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/brt.h> 45 #include <sys/ddt.h> 46 #include <sys/blkptr.h> 47 #include <sys/zfeature.h> 48 #include <sys/dsl_scan.h> 49 #include <sys/metaslab_impl.h> 50 #include <sys/time.h> 51 #include <sys/trace_zfs.h> 52 #include <sys/abd.h> 53 #include <sys/dsl_crypt.h> 54 #include <cityhash.h> 55 56 /* 57 * ========================================================================== 58 * I/O type descriptions 59 * ========================================================================== 60 */ 61 const char *const zio_type_name[ZIO_TYPES] = { 62 /* 63 * Note: Linux kernel thread name length is limited 64 * so these names will differ from upstream open zfs. 65 */ 66 "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl", "z_trim" 67 }; 68 69 int zio_dva_throttle_enabled = B_TRUE; 70 static int zio_deadman_log_all = B_FALSE; 71 72 /* 73 * ========================================================================== 74 * I/O kmem caches 75 * ========================================================================== 76 */ 77 static kmem_cache_t *zio_cache; 78 static kmem_cache_t *zio_link_cache; 79 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; 80 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; 81 #if defined(ZFS_DEBUG) && !defined(_KERNEL) 82 static uint64_t zio_buf_cache_allocs[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; 83 static uint64_t zio_buf_cache_frees[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; 84 #endif 85 86 /* Mark IOs as "slow" if they take longer than 30 seconds */ 87 static uint_t zio_slow_io_ms = (30 * MILLISEC); 88 89 #define BP_SPANB(indblkshift, level) \ 90 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT))) 91 #define COMPARE_META_LEVEL 0x80000000ul 92 /* 93 * The following actions directly effect the spa's sync-to-convergence logic. 94 * The values below define the sync pass when we start performing the action. 95 * Care should be taken when changing these values as they directly impact 96 * spa_sync() performance. Tuning these values may introduce subtle performance 97 * pathologies and should only be done in the context of performance analysis. 98 * These tunables will eventually be removed and replaced with #defines once 99 * enough analysis has been done to determine optimal values. 100 * 101 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that 102 * regular blocks are not deferred. 103 * 104 * Starting in sync pass 8 (zfs_sync_pass_dont_compress), we disable 105 * compression (including of metadata). In practice, we don't have this 106 * many sync passes, so this has no effect. 107 * 108 * The original intent was that disabling compression would help the sync 109 * passes to converge. However, in practice disabling compression increases 110 * the average number of sync passes, because when we turn compression off, a 111 * lot of block's size will change and thus we have to re-allocate (not 112 * overwrite) them. It also increases the number of 128KB allocations (e.g. 113 * for indirect blocks and spacemaps) because these will not be compressed. 114 * The 128K allocations are especially detrimental to performance on highly 115 * fragmented systems, which may have very few free segments of this size, 116 * and may need to load new metaslabs to satisfy 128K allocations. 117 */ 118 119 /* defer frees starting in this pass */ 120 uint_t zfs_sync_pass_deferred_free = 2; 121 122 /* don't compress starting in this pass */ 123 static uint_t zfs_sync_pass_dont_compress = 8; 124 125 /* rewrite new bps starting in this pass */ 126 static uint_t zfs_sync_pass_rewrite = 2; 127 128 /* 129 * An allocating zio is one that either currently has the DVA allocate 130 * stage set or will have it later in its lifetime. 131 */ 132 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE) 133 134 /* 135 * Enable smaller cores by excluding metadata 136 * allocations as well. 137 */ 138 int zio_exclude_metadata = 0; 139 static int zio_requeue_io_start_cut_in_line = 1; 140 141 #ifdef ZFS_DEBUG 142 static const int zio_buf_debug_limit = 16384; 143 #else 144 static const int zio_buf_debug_limit = 0; 145 #endif 146 147 static inline void __zio_execute(zio_t *zio); 148 149 static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t); 150 151 void 152 zio_init(void) 153 { 154 size_t c; 155 156 zio_cache = kmem_cache_create("zio_cache", 157 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0); 158 zio_link_cache = kmem_cache_create("zio_link_cache", 159 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0); 160 161 /* 162 * For small buffers, we want a cache for each multiple of 163 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache 164 * for each quarter-power of 2. 165 */ 166 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { 167 size_t size = (c + 1) << SPA_MINBLOCKSHIFT; 168 size_t p2 = size; 169 size_t align = 0; 170 size_t data_cflags, cflags; 171 172 data_cflags = KMC_NODEBUG; 173 cflags = (zio_exclude_metadata || size > zio_buf_debug_limit) ? 174 KMC_NODEBUG : 0; 175 176 while (!ISP2(p2)) 177 p2 &= p2 - 1; 178 179 #ifndef _KERNEL 180 /* 181 * If we are using watchpoints, put each buffer on its own page, 182 * to eliminate the performance overhead of trapping to the 183 * kernel when modifying a non-watched buffer that shares the 184 * page with a watched buffer. 185 */ 186 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE)) 187 continue; 188 /* 189 * Here's the problem - on 4K native devices in userland on 190 * Linux using O_DIRECT, buffers must be 4K aligned or I/O 191 * will fail with EINVAL, causing zdb (and others) to coredump. 192 * Since userland probably doesn't need optimized buffer caches, 193 * we just force 4K alignment on everything. 194 */ 195 align = 8 * SPA_MINBLOCKSIZE; 196 #else 197 if (size < PAGESIZE) { 198 align = SPA_MINBLOCKSIZE; 199 } else if (IS_P2ALIGNED(size, p2 >> 2)) { 200 align = PAGESIZE; 201 } 202 #endif 203 204 if (align != 0) { 205 char name[36]; 206 if (cflags == data_cflags) { 207 /* 208 * Resulting kmem caches would be identical. 209 * Save memory by creating only one. 210 */ 211 (void) snprintf(name, sizeof (name), 212 "zio_buf_comb_%lu", (ulong_t)size); 213 zio_buf_cache[c] = kmem_cache_create(name, 214 size, align, NULL, NULL, NULL, NULL, NULL, 215 cflags); 216 zio_data_buf_cache[c] = zio_buf_cache[c]; 217 continue; 218 } 219 (void) snprintf(name, sizeof (name), "zio_buf_%lu", 220 (ulong_t)size); 221 zio_buf_cache[c] = kmem_cache_create(name, size, 222 align, NULL, NULL, NULL, NULL, NULL, cflags); 223 224 (void) snprintf(name, sizeof (name), "zio_data_buf_%lu", 225 (ulong_t)size); 226 zio_data_buf_cache[c] = kmem_cache_create(name, size, 227 align, NULL, NULL, NULL, NULL, NULL, data_cflags); 228 } 229 } 230 231 while (--c != 0) { 232 ASSERT(zio_buf_cache[c] != NULL); 233 if (zio_buf_cache[c - 1] == NULL) 234 zio_buf_cache[c - 1] = zio_buf_cache[c]; 235 236 ASSERT(zio_data_buf_cache[c] != NULL); 237 if (zio_data_buf_cache[c - 1] == NULL) 238 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c]; 239 } 240 241 zio_inject_init(); 242 243 lz4_init(); 244 } 245 246 void 247 zio_fini(void) 248 { 249 size_t n = SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; 250 251 #if defined(ZFS_DEBUG) && !defined(_KERNEL) 252 for (size_t i = 0; i < n; i++) { 253 if (zio_buf_cache_allocs[i] != zio_buf_cache_frees[i]) 254 (void) printf("zio_fini: [%d] %llu != %llu\n", 255 (int)((i + 1) << SPA_MINBLOCKSHIFT), 256 (long long unsigned)zio_buf_cache_allocs[i], 257 (long long unsigned)zio_buf_cache_frees[i]); 258 } 259 #endif 260 261 /* 262 * The same kmem cache can show up multiple times in both zio_buf_cache 263 * and zio_data_buf_cache. Do a wasteful but trivially correct scan to 264 * sort it out. 265 */ 266 for (size_t i = 0; i < n; i++) { 267 kmem_cache_t *cache = zio_buf_cache[i]; 268 if (cache == NULL) 269 continue; 270 for (size_t j = i; j < n; j++) { 271 if (cache == zio_buf_cache[j]) 272 zio_buf_cache[j] = NULL; 273 if (cache == zio_data_buf_cache[j]) 274 zio_data_buf_cache[j] = NULL; 275 } 276 kmem_cache_destroy(cache); 277 } 278 279 for (size_t i = 0; i < n; i++) { 280 kmem_cache_t *cache = zio_data_buf_cache[i]; 281 if (cache == NULL) 282 continue; 283 for (size_t j = i; j < n; j++) { 284 if (cache == zio_data_buf_cache[j]) 285 zio_data_buf_cache[j] = NULL; 286 } 287 kmem_cache_destroy(cache); 288 } 289 290 for (size_t i = 0; i < n; i++) { 291 VERIFY3P(zio_buf_cache[i], ==, NULL); 292 VERIFY3P(zio_data_buf_cache[i], ==, NULL); 293 } 294 295 kmem_cache_destroy(zio_link_cache); 296 kmem_cache_destroy(zio_cache); 297 298 zio_inject_fini(); 299 300 lz4_fini(); 301 } 302 303 /* 304 * ========================================================================== 305 * Allocate and free I/O buffers 306 * ========================================================================== 307 */ 308 309 /* 310 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a 311 * crashdump if the kernel panics, so use it judiciously. Obviously, it's 312 * useful to inspect ZFS metadata, but if possible, we should avoid keeping 313 * excess / transient data in-core during a crashdump. 314 */ 315 void * 316 zio_buf_alloc(size_t size) 317 { 318 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 319 320 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 321 #if defined(ZFS_DEBUG) && !defined(_KERNEL) 322 atomic_add_64(&zio_buf_cache_allocs[c], 1); 323 #endif 324 325 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE)); 326 } 327 328 /* 329 * Use zio_data_buf_alloc to allocate data. The data will not appear in a 330 * crashdump if the kernel panics. This exists so that we will limit the amount 331 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount 332 * of kernel heap dumped to disk when the kernel panics) 333 */ 334 void * 335 zio_data_buf_alloc(size_t size) 336 { 337 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 338 339 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 340 341 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE)); 342 } 343 344 void 345 zio_buf_free(void *buf, size_t size) 346 { 347 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 348 349 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 350 #if defined(ZFS_DEBUG) && !defined(_KERNEL) 351 atomic_add_64(&zio_buf_cache_frees[c], 1); 352 #endif 353 354 kmem_cache_free(zio_buf_cache[c], buf); 355 } 356 357 void 358 zio_data_buf_free(void *buf, size_t size) 359 { 360 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 361 362 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 363 364 kmem_cache_free(zio_data_buf_cache[c], buf); 365 } 366 367 static void 368 zio_abd_free(void *abd, size_t size) 369 { 370 (void) size; 371 abd_free((abd_t *)abd); 372 } 373 374 /* 375 * ========================================================================== 376 * Push and pop I/O transform buffers 377 * ========================================================================== 378 */ 379 void 380 zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize, 381 zio_transform_func_t *transform) 382 { 383 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP); 384 385 zt->zt_orig_abd = zio->io_abd; 386 zt->zt_orig_size = zio->io_size; 387 zt->zt_bufsize = bufsize; 388 zt->zt_transform = transform; 389 390 zt->zt_next = zio->io_transform_stack; 391 zio->io_transform_stack = zt; 392 393 zio->io_abd = data; 394 zio->io_size = size; 395 } 396 397 void 398 zio_pop_transforms(zio_t *zio) 399 { 400 zio_transform_t *zt; 401 402 while ((zt = zio->io_transform_stack) != NULL) { 403 if (zt->zt_transform != NULL) 404 zt->zt_transform(zio, 405 zt->zt_orig_abd, zt->zt_orig_size); 406 407 if (zt->zt_bufsize != 0) 408 abd_free(zio->io_abd); 409 410 zio->io_abd = zt->zt_orig_abd; 411 zio->io_size = zt->zt_orig_size; 412 zio->io_transform_stack = zt->zt_next; 413 414 kmem_free(zt, sizeof (zio_transform_t)); 415 } 416 } 417 418 /* 419 * ========================================================================== 420 * I/O transform callbacks for subblocks, decompression, and decryption 421 * ========================================================================== 422 */ 423 static void 424 zio_subblock(zio_t *zio, abd_t *data, uint64_t size) 425 { 426 ASSERT(zio->io_size > size); 427 428 if (zio->io_type == ZIO_TYPE_READ) 429 abd_copy(data, zio->io_abd, size); 430 } 431 432 static void 433 zio_decompress(zio_t *zio, abd_t *data, uint64_t size) 434 { 435 if (zio->io_error == 0) { 436 void *tmp = abd_borrow_buf(data, size); 437 int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp), 438 zio->io_abd, tmp, zio->io_size, size, 439 &zio->io_prop.zp_complevel); 440 abd_return_buf_copy(data, tmp, size); 441 442 if (zio_injection_enabled && ret == 0) 443 ret = zio_handle_fault_injection(zio, EINVAL); 444 445 if (ret != 0) 446 zio->io_error = SET_ERROR(EIO); 447 } 448 } 449 450 static void 451 zio_decrypt(zio_t *zio, abd_t *data, uint64_t size) 452 { 453 int ret; 454 void *tmp; 455 blkptr_t *bp = zio->io_bp; 456 spa_t *spa = zio->io_spa; 457 uint64_t dsobj = zio->io_bookmark.zb_objset; 458 uint64_t lsize = BP_GET_LSIZE(bp); 459 dmu_object_type_t ot = BP_GET_TYPE(bp); 460 uint8_t salt[ZIO_DATA_SALT_LEN]; 461 uint8_t iv[ZIO_DATA_IV_LEN]; 462 uint8_t mac[ZIO_DATA_MAC_LEN]; 463 boolean_t no_crypt = B_FALSE; 464 465 ASSERT(BP_USES_CRYPT(bp)); 466 ASSERT3U(size, !=, 0); 467 468 if (zio->io_error != 0) 469 return; 470 471 /* 472 * Verify the cksum of MACs stored in an indirect bp. It will always 473 * be possible to verify this since it does not require an encryption 474 * key. 475 */ 476 if (BP_HAS_INDIRECT_MAC_CKSUM(bp)) { 477 zio_crypt_decode_mac_bp(bp, mac); 478 479 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF) { 480 /* 481 * We haven't decompressed the data yet, but 482 * zio_crypt_do_indirect_mac_checksum() requires 483 * decompressed data to be able to parse out the MACs 484 * from the indirect block. We decompress it now and 485 * throw away the result after we are finished. 486 */ 487 tmp = zio_buf_alloc(lsize); 488 ret = zio_decompress_data(BP_GET_COMPRESS(bp), 489 zio->io_abd, tmp, zio->io_size, lsize, 490 &zio->io_prop.zp_complevel); 491 if (ret != 0) { 492 ret = SET_ERROR(EIO); 493 goto error; 494 } 495 ret = zio_crypt_do_indirect_mac_checksum(B_FALSE, 496 tmp, lsize, BP_SHOULD_BYTESWAP(bp), mac); 497 zio_buf_free(tmp, lsize); 498 } else { 499 ret = zio_crypt_do_indirect_mac_checksum_abd(B_FALSE, 500 zio->io_abd, size, BP_SHOULD_BYTESWAP(bp), mac); 501 } 502 abd_copy(data, zio->io_abd, size); 503 504 if (zio_injection_enabled && ot != DMU_OT_DNODE && ret == 0) { 505 ret = zio_handle_decrypt_injection(spa, 506 &zio->io_bookmark, ot, ECKSUM); 507 } 508 if (ret != 0) 509 goto error; 510 511 return; 512 } 513 514 /* 515 * If this is an authenticated block, just check the MAC. It would be 516 * nice to separate this out into its own flag, but when this was done, 517 * we had run out of bits in what is now zio_flag_t. Future cleanup 518 * could make this a flag bit. 519 */ 520 if (BP_IS_AUTHENTICATED(bp)) { 521 if (ot == DMU_OT_OBJSET) { 522 ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa, 523 dsobj, zio->io_abd, size, BP_SHOULD_BYTESWAP(bp)); 524 } else { 525 zio_crypt_decode_mac_bp(bp, mac); 526 ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj, 527 zio->io_abd, size, mac); 528 if (zio_injection_enabled && ret == 0) { 529 ret = zio_handle_decrypt_injection(spa, 530 &zio->io_bookmark, ot, ECKSUM); 531 } 532 } 533 abd_copy(data, zio->io_abd, size); 534 535 if (ret != 0) 536 goto error; 537 538 return; 539 } 540 541 zio_crypt_decode_params_bp(bp, salt, iv); 542 543 if (ot == DMU_OT_INTENT_LOG) { 544 tmp = abd_borrow_buf_copy(zio->io_abd, sizeof (zil_chain_t)); 545 zio_crypt_decode_mac_zil(tmp, mac); 546 abd_return_buf(zio->io_abd, tmp, sizeof (zil_chain_t)); 547 } else { 548 zio_crypt_decode_mac_bp(bp, mac); 549 } 550 551 ret = spa_do_crypt_abd(B_FALSE, spa, &zio->io_bookmark, BP_GET_TYPE(bp), 552 BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp), salt, iv, mac, size, data, 553 zio->io_abd, &no_crypt); 554 if (no_crypt) 555 abd_copy(data, zio->io_abd, size); 556 557 if (ret != 0) 558 goto error; 559 560 return; 561 562 error: 563 /* assert that the key was found unless this was speculative */ 564 ASSERT(ret != EACCES || (zio->io_flags & ZIO_FLAG_SPECULATIVE)); 565 566 /* 567 * If there was a decryption / authentication error return EIO as 568 * the io_error. If this was not a speculative zio, create an ereport. 569 */ 570 if (ret == ECKSUM) { 571 zio->io_error = SET_ERROR(EIO); 572 if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) { 573 spa_log_error(spa, &zio->io_bookmark, 574 &zio->io_bp->blk_birth); 575 (void) zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION, 576 spa, NULL, &zio->io_bookmark, zio, 0); 577 } 578 } else { 579 zio->io_error = ret; 580 } 581 } 582 583 /* 584 * ========================================================================== 585 * I/O parent/child relationships and pipeline interlocks 586 * ========================================================================== 587 */ 588 zio_t * 589 zio_walk_parents(zio_t *cio, zio_link_t **zl) 590 { 591 list_t *pl = &cio->io_parent_list; 592 593 *zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl); 594 if (*zl == NULL) 595 return (NULL); 596 597 ASSERT((*zl)->zl_child == cio); 598 return ((*zl)->zl_parent); 599 } 600 601 zio_t * 602 zio_walk_children(zio_t *pio, zio_link_t **zl) 603 { 604 list_t *cl = &pio->io_child_list; 605 606 ASSERT(MUTEX_HELD(&pio->io_lock)); 607 608 *zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl); 609 if (*zl == NULL) 610 return (NULL); 611 612 ASSERT((*zl)->zl_parent == pio); 613 return ((*zl)->zl_child); 614 } 615 616 zio_t * 617 zio_unique_parent(zio_t *cio) 618 { 619 zio_link_t *zl = NULL; 620 zio_t *pio = zio_walk_parents(cio, &zl); 621 622 VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL); 623 return (pio); 624 } 625 626 void 627 zio_add_child(zio_t *pio, zio_t *cio) 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 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP); 638 zl->zl_parent = pio; 639 zl->zl_child = cio; 640 641 mutex_enter(&pio->io_lock); 642 mutex_enter(&cio->io_lock); 643 644 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0); 645 646 uint64_t *countp = pio->io_children[cio->io_child_type]; 647 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 648 countp[w] += !cio->io_state[w]; 649 650 list_insert_head(&pio->io_child_list, zl); 651 list_insert_head(&cio->io_parent_list, zl); 652 653 mutex_exit(&cio->io_lock); 654 mutex_exit(&pio->io_lock); 655 } 656 657 void 658 zio_add_child_first(zio_t *pio, zio_t *cio) 659 { 660 /* 661 * Logical I/Os can have logical, gang, or vdev children. 662 * Gang I/Os can have gang or vdev children. 663 * Vdev I/Os can only have vdev children. 664 * The following ASSERT captures all of these constraints. 665 */ 666 ASSERT3S(cio->io_child_type, <=, pio->io_child_type); 667 668 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP); 669 zl->zl_parent = pio; 670 zl->zl_child = cio; 671 672 ASSERT(list_is_empty(&cio->io_parent_list)); 673 list_insert_head(&cio->io_parent_list, zl); 674 675 mutex_enter(&pio->io_lock); 676 677 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0); 678 679 uint64_t *countp = pio->io_children[cio->io_child_type]; 680 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 681 countp[w] += !cio->io_state[w]; 682 683 list_insert_head(&pio->io_child_list, zl); 684 685 mutex_exit(&pio->io_lock); 686 } 687 688 static void 689 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl) 690 { 691 ASSERT(zl->zl_parent == pio); 692 ASSERT(zl->zl_child == cio); 693 694 mutex_enter(&pio->io_lock); 695 mutex_enter(&cio->io_lock); 696 697 list_remove(&pio->io_child_list, zl); 698 list_remove(&cio->io_parent_list, zl); 699 700 mutex_exit(&cio->io_lock); 701 mutex_exit(&pio->io_lock); 702 kmem_cache_free(zio_link_cache, zl); 703 } 704 705 static boolean_t 706 zio_wait_for_children(zio_t *zio, uint8_t childbits, enum zio_wait_type wait) 707 { 708 boolean_t waiting = B_FALSE; 709 710 mutex_enter(&zio->io_lock); 711 ASSERT(zio->io_stall == NULL); 712 for (int c = 0; c < ZIO_CHILD_TYPES; c++) { 713 if (!(ZIO_CHILD_BIT_IS_SET(childbits, c))) 714 continue; 715 716 uint64_t *countp = &zio->io_children[c][wait]; 717 if (*countp != 0) { 718 zio->io_stage >>= 1; 719 ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN); 720 zio->io_stall = countp; 721 waiting = B_TRUE; 722 break; 723 } 724 } 725 mutex_exit(&zio->io_lock); 726 return (waiting); 727 } 728 729 __attribute__((always_inline)) 730 static inline void 731 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait, 732 zio_t **next_to_executep) 733 { 734 uint64_t *countp = &pio->io_children[zio->io_child_type][wait]; 735 int *errorp = &pio->io_child_error[zio->io_child_type]; 736 737 mutex_enter(&pio->io_lock); 738 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE)) 739 *errorp = zio_worst_error(*errorp, zio->io_error); 740 pio->io_reexecute |= zio->io_reexecute; 741 ASSERT3U(*countp, >, 0); 742 743 (*countp)--; 744 745 if (*countp == 0 && pio->io_stall == countp) { 746 zio_taskq_type_t type = 747 pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE : 748 ZIO_TASKQ_INTERRUPT; 749 pio->io_stall = NULL; 750 mutex_exit(&pio->io_lock); 751 752 /* 753 * If we can tell the caller to execute this parent next, do 754 * so. We only do this if the parent's zio type matches the 755 * child's type. Otherwise dispatch the parent zio in its 756 * own taskq. 757 * 758 * Having the caller execute the parent when possible reduces 759 * locking on the zio taskq's, reduces context switch 760 * overhead, and has no recursion penalty. Note that one 761 * read from disk typically causes at least 3 zio's: a 762 * zio_null(), the logical zio_read(), and then a physical 763 * zio. When the physical ZIO completes, we are able to call 764 * zio_done() on all 3 of these zio's from one invocation of 765 * zio_execute() by returning the parent back to 766 * zio_execute(). Since the parent isn't executed until this 767 * thread returns back to zio_execute(), the caller should do 768 * so promptly. 769 * 770 * In other cases, dispatching the parent prevents 771 * overflowing the stack when we have deeply nested 772 * parent-child relationships, as we do with the "mega zio" 773 * of writes for spa_sync(), and the chain of ZIL blocks. 774 */ 775 if (next_to_executep != NULL && *next_to_executep == NULL && 776 pio->io_type == zio->io_type) { 777 *next_to_executep = pio; 778 } else { 779 zio_taskq_dispatch(pio, type, B_FALSE); 780 } 781 } else { 782 mutex_exit(&pio->io_lock); 783 } 784 } 785 786 static void 787 zio_inherit_child_errors(zio_t *zio, enum zio_child c) 788 { 789 if (zio->io_child_error[c] != 0 && zio->io_error == 0) 790 zio->io_error = zio->io_child_error[c]; 791 } 792 793 int 794 zio_bookmark_compare(const void *x1, const void *x2) 795 { 796 const zio_t *z1 = x1; 797 const zio_t *z2 = x2; 798 799 if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset) 800 return (-1); 801 if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset) 802 return (1); 803 804 if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object) 805 return (-1); 806 if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object) 807 return (1); 808 809 if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level) 810 return (-1); 811 if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level) 812 return (1); 813 814 if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid) 815 return (-1); 816 if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid) 817 return (1); 818 819 if (z1 < z2) 820 return (-1); 821 if (z1 > z2) 822 return (1); 823 824 return (0); 825 } 826 827 /* 828 * ========================================================================== 829 * Create the various types of I/O (read, write, free, etc) 830 * ========================================================================== 831 */ 832 static zio_t * 833 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 834 abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done, 835 void *private, zio_type_t type, zio_priority_t priority, 836 zio_flag_t flags, vdev_t *vd, uint64_t offset, 837 const zbookmark_phys_t *zb, enum zio_stage stage, 838 enum zio_stage pipeline) 839 { 840 zio_t *zio; 841 842 IMPLY(type != ZIO_TYPE_TRIM, psize <= SPA_MAXBLOCKSIZE); 843 ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0); 844 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0); 845 846 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER)); 847 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER)); 848 ASSERT(vd || stage == ZIO_STAGE_OPEN); 849 850 IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW_COMPRESS) != 0); 851 852 zio = kmem_cache_alloc(zio_cache, KM_SLEEP); 853 memset(zio, 0, sizeof (zio_t)); 854 855 mutex_init(&zio->io_lock, NULL, MUTEX_NOLOCKDEP, NULL); 856 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL); 857 858 list_create(&zio->io_parent_list, sizeof (zio_link_t), 859 offsetof(zio_link_t, zl_parent_node)); 860 list_create(&zio->io_child_list, sizeof (zio_link_t), 861 offsetof(zio_link_t, zl_child_node)); 862 metaslab_trace_init(&zio->io_alloc_list); 863 864 if (vd != NULL) 865 zio->io_child_type = ZIO_CHILD_VDEV; 866 else if (flags & ZIO_FLAG_GANG_CHILD) 867 zio->io_child_type = ZIO_CHILD_GANG; 868 else if (flags & ZIO_FLAG_DDT_CHILD) 869 zio->io_child_type = ZIO_CHILD_DDT; 870 else 871 zio->io_child_type = ZIO_CHILD_LOGICAL; 872 873 if (bp != NULL) { 874 if (type != ZIO_TYPE_WRITE || 875 zio->io_child_type == ZIO_CHILD_DDT) { 876 zio->io_bp_copy = *bp; 877 zio->io_bp = &zio->io_bp_copy; /* so caller can free */ 878 } else { 879 zio->io_bp = (blkptr_t *)bp; 880 } 881 zio->io_bp_orig = *bp; 882 if (zio->io_child_type == ZIO_CHILD_LOGICAL) 883 zio->io_logical = zio; 884 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp)) 885 pipeline |= ZIO_GANG_STAGES; 886 } 887 888 zio->io_spa = spa; 889 zio->io_txg = txg; 890 zio->io_done = done; 891 zio->io_private = private; 892 zio->io_type = type; 893 zio->io_priority = priority; 894 zio->io_vd = vd; 895 zio->io_offset = offset; 896 zio->io_orig_abd = zio->io_abd = data; 897 zio->io_orig_size = zio->io_size = psize; 898 zio->io_lsize = lsize; 899 zio->io_orig_flags = zio->io_flags = flags; 900 zio->io_orig_stage = zio->io_stage = stage; 901 zio->io_orig_pipeline = zio->io_pipeline = pipeline; 902 zio->io_pipeline_trace = ZIO_STAGE_OPEN; 903 904 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY); 905 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE); 906 907 if (zb != NULL) 908 zio->io_bookmark = *zb; 909 910 if (pio != NULL) { 911 zio->io_metaslab_class = pio->io_metaslab_class; 912 if (zio->io_logical == NULL) 913 zio->io_logical = pio->io_logical; 914 if (zio->io_child_type == ZIO_CHILD_GANG) 915 zio->io_gang_leader = pio->io_gang_leader; 916 zio_add_child_first(pio, zio); 917 } 918 919 taskq_init_ent(&zio->io_tqent); 920 921 return (zio); 922 } 923 924 void 925 zio_destroy(zio_t *zio) 926 { 927 metaslab_trace_fini(&zio->io_alloc_list); 928 list_destroy(&zio->io_parent_list); 929 list_destroy(&zio->io_child_list); 930 mutex_destroy(&zio->io_lock); 931 cv_destroy(&zio->io_cv); 932 kmem_cache_free(zio_cache, zio); 933 } 934 935 zio_t * 936 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done, 937 void *private, zio_flag_t flags) 938 { 939 zio_t *zio; 940 941 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private, 942 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, 943 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE); 944 945 return (zio); 946 } 947 948 zio_t * 949 zio_root(spa_t *spa, zio_done_func_t *done, void *private, zio_flag_t flags) 950 { 951 return (zio_null(NULL, spa, NULL, done, private, flags)); 952 } 953 954 static int 955 zfs_blkptr_verify_log(spa_t *spa, const blkptr_t *bp, 956 enum blk_verify_flag blk_verify, const char *fmt, ...) 957 { 958 va_list adx; 959 char buf[256]; 960 961 va_start(adx, fmt); 962 (void) vsnprintf(buf, sizeof (buf), fmt, adx); 963 va_end(adx); 964 965 zfs_dbgmsg("bad blkptr at %px: " 966 "DVA[0]=%#llx/%#llx " 967 "DVA[1]=%#llx/%#llx " 968 "DVA[2]=%#llx/%#llx " 969 "prop=%#llx " 970 "pad=%#llx,%#llx " 971 "phys_birth=%#llx " 972 "birth=%#llx " 973 "fill=%#llx " 974 "cksum=%#llx/%#llx/%#llx/%#llx", 975 bp, 976 (long long)bp->blk_dva[0].dva_word[0], 977 (long long)bp->blk_dva[0].dva_word[1], 978 (long long)bp->blk_dva[1].dva_word[0], 979 (long long)bp->blk_dva[1].dva_word[1], 980 (long long)bp->blk_dva[2].dva_word[0], 981 (long long)bp->blk_dva[2].dva_word[1], 982 (long long)bp->blk_prop, 983 (long long)bp->blk_pad[0], 984 (long long)bp->blk_pad[1], 985 (long long)bp->blk_phys_birth, 986 (long long)bp->blk_birth, 987 (long long)bp->blk_fill, 988 (long long)bp->blk_cksum.zc_word[0], 989 (long long)bp->blk_cksum.zc_word[1], 990 (long long)bp->blk_cksum.zc_word[2], 991 (long long)bp->blk_cksum.zc_word[3]); 992 switch (blk_verify) { 993 case BLK_VERIFY_HALT: 994 zfs_panic_recover("%s: %s", spa_name(spa), buf); 995 break; 996 case BLK_VERIFY_LOG: 997 zfs_dbgmsg("%s: %s", spa_name(spa), buf); 998 break; 999 case BLK_VERIFY_ONLY: 1000 break; 1001 } 1002 1003 return (1); 1004 } 1005 1006 /* 1007 * Verify the block pointer fields contain reasonable values. This means 1008 * it only contains known object types, checksum/compression identifiers, 1009 * block sizes within the maximum allowed limits, valid DVAs, etc. 1010 * 1011 * If everything checks out B_TRUE is returned. The zfs_blkptr_verify 1012 * argument controls the behavior when an invalid field is detected. 1013 * 1014 * Values for blk_verify_flag: 1015 * BLK_VERIFY_ONLY: evaluate the block 1016 * BLK_VERIFY_LOG: evaluate the block and log problems 1017 * BLK_VERIFY_HALT: call zfs_panic_recover on error 1018 * 1019 * Values for blk_config_flag: 1020 * BLK_CONFIG_HELD: caller holds SCL_VDEV for writer 1021 * BLK_CONFIG_NEEDED: caller holds no config lock, SCL_VDEV will be 1022 * obtained for reader 1023 * BLK_CONFIG_SKIP: skip checks which require SCL_VDEV, for better 1024 * performance 1025 */ 1026 boolean_t 1027 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp, 1028 enum blk_config_flag blk_config, enum blk_verify_flag blk_verify) 1029 { 1030 int errors = 0; 1031 1032 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) { 1033 errors += zfs_blkptr_verify_log(spa, bp, blk_verify, 1034 "blkptr at %px has invalid TYPE %llu", 1035 bp, (longlong_t)BP_GET_TYPE(bp)); 1036 } 1037 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS) { 1038 errors += zfs_blkptr_verify_log(spa, bp, blk_verify, 1039 "blkptr at %px has invalid CHECKSUM %llu", 1040 bp, (longlong_t)BP_GET_CHECKSUM(bp)); 1041 } 1042 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS) { 1043 errors += zfs_blkptr_verify_log(spa, bp, blk_verify, 1044 "blkptr at %px has invalid COMPRESS %llu", 1045 bp, (longlong_t)BP_GET_COMPRESS(bp)); 1046 } 1047 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) { 1048 errors += zfs_blkptr_verify_log(spa, bp, blk_verify, 1049 "blkptr at %px has invalid LSIZE %llu", 1050 bp, (longlong_t)BP_GET_LSIZE(bp)); 1051 } 1052 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) { 1053 errors += zfs_blkptr_verify_log(spa, bp, blk_verify, 1054 "blkptr at %px has invalid PSIZE %llu", 1055 bp, (longlong_t)BP_GET_PSIZE(bp)); 1056 } 1057 1058 if (BP_IS_EMBEDDED(bp)) { 1059 if (BPE_GET_ETYPE(bp) >= NUM_BP_EMBEDDED_TYPES) { 1060 errors += zfs_blkptr_verify_log(spa, bp, blk_verify, 1061 "blkptr at %px has invalid ETYPE %llu", 1062 bp, (longlong_t)BPE_GET_ETYPE(bp)); 1063 } 1064 } 1065 1066 /* 1067 * Do not verify individual DVAs if the config is not trusted. This 1068 * will be done once the zio is executed in vdev_mirror_map_alloc. 1069 */ 1070 if (!spa->spa_trust_config) 1071 return (errors == 0); 1072 1073 switch (blk_config) { 1074 case BLK_CONFIG_HELD: 1075 ASSERT(spa_config_held(spa, SCL_VDEV, RW_WRITER)); 1076 break; 1077 case BLK_CONFIG_NEEDED: 1078 spa_config_enter(spa, SCL_VDEV, bp, RW_READER); 1079 break; 1080 case BLK_CONFIG_SKIP: 1081 return (errors == 0); 1082 default: 1083 panic("invalid blk_config %u", blk_config); 1084 } 1085 1086 /* 1087 * Pool-specific checks. 1088 * 1089 * Note: it would be nice to verify that the blk_birth and 1090 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze() 1091 * allows the birth time of log blocks (and dmu_sync()-ed blocks 1092 * that are in the log) to be arbitrarily large. 1093 */ 1094 for (int i = 0; i < BP_GET_NDVAS(bp); i++) { 1095 const dva_t *dva = &bp->blk_dva[i]; 1096 uint64_t vdevid = DVA_GET_VDEV(dva); 1097 1098 if (vdevid >= spa->spa_root_vdev->vdev_children) { 1099 errors += zfs_blkptr_verify_log(spa, bp, blk_verify, 1100 "blkptr at %px DVA %u has invalid VDEV %llu", 1101 bp, i, (longlong_t)vdevid); 1102 continue; 1103 } 1104 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid]; 1105 if (vd == NULL) { 1106 errors += zfs_blkptr_verify_log(spa, bp, blk_verify, 1107 "blkptr at %px DVA %u has invalid VDEV %llu", 1108 bp, i, (longlong_t)vdevid); 1109 continue; 1110 } 1111 if (vd->vdev_ops == &vdev_hole_ops) { 1112 errors += zfs_blkptr_verify_log(spa, bp, blk_verify, 1113 "blkptr at %px DVA %u has hole VDEV %llu", 1114 bp, i, (longlong_t)vdevid); 1115 continue; 1116 } 1117 if (vd->vdev_ops == &vdev_missing_ops) { 1118 /* 1119 * "missing" vdevs are valid during import, but we 1120 * don't have their detailed info (e.g. asize), so 1121 * we can't perform any more checks on them. 1122 */ 1123 continue; 1124 } 1125 uint64_t offset = DVA_GET_OFFSET(dva); 1126 uint64_t asize = DVA_GET_ASIZE(dva); 1127 if (DVA_GET_GANG(dva)) 1128 asize = vdev_gang_header_asize(vd); 1129 if (offset + asize > vd->vdev_asize) { 1130 errors += zfs_blkptr_verify_log(spa, bp, blk_verify, 1131 "blkptr at %px DVA %u has invalid OFFSET %llu", 1132 bp, i, (longlong_t)offset); 1133 } 1134 } 1135 if (blk_config == BLK_CONFIG_NEEDED) 1136 spa_config_exit(spa, SCL_VDEV, bp); 1137 1138 return (errors == 0); 1139 } 1140 1141 boolean_t 1142 zfs_dva_valid(spa_t *spa, const dva_t *dva, const blkptr_t *bp) 1143 { 1144 (void) bp; 1145 uint64_t vdevid = DVA_GET_VDEV(dva); 1146 1147 if (vdevid >= spa->spa_root_vdev->vdev_children) 1148 return (B_FALSE); 1149 1150 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid]; 1151 if (vd == NULL) 1152 return (B_FALSE); 1153 1154 if (vd->vdev_ops == &vdev_hole_ops) 1155 return (B_FALSE); 1156 1157 if (vd->vdev_ops == &vdev_missing_ops) { 1158 return (B_FALSE); 1159 } 1160 1161 uint64_t offset = DVA_GET_OFFSET(dva); 1162 uint64_t asize = DVA_GET_ASIZE(dva); 1163 1164 if (DVA_GET_GANG(dva)) 1165 asize = vdev_gang_header_asize(vd); 1166 if (offset + asize > vd->vdev_asize) 1167 return (B_FALSE); 1168 1169 return (B_TRUE); 1170 } 1171 1172 zio_t * 1173 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, 1174 abd_t *data, uint64_t size, zio_done_func_t *done, void *private, 1175 zio_priority_t priority, zio_flag_t flags, const zbookmark_phys_t *zb) 1176 { 1177 zio_t *zio; 1178 1179 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp, 1180 data, size, size, done, private, 1181 ZIO_TYPE_READ, priority, flags, NULL, 0, zb, 1182 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? 1183 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE); 1184 1185 return (zio); 1186 } 1187 1188 zio_t * 1189 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, 1190 abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp, 1191 zio_done_func_t *ready, zio_done_func_t *children_ready, 1192 zio_done_func_t *done, void *private, zio_priority_t priority, 1193 zio_flag_t flags, const zbookmark_phys_t *zb) 1194 { 1195 zio_t *zio; 1196 1197 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF && 1198 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS && 1199 zp->zp_compress >= ZIO_COMPRESS_OFF && 1200 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS && 1201 DMU_OT_IS_VALID(zp->zp_type) && 1202 zp->zp_level < 32 && 1203 zp->zp_copies > 0 && 1204 zp->zp_copies <= spa_max_replication(spa)); 1205 1206 zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private, 1207 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, 1208 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? 1209 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE); 1210 1211 zio->io_ready = ready; 1212 zio->io_children_ready = children_ready; 1213 zio->io_prop = *zp; 1214 1215 /* 1216 * Data can be NULL if we are going to call zio_write_override() to 1217 * provide the already-allocated BP. But we may need the data to 1218 * verify a dedup hit (if requested). In this case, don't try to 1219 * dedup (just take the already-allocated BP verbatim). Encrypted 1220 * dedup blocks need data as well so we also disable dedup in this 1221 * case. 1222 */ 1223 if (data == NULL && 1224 (zio->io_prop.zp_dedup_verify || zio->io_prop.zp_encrypt)) { 1225 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE; 1226 } 1227 1228 return (zio); 1229 } 1230 1231 zio_t * 1232 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data, 1233 uint64_t size, zio_done_func_t *done, void *private, 1234 zio_priority_t priority, zio_flag_t flags, zbookmark_phys_t *zb) 1235 { 1236 zio_t *zio; 1237 1238 zio = zio_create(pio, spa, txg, bp, data, size, size, done, private, 1239 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb, 1240 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE); 1241 1242 return (zio); 1243 } 1244 1245 void 1246 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite, 1247 boolean_t brtwrite) 1248 { 1249 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 1250 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1251 ASSERT(zio->io_stage == ZIO_STAGE_OPEN); 1252 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa)); 1253 ASSERT(!brtwrite || !nopwrite); 1254 1255 /* 1256 * We must reset the io_prop to match the values that existed 1257 * when the bp was first written by dmu_sync() keeping in mind 1258 * that nopwrite and dedup are mutually exclusive. 1259 */ 1260 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup; 1261 zio->io_prop.zp_nopwrite = nopwrite; 1262 zio->io_prop.zp_brtwrite = brtwrite; 1263 zio->io_prop.zp_copies = copies; 1264 zio->io_bp_override = bp; 1265 } 1266 1267 void 1268 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp) 1269 { 1270 1271 (void) zfs_blkptr_verify(spa, bp, BLK_CONFIG_NEEDED, BLK_VERIFY_HALT); 1272 1273 /* 1274 * The check for EMBEDDED is a performance optimization. We 1275 * process the free here (by ignoring it) rather than 1276 * putting it on the list and then processing it in zio_free_sync(). 1277 */ 1278 if (BP_IS_EMBEDDED(bp)) 1279 return; 1280 1281 /* 1282 * Frees that are for the currently-syncing txg, are not going to be 1283 * deferred, and which will not need to do a read (i.e. not GANG or 1284 * DEDUP), can be processed immediately. Otherwise, put them on the 1285 * in-memory list for later processing. 1286 * 1287 * Note that we only defer frees after zfs_sync_pass_deferred_free 1288 * when the log space map feature is disabled. [see relevant comment 1289 * in spa_sync_iterate_to_convergence()] 1290 */ 1291 if (BP_IS_GANG(bp) || 1292 BP_GET_DEDUP(bp) || 1293 txg != spa->spa_syncing_txg || 1294 (spa_sync_pass(spa) >= zfs_sync_pass_deferred_free && 1295 !spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP)) || 1296 brt_maybe_exists(spa, bp)) { 1297 metaslab_check_free(spa, bp); 1298 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp); 1299 } else { 1300 VERIFY3P(zio_free_sync(NULL, spa, txg, bp, 0), ==, NULL); 1301 } 1302 } 1303 1304 /* 1305 * To improve performance, this function may return NULL if we were able 1306 * to do the free immediately. This avoids the cost of creating a zio 1307 * (and linking it to the parent, etc). 1308 */ 1309 zio_t * 1310 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 1311 zio_flag_t flags) 1312 { 1313 ASSERT(!BP_IS_HOLE(bp)); 1314 ASSERT(spa_syncing_txg(spa) == txg); 1315 1316 if (BP_IS_EMBEDDED(bp)) 1317 return (NULL); 1318 1319 metaslab_check_free(spa, bp); 1320 arc_freed(spa, bp); 1321 dsl_scan_freed(spa, bp); 1322 1323 if (BP_IS_GANG(bp) || 1324 BP_GET_DEDUP(bp) || 1325 brt_maybe_exists(spa, bp)) { 1326 /* 1327 * GANG, DEDUP and BRT blocks can induce a read (for the gang 1328 * block header, the DDT or the BRT), so issue them 1329 * asynchronously so that this thread is not tied up. 1330 */ 1331 enum zio_stage stage = 1332 ZIO_FREE_PIPELINE | ZIO_STAGE_ISSUE_ASYNC; 1333 1334 return (zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), 1335 BP_GET_PSIZE(bp), NULL, NULL, 1336 ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, 1337 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage)); 1338 } else { 1339 metaslab_free(spa, bp, txg, B_FALSE); 1340 return (NULL); 1341 } 1342 } 1343 1344 zio_t * 1345 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 1346 zio_done_func_t *done, void *private, zio_flag_t flags) 1347 { 1348 zio_t *zio; 1349 1350 (void) zfs_blkptr_verify(spa, bp, (flags & ZIO_FLAG_CONFIG_WRITER) ? 1351 BLK_CONFIG_HELD : BLK_CONFIG_NEEDED, BLK_VERIFY_HALT); 1352 1353 if (BP_IS_EMBEDDED(bp)) 1354 return (zio_null(pio, spa, NULL, NULL, NULL, 0)); 1355 1356 /* 1357 * A claim is an allocation of a specific block. Claims are needed 1358 * to support immediate writes in the intent log. The issue is that 1359 * immediate writes contain committed data, but in a txg that was 1360 * *not* committed. Upon opening the pool after an unclean shutdown, 1361 * the intent log claims all blocks that contain immediate write data 1362 * so that the SPA knows they're in use. 1363 * 1364 * All claims *must* be resolved in the first txg -- before the SPA 1365 * starts allocating blocks -- so that nothing is allocated twice. 1366 * If txg == 0 we just verify that the block is claimable. 1367 */ 1368 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, 1369 spa_min_claim_txg(spa)); 1370 ASSERT(txg == spa_min_claim_txg(spa) || txg == 0); 1371 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(8) */ 1372 1373 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), 1374 BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, 1375 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE); 1376 ASSERT0(zio->io_queued_timestamp); 1377 1378 return (zio); 1379 } 1380 1381 zio_t * 1382 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, 1383 zio_done_func_t *done, void *private, zio_flag_t flags) 1384 { 1385 zio_t *zio; 1386 int c; 1387 1388 if (vd->vdev_children == 0) { 1389 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private, 1390 ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, 1391 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE); 1392 1393 zio->io_cmd = cmd; 1394 } else { 1395 zio = zio_null(pio, spa, NULL, NULL, NULL, flags); 1396 1397 for (c = 0; c < vd->vdev_children; c++) 1398 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd, 1399 done, private, flags)); 1400 } 1401 1402 return (zio); 1403 } 1404 1405 zio_t * 1406 zio_trim(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 1407 zio_done_func_t *done, void *private, zio_priority_t priority, 1408 zio_flag_t flags, enum trim_flag trim_flags) 1409 { 1410 zio_t *zio; 1411 1412 ASSERT0(vd->vdev_children); 1413 ASSERT0(P2PHASE(offset, 1ULL << vd->vdev_ashift)); 1414 ASSERT0(P2PHASE(size, 1ULL << vd->vdev_ashift)); 1415 ASSERT3U(size, !=, 0); 1416 1417 zio = zio_create(pio, vd->vdev_spa, 0, NULL, NULL, size, size, done, 1418 private, ZIO_TYPE_TRIM, priority, flags | ZIO_FLAG_PHYSICAL, 1419 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_TRIM_PIPELINE); 1420 zio->io_trim_flags = trim_flags; 1421 1422 return (zio); 1423 } 1424 1425 zio_t * 1426 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 1427 abd_t *data, int checksum, zio_done_func_t *done, void *private, 1428 zio_priority_t priority, zio_flag_t flags, boolean_t labels) 1429 { 1430 zio_t *zio; 1431 1432 ASSERT(vd->vdev_children == 0); 1433 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || 1434 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); 1435 ASSERT3U(offset + size, <=, vd->vdev_psize); 1436 1437 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done, 1438 private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, 1439 offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE); 1440 1441 zio->io_prop.zp_checksum = checksum; 1442 1443 return (zio); 1444 } 1445 1446 zio_t * 1447 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 1448 abd_t *data, int checksum, zio_done_func_t *done, void *private, 1449 zio_priority_t priority, zio_flag_t flags, boolean_t labels) 1450 { 1451 zio_t *zio; 1452 1453 ASSERT(vd->vdev_children == 0); 1454 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || 1455 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); 1456 ASSERT3U(offset + size, <=, vd->vdev_psize); 1457 1458 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done, 1459 private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, 1460 offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE); 1461 1462 zio->io_prop.zp_checksum = checksum; 1463 1464 if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) { 1465 /* 1466 * zec checksums are necessarily destructive -- they modify 1467 * the end of the write buffer to hold the verifier/checksum. 1468 * Therefore, we must make a local copy in case the data is 1469 * being written to multiple places in parallel. 1470 */ 1471 abd_t *wbuf = abd_alloc_sametype(data, size); 1472 abd_copy(wbuf, data, size); 1473 1474 zio_push_transform(zio, wbuf, size, size, NULL); 1475 } 1476 1477 return (zio); 1478 } 1479 1480 /* 1481 * Create a child I/O to do some work for us. 1482 */ 1483 zio_t * 1484 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset, 1485 abd_t *data, uint64_t size, int type, zio_priority_t priority, 1486 zio_flag_t flags, zio_done_func_t *done, void *private) 1487 { 1488 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE; 1489 zio_t *zio; 1490 1491 /* 1492 * vdev child I/Os do not propagate their error to the parent. 1493 * Therefore, for correct operation the caller *must* check for 1494 * and handle the error in the child i/o's done callback. 1495 * The only exceptions are i/os that we don't care about 1496 * (OPTIONAL or REPAIR). 1497 */ 1498 ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) || 1499 done != NULL); 1500 1501 if (type == ZIO_TYPE_READ && bp != NULL) { 1502 /* 1503 * If we have the bp, then the child should perform the 1504 * checksum and the parent need not. This pushes error 1505 * detection as close to the leaves as possible and 1506 * eliminates redundant checksums in the interior nodes. 1507 */ 1508 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY; 1509 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; 1510 } 1511 1512 if (vd->vdev_ops->vdev_op_leaf) { 1513 ASSERT0(vd->vdev_children); 1514 offset += VDEV_LABEL_START_SIZE; 1515 } 1516 1517 flags |= ZIO_VDEV_CHILD_FLAGS(pio); 1518 1519 /* 1520 * If we've decided to do a repair, the write is not speculative -- 1521 * even if the original read was. 1522 */ 1523 if (flags & ZIO_FLAG_IO_REPAIR) 1524 flags &= ~ZIO_FLAG_SPECULATIVE; 1525 1526 /* 1527 * If we're creating a child I/O that is not associated with a 1528 * top-level vdev, then the child zio is not an allocating I/O. 1529 * If this is a retried I/O then we ignore it since we will 1530 * have already processed the original allocating I/O. 1531 */ 1532 if (flags & ZIO_FLAG_IO_ALLOCATING && 1533 (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) { 1534 ASSERT(pio->io_metaslab_class != NULL); 1535 ASSERT(pio->io_metaslab_class->mc_alloc_throttle_enabled); 1536 ASSERT(type == ZIO_TYPE_WRITE); 1537 ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE); 1538 ASSERT(!(flags & ZIO_FLAG_IO_REPAIR)); 1539 ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) || 1540 pio->io_child_type == ZIO_CHILD_GANG); 1541 1542 flags &= ~ZIO_FLAG_IO_ALLOCATING; 1543 } 1544 1545 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size, 1546 done, private, type, priority, flags, vd, offset, &pio->io_bookmark, 1547 ZIO_STAGE_VDEV_IO_START >> 1, pipeline); 1548 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV); 1549 1550 return (zio); 1551 } 1552 1553 zio_t * 1554 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size, 1555 zio_type_t type, zio_priority_t priority, zio_flag_t flags, 1556 zio_done_func_t *done, void *private) 1557 { 1558 zio_t *zio; 1559 1560 ASSERT(vd->vdev_ops->vdev_op_leaf); 1561 1562 zio = zio_create(NULL, vd->vdev_spa, 0, NULL, 1563 data, size, size, done, private, type, priority, 1564 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED, 1565 vd, offset, NULL, 1566 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE); 1567 1568 return (zio); 1569 } 1570 1571 void 1572 zio_flush(zio_t *zio, vdev_t *vd) 1573 { 1574 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 1575 NULL, NULL, 1576 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY)); 1577 } 1578 1579 void 1580 zio_shrink(zio_t *zio, uint64_t size) 1581 { 1582 ASSERT3P(zio->io_executor, ==, NULL); 1583 ASSERT3U(zio->io_orig_size, ==, zio->io_size); 1584 ASSERT3U(size, <=, zio->io_size); 1585 1586 /* 1587 * We don't shrink for raidz because of problems with the 1588 * reconstruction when reading back less than the block size. 1589 * Note, BP_IS_RAIDZ() assumes no compression. 1590 */ 1591 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 1592 if (!BP_IS_RAIDZ(zio->io_bp)) { 1593 /* we are not doing a raw write */ 1594 ASSERT3U(zio->io_size, ==, zio->io_lsize); 1595 zio->io_orig_size = zio->io_size = zio->io_lsize = size; 1596 } 1597 } 1598 1599 /* 1600 * Round provided allocation size up to a value that can be allocated 1601 * by at least some vdev(s) in the pool with minimum or no additional 1602 * padding and without extra space usage on others 1603 */ 1604 static uint64_t 1605 zio_roundup_alloc_size(spa_t *spa, uint64_t size) 1606 { 1607 if (size > spa->spa_min_alloc) 1608 return (roundup(size, spa->spa_gcd_alloc)); 1609 return (spa->spa_min_alloc); 1610 } 1611 1612 /* 1613 * ========================================================================== 1614 * Prepare to read and write logical blocks 1615 * ========================================================================== 1616 */ 1617 1618 static zio_t * 1619 zio_read_bp_init(zio_t *zio) 1620 { 1621 blkptr_t *bp = zio->io_bp; 1622 uint64_t psize = 1623 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp); 1624 1625 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy); 1626 1627 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF && 1628 zio->io_child_type == ZIO_CHILD_LOGICAL && 1629 !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) { 1630 zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize), 1631 psize, psize, zio_decompress); 1632 } 1633 1634 if (((BP_IS_PROTECTED(bp) && !(zio->io_flags & ZIO_FLAG_RAW_ENCRYPT)) || 1635 BP_HAS_INDIRECT_MAC_CKSUM(bp)) && 1636 zio->io_child_type == ZIO_CHILD_LOGICAL) { 1637 zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize), 1638 psize, psize, zio_decrypt); 1639 } 1640 1641 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) { 1642 int psize = BPE_GET_PSIZE(bp); 1643 void *data = abd_borrow_buf(zio->io_abd, psize); 1644 1645 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1646 decode_embedded_bp_compressed(bp, data); 1647 abd_return_buf_copy(zio->io_abd, data, psize); 1648 } else { 1649 ASSERT(!BP_IS_EMBEDDED(bp)); 1650 } 1651 1652 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL) 1653 zio->io_pipeline = ZIO_DDT_READ_PIPELINE; 1654 1655 return (zio); 1656 } 1657 1658 static zio_t * 1659 zio_write_bp_init(zio_t *zio) 1660 { 1661 if (!IO_IS_ALLOCATING(zio)) 1662 return (zio); 1663 1664 ASSERT(zio->io_child_type != ZIO_CHILD_DDT); 1665 1666 if (zio->io_bp_override) { 1667 blkptr_t *bp = zio->io_bp; 1668 zio_prop_t *zp = &zio->io_prop; 1669 1670 ASSERT(bp->blk_birth != zio->io_txg); 1671 1672 *bp = *zio->io_bp_override; 1673 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1674 1675 if (zp->zp_brtwrite) 1676 return (zio); 1677 1678 ASSERT(!BP_GET_DEDUP(zio->io_bp_override)); 1679 1680 if (BP_IS_EMBEDDED(bp)) 1681 return (zio); 1682 1683 /* 1684 * If we've been overridden and nopwrite is set then 1685 * set the flag accordingly to indicate that a nopwrite 1686 * has already occurred. 1687 */ 1688 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) { 1689 ASSERT(!zp->zp_dedup); 1690 ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum); 1691 zio->io_flags |= ZIO_FLAG_NOPWRITE; 1692 return (zio); 1693 } 1694 1695 ASSERT(!zp->zp_nopwrite); 1696 1697 if (BP_IS_HOLE(bp) || !zp->zp_dedup) 1698 return (zio); 1699 1700 ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags & 1701 ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify); 1702 1703 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum && 1704 !zp->zp_encrypt) { 1705 BP_SET_DEDUP(bp, 1); 1706 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE; 1707 return (zio); 1708 } 1709 1710 /* 1711 * We were unable to handle this as an override bp, treat 1712 * it as a regular write I/O. 1713 */ 1714 zio->io_bp_override = NULL; 1715 *bp = zio->io_bp_orig; 1716 zio->io_pipeline = zio->io_orig_pipeline; 1717 } 1718 1719 return (zio); 1720 } 1721 1722 static zio_t * 1723 zio_write_compress(zio_t *zio) 1724 { 1725 spa_t *spa = zio->io_spa; 1726 zio_prop_t *zp = &zio->io_prop; 1727 enum zio_compress compress = zp->zp_compress; 1728 blkptr_t *bp = zio->io_bp; 1729 uint64_t lsize = zio->io_lsize; 1730 uint64_t psize = zio->io_size; 1731 uint32_t pass = 1; 1732 1733 /* 1734 * If our children haven't all reached the ready stage, 1735 * wait for them and then repeat this pipeline stage. 1736 */ 1737 if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT | 1738 ZIO_CHILD_GANG_BIT, ZIO_WAIT_READY)) { 1739 return (NULL); 1740 } 1741 1742 if (!IO_IS_ALLOCATING(zio)) 1743 return (zio); 1744 1745 if (zio->io_children_ready != NULL) { 1746 /* 1747 * Now that all our children are ready, run the callback 1748 * associated with this zio in case it wants to modify the 1749 * data to be written. 1750 */ 1751 ASSERT3U(zp->zp_level, >, 0); 1752 zio->io_children_ready(zio); 1753 } 1754 1755 ASSERT(zio->io_child_type != ZIO_CHILD_DDT); 1756 ASSERT(zio->io_bp_override == NULL); 1757 1758 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) { 1759 /* 1760 * We're rewriting an existing block, which means we're 1761 * working on behalf of spa_sync(). For spa_sync() to 1762 * converge, it must eventually be the case that we don't 1763 * have to allocate new blocks. But compression changes 1764 * the blocksize, which forces a reallocate, and makes 1765 * convergence take longer. Therefore, after the first 1766 * few passes, stop compressing to ensure convergence. 1767 */ 1768 pass = spa_sync_pass(spa); 1769 1770 ASSERT(zio->io_txg == spa_syncing_txg(spa)); 1771 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1772 ASSERT(!BP_GET_DEDUP(bp)); 1773 1774 if (pass >= zfs_sync_pass_dont_compress) 1775 compress = ZIO_COMPRESS_OFF; 1776 1777 /* Make sure someone doesn't change their mind on overwrites */ 1778 ASSERT(BP_IS_EMBEDDED(bp) || BP_IS_GANG(bp) || 1779 MIN(zp->zp_copies, spa_max_replication(spa)) 1780 == BP_GET_NDVAS(bp)); 1781 } 1782 1783 /* If it's a compressed write that is not raw, compress the buffer. */ 1784 if (compress != ZIO_COMPRESS_OFF && 1785 !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) { 1786 void *cbuf = NULL; 1787 psize = zio_compress_data(compress, zio->io_abd, &cbuf, lsize, 1788 zp->zp_complevel); 1789 if (psize == 0) { 1790 compress = ZIO_COMPRESS_OFF; 1791 } else if (psize >= lsize) { 1792 compress = ZIO_COMPRESS_OFF; 1793 if (cbuf != NULL) 1794 zio_buf_free(cbuf, lsize); 1795 } else if (!zp->zp_dedup && !zp->zp_encrypt && 1796 psize <= BPE_PAYLOAD_SIZE && 1797 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) && 1798 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) { 1799 encode_embedded_bp_compressed(bp, 1800 cbuf, compress, lsize, psize); 1801 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA); 1802 BP_SET_TYPE(bp, zio->io_prop.zp_type); 1803 BP_SET_LEVEL(bp, zio->io_prop.zp_level); 1804 zio_buf_free(cbuf, lsize); 1805 bp->blk_birth = zio->io_txg; 1806 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1807 ASSERT(spa_feature_is_active(spa, 1808 SPA_FEATURE_EMBEDDED_DATA)); 1809 return (zio); 1810 } else { 1811 /* 1812 * Round compressed size up to the minimum allocation 1813 * size of the smallest-ashift device, and zero the 1814 * tail. This ensures that the compressed size of the 1815 * BP (and thus compressratio property) are correct, 1816 * in that we charge for the padding used to fill out 1817 * the last sector. 1818 */ 1819 size_t rounded = (size_t)zio_roundup_alloc_size(spa, 1820 psize); 1821 if (rounded >= lsize) { 1822 compress = ZIO_COMPRESS_OFF; 1823 zio_buf_free(cbuf, lsize); 1824 psize = lsize; 1825 } else { 1826 abd_t *cdata = abd_get_from_buf(cbuf, lsize); 1827 abd_take_ownership_of_buf(cdata, B_TRUE); 1828 abd_zero_off(cdata, psize, rounded - psize); 1829 psize = rounded; 1830 zio_push_transform(zio, cdata, 1831 psize, lsize, NULL); 1832 } 1833 } 1834 1835 /* 1836 * We were unable to handle this as an override bp, treat 1837 * it as a regular write I/O. 1838 */ 1839 zio->io_bp_override = NULL; 1840 *bp = zio->io_bp_orig; 1841 zio->io_pipeline = zio->io_orig_pipeline; 1842 1843 } else if ((zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) != 0 && 1844 zp->zp_type == DMU_OT_DNODE) { 1845 /* 1846 * The DMU actually relies on the zio layer's compression 1847 * to free metadnode blocks that have had all contained 1848 * dnodes freed. As a result, even when doing a raw 1849 * receive, we must check whether the block can be compressed 1850 * to a hole. 1851 */ 1852 psize = zio_compress_data(ZIO_COMPRESS_EMPTY, 1853 zio->io_abd, NULL, lsize, zp->zp_complevel); 1854 if (psize == 0 || psize >= lsize) 1855 compress = ZIO_COMPRESS_OFF; 1856 } else if (zio->io_flags & ZIO_FLAG_RAW_COMPRESS && 1857 !(zio->io_flags & ZIO_FLAG_RAW_ENCRYPT)) { 1858 /* 1859 * If we are raw receiving an encrypted dataset we should not 1860 * take this codepath because it will change the on-disk block 1861 * and decryption will fail. 1862 */ 1863 size_t rounded = MIN((size_t)zio_roundup_alloc_size(spa, psize), 1864 lsize); 1865 1866 if (rounded != psize) { 1867 abd_t *cdata = abd_alloc_linear(rounded, B_TRUE); 1868 abd_zero_off(cdata, psize, rounded - psize); 1869 abd_copy_off(cdata, zio->io_abd, 0, 0, psize); 1870 psize = rounded; 1871 zio_push_transform(zio, cdata, 1872 psize, rounded, NULL); 1873 } 1874 } else { 1875 ASSERT3U(psize, !=, 0); 1876 } 1877 1878 /* 1879 * The final pass of spa_sync() must be all rewrites, but the first 1880 * few passes offer a trade-off: allocating blocks defers convergence, 1881 * but newly allocated blocks are sequential, so they can be written 1882 * to disk faster. Therefore, we allow the first few passes of 1883 * spa_sync() to allocate new blocks, but force rewrites after that. 1884 * There should only be a handful of blocks after pass 1 in any case. 1885 */ 1886 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg && 1887 BP_GET_PSIZE(bp) == psize && 1888 pass >= zfs_sync_pass_rewrite) { 1889 VERIFY3U(psize, !=, 0); 1890 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES; 1891 1892 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages; 1893 zio->io_flags |= ZIO_FLAG_IO_REWRITE; 1894 } else { 1895 BP_ZERO(bp); 1896 zio->io_pipeline = ZIO_WRITE_PIPELINE; 1897 } 1898 1899 if (psize == 0) { 1900 if (zio->io_bp_orig.blk_birth != 0 && 1901 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) { 1902 BP_SET_LSIZE(bp, lsize); 1903 BP_SET_TYPE(bp, zp->zp_type); 1904 BP_SET_LEVEL(bp, zp->zp_level); 1905 BP_SET_BIRTH(bp, zio->io_txg, 0); 1906 } 1907 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1908 } else { 1909 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER); 1910 BP_SET_LSIZE(bp, lsize); 1911 BP_SET_TYPE(bp, zp->zp_type); 1912 BP_SET_LEVEL(bp, zp->zp_level); 1913 BP_SET_PSIZE(bp, psize); 1914 BP_SET_COMPRESS(bp, compress); 1915 BP_SET_CHECKSUM(bp, zp->zp_checksum); 1916 BP_SET_DEDUP(bp, zp->zp_dedup); 1917 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER); 1918 if (zp->zp_dedup) { 1919 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1920 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 1921 ASSERT(!zp->zp_encrypt || 1922 DMU_OT_IS_ENCRYPTED(zp->zp_type)); 1923 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE; 1924 } 1925 if (zp->zp_nopwrite) { 1926 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1927 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 1928 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE; 1929 } 1930 } 1931 return (zio); 1932 } 1933 1934 static zio_t * 1935 zio_free_bp_init(zio_t *zio) 1936 { 1937 blkptr_t *bp = zio->io_bp; 1938 1939 if (zio->io_child_type == ZIO_CHILD_LOGICAL) { 1940 if (BP_GET_DEDUP(bp)) 1941 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE; 1942 } 1943 1944 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy); 1945 1946 return (zio); 1947 } 1948 1949 /* 1950 * ========================================================================== 1951 * Execute the I/O pipeline 1952 * ========================================================================== 1953 */ 1954 1955 static void 1956 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline) 1957 { 1958 spa_t *spa = zio->io_spa; 1959 zio_type_t t = zio->io_type; 1960 int flags = (cutinline ? TQ_FRONT : 0); 1961 1962 /* 1963 * If we're a config writer or a probe, the normal issue and 1964 * interrupt threads may all be blocked waiting for the config lock. 1965 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL. 1966 */ 1967 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE)) 1968 t = ZIO_TYPE_NULL; 1969 1970 /* 1971 * A similar issue exists for the L2ARC write thread until L2ARC 2.0. 1972 */ 1973 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux) 1974 t = ZIO_TYPE_NULL; 1975 1976 /* 1977 * If this is a high priority I/O, then use the high priority taskq if 1978 * available. 1979 */ 1980 if ((zio->io_priority == ZIO_PRIORITY_NOW || 1981 zio->io_priority == ZIO_PRIORITY_SYNC_WRITE) && 1982 spa->spa_zio_taskq[t][q + 1].stqs_count != 0) 1983 q++; 1984 1985 ASSERT3U(q, <, ZIO_TASKQ_TYPES); 1986 1987 /* 1988 * NB: We are assuming that the zio can only be dispatched 1989 * to a single taskq at a time. It would be a grievous error 1990 * to dispatch the zio to another taskq at the same time. 1991 */ 1992 ASSERT(taskq_empty_ent(&zio->io_tqent)); 1993 spa_taskq_dispatch_ent(spa, t, q, zio_execute, zio, flags, 1994 &zio->io_tqent); 1995 } 1996 1997 static boolean_t 1998 zio_taskq_member(zio_t *zio, zio_taskq_type_t q) 1999 { 2000 spa_t *spa = zio->io_spa; 2001 2002 taskq_t *tq = taskq_of_curthread(); 2003 2004 for (zio_type_t t = 0; t < ZIO_TYPES; t++) { 2005 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q]; 2006 uint_t i; 2007 for (i = 0; i < tqs->stqs_count; i++) { 2008 if (tqs->stqs_taskq[i] == tq) 2009 return (B_TRUE); 2010 } 2011 } 2012 2013 return (B_FALSE); 2014 } 2015 2016 static zio_t * 2017 zio_issue_async(zio_t *zio) 2018 { 2019 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); 2020 2021 return (NULL); 2022 } 2023 2024 void 2025 zio_interrupt(void *zio) 2026 { 2027 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE); 2028 } 2029 2030 void 2031 zio_delay_interrupt(zio_t *zio) 2032 { 2033 /* 2034 * The timeout_generic() function isn't defined in userspace, so 2035 * rather than trying to implement the function, the zio delay 2036 * functionality has been disabled for userspace builds. 2037 */ 2038 2039 #ifdef _KERNEL 2040 /* 2041 * If io_target_timestamp is zero, then no delay has been registered 2042 * for this IO, thus jump to the end of this function and "skip" the 2043 * delay; issuing it directly to the zio layer. 2044 */ 2045 if (zio->io_target_timestamp != 0) { 2046 hrtime_t now = gethrtime(); 2047 2048 if (now >= zio->io_target_timestamp) { 2049 /* 2050 * This IO has already taken longer than the target 2051 * delay to complete, so we don't want to delay it 2052 * any longer; we "miss" the delay and issue it 2053 * directly to the zio layer. This is likely due to 2054 * the target latency being set to a value less than 2055 * the underlying hardware can satisfy (e.g. delay 2056 * set to 1ms, but the disks take 10ms to complete an 2057 * IO request). 2058 */ 2059 2060 DTRACE_PROBE2(zio__delay__miss, zio_t *, zio, 2061 hrtime_t, now); 2062 2063 zio_interrupt(zio); 2064 } else { 2065 taskqid_t tid; 2066 hrtime_t diff = zio->io_target_timestamp - now; 2067 clock_t expire_at_tick = ddi_get_lbolt() + 2068 NSEC_TO_TICK(diff); 2069 2070 DTRACE_PROBE3(zio__delay__hit, zio_t *, zio, 2071 hrtime_t, now, hrtime_t, diff); 2072 2073 if (NSEC_TO_TICK(diff) == 0) { 2074 /* Our delay is less than a jiffy - just spin */ 2075 zfs_sleep_until(zio->io_target_timestamp); 2076 zio_interrupt(zio); 2077 } else { 2078 /* 2079 * Use taskq_dispatch_delay() in the place of 2080 * OpenZFS's timeout_generic(). 2081 */ 2082 tid = taskq_dispatch_delay(system_taskq, 2083 zio_interrupt, zio, TQ_NOSLEEP, 2084 expire_at_tick); 2085 if (tid == TASKQID_INVALID) { 2086 /* 2087 * Couldn't allocate a task. Just 2088 * finish the zio without a delay. 2089 */ 2090 zio_interrupt(zio); 2091 } 2092 } 2093 } 2094 return; 2095 } 2096 #endif 2097 DTRACE_PROBE1(zio__delay__skip, zio_t *, zio); 2098 zio_interrupt(zio); 2099 } 2100 2101 static void 2102 zio_deadman_impl(zio_t *pio, int ziodepth) 2103 { 2104 zio_t *cio, *cio_next; 2105 zio_link_t *zl = NULL; 2106 vdev_t *vd = pio->io_vd; 2107 2108 if (zio_deadman_log_all || (vd != NULL && vd->vdev_ops->vdev_op_leaf)) { 2109 vdev_queue_t *vq = vd ? &vd->vdev_queue : NULL; 2110 zbookmark_phys_t *zb = &pio->io_bookmark; 2111 uint64_t delta = gethrtime() - pio->io_timestamp; 2112 uint64_t failmode = spa_get_deadman_failmode(pio->io_spa); 2113 2114 zfs_dbgmsg("slow zio[%d]: zio=%px timestamp=%llu " 2115 "delta=%llu queued=%llu io=%llu " 2116 "path=%s " 2117 "last=%llu type=%d " 2118 "priority=%d flags=0x%llx stage=0x%x " 2119 "pipeline=0x%x pipeline-trace=0x%x " 2120 "objset=%llu object=%llu " 2121 "level=%llu blkid=%llu " 2122 "offset=%llu size=%llu " 2123 "error=%d", 2124 ziodepth, pio, pio->io_timestamp, 2125 (u_longlong_t)delta, pio->io_delta, pio->io_delay, 2126 vd ? vd->vdev_path : "NULL", 2127 vq ? vq->vq_io_complete_ts : 0, pio->io_type, 2128 pio->io_priority, (u_longlong_t)pio->io_flags, 2129 pio->io_stage, pio->io_pipeline, pio->io_pipeline_trace, 2130 (u_longlong_t)zb->zb_objset, (u_longlong_t)zb->zb_object, 2131 (u_longlong_t)zb->zb_level, (u_longlong_t)zb->zb_blkid, 2132 (u_longlong_t)pio->io_offset, (u_longlong_t)pio->io_size, 2133 pio->io_error); 2134 (void) zfs_ereport_post(FM_EREPORT_ZFS_DEADMAN, 2135 pio->io_spa, vd, zb, pio, 0); 2136 2137 if (failmode == ZIO_FAILURE_MODE_CONTINUE && 2138 taskq_empty_ent(&pio->io_tqent)) { 2139 zio_interrupt(pio); 2140 } 2141 } 2142 2143 mutex_enter(&pio->io_lock); 2144 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) { 2145 cio_next = zio_walk_children(pio, &zl); 2146 zio_deadman_impl(cio, ziodepth + 1); 2147 } 2148 mutex_exit(&pio->io_lock); 2149 } 2150 2151 /* 2152 * Log the critical information describing this zio and all of its children 2153 * using the zfs_dbgmsg() interface then post deadman event for the ZED. 2154 */ 2155 void 2156 zio_deadman(zio_t *pio, const char *tag) 2157 { 2158 spa_t *spa = pio->io_spa; 2159 char *name = spa_name(spa); 2160 2161 if (!zfs_deadman_enabled || spa_suspended(spa)) 2162 return; 2163 2164 zio_deadman_impl(pio, 0); 2165 2166 switch (spa_get_deadman_failmode(spa)) { 2167 case ZIO_FAILURE_MODE_WAIT: 2168 zfs_dbgmsg("%s waiting for hung I/O to pool '%s'", tag, name); 2169 break; 2170 2171 case ZIO_FAILURE_MODE_CONTINUE: 2172 zfs_dbgmsg("%s restarting hung I/O for pool '%s'", tag, name); 2173 break; 2174 2175 case ZIO_FAILURE_MODE_PANIC: 2176 fm_panic("%s determined I/O to pool '%s' is hung.", tag, name); 2177 break; 2178 } 2179 } 2180 2181 /* 2182 * Execute the I/O pipeline until one of the following occurs: 2183 * (1) the I/O completes; (2) the pipeline stalls waiting for 2184 * dependent child I/Os; (3) the I/O issues, so we're waiting 2185 * for an I/O completion interrupt; (4) the I/O is delegated by 2186 * vdev-level caching or aggregation; (5) the I/O is deferred 2187 * due to vdev-level queueing; (6) the I/O is handed off to 2188 * another thread. In all cases, the pipeline stops whenever 2189 * there's no CPU work; it never burns a thread in cv_wait_io(). 2190 * 2191 * There's no locking on io_stage because there's no legitimate way 2192 * for multiple threads to be attempting to process the same I/O. 2193 */ 2194 static zio_pipe_stage_t *zio_pipeline[]; 2195 2196 /* 2197 * zio_execute() is a wrapper around the static function 2198 * __zio_execute() so that we can force __zio_execute() to be 2199 * inlined. This reduces stack overhead which is important 2200 * because __zio_execute() is called recursively in several zio 2201 * code paths. zio_execute() itself cannot be inlined because 2202 * it is externally visible. 2203 */ 2204 void 2205 zio_execute(void *zio) 2206 { 2207 fstrans_cookie_t cookie; 2208 2209 cookie = spl_fstrans_mark(); 2210 __zio_execute(zio); 2211 spl_fstrans_unmark(cookie); 2212 } 2213 2214 /* 2215 * Used to determine if in the current context the stack is sized large 2216 * enough to allow zio_execute() to be called recursively. A minimum 2217 * stack size of 16K is required to avoid needing to re-dispatch the zio. 2218 */ 2219 static boolean_t 2220 zio_execute_stack_check(zio_t *zio) 2221 { 2222 #if !defined(HAVE_LARGE_STACKS) 2223 dsl_pool_t *dp = spa_get_dsl(zio->io_spa); 2224 2225 /* Executing in txg_sync_thread() context. */ 2226 if (dp && curthread == dp->dp_tx.tx_sync_thread) 2227 return (B_TRUE); 2228 2229 /* Pool initialization outside of zio_taskq context. */ 2230 if (dp && spa_is_initializing(dp->dp_spa) && 2231 !zio_taskq_member(zio, ZIO_TASKQ_ISSUE) && 2232 !zio_taskq_member(zio, ZIO_TASKQ_ISSUE_HIGH)) 2233 return (B_TRUE); 2234 #else 2235 (void) zio; 2236 #endif /* HAVE_LARGE_STACKS */ 2237 2238 return (B_FALSE); 2239 } 2240 2241 __attribute__((always_inline)) 2242 static inline void 2243 __zio_execute(zio_t *zio) 2244 { 2245 ASSERT3U(zio->io_queued_timestamp, >, 0); 2246 2247 while (zio->io_stage < ZIO_STAGE_DONE) { 2248 enum zio_stage pipeline = zio->io_pipeline; 2249 enum zio_stage stage = zio->io_stage; 2250 2251 zio->io_executor = curthread; 2252 2253 ASSERT(!MUTEX_HELD(&zio->io_lock)); 2254 ASSERT(ISP2(stage)); 2255 ASSERT(zio->io_stall == NULL); 2256 2257 do { 2258 stage <<= 1; 2259 } while ((stage & pipeline) == 0); 2260 2261 ASSERT(stage <= ZIO_STAGE_DONE); 2262 2263 /* 2264 * If we are in interrupt context and this pipeline stage 2265 * will grab a config lock that is held across I/O, 2266 * or may wait for an I/O that needs an interrupt thread 2267 * to complete, issue async to avoid deadlock. 2268 * 2269 * For VDEV_IO_START, we cut in line so that the io will 2270 * be sent to disk promptly. 2271 */ 2272 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL && 2273 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) { 2274 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? 2275 zio_requeue_io_start_cut_in_line : B_FALSE; 2276 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); 2277 return; 2278 } 2279 2280 /* 2281 * If the current context doesn't have large enough stacks 2282 * the zio must be issued asynchronously to prevent overflow. 2283 */ 2284 if (zio_execute_stack_check(zio)) { 2285 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? 2286 zio_requeue_io_start_cut_in_line : B_FALSE; 2287 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); 2288 return; 2289 } 2290 2291 zio->io_stage = stage; 2292 zio->io_pipeline_trace |= zio->io_stage; 2293 2294 /* 2295 * The zio pipeline stage returns the next zio to execute 2296 * (typically the same as this one), or NULL if we should 2297 * stop. 2298 */ 2299 zio = zio_pipeline[highbit64(stage) - 1](zio); 2300 2301 if (zio == NULL) 2302 return; 2303 } 2304 } 2305 2306 2307 /* 2308 * ========================================================================== 2309 * Initiate I/O, either sync or async 2310 * ========================================================================== 2311 */ 2312 int 2313 zio_wait(zio_t *zio) 2314 { 2315 /* 2316 * Some routines, like zio_free_sync(), may return a NULL zio 2317 * to avoid the performance overhead of creating and then destroying 2318 * an unneeded zio. For the callers' simplicity, we accept a NULL 2319 * zio and ignore it. 2320 */ 2321 if (zio == NULL) 2322 return (0); 2323 2324 long timeout = MSEC_TO_TICK(zfs_deadman_ziotime_ms); 2325 int error; 2326 2327 ASSERT3S(zio->io_stage, ==, ZIO_STAGE_OPEN); 2328 ASSERT3P(zio->io_executor, ==, NULL); 2329 2330 zio->io_waiter = curthread; 2331 ASSERT0(zio->io_queued_timestamp); 2332 zio->io_queued_timestamp = gethrtime(); 2333 2334 __zio_execute(zio); 2335 2336 mutex_enter(&zio->io_lock); 2337 while (zio->io_executor != NULL) { 2338 error = cv_timedwait_io(&zio->io_cv, &zio->io_lock, 2339 ddi_get_lbolt() + timeout); 2340 2341 if (zfs_deadman_enabled && error == -1 && 2342 gethrtime() - zio->io_queued_timestamp > 2343 spa_deadman_ziotime(zio->io_spa)) { 2344 mutex_exit(&zio->io_lock); 2345 timeout = MSEC_TO_TICK(zfs_deadman_checktime_ms); 2346 zio_deadman(zio, FTAG); 2347 mutex_enter(&zio->io_lock); 2348 } 2349 } 2350 mutex_exit(&zio->io_lock); 2351 2352 error = zio->io_error; 2353 zio_destroy(zio); 2354 2355 return (error); 2356 } 2357 2358 void 2359 zio_nowait(zio_t *zio) 2360 { 2361 /* 2362 * See comment in zio_wait(). 2363 */ 2364 if (zio == NULL) 2365 return; 2366 2367 ASSERT3P(zio->io_executor, ==, NULL); 2368 2369 if (zio->io_child_type == ZIO_CHILD_LOGICAL && 2370 list_is_empty(&zio->io_parent_list)) { 2371 zio_t *pio; 2372 2373 /* 2374 * This is a logical async I/O with no parent to wait for it. 2375 * We add it to the spa_async_root_zio "Godfather" I/O which 2376 * will ensure they complete prior to unloading the pool. 2377 */ 2378 spa_t *spa = zio->io_spa; 2379 pio = spa->spa_async_zio_root[CPU_SEQID_UNSTABLE]; 2380 2381 zio_add_child(pio, zio); 2382 } 2383 2384 ASSERT0(zio->io_queued_timestamp); 2385 zio->io_queued_timestamp = gethrtime(); 2386 __zio_execute(zio); 2387 } 2388 2389 /* 2390 * ========================================================================== 2391 * Reexecute, cancel, or suspend/resume failed I/O 2392 * ========================================================================== 2393 */ 2394 2395 static void 2396 zio_reexecute(void *arg) 2397 { 2398 zio_t *pio = arg; 2399 zio_t *cio, *cio_next; 2400 2401 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL); 2402 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN); 2403 ASSERT(pio->io_gang_leader == NULL); 2404 ASSERT(pio->io_gang_tree == NULL); 2405 2406 pio->io_flags = pio->io_orig_flags; 2407 pio->io_stage = pio->io_orig_stage; 2408 pio->io_pipeline = pio->io_orig_pipeline; 2409 pio->io_reexecute = 0; 2410 pio->io_flags |= ZIO_FLAG_REEXECUTED; 2411 pio->io_pipeline_trace = 0; 2412 pio->io_error = 0; 2413 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 2414 pio->io_state[w] = 0; 2415 for (int c = 0; c < ZIO_CHILD_TYPES; c++) 2416 pio->io_child_error[c] = 0; 2417 2418 if (IO_IS_ALLOCATING(pio)) 2419 BP_ZERO(pio->io_bp); 2420 2421 /* 2422 * As we reexecute pio's children, new children could be created. 2423 * New children go to the head of pio's io_child_list, however, 2424 * so we will (correctly) not reexecute them. The key is that 2425 * the remainder of pio's io_child_list, from 'cio_next' onward, 2426 * cannot be affected by any side effects of reexecuting 'cio'. 2427 */ 2428 zio_link_t *zl = NULL; 2429 mutex_enter(&pio->io_lock); 2430 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) { 2431 cio_next = zio_walk_children(pio, &zl); 2432 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 2433 pio->io_children[cio->io_child_type][w]++; 2434 mutex_exit(&pio->io_lock); 2435 zio_reexecute(cio); 2436 mutex_enter(&pio->io_lock); 2437 } 2438 mutex_exit(&pio->io_lock); 2439 2440 /* 2441 * Now that all children have been reexecuted, execute the parent. 2442 * We don't reexecute "The Godfather" I/O here as it's the 2443 * responsibility of the caller to wait on it. 2444 */ 2445 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) { 2446 pio->io_queued_timestamp = gethrtime(); 2447 __zio_execute(pio); 2448 } 2449 } 2450 2451 void 2452 zio_suspend(spa_t *spa, zio_t *zio, zio_suspend_reason_t reason) 2453 { 2454 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC) 2455 fm_panic("Pool '%s' has encountered an uncorrectable I/O " 2456 "failure and the failure mode property for this pool " 2457 "is set to panic.", spa_name(spa)); 2458 2459 cmn_err(CE_WARN, "Pool '%s' has encountered an uncorrectable I/O " 2460 "failure and has been suspended.\n", spa_name(spa)); 2461 2462 (void) zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, 2463 NULL, NULL, 0); 2464 2465 mutex_enter(&spa->spa_suspend_lock); 2466 2467 if (spa->spa_suspend_zio_root == NULL) 2468 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL, 2469 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | 2470 ZIO_FLAG_GODFATHER); 2471 2472 spa->spa_suspended = reason; 2473 2474 if (zio != NULL) { 2475 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); 2476 ASSERT(zio != spa->spa_suspend_zio_root); 2477 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2478 ASSERT(zio_unique_parent(zio) == NULL); 2479 ASSERT(zio->io_stage == ZIO_STAGE_DONE); 2480 zio_add_child(spa->spa_suspend_zio_root, zio); 2481 } 2482 2483 mutex_exit(&spa->spa_suspend_lock); 2484 } 2485 2486 int 2487 zio_resume(spa_t *spa) 2488 { 2489 zio_t *pio; 2490 2491 /* 2492 * Reexecute all previously suspended i/o. 2493 */ 2494 mutex_enter(&spa->spa_suspend_lock); 2495 spa->spa_suspended = ZIO_SUSPEND_NONE; 2496 cv_broadcast(&spa->spa_suspend_cv); 2497 pio = spa->spa_suspend_zio_root; 2498 spa->spa_suspend_zio_root = NULL; 2499 mutex_exit(&spa->spa_suspend_lock); 2500 2501 if (pio == NULL) 2502 return (0); 2503 2504 zio_reexecute(pio); 2505 return (zio_wait(pio)); 2506 } 2507 2508 void 2509 zio_resume_wait(spa_t *spa) 2510 { 2511 mutex_enter(&spa->spa_suspend_lock); 2512 while (spa_suspended(spa)) 2513 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock); 2514 mutex_exit(&spa->spa_suspend_lock); 2515 } 2516 2517 /* 2518 * ========================================================================== 2519 * Gang blocks. 2520 * 2521 * A gang block is a collection of small blocks that looks to the DMU 2522 * like one large block. When zio_dva_allocate() cannot find a block 2523 * of the requested size, due to either severe fragmentation or the pool 2524 * being nearly full, it calls zio_write_gang_block() to construct the 2525 * block from smaller fragments. 2526 * 2527 * A gang block consists of a gang header (zio_gbh_phys_t) and up to 2528 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like 2529 * an indirect block: it's an array of block pointers. It consumes 2530 * only one sector and hence is allocatable regardless of fragmentation. 2531 * The gang header's bps point to its gang members, which hold the data. 2532 * 2533 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg> 2534 * as the verifier to ensure uniqueness of the SHA256 checksum. 2535 * Critically, the gang block bp's blk_cksum is the checksum of the data, 2536 * not the gang header. This ensures that data block signatures (needed for 2537 * deduplication) are independent of how the block is physically stored. 2538 * 2539 * Gang blocks can be nested: a gang member may itself be a gang block. 2540 * Thus every gang block is a tree in which root and all interior nodes are 2541 * gang headers, and the leaves are normal blocks that contain user data. 2542 * The root of the gang tree is called the gang leader. 2543 * 2544 * To perform any operation (read, rewrite, free, claim) on a gang block, 2545 * zio_gang_assemble() first assembles the gang tree (minus data leaves) 2546 * in the io_gang_tree field of the original logical i/o by recursively 2547 * reading the gang leader and all gang headers below it. This yields 2548 * an in-core tree containing the contents of every gang header and the 2549 * bps for every constituent of the gang block. 2550 * 2551 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree 2552 * and invokes a callback on each bp. To free a gang block, zio_gang_issue() 2553 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp. 2554 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim(). 2555 * zio_read_gang() is a wrapper around zio_read() that omits reading gang 2556 * headers, since we already have those in io_gang_tree. zio_rewrite_gang() 2557 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite() 2558 * of the gang header plus zio_checksum_compute() of the data to update the 2559 * gang header's blk_cksum as described above. 2560 * 2561 * The two-phase assemble/issue model solves the problem of partial failure -- 2562 * what if you'd freed part of a gang block but then couldn't read the 2563 * gang header for another part? Assembling the entire gang tree first 2564 * ensures that all the necessary gang header I/O has succeeded before 2565 * starting the actual work of free, claim, or write. Once the gang tree 2566 * is assembled, free and claim are in-memory operations that cannot fail. 2567 * 2568 * In the event that a gang write fails, zio_dva_unallocate() walks the 2569 * gang tree to immediately free (i.e. insert back into the space map) 2570 * everything we've allocated. This ensures that we don't get ENOSPC 2571 * errors during repeated suspend/resume cycles due to a flaky device. 2572 * 2573 * Gang rewrites only happen during sync-to-convergence. If we can't assemble 2574 * the gang tree, we won't modify the block, so we can safely defer the free 2575 * (knowing that the block is still intact). If we *can* assemble the gang 2576 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free 2577 * each constituent bp and we can allocate a new block on the next sync pass. 2578 * 2579 * In all cases, the gang tree allows complete recovery from partial failure. 2580 * ========================================================================== 2581 */ 2582 2583 static void 2584 zio_gang_issue_func_done(zio_t *zio) 2585 { 2586 abd_free(zio->io_abd); 2587 } 2588 2589 static zio_t * 2590 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, 2591 uint64_t offset) 2592 { 2593 if (gn != NULL) 2594 return (pio); 2595 2596 return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset), 2597 BP_GET_PSIZE(bp), zio_gang_issue_func_done, 2598 NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), 2599 &pio->io_bookmark)); 2600 } 2601 2602 static zio_t * 2603 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, 2604 uint64_t offset) 2605 { 2606 zio_t *zio; 2607 2608 if (gn != NULL) { 2609 abd_t *gbh_abd = 2610 abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE); 2611 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, 2612 gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL, 2613 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), 2614 &pio->io_bookmark); 2615 /* 2616 * As we rewrite each gang header, the pipeline will compute 2617 * a new gang block header checksum for it; but no one will 2618 * compute a new data checksum, so we do that here. The one 2619 * exception is the gang leader: the pipeline already computed 2620 * its data checksum because that stage precedes gang assembly. 2621 * (Presently, nothing actually uses interior data checksums; 2622 * this is just good hygiene.) 2623 */ 2624 if (gn != pio->io_gang_leader->io_gang_tree) { 2625 abd_t *buf = abd_get_offset(data, offset); 2626 2627 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp), 2628 buf, BP_GET_PSIZE(bp)); 2629 2630 abd_free(buf); 2631 } 2632 /* 2633 * If we are here to damage data for testing purposes, 2634 * leave the GBH alone so that we can detect the damage. 2635 */ 2636 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE) 2637 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; 2638 } else { 2639 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, 2640 abd_get_offset(data, offset), BP_GET_PSIZE(bp), 2641 zio_gang_issue_func_done, NULL, pio->io_priority, 2642 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 2643 } 2644 2645 return (zio); 2646 } 2647 2648 static zio_t * 2649 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, 2650 uint64_t offset) 2651 { 2652 (void) gn, (void) data, (void) offset; 2653 2654 zio_t *zio = zio_free_sync(pio, pio->io_spa, pio->io_txg, bp, 2655 ZIO_GANG_CHILD_FLAGS(pio)); 2656 if (zio == NULL) { 2657 zio = zio_null(pio, pio->io_spa, 2658 NULL, NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)); 2659 } 2660 return (zio); 2661 } 2662 2663 static zio_t * 2664 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, 2665 uint64_t offset) 2666 { 2667 (void) gn, (void) data, (void) offset; 2668 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp, 2669 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio))); 2670 } 2671 2672 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = { 2673 NULL, 2674 zio_read_gang, 2675 zio_rewrite_gang, 2676 zio_free_gang, 2677 zio_claim_gang, 2678 NULL 2679 }; 2680 2681 static void zio_gang_tree_assemble_done(zio_t *zio); 2682 2683 static zio_gang_node_t * 2684 zio_gang_node_alloc(zio_gang_node_t **gnpp) 2685 { 2686 zio_gang_node_t *gn; 2687 2688 ASSERT(*gnpp == NULL); 2689 2690 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP); 2691 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE); 2692 *gnpp = gn; 2693 2694 return (gn); 2695 } 2696 2697 static void 2698 zio_gang_node_free(zio_gang_node_t **gnpp) 2699 { 2700 zio_gang_node_t *gn = *gnpp; 2701 2702 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) 2703 ASSERT(gn->gn_child[g] == NULL); 2704 2705 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE); 2706 kmem_free(gn, sizeof (*gn)); 2707 *gnpp = NULL; 2708 } 2709 2710 static void 2711 zio_gang_tree_free(zio_gang_node_t **gnpp) 2712 { 2713 zio_gang_node_t *gn = *gnpp; 2714 2715 if (gn == NULL) 2716 return; 2717 2718 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) 2719 zio_gang_tree_free(&gn->gn_child[g]); 2720 2721 zio_gang_node_free(gnpp); 2722 } 2723 2724 static void 2725 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp) 2726 { 2727 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp); 2728 abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE); 2729 2730 ASSERT(gio->io_gang_leader == gio); 2731 ASSERT(BP_IS_GANG(bp)); 2732 2733 zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE, 2734 zio_gang_tree_assemble_done, gn, gio->io_priority, 2735 ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark)); 2736 } 2737 2738 static void 2739 zio_gang_tree_assemble_done(zio_t *zio) 2740 { 2741 zio_t *gio = zio->io_gang_leader; 2742 zio_gang_node_t *gn = zio->io_private; 2743 blkptr_t *bp = zio->io_bp; 2744 2745 ASSERT(gio == zio_unique_parent(zio)); 2746 ASSERT(list_is_empty(&zio->io_child_list)); 2747 2748 if (zio->io_error) 2749 return; 2750 2751 /* this ABD was created from a linear buf in zio_gang_tree_assemble */ 2752 if (BP_SHOULD_BYTESWAP(bp)) 2753 byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size); 2754 2755 ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh); 2756 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE); 2757 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); 2758 2759 abd_free(zio->io_abd); 2760 2761 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 2762 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; 2763 if (!BP_IS_GANG(gbp)) 2764 continue; 2765 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]); 2766 } 2767 } 2768 2769 static void 2770 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data, 2771 uint64_t offset) 2772 { 2773 zio_t *gio = pio->io_gang_leader; 2774 zio_t *zio; 2775 2776 ASSERT(BP_IS_GANG(bp) == !!gn); 2777 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp)); 2778 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree); 2779 2780 /* 2781 * If you're a gang header, your data is in gn->gn_gbh. 2782 * If you're a gang member, your data is in 'data' and gn == NULL. 2783 */ 2784 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset); 2785 2786 if (gn != NULL) { 2787 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); 2788 2789 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 2790 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; 2791 if (BP_IS_HOLE(gbp)) 2792 continue; 2793 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data, 2794 offset); 2795 offset += BP_GET_PSIZE(gbp); 2796 } 2797 } 2798 2799 if (gn == gio->io_gang_tree) 2800 ASSERT3U(gio->io_size, ==, offset); 2801 2802 if (zio != pio) 2803 zio_nowait(zio); 2804 } 2805 2806 static zio_t * 2807 zio_gang_assemble(zio_t *zio) 2808 { 2809 blkptr_t *bp = zio->io_bp; 2810 2811 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL); 2812 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 2813 2814 zio->io_gang_leader = zio; 2815 2816 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree); 2817 2818 return (zio); 2819 } 2820 2821 static zio_t * 2822 zio_gang_issue(zio_t *zio) 2823 { 2824 blkptr_t *bp = zio->io_bp; 2825 2826 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT, ZIO_WAIT_DONE)) { 2827 return (NULL); 2828 } 2829 2830 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio); 2831 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 2832 2833 if (zio->io_child_error[ZIO_CHILD_GANG] == 0) 2834 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd, 2835 0); 2836 else 2837 zio_gang_tree_free(&zio->io_gang_tree); 2838 2839 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2840 2841 return (zio); 2842 } 2843 2844 static void 2845 zio_write_gang_member_ready(zio_t *zio) 2846 { 2847 zio_t *pio = zio_unique_parent(zio); 2848 dva_t *cdva = zio->io_bp->blk_dva; 2849 dva_t *pdva = pio->io_bp->blk_dva; 2850 uint64_t asize; 2851 zio_t *gio __maybe_unused = zio->io_gang_leader; 2852 2853 if (BP_IS_HOLE(zio->io_bp)) 2854 return; 2855 2856 ASSERT(BP_IS_HOLE(&zio->io_bp_orig)); 2857 2858 ASSERT(zio->io_child_type == ZIO_CHILD_GANG); 2859 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies); 2860 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp)); 2861 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp)); 2862 VERIFY3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp)); 2863 2864 mutex_enter(&pio->io_lock); 2865 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) { 2866 ASSERT(DVA_GET_GANG(&pdva[d])); 2867 asize = DVA_GET_ASIZE(&pdva[d]); 2868 asize += DVA_GET_ASIZE(&cdva[d]); 2869 DVA_SET_ASIZE(&pdva[d], asize); 2870 } 2871 mutex_exit(&pio->io_lock); 2872 } 2873 2874 static void 2875 zio_write_gang_done(zio_t *zio) 2876 { 2877 /* 2878 * The io_abd field will be NULL for a zio with no data. The io_flags 2879 * will initially have the ZIO_FLAG_NODATA bit flag set, but we can't 2880 * check for it here as it is cleared in zio_ready. 2881 */ 2882 if (zio->io_abd != NULL) 2883 abd_free(zio->io_abd); 2884 } 2885 2886 static zio_t * 2887 zio_write_gang_block(zio_t *pio, metaslab_class_t *mc) 2888 { 2889 spa_t *spa = pio->io_spa; 2890 blkptr_t *bp = pio->io_bp; 2891 zio_t *gio = pio->io_gang_leader; 2892 zio_t *zio; 2893 zio_gang_node_t *gn, **gnpp; 2894 zio_gbh_phys_t *gbh; 2895 abd_t *gbh_abd; 2896 uint64_t txg = pio->io_txg; 2897 uint64_t resid = pio->io_size; 2898 uint64_t lsize; 2899 int copies = gio->io_prop.zp_copies; 2900 zio_prop_t zp; 2901 int error; 2902 boolean_t has_data = !(pio->io_flags & ZIO_FLAG_NODATA); 2903 2904 /* 2905 * If one copy was requested, store 2 copies of the GBH, so that we 2906 * can still traverse all the data (e.g. to free or scrub) even if a 2907 * block is damaged. Note that we can't store 3 copies of the GBH in 2908 * all cases, e.g. with encryption, which uses DVA[2] for the IV+salt. 2909 */ 2910 int gbh_copies = copies; 2911 if (gbh_copies == 1) { 2912 gbh_copies = MIN(2, spa_max_replication(spa)); 2913 } 2914 2915 int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER; 2916 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { 2917 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); 2918 ASSERT(has_data); 2919 2920 flags |= METASLAB_ASYNC_ALLOC; 2921 VERIFY(zfs_refcount_held(&mc->mc_allocator[pio->io_allocator]. 2922 mca_alloc_slots, pio)); 2923 2924 /* 2925 * The logical zio has already placed a reservation for 2926 * 'copies' allocation slots but gang blocks may require 2927 * additional copies. These additional copies 2928 * (i.e. gbh_copies - copies) are guaranteed to succeed 2929 * since metaslab_class_throttle_reserve() always allows 2930 * additional reservations for gang blocks. 2931 */ 2932 VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies, 2933 pio->io_allocator, pio, flags)); 2934 } 2935 2936 error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE, 2937 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags, 2938 &pio->io_alloc_list, pio, pio->io_allocator); 2939 if (error) { 2940 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { 2941 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); 2942 ASSERT(has_data); 2943 2944 /* 2945 * If we failed to allocate the gang block header then 2946 * we remove any additional allocation reservations that 2947 * we placed here. The original reservation will 2948 * be removed when the logical I/O goes to the ready 2949 * stage. 2950 */ 2951 metaslab_class_throttle_unreserve(mc, 2952 gbh_copies - copies, pio->io_allocator, pio); 2953 } 2954 2955 pio->io_error = error; 2956 return (pio); 2957 } 2958 2959 if (pio == gio) { 2960 gnpp = &gio->io_gang_tree; 2961 } else { 2962 gnpp = pio->io_private; 2963 ASSERT(pio->io_ready == zio_write_gang_member_ready); 2964 } 2965 2966 gn = zio_gang_node_alloc(gnpp); 2967 gbh = gn->gn_gbh; 2968 memset(gbh, 0, SPA_GANGBLOCKSIZE); 2969 gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE); 2970 2971 /* 2972 * Create the gang header. 2973 */ 2974 zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE, 2975 zio_write_gang_done, NULL, pio->io_priority, 2976 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 2977 2978 /* 2979 * Create and nowait the gang children. 2980 */ 2981 for (int g = 0; resid != 0; resid -= lsize, g++) { 2982 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g), 2983 SPA_MINBLOCKSIZE); 2984 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid); 2985 2986 zp.zp_checksum = gio->io_prop.zp_checksum; 2987 zp.zp_compress = ZIO_COMPRESS_OFF; 2988 zp.zp_complevel = gio->io_prop.zp_complevel; 2989 zp.zp_type = DMU_OT_NONE; 2990 zp.zp_level = 0; 2991 zp.zp_copies = gio->io_prop.zp_copies; 2992 zp.zp_dedup = B_FALSE; 2993 zp.zp_dedup_verify = B_FALSE; 2994 zp.zp_nopwrite = B_FALSE; 2995 zp.zp_encrypt = gio->io_prop.zp_encrypt; 2996 zp.zp_byteorder = gio->io_prop.zp_byteorder; 2997 memset(zp.zp_salt, 0, ZIO_DATA_SALT_LEN); 2998 memset(zp.zp_iv, 0, ZIO_DATA_IV_LEN); 2999 memset(zp.zp_mac, 0, ZIO_DATA_MAC_LEN); 3000 3001 zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g], 3002 has_data ? abd_get_offset(pio->io_abd, pio->io_size - 3003 resid) : NULL, lsize, lsize, &zp, 3004 zio_write_gang_member_ready, NULL, 3005 zio_write_gang_done, &gn->gn_child[g], pio->io_priority, 3006 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 3007 3008 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { 3009 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); 3010 ASSERT(has_data); 3011 3012 /* 3013 * Gang children won't throttle but we should 3014 * account for their work, so reserve an allocation 3015 * slot for them here. 3016 */ 3017 VERIFY(metaslab_class_throttle_reserve(mc, 3018 zp.zp_copies, cio->io_allocator, cio, flags)); 3019 } 3020 zio_nowait(cio); 3021 } 3022 3023 /* 3024 * Set pio's pipeline to just wait for zio to finish. 3025 */ 3026 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 3027 3028 zio_nowait(zio); 3029 3030 return (pio); 3031 } 3032 3033 /* 3034 * The zio_nop_write stage in the pipeline determines if allocating a 3035 * new bp is necessary. The nopwrite feature can handle writes in 3036 * either syncing or open context (i.e. zil writes) and as a result is 3037 * mutually exclusive with dedup. 3038 * 3039 * By leveraging a cryptographically secure checksum, such as SHA256, we 3040 * can compare the checksums of the new data and the old to determine if 3041 * allocating a new block is required. Note that our requirements for 3042 * cryptographic strength are fairly weak: there can't be any accidental 3043 * hash collisions, but we don't need to be secure against intentional 3044 * (malicious) collisions. To trigger a nopwrite, you have to be able 3045 * to write the file to begin with, and triggering an incorrect (hash 3046 * collision) nopwrite is no worse than simply writing to the file. 3047 * That said, there are no known attacks against the checksum algorithms 3048 * used for nopwrite, assuming that the salt and the checksums 3049 * themselves remain secret. 3050 */ 3051 static zio_t * 3052 zio_nop_write(zio_t *zio) 3053 { 3054 blkptr_t *bp = zio->io_bp; 3055 blkptr_t *bp_orig = &zio->io_bp_orig; 3056 zio_prop_t *zp = &zio->io_prop; 3057 3058 ASSERT(BP_IS_HOLE(bp)); 3059 ASSERT(BP_GET_LEVEL(bp) == 0); 3060 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 3061 ASSERT(zp->zp_nopwrite); 3062 ASSERT(!zp->zp_dedup); 3063 ASSERT(zio->io_bp_override == NULL); 3064 ASSERT(IO_IS_ALLOCATING(zio)); 3065 3066 /* 3067 * Check to see if the original bp and the new bp have matching 3068 * characteristics (i.e. same checksum, compression algorithms, etc). 3069 * If they don't then just continue with the pipeline which will 3070 * allocate a new bp. 3071 */ 3072 if (BP_IS_HOLE(bp_orig) || 3073 !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags & 3074 ZCHECKSUM_FLAG_NOPWRITE) || 3075 BP_IS_ENCRYPTED(bp) || BP_IS_ENCRYPTED(bp_orig) || 3076 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) || 3077 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) || 3078 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) || 3079 zp->zp_copies != BP_GET_NDVAS(bp_orig)) 3080 return (zio); 3081 3082 /* 3083 * If the checksums match then reset the pipeline so that we 3084 * avoid allocating a new bp and issuing any I/O. 3085 */ 3086 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) { 3087 ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags & 3088 ZCHECKSUM_FLAG_NOPWRITE); 3089 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig)); 3090 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig)); 3091 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF); 3092 ASSERT3U(bp->blk_prop, ==, bp_orig->blk_prop); 3093 3094 /* 3095 * If we're overwriting a block that is currently on an 3096 * indirect vdev, then ignore the nopwrite request and 3097 * allow a new block to be allocated on a concrete vdev. 3098 */ 3099 spa_config_enter(zio->io_spa, SCL_VDEV, FTAG, RW_READER); 3100 for (int d = 0; d < BP_GET_NDVAS(bp_orig); d++) { 3101 vdev_t *tvd = vdev_lookup_top(zio->io_spa, 3102 DVA_GET_VDEV(&bp_orig->blk_dva[d])); 3103 if (tvd->vdev_ops == &vdev_indirect_ops) { 3104 spa_config_exit(zio->io_spa, SCL_VDEV, FTAG); 3105 return (zio); 3106 } 3107 } 3108 spa_config_exit(zio->io_spa, SCL_VDEV, FTAG); 3109 3110 *bp = *bp_orig; 3111 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 3112 zio->io_flags |= ZIO_FLAG_NOPWRITE; 3113 } 3114 3115 return (zio); 3116 } 3117 3118 /* 3119 * ========================================================================== 3120 * Block Reference Table 3121 * ========================================================================== 3122 */ 3123 static zio_t * 3124 zio_brt_free(zio_t *zio) 3125 { 3126 blkptr_t *bp; 3127 3128 bp = zio->io_bp; 3129 3130 if (BP_GET_LEVEL(bp) > 0 || 3131 BP_IS_METADATA(bp) || 3132 !brt_maybe_exists(zio->io_spa, bp)) { 3133 return (zio); 3134 } 3135 3136 if (!brt_entry_decref(zio->io_spa, bp)) { 3137 /* 3138 * This isn't the last reference, so we cannot free 3139 * the data yet. 3140 */ 3141 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 3142 } 3143 3144 return (zio); 3145 } 3146 3147 /* 3148 * ========================================================================== 3149 * Dedup 3150 * ========================================================================== 3151 */ 3152 static void 3153 zio_ddt_child_read_done(zio_t *zio) 3154 { 3155 blkptr_t *bp = zio->io_bp; 3156 ddt_entry_t *dde = zio->io_private; 3157 ddt_phys_t *ddp; 3158 zio_t *pio = zio_unique_parent(zio); 3159 3160 mutex_enter(&pio->io_lock); 3161 ddp = ddt_phys_select(dde, bp); 3162 if (zio->io_error == 0) 3163 ddt_phys_clear(ddp); /* this ddp doesn't need repair */ 3164 3165 if (zio->io_error == 0 && dde->dde_repair_abd == NULL) 3166 dde->dde_repair_abd = zio->io_abd; 3167 else 3168 abd_free(zio->io_abd); 3169 mutex_exit(&pio->io_lock); 3170 } 3171 3172 static zio_t * 3173 zio_ddt_read_start(zio_t *zio) 3174 { 3175 blkptr_t *bp = zio->io_bp; 3176 3177 ASSERT(BP_GET_DEDUP(bp)); 3178 ASSERT(BP_GET_PSIZE(bp) == zio->io_size); 3179 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 3180 3181 if (zio->io_child_error[ZIO_CHILD_DDT]) { 3182 ddt_t *ddt = ddt_select(zio->io_spa, bp); 3183 ddt_entry_t *dde = ddt_repair_start(ddt, bp); 3184 ddt_phys_t *ddp = dde->dde_phys; 3185 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp); 3186 blkptr_t blk; 3187 3188 ASSERT(zio->io_vsd == NULL); 3189 zio->io_vsd = dde; 3190 3191 if (ddp_self == NULL) 3192 return (zio); 3193 3194 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) { 3195 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self) 3196 continue; 3197 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp, 3198 &blk); 3199 zio_nowait(zio_read(zio, zio->io_spa, &blk, 3200 abd_alloc_for_io(zio->io_size, B_TRUE), 3201 zio->io_size, zio_ddt_child_read_done, dde, 3202 zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) | 3203 ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark)); 3204 } 3205 return (zio); 3206 } 3207 3208 zio_nowait(zio_read(zio, zio->io_spa, bp, 3209 zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority, 3210 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark)); 3211 3212 return (zio); 3213 } 3214 3215 static zio_t * 3216 zio_ddt_read_done(zio_t *zio) 3217 { 3218 blkptr_t *bp = zio->io_bp; 3219 3220 if (zio_wait_for_children(zio, ZIO_CHILD_DDT_BIT, ZIO_WAIT_DONE)) { 3221 return (NULL); 3222 } 3223 3224 ASSERT(BP_GET_DEDUP(bp)); 3225 ASSERT(BP_GET_PSIZE(bp) == zio->io_size); 3226 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 3227 3228 if (zio->io_child_error[ZIO_CHILD_DDT]) { 3229 ddt_t *ddt = ddt_select(zio->io_spa, bp); 3230 ddt_entry_t *dde = zio->io_vsd; 3231 if (ddt == NULL) { 3232 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE); 3233 return (zio); 3234 } 3235 if (dde == NULL) { 3236 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1; 3237 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); 3238 return (NULL); 3239 } 3240 if (dde->dde_repair_abd != NULL) { 3241 abd_copy(zio->io_abd, dde->dde_repair_abd, 3242 zio->io_size); 3243 zio->io_child_error[ZIO_CHILD_DDT] = 0; 3244 } 3245 ddt_repair_done(ddt, dde); 3246 zio->io_vsd = NULL; 3247 } 3248 3249 ASSERT(zio->io_vsd == NULL); 3250 3251 return (zio); 3252 } 3253 3254 static boolean_t 3255 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde) 3256 { 3257 spa_t *spa = zio->io_spa; 3258 boolean_t do_raw = !!(zio->io_flags & ZIO_FLAG_RAW); 3259 3260 ASSERT(!(zio->io_bp_override && do_raw)); 3261 3262 /* 3263 * Note: we compare the original data, not the transformed data, 3264 * because when zio->io_bp is an override bp, we will not have 3265 * pushed the I/O transforms. That's an important optimization 3266 * because otherwise we'd compress/encrypt all dmu_sync() data twice. 3267 * However, we should never get a raw, override zio so in these 3268 * cases we can compare the io_abd directly. This is useful because 3269 * it allows us to do dedup verification even if we don't have access 3270 * to the original data (for instance, if the encryption keys aren't 3271 * loaded). 3272 */ 3273 3274 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { 3275 zio_t *lio = dde->dde_lead_zio[p]; 3276 3277 if (lio != NULL && do_raw) { 3278 return (lio->io_size != zio->io_size || 3279 abd_cmp(zio->io_abd, lio->io_abd) != 0); 3280 } else if (lio != NULL) { 3281 return (lio->io_orig_size != zio->io_orig_size || 3282 abd_cmp(zio->io_orig_abd, lio->io_orig_abd) != 0); 3283 } 3284 } 3285 3286 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { 3287 ddt_phys_t *ddp = &dde->dde_phys[p]; 3288 3289 if (ddp->ddp_phys_birth != 0 && do_raw) { 3290 blkptr_t blk = *zio->io_bp; 3291 uint64_t psize; 3292 abd_t *tmpabd; 3293 int error; 3294 3295 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth); 3296 psize = BP_GET_PSIZE(&blk); 3297 3298 if (psize != zio->io_size) 3299 return (B_TRUE); 3300 3301 ddt_exit(ddt); 3302 3303 tmpabd = abd_alloc_for_io(psize, B_TRUE); 3304 3305 error = zio_wait(zio_read(NULL, spa, &blk, tmpabd, 3306 psize, NULL, NULL, ZIO_PRIORITY_SYNC_READ, 3307 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | 3308 ZIO_FLAG_RAW, &zio->io_bookmark)); 3309 3310 if (error == 0) { 3311 if (abd_cmp(tmpabd, zio->io_abd) != 0) 3312 error = SET_ERROR(ENOENT); 3313 } 3314 3315 abd_free(tmpabd); 3316 ddt_enter(ddt); 3317 return (error != 0); 3318 } else if (ddp->ddp_phys_birth != 0) { 3319 arc_buf_t *abuf = NULL; 3320 arc_flags_t aflags = ARC_FLAG_WAIT; 3321 blkptr_t blk = *zio->io_bp; 3322 int error; 3323 3324 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth); 3325 3326 if (BP_GET_LSIZE(&blk) != zio->io_orig_size) 3327 return (B_TRUE); 3328 3329 ddt_exit(ddt); 3330 3331 error = arc_read(NULL, spa, &blk, 3332 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ, 3333 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 3334 &aflags, &zio->io_bookmark); 3335 3336 if (error == 0) { 3337 if (abd_cmp_buf(zio->io_orig_abd, abuf->b_data, 3338 zio->io_orig_size) != 0) 3339 error = SET_ERROR(ENOENT); 3340 arc_buf_destroy(abuf, &abuf); 3341 } 3342 3343 ddt_enter(ddt); 3344 return (error != 0); 3345 } 3346 } 3347 3348 return (B_FALSE); 3349 } 3350 3351 static void 3352 zio_ddt_child_write_ready(zio_t *zio) 3353 { 3354 int p = zio->io_prop.zp_copies; 3355 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); 3356 ddt_entry_t *dde = zio->io_private; 3357 ddt_phys_t *ddp = &dde->dde_phys[p]; 3358 zio_t *pio; 3359 3360 if (zio->io_error) 3361 return; 3362 3363 ddt_enter(ddt); 3364 3365 ASSERT(dde->dde_lead_zio[p] == zio); 3366 3367 ddt_phys_fill(ddp, zio->io_bp); 3368 3369 zio_link_t *zl = NULL; 3370 while ((pio = zio_walk_parents(zio, &zl)) != NULL) 3371 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg); 3372 3373 ddt_exit(ddt); 3374 } 3375 3376 static void 3377 zio_ddt_child_write_done(zio_t *zio) 3378 { 3379 int p = zio->io_prop.zp_copies; 3380 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); 3381 ddt_entry_t *dde = zio->io_private; 3382 ddt_phys_t *ddp = &dde->dde_phys[p]; 3383 3384 ddt_enter(ddt); 3385 3386 ASSERT(ddp->ddp_refcnt == 0); 3387 ASSERT(dde->dde_lead_zio[p] == zio); 3388 dde->dde_lead_zio[p] = NULL; 3389 3390 if (zio->io_error == 0) { 3391 zio_link_t *zl = NULL; 3392 while (zio_walk_parents(zio, &zl) != NULL) 3393 ddt_phys_addref(ddp); 3394 } else { 3395 ddt_phys_clear(ddp); 3396 } 3397 3398 ddt_exit(ddt); 3399 } 3400 3401 static zio_t * 3402 zio_ddt_write(zio_t *zio) 3403 { 3404 spa_t *spa = zio->io_spa; 3405 blkptr_t *bp = zio->io_bp; 3406 uint64_t txg = zio->io_txg; 3407 zio_prop_t *zp = &zio->io_prop; 3408 int p = zp->zp_copies; 3409 zio_t *cio = NULL; 3410 ddt_t *ddt = ddt_select(spa, bp); 3411 ddt_entry_t *dde; 3412 ddt_phys_t *ddp; 3413 3414 ASSERT(BP_GET_DEDUP(bp)); 3415 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum); 3416 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override); 3417 ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW))); 3418 3419 ddt_enter(ddt); 3420 dde = ddt_lookup(ddt, bp, B_TRUE); 3421 ddp = &dde->dde_phys[p]; 3422 3423 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) { 3424 /* 3425 * If we're using a weak checksum, upgrade to a strong checksum 3426 * and try again. If we're already using a strong checksum, 3427 * we can't resolve it, so just convert to an ordinary write. 3428 * (And automatically e-mail a paper to Nature?) 3429 */ 3430 if (!(zio_checksum_table[zp->zp_checksum].ci_flags & 3431 ZCHECKSUM_FLAG_DEDUP)) { 3432 zp->zp_checksum = spa_dedup_checksum(spa); 3433 zio_pop_transforms(zio); 3434 zio->io_stage = ZIO_STAGE_OPEN; 3435 BP_ZERO(bp); 3436 } else { 3437 zp->zp_dedup = B_FALSE; 3438 BP_SET_DEDUP(bp, B_FALSE); 3439 } 3440 ASSERT(!BP_GET_DEDUP(bp)); 3441 zio->io_pipeline = ZIO_WRITE_PIPELINE; 3442 ddt_exit(ddt); 3443 return (zio); 3444 } 3445 3446 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) { 3447 if (ddp->ddp_phys_birth != 0) 3448 ddt_bp_fill(ddp, bp, txg); 3449 if (dde->dde_lead_zio[p] != NULL) 3450 zio_add_child(zio, dde->dde_lead_zio[p]); 3451 else 3452 ddt_phys_addref(ddp); 3453 } else if (zio->io_bp_override) { 3454 ASSERT(bp->blk_birth == txg); 3455 ASSERT(BP_EQUAL(bp, zio->io_bp_override)); 3456 ddt_phys_fill(ddp, bp); 3457 ddt_phys_addref(ddp); 3458 } else { 3459 cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd, 3460 zio->io_orig_size, zio->io_orig_size, zp, 3461 zio_ddt_child_write_ready, NULL, 3462 zio_ddt_child_write_done, dde, zio->io_priority, 3463 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); 3464 3465 zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL); 3466 dde->dde_lead_zio[p] = cio; 3467 } 3468 3469 ddt_exit(ddt); 3470 3471 zio_nowait(cio); 3472 3473 return (zio); 3474 } 3475 3476 static ddt_entry_t *freedde; /* for debugging */ 3477 3478 static zio_t * 3479 zio_ddt_free(zio_t *zio) 3480 { 3481 spa_t *spa = zio->io_spa; 3482 blkptr_t *bp = zio->io_bp; 3483 ddt_t *ddt = ddt_select(spa, bp); 3484 ddt_entry_t *dde; 3485 ddt_phys_t *ddp; 3486 3487 ASSERT(BP_GET_DEDUP(bp)); 3488 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 3489 3490 ddt_enter(ddt); 3491 freedde = dde = ddt_lookup(ddt, bp, B_TRUE); 3492 if (dde) { 3493 ddp = ddt_phys_select(dde, bp); 3494 if (ddp) 3495 ddt_phys_decref(ddp); 3496 } 3497 ddt_exit(ddt); 3498 3499 return (zio); 3500 } 3501 3502 /* 3503 * ========================================================================== 3504 * Allocate and free blocks 3505 * ========================================================================== 3506 */ 3507 3508 static zio_t * 3509 zio_io_to_allocate(spa_t *spa, int allocator) 3510 { 3511 zio_t *zio; 3512 3513 ASSERT(MUTEX_HELD(&spa->spa_allocs[allocator].spaa_lock)); 3514 3515 zio = avl_first(&spa->spa_allocs[allocator].spaa_tree); 3516 if (zio == NULL) 3517 return (NULL); 3518 3519 ASSERT(IO_IS_ALLOCATING(zio)); 3520 3521 /* 3522 * Try to place a reservation for this zio. If we're unable to 3523 * reserve then we throttle. 3524 */ 3525 ASSERT3U(zio->io_allocator, ==, allocator); 3526 if (!metaslab_class_throttle_reserve(zio->io_metaslab_class, 3527 zio->io_prop.zp_copies, allocator, zio, 0)) { 3528 return (NULL); 3529 } 3530 3531 avl_remove(&spa->spa_allocs[allocator].spaa_tree, zio); 3532 ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE); 3533 3534 return (zio); 3535 } 3536 3537 static zio_t * 3538 zio_dva_throttle(zio_t *zio) 3539 { 3540 spa_t *spa = zio->io_spa; 3541 zio_t *nio; 3542 metaslab_class_t *mc; 3543 3544 /* locate an appropriate allocation class */ 3545 mc = spa_preferred_class(spa, zio->io_size, zio->io_prop.zp_type, 3546 zio->io_prop.zp_level, zio->io_prop.zp_zpl_smallblk); 3547 3548 if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE || 3549 !mc->mc_alloc_throttle_enabled || 3550 zio->io_child_type == ZIO_CHILD_GANG || 3551 zio->io_flags & ZIO_FLAG_NODATA) { 3552 return (zio); 3553 } 3554 3555 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 3556 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 3557 ASSERT3U(zio->io_queued_timestamp, >, 0); 3558 ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE); 3559 3560 zbookmark_phys_t *bm = &zio->io_bookmark; 3561 /* 3562 * We want to try to use as many allocators as possible to help improve 3563 * performance, but we also want logically adjacent IOs to be physically 3564 * adjacent to improve sequential read performance. We chunk each object 3565 * into 2^20 block regions, and then hash based on the objset, object, 3566 * level, and region to accomplish both of these goals. 3567 */ 3568 int allocator = (uint_t)cityhash4(bm->zb_objset, bm->zb_object, 3569 bm->zb_level, bm->zb_blkid >> 20) % spa->spa_alloc_count; 3570 zio->io_allocator = allocator; 3571 zio->io_metaslab_class = mc; 3572 mutex_enter(&spa->spa_allocs[allocator].spaa_lock); 3573 avl_add(&spa->spa_allocs[allocator].spaa_tree, zio); 3574 nio = zio_io_to_allocate(spa, allocator); 3575 mutex_exit(&spa->spa_allocs[allocator].spaa_lock); 3576 return (nio); 3577 } 3578 3579 static void 3580 zio_allocate_dispatch(spa_t *spa, int allocator) 3581 { 3582 zio_t *zio; 3583 3584 mutex_enter(&spa->spa_allocs[allocator].spaa_lock); 3585 zio = zio_io_to_allocate(spa, allocator); 3586 mutex_exit(&spa->spa_allocs[allocator].spaa_lock); 3587 if (zio == NULL) 3588 return; 3589 3590 ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE); 3591 ASSERT0(zio->io_error); 3592 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE); 3593 } 3594 3595 static zio_t * 3596 zio_dva_allocate(zio_t *zio) 3597 { 3598 spa_t *spa = zio->io_spa; 3599 metaslab_class_t *mc; 3600 blkptr_t *bp = zio->io_bp; 3601 int error; 3602 int flags = 0; 3603 3604 if (zio->io_gang_leader == NULL) { 3605 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 3606 zio->io_gang_leader = zio; 3607 } 3608 3609 ASSERT(BP_IS_HOLE(bp)); 3610 ASSERT0(BP_GET_NDVAS(bp)); 3611 ASSERT3U(zio->io_prop.zp_copies, >, 0); 3612 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa)); 3613 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp)); 3614 3615 if (zio->io_flags & ZIO_FLAG_NODATA) 3616 flags |= METASLAB_DONT_THROTTLE; 3617 if (zio->io_flags & ZIO_FLAG_GANG_CHILD) 3618 flags |= METASLAB_GANG_CHILD; 3619 if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE) 3620 flags |= METASLAB_ASYNC_ALLOC; 3621 3622 /* 3623 * if not already chosen, locate an appropriate allocation class 3624 */ 3625 mc = zio->io_metaslab_class; 3626 if (mc == NULL) { 3627 mc = spa_preferred_class(spa, zio->io_size, 3628 zio->io_prop.zp_type, zio->io_prop.zp_level, 3629 zio->io_prop.zp_zpl_smallblk); 3630 zio->io_metaslab_class = mc; 3631 } 3632 3633 /* 3634 * Try allocating the block in the usual metaslab class. 3635 * If that's full, allocate it in the normal class. 3636 * If that's full, allocate as a gang block, 3637 * and if all are full, the allocation fails (which shouldn't happen). 3638 * 3639 * Note that we do not fall back on embedded slog (ZIL) space, to 3640 * preserve unfragmented slog space, which is critical for decent 3641 * sync write performance. If a log allocation fails, we will fall 3642 * back to spa_sync() which is abysmal for performance. 3643 */ 3644 error = metaslab_alloc(spa, mc, zio->io_size, bp, 3645 zio->io_prop.zp_copies, zio->io_txg, NULL, flags, 3646 &zio->io_alloc_list, zio, zio->io_allocator); 3647 3648 /* 3649 * Fallback to normal class when an alloc class is full 3650 */ 3651 if (error == ENOSPC && mc != spa_normal_class(spa)) { 3652 /* 3653 * If throttling, transfer reservation over to normal class. 3654 * The io_allocator slot can remain the same even though we 3655 * are switching classes. 3656 */ 3657 if (mc->mc_alloc_throttle_enabled && 3658 (zio->io_flags & ZIO_FLAG_IO_ALLOCATING)) { 3659 metaslab_class_throttle_unreserve(mc, 3660 zio->io_prop.zp_copies, zio->io_allocator, zio); 3661 zio->io_flags &= ~ZIO_FLAG_IO_ALLOCATING; 3662 3663 VERIFY(metaslab_class_throttle_reserve( 3664 spa_normal_class(spa), 3665 zio->io_prop.zp_copies, zio->io_allocator, zio, 3666 flags | METASLAB_MUST_RESERVE)); 3667 } 3668 zio->io_metaslab_class = mc = spa_normal_class(spa); 3669 if (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC) { 3670 zfs_dbgmsg("%s: metaslab allocation failure, " 3671 "trying normal class: zio %px, size %llu, error %d", 3672 spa_name(spa), zio, (u_longlong_t)zio->io_size, 3673 error); 3674 } 3675 3676 error = metaslab_alloc(spa, mc, zio->io_size, bp, 3677 zio->io_prop.zp_copies, zio->io_txg, NULL, flags, 3678 &zio->io_alloc_list, zio, zio->io_allocator); 3679 } 3680 3681 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE) { 3682 if (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC) { 3683 zfs_dbgmsg("%s: metaslab allocation failure, " 3684 "trying ganging: zio %px, size %llu, error %d", 3685 spa_name(spa), zio, (u_longlong_t)zio->io_size, 3686 error); 3687 } 3688 return (zio_write_gang_block(zio, mc)); 3689 } 3690 if (error != 0) { 3691 if (error != ENOSPC || 3692 (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC)) { 3693 zfs_dbgmsg("%s: metaslab allocation failure: zio %px, " 3694 "size %llu, error %d", 3695 spa_name(spa), zio, (u_longlong_t)zio->io_size, 3696 error); 3697 } 3698 zio->io_error = error; 3699 } 3700 3701 return (zio); 3702 } 3703 3704 static zio_t * 3705 zio_dva_free(zio_t *zio) 3706 { 3707 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE); 3708 3709 return (zio); 3710 } 3711 3712 static zio_t * 3713 zio_dva_claim(zio_t *zio) 3714 { 3715 int error; 3716 3717 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg); 3718 if (error) 3719 zio->io_error = error; 3720 3721 return (zio); 3722 } 3723 3724 /* 3725 * Undo an allocation. This is used by zio_done() when an I/O fails 3726 * and we want to give back the block we just allocated. 3727 * This handles both normal blocks and gang blocks. 3728 */ 3729 static void 3730 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp) 3731 { 3732 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp)); 3733 ASSERT(zio->io_bp_override == NULL); 3734 3735 if (!BP_IS_HOLE(bp)) 3736 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE); 3737 3738 if (gn != NULL) { 3739 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 3740 zio_dva_unallocate(zio, gn->gn_child[g], 3741 &gn->gn_gbh->zg_blkptr[g]); 3742 } 3743 } 3744 } 3745 3746 /* 3747 * Try to allocate an intent log block. Return 0 on success, errno on failure. 3748 */ 3749 int 3750 zio_alloc_zil(spa_t *spa, objset_t *os, uint64_t txg, blkptr_t *new_bp, 3751 uint64_t size, boolean_t *slog) 3752 { 3753 int error = 1; 3754 zio_alloc_list_t io_alloc_list; 3755 3756 ASSERT(txg > spa_syncing_txg(spa)); 3757 3758 metaslab_trace_init(&io_alloc_list); 3759 3760 /* 3761 * Block pointer fields are useful to metaslabs for stats and debugging. 3762 * Fill in the obvious ones before calling into metaslab_alloc(). 3763 */ 3764 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); 3765 BP_SET_PSIZE(new_bp, size); 3766 BP_SET_LEVEL(new_bp, 0); 3767 3768 /* 3769 * When allocating a zil block, we don't have information about 3770 * the final destination of the block except the objset it's part 3771 * of, so we just hash the objset ID to pick the allocator to get 3772 * some parallelism. 3773 */ 3774 int flags = METASLAB_ZIL; 3775 int allocator = (uint_t)cityhash4(0, 0, 0, 3776 os->os_dsl_dataset->ds_object) % spa->spa_alloc_count; 3777 error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1, 3778 txg, NULL, flags, &io_alloc_list, NULL, allocator); 3779 *slog = (error == 0); 3780 if (error != 0) { 3781 error = metaslab_alloc(spa, spa_embedded_log_class(spa), size, 3782 new_bp, 1, txg, NULL, flags, 3783 &io_alloc_list, NULL, allocator); 3784 } 3785 if (error != 0) { 3786 error = metaslab_alloc(spa, spa_normal_class(spa), size, 3787 new_bp, 1, txg, NULL, flags, 3788 &io_alloc_list, NULL, allocator); 3789 } 3790 metaslab_trace_fini(&io_alloc_list); 3791 3792 if (error == 0) { 3793 BP_SET_LSIZE(new_bp, size); 3794 BP_SET_PSIZE(new_bp, size); 3795 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF); 3796 BP_SET_CHECKSUM(new_bp, 3797 spa_version(spa) >= SPA_VERSION_SLIM_ZIL 3798 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG); 3799 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); 3800 BP_SET_LEVEL(new_bp, 0); 3801 BP_SET_DEDUP(new_bp, 0); 3802 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER); 3803 3804 /* 3805 * encrypted blocks will require an IV and salt. We generate 3806 * these now since we will not be rewriting the bp at 3807 * rewrite time. 3808 */ 3809 if (os->os_encrypted) { 3810 uint8_t iv[ZIO_DATA_IV_LEN]; 3811 uint8_t salt[ZIO_DATA_SALT_LEN]; 3812 3813 BP_SET_CRYPT(new_bp, B_TRUE); 3814 VERIFY0(spa_crypt_get_salt(spa, 3815 dmu_objset_id(os), salt)); 3816 VERIFY0(zio_crypt_generate_iv(iv)); 3817 3818 zio_crypt_encode_params_bp(new_bp, salt, iv); 3819 } 3820 } else { 3821 zfs_dbgmsg("%s: zil block allocation failure: " 3822 "size %llu, error %d", spa_name(spa), (u_longlong_t)size, 3823 error); 3824 } 3825 3826 return (error); 3827 } 3828 3829 /* 3830 * ========================================================================== 3831 * Read and write to physical devices 3832 * ========================================================================== 3833 */ 3834 3835 /* 3836 * Issue an I/O to the underlying vdev. Typically the issue pipeline 3837 * stops after this stage and will resume upon I/O completion. 3838 * However, there are instances where the vdev layer may need to 3839 * continue the pipeline when an I/O was not issued. Since the I/O 3840 * that was sent to the vdev layer might be different than the one 3841 * currently active in the pipeline (see vdev_queue_io()), we explicitly 3842 * force the underlying vdev layers to call either zio_execute() or 3843 * zio_interrupt() to ensure that the pipeline continues with the correct I/O. 3844 */ 3845 static zio_t * 3846 zio_vdev_io_start(zio_t *zio) 3847 { 3848 vdev_t *vd = zio->io_vd; 3849 uint64_t align; 3850 spa_t *spa = zio->io_spa; 3851 3852 zio->io_delay = 0; 3853 3854 ASSERT(zio->io_error == 0); 3855 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0); 3856 3857 if (vd == NULL) { 3858 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) 3859 spa_config_enter(spa, SCL_ZIO, zio, RW_READER); 3860 3861 /* 3862 * The mirror_ops handle multiple DVAs in a single BP. 3863 */ 3864 vdev_mirror_ops.vdev_op_io_start(zio); 3865 return (NULL); 3866 } 3867 3868 ASSERT3P(zio->io_logical, !=, zio); 3869 if (zio->io_type == ZIO_TYPE_WRITE) { 3870 ASSERT(spa->spa_trust_config); 3871 3872 /* 3873 * Note: the code can handle other kinds of writes, 3874 * but we don't expect them. 3875 */ 3876 if (zio->io_vd->vdev_noalloc) { 3877 ASSERT(zio->io_flags & 3878 (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL | 3879 ZIO_FLAG_RESILVER | ZIO_FLAG_INDUCE_DAMAGE)); 3880 } 3881 } 3882 3883 align = 1ULL << vd->vdev_top->vdev_ashift; 3884 3885 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) && 3886 P2PHASE(zio->io_size, align) != 0) { 3887 /* Transform logical writes to be a full physical block size. */ 3888 uint64_t asize = P2ROUNDUP(zio->io_size, align); 3889 abd_t *abuf = abd_alloc_sametype(zio->io_abd, asize); 3890 ASSERT(vd == vd->vdev_top); 3891 if (zio->io_type == ZIO_TYPE_WRITE) { 3892 abd_copy(abuf, zio->io_abd, zio->io_size); 3893 abd_zero_off(abuf, zio->io_size, asize - zio->io_size); 3894 } 3895 zio_push_transform(zio, abuf, asize, asize, zio_subblock); 3896 } 3897 3898 /* 3899 * If this is not a physical io, make sure that it is properly aligned 3900 * before proceeding. 3901 */ 3902 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) { 3903 ASSERT0(P2PHASE(zio->io_offset, align)); 3904 ASSERT0(P2PHASE(zio->io_size, align)); 3905 } else { 3906 /* 3907 * For physical writes, we allow 512b aligned writes and assume 3908 * the device will perform a read-modify-write as necessary. 3909 */ 3910 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE)); 3911 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE)); 3912 } 3913 3914 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa)); 3915 3916 /* 3917 * If this is a repair I/O, and there's no self-healing involved -- 3918 * that is, we're just resilvering what we expect to resilver -- 3919 * then don't do the I/O unless zio's txg is actually in vd's DTL. 3920 * This prevents spurious resilvering. 3921 * 3922 * There are a few ways that we can end up creating these spurious 3923 * resilver i/os: 3924 * 3925 * 1. A resilver i/o will be issued if any DVA in the BP has a 3926 * dirty DTL. The mirror code will issue resilver writes to 3927 * each DVA, including the one(s) that are not on vdevs with dirty 3928 * DTLs. 3929 * 3930 * 2. With nested replication, which happens when we have a 3931 * "replacing" or "spare" vdev that's a child of a mirror or raidz. 3932 * For example, given mirror(replacing(A+B), C), it's likely that 3933 * only A is out of date (it's the new device). In this case, we'll 3934 * read from C, then use the data to resilver A+B -- but we don't 3935 * actually want to resilver B, just A. The top-level mirror has no 3936 * way to know this, so instead we just discard unnecessary repairs 3937 * as we work our way down the vdev tree. 3938 * 3939 * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc. 3940 * The same logic applies to any form of nested replication: ditto 3941 * + mirror, RAID-Z + replacing, etc. 3942 * 3943 * However, indirect vdevs point off to other vdevs which may have 3944 * DTL's, so we never bypass them. The child i/os on concrete vdevs 3945 * will be properly bypassed instead. 3946 * 3947 * Leaf DTL_PARTIAL can be empty when a legitimate write comes from 3948 * a dRAID spare vdev. For example, when a dRAID spare is first 3949 * used, its spare blocks need to be written to but the leaf vdev's 3950 * of such blocks can have empty DTL_PARTIAL. 3951 * 3952 * There seemed no clean way to allow such writes while bypassing 3953 * spurious ones. At this point, just avoid all bypassing for dRAID 3954 * for correctness. 3955 */ 3956 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) && 3957 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) && 3958 zio->io_txg != 0 && /* not a delegated i/o */ 3959 vd->vdev_ops != &vdev_indirect_ops && 3960 vd->vdev_top->vdev_ops != &vdev_draid_ops && 3961 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) { 3962 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 3963 zio_vdev_io_bypass(zio); 3964 return (zio); 3965 } 3966 3967 /* 3968 * Select the next best leaf I/O to process. Distributed spares are 3969 * excluded since they dispatch the I/O directly to a leaf vdev after 3970 * applying the dRAID mapping. 3971 */ 3972 if (vd->vdev_ops->vdev_op_leaf && 3973 vd->vdev_ops != &vdev_draid_spare_ops && 3974 (zio->io_type == ZIO_TYPE_READ || 3975 zio->io_type == ZIO_TYPE_WRITE || 3976 zio->io_type == ZIO_TYPE_TRIM)) { 3977 3978 if ((zio = vdev_queue_io(zio)) == NULL) 3979 return (NULL); 3980 3981 if (!vdev_accessible(vd, zio)) { 3982 zio->io_error = SET_ERROR(ENXIO); 3983 zio_interrupt(zio); 3984 return (NULL); 3985 } 3986 zio->io_delay = gethrtime(); 3987 } 3988 3989 vd->vdev_ops->vdev_op_io_start(zio); 3990 return (NULL); 3991 } 3992 3993 static zio_t * 3994 zio_vdev_io_done(zio_t *zio) 3995 { 3996 vdev_t *vd = zio->io_vd; 3997 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops; 3998 boolean_t unexpected_error = B_FALSE; 3999 4000 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) { 4001 return (NULL); 4002 } 4003 4004 ASSERT(zio->io_type == ZIO_TYPE_READ || 4005 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_TRIM); 4006 4007 if (zio->io_delay) 4008 zio->io_delay = gethrtime() - zio->io_delay; 4009 4010 if (vd != NULL && vd->vdev_ops->vdev_op_leaf && 4011 vd->vdev_ops != &vdev_draid_spare_ops) { 4012 vdev_queue_io_done(zio); 4013 4014 if (zio_injection_enabled && zio->io_error == 0) 4015 zio->io_error = zio_handle_device_injections(vd, zio, 4016 EIO, EILSEQ); 4017 4018 if (zio_injection_enabled && zio->io_error == 0) 4019 zio->io_error = zio_handle_label_injection(zio, EIO); 4020 4021 if (zio->io_error && zio->io_type != ZIO_TYPE_TRIM) { 4022 if (!vdev_accessible(vd, zio)) { 4023 zio->io_error = SET_ERROR(ENXIO); 4024 } else { 4025 unexpected_error = B_TRUE; 4026 } 4027 } 4028 } 4029 4030 ops->vdev_op_io_done(zio); 4031 4032 if (unexpected_error && vd->vdev_remove_wanted == B_FALSE) 4033 VERIFY(vdev_probe(vd, zio) == NULL); 4034 4035 return (zio); 4036 } 4037 4038 /* 4039 * This function is used to change the priority of an existing zio that is 4040 * currently in-flight. This is used by the arc to upgrade priority in the 4041 * event that a demand read is made for a block that is currently queued 4042 * as a scrub or async read IO. Otherwise, the high priority read request 4043 * would end up having to wait for the lower priority IO. 4044 */ 4045 void 4046 zio_change_priority(zio_t *pio, zio_priority_t priority) 4047 { 4048 zio_t *cio, *cio_next; 4049 zio_link_t *zl = NULL; 4050 4051 ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 4052 4053 if (pio->io_vd != NULL && pio->io_vd->vdev_ops->vdev_op_leaf) { 4054 vdev_queue_change_io_priority(pio, priority); 4055 } else { 4056 pio->io_priority = priority; 4057 } 4058 4059 mutex_enter(&pio->io_lock); 4060 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) { 4061 cio_next = zio_walk_children(pio, &zl); 4062 zio_change_priority(cio, priority); 4063 } 4064 mutex_exit(&pio->io_lock); 4065 } 4066 4067 /* 4068 * For non-raidz ZIOs, we can just copy aside the bad data read from the 4069 * disk, and use that to finish the checksum ereport later. 4070 */ 4071 static void 4072 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr, 4073 const abd_t *good_buf) 4074 { 4075 /* no processing needed */ 4076 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE); 4077 } 4078 4079 void 4080 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr) 4081 { 4082 void *abd = abd_alloc_sametype(zio->io_abd, zio->io_size); 4083 4084 abd_copy(abd, zio->io_abd, zio->io_size); 4085 4086 zcr->zcr_cbinfo = zio->io_size; 4087 zcr->zcr_cbdata = abd; 4088 zcr->zcr_finish = zio_vsd_default_cksum_finish; 4089 zcr->zcr_free = zio_abd_free; 4090 } 4091 4092 static zio_t * 4093 zio_vdev_io_assess(zio_t *zio) 4094 { 4095 vdev_t *vd = zio->io_vd; 4096 4097 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) { 4098 return (NULL); 4099 } 4100 4101 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) 4102 spa_config_exit(zio->io_spa, SCL_ZIO, zio); 4103 4104 if (zio->io_vsd != NULL) { 4105 zio->io_vsd_ops->vsd_free(zio); 4106 zio->io_vsd = NULL; 4107 } 4108 4109 if (zio_injection_enabled && zio->io_error == 0) 4110 zio->io_error = zio_handle_fault_injection(zio, EIO); 4111 4112 /* 4113 * If the I/O failed, determine whether we should attempt to retry it. 4114 * 4115 * On retry, we cut in line in the issue queue, since we don't want 4116 * compression/checksumming/etc. work to prevent our (cheap) IO reissue. 4117 */ 4118 if (zio->io_error && vd == NULL && 4119 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) { 4120 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */ 4121 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */ 4122 zio->io_error = 0; 4123 zio->io_flags |= ZIO_FLAG_IO_RETRY | ZIO_FLAG_DONT_AGGREGATE; 4124 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1; 4125 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, 4126 zio_requeue_io_start_cut_in_line); 4127 return (NULL); 4128 } 4129 4130 /* 4131 * If we got an error on a leaf device, convert it to ENXIO 4132 * if the device is not accessible at all. 4133 */ 4134 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf && 4135 !vdev_accessible(vd, zio)) 4136 zio->io_error = SET_ERROR(ENXIO); 4137 4138 /* 4139 * If we can't write to an interior vdev (mirror or RAID-Z), 4140 * set vdev_cant_write so that we stop trying to allocate from it. 4141 */ 4142 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE && 4143 vd != NULL && !vd->vdev_ops->vdev_op_leaf) { 4144 vdev_dbgmsg(vd, "zio_vdev_io_assess(zio=%px) setting " 4145 "cant_write=TRUE due to write failure with ENXIO", 4146 zio); 4147 vd->vdev_cant_write = B_TRUE; 4148 } 4149 4150 /* 4151 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future 4152 * attempts will ever succeed. In this case we set a persistent 4153 * boolean flag so that we don't bother with it in the future. 4154 */ 4155 if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) && 4156 zio->io_type == ZIO_TYPE_IOCTL && 4157 zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL) 4158 vd->vdev_nowritecache = B_TRUE; 4159 4160 if (zio->io_error) 4161 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 4162 4163 return (zio); 4164 } 4165 4166 void 4167 zio_vdev_io_reissue(zio_t *zio) 4168 { 4169 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); 4170 ASSERT(zio->io_error == 0); 4171 4172 zio->io_stage >>= 1; 4173 } 4174 4175 void 4176 zio_vdev_io_redone(zio_t *zio) 4177 { 4178 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE); 4179 4180 zio->io_stage >>= 1; 4181 } 4182 4183 void 4184 zio_vdev_io_bypass(zio_t *zio) 4185 { 4186 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); 4187 ASSERT(zio->io_error == 0); 4188 4189 zio->io_flags |= ZIO_FLAG_IO_BYPASS; 4190 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1; 4191 } 4192 4193 /* 4194 * ========================================================================== 4195 * Encrypt and store encryption parameters 4196 * ========================================================================== 4197 */ 4198 4199 4200 /* 4201 * This function is used for ZIO_STAGE_ENCRYPT. It is responsible for 4202 * managing the storage of encryption parameters and passing them to the 4203 * lower-level encryption functions. 4204 */ 4205 static zio_t * 4206 zio_encrypt(zio_t *zio) 4207 { 4208 zio_prop_t *zp = &zio->io_prop; 4209 spa_t *spa = zio->io_spa; 4210 blkptr_t *bp = zio->io_bp; 4211 uint64_t psize = BP_GET_PSIZE(bp); 4212 uint64_t dsobj = zio->io_bookmark.zb_objset; 4213 dmu_object_type_t ot = BP_GET_TYPE(bp); 4214 void *enc_buf = NULL; 4215 abd_t *eabd = NULL; 4216 uint8_t salt[ZIO_DATA_SALT_LEN]; 4217 uint8_t iv[ZIO_DATA_IV_LEN]; 4218 uint8_t mac[ZIO_DATA_MAC_LEN]; 4219 boolean_t no_crypt = B_FALSE; 4220 4221 /* the root zio already encrypted the data */ 4222 if (zio->io_child_type == ZIO_CHILD_GANG) 4223 return (zio); 4224 4225 /* only ZIL blocks are re-encrypted on rewrite */ 4226 if (!IO_IS_ALLOCATING(zio) && ot != DMU_OT_INTENT_LOG) 4227 return (zio); 4228 4229 if (!(zp->zp_encrypt || BP_IS_ENCRYPTED(bp))) { 4230 BP_SET_CRYPT(bp, B_FALSE); 4231 return (zio); 4232 } 4233 4234 /* if we are doing raw encryption set the provided encryption params */ 4235 if (zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) { 4236 ASSERT0(BP_GET_LEVEL(bp)); 4237 BP_SET_CRYPT(bp, B_TRUE); 4238 BP_SET_BYTEORDER(bp, zp->zp_byteorder); 4239 if (ot != DMU_OT_OBJSET) 4240 zio_crypt_encode_mac_bp(bp, zp->zp_mac); 4241 4242 /* dnode blocks must be written out in the provided byteorder */ 4243 if (zp->zp_byteorder != ZFS_HOST_BYTEORDER && 4244 ot == DMU_OT_DNODE) { 4245 void *bswap_buf = zio_buf_alloc(psize); 4246 abd_t *babd = abd_get_from_buf(bswap_buf, psize); 4247 4248 ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF); 4249 abd_copy_to_buf(bswap_buf, zio->io_abd, psize); 4250 dmu_ot_byteswap[DMU_OT_BYTESWAP(ot)].ob_func(bswap_buf, 4251 psize); 4252 4253 abd_take_ownership_of_buf(babd, B_TRUE); 4254 zio_push_transform(zio, babd, psize, psize, NULL); 4255 } 4256 4257 if (DMU_OT_IS_ENCRYPTED(ot)) 4258 zio_crypt_encode_params_bp(bp, zp->zp_salt, zp->zp_iv); 4259 return (zio); 4260 } 4261 4262 /* indirect blocks only maintain a cksum of the lower level MACs */ 4263 if (BP_GET_LEVEL(bp) > 0) { 4264 BP_SET_CRYPT(bp, B_TRUE); 4265 VERIFY0(zio_crypt_do_indirect_mac_checksum_abd(B_TRUE, 4266 zio->io_orig_abd, BP_GET_LSIZE(bp), BP_SHOULD_BYTESWAP(bp), 4267 mac)); 4268 zio_crypt_encode_mac_bp(bp, mac); 4269 return (zio); 4270 } 4271 4272 /* 4273 * Objset blocks are a special case since they have 2 256-bit MACs 4274 * embedded within them. 4275 */ 4276 if (ot == DMU_OT_OBJSET) { 4277 ASSERT0(DMU_OT_IS_ENCRYPTED(ot)); 4278 ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF); 4279 BP_SET_CRYPT(bp, B_TRUE); 4280 VERIFY0(spa_do_crypt_objset_mac_abd(B_TRUE, spa, dsobj, 4281 zio->io_abd, psize, BP_SHOULD_BYTESWAP(bp))); 4282 return (zio); 4283 } 4284 4285 /* unencrypted object types are only authenticated with a MAC */ 4286 if (!DMU_OT_IS_ENCRYPTED(ot)) { 4287 BP_SET_CRYPT(bp, B_TRUE); 4288 VERIFY0(spa_do_crypt_mac_abd(B_TRUE, spa, dsobj, 4289 zio->io_abd, psize, mac)); 4290 zio_crypt_encode_mac_bp(bp, mac); 4291 return (zio); 4292 } 4293 4294 /* 4295 * Later passes of sync-to-convergence may decide to rewrite data 4296 * in place to avoid more disk reallocations. This presents a problem 4297 * for encryption because this constitutes rewriting the new data with 4298 * the same encryption key and IV. However, this only applies to blocks 4299 * in the MOS (particularly the spacemaps) and we do not encrypt the 4300 * MOS. We assert that the zio is allocating or an intent log write 4301 * to enforce this. 4302 */ 4303 ASSERT(IO_IS_ALLOCATING(zio) || ot == DMU_OT_INTENT_LOG); 4304 ASSERT(BP_GET_LEVEL(bp) == 0 || ot == DMU_OT_INTENT_LOG); 4305 ASSERT(spa_feature_is_active(spa, SPA_FEATURE_ENCRYPTION)); 4306 ASSERT3U(psize, !=, 0); 4307 4308 enc_buf = zio_buf_alloc(psize); 4309 eabd = abd_get_from_buf(enc_buf, psize); 4310 abd_take_ownership_of_buf(eabd, B_TRUE); 4311 4312 /* 4313 * For an explanation of what encryption parameters are stored 4314 * where, see the block comment in zio_crypt.c. 4315 */ 4316 if (ot == DMU_OT_INTENT_LOG) { 4317 zio_crypt_decode_params_bp(bp, salt, iv); 4318 } else { 4319 BP_SET_CRYPT(bp, B_TRUE); 4320 } 4321 4322 /* Perform the encryption. This should not fail */ 4323 VERIFY0(spa_do_crypt_abd(B_TRUE, spa, &zio->io_bookmark, 4324 BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp), 4325 salt, iv, mac, psize, zio->io_abd, eabd, &no_crypt)); 4326 4327 /* encode encryption metadata into the bp */ 4328 if (ot == DMU_OT_INTENT_LOG) { 4329 /* 4330 * ZIL blocks store the MAC in the embedded checksum, so the 4331 * transform must always be applied. 4332 */ 4333 zio_crypt_encode_mac_zil(enc_buf, mac); 4334 zio_push_transform(zio, eabd, psize, psize, NULL); 4335 } else { 4336 BP_SET_CRYPT(bp, B_TRUE); 4337 zio_crypt_encode_params_bp(bp, salt, iv); 4338 zio_crypt_encode_mac_bp(bp, mac); 4339 4340 if (no_crypt) { 4341 ASSERT3U(ot, ==, DMU_OT_DNODE); 4342 abd_free(eabd); 4343 } else { 4344 zio_push_transform(zio, eabd, psize, psize, NULL); 4345 } 4346 } 4347 4348 return (zio); 4349 } 4350 4351 /* 4352 * ========================================================================== 4353 * Generate and verify checksums 4354 * ========================================================================== 4355 */ 4356 static zio_t * 4357 zio_checksum_generate(zio_t *zio) 4358 { 4359 blkptr_t *bp = zio->io_bp; 4360 enum zio_checksum checksum; 4361 4362 if (bp == NULL) { 4363 /* 4364 * This is zio_write_phys(). 4365 * We're either generating a label checksum, or none at all. 4366 */ 4367 checksum = zio->io_prop.zp_checksum; 4368 4369 if (checksum == ZIO_CHECKSUM_OFF) 4370 return (zio); 4371 4372 ASSERT(checksum == ZIO_CHECKSUM_LABEL); 4373 } else { 4374 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) { 4375 ASSERT(!IO_IS_ALLOCATING(zio)); 4376 checksum = ZIO_CHECKSUM_GANG_HEADER; 4377 } else { 4378 checksum = BP_GET_CHECKSUM(bp); 4379 } 4380 } 4381 4382 zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size); 4383 4384 return (zio); 4385 } 4386 4387 static zio_t * 4388 zio_checksum_verify(zio_t *zio) 4389 { 4390 zio_bad_cksum_t info; 4391 blkptr_t *bp = zio->io_bp; 4392 int error; 4393 4394 ASSERT(zio->io_vd != NULL); 4395 4396 if (bp == NULL) { 4397 /* 4398 * This is zio_read_phys(). 4399 * We're either verifying a label checksum, or nothing at all. 4400 */ 4401 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF) 4402 return (zio); 4403 4404 ASSERT3U(zio->io_prop.zp_checksum, ==, ZIO_CHECKSUM_LABEL); 4405 } 4406 4407 if ((error = zio_checksum_error(zio, &info)) != 0) { 4408 zio->io_error = error; 4409 if (error == ECKSUM && 4410 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) { 4411 mutex_enter(&zio->io_vd->vdev_stat_lock); 4412 zio->io_vd->vdev_stat.vs_checksum_errors++; 4413 mutex_exit(&zio->io_vd->vdev_stat_lock); 4414 (void) zfs_ereport_start_checksum(zio->io_spa, 4415 zio->io_vd, &zio->io_bookmark, zio, 4416 zio->io_offset, zio->io_size, &info); 4417 } 4418 } 4419 4420 return (zio); 4421 } 4422 4423 /* 4424 * Called by RAID-Z to ensure we don't compute the checksum twice. 4425 */ 4426 void 4427 zio_checksum_verified(zio_t *zio) 4428 { 4429 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; 4430 } 4431 4432 /* 4433 * ========================================================================== 4434 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other. 4435 * An error of 0 indicates success. ENXIO indicates whole-device failure, 4436 * which may be transient (e.g. unplugged) or permanent. ECKSUM and EIO 4437 * indicate errors that are specific to one I/O, and most likely permanent. 4438 * Any other error is presumed to be worse because we weren't expecting it. 4439 * ========================================================================== 4440 */ 4441 int 4442 zio_worst_error(int e1, int e2) 4443 { 4444 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO }; 4445 int r1, r2; 4446 4447 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++) 4448 if (e1 == zio_error_rank[r1]) 4449 break; 4450 4451 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++) 4452 if (e2 == zio_error_rank[r2]) 4453 break; 4454 4455 return (r1 > r2 ? e1 : e2); 4456 } 4457 4458 /* 4459 * ========================================================================== 4460 * I/O completion 4461 * ========================================================================== 4462 */ 4463 static zio_t * 4464 zio_ready(zio_t *zio) 4465 { 4466 blkptr_t *bp = zio->io_bp; 4467 zio_t *pio, *pio_next; 4468 zio_link_t *zl = NULL; 4469 4470 if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT | 4471 ZIO_CHILD_GANG_BIT | ZIO_CHILD_DDT_BIT, ZIO_WAIT_READY)) { 4472 return (NULL); 4473 } 4474 4475 if (zio->io_ready) { 4476 ASSERT(IO_IS_ALLOCATING(zio)); 4477 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) || 4478 (zio->io_flags & ZIO_FLAG_NOPWRITE)); 4479 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0); 4480 4481 zio->io_ready(zio); 4482 } 4483 4484 #ifdef ZFS_DEBUG 4485 if (bp != NULL && bp != &zio->io_bp_copy) 4486 zio->io_bp_copy = *bp; 4487 #endif 4488 4489 if (zio->io_error != 0) { 4490 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 4491 4492 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) { 4493 ASSERT(IO_IS_ALLOCATING(zio)); 4494 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); 4495 ASSERT(zio->io_metaslab_class != NULL); 4496 4497 /* 4498 * We were unable to allocate anything, unreserve and 4499 * issue the next I/O to allocate. 4500 */ 4501 metaslab_class_throttle_unreserve( 4502 zio->io_metaslab_class, zio->io_prop.zp_copies, 4503 zio->io_allocator, zio); 4504 zio_allocate_dispatch(zio->io_spa, zio->io_allocator); 4505 } 4506 } 4507 4508 mutex_enter(&zio->io_lock); 4509 zio->io_state[ZIO_WAIT_READY] = 1; 4510 pio = zio_walk_parents(zio, &zl); 4511 mutex_exit(&zio->io_lock); 4512 4513 /* 4514 * As we notify zio's parents, new parents could be added. 4515 * New parents go to the head of zio's io_parent_list, however, 4516 * so we will (correctly) not notify them. The remainder of zio's 4517 * io_parent_list, from 'pio_next' onward, cannot change because 4518 * all parents must wait for us to be done before they can be done. 4519 */ 4520 for (; pio != NULL; pio = pio_next) { 4521 pio_next = zio_walk_parents(zio, &zl); 4522 zio_notify_parent(pio, zio, ZIO_WAIT_READY, NULL); 4523 } 4524 4525 if (zio->io_flags & ZIO_FLAG_NODATA) { 4526 if (bp != NULL && BP_IS_GANG(bp)) { 4527 zio->io_flags &= ~ZIO_FLAG_NODATA; 4528 } else { 4529 ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE); 4530 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; 4531 } 4532 } 4533 4534 if (zio_injection_enabled && 4535 zio->io_spa->spa_syncing_txg == zio->io_txg) 4536 zio_handle_ignored_writes(zio); 4537 4538 return (zio); 4539 } 4540 4541 /* 4542 * Update the allocation throttle accounting. 4543 */ 4544 static void 4545 zio_dva_throttle_done(zio_t *zio) 4546 { 4547 zio_t *lio __maybe_unused = zio->io_logical; 4548 zio_t *pio = zio_unique_parent(zio); 4549 vdev_t *vd = zio->io_vd; 4550 int flags = METASLAB_ASYNC_ALLOC; 4551 4552 ASSERT3P(zio->io_bp, !=, NULL); 4553 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE); 4554 ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE); 4555 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV); 4556 ASSERT(vd != NULL); 4557 ASSERT3P(vd, ==, vd->vdev_top); 4558 ASSERT(zio_injection_enabled || !(zio->io_flags & ZIO_FLAG_IO_RETRY)); 4559 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR)); 4560 ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING); 4561 ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE)); 4562 ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA)); 4563 4564 /* 4565 * Parents of gang children can have two flavors -- ones that 4566 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set) 4567 * and ones that allocated the constituent blocks. The allocation 4568 * throttle needs to know the allocating parent zio so we must find 4569 * it here. 4570 */ 4571 if (pio->io_child_type == ZIO_CHILD_GANG) { 4572 /* 4573 * If our parent is a rewrite gang child then our grandparent 4574 * would have been the one that performed the allocation. 4575 */ 4576 if (pio->io_flags & ZIO_FLAG_IO_REWRITE) 4577 pio = zio_unique_parent(pio); 4578 flags |= METASLAB_GANG_CHILD; 4579 } 4580 4581 ASSERT(IO_IS_ALLOCATING(pio)); 4582 ASSERT3P(zio, !=, zio->io_logical); 4583 ASSERT(zio->io_logical != NULL); 4584 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR)); 4585 ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE); 4586 ASSERT(zio->io_metaslab_class != NULL); 4587 4588 mutex_enter(&pio->io_lock); 4589 metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags, 4590 pio->io_allocator, B_TRUE); 4591 mutex_exit(&pio->io_lock); 4592 4593 metaslab_class_throttle_unreserve(zio->io_metaslab_class, 1, 4594 pio->io_allocator, pio); 4595 4596 /* 4597 * Call into the pipeline to see if there is more work that 4598 * needs to be done. If there is work to be done it will be 4599 * dispatched to another taskq thread. 4600 */ 4601 zio_allocate_dispatch(zio->io_spa, pio->io_allocator); 4602 } 4603 4604 static zio_t * 4605 zio_done(zio_t *zio) 4606 { 4607 /* 4608 * Always attempt to keep stack usage minimal here since 4609 * we can be called recursively up to 19 levels deep. 4610 */ 4611 const uint64_t psize = zio->io_size; 4612 zio_t *pio, *pio_next; 4613 zio_link_t *zl = NULL; 4614 4615 /* 4616 * If our children haven't all completed, 4617 * wait for them and then repeat this pipeline stage. 4618 */ 4619 if (zio_wait_for_children(zio, ZIO_CHILD_ALL_BITS, ZIO_WAIT_DONE)) { 4620 return (NULL); 4621 } 4622 4623 /* 4624 * If the allocation throttle is enabled, then update the accounting. 4625 * We only track child I/Os that are part of an allocating async 4626 * write. We must do this since the allocation is performed 4627 * by the logical I/O but the actual write is done by child I/Os. 4628 */ 4629 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING && 4630 zio->io_child_type == ZIO_CHILD_VDEV) { 4631 ASSERT(zio->io_metaslab_class != NULL); 4632 ASSERT(zio->io_metaslab_class->mc_alloc_throttle_enabled); 4633 zio_dva_throttle_done(zio); 4634 } 4635 4636 /* 4637 * If the allocation throttle is enabled, verify that 4638 * we have decremented the refcounts for every I/O that was throttled. 4639 */ 4640 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) { 4641 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 4642 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); 4643 ASSERT(zio->io_bp != NULL); 4644 4645 metaslab_group_alloc_verify(zio->io_spa, zio->io_bp, zio, 4646 zio->io_allocator); 4647 VERIFY(zfs_refcount_not_held(&zio->io_metaslab_class-> 4648 mc_allocator[zio->io_allocator].mca_alloc_slots, zio)); 4649 } 4650 4651 4652 for (int c = 0; c < ZIO_CHILD_TYPES; c++) 4653 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 4654 ASSERT(zio->io_children[c][w] == 0); 4655 4656 if (zio->io_bp != NULL && !BP_IS_EMBEDDED(zio->io_bp)) { 4657 ASSERT(zio->io_bp->blk_pad[0] == 0); 4658 ASSERT(zio->io_bp->blk_pad[1] == 0); 4659 ASSERT(memcmp(zio->io_bp, &zio->io_bp_copy, 4660 sizeof (blkptr_t)) == 0 || 4661 (zio->io_bp == zio_unique_parent(zio)->io_bp)); 4662 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) && 4663 zio->io_bp_override == NULL && 4664 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) { 4665 ASSERT3U(zio->io_prop.zp_copies, <=, 4666 BP_GET_NDVAS(zio->io_bp)); 4667 ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 || 4668 (BP_COUNT_GANG(zio->io_bp) == 4669 BP_GET_NDVAS(zio->io_bp))); 4670 } 4671 if (zio->io_flags & ZIO_FLAG_NOPWRITE) 4672 VERIFY(BP_EQUAL(zio->io_bp, &zio->io_bp_orig)); 4673 } 4674 4675 /* 4676 * If there were child vdev/gang/ddt errors, they apply to us now. 4677 */ 4678 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV); 4679 zio_inherit_child_errors(zio, ZIO_CHILD_GANG); 4680 zio_inherit_child_errors(zio, ZIO_CHILD_DDT); 4681 4682 /* 4683 * If the I/O on the transformed data was successful, generate any 4684 * checksum reports now while we still have the transformed data. 4685 */ 4686 if (zio->io_error == 0) { 4687 while (zio->io_cksum_report != NULL) { 4688 zio_cksum_report_t *zcr = zio->io_cksum_report; 4689 uint64_t align = zcr->zcr_align; 4690 uint64_t asize = P2ROUNDUP(psize, align); 4691 abd_t *adata = zio->io_abd; 4692 4693 if (adata != NULL && asize != psize) { 4694 adata = abd_alloc(asize, B_TRUE); 4695 abd_copy(adata, zio->io_abd, psize); 4696 abd_zero_off(adata, psize, asize - psize); 4697 } 4698 4699 zio->io_cksum_report = zcr->zcr_next; 4700 zcr->zcr_next = NULL; 4701 zcr->zcr_finish(zcr, adata); 4702 zfs_ereport_free_checksum(zcr); 4703 4704 if (adata != NULL && asize != psize) 4705 abd_free(adata); 4706 } 4707 } 4708 4709 zio_pop_transforms(zio); /* note: may set zio->io_error */ 4710 4711 vdev_stat_update(zio, psize); 4712 4713 /* 4714 * If this I/O is attached to a particular vdev is slow, exceeding 4715 * 30 seconds to complete, post an error described the I/O delay. 4716 * We ignore these errors if the device is currently unavailable. 4717 */ 4718 if (zio->io_delay >= MSEC2NSEC(zio_slow_io_ms)) { 4719 if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd)) { 4720 /* 4721 * We want to only increment our slow IO counters if 4722 * the IO is valid (i.e. not if the drive is removed). 4723 * 4724 * zfs_ereport_post() will also do these checks, but 4725 * it can also ratelimit and have other failures, so we 4726 * need to increment the slow_io counters independent 4727 * of it. 4728 */ 4729 if (zfs_ereport_is_valid(FM_EREPORT_ZFS_DELAY, 4730 zio->io_spa, zio->io_vd, zio)) { 4731 mutex_enter(&zio->io_vd->vdev_stat_lock); 4732 zio->io_vd->vdev_stat.vs_slow_ios++; 4733 mutex_exit(&zio->io_vd->vdev_stat_lock); 4734 4735 (void) zfs_ereport_post(FM_EREPORT_ZFS_DELAY, 4736 zio->io_spa, zio->io_vd, &zio->io_bookmark, 4737 zio, 0); 4738 } 4739 } 4740 } 4741 4742 if (zio->io_error) { 4743 /* 4744 * If this I/O is attached to a particular vdev, 4745 * generate an error message describing the I/O failure 4746 * at the block level. We ignore these errors if the 4747 * device is currently unavailable. 4748 */ 4749 if (zio->io_error != ECKSUM && zio->io_vd != NULL && 4750 !vdev_is_dead(zio->io_vd)) { 4751 int ret = zfs_ereport_post(FM_EREPORT_ZFS_IO, 4752 zio->io_spa, zio->io_vd, &zio->io_bookmark, zio, 0); 4753 if (ret != EALREADY) { 4754 mutex_enter(&zio->io_vd->vdev_stat_lock); 4755 if (zio->io_type == ZIO_TYPE_READ) 4756 zio->io_vd->vdev_stat.vs_read_errors++; 4757 else if (zio->io_type == ZIO_TYPE_WRITE) 4758 zio->io_vd->vdev_stat.vs_write_errors++; 4759 mutex_exit(&zio->io_vd->vdev_stat_lock); 4760 } 4761 } 4762 4763 if ((zio->io_error == EIO || !(zio->io_flags & 4764 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) && 4765 zio == zio->io_logical) { 4766 /* 4767 * For logical I/O requests, tell the SPA to log the 4768 * error and generate a logical data ereport. 4769 */ 4770 spa_log_error(zio->io_spa, &zio->io_bookmark, 4771 &zio->io_bp->blk_birth); 4772 (void) zfs_ereport_post(FM_EREPORT_ZFS_DATA, 4773 zio->io_spa, NULL, &zio->io_bookmark, zio, 0); 4774 } 4775 } 4776 4777 if (zio->io_error && zio == zio->io_logical) { 4778 /* 4779 * Determine whether zio should be reexecuted. This will 4780 * propagate all the way to the root via zio_notify_parent(). 4781 */ 4782 ASSERT(zio->io_vd == NULL && zio->io_bp != NULL); 4783 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 4784 4785 if (IO_IS_ALLOCATING(zio) && 4786 !(zio->io_flags & ZIO_FLAG_CANFAIL)) { 4787 if (zio->io_error != ENOSPC) 4788 zio->io_reexecute |= ZIO_REEXECUTE_NOW; 4789 else 4790 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 4791 } 4792 4793 if ((zio->io_type == ZIO_TYPE_READ || 4794 zio->io_type == ZIO_TYPE_FREE) && 4795 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && 4796 zio->io_error == ENXIO && 4797 spa_load_state(zio->io_spa) == SPA_LOAD_NONE && 4798 spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE) 4799 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 4800 4801 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute) 4802 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 4803 4804 /* 4805 * Here is a possibly good place to attempt to do 4806 * either combinatorial reconstruction or error correction 4807 * based on checksums. It also might be a good place 4808 * to send out preliminary ereports before we suspend 4809 * processing. 4810 */ 4811 } 4812 4813 /* 4814 * If there were logical child errors, they apply to us now. 4815 * We defer this until now to avoid conflating logical child 4816 * errors with errors that happened to the zio itself when 4817 * updating vdev stats and reporting FMA events above. 4818 */ 4819 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL); 4820 4821 if ((zio->io_error || zio->io_reexecute) && 4822 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio && 4823 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE))) 4824 zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp); 4825 4826 zio_gang_tree_free(&zio->io_gang_tree); 4827 4828 /* 4829 * Godfather I/Os should never suspend. 4830 */ 4831 if ((zio->io_flags & ZIO_FLAG_GODFATHER) && 4832 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) 4833 zio->io_reexecute &= ~ZIO_REEXECUTE_SUSPEND; 4834 4835 if (zio->io_reexecute) { 4836 /* 4837 * This is a logical I/O that wants to reexecute. 4838 * 4839 * Reexecute is top-down. When an i/o fails, if it's not 4840 * the root, it simply notifies its parent and sticks around. 4841 * The parent, seeing that it still has children in zio_done(), 4842 * does the same. This percolates all the way up to the root. 4843 * The root i/o will reexecute or suspend the entire tree. 4844 * 4845 * This approach ensures that zio_reexecute() honors 4846 * all the original i/o dependency relationships, e.g. 4847 * parents not executing until children are ready. 4848 */ 4849 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 4850 4851 zio->io_gang_leader = NULL; 4852 4853 mutex_enter(&zio->io_lock); 4854 zio->io_state[ZIO_WAIT_DONE] = 1; 4855 mutex_exit(&zio->io_lock); 4856 4857 /* 4858 * "The Godfather" I/O monitors its children but is 4859 * not a true parent to them. It will track them through 4860 * the pipeline but severs its ties whenever they get into 4861 * trouble (e.g. suspended). This allows "The Godfather" 4862 * I/O to return status without blocking. 4863 */ 4864 zl = NULL; 4865 for (pio = zio_walk_parents(zio, &zl); pio != NULL; 4866 pio = pio_next) { 4867 zio_link_t *remove_zl = zl; 4868 pio_next = zio_walk_parents(zio, &zl); 4869 4870 if ((pio->io_flags & ZIO_FLAG_GODFATHER) && 4871 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) { 4872 zio_remove_child(pio, zio, remove_zl); 4873 /* 4874 * This is a rare code path, so we don't 4875 * bother with "next_to_execute". 4876 */ 4877 zio_notify_parent(pio, zio, ZIO_WAIT_DONE, 4878 NULL); 4879 } 4880 } 4881 4882 if ((pio = zio_unique_parent(zio)) != NULL) { 4883 /* 4884 * We're not a root i/o, so there's nothing to do 4885 * but notify our parent. Don't propagate errors 4886 * upward since we haven't permanently failed yet. 4887 */ 4888 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); 4889 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE; 4890 /* 4891 * This is a rare code path, so we don't bother with 4892 * "next_to_execute". 4893 */ 4894 zio_notify_parent(pio, zio, ZIO_WAIT_DONE, NULL); 4895 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) { 4896 /* 4897 * We'd fail again if we reexecuted now, so suspend 4898 * until conditions improve (e.g. device comes online). 4899 */ 4900 zio_suspend(zio->io_spa, zio, ZIO_SUSPEND_IOERR); 4901 } else { 4902 /* 4903 * Reexecution is potentially a huge amount of work. 4904 * Hand it off to the otherwise-unused claim taskq. 4905 */ 4906 ASSERT(taskq_empty_ent(&zio->io_tqent)); 4907 spa_taskq_dispatch_ent(zio->io_spa, 4908 ZIO_TYPE_CLAIM, ZIO_TASKQ_ISSUE, 4909 zio_reexecute, zio, 0, &zio->io_tqent); 4910 } 4911 return (NULL); 4912 } 4913 4914 ASSERT(list_is_empty(&zio->io_child_list)); 4915 ASSERT(zio->io_reexecute == 0); 4916 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL)); 4917 4918 /* 4919 * Report any checksum errors, since the I/O is complete. 4920 */ 4921 while (zio->io_cksum_report != NULL) { 4922 zio_cksum_report_t *zcr = zio->io_cksum_report; 4923 zio->io_cksum_report = zcr->zcr_next; 4924 zcr->zcr_next = NULL; 4925 zcr->zcr_finish(zcr, NULL); 4926 zfs_ereport_free_checksum(zcr); 4927 } 4928 4929 /* 4930 * It is the responsibility of the done callback to ensure that this 4931 * particular zio is no longer discoverable for adoption, and as 4932 * such, cannot acquire any new parents. 4933 */ 4934 if (zio->io_done) 4935 zio->io_done(zio); 4936 4937 mutex_enter(&zio->io_lock); 4938 zio->io_state[ZIO_WAIT_DONE] = 1; 4939 mutex_exit(&zio->io_lock); 4940 4941 /* 4942 * We are done executing this zio. We may want to execute a parent 4943 * next. See the comment in zio_notify_parent(). 4944 */ 4945 zio_t *next_to_execute = NULL; 4946 zl = NULL; 4947 for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) { 4948 zio_link_t *remove_zl = zl; 4949 pio_next = zio_walk_parents(zio, &zl); 4950 zio_remove_child(pio, zio, remove_zl); 4951 zio_notify_parent(pio, zio, ZIO_WAIT_DONE, &next_to_execute); 4952 } 4953 4954 if (zio->io_waiter != NULL) { 4955 mutex_enter(&zio->io_lock); 4956 zio->io_executor = NULL; 4957 cv_broadcast(&zio->io_cv); 4958 mutex_exit(&zio->io_lock); 4959 } else { 4960 zio_destroy(zio); 4961 } 4962 4963 return (next_to_execute); 4964 } 4965 4966 /* 4967 * ========================================================================== 4968 * I/O pipeline definition 4969 * ========================================================================== 4970 */ 4971 static zio_pipe_stage_t *zio_pipeline[] = { 4972 NULL, 4973 zio_read_bp_init, 4974 zio_write_bp_init, 4975 zio_free_bp_init, 4976 zio_issue_async, 4977 zio_write_compress, 4978 zio_encrypt, 4979 zio_checksum_generate, 4980 zio_nop_write, 4981 zio_brt_free, 4982 zio_ddt_read_start, 4983 zio_ddt_read_done, 4984 zio_ddt_write, 4985 zio_ddt_free, 4986 zio_gang_assemble, 4987 zio_gang_issue, 4988 zio_dva_throttle, 4989 zio_dva_allocate, 4990 zio_dva_free, 4991 zio_dva_claim, 4992 zio_ready, 4993 zio_vdev_io_start, 4994 zio_vdev_io_done, 4995 zio_vdev_io_assess, 4996 zio_checksum_verify, 4997 zio_done 4998 }; 4999 5000 5001 5002 5003 /* 5004 * Compare two zbookmark_phys_t's to see which we would reach first in a 5005 * pre-order traversal of the object tree. 5006 * 5007 * This is simple in every case aside from the meta-dnode object. For all other 5008 * objects, we traverse them in order (object 1 before object 2, and so on). 5009 * However, all of these objects are traversed while traversing object 0, since 5010 * the data it points to is the list of objects. Thus, we need to convert to a 5011 * canonical representation so we can compare meta-dnode bookmarks to 5012 * non-meta-dnode bookmarks. 5013 * 5014 * We do this by calculating "equivalents" for each field of the zbookmark. 5015 * zbookmarks outside of the meta-dnode use their own object and level, and 5016 * calculate the level 0 equivalent (the first L0 blkid that is contained in the 5017 * blocks this bookmark refers to) by multiplying their blkid by their span 5018 * (the number of L0 blocks contained within one block at their level). 5019 * zbookmarks inside the meta-dnode calculate their object equivalent 5020 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use 5021 * level + 1<<31 (any value larger than a level could ever be) for their level. 5022 * This causes them to always compare before a bookmark in their object 5023 * equivalent, compare appropriately to bookmarks in other objects, and to 5024 * compare appropriately to other bookmarks in the meta-dnode. 5025 */ 5026 int 5027 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2, 5028 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2) 5029 { 5030 /* 5031 * These variables represent the "equivalent" values for the zbookmark, 5032 * after converting zbookmarks inside the meta dnode to their 5033 * normal-object equivalents. 5034 */ 5035 uint64_t zb1obj, zb2obj; 5036 uint64_t zb1L0, zb2L0; 5037 uint64_t zb1level, zb2level; 5038 5039 if (zb1->zb_object == zb2->zb_object && 5040 zb1->zb_level == zb2->zb_level && 5041 zb1->zb_blkid == zb2->zb_blkid) 5042 return (0); 5043 5044 IMPLY(zb1->zb_level > 0, ibs1 >= SPA_MINBLOCKSHIFT); 5045 IMPLY(zb2->zb_level > 0, ibs2 >= SPA_MINBLOCKSHIFT); 5046 5047 /* 5048 * BP_SPANB calculates the span in blocks. 5049 */ 5050 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level); 5051 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level); 5052 5053 if (zb1->zb_object == DMU_META_DNODE_OBJECT) { 5054 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT)); 5055 zb1L0 = 0; 5056 zb1level = zb1->zb_level + COMPARE_META_LEVEL; 5057 } else { 5058 zb1obj = zb1->zb_object; 5059 zb1level = zb1->zb_level; 5060 } 5061 5062 if (zb2->zb_object == DMU_META_DNODE_OBJECT) { 5063 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT)); 5064 zb2L0 = 0; 5065 zb2level = zb2->zb_level + COMPARE_META_LEVEL; 5066 } else { 5067 zb2obj = zb2->zb_object; 5068 zb2level = zb2->zb_level; 5069 } 5070 5071 /* Now that we have a canonical representation, do the comparison. */ 5072 if (zb1obj != zb2obj) 5073 return (zb1obj < zb2obj ? -1 : 1); 5074 else if (zb1L0 != zb2L0) 5075 return (zb1L0 < zb2L0 ? -1 : 1); 5076 else if (zb1level != zb2level) 5077 return (zb1level > zb2level ? -1 : 1); 5078 /* 5079 * This can (theoretically) happen if the bookmarks have the same object 5080 * and level, but different blkids, if the block sizes are not the same. 5081 * There is presently no way to change the indirect block sizes 5082 */ 5083 return (0); 5084 } 5085 5086 /* 5087 * This function checks the following: given that last_block is the place that 5088 * our traversal stopped last time, does that guarantee that we've visited 5089 * every node under subtree_root? Therefore, we can't just use the raw output 5090 * of zbookmark_compare. We have to pass in a modified version of 5091 * subtree_root; by incrementing the block id, and then checking whether 5092 * last_block is before or equal to that, we can tell whether or not having 5093 * visited last_block implies that all of subtree_root's children have been 5094 * visited. 5095 */ 5096 boolean_t 5097 zbookmark_subtree_completed(const dnode_phys_t *dnp, 5098 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block) 5099 { 5100 zbookmark_phys_t mod_zb = *subtree_root; 5101 mod_zb.zb_blkid++; 5102 ASSERT0(last_block->zb_level); 5103 5104 /* The objset_phys_t isn't before anything. */ 5105 if (dnp == NULL) 5106 return (B_FALSE); 5107 5108 /* 5109 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the 5110 * data block size in sectors, because that variable is only used if 5111 * the bookmark refers to a block in the meta-dnode. Since we don't 5112 * know without examining it what object it refers to, and there's no 5113 * harm in passing in this value in other cases, we always pass it in. 5114 * 5115 * We pass in 0 for the indirect block size shift because zb2 must be 5116 * level 0. The indirect block size is only used to calculate the span 5117 * of the bookmark, but since the bookmark must be level 0, the span is 5118 * always 1, so the math works out. 5119 * 5120 * If you make changes to how the zbookmark_compare code works, be sure 5121 * to make sure that this code still works afterwards. 5122 */ 5123 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift, 5124 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb, 5125 last_block) <= 0); 5126 } 5127 5128 /* 5129 * This function is similar to zbookmark_subtree_completed(), but returns true 5130 * if subtree_root is equal or ahead of last_block, i.e. still to be done. 5131 */ 5132 boolean_t 5133 zbookmark_subtree_tbd(const dnode_phys_t *dnp, 5134 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block) 5135 { 5136 ASSERT0(last_block->zb_level); 5137 if (dnp == NULL) 5138 return (B_FALSE); 5139 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift, 5140 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, subtree_root, 5141 last_block) >= 0); 5142 } 5143 5144 EXPORT_SYMBOL(zio_type_name); 5145 EXPORT_SYMBOL(zio_buf_alloc); 5146 EXPORT_SYMBOL(zio_data_buf_alloc); 5147 EXPORT_SYMBOL(zio_buf_free); 5148 EXPORT_SYMBOL(zio_data_buf_free); 5149 5150 ZFS_MODULE_PARAM(zfs_zio, zio_, slow_io_ms, INT, ZMOD_RW, 5151 "Max I/O completion time (milliseconds) before marking it as slow"); 5152 5153 ZFS_MODULE_PARAM(zfs_zio, zio_, requeue_io_start_cut_in_line, INT, ZMOD_RW, 5154 "Prioritize requeued I/O"); 5155 5156 ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_deferred_free, UINT, ZMOD_RW, 5157 "Defer frees starting in this pass"); 5158 5159 ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_dont_compress, UINT, ZMOD_RW, 5160 "Don't compress starting in this pass"); 5161 5162 ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_rewrite, UINT, ZMOD_RW, 5163 "Rewrite new bps starting in this pass"); 5164 5165 ZFS_MODULE_PARAM(zfs_zio, zio_, dva_throttle_enabled, INT, ZMOD_RW, 5166 "Throttle block allocations in the ZIO pipeline"); 5167 5168 ZFS_MODULE_PARAM(zfs_zio, zio_, deadman_log_all, INT, ZMOD_RW, 5169 "Log all slow ZIOs, not just those with vdevs"); 5170