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, 2017 by Delphix. All rights reserved. 24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved. 25 * Copyright (c) 2014 Integros [integros.com] 26 */ 27 28 #include <sys/sysmacros.h> 29 #include <sys/zfs_context.h> 30 #include <sys/fm/fs/zfs.h> 31 #include <sys/spa.h> 32 #include <sys/txg.h> 33 #include <sys/spa_impl.h> 34 #include <sys/vdev_impl.h> 35 #include <sys/zio_impl.h> 36 #include <sys/zio_compress.h> 37 #include <sys/zio_checksum.h> 38 #include <sys/dmu_objset.h> 39 #include <sys/arc.h> 40 #include <sys/ddt.h> 41 #include <sys/blkptr.h> 42 #include <sys/zfeature.h> 43 #include <sys/metaslab_impl.h> 44 #include <sys/abd.h> 45 46 /* 47 * ========================================================================== 48 * I/O type descriptions 49 * ========================================================================== 50 */ 51 const char *zio_type_name[ZIO_TYPES] = { 52 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim", 53 "zio_ioctl" 54 }; 55 56 boolean_t zio_dva_throttle_enabled = B_TRUE; 57 58 /* 59 * ========================================================================== 60 * I/O kmem caches 61 * ========================================================================== 62 */ 63 kmem_cache_t *zio_cache; 64 kmem_cache_t *zio_link_cache; 65 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; 66 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; 67 68 #ifdef _KERNEL 69 extern vmem_t *zio_alloc_arena; 70 #endif 71 72 #define ZIO_PIPELINE_CONTINUE 0x100 73 #define ZIO_PIPELINE_STOP 0x101 74 75 #define BP_SPANB(indblkshift, level) \ 76 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT))) 77 #define COMPARE_META_LEVEL 0x80000000ul 78 /* 79 * The following actions directly effect the spa's sync-to-convergence logic. 80 * The values below define the sync pass when we start performing the action. 81 * Care should be taken when changing these values as they directly impact 82 * spa_sync() performance. Tuning these values may introduce subtle performance 83 * pathologies and should only be done in the context of performance analysis. 84 * These tunables will eventually be removed and replaced with #defines once 85 * enough analysis has been done to determine optimal values. 86 * 87 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that 88 * regular blocks are not deferred. 89 */ 90 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */ 91 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */ 92 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */ 93 94 /* 95 * An allocating zio is one that either currently has the DVA allocate 96 * stage set or will have it later in its lifetime. 97 */ 98 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE) 99 100 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE; 101 102 #ifdef ZFS_DEBUG 103 int zio_buf_debug_limit = 16384; 104 #else 105 int zio_buf_debug_limit = 0; 106 #endif 107 108 static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t); 109 110 void 111 zio_init(void) 112 { 113 size_t c; 114 vmem_t *data_alloc_arena = NULL; 115 116 #ifdef _KERNEL 117 data_alloc_arena = zio_alloc_arena; 118 #endif 119 zio_cache = kmem_cache_create("zio_cache", 120 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0); 121 zio_link_cache = kmem_cache_create("zio_link_cache", 122 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0); 123 124 /* 125 * For small buffers, we want a cache for each multiple of 126 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache 127 * for each quarter-power of 2. 128 */ 129 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { 130 size_t size = (c + 1) << SPA_MINBLOCKSHIFT; 131 size_t p2 = size; 132 size_t align = 0; 133 size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0; 134 135 while (!ISP2(p2)) 136 p2 &= p2 - 1; 137 138 #ifndef _KERNEL 139 /* 140 * If we are using watchpoints, put each buffer on its own page, 141 * to eliminate the performance overhead of trapping to the 142 * kernel when modifying a non-watched buffer that shares the 143 * page with a watched buffer. 144 */ 145 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE)) 146 continue; 147 #endif 148 if (size <= 4 * SPA_MINBLOCKSIZE) { 149 align = SPA_MINBLOCKSIZE; 150 } else if (IS_P2ALIGNED(size, p2 >> 2)) { 151 align = MIN(p2 >> 2, PAGESIZE); 152 } 153 154 if (align != 0) { 155 char name[36]; 156 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size); 157 zio_buf_cache[c] = kmem_cache_create(name, size, 158 align, NULL, NULL, NULL, NULL, NULL, cflags); 159 160 /* 161 * Since zio_data bufs do not appear in crash dumps, we 162 * pass KMC_NOTOUCH so that no allocator metadata is 163 * stored with the buffers. 164 */ 165 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size); 166 zio_data_buf_cache[c] = kmem_cache_create(name, size, 167 align, NULL, NULL, NULL, NULL, data_alloc_arena, 168 cflags | KMC_NOTOUCH); 169 } 170 } 171 172 while (--c != 0) { 173 ASSERT(zio_buf_cache[c] != NULL); 174 if (zio_buf_cache[c - 1] == NULL) 175 zio_buf_cache[c - 1] = zio_buf_cache[c]; 176 177 ASSERT(zio_data_buf_cache[c] != NULL); 178 if (zio_data_buf_cache[c - 1] == NULL) 179 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c]; 180 } 181 182 zio_inject_init(); 183 } 184 185 void 186 zio_fini(void) 187 { 188 size_t c; 189 kmem_cache_t *last_cache = NULL; 190 kmem_cache_t *last_data_cache = NULL; 191 192 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { 193 if (zio_buf_cache[c] != last_cache) { 194 last_cache = zio_buf_cache[c]; 195 kmem_cache_destroy(zio_buf_cache[c]); 196 } 197 zio_buf_cache[c] = NULL; 198 199 if (zio_data_buf_cache[c] != last_data_cache) { 200 last_data_cache = zio_data_buf_cache[c]; 201 kmem_cache_destroy(zio_data_buf_cache[c]); 202 } 203 zio_data_buf_cache[c] = NULL; 204 } 205 206 kmem_cache_destroy(zio_link_cache); 207 kmem_cache_destroy(zio_cache); 208 209 zio_inject_fini(); 210 } 211 212 /* 213 * ========================================================================== 214 * Allocate and free I/O buffers 215 * ========================================================================== 216 */ 217 218 /* 219 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a 220 * crashdump if the kernel panics, so use it judiciously. Obviously, it's 221 * useful to inspect ZFS metadata, but if possible, we should avoid keeping 222 * excess / transient data in-core during a crashdump. 223 */ 224 void * 225 zio_buf_alloc(size_t size) 226 { 227 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 228 229 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 230 231 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE)); 232 } 233 234 /* 235 * Use zio_data_buf_alloc to allocate data. The data will not appear in a 236 * crashdump if the kernel panics. This exists so that we will limit the amount 237 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount 238 * of kernel heap dumped to disk when the kernel panics) 239 */ 240 void * 241 zio_data_buf_alloc(size_t size) 242 { 243 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 244 245 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 246 247 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE)); 248 } 249 250 void 251 zio_buf_free(void *buf, size_t size) 252 { 253 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 254 255 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 256 257 kmem_cache_free(zio_buf_cache[c], buf); 258 } 259 260 void 261 zio_data_buf_free(void *buf, size_t size) 262 { 263 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 264 265 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 266 267 kmem_cache_free(zio_data_buf_cache[c], buf); 268 } 269 270 /* 271 * ========================================================================== 272 * Push and pop I/O transform buffers 273 * ========================================================================== 274 */ 275 void 276 zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize, 277 zio_transform_func_t *transform) 278 { 279 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP); 280 281 /* 282 * Ensure that anyone expecting this zio to contain a linear ABD isn't 283 * going to get a nasty surprise when they try to access the data. 284 */ 285 IMPLY(abd_is_linear(zio->io_abd), abd_is_linear(data)); 286 287 zt->zt_orig_abd = zio->io_abd; 288 zt->zt_orig_size = zio->io_size; 289 zt->zt_bufsize = bufsize; 290 zt->zt_transform = transform; 291 292 zt->zt_next = zio->io_transform_stack; 293 zio->io_transform_stack = zt; 294 295 zio->io_abd = data; 296 zio->io_size = size; 297 } 298 299 void 300 zio_pop_transforms(zio_t *zio) 301 { 302 zio_transform_t *zt; 303 304 while ((zt = zio->io_transform_stack) != NULL) { 305 if (zt->zt_transform != NULL) 306 zt->zt_transform(zio, 307 zt->zt_orig_abd, zt->zt_orig_size); 308 309 if (zt->zt_bufsize != 0) 310 abd_free(zio->io_abd); 311 312 zio->io_abd = zt->zt_orig_abd; 313 zio->io_size = zt->zt_orig_size; 314 zio->io_transform_stack = zt->zt_next; 315 316 kmem_free(zt, sizeof (zio_transform_t)); 317 } 318 } 319 320 /* 321 * ========================================================================== 322 * I/O transform callbacks for subblocks and decompression 323 * ========================================================================== 324 */ 325 static void 326 zio_subblock(zio_t *zio, abd_t *data, uint64_t size) 327 { 328 ASSERT(zio->io_size > size); 329 330 if (zio->io_type == ZIO_TYPE_READ) 331 abd_copy(data, zio->io_abd, size); 332 } 333 334 static void 335 zio_decompress(zio_t *zio, abd_t *data, uint64_t size) 336 { 337 if (zio->io_error == 0) { 338 void *tmp = abd_borrow_buf(data, size); 339 int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp), 340 zio->io_abd, tmp, zio->io_size, size); 341 abd_return_buf_copy(data, tmp, size); 342 343 if (ret != 0) 344 zio->io_error = SET_ERROR(EIO); 345 } 346 } 347 348 /* 349 * ========================================================================== 350 * I/O parent/child relationships and pipeline interlocks 351 * ========================================================================== 352 */ 353 zio_t * 354 zio_walk_parents(zio_t *cio, zio_link_t **zl) 355 { 356 list_t *pl = &cio->io_parent_list; 357 358 *zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl); 359 if (*zl == NULL) 360 return (NULL); 361 362 ASSERT((*zl)->zl_child == cio); 363 return ((*zl)->zl_parent); 364 } 365 366 zio_t * 367 zio_walk_children(zio_t *pio, zio_link_t **zl) 368 { 369 list_t *cl = &pio->io_child_list; 370 371 *zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl); 372 if (*zl == NULL) 373 return (NULL); 374 375 ASSERT((*zl)->zl_parent == pio); 376 return ((*zl)->zl_child); 377 } 378 379 zio_t * 380 zio_unique_parent(zio_t *cio) 381 { 382 zio_link_t *zl = NULL; 383 zio_t *pio = zio_walk_parents(cio, &zl); 384 385 VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL); 386 return (pio); 387 } 388 389 void 390 zio_add_child(zio_t *pio, zio_t *cio) 391 { 392 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP); 393 394 /* 395 * Logical I/Os can have logical, gang, or vdev children. 396 * Gang I/Os can have gang or vdev children. 397 * Vdev I/Os can only have vdev children. 398 * The following ASSERT captures all of these constraints. 399 */ 400 ASSERT3S(cio->io_child_type, <=, pio->io_child_type); 401 402 zl->zl_parent = pio; 403 zl->zl_child = cio; 404 405 mutex_enter(&cio->io_lock); 406 mutex_enter(&pio->io_lock); 407 408 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0); 409 410 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 411 pio->io_children[cio->io_child_type][w] += !cio->io_state[w]; 412 413 list_insert_head(&pio->io_child_list, zl); 414 list_insert_head(&cio->io_parent_list, zl); 415 416 pio->io_child_count++; 417 cio->io_parent_count++; 418 419 mutex_exit(&pio->io_lock); 420 mutex_exit(&cio->io_lock); 421 } 422 423 static void 424 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl) 425 { 426 ASSERT(zl->zl_parent == pio); 427 ASSERT(zl->zl_child == cio); 428 429 mutex_enter(&cio->io_lock); 430 mutex_enter(&pio->io_lock); 431 432 list_remove(&pio->io_child_list, zl); 433 list_remove(&cio->io_parent_list, zl); 434 435 pio->io_child_count--; 436 cio->io_parent_count--; 437 438 mutex_exit(&pio->io_lock); 439 mutex_exit(&cio->io_lock); 440 441 kmem_cache_free(zio_link_cache, zl); 442 } 443 444 static boolean_t 445 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait) 446 { 447 uint64_t *countp = &zio->io_children[child][wait]; 448 boolean_t waiting = B_FALSE; 449 450 mutex_enter(&zio->io_lock); 451 ASSERT(zio->io_stall == NULL); 452 if (*countp != 0) { 453 zio->io_stage >>= 1; 454 ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN); 455 zio->io_stall = countp; 456 waiting = B_TRUE; 457 } 458 mutex_exit(&zio->io_lock); 459 460 return (waiting); 461 } 462 463 static void 464 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait) 465 { 466 uint64_t *countp = &pio->io_children[zio->io_child_type][wait]; 467 int *errorp = &pio->io_child_error[zio->io_child_type]; 468 469 mutex_enter(&pio->io_lock); 470 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE)) 471 *errorp = zio_worst_error(*errorp, zio->io_error); 472 pio->io_reexecute |= zio->io_reexecute; 473 ASSERT3U(*countp, >, 0); 474 475 (*countp)--; 476 477 if (*countp == 0 && pio->io_stall == countp) { 478 zio_taskq_type_t type = 479 pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE : 480 ZIO_TASKQ_INTERRUPT; 481 pio->io_stall = NULL; 482 mutex_exit(&pio->io_lock); 483 /* 484 * Dispatch the parent zio in its own taskq so that 485 * the child can continue to make progress. This also 486 * prevents overflowing the stack when we have deeply nested 487 * parent-child relationships. 488 */ 489 zio_taskq_dispatch(pio, type, B_FALSE); 490 } else { 491 mutex_exit(&pio->io_lock); 492 } 493 } 494 495 static void 496 zio_inherit_child_errors(zio_t *zio, enum zio_child c) 497 { 498 if (zio->io_child_error[c] != 0 && zio->io_error == 0) 499 zio->io_error = zio->io_child_error[c]; 500 } 501 502 int 503 zio_bookmark_compare(const void *x1, const void *x2) 504 { 505 const zio_t *z1 = x1; 506 const zio_t *z2 = x2; 507 508 if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset) 509 return (-1); 510 if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset) 511 return (1); 512 513 if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object) 514 return (-1); 515 if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object) 516 return (1); 517 518 if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level) 519 return (-1); 520 if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level) 521 return (1); 522 523 if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid) 524 return (-1); 525 if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid) 526 return (1); 527 528 if (z1 < z2) 529 return (-1); 530 if (z1 > z2) 531 return (1); 532 533 return (0); 534 } 535 536 /* 537 * ========================================================================== 538 * Create the various types of I/O (read, write, free, etc) 539 * ========================================================================== 540 */ 541 static zio_t * 542 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 543 abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done, 544 void *private, zio_type_t type, zio_priority_t priority, 545 enum zio_flag flags, vdev_t *vd, uint64_t offset, 546 const zbookmark_phys_t *zb, enum zio_stage stage, enum zio_stage pipeline) 547 { 548 zio_t *zio; 549 550 ASSERT3U(psize, <=, SPA_MAXBLOCKSIZE); 551 ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0); 552 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0); 553 554 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER)); 555 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER)); 556 ASSERT(vd || stage == ZIO_STAGE_OPEN); 557 558 IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW) != 0); 559 560 zio = kmem_cache_alloc(zio_cache, KM_SLEEP); 561 bzero(zio, sizeof (zio_t)); 562 563 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL); 564 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL); 565 566 list_create(&zio->io_parent_list, sizeof (zio_link_t), 567 offsetof(zio_link_t, zl_parent_node)); 568 list_create(&zio->io_child_list, sizeof (zio_link_t), 569 offsetof(zio_link_t, zl_child_node)); 570 metaslab_trace_init(&zio->io_alloc_list); 571 572 if (vd != NULL) 573 zio->io_child_type = ZIO_CHILD_VDEV; 574 else if (flags & ZIO_FLAG_GANG_CHILD) 575 zio->io_child_type = ZIO_CHILD_GANG; 576 else if (flags & ZIO_FLAG_DDT_CHILD) 577 zio->io_child_type = ZIO_CHILD_DDT; 578 else 579 zio->io_child_type = ZIO_CHILD_LOGICAL; 580 581 if (bp != NULL) { 582 zio->io_bp = (blkptr_t *)bp; 583 zio->io_bp_copy = *bp; 584 zio->io_bp_orig = *bp; 585 if (type != ZIO_TYPE_WRITE || 586 zio->io_child_type == ZIO_CHILD_DDT) 587 zio->io_bp = &zio->io_bp_copy; /* so caller can free */ 588 if (zio->io_child_type == ZIO_CHILD_LOGICAL) 589 zio->io_logical = zio; 590 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp)) 591 pipeline |= ZIO_GANG_STAGES; 592 } 593 594 zio->io_spa = spa; 595 zio->io_txg = txg; 596 zio->io_done = done; 597 zio->io_private = private; 598 zio->io_type = type; 599 zio->io_priority = priority; 600 zio->io_vd = vd; 601 zio->io_offset = offset; 602 zio->io_orig_abd = zio->io_abd = data; 603 zio->io_orig_size = zio->io_size = psize; 604 zio->io_lsize = lsize; 605 zio->io_orig_flags = zio->io_flags = flags; 606 zio->io_orig_stage = zio->io_stage = stage; 607 zio->io_orig_pipeline = zio->io_pipeline = pipeline; 608 zio->io_pipeline_trace = ZIO_STAGE_OPEN; 609 610 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY); 611 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE); 612 613 if (zb != NULL) 614 zio->io_bookmark = *zb; 615 616 if (pio != NULL) { 617 if (zio->io_logical == NULL) 618 zio->io_logical = pio->io_logical; 619 if (zio->io_child_type == ZIO_CHILD_GANG) 620 zio->io_gang_leader = pio->io_gang_leader; 621 zio_add_child(pio, zio); 622 } 623 624 return (zio); 625 } 626 627 static void 628 zio_destroy(zio_t *zio) 629 { 630 metaslab_trace_fini(&zio->io_alloc_list); 631 list_destroy(&zio->io_parent_list); 632 list_destroy(&zio->io_child_list); 633 mutex_destroy(&zio->io_lock); 634 cv_destroy(&zio->io_cv); 635 kmem_cache_free(zio_cache, zio); 636 } 637 638 zio_t * 639 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done, 640 void *private, enum zio_flag flags) 641 { 642 zio_t *zio; 643 644 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private, 645 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, 646 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE); 647 648 return (zio); 649 } 650 651 zio_t * 652 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags) 653 { 654 return (zio_null(NULL, spa, NULL, done, private, flags)); 655 } 656 657 void 658 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp) 659 { 660 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) { 661 zfs_panic_recover("blkptr at %p has invalid TYPE %llu", 662 bp, (longlong_t)BP_GET_TYPE(bp)); 663 } 664 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS || 665 BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) { 666 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu", 667 bp, (longlong_t)BP_GET_CHECKSUM(bp)); 668 } 669 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS || 670 BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) { 671 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu", 672 bp, (longlong_t)BP_GET_COMPRESS(bp)); 673 } 674 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) { 675 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu", 676 bp, (longlong_t)BP_GET_LSIZE(bp)); 677 } 678 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) { 679 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu", 680 bp, (longlong_t)BP_GET_PSIZE(bp)); 681 } 682 683 if (BP_IS_EMBEDDED(bp)) { 684 if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) { 685 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu", 686 bp, (longlong_t)BPE_GET_ETYPE(bp)); 687 } 688 } 689 690 /* 691 * Do not verify individual DVAs if the config is not trusted. This 692 * will be done once the zio is executed in vdev_mirror_map_alloc. 693 */ 694 if (!spa->spa_trust_config) 695 return; 696 697 /* 698 * Pool-specific checks. 699 * 700 * Note: it would be nice to verify that the blk_birth and 701 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze() 702 * allows the birth time of log blocks (and dmu_sync()-ed blocks 703 * that are in the log) to be arbitrarily large. 704 */ 705 for (int i = 0; i < BP_GET_NDVAS(bp); i++) { 706 uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]); 707 if (vdevid >= spa->spa_root_vdev->vdev_children) { 708 zfs_panic_recover("blkptr at %p DVA %u has invalid " 709 "VDEV %llu", 710 bp, i, (longlong_t)vdevid); 711 continue; 712 } 713 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid]; 714 if (vd == NULL) { 715 zfs_panic_recover("blkptr at %p DVA %u has invalid " 716 "VDEV %llu", 717 bp, i, (longlong_t)vdevid); 718 continue; 719 } 720 if (vd->vdev_ops == &vdev_hole_ops) { 721 zfs_panic_recover("blkptr at %p DVA %u has hole " 722 "VDEV %llu", 723 bp, i, (longlong_t)vdevid); 724 continue; 725 } 726 if (vd->vdev_ops == &vdev_missing_ops) { 727 /* 728 * "missing" vdevs are valid during import, but we 729 * don't have their detailed info (e.g. asize), so 730 * we can't perform any more checks on them. 731 */ 732 continue; 733 } 734 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]); 735 uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]); 736 if (BP_IS_GANG(bp)) 737 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE); 738 if (offset + asize > vd->vdev_asize) { 739 zfs_panic_recover("blkptr at %p DVA %u has invalid " 740 "OFFSET %llu", 741 bp, i, (longlong_t)offset); 742 } 743 } 744 } 745 746 boolean_t 747 zfs_dva_valid(spa_t *spa, const dva_t *dva, const blkptr_t *bp) 748 { 749 uint64_t vdevid = DVA_GET_VDEV(dva); 750 751 if (vdevid >= spa->spa_root_vdev->vdev_children) 752 return (B_FALSE); 753 754 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid]; 755 if (vd == NULL) 756 return (B_FALSE); 757 758 if (vd->vdev_ops == &vdev_hole_ops) 759 return (B_FALSE); 760 761 if (vd->vdev_ops == &vdev_missing_ops) { 762 return (B_FALSE); 763 } 764 765 uint64_t offset = DVA_GET_OFFSET(dva); 766 uint64_t asize = DVA_GET_ASIZE(dva); 767 768 if (BP_IS_GANG(bp)) 769 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE); 770 if (offset + asize > vd->vdev_asize) 771 return (B_FALSE); 772 773 return (B_TRUE); 774 } 775 776 zio_t * 777 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, 778 abd_t *data, uint64_t size, zio_done_func_t *done, void *private, 779 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb) 780 { 781 zio_t *zio; 782 783 zfs_blkptr_verify(spa, bp); 784 785 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp, 786 data, size, size, done, private, 787 ZIO_TYPE_READ, priority, flags, NULL, 0, zb, 788 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? 789 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE); 790 791 return (zio); 792 } 793 794 zio_t * 795 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, 796 abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp, 797 zio_done_func_t *ready, zio_done_func_t *children_ready, 798 zio_done_func_t *physdone, zio_done_func_t *done, 799 void *private, zio_priority_t priority, enum zio_flag flags, 800 const zbookmark_phys_t *zb) 801 { 802 zio_t *zio; 803 804 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF && 805 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS && 806 zp->zp_compress >= ZIO_COMPRESS_OFF && 807 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS && 808 DMU_OT_IS_VALID(zp->zp_type) && 809 zp->zp_level < 32 && 810 zp->zp_copies > 0 && 811 zp->zp_copies <= spa_max_replication(spa)); 812 813 zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private, 814 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, 815 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? 816 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE); 817 818 zio->io_ready = ready; 819 zio->io_children_ready = children_ready; 820 zio->io_physdone = physdone; 821 zio->io_prop = *zp; 822 823 /* 824 * Data can be NULL if we are going to call zio_write_override() to 825 * provide the already-allocated BP. But we may need the data to 826 * verify a dedup hit (if requested). In this case, don't try to 827 * dedup (just take the already-allocated BP verbatim). 828 */ 829 if (data == NULL && zio->io_prop.zp_dedup_verify) { 830 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE; 831 } 832 833 return (zio); 834 } 835 836 zio_t * 837 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data, 838 uint64_t size, zio_done_func_t *done, void *private, 839 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb) 840 { 841 zio_t *zio; 842 843 zio = zio_create(pio, spa, txg, bp, data, size, size, done, private, 844 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb, 845 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE); 846 847 return (zio); 848 } 849 850 void 851 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite) 852 { 853 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 854 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 855 ASSERT(zio->io_stage == ZIO_STAGE_OPEN); 856 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa)); 857 858 /* 859 * We must reset the io_prop to match the values that existed 860 * when the bp was first written by dmu_sync() keeping in mind 861 * that nopwrite and dedup are mutually exclusive. 862 */ 863 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup; 864 zio->io_prop.zp_nopwrite = nopwrite; 865 zio->io_prop.zp_copies = copies; 866 zio->io_bp_override = bp; 867 } 868 869 void 870 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp) 871 { 872 873 zfs_blkptr_verify(spa, bp); 874 875 /* 876 * The check for EMBEDDED is a performance optimization. We 877 * process the free here (by ignoring it) rather than 878 * putting it on the list and then processing it in zio_free_sync(). 879 */ 880 if (BP_IS_EMBEDDED(bp)) 881 return; 882 metaslab_check_free(spa, bp); 883 884 /* 885 * Frees that are for the currently-syncing txg, are not going to be 886 * deferred, and which will not need to do a read (i.e. not GANG or 887 * DEDUP), can be processed immediately. Otherwise, put them on the 888 * in-memory list for later processing. 889 */ 890 if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp) || 891 txg != spa->spa_syncing_txg || 892 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) { 893 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp); 894 } else { 895 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 0))); 896 } 897 } 898 899 zio_t * 900 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 901 enum zio_flag flags) 902 { 903 zio_t *zio; 904 enum zio_stage stage = ZIO_FREE_PIPELINE; 905 906 ASSERT(!BP_IS_HOLE(bp)); 907 ASSERT(spa_syncing_txg(spa) == txg); 908 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free); 909 910 if (BP_IS_EMBEDDED(bp)) 911 return (zio_null(pio, spa, NULL, NULL, NULL, 0)); 912 913 metaslab_check_free(spa, bp); 914 arc_freed(spa, bp); 915 916 /* 917 * GANG and DEDUP blocks can induce a read (for the gang block header, 918 * or the DDT), so issue them asynchronously so that this thread is 919 * not tied up. 920 */ 921 if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp)) 922 stage |= ZIO_STAGE_ISSUE_ASYNC; 923 924 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), 925 BP_GET_PSIZE(bp), NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, 926 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage); 927 928 return (zio); 929 } 930 931 zio_t * 932 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 933 zio_done_func_t *done, void *private, enum zio_flag flags) 934 { 935 zio_t *zio; 936 937 zfs_blkptr_verify(spa, bp); 938 939 if (BP_IS_EMBEDDED(bp)) 940 return (zio_null(pio, spa, NULL, NULL, NULL, 0)); 941 942 /* 943 * A claim is an allocation of a specific block. Claims are needed 944 * to support immediate writes in the intent log. The issue is that 945 * immediate writes contain committed data, but in a txg that was 946 * *not* committed. Upon opening the pool after an unclean shutdown, 947 * the intent log claims all blocks that contain immediate write data 948 * so that the SPA knows they're in use. 949 * 950 * All claims *must* be resolved in the first txg -- before the SPA 951 * starts allocating blocks -- so that nothing is allocated twice. 952 * If txg == 0 we just verify that the block is claimable. 953 */ 954 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa)); 955 ASSERT(txg == spa_first_txg(spa) || txg == 0); 956 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */ 957 958 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), 959 BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, 960 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE); 961 ASSERT0(zio->io_queued_timestamp); 962 963 return (zio); 964 } 965 966 zio_t * 967 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, 968 zio_done_func_t *done, void *private, enum zio_flag flags) 969 { 970 zio_t *zio; 971 int c; 972 973 if (vd->vdev_children == 0) { 974 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private, 975 ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, 976 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE); 977 978 zio->io_cmd = cmd; 979 } else { 980 zio = zio_null(pio, spa, NULL, NULL, NULL, flags); 981 982 for (c = 0; c < vd->vdev_children; c++) 983 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd, 984 done, private, flags)); 985 } 986 987 return (zio); 988 } 989 990 zio_t * 991 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 992 abd_t *data, int checksum, zio_done_func_t *done, void *private, 993 zio_priority_t priority, enum zio_flag flags, boolean_t labels) 994 { 995 zio_t *zio; 996 997 ASSERT(vd->vdev_children == 0); 998 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || 999 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); 1000 ASSERT3U(offset + size, <=, vd->vdev_psize); 1001 1002 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done, 1003 private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, 1004 offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE); 1005 1006 zio->io_prop.zp_checksum = checksum; 1007 1008 return (zio); 1009 } 1010 1011 zio_t * 1012 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 1013 abd_t *data, int checksum, zio_done_func_t *done, void *private, 1014 zio_priority_t priority, enum zio_flag flags, boolean_t labels) 1015 { 1016 zio_t *zio; 1017 1018 ASSERT(vd->vdev_children == 0); 1019 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || 1020 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); 1021 ASSERT3U(offset + size, <=, vd->vdev_psize); 1022 1023 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done, 1024 private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, 1025 offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE); 1026 1027 zio->io_prop.zp_checksum = checksum; 1028 1029 if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) { 1030 /* 1031 * zec checksums are necessarily destructive -- they modify 1032 * the end of the write buffer to hold the verifier/checksum. 1033 * Therefore, we must make a local copy in case the data is 1034 * being written to multiple places in parallel. 1035 */ 1036 abd_t *wbuf = abd_alloc_sametype(data, size); 1037 abd_copy(wbuf, data, size); 1038 1039 zio_push_transform(zio, wbuf, size, size, NULL); 1040 } 1041 1042 return (zio); 1043 } 1044 1045 /* 1046 * Create a child I/O to do some work for us. 1047 */ 1048 zio_t * 1049 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset, 1050 abd_t *data, uint64_t size, int type, zio_priority_t priority, 1051 enum zio_flag flags, zio_done_func_t *done, void *private) 1052 { 1053 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE; 1054 zio_t *zio; 1055 1056 /* 1057 * vdev child I/Os do not propagate their error to the parent. 1058 * Therefore, for correct operation the caller *must* check for 1059 * and handle the error in the child i/o's done callback. 1060 * The only exceptions are i/os that we don't care about 1061 * (OPTIONAL or REPAIR). 1062 */ 1063 ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) || 1064 done != NULL); 1065 1066 /* 1067 * In the common case, where the parent zio was to a normal vdev, 1068 * the child zio must be to a child vdev of that vdev. Otherwise, 1069 * the child zio must be to a top-level vdev. 1070 */ 1071 if (pio->io_vd != NULL && pio->io_vd->vdev_ops != &vdev_indirect_ops) { 1072 ASSERT3P(vd->vdev_parent, ==, pio->io_vd); 1073 } else { 1074 ASSERT3P(vd, ==, vd->vdev_top); 1075 } 1076 1077 if (type == ZIO_TYPE_READ && bp != NULL) { 1078 /* 1079 * If we have the bp, then the child should perform the 1080 * checksum and the parent need not. This pushes error 1081 * detection as close to the leaves as possible and 1082 * eliminates redundant checksums in the interior nodes. 1083 */ 1084 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY; 1085 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; 1086 } 1087 1088 if (vd->vdev_ops->vdev_op_leaf) { 1089 ASSERT0(vd->vdev_children); 1090 offset += VDEV_LABEL_START_SIZE; 1091 } 1092 1093 flags |= ZIO_VDEV_CHILD_FLAGS(pio); 1094 1095 /* 1096 * If we've decided to do a repair, the write is not speculative -- 1097 * even if the original read was. 1098 */ 1099 if (flags & ZIO_FLAG_IO_REPAIR) 1100 flags &= ~ZIO_FLAG_SPECULATIVE; 1101 1102 /* 1103 * If we're creating a child I/O that is not associated with a 1104 * top-level vdev, then the child zio is not an allocating I/O. 1105 * If this is a retried I/O then we ignore it since we will 1106 * have already processed the original allocating I/O. 1107 */ 1108 if (flags & ZIO_FLAG_IO_ALLOCATING && 1109 (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) { 1110 metaslab_class_t *mc = spa_normal_class(pio->io_spa); 1111 1112 ASSERT(mc->mc_alloc_throttle_enabled); 1113 ASSERT(type == ZIO_TYPE_WRITE); 1114 ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE); 1115 ASSERT(!(flags & ZIO_FLAG_IO_REPAIR)); 1116 ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) || 1117 pio->io_child_type == ZIO_CHILD_GANG); 1118 1119 flags &= ~ZIO_FLAG_IO_ALLOCATING; 1120 } 1121 1122 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size, 1123 done, private, type, priority, flags, vd, offset, &pio->io_bookmark, 1124 ZIO_STAGE_VDEV_IO_START >> 1, pipeline); 1125 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV); 1126 1127 zio->io_physdone = pio->io_physdone; 1128 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL) 1129 zio->io_logical->io_phys_children++; 1130 1131 return (zio); 1132 } 1133 1134 zio_t * 1135 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size, 1136 int type, zio_priority_t priority, enum zio_flag flags, 1137 zio_done_func_t *done, void *private) 1138 { 1139 zio_t *zio; 1140 1141 ASSERT(vd->vdev_ops->vdev_op_leaf); 1142 1143 zio = zio_create(NULL, vd->vdev_spa, 0, NULL, 1144 data, size, size, done, private, type, priority, 1145 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED, 1146 vd, offset, NULL, 1147 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE); 1148 1149 return (zio); 1150 } 1151 1152 void 1153 zio_flush(zio_t *zio, vdev_t *vd) 1154 { 1155 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 1156 NULL, NULL, 1157 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY)); 1158 } 1159 1160 void 1161 zio_shrink(zio_t *zio, uint64_t size) 1162 { 1163 ASSERT3P(zio->io_executor, ==, NULL); 1164 ASSERT3P(zio->io_orig_size, ==, zio->io_size); 1165 ASSERT3U(size, <=, zio->io_size); 1166 1167 /* 1168 * We don't shrink for raidz because of problems with the 1169 * reconstruction when reading back less than the block size. 1170 * Note, BP_IS_RAIDZ() assumes no compression. 1171 */ 1172 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 1173 if (!BP_IS_RAIDZ(zio->io_bp)) { 1174 /* we are not doing a raw write */ 1175 ASSERT3U(zio->io_size, ==, zio->io_lsize); 1176 zio->io_orig_size = zio->io_size = zio->io_lsize = size; 1177 } 1178 } 1179 1180 /* 1181 * ========================================================================== 1182 * Prepare to read and write logical blocks 1183 * ========================================================================== 1184 */ 1185 1186 static int 1187 zio_read_bp_init(zio_t *zio) 1188 { 1189 blkptr_t *bp = zio->io_bp; 1190 1191 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy); 1192 1193 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF && 1194 zio->io_child_type == ZIO_CHILD_LOGICAL && 1195 !(zio->io_flags & ZIO_FLAG_RAW)) { 1196 uint64_t psize = 1197 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp); 1198 zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize), 1199 psize, psize, zio_decompress); 1200 } 1201 1202 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) { 1203 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1204 1205 int psize = BPE_GET_PSIZE(bp); 1206 void *data = abd_borrow_buf(zio->io_abd, psize); 1207 decode_embedded_bp_compressed(bp, data); 1208 abd_return_buf_copy(zio->io_abd, data, psize); 1209 } else { 1210 ASSERT(!BP_IS_EMBEDDED(bp)); 1211 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy); 1212 } 1213 1214 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0) 1215 zio->io_flags |= ZIO_FLAG_DONT_CACHE; 1216 1217 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP) 1218 zio->io_flags |= ZIO_FLAG_DONT_CACHE; 1219 1220 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL) 1221 zio->io_pipeline = ZIO_DDT_READ_PIPELINE; 1222 1223 return (ZIO_PIPELINE_CONTINUE); 1224 } 1225 1226 static int 1227 zio_write_bp_init(zio_t *zio) 1228 { 1229 if (!IO_IS_ALLOCATING(zio)) 1230 return (ZIO_PIPELINE_CONTINUE); 1231 1232 ASSERT(zio->io_child_type != ZIO_CHILD_DDT); 1233 1234 if (zio->io_bp_override) { 1235 blkptr_t *bp = zio->io_bp; 1236 zio_prop_t *zp = &zio->io_prop; 1237 1238 ASSERT(bp->blk_birth != zio->io_txg); 1239 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0); 1240 1241 *bp = *zio->io_bp_override; 1242 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1243 1244 if (BP_IS_EMBEDDED(bp)) 1245 return (ZIO_PIPELINE_CONTINUE); 1246 1247 /* 1248 * If we've been overridden and nopwrite is set then 1249 * set the flag accordingly to indicate that a nopwrite 1250 * has already occurred. 1251 */ 1252 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) { 1253 ASSERT(!zp->zp_dedup); 1254 ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum); 1255 zio->io_flags |= ZIO_FLAG_NOPWRITE; 1256 return (ZIO_PIPELINE_CONTINUE); 1257 } 1258 1259 ASSERT(!zp->zp_nopwrite); 1260 1261 if (BP_IS_HOLE(bp) || !zp->zp_dedup) 1262 return (ZIO_PIPELINE_CONTINUE); 1263 1264 ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags & 1265 ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify); 1266 1267 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) { 1268 BP_SET_DEDUP(bp, 1); 1269 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE; 1270 return (ZIO_PIPELINE_CONTINUE); 1271 } 1272 1273 /* 1274 * We were unable to handle this as an override bp, treat 1275 * it as a regular write I/O. 1276 */ 1277 zio->io_bp_override = NULL; 1278 *bp = zio->io_bp_orig; 1279 zio->io_pipeline = zio->io_orig_pipeline; 1280 } 1281 1282 return (ZIO_PIPELINE_CONTINUE); 1283 } 1284 1285 static int 1286 zio_write_compress(zio_t *zio) 1287 { 1288 spa_t *spa = zio->io_spa; 1289 zio_prop_t *zp = &zio->io_prop; 1290 enum zio_compress compress = zp->zp_compress; 1291 blkptr_t *bp = zio->io_bp; 1292 uint64_t lsize = zio->io_lsize; 1293 uint64_t psize = zio->io_size; 1294 int pass = 1; 1295 1296 EQUIV(lsize != psize, (zio->io_flags & ZIO_FLAG_RAW) != 0); 1297 1298 /* 1299 * If our children haven't all reached the ready stage, 1300 * wait for them and then repeat this pipeline stage. 1301 */ 1302 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) || 1303 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY)) 1304 return (ZIO_PIPELINE_STOP); 1305 1306 if (!IO_IS_ALLOCATING(zio)) 1307 return (ZIO_PIPELINE_CONTINUE); 1308 1309 if (zio->io_children_ready != NULL) { 1310 /* 1311 * Now that all our children are ready, run the callback 1312 * associated with this zio in case it wants to modify the 1313 * data to be written. 1314 */ 1315 ASSERT3U(zp->zp_level, >, 0); 1316 zio->io_children_ready(zio); 1317 } 1318 1319 ASSERT(zio->io_child_type != ZIO_CHILD_DDT); 1320 ASSERT(zio->io_bp_override == NULL); 1321 1322 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) { 1323 /* 1324 * We're rewriting an existing block, which means we're 1325 * working on behalf of spa_sync(). For spa_sync() to 1326 * converge, it must eventually be the case that we don't 1327 * have to allocate new blocks. But compression changes 1328 * the blocksize, which forces a reallocate, and makes 1329 * convergence take longer. Therefore, after the first 1330 * few passes, stop compressing to ensure convergence. 1331 */ 1332 pass = spa_sync_pass(spa); 1333 1334 ASSERT(zio->io_txg == spa_syncing_txg(spa)); 1335 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1336 ASSERT(!BP_GET_DEDUP(bp)); 1337 1338 if (pass >= zfs_sync_pass_dont_compress) 1339 compress = ZIO_COMPRESS_OFF; 1340 1341 /* Make sure someone doesn't change their mind on overwrites */ 1342 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp), 1343 spa_max_replication(spa)) == BP_GET_NDVAS(bp)); 1344 } 1345 1346 /* If it's a compressed write that is not raw, compress the buffer. */ 1347 if (compress != ZIO_COMPRESS_OFF && psize == lsize) { 1348 void *cbuf = zio_buf_alloc(lsize); 1349 psize = zio_compress_data(compress, zio->io_abd, cbuf, lsize); 1350 if (psize == 0 || psize == lsize) { 1351 compress = ZIO_COMPRESS_OFF; 1352 zio_buf_free(cbuf, lsize); 1353 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE && 1354 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) && 1355 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) { 1356 encode_embedded_bp_compressed(bp, 1357 cbuf, compress, lsize, psize); 1358 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA); 1359 BP_SET_TYPE(bp, zio->io_prop.zp_type); 1360 BP_SET_LEVEL(bp, zio->io_prop.zp_level); 1361 zio_buf_free(cbuf, lsize); 1362 bp->blk_birth = zio->io_txg; 1363 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1364 ASSERT(spa_feature_is_active(spa, 1365 SPA_FEATURE_EMBEDDED_DATA)); 1366 return (ZIO_PIPELINE_CONTINUE); 1367 } else { 1368 /* 1369 * Round up compressed size up to the ashift 1370 * of the smallest-ashift device, and zero the tail. 1371 * This ensures that the compressed size of the BP 1372 * (and thus compressratio property) are correct, 1373 * in that we charge for the padding used to fill out 1374 * the last sector. 1375 */ 1376 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT); 1377 size_t rounded = (size_t)P2ROUNDUP(psize, 1378 1ULL << spa->spa_min_ashift); 1379 if (rounded >= lsize) { 1380 compress = ZIO_COMPRESS_OFF; 1381 zio_buf_free(cbuf, lsize); 1382 psize = lsize; 1383 } else { 1384 abd_t *cdata = abd_get_from_buf(cbuf, lsize); 1385 abd_take_ownership_of_buf(cdata, B_TRUE); 1386 abd_zero_off(cdata, psize, rounded - psize); 1387 psize = rounded; 1388 zio_push_transform(zio, cdata, 1389 psize, lsize, NULL); 1390 } 1391 } 1392 1393 /* 1394 * We were unable to handle this as an override bp, treat 1395 * it as a regular write I/O. 1396 */ 1397 zio->io_bp_override = NULL; 1398 *bp = zio->io_bp_orig; 1399 zio->io_pipeline = zio->io_orig_pipeline; 1400 } else { 1401 ASSERT3U(psize, !=, 0); 1402 } 1403 1404 /* 1405 * The final pass of spa_sync() must be all rewrites, but the first 1406 * few passes offer a trade-off: allocating blocks defers convergence, 1407 * but newly allocated blocks are sequential, so they can be written 1408 * to disk faster. Therefore, we allow the first few passes of 1409 * spa_sync() to allocate new blocks, but force rewrites after that. 1410 * There should only be a handful of blocks after pass 1 in any case. 1411 */ 1412 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg && 1413 BP_GET_PSIZE(bp) == psize && 1414 pass >= zfs_sync_pass_rewrite) { 1415 ASSERT(psize != 0); 1416 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES; 1417 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages; 1418 zio->io_flags |= ZIO_FLAG_IO_REWRITE; 1419 } else { 1420 BP_ZERO(bp); 1421 zio->io_pipeline = ZIO_WRITE_PIPELINE; 1422 } 1423 1424 if (psize == 0) { 1425 if (zio->io_bp_orig.blk_birth != 0 && 1426 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) { 1427 BP_SET_LSIZE(bp, lsize); 1428 BP_SET_TYPE(bp, zp->zp_type); 1429 BP_SET_LEVEL(bp, zp->zp_level); 1430 BP_SET_BIRTH(bp, zio->io_txg, 0); 1431 } 1432 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1433 } else { 1434 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER); 1435 BP_SET_LSIZE(bp, lsize); 1436 BP_SET_TYPE(bp, zp->zp_type); 1437 BP_SET_LEVEL(bp, zp->zp_level); 1438 BP_SET_PSIZE(bp, psize); 1439 BP_SET_COMPRESS(bp, compress); 1440 BP_SET_CHECKSUM(bp, zp->zp_checksum); 1441 BP_SET_DEDUP(bp, zp->zp_dedup); 1442 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER); 1443 if (zp->zp_dedup) { 1444 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1445 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 1446 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE; 1447 } 1448 if (zp->zp_nopwrite) { 1449 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1450 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 1451 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE; 1452 } 1453 } 1454 return (ZIO_PIPELINE_CONTINUE); 1455 } 1456 1457 static int 1458 zio_free_bp_init(zio_t *zio) 1459 { 1460 blkptr_t *bp = zio->io_bp; 1461 1462 if (zio->io_child_type == ZIO_CHILD_LOGICAL) { 1463 if (BP_GET_DEDUP(bp)) 1464 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE; 1465 } 1466 1467 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy); 1468 1469 return (ZIO_PIPELINE_CONTINUE); 1470 } 1471 1472 /* 1473 * ========================================================================== 1474 * Execute the I/O pipeline 1475 * ========================================================================== 1476 */ 1477 1478 static void 1479 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline) 1480 { 1481 spa_t *spa = zio->io_spa; 1482 zio_type_t t = zio->io_type; 1483 int flags = (cutinline ? TQ_FRONT : 0); 1484 1485 /* 1486 * If we're a config writer or a probe, the normal issue and 1487 * interrupt threads may all be blocked waiting for the config lock. 1488 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL. 1489 */ 1490 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE)) 1491 t = ZIO_TYPE_NULL; 1492 1493 /* 1494 * A similar issue exists for the L2ARC write thread until L2ARC 2.0. 1495 */ 1496 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux) 1497 t = ZIO_TYPE_NULL; 1498 1499 /* 1500 * If this is a high priority I/O, then use the high priority taskq if 1501 * available. 1502 */ 1503 if (zio->io_priority == ZIO_PRIORITY_NOW && 1504 spa->spa_zio_taskq[t][q + 1].stqs_count != 0) 1505 q++; 1506 1507 ASSERT3U(q, <, ZIO_TASKQ_TYPES); 1508 1509 /* 1510 * NB: We are assuming that the zio can only be dispatched 1511 * to a single taskq at a time. It would be a grievous error 1512 * to dispatch the zio to another taskq at the same time. 1513 */ 1514 ASSERT(zio->io_tqent.tqent_next == NULL); 1515 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio, 1516 flags, &zio->io_tqent); 1517 } 1518 1519 static boolean_t 1520 zio_taskq_member(zio_t *zio, zio_taskq_type_t q) 1521 { 1522 kthread_t *executor = zio->io_executor; 1523 spa_t *spa = zio->io_spa; 1524 1525 for (zio_type_t t = 0; t < ZIO_TYPES; t++) { 1526 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q]; 1527 uint_t i; 1528 for (i = 0; i < tqs->stqs_count; i++) { 1529 if (taskq_member(tqs->stqs_taskq[i], executor)) 1530 return (B_TRUE); 1531 } 1532 } 1533 1534 return (B_FALSE); 1535 } 1536 1537 static int 1538 zio_issue_async(zio_t *zio) 1539 { 1540 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); 1541 1542 return (ZIO_PIPELINE_STOP); 1543 } 1544 1545 void 1546 zio_interrupt(zio_t *zio) 1547 { 1548 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE); 1549 } 1550 1551 void 1552 zio_delay_interrupt(zio_t *zio) 1553 { 1554 /* 1555 * The timeout_generic() function isn't defined in userspace, so 1556 * rather than trying to implement the function, the zio delay 1557 * functionality has been disabled for userspace builds. 1558 */ 1559 1560 #ifdef _KERNEL 1561 /* 1562 * If io_target_timestamp is zero, then no delay has been registered 1563 * for this IO, thus jump to the end of this function and "skip" the 1564 * delay; issuing it directly to the zio layer. 1565 */ 1566 if (zio->io_target_timestamp != 0) { 1567 hrtime_t now = gethrtime(); 1568 1569 if (now >= zio->io_target_timestamp) { 1570 /* 1571 * This IO has already taken longer than the target 1572 * delay to complete, so we don't want to delay it 1573 * any longer; we "miss" the delay and issue it 1574 * directly to the zio layer. This is likely due to 1575 * the target latency being set to a value less than 1576 * the underlying hardware can satisfy (e.g. delay 1577 * set to 1ms, but the disks take 10ms to complete an 1578 * IO request). 1579 */ 1580 1581 DTRACE_PROBE2(zio__delay__miss, zio_t *, zio, 1582 hrtime_t, now); 1583 1584 zio_interrupt(zio); 1585 } else { 1586 hrtime_t diff = zio->io_target_timestamp - now; 1587 1588 DTRACE_PROBE3(zio__delay__hit, zio_t *, zio, 1589 hrtime_t, now, hrtime_t, diff); 1590 1591 (void) timeout_generic(CALLOUT_NORMAL, 1592 (void (*)(void *))zio_interrupt, zio, diff, 1, 0); 1593 } 1594 1595 return; 1596 } 1597 #endif 1598 1599 DTRACE_PROBE1(zio__delay__skip, zio_t *, zio); 1600 zio_interrupt(zio); 1601 } 1602 1603 /* 1604 * Execute the I/O pipeline until one of the following occurs: 1605 * 1606 * (1) the I/O completes 1607 * (2) the pipeline stalls waiting for dependent child I/Os 1608 * (3) the I/O issues, so we're waiting for an I/O completion interrupt 1609 * (4) the I/O is delegated by vdev-level caching or aggregation 1610 * (5) the I/O is deferred due to vdev-level queueing 1611 * (6) the I/O is handed off to another thread. 1612 * 1613 * In all cases, the pipeline stops whenever there's no CPU work; it never 1614 * burns a thread in cv_wait(). 1615 * 1616 * There's no locking on io_stage because there's no legitimate way 1617 * for multiple threads to be attempting to process the same I/O. 1618 */ 1619 static zio_pipe_stage_t *zio_pipeline[]; 1620 1621 void 1622 zio_execute(zio_t *zio) 1623 { 1624 zio->io_executor = curthread; 1625 1626 ASSERT3U(zio->io_queued_timestamp, >, 0); 1627 1628 while (zio->io_stage < ZIO_STAGE_DONE) { 1629 enum zio_stage pipeline = zio->io_pipeline; 1630 enum zio_stage stage = zio->io_stage; 1631 int rv; 1632 1633 ASSERT(!MUTEX_HELD(&zio->io_lock)); 1634 ASSERT(ISP2(stage)); 1635 ASSERT(zio->io_stall == NULL); 1636 1637 do { 1638 stage <<= 1; 1639 } while ((stage & pipeline) == 0); 1640 1641 ASSERT(stage <= ZIO_STAGE_DONE); 1642 1643 /* 1644 * If we are in interrupt context and this pipeline stage 1645 * will grab a config lock that is held across I/O, 1646 * or may wait for an I/O that needs an interrupt thread 1647 * to complete, issue async to avoid deadlock. 1648 * 1649 * For VDEV_IO_START, we cut in line so that the io will 1650 * be sent to disk promptly. 1651 */ 1652 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL && 1653 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) { 1654 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? 1655 zio_requeue_io_start_cut_in_line : B_FALSE; 1656 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); 1657 return; 1658 } 1659 1660 zio->io_stage = stage; 1661 zio->io_pipeline_trace |= zio->io_stage; 1662 rv = zio_pipeline[highbit64(stage) - 1](zio); 1663 1664 if (rv == ZIO_PIPELINE_STOP) 1665 return; 1666 1667 ASSERT(rv == ZIO_PIPELINE_CONTINUE); 1668 } 1669 } 1670 1671 /* 1672 * ========================================================================== 1673 * Initiate I/O, either sync or async 1674 * ========================================================================== 1675 */ 1676 int 1677 zio_wait(zio_t *zio) 1678 { 1679 int error; 1680 1681 ASSERT3P(zio->io_stage, ==, ZIO_STAGE_OPEN); 1682 ASSERT3P(zio->io_executor, ==, NULL); 1683 1684 zio->io_waiter = curthread; 1685 ASSERT0(zio->io_queued_timestamp); 1686 zio->io_queued_timestamp = gethrtime(); 1687 1688 zio_execute(zio); 1689 1690 mutex_enter(&zio->io_lock); 1691 while (zio->io_executor != NULL) 1692 cv_wait(&zio->io_cv, &zio->io_lock); 1693 mutex_exit(&zio->io_lock); 1694 1695 error = zio->io_error; 1696 zio_destroy(zio); 1697 1698 return (error); 1699 } 1700 1701 void 1702 zio_nowait(zio_t *zio) 1703 { 1704 ASSERT3P(zio->io_executor, ==, NULL); 1705 1706 if (zio->io_child_type == ZIO_CHILD_LOGICAL && 1707 zio_unique_parent(zio) == NULL) { 1708 /* 1709 * This is a logical async I/O with no parent to wait for it. 1710 * We add it to the spa_async_root_zio "Godfather" I/O which 1711 * will ensure they complete prior to unloading the pool. 1712 */ 1713 spa_t *spa = zio->io_spa; 1714 1715 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio); 1716 } 1717 1718 ASSERT0(zio->io_queued_timestamp); 1719 zio->io_queued_timestamp = gethrtime(); 1720 zio_execute(zio); 1721 } 1722 1723 /* 1724 * ========================================================================== 1725 * Reexecute, cancel, or suspend/resume failed I/O 1726 * ========================================================================== 1727 */ 1728 1729 static void 1730 zio_reexecute(zio_t *pio) 1731 { 1732 zio_t *cio, *cio_next; 1733 1734 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL); 1735 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN); 1736 ASSERT(pio->io_gang_leader == NULL); 1737 ASSERT(pio->io_gang_tree == NULL); 1738 1739 pio->io_flags = pio->io_orig_flags; 1740 pio->io_stage = pio->io_orig_stage; 1741 pio->io_pipeline = pio->io_orig_pipeline; 1742 pio->io_reexecute = 0; 1743 pio->io_flags |= ZIO_FLAG_REEXECUTED; 1744 pio->io_pipeline_trace = 0; 1745 pio->io_error = 0; 1746 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 1747 pio->io_state[w] = 0; 1748 for (int c = 0; c < ZIO_CHILD_TYPES; c++) 1749 pio->io_child_error[c] = 0; 1750 1751 if (IO_IS_ALLOCATING(pio)) 1752 BP_ZERO(pio->io_bp); 1753 1754 /* 1755 * As we reexecute pio's children, new children could be created. 1756 * New children go to the head of pio's io_child_list, however, 1757 * so we will (correctly) not reexecute them. The key is that 1758 * the remainder of pio's io_child_list, from 'cio_next' onward, 1759 * cannot be affected by any side effects of reexecuting 'cio'. 1760 */ 1761 zio_link_t *zl = NULL; 1762 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) { 1763 cio_next = zio_walk_children(pio, &zl); 1764 mutex_enter(&pio->io_lock); 1765 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 1766 pio->io_children[cio->io_child_type][w]++; 1767 mutex_exit(&pio->io_lock); 1768 zio_reexecute(cio); 1769 } 1770 1771 /* 1772 * Now that all children have been reexecuted, execute the parent. 1773 * We don't reexecute "The Godfather" I/O here as it's the 1774 * responsibility of the caller to wait on it. 1775 */ 1776 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) { 1777 pio->io_queued_timestamp = gethrtime(); 1778 zio_execute(pio); 1779 } 1780 } 1781 1782 void 1783 zio_suspend(spa_t *spa, zio_t *zio) 1784 { 1785 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC) 1786 fm_panic("Pool '%s' has encountered an uncorrectable I/O " 1787 "failure and the failure mode property for this pool " 1788 "is set to panic.", spa_name(spa)); 1789 1790 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0); 1791 1792 mutex_enter(&spa->spa_suspend_lock); 1793 1794 if (spa->spa_suspend_zio_root == NULL) 1795 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL, 1796 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | 1797 ZIO_FLAG_GODFATHER); 1798 1799 spa->spa_suspended = B_TRUE; 1800 1801 if (zio != NULL) { 1802 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); 1803 ASSERT(zio != spa->spa_suspend_zio_root); 1804 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1805 ASSERT(zio_unique_parent(zio) == NULL); 1806 ASSERT(zio->io_stage == ZIO_STAGE_DONE); 1807 zio_add_child(spa->spa_suspend_zio_root, zio); 1808 } 1809 1810 mutex_exit(&spa->spa_suspend_lock); 1811 } 1812 1813 int 1814 zio_resume(spa_t *spa) 1815 { 1816 zio_t *pio; 1817 1818 /* 1819 * Reexecute all previously suspended i/o. 1820 */ 1821 mutex_enter(&spa->spa_suspend_lock); 1822 spa->spa_suspended = B_FALSE; 1823 cv_broadcast(&spa->spa_suspend_cv); 1824 pio = spa->spa_suspend_zio_root; 1825 spa->spa_suspend_zio_root = NULL; 1826 mutex_exit(&spa->spa_suspend_lock); 1827 1828 if (pio == NULL) 1829 return (0); 1830 1831 zio_reexecute(pio); 1832 return (zio_wait(pio)); 1833 } 1834 1835 void 1836 zio_resume_wait(spa_t *spa) 1837 { 1838 mutex_enter(&spa->spa_suspend_lock); 1839 while (spa_suspended(spa)) 1840 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock); 1841 mutex_exit(&spa->spa_suspend_lock); 1842 } 1843 1844 /* 1845 * ========================================================================== 1846 * Gang blocks. 1847 * 1848 * A gang block is a collection of small blocks that looks to the DMU 1849 * like one large block. When zio_dva_allocate() cannot find a block 1850 * of the requested size, due to either severe fragmentation or the pool 1851 * being nearly full, it calls zio_write_gang_block() to construct the 1852 * block from smaller fragments. 1853 * 1854 * A gang block consists of a gang header (zio_gbh_phys_t) and up to 1855 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like 1856 * an indirect block: it's an array of block pointers. It consumes 1857 * only one sector and hence is allocatable regardless of fragmentation. 1858 * The gang header's bps point to its gang members, which hold the data. 1859 * 1860 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg> 1861 * as the verifier to ensure uniqueness of the SHA256 checksum. 1862 * Critically, the gang block bp's blk_cksum is the checksum of the data, 1863 * not the gang header. This ensures that data block signatures (needed for 1864 * deduplication) are independent of how the block is physically stored. 1865 * 1866 * Gang blocks can be nested: a gang member may itself be a gang block. 1867 * Thus every gang block is a tree in which root and all interior nodes are 1868 * gang headers, and the leaves are normal blocks that contain user data. 1869 * The root of the gang tree is called the gang leader. 1870 * 1871 * To perform any operation (read, rewrite, free, claim) on a gang block, 1872 * zio_gang_assemble() first assembles the gang tree (minus data leaves) 1873 * in the io_gang_tree field of the original logical i/o by recursively 1874 * reading the gang leader and all gang headers below it. This yields 1875 * an in-core tree containing the contents of every gang header and the 1876 * bps for every constituent of the gang block. 1877 * 1878 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree 1879 * and invokes a callback on each bp. To free a gang block, zio_gang_issue() 1880 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp. 1881 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim(). 1882 * zio_read_gang() is a wrapper around zio_read() that omits reading gang 1883 * headers, since we already have those in io_gang_tree. zio_rewrite_gang() 1884 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite() 1885 * of the gang header plus zio_checksum_compute() of the data to update the 1886 * gang header's blk_cksum as described above. 1887 * 1888 * The two-phase assemble/issue model solves the problem of partial failure -- 1889 * what if you'd freed part of a gang block but then couldn't read the 1890 * gang header for another part? Assembling the entire gang tree first 1891 * ensures that all the necessary gang header I/O has succeeded before 1892 * starting the actual work of free, claim, or write. Once the gang tree 1893 * is assembled, free and claim are in-memory operations that cannot fail. 1894 * 1895 * In the event that a gang write fails, zio_dva_unallocate() walks the 1896 * gang tree to immediately free (i.e. insert back into the space map) 1897 * everything we've allocated. This ensures that we don't get ENOSPC 1898 * errors during repeated suspend/resume cycles due to a flaky device. 1899 * 1900 * Gang rewrites only happen during sync-to-convergence. If we can't assemble 1901 * the gang tree, we won't modify the block, so we can safely defer the free 1902 * (knowing that the block is still intact). If we *can* assemble the gang 1903 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free 1904 * each constituent bp and we can allocate a new block on the next sync pass. 1905 * 1906 * In all cases, the gang tree allows complete recovery from partial failure. 1907 * ========================================================================== 1908 */ 1909 1910 static void 1911 zio_gang_issue_func_done(zio_t *zio) 1912 { 1913 abd_put(zio->io_abd); 1914 } 1915 1916 static zio_t * 1917 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, 1918 uint64_t offset) 1919 { 1920 if (gn != NULL) 1921 return (pio); 1922 1923 return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset), 1924 BP_GET_PSIZE(bp), zio_gang_issue_func_done, 1925 NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), 1926 &pio->io_bookmark)); 1927 } 1928 1929 static zio_t * 1930 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, 1931 uint64_t offset) 1932 { 1933 zio_t *zio; 1934 1935 if (gn != NULL) { 1936 abd_t *gbh_abd = 1937 abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE); 1938 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, 1939 gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL, 1940 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), 1941 &pio->io_bookmark); 1942 /* 1943 * As we rewrite each gang header, the pipeline will compute 1944 * a new gang block header checksum for it; but no one will 1945 * compute a new data checksum, so we do that here. The one 1946 * exception is the gang leader: the pipeline already computed 1947 * its data checksum because that stage precedes gang assembly. 1948 * (Presently, nothing actually uses interior data checksums; 1949 * this is just good hygiene.) 1950 */ 1951 if (gn != pio->io_gang_leader->io_gang_tree) { 1952 abd_t *buf = abd_get_offset(data, offset); 1953 1954 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp), 1955 buf, BP_GET_PSIZE(bp)); 1956 1957 abd_put(buf); 1958 } 1959 /* 1960 * If we are here to damage data for testing purposes, 1961 * leave the GBH alone so that we can detect the damage. 1962 */ 1963 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE) 1964 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; 1965 } else { 1966 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, 1967 abd_get_offset(data, offset), BP_GET_PSIZE(bp), 1968 zio_gang_issue_func_done, NULL, pio->io_priority, 1969 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 1970 } 1971 1972 return (zio); 1973 } 1974 1975 /* ARGSUSED */ 1976 static zio_t * 1977 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, 1978 uint64_t offset) 1979 { 1980 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp, 1981 ZIO_GANG_CHILD_FLAGS(pio))); 1982 } 1983 1984 /* ARGSUSED */ 1985 static zio_t * 1986 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, 1987 uint64_t offset) 1988 { 1989 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp, 1990 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio))); 1991 } 1992 1993 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = { 1994 NULL, 1995 zio_read_gang, 1996 zio_rewrite_gang, 1997 zio_free_gang, 1998 zio_claim_gang, 1999 NULL 2000 }; 2001 2002 static void zio_gang_tree_assemble_done(zio_t *zio); 2003 2004 static zio_gang_node_t * 2005 zio_gang_node_alloc(zio_gang_node_t **gnpp) 2006 { 2007 zio_gang_node_t *gn; 2008 2009 ASSERT(*gnpp == NULL); 2010 2011 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP); 2012 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE); 2013 *gnpp = gn; 2014 2015 return (gn); 2016 } 2017 2018 static void 2019 zio_gang_node_free(zio_gang_node_t **gnpp) 2020 { 2021 zio_gang_node_t *gn = *gnpp; 2022 2023 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) 2024 ASSERT(gn->gn_child[g] == NULL); 2025 2026 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE); 2027 kmem_free(gn, sizeof (*gn)); 2028 *gnpp = NULL; 2029 } 2030 2031 static void 2032 zio_gang_tree_free(zio_gang_node_t **gnpp) 2033 { 2034 zio_gang_node_t *gn = *gnpp; 2035 2036 if (gn == NULL) 2037 return; 2038 2039 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) 2040 zio_gang_tree_free(&gn->gn_child[g]); 2041 2042 zio_gang_node_free(gnpp); 2043 } 2044 2045 static void 2046 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp) 2047 { 2048 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp); 2049 abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE); 2050 2051 ASSERT(gio->io_gang_leader == gio); 2052 ASSERT(BP_IS_GANG(bp)); 2053 2054 zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE, 2055 zio_gang_tree_assemble_done, gn, gio->io_priority, 2056 ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark)); 2057 } 2058 2059 static void 2060 zio_gang_tree_assemble_done(zio_t *zio) 2061 { 2062 zio_t *gio = zio->io_gang_leader; 2063 zio_gang_node_t *gn = zio->io_private; 2064 blkptr_t *bp = zio->io_bp; 2065 2066 ASSERT(gio == zio_unique_parent(zio)); 2067 ASSERT(zio->io_child_count == 0); 2068 2069 if (zio->io_error) 2070 return; 2071 2072 /* this ABD was created from a linear buf in zio_gang_tree_assemble */ 2073 if (BP_SHOULD_BYTESWAP(bp)) 2074 byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size); 2075 2076 ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh); 2077 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE); 2078 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); 2079 2080 abd_put(zio->io_abd); 2081 2082 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 2083 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; 2084 if (!BP_IS_GANG(gbp)) 2085 continue; 2086 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]); 2087 } 2088 } 2089 2090 static void 2091 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data, 2092 uint64_t offset) 2093 { 2094 zio_t *gio = pio->io_gang_leader; 2095 zio_t *zio; 2096 2097 ASSERT(BP_IS_GANG(bp) == !!gn); 2098 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp)); 2099 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree); 2100 2101 /* 2102 * If you're a gang header, your data is in gn->gn_gbh. 2103 * If you're a gang member, your data is in 'data' and gn == NULL. 2104 */ 2105 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset); 2106 2107 if (gn != NULL) { 2108 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); 2109 2110 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 2111 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; 2112 if (BP_IS_HOLE(gbp)) 2113 continue; 2114 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data, 2115 offset); 2116 offset += BP_GET_PSIZE(gbp); 2117 } 2118 } 2119 2120 if (gn == gio->io_gang_tree) 2121 ASSERT3U(gio->io_size, ==, offset); 2122 2123 if (zio != pio) 2124 zio_nowait(zio); 2125 } 2126 2127 static int 2128 zio_gang_assemble(zio_t *zio) 2129 { 2130 blkptr_t *bp = zio->io_bp; 2131 2132 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL); 2133 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 2134 2135 zio->io_gang_leader = zio; 2136 2137 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree); 2138 2139 return (ZIO_PIPELINE_CONTINUE); 2140 } 2141 2142 static int 2143 zio_gang_issue(zio_t *zio) 2144 { 2145 blkptr_t *bp = zio->io_bp; 2146 2147 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE)) 2148 return (ZIO_PIPELINE_STOP); 2149 2150 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio); 2151 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 2152 2153 if (zio->io_child_error[ZIO_CHILD_GANG] == 0) 2154 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd, 2155 0); 2156 else 2157 zio_gang_tree_free(&zio->io_gang_tree); 2158 2159 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2160 2161 return (ZIO_PIPELINE_CONTINUE); 2162 } 2163 2164 static void 2165 zio_write_gang_member_ready(zio_t *zio) 2166 { 2167 zio_t *pio = zio_unique_parent(zio); 2168 zio_t *gio = zio->io_gang_leader; 2169 dva_t *cdva = zio->io_bp->blk_dva; 2170 dva_t *pdva = pio->io_bp->blk_dva; 2171 uint64_t asize; 2172 2173 if (BP_IS_HOLE(zio->io_bp)) 2174 return; 2175 2176 ASSERT(BP_IS_HOLE(&zio->io_bp_orig)); 2177 2178 ASSERT(zio->io_child_type == ZIO_CHILD_GANG); 2179 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies); 2180 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp)); 2181 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp)); 2182 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp)); 2183 2184 mutex_enter(&pio->io_lock); 2185 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) { 2186 ASSERT(DVA_GET_GANG(&pdva[d])); 2187 asize = DVA_GET_ASIZE(&pdva[d]); 2188 asize += DVA_GET_ASIZE(&cdva[d]); 2189 DVA_SET_ASIZE(&pdva[d], asize); 2190 } 2191 mutex_exit(&pio->io_lock); 2192 } 2193 2194 static void 2195 zio_write_gang_done(zio_t *zio) 2196 { 2197 abd_put(zio->io_abd); 2198 } 2199 2200 static int 2201 zio_write_gang_block(zio_t *pio) 2202 { 2203 spa_t *spa = pio->io_spa; 2204 metaslab_class_t *mc = spa_normal_class(spa); 2205 blkptr_t *bp = pio->io_bp; 2206 zio_t *gio = pio->io_gang_leader; 2207 zio_t *zio; 2208 zio_gang_node_t *gn, **gnpp; 2209 zio_gbh_phys_t *gbh; 2210 abd_t *gbh_abd; 2211 uint64_t txg = pio->io_txg; 2212 uint64_t resid = pio->io_size; 2213 uint64_t lsize; 2214 int copies = gio->io_prop.zp_copies; 2215 int gbh_copies = MIN(copies + 1, spa_max_replication(spa)); 2216 zio_prop_t zp; 2217 int error; 2218 2219 int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER; 2220 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { 2221 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); 2222 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA)); 2223 2224 flags |= METASLAB_ASYNC_ALLOC; 2225 VERIFY(refcount_held(&mc->mc_alloc_slots, pio)); 2226 2227 /* 2228 * The logical zio has already placed a reservation for 2229 * 'copies' allocation slots but gang blocks may require 2230 * additional copies. These additional copies 2231 * (i.e. gbh_copies - copies) are guaranteed to succeed 2232 * since metaslab_class_throttle_reserve() always allows 2233 * additional reservations for gang blocks. 2234 */ 2235 VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies, 2236 pio, flags)); 2237 } 2238 2239 error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE, 2240 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags, 2241 &pio->io_alloc_list, pio); 2242 if (error) { 2243 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { 2244 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); 2245 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA)); 2246 2247 /* 2248 * If we failed to allocate the gang block header then 2249 * we remove any additional allocation reservations that 2250 * we placed here. The original reservation will 2251 * be removed when the logical I/O goes to the ready 2252 * stage. 2253 */ 2254 metaslab_class_throttle_unreserve(mc, 2255 gbh_copies - copies, pio); 2256 } 2257 pio->io_error = error; 2258 return (ZIO_PIPELINE_CONTINUE); 2259 } 2260 2261 if (pio == gio) { 2262 gnpp = &gio->io_gang_tree; 2263 } else { 2264 gnpp = pio->io_private; 2265 ASSERT(pio->io_ready == zio_write_gang_member_ready); 2266 } 2267 2268 gn = zio_gang_node_alloc(gnpp); 2269 gbh = gn->gn_gbh; 2270 bzero(gbh, SPA_GANGBLOCKSIZE); 2271 gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE); 2272 2273 /* 2274 * Create the gang header. 2275 */ 2276 zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE, 2277 zio_write_gang_done, NULL, pio->io_priority, 2278 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 2279 2280 /* 2281 * Create and nowait the gang children. 2282 */ 2283 for (int g = 0; resid != 0; resid -= lsize, g++) { 2284 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g), 2285 SPA_MINBLOCKSIZE); 2286 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid); 2287 2288 zp.zp_checksum = gio->io_prop.zp_checksum; 2289 zp.zp_compress = ZIO_COMPRESS_OFF; 2290 zp.zp_type = DMU_OT_NONE; 2291 zp.zp_level = 0; 2292 zp.zp_copies = gio->io_prop.zp_copies; 2293 zp.zp_dedup = B_FALSE; 2294 zp.zp_dedup_verify = B_FALSE; 2295 zp.zp_nopwrite = B_FALSE; 2296 2297 zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g], 2298 abd_get_offset(pio->io_abd, pio->io_size - resid), lsize, 2299 lsize, &zp, zio_write_gang_member_ready, NULL, NULL, 2300 zio_write_gang_done, &gn->gn_child[g], pio->io_priority, 2301 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 2302 2303 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { 2304 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); 2305 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA)); 2306 2307 /* 2308 * Gang children won't throttle but we should 2309 * account for their work, so reserve an allocation 2310 * slot for them here. 2311 */ 2312 VERIFY(metaslab_class_throttle_reserve(mc, 2313 zp.zp_copies, cio, flags)); 2314 } 2315 zio_nowait(cio); 2316 } 2317 2318 /* 2319 * Set pio's pipeline to just wait for zio to finish. 2320 */ 2321 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2322 2323 zio_nowait(zio); 2324 2325 return (ZIO_PIPELINE_CONTINUE); 2326 } 2327 2328 /* 2329 * The zio_nop_write stage in the pipeline determines if allocating a 2330 * new bp is necessary. The nopwrite feature can handle writes in 2331 * either syncing or open context (i.e. zil writes) and as a result is 2332 * mutually exclusive with dedup. 2333 * 2334 * By leveraging a cryptographically secure checksum, such as SHA256, we 2335 * can compare the checksums of the new data and the old to determine if 2336 * allocating a new block is required. Note that our requirements for 2337 * cryptographic strength are fairly weak: there can't be any accidental 2338 * hash collisions, but we don't need to be secure against intentional 2339 * (malicious) collisions. To trigger a nopwrite, you have to be able 2340 * to write the file to begin with, and triggering an incorrect (hash 2341 * collision) nopwrite is no worse than simply writing to the file. 2342 * That said, there are no known attacks against the checksum algorithms 2343 * used for nopwrite, assuming that the salt and the checksums 2344 * themselves remain secret. 2345 */ 2346 static int 2347 zio_nop_write(zio_t *zio) 2348 { 2349 blkptr_t *bp = zio->io_bp; 2350 blkptr_t *bp_orig = &zio->io_bp_orig; 2351 zio_prop_t *zp = &zio->io_prop; 2352 2353 ASSERT(BP_GET_LEVEL(bp) == 0); 2354 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 2355 ASSERT(zp->zp_nopwrite); 2356 ASSERT(!zp->zp_dedup); 2357 ASSERT(zio->io_bp_override == NULL); 2358 ASSERT(IO_IS_ALLOCATING(zio)); 2359 2360 /* 2361 * Check to see if the original bp and the new bp have matching 2362 * characteristics (i.e. same checksum, compression algorithms, etc). 2363 * If they don't then just continue with the pipeline which will 2364 * allocate a new bp. 2365 */ 2366 if (BP_IS_HOLE(bp_orig) || 2367 !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags & 2368 ZCHECKSUM_FLAG_NOPWRITE) || 2369 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) || 2370 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) || 2371 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) || 2372 zp->zp_copies != BP_GET_NDVAS(bp_orig)) 2373 return (ZIO_PIPELINE_CONTINUE); 2374 2375 /* 2376 * If the checksums match then reset the pipeline so that we 2377 * avoid allocating a new bp and issuing any I/O. 2378 */ 2379 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) { 2380 ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags & 2381 ZCHECKSUM_FLAG_NOPWRITE); 2382 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig)); 2383 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig)); 2384 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF); 2385 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop, 2386 sizeof (uint64_t)) == 0); 2387 2388 *bp = *bp_orig; 2389 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2390 zio->io_flags |= ZIO_FLAG_NOPWRITE; 2391 } 2392 2393 return (ZIO_PIPELINE_CONTINUE); 2394 } 2395 2396 /* 2397 * ========================================================================== 2398 * Dedup 2399 * ========================================================================== 2400 */ 2401 static void 2402 zio_ddt_child_read_done(zio_t *zio) 2403 { 2404 blkptr_t *bp = zio->io_bp; 2405 ddt_entry_t *dde = zio->io_private; 2406 ddt_phys_t *ddp; 2407 zio_t *pio = zio_unique_parent(zio); 2408 2409 mutex_enter(&pio->io_lock); 2410 ddp = ddt_phys_select(dde, bp); 2411 if (zio->io_error == 0) 2412 ddt_phys_clear(ddp); /* this ddp doesn't need repair */ 2413 2414 if (zio->io_error == 0 && dde->dde_repair_abd == NULL) 2415 dde->dde_repair_abd = zio->io_abd; 2416 else 2417 abd_free(zio->io_abd); 2418 mutex_exit(&pio->io_lock); 2419 } 2420 2421 static int 2422 zio_ddt_read_start(zio_t *zio) 2423 { 2424 blkptr_t *bp = zio->io_bp; 2425 2426 ASSERT(BP_GET_DEDUP(bp)); 2427 ASSERT(BP_GET_PSIZE(bp) == zio->io_size); 2428 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2429 2430 if (zio->io_child_error[ZIO_CHILD_DDT]) { 2431 ddt_t *ddt = ddt_select(zio->io_spa, bp); 2432 ddt_entry_t *dde = ddt_repair_start(ddt, bp); 2433 ddt_phys_t *ddp = dde->dde_phys; 2434 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp); 2435 blkptr_t blk; 2436 2437 ASSERT(zio->io_vsd == NULL); 2438 zio->io_vsd = dde; 2439 2440 if (ddp_self == NULL) 2441 return (ZIO_PIPELINE_CONTINUE); 2442 2443 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) { 2444 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self) 2445 continue; 2446 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp, 2447 &blk); 2448 zio_nowait(zio_read(zio, zio->io_spa, &blk, 2449 abd_alloc_for_io(zio->io_size, B_TRUE), 2450 zio->io_size, zio_ddt_child_read_done, dde, 2451 zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) | 2452 ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark)); 2453 } 2454 return (ZIO_PIPELINE_CONTINUE); 2455 } 2456 2457 zio_nowait(zio_read(zio, zio->io_spa, bp, 2458 zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority, 2459 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark)); 2460 2461 return (ZIO_PIPELINE_CONTINUE); 2462 } 2463 2464 static int 2465 zio_ddt_read_done(zio_t *zio) 2466 { 2467 blkptr_t *bp = zio->io_bp; 2468 2469 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE)) 2470 return (ZIO_PIPELINE_STOP); 2471 2472 ASSERT(BP_GET_DEDUP(bp)); 2473 ASSERT(BP_GET_PSIZE(bp) == zio->io_size); 2474 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2475 2476 if (zio->io_child_error[ZIO_CHILD_DDT]) { 2477 ddt_t *ddt = ddt_select(zio->io_spa, bp); 2478 ddt_entry_t *dde = zio->io_vsd; 2479 if (ddt == NULL) { 2480 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE); 2481 return (ZIO_PIPELINE_CONTINUE); 2482 } 2483 if (dde == NULL) { 2484 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1; 2485 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); 2486 return (ZIO_PIPELINE_STOP); 2487 } 2488 if (dde->dde_repair_abd != NULL) { 2489 abd_copy(zio->io_abd, dde->dde_repair_abd, 2490 zio->io_size); 2491 zio->io_child_error[ZIO_CHILD_DDT] = 0; 2492 } 2493 ddt_repair_done(ddt, dde); 2494 zio->io_vsd = NULL; 2495 } 2496 2497 ASSERT(zio->io_vsd == NULL); 2498 2499 return (ZIO_PIPELINE_CONTINUE); 2500 } 2501 2502 static boolean_t 2503 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde) 2504 { 2505 spa_t *spa = zio->io_spa; 2506 boolean_t do_raw = (zio->io_flags & ZIO_FLAG_RAW); 2507 2508 /* We should never get a raw, override zio */ 2509 ASSERT(!(zio->io_bp_override && do_raw)); 2510 2511 /* 2512 * Note: we compare the original data, not the transformed data, 2513 * because when zio->io_bp is an override bp, we will not have 2514 * pushed the I/O transforms. That's an important optimization 2515 * because otherwise we'd compress/encrypt all dmu_sync() data twice. 2516 */ 2517 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { 2518 zio_t *lio = dde->dde_lead_zio[p]; 2519 2520 if (lio != NULL) { 2521 return (lio->io_orig_size != zio->io_orig_size || 2522 abd_cmp(zio->io_orig_abd, lio->io_orig_abd, 2523 zio->io_orig_size) != 0); 2524 } 2525 } 2526 2527 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { 2528 ddt_phys_t *ddp = &dde->dde_phys[p]; 2529 2530 if (ddp->ddp_phys_birth != 0) { 2531 arc_buf_t *abuf = NULL; 2532 arc_flags_t aflags = ARC_FLAG_WAIT; 2533 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE; 2534 blkptr_t blk = *zio->io_bp; 2535 int error; 2536 2537 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth); 2538 2539 ddt_exit(ddt); 2540 2541 /* 2542 * Intuitively, it would make more sense to compare 2543 * io_abd than io_orig_abd in the raw case since you 2544 * don't want to look at any transformations that have 2545 * happened to the data. However, for raw I/Os the 2546 * data will actually be the same in io_abd and 2547 * io_orig_abd, so all we have to do is issue this as 2548 * a raw ARC read. 2549 */ 2550 if (do_raw) { 2551 zio_flags |= ZIO_FLAG_RAW; 2552 ASSERT3U(zio->io_size, ==, zio->io_orig_size); 2553 ASSERT0(abd_cmp(zio->io_abd, zio->io_orig_abd, 2554 zio->io_size)); 2555 ASSERT3P(zio->io_transform_stack, ==, NULL); 2556 } 2557 2558 error = arc_read(NULL, spa, &blk, 2559 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ, 2560 zio_flags, &aflags, &zio->io_bookmark); 2561 2562 if (error == 0) { 2563 if (arc_buf_size(abuf) != zio->io_orig_size || 2564 abd_cmp_buf(zio->io_orig_abd, abuf->b_data, 2565 zio->io_orig_size) != 0) 2566 error = SET_ERROR(EEXIST); 2567 arc_buf_destroy(abuf, &abuf); 2568 } 2569 2570 ddt_enter(ddt); 2571 return (error != 0); 2572 } 2573 } 2574 2575 return (B_FALSE); 2576 } 2577 2578 static void 2579 zio_ddt_child_write_ready(zio_t *zio) 2580 { 2581 int p = zio->io_prop.zp_copies; 2582 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); 2583 ddt_entry_t *dde = zio->io_private; 2584 ddt_phys_t *ddp = &dde->dde_phys[p]; 2585 zio_t *pio; 2586 2587 if (zio->io_error) 2588 return; 2589 2590 ddt_enter(ddt); 2591 2592 ASSERT(dde->dde_lead_zio[p] == zio); 2593 2594 ddt_phys_fill(ddp, zio->io_bp); 2595 2596 zio_link_t *zl = NULL; 2597 while ((pio = zio_walk_parents(zio, &zl)) != NULL) 2598 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg); 2599 2600 ddt_exit(ddt); 2601 } 2602 2603 static void 2604 zio_ddt_child_write_done(zio_t *zio) 2605 { 2606 int p = zio->io_prop.zp_copies; 2607 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); 2608 ddt_entry_t *dde = zio->io_private; 2609 ddt_phys_t *ddp = &dde->dde_phys[p]; 2610 2611 ddt_enter(ddt); 2612 2613 ASSERT(ddp->ddp_refcnt == 0); 2614 ASSERT(dde->dde_lead_zio[p] == zio); 2615 dde->dde_lead_zio[p] = NULL; 2616 2617 if (zio->io_error == 0) { 2618 zio_link_t *zl = NULL; 2619 while (zio_walk_parents(zio, &zl) != NULL) 2620 ddt_phys_addref(ddp); 2621 } else { 2622 ddt_phys_clear(ddp); 2623 } 2624 2625 ddt_exit(ddt); 2626 } 2627 2628 static void 2629 zio_ddt_ditto_write_done(zio_t *zio) 2630 { 2631 int p = DDT_PHYS_DITTO; 2632 zio_prop_t *zp = &zio->io_prop; 2633 blkptr_t *bp = zio->io_bp; 2634 ddt_t *ddt = ddt_select(zio->io_spa, bp); 2635 ddt_entry_t *dde = zio->io_private; 2636 ddt_phys_t *ddp = &dde->dde_phys[p]; 2637 ddt_key_t *ddk = &dde->dde_key; 2638 2639 ddt_enter(ddt); 2640 2641 ASSERT(ddp->ddp_refcnt == 0); 2642 ASSERT(dde->dde_lead_zio[p] == zio); 2643 dde->dde_lead_zio[p] = NULL; 2644 2645 if (zio->io_error == 0) { 2646 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum)); 2647 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP); 2648 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp)); 2649 if (ddp->ddp_phys_birth != 0) 2650 ddt_phys_free(ddt, ddk, ddp, zio->io_txg); 2651 ddt_phys_fill(ddp, bp); 2652 } 2653 2654 ddt_exit(ddt); 2655 } 2656 2657 static int 2658 zio_ddt_write(zio_t *zio) 2659 { 2660 spa_t *spa = zio->io_spa; 2661 blkptr_t *bp = zio->io_bp; 2662 uint64_t txg = zio->io_txg; 2663 zio_prop_t *zp = &zio->io_prop; 2664 int p = zp->zp_copies; 2665 int ditto_copies; 2666 zio_t *cio = NULL; 2667 zio_t *dio = NULL; 2668 ddt_t *ddt = ddt_select(spa, bp); 2669 ddt_entry_t *dde; 2670 ddt_phys_t *ddp; 2671 2672 ASSERT(BP_GET_DEDUP(bp)); 2673 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum); 2674 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override); 2675 ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW))); 2676 2677 ddt_enter(ddt); 2678 dde = ddt_lookup(ddt, bp, B_TRUE); 2679 ddp = &dde->dde_phys[p]; 2680 2681 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) { 2682 /* 2683 * If we're using a weak checksum, upgrade to a strong checksum 2684 * and try again. If we're already using a strong checksum, 2685 * we can't resolve it, so just convert to an ordinary write. 2686 * (And automatically e-mail a paper to Nature?) 2687 */ 2688 if (!(zio_checksum_table[zp->zp_checksum].ci_flags & 2689 ZCHECKSUM_FLAG_DEDUP)) { 2690 zp->zp_checksum = spa_dedup_checksum(spa); 2691 zio_pop_transforms(zio); 2692 zio->io_stage = ZIO_STAGE_OPEN; 2693 BP_ZERO(bp); 2694 } else { 2695 zp->zp_dedup = B_FALSE; 2696 BP_SET_DEDUP(bp, B_FALSE); 2697 } 2698 ASSERT(!BP_GET_DEDUP(bp)); 2699 zio->io_pipeline = ZIO_WRITE_PIPELINE; 2700 ddt_exit(ddt); 2701 return (ZIO_PIPELINE_CONTINUE); 2702 } 2703 2704 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp); 2705 ASSERT(ditto_copies < SPA_DVAS_PER_BP); 2706 2707 if (ditto_copies > ddt_ditto_copies_present(dde) && 2708 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) { 2709 zio_prop_t czp = *zp; 2710 2711 czp.zp_copies = ditto_copies; 2712 2713 /* 2714 * If we arrived here with an override bp, we won't have run 2715 * the transform stack, so we won't have the data we need to 2716 * generate a child i/o. So, toss the override bp and restart. 2717 * This is safe, because using the override bp is just an 2718 * optimization; and it's rare, so the cost doesn't matter. 2719 */ 2720 if (zio->io_bp_override) { 2721 zio_pop_transforms(zio); 2722 zio->io_stage = ZIO_STAGE_OPEN; 2723 zio->io_pipeline = ZIO_WRITE_PIPELINE; 2724 zio->io_bp_override = NULL; 2725 BP_ZERO(bp); 2726 ddt_exit(ddt); 2727 return (ZIO_PIPELINE_CONTINUE); 2728 } 2729 2730 dio = zio_write(zio, spa, txg, bp, zio->io_orig_abd, 2731 zio->io_orig_size, zio->io_orig_size, &czp, NULL, NULL, 2732 NULL, zio_ddt_ditto_write_done, dde, zio->io_priority, 2733 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); 2734 2735 zio_push_transform(dio, zio->io_abd, zio->io_size, 0, NULL); 2736 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio; 2737 } 2738 2739 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) { 2740 if (ddp->ddp_phys_birth != 0) 2741 ddt_bp_fill(ddp, bp, txg); 2742 if (dde->dde_lead_zio[p] != NULL) 2743 zio_add_child(zio, dde->dde_lead_zio[p]); 2744 else 2745 ddt_phys_addref(ddp); 2746 } else if (zio->io_bp_override) { 2747 ASSERT(bp->blk_birth == txg); 2748 ASSERT(BP_EQUAL(bp, zio->io_bp_override)); 2749 ddt_phys_fill(ddp, bp); 2750 ddt_phys_addref(ddp); 2751 } else { 2752 cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd, 2753 zio->io_orig_size, zio->io_orig_size, zp, 2754 zio_ddt_child_write_ready, NULL, NULL, 2755 zio_ddt_child_write_done, dde, zio->io_priority, 2756 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); 2757 2758 zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL); 2759 dde->dde_lead_zio[p] = cio; 2760 } 2761 2762 ddt_exit(ddt); 2763 2764 if (cio) 2765 zio_nowait(cio); 2766 if (dio) 2767 zio_nowait(dio); 2768 2769 return (ZIO_PIPELINE_CONTINUE); 2770 } 2771 2772 ddt_entry_t *freedde; /* for debugging */ 2773 2774 static int 2775 zio_ddt_free(zio_t *zio) 2776 { 2777 spa_t *spa = zio->io_spa; 2778 blkptr_t *bp = zio->io_bp; 2779 ddt_t *ddt = ddt_select(spa, bp); 2780 ddt_entry_t *dde; 2781 ddt_phys_t *ddp; 2782 2783 ASSERT(BP_GET_DEDUP(bp)); 2784 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2785 2786 ddt_enter(ddt); 2787 freedde = dde = ddt_lookup(ddt, bp, B_TRUE); 2788 ddp = ddt_phys_select(dde, bp); 2789 ddt_phys_decref(ddp); 2790 ddt_exit(ddt); 2791 2792 return (ZIO_PIPELINE_CONTINUE); 2793 } 2794 2795 /* 2796 * ========================================================================== 2797 * Allocate and free blocks 2798 * ========================================================================== 2799 */ 2800 2801 static zio_t * 2802 zio_io_to_allocate(spa_t *spa) 2803 { 2804 zio_t *zio; 2805 2806 ASSERT(MUTEX_HELD(&spa->spa_alloc_lock)); 2807 2808 zio = avl_first(&spa->spa_alloc_tree); 2809 if (zio == NULL) 2810 return (NULL); 2811 2812 ASSERT(IO_IS_ALLOCATING(zio)); 2813 2814 /* 2815 * Try to place a reservation for this zio. If we're unable to 2816 * reserve then we throttle. 2817 */ 2818 if (!metaslab_class_throttle_reserve(spa_normal_class(spa), 2819 zio->io_prop.zp_copies, zio, 0)) { 2820 return (NULL); 2821 } 2822 2823 avl_remove(&spa->spa_alloc_tree, zio); 2824 ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE); 2825 2826 return (zio); 2827 } 2828 2829 static int 2830 zio_dva_throttle(zio_t *zio) 2831 { 2832 spa_t *spa = zio->io_spa; 2833 zio_t *nio; 2834 2835 if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE || 2836 !spa_normal_class(zio->io_spa)->mc_alloc_throttle_enabled || 2837 zio->io_child_type == ZIO_CHILD_GANG || 2838 zio->io_flags & ZIO_FLAG_NODATA) { 2839 return (ZIO_PIPELINE_CONTINUE); 2840 } 2841 2842 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 2843 2844 ASSERT3U(zio->io_queued_timestamp, >, 0); 2845 ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE); 2846 2847 mutex_enter(&spa->spa_alloc_lock); 2848 2849 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 2850 avl_add(&spa->spa_alloc_tree, zio); 2851 2852 nio = zio_io_to_allocate(zio->io_spa); 2853 mutex_exit(&spa->spa_alloc_lock); 2854 2855 if (nio == zio) 2856 return (ZIO_PIPELINE_CONTINUE); 2857 2858 if (nio != NULL) { 2859 ASSERT(nio->io_stage == ZIO_STAGE_DVA_THROTTLE); 2860 /* 2861 * We are passing control to a new zio so make sure that 2862 * it is processed by a different thread. We do this to 2863 * avoid stack overflows that can occur when parents are 2864 * throttled and children are making progress. We allow 2865 * it to go to the head of the taskq since it's already 2866 * been waiting. 2867 */ 2868 zio_taskq_dispatch(nio, ZIO_TASKQ_ISSUE, B_TRUE); 2869 } 2870 return (ZIO_PIPELINE_STOP); 2871 } 2872 2873 void 2874 zio_allocate_dispatch(spa_t *spa) 2875 { 2876 zio_t *zio; 2877 2878 mutex_enter(&spa->spa_alloc_lock); 2879 zio = zio_io_to_allocate(spa); 2880 mutex_exit(&spa->spa_alloc_lock); 2881 if (zio == NULL) 2882 return; 2883 2884 ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE); 2885 ASSERT0(zio->io_error); 2886 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE); 2887 } 2888 2889 static int 2890 zio_dva_allocate(zio_t *zio) 2891 { 2892 spa_t *spa = zio->io_spa; 2893 metaslab_class_t *mc = spa_normal_class(spa); 2894 blkptr_t *bp = zio->io_bp; 2895 int error; 2896 int flags = 0; 2897 2898 if (zio->io_gang_leader == NULL) { 2899 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 2900 zio->io_gang_leader = zio; 2901 } 2902 2903 ASSERT(BP_IS_HOLE(bp)); 2904 ASSERT0(BP_GET_NDVAS(bp)); 2905 ASSERT3U(zio->io_prop.zp_copies, >, 0); 2906 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa)); 2907 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp)); 2908 2909 if (zio->io_flags & ZIO_FLAG_NODATA) { 2910 flags |= METASLAB_DONT_THROTTLE; 2911 } 2912 if (zio->io_flags & ZIO_FLAG_GANG_CHILD) { 2913 flags |= METASLAB_GANG_CHILD; 2914 } 2915 if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE) { 2916 flags |= METASLAB_ASYNC_ALLOC; 2917 } 2918 2919 error = metaslab_alloc(spa, mc, zio->io_size, bp, 2920 zio->io_prop.zp_copies, zio->io_txg, NULL, flags, 2921 &zio->io_alloc_list, zio); 2922 2923 if (error != 0) { 2924 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, " 2925 "size %llu, error %d", spa_name(spa), zio, zio->io_size, 2926 error); 2927 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE) 2928 return (zio_write_gang_block(zio)); 2929 zio->io_error = error; 2930 } 2931 2932 return (ZIO_PIPELINE_CONTINUE); 2933 } 2934 2935 static int 2936 zio_dva_free(zio_t *zio) 2937 { 2938 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE); 2939 2940 return (ZIO_PIPELINE_CONTINUE); 2941 } 2942 2943 static int 2944 zio_dva_claim(zio_t *zio) 2945 { 2946 int error; 2947 2948 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg); 2949 if (error) 2950 zio->io_error = error; 2951 2952 return (ZIO_PIPELINE_CONTINUE); 2953 } 2954 2955 /* 2956 * Undo an allocation. This is used by zio_done() when an I/O fails 2957 * and we want to give back the block we just allocated. 2958 * This handles both normal blocks and gang blocks. 2959 */ 2960 static void 2961 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp) 2962 { 2963 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp)); 2964 ASSERT(zio->io_bp_override == NULL); 2965 2966 if (!BP_IS_HOLE(bp)) 2967 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE); 2968 2969 if (gn != NULL) { 2970 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 2971 zio_dva_unallocate(zio, gn->gn_child[g], 2972 &gn->gn_gbh->zg_blkptr[g]); 2973 } 2974 } 2975 } 2976 2977 /* 2978 * Try to allocate an intent log block. Return 0 on success, errno on failure. 2979 */ 2980 int 2981 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp, 2982 uint64_t size, boolean_t *slog) 2983 { 2984 int error = 1; 2985 zio_alloc_list_t io_alloc_list; 2986 2987 ASSERT(txg > spa_syncing_txg(spa)); 2988 2989 metaslab_trace_init(&io_alloc_list); 2990 error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1, 2991 txg, old_bp, METASLAB_HINTBP_AVOID, &io_alloc_list, NULL); 2992 if (error == 0) { 2993 *slog = TRUE; 2994 } else { 2995 error = metaslab_alloc(spa, spa_normal_class(spa), size, 2996 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID, 2997 &io_alloc_list, NULL); 2998 if (error == 0) 2999 *slog = FALSE; 3000 } 3001 metaslab_trace_fini(&io_alloc_list); 3002 3003 if (error == 0) { 3004 BP_SET_LSIZE(new_bp, size); 3005 BP_SET_PSIZE(new_bp, size); 3006 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF); 3007 BP_SET_CHECKSUM(new_bp, 3008 spa_version(spa) >= SPA_VERSION_SLIM_ZIL 3009 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG); 3010 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); 3011 BP_SET_LEVEL(new_bp, 0); 3012 BP_SET_DEDUP(new_bp, 0); 3013 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER); 3014 } else { 3015 zfs_dbgmsg("%s: zil block allocation failure: " 3016 "size %llu, error %d", spa_name(spa), size, error); 3017 } 3018 3019 return (error); 3020 } 3021 3022 /* 3023 * Free an intent log block. 3024 */ 3025 void 3026 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp) 3027 { 3028 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG); 3029 ASSERT(!BP_IS_GANG(bp)); 3030 3031 zio_free(spa, txg, bp); 3032 } 3033 3034 /* 3035 * ========================================================================== 3036 * Read and write to physical devices 3037 * ========================================================================== 3038 */ 3039 3040 3041 /* 3042 * Issue an I/O to the underlying vdev. Typically the issue pipeline 3043 * stops after this stage and will resume upon I/O completion. 3044 * However, there are instances where the vdev layer may need to 3045 * continue the pipeline when an I/O was not issued. Since the I/O 3046 * that was sent to the vdev layer might be different than the one 3047 * currently active in the pipeline (see vdev_queue_io()), we explicitly 3048 * force the underlying vdev layers to call either zio_execute() or 3049 * zio_interrupt() to ensure that the pipeline continues with the correct I/O. 3050 */ 3051 static int 3052 zio_vdev_io_start(zio_t *zio) 3053 { 3054 vdev_t *vd = zio->io_vd; 3055 uint64_t align; 3056 spa_t *spa = zio->io_spa; 3057 3058 ASSERT(zio->io_error == 0); 3059 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0); 3060 3061 if (vd == NULL) { 3062 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) 3063 spa_config_enter(spa, SCL_ZIO, zio, RW_READER); 3064 3065 /* 3066 * The mirror_ops handle multiple DVAs in a single BP. 3067 */ 3068 vdev_mirror_ops.vdev_op_io_start(zio); 3069 return (ZIO_PIPELINE_STOP); 3070 } 3071 3072 ASSERT3P(zio->io_logical, !=, zio); 3073 if (zio->io_type == ZIO_TYPE_WRITE) { 3074 ASSERT(spa->spa_trust_config); 3075 3076 if (zio->io_vd->vdev_removing) { 3077 ASSERT(zio->io_flags & 3078 (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL | 3079 ZIO_FLAG_INDUCE_DAMAGE)); 3080 } 3081 } 3082 3083 /* 3084 * We keep track of time-sensitive I/Os so that the scan thread 3085 * can quickly react to certain workloads. In particular, we care 3086 * about non-scrubbing, top-level reads and writes with the following 3087 * characteristics: 3088 * - synchronous writes of user data to non-slog devices 3089 * - any reads of user data 3090 * When these conditions are met, adjust the timestamp of spa_last_io 3091 * which allows the scan thread to adjust its workload accordingly. 3092 */ 3093 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL && 3094 vd == vd->vdev_top && !vd->vdev_islog && 3095 zio->io_bookmark.zb_objset != DMU_META_OBJSET && 3096 zio->io_txg != spa_syncing_txg(spa)) { 3097 uint64_t old = spa->spa_last_io; 3098 uint64_t new = ddi_get_lbolt64(); 3099 if (old != new) 3100 (void) atomic_cas_64(&spa->spa_last_io, old, new); 3101 } 3102 3103 align = 1ULL << vd->vdev_top->vdev_ashift; 3104 3105 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) && 3106 P2PHASE(zio->io_size, align) != 0) { 3107 /* Transform logical writes to be a full physical block size. */ 3108 uint64_t asize = P2ROUNDUP(zio->io_size, align); 3109 abd_t *abuf = abd_alloc_sametype(zio->io_abd, asize); 3110 ASSERT(vd == vd->vdev_top); 3111 if (zio->io_type == ZIO_TYPE_WRITE) { 3112 abd_copy(abuf, zio->io_abd, zio->io_size); 3113 abd_zero_off(abuf, zio->io_size, asize - zio->io_size); 3114 } 3115 zio_push_transform(zio, abuf, asize, asize, zio_subblock); 3116 } 3117 3118 /* 3119 * If this is not a physical io, make sure that it is properly aligned 3120 * before proceeding. 3121 */ 3122 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) { 3123 ASSERT0(P2PHASE(zio->io_offset, align)); 3124 ASSERT0(P2PHASE(zio->io_size, align)); 3125 } else { 3126 /* 3127 * For physical writes, we allow 512b aligned writes and assume 3128 * the device will perform a read-modify-write as necessary. 3129 */ 3130 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE)); 3131 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE)); 3132 } 3133 3134 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa)); 3135 3136 /* 3137 * If this is a repair I/O, and there's no self-healing involved -- 3138 * that is, we're just resilvering what we expect to resilver -- 3139 * then don't do the I/O unless zio's txg is actually in vd's DTL. 3140 * This prevents spurious resilvering with nested replication. 3141 * For example, given a mirror of mirrors, (A+B)+(C+D), if only 3142 * A is out of date, we'll read from C+D, then use the data to 3143 * resilver A+B -- but we don't actually want to resilver B, just A. 3144 * The top-level mirror has no way to know this, so instead we just 3145 * discard unnecessary repairs as we work our way down the vdev tree. 3146 * The same logic applies to any form of nested replication: 3147 * ditto + mirror, RAID-Z + replacing, etc. This covers them all. 3148 */ 3149 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) && 3150 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) && 3151 zio->io_txg != 0 && /* not a delegated i/o */ 3152 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) { 3153 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 3154 zio_vdev_io_bypass(zio); 3155 return (ZIO_PIPELINE_CONTINUE); 3156 } 3157 3158 if (vd->vdev_ops->vdev_op_leaf && 3159 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) { 3160 3161 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio)) 3162 return (ZIO_PIPELINE_CONTINUE); 3163 3164 if ((zio = vdev_queue_io(zio)) == NULL) 3165 return (ZIO_PIPELINE_STOP); 3166 3167 if (!vdev_accessible(vd, zio)) { 3168 zio->io_error = SET_ERROR(ENXIO); 3169 zio_interrupt(zio); 3170 return (ZIO_PIPELINE_STOP); 3171 } 3172 } 3173 3174 vd->vdev_ops->vdev_op_io_start(zio); 3175 return (ZIO_PIPELINE_STOP); 3176 } 3177 3178 static int 3179 zio_vdev_io_done(zio_t *zio) 3180 { 3181 vdev_t *vd = zio->io_vd; 3182 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops; 3183 boolean_t unexpected_error = B_FALSE; 3184 3185 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE)) 3186 return (ZIO_PIPELINE_STOP); 3187 3188 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE); 3189 3190 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) { 3191 3192 vdev_queue_io_done(zio); 3193 3194 if (zio->io_type == ZIO_TYPE_WRITE) 3195 vdev_cache_write(zio); 3196 3197 if (zio_injection_enabled && zio->io_error == 0) 3198 zio->io_error = zio_handle_device_injection(vd, 3199 zio, EIO); 3200 3201 if (zio_injection_enabled && zio->io_error == 0) 3202 zio->io_error = zio_handle_label_injection(zio, EIO); 3203 3204 if (zio->io_error) { 3205 if (!vdev_accessible(vd, zio)) { 3206 zio->io_error = SET_ERROR(ENXIO); 3207 } else { 3208 unexpected_error = B_TRUE; 3209 } 3210 } 3211 } 3212 3213 ops->vdev_op_io_done(zio); 3214 3215 if (unexpected_error) 3216 VERIFY(vdev_probe(vd, zio) == NULL); 3217 3218 return (ZIO_PIPELINE_CONTINUE); 3219 } 3220 3221 /* 3222 * For non-raidz ZIOs, we can just copy aside the bad data read from the 3223 * disk, and use that to finish the checksum ereport later. 3224 */ 3225 static void 3226 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr, 3227 const void *good_buf) 3228 { 3229 /* no processing needed */ 3230 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE); 3231 } 3232 3233 /*ARGSUSED*/ 3234 void 3235 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored) 3236 { 3237 void *buf = zio_buf_alloc(zio->io_size); 3238 3239 abd_copy_to_buf(buf, zio->io_abd, zio->io_size); 3240 3241 zcr->zcr_cbinfo = zio->io_size; 3242 zcr->zcr_cbdata = buf; 3243 zcr->zcr_finish = zio_vsd_default_cksum_finish; 3244 zcr->zcr_free = zio_buf_free; 3245 } 3246 3247 static int 3248 zio_vdev_io_assess(zio_t *zio) 3249 { 3250 vdev_t *vd = zio->io_vd; 3251 3252 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE)) 3253 return (ZIO_PIPELINE_STOP); 3254 3255 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) 3256 spa_config_exit(zio->io_spa, SCL_ZIO, zio); 3257 3258 if (zio->io_vsd != NULL) { 3259 zio->io_vsd_ops->vsd_free(zio); 3260 zio->io_vsd = NULL; 3261 } 3262 3263 if (zio_injection_enabled && zio->io_error == 0) 3264 zio->io_error = zio_handle_fault_injection(zio, EIO); 3265 3266 /* 3267 * If the I/O failed, determine whether we should attempt to retry it. 3268 * 3269 * On retry, we cut in line in the issue queue, since we don't want 3270 * compression/checksumming/etc. work to prevent our (cheap) IO reissue. 3271 */ 3272 if (zio->io_error && vd == NULL && 3273 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) { 3274 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */ 3275 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */ 3276 zio->io_error = 0; 3277 zio->io_flags |= ZIO_FLAG_IO_RETRY | 3278 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE; 3279 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1; 3280 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, 3281 zio_requeue_io_start_cut_in_line); 3282 return (ZIO_PIPELINE_STOP); 3283 } 3284 3285 /* 3286 * If we got an error on a leaf device, convert it to ENXIO 3287 * if the device is not accessible at all. 3288 */ 3289 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf && 3290 !vdev_accessible(vd, zio)) 3291 zio->io_error = SET_ERROR(ENXIO); 3292 3293 /* 3294 * If we can't write to an interior vdev (mirror or RAID-Z), 3295 * set vdev_cant_write so that we stop trying to allocate from it. 3296 */ 3297 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE && 3298 vd != NULL && !vd->vdev_ops->vdev_op_leaf) { 3299 vd->vdev_cant_write = B_TRUE; 3300 } 3301 3302 /* 3303 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future 3304 * attempts will ever succeed. In this case we set a persistent bit so 3305 * that we don't bother with it in the future. 3306 */ 3307 if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) && 3308 zio->io_type == ZIO_TYPE_IOCTL && 3309 zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL) 3310 vd->vdev_nowritecache = B_TRUE; 3311 3312 if (zio->io_error) 3313 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 3314 3315 if (vd != NULL && vd->vdev_ops->vdev_op_leaf && 3316 zio->io_physdone != NULL) { 3317 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED)); 3318 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV); 3319 zio->io_physdone(zio->io_logical); 3320 } 3321 3322 return (ZIO_PIPELINE_CONTINUE); 3323 } 3324 3325 void 3326 zio_vdev_io_reissue(zio_t *zio) 3327 { 3328 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); 3329 ASSERT(zio->io_error == 0); 3330 3331 zio->io_stage >>= 1; 3332 } 3333 3334 void 3335 zio_vdev_io_redone(zio_t *zio) 3336 { 3337 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE); 3338 3339 zio->io_stage >>= 1; 3340 } 3341 3342 void 3343 zio_vdev_io_bypass(zio_t *zio) 3344 { 3345 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); 3346 ASSERT(zio->io_error == 0); 3347 3348 zio->io_flags |= ZIO_FLAG_IO_BYPASS; 3349 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1; 3350 } 3351 3352 /* 3353 * ========================================================================== 3354 * Generate and verify checksums 3355 * ========================================================================== 3356 */ 3357 static int 3358 zio_checksum_generate(zio_t *zio) 3359 { 3360 blkptr_t *bp = zio->io_bp; 3361 enum zio_checksum checksum; 3362 3363 if (bp == NULL) { 3364 /* 3365 * This is zio_write_phys(). 3366 * We're either generating a label checksum, or none at all. 3367 */ 3368 checksum = zio->io_prop.zp_checksum; 3369 3370 if (checksum == ZIO_CHECKSUM_OFF) 3371 return (ZIO_PIPELINE_CONTINUE); 3372 3373 ASSERT(checksum == ZIO_CHECKSUM_LABEL); 3374 } else { 3375 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) { 3376 ASSERT(!IO_IS_ALLOCATING(zio)); 3377 checksum = ZIO_CHECKSUM_GANG_HEADER; 3378 } else { 3379 checksum = BP_GET_CHECKSUM(bp); 3380 } 3381 } 3382 3383 zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size); 3384 3385 return (ZIO_PIPELINE_CONTINUE); 3386 } 3387 3388 static int 3389 zio_checksum_verify(zio_t *zio) 3390 { 3391 zio_bad_cksum_t info; 3392 blkptr_t *bp = zio->io_bp; 3393 int error; 3394 3395 ASSERT(zio->io_vd != NULL); 3396 3397 if (bp == NULL) { 3398 /* 3399 * This is zio_read_phys(). 3400 * We're either verifying a label checksum, or nothing at all. 3401 */ 3402 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF) 3403 return (ZIO_PIPELINE_CONTINUE); 3404 3405 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL); 3406 } 3407 3408 if ((error = zio_checksum_error(zio, &info)) != 0) { 3409 zio->io_error = error; 3410 if (error == ECKSUM && 3411 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) { 3412 zfs_ereport_start_checksum(zio->io_spa, 3413 zio->io_vd, zio, zio->io_offset, 3414 zio->io_size, NULL, &info); 3415 } 3416 } 3417 3418 return (ZIO_PIPELINE_CONTINUE); 3419 } 3420 3421 /* 3422 * Called by RAID-Z to ensure we don't compute the checksum twice. 3423 */ 3424 void 3425 zio_checksum_verified(zio_t *zio) 3426 { 3427 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; 3428 } 3429 3430 /* 3431 * ========================================================================== 3432 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other. 3433 * An error of 0 indicates success. ENXIO indicates whole-device failure, 3434 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO 3435 * indicate errors that are specific to one I/O, and most likely permanent. 3436 * Any other error is presumed to be worse because we weren't expecting it. 3437 * ========================================================================== 3438 */ 3439 int 3440 zio_worst_error(int e1, int e2) 3441 { 3442 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO }; 3443 int r1, r2; 3444 3445 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++) 3446 if (e1 == zio_error_rank[r1]) 3447 break; 3448 3449 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++) 3450 if (e2 == zio_error_rank[r2]) 3451 break; 3452 3453 return (r1 > r2 ? e1 : e2); 3454 } 3455 3456 /* 3457 * ========================================================================== 3458 * I/O completion 3459 * ========================================================================== 3460 */ 3461 static int 3462 zio_ready(zio_t *zio) 3463 { 3464 blkptr_t *bp = zio->io_bp; 3465 zio_t *pio, *pio_next; 3466 zio_link_t *zl = NULL; 3467 3468 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) || 3469 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY)) 3470 return (ZIO_PIPELINE_STOP); 3471 3472 if (zio->io_ready) { 3473 ASSERT(IO_IS_ALLOCATING(zio)); 3474 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) || 3475 (zio->io_flags & ZIO_FLAG_NOPWRITE)); 3476 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0); 3477 3478 zio->io_ready(zio); 3479 } 3480 3481 if (bp != NULL && bp != &zio->io_bp_copy) 3482 zio->io_bp_copy = *bp; 3483 3484 if (zio->io_error != 0) { 3485 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 3486 3487 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) { 3488 ASSERT(IO_IS_ALLOCATING(zio)); 3489 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); 3490 /* 3491 * We were unable to allocate anything, unreserve and 3492 * issue the next I/O to allocate. 3493 */ 3494 metaslab_class_throttle_unreserve( 3495 spa_normal_class(zio->io_spa), 3496 zio->io_prop.zp_copies, zio); 3497 zio_allocate_dispatch(zio->io_spa); 3498 } 3499 } 3500 3501 mutex_enter(&zio->io_lock); 3502 zio->io_state[ZIO_WAIT_READY] = 1; 3503 pio = zio_walk_parents(zio, &zl); 3504 mutex_exit(&zio->io_lock); 3505 3506 /* 3507 * As we notify zio's parents, new parents could be added. 3508 * New parents go to the head of zio's io_parent_list, however, 3509 * so we will (correctly) not notify them. The remainder of zio's 3510 * io_parent_list, from 'pio_next' onward, cannot change because 3511 * all parents must wait for us to be done before they can be done. 3512 */ 3513 for (; pio != NULL; pio = pio_next) { 3514 pio_next = zio_walk_parents(zio, &zl); 3515 zio_notify_parent(pio, zio, ZIO_WAIT_READY); 3516 } 3517 3518 if (zio->io_flags & ZIO_FLAG_NODATA) { 3519 if (BP_IS_GANG(bp)) { 3520 zio->io_flags &= ~ZIO_FLAG_NODATA; 3521 } else { 3522 ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE); 3523 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; 3524 } 3525 } 3526 3527 if (zio_injection_enabled && 3528 zio->io_spa->spa_syncing_txg == zio->io_txg) 3529 zio_handle_ignored_writes(zio); 3530 3531 return (ZIO_PIPELINE_CONTINUE); 3532 } 3533 3534 /* 3535 * Update the allocation throttle accounting. 3536 */ 3537 static void 3538 zio_dva_throttle_done(zio_t *zio) 3539 { 3540 zio_t *lio = zio->io_logical; 3541 zio_t *pio = zio_unique_parent(zio); 3542 vdev_t *vd = zio->io_vd; 3543 int flags = METASLAB_ASYNC_ALLOC; 3544 3545 ASSERT3P(zio->io_bp, !=, NULL); 3546 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE); 3547 ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE); 3548 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV); 3549 ASSERT(vd != NULL); 3550 ASSERT3P(vd, ==, vd->vdev_top); 3551 ASSERT(!(zio->io_flags & (ZIO_FLAG_IO_REPAIR | ZIO_FLAG_IO_RETRY))); 3552 ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING); 3553 ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE)); 3554 ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA)); 3555 3556 /* 3557 * Parents of gang children can have two flavors -- ones that 3558 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set) 3559 * and ones that allocated the constituent blocks. The allocation 3560 * throttle needs to know the allocating parent zio so we must find 3561 * it here. 3562 */ 3563 if (pio->io_child_type == ZIO_CHILD_GANG) { 3564 /* 3565 * If our parent is a rewrite gang child then our grandparent 3566 * would have been the one that performed the allocation. 3567 */ 3568 if (pio->io_flags & ZIO_FLAG_IO_REWRITE) 3569 pio = zio_unique_parent(pio); 3570 flags |= METASLAB_GANG_CHILD; 3571 } 3572 3573 ASSERT(IO_IS_ALLOCATING(pio)); 3574 ASSERT3P(zio, !=, zio->io_logical); 3575 ASSERT(zio->io_logical != NULL); 3576 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR)); 3577 ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE); 3578 3579 mutex_enter(&pio->io_lock); 3580 metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags); 3581 mutex_exit(&pio->io_lock); 3582 3583 metaslab_class_throttle_unreserve(spa_normal_class(zio->io_spa), 3584 1, pio); 3585 3586 /* 3587 * Call into the pipeline to see if there is more work that 3588 * needs to be done. If there is work to be done it will be 3589 * dispatched to another taskq thread. 3590 */ 3591 zio_allocate_dispatch(zio->io_spa); 3592 } 3593 3594 static int 3595 zio_done(zio_t *zio) 3596 { 3597 spa_t *spa = zio->io_spa; 3598 zio_t *lio = zio->io_logical; 3599 blkptr_t *bp = zio->io_bp; 3600 vdev_t *vd = zio->io_vd; 3601 uint64_t psize = zio->io_size; 3602 zio_t *pio, *pio_next; 3603 metaslab_class_t *mc = spa_normal_class(spa); 3604 zio_link_t *zl = NULL; 3605 3606 /* 3607 * If our children haven't all completed, 3608 * wait for them and then repeat this pipeline stage. 3609 */ 3610 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) || 3611 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) || 3612 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) || 3613 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE)) 3614 return (ZIO_PIPELINE_STOP); 3615 3616 /* 3617 * If the allocation throttle is enabled, then update the accounting. 3618 * We only track child I/Os that are part of an allocating async 3619 * write. We must do this since the allocation is performed 3620 * by the logical I/O but the actual write is done by child I/Os. 3621 */ 3622 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING && 3623 zio->io_child_type == ZIO_CHILD_VDEV) { 3624 ASSERT(mc->mc_alloc_throttle_enabled); 3625 zio_dva_throttle_done(zio); 3626 } 3627 3628 /* 3629 * If the allocation throttle is enabled, verify that 3630 * we have decremented the refcounts for every I/O that was throttled. 3631 */ 3632 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) { 3633 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 3634 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); 3635 ASSERT(bp != NULL); 3636 metaslab_group_alloc_verify(spa, zio->io_bp, zio); 3637 VERIFY(refcount_not_held(&mc->mc_alloc_slots, zio)); 3638 } 3639 3640 for (int c = 0; c < ZIO_CHILD_TYPES; c++) 3641 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 3642 ASSERT(zio->io_children[c][w] == 0); 3643 3644 if (bp != NULL && !BP_IS_EMBEDDED(bp)) { 3645 ASSERT(bp->blk_pad[0] == 0); 3646 ASSERT(bp->blk_pad[1] == 0); 3647 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 || 3648 (bp == zio_unique_parent(zio)->io_bp)); 3649 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) && 3650 zio->io_bp_override == NULL && 3651 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) { 3652 ASSERT(!BP_SHOULD_BYTESWAP(bp)); 3653 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp)); 3654 ASSERT(BP_COUNT_GANG(bp) == 0 || 3655 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp))); 3656 } 3657 if (zio->io_flags & ZIO_FLAG_NOPWRITE) 3658 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig)); 3659 } 3660 3661 /* 3662 * If there were child vdev/gang/ddt errors, they apply to us now. 3663 */ 3664 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV); 3665 zio_inherit_child_errors(zio, ZIO_CHILD_GANG); 3666 zio_inherit_child_errors(zio, ZIO_CHILD_DDT); 3667 3668 /* 3669 * If the I/O on the transformed data was successful, generate any 3670 * checksum reports now while we still have the transformed data. 3671 */ 3672 if (zio->io_error == 0) { 3673 while (zio->io_cksum_report != NULL) { 3674 zio_cksum_report_t *zcr = zio->io_cksum_report; 3675 uint64_t align = zcr->zcr_align; 3676 uint64_t asize = P2ROUNDUP(psize, align); 3677 char *abuf = NULL; 3678 abd_t *adata = zio->io_abd; 3679 3680 if (asize != psize) { 3681 adata = abd_alloc_linear(asize, B_TRUE); 3682 abd_copy(adata, zio->io_abd, psize); 3683 abd_zero_off(adata, psize, asize - psize); 3684 } 3685 3686 if (adata != NULL) 3687 abuf = abd_borrow_buf_copy(adata, asize); 3688 3689 zio->io_cksum_report = zcr->zcr_next; 3690 zcr->zcr_next = NULL; 3691 zcr->zcr_finish(zcr, abuf); 3692 zfs_ereport_free_checksum(zcr); 3693 3694 if (adata != NULL) 3695 abd_return_buf(adata, abuf, asize); 3696 3697 if (asize != psize) 3698 abd_free(adata); 3699 } 3700 } 3701 3702 zio_pop_transforms(zio); /* note: may set zio->io_error */ 3703 3704 vdev_stat_update(zio, psize); 3705 3706 if (zio->io_error) { 3707 /* 3708 * If this I/O is attached to a particular vdev, 3709 * generate an error message describing the I/O failure 3710 * at the block level. We ignore these errors if the 3711 * device is currently unavailable. 3712 */ 3713 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd)) 3714 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0); 3715 3716 if ((zio->io_error == EIO || !(zio->io_flags & 3717 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) && 3718 zio == lio) { 3719 /* 3720 * For logical I/O requests, tell the SPA to log the 3721 * error and generate a logical data ereport. 3722 */ 3723 spa_log_error(spa, zio); 3724 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio, 3725 0, 0); 3726 } 3727 } 3728 3729 if (zio->io_error && zio == lio) { 3730 /* 3731 * Determine whether zio should be reexecuted. This will 3732 * propagate all the way to the root via zio_notify_parent(). 3733 */ 3734 ASSERT(vd == NULL && bp != NULL); 3735 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 3736 3737 if (IO_IS_ALLOCATING(zio) && 3738 !(zio->io_flags & ZIO_FLAG_CANFAIL)) { 3739 if (zio->io_error != ENOSPC) 3740 zio->io_reexecute |= ZIO_REEXECUTE_NOW; 3741 else 3742 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 3743 } 3744 3745 if ((zio->io_type == ZIO_TYPE_READ || 3746 zio->io_type == ZIO_TYPE_FREE) && 3747 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && 3748 zio->io_error == ENXIO && 3749 spa_load_state(spa) == SPA_LOAD_NONE && 3750 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE) 3751 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 3752 3753 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute) 3754 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 3755 3756 /* 3757 * Here is a possibly good place to attempt to do 3758 * either combinatorial reconstruction or error correction 3759 * based on checksums. It also might be a good place 3760 * to send out preliminary ereports before we suspend 3761 * processing. 3762 */ 3763 } 3764 3765 /* 3766 * If there were logical child errors, they apply to us now. 3767 * We defer this until now to avoid conflating logical child 3768 * errors with errors that happened to the zio itself when 3769 * updating vdev stats and reporting FMA events above. 3770 */ 3771 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL); 3772 3773 if ((zio->io_error || zio->io_reexecute) && 3774 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio && 3775 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE))) 3776 zio_dva_unallocate(zio, zio->io_gang_tree, bp); 3777 3778 zio_gang_tree_free(&zio->io_gang_tree); 3779 3780 /* 3781 * Godfather I/Os should never suspend. 3782 */ 3783 if ((zio->io_flags & ZIO_FLAG_GODFATHER) && 3784 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) 3785 zio->io_reexecute = 0; 3786 3787 if (zio->io_reexecute) { 3788 /* 3789 * This is a logical I/O that wants to reexecute. 3790 * 3791 * Reexecute is top-down. When an i/o fails, if it's not 3792 * the root, it simply notifies its parent and sticks around. 3793 * The parent, seeing that it still has children in zio_done(), 3794 * does the same. This percolates all the way up to the root. 3795 * The root i/o will reexecute or suspend the entire tree. 3796 * 3797 * This approach ensures that zio_reexecute() honors 3798 * all the original i/o dependency relationships, e.g. 3799 * parents not executing until children are ready. 3800 */ 3801 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 3802 3803 zio->io_gang_leader = NULL; 3804 3805 mutex_enter(&zio->io_lock); 3806 zio->io_state[ZIO_WAIT_DONE] = 1; 3807 mutex_exit(&zio->io_lock); 3808 3809 /* 3810 * "The Godfather" I/O monitors its children but is 3811 * not a true parent to them. It will track them through 3812 * the pipeline but severs its ties whenever they get into 3813 * trouble (e.g. suspended). This allows "The Godfather" 3814 * I/O to return status without blocking. 3815 */ 3816 zl = NULL; 3817 for (pio = zio_walk_parents(zio, &zl); pio != NULL; 3818 pio = pio_next) { 3819 zio_link_t *remove_zl = zl; 3820 pio_next = zio_walk_parents(zio, &zl); 3821 3822 if ((pio->io_flags & ZIO_FLAG_GODFATHER) && 3823 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) { 3824 zio_remove_child(pio, zio, remove_zl); 3825 zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 3826 } 3827 } 3828 3829 if ((pio = zio_unique_parent(zio)) != NULL) { 3830 /* 3831 * We're not a root i/o, so there's nothing to do 3832 * but notify our parent. Don't propagate errors 3833 * upward since we haven't permanently failed yet. 3834 */ 3835 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); 3836 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE; 3837 zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 3838 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) { 3839 /* 3840 * We'd fail again if we reexecuted now, so suspend 3841 * until conditions improve (e.g. device comes online). 3842 */ 3843 zio_suspend(spa, zio); 3844 } else { 3845 /* 3846 * Reexecution is potentially a huge amount of work. 3847 * Hand it off to the otherwise-unused claim taskq. 3848 */ 3849 ASSERT(zio->io_tqent.tqent_next == NULL); 3850 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM, 3851 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio, 3852 0, &zio->io_tqent); 3853 } 3854 return (ZIO_PIPELINE_STOP); 3855 } 3856 3857 ASSERT(zio->io_child_count == 0); 3858 ASSERT(zio->io_reexecute == 0); 3859 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL)); 3860 3861 /* 3862 * Report any checksum errors, since the I/O is complete. 3863 */ 3864 while (zio->io_cksum_report != NULL) { 3865 zio_cksum_report_t *zcr = zio->io_cksum_report; 3866 zio->io_cksum_report = zcr->zcr_next; 3867 zcr->zcr_next = NULL; 3868 zcr->zcr_finish(zcr, NULL); 3869 zfs_ereport_free_checksum(zcr); 3870 } 3871 3872 /* 3873 * It is the responsibility of the done callback to ensure that this 3874 * particular zio is no longer discoverable for adoption, and as 3875 * such, cannot acquire any new parents. 3876 */ 3877 if (zio->io_done) 3878 zio->io_done(zio); 3879 3880 mutex_enter(&zio->io_lock); 3881 zio->io_state[ZIO_WAIT_DONE] = 1; 3882 mutex_exit(&zio->io_lock); 3883 3884 zl = NULL; 3885 for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) { 3886 zio_link_t *remove_zl = zl; 3887 pio_next = zio_walk_parents(zio, &zl); 3888 zio_remove_child(pio, zio, remove_zl); 3889 zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 3890 } 3891 3892 if (zio->io_waiter != NULL) { 3893 mutex_enter(&zio->io_lock); 3894 zio->io_executor = NULL; 3895 cv_broadcast(&zio->io_cv); 3896 mutex_exit(&zio->io_lock); 3897 } else { 3898 zio_destroy(zio); 3899 } 3900 3901 return (ZIO_PIPELINE_STOP); 3902 } 3903 3904 /* 3905 * ========================================================================== 3906 * I/O pipeline definition 3907 * ========================================================================== 3908 */ 3909 static zio_pipe_stage_t *zio_pipeline[] = { 3910 NULL, 3911 zio_read_bp_init, 3912 zio_write_bp_init, 3913 zio_free_bp_init, 3914 zio_issue_async, 3915 zio_write_compress, 3916 zio_checksum_generate, 3917 zio_nop_write, 3918 zio_ddt_read_start, 3919 zio_ddt_read_done, 3920 zio_ddt_write, 3921 zio_ddt_free, 3922 zio_gang_assemble, 3923 zio_gang_issue, 3924 zio_dva_throttle, 3925 zio_dva_allocate, 3926 zio_dva_free, 3927 zio_dva_claim, 3928 zio_ready, 3929 zio_vdev_io_start, 3930 zio_vdev_io_done, 3931 zio_vdev_io_assess, 3932 zio_checksum_verify, 3933 zio_done 3934 }; 3935 3936 3937 3938 3939 /* 3940 * Compare two zbookmark_phys_t's to see which we would reach first in a 3941 * pre-order traversal of the object tree. 3942 * 3943 * This is simple in every case aside from the meta-dnode object. For all other 3944 * objects, we traverse them in order (object 1 before object 2, and so on). 3945 * However, all of these objects are traversed while traversing object 0, since 3946 * the data it points to is the list of objects. Thus, we need to convert to a 3947 * canonical representation so we can compare meta-dnode bookmarks to 3948 * non-meta-dnode bookmarks. 3949 * 3950 * We do this by calculating "equivalents" for each field of the zbookmark. 3951 * zbookmarks outside of the meta-dnode use their own object and level, and 3952 * calculate the level 0 equivalent (the first L0 blkid that is contained in the 3953 * blocks this bookmark refers to) by multiplying their blkid by their span 3954 * (the number of L0 blocks contained within one block at their level). 3955 * zbookmarks inside the meta-dnode calculate their object equivalent 3956 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use 3957 * level + 1<<31 (any value larger than a level could ever be) for their level. 3958 * This causes them to always compare before a bookmark in their object 3959 * equivalent, compare appropriately to bookmarks in other objects, and to 3960 * compare appropriately to other bookmarks in the meta-dnode. 3961 */ 3962 int 3963 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2, 3964 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2) 3965 { 3966 /* 3967 * These variables represent the "equivalent" values for the zbookmark, 3968 * after converting zbookmarks inside the meta dnode to their 3969 * normal-object equivalents. 3970 */ 3971 uint64_t zb1obj, zb2obj; 3972 uint64_t zb1L0, zb2L0; 3973 uint64_t zb1level, zb2level; 3974 3975 if (zb1->zb_object == zb2->zb_object && 3976 zb1->zb_level == zb2->zb_level && 3977 zb1->zb_blkid == zb2->zb_blkid) 3978 return (0); 3979 3980 /* 3981 * BP_SPANB calculates the span in blocks. 3982 */ 3983 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level); 3984 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level); 3985 3986 if (zb1->zb_object == DMU_META_DNODE_OBJECT) { 3987 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT)); 3988 zb1L0 = 0; 3989 zb1level = zb1->zb_level + COMPARE_META_LEVEL; 3990 } else { 3991 zb1obj = zb1->zb_object; 3992 zb1level = zb1->zb_level; 3993 } 3994 3995 if (zb2->zb_object == DMU_META_DNODE_OBJECT) { 3996 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT)); 3997 zb2L0 = 0; 3998 zb2level = zb2->zb_level + COMPARE_META_LEVEL; 3999 } else { 4000 zb2obj = zb2->zb_object; 4001 zb2level = zb2->zb_level; 4002 } 4003 4004 /* Now that we have a canonical representation, do the comparison. */ 4005 if (zb1obj != zb2obj) 4006 return (zb1obj < zb2obj ? -1 : 1); 4007 else if (zb1L0 != zb2L0) 4008 return (zb1L0 < zb2L0 ? -1 : 1); 4009 else if (zb1level != zb2level) 4010 return (zb1level > zb2level ? -1 : 1); 4011 /* 4012 * This can (theoretically) happen if the bookmarks have the same object 4013 * and level, but different blkids, if the block sizes are not the same. 4014 * There is presently no way to change the indirect block sizes 4015 */ 4016 return (0); 4017 } 4018 4019 /* 4020 * This function checks the following: given that last_block is the place that 4021 * our traversal stopped last time, does that guarantee that we've visited 4022 * every node under subtree_root? Therefore, we can't just use the raw output 4023 * of zbookmark_compare. We have to pass in a modified version of 4024 * subtree_root; by incrementing the block id, and then checking whether 4025 * last_block is before or equal to that, we can tell whether or not having 4026 * visited last_block implies that all of subtree_root's children have been 4027 * visited. 4028 */ 4029 boolean_t 4030 zbookmark_subtree_completed(const dnode_phys_t *dnp, 4031 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block) 4032 { 4033 zbookmark_phys_t mod_zb = *subtree_root; 4034 mod_zb.zb_blkid++; 4035 ASSERT(last_block->zb_level == 0); 4036 4037 /* The objset_phys_t isn't before anything. */ 4038 if (dnp == NULL) 4039 return (B_FALSE); 4040 4041 /* 4042 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the 4043 * data block size in sectors, because that variable is only used if 4044 * the bookmark refers to a block in the meta-dnode. Since we don't 4045 * know without examining it what object it refers to, and there's no 4046 * harm in passing in this value in other cases, we always pass it in. 4047 * 4048 * We pass in 0 for the indirect block size shift because zb2 must be 4049 * level 0. The indirect block size is only used to calculate the span 4050 * of the bookmark, but since the bookmark must be level 0, the span is 4051 * always 1, so the math works out. 4052 * 4053 * If you make changes to how the zbookmark_compare code works, be sure 4054 * to make sure that this code still works afterwards. 4055 */ 4056 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift, 4057 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb, 4058 last_block) <= 0); 4059 } 4060