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) 2013 by Delphix. All rights reserved.
  24  * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
  25  */
  26 
  27 #include <sys/zfs_context.h>
  28 #include <sys/fm/fs/zfs.h>
  29 #include <sys/spa.h>
  30 #include <sys/txg.h>
  31 #include <sys/spa_impl.h>
  32 #include <sys/vdev_impl.h>
  33 #include <sys/zio_impl.h>
  34 #include <sys/zio_compress.h>
  35 #include <sys/zio_checksum.h>
  36 #include <sys/dmu_objset.h>
  37 #include <sys/arc.h>
  38 #include <sys/ddt.h>
  39 
  40 /*
  41  * ==========================================================================
  42  * I/O type descriptions
  43  * ==========================================================================
  44  */
  45 const char *zio_type_name[ZIO_TYPES] = {
  46         "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
  47         "zio_ioctl"
  48 };
  49 
  50 /*
  51  * ==========================================================================
  52  * I/O kmem caches
  53  * ==========================================================================
  54  */
  55 kmem_cache_t *zio_cache;
  56 kmem_cache_t *zio_link_cache;
  57 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
  58 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
  59 
  60 #ifdef _KERNEL
  61 extern vmem_t *zio_alloc_arena;
  62 #endif
  63 extern int zfs_mg_alloc_failures;
  64 
  65 /*
  66  * The following actions directly effect the spa's sync-to-convergence logic.
  67  * The values below define the sync pass when we start performing the action.
  68  * Care should be taken when changing these values as they directly impact
  69  * spa_sync() performance. Tuning these values may introduce subtle performance
  70  * pathologies and should only be done in the context of performance analysis.
  71  * These tunables will eventually be removed and replaced with #defines once
  72  * enough analysis has been done to determine optimal values.
  73  *
  74  * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
  75  * regular blocks are not deferred.
  76  */
  77 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
  78 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
  79 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
  80 
  81 /*
  82  * An allocating zio is one that either currently has the DVA allocate
  83  * stage set or will have it later in its lifetime.
  84  */
  85 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
  86 
  87 boolean_t       zio_requeue_io_start_cut_in_line = B_TRUE;
  88 
  89 #ifdef ZFS_DEBUG
  90 int zio_buf_debug_limit = 16384;
  91 #else
  92 int zio_buf_debug_limit = 0;
  93 #endif
  94 
  95 void
  96 zio_init(void)
  97 {
  98         size_t c;
  99         vmem_t *data_alloc_arena = NULL;
 100 
 101 #ifdef _KERNEL
 102         data_alloc_arena = zio_alloc_arena;
 103 #endif
 104         zio_cache = kmem_cache_create("zio_cache",
 105             sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
 106         zio_link_cache = kmem_cache_create("zio_link_cache",
 107             sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
 108 
 109         /*
 110          * For small buffers, we want a cache for each multiple of
 111          * SPA_MINBLOCKSIZE.  For medium-size buffers, we want a cache
 112          * for each quarter-power of 2.  For large buffers, we want
 113          * a cache for each multiple of PAGESIZE.
 114          */
 115         for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
 116                 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
 117                 size_t p2 = size;
 118                 size_t align = 0;
 119                 size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
 120 
 121                 while (p2 & (p2 - 1))
 122                         p2 &= p2 - 1;
 123 
 124 #ifndef _KERNEL
 125                 /*
 126                  * If we are using watchpoints, put each buffer on its own page,
 127                  * to eliminate the performance overhead of trapping to the
 128                  * kernel when modifying a non-watched buffer that shares the
 129                  * page with a watched buffer.
 130                  */
 131                 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
 132                         continue;
 133 #endif
 134                 if (size <= 4 * SPA_MINBLOCKSIZE) {
 135                         align = SPA_MINBLOCKSIZE;
 136                 } else if (IS_P2ALIGNED(size, PAGESIZE)) {
 137                         align = PAGESIZE;
 138                 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
 139                         align = p2 >> 2;
 140                 }
 141 
 142                 if (align != 0) {
 143                         char name[36];
 144                         (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
 145                         zio_buf_cache[c] = kmem_cache_create(name, size,
 146                             align, NULL, NULL, NULL, NULL, NULL, cflags);
 147 
 148                         /*
 149                          * Since zio_data bufs do not appear in crash dumps, we
 150                          * pass KMC_NOTOUCH so that no allocator metadata is
 151                          * stored with the buffers.
 152                          */
 153                         (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
 154                         zio_data_buf_cache[c] = kmem_cache_create(name, size,
 155                             align, NULL, NULL, NULL, NULL, data_alloc_arena,
 156                             cflags | KMC_NOTOUCH);
 157                 }
 158         }
 159 
 160         while (--c != 0) {
 161                 ASSERT(zio_buf_cache[c] != NULL);
 162                 if (zio_buf_cache[c - 1] == NULL)
 163                         zio_buf_cache[c - 1] = zio_buf_cache[c];
 164 
 165                 ASSERT(zio_data_buf_cache[c] != NULL);
 166                 if (zio_data_buf_cache[c - 1] == NULL)
 167                         zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
 168         }
 169 
 170         /*
 171          * The zio write taskqs have 1 thread per cpu, allow 1/2 of the taskqs
 172          * to fail 3 times per txg or 8 failures, whichever is greater.
 173          */
 174         zfs_mg_alloc_failures = MAX((3 * max_ncpus / 2), 8);
 175 
 176         zio_inject_init();
 177 }
 178 
 179 void
 180 zio_fini(void)
 181 {
 182         size_t c;
 183         kmem_cache_t *last_cache = NULL;
 184         kmem_cache_t *last_data_cache = NULL;
 185 
 186         for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
 187                 if (zio_buf_cache[c] != last_cache) {
 188                         last_cache = zio_buf_cache[c];
 189                         kmem_cache_destroy(zio_buf_cache[c]);
 190                 }
 191                 zio_buf_cache[c] = NULL;
 192 
 193                 if (zio_data_buf_cache[c] != last_data_cache) {
 194                         last_data_cache = zio_data_buf_cache[c];
 195                         kmem_cache_destroy(zio_data_buf_cache[c]);
 196                 }
 197                 zio_data_buf_cache[c] = NULL;
 198         }
 199 
 200         kmem_cache_destroy(zio_link_cache);
 201         kmem_cache_destroy(zio_cache);
 202 
 203         zio_inject_fini();
 204 }
 205 
 206 /*
 207  * ==========================================================================
 208  * Allocate and free I/O buffers
 209  * ==========================================================================
 210  */
 211 
 212 /*
 213  * Use zio_buf_alloc to allocate ZFS metadata.  This data will appear in a
 214  * crashdump if the kernel panics, so use it judiciously.  Obviously, it's
 215  * useful to inspect ZFS metadata, but if possible, we should avoid keeping
 216  * excess / transient data in-core during a crashdump.
 217  */
 218 void *
 219 zio_buf_alloc(size_t size)
 220 {
 221         size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
 222 
 223         ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
 224 
 225         return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
 226 }
 227 
 228 /*
 229  * Use zio_data_buf_alloc to allocate data.  The data will not appear in a
 230  * crashdump if the kernel panics.  This exists so that we will limit the amount
 231  * of ZFS data that shows up in a kernel crashdump.  (Thus reducing the amount
 232  * of kernel heap dumped to disk when the kernel panics)
 233  */
 234 void *
 235 zio_data_buf_alloc(size_t size)
 236 {
 237         size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
 238 
 239         ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
 240 
 241         return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
 242 }
 243 
 244 void
 245 zio_buf_free(void *buf, size_t size)
 246 {
 247         size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
 248 
 249         ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
 250 
 251         kmem_cache_free(zio_buf_cache[c], buf);
 252 }
 253 
 254 void
 255 zio_data_buf_free(void *buf, size_t size)
 256 {
 257         size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
 258 
 259         ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
 260 
 261         kmem_cache_free(zio_data_buf_cache[c], buf);
 262 }
 263 
 264 /*
 265  * ==========================================================================
 266  * Push and pop I/O transform buffers
 267  * ==========================================================================
 268  */
 269 static void
 270 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
 271         zio_transform_func_t *transform)
 272 {
 273         zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
 274 
 275         zt->zt_orig_data = zio->io_data;
 276         zt->zt_orig_size = zio->io_size;
 277         zt->zt_bufsize = bufsize;
 278         zt->zt_transform = transform;
 279 
 280         zt->zt_next = zio->io_transform_stack;
 281         zio->io_transform_stack = zt;
 282 
 283         zio->io_data = data;
 284         zio->io_size = size;
 285 }
 286 
 287 static void
 288 zio_pop_transforms(zio_t *zio)
 289 {
 290         zio_transform_t *zt;
 291 
 292         while ((zt = zio->io_transform_stack) != NULL) {
 293                 if (zt->zt_transform != NULL)
 294                         zt->zt_transform(zio,
 295                             zt->zt_orig_data, zt->zt_orig_size);
 296 
 297                 if (zt->zt_bufsize != 0)
 298                         zio_buf_free(zio->io_data, zt->zt_bufsize);
 299 
 300                 zio->io_data = zt->zt_orig_data;
 301                 zio->io_size = zt->zt_orig_size;
 302                 zio->io_transform_stack = zt->zt_next;
 303 
 304                 kmem_free(zt, sizeof (zio_transform_t));
 305         }
 306 }
 307 
 308 /*
 309  * ==========================================================================
 310  * I/O transform callbacks for subblocks and decompression
 311  * ==========================================================================
 312  */
 313 static void
 314 zio_subblock(zio_t *zio, void *data, uint64_t size)
 315 {
 316         ASSERT(zio->io_size > size);
 317 
 318         if (zio->io_type == ZIO_TYPE_READ)
 319                 bcopy(zio->io_data, data, size);
 320 }
 321 
 322 static void
 323 zio_decompress(zio_t *zio, void *data, uint64_t size)
 324 {
 325         if (zio->io_error == 0 &&
 326             zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
 327             zio->io_data, data, zio->io_size, size) != 0)
 328                 zio->io_error = SET_ERROR(EIO);
 329 }
 330 
 331 /*
 332  * ==========================================================================
 333  * I/O parent/child relationships and pipeline interlocks
 334  * ==========================================================================
 335  */
 336 /*
 337  * NOTE - Callers to zio_walk_parents() and zio_walk_children must
 338  *        continue calling these functions until they return NULL.
 339  *        Otherwise, the next caller will pick up the list walk in
 340  *        some indeterminate state.  (Otherwise every caller would
 341  *        have to pass in a cookie to keep the state represented by
 342  *        io_walk_link, which gets annoying.)
 343  */
 344 zio_t *
 345 zio_walk_parents(zio_t *cio)
 346 {
 347         zio_link_t *zl = cio->io_walk_link;
 348         list_t *pl = &cio->io_parent_list;
 349 
 350         zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
 351         cio->io_walk_link = zl;
 352 
 353         if (zl == NULL)
 354                 return (NULL);
 355 
 356         ASSERT(zl->zl_child == cio);
 357         return (zl->zl_parent);
 358 }
 359 
 360 zio_t *
 361 zio_walk_children(zio_t *pio)
 362 {
 363         zio_link_t *zl = pio->io_walk_link;
 364         list_t *cl = &pio->io_child_list;
 365 
 366         zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
 367         pio->io_walk_link = zl;
 368 
 369         if (zl == NULL)
 370                 return (NULL);
 371 
 372         ASSERT(zl->zl_parent == pio);
 373         return (zl->zl_child);
 374 }
 375 
 376 zio_t *
 377 zio_unique_parent(zio_t *cio)
 378 {
 379         zio_t *pio = zio_walk_parents(cio);
 380 
 381         VERIFY(zio_walk_parents(cio) == NULL);
 382         return (pio);
 383 }
 384 
 385 void
 386 zio_add_child(zio_t *pio, zio_t *cio)
 387 {
 388         zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
 389 
 390         /*
 391          * Logical I/Os can have logical, gang, or vdev children.
 392          * Gang I/Os can have gang or vdev children.
 393          * Vdev I/Os can only have vdev children.
 394          * The following ASSERT captures all of these constraints.
 395          */
 396         ASSERT(cio->io_child_type <= pio->io_child_type);
 397 
 398         zl->zl_parent = pio;
 399         zl->zl_child = cio;
 400 
 401         mutex_enter(&cio->io_lock);
 402         mutex_enter(&pio->io_lock);
 403 
 404         ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
 405 
 406         for (int w = 0; w < ZIO_WAIT_TYPES; w++)
 407                 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
 408 
 409         list_insert_head(&pio->io_child_list, zl);
 410         list_insert_head(&cio->io_parent_list, zl);
 411 
 412         pio->io_child_count++;
 413         cio->io_parent_count++;
 414 
 415         mutex_exit(&pio->io_lock);
 416         mutex_exit(&cio->io_lock);
 417 }
 418 
 419 static void
 420 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
 421 {
 422         ASSERT(zl->zl_parent == pio);
 423         ASSERT(zl->zl_child == cio);
 424 
 425         mutex_enter(&cio->io_lock);
 426         mutex_enter(&pio->io_lock);
 427 
 428         list_remove(&pio->io_child_list, zl);
 429         list_remove(&cio->io_parent_list, zl);
 430 
 431         pio->io_child_count--;
 432         cio->io_parent_count--;
 433 
 434         mutex_exit(&pio->io_lock);
 435         mutex_exit(&cio->io_lock);
 436 
 437         kmem_cache_free(zio_link_cache, zl);
 438 }
 439 
 440 static boolean_t
 441 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
 442 {
 443         uint64_t *countp = &zio->io_children[child][wait];
 444         boolean_t waiting = B_FALSE;
 445 
 446         mutex_enter(&zio->io_lock);
 447         ASSERT(zio->io_stall == NULL);
 448         if (*countp != 0) {
 449                 zio->io_stage >>= 1;
 450                 zio->io_stall = countp;
 451                 waiting = B_TRUE;
 452         }
 453         mutex_exit(&zio->io_lock);
 454 
 455         return (waiting);
 456 }
 457 
 458 static void
 459 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
 460 {
 461         uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
 462         int *errorp = &pio->io_child_error[zio->io_child_type];
 463 
 464         mutex_enter(&pio->io_lock);
 465         if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
 466                 *errorp = zio_worst_error(*errorp, zio->io_error);
 467         pio->io_reexecute |= zio->io_reexecute;
 468         ASSERT3U(*countp, >, 0);
 469 
 470         (*countp)--;
 471 
 472         if (*countp == 0 && pio->io_stall == countp) {
 473                 pio->io_stall = NULL;
 474                 mutex_exit(&pio->io_lock);
 475                 zio_execute(pio);
 476         } else {
 477                 mutex_exit(&pio->io_lock);
 478         }
 479 }
 480 
 481 static void
 482 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
 483 {
 484         if (zio->io_child_error[c] != 0 && zio->io_error == 0)
 485                 zio->io_error = zio->io_child_error[c];
 486 }
 487 
 488 /*
 489  * ==========================================================================
 490  * Create the various types of I/O (read, write, free, etc)
 491  * ==========================================================================
 492  */
 493 static zio_t *
 494 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
 495     void *data, uint64_t size, zio_done_func_t *done, void *private,
 496     zio_type_t type, zio_priority_t priority, enum zio_flag flags,
 497     vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
 498     enum zio_stage stage, enum zio_stage pipeline)
 499 {
 500         zio_t *zio;
 501 
 502         ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
 503         ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
 504         ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
 505 
 506         ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
 507         ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
 508         ASSERT(vd || stage == ZIO_STAGE_OPEN);
 509 
 510         zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
 511         bzero(zio, sizeof (zio_t));
 512 
 513         mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
 514         cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
 515 
 516         list_create(&zio->io_parent_list, sizeof (zio_link_t),
 517             offsetof(zio_link_t, zl_parent_node));
 518         list_create(&zio->io_child_list, sizeof (zio_link_t),
 519             offsetof(zio_link_t, zl_child_node));
 520 
 521         if (vd != NULL)
 522                 zio->io_child_type = ZIO_CHILD_VDEV;
 523         else if (flags & ZIO_FLAG_GANG_CHILD)
 524                 zio->io_child_type = ZIO_CHILD_GANG;
 525         else if (flags & ZIO_FLAG_DDT_CHILD)
 526                 zio->io_child_type = ZIO_CHILD_DDT;
 527         else
 528                 zio->io_child_type = ZIO_CHILD_LOGICAL;
 529 
 530         if (bp != NULL) {
 531                 zio->io_bp = (blkptr_t *)bp;
 532                 zio->io_bp_copy = *bp;
 533                 zio->io_bp_orig = *bp;
 534                 if (type != ZIO_TYPE_WRITE ||
 535                     zio->io_child_type == ZIO_CHILD_DDT)
 536                         zio->io_bp = &zio->io_bp_copy;        /* so caller can free */
 537                 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
 538                         zio->io_logical = zio;
 539                 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
 540                         pipeline |= ZIO_GANG_STAGES;
 541         }
 542 
 543         zio->io_spa = spa;
 544         zio->io_txg = txg;
 545         zio->io_done = done;
 546         zio->io_private = private;
 547         zio->io_type = type;
 548         zio->io_priority = priority;
 549         zio->io_vd = vd;
 550         zio->io_offset = offset;
 551         zio->io_orig_data = zio->io_data = data;
 552         zio->io_orig_size = zio->io_size = size;
 553         zio->io_orig_flags = zio->io_flags = flags;
 554         zio->io_orig_stage = zio->io_stage = stage;
 555         zio->io_orig_pipeline = zio->io_pipeline = pipeline;
 556 
 557         zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
 558         zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
 559 
 560         if (zb != NULL)
 561                 zio->io_bookmark = *zb;
 562 
 563         if (pio != NULL) {
 564                 if (zio->io_logical == NULL)
 565                         zio->io_logical = pio->io_logical;
 566                 if (zio->io_child_type == ZIO_CHILD_GANG)
 567                         zio->io_gang_leader = pio->io_gang_leader;
 568                 zio_add_child(pio, zio);
 569         }
 570 
 571         return (zio);
 572 }
 573 
 574 static void
 575 zio_destroy(zio_t *zio)
 576 {
 577         list_destroy(&zio->io_parent_list);
 578         list_destroy(&zio->io_child_list);
 579         mutex_destroy(&zio->io_lock);
 580         cv_destroy(&zio->io_cv);
 581         kmem_cache_free(zio_cache, zio);
 582 }
 583 
 584 zio_t *
 585 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
 586     void *private, enum zio_flag flags)
 587 {
 588         zio_t *zio;
 589 
 590         zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
 591             ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
 592             ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
 593 
 594         return (zio);
 595 }
 596 
 597 zio_t *
 598 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
 599 {
 600         return (zio_null(NULL, spa, NULL, done, private, flags));
 601 }
 602 
 603 zio_t *
 604 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
 605     void *data, uint64_t size, zio_done_func_t *done, void *private,
 606     zio_priority_t priority, enum zio_flag flags, const zbookmark_t *zb)
 607 {
 608         zio_t *zio;
 609 
 610         zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
 611             data, size, done, private,
 612             ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
 613             ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
 614             ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
 615 
 616         return (zio);
 617 }
 618 
 619 zio_t *
 620 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
 621     void *data, uint64_t size, const zio_prop_t *zp,
 622     zio_done_func_t *ready, zio_done_func_t *physdone, zio_done_func_t *done,
 623     void *private,
 624     zio_priority_t priority, enum zio_flag flags, const zbookmark_t *zb)
 625 {
 626         zio_t *zio;
 627 
 628         ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
 629             zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
 630             zp->zp_compress >= ZIO_COMPRESS_OFF &&
 631             zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
 632             DMU_OT_IS_VALID(zp->zp_type) &&
 633             zp->zp_level < 32 &&
 634             zp->zp_copies > 0 &&
 635             zp->zp_copies <= spa_max_replication(spa));
 636 
 637         zio = zio_create(pio, spa, txg, bp, data, size, done, private,
 638             ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
 639             ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
 640             ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
 641 
 642         zio->io_ready = ready;
 643         zio->io_physdone = physdone;
 644         zio->io_prop = *zp;
 645 
 646         return (zio);
 647 }
 648 
 649 zio_t *
 650 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
 651     uint64_t size, zio_done_func_t *done, void *private,
 652     zio_priority_t priority, enum zio_flag flags, zbookmark_t *zb)
 653 {
 654         zio_t *zio;
 655 
 656         zio = zio_create(pio, spa, txg, bp, data, size, done, private,
 657             ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
 658             ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
 659 
 660         return (zio);
 661 }
 662 
 663 void
 664 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
 665 {
 666         ASSERT(zio->io_type == ZIO_TYPE_WRITE);
 667         ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
 668         ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
 669         ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
 670 
 671         /*
 672          * We must reset the io_prop to match the values that existed
 673          * when the bp was first written by dmu_sync() keeping in mind
 674          * that nopwrite and dedup are mutually exclusive.
 675          */
 676         zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
 677         zio->io_prop.zp_nopwrite = nopwrite;
 678         zio->io_prop.zp_copies = copies;
 679         zio->io_bp_override = bp;
 680 }
 681 
 682 void
 683 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
 684 {
 685         metaslab_check_free(spa, bp);
 686 
 687         /*
 688          * Frees that are for the currently-syncing txg, are not going to be
 689          * deferred, and which will not need to do a read (i.e. not GANG or
 690          * DEDUP), can be processed immediately.  Otherwise, put them on the
 691          * in-memory list for later processing.
 692          */
 693         if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
 694             txg != spa->spa_syncing_txg ||
 695             spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
 696                 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
 697         } else {
 698                 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 0)));
 699         }
 700 }
 701 
 702 zio_t *
 703 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
 704     enum zio_flag flags)
 705 {
 706         zio_t *zio;
 707         enum zio_stage stage = ZIO_FREE_PIPELINE;
 708 
 709         dprintf_bp(bp, "freeing in txg %llu, pass %u",
 710             (longlong_t)txg, spa->spa_sync_pass);
 711 
 712         ASSERT(!BP_IS_HOLE(bp));
 713         ASSERT(spa_syncing_txg(spa) == txg);
 714         ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
 715 
 716         metaslab_check_free(spa, bp);
 717         arc_freed(spa, bp);
 718 
 719         /*
 720          * GANG and DEDUP blocks can induce a read (for the gang block header,
 721          * or the DDT), so issue them asynchronously so that this thread is
 722          * not tied up.
 723          */
 724         if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
 725                 stage |= ZIO_STAGE_ISSUE_ASYNC;
 726 
 727         zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
 728             NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags,
 729             NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
 730 
 731 
 732         return (zio);
 733 }
 734 
 735 zio_t *
 736 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
 737     zio_done_func_t *done, void *private, enum zio_flag flags)
 738 {
 739         zio_t *zio;
 740 
 741         /*
 742          * A claim is an allocation of a specific block.  Claims are needed
 743          * to support immediate writes in the intent log.  The issue is that
 744          * immediate writes contain committed data, but in a txg that was
 745          * *not* committed.  Upon opening the pool after an unclean shutdown,
 746          * the intent log claims all blocks that contain immediate write data
 747          * so that the SPA knows they're in use.
 748          *
 749          * All claims *must* be resolved in the first txg -- before the SPA
 750          * starts allocating blocks -- so that nothing is allocated twice.
 751          * If txg == 0 we just verify that the block is claimable.
 752          */
 753         ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
 754         ASSERT(txg == spa_first_txg(spa) || txg == 0);
 755         ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa));       /* zdb(1M) */
 756 
 757         zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
 758             done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
 759             NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
 760 
 761         return (zio);
 762 }
 763 
 764 zio_t *
 765 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
 766     zio_done_func_t *done, void *private, enum zio_flag flags)
 767 {
 768         zio_t *zio;
 769         int c;
 770 
 771         if (vd->vdev_children == 0) {
 772                 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
 773                     ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
 774                     ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
 775 
 776                 zio->io_cmd = cmd;
 777         } else {
 778                 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
 779 
 780                 for (c = 0; c < vd->vdev_children; c++)
 781                         zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
 782                             done, private, flags));
 783         }
 784 
 785         return (zio);
 786 }
 787 
 788 zio_t *
 789 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
 790     void *data, int checksum, zio_done_func_t *done, void *private,
 791     zio_priority_t priority, enum zio_flag flags, boolean_t labels)
 792 {
 793         zio_t *zio;
 794 
 795         ASSERT(vd->vdev_children == 0);
 796         ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
 797             offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
 798         ASSERT3U(offset + size, <=, vd->vdev_psize);
 799 
 800         zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
 801             ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
 802             ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
 803 
 804         zio->io_prop.zp_checksum = checksum;
 805 
 806         return (zio);
 807 }
 808 
 809 zio_t *
 810 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
 811     void *data, int checksum, zio_done_func_t *done, void *private,
 812     zio_priority_t priority, enum zio_flag flags, boolean_t labels)
 813 {
 814         zio_t *zio;
 815 
 816         ASSERT(vd->vdev_children == 0);
 817         ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
 818             offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
 819         ASSERT3U(offset + size, <=, vd->vdev_psize);
 820 
 821         zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
 822             ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
 823             ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
 824 
 825         zio->io_prop.zp_checksum = checksum;
 826 
 827         if (zio_checksum_table[checksum].ci_eck) {
 828                 /*
 829                  * zec checksums are necessarily destructive -- they modify
 830                  * the end of the write buffer to hold the verifier/checksum.
 831                  * Therefore, we must make a local copy in case the data is
 832                  * being written to multiple places in parallel.
 833                  */
 834                 void *wbuf = zio_buf_alloc(size);
 835                 bcopy(data, wbuf, size);
 836                 zio_push_transform(zio, wbuf, size, size, NULL);
 837         }
 838 
 839         return (zio);
 840 }
 841 
 842 /*
 843  * Create a child I/O to do some work for us.
 844  */
 845 zio_t *
 846 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
 847         void *data, uint64_t size, int type, zio_priority_t priority,
 848         enum zio_flag flags, zio_done_func_t *done, void *private)
 849 {
 850         enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
 851         zio_t *zio;
 852 
 853         ASSERT(vd->vdev_parent ==
 854             (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
 855 
 856         if (type == ZIO_TYPE_READ && bp != NULL) {
 857                 /*
 858                  * If we have the bp, then the child should perform the
 859                  * checksum and the parent need not.  This pushes error
 860                  * detection as close to the leaves as possible and
 861                  * eliminates redundant checksums in the interior nodes.
 862                  */
 863                 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
 864                 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
 865         }
 866 
 867         if (vd->vdev_children == 0)
 868                 offset += VDEV_LABEL_START_SIZE;
 869 
 870         flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
 871 
 872         /*
 873          * If we've decided to do a repair, the write is not speculative --
 874          * even if the original read was.
 875          */
 876         if (flags & ZIO_FLAG_IO_REPAIR)
 877                 flags &= ~ZIO_FLAG_SPECULATIVE;
 878 
 879         zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
 880             done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
 881             ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
 882 
 883         zio->io_physdone = pio->io_physdone;
 884         if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
 885                 zio->io_logical->io_phys_children++;
 886 
 887         return (zio);
 888 }
 889 
 890 zio_t *
 891 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
 892         int type, zio_priority_t priority, enum zio_flag flags,
 893         zio_done_func_t *done, void *private)
 894 {
 895         zio_t *zio;
 896 
 897         ASSERT(vd->vdev_ops->vdev_op_leaf);
 898 
 899         zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
 900             data, size, done, private, type, priority,
 901             flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
 902             vd, offset, NULL,
 903             ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
 904 
 905         return (zio);
 906 }
 907 
 908 void
 909 zio_flush(zio_t *zio, vdev_t *vd)
 910 {
 911         zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
 912             NULL, NULL,
 913             ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
 914 }
 915 
 916 void
 917 zio_shrink(zio_t *zio, uint64_t size)
 918 {
 919         ASSERT(zio->io_executor == NULL);
 920         ASSERT(zio->io_orig_size == zio->io_size);
 921         ASSERT(size <= zio->io_size);
 922 
 923         /*
 924          * We don't shrink for raidz because of problems with the
 925          * reconstruction when reading back less than the block size.
 926          * Note, BP_IS_RAIDZ() assumes no compression.
 927          */
 928         ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
 929         if (!BP_IS_RAIDZ(zio->io_bp))
 930                 zio->io_orig_size = zio->io_size = size;
 931 }
 932 
 933 /*
 934  * ==========================================================================
 935  * Prepare to read and write logical blocks
 936  * ==========================================================================
 937  */
 938 
 939 static int
 940 zio_read_bp_init(zio_t *zio)
 941 {
 942         blkptr_t *bp = zio->io_bp;
 943 
 944         if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
 945             zio->io_child_type == ZIO_CHILD_LOGICAL &&
 946             !(zio->io_flags & ZIO_FLAG_RAW)) {
 947                 uint64_t psize = BP_GET_PSIZE(bp);
 948                 void *cbuf = zio_buf_alloc(psize);
 949 
 950                 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
 951         }
 952 
 953         if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
 954                 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
 955 
 956         if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
 957                 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
 958 
 959         if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
 960                 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
 961 
 962         return (ZIO_PIPELINE_CONTINUE);
 963 }
 964 
 965 static int
 966 zio_write_bp_init(zio_t *zio)
 967 {
 968         spa_t *spa = zio->io_spa;
 969         zio_prop_t *zp = &zio->io_prop;
 970         enum zio_compress compress = zp->zp_compress;
 971         blkptr_t *bp = zio->io_bp;
 972         uint64_t lsize = zio->io_size;
 973         uint64_t psize = lsize;
 974         int pass = 1;
 975 
 976         /*
 977          * If our children haven't all reached the ready stage,
 978          * wait for them and then repeat this pipeline stage.
 979          */
 980         if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
 981             zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
 982                 return (ZIO_PIPELINE_STOP);
 983 
 984         if (!IO_IS_ALLOCATING(zio))
 985                 return (ZIO_PIPELINE_CONTINUE);
 986 
 987         ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
 988 
 989         if (zio->io_bp_override) {
 990                 ASSERT(bp->blk_birth != zio->io_txg);
 991                 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
 992 
 993                 *bp = *zio->io_bp_override;
 994                 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
 995 
 996                 /*
 997                  * If we've been overridden and nopwrite is set then
 998                  * set the flag accordingly to indicate that a nopwrite
 999                  * has already occurred.
1000                  */
1001                 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1002                         ASSERT(!zp->zp_dedup);
1003                         zio->io_flags |= ZIO_FLAG_NOPWRITE;
1004                         return (ZIO_PIPELINE_CONTINUE);
1005                 }
1006 
1007                 ASSERT(!zp->zp_nopwrite);
1008 
1009                 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1010                         return (ZIO_PIPELINE_CONTINUE);
1011 
1012                 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1013                     zp->zp_dedup_verify);
1014 
1015                 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1016                         BP_SET_DEDUP(bp, 1);
1017                         zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1018                         return (ZIO_PIPELINE_CONTINUE);
1019                 }
1020                 zio->io_bp_override = NULL;
1021                 BP_ZERO(bp);
1022         }
1023 
1024         if (bp->blk_birth == zio->io_txg) {
1025                 /*
1026                  * We're rewriting an existing block, which means we're
1027                  * working on behalf of spa_sync().  For spa_sync() to
1028                  * converge, it must eventually be the case that we don't
1029                  * have to allocate new blocks.  But compression changes
1030                  * the blocksize, which forces a reallocate, and makes
1031                  * convergence take longer.  Therefore, after the first
1032                  * few passes, stop compressing to ensure convergence.
1033                  */
1034                 pass = spa_sync_pass(spa);
1035 
1036                 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1037                 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1038                 ASSERT(!BP_GET_DEDUP(bp));
1039 
1040                 if (pass >= zfs_sync_pass_dont_compress)
1041                         compress = ZIO_COMPRESS_OFF;
1042 
1043                 /* Make sure someone doesn't change their mind on overwrites */
1044                 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
1045                     spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1046         }
1047 
1048         if (compress != ZIO_COMPRESS_OFF) {
1049                 void *cbuf = zio_buf_alloc(lsize);
1050                 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1051                 if (psize == 0 || psize == lsize) {
1052                         compress = ZIO_COMPRESS_OFF;
1053                         zio_buf_free(cbuf, lsize);
1054                 } else {
1055                         ASSERT(psize < lsize);
1056                         zio_push_transform(zio, cbuf, psize, lsize, NULL);
1057                 }
1058         }
1059 
1060         /*
1061          * The final pass of spa_sync() must be all rewrites, but the first
1062          * few passes offer a trade-off: allocating blocks defers convergence,
1063          * but newly allocated blocks are sequential, so they can be written
1064          * to disk faster.  Therefore, we allow the first few passes of
1065          * spa_sync() to allocate new blocks, but force rewrites after that.
1066          * There should only be a handful of blocks after pass 1 in any case.
1067          */
1068         if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
1069             pass >= zfs_sync_pass_rewrite) {
1070                 ASSERT(psize != 0);
1071                 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1072                 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1073                 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1074         } else {
1075                 BP_ZERO(bp);
1076                 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1077         }
1078 
1079         if (psize == 0) {
1080                 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1081         } else {
1082                 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1083                 BP_SET_LSIZE(bp, lsize);
1084                 BP_SET_PSIZE(bp, psize);
1085                 BP_SET_COMPRESS(bp, compress);
1086                 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1087                 BP_SET_TYPE(bp, zp->zp_type);
1088                 BP_SET_LEVEL(bp, zp->zp_level);
1089                 BP_SET_DEDUP(bp, zp->zp_dedup);
1090                 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1091                 if (zp->zp_dedup) {
1092                         ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1093                         ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1094                         zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1095                 }
1096                 if (zp->zp_nopwrite) {
1097                         ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1098                         ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1099                         zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1100                 }
1101         }
1102 
1103         return (ZIO_PIPELINE_CONTINUE);
1104 }
1105 
1106 static int
1107 zio_free_bp_init(zio_t *zio)
1108 {
1109         blkptr_t *bp = zio->io_bp;
1110 
1111         if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1112                 if (BP_GET_DEDUP(bp))
1113                         zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1114         }
1115 
1116         return (ZIO_PIPELINE_CONTINUE);
1117 }
1118 
1119 /*
1120  * ==========================================================================
1121  * Execute the I/O pipeline
1122  * ==========================================================================
1123  */
1124 
1125 static void
1126 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1127 {
1128         spa_t *spa = zio->io_spa;
1129         zio_type_t t = zio->io_type;
1130         int flags = (cutinline ? TQ_FRONT : 0);
1131 
1132         /*
1133          * If we're a config writer or a probe, the normal issue and
1134          * interrupt threads may all be blocked waiting for the config lock.
1135          * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1136          */
1137         if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1138                 t = ZIO_TYPE_NULL;
1139 
1140         /*
1141          * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1142          */
1143         if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1144                 t = ZIO_TYPE_NULL;
1145 
1146         /*
1147          * If this is a high priority I/O, then use the high priority taskq if
1148          * available.
1149          */
1150         if (zio->io_priority == ZIO_PRIORITY_NOW &&
1151             spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1152                 q++;
1153 
1154         ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1155 
1156         /*
1157          * NB: We are assuming that the zio can only be dispatched
1158          * to a single taskq at a time.  It would be a grievous error
1159          * to dispatch the zio to another taskq at the same time.
1160          */
1161         ASSERT(zio->io_tqent.tqent_next == NULL);
1162         spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1163             flags, &zio->io_tqent);
1164 }
1165 
1166 static boolean_t
1167 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1168 {
1169         kthread_t *executor = zio->io_executor;
1170         spa_t *spa = zio->io_spa;
1171 
1172         for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1173                 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1174                 uint_t i;
1175                 for (i = 0; i < tqs->stqs_count; i++) {
1176                         if (taskq_member(tqs->stqs_taskq[i], executor))
1177                                 return (B_TRUE);
1178                 }
1179         }
1180 
1181         return (B_FALSE);
1182 }
1183 
1184 static int
1185 zio_issue_async(zio_t *zio)
1186 {
1187         zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1188 
1189         return (ZIO_PIPELINE_STOP);
1190 }
1191 
1192 void
1193 zio_interrupt(zio_t *zio)
1194 {
1195         zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1196 }
1197 
1198 /*
1199  * Execute the I/O pipeline until one of the following occurs:
1200  *
1201  *      (1) the I/O completes
1202  *      (2) the pipeline stalls waiting for dependent child I/Os
1203  *      (3) the I/O issues, so we're waiting for an I/O completion interrupt
1204  *      (4) the I/O is delegated by vdev-level caching or aggregation
1205  *      (5) the I/O is deferred due to vdev-level queueing
1206  *      (6) the I/O is handed off to another thread.
1207  *
1208  * In all cases, the pipeline stops whenever there's no CPU work; it never
1209  * burns a thread in cv_wait().
1210  *
1211  * There's no locking on io_stage because there's no legitimate way
1212  * for multiple threads to be attempting to process the same I/O.
1213  */
1214 static zio_pipe_stage_t *zio_pipeline[];
1215 
1216 void
1217 zio_execute(zio_t *zio)
1218 {
1219         zio->io_executor = curthread;
1220 
1221         while (zio->io_stage < ZIO_STAGE_DONE) {
1222                 enum zio_stage pipeline = zio->io_pipeline;
1223                 enum zio_stage stage = zio->io_stage;
1224                 int rv;
1225 
1226                 ASSERT(!MUTEX_HELD(&zio->io_lock));
1227                 ASSERT(ISP2(stage));
1228                 ASSERT(zio->io_stall == NULL);
1229 
1230                 do {
1231                         stage <<= 1;
1232                 } while ((stage & pipeline) == 0);
1233 
1234                 ASSERT(stage <= ZIO_STAGE_DONE);
1235 
1236                 /*
1237                  * If we are in interrupt context and this pipeline stage
1238                  * will grab a config lock that is held across I/O,
1239                  * or may wait for an I/O that needs an interrupt thread
1240                  * to complete, issue async to avoid deadlock.
1241                  *
1242                  * For VDEV_IO_START, we cut in line so that the io will
1243                  * be sent to disk promptly.
1244                  */
1245                 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1246                     zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1247                         boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1248                             zio_requeue_io_start_cut_in_line : B_FALSE;
1249                         zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1250                         return;
1251                 }
1252 
1253                 zio->io_stage = stage;
1254                 rv = zio_pipeline[highbit(stage) - 1](zio);
1255 
1256                 if (rv == ZIO_PIPELINE_STOP)
1257                         return;
1258 
1259                 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1260         }
1261 }
1262 
1263 /*
1264  * ==========================================================================
1265  * Initiate I/O, either sync or async
1266  * ==========================================================================
1267  */
1268 int
1269 zio_wait(zio_t *zio)
1270 {
1271         int error;
1272 
1273         ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1274         ASSERT(zio->io_executor == NULL);
1275 
1276         zio->io_waiter = curthread;
1277 
1278         zio_execute(zio);
1279 
1280         mutex_enter(&zio->io_lock);
1281         while (zio->io_executor != NULL)
1282                 cv_wait(&zio->io_cv, &zio->io_lock);
1283         mutex_exit(&zio->io_lock);
1284 
1285         error = zio->io_error;
1286         zio_destroy(zio);
1287 
1288         return (error);
1289 }
1290 
1291 void
1292 zio_nowait(zio_t *zio)
1293 {
1294         ASSERT(zio->io_executor == NULL);
1295 
1296         if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1297             zio_unique_parent(zio) == NULL) {
1298                 /*
1299                  * This is a logical async I/O with no parent to wait for it.
1300                  * We add it to the spa_async_root_zio "Godfather" I/O which
1301                  * will ensure they complete prior to unloading the pool.
1302                  */
1303                 spa_t *spa = zio->io_spa;
1304 
1305                 zio_add_child(spa->spa_async_zio_root, zio);
1306         }
1307 
1308         zio_execute(zio);
1309 }
1310 
1311 /*
1312  * ==========================================================================
1313  * Reexecute or suspend/resume failed I/O
1314  * ==========================================================================
1315  */
1316 
1317 static void
1318 zio_reexecute(zio_t *pio)
1319 {
1320         zio_t *cio, *cio_next;
1321 
1322         ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1323         ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1324         ASSERT(pio->io_gang_leader == NULL);
1325         ASSERT(pio->io_gang_tree == NULL);
1326 
1327         pio->io_flags = pio->io_orig_flags;
1328         pio->io_stage = pio->io_orig_stage;
1329         pio->io_pipeline = pio->io_orig_pipeline;
1330         pio->io_reexecute = 0;
1331         pio->io_flags |= ZIO_FLAG_REEXECUTED;
1332         pio->io_error = 0;
1333         for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1334                 pio->io_state[w] = 0;
1335         for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1336                 pio->io_child_error[c] = 0;
1337 
1338         if (IO_IS_ALLOCATING(pio))
1339                 BP_ZERO(pio->io_bp);
1340 
1341         /*
1342          * As we reexecute pio's children, new children could be created.
1343          * New children go to the head of pio's io_child_list, however,
1344          * so we will (correctly) not reexecute them.  The key is that
1345          * the remainder of pio's io_child_list, from 'cio_next' onward,
1346          * cannot be affected by any side effects of reexecuting 'cio'.
1347          */
1348         for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1349                 cio_next = zio_walk_children(pio);
1350                 mutex_enter(&pio->io_lock);
1351                 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1352                         pio->io_children[cio->io_child_type][w]++;
1353                 mutex_exit(&pio->io_lock);
1354                 zio_reexecute(cio);
1355         }
1356 
1357         /*
1358          * Now that all children have been reexecuted, execute the parent.
1359          * We don't reexecute "The Godfather" I/O here as it's the
1360          * responsibility of the caller to wait on him.
1361          */
1362         if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1363                 zio_execute(pio);
1364 }
1365 
1366 void
1367 zio_suspend(spa_t *spa, zio_t *zio)
1368 {
1369         if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1370                 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1371                     "failure and the failure mode property for this pool "
1372                     "is set to panic.", spa_name(spa));
1373 
1374         zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1375 
1376         mutex_enter(&spa->spa_suspend_lock);
1377 
1378         if (spa->spa_suspend_zio_root == NULL)
1379                 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1380                     ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1381                     ZIO_FLAG_GODFATHER);
1382 
1383         spa->spa_suspended = B_TRUE;
1384 
1385         if (zio != NULL) {
1386                 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1387                 ASSERT(zio != spa->spa_suspend_zio_root);
1388                 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1389                 ASSERT(zio_unique_parent(zio) == NULL);
1390                 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1391                 zio_add_child(spa->spa_suspend_zio_root, zio);
1392         }
1393 
1394         mutex_exit(&spa->spa_suspend_lock);
1395 }
1396 
1397 int
1398 zio_resume(spa_t *spa)
1399 {
1400         zio_t *pio;
1401 
1402         /*
1403          * Reexecute all previously suspended i/o.
1404          */
1405         mutex_enter(&spa->spa_suspend_lock);
1406         spa->spa_suspended = B_FALSE;
1407         cv_broadcast(&spa->spa_suspend_cv);
1408         pio = spa->spa_suspend_zio_root;
1409         spa->spa_suspend_zio_root = NULL;
1410         mutex_exit(&spa->spa_suspend_lock);
1411 
1412         if (pio == NULL)
1413                 return (0);
1414 
1415         zio_reexecute(pio);
1416         return (zio_wait(pio));
1417 }
1418 
1419 void
1420 zio_resume_wait(spa_t *spa)
1421 {
1422         mutex_enter(&spa->spa_suspend_lock);
1423         while (spa_suspended(spa))
1424                 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1425         mutex_exit(&spa->spa_suspend_lock);
1426 }
1427 
1428 /*
1429  * ==========================================================================
1430  * Gang blocks.
1431  *
1432  * A gang block is a collection of small blocks that looks to the DMU
1433  * like one large block.  When zio_dva_allocate() cannot find a block
1434  * of the requested size, due to either severe fragmentation or the pool
1435  * being nearly full, it calls zio_write_gang_block() to construct the
1436  * block from smaller fragments.
1437  *
1438  * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1439  * three (SPA_GBH_NBLKPTRS) gang members.  The gang header is just like
1440  * an indirect block: it's an array of block pointers.  It consumes
1441  * only one sector and hence is allocatable regardless of fragmentation.
1442  * The gang header's bps point to its gang members, which hold the data.
1443  *
1444  * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1445  * as the verifier to ensure uniqueness of the SHA256 checksum.
1446  * Critically, the gang block bp's blk_cksum is the checksum of the data,
1447  * not the gang header.  This ensures that data block signatures (needed for
1448  * deduplication) are independent of how the block is physically stored.
1449  *
1450  * Gang blocks can be nested: a gang member may itself be a gang block.
1451  * Thus every gang block is a tree in which root and all interior nodes are
1452  * gang headers, and the leaves are normal blocks that contain user data.
1453  * The root of the gang tree is called the gang leader.
1454  *
1455  * To perform any operation (read, rewrite, free, claim) on a gang block,
1456  * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1457  * in the io_gang_tree field of the original logical i/o by recursively
1458  * reading the gang leader and all gang headers below it.  This yields
1459  * an in-core tree containing the contents of every gang header and the
1460  * bps for every constituent of the gang block.
1461  *
1462  * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1463  * and invokes a callback on each bp.  To free a gang block, zio_gang_issue()
1464  * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1465  * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1466  * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1467  * headers, since we already have those in io_gang_tree.  zio_rewrite_gang()
1468  * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1469  * of the gang header plus zio_checksum_compute() of the data to update the
1470  * gang header's blk_cksum as described above.
1471  *
1472  * The two-phase assemble/issue model solves the problem of partial failure --
1473  * what if you'd freed part of a gang block but then couldn't read the
1474  * gang header for another part?  Assembling the entire gang tree first
1475  * ensures that all the necessary gang header I/O has succeeded before
1476  * starting the actual work of free, claim, or write.  Once the gang tree
1477  * is assembled, free and claim are in-memory operations that cannot fail.
1478  *
1479  * In the event that a gang write fails, zio_dva_unallocate() walks the
1480  * gang tree to immediately free (i.e. insert back into the space map)
1481  * everything we've allocated.  This ensures that we don't get ENOSPC
1482  * errors during repeated suspend/resume cycles due to a flaky device.
1483  *
1484  * Gang rewrites only happen during sync-to-convergence.  If we can't assemble
1485  * the gang tree, we won't modify the block, so we can safely defer the free
1486  * (knowing that the block is still intact).  If we *can* assemble the gang
1487  * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1488  * each constituent bp and we can allocate a new block on the next sync pass.
1489  *
1490  * In all cases, the gang tree allows complete recovery from partial failure.
1491  * ==========================================================================
1492  */
1493 
1494 static zio_t *
1495 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1496 {
1497         if (gn != NULL)
1498                 return (pio);
1499 
1500         return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1501             NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1502             &pio->io_bookmark));
1503 }
1504 
1505 zio_t *
1506 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1507 {
1508         zio_t *zio;
1509 
1510         if (gn != NULL) {
1511                 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1512                     gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1513                     ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1514                 /*
1515                  * As we rewrite each gang header, the pipeline will compute
1516                  * a new gang block header checksum for it; but no one will
1517                  * compute a new data checksum, so we do that here.  The one
1518                  * exception is the gang leader: the pipeline already computed
1519                  * its data checksum because that stage precedes gang assembly.
1520                  * (Presently, nothing actually uses interior data checksums;
1521                  * this is just good hygiene.)
1522                  */
1523                 if (gn != pio->io_gang_leader->io_gang_tree) {
1524                         zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1525                             data, BP_GET_PSIZE(bp));
1526                 }
1527                 /*
1528                  * If we are here to damage data for testing purposes,
1529                  * leave the GBH alone so that we can detect the damage.
1530                  */
1531                 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1532                         zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1533         } else {
1534                 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1535                     data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1536                     ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1537         }
1538 
1539         return (zio);
1540 }
1541 
1542 /* ARGSUSED */
1543 zio_t *
1544 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1545 {
1546         return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1547             ZIO_GANG_CHILD_FLAGS(pio)));
1548 }
1549 
1550 /* ARGSUSED */
1551 zio_t *
1552 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1553 {
1554         return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1555             NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1556 }
1557 
1558 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1559         NULL,
1560         zio_read_gang,
1561         zio_rewrite_gang,
1562         zio_free_gang,
1563         zio_claim_gang,
1564         NULL
1565 };
1566 
1567 static void zio_gang_tree_assemble_done(zio_t *zio);
1568 
1569 static zio_gang_node_t *
1570 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1571 {
1572         zio_gang_node_t *gn;
1573 
1574         ASSERT(*gnpp == NULL);
1575 
1576         gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1577         gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1578         *gnpp = gn;
1579 
1580         return (gn);
1581 }
1582 
1583 static void
1584 zio_gang_node_free(zio_gang_node_t **gnpp)
1585 {
1586         zio_gang_node_t *gn = *gnpp;
1587 
1588         for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1589                 ASSERT(gn->gn_child[g] == NULL);
1590 
1591         zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1592         kmem_free(gn, sizeof (*gn));
1593         *gnpp = NULL;
1594 }
1595 
1596 static void
1597 zio_gang_tree_free(zio_gang_node_t **gnpp)
1598 {
1599         zio_gang_node_t *gn = *gnpp;
1600 
1601         if (gn == NULL)
1602                 return;
1603 
1604         for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1605                 zio_gang_tree_free(&gn->gn_child[g]);
1606 
1607         zio_gang_node_free(gnpp);
1608 }
1609 
1610 static void
1611 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1612 {
1613         zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1614 
1615         ASSERT(gio->io_gang_leader == gio);
1616         ASSERT(BP_IS_GANG(bp));
1617 
1618         zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1619             SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1620             gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1621 }
1622 
1623 static void
1624 zio_gang_tree_assemble_done(zio_t *zio)
1625 {
1626         zio_t *gio = zio->io_gang_leader;
1627         zio_gang_node_t *gn = zio->io_private;
1628         blkptr_t *bp = zio->io_bp;
1629 
1630         ASSERT(gio == zio_unique_parent(zio));
1631         ASSERT(zio->io_child_count == 0);
1632 
1633         if (zio->io_error)
1634                 return;
1635 
1636         if (BP_SHOULD_BYTESWAP(bp))
1637                 byteswap_uint64_array(zio->io_data, zio->io_size);
1638 
1639         ASSERT(zio->io_data == gn->gn_gbh);
1640         ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1641         ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1642 
1643         for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1644                 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1645                 if (!BP_IS_GANG(gbp))
1646                         continue;
1647                 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1648         }
1649 }
1650 
1651 static void
1652 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1653 {
1654         zio_t *gio = pio->io_gang_leader;
1655         zio_t *zio;
1656 
1657         ASSERT(BP_IS_GANG(bp) == !!gn);
1658         ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1659         ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1660 
1661         /*
1662          * If you're a gang header, your data is in gn->gn_gbh.
1663          * If you're a gang member, your data is in 'data' and gn == NULL.
1664          */
1665         zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1666 
1667         if (gn != NULL) {
1668                 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1669 
1670                 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1671                         blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1672                         if (BP_IS_HOLE(gbp))
1673                                 continue;
1674                         zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1675                         data = (char *)data + BP_GET_PSIZE(gbp);
1676                 }
1677         }
1678 
1679         if (gn == gio->io_gang_tree)
1680                 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1681 
1682         if (zio != pio)
1683                 zio_nowait(zio);
1684 }
1685 
1686 static int
1687 zio_gang_assemble(zio_t *zio)
1688 {
1689         blkptr_t *bp = zio->io_bp;
1690 
1691         ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1692         ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1693 
1694         zio->io_gang_leader = zio;
1695 
1696         zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1697 
1698         return (ZIO_PIPELINE_CONTINUE);
1699 }
1700 
1701 static int
1702 zio_gang_issue(zio_t *zio)
1703 {
1704         blkptr_t *bp = zio->io_bp;
1705 
1706         if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1707                 return (ZIO_PIPELINE_STOP);
1708 
1709         ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1710         ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1711 
1712         if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1713                 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1714         else
1715                 zio_gang_tree_free(&zio->io_gang_tree);
1716 
1717         zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1718 
1719         return (ZIO_PIPELINE_CONTINUE);
1720 }
1721 
1722 static void
1723 zio_write_gang_member_ready(zio_t *zio)
1724 {
1725         zio_t *pio = zio_unique_parent(zio);
1726         zio_t *gio = zio->io_gang_leader;
1727         dva_t *cdva = zio->io_bp->blk_dva;
1728         dva_t *pdva = pio->io_bp->blk_dva;
1729         uint64_t asize;
1730 
1731         if (BP_IS_HOLE(zio->io_bp))
1732                 return;
1733 
1734         ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1735 
1736         ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1737         ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1738         ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1739         ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1740         ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1741 
1742         mutex_enter(&pio->io_lock);
1743         for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1744                 ASSERT(DVA_GET_GANG(&pdva[d]));
1745                 asize = DVA_GET_ASIZE(&pdva[d]);
1746                 asize += DVA_GET_ASIZE(&cdva[d]);
1747                 DVA_SET_ASIZE(&pdva[d], asize);
1748         }
1749         mutex_exit(&pio->io_lock);
1750 }
1751 
1752 static int
1753 zio_write_gang_block(zio_t *pio)
1754 {
1755         spa_t *spa = pio->io_spa;
1756         blkptr_t *bp = pio->io_bp;
1757         zio_t *gio = pio->io_gang_leader;
1758         zio_t *zio;
1759         zio_gang_node_t *gn, **gnpp;
1760         zio_gbh_phys_t *gbh;
1761         uint64_t txg = pio->io_txg;
1762         uint64_t resid = pio->io_size;
1763         uint64_t lsize;
1764         int copies = gio->io_prop.zp_copies;
1765         int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1766         zio_prop_t zp;
1767         int error;
1768 
1769         error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1770             bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1771             METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1772         if (error) {
1773                 pio->io_error = error;
1774                 return (ZIO_PIPELINE_CONTINUE);
1775         }
1776 
1777         if (pio == gio) {
1778                 gnpp = &gio->io_gang_tree;
1779         } else {
1780                 gnpp = pio->io_private;
1781                 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1782         }
1783 
1784         gn = zio_gang_node_alloc(gnpp);
1785         gbh = gn->gn_gbh;
1786         bzero(gbh, SPA_GANGBLOCKSIZE);
1787 
1788         /*
1789          * Create the gang header.
1790          */
1791         zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1792             pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1793 
1794         /*
1795          * Create and nowait the gang children.
1796          */
1797         for (int g = 0; resid != 0; resid -= lsize, g++) {
1798                 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1799                     SPA_MINBLOCKSIZE);
1800                 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1801 
1802                 zp.zp_checksum = gio->io_prop.zp_checksum;
1803                 zp.zp_compress = ZIO_COMPRESS_OFF;
1804                 zp.zp_type = DMU_OT_NONE;
1805                 zp.zp_level = 0;
1806                 zp.zp_copies = gio->io_prop.zp_copies;
1807                 zp.zp_dedup = B_FALSE;
1808                 zp.zp_dedup_verify = B_FALSE;
1809                 zp.zp_nopwrite = B_FALSE;
1810 
1811                 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1812                     (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1813                     zio_write_gang_member_ready, NULL, NULL, &gn->gn_child[g],
1814                     pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1815                     &pio->io_bookmark));
1816         }
1817 
1818         /*
1819          * Set pio's pipeline to just wait for zio to finish.
1820          */
1821         pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1822 
1823         zio_nowait(zio);
1824 
1825         return (ZIO_PIPELINE_CONTINUE);
1826 }
1827 
1828 /*
1829  * The zio_nop_write stage in the pipeline determines if allocating
1830  * a new bp is necessary.  By leveraging a cryptographically secure checksum,
1831  * such as SHA256, we can compare the checksums of the new data and the old
1832  * to determine if allocating a new block is required.  The nopwrite
1833  * feature can handle writes in either syncing or open context (i.e. zil
1834  * writes) and as a result is mutually exclusive with dedup.
1835  */
1836 static int
1837 zio_nop_write(zio_t *zio)
1838 {
1839         blkptr_t *bp = zio->io_bp;
1840         blkptr_t *bp_orig = &zio->io_bp_orig;
1841         zio_prop_t *zp = &zio->io_prop;
1842 
1843         ASSERT(BP_GET_LEVEL(bp) == 0);
1844         ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1845         ASSERT(zp->zp_nopwrite);
1846         ASSERT(!zp->zp_dedup);
1847         ASSERT(zio->io_bp_override == NULL);
1848         ASSERT(IO_IS_ALLOCATING(zio));
1849 
1850         /*
1851          * Check to see if the original bp and the new bp have matching
1852          * characteristics (i.e. same checksum, compression algorithms, etc).
1853          * If they don't then just continue with the pipeline which will
1854          * allocate a new bp.
1855          */
1856         if (BP_IS_HOLE(bp_orig) ||
1857             !zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup ||
1858             BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
1859             BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
1860             BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
1861             zp->zp_copies != BP_GET_NDVAS(bp_orig))
1862                 return (ZIO_PIPELINE_CONTINUE);
1863 
1864         /*
1865          * If the checksums match then reset the pipeline so that we
1866          * avoid allocating a new bp and issuing any I/O.
1867          */
1868         if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
1869                 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup);
1870                 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
1871                 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
1872                 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
1873                 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
1874                     sizeof (uint64_t)) == 0);
1875 
1876                 *bp = *bp_orig;
1877                 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1878                 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1879         }
1880 
1881         return (ZIO_PIPELINE_CONTINUE);
1882 }
1883 
1884 /*
1885  * ==========================================================================
1886  * Dedup
1887  * ==========================================================================
1888  */
1889 static void
1890 zio_ddt_child_read_done(zio_t *zio)
1891 {
1892         blkptr_t *bp = zio->io_bp;
1893         ddt_entry_t *dde = zio->io_private;
1894         ddt_phys_t *ddp;
1895         zio_t *pio = zio_unique_parent(zio);
1896 
1897         mutex_enter(&pio->io_lock);
1898         ddp = ddt_phys_select(dde, bp);
1899         if (zio->io_error == 0)
1900                 ddt_phys_clear(ddp);    /* this ddp doesn't need repair */
1901         if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1902                 dde->dde_repair_data = zio->io_data;
1903         else
1904                 zio_buf_free(zio->io_data, zio->io_size);
1905         mutex_exit(&pio->io_lock);
1906 }
1907 
1908 static int
1909 zio_ddt_read_start(zio_t *zio)
1910 {
1911         blkptr_t *bp = zio->io_bp;
1912 
1913         ASSERT(BP_GET_DEDUP(bp));
1914         ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1915         ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1916 
1917         if (zio->io_child_error[ZIO_CHILD_DDT]) {
1918                 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1919                 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1920                 ddt_phys_t *ddp = dde->dde_phys;
1921                 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1922                 blkptr_t blk;
1923 
1924                 ASSERT(zio->io_vsd == NULL);
1925                 zio->io_vsd = dde;
1926 
1927                 if (ddp_self == NULL)
1928                         return (ZIO_PIPELINE_CONTINUE);
1929 
1930                 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1931                         if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1932                                 continue;
1933                         ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1934                             &blk);
1935                         zio_nowait(zio_read(zio, zio->io_spa, &blk,
1936                             zio_buf_alloc(zio->io_size), zio->io_size,
1937                             zio_ddt_child_read_done, dde, zio->io_priority,
1938                             ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1939                             &zio->io_bookmark));
1940                 }
1941                 return (ZIO_PIPELINE_CONTINUE);
1942         }
1943 
1944         zio_nowait(zio_read(zio, zio->io_spa, bp,
1945             zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1946             ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1947 
1948         return (ZIO_PIPELINE_CONTINUE);
1949 }
1950 
1951 static int
1952 zio_ddt_read_done(zio_t *zio)
1953 {
1954         blkptr_t *bp = zio->io_bp;
1955 
1956         if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1957                 return (ZIO_PIPELINE_STOP);
1958 
1959         ASSERT(BP_GET_DEDUP(bp));
1960         ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1961         ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1962 
1963         if (zio->io_child_error[ZIO_CHILD_DDT]) {
1964                 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1965                 ddt_entry_t *dde = zio->io_vsd;
1966                 if (ddt == NULL) {
1967                         ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
1968                         return (ZIO_PIPELINE_CONTINUE);
1969                 }
1970                 if (dde == NULL) {
1971                         zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
1972                         zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1973                         return (ZIO_PIPELINE_STOP);
1974                 }
1975                 if (dde->dde_repair_data != NULL) {
1976                         bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
1977                         zio->io_child_error[ZIO_CHILD_DDT] = 0;
1978                 }
1979                 ddt_repair_done(ddt, dde);
1980                 zio->io_vsd = NULL;
1981         }
1982 
1983         ASSERT(zio->io_vsd == NULL);
1984 
1985         return (ZIO_PIPELINE_CONTINUE);
1986 }
1987 
1988 static boolean_t
1989 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
1990 {
1991         spa_t *spa = zio->io_spa;
1992 
1993         /*
1994          * Note: we compare the original data, not the transformed data,
1995          * because when zio->io_bp is an override bp, we will not have
1996          * pushed the I/O transforms.  That's an important optimization
1997          * because otherwise we'd compress/encrypt all dmu_sync() data twice.
1998          */
1999         for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2000                 zio_t *lio = dde->dde_lead_zio[p];
2001 
2002                 if (lio != NULL) {
2003                         return (lio->io_orig_size != zio->io_orig_size ||
2004                             bcmp(zio->io_orig_data, lio->io_orig_data,
2005                             zio->io_orig_size) != 0);
2006                 }
2007         }
2008 
2009         for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2010                 ddt_phys_t *ddp = &dde->dde_phys[p];
2011 
2012                 if (ddp->ddp_phys_birth != 0) {
2013                         arc_buf_t *abuf = NULL;
2014                         uint32_t aflags = ARC_WAIT;
2015                         blkptr_t blk = *zio->io_bp;
2016                         int error;
2017 
2018                         ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2019 
2020                         ddt_exit(ddt);
2021 
2022                         error = arc_read(NULL, spa, &blk,
2023                             arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2024                             ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2025                             &aflags, &zio->io_bookmark);
2026 
2027                         if (error == 0) {
2028                                 if (arc_buf_size(abuf) != zio->io_orig_size ||
2029                                     bcmp(abuf->b_data, zio->io_orig_data,
2030                                     zio->io_orig_size) != 0)
2031                                         error = SET_ERROR(EEXIST);
2032                                 VERIFY(arc_buf_remove_ref(abuf, &abuf));
2033                         }
2034 
2035                         ddt_enter(ddt);
2036                         return (error != 0);
2037                 }
2038         }
2039 
2040         return (B_FALSE);
2041 }
2042 
2043 static void
2044 zio_ddt_child_write_ready(zio_t *zio)
2045 {
2046         int p = zio->io_prop.zp_copies;
2047         ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2048         ddt_entry_t *dde = zio->io_private;
2049         ddt_phys_t *ddp = &dde->dde_phys[p];
2050         zio_t *pio;
2051 
2052         if (zio->io_error)
2053                 return;
2054 
2055         ddt_enter(ddt);
2056 
2057         ASSERT(dde->dde_lead_zio[p] == zio);
2058 
2059         ddt_phys_fill(ddp, zio->io_bp);
2060 
2061         while ((pio = zio_walk_parents(zio)) != NULL)
2062                 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2063 
2064         ddt_exit(ddt);
2065 }
2066 
2067 static void
2068 zio_ddt_child_write_done(zio_t *zio)
2069 {
2070         int p = zio->io_prop.zp_copies;
2071         ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2072         ddt_entry_t *dde = zio->io_private;
2073         ddt_phys_t *ddp = &dde->dde_phys[p];
2074 
2075         ddt_enter(ddt);
2076 
2077         ASSERT(ddp->ddp_refcnt == 0);
2078         ASSERT(dde->dde_lead_zio[p] == zio);
2079         dde->dde_lead_zio[p] = NULL;
2080 
2081         if (zio->io_error == 0) {
2082                 while (zio_walk_parents(zio) != NULL)
2083                         ddt_phys_addref(ddp);
2084         } else {
2085                 ddt_phys_clear(ddp);
2086         }
2087 
2088         ddt_exit(ddt);
2089 }
2090 
2091 static void
2092 zio_ddt_ditto_write_done(zio_t *zio)
2093 {
2094         int p = DDT_PHYS_DITTO;
2095         zio_prop_t *zp = &zio->io_prop;
2096         blkptr_t *bp = zio->io_bp;
2097         ddt_t *ddt = ddt_select(zio->io_spa, bp);
2098         ddt_entry_t *dde = zio->io_private;
2099         ddt_phys_t *ddp = &dde->dde_phys[p];
2100         ddt_key_t *ddk = &dde->dde_key;
2101 
2102         ddt_enter(ddt);
2103 
2104         ASSERT(ddp->ddp_refcnt == 0);
2105         ASSERT(dde->dde_lead_zio[p] == zio);
2106         dde->dde_lead_zio[p] = NULL;
2107 
2108         if (zio->io_error == 0) {
2109                 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2110                 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2111                 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2112                 if (ddp->ddp_phys_birth != 0)
2113                         ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2114                 ddt_phys_fill(ddp, bp);
2115         }
2116 
2117         ddt_exit(ddt);
2118 }
2119 
2120 static int
2121 zio_ddt_write(zio_t *zio)
2122 {
2123         spa_t *spa = zio->io_spa;
2124         blkptr_t *bp = zio->io_bp;
2125         uint64_t txg = zio->io_txg;
2126         zio_prop_t *zp = &zio->io_prop;
2127         int p = zp->zp_copies;
2128         int ditto_copies;
2129         zio_t *cio = NULL;
2130         zio_t *dio = NULL;
2131         ddt_t *ddt = ddt_select(spa, bp);
2132         ddt_entry_t *dde;
2133         ddt_phys_t *ddp;
2134 
2135         ASSERT(BP_GET_DEDUP(bp));
2136         ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2137         ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2138 
2139         ddt_enter(ddt);
2140         dde = ddt_lookup(ddt, bp, B_TRUE);
2141         ddp = &dde->dde_phys[p];
2142 
2143         if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2144                 /*
2145                  * If we're using a weak checksum, upgrade to a strong checksum
2146                  * and try again.  If we're already using a strong checksum,
2147                  * we can't resolve it, so just convert to an ordinary write.
2148                  * (And automatically e-mail a paper to Nature?)
2149                  */
2150                 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2151                         zp->zp_checksum = spa_dedup_checksum(spa);
2152                         zio_pop_transforms(zio);
2153                         zio->io_stage = ZIO_STAGE_OPEN;
2154                         BP_ZERO(bp);
2155                 } else {
2156                         zp->zp_dedup = B_FALSE;
2157                 }
2158                 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2159                 ddt_exit(ddt);
2160                 return (ZIO_PIPELINE_CONTINUE);
2161         }
2162 
2163         ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2164         ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2165 
2166         if (ditto_copies > ddt_ditto_copies_present(dde) &&
2167             dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2168                 zio_prop_t czp = *zp;
2169 
2170                 czp.zp_copies = ditto_copies;
2171 
2172                 /*
2173                  * If we arrived here with an override bp, we won't have run
2174                  * the transform stack, so we won't have the data we need to
2175                  * generate a child i/o.  So, toss the override bp and restart.
2176                  * This is safe, because using the override bp is just an
2177                  * optimization; and it's rare, so the cost doesn't matter.
2178                  */
2179                 if (zio->io_bp_override) {
2180                         zio_pop_transforms(zio);
2181                         zio->io_stage = ZIO_STAGE_OPEN;
2182                         zio->io_pipeline = ZIO_WRITE_PIPELINE;
2183                         zio->io_bp_override = NULL;
2184                         BP_ZERO(bp);
2185                         ddt_exit(ddt);
2186                         return (ZIO_PIPELINE_CONTINUE);
2187                 }
2188 
2189                 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2190                     zio->io_orig_size, &czp, NULL, NULL,
2191                     zio_ddt_ditto_write_done, dde, zio->io_priority,
2192                     ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2193 
2194                 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2195                 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2196         }
2197 
2198         if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2199                 if (ddp->ddp_phys_birth != 0)
2200                         ddt_bp_fill(ddp, bp, txg);
2201                 if (dde->dde_lead_zio[p] != NULL)
2202                         zio_add_child(zio, dde->dde_lead_zio[p]);
2203                 else
2204                         ddt_phys_addref(ddp);
2205         } else if (zio->io_bp_override) {
2206                 ASSERT(bp->blk_birth == txg);
2207                 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2208                 ddt_phys_fill(ddp, bp);
2209                 ddt_phys_addref(ddp);
2210         } else {
2211                 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2212                     zio->io_orig_size, zp, zio_ddt_child_write_ready, NULL,
2213                     zio_ddt_child_write_done, dde, zio->io_priority,
2214                     ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2215 
2216                 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2217                 dde->dde_lead_zio[p] = cio;
2218         }
2219 
2220         ddt_exit(ddt);
2221 
2222         if (cio)
2223                 zio_nowait(cio);
2224         if (dio)
2225                 zio_nowait(dio);
2226 
2227         return (ZIO_PIPELINE_CONTINUE);
2228 }
2229 
2230 ddt_entry_t *freedde; /* for debugging */
2231 
2232 static int
2233 zio_ddt_free(zio_t *zio)
2234 {
2235         spa_t *spa = zio->io_spa;
2236         blkptr_t *bp = zio->io_bp;
2237         ddt_t *ddt = ddt_select(spa, bp);
2238         ddt_entry_t *dde;
2239         ddt_phys_t *ddp;
2240 
2241         ASSERT(BP_GET_DEDUP(bp));
2242         ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2243 
2244         ddt_enter(ddt);
2245         freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2246         ddp = ddt_phys_select(dde, bp);
2247         ddt_phys_decref(ddp);
2248         ddt_exit(ddt);
2249 
2250         return (ZIO_PIPELINE_CONTINUE);
2251 }
2252 
2253 /*
2254  * ==========================================================================
2255  * Allocate and free blocks
2256  * ==========================================================================
2257  */
2258 static int
2259 zio_dva_allocate(zio_t *zio)
2260 {
2261         spa_t *spa = zio->io_spa;
2262         metaslab_class_t *mc = spa_normal_class(spa);
2263         blkptr_t *bp = zio->io_bp;
2264         int error;
2265         int flags = 0;
2266 
2267         if (zio->io_gang_leader == NULL) {
2268                 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2269                 zio->io_gang_leader = zio;
2270         }
2271 
2272         ASSERT(BP_IS_HOLE(bp));
2273         ASSERT0(BP_GET_NDVAS(bp));
2274         ASSERT3U(zio->io_prop.zp_copies, >, 0);
2275         ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2276         ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2277 
2278         /*
2279          * The dump device does not support gang blocks so allocation on
2280          * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2281          * the "fast" gang feature.
2282          */
2283         flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2284         flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2285             METASLAB_GANG_CHILD : 0;
2286         error = metaslab_alloc(spa, mc, zio->io_size, bp,
2287             zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2288 
2289         if (error) {
2290                 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2291                     "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2292                     error);
2293                 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2294                         return (zio_write_gang_block(zio));
2295                 zio->io_error = error;
2296         }
2297 
2298         return (ZIO_PIPELINE_CONTINUE);
2299 }
2300 
2301 static int
2302 zio_dva_free(zio_t *zio)
2303 {
2304         metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2305 
2306         return (ZIO_PIPELINE_CONTINUE);
2307 }
2308 
2309 static int
2310 zio_dva_claim(zio_t *zio)
2311 {
2312         int error;
2313 
2314         error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2315         if (error)
2316                 zio->io_error = error;
2317 
2318         return (ZIO_PIPELINE_CONTINUE);
2319 }
2320 
2321 /*
2322  * Undo an allocation.  This is used by zio_done() when an I/O fails
2323  * and we want to give back the block we just allocated.
2324  * This handles both normal blocks and gang blocks.
2325  */
2326 static void
2327 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2328 {
2329         ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2330         ASSERT(zio->io_bp_override == NULL);
2331 
2332         if (!BP_IS_HOLE(bp))
2333                 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2334 
2335         if (gn != NULL) {
2336                 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2337                         zio_dva_unallocate(zio, gn->gn_child[g],
2338                             &gn->gn_gbh->zg_blkptr[g]);
2339                 }
2340         }
2341 }
2342 
2343 /*
2344  * Try to allocate an intent log block.  Return 0 on success, errno on failure.
2345  */
2346 int
2347 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2348     uint64_t size, boolean_t use_slog)
2349 {
2350         int error = 1;
2351 
2352         ASSERT(txg > spa_syncing_txg(spa));
2353 
2354         /*
2355          * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2356          * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2357          * when allocating them.
2358          */
2359         if (use_slog) {
2360                 error = metaslab_alloc(spa, spa_log_class(spa), size,
2361                     new_bp, 1, txg, old_bp,
2362                     METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2363         }
2364 
2365         if (error) {
2366                 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2367                     new_bp, 1, txg, old_bp,
2368                     METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2369         }
2370 
2371         if (error == 0) {
2372                 BP_SET_LSIZE(new_bp, size);
2373                 BP_SET_PSIZE(new_bp, size);
2374                 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2375                 BP_SET_CHECKSUM(new_bp,
2376                     spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2377                     ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2378                 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2379                 BP_SET_LEVEL(new_bp, 0);
2380                 BP_SET_DEDUP(new_bp, 0);
2381                 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2382         }
2383 
2384         return (error);
2385 }
2386 
2387 /*
2388  * Free an intent log block.
2389  */
2390 void
2391 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2392 {
2393         ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2394         ASSERT(!BP_IS_GANG(bp));
2395 
2396         zio_free(spa, txg, bp);
2397 }
2398 
2399 /*
2400  * ==========================================================================
2401  * Read and write to physical devices
2402  * ==========================================================================
2403  */
2404 static int
2405 zio_vdev_io_start(zio_t *zio)
2406 {
2407         vdev_t *vd = zio->io_vd;
2408         uint64_t align;
2409         spa_t *spa = zio->io_spa;
2410 
2411         ASSERT(zio->io_error == 0);
2412         ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2413 
2414         if (vd == NULL) {
2415                 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2416                         spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2417 
2418                 /*
2419                  * The mirror_ops handle multiple DVAs in a single BP.
2420                  */
2421                 return (vdev_mirror_ops.vdev_op_io_start(zio));
2422         }
2423 
2424         /*
2425          * We keep track of time-sensitive I/Os so that the scan thread
2426          * can quickly react to certain workloads.  In particular, we care
2427          * about non-scrubbing, top-level reads and writes with the following
2428          * characteristics:
2429          *      - synchronous writes of user data to non-slog devices
2430          *      - any reads of user data
2431          * When these conditions are met, adjust the timestamp of spa_last_io
2432          * which allows the scan thread to adjust its workload accordingly.
2433          */
2434         if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2435             vd == vd->vdev_top && !vd->vdev_islog &&
2436             zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2437             zio->io_txg != spa_syncing_txg(spa)) {
2438                 uint64_t old = spa->spa_last_io;
2439                 uint64_t new = ddi_get_lbolt64();
2440                 if (old != new)
2441                         (void) atomic_cas_64(&spa->spa_last_io, old, new);
2442         }
2443 
2444         align = 1ULL << vd->vdev_top->vdev_ashift;
2445 
2446         if (P2PHASE(zio->io_size, align) != 0) {
2447                 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2448                 char *abuf = zio_buf_alloc(asize);
2449                 ASSERT(vd == vd->vdev_top);
2450                 if (zio->io_type == ZIO_TYPE_WRITE) {
2451                         bcopy(zio->io_data, abuf, zio->io_size);
2452                         bzero(abuf + zio->io_size, asize - zio->io_size);
2453                 }
2454                 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2455         }
2456 
2457         ASSERT(P2PHASE(zio->io_offset, align) == 0);
2458         ASSERT(P2PHASE(zio->io_size, align) == 0);
2459         VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2460 
2461         /*
2462          * If this is a repair I/O, and there's no self-healing involved --
2463          * that is, we're just resilvering what we expect to resilver --
2464          * then don't do the I/O unless zio's txg is actually in vd's DTL.
2465          * This prevents spurious resilvering with nested replication.
2466          * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2467          * A is out of date, we'll read from C+D, then use the data to
2468          * resilver A+B -- but we don't actually want to resilver B, just A.
2469          * The top-level mirror has no way to know this, so instead we just
2470          * discard unnecessary repairs as we work our way down the vdev tree.
2471          * The same logic applies to any form of nested replication:
2472          * ditto + mirror, RAID-Z + replacing, etc.  This covers them all.
2473          */
2474         if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2475             !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2476             zio->io_txg != 0 &&      /* not a delegated i/o */
2477             !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2478                 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2479                 zio_vdev_io_bypass(zio);
2480                 return (ZIO_PIPELINE_CONTINUE);
2481         }
2482 
2483         if (vd->vdev_ops->vdev_op_leaf &&
2484             (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2485 
2486                 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2487                         return (ZIO_PIPELINE_CONTINUE);
2488 
2489                 if ((zio = vdev_queue_io(zio)) == NULL)
2490                         return (ZIO_PIPELINE_STOP);
2491 
2492                 if (!vdev_accessible(vd, zio)) {
2493                         zio->io_error = SET_ERROR(ENXIO);
2494                         zio_interrupt(zio);
2495                         return (ZIO_PIPELINE_STOP);
2496                 }
2497         }
2498 
2499         return (vd->vdev_ops->vdev_op_io_start(zio));
2500 }
2501 
2502 static int
2503 zio_vdev_io_done(zio_t *zio)
2504 {
2505         vdev_t *vd = zio->io_vd;
2506         vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2507         boolean_t unexpected_error = B_FALSE;
2508 
2509         if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2510                 return (ZIO_PIPELINE_STOP);
2511 
2512         ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2513 
2514         if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2515 
2516                 vdev_queue_io_done(zio);
2517 
2518                 if (zio->io_type == ZIO_TYPE_WRITE)
2519                         vdev_cache_write(zio);
2520 
2521                 if (zio_injection_enabled && zio->io_error == 0)
2522                         zio->io_error = zio_handle_device_injection(vd,
2523                             zio, EIO);
2524 
2525                 if (zio_injection_enabled && zio->io_error == 0)
2526                         zio->io_error = zio_handle_label_injection(zio, EIO);
2527 
2528                 if (zio->io_error) {
2529                         if (!vdev_accessible(vd, zio)) {
2530                                 zio->io_error = SET_ERROR(ENXIO);
2531                         } else {
2532                                 unexpected_error = B_TRUE;
2533                         }
2534                 }
2535         }
2536 
2537         ops->vdev_op_io_done(zio);
2538 
2539         if (unexpected_error)
2540                 VERIFY(vdev_probe(vd, zio) == NULL);
2541 
2542         return (ZIO_PIPELINE_CONTINUE);
2543 }
2544 
2545 /*
2546  * For non-raidz ZIOs, we can just copy aside the bad data read from the
2547  * disk, and use that to finish the checksum ereport later.
2548  */
2549 static void
2550 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2551     const void *good_buf)
2552 {
2553         /* no processing needed */
2554         zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2555 }
2556 
2557 /*ARGSUSED*/
2558 void
2559 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2560 {
2561         void *buf = zio_buf_alloc(zio->io_size);
2562 
2563         bcopy(zio->io_data, buf, zio->io_size);
2564 
2565         zcr->zcr_cbinfo = zio->io_size;
2566         zcr->zcr_cbdata = buf;
2567         zcr->zcr_finish = zio_vsd_default_cksum_finish;
2568         zcr->zcr_free = zio_buf_free;
2569 }
2570 
2571 static int
2572 zio_vdev_io_assess(zio_t *zio)
2573 {
2574         vdev_t *vd = zio->io_vd;
2575 
2576         if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2577                 return (ZIO_PIPELINE_STOP);
2578 
2579         if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2580                 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2581 
2582         if (zio->io_vsd != NULL) {
2583                 zio->io_vsd_ops->vsd_free(zio);
2584                 zio->io_vsd = NULL;
2585         }
2586 
2587         if (zio_injection_enabled && zio->io_error == 0)
2588                 zio->io_error = zio_handle_fault_injection(zio, EIO);
2589 
2590         /*
2591          * If the I/O failed, determine whether we should attempt to retry it.
2592          *
2593          * On retry, we cut in line in the issue queue, since we don't want
2594          * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2595          */
2596         if (zio->io_error && vd == NULL &&
2597             !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2598                 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE));  /* not a leaf */
2599                 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS));   /* not a leaf */
2600                 zio->io_error = 0;
2601                 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2602                     ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2603                 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2604                 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2605                     zio_requeue_io_start_cut_in_line);
2606                 return (ZIO_PIPELINE_STOP);
2607         }
2608 
2609         /*
2610          * If we got an error on a leaf device, convert it to ENXIO
2611          * if the device is not accessible at all.
2612          */
2613         if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2614             !vdev_accessible(vd, zio))
2615                 zio->io_error = SET_ERROR(ENXIO);
2616 
2617         /*
2618          * If we can't write to an interior vdev (mirror or RAID-Z),
2619          * set vdev_cant_write so that we stop trying to allocate from it.
2620          */
2621         if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2622             vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
2623                 vd->vdev_cant_write = B_TRUE;
2624         }
2625 
2626         if (zio->io_error)
2627                 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2628 
2629         if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2630             zio->io_physdone != NULL) {
2631                 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
2632                 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
2633                 zio->io_physdone(zio->io_logical);
2634         }
2635 
2636         return (ZIO_PIPELINE_CONTINUE);
2637 }
2638 
2639 void
2640 zio_vdev_io_reissue(zio_t *zio)
2641 {
2642         ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2643         ASSERT(zio->io_error == 0);
2644 
2645         zio->io_stage >>= 1;
2646 }
2647 
2648 void
2649 zio_vdev_io_redone(zio_t *zio)
2650 {
2651         ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2652 
2653         zio->io_stage >>= 1;
2654 }
2655 
2656 void
2657 zio_vdev_io_bypass(zio_t *zio)
2658 {
2659         ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2660         ASSERT(zio->io_error == 0);
2661 
2662         zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2663         zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2664 }
2665 
2666 /*
2667  * ==========================================================================
2668  * Generate and verify checksums
2669  * ==========================================================================
2670  */
2671 static int
2672 zio_checksum_generate(zio_t *zio)
2673 {
2674         blkptr_t *bp = zio->io_bp;
2675         enum zio_checksum checksum;
2676 
2677         if (bp == NULL) {
2678                 /*
2679                  * This is zio_write_phys().
2680                  * We're either generating a label checksum, or none at all.
2681                  */
2682                 checksum = zio->io_prop.zp_checksum;
2683 
2684                 if (checksum == ZIO_CHECKSUM_OFF)
2685                         return (ZIO_PIPELINE_CONTINUE);
2686 
2687                 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2688         } else {
2689                 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2690                         ASSERT(!IO_IS_ALLOCATING(zio));
2691                         checksum = ZIO_CHECKSUM_GANG_HEADER;
2692                 } else {
2693                         checksum = BP_GET_CHECKSUM(bp);
2694                 }
2695         }
2696 
2697         zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2698 
2699         return (ZIO_PIPELINE_CONTINUE);
2700 }
2701 
2702 static int
2703 zio_checksum_verify(zio_t *zio)
2704 {
2705         zio_bad_cksum_t info;
2706         blkptr_t *bp = zio->io_bp;
2707         int error;
2708 
2709         ASSERT(zio->io_vd != NULL);
2710 
2711         if (bp == NULL) {
2712                 /*
2713                  * This is zio_read_phys().
2714                  * We're either verifying a label checksum, or nothing at all.
2715                  */
2716                 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2717                         return (ZIO_PIPELINE_CONTINUE);
2718 
2719                 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2720         }
2721 
2722         if ((error = zio_checksum_error(zio, &info)) != 0) {
2723                 zio->io_error = error;
2724                 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2725                         zfs_ereport_start_checksum(zio->io_spa,
2726                             zio->io_vd, zio, zio->io_offset,
2727                             zio->io_size, NULL, &info);
2728                 }
2729         }
2730 
2731         return (ZIO_PIPELINE_CONTINUE);
2732 }
2733 
2734 /*
2735  * Called by RAID-Z to ensure we don't compute the checksum twice.
2736  */
2737 void
2738 zio_checksum_verified(zio_t *zio)
2739 {
2740         zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2741 }
2742 
2743 /*
2744  * ==========================================================================
2745  * Error rank.  Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2746  * An error of 0 indictes success.  ENXIO indicates whole-device failure,
2747  * which may be transient (e.g. unplugged) or permament.  ECKSUM and EIO
2748  * indicate errors that are specific to one I/O, and most likely permanent.
2749  * Any other error is presumed to be worse because we weren't expecting it.
2750  * ==========================================================================
2751  */
2752 int
2753 zio_worst_error(int e1, int e2)
2754 {
2755         static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2756         int r1, r2;
2757 
2758         for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2759                 if (e1 == zio_error_rank[r1])
2760                         break;
2761 
2762         for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2763                 if (e2 == zio_error_rank[r2])
2764                         break;
2765 
2766         return (r1 > r2 ? e1 : e2);
2767 }
2768 
2769 /*
2770  * ==========================================================================
2771  * I/O completion
2772  * ==========================================================================
2773  */
2774 static int
2775 zio_ready(zio_t *zio)
2776 {
2777         blkptr_t *bp = zio->io_bp;
2778         zio_t *pio, *pio_next;
2779 
2780         if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2781             zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2782                 return (ZIO_PIPELINE_STOP);
2783 
2784         if (zio->io_ready) {
2785                 ASSERT(IO_IS_ALLOCATING(zio));
2786                 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
2787                     (zio->io_flags & ZIO_FLAG_NOPWRITE));
2788                 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2789 
2790                 zio->io_ready(zio);
2791         }
2792 
2793         if (bp != NULL && bp != &zio->io_bp_copy)
2794                 zio->io_bp_copy = *bp;
2795 
2796         if (zio->io_error)
2797                 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2798 
2799         mutex_enter(&zio->io_lock);
2800         zio->io_state[ZIO_WAIT_READY] = 1;
2801         pio = zio_walk_parents(zio);
2802         mutex_exit(&zio->io_lock);
2803 
2804         /*
2805          * As we notify zio's parents, new parents could be added.
2806          * New parents go to the head of zio's io_parent_list, however,
2807          * so we will (correctly) not notify them.  The remainder of zio's
2808          * io_parent_list, from 'pio_next' onward, cannot change because
2809          * all parents must wait for us to be done before they can be done.
2810          */
2811         for (; pio != NULL; pio = pio_next) {
2812                 pio_next = zio_walk_parents(zio);
2813                 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2814         }
2815 
2816         if (zio->io_flags & ZIO_FLAG_NODATA) {
2817                 if (BP_IS_GANG(bp)) {
2818                         zio->io_flags &= ~ZIO_FLAG_NODATA;
2819                 } else {
2820                         ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2821                         zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2822                 }
2823         }
2824 
2825         if (zio_injection_enabled &&
2826             zio->io_spa->spa_syncing_txg == zio->io_txg)
2827                 zio_handle_ignored_writes(zio);
2828 
2829         return (ZIO_PIPELINE_CONTINUE);
2830 }
2831 
2832 static int
2833 zio_done(zio_t *zio)
2834 {
2835         spa_t *spa = zio->io_spa;
2836         zio_t *lio = zio->io_logical;
2837         blkptr_t *bp = zio->io_bp;
2838         vdev_t *vd = zio->io_vd;
2839         uint64_t psize = zio->io_size;
2840         zio_t *pio, *pio_next;
2841 
2842         /*
2843          * If our children haven't all completed,
2844          * wait for them and then repeat this pipeline stage.
2845          */
2846         if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2847             zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2848             zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2849             zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2850                 return (ZIO_PIPELINE_STOP);
2851 
2852         for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2853                 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2854                         ASSERT(zio->io_children[c][w] == 0);
2855 
2856         if (bp != NULL) {
2857                 ASSERT(bp->blk_pad[0] == 0);
2858                 ASSERT(bp->blk_pad[1] == 0);
2859                 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2860                     (bp == zio_unique_parent(zio)->io_bp));
2861                 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
2862                     zio->io_bp_override == NULL &&
2863                     !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2864                         ASSERT(!BP_SHOULD_BYTESWAP(bp));
2865                         ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
2866                         ASSERT(BP_COUNT_GANG(bp) == 0 ||
2867                             (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
2868                 }
2869                 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
2870                         VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
2871         }
2872 
2873         /*
2874          * If there were child vdev/gang/ddt errors, they apply to us now.
2875          */
2876         zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2877         zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2878         zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2879 
2880         /*
2881          * If the I/O on the transformed data was successful, generate any
2882          * checksum reports now while we still have the transformed data.
2883          */
2884         if (zio->io_error == 0) {
2885                 while (zio->io_cksum_report != NULL) {
2886                         zio_cksum_report_t *zcr = zio->io_cksum_report;
2887                         uint64_t align = zcr->zcr_align;
2888                         uint64_t asize = P2ROUNDUP(psize, align);
2889                         char *abuf = zio->io_data;
2890 
2891                         if (asize != psize) {
2892                                 abuf = zio_buf_alloc(asize);
2893                                 bcopy(zio->io_data, abuf, psize);
2894                                 bzero(abuf + psize, asize - psize);
2895                         }
2896 
2897                         zio->io_cksum_report = zcr->zcr_next;
2898                         zcr->zcr_next = NULL;
2899                         zcr->zcr_finish(zcr, abuf);
2900                         zfs_ereport_free_checksum(zcr);
2901 
2902                         if (asize != psize)
2903                                 zio_buf_free(abuf, asize);
2904                 }
2905         }
2906 
2907         zio_pop_transforms(zio);        /* note: may set zio->io_error */
2908 
2909         vdev_stat_update(zio, psize);
2910 
2911         if (zio->io_error) {
2912                 /*
2913                  * If this I/O is attached to a particular vdev,
2914                  * generate an error message describing the I/O failure
2915                  * at the block level.  We ignore these errors if the
2916                  * device is currently unavailable.
2917                  */
2918                 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
2919                         zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
2920 
2921                 if ((zio->io_error == EIO || !(zio->io_flags &
2922                     (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2923                     zio == lio) {
2924                         /*
2925                          * For logical I/O requests, tell the SPA to log the
2926                          * error and generate a logical data ereport.
2927                          */
2928                         spa_log_error(spa, zio);
2929                         zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
2930                             0, 0);
2931                 }
2932         }
2933 
2934         if (zio->io_error && zio == lio) {
2935                 /*
2936                  * Determine whether zio should be reexecuted.  This will
2937                  * propagate all the way to the root via zio_notify_parent().
2938                  */
2939                 ASSERT(vd == NULL && bp != NULL);
2940                 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2941 
2942                 if (IO_IS_ALLOCATING(zio) &&
2943                     !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2944                         if (zio->io_error != ENOSPC)
2945                                 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2946                         else
2947                                 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2948                 }
2949 
2950                 if ((zio->io_type == ZIO_TYPE_READ ||
2951                     zio->io_type == ZIO_TYPE_FREE) &&
2952                     !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
2953                     zio->io_error == ENXIO &&
2954                     spa_load_state(spa) == SPA_LOAD_NONE &&
2955                     spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
2956                         zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2957 
2958                 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2959                         zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2960 
2961                 /*
2962                  * Here is a possibly good place to attempt to do
2963                  * either combinatorial reconstruction or error correction
2964                  * based on checksums.  It also might be a good place
2965                  * to send out preliminary ereports before we suspend
2966                  * processing.
2967                  */
2968         }
2969 
2970         /*
2971          * If there were logical child errors, they apply to us now.
2972          * We defer this until now to avoid conflating logical child
2973          * errors with errors that happened to the zio itself when
2974          * updating vdev stats and reporting FMA events above.
2975          */
2976         zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2977 
2978         if ((zio->io_error || zio->io_reexecute) &&
2979             IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
2980             !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
2981                 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
2982 
2983         zio_gang_tree_free(&zio->io_gang_tree);
2984 
2985         /*
2986          * Godfather I/Os should never suspend.
2987          */
2988         if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
2989             (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
2990                 zio->io_reexecute = 0;
2991 
2992         if (zio->io_reexecute) {
2993                 /*
2994                  * This is a logical I/O that wants to reexecute.
2995                  *
2996                  * Reexecute is top-down.  When an i/o fails, if it's not
2997                  * the root, it simply notifies its parent and sticks around.
2998                  * The parent, seeing that it still has children in zio_done(),
2999                  * does the same.  This percolates all the way up to the root.
3000                  * The root i/o will reexecute or suspend the entire tree.
3001                  *
3002                  * This approach ensures that zio_reexecute() honors
3003                  * all the original i/o dependency relationships, e.g.
3004                  * parents not executing until children are ready.
3005                  */
3006                 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3007 
3008                 zio->io_gang_leader = NULL;
3009 
3010                 mutex_enter(&zio->io_lock);
3011                 zio->io_state[ZIO_WAIT_DONE] = 1;
3012                 mutex_exit(&zio->io_lock);
3013 
3014                 /*
3015                  * "The Godfather" I/O monitors its children but is
3016                  * not a true parent to them. It will track them through
3017                  * the pipeline but severs its ties whenever they get into
3018                  * trouble (e.g. suspended). This allows "The Godfather"
3019                  * I/O to return status without blocking.
3020                  */
3021                 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3022                         zio_link_t *zl = zio->io_walk_link;
3023                         pio_next = zio_walk_parents(zio);
3024 
3025                         if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3026                             (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3027                                 zio_remove_child(pio, zio, zl);
3028                                 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3029                         }
3030                 }
3031 
3032                 if ((pio = zio_unique_parent(zio)) != NULL) {
3033                         /*
3034                          * We're not a root i/o, so there's nothing to do
3035                          * but notify our parent.  Don't propagate errors
3036                          * upward since we haven't permanently failed yet.
3037                          */
3038                         ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3039                         zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3040                         zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3041                 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3042                         /*
3043                          * We'd fail again if we reexecuted now, so suspend
3044                          * until conditions improve (e.g. device comes online).
3045                          */
3046                         zio_suspend(spa, zio);
3047                 } else {
3048                         /*
3049                          * Reexecution is potentially a huge amount of work.
3050                          * Hand it off to the otherwise-unused claim taskq.
3051                          */
3052                         ASSERT(zio->io_tqent.tqent_next == NULL);
3053                         spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3054                             ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3055                             0, &zio->io_tqent);
3056                 }
3057                 return (ZIO_PIPELINE_STOP);
3058         }
3059 
3060         ASSERT(zio->io_child_count == 0);
3061         ASSERT(zio->io_reexecute == 0);
3062         ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3063 
3064         /*
3065          * Report any checksum errors, since the I/O is complete.
3066          */
3067         while (zio->io_cksum_report != NULL) {
3068                 zio_cksum_report_t *zcr = zio->io_cksum_report;
3069                 zio->io_cksum_report = zcr->zcr_next;
3070                 zcr->zcr_next = NULL;
3071                 zcr->zcr_finish(zcr, NULL);
3072                 zfs_ereport_free_checksum(zcr);
3073         }
3074 
3075         /*
3076          * It is the responsibility of the done callback to ensure that this
3077          * particular zio is no longer discoverable for adoption, and as
3078          * such, cannot acquire any new parents.
3079          */
3080         if (zio->io_done)
3081                 zio->io_done(zio);
3082 
3083         mutex_enter(&zio->io_lock);
3084         zio->io_state[ZIO_WAIT_DONE] = 1;
3085         mutex_exit(&zio->io_lock);
3086 
3087         for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3088                 zio_link_t *zl = zio->io_walk_link;
3089                 pio_next = zio_walk_parents(zio);
3090                 zio_remove_child(pio, zio, zl);
3091                 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3092         }
3093 
3094         if (zio->io_waiter != NULL) {
3095                 mutex_enter(&zio->io_lock);
3096                 zio->io_executor = NULL;
3097                 cv_broadcast(&zio->io_cv);
3098                 mutex_exit(&zio->io_lock);
3099         } else {
3100                 zio_destroy(zio);
3101         }
3102 
3103         return (ZIO_PIPELINE_STOP);
3104 }
3105 
3106 /*
3107  * ==========================================================================
3108  * I/O pipeline definition
3109  * ==========================================================================
3110  */
3111 static zio_pipe_stage_t *zio_pipeline[] = {
3112         NULL,
3113         zio_read_bp_init,
3114         zio_free_bp_init,
3115         zio_issue_async,
3116         zio_write_bp_init,
3117         zio_checksum_generate,
3118         zio_nop_write,
3119         zio_ddt_read_start,
3120         zio_ddt_read_done,
3121         zio_ddt_write,
3122         zio_ddt_free,
3123         zio_gang_assemble,
3124         zio_gang_issue,
3125         zio_dva_allocate,
3126         zio_dva_free,
3127         zio_dva_claim,
3128         zio_ready,
3129         zio_vdev_io_start,
3130         zio_vdev_io_done,
3131         zio_vdev_io_assess,
3132         zio_checksum_verify,
3133         zio_done
3134 };
3135 
3136 /* dnp is the dnode for zb1->zb_object */
3137 boolean_t
3138 zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_t *zb1,
3139     const zbookmark_t *zb2)
3140 {
3141         uint64_t zb1nextL0, zb2thisobj;
3142 
3143         ASSERT(zb1->zb_objset == zb2->zb_objset);
3144         ASSERT(zb2->zb_level == 0);
3145 
3146         /*
3147          * A bookmark in the deadlist is considered to be after
3148          * everything else.
3149          */
3150         if (zb2->zb_object == DMU_DEADLIST_OBJECT)
3151                 return (B_TRUE);
3152 
3153         /* The objset_phys_t isn't before anything. */
3154         if (dnp == NULL)
3155                 return (B_FALSE);
3156 
3157         zb1nextL0 = (zb1->zb_blkid + 1) <<
3158             ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
3159 
3160         zb2thisobj = zb2->zb_object ? zb2->zb_object :
3161             zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
3162 
3163         if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3164                 uint64_t nextobj = zb1nextL0 *
3165                     (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
3166                 return (nextobj <= zb2thisobj);
3167         }
3168 
3169         if (zb1->zb_object < zb2thisobj)
3170                 return (B_TRUE);
3171         if (zb1->zb_object > zb2thisobj)
3172                 return (B_FALSE);
3173         if (zb2->zb_object == DMU_META_DNODE_OBJECT)
3174                 return (B_FALSE);
3175         return (zb1nextL0 <= zb2->zb_blkid);
3176 }