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