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, 2018 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 #include <sys/cityhash.h> 46 47 /* 48 * ========================================================================== 49 * I/O type descriptions 50 * ========================================================================== 51 */ 52 const char *zio_type_name[ZIO_TYPES] = { 53 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim", 54 "zio_ioctl" 55 }; 56 57 boolean_t zio_dva_throttle_enabled = B_TRUE; 58 59 /* 60 * ========================================================================== 61 * I/O kmem caches 62 * ========================================================================== 63 */ 64 kmem_cache_t *zio_cache; 65 kmem_cache_t *zio_link_cache; 66 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; 67 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; 68 69 #ifdef _KERNEL 70 extern vmem_t *zio_alloc_arena; 71 #endif 72 73 #define ZIO_PIPELINE_CONTINUE 0x100 74 #define ZIO_PIPELINE_STOP 0x101 75 76 #define BP_SPANB(indblkshift, level) \ 77 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT))) 78 #define COMPARE_META_LEVEL 0x80000000ul 79 /* 80 * The following actions directly effect the spa's sync-to-convergence logic. 81 * The values below define the sync pass when we start performing the action. 82 * Care should be taken when changing these values as they directly impact 83 * spa_sync() performance. Tuning these values may introduce subtle performance 84 * pathologies and should only be done in the context of performance analysis. 85 * These tunables will eventually be removed and replaced with #defines once 86 * enough analysis has been done to determine optimal values. 87 * 88 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that 89 * regular blocks are not deferred. 90 */ 91 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */ 92 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */ 93 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */ 94 95 /* 96 * An allocating zio is one that either currently has the DVA allocate 97 * stage set or will have it later in its lifetime. 98 */ 99 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE) 100 101 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE; 102 103 #ifdef ZFS_DEBUG 104 int zio_buf_debug_limit = 16384; 105 #else 106 int zio_buf_debug_limit = 0; 107 #endif 108 109 static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t); 110 111 void 112 zio_init(void) 113 { 114 size_t c; 115 vmem_t *data_alloc_arena = NULL; 116 117 #ifdef _KERNEL 118 data_alloc_arena = zio_alloc_arena; 119 #endif 120 zio_cache = kmem_cache_create("zio_cache", 121 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0); 122 zio_link_cache = kmem_cache_create("zio_link_cache", 123 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0); 124 125 /* 126 * For small buffers, we want a cache for each multiple of 127 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache 128 * for each quarter-power of 2. 129 */ 130 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { 131 size_t size = (c + 1) << SPA_MINBLOCKSHIFT; 132 size_t p2 = size; 133 size_t align = 0; 134 size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0; 135 136 while (!ISP2(p2)) 137 p2 &= p2 - 1; 138 139 #ifndef _KERNEL 140 /* 141 * If we are using watchpoints, put each buffer on its own page, 142 * to eliminate the performance overhead of trapping to the 143 * kernel when modifying a non-watched buffer that shares the 144 * page with a watched buffer. 145 */ 146 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE)) 147 continue; 148 #endif 149 if (size <= 4 * SPA_MINBLOCKSIZE) { 150 align = SPA_MINBLOCKSIZE; 151 } else if (IS_P2ALIGNED(size, p2 >> 2)) { 152 align = MIN(p2 >> 2, PAGESIZE); 153 } 154 155 if (align != 0) { 156 char name[36]; 157 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size); 158 zio_buf_cache[c] = kmem_cache_create(name, size, 159 align, NULL, NULL, NULL, NULL, NULL, cflags); 160 161 /* 162 * Since zio_data bufs do not appear in crash dumps, we 163 * pass KMC_NOTOUCH so that no allocator metadata is 164 * stored with the buffers. 165 */ 166 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size); 167 zio_data_buf_cache[c] = kmem_cache_create(name, size, 168 align, NULL, NULL, NULL, NULL, data_alloc_arena, 169 cflags | KMC_NOTOUCH); 170 } 171 } 172 173 while (--c != 0) { 174 ASSERT(zio_buf_cache[c] != NULL); 175 if (zio_buf_cache[c - 1] == NULL) 176 zio_buf_cache[c - 1] = zio_buf_cache[c]; 177 178 ASSERT(zio_data_buf_cache[c] != NULL); 179 if (zio_data_buf_cache[c - 1] == NULL) 180 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c]; 181 } 182 183 zio_inject_init(); 184 } 185 186 void 187 zio_fini(void) 188 { 189 size_t c; 190 kmem_cache_t *last_cache = NULL; 191 kmem_cache_t *last_data_cache = NULL; 192 193 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { 194 if (zio_buf_cache[c] != last_cache) { 195 last_cache = zio_buf_cache[c]; 196 kmem_cache_destroy(zio_buf_cache[c]); 197 } 198 zio_buf_cache[c] = NULL; 199 200 if (zio_data_buf_cache[c] != last_data_cache) { 201 last_data_cache = zio_data_buf_cache[c]; 202 kmem_cache_destroy(zio_data_buf_cache[c]); 203 } 204 zio_data_buf_cache[c] = NULL; 205 } 206 207 kmem_cache_destroy(zio_link_cache); 208 kmem_cache_destroy(zio_cache); 209 210 zio_inject_fini(); 211 } 212 213 /* 214 * ========================================================================== 215 * Allocate and free I/O buffers 216 * ========================================================================== 217 */ 218 219 /* 220 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a 221 * crashdump if the kernel panics, so use it judiciously. Obviously, it's 222 * useful to inspect ZFS metadata, but if possible, we should avoid keeping 223 * excess / transient data in-core during a crashdump. 224 */ 225 void * 226 zio_buf_alloc(size_t size) 227 { 228 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 229 230 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 231 232 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE)); 233 } 234 235 /* 236 * Use zio_data_buf_alloc to allocate data. The data will not appear in a 237 * crashdump if the kernel panics. This exists so that we will limit the amount 238 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount 239 * of kernel heap dumped to disk when the kernel panics) 240 */ 241 void * 242 zio_data_buf_alloc(size_t size) 243 { 244 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 245 246 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 247 248 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE)); 249 } 250 251 void 252 zio_buf_free(void *buf, size_t size) 253 { 254 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 255 256 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 257 258 kmem_cache_free(zio_buf_cache[c], buf); 259 } 260 261 void 262 zio_data_buf_free(void *buf, size_t size) 263 { 264 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 265 266 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 267 268 kmem_cache_free(zio_data_buf_cache[c], buf); 269 } 270 271 /* 272 * ========================================================================== 273 * Push and pop I/O transform buffers 274 * ========================================================================== 275 */ 276 void 277 zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize, 278 zio_transform_func_t *transform) 279 { 280 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP); 281 282 /* 283 * Ensure that anyone expecting this zio to contain a linear ABD isn't 284 * going to get a nasty surprise when they try to access the data. 285 */ 286 IMPLY(abd_is_linear(zio->io_abd), abd_is_linear(data)); 287 288 zt->zt_orig_abd = zio->io_abd; 289 zt->zt_orig_size = zio->io_size; 290 zt->zt_bufsize = bufsize; 291 zt->zt_transform = transform; 292 293 zt->zt_next = zio->io_transform_stack; 294 zio->io_transform_stack = zt; 295 296 zio->io_abd = data; 297 zio->io_size = size; 298 } 299 300 void 301 zio_pop_transforms(zio_t *zio) 302 { 303 zio_transform_t *zt; 304 305 while ((zt = zio->io_transform_stack) != NULL) { 306 if (zt->zt_transform != NULL) 307 zt->zt_transform(zio, 308 zt->zt_orig_abd, zt->zt_orig_size); 309 310 if (zt->zt_bufsize != 0) 311 abd_free(zio->io_abd); 312 313 zio->io_abd = zt->zt_orig_abd; 314 zio->io_size = zt->zt_orig_size; 315 zio->io_transform_stack = zt->zt_next; 316 317 kmem_free(zt, sizeof (zio_transform_t)); 318 } 319 } 320 321 /* 322 * ========================================================================== 323 * I/O transform callbacks for subblocks and decompression 324 * ========================================================================== 325 */ 326 static void 327 zio_subblock(zio_t *zio, abd_t *data, uint64_t size) 328 { 329 ASSERT(zio->io_size > size); 330 331 if (zio->io_type == ZIO_TYPE_READ) 332 abd_copy(data, zio->io_abd, size); 333 } 334 335 static void 336 zio_decompress(zio_t *zio, abd_t *data, uint64_t size) 337 { 338 if (zio->io_error == 0) { 339 void *tmp = abd_borrow_buf(data, size); 340 int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp), 341 zio->io_abd, tmp, zio->io_size, size); 342 abd_return_buf_copy(data, tmp, size); 343 344 if (ret != 0) 345 zio->io_error = SET_ERROR(EIO); 346 } 347 } 348 349 /* 350 * ========================================================================== 351 * I/O parent/child relationships and pipeline interlocks 352 * ========================================================================== 353 */ 354 zio_t * 355 zio_walk_parents(zio_t *cio, zio_link_t **zl) 356 { 357 list_t *pl = &cio->io_parent_list; 358 359 *zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl); 360 if (*zl == NULL) 361 return (NULL); 362 363 ASSERT((*zl)->zl_child == cio); 364 return ((*zl)->zl_parent); 365 } 366 367 zio_t * 368 zio_walk_children(zio_t *pio, zio_link_t **zl) 369 { 370 list_t *cl = &pio->io_child_list; 371 372 *zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl); 373 if (*zl == NULL) 374 return (NULL); 375 376 ASSERT((*zl)->zl_parent == pio); 377 return ((*zl)->zl_child); 378 } 379 380 zio_t * 381 zio_unique_parent(zio_t *cio) 382 { 383 zio_link_t *zl = NULL; 384 zio_t *pio = zio_walk_parents(cio, &zl); 385 386 VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL); 387 return (pio); 388 } 389 390 void 391 zio_add_child(zio_t *pio, zio_t *cio) 392 { 393 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP); 394 395 /* 396 * Logical I/Os can have logical, gang, or vdev children. 397 * Gang I/Os can have gang or vdev children. 398 * Vdev I/Os can only have vdev children. 399 * The following ASSERT captures all of these constraints. 400 */ 401 ASSERT3S(cio->io_child_type, <=, pio->io_child_type); 402 403 zl->zl_parent = pio; 404 zl->zl_child = cio; 405 406 mutex_enter(&cio->io_lock); 407 mutex_enter(&pio->io_lock); 408 409 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0); 410 411 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 412 pio->io_children[cio->io_child_type][w] += !cio->io_state[w]; 413 414 list_insert_head(&pio->io_child_list, zl); 415 list_insert_head(&cio->io_parent_list, zl); 416 417 pio->io_child_count++; 418 cio->io_parent_count++; 419 420 mutex_exit(&pio->io_lock); 421 mutex_exit(&cio->io_lock); 422 } 423 424 static void 425 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl) 426 { 427 ASSERT(zl->zl_parent == pio); 428 ASSERT(zl->zl_child == cio); 429 430 mutex_enter(&cio->io_lock); 431 mutex_enter(&pio->io_lock); 432 433 list_remove(&pio->io_child_list, zl); 434 list_remove(&cio->io_parent_list, zl); 435 436 pio->io_child_count--; 437 cio->io_parent_count--; 438 439 mutex_exit(&pio->io_lock); 440 mutex_exit(&cio->io_lock); 441 442 kmem_cache_free(zio_link_cache, zl); 443 } 444 445 static boolean_t 446 zio_wait_for_children(zio_t *zio, uint8_t childbits, enum zio_wait_type wait) 447 { 448 boolean_t waiting = B_FALSE; 449 450 mutex_enter(&zio->io_lock); 451 ASSERT(zio->io_stall == NULL); 452 for (int c = 0; c < ZIO_CHILD_TYPES; c++) { 453 if (!(ZIO_CHILD_BIT_IS_SET(childbits, c))) 454 continue; 455 456 uint64_t *countp = &zio->io_children[c][wait]; 457 if (*countp != 0) { 458 zio->io_stage >>= 1; 459 ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN); 460 zio->io_stall = countp; 461 waiting = B_TRUE; 462 break; 463 } 464 } 465 mutex_exit(&zio->io_lock); 466 return (waiting); 467 } 468 469 static void 470 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait) 471 { 472 uint64_t *countp = &pio->io_children[zio->io_child_type][wait]; 473 int *errorp = &pio->io_child_error[zio->io_child_type]; 474 475 mutex_enter(&pio->io_lock); 476 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE)) 477 *errorp = zio_worst_error(*errorp, zio->io_error); 478 pio->io_reexecute |= zio->io_reexecute; 479 ASSERT3U(*countp, >, 0); 480 481 (*countp)--; 482 483 if (*countp == 0 && pio->io_stall == countp) { 484 zio_taskq_type_t type = 485 pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE : 486 ZIO_TASKQ_INTERRUPT; 487 pio->io_stall = NULL; 488 mutex_exit(&pio->io_lock); 489 /* 490 * Dispatch the parent zio in its own taskq so that 491 * the child can continue to make progress. This also 492 * prevents overflowing the stack when we have deeply nested 493 * parent-child relationships. 494 */ 495 zio_taskq_dispatch(pio, type, B_FALSE); 496 } else { 497 mutex_exit(&pio->io_lock); 498 } 499 } 500 501 static void 502 zio_inherit_child_errors(zio_t *zio, enum zio_child c) 503 { 504 if (zio->io_child_error[c] != 0 && zio->io_error == 0) 505 zio->io_error = zio->io_child_error[c]; 506 } 507 508 int 509 zio_bookmark_compare(const void *x1, const void *x2) 510 { 511 const zio_t *z1 = x1; 512 const zio_t *z2 = x2; 513 514 if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset) 515 return (-1); 516 if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset) 517 return (1); 518 519 if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object) 520 return (-1); 521 if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object) 522 return (1); 523 524 if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level) 525 return (-1); 526 if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level) 527 return (1); 528 529 if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid) 530 return (-1); 531 if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid) 532 return (1); 533 534 if (z1 < z2) 535 return (-1); 536 if (z1 > z2) 537 return (1); 538 539 return (0); 540 } 541 542 /* 543 * ========================================================================== 544 * Create the various types of I/O (read, write, free, etc) 545 * ========================================================================== 546 */ 547 static zio_t * 548 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 549 abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done, 550 void *private, zio_type_t type, zio_priority_t priority, 551 enum zio_flag flags, vdev_t *vd, uint64_t offset, 552 const zbookmark_phys_t *zb, enum zio_stage stage, enum zio_stage pipeline) 553 { 554 zio_t *zio; 555 556 ASSERT3U(psize, <=, SPA_MAXBLOCKSIZE); 557 ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0); 558 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0); 559 560 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER)); 561 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER)); 562 ASSERT(vd || stage == ZIO_STAGE_OPEN); 563 564 IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW) != 0); 565 566 zio = kmem_cache_alloc(zio_cache, KM_SLEEP); 567 bzero(zio, sizeof (zio_t)); 568 569 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL); 570 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL); 571 572 list_create(&zio->io_parent_list, sizeof (zio_link_t), 573 offsetof(zio_link_t, zl_parent_node)); 574 list_create(&zio->io_child_list, sizeof (zio_link_t), 575 offsetof(zio_link_t, zl_child_node)); 576 metaslab_trace_init(&zio->io_alloc_list); 577 578 if (vd != NULL) 579 zio->io_child_type = ZIO_CHILD_VDEV; 580 else if (flags & ZIO_FLAG_GANG_CHILD) 581 zio->io_child_type = ZIO_CHILD_GANG; 582 else if (flags & ZIO_FLAG_DDT_CHILD) 583 zio->io_child_type = ZIO_CHILD_DDT; 584 else 585 zio->io_child_type = ZIO_CHILD_LOGICAL; 586 587 if (bp != NULL) { 588 zio->io_bp = (blkptr_t *)bp; 589 zio->io_bp_copy = *bp; 590 zio->io_bp_orig = *bp; 591 if (type != ZIO_TYPE_WRITE || 592 zio->io_child_type == ZIO_CHILD_DDT) 593 zio->io_bp = &zio->io_bp_copy; /* so caller can free */ 594 if (zio->io_child_type == ZIO_CHILD_LOGICAL) 595 zio->io_logical = zio; 596 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp)) 597 pipeline |= ZIO_GANG_STAGES; 598 } 599 600 zio->io_spa = spa; 601 zio->io_txg = txg; 602 zio->io_done = done; 603 zio->io_private = private; 604 zio->io_type = type; 605 zio->io_priority = priority; 606 zio->io_vd = vd; 607 zio->io_offset = offset; 608 zio->io_orig_abd = zio->io_abd = data; 609 zio->io_orig_size = zio->io_size = psize; 610 zio->io_lsize = lsize; 611 zio->io_orig_flags = zio->io_flags = flags; 612 zio->io_orig_stage = zio->io_stage = stage; 613 zio->io_orig_pipeline = zio->io_pipeline = pipeline; 614 zio->io_pipeline_trace = ZIO_STAGE_OPEN; 615 616 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY); 617 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE); 618 619 if (zb != NULL) 620 zio->io_bookmark = *zb; 621 622 if (pio != NULL) { 623 if (zio->io_logical == NULL) 624 zio->io_logical = pio->io_logical; 625 if (zio->io_child_type == ZIO_CHILD_GANG) 626 zio->io_gang_leader = pio->io_gang_leader; 627 zio_add_child(pio, zio); 628 } 629 630 return (zio); 631 } 632 633 static void 634 zio_destroy(zio_t *zio) 635 { 636 metaslab_trace_fini(&zio->io_alloc_list); 637 list_destroy(&zio->io_parent_list); 638 list_destroy(&zio->io_child_list); 639 mutex_destroy(&zio->io_lock); 640 cv_destroy(&zio->io_cv); 641 kmem_cache_free(zio_cache, zio); 642 } 643 644 zio_t * 645 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done, 646 void *private, enum zio_flag flags) 647 { 648 zio_t *zio; 649 650 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private, 651 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, 652 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE); 653 654 return (zio); 655 } 656 657 zio_t * 658 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags) 659 { 660 return (zio_null(NULL, spa, NULL, done, private, flags)); 661 } 662 663 void 664 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp) 665 { 666 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) { 667 zfs_panic_recover("blkptr at %p has invalid TYPE %llu", 668 bp, (longlong_t)BP_GET_TYPE(bp)); 669 } 670 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS || 671 BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) { 672 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu", 673 bp, (longlong_t)BP_GET_CHECKSUM(bp)); 674 } 675 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS || 676 BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) { 677 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu", 678 bp, (longlong_t)BP_GET_COMPRESS(bp)); 679 } 680 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) { 681 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu", 682 bp, (longlong_t)BP_GET_LSIZE(bp)); 683 } 684 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) { 685 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu", 686 bp, (longlong_t)BP_GET_PSIZE(bp)); 687 } 688 689 if (BP_IS_EMBEDDED(bp)) { 690 if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) { 691 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu", 692 bp, (longlong_t)BPE_GET_ETYPE(bp)); 693 } 694 } 695 696 /* 697 * Do not verify individual DVAs if the config is not trusted. This 698 * will be done once the zio is executed in vdev_mirror_map_alloc. 699 */ 700 if (!spa->spa_trust_config) 701 return; 702 703 /* 704 * Pool-specific checks. 705 * 706 * Note: it would be nice to verify that the blk_birth and 707 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze() 708 * allows the birth time of log blocks (and dmu_sync()-ed blocks 709 * that are in the log) to be arbitrarily large. 710 */ 711 for (int i = 0; i < BP_GET_NDVAS(bp); i++) { 712 uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]); 713 if (vdevid >= spa->spa_root_vdev->vdev_children) { 714 zfs_panic_recover("blkptr at %p DVA %u has invalid " 715 "VDEV %llu", 716 bp, i, (longlong_t)vdevid); 717 continue; 718 } 719 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid]; 720 if (vd == NULL) { 721 zfs_panic_recover("blkptr at %p DVA %u has invalid " 722 "VDEV %llu", 723 bp, i, (longlong_t)vdevid); 724 continue; 725 } 726 if (vd->vdev_ops == &vdev_hole_ops) { 727 zfs_panic_recover("blkptr at %p DVA %u has hole " 728 "VDEV %llu", 729 bp, i, (longlong_t)vdevid); 730 continue; 731 } 732 if (vd->vdev_ops == &vdev_missing_ops) { 733 /* 734 * "missing" vdevs are valid during import, but we 735 * don't have their detailed info (e.g. asize), so 736 * we can't perform any more checks on them. 737 */ 738 continue; 739 } 740 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]); 741 uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]); 742 if (BP_IS_GANG(bp)) 743 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE); 744 if (offset + asize > vd->vdev_asize) { 745 zfs_panic_recover("blkptr at %p DVA %u has invalid " 746 "OFFSET %llu", 747 bp, i, (longlong_t)offset); 748 } 749 } 750 } 751 752 boolean_t 753 zfs_dva_valid(spa_t *spa, const dva_t *dva, const blkptr_t *bp) 754 { 755 uint64_t vdevid = DVA_GET_VDEV(dva); 756 757 if (vdevid >= spa->spa_root_vdev->vdev_children) 758 return (B_FALSE); 759 760 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid]; 761 if (vd == NULL) 762 return (B_FALSE); 763 764 if (vd->vdev_ops == &vdev_hole_ops) 765 return (B_FALSE); 766 767 if (vd->vdev_ops == &vdev_missing_ops) { 768 return (B_FALSE); 769 } 770 771 uint64_t offset = DVA_GET_OFFSET(dva); 772 uint64_t asize = DVA_GET_ASIZE(dva); 773 774 if (BP_IS_GANG(bp)) 775 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE); 776 if (offset + asize > vd->vdev_asize) 777 return (B_FALSE); 778 779 return (B_TRUE); 780 } 781 782 zio_t * 783 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, 784 abd_t *data, uint64_t size, zio_done_func_t *done, void *private, 785 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb) 786 { 787 zio_t *zio; 788 789 zfs_blkptr_verify(spa, bp); 790 791 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp, 792 data, size, size, done, private, 793 ZIO_TYPE_READ, priority, flags, NULL, 0, zb, 794 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? 795 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE); 796 797 return (zio); 798 } 799 800 zio_t * 801 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, 802 abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp, 803 zio_done_func_t *ready, zio_done_func_t *children_ready, 804 zio_done_func_t *physdone, zio_done_func_t *done, 805 void *private, zio_priority_t priority, enum zio_flag flags, 806 const zbookmark_phys_t *zb) 807 { 808 zio_t *zio; 809 810 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF && 811 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS && 812 zp->zp_compress >= ZIO_COMPRESS_OFF && 813 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS && 814 DMU_OT_IS_VALID(zp->zp_type) && 815 zp->zp_level < 32 && 816 zp->zp_copies > 0 && 817 zp->zp_copies <= spa_max_replication(spa)); 818 819 zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private, 820 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, 821 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? 822 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE); 823 824 zio->io_ready = ready; 825 zio->io_children_ready = children_ready; 826 zio->io_physdone = physdone; 827 zio->io_prop = *zp; 828 829 /* 830 * Data can be NULL if we are going to call zio_write_override() to 831 * provide the already-allocated BP. But we may need the data to 832 * verify a dedup hit (if requested). In this case, don't try to 833 * dedup (just take the already-allocated BP verbatim). 834 */ 835 if (data == NULL && zio->io_prop.zp_dedup_verify) { 836 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE; 837 } 838 839 return (zio); 840 } 841 842 zio_t * 843 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data, 844 uint64_t size, zio_done_func_t *done, void *private, 845 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb) 846 { 847 zio_t *zio; 848 849 zio = zio_create(pio, spa, txg, bp, data, size, size, done, private, 850 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb, 851 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE); 852 853 return (zio); 854 } 855 856 void 857 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite) 858 { 859 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 860 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 861 ASSERT(zio->io_stage == ZIO_STAGE_OPEN); 862 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa)); 863 864 /* 865 * We must reset the io_prop to match the values that existed 866 * when the bp was first written by dmu_sync() keeping in mind 867 * that nopwrite and dedup are mutually exclusive. 868 */ 869 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup; 870 zio->io_prop.zp_nopwrite = nopwrite; 871 zio->io_prop.zp_copies = copies; 872 zio->io_bp_override = bp; 873 } 874 875 void 876 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp) 877 { 878 879 zfs_blkptr_verify(spa, bp); 880 881 /* 882 * The check for EMBEDDED is a performance optimization. We 883 * process the free here (by ignoring it) rather than 884 * putting it on the list and then processing it in zio_free_sync(). 885 */ 886 if (BP_IS_EMBEDDED(bp)) 887 return; 888 metaslab_check_free(spa, bp); 889 890 /* 891 * Frees that are for the currently-syncing txg, are not going to be 892 * deferred, and which will not need to do a read (i.e. not GANG or 893 * DEDUP), can be processed immediately. Otherwise, put them on the 894 * in-memory list for later processing. 895 */ 896 if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp) || 897 txg != spa->spa_syncing_txg || 898 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) { 899 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp); 900 } else { 901 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 0))); 902 } 903 } 904 905 zio_t * 906 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 907 enum zio_flag flags) 908 { 909 zio_t *zio; 910 enum zio_stage stage = ZIO_FREE_PIPELINE; 911 912 ASSERT(!BP_IS_HOLE(bp)); 913 ASSERT(spa_syncing_txg(spa) == txg); 914 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free); 915 916 if (BP_IS_EMBEDDED(bp)) 917 return (zio_null(pio, spa, NULL, NULL, NULL, 0)); 918 919 metaslab_check_free(spa, bp); 920 arc_freed(spa, bp); 921 922 /* 923 * GANG and DEDUP blocks can induce a read (for the gang block header, 924 * or the DDT), so issue them asynchronously so that this thread is 925 * not tied up. 926 */ 927 if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp)) 928 stage |= ZIO_STAGE_ISSUE_ASYNC; 929 930 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), 931 BP_GET_PSIZE(bp), NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, 932 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage); 933 934 return (zio); 935 } 936 937 zio_t * 938 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 939 zio_done_func_t *done, void *private, enum zio_flag flags) 940 { 941 zio_t *zio; 942 943 zfs_blkptr_verify(spa, bp); 944 945 if (BP_IS_EMBEDDED(bp)) 946 return (zio_null(pio, spa, NULL, NULL, NULL, 0)); 947 948 /* 949 * A claim is an allocation of a specific block. Claims are needed 950 * to support immediate writes in the intent log. The issue is that 951 * immediate writes contain committed data, but in a txg that was 952 * *not* committed. Upon opening the pool after an unclean shutdown, 953 * the intent log claims all blocks that contain immediate write data 954 * so that the SPA knows they're in use. 955 * 956 * All claims *must* be resolved in the first txg -- before the SPA 957 * starts allocating blocks -- so that nothing is allocated twice. 958 * If txg == 0 we just verify that the block is claimable. 959 */ 960 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, 961 spa_min_claim_txg(spa)); 962 ASSERT(txg == spa_min_claim_txg(spa) || txg == 0); 963 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */ 964 965 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), 966 BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, 967 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE); 968 ASSERT0(zio->io_queued_timestamp); 969 970 return (zio); 971 } 972 973 zio_t * 974 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, 975 zio_done_func_t *done, void *private, enum zio_flag flags) 976 { 977 zio_t *zio; 978 int c; 979 980 if (vd->vdev_children == 0) { 981 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private, 982 ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, 983 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE); 984 985 zio->io_cmd = cmd; 986 } else { 987 zio = zio_null(pio, spa, NULL, NULL, NULL, flags); 988 989 for (c = 0; c < vd->vdev_children; c++) 990 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd, 991 done, private, flags)); 992 } 993 994 return (zio); 995 } 996 997 zio_t * 998 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 999 abd_t *data, int checksum, zio_done_func_t *done, void *private, 1000 zio_priority_t priority, enum zio_flag flags, boolean_t labels) 1001 { 1002 zio_t *zio; 1003 1004 ASSERT(vd->vdev_children == 0); 1005 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || 1006 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); 1007 ASSERT3U(offset + size, <=, vd->vdev_psize); 1008 1009 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done, 1010 private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, 1011 offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE); 1012 1013 zio->io_prop.zp_checksum = checksum; 1014 1015 return (zio); 1016 } 1017 1018 zio_t * 1019 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 1020 abd_t *data, int checksum, zio_done_func_t *done, void *private, 1021 zio_priority_t priority, enum zio_flag flags, boolean_t labels) 1022 { 1023 zio_t *zio; 1024 1025 ASSERT(vd->vdev_children == 0); 1026 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || 1027 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); 1028 ASSERT3U(offset + size, <=, vd->vdev_psize); 1029 1030 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done, 1031 private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, 1032 offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE); 1033 1034 zio->io_prop.zp_checksum = checksum; 1035 1036 if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) { 1037 /* 1038 * zec checksums are necessarily destructive -- they modify 1039 * the end of the write buffer to hold the verifier/checksum. 1040 * Therefore, we must make a local copy in case the data is 1041 * being written to multiple places in parallel. 1042 */ 1043 abd_t *wbuf = abd_alloc_sametype(data, size); 1044 abd_copy(wbuf, data, size); 1045 1046 zio_push_transform(zio, wbuf, size, size, NULL); 1047 } 1048 1049 return (zio); 1050 } 1051 1052 /* 1053 * Create a child I/O to do some work for us. 1054 */ 1055 zio_t * 1056 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset, 1057 abd_t *data, uint64_t size, int type, zio_priority_t priority, 1058 enum zio_flag flags, zio_done_func_t *done, void *private) 1059 { 1060 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE; 1061 zio_t *zio; 1062 1063 /* 1064 * vdev child I/Os do not propagate their error to the parent. 1065 * Therefore, for correct operation the caller *must* check for 1066 * and handle the error in the child i/o's done callback. 1067 * The only exceptions are i/os that we don't care about 1068 * (OPTIONAL or REPAIR). 1069 */ 1070 ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) || 1071 done != NULL); 1072 1073 if (type == ZIO_TYPE_READ && bp != NULL) { 1074 /* 1075 * If we have the bp, then the child should perform the 1076 * checksum and the parent need not. This pushes error 1077 * detection as close to the leaves as possible and 1078 * eliminates redundant checksums in the interior nodes. 1079 */ 1080 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY; 1081 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; 1082 } 1083 1084 if (vd->vdev_ops->vdev_op_leaf) { 1085 ASSERT0(vd->vdev_children); 1086 offset += VDEV_LABEL_START_SIZE; 1087 } 1088 1089 flags |= ZIO_VDEV_CHILD_FLAGS(pio); 1090 1091 /* 1092 * If we've decided to do a repair, the write is not speculative -- 1093 * even if the original read was. 1094 */ 1095 if (flags & ZIO_FLAG_IO_REPAIR) 1096 flags &= ~ZIO_FLAG_SPECULATIVE; 1097 1098 /* 1099 * If we're creating a child I/O that is not associated with a 1100 * top-level vdev, then the child zio is not an allocating I/O. 1101 * If this is a retried I/O then we ignore it since we will 1102 * have already processed the original allocating I/O. 1103 */ 1104 if (flags & ZIO_FLAG_IO_ALLOCATING && 1105 (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) { 1106 metaslab_class_t *mc = spa_normal_class(pio->io_spa); 1107 1108 ASSERT(mc->mc_alloc_throttle_enabled); 1109 ASSERT(type == ZIO_TYPE_WRITE); 1110 ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE); 1111 ASSERT(!(flags & ZIO_FLAG_IO_REPAIR)); 1112 ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) || 1113 pio->io_child_type == ZIO_CHILD_GANG); 1114 1115 flags &= ~ZIO_FLAG_IO_ALLOCATING; 1116 } 1117 1118 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size, 1119 done, private, type, priority, flags, vd, offset, &pio->io_bookmark, 1120 ZIO_STAGE_VDEV_IO_START >> 1, pipeline); 1121 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV); 1122 1123 zio->io_physdone = pio->io_physdone; 1124 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL) 1125 zio->io_logical->io_phys_children++; 1126 1127 return (zio); 1128 } 1129 1130 zio_t * 1131 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size, 1132 zio_type_t type, zio_priority_t priority, enum zio_flag flags, 1133 zio_done_func_t *done, void *private) 1134 { 1135 zio_t *zio; 1136 1137 ASSERT(vd->vdev_ops->vdev_op_leaf); 1138 1139 zio = zio_create(NULL, vd->vdev_spa, 0, NULL, 1140 data, size, size, done, private, type, priority, 1141 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED, 1142 vd, offset, NULL, 1143 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE); 1144 1145 return (zio); 1146 } 1147 1148 void 1149 zio_flush(zio_t *zio, vdev_t *vd) 1150 { 1151 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 1152 NULL, NULL, 1153 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY)); 1154 } 1155 1156 void 1157 zio_shrink(zio_t *zio, uint64_t size) 1158 { 1159 ASSERT3P(zio->io_executor, ==, NULL); 1160 ASSERT3P(zio->io_orig_size, ==, zio->io_size); 1161 ASSERT3U(size, <=, zio->io_size); 1162 1163 /* 1164 * We don't shrink for raidz because of problems with the 1165 * reconstruction when reading back less than the block size. 1166 * Note, BP_IS_RAIDZ() assumes no compression. 1167 */ 1168 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 1169 if (!BP_IS_RAIDZ(zio->io_bp)) { 1170 /* we are not doing a raw write */ 1171 ASSERT3U(zio->io_size, ==, zio->io_lsize); 1172 zio->io_orig_size = zio->io_size = zio->io_lsize = size; 1173 } 1174 } 1175 1176 /* 1177 * ========================================================================== 1178 * Prepare to read and write logical blocks 1179 * ========================================================================== 1180 */ 1181 1182 static int 1183 zio_read_bp_init(zio_t *zio) 1184 { 1185 blkptr_t *bp = zio->io_bp; 1186 1187 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy); 1188 1189 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF && 1190 zio->io_child_type == ZIO_CHILD_LOGICAL && 1191 !(zio->io_flags & ZIO_FLAG_RAW)) { 1192 uint64_t psize = 1193 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp); 1194 zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize), 1195 psize, psize, zio_decompress); 1196 } 1197 1198 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) { 1199 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1200 1201 int psize = BPE_GET_PSIZE(bp); 1202 void *data = abd_borrow_buf(zio->io_abd, psize); 1203 decode_embedded_bp_compressed(bp, data); 1204 abd_return_buf_copy(zio->io_abd, data, psize); 1205 } else { 1206 ASSERT(!BP_IS_EMBEDDED(bp)); 1207 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy); 1208 } 1209 1210 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0) 1211 zio->io_flags |= ZIO_FLAG_DONT_CACHE; 1212 1213 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP) 1214 zio->io_flags |= ZIO_FLAG_DONT_CACHE; 1215 1216 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL) 1217 zio->io_pipeline = ZIO_DDT_READ_PIPELINE; 1218 1219 return (ZIO_PIPELINE_CONTINUE); 1220 } 1221 1222 static int 1223 zio_write_bp_init(zio_t *zio) 1224 { 1225 if (!IO_IS_ALLOCATING(zio)) 1226 return (ZIO_PIPELINE_CONTINUE); 1227 1228 ASSERT(zio->io_child_type != ZIO_CHILD_DDT); 1229 1230 if (zio->io_bp_override) { 1231 blkptr_t *bp = zio->io_bp; 1232 zio_prop_t *zp = &zio->io_prop; 1233 1234 ASSERT(bp->blk_birth != zio->io_txg); 1235 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0); 1236 1237 *bp = *zio->io_bp_override; 1238 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1239 1240 if (BP_IS_EMBEDDED(bp)) 1241 return (ZIO_PIPELINE_CONTINUE); 1242 1243 /* 1244 * If we've been overridden and nopwrite is set then 1245 * set the flag accordingly to indicate that a nopwrite 1246 * has already occurred. 1247 */ 1248 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) { 1249 ASSERT(!zp->zp_dedup); 1250 ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum); 1251 zio->io_flags |= ZIO_FLAG_NOPWRITE; 1252 return (ZIO_PIPELINE_CONTINUE); 1253 } 1254 1255 ASSERT(!zp->zp_nopwrite); 1256 1257 if (BP_IS_HOLE(bp) || !zp->zp_dedup) 1258 return (ZIO_PIPELINE_CONTINUE); 1259 1260 ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags & 1261 ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify); 1262 1263 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) { 1264 BP_SET_DEDUP(bp, 1); 1265 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE; 1266 return (ZIO_PIPELINE_CONTINUE); 1267 } 1268 1269 /* 1270 * We were unable to handle this as an override bp, treat 1271 * it as a regular write I/O. 1272 */ 1273 zio->io_bp_override = NULL; 1274 *bp = zio->io_bp_orig; 1275 zio->io_pipeline = zio->io_orig_pipeline; 1276 } 1277 1278 return (ZIO_PIPELINE_CONTINUE); 1279 } 1280 1281 static int 1282 zio_write_compress(zio_t *zio) 1283 { 1284 spa_t *spa = zio->io_spa; 1285 zio_prop_t *zp = &zio->io_prop; 1286 enum zio_compress compress = zp->zp_compress; 1287 blkptr_t *bp = zio->io_bp; 1288 uint64_t lsize = zio->io_lsize; 1289 uint64_t psize = zio->io_size; 1290 int pass = 1; 1291 1292 EQUIV(lsize != psize, (zio->io_flags & ZIO_FLAG_RAW) != 0); 1293 1294 /* 1295 * If our children haven't all reached the ready stage, 1296 * wait for them and then repeat this pipeline stage. 1297 */ 1298 if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT | 1299 ZIO_CHILD_GANG_BIT, ZIO_WAIT_READY)) { 1300 return (ZIO_PIPELINE_STOP); 1301 } 1302 1303 if (!IO_IS_ALLOCATING(zio)) 1304 return (ZIO_PIPELINE_CONTINUE); 1305 1306 if (zio->io_children_ready != NULL) { 1307 /* 1308 * Now that all our children are ready, run the callback 1309 * associated with this zio in case it wants to modify the 1310 * data to be written. 1311 */ 1312 ASSERT3U(zp->zp_level, >, 0); 1313 zio->io_children_ready(zio); 1314 } 1315 1316 ASSERT(zio->io_child_type != ZIO_CHILD_DDT); 1317 ASSERT(zio->io_bp_override == NULL); 1318 1319 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) { 1320 /* 1321 * We're rewriting an existing block, which means we're 1322 * working on behalf of spa_sync(). For spa_sync() to 1323 * converge, it must eventually be the case that we don't 1324 * have to allocate new blocks. But compression changes 1325 * the blocksize, which forces a reallocate, and makes 1326 * convergence take longer. Therefore, after the first 1327 * few passes, stop compressing to ensure convergence. 1328 */ 1329 pass = spa_sync_pass(spa); 1330 1331 ASSERT(zio->io_txg == spa_syncing_txg(spa)); 1332 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1333 ASSERT(!BP_GET_DEDUP(bp)); 1334 1335 if (pass >= zfs_sync_pass_dont_compress) 1336 compress = ZIO_COMPRESS_OFF; 1337 1338 /* Make sure someone doesn't change their mind on overwrites */ 1339 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp), 1340 spa_max_replication(spa)) == BP_GET_NDVAS(bp)); 1341 } 1342 1343 /* If it's a compressed write that is not raw, compress the buffer. */ 1344 if (compress != ZIO_COMPRESS_OFF && psize == lsize) { 1345 void *cbuf = zio_buf_alloc(lsize); 1346 psize = zio_compress_data(compress, zio->io_abd, cbuf, lsize); 1347 if (psize == 0 || psize == lsize) { 1348 compress = ZIO_COMPRESS_OFF; 1349 zio_buf_free(cbuf, lsize); 1350 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE && 1351 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) && 1352 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) { 1353 encode_embedded_bp_compressed(bp, 1354 cbuf, compress, lsize, psize); 1355 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA); 1356 BP_SET_TYPE(bp, zio->io_prop.zp_type); 1357 BP_SET_LEVEL(bp, zio->io_prop.zp_level); 1358 zio_buf_free(cbuf, lsize); 1359 bp->blk_birth = zio->io_txg; 1360 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1361 ASSERT(spa_feature_is_active(spa, 1362 SPA_FEATURE_EMBEDDED_DATA)); 1363 return (ZIO_PIPELINE_CONTINUE); 1364 } else { 1365 /* 1366 * Round up compressed size up to the ashift 1367 * of the smallest-ashift device, and zero the tail. 1368 * This ensures that the compressed size of the BP 1369 * (and thus compressratio property) are correct, 1370 * in that we charge for the padding used to fill out 1371 * the last sector. 1372 */ 1373 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT); 1374 size_t rounded = (size_t)P2ROUNDUP(psize, 1375 1ULL << spa->spa_min_ashift); 1376 if (rounded >= lsize) { 1377 compress = ZIO_COMPRESS_OFF; 1378 zio_buf_free(cbuf, lsize); 1379 psize = lsize; 1380 } else { 1381 abd_t *cdata = abd_get_from_buf(cbuf, lsize); 1382 abd_take_ownership_of_buf(cdata, B_TRUE); 1383 abd_zero_off(cdata, psize, rounded - psize); 1384 psize = rounded; 1385 zio_push_transform(zio, cdata, 1386 psize, lsize, NULL); 1387 } 1388 } 1389 1390 /* 1391 * We were unable to handle this as an override bp, treat 1392 * it as a regular write I/O. 1393 */ 1394 zio->io_bp_override = NULL; 1395 *bp = zio->io_bp_orig; 1396 zio->io_pipeline = zio->io_orig_pipeline; 1397 } else { 1398 ASSERT3U(psize, !=, 0); 1399 } 1400 1401 /* 1402 * The final pass of spa_sync() must be all rewrites, but the first 1403 * few passes offer a trade-off: allocating blocks defers convergence, 1404 * but newly allocated blocks are sequential, so they can be written 1405 * to disk faster. Therefore, we allow the first few passes of 1406 * spa_sync() to allocate new blocks, but force rewrites after that. 1407 * There should only be a handful of blocks after pass 1 in any case. 1408 */ 1409 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg && 1410 BP_GET_PSIZE(bp) == psize && 1411 pass >= zfs_sync_pass_rewrite) { 1412 ASSERT(psize != 0); 1413 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES; 1414 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages; 1415 zio->io_flags |= ZIO_FLAG_IO_REWRITE; 1416 } else { 1417 BP_ZERO(bp); 1418 zio->io_pipeline = ZIO_WRITE_PIPELINE; 1419 } 1420 1421 if (psize == 0) { 1422 if (zio->io_bp_orig.blk_birth != 0 && 1423 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) { 1424 BP_SET_LSIZE(bp, lsize); 1425 BP_SET_TYPE(bp, zp->zp_type); 1426 BP_SET_LEVEL(bp, zp->zp_level); 1427 BP_SET_BIRTH(bp, zio->io_txg, 0); 1428 } 1429 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1430 } else { 1431 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER); 1432 BP_SET_LSIZE(bp, lsize); 1433 BP_SET_TYPE(bp, zp->zp_type); 1434 BP_SET_LEVEL(bp, zp->zp_level); 1435 BP_SET_PSIZE(bp, psize); 1436 BP_SET_COMPRESS(bp, compress); 1437 BP_SET_CHECKSUM(bp, zp->zp_checksum); 1438 BP_SET_DEDUP(bp, zp->zp_dedup); 1439 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER); 1440 if (zp->zp_dedup) { 1441 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1442 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 1443 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE; 1444 } 1445 if (zp->zp_nopwrite) { 1446 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1447 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 1448 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE; 1449 } 1450 } 1451 return (ZIO_PIPELINE_CONTINUE); 1452 } 1453 1454 static int 1455 zio_free_bp_init(zio_t *zio) 1456 { 1457 blkptr_t *bp = zio->io_bp; 1458 1459 if (zio->io_child_type == ZIO_CHILD_LOGICAL) { 1460 if (BP_GET_DEDUP(bp)) 1461 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE; 1462 } 1463 1464 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy); 1465 1466 return (ZIO_PIPELINE_CONTINUE); 1467 } 1468 1469 /* 1470 * ========================================================================== 1471 * Execute the I/O pipeline 1472 * ========================================================================== 1473 */ 1474 1475 static void 1476 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline) 1477 { 1478 spa_t *spa = zio->io_spa; 1479 zio_type_t t = zio->io_type; 1480 int flags = (cutinline ? TQ_FRONT : 0); 1481 1482 /* 1483 * If we're a config writer or a probe, the normal issue and 1484 * interrupt threads may all be blocked waiting for the config lock. 1485 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL. 1486 */ 1487 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE)) 1488 t = ZIO_TYPE_NULL; 1489 1490 /* 1491 * A similar issue exists for the L2ARC write thread until L2ARC 2.0. 1492 */ 1493 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux) 1494 t = ZIO_TYPE_NULL; 1495 1496 /* 1497 * If this is a high priority I/O, then use the high priority taskq if 1498 * available. 1499 */ 1500 if (zio->io_priority == ZIO_PRIORITY_NOW && 1501 spa->spa_zio_taskq[t][q + 1].stqs_count != 0) 1502 q++; 1503 1504 ASSERT3U(q, <, ZIO_TASKQ_TYPES); 1505 1506 /* 1507 * NB: We are assuming that the zio can only be dispatched 1508 * to a single taskq at a time. It would be a grievous error 1509 * to dispatch the zio to another taskq at the same time. 1510 */ 1511 ASSERT(zio->io_tqent.tqent_next == NULL); 1512 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio, 1513 flags, &zio->io_tqent); 1514 } 1515 1516 static boolean_t 1517 zio_taskq_member(zio_t *zio, zio_taskq_type_t q) 1518 { 1519 kthread_t *executor = zio->io_executor; 1520 spa_t *spa = zio->io_spa; 1521 1522 for (zio_type_t t = 0; t < ZIO_TYPES; t++) { 1523 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q]; 1524 uint_t i; 1525 for (i = 0; i < tqs->stqs_count; i++) { 1526 if (taskq_member(tqs->stqs_taskq[i], executor)) 1527 return (B_TRUE); 1528 } 1529 } 1530 1531 return (B_FALSE); 1532 } 1533 1534 static int 1535 zio_issue_async(zio_t *zio) 1536 { 1537 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); 1538 1539 return (ZIO_PIPELINE_STOP); 1540 } 1541 1542 void 1543 zio_interrupt(zio_t *zio) 1544 { 1545 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE); 1546 } 1547 1548 void 1549 zio_delay_interrupt(zio_t *zio) 1550 { 1551 /* 1552 * The timeout_generic() function isn't defined in userspace, so 1553 * rather than trying to implement the function, the zio delay 1554 * functionality has been disabled for userspace builds. 1555 */ 1556 1557 #ifdef _KERNEL 1558 /* 1559 * If io_target_timestamp is zero, then no delay has been registered 1560 * for this IO, thus jump to the end of this function and "skip" the 1561 * delay; issuing it directly to the zio layer. 1562 */ 1563 if (zio->io_target_timestamp != 0) { 1564 hrtime_t now = gethrtime(); 1565 1566 if (now >= zio->io_target_timestamp) { 1567 /* 1568 * This IO has already taken longer than the target 1569 * delay to complete, so we don't want to delay it 1570 * any longer; we "miss" the delay and issue it 1571 * directly to the zio layer. This is likely due to 1572 * the target latency being set to a value less than 1573 * the underlying hardware can satisfy (e.g. delay 1574 * set to 1ms, but the disks take 10ms to complete an 1575 * IO request). 1576 */ 1577 1578 DTRACE_PROBE2(zio__delay__miss, zio_t *, zio, 1579 hrtime_t, now); 1580 1581 zio_interrupt(zio); 1582 } else { 1583 hrtime_t diff = zio->io_target_timestamp - now; 1584 1585 DTRACE_PROBE3(zio__delay__hit, zio_t *, zio, 1586 hrtime_t, now, hrtime_t, diff); 1587 1588 (void) timeout_generic(CALLOUT_NORMAL, 1589 (void (*)(void *))zio_interrupt, zio, diff, 1, 0); 1590 } 1591 1592 return; 1593 } 1594 #endif 1595 1596 DTRACE_PROBE1(zio__delay__skip, zio_t *, zio); 1597 zio_interrupt(zio); 1598 } 1599 1600 /* 1601 * Execute the I/O pipeline until one of the following occurs: 1602 * 1603 * (1) the I/O completes 1604 * (2) the pipeline stalls waiting for dependent child I/Os 1605 * (3) the I/O issues, so we're waiting for an I/O completion interrupt 1606 * (4) the I/O is delegated by vdev-level caching or aggregation 1607 * (5) the I/O is deferred due to vdev-level queueing 1608 * (6) the I/O is handed off to another thread. 1609 * 1610 * In all cases, the pipeline stops whenever there's no CPU work; it never 1611 * burns a thread in cv_wait(). 1612 * 1613 * There's no locking on io_stage because there's no legitimate way 1614 * for multiple threads to be attempting to process the same I/O. 1615 */ 1616 static zio_pipe_stage_t *zio_pipeline[]; 1617 1618 void 1619 zio_execute(zio_t *zio) 1620 { 1621 zio->io_executor = curthread; 1622 1623 ASSERT3U(zio->io_queued_timestamp, >, 0); 1624 1625 while (zio->io_stage < ZIO_STAGE_DONE) { 1626 enum zio_stage pipeline = zio->io_pipeline; 1627 enum zio_stage stage = zio->io_stage; 1628 int rv; 1629 1630 ASSERT(!MUTEX_HELD(&zio->io_lock)); 1631 ASSERT(ISP2(stage)); 1632 ASSERT(zio->io_stall == NULL); 1633 1634 do { 1635 stage <<= 1; 1636 } while ((stage & pipeline) == 0); 1637 1638 ASSERT(stage <= ZIO_STAGE_DONE); 1639 1640 /* 1641 * If we are in interrupt context and this pipeline stage 1642 * will grab a config lock that is held across I/O, 1643 * or may wait for an I/O that needs an interrupt thread 1644 * to complete, issue async to avoid deadlock. 1645 * 1646 * For VDEV_IO_START, we cut in line so that the io will 1647 * be sent to disk promptly. 1648 */ 1649 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL && 1650 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) { 1651 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? 1652 zio_requeue_io_start_cut_in_line : B_FALSE; 1653 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); 1654 return; 1655 } 1656 1657 zio->io_stage = stage; 1658 zio->io_pipeline_trace |= zio->io_stage; 1659 rv = zio_pipeline[highbit64(stage) - 1](zio); 1660 1661 if (rv == ZIO_PIPELINE_STOP) 1662 return; 1663 1664 ASSERT(rv == ZIO_PIPELINE_CONTINUE); 1665 } 1666 } 1667 1668 /* 1669 * ========================================================================== 1670 * Initiate I/O, either sync or async 1671 * ========================================================================== 1672 */ 1673 int 1674 zio_wait(zio_t *zio) 1675 { 1676 int error; 1677 1678 ASSERT3P(zio->io_stage, ==, ZIO_STAGE_OPEN); 1679 ASSERT3P(zio->io_executor, ==, NULL); 1680 1681 zio->io_waiter = curthread; 1682 ASSERT0(zio->io_queued_timestamp); 1683 zio->io_queued_timestamp = gethrtime(); 1684 1685 zio_execute(zio); 1686 1687 mutex_enter(&zio->io_lock); 1688 while (zio->io_executor != NULL) 1689 cv_wait(&zio->io_cv, &zio->io_lock); 1690 mutex_exit(&zio->io_lock); 1691 1692 error = zio->io_error; 1693 zio_destroy(zio); 1694 1695 return (error); 1696 } 1697 1698 void 1699 zio_nowait(zio_t *zio) 1700 { 1701 ASSERT3P(zio->io_executor, ==, NULL); 1702 1703 if (zio->io_child_type == ZIO_CHILD_LOGICAL && 1704 zio_unique_parent(zio) == NULL) { 1705 /* 1706 * This is a logical async I/O with no parent to wait for it. 1707 * We add it to the spa_async_root_zio "Godfather" I/O which 1708 * will ensure they complete prior to unloading the pool. 1709 */ 1710 spa_t *spa = zio->io_spa; 1711 1712 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio); 1713 } 1714 1715 ASSERT0(zio->io_queued_timestamp); 1716 zio->io_queued_timestamp = gethrtime(); 1717 zio_execute(zio); 1718 } 1719 1720 /* 1721 * ========================================================================== 1722 * Reexecute, cancel, or suspend/resume failed I/O 1723 * ========================================================================== 1724 */ 1725 1726 static void 1727 zio_reexecute(zio_t *pio) 1728 { 1729 zio_t *cio, *cio_next; 1730 1731 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL); 1732 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN); 1733 ASSERT(pio->io_gang_leader == NULL); 1734 ASSERT(pio->io_gang_tree == NULL); 1735 1736 pio->io_flags = pio->io_orig_flags; 1737 pio->io_stage = pio->io_orig_stage; 1738 pio->io_pipeline = pio->io_orig_pipeline; 1739 pio->io_reexecute = 0; 1740 pio->io_flags |= ZIO_FLAG_REEXECUTED; 1741 pio->io_pipeline_trace = 0; 1742 pio->io_error = 0; 1743 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 1744 pio->io_state[w] = 0; 1745 for (int c = 0; c < ZIO_CHILD_TYPES; c++) 1746 pio->io_child_error[c] = 0; 1747 1748 if (IO_IS_ALLOCATING(pio)) 1749 BP_ZERO(pio->io_bp); 1750 1751 /* 1752 * As we reexecute pio's children, new children could be created. 1753 * New children go to the head of pio's io_child_list, however, 1754 * so we will (correctly) not reexecute them. The key is that 1755 * the remainder of pio's io_child_list, from 'cio_next' onward, 1756 * cannot be affected by any side effects of reexecuting 'cio'. 1757 */ 1758 zio_link_t *zl = NULL; 1759 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) { 1760 cio_next = zio_walk_children(pio, &zl); 1761 mutex_enter(&pio->io_lock); 1762 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 1763 pio->io_children[cio->io_child_type][w]++; 1764 mutex_exit(&pio->io_lock); 1765 zio_reexecute(cio); 1766 } 1767 1768 /* 1769 * Now that all children have been reexecuted, execute the parent. 1770 * We don't reexecute "The Godfather" I/O here as it's the 1771 * responsibility of the caller to wait on it. 1772 */ 1773 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) { 1774 pio->io_queued_timestamp = gethrtime(); 1775 zio_execute(pio); 1776 } 1777 } 1778 1779 void 1780 zio_suspend(spa_t *spa, zio_t *zio) 1781 { 1782 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC) 1783 fm_panic("Pool '%s' has encountered an uncorrectable I/O " 1784 "failure and the failure mode property for this pool " 1785 "is set to panic.", spa_name(spa)); 1786 1787 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0); 1788 1789 mutex_enter(&spa->spa_suspend_lock); 1790 1791 if (spa->spa_suspend_zio_root == NULL) 1792 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL, 1793 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | 1794 ZIO_FLAG_GODFATHER); 1795 1796 spa->spa_suspended = B_TRUE; 1797 1798 if (zio != NULL) { 1799 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); 1800 ASSERT(zio != spa->spa_suspend_zio_root); 1801 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1802 ASSERT(zio_unique_parent(zio) == NULL); 1803 ASSERT(zio->io_stage == ZIO_STAGE_DONE); 1804 zio_add_child(spa->spa_suspend_zio_root, zio); 1805 } 1806 1807 mutex_exit(&spa->spa_suspend_lock); 1808 } 1809 1810 int 1811 zio_resume(spa_t *spa) 1812 { 1813 zio_t *pio; 1814 1815 /* 1816 * Reexecute all previously suspended i/o. 1817 */ 1818 mutex_enter(&spa->spa_suspend_lock); 1819 spa->spa_suspended = B_FALSE; 1820 cv_broadcast(&spa->spa_suspend_cv); 1821 pio = spa->spa_suspend_zio_root; 1822 spa->spa_suspend_zio_root = NULL; 1823 mutex_exit(&spa->spa_suspend_lock); 1824 1825 if (pio == NULL) 1826 return (0); 1827 1828 zio_reexecute(pio); 1829 return (zio_wait(pio)); 1830 } 1831 1832 void 1833 zio_resume_wait(spa_t *spa) 1834 { 1835 mutex_enter(&spa->spa_suspend_lock); 1836 while (spa_suspended(spa)) 1837 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock); 1838 mutex_exit(&spa->spa_suspend_lock); 1839 } 1840 1841 /* 1842 * ========================================================================== 1843 * Gang blocks. 1844 * 1845 * A gang block is a collection of small blocks that looks to the DMU 1846 * like one large block. When zio_dva_allocate() cannot find a block 1847 * of the requested size, due to either severe fragmentation or the pool 1848 * being nearly full, it calls zio_write_gang_block() to construct the 1849 * block from smaller fragments. 1850 * 1851 * A gang block consists of a gang header (zio_gbh_phys_t) and up to 1852 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like 1853 * an indirect block: it's an array of block pointers. It consumes 1854 * only one sector and hence is allocatable regardless of fragmentation. 1855 * The gang header's bps point to its gang members, which hold the data. 1856 * 1857 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg> 1858 * as the verifier to ensure uniqueness of the SHA256 checksum. 1859 * Critically, the gang block bp's blk_cksum is the checksum of the data, 1860 * not the gang header. This ensures that data block signatures (needed for 1861 * deduplication) are independent of how the block is physically stored. 1862 * 1863 * Gang blocks can be nested: a gang member may itself be a gang block. 1864 * Thus every gang block is a tree in which root and all interior nodes are 1865 * gang headers, and the leaves are normal blocks that contain user data. 1866 * The root of the gang tree is called the gang leader. 1867 * 1868 * To perform any operation (read, rewrite, free, claim) on a gang block, 1869 * zio_gang_assemble() first assembles the gang tree (minus data leaves) 1870 * in the io_gang_tree field of the original logical i/o by recursively 1871 * reading the gang leader and all gang headers below it. This yields 1872 * an in-core tree containing the contents of every gang header and the 1873 * bps for every constituent of the gang block. 1874 * 1875 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree 1876 * and invokes a callback on each bp. To free a gang block, zio_gang_issue() 1877 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp. 1878 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim(). 1879 * zio_read_gang() is a wrapper around zio_read() that omits reading gang 1880 * headers, since we already have those in io_gang_tree. zio_rewrite_gang() 1881 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite() 1882 * of the gang header plus zio_checksum_compute() of the data to update the 1883 * gang header's blk_cksum as described above. 1884 * 1885 * The two-phase assemble/issue model solves the problem of partial failure -- 1886 * what if you'd freed part of a gang block but then couldn't read the 1887 * gang header for another part? Assembling the entire gang tree first 1888 * ensures that all the necessary gang header I/O has succeeded before 1889 * starting the actual work of free, claim, or write. Once the gang tree 1890 * is assembled, free and claim are in-memory operations that cannot fail. 1891 * 1892 * In the event that a gang write fails, zio_dva_unallocate() walks the 1893 * gang tree to immediately free (i.e. insert back into the space map) 1894 * everything we've allocated. This ensures that we don't get ENOSPC 1895 * errors during repeated suspend/resume cycles due to a flaky device. 1896 * 1897 * Gang rewrites only happen during sync-to-convergence. If we can't assemble 1898 * the gang tree, we won't modify the block, so we can safely defer the free 1899 * (knowing that the block is still intact). If we *can* assemble the gang 1900 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free 1901 * each constituent bp and we can allocate a new block on the next sync pass. 1902 * 1903 * In all cases, the gang tree allows complete recovery from partial failure. 1904 * ========================================================================== 1905 */ 1906 1907 static void 1908 zio_gang_issue_func_done(zio_t *zio) 1909 { 1910 abd_put(zio->io_abd); 1911 } 1912 1913 static zio_t * 1914 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, 1915 uint64_t offset) 1916 { 1917 if (gn != NULL) 1918 return (pio); 1919 1920 return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset), 1921 BP_GET_PSIZE(bp), zio_gang_issue_func_done, 1922 NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), 1923 &pio->io_bookmark)); 1924 } 1925 1926 static zio_t * 1927 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, 1928 uint64_t offset) 1929 { 1930 zio_t *zio; 1931 1932 if (gn != NULL) { 1933 abd_t *gbh_abd = 1934 abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE); 1935 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, 1936 gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL, 1937 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), 1938 &pio->io_bookmark); 1939 /* 1940 * As we rewrite each gang header, the pipeline will compute 1941 * a new gang block header checksum for it; but no one will 1942 * compute a new data checksum, so we do that here. The one 1943 * exception is the gang leader: the pipeline already computed 1944 * its data checksum because that stage precedes gang assembly. 1945 * (Presently, nothing actually uses interior data checksums; 1946 * this is just good hygiene.) 1947 */ 1948 if (gn != pio->io_gang_leader->io_gang_tree) { 1949 abd_t *buf = abd_get_offset(data, offset); 1950 1951 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp), 1952 buf, BP_GET_PSIZE(bp)); 1953 1954 abd_put(buf); 1955 } 1956 /* 1957 * If we are here to damage data for testing purposes, 1958 * leave the GBH alone so that we can detect the damage. 1959 */ 1960 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE) 1961 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; 1962 } else { 1963 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, 1964 abd_get_offset(data, offset), BP_GET_PSIZE(bp), 1965 zio_gang_issue_func_done, NULL, pio->io_priority, 1966 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 1967 } 1968 1969 return (zio); 1970 } 1971 1972 /* ARGSUSED */ 1973 static zio_t * 1974 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, 1975 uint64_t offset) 1976 { 1977 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp, 1978 ZIO_GANG_CHILD_FLAGS(pio))); 1979 } 1980 1981 /* ARGSUSED */ 1982 static zio_t * 1983 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, 1984 uint64_t offset) 1985 { 1986 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp, 1987 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio))); 1988 } 1989 1990 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = { 1991 NULL, 1992 zio_read_gang, 1993 zio_rewrite_gang, 1994 zio_free_gang, 1995 zio_claim_gang, 1996 NULL 1997 }; 1998 1999 static void zio_gang_tree_assemble_done(zio_t *zio); 2000 2001 static zio_gang_node_t * 2002 zio_gang_node_alloc(zio_gang_node_t **gnpp) 2003 { 2004 zio_gang_node_t *gn; 2005 2006 ASSERT(*gnpp == NULL); 2007 2008 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP); 2009 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE); 2010 *gnpp = gn; 2011 2012 return (gn); 2013 } 2014 2015 static void 2016 zio_gang_node_free(zio_gang_node_t **gnpp) 2017 { 2018 zio_gang_node_t *gn = *gnpp; 2019 2020 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) 2021 ASSERT(gn->gn_child[g] == NULL); 2022 2023 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE); 2024 kmem_free(gn, sizeof (*gn)); 2025 *gnpp = NULL; 2026 } 2027 2028 static void 2029 zio_gang_tree_free(zio_gang_node_t **gnpp) 2030 { 2031 zio_gang_node_t *gn = *gnpp; 2032 2033 if (gn == NULL) 2034 return; 2035 2036 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) 2037 zio_gang_tree_free(&gn->gn_child[g]); 2038 2039 zio_gang_node_free(gnpp); 2040 } 2041 2042 static void 2043 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp) 2044 { 2045 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp); 2046 abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE); 2047 2048 ASSERT(gio->io_gang_leader == gio); 2049 ASSERT(BP_IS_GANG(bp)); 2050 2051 zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE, 2052 zio_gang_tree_assemble_done, gn, gio->io_priority, 2053 ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark)); 2054 } 2055 2056 static void 2057 zio_gang_tree_assemble_done(zio_t *zio) 2058 { 2059 zio_t *gio = zio->io_gang_leader; 2060 zio_gang_node_t *gn = zio->io_private; 2061 blkptr_t *bp = zio->io_bp; 2062 2063 ASSERT(gio == zio_unique_parent(zio)); 2064 ASSERT(zio->io_child_count == 0); 2065 2066 if (zio->io_error) 2067 return; 2068 2069 /* this ABD was created from a linear buf in zio_gang_tree_assemble */ 2070 if (BP_SHOULD_BYTESWAP(bp)) 2071 byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size); 2072 2073 ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh); 2074 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE); 2075 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); 2076 2077 abd_put(zio->io_abd); 2078 2079 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 2080 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; 2081 if (!BP_IS_GANG(gbp)) 2082 continue; 2083 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]); 2084 } 2085 } 2086 2087 static void 2088 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data, 2089 uint64_t offset) 2090 { 2091 zio_t *gio = pio->io_gang_leader; 2092 zio_t *zio; 2093 2094 ASSERT(BP_IS_GANG(bp) == !!gn); 2095 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp)); 2096 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree); 2097 2098 /* 2099 * If you're a gang header, your data is in gn->gn_gbh. 2100 * If you're a gang member, your data is in 'data' and gn == NULL. 2101 */ 2102 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset); 2103 2104 if (gn != NULL) { 2105 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); 2106 2107 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 2108 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; 2109 if (BP_IS_HOLE(gbp)) 2110 continue; 2111 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data, 2112 offset); 2113 offset += BP_GET_PSIZE(gbp); 2114 } 2115 } 2116 2117 if (gn == gio->io_gang_tree) 2118 ASSERT3U(gio->io_size, ==, offset); 2119 2120 if (zio != pio) 2121 zio_nowait(zio); 2122 } 2123 2124 static int 2125 zio_gang_assemble(zio_t *zio) 2126 { 2127 blkptr_t *bp = zio->io_bp; 2128 2129 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL); 2130 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 2131 2132 zio->io_gang_leader = zio; 2133 2134 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree); 2135 2136 return (ZIO_PIPELINE_CONTINUE); 2137 } 2138 2139 static int 2140 zio_gang_issue(zio_t *zio) 2141 { 2142 blkptr_t *bp = zio->io_bp; 2143 2144 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT, ZIO_WAIT_DONE)) { 2145 return (ZIO_PIPELINE_STOP); 2146 } 2147 2148 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio); 2149 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 2150 2151 if (zio->io_child_error[ZIO_CHILD_GANG] == 0) 2152 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd, 2153 0); 2154 else 2155 zio_gang_tree_free(&zio->io_gang_tree); 2156 2157 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2158 2159 return (ZIO_PIPELINE_CONTINUE); 2160 } 2161 2162 static void 2163 zio_write_gang_member_ready(zio_t *zio) 2164 { 2165 zio_t *pio = zio_unique_parent(zio); 2166 zio_t *gio = zio->io_gang_leader; 2167 dva_t *cdva = zio->io_bp->blk_dva; 2168 dva_t *pdva = pio->io_bp->blk_dva; 2169 uint64_t asize; 2170 2171 if (BP_IS_HOLE(zio->io_bp)) 2172 return; 2173 2174 ASSERT(BP_IS_HOLE(&zio->io_bp_orig)); 2175 2176 ASSERT(zio->io_child_type == ZIO_CHILD_GANG); 2177 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies); 2178 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp)); 2179 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp)); 2180 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp)); 2181 2182 mutex_enter(&pio->io_lock); 2183 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) { 2184 ASSERT(DVA_GET_GANG(&pdva[d])); 2185 asize = DVA_GET_ASIZE(&pdva[d]); 2186 asize += DVA_GET_ASIZE(&cdva[d]); 2187 DVA_SET_ASIZE(&pdva[d], asize); 2188 } 2189 mutex_exit(&pio->io_lock); 2190 } 2191 2192 static void 2193 zio_write_gang_done(zio_t *zio) 2194 { 2195 abd_put(zio->io_abd); 2196 } 2197 2198 static int 2199 zio_write_gang_block(zio_t *pio) 2200 { 2201 spa_t *spa = pio->io_spa; 2202 metaslab_class_t *mc = spa_normal_class(spa); 2203 blkptr_t *bp = pio->io_bp; 2204 zio_t *gio = pio->io_gang_leader; 2205 zio_t *zio; 2206 zio_gang_node_t *gn, **gnpp; 2207 zio_gbh_phys_t *gbh; 2208 abd_t *gbh_abd; 2209 uint64_t txg = pio->io_txg; 2210 uint64_t resid = pio->io_size; 2211 uint64_t lsize; 2212 int copies = gio->io_prop.zp_copies; 2213 int gbh_copies = MIN(copies + 1, spa_max_replication(spa)); 2214 zio_prop_t zp; 2215 int error; 2216 2217 int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER; 2218 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { 2219 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); 2220 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA)); 2221 2222 flags |= METASLAB_ASYNC_ALLOC; 2223 VERIFY(refcount_held(&mc->mc_alloc_slots[pio->io_allocator], 2224 pio)); 2225 2226 /* 2227 * The logical zio has already placed a reservation for 2228 * 'copies' allocation slots but gang blocks may require 2229 * additional copies. These additional copies 2230 * (i.e. gbh_copies - copies) are guaranteed to succeed 2231 * since metaslab_class_throttle_reserve() always allows 2232 * additional reservations for gang blocks. 2233 */ 2234 VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies, 2235 pio->io_allocator, pio, flags)); 2236 } 2237 2238 error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE, 2239 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags, 2240 &pio->io_alloc_list, pio, pio->io_allocator); 2241 if (error) { 2242 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { 2243 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); 2244 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA)); 2245 2246 /* 2247 * If we failed to allocate the gang block header then 2248 * we remove any additional allocation reservations that 2249 * we placed here. The original reservation will 2250 * be removed when the logical I/O goes to the ready 2251 * stage. 2252 */ 2253 metaslab_class_throttle_unreserve(mc, 2254 gbh_copies - copies, pio->io_allocator, pio); 2255 } 2256 pio->io_error = error; 2257 return (ZIO_PIPELINE_CONTINUE); 2258 } 2259 2260 if (pio == gio) { 2261 gnpp = &gio->io_gang_tree; 2262 } else { 2263 gnpp = pio->io_private; 2264 ASSERT(pio->io_ready == zio_write_gang_member_ready); 2265 } 2266 2267 gn = zio_gang_node_alloc(gnpp); 2268 gbh = gn->gn_gbh; 2269 bzero(gbh, SPA_GANGBLOCKSIZE); 2270 gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE); 2271 2272 /* 2273 * Create the gang header. 2274 */ 2275 zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE, 2276 zio_write_gang_done, NULL, pio->io_priority, 2277 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 2278 2279 /* 2280 * Create and nowait the gang children. 2281 */ 2282 for (int g = 0; resid != 0; resid -= lsize, g++) { 2283 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g), 2284 SPA_MINBLOCKSIZE); 2285 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid); 2286 2287 zp.zp_checksum = gio->io_prop.zp_checksum; 2288 zp.zp_compress = ZIO_COMPRESS_OFF; 2289 zp.zp_type = DMU_OT_NONE; 2290 zp.zp_level = 0; 2291 zp.zp_copies = gio->io_prop.zp_copies; 2292 zp.zp_dedup = B_FALSE; 2293 zp.zp_dedup_verify = B_FALSE; 2294 zp.zp_nopwrite = B_FALSE; 2295 2296 zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g], 2297 abd_get_offset(pio->io_abd, pio->io_size - resid), lsize, 2298 lsize, &zp, zio_write_gang_member_ready, NULL, NULL, 2299 zio_write_gang_done, &gn->gn_child[g], pio->io_priority, 2300 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 2301 2302 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { 2303 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); 2304 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA)); 2305 2306 /* 2307 * Gang children won't throttle but we should 2308 * account for their work, so reserve an allocation 2309 * slot for them here. 2310 */ 2311 VERIFY(metaslab_class_throttle_reserve(mc, 2312 zp.zp_copies, cio->io_allocator, cio, flags)); 2313 } 2314 zio_nowait(cio); 2315 } 2316 2317 /* 2318 * Set pio's pipeline to just wait for zio to finish. 2319 */ 2320 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2321 2322 zio_nowait(zio); 2323 2324 return (ZIO_PIPELINE_CONTINUE); 2325 } 2326 2327 /* 2328 * The zio_nop_write stage in the pipeline determines if allocating a 2329 * new bp is necessary. The nopwrite feature can handle writes in 2330 * either syncing or open context (i.e. zil writes) and as a result is 2331 * mutually exclusive with dedup. 2332 * 2333 * By leveraging a cryptographically secure checksum, such as SHA256, we 2334 * can compare the checksums of the new data and the old to determine if 2335 * allocating a new block is required. Note that our requirements for 2336 * cryptographic strength are fairly weak: there can't be any accidental 2337 * hash collisions, but we don't need to be secure against intentional 2338 * (malicious) collisions. To trigger a nopwrite, you have to be able 2339 * to write the file to begin with, and triggering an incorrect (hash 2340 * collision) nopwrite is no worse than simply writing to the file. 2341 * That said, there are no known attacks against the checksum algorithms 2342 * used for nopwrite, assuming that the salt and the checksums 2343 * themselves remain secret. 2344 */ 2345 static int 2346 zio_nop_write(zio_t *zio) 2347 { 2348 blkptr_t *bp = zio->io_bp; 2349 blkptr_t *bp_orig = &zio->io_bp_orig; 2350 zio_prop_t *zp = &zio->io_prop; 2351 2352 ASSERT(BP_GET_LEVEL(bp) == 0); 2353 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 2354 ASSERT(zp->zp_nopwrite); 2355 ASSERT(!zp->zp_dedup); 2356 ASSERT(zio->io_bp_override == NULL); 2357 ASSERT(IO_IS_ALLOCATING(zio)); 2358 2359 /* 2360 * Check to see if the original bp and the new bp have matching 2361 * characteristics (i.e. same checksum, compression algorithms, etc). 2362 * If they don't then just continue with the pipeline which will 2363 * allocate a new bp. 2364 */ 2365 if (BP_IS_HOLE(bp_orig) || 2366 !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags & 2367 ZCHECKSUM_FLAG_NOPWRITE) || 2368 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) || 2369 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) || 2370 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) || 2371 zp->zp_copies != BP_GET_NDVAS(bp_orig)) 2372 return (ZIO_PIPELINE_CONTINUE); 2373 2374 /* 2375 * If the checksums match then reset the pipeline so that we 2376 * avoid allocating a new bp and issuing any I/O. 2377 */ 2378 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) { 2379 ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags & 2380 ZCHECKSUM_FLAG_NOPWRITE); 2381 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig)); 2382 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig)); 2383 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF); 2384 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop, 2385 sizeof (uint64_t)) == 0); 2386 2387 *bp = *bp_orig; 2388 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2389 zio->io_flags |= ZIO_FLAG_NOPWRITE; 2390 } 2391 2392 return (ZIO_PIPELINE_CONTINUE); 2393 } 2394 2395 /* 2396 * ========================================================================== 2397 * Dedup 2398 * ========================================================================== 2399 */ 2400 static void 2401 zio_ddt_child_read_done(zio_t *zio) 2402 { 2403 blkptr_t *bp = zio->io_bp; 2404 ddt_entry_t *dde = zio->io_private; 2405 ddt_phys_t *ddp; 2406 zio_t *pio = zio_unique_parent(zio); 2407 2408 mutex_enter(&pio->io_lock); 2409 ddp = ddt_phys_select(dde, bp); 2410 if (zio->io_error == 0) 2411 ddt_phys_clear(ddp); /* this ddp doesn't need repair */ 2412 2413 if (zio->io_error == 0 && dde->dde_repair_abd == NULL) 2414 dde->dde_repair_abd = zio->io_abd; 2415 else 2416 abd_free(zio->io_abd); 2417 mutex_exit(&pio->io_lock); 2418 } 2419 2420 static int 2421 zio_ddt_read_start(zio_t *zio) 2422 { 2423 blkptr_t *bp = zio->io_bp; 2424 2425 ASSERT(BP_GET_DEDUP(bp)); 2426 ASSERT(BP_GET_PSIZE(bp) == zio->io_size); 2427 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2428 2429 if (zio->io_child_error[ZIO_CHILD_DDT]) { 2430 ddt_t *ddt = ddt_select(zio->io_spa, bp); 2431 ddt_entry_t *dde = ddt_repair_start(ddt, bp); 2432 ddt_phys_t *ddp = dde->dde_phys; 2433 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp); 2434 blkptr_t blk; 2435 2436 ASSERT(zio->io_vsd == NULL); 2437 zio->io_vsd = dde; 2438 2439 if (ddp_self == NULL) 2440 return (ZIO_PIPELINE_CONTINUE); 2441 2442 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) { 2443 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self) 2444 continue; 2445 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp, 2446 &blk); 2447 zio_nowait(zio_read(zio, zio->io_spa, &blk, 2448 abd_alloc_for_io(zio->io_size, B_TRUE), 2449 zio->io_size, zio_ddt_child_read_done, dde, 2450 zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) | 2451 ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark)); 2452 } 2453 return (ZIO_PIPELINE_CONTINUE); 2454 } 2455 2456 zio_nowait(zio_read(zio, zio->io_spa, bp, 2457 zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority, 2458 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark)); 2459 2460 return (ZIO_PIPELINE_CONTINUE); 2461 } 2462 2463 static int 2464 zio_ddt_read_done(zio_t *zio) 2465 { 2466 blkptr_t *bp = zio->io_bp; 2467 2468 if (zio_wait_for_children(zio, ZIO_CHILD_DDT_BIT, ZIO_WAIT_DONE)) { 2469 return (ZIO_PIPELINE_STOP); 2470 } 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, int allocator) 2803 { 2804 zio_t *zio; 2805 2806 ASSERT(MUTEX_HELD(&spa->spa_alloc_locks[allocator])); 2807 2808 zio = avl_first(&spa->spa_alloc_trees[allocator]); 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 ASSERT3U(zio->io_allocator, ==, allocator); 2819 if (!metaslab_class_throttle_reserve(spa_normal_class(spa), 2820 zio->io_prop.zp_copies, zio->io_allocator, zio, 0)) { 2821 return (NULL); 2822 } 2823 2824 avl_remove(&spa->spa_alloc_trees[allocator], zio); 2825 ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE); 2826 2827 return (zio); 2828 } 2829 2830 static int 2831 zio_dva_throttle(zio_t *zio) 2832 { 2833 spa_t *spa = zio->io_spa; 2834 zio_t *nio; 2835 2836 if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE || 2837 !spa_normal_class(zio->io_spa)->mc_alloc_throttle_enabled || 2838 zio->io_child_type == ZIO_CHILD_GANG || 2839 zio->io_flags & ZIO_FLAG_NODATA) { 2840 return (ZIO_PIPELINE_CONTINUE); 2841 } 2842 2843 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 2844 2845 ASSERT3U(zio->io_queued_timestamp, >, 0); 2846 ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE); 2847 2848 zbookmark_phys_t *bm = &zio->io_bookmark; 2849 /* 2850 * We want to try to use as many allocators as possible to help improve 2851 * performance, but we also want logically adjacent IOs to be physically 2852 * adjacent to improve sequential read performance. We chunk each object 2853 * into 2^20 block regions, and then hash based on the objset, object, 2854 * level, and region to accomplish both of these goals. 2855 */ 2856 zio->io_allocator = cityhash4(bm->zb_objset, bm->zb_object, 2857 bm->zb_level, bm->zb_blkid >> 20) % spa->spa_alloc_count; 2858 mutex_enter(&spa->spa_alloc_locks[zio->io_allocator]); 2859 2860 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 2861 avl_add(&spa->spa_alloc_trees[zio->io_allocator], zio); 2862 2863 nio = zio_io_to_allocate(zio->io_spa, zio->io_allocator); 2864 mutex_exit(&spa->spa_alloc_locks[zio->io_allocator]); 2865 2866 if (nio == zio) 2867 return (ZIO_PIPELINE_CONTINUE); 2868 2869 if (nio != NULL) { 2870 ASSERT(nio->io_stage == ZIO_STAGE_DVA_THROTTLE); 2871 /* 2872 * We are passing control to a new zio so make sure that 2873 * it is processed by a different thread. We do this to 2874 * avoid stack overflows that can occur when parents are 2875 * throttled and children are making progress. We allow 2876 * it to go to the head of the taskq since it's already 2877 * been waiting. 2878 */ 2879 zio_taskq_dispatch(nio, ZIO_TASKQ_ISSUE, B_TRUE); 2880 } 2881 return (ZIO_PIPELINE_STOP); 2882 } 2883 2884 void 2885 zio_allocate_dispatch(spa_t *spa, int allocator) 2886 { 2887 zio_t *zio; 2888 2889 mutex_enter(&spa->spa_alloc_locks[allocator]); 2890 zio = zio_io_to_allocate(spa, allocator); 2891 mutex_exit(&spa->spa_alloc_locks[allocator]); 2892 if (zio == NULL) 2893 return; 2894 2895 ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE); 2896 ASSERT0(zio->io_error); 2897 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE); 2898 } 2899 2900 static int 2901 zio_dva_allocate(zio_t *zio) 2902 { 2903 spa_t *spa = zio->io_spa; 2904 metaslab_class_t *mc = spa_normal_class(spa); 2905 blkptr_t *bp = zio->io_bp; 2906 int error; 2907 int flags = 0; 2908 2909 if (zio->io_gang_leader == NULL) { 2910 ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 2911 zio->io_gang_leader = zio; 2912 } 2913 2914 ASSERT(BP_IS_HOLE(bp)); 2915 ASSERT0(BP_GET_NDVAS(bp)); 2916 ASSERT3U(zio->io_prop.zp_copies, >, 0); 2917 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa)); 2918 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp)); 2919 2920 if (zio->io_flags & ZIO_FLAG_NODATA) { 2921 flags |= METASLAB_DONT_THROTTLE; 2922 } 2923 if (zio->io_flags & ZIO_FLAG_GANG_CHILD) { 2924 flags |= METASLAB_GANG_CHILD; 2925 } 2926 if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE) { 2927 flags |= METASLAB_ASYNC_ALLOC; 2928 } 2929 2930 error = metaslab_alloc(spa, mc, zio->io_size, bp, 2931 zio->io_prop.zp_copies, zio->io_txg, NULL, flags, 2932 &zio->io_alloc_list, zio, zio->io_allocator); 2933 2934 if (error != 0) { 2935 zfs_dbgmsg("%s: metaslab allocation failure: zio %p, " 2936 "size %llu, error %d", spa_name(spa), zio, zio->io_size, 2937 error); 2938 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE) 2939 return (zio_write_gang_block(zio)); 2940 zio->io_error = error; 2941 } 2942 2943 return (ZIO_PIPELINE_CONTINUE); 2944 } 2945 2946 static int 2947 zio_dva_free(zio_t *zio) 2948 { 2949 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE); 2950 2951 return (ZIO_PIPELINE_CONTINUE); 2952 } 2953 2954 static int 2955 zio_dva_claim(zio_t *zio) 2956 { 2957 int error; 2958 2959 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg); 2960 if (error) 2961 zio->io_error = error; 2962 2963 return (ZIO_PIPELINE_CONTINUE); 2964 } 2965 2966 /* 2967 * Undo an allocation. This is used by zio_done() when an I/O fails 2968 * and we want to give back the block we just allocated. 2969 * This handles both normal blocks and gang blocks. 2970 */ 2971 static void 2972 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp) 2973 { 2974 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp)); 2975 ASSERT(zio->io_bp_override == NULL); 2976 2977 if (!BP_IS_HOLE(bp)) 2978 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE); 2979 2980 if (gn != NULL) { 2981 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 2982 zio_dva_unallocate(zio, gn->gn_child[g], 2983 &gn->gn_gbh->zg_blkptr[g]); 2984 } 2985 } 2986 } 2987 2988 /* 2989 * Try to allocate an intent log block. Return 0 on success, errno on failure. 2990 */ 2991 int 2992 zio_alloc_zil(spa_t *spa, uint64_t objset, uint64_t txg, blkptr_t *new_bp, 2993 blkptr_t *old_bp, uint64_t size, boolean_t *slog) 2994 { 2995 int error = 1; 2996 zio_alloc_list_t io_alloc_list; 2997 2998 ASSERT(txg > spa_syncing_txg(spa)); 2999 3000 metaslab_trace_init(&io_alloc_list); 3001 /* 3002 * When allocating a zil block, we don't have information about 3003 * the final destination of the block except the objset it's part 3004 * of, so we just hash the objset ID to pick the allocator to get 3005 * some parallelism. 3006 */ 3007 error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1, 3008 txg, old_bp, METASLAB_HINTBP_AVOID, &io_alloc_list, NULL, 3009 cityhash4(0, 0, 0, objset) % spa->spa_alloc_count); 3010 if (error == 0) { 3011 *slog = TRUE; 3012 } else { 3013 error = metaslab_alloc(spa, spa_normal_class(spa), size, 3014 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID, 3015 &io_alloc_list, NULL, cityhash4(0, 0, 0, objset) % 3016 spa->spa_alloc_count); 3017 if (error == 0) 3018 *slog = FALSE; 3019 } 3020 metaslab_trace_fini(&io_alloc_list); 3021 3022 if (error == 0) { 3023 BP_SET_LSIZE(new_bp, size); 3024 BP_SET_PSIZE(new_bp, size); 3025 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF); 3026 BP_SET_CHECKSUM(new_bp, 3027 spa_version(spa) >= SPA_VERSION_SLIM_ZIL 3028 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG); 3029 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); 3030 BP_SET_LEVEL(new_bp, 0); 3031 BP_SET_DEDUP(new_bp, 0); 3032 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER); 3033 } else { 3034 zfs_dbgmsg("%s: zil block allocation failure: " 3035 "size %llu, error %d", spa_name(spa), size, error); 3036 } 3037 3038 return (error); 3039 } 3040 3041 /* 3042 * ========================================================================== 3043 * Read and write to physical devices 3044 * ========================================================================== 3045 */ 3046 3047 3048 /* 3049 * Issue an I/O to the underlying vdev. Typically the issue pipeline 3050 * stops after this stage and will resume upon I/O completion. 3051 * However, there are instances where the vdev layer may need to 3052 * continue the pipeline when an I/O was not issued. Since the I/O 3053 * that was sent to the vdev layer might be different than the one 3054 * currently active in the pipeline (see vdev_queue_io()), we explicitly 3055 * force the underlying vdev layers to call either zio_execute() or 3056 * zio_interrupt() to ensure that the pipeline continues with the correct I/O. 3057 */ 3058 static int 3059 zio_vdev_io_start(zio_t *zio) 3060 { 3061 vdev_t *vd = zio->io_vd; 3062 uint64_t align; 3063 spa_t *spa = zio->io_spa; 3064 3065 ASSERT(zio->io_error == 0); 3066 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0); 3067 3068 if (vd == NULL) { 3069 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) 3070 spa_config_enter(spa, SCL_ZIO, zio, RW_READER); 3071 3072 /* 3073 * The mirror_ops handle multiple DVAs in a single BP. 3074 */ 3075 vdev_mirror_ops.vdev_op_io_start(zio); 3076 return (ZIO_PIPELINE_STOP); 3077 } 3078 3079 ASSERT3P(zio->io_logical, !=, zio); 3080 if (zio->io_type == ZIO_TYPE_WRITE) { 3081 ASSERT(spa->spa_trust_config); 3082 3083 if (zio->io_vd->vdev_removing) { 3084 /* 3085 * Note: the code can handle other kinds of writes, 3086 * but we don't expect them. 3087 */ 3088 ASSERT(zio->io_flags & 3089 (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL | 3090 ZIO_FLAG_RESILVER | ZIO_FLAG_INDUCE_DAMAGE)); 3091 } 3092 } 3093 3094 /* 3095 * We keep track of time-sensitive I/Os so that the scan thread 3096 * can quickly react to certain workloads. In particular, we care 3097 * about non-scrubbing, top-level reads and writes with the following 3098 * characteristics: 3099 * - synchronous writes of user data to non-slog devices 3100 * - any reads of user data 3101 * When these conditions are met, adjust the timestamp of spa_last_io 3102 * which allows the scan thread to adjust its workload accordingly. 3103 */ 3104 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL && 3105 vd == vd->vdev_top && !vd->vdev_islog && 3106 zio->io_bookmark.zb_objset != DMU_META_OBJSET && 3107 zio->io_txg != spa_syncing_txg(spa)) { 3108 uint64_t old = spa->spa_last_io; 3109 uint64_t new = ddi_get_lbolt64(); 3110 if (old != new) 3111 (void) atomic_cas_64(&spa->spa_last_io, old, new); 3112 } 3113 3114 align = 1ULL << vd->vdev_top->vdev_ashift; 3115 3116 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) && 3117 P2PHASE(zio->io_size, align) != 0) { 3118 /* Transform logical writes to be a full physical block size. */ 3119 uint64_t asize = P2ROUNDUP(zio->io_size, align); 3120 abd_t *abuf = abd_alloc_sametype(zio->io_abd, asize); 3121 ASSERT(vd == vd->vdev_top); 3122 if (zio->io_type == ZIO_TYPE_WRITE) { 3123 abd_copy(abuf, zio->io_abd, zio->io_size); 3124 abd_zero_off(abuf, zio->io_size, asize - zio->io_size); 3125 } 3126 zio_push_transform(zio, abuf, asize, asize, zio_subblock); 3127 } 3128 3129 /* 3130 * If this is not a physical io, make sure that it is properly aligned 3131 * before proceeding. 3132 */ 3133 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) { 3134 ASSERT0(P2PHASE(zio->io_offset, align)); 3135 ASSERT0(P2PHASE(zio->io_size, align)); 3136 } else { 3137 /* 3138 * For physical writes, we allow 512b aligned writes and assume 3139 * the device will perform a read-modify-write as necessary. 3140 */ 3141 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE)); 3142 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE)); 3143 } 3144 3145 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa)); 3146 3147 /* 3148 * If this is a repair I/O, and there's no self-healing involved -- 3149 * that is, we're just resilvering what we expect to resilver -- 3150 * then don't do the I/O unless zio's txg is actually in vd's DTL. 3151 * This prevents spurious resilvering. 3152 * 3153 * There are a few ways that we can end up creating these spurious 3154 * resilver i/os: 3155 * 3156 * 1. A resilver i/o will be issued if any DVA in the BP has a 3157 * dirty DTL. The mirror code will issue resilver writes to 3158 * each DVA, including the one(s) that are not on vdevs with dirty 3159 * DTLs. 3160 * 3161 * 2. With nested replication, which happens when we have a 3162 * "replacing" or "spare" vdev that's a child of a mirror or raidz. 3163 * For example, given mirror(replacing(A+B), C), it's likely that 3164 * only A is out of date (it's the new device). In this case, we'll 3165 * read from C, then use the data to resilver A+B -- but we don't 3166 * actually want to resilver B, just A. The top-level mirror has no 3167 * way to know this, so instead we just discard unnecessary repairs 3168 * as we work our way down the vdev tree. 3169 * 3170 * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc. 3171 * The same logic applies to any form of nested replication: ditto 3172 * + mirror, RAID-Z + replacing, etc. 3173 * 3174 * However, indirect vdevs point off to other vdevs which may have 3175 * DTL's, so we never bypass them. The child i/os on concrete vdevs 3176 * will be properly bypassed instead. 3177 */ 3178 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) && 3179 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) && 3180 zio->io_txg != 0 && /* not a delegated i/o */ 3181 vd->vdev_ops != &vdev_indirect_ops && 3182 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) { 3183 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 3184 zio_vdev_io_bypass(zio); 3185 return (ZIO_PIPELINE_CONTINUE); 3186 } 3187 3188 if (vd->vdev_ops->vdev_op_leaf && 3189 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) { 3190 3191 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio)) 3192 return (ZIO_PIPELINE_CONTINUE); 3193 3194 if ((zio = vdev_queue_io(zio)) == NULL) 3195 return (ZIO_PIPELINE_STOP); 3196 3197 if (!vdev_accessible(vd, zio)) { 3198 zio->io_error = SET_ERROR(ENXIO); 3199 zio_interrupt(zio); 3200 return (ZIO_PIPELINE_STOP); 3201 } 3202 } 3203 3204 vd->vdev_ops->vdev_op_io_start(zio); 3205 return (ZIO_PIPELINE_STOP); 3206 } 3207 3208 static int 3209 zio_vdev_io_done(zio_t *zio) 3210 { 3211 vdev_t *vd = zio->io_vd; 3212 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops; 3213 boolean_t unexpected_error = B_FALSE; 3214 3215 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) { 3216 return (ZIO_PIPELINE_STOP); 3217 } 3218 3219 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE); 3220 3221 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) { 3222 3223 vdev_queue_io_done(zio); 3224 3225 if (zio->io_type == ZIO_TYPE_WRITE) 3226 vdev_cache_write(zio); 3227 3228 if (zio_injection_enabled && zio->io_error == 0) 3229 zio->io_error = zio_handle_device_injection(vd, 3230 zio, EIO); 3231 3232 if (zio_injection_enabled && zio->io_error == 0) 3233 zio->io_error = zio_handle_label_injection(zio, EIO); 3234 3235 if (zio->io_error) { 3236 if (!vdev_accessible(vd, zio)) { 3237 zio->io_error = SET_ERROR(ENXIO); 3238 } else { 3239 unexpected_error = B_TRUE; 3240 } 3241 } 3242 } 3243 3244 ops->vdev_op_io_done(zio); 3245 3246 if (unexpected_error) 3247 VERIFY(vdev_probe(vd, zio) == NULL); 3248 3249 return (ZIO_PIPELINE_CONTINUE); 3250 } 3251 3252 /* 3253 * For non-raidz ZIOs, we can just copy aside the bad data read from the 3254 * disk, and use that to finish the checksum ereport later. 3255 */ 3256 static void 3257 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr, 3258 const void *good_buf) 3259 { 3260 /* no processing needed */ 3261 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE); 3262 } 3263 3264 /*ARGSUSED*/ 3265 void 3266 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored) 3267 { 3268 void *buf = zio_buf_alloc(zio->io_size); 3269 3270 abd_copy_to_buf(buf, zio->io_abd, zio->io_size); 3271 3272 zcr->zcr_cbinfo = zio->io_size; 3273 zcr->zcr_cbdata = buf; 3274 zcr->zcr_finish = zio_vsd_default_cksum_finish; 3275 zcr->zcr_free = zio_buf_free; 3276 } 3277 3278 static int 3279 zio_vdev_io_assess(zio_t *zio) 3280 { 3281 vdev_t *vd = zio->io_vd; 3282 3283 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) { 3284 return (ZIO_PIPELINE_STOP); 3285 } 3286 3287 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) 3288 spa_config_exit(zio->io_spa, SCL_ZIO, zio); 3289 3290 if (zio->io_vsd != NULL) { 3291 zio->io_vsd_ops->vsd_free(zio); 3292 zio->io_vsd = NULL; 3293 } 3294 3295 if (zio_injection_enabled && zio->io_error == 0) 3296 zio->io_error = zio_handle_fault_injection(zio, EIO); 3297 3298 /* 3299 * If the I/O failed, determine whether we should attempt to retry it. 3300 * 3301 * On retry, we cut in line in the issue queue, since we don't want 3302 * compression/checksumming/etc. work to prevent our (cheap) IO reissue. 3303 */ 3304 if (zio->io_error && vd == NULL && 3305 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) { 3306 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */ 3307 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */ 3308 zio->io_error = 0; 3309 zio->io_flags |= ZIO_FLAG_IO_RETRY | 3310 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE; 3311 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1; 3312 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, 3313 zio_requeue_io_start_cut_in_line); 3314 return (ZIO_PIPELINE_STOP); 3315 } 3316 3317 /* 3318 * If we got an error on a leaf device, convert it to ENXIO 3319 * if the device is not accessible at all. 3320 */ 3321 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf && 3322 !vdev_accessible(vd, zio)) 3323 zio->io_error = SET_ERROR(ENXIO); 3324 3325 /* 3326 * If we can't write to an interior vdev (mirror or RAID-Z), 3327 * set vdev_cant_write so that we stop trying to allocate from it. 3328 */ 3329 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE && 3330 vd != NULL && !vd->vdev_ops->vdev_op_leaf) { 3331 vd->vdev_cant_write = B_TRUE; 3332 } 3333 3334 /* 3335 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future 3336 * attempts will ever succeed. In this case we set a persistent bit so 3337 * that we don't bother with it in the future. 3338 */ 3339 if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) && 3340 zio->io_type == ZIO_TYPE_IOCTL && 3341 zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL) 3342 vd->vdev_nowritecache = B_TRUE; 3343 3344 if (zio->io_error) 3345 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 3346 3347 if (vd != NULL && vd->vdev_ops->vdev_op_leaf && 3348 zio->io_physdone != NULL) { 3349 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED)); 3350 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV); 3351 zio->io_physdone(zio->io_logical); 3352 } 3353 3354 return (ZIO_PIPELINE_CONTINUE); 3355 } 3356 3357 void 3358 zio_vdev_io_reissue(zio_t *zio) 3359 { 3360 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); 3361 ASSERT(zio->io_error == 0); 3362 3363 zio->io_stage >>= 1; 3364 } 3365 3366 void 3367 zio_vdev_io_redone(zio_t *zio) 3368 { 3369 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE); 3370 3371 zio->io_stage >>= 1; 3372 } 3373 3374 void 3375 zio_vdev_io_bypass(zio_t *zio) 3376 { 3377 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); 3378 ASSERT(zio->io_error == 0); 3379 3380 zio->io_flags |= ZIO_FLAG_IO_BYPASS; 3381 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1; 3382 } 3383 3384 /* 3385 * ========================================================================== 3386 * Generate and verify checksums 3387 * ========================================================================== 3388 */ 3389 static int 3390 zio_checksum_generate(zio_t *zio) 3391 { 3392 blkptr_t *bp = zio->io_bp; 3393 enum zio_checksum checksum; 3394 3395 if (bp == NULL) { 3396 /* 3397 * This is zio_write_phys(). 3398 * We're either generating a label checksum, or none at all. 3399 */ 3400 checksum = zio->io_prop.zp_checksum; 3401 3402 if (checksum == ZIO_CHECKSUM_OFF) 3403 return (ZIO_PIPELINE_CONTINUE); 3404 3405 ASSERT(checksum == ZIO_CHECKSUM_LABEL); 3406 } else { 3407 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) { 3408 ASSERT(!IO_IS_ALLOCATING(zio)); 3409 checksum = ZIO_CHECKSUM_GANG_HEADER; 3410 } else { 3411 checksum = BP_GET_CHECKSUM(bp); 3412 } 3413 } 3414 3415 zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size); 3416 3417 return (ZIO_PIPELINE_CONTINUE); 3418 } 3419 3420 static int 3421 zio_checksum_verify(zio_t *zio) 3422 { 3423 zio_bad_cksum_t info; 3424 blkptr_t *bp = zio->io_bp; 3425 int error; 3426 3427 ASSERT(zio->io_vd != NULL); 3428 3429 if (bp == NULL) { 3430 /* 3431 * This is zio_read_phys(). 3432 * We're either verifying a label checksum, or nothing at all. 3433 */ 3434 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF) 3435 return (ZIO_PIPELINE_CONTINUE); 3436 3437 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL); 3438 } 3439 3440 if ((error = zio_checksum_error(zio, &info)) != 0) { 3441 zio->io_error = error; 3442 if (error == ECKSUM && 3443 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) { 3444 zfs_ereport_start_checksum(zio->io_spa, 3445 zio->io_vd, zio, zio->io_offset, 3446 zio->io_size, NULL, &info); 3447 } 3448 } 3449 3450 return (ZIO_PIPELINE_CONTINUE); 3451 } 3452 3453 /* 3454 * Called by RAID-Z to ensure we don't compute the checksum twice. 3455 */ 3456 void 3457 zio_checksum_verified(zio_t *zio) 3458 { 3459 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; 3460 } 3461 3462 /* 3463 * ========================================================================== 3464 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other. 3465 * An error of 0 indicates success. ENXIO indicates whole-device failure, 3466 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO 3467 * indicate errors that are specific to one I/O, and most likely permanent. 3468 * Any other error is presumed to be worse because we weren't expecting it. 3469 * ========================================================================== 3470 */ 3471 int 3472 zio_worst_error(int e1, int e2) 3473 { 3474 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO }; 3475 int r1, r2; 3476 3477 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++) 3478 if (e1 == zio_error_rank[r1]) 3479 break; 3480 3481 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++) 3482 if (e2 == zio_error_rank[r2]) 3483 break; 3484 3485 return (r1 > r2 ? e1 : e2); 3486 } 3487 3488 /* 3489 * ========================================================================== 3490 * I/O completion 3491 * ========================================================================== 3492 */ 3493 static int 3494 zio_ready(zio_t *zio) 3495 { 3496 blkptr_t *bp = zio->io_bp; 3497 zio_t *pio, *pio_next; 3498 zio_link_t *zl = NULL; 3499 3500 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT | ZIO_CHILD_DDT_BIT, 3501 ZIO_WAIT_READY)) { 3502 return (ZIO_PIPELINE_STOP); 3503 } 3504 3505 if (zio->io_ready) { 3506 ASSERT(IO_IS_ALLOCATING(zio)); 3507 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) || 3508 (zio->io_flags & ZIO_FLAG_NOPWRITE)); 3509 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0); 3510 3511 zio->io_ready(zio); 3512 } 3513 3514 if (bp != NULL && bp != &zio->io_bp_copy) 3515 zio->io_bp_copy = *bp; 3516 3517 if (zio->io_error != 0) { 3518 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 3519 3520 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) { 3521 ASSERT(IO_IS_ALLOCATING(zio)); 3522 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); 3523 /* 3524 * We were unable to allocate anything, unreserve and 3525 * issue the next I/O to allocate. 3526 */ 3527 metaslab_class_throttle_unreserve( 3528 spa_normal_class(zio->io_spa), 3529 zio->io_prop.zp_copies, zio->io_allocator, zio); 3530 zio_allocate_dispatch(zio->io_spa, zio->io_allocator); 3531 } 3532 } 3533 3534 mutex_enter(&zio->io_lock); 3535 zio->io_state[ZIO_WAIT_READY] = 1; 3536 pio = zio_walk_parents(zio, &zl); 3537 mutex_exit(&zio->io_lock); 3538 3539 /* 3540 * As we notify zio's parents, new parents could be added. 3541 * New parents go to the head of zio's io_parent_list, however, 3542 * so we will (correctly) not notify them. The remainder of zio's 3543 * io_parent_list, from 'pio_next' onward, cannot change because 3544 * all parents must wait for us to be done before they can be done. 3545 */ 3546 for (; pio != NULL; pio = pio_next) { 3547 pio_next = zio_walk_parents(zio, &zl); 3548 zio_notify_parent(pio, zio, ZIO_WAIT_READY); 3549 } 3550 3551 if (zio->io_flags & ZIO_FLAG_NODATA) { 3552 if (BP_IS_GANG(bp)) { 3553 zio->io_flags &= ~ZIO_FLAG_NODATA; 3554 } else { 3555 ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE); 3556 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; 3557 } 3558 } 3559 3560 if (zio_injection_enabled && 3561 zio->io_spa->spa_syncing_txg == zio->io_txg) 3562 zio_handle_ignored_writes(zio); 3563 3564 return (ZIO_PIPELINE_CONTINUE); 3565 } 3566 3567 /* 3568 * Update the allocation throttle accounting. 3569 */ 3570 static void 3571 zio_dva_throttle_done(zio_t *zio) 3572 { 3573 zio_t *lio = zio->io_logical; 3574 zio_t *pio = zio_unique_parent(zio); 3575 vdev_t *vd = zio->io_vd; 3576 int flags = METASLAB_ASYNC_ALLOC; 3577 3578 ASSERT3P(zio->io_bp, !=, NULL); 3579 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE); 3580 ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE); 3581 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV); 3582 ASSERT(vd != NULL); 3583 ASSERT3P(vd, ==, vd->vdev_top); 3584 ASSERT(!(zio->io_flags & (ZIO_FLAG_IO_REPAIR | ZIO_FLAG_IO_RETRY))); 3585 ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING); 3586 ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE)); 3587 ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA)); 3588 3589 /* 3590 * Parents of gang children can have two flavors -- ones that 3591 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set) 3592 * and ones that allocated the constituent blocks. The allocation 3593 * throttle needs to know the allocating parent zio so we must find 3594 * it here. 3595 */ 3596 if (pio->io_child_type == ZIO_CHILD_GANG) { 3597 /* 3598 * If our parent is a rewrite gang child then our grandparent 3599 * would have been the one that performed the allocation. 3600 */ 3601 if (pio->io_flags & ZIO_FLAG_IO_REWRITE) 3602 pio = zio_unique_parent(pio); 3603 flags |= METASLAB_GANG_CHILD; 3604 } 3605 3606 ASSERT(IO_IS_ALLOCATING(pio)); 3607 ASSERT3P(zio, !=, zio->io_logical); 3608 ASSERT(zio->io_logical != NULL); 3609 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR)); 3610 ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE); 3611 3612 mutex_enter(&pio->io_lock); 3613 metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags, 3614 pio->io_allocator, B_TRUE); 3615 mutex_exit(&pio->io_lock); 3616 3617 metaslab_class_throttle_unreserve(spa_normal_class(zio->io_spa), 3618 1, pio->io_allocator, pio); 3619 3620 /* 3621 * Call into the pipeline to see if there is more work that 3622 * needs to be done. If there is work to be done it will be 3623 * dispatched to another taskq thread. 3624 */ 3625 zio_allocate_dispatch(zio->io_spa, pio->io_allocator); 3626 } 3627 3628 static int 3629 zio_done(zio_t *zio) 3630 { 3631 spa_t *spa = zio->io_spa; 3632 zio_t *lio = zio->io_logical; 3633 blkptr_t *bp = zio->io_bp; 3634 vdev_t *vd = zio->io_vd; 3635 uint64_t psize = zio->io_size; 3636 zio_t *pio, *pio_next; 3637 metaslab_class_t *mc = spa_normal_class(spa); 3638 zio_link_t *zl = NULL; 3639 3640 /* 3641 * If our children haven't all completed, 3642 * wait for them and then repeat this pipeline stage. 3643 */ 3644 if (zio_wait_for_children(zio, ZIO_CHILD_ALL_BITS, ZIO_WAIT_DONE)) { 3645 return (ZIO_PIPELINE_STOP); 3646 } 3647 3648 /* 3649 * If the allocation throttle is enabled, then update the accounting. 3650 * We only track child I/Os that are part of an allocating async 3651 * write. We must do this since the allocation is performed 3652 * by the logical I/O but the actual write is done by child I/Os. 3653 */ 3654 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING && 3655 zio->io_child_type == ZIO_CHILD_VDEV) { 3656 ASSERT(mc->mc_alloc_throttle_enabled); 3657 zio_dva_throttle_done(zio); 3658 } 3659 3660 /* 3661 * If the allocation throttle is enabled, verify that 3662 * we have decremented the refcounts for every I/O that was throttled. 3663 */ 3664 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) { 3665 ASSERT(zio->io_type == ZIO_TYPE_WRITE); 3666 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); 3667 ASSERT(bp != NULL); 3668 metaslab_group_alloc_verify(spa, zio->io_bp, zio, 3669 zio->io_allocator); 3670 VERIFY(refcount_not_held(&mc->mc_alloc_slots[zio->io_allocator], 3671 zio)); 3672 } 3673 3674 for (int c = 0; c < ZIO_CHILD_TYPES; c++) 3675 for (int w = 0; w < ZIO_WAIT_TYPES; w++) 3676 ASSERT(zio->io_children[c][w] == 0); 3677 3678 if (bp != NULL && !BP_IS_EMBEDDED(bp)) { 3679 ASSERT(bp->blk_pad[0] == 0); 3680 ASSERT(bp->blk_pad[1] == 0); 3681 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 || 3682 (bp == zio_unique_parent(zio)->io_bp)); 3683 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) && 3684 zio->io_bp_override == NULL && 3685 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) { 3686 ASSERT(!BP_SHOULD_BYTESWAP(bp)); 3687 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp)); 3688 ASSERT(BP_COUNT_GANG(bp) == 0 || 3689 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp))); 3690 } 3691 if (zio->io_flags & ZIO_FLAG_NOPWRITE) 3692 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig)); 3693 } 3694 3695 /* 3696 * If there were child vdev/gang/ddt errors, they apply to us now. 3697 */ 3698 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV); 3699 zio_inherit_child_errors(zio, ZIO_CHILD_GANG); 3700 zio_inherit_child_errors(zio, ZIO_CHILD_DDT); 3701 3702 /* 3703 * If the I/O on the transformed data was successful, generate any 3704 * checksum reports now while we still have the transformed data. 3705 */ 3706 if (zio->io_error == 0) { 3707 while (zio->io_cksum_report != NULL) { 3708 zio_cksum_report_t *zcr = zio->io_cksum_report; 3709 uint64_t align = zcr->zcr_align; 3710 uint64_t asize = P2ROUNDUP(psize, align); 3711 char *abuf = NULL; 3712 abd_t *adata = zio->io_abd; 3713 3714 if (asize != psize) { 3715 adata = abd_alloc_linear(asize, B_TRUE); 3716 abd_copy(adata, zio->io_abd, psize); 3717 abd_zero_off(adata, psize, asize - psize); 3718 } 3719 3720 if (adata != NULL) 3721 abuf = abd_borrow_buf_copy(adata, asize); 3722 3723 zio->io_cksum_report = zcr->zcr_next; 3724 zcr->zcr_next = NULL; 3725 zcr->zcr_finish(zcr, abuf); 3726 zfs_ereport_free_checksum(zcr); 3727 3728 if (adata != NULL) 3729 abd_return_buf(adata, abuf, asize); 3730 3731 if (asize != psize) 3732 abd_free(adata); 3733 } 3734 } 3735 3736 zio_pop_transforms(zio); /* note: may set zio->io_error */ 3737 3738 vdev_stat_update(zio, psize); 3739 3740 if (zio->io_error) { 3741 /* 3742 * If this I/O is attached to a particular vdev, 3743 * generate an error message describing the I/O failure 3744 * at the block level. We ignore these errors if the 3745 * device is currently unavailable. 3746 */ 3747 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd)) 3748 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0); 3749 3750 if ((zio->io_error == EIO || !(zio->io_flags & 3751 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) && 3752 zio == lio) { 3753 /* 3754 * For logical I/O requests, tell the SPA to log the 3755 * error and generate a logical data ereport. 3756 */ 3757 spa_log_error(spa, zio); 3758 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio, 3759 0, 0); 3760 } 3761 } 3762 3763 if (zio->io_error && zio == lio) { 3764 /* 3765 * Determine whether zio should be reexecuted. This will 3766 * propagate all the way to the root via zio_notify_parent(). 3767 */ 3768 ASSERT(vd == NULL && bp != NULL); 3769 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 3770 3771 if (IO_IS_ALLOCATING(zio) && 3772 !(zio->io_flags & ZIO_FLAG_CANFAIL)) { 3773 if (zio->io_error != ENOSPC) 3774 zio->io_reexecute |= ZIO_REEXECUTE_NOW; 3775 else 3776 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 3777 } 3778 3779 if ((zio->io_type == ZIO_TYPE_READ || 3780 zio->io_type == ZIO_TYPE_FREE) && 3781 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && 3782 zio->io_error == ENXIO && 3783 spa_load_state(spa) == SPA_LOAD_NONE && 3784 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE) 3785 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 3786 3787 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute) 3788 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 3789 3790 /* 3791 * Here is a possibly good place to attempt to do 3792 * either combinatorial reconstruction or error correction 3793 * based on checksums. It also might be a good place 3794 * to send out preliminary ereports before we suspend 3795 * processing. 3796 */ 3797 } 3798 3799 /* 3800 * If there were logical child errors, they apply to us now. 3801 * We defer this until now to avoid conflating logical child 3802 * errors with errors that happened to the zio itself when 3803 * updating vdev stats and reporting FMA events above. 3804 */ 3805 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL); 3806 3807 if ((zio->io_error || zio->io_reexecute) && 3808 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio && 3809 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE))) 3810 zio_dva_unallocate(zio, zio->io_gang_tree, bp); 3811 3812 zio_gang_tree_free(&zio->io_gang_tree); 3813 3814 /* 3815 * Godfather I/Os should never suspend. 3816 */ 3817 if ((zio->io_flags & ZIO_FLAG_GODFATHER) && 3818 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) 3819 zio->io_reexecute = 0; 3820 3821 if (zio->io_reexecute) { 3822 /* 3823 * This is a logical I/O that wants to reexecute. 3824 * 3825 * Reexecute is top-down. When an i/o fails, if it's not 3826 * the root, it simply notifies its parent and sticks around. 3827 * The parent, seeing that it still has children in zio_done(), 3828 * does the same. This percolates all the way up to the root. 3829 * The root i/o will reexecute or suspend the entire tree. 3830 * 3831 * This approach ensures that zio_reexecute() honors 3832 * all the original i/o dependency relationships, e.g. 3833 * parents not executing until children are ready. 3834 */ 3835 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 3836 3837 zio->io_gang_leader = NULL; 3838 3839 mutex_enter(&zio->io_lock); 3840 zio->io_state[ZIO_WAIT_DONE] = 1; 3841 mutex_exit(&zio->io_lock); 3842 3843 /* 3844 * "The Godfather" I/O monitors its children but is 3845 * not a true parent to them. It will track them through 3846 * the pipeline but severs its ties whenever they get into 3847 * trouble (e.g. suspended). This allows "The Godfather" 3848 * I/O to return status without blocking. 3849 */ 3850 zl = NULL; 3851 for (pio = zio_walk_parents(zio, &zl); pio != NULL; 3852 pio = pio_next) { 3853 zio_link_t *remove_zl = zl; 3854 pio_next = zio_walk_parents(zio, &zl); 3855 3856 if ((pio->io_flags & ZIO_FLAG_GODFATHER) && 3857 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) { 3858 zio_remove_child(pio, zio, remove_zl); 3859 zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 3860 } 3861 } 3862 3863 if ((pio = zio_unique_parent(zio)) != NULL) { 3864 /* 3865 * We're not a root i/o, so there's nothing to do 3866 * but notify our parent. Don't propagate errors 3867 * upward since we haven't permanently failed yet. 3868 */ 3869 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); 3870 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE; 3871 zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 3872 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) { 3873 /* 3874 * We'd fail again if we reexecuted now, so suspend 3875 * until conditions improve (e.g. device comes online). 3876 */ 3877 zio_suspend(spa, zio); 3878 } else { 3879 /* 3880 * Reexecution is potentially a huge amount of work. 3881 * Hand it off to the otherwise-unused claim taskq. 3882 */ 3883 ASSERT(zio->io_tqent.tqent_next == NULL); 3884 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM, 3885 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio, 3886 0, &zio->io_tqent); 3887 } 3888 return (ZIO_PIPELINE_STOP); 3889 } 3890 3891 ASSERT(zio->io_child_count == 0); 3892 ASSERT(zio->io_reexecute == 0); 3893 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL)); 3894 3895 /* 3896 * Report any checksum errors, since the I/O is complete. 3897 */ 3898 while (zio->io_cksum_report != NULL) { 3899 zio_cksum_report_t *zcr = zio->io_cksum_report; 3900 zio->io_cksum_report = zcr->zcr_next; 3901 zcr->zcr_next = NULL; 3902 zcr->zcr_finish(zcr, NULL); 3903 zfs_ereport_free_checksum(zcr); 3904 } 3905 3906 /* 3907 * It is the responsibility of the done callback to ensure that this 3908 * particular zio is no longer discoverable for adoption, and as 3909 * such, cannot acquire any new parents. 3910 */ 3911 if (zio->io_done) 3912 zio->io_done(zio); 3913 3914 mutex_enter(&zio->io_lock); 3915 zio->io_state[ZIO_WAIT_DONE] = 1; 3916 mutex_exit(&zio->io_lock); 3917 3918 zl = NULL; 3919 for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) { 3920 zio_link_t *remove_zl = zl; 3921 pio_next = zio_walk_parents(zio, &zl); 3922 zio_remove_child(pio, zio, remove_zl); 3923 zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 3924 } 3925 3926 if (zio->io_waiter != NULL) { 3927 mutex_enter(&zio->io_lock); 3928 zio->io_executor = NULL; 3929 cv_broadcast(&zio->io_cv); 3930 mutex_exit(&zio->io_lock); 3931 } else { 3932 zio_destroy(zio); 3933 } 3934 3935 return (ZIO_PIPELINE_STOP); 3936 } 3937 3938 /* 3939 * ========================================================================== 3940 * I/O pipeline definition 3941 * ========================================================================== 3942 */ 3943 static zio_pipe_stage_t *zio_pipeline[] = { 3944 NULL, 3945 zio_read_bp_init, 3946 zio_write_bp_init, 3947 zio_free_bp_init, 3948 zio_issue_async, 3949 zio_write_compress, 3950 zio_checksum_generate, 3951 zio_nop_write, 3952 zio_ddt_read_start, 3953 zio_ddt_read_done, 3954 zio_ddt_write, 3955 zio_ddt_free, 3956 zio_gang_assemble, 3957 zio_gang_issue, 3958 zio_dva_throttle, 3959 zio_dva_allocate, 3960 zio_dva_free, 3961 zio_dva_claim, 3962 zio_ready, 3963 zio_vdev_io_start, 3964 zio_vdev_io_done, 3965 zio_vdev_io_assess, 3966 zio_checksum_verify, 3967 zio_done 3968 }; 3969 3970 3971 3972 3973 /* 3974 * Compare two zbookmark_phys_t's to see which we would reach first in a 3975 * pre-order traversal of the object tree. 3976 * 3977 * This is simple in every case aside from the meta-dnode object. For all other 3978 * objects, we traverse them in order (object 1 before object 2, and so on). 3979 * However, all of these objects are traversed while traversing object 0, since 3980 * the data it points to is the list of objects. Thus, we need to convert to a 3981 * canonical representation so we can compare meta-dnode bookmarks to 3982 * non-meta-dnode bookmarks. 3983 * 3984 * We do this by calculating "equivalents" for each field of the zbookmark. 3985 * zbookmarks outside of the meta-dnode use their own object and level, and 3986 * calculate the level 0 equivalent (the first L0 blkid that is contained in the 3987 * blocks this bookmark refers to) by multiplying their blkid by their span 3988 * (the number of L0 blocks contained within one block at their level). 3989 * zbookmarks inside the meta-dnode calculate their object equivalent 3990 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use 3991 * level + 1<<31 (any value larger than a level could ever be) for their level. 3992 * This causes them to always compare before a bookmark in their object 3993 * equivalent, compare appropriately to bookmarks in other objects, and to 3994 * compare appropriately to other bookmarks in the meta-dnode. 3995 */ 3996 int 3997 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2, 3998 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2) 3999 { 4000 /* 4001 * These variables represent the "equivalent" values for the zbookmark, 4002 * after converting zbookmarks inside the meta dnode to their 4003 * normal-object equivalents. 4004 */ 4005 uint64_t zb1obj, zb2obj; 4006 uint64_t zb1L0, zb2L0; 4007 uint64_t zb1level, zb2level; 4008 4009 if (zb1->zb_object == zb2->zb_object && 4010 zb1->zb_level == zb2->zb_level && 4011 zb1->zb_blkid == zb2->zb_blkid) 4012 return (0); 4013 4014 /* 4015 * BP_SPANB calculates the span in blocks. 4016 */ 4017 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level); 4018 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level); 4019 4020 if (zb1->zb_object == DMU_META_DNODE_OBJECT) { 4021 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT)); 4022 zb1L0 = 0; 4023 zb1level = zb1->zb_level + COMPARE_META_LEVEL; 4024 } else { 4025 zb1obj = zb1->zb_object; 4026 zb1level = zb1->zb_level; 4027 } 4028 4029 if (zb2->zb_object == DMU_META_DNODE_OBJECT) { 4030 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT)); 4031 zb2L0 = 0; 4032 zb2level = zb2->zb_level + COMPARE_META_LEVEL; 4033 } else { 4034 zb2obj = zb2->zb_object; 4035 zb2level = zb2->zb_level; 4036 } 4037 4038 /* Now that we have a canonical representation, do the comparison. */ 4039 if (zb1obj != zb2obj) 4040 return (zb1obj < zb2obj ? -1 : 1); 4041 else if (zb1L0 != zb2L0) 4042 return (zb1L0 < zb2L0 ? -1 : 1); 4043 else if (zb1level != zb2level) 4044 return (zb1level > zb2level ? -1 : 1); 4045 /* 4046 * This can (theoretically) happen if the bookmarks have the same object 4047 * and level, but different blkids, if the block sizes are not the same. 4048 * There is presently no way to change the indirect block sizes 4049 */ 4050 return (0); 4051 } 4052 4053 /* 4054 * This function checks the following: given that last_block is the place that 4055 * our traversal stopped last time, does that guarantee that we've visited 4056 * every node under subtree_root? Therefore, we can't just use the raw output 4057 * of zbookmark_compare. We have to pass in a modified version of 4058 * subtree_root; by incrementing the block id, and then checking whether 4059 * last_block is before or equal to that, we can tell whether or not having 4060 * visited last_block implies that all of subtree_root's children have been 4061 * visited. 4062 */ 4063 boolean_t 4064 zbookmark_subtree_completed(const dnode_phys_t *dnp, 4065 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block) 4066 { 4067 zbookmark_phys_t mod_zb = *subtree_root; 4068 mod_zb.zb_blkid++; 4069 ASSERT(last_block->zb_level == 0); 4070 4071 /* The objset_phys_t isn't before anything. */ 4072 if (dnp == NULL) 4073 return (B_FALSE); 4074 4075 /* 4076 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the 4077 * data block size in sectors, because that variable is only used if 4078 * the bookmark refers to a block in the meta-dnode. Since we don't 4079 * know without examining it what object it refers to, and there's no 4080 * harm in passing in this value in other cases, we always pass it in. 4081 * 4082 * We pass in 0 for the indirect block size shift because zb2 must be 4083 * level 0. The indirect block size is only used to calculate the span 4084 * of the bookmark, but since the bookmark must be level 0, the span is 4085 * always 1, so the math works out. 4086 * 4087 * If you make changes to how the zbookmark_compare code works, be sure 4088 * to make sure that this code still works afterwards. 4089 */ 4090 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift, 4091 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb, 4092 last_block) <= 0); 4093 } 4094