top(zio->io_spa, 3024 DVA_GET_VDEV(&bp_orig->blk_dva[d])); 3025 if (tvd->vdev_ops == &vdev_indirect_ops) { 3026 spa_config_exit(zio->io_spa, SCL_VDEV, FTAG); 3027 return (zio); 3028 } 3029 } 3030 spa_config_exit(zio->io_spa, SCL_VDEV, FTAG); 3031 3032 *bp = *bp_orig; 3033 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 3034 zio->io_flags |= ZIO_FLAG_NOPWRITE; 3035 } 3036 3037 return (zio); 3038 } 3039 3040 /* 3041 * ========================================================================== 3042 * Dedup 3043 * ========================================================================== 3044 */ 3045 static void 3046 zio_ddt_child_read_done(zio_t *zio) 3047 { 3048 blkptr_t *bp = zio->io_bp; 3049 ddt_entry_t *dde = zio->io_private; 3050 ddt_phys_t *ddp; 3051 zio_t *pio = zio_unique_parent(zio); 3052 3053 mutex_enter(&pio->io_lock); 3054 ddp = ddt_phys_select(dde, bp); 3055 if (zio->io_error == 0) 3056 ddt_phys_clear(ddp); /* this ddp doesn't need repair */ 3057 3058 if (zio->io_error == 0 && dde->dde_repair_abd == NULL) 3059 dde->dde_repair_abd = zio->io_abd; 3060 else 3061 abd_free(zio->io_abd); 3062 mutex_exit(&pio->io_lock); 3063 } 3064 3065 static zio_t * 3066 zio_ddt_read_start(zio_t *zio) 3067 { 3068 blkptr_t *bp = zio->io_bp; 3069 3070 ASSERT(BP_GET_DEDUP(bp)); 3071 ASSERT(BP_GET_PSIZE(bp) == zio->io_size); 3072 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 3073 3074 if (zio->io_child_error[ZIO_CHILD_DDT]) { 3075 ddt_t *ddt = ddt_select(zio->io_spa, bp); 3076 ddt_entry_t *dde = ddt_repair_start(ddt, bp); 3077 ddt_phys_t *ddp = dde->dde_phys; 3078 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp); 3079 blkptr_t blk; 3080 3081 ASSERT(zio->io_vsd == NULL); 3082 zio->io_vsd = dde; 3083 3084 if (ddp_self == NULL) 3085 return (zio); 3086 3087 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) { 3088 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self) 3089 continue; 3090 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp, 3091 &blk); 3092 zio_nowait(zio_read(zio, zio->io_spa, &blk, 3093 abd_alloc_for_io(zio->io_size, B_TRUE), 3094 zio->io_size, zio_ddt_child_read_done, dde, 3095 zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) | 3096 ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark)); 3097 } 3098 return (zio); 3099 } 3100 3101 zio_nowait(zio_read(zio, zio->io_spa, bp, 3102 zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority, 3103 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark)); 3104 3105 return (zio); 3106 } 3107 3108 static zio_t * 3109 zio_ddt_read_done(zio_t *zio) 3110 { 3111 blkptr_t *bp = zio->io_bp; 3112 3113 if (zio_wait_for_children(zio, ZIO_CHILD_DDT_BIT, ZIO_WAIT_DONE)) { 3114 return (NULL); 3115 } 3116 3117 ASSERT(BP_GET_DEDUP(bp)); 3118 ASSERT(BP_GET_PSIZE(bp) == zio->io_size); 3119 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 3120 3121 if (zio->io_child_error[ZIO_CHILD_DDT]) { 3122 ddt_t *ddt = ddt_select(zio->io_spa, bp); 3123 ddt_entry_t *dde = zio->io_vsd; 3124 if (ddt == NULL) { 3125 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE); 3126 return (zio); 3127 } 3128 if (dde == NULL) { 3129 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1; 3130 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); 3131 return (NULL); 3132 } 3133 if (dde->dde_repair_abd != NULL) { 3134 abd_copy(zio->io_abd, dde->dde_repair_abd, 3135 zio->io_size); 3136 zio->io_child_error[ZIO_CHILD_DDT] = 0; 3137 } 3138 ddt_repair_done(ddt, dde); 3139 zio->io_vsd = NULL; 3140 } 3141 3142 ASSERT(zio->io_vsd == NULL); 3143 3144 return (zio); 3145 } 3146 3147 static boolean_t 3148 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde) 3149 { 3150 spa_t *spa = zio->io_spa; 3151 boolean_t do_raw = !!(zio->io_flags & ZIO_FLAG_RAW); 3152 3153 ASSERT(!(zio->io_bp_override && do_raw)); 3154 3155 /* 3156 * Note: we compare the original data, not the transformed data, 3157 * because when zio->io_bp is an override bp, we will not have 3158 * pushed the I/O transforms. That's an important optimization 3159 * because otherwise we'd compress/encrypt all dmu_sync() data twice. 3160 * However, we should never get a raw, override zio so in these 3161 * cases we can compare the io_abd directly. This is useful because 3162 * it allows us to do dedup verification even if we don't have access 3163 * to the original data (for instance, if the encryption keys aren't 3164 * loaded). 3165 */ 3166 3167 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { 3168 zio_t *lio = dde->dde_lead_zio[p]; 3169 3170 if (lio != NULL && do_raw) { 3171 return (lio->io_size != zio->io_size || 3172 abd_cmp(zio->io_abd, lio->io_abd) != 0); 3173 } else if (lio != NULL) { 3174 return (lio->io_orig_size != zio->io_orig_size || 3175 abd_cmp(zio->io_orig_abd, lio->io_orig_abd) != 0); 3176 } 3177 } 3178 3179 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { 3180 ddt_phys_t *ddp = &dde->dde_phys[p]; 3181 3182 if (ddp->ddp_phys_birth != 0 && do_raw) { 3183 blkptr_t blk = *zio->io_bp; 3184 uint64_t psize; 3185 abd_t *tmpabd; 3186 int error; 3187 3188 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth); 3189 psize = BP_GET_PSIZE(&blk); 3190 3191 if (psize != zio->io_size) 3192 return (B_TRUE); 3193 3194 ddt_exit(ddt); 3195 3196 tmpabd = abd_alloc_for_io(psize, B_TRUE); 3197 3198 error = zio_wait(zio_read(NULL, spa, &blk, tmpabd, 3199 psize, NULL, NULL, ZIO_PRIORITY_SYNC_READ, 3200 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | 3201 ZIO_FLAG_RAW, &zio->io_bookmark)); 3202 3203 if (error == 0) { 3204 if (abd_cmp(tmpabd, zio->io_abd) != 0) 3205 error = SET_ERROR(ENOENT); 3206 } 3207 3208 abd_free(tmpabd); 3209 ddt_enter(ddt); 3210 return (error != 0); 3211 } else if (ddp->ddp_phys_birth != 0) { 3212 arc_buf_t *abuf = NULL; 3213 arc_flags_t aflags = ARC_FLAG_WAIT; 3214 blkptr_t blk = *zio->io_bp; 3215 int error; 3216 3217 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth); 3218 3219 if (BP_GET_LSIZE(&blk) != zio->io_orig_size) 3220 return (B_TRUE); 3221 3222 ddt_exit(ddt); 3223 3224 error = arc_read(NULL, spa, &blk, 3225 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ, 3226 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 3227 &aflags, &zio->io_bookmark); 3228 3229 if (error == 0) { 3230 if (abd_cmp_buf(zio->io_orig_abd, abuf->b_data, 3231 zio->io_orig_size) != 0) 3232 error = SET_ERROR(ENOENT); 3233 arc_buf_destroy(abuf, &abuf); 3234 } 3235 3236 ddt_enter(ddt); 3237 return (error != 0); 3238 } 3239 } 3240 3241 return (B_FALSE); 3242 } 3243 3244 static void 3245 zio_ddt_child_write_ready(zio_t *zio) 3246 { 3247 int p = zio->io_prop.zp_copies; 3248 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); 3249 ddt_entry_t *dde = zio->io_private; 3250 ddt_phys_t *ddp = &dde->dde_phys[p]; 3251 zio_t *pio; 3252 3253 if (zio->io_error) 3254 return; 3255 3256 ddt_enter(ddt); 3257 3258 ASSERT(dde->dde_lead_zio[p] == zio); 3259 3260 ddt_phys_fill(ddp, zio->io_bp); 3261 3262 zio_link_t *zl = NULL; 3263 while ((pio = zio_walk_parents(zio, &zl)) != NULL) 3264 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg); 3265 3266 ddt_exit(ddt); 3267 } 3268 3269 static void 3270 zio_ddt_child_write_done(zio_t *zio) 3271 { 3272 int p = zio->io_prop.zp_copies; 3273 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); 3274 ddt_entry_t *dde = zio->io_private; 3275 ddt_phys_t *ddp = &dde->dde_phys[p]; 3276 3277 ddt_enter(ddt); 3278 3279 ASSERT(ddp->ddp_refcnt == 0); 3280 ASSERT(dde->dde_lead_zio[p] == zio); 3281 dde->dde_lead_zio[p] = NULL; 3282 3283 if (zio->io_error == 0) { 3284 zio_link_t *zl = NULL; 3285 while (zio_walk_parents(zio, &zl) != NULL) 3286 ddt_phys_addref(ddp); 3287 } else { 3288 ddt_phys_clear(ddp); 3289 } 3290 3291 ddt_exit(ddt); 3292 } 3293 3294 static zio_t * 3295 zio_ddt_write(zio_t *zio) 3296 { 3297 spa_t *spa = zio->io_spa; 3298 blkptr_t *bp = zio->io_bp; 3299 uint64_t txg = zio->io_txg; 3300 zio_prop_t *zp = &zio->io_prop; 3301 int p = zp->zp_copies; 3302 zio_t *cio = NULL; 3303 ddt_t *ddt = ddt_select(spa, bp); 3304 ddt_entry_t *dde; 3305 ddt_phys_t *ddp; 3306 3307 ASSERT(BP_GET_DEDUP(bp)); 3308 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum); 3309 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override); 3310 ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW))); 3311 3312 ddt_enter(ddt); 3313 dde = ddt_lookup(ddt, bp, B_TRUE); 3314 ddp = &dde->dde_phys[p]; 3315 3316 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) { 3317 /* 3318 * If we're using a weak checksum, upgrade to a strong checksum 3319 * and try again. If we're already using a strong checksum, 3320 * we can't resolve it, so just convert to an ordinary write. 3321 * (And automatically e-mail a paper to Nature?) 3322 */ 3323 if (!(zio_checksum_table[zp->zp_checksum].ci_flags & 3324 ZCHECKSUM_FLAG_DEDUP)) { 3325 zp->zp_checksum = spa_dedup_checksum(spa); 3326 zio_pop_transforms(zio); 3327 zio->io_stage = ZIO_STAGE_OPEN; 3328 BP_ZERO(bp); 3329 } else { 3330 zp->zp_dedup = B_FALSE; 3331 BP_SET_DEDUP(bp, B_FALSE); 3332 } 3333 ASSERT(!BP_GET_DEDUP(bp)); 3334 zio->io_pipeline = ZIO_WRITE_PIPELINE; 3335 ddt_exit(ddt); 3336 return (zio); 3337 } 3338 3339 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) { 3340 if (ddp->ddp_phys_birth != 0) 3341 ddt_bp_fill(ddp, bp, txg); 3342 if (dde->dde_lead_zio[p] != NULL) 3343 zio_add_child(zio, dde->dde_lead_zio[p]); 3344 else 3345 ddt_phys_addref(ddp); 3346 } else if (zio->io_bp_override) { 3347 ASSERT(bp->blk_birth == txg); 3348 ASSERT(BP_EQUAL(bp, zio->io_bp_override)); 3349 ddt_phys_fill(ddp, bp); 3350 ddt_phys_addref(ddp); 3351 } else { 3352 cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd, 3353 zio->io_orig_size, zio->io_orig_size, zp, 3354 zio_ddt_child_write_ready, NULL, NULL, 3355 zio_ddt_child_write_done, dde, zio->io_priority, 3356 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); 3357 3358 zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL); 3359 dde->dde_lead_zio[p] = cio; 3360 } 3361 3362 ddt_exit(ddt); 3363 3364 zio_nowait(cio); 3365 3366 return (zio); 3367 } 3368 3369 static ddt_entry_t *freedde; /* for debugging */ 3370 3371 static zio_t * 3372 zio_ddt_free(zio_t *zio) 3373 { 3374 spa_t *spa = zio->io_spa; 3375 blkptr_t *bp = zio->io_bp; 3376 ddt_t *ddt = ddt_select(spa, bp); 3377 ddt_entry_t *dde; 3378 ddt_phys_t *ddp; 3379 3380 ASSERT(BP_GET_DEDUP(bp)); 3381 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 3382 3383 ddt_enter(ddt); 3384 freedde = dde = ddt_lookup(ddt, bp, B_TRUE); 3385 if (dde) { 3386 ddp = ddt_phys_select(dde, bp); 3387 if (ddp) 3388 ddt_phys_decref(ddp); 3389 } 3390 ddt_exit(ddt); 3391 3392 return (zio); 3393 } 3394 3395 /* 3396 * ========================================================================== 3397 * Allocate and free blocks 3398 * ========================================================================== 3399 */ 3400 3401 static zio_t * 3402 zio_io_to_allocate(spa_t *spa, int allocator) 3403 { 3404 zio_t *zio; 3405 3406 ASSERT(MUTEX_HELD(&spa->spa_allocs[allocator].spaa_lock)); 3407 3408 zio = avl_first(&spa->spa_allocs[allocator].spaa_tree); 3409 if (zio == NULL) 3410 return (NULL); 3411 3412 ASSERT(IO_IS_ALLOCATING(zio)); 3413 3414 /* 3415 * Try to place a reservation for this zio. If we're unable to 3416 * reserve then we throttle. 3417 */ 3418 ASSERT3U(zio->io_allocator, ==, allocator); 3419 if (!metaslab_class_throttle_reserve(zio->io_metaslab_class, 3420 zio->io_prop.zp_copies, allocator, zio, 0)) { 3421 return (NULL); 3422 } 3423 3424 avl_remove(&spa->spa_allocs[allocator].spaa_tree, zio); 3425 ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE); 3426 3427 return (zio); 3428 } 3429 3430 static zio_t * 3431 zio_dva_throttle(zio_t *zio) 3432 { 3433 spa_t *spa = zio->io_spa; 3434 zio_t *nio; 3435 metaslab_class_t *mc; 3436 3437 /* locate an appropriate allocation class */ 3438 mc = spa_preferred_class(spa, zio->io_size, zio->io_prop.zp_type, 3439 zio->io_prop.zp_level, zio->io_prop.zp_zpl_smallblk); 3440 3441 if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE || 3442 !mc->mc_alloc_throttle_enabled || 3443 zio->io_child_type == ZIO_CHILD_GANG || 3444 zio->io_flags & ZIO_FLAG_NODATA) { 3445 return (zio); 3446 } 3447 3448 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 3449 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 3450 ASSERT3U(zio->io_queued_timestamp, >, 0); 3451 ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE); 3452 3453 zbookmark_phys_t *bm = &zio->io_bookmark; 3454 /* 3455 * We want to try to use as many allocators as possible to help improve 3456 * performance, but we also want logically adjacent IOs to be physically 3457 * adjacent to improve sequential read performance. We chunk each object 3458 * into 2^20 block regions, and then hash based on the objset, object, 3459 * level, and region to accomplish both of these goals. 3460 */ 3461 int allocator = (uint_t)cityhash4(bm->zb_objset, bm->zb_object, 3462 bm->zb_level, bm->zb_blkid >> 20) % spa->spa_alloc_count; 3463 zio->io_allocator = allocator; 3464 zio->io_metaslab_class = mc; 3465 mutex_enter(&spa->spa_allocs[allocator].spaa_lock); 3466 avl_add(&spa->spa_allocs[allocator].spaa_tree, zio); 3467 nio = zio_io_to_allocate(spa, allocator); 3468 mutex_exit(&spa->spa_allocs[allocator].spaa_lock); 3469 return (nio); 3470 } 3471 3472 static void 3473 zio_allocate_dispatch(spa_t *spa, int allocator) 3474 { 3475 zio_t *zio; 3476 3477 mutex_enter(&spa->spa_allocs[allocator].spaa_lock); 3478 zio = zio_io_to_allocate(spa, allocator); 3479 mutex_exit(&spa->spa_allocs[allocator].spaa_lock); 3480 if (zio == NULL) 3481 return; 3482 3483 ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE); 3484 ASSERT0(zio->io_error); 3485 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE); 3486 } 3487 3488 static zio_t * 3489 zio_dva_allocate(zio_t *zio) 3490 { 3491 spa_t *spa = zio->io_spa; 3492 metaslab_class_t *mc; 3493 blkptr_t *bp = zio->io_bp; 3494 int error; 3495 int flags = 0; 3496 3497 if (zio->io_gang_leader == NULL) { 3498 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 3499 zio->io_gang_leader = zio; 3500 } 3501 3502 ASSERT(BP_IS_HOLE(bp)); 3503 ASSERT0(BP_GET_NDVAS(bp)); 3504 ASSERT3U(zio->io_prop.zp_copies, >, 0); 3505 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa)); 3506 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp)); 3507 3508 flags |= (zio->io_flags & ZIO_FLAG_FASTWRITE) ? METASLAB_FASTWRITE : 0; 3509 if (zio->io_flags & ZIO_FLAG_NODATA) 3510 flags |= METASLAB_DONT_THROTTLE; 3511 if (zio->io_flags & ZIO_FLAG_GANG_CHILD) 3512 flags |= METASLAB_GANG_CHILD; 3513 if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE) 3514 flags |= METASLAB_ASYNC_ALLOC; 3515 3516 /* 3517 * if not already chosen, locate an appropriate allocation class 3518 */ 3519 mc = zio->io_metaslab_class; 3520 if (mc == NULL) { 3521 mc = spa_preferred_class(spa, zio->io_size, 3522 zio->io_prop.zp_type, zio->io_prop.zp_level, 3523 zio->io_prop.zp_zpl_smallblk); 3524 zio->io_metaslab_class = mc; 3525 } 3526 3527 /* 3528 * Try allocating the block in the usual metaslab class. 3529 * If that's full, allocate it in the normal class. 3530 * If that's full, allocate as a gang block, 3531 * and if all are full, the allocation fails (which shouldn't happen). 3532 * 3533 * Note that we do not fall back on embedded slog (ZIL) space, to 3534 * preserve unfragmented slog space, which is critical for decent 3535 * sync write performance. If a log allocation fails, we will fall 3536 * back to spa_sync() which is abysmal for performance. 3537 */ 3538 error = metaslab_alloc(spa, mc, zio->io_size, bp, 3539 zio->io_prop.zp_copies, zio->io_txg, NULL, flags, 3540 &zio->io_alloc_list, zio, zio->io_allocator); 3541 3542 /* 3543 * Fallback to normal class when an alloc class is full 3544 */ 3545 if (error == ENOSPC && mc != spa_normal_class(spa)) { 3546 /* 3547 * If throttling, transfer reservation over to normal class. 3548 * The io_allocator slot can remain the same even though we 3549 * are switching classes. 3550 */ 3551 if (mc->mc_alloc_throttle_enabled && 3552 (zio->io_flags & ZIO_FLAG_IO_ALLOCATING)) { 3553 metaslab_class_throttle_unreserve(mc, 3554 zio->io_prop.zp_copies, zio->io_allocator, zio); 3555 zio->io_flags &= ~ZIO_FLAG_IO_ALLOCATING; 3556 3557 VERIFY(metaslab_class_throttle_reserve( 3558 spa_normal_class(spa), 3559 zio->io_prop.zp_copies, zio->io_allocator, zio, 3560 flags | METASLAB_MUST_RESERVE)); 3561 } 3562 zio->io_metaslab_class = mc = spa_normal_class(spa); 3563 if (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC) { 3564 zfs_dbgmsg("%s: metaslab allocation failure, " 3565 "trying normal class: zio %px, size %llu, error %d", 3566 spa_name(spa), zio, (u_longlong_t)zio->io_size, 3567 error); 3568 } 3569 3570 error = metaslab_alloc(spa, mc, zio->io_size, bp, 3571 zio->io_prop.zp_copies, zio->io_txg, NULL, flags, 3572 &zio->io_alloc_list, zio, zio->io_allocator); 3573 } 3574 3575 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE) { 3576 if (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC) { 3577 zfs_dbgmsg("%s: metaslab allocation failure, " 3578 "trying ganging: zio %px, size %llu, error %d", 3579 spa_name(spa), zio, (u_longlong_t)zio->io_size, 3580 error); 3581 } 3582 return (zio_write_gang_block(zio, mc)); 3583 } 3584 if (error != 0) { 3585 if (error != ENOSPC || 3586 (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC)) { 3587 zfs_dbgmsg("%s: metaslab allocation failure: zio %px, " 3588 "size %llu, error %d", 3589 spa_name(spa), zio, (u_longlong_t)zio->io_size, 3590 error); 3591 } 3592 zio->io_error = error; 3593 } 3594 3595 return (zio); 3596 } 3597 3598 static zio_t * 3599 zio_dva_free(zio_t *zio) 3600 { 3601 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE); 3602 3603 return (zio); 3604 } 3605 3606 static zio_t * 3607 zio_dva_claim(zio_t *zio) 3608 { 3609 int error; 3610 3611 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg); 3612 if (error) 3613 zio->io_error = error; 3614 3615 return (zio); 3616 } 3617 3618 /* 3619 * Undo an allocation. This is used by zio_done() when an I/O fails 3620 * and we want to give back the block we just allocated. 3621 * This handles both normal blocks and gang blocks. 3622 */ 3623 static void 3624 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp) 3625 { 3626 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp)); 3627 ASSERT(zio->io_bp_override == NULL); 3628 3629 if (!BP_IS_HOLE(bp)) 3630 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE); 3631 3632 if (gn != NULL) { 3633 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 3634 zio_dva_unallocate(zio, gn->gn_child[g], 3635 &gn->gn_gbh->zg_blkptr[g]); 3636 } 3637 } 3638 } 3639 3640 /* 3641 * Try to allocate an intent log block. Return 0 on success, errno on failure. 3642 */ 3643 int 3644 zio_alloc_zil(spa_t *spa, objset_t *os, uint64_t txg, blkptr_t *new_bp, 3645 uint64_t size, boolean_t *slog) 3646 { 3647 int error = 1; 3648 zio_alloc_list_t io_alloc_list; 3649 3650 ASSERT(txg > spa_syncing_txg(spa)); 3651 3652 metaslab_trace_init(&io_alloc_list); 3653 3654 /* 3655 * Block pointer fields are useful to metaslabs for stats and debugging. 3656 * Fill in the obvious ones before calling into metaslab_alloc(). 3657 */ 3658 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); 3659 BP_SET_PSIZE(new_bp, size); 3660 BP_SET_LEVEL(new_bp, 0); 3661 3662 /* 3663 * When allocating a zil block, we don't have information about 3664 * the final destination of the block except the objset it's part 3665 * of, so we just hash the objset ID to pick the allocator to get 3666 * some parallelism. 3667 */ 3668 int flags = METASLAB_FASTWRITE | METASLAB_ZIL; 3669 int allocator = (uint_t)cityhash4(0, 0, 0, 3670 os->os_dsl_dataset->ds_object) % spa->spa_alloc_count; 3671 error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1, 3672 txg, NULL, flags, &io_alloc_list, NULL, allocator); 3673 *slog = (error == 0); 3674 if (error != 0) { 3675 error = metaslab_alloc(spa, spa_embedded_log_class(spa), size, 3676 new_bp, 1, txg, NULL, flags, 3677 &io_alloc_list, NULL, allocator); 3678 } 3679 if (error != 0) { 3680 error = metaslab_alloc(spa, spa_normal_class(spa), size, 3681 new_bp, 1, txg, NULL, flags, 3682 &io_alloc_list, NULL, allocator); 3683 } 3684 metaslab_trace_fini(&io_alloc_list); 3685 3686 if (error == 0) { 3687 BP_SET_LSIZE(new_bp, size); 3688 BP_SET_PSIZE(new_bp, size); 3689 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF); 3690 BP_SET_CHECKSUM(new_bp, 3691 spa_version(spa) >= SPA_VERSION_SLIM_ZIL 3692 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG); 3693 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); 3694 BP_SET_LEVEL(new_bp, 0); 3695 BP_SET_DEDUP(new_bp, 0); 3696 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER); 3697 3698 /* 3699 * encrypted blocks will require an IV and salt. We generate 3700 * these now since we will not be rewriting the bp at 3701 * rewrite time. 3702 */ 3703 if (os->os_encrypted) { 3704 uint8_t iv[ZIO_DATA_IV_LEN]; 3705 uint8_t salt[ZIO_DATA_SALT_LEN]; 3706 3707 BP_SET_CRYPT(new_bp, B_TRUE); 3708 VERIFY0(spa_crypt_get_salt(spa, 3709 dmu_objset_id(os), salt)); 3710 VERIFY0(zio_crypt_generate_iv(iv)); 3711 3712 zio_crypt_encode_params_bp(new_bp, salt, iv); 3713 } 3714 } else { 3715 zfs_dbgmsg("%s: zil block allocation failure: " 3716 "size %llu, error %d", spa_name(spa), (u_longlong_t)size, 3717 error); 3718 } 3719 3720 return (error); 3721 } 3722 3723 /* 3724 * ========================================================================== 3725 * Read and write to physical devices 3726 * ========================================================================== 3727 */ 3728 3729 /* 3730 * Issue an I/O to the underlying vdev. Typically the issue pipeline 3731 * stops after this stage and will resume upon I/O completion. 3732 * However, there are instances where the vdev layer may need to 3733 * continue the pipeline when an I/O was not issued. Since the I/O 3734 * that was sent to the vdev layer might be different than the one 3735 * currently active in the pipeline (see vdev_queue_io()), we explicitly 3736 * force the underlying vdev layers to call either zio_execute() or 3737 * zio_interrupt() to ensure that the pipeline continues with the correct I/O. 3738 */ 3739 static zio_t * 3740 zio_vdev_io_start(zio_t *zio) 3741 { 3742 vdev_t *vd = zio->io_vd; 3743 uint64_t align; 3744 spa_t *spa = zio->io_spa; 3745 3746 zio->io_delay = 0; 3747 3748 ASSERT(zio->io_error == 0); 3749 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0); 3750 3751 if (vd == NULL) { 3752 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) 3753 spa_config_enter(spa, SCL_ZIO, zio, RW_READER); 3754 3755 /* 3756 * The mirror_ops handle multiple DVAs in a single BP. 3757 */ 3758 vdev_mirror_ops.vdev_op_io_start(zio); 3759 return (NULL); 3760 } 3761 3762 ASSERT3P(zio->io_logical, !=, zio); 3763 if (zio->io_type == ZIO_TYPE_WRITE) { 3764 ASSERT(spa->spa_trust_config); 3765 3766 /* 3767 * Note: the code can handle other kinds of writes, 3768 * but we don't expect them. 3769 */ 3770 if (zio->io_vd->vdev_noalloc) { 3771 ASSERT(zio->io_flags & 3772 (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL | 3773 ZIO_FLAG_RESILVER | ZIO_FLAG_INDUCE_DAMAGE)); 3774 } 3775 } 3776 3777 align = 1ULL << vd->vdev_top->vdev_ashift; 3778 3779 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) && 3780 P2PHASE(zio->io_size, align) != 0) { 3781 /* Transform logical writes to be a full physical block size. */ 3782 uint64_t asize = P2ROUNDUP(zio->io_size, align); 3783 abd_t *abuf = abd_alloc_sametype(zio->io_abd, asize); 3784 ASSERT(vd == vd->vdev_top); 3785 if (zio->io_type == ZIO_TYPE_WRITE) { 3786 abd_copy(abuf, zio->io_abd, zio->io_size); 3787 abd_zero_off(abuf, zio->io_size, asize - zio->io_size); 3788 } 3789 zio_push_transform(zio, abuf, asize, asize, zio_subblock); 3790 } 3791 3792 /* 3793 * If this is not a physical io, make sure that it is properly aligned 3794 * before proceeding. 3795 */ 3796 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) { 3797 ASSERT0(P2PHASE(zio->io_offset, align)); 3798 ASSERT0(P2PHASE(zio->io_size, align)); 3799 } else { 3800 /* 3801 * For physical writes, we allow 512b aligned writes and assume 3802 * the device will perform a read-modify-write as necessary. 3803 */ 3804 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE)); 3805 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE)); 3806 } 3807 3808 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa)); 3809 3810 /* 3811 * If this is a repair I/O, and there's no self-healing involved -- 3812 * that is, we're just resilvering what we expect to resilver -- 3813 * then don't do the I/O unless zio's txg is actually in vd's DTL. 3814 * This prevents spurious resilvering. 3815 * 3816 * There are a few ways that we can end up creating these spurious 3817 * resilver i/os: 3818 * 3819 * 1. A resilver i/o will be issued if any DVA in the BP has a 3820 * dirty DTL. The mirror code will issue resilver writes to 3821 * each DVA, including the one(s) that are not on vdevs with dirty 3822 * DTLs. 3823 * 3824 * 2. With nested replication, which happens when we have a 3825 * "replacing" or "spare" vdev that's a child of a mirror or raidz. 3826 * For example, given mirror(replacing(A+B), C), it's likely that 3827 * only A is out of date (it's the new device). In this case, we'll 3828 * read from C, then use the data to resilver A+B -- but we don't 3829 * actually want to resilver B, just A. The top-level mirror has no 3830 * way to know this, so instead we just discard unnecessary repairs 3831 * as we work our way down the vdev tree. 3832 * 3833 * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc. 3834 * The same logic applies to any form of nested replication: ditto 3835 * + mirror, RAID-Z + replacing, etc. 3836 * 3837 * However, indirect vdevs point off to other vdevs which may have 3838 * DTL's, so we never bypass them. The child i/os on concrete vdevs 3839 * will be properly bypassed instead. 3840 * 3841 * Leaf DTL_PARTIAL can be empty when a legitimate write comes from 3842 * a dRAID spare vdev. For example, when a dRAID spare is first 3843 * used, its spare blocks need to be written to but the leaf vdev's 3844 * of such blocks can have empty DTL_PARTIAL. 3845 * 3846 * There seemed no clean way to allow such writes while bypassing 3847 * spurious ones. At this point, just avoid all bypassing for dRAID 3848 * for correctness. 3849 */ 3850 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) && 3851 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) && 3852 zio->io_txg != 0 && /* not a delegated i/o */ 3853 vd->vdev_ops != &vdev_indirect_ops && 3854 vd->vdev_top->vdev_ops != &vdev_draid_ops && 3855 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) { 3856 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 3857 zio_vdev_io_bypass(zio); 3858 return (zio); 3859 } 3860 3861 /* 3862 * Select the next best leaf I/O to process. Distributed spares are 3863 * excluded since they dispatch the I/O directly to a leaf vdev after 3864 * applying the dRAID mapping. 3865 */ 3866 if (vd->vdev_ops->vdev_op_leaf && 3867 vd->vdev_ops != &vdev_draid_spare_ops && 3868 (zio->io_type == ZIO_TYPE_READ || 3869 zio->io_type == ZIO_TYPE_WRITE || 3870 zio->io_type == ZIO_TYPE_TRIM)) { 3871 3872 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio)) 3873 return (zio); 3874 3875 if ((zio = vdev_queue_io(zio)) == NULL) 3876 return (NULL); 3877 3878 if (!vdev_accessible(vd, zio)) { 3879 zio->io_error = SET_ERROR(ENXIO); 3880 zio_interrupt(zio); 3881 return (NULL); 3882 } 3883 zio->io_delay = gethrtime(); 3884 } 3885 3886 vd->vdev_ops->vdev_op_io_start(zio); 3887 return (NULL); 3888 } 3889 3890 static zio_t * 3891 zio_vdev_io_done(zio_t *zio) 3892 { 3893 vdev_t *vd = zio->io_vd; 3894 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops; 3895 boolean_t unexpected_error = B_FALSE; 3896 3897 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) { 3898 return (NULL); 3899 } 3900 3901 ASSERT(zio->io_type == ZIO_TYPE_READ || 3902 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_TRIM); 3903 3904 if (zio->io_delay) 3905 zio->io_delay = gethrtime() - zio->io_delay; 3906 3907 if (vd != NULL && vd->vdev_ops->vdev_op_leaf && 3908 vd->vdev_ops != &vdev_draid_spare_ops) { 3909 vdev_queue_io_done(zio); 3910 3911 if (zio->io_type == ZIO_TYPE_WRITE) 3912 vdev_cache_write(zio); 3913 3914 if (zio_injection_enabled && zio->io_error == 0) 3915 zio->io_error = zio_handle_device_injections(vd, zio, 3916 EIO, EILSEQ); 3917 3918 if (zio_injection_enabled && zio->io_error == 0) 3919 zio->io_error = zio_handle_label_injection(zio, EIO); 3920 3921 if (zio->io_error && zio->io_type != ZIO_TYPE_TRIM) { 3922 if (!vdev_accessible(vd, zio)) { 3923 zio->io_error = SET_ERROR(ENXIO); 3924 } else { 3925 unexpected_error = B_TRUE; 3926 } 3927 } 3928 } 3929 3930 ops->vdev_op_io_done(zio); 3931 3932 if (unexpected_error && vd->vdev_remove_wanted == B_FALSE) 3933 VERIFY(vdev_probe(vd, zio) == NULL); 3934 3935 return (zio); 3936 } 3937 3938 /* 3939 * This function is used to change the priority of an existing zio that is 3940 * currently in-flight. This is used by the arc to upgrade priority in the 3941 * event that a demand read is made for a block that is currently queued 3942 * as a scrub or async read IO. Otherwise, the high priority read request 3943 * would end up having to wait for the lower priority IO. 3944 */ 3945 void 3946 zio_change_priority(zio_t *pio, zio_priority_t priority) 3947 { 3948 zio_t *cio, *cio_next; 3949 zio_link_t *zl = NULL; 3950 3951 ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 3952 3953 if (pio->io_vd != NULL && pio->io_vd->vdev_ops->vdev_op_leaf) { 3954 vdev_queue_change_io_priority(pio, priority); 3955 } else { 3956 pio->io_priority = priority; 3957 } 3958 3959 mutex_enter(&pio->io_lock); 3960 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) { 3961 cio_next = zio_walk_children(pio, &zl); 3962 zio_change_priority(cio, priority); 3963 } 3964 mutex_exit(&pio->io_lock); 3965 } 3966 3967 /* 3968 * For non-raidz ZIOs, we can just copy aside the bad data read from the 3969 * disk, and use that to finish the checksum ereport later. 3970 */ 3971 static void 3972 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr, 3973 const abd_t *good_buf) 3974 { 3975 /* no processing needed */ 3976 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE); 3977 } 3978 3979 void 3980 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr) 3981 { 3982 void *abd = abd_alloc_sametype(zio->io_abd, zio->io_size); 3983 3984 abd_copy(abd, zio->io_abd, zio->io_size); 3985 3986 zcr->zcr_cbinfo = zio->io_size; 3987 zcr->zcr_cbdata = abd; 3988 zcr->zcr_finish = zio_vsd_default_cksum_finish; 3989 zcr->zcr_free = zio_abd_free; 3990 } 3991 3992 static zio_t * 3993 zio_vdev_io_assess(zio_t *zio) 3994 { 3995 vdev_t *vd = zio->io_vd; 3996 3997 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) { 3998 return (NULL); 3999 } 4000 4001 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) 4002 spa_config_exit(zio->io_spa, SCL_ZIO, zio); 4003 4004 if (zio->io_vsd != NULL) { 4005 zio->io_vsd_ops->vsd_free(zio); 4006 zio->io_vsd = NULL; 4007 } 4008 4009 if (zio_injection_enabled && zio->io_error == 0) 4010 zio->io_error = zio_handle_fault_injection(zio, EIO); 4011 4012 /* 4013 * If the I/O failed, determine whether we should attempt to retry it. 4014 * 4015 * On retry, we cut in line in the issue queue, since we don't want 4016 * compression/checksumming/etc. work to prevent our (cheap) IO reissue. 4017 */ 4018 if (zio->io_error && vd == NULL && 4019 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) { 4020 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */ 4021 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */ 4022 zio->io_error = 0; 4023 zio->io_flags |= ZIO_FLAG_IO_RETRY | 4024 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE; 4025 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1; 4026 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, 4027 zio_requeue_io_start_cut_in_line); 4028 return (NULL); 4029 } 4030 4031 /* 4032 * If we got an error on a leaf device, convert it to ENXIO 4033 * if the device is not accessible at all. 4034 */ 4035 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf && 4036 !vdev_accessible(vd, zio)) 4037 zio->io_error = SET_ERROR(ENXIO); 4038 4039 /* 4040 * If we can't write to an interior vdev (mirror or RAID-Z), 4041 * set vdev_cant_write so that we stop trying to allocate from it. 4042 */ 4043 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE && 4044 vd != NULL && !vd->vdev_ops->vdev_op_leaf) { 4045 vdev_dbgmsg(vd, "zio_vdev_io_assess(zio=%px) setting " 4046 "cant_write=TRUE due to write failure with ENXIO", 4047 zio); 4048 vd->vdev_cant_write = B_TRUE; 4049 } 4050 4051 /* 4052 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future 4053 * attempts will ever succeed. In this case we set a persistent 4054 * boolean flag so that we don't bother with it in the future. 4055 */ 4056 if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) && 4057 zio->io_type == ZIO_TYPE_IOCTL && 4058 zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL) 4059 vd->vdev_nowritecache = B_TRUE; 4060 4061 if (zio->io_error) 4062 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 4063 4064 if (vd != NULL && vd->vdev_ops->vdev_op_leaf && 4065 zio->io_physdone != NULL) { 4066 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED)); 4067 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV); 4068 zio->io_physdone(zio->io_logical); 4069 } 4070 4071 return (zio); 4072 } 4073 4074 void 4075 zio_vdev_io_reissue(zio_t *zio) 4076 { 4077 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); 4078 ASSERT(zio->io_error == 0); 4079 4080 zio->io_stage >>= 1; 4081 } 4082 4083 void 4084 zio_vdev_io_redone(zio_t *zio) 4085 { 4086 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE); 4087 4088 zio->io_stage >>= 1; 4089 } 4090 4091 void 4092 zio_vdev_io_bypass(zio_t *zio) 4093 { 4094 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); 4095 ASSERT(zio->io_error == 0); 4096 4097 zio->io_flags |= ZIO_FLAG_IO_BYPASS; 4098 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1; 4099 } 4100 4101 /* 4102 * ========================================================================== 4103 * Encrypt and store encryption parameters 4104 * ========================================================================== 4105 */ 4106 4107 4108 /* 4109 * This function is used for ZIO_STAGE_ENCRYPT. It is responsible for 4110 * managing the storage of encryption parameters and passing them to the 4111 * lower-level encryption functions. 4112 */ 4113 static zio_t * 4114 zio_encrypt(zio_t *zio) 4115 { 4116 zio_prop_t *zp = &zio->io_prop; 4117 spa_t *spa = zio->io_spa; 4118 blkptr_t *bp = zio->io_bp; 4119 uint64_t psize = BP_GET_PSIZE(bp); 4120 uint64_t dsobj = zio->io_bookmark.zb_objset; 4121 dmu_object_type_t ot = BP_GET_TYPE(bp); 4122 void *enc_buf = NULL; 4123 abd_t *eabd = NULL; 4124 uint8_t salt[ZIO_DATA_SALT_LEN]; 4125 uint8_t iv[ZIO_DATA_IV_LEN]; 4126 uint8_t mac[ZIO_DATA_MAC_LEN]; 4127 boolean_t no_crypt = B_FALSE; 4128 4129 /* the root zio already encrypted the data */ 4130 if (zio->io_child_type == ZIO_CHILD_GANG) 4131 return (zio); 4132 4133 /* only ZIL blocks are re-encrypted on rewrite */ 4134 if (!IO_IS_ALLOCATING(zio) && ot != DMU_OT_INTENT_LOG) 4135 return (zio); 4136 4137 if (!(zp->zp_encrypt || BP_IS_ENCRYPTED(bp))) { 4138 BP_SET_CRYPT(bp, B_FALSE); 4139 return (zio); 4140 } 4141 4142 /* if we are doing raw encryption set the provided encryption params */ 4143 if (zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) { 4144 ASSERT0(BP_GET_LEVEL(bp)); 4145 BP_SET_CRYPT(bp, B_TRUE); 4146 BP_SET_BYTEORDER(bp, zp->zp_byteorder); 4147 if (ot != DMU_OT_OBJSET) 4148 zio_crypt_encode_mac_bp(bp, zp->zp_mac); 4149 4150 /* dnode blocks must be written out in the provided byteorder */ 4151 if (zp->zp_byteorder != ZFS_HOST_BYTEORDER && 4152 ot == DMU_OT_DNODE) { 4153 void *bswap_buf = zio_buf_alloc(psize); 4154 abd_t *babd = abd_get_from_buf(bswap_buf, psize); 4155 4156 ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF); 4157 abd_copy_to_buf(bswap_buf, zio->io_abd, psize); 4158 dmu_ot_byteswap[DMU_OT_BYTESWAP(ot)].ob_func(bswap_buf, 4159 psize); 4160 4161 abd_take_ownership_of_buf(babd, B_TRUE); 4162 zio_push_transform(zio, babd, psize, psize, NULL); 4163 } 4164 4165 if (DMU_OT_IS_ENCRYPTED(ot)) 4166 zio_crypt_encode_params_bp(bp, zp->zp_salt, zp->zp_iv); 4167 return (zio); 4168 } 4169 4170 /* indirect blocks only maintain a cksum of the lower level MACs */ 4171 if (BP_GET_LEVEL(bp) > 0) { 4172 BP_SET_CRYPT(bp, B_TRUE); 4173 VERIFY0(zio_crypt_do_indirect_mac_checksum_abd(B_TRUE, 4174 zio->io_orig_abd, BP_GET_LSIZE(bp), BP_SHOULD_BYTESWAP(bp), 4175 mac)); 4176 zio_crypt_encode_mac_bp(bp, mac); 4177 return (zio); 4178 } 4179 4180 /* 4181 * Objset blocks are a special case since they have 2 256-bit MACs 4182 * embedded within them. 4183 */ 4184 if (ot == DMU_OT_OBJSET) { 4185 ASSERT0(DMU_OT_IS_ENCRYPTED(ot)); 4186 ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF); 4187 BP_SET_CRYPT(bp, B_TRUE); 4188 VERIFY0(spa_do_crypt_objset_mac_abd(B_TRUE, spa, dsobj, 4189 zio->io_abd, psize, BP_SHOULD_BYTESWAP(bp))); 4190 return (zio); 4191 } 4192 4193 /* unencrypted object types are only authenticated with a MAC */ 4194 if (!DMU_OT_IS_ENCRYPTED(ot)) { 4195 BP_SET_CRYPT(bp, B_TRUE); 4196 VERIFY0(spa_do_crypt_mac_abd(B_TRUE, spa, dsobj, 4197 zio->io_abd, psize, mac)); 4198 zio_crypt_encode_mac_bp(bp, mac); 4199 return (zio); 4200 } 4201 4202 /* 4203 * Later passes of sync-to-convergence may decide to rewrite data 4204 * in place to avoid more disk reallocations. This presents a problem 4205 * for encryption because this constitutes rewriting the new data with 4206 * the same encryption key and IV. However, this only applies to blocks 4207 * in the MOS (particularly the spacemaps) and we do not encrypt the 4208 * MOS. We assert that the zio is allocating or an intent log write 4209 * to enforce this. 4210 */ 4211 ASSERT(IO_IS_ALLOCATING(zio) || ot == DMU_OT_INTENT_LOG); 4212 ASSERT(BP_GET_LEVEL(bp) == 0 || ot == DMU_OT_INTENT_LOG); 4213 ASSERT(spa_feature_is_active(spa, SPA_FEATURE_ENCRYPTION)); 4214 ASSERT3U(psize, !=, 0); 4215 4216 enc_buf = zio_buf_alloc(psize); 4217 eabd = abd_get_from_buf(enc_buf, psize); 4218 abd_take_ownership_of_buf(eabd, B_TRUE); 4219 4220 /* 4221 * For an explanation of what encryption parameters are stored 4222 * where, see the block comment in zio_crypt.c. 4223 */ 4224 if (ot == DMU_OT_INTENT_LOG) { 4225 zio_crypt_decode_params_bp(bp, salt, iv); 4226 } else { 4227 BP_SET_CRYPT(bp, B_TRUE); 4228 } 4229 4230 /* Perform the encryption. This should not fail */ 4231 VERIFY0(spa_do_crypt_abd(B_TRUE, spa, &zio->io_bookmark, 4232 BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp), 4233 salt, iv, mac, psize, zio->io_abd, eabd, &no_crypt)); 4234 4235 /* encode encryption metadata into the bp */ 4236 if (ot == DMU_OT_INTENT_LOG) { 4237 /* 4238 * ZIL blocks store the MAC in the embedded checksum, so the 4239 * transform must always be applied. 4240 */ 4241 zio_crypt_encode_mac_zil(enc_buf, mac); 4242 zio_push_transform(zio, eabd, psize, psize, NULL); 4243 } else { 4244 BP_SET_CRYPT(bp, B_TRUE); 4245 zio_crypt_encode_params_bp(bp, salt, iv); 4246 zio_crypt_encode_mac_bp(bp, mac); 4247 4248 if (no_crypt) { 4249 ASSERT3U(ot, ==, DMU_OT_DNODE); 4250 abd_free(eabd); 4251 } else { 4252 zio_push_transform(zio, eabd, psize, psize, NULL); 4253 } 4254 } 4255 4256 return (zio); 4257 } 4258 4259 /* 4260 * ========================================================================== 4261 * Generate and verify checksums 4262 * ========================================================================== 4263 */ 4264 static zio_t * 4265 zio_checksum_generate(zio_t *zio) 4266 { 4267 blkptr_t *bp = zio->io_bp; 4268 enum zio_checksum checksum; 4269 4270 if (bp == NULL) { 4271 /* 4272 * This is zio_write_phys(). 4273 * We're either generating a label checksum, or none at all. 4274 */ 4275 checksum = zio->io_prop.zp_checksum; 4276 4277 if (checksum == ZIO_CHECKSUM_OFF) 4278 return (zio); 4279 4280 ASSERT(checksum == ZIO_CHECKSUM_LABEL); 4281 } else { 4282 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) { 4283 ASSERT(!IO_IS_ALLOCATING(zio)); 4284 checksum = ZIO_CHECKSUM_GANG_HEADER; 4285 } else { 4286 checksum = BP_GET_CHECKSUM(bp); 4287 } 4288 } 4289 4290 zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size); 4291 4292 return (zio); 4293 } 4294 4295 static zio_t * 4296 zio_checksum_verify(zio_t *zio) 4297 { 4298 zio_bad_cksum_t info; 4299 blkptr_t *bp = zio->io_bp; 4300 int error; 4301 4302 ASSERT(zio->io_vd != NULL); 4303 4304 if (bp == NULL) { 4305 /* 4306 * This is zio_read_phys(). 4307 * We're either verifying a label checksum, or nothing at all. 4308 */ 4309 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF) 4310 return (zio); 4311 4312 ASSERT3U(zio->io_prop.zp_checksum, ==, ZIO_CHECKSUM_LABEL); 4313 } 4314 4315 if ((error = zio_checksum_error(zio, &info)) != 0) { 4316 zio->io_error = error; 4317 if (error == ECKSUM && 4318 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) { 4319 mutex_enter(&zio->io_vd->vdev_stat_lock); 4320 zio->io_vd->vdev_stat.vs_checksum_errors++; 4321 mutex_exit(&zio->io_vd->vdev_stat_lock); 4322 (void) zfs_ereport_start_checksum(zio->io_spa, 4323 zio->io_vd, &zio->io_bookmark, zio, 4324 zio->io_offset, zio->io_size, &info); 4325 } 4326 } 4327 4328 return (zio); 4329 } 4330 4331 /* 4332 * Called by RAID-Z to ensure we don't compute the checksum twice. 4333 */ 4334 void 4335 zio_checksum_verified(zio_t *zio) 4336 { 4337 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; 4338 } 4339 4340 /* 4341 * ========================================================================== 4342 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other. 4343 * An error of 0 indicates success. ENXIO indicates whole-device failure, 4344 * which may be transient (e.g. unplugged) or permanent. ECKSUM and EIO 4345 * indicate errors that are specific to one I/O, and most likely permanent. 4346 * Any other error is presumed to be worse because we weren't expecting it. 4347 * ========================================================================== 4348 */ 4349 int 4350 zio_worst_error(int e1, int e2) 4351 { 4352 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO }; 4353 int r1, r2; 4354 4355 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++) 4356 if (e1 == zio_error_rank[r1]) 4357 break; 4358 4359 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++) 4360 if (e2 == zio_error_rank[r2]) 4361 break; 4362 4363 return (r1 > r2 ? e1 : e2); 4364 } 4365 4366 /* 4367 * ========================================================================== 4368 * I/O completion 4369 * ========================================================================== 4370 */ 4371 static zio_t * 4372 zio_ready(zio_t *zio) 4373 { 4374 blkptr_t *bp = zio->io_bp; 4375 zio_t *pio, *pio_next; 4376 zio_link_t *zl = NULL; 4377 4378 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT | ZIO_CHILD_DDT_BIT, 4379 ZIO_WAIT_READY)) { 4380 return (NULL); 4381 } 4382 4383 if (zio->io_ready) { 4384 ASSERT(IO_IS_ALLOCATING(zio)); 4385 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) || 4386 (zio->io_flags & ZIO_FLAG_NOPWRITE)); 4387 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0); 4388 4389 zio->io_ready(zio); 4390 } 4391 4392 if (bp != NULL && bp != &zio->io_bp_copy) 4393 zio->io_bp_copy = *bp; 4394 4395 if (zio->io_error != 0) { 4396 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 4397 4398 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) { 4399 ASSERT(IO_IS_ALLOCATING(zio)); 4400 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); 4401 ASSERT(zio->io_metaslab_class != NULL); 4402 4403 /* 4404 * We were unable to allocate anything, unreserve and 4405 * issue the next I/O to allocate. 4406 */ 4407 metaslab_class_throttle_unreserve( 4408 zio->io_metaslab_class, zio->io_prop.zp_copies, 4409 zio->io_allocator, zio); 4410 zio_allocate_dispatch(zio->io_spa, zio->io_allocator); 4411 } 4412 } 4413 4414 mutex_enter(&zio->io_lock); 4415 zio->io_state[ZIO_WAIT_READY] = 1; 4416 pio = zio_walk_parents(zio, &zl); 4417 mutex_exit(&zio->io_lock); 4418 4419 /* 4420 * As we notify zio's parents, new parents could be added. 4421 * New parents go to the head of zio's io_parent_list, however, 4422 * so we will (correctly) not notify them. The remainder of zio's 4423 * io_parent_list, from 'pio_next' onward, cannot change because 4424 * all parents must wait for us to be done before they can be done. 4425 */ 4426 for (; pio != NULL; pio = pio_next) { 4427 pio_next = zio_walk_parents(zio, &zl); 4428 zio_notify_parent(pio, zio, ZIO_WAIT_READY, NULL); 4429 } 4430 4431 if (zio->io_flags & ZIO_FLAG_NODATA) { 4432 if (BP_IS_GANG(bp)) { 4433 zio->io_flags &= ~ZIO_FLAG_NODATA; 4434 } else { 4435 ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE); 4436 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; 4437 } 4438 } 4439 4440 if (zio_injection_enabled && 4441 zio->io_spa->spa_syncing_txg == zio->io_txg) 4442 zio_handle_ignored_writes(zio); 4443 4444 return (zio); 4445 } 4446 4447 /* 4448 * Update the allocation throttle accounting. 4449 */ 4450 static void 4451 zio_dva_throttle_done(zio_t *zio) 4452 { 4453 zio_t *lio __maybe_unused = zio->io_logical; 4454 zio_t *pio = zio_unique_parent(zio); 4455 vdev_t *vd = zio->io_vd; 4456 int flags = METASLAB_ASYNC_ALLOC; 4457 4458 ASSERT3P(zio->io_bp, !=, NULL); 4459 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE); 4460 ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE); 4461 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV); 4462 ASSERT(vd != NULL); 4463 ASSERT3P(vd, ==, vd->vdev_top); 4464 ASSERT(zio_injection_enabled || !(zio->io_flags & ZIO_FLAG_IO_RETRY)); 4465 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR)); 4466 ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING); 4467 ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE)); 4468 ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA)); 4469 4470 /* 4471 * Parents of gang children can have two flavors -- ones that 4472 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set) 4473 * and ones that allocated the constituent blocks. The allocation 4474 * throttle needs to know the allocating parent zio so we must find 4475 * it here. 4476 */ 4477 if (pio->io_child_type == ZIO_CHILD_GANG) { 4478 /* 4479 * If our parent is a rewrite gang child then our grandparent 4480 * would have been the one that performed the allocation. 4481 */ 4482 if (pio->io_flags & ZIO_FLAG_IO_REWRITE) 4483 pio = zio_unique_parent(pio); 4484 flags |= METASLAB_GANG_CHILD; 4485 } 4486 4487 ASSERT(IO_IS_ALLOCATING(pio)); 4488 ASSERT3P(zio, !=, zio->io_logical); 4489 ASSERT(zio->io_logical != NULL); 4490 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR)); 4491 ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE); 4492 ASSERT(zio->io_metaslab_class != NULL); 4493 4494 mutex_enter(&pio->io_lock); 4495 metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags, 4496 pio->io_allocator, B_TRUE); 4497 mutex_exit(&pio->io_lock); 4498 4499 metaslab_class_throttle_unreserve(zio->io_metaslab_class, 1, 4500 pio->io_allocator, pio); 4501 4502 /* 4503 * Call into the pipeline to see if there is more work that 4504 * needs to be done. If there is work to be done it will be 4505 * dispatched to another taskq thread. 4506 */ 4507 zio_allocate_dispatch(zio->io_spa, pio->io_allocator); 4508 } 4509 4510 static zio_t * 4511 zio_done(zio_t *zio) 4512 { 4513 /* 4514 * Always attempt to keep stack usage minimal here since 4515 * we can be called recursively up to 19 levels deep. 4516 */ 4517 const uint64_t psize = zio->io_size; 4518 zio_t *pio, *pio_next; 4519 zio_link_t *zl = NULL; 4520 4521 /* 4522 * If our children haven't all completed, 4523 * wait for them and then repeat this pipeline stage. 4524 */ 4525 if (zio_wait_for_children(zio, ZIO_CHILD_ALL_BITS, ZIO_WAIT_DONE)) { 4526 return (NULL); 4527 } 4528 4529 /* 4530 * If the allocation throttle is enabled, then update the accounting. 4531 * We only track child I/Os that are part of an allocating async 4532 * write. We must do this since the allocation is performed 4533 * by the logical I/O but the actual write is done by child I/Os. 4534 */ 4535 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING && 4536 zio->io_child_type == ZIO_CHILD_VDEV) { 4537 ASSERT(zio->io_metaslab_class != NULL); 4538 ASSERT(zio->io_metaslab_class->mc_alloc_throttle_enabled); 4539 zio_dva_throttle_done(zio); 4540 } 4541 4542 /* 4543 * If the allocation throttle is enabled, verify that 4544 * we have decremented the refcounts for every I/O that was throttled. 4545 */ 4546 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) { 4547 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 4548 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); 4549 ASSERT(zio->io_bp != NULL); 4550 4551 metaslab_group_alloc_verify(zio->io_spa, zio->io_bp, zio, 4552 zio->io_allocator); 4553 VERIFY(zfs_refcount_not_held(&zio->io_metaslab_class-> 4554 mc_allocator[zio->io_allocator].mca_alloc_slots, zio)); 4555 } 4556 4557 4558 for (int c = 0; c < ZIO_CHILD_TYPES; c++) 4559 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 4560 ASSERT(zio->io_children[c][w] == 0); 4561 4562 if (zio->io_bp != NULL && !BP_IS_EMBEDDED(zio->io_bp)) { 4563 ASSERT(zio->io_bp->blk_pad[0] == 0); 4564 ASSERT(zio->io_bp->blk_pad[1] == 0); 4565 ASSERT(memcmp(zio->io_bp, &zio->io_bp_copy, 4566 sizeof (blkptr_t)) == 0 || 4567 (zio->io_bp == zio_unique_parent(zio)->io_bp)); 4568 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) && 4569 zio->io_bp_override == NULL && 4570 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) { 4571 ASSERT3U(zio->io_prop.zp_copies, <=, 4572 BP_GET_NDVAS(zio->io_bp)); 4573 ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 || 4574 (BP_COUNT_GANG(zio->io_bp) == 4575 BP_GET_NDVAS(zio->io_bp))); 4576 } 4577 if (zio->io_flags & ZIO_FLAG_NOPWRITE) 4578 VERIFY(BP_EQUAL(zio->io_bp, &zio->io_bp_orig)); 4579 } 4580 4581 /* 4582 * If there were child vdev/gang/ddt errors, they apply to us now. 4583 */ 4584 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV); 4585 zio_inherit_child_errors(zio, ZIO_CHILD_GANG); 4586 zio_inherit_child_errors(zio, ZIO_CHILD_DDT); 4587 4588 /* 4589 * If the I/O on the transformed data was successful, generate any 4590 * checksum reports now while we still have the transformed data. 4591 */ 4592 if (zio->io_error == 0) { 4593 while (zio->io_cksum_report != NULL) { 4594 zio_cksum_report_t *zcr = zio->io_cksum_report; 4595 uint64_t align = zcr->zcr_align; 4596 uint64_t asize = P2ROUNDUP(psize, align); 4597 abd_t *adata = zio->io_abd; 4598 4599 if (adata != NULL && asize != psize) { 4600 adata = abd_alloc(asize, B_TRUE); 4601 abd_copy(adata, zio->io_abd, psize); 4602 abd_zero_off(adata, psize, asize - psize); 4603 } 4604 4605 zio->io_cksum_report = zcr->zcr_next; 4606 zcr->zcr_next = NULL; 4607 zcr->zcr_finish(zcr, adata); 4608 zfs_ereport_free_checksum(zcr); 4609 4610 if (adata != NULL && asize != psize) 4611 abd_free(adata); 4612 } 4613 } 4614 4615 zio_pop_transforms(zio); /* note: may set zio->io_error */ 4616 4617 vdev_stat_update(zio, psize); 4618 4619 /* 4620 * If this I/O is attached to a particular vdev is slow, exceeding 4621 * 30 seconds to complete, post an error described the I/O delay. 4622 * We ignore these errors if the device is currently unavailable. 4623 */ 4624 if (zio->io_delay >= MSEC2NSEC(zio_slow_io_ms)) { 4625 if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd)) { 4626 /* 4627 * We want to only increment our slow IO counters if 4628 * the IO is valid (i.e. not if the drive is removed). 4629 * 4630 * zfs_ereport_post() will also do these checks, but 4631 * it can also ratelimit and have other failures, so we 4632 * need to increment the slow_io counters independent 4633 * of it. 4634 */ 4635 if (zfs_ereport_is_valid(FM_EREPORT_ZFS_DELAY, 4636 zio->io_spa, zio->io_vd, zio)) { 4637 mutex_enter(&zio->io_vd->vdev_stat_lock); 4638 zio->io_vd->vdev_stat.vs_slow_ios++; 4639 mutex_exit(&zio->io_vd->vdev_stat_lock); 4640 4641 (void) zfs_ereport_post(FM_EREPORT_ZFS_DELAY, 4642 zio->io_spa, zio->io_vd, &zio->io_bookmark, 4643 zio, 0); 4644 } 4645 } 4646 } 4647 4648 if (zio->io_error) { 4649 /* 4650 * If this I/O is attached to a particular vdev, 4651 * generate an error message describing the I/O failure 4652 * at the block level. We ignore these errors if the 4653 * device is currently unavailable. 4654 */ 4655 if (zio->io_error != ECKSUM && zio->io_vd != NULL && 4656 !vdev_is_dead(zio->io_vd)) { 4657 int ret = zfs_ereport_post(FM_EREPORT_ZFS_IO, 4658 zio->io_spa, zio->io_vd, &zio->io_bookmark, zio, 0); 4659 if (ret != EALREADY) { 4660 mutex_enter(&zio->io_vd->vdev_stat_lock); 4661 if (zio->io_type == ZIO_TYPE_READ) 4662 zio->io_vd->vdev_stat.vs_read_errors++; 4663 else if (zio->io_type == ZIO_TYPE_WRITE) 4664 zio->io_vd->vdev_stat.vs_write_errors++; 4665 mutex_exit(&zio->io_vd->vdev_stat_lock); 4666 } 4667 } 4668 4669 if ((zio->io_error == EIO || !(zio->io_flags & 4670 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) && 4671 zio == zio->io_logical) { 4672 /* 4673 * For logical I/O requests, tell the SPA to log the 4674 * error and generate a logical data ereport. 4675 */ 4676 spa_log_error(zio->io_spa, &zio->io_bookmark); 4677 (void) zfs_ereport_post(FM_EREPORT_ZFS_DATA, 4678 zio->io_spa, NULL, &zio->io_bookmark, zio, 0); 4679 } 4680 } 4681 4682 if (zio->io_error && zio == zio->io_logical) { 4683 /* 4684 * Determine whether zio should be reexecuted. This will 4685 * propagate all the way to the root via zio_notify_parent(). 4686 */ 4687 ASSERT(zio->io_vd == NULL && zio->io_bp != NULL); 4688 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 4689 4690 if (IO_IS_ALLOCATING(zio) && 4691 !(zio->io_flags & ZIO_FLAG_CANFAIL)) { 4692 if (zio->io_error != ENOSPC) 4693 zio->io_reexecute |= ZIO_REEXECUTE_NOW; 4694 else 4695 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 4696 } 4697 4698 if ((zio->io_type == ZIO_TYPE_READ || 4699 zio->io_type == ZIO_TYPE_FREE) && 4700 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && 4701 zio->io_error == ENXIO && 4702 spa_load_state(zio->io_spa) == SPA_LOAD_NONE && 4703 spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE) 4704 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 4705 4706 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute) 4707 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 4708 4709 /* 4710 * Here is a possibly good place to attempt to do 4711 * either combinatorial reconstruction or error correction 4712 * based on checksums. It also might be a good place 4713 * to send out preliminary ereports before we suspend 4714 * processing. 4715 */ 4716 } 4717 4718 /* 4719 * If there were logical child errors, they apply to us now. 4720 * We defer this until now to avoid conflating logical child 4721 * errors with errors that happened to the zio itself when 4722 * updating vdev stats and reporting FMA events above. 4723 */ 4724 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL); 4725 4726 if ((zio->io_error || zio->io_reexecute) && 4727 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio && 4728 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE))) 4729 zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp); 4730 4731 zio_gang_tree_free(&zio->io_gang_tree); 4732 4733 /* 4734 * Godfather I/Os should never suspend. 4735 */ 4736 if ((zio->io_flags & ZIO_FLAG_GODFATHER) && 4737 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) 4738 zio->io_reexecute &= ~ZIO_REEXECUTE_SUSPEND; 4739 4740 if (zio->io_reexecute) { 4741 /* 4742 * This is a logical I/O that wants to reexecute. 4743 * 4744 * Reexecute is top-down. When an i/o fails, if it's not 4745 * the root, it simply notifies its parent and sticks around. 4746 * The parent, seeing that it still has children in zio_done(), 4747 * does the same. This percolates all the way up to the root. 4748 * The root i/o will reexecute or suspend the entire tree. 4749 * 4750 * This approach ensures that zio_reexecute() honors 4751 * all the original i/o dependency relationships, e.g. 4752 * parents not executing until children are ready. 4753 */ 4754 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 4755 4756 zio->io_gang_leader = NULL; 4757 4758 mutex_enter(&zio->io_lock); 4759 zio->io_state[ZIO_WAIT_DONE] = 1; 4760 mutex_exit(&zio->io_lock); 4761 4762 /* 4763 * "The Godfather" I/O monitors its children but is 4764 * not a true parent to them. It will track them through 4765 * the pipeline but severs its ties whenever they get into 4766 * trouble (e.g. suspended). This allows "The Godfather" 4767 * I/O to return status without blocking. 4768 */ 4769 zl = NULL; 4770 for (pio = zio_walk_parents(zio, &zl); pio != NULL; 4771 pio = pio_next) { 4772 zio_link_t *remove_zl = zl; 4773 pio_next = zio_walk_parents(zio, &zl); 4774 4775 if ((pio->io_flags & ZIO_FLAG_GODFATHER) && 4776 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) { 4777 zio_remove_child(pio, zio, remove_zl); 4778 /* 4779 * This is a rare code path, so we don't 4780 * bother with "next_to_execute". 4781 */ 4782 zio_notify_parent(pio, zio, ZIO_WAIT_DONE, 4783 NULL); 4784 } 4785 } 4786 4787 if ((pio = zio_unique_parent(zio)) != NULL) { 4788 /* 4789 * We're not a root i/o, so there's nothing to do 4790 * but notify our parent. Don't propagate errors 4791 * upward since we haven't permanently failed yet. 4792 */ 4793 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); 4794 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE; 4795 /* 4796 * This is a rare code path, so we don't bother with 4797 * "next_to_execute". 4798 */ 4799 zio_notify_parent(pio, zio, ZIO_WAIT_DONE, NULL); 4800 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) { 4801 /* 4802 * We'd fail again if we reexecuted now, so suspend 4803 * until conditions improve (e.g. device comes online). 4804 */ 4805 zio_suspend(zio->io_spa, zio, ZIO_SUSPEND_IOERR); 4806 } else { 4807 /* 4808 * Reexecution is potentially a huge amount of work. 4809 * Hand it off to the otherwise-unused claim taskq. 4810 */ 4811 ASSERT(taskq_empty_ent(&zio->io_tqent)); 4812 spa_taskq_dispatch_ent(zio->io_spa, 4813 ZIO_TYPE_CLAIM, ZIO_TASKQ_ISSUE, 4814 zio_reexecute, zio, 0, &zio->io_tqent); 4815 } 4816 return (NULL); 4817 } 4818 4819 ASSERT(zio->io_child_count == 0); 4820 ASSERT(zio->io_reexecute == 0); 4821 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL)); 4822 4823 /* 4824 * Report any checksum errors, since the I/O is complete. 4825 */ 4826 while (zio->io_cksum_report != NULL) { 4827 zio_cksum_report_t *zcr = zio->io_cksum_report; 4828 zio->io_cksum_report = zcr->zcr_next; 4829 zcr->zcr_next = NULL; 4830 zcr->zcr_finish(zcr, NULL); 4831 zfs_ereport_free_checksum(zcr); 4832 } 4833 4834 if (zio->io_flags & ZIO_FLAG_FASTWRITE && zio->io_bp && 4835 !BP_IS_HOLE(zio->io_bp) && !BP_IS_EMBEDDED(zio->io_bp) && 4836 !(zio->io_flags & ZIO_FLAG_NOPWRITE)) { 4837 metaslab_fastwrite_unmark(zio->io_spa, zio->io_bp); 4838 } 4839 4840 /* 4841 * It is the responsibility of the done callback to ensure that this 4842 * particular zio is no longer discoverable for adoption, and as 4843 * such, cannot acquire any new parents. 4844 */ 4845 if (zio->io_done) 4846 zio->io_done(zio); 4847 4848 mutex_enter(&zio->io_lock); 4849 zio->io_state[ZIO_WAIT_DONE] = 1; 4850 mutex_exit(&zio->io_lock); 4851 4852 /* 4853 * We are done executing this zio. We may want to execute a parent 4854 * next. See the comment in zio_notify_parent(). 4855 */ 4856 zio_t *next_to_execute = NULL; 4857 zl = NULL; 4858 for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) { 4859 zio_link_t *remove_zl = zl; 4860 pio_next = zio_walk_parents(zio, &zl); 4861 zio_remove_child(pio, zio, remove_zl); 4862 zio_notify_parent(pio, zio, ZIO_WAIT_DONE, &next_to_execute); 4863 } 4864 4865 if (zio->io_waiter != NULL) { 4866 mutex_enter(&zio->io_lock); 4867 zio->io_executor = NULL; 4868 cv_broadcast(&zio->io_cv); 4869 mutex_exit(&zio->io_lock); 4870 } else { 4871 zio_destroy(zio); 4872 } 4873 4874 return (next_to_execute); 4875 } 4876 4877 /* 4878 * ========================================================================== 4879 * I/O pipeline definition 4880 * ========================================================================== 4881 */ 4882 static zio_pipe_stage_t *zio_pipeline[] = { 4883 NULL, 4884 zio_read_bp_init, 4885 zio_write_bp_init, 4886 zio_free_bp_init, 4887 zio_issue_async, 4888 zio_write_compress, 4889 zio_encrypt, 4890 zio_checksum_generate, 4891 zio_nop_write, 4892 zio_ddt_read_start, 4893 zio_ddt_read_done, 4894 zio_ddt_write, 4895 zio_ddt_free, 4896 zio_gang_assemble, 4897 zio_gang_issue, 4898 zio_dva_throttle, 4899 zio_dva_allocate, 4900 zio_dva_free, 4901 zio_dva_claim, 4902 zio_ready, 4903 zio_vdev_io_start, 4904 zio_vdev_io_done, 4905 zio_vdev_io_assess, 4906 zio_checksum_verify, 4907 zio_done 4908 }; 4909 4910 4911 4912 4913 /* 4914 * Compare two zbookmark_phys_t's to see which we would reach first in a 4915 * pre-order traversal of the object tree. 4916 * 4917 * This is simple in every case aside from the meta-dnode object. For all other 4918 * objects, we traverse them in order (object 1 before object 2, and so on). 4919 * However, all of these objects are traversed while traversing object 0, since 4920 * the data it points to is the list of objects. Thus, we need to convert to a 4921 * canonical representation so we can compare meta-dnode bookmarks to 4922 * non-meta-dnode bookmarks. 4923 * 4924 * We do this by calculating "equivalents" for each field of the zbookmark. 4925 * zbookmarks outside of the meta-dnode use their own object and level, and 4926 * calculate the level 0 equivalent (the first L0 blkid that is contained in the 4927 * blocks this bookmark refers to) by multiplying their blkid by their span 4928 * (the number of L0 blocks contained within one block at their level). 4929 * zbookmarks inside the meta-dnode calculate their object equivalent 4930 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use 4931 * level + 1<<31 (any value larger than a level could ever be) for their level. 4932 * This causes them to always compare before a bookmark in their object 4933 * equivalent, compare appropriately to bookmarks in other objects, and to 4934 * compare appropriately to other bookmarks in the meta-dnode. 4935 */ 4936 int 4937 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2, 4938 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2) 4939 { 4940 /* 4941 * These variables represent the "equivalent" values for the zbookmark, 4942 * after converting zbookmarks inside the meta dnode to their 4943 * normal-object equivalents. 4944 */ 4945 uint64_t zb1obj, zb2obj; 4946 uint64_t zb1L0, zb2L0; 4947 uint64_t zb1level, zb2level; 4948 4949 if (zb1->zb_object == zb2->zb_object && 4950 zb1->zb_level == zb2->zb_level && 4951 zb1->zb_blkid == zb2->zb_blkid) 4952 return (0); 4953 4954 IMPLY(zb1->zb_level > 0, ibs1 >= SPA_MINBLOCKSHIFT); 4955 IMPLY(zb2->zb_level > 0, ibs2 >= SPA_MINBLOCKSHIFT); 4956 4957 /* 4958 * BP_SPANB calculates the span in blocks. 4959 */ 4960 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level); 4961 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level); 4962 4963 if (zb1->zb_object == DMU_META_DNODE_OBJECT) { 4964 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT)); 4965 zb1L0 = 0; 4966 zb1level = zb1->zb_level + COMPARE_META_LEVEL; 4967 } else { 4968 zb1obj = zb1->zb_object; 4969 zb1level = zb1->zb_level; 4970 } 4971 4972 if (zb2->zb_object == DMU_META_DNODE_OBJECT) { 4973 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT)); 4974 zb2L0 = 0; 4975 zb2level = zb2->zb_level + COMPARE_META_LEVEL; 4976 } else { 4977 zb2obj = zb2->zb_object; 4978 zb2level = zb2->zb_level; 4979 } 4980 4981 /* Now that we have a canonical representation, do the comparison. */ 4982 if (zb1obj != zb2obj) 4983 return (zb1obj < zb2obj ? -1 : 1); 4984 else if (zb1L0 != zb2L0) 4985 return (zb1L0 < zb2L0 ? -1 : 1); 4986 else if (zb1level != zb2level) 4987 return (zb1level > zb2level ? -1 : 1); 4988 /* 4989 * This can (theoretically) happen if the bookmarks have the same object 4990 * and level, but different blkids, if the block sizes are not the same. 4991 * There is presently no way to change the indirect block sizes 4992 */ 4993 return (0); 4994 } 4995 4996 /* 4997 * This function checks the following: given that last_block is the place that 4998 * our traversal stopped last time, does that guarantee that we've visited 4999 * every node under subtree_root? Therefore, we can't just use the raw output 5000 * of zbookmark_compare. We have to pass in a modified version of 5001 * subtree_root; by incrementing the block id, and then checking whether 5002 * last_block is before or equal to that, we can tell whether or not having 5003 * visited last_block implies that all of subtree_root's children have been 5004 * visited. 5005 */ 5006 boolean_t 5007 zbookmark_subtree_completed(const dnode_phys_t *dnp, 5008 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block) 5009 { 5010 zbookmark_phys_t mod_zb = *subtree_root; 5011 mod_zb.zb_blkid++; 5012 ASSERT0(last_block->zb_level); 5013 5014 /* The objset_phys_t isn't before anything. */ 5015 if (dnp == NULL) 5016 return (B_FALSE); 5017 5018 /* 5019 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the 5020 * data block size in sectors, because that variable is only used if 5021 * the bookmark refers to a block in the meta-dnode. Since we don't 5022 * know without examining it what object it refers to, and there's no 5023 * harm in passing in this value in other cases, we always pass it in. 5024 * 5025 * We pass in 0 for the indirect block size shift because zb2 must be 5026 * level 0. The indirect block size is only used to calculate the span 5027 * of the bookmark, but since the bookmark must be level 0, the span is 5028 * always 1, so the math works out. 5029 * 5030 * If you make changes to how the zbookmark_compare code works, be sure 5031 * to make sure that this code still works afterwards. 5032 */ 5033 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift, 5034 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb, 5035 last_block) <= 0); 5036 } 5037 5038 /* 5039 * This function is similar to zbookmark_subtree_completed(), but returns true 5040 * if subtree_root is equal or ahead of last_block, i.e. still to be done. 5041 */ 5042 boolean_t 5043 zbookmark_subtree_tbd(const dnode_phys_t *dnp, 5044 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block) 5045 { 5046 ASSERT0(last_block->zb_level); 5047 if (dnp == NULL) 5048 return (B_FALSE); 5049 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift, 5050 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, subtree_root, 5051 last_block) >= 0); 5052 } 5053 5054 EXPORT_SYMBOL(zio_type_name); 5055 EXPORT_SYMBOL(zio_buf_alloc); 5056 EXPORT_SYMBOL(zio_data_buf_alloc); 5057 EXPORT_SYMBOL(zio_buf_free); 5058 EXPORT_SYMBOL(zio_data_buf_free); 5059 5060 ZFS_MODULE_PARAM(zfs_zio, zio_, slow_io_ms, INT, ZMOD_RW, 5061 "Max I/O completion time (milliseconds) before marking it as slow"); 5062 5063 ZFS_MODULE_PARAM(zfs_zio, zio_, requeue_io_start_cut_in_line, INT, ZMOD_RW, 5064 "Prioritize requeued I/O"); 5065 5066 ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_deferred_free, UINT, ZMOD_RW, 5067 "Defer frees starting in this pass"); 5068 5069 ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_dont_compress, UINT, ZMOD_RW, 5070 "Don't compress starting in this pass"); 5071 5072 ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_rewrite, UINT, ZMOD_RW, 5073 "Rewrite new bps starting in this pass"); 5074 5075 ZFS_MODULE_PARAM(zfs_zio, zio_, dva_throttle_enabled, INT, ZMOD_RW, 5076 "Throttle block allocations in the ZIO pipeline"); 5077 5078 ZFS_MODULE_PARAM(zfs_zio, zio_, deadman_log_all, INT, ZMOD_RW, 5079 "Log all slow ZIOs, not just those with vdevs"); 5080