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.
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.
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]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011 by Delphix. All rights reserved.
26 #include <sys/zfs_context.h>
27 #include <sys/fm/fs/zfs.h>
30 #include <sys/spa_impl.h>
31 #include <sys/vdev_impl.h>
32 #include <sys/zio_impl.h>
33 #include <sys/zio_compress.h>
34 #include <sys/zio_checksum.h>
35 #include <sys/dmu_objset.h>
39 SYSCTL_DECL(_vfs_zfs);
40 SYSCTL_NODE(_vfs_zfs, OID_AUTO, zio, CTLFLAG_RW, 0, "ZFS ZIO");
41 static int zio_use_uma = 0;
42 TUNABLE_INT("vfs.zfs.zio.use_uma", &zio_use_uma);
43 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, use_uma, CTLFLAG_RDTUN, &zio_use_uma, 0,
44 "Use uma(9) for ZIO allocations");
47 * ==========================================================================
49 * ==========================================================================
51 uint8_t zio_priority_table[ZIO_PRIORITY_TABLE_SIZE] = {
52 0, /* ZIO_PRIORITY_NOW */
53 0, /* ZIO_PRIORITY_SYNC_READ */
54 0, /* ZIO_PRIORITY_SYNC_WRITE */
55 0, /* ZIO_PRIORITY_LOG_WRITE */
56 1, /* ZIO_PRIORITY_CACHE_FILL */
57 1, /* ZIO_PRIORITY_AGG */
58 4, /* ZIO_PRIORITY_FREE */
59 4, /* ZIO_PRIORITY_ASYNC_WRITE */
60 6, /* ZIO_PRIORITY_ASYNC_READ */
61 10, /* ZIO_PRIORITY_RESILVER */
62 20, /* ZIO_PRIORITY_SCRUB */
63 2, /* ZIO_PRIORITY_DDT_PREFETCH */
67 * ==========================================================================
68 * I/O type descriptions
69 * ==========================================================================
71 char *zio_type_name[ZIO_TYPES] = {
72 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
77 * ==========================================================================
79 * ==========================================================================
81 kmem_cache_t *zio_cache;
82 kmem_cache_t *zio_link_cache;
83 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
84 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
87 extern vmem_t *zio_alloc_arena;
89 extern int zfs_mg_alloc_failures;
92 * An allocating zio is one that either currently has the DVA allocate
93 * stage set or will have it later in its lifetime.
95 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
97 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
100 int zio_buf_debug_limit = 16384;
102 int zio_buf_debug_limit = 0;
109 zio_cache = kmem_cache_create("zio_cache",
110 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
111 zio_link_cache = kmem_cache_create("zio_link_cache",
112 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
115 * For small buffers, we want a cache for each multiple of
116 * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache
117 * for each quarter-power of 2. For large buffers, we want
118 * a cache for each multiple of PAGESIZE.
120 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
121 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
124 size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
126 while (p2 & (p2 - 1))
129 if (size <= 4 * SPA_MINBLOCKSIZE) {
130 align = SPA_MINBLOCKSIZE;
131 } else if (P2PHASE(size, PAGESIZE) == 0) {
133 } else if (P2PHASE(size, p2 >> 2) == 0) {
139 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
140 zio_buf_cache[c] = kmem_cache_create(name, size,
141 align, NULL, NULL, NULL, NULL, NULL, cflags);
144 * Since zio_data bufs do not appear in crash dumps, we
145 * pass KMC_NOTOUCH so that no allocator metadata is
146 * stored with the buffers.
148 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
149 zio_data_buf_cache[c] = kmem_cache_create(name, size,
150 align, NULL, NULL, NULL, NULL, NULL,
151 cflags | KMC_NOTOUCH);
156 ASSERT(zio_buf_cache[c] != NULL);
157 if (zio_buf_cache[c - 1] == NULL)
158 zio_buf_cache[c - 1] = zio_buf_cache[c];
160 ASSERT(zio_data_buf_cache[c] != NULL);
161 if (zio_data_buf_cache[c - 1] == NULL)
162 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
166 * The zio write taskqs have 1 thread per cpu, allow 1/2 of the taskqs
167 * to fail 3 times per txg or 8 failures, whichever is greater.
169 if (zfs_mg_alloc_failures == 0)
170 zfs_mg_alloc_failures = MAX((3 * max_ncpus / 2), 8);
171 else if (zfs_mg_alloc_failures < 8)
172 zfs_mg_alloc_failures = 8;
181 kmem_cache_t *last_cache = NULL;
182 kmem_cache_t *last_data_cache = NULL;
184 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
185 if (zio_buf_cache[c] != last_cache) {
186 last_cache = zio_buf_cache[c];
187 kmem_cache_destroy(zio_buf_cache[c]);
189 zio_buf_cache[c] = NULL;
191 if (zio_data_buf_cache[c] != last_data_cache) {
192 last_data_cache = zio_data_buf_cache[c];
193 kmem_cache_destroy(zio_data_buf_cache[c]);
195 zio_data_buf_cache[c] = NULL;
198 kmem_cache_destroy(zio_link_cache);
199 kmem_cache_destroy(zio_cache);
205 * ==========================================================================
206 * Allocate and free I/O buffers
207 * ==========================================================================
211 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
212 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
213 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
214 * excess / transient data in-core during a crashdump.
217 zio_buf_alloc(size_t size)
219 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
221 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
224 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
226 return (kmem_alloc(size, KM_SLEEP));
230 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
231 * crashdump if the kernel panics. This exists so that we will limit the amount
232 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
233 * of kernel heap dumped to disk when the kernel panics)
236 zio_data_buf_alloc(size_t size)
238 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
240 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
243 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
245 return (kmem_alloc(size, KM_SLEEP));
249 zio_buf_free(void *buf, size_t size)
251 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
253 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
256 kmem_cache_free(zio_buf_cache[c], buf);
258 kmem_free(buf, size);
262 zio_data_buf_free(void *buf, size_t size)
264 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
266 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
269 kmem_cache_free(zio_data_buf_cache[c], buf);
271 kmem_free(buf, size);
275 * ==========================================================================
276 * Push and pop I/O transform buffers
277 * ==========================================================================
280 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
281 zio_transform_func_t *transform)
283 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
285 zt->zt_orig_data = zio->io_data;
286 zt->zt_orig_size = zio->io_size;
287 zt->zt_bufsize = bufsize;
288 zt->zt_transform = transform;
290 zt->zt_next = zio->io_transform_stack;
291 zio->io_transform_stack = zt;
298 zio_pop_transforms(zio_t *zio)
302 while ((zt = zio->io_transform_stack) != NULL) {
303 if (zt->zt_transform != NULL)
304 zt->zt_transform(zio,
305 zt->zt_orig_data, zt->zt_orig_size);
307 if (zt->zt_bufsize != 0)
308 zio_buf_free(zio->io_data, zt->zt_bufsize);
310 zio->io_data = zt->zt_orig_data;
311 zio->io_size = zt->zt_orig_size;
312 zio->io_transform_stack = zt->zt_next;
314 kmem_free(zt, sizeof (zio_transform_t));
319 * ==========================================================================
320 * I/O transform callbacks for subblocks and decompression
321 * ==========================================================================
324 zio_subblock(zio_t *zio, void *data, uint64_t size)
326 ASSERT(zio->io_size > size);
328 if (zio->io_type == ZIO_TYPE_READ)
329 bcopy(zio->io_data, data, size);
333 zio_decompress(zio_t *zio, void *data, uint64_t size)
335 if (zio->io_error == 0 &&
336 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
337 zio->io_data, data, zio->io_size, size) != 0)
342 * ==========================================================================
343 * I/O parent/child relationships and pipeline interlocks
344 * ==========================================================================
347 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
348 * continue calling these functions until they return NULL.
349 * Otherwise, the next caller will pick up the list walk in
350 * some indeterminate state. (Otherwise every caller would
351 * have to pass in a cookie to keep the state represented by
352 * io_walk_link, which gets annoying.)
355 zio_walk_parents(zio_t *cio)
357 zio_link_t *zl = cio->io_walk_link;
358 list_t *pl = &cio->io_parent_list;
360 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
361 cio->io_walk_link = zl;
366 ASSERT(zl->zl_child == cio);
367 return (zl->zl_parent);
371 zio_walk_children(zio_t *pio)
373 zio_link_t *zl = pio->io_walk_link;
374 list_t *cl = &pio->io_child_list;
376 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
377 pio->io_walk_link = zl;
382 ASSERT(zl->zl_parent == pio);
383 return (zl->zl_child);
387 zio_unique_parent(zio_t *cio)
389 zio_t *pio = zio_walk_parents(cio);
391 VERIFY(zio_walk_parents(cio) == NULL);
396 zio_add_child(zio_t *pio, zio_t *cio)
398 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
401 * Logical I/Os can have logical, gang, or vdev children.
402 * Gang I/Os can have gang or vdev children.
403 * Vdev I/Os can only have vdev children.
404 * The following ASSERT captures all of these constraints.
406 ASSERT(cio->io_child_type <= pio->io_child_type);
411 mutex_enter(&cio->io_lock);
412 mutex_enter(&pio->io_lock);
414 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
416 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
417 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
419 list_insert_head(&pio->io_child_list, zl);
420 list_insert_head(&cio->io_parent_list, zl);
422 pio->io_child_count++;
423 cio->io_parent_count++;
425 mutex_exit(&pio->io_lock);
426 mutex_exit(&cio->io_lock);
430 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
432 ASSERT(zl->zl_parent == pio);
433 ASSERT(zl->zl_child == cio);
435 mutex_enter(&cio->io_lock);
436 mutex_enter(&pio->io_lock);
438 list_remove(&pio->io_child_list, zl);
439 list_remove(&cio->io_parent_list, zl);
441 pio->io_child_count--;
442 cio->io_parent_count--;
444 mutex_exit(&pio->io_lock);
445 mutex_exit(&cio->io_lock);
447 kmem_cache_free(zio_link_cache, zl);
451 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
453 uint64_t *countp = &zio->io_children[child][wait];
454 boolean_t waiting = B_FALSE;
456 mutex_enter(&zio->io_lock);
457 ASSERT(zio->io_stall == NULL);
460 zio->io_stall = countp;
463 mutex_exit(&zio->io_lock);
469 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
471 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
472 int *errorp = &pio->io_child_error[zio->io_child_type];
474 mutex_enter(&pio->io_lock);
475 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
476 *errorp = zio_worst_error(*errorp, zio->io_error);
477 pio->io_reexecute |= zio->io_reexecute;
478 ASSERT3U(*countp, >, 0);
479 if (--*countp == 0 && pio->io_stall == countp) {
480 pio->io_stall = NULL;
481 mutex_exit(&pio->io_lock);
484 mutex_exit(&pio->io_lock);
489 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
491 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
492 zio->io_error = zio->io_child_error[c];
496 * ==========================================================================
497 * Create the various types of I/O (read, write, free, etc)
498 * ==========================================================================
501 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
502 void *data, uint64_t size, zio_done_func_t *done, void *private,
503 zio_type_t type, int priority, enum zio_flag flags,
504 vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
505 enum zio_stage stage, enum zio_stage pipeline)
509 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
510 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
511 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
513 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
514 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
515 ASSERT(vd || stage == ZIO_STAGE_OPEN);
517 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
518 bzero(zio, sizeof (zio_t));
520 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
521 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
523 list_create(&zio->io_parent_list, sizeof (zio_link_t),
524 offsetof(zio_link_t, zl_parent_node));
525 list_create(&zio->io_child_list, sizeof (zio_link_t),
526 offsetof(zio_link_t, zl_child_node));
529 zio->io_child_type = ZIO_CHILD_VDEV;
530 else if (flags & ZIO_FLAG_GANG_CHILD)
531 zio->io_child_type = ZIO_CHILD_GANG;
532 else if (flags & ZIO_FLAG_DDT_CHILD)
533 zio->io_child_type = ZIO_CHILD_DDT;
535 zio->io_child_type = ZIO_CHILD_LOGICAL;
538 zio->io_bp = (blkptr_t *)bp;
539 zio->io_bp_copy = *bp;
540 zio->io_bp_orig = *bp;
541 if (type != ZIO_TYPE_WRITE ||
542 zio->io_child_type == ZIO_CHILD_DDT)
543 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
544 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
545 zio->io_logical = zio;
546 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
547 pipeline |= ZIO_GANG_STAGES;
553 zio->io_private = private;
555 zio->io_priority = priority;
557 zio->io_offset = offset;
558 zio->io_orig_data = zio->io_data = data;
559 zio->io_orig_size = zio->io_size = size;
560 zio->io_orig_flags = zio->io_flags = flags;
561 zio->io_orig_stage = zio->io_stage = stage;
562 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
564 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
565 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
568 zio->io_bookmark = *zb;
571 if (zio->io_logical == NULL)
572 zio->io_logical = pio->io_logical;
573 if (zio->io_child_type == ZIO_CHILD_GANG)
574 zio->io_gang_leader = pio->io_gang_leader;
575 zio_add_child(pio, zio);
582 zio_destroy(zio_t *zio)
584 list_destroy(&zio->io_parent_list);
585 list_destroy(&zio->io_child_list);
586 mutex_destroy(&zio->io_lock);
587 cv_destroy(&zio->io_cv);
588 kmem_cache_free(zio_cache, zio);
592 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
593 void *private, enum zio_flag flags)
597 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
598 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
599 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
605 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
607 return (zio_null(NULL, spa, NULL, done, private, flags));
611 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
612 void *data, uint64_t size, zio_done_func_t *done, void *private,
613 int priority, enum zio_flag flags, const zbookmark_t *zb)
617 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
618 data, size, done, private,
619 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
620 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
621 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
627 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
628 void *data, uint64_t size, const zio_prop_t *zp,
629 zio_done_func_t *ready, zio_done_func_t *done, void *private,
630 int priority, enum zio_flag flags, const zbookmark_t *zb)
634 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
635 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
636 zp->zp_compress >= ZIO_COMPRESS_OFF &&
637 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
638 zp->zp_type < DMU_OT_NUMTYPES &&
641 zp->zp_copies <= spa_max_replication(spa) &&
643 zp->zp_dedup_verify <= 1);
645 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
646 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
647 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
648 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
650 zio->io_ready = ready;
657 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
658 uint64_t size, zio_done_func_t *done, void *private, int priority,
659 enum zio_flag flags, zbookmark_t *zb)
663 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
664 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
665 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
671 zio_write_override(zio_t *zio, blkptr_t *bp, int copies)
673 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
674 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
675 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
676 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
678 zio->io_prop.zp_copies = copies;
679 zio->io_bp_override = bp;
683 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
685 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
689 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
694 dprintf_bp(bp, "freeing in txg %llu, pass %u",
695 (longlong_t)txg, spa->spa_sync_pass);
697 ASSERT(!BP_IS_HOLE(bp));
698 ASSERT(spa_syncing_txg(spa) == txg);
699 ASSERT(spa_sync_pass(spa) <= SYNC_PASS_DEFERRED_FREE);
701 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
702 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
703 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
709 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
710 zio_done_func_t *done, void *private, enum zio_flag flags)
715 * A claim is an allocation of a specific block. Claims are needed
716 * to support immediate writes in the intent log. The issue is that
717 * immediate writes contain committed data, but in a txg that was
718 * *not* committed. Upon opening the pool after an unclean shutdown,
719 * the intent log claims all blocks that contain immediate write data
720 * so that the SPA knows they're in use.
722 * All claims *must* be resolved in the first txg -- before the SPA
723 * starts allocating blocks -- so that nothing is allocated twice.
724 * If txg == 0 we just verify that the block is claimable.
726 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
727 ASSERT(txg == spa_first_txg(spa) || txg == 0);
728 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
730 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
731 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
732 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
738 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
739 zio_done_func_t *done, void *private, int priority, enum zio_flag flags)
744 if (vd->vdev_children == 0) {
745 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
746 ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
747 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
751 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
753 for (c = 0; c < vd->vdev_children; c++)
754 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
755 done, private, priority, flags));
762 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
763 void *data, int checksum, zio_done_func_t *done, void *private,
764 int priority, enum zio_flag flags, boolean_t labels)
768 ASSERT(vd->vdev_children == 0);
769 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
770 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
771 ASSERT3U(offset + size, <=, vd->vdev_psize);
773 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
774 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
775 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
777 zio->io_prop.zp_checksum = checksum;
783 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
784 void *data, int checksum, zio_done_func_t *done, void *private,
785 int priority, enum zio_flag flags, boolean_t labels)
789 ASSERT(vd->vdev_children == 0);
790 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
791 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
792 ASSERT3U(offset + size, <=, vd->vdev_psize);
794 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
795 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
796 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
798 zio->io_prop.zp_checksum = checksum;
800 if (zio_checksum_table[checksum].ci_eck) {
802 * zec checksums are necessarily destructive -- they modify
803 * the end of the write buffer to hold the verifier/checksum.
804 * Therefore, we must make a local copy in case the data is
805 * being written to multiple places in parallel.
807 void *wbuf = zio_buf_alloc(size);
808 bcopy(data, wbuf, size);
809 zio_push_transform(zio, wbuf, size, size, NULL);
816 * Create a child I/O to do some work for us.
819 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
820 void *data, uint64_t size, int type, int priority, enum zio_flag flags,
821 zio_done_func_t *done, void *private)
823 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
826 ASSERT(vd->vdev_parent ==
827 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
829 if (type == ZIO_TYPE_READ && bp != NULL) {
831 * If we have the bp, then the child should perform the
832 * checksum and the parent need not. This pushes error
833 * detection as close to the leaves as possible and
834 * eliminates redundant checksums in the interior nodes.
836 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
837 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
840 if (vd->vdev_children == 0)
841 offset += VDEV_LABEL_START_SIZE;
843 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
846 * If we've decided to do a repair, the write is not speculative --
847 * even if the original read was.
849 if (flags & ZIO_FLAG_IO_REPAIR)
850 flags &= ~ZIO_FLAG_SPECULATIVE;
852 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
853 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
854 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
860 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
861 int type, int priority, enum zio_flag flags,
862 zio_done_func_t *done, void *private)
866 ASSERT(vd->vdev_ops->vdev_op_leaf);
868 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
869 data, size, done, private, type, priority,
870 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
872 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
878 zio_flush(zio_t *zio, vdev_t *vd)
880 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
881 NULL, NULL, ZIO_PRIORITY_NOW,
882 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
886 zio_shrink(zio_t *zio, uint64_t size)
888 ASSERT(zio->io_executor == NULL);
889 ASSERT(zio->io_orig_size == zio->io_size);
890 ASSERT(size <= zio->io_size);
893 * We don't shrink for raidz because of problems with the
894 * reconstruction when reading back less than the block size.
895 * Note, BP_IS_RAIDZ() assumes no compression.
897 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
898 if (!BP_IS_RAIDZ(zio->io_bp))
899 zio->io_orig_size = zio->io_size = size;
903 * ==========================================================================
904 * Prepare to read and write logical blocks
905 * ==========================================================================
909 zio_read_bp_init(zio_t *zio)
911 blkptr_t *bp = zio->io_bp;
913 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
914 zio->io_child_type == ZIO_CHILD_LOGICAL &&
915 !(zio->io_flags & ZIO_FLAG_RAW)) {
916 uint64_t psize = BP_GET_PSIZE(bp);
917 void *cbuf = zio_buf_alloc(psize);
919 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
922 if (!dmu_ot[BP_GET_TYPE(bp)].ot_metadata && BP_GET_LEVEL(bp) == 0)
923 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
925 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
926 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
928 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
929 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
931 return (ZIO_PIPELINE_CONTINUE);
935 zio_write_bp_init(zio_t *zio)
937 spa_t *spa = zio->io_spa;
938 zio_prop_t *zp = &zio->io_prop;
939 enum zio_compress compress = zp->zp_compress;
940 blkptr_t *bp = zio->io_bp;
941 uint64_t lsize = zio->io_size;
942 uint64_t psize = lsize;
946 * If our children haven't all reached the ready stage,
947 * wait for them and then repeat this pipeline stage.
949 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
950 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
951 return (ZIO_PIPELINE_STOP);
953 if (!IO_IS_ALLOCATING(zio))
954 return (ZIO_PIPELINE_CONTINUE);
956 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
958 if (zio->io_bp_override) {
959 ASSERT(bp->blk_birth != zio->io_txg);
960 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
962 *bp = *zio->io_bp_override;
963 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
965 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
966 return (ZIO_PIPELINE_CONTINUE);
968 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
969 zp->zp_dedup_verify);
971 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
973 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
974 return (ZIO_PIPELINE_CONTINUE);
976 zio->io_bp_override = NULL;
980 if (bp->blk_birth == zio->io_txg) {
982 * We're rewriting an existing block, which means we're
983 * working on behalf of spa_sync(). For spa_sync() to
984 * converge, it must eventually be the case that we don't
985 * have to allocate new blocks. But compression changes
986 * the blocksize, which forces a reallocate, and makes
987 * convergence take longer. Therefore, after the first
988 * few passes, stop compressing to ensure convergence.
990 pass = spa_sync_pass(spa);
992 ASSERT(zio->io_txg == spa_syncing_txg(spa));
993 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
994 ASSERT(!BP_GET_DEDUP(bp));
996 if (pass > SYNC_PASS_DONT_COMPRESS)
997 compress = ZIO_COMPRESS_OFF;
999 /* Make sure someone doesn't change their mind on overwrites */
1000 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
1001 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1004 if (compress != ZIO_COMPRESS_OFF) {
1005 void *cbuf = zio_buf_alloc(lsize);
1006 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1007 if (psize == 0 || psize == lsize) {
1008 compress = ZIO_COMPRESS_OFF;
1009 zio_buf_free(cbuf, lsize);
1011 ASSERT(psize < lsize);
1012 zio_push_transform(zio, cbuf, psize, lsize, NULL);
1017 * The final pass of spa_sync() must be all rewrites, but the first
1018 * few passes offer a trade-off: allocating blocks defers convergence,
1019 * but newly allocated blocks are sequential, so they can be written
1020 * to disk faster. Therefore, we allow the first few passes of
1021 * spa_sync() to allocate new blocks, but force rewrites after that.
1022 * There should only be a handful of blocks after pass 1 in any case.
1024 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
1025 pass > SYNC_PASS_REWRITE) {
1027 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1028 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1029 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1032 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1036 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1038 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1039 BP_SET_LSIZE(bp, lsize);
1040 BP_SET_PSIZE(bp, psize);
1041 BP_SET_COMPRESS(bp, compress);
1042 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1043 BP_SET_TYPE(bp, zp->zp_type);
1044 BP_SET_LEVEL(bp, zp->zp_level);
1045 BP_SET_DEDUP(bp, zp->zp_dedup);
1046 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1048 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1049 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1050 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1054 return (ZIO_PIPELINE_CONTINUE);
1058 zio_free_bp_init(zio_t *zio)
1060 blkptr_t *bp = zio->io_bp;
1062 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1063 if (BP_GET_DEDUP(bp))
1064 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1067 return (ZIO_PIPELINE_CONTINUE);
1071 * ==========================================================================
1072 * Execute the I/O pipeline
1073 * ==========================================================================
1077 zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q, boolean_t cutinline)
1079 spa_t *spa = zio->io_spa;
1080 zio_type_t t = zio->io_type;
1081 int flags = TQ_SLEEP | (cutinline ? TQ_FRONT : 0);
1083 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1086 * If we're a config writer or a probe, the normal issue and
1087 * interrupt threads may all be blocked waiting for the config lock.
1088 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1090 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1094 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1096 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1100 * If this is a high priority I/O, then use the high priority taskq.
1102 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1103 spa->spa_zio_taskq[t][q + 1] != NULL)
1106 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1108 (void) taskq_dispatch_safe(spa->spa_zio_taskq[t][q],
1109 (task_func_t *)zio_execute, zio, flags, &zio->io_task);
1111 (void) taskq_dispatch(spa->spa_zio_taskq[t][q],
1112 (task_func_t *)zio_execute, zio, flags);
1117 zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
1119 kthread_t *executor = zio->io_executor;
1120 spa_t *spa = zio->io_spa;
1122 for (zio_type_t t = 0; t < ZIO_TYPES; t++)
1123 if (taskq_member(spa->spa_zio_taskq[t][q], executor))
1130 zio_issue_async(zio_t *zio)
1132 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1134 return (ZIO_PIPELINE_STOP);
1138 zio_interrupt(zio_t *zio)
1140 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1144 * Execute the I/O pipeline until one of the following occurs:
1145 * (1) the I/O completes; (2) the pipeline stalls waiting for
1146 * dependent child I/Os; (3) the I/O issues, so we're waiting
1147 * for an I/O completion interrupt; (4) the I/O is delegated by
1148 * vdev-level caching or aggregation; (5) the I/O is deferred
1149 * due to vdev-level queueing; (6) the I/O is handed off to
1150 * another thread. In all cases, the pipeline stops whenever
1151 * there's no CPU work; it never burns a thread in cv_wait().
1153 * There's no locking on io_stage because there's no legitimate way
1154 * for multiple threads to be attempting to process the same I/O.
1156 static zio_pipe_stage_t *zio_pipeline[];
1159 zio_execute(zio_t *zio)
1161 zio->io_executor = curthread;
1163 while (zio->io_stage < ZIO_STAGE_DONE) {
1164 enum zio_stage pipeline = zio->io_pipeline;
1165 enum zio_stage stage = zio->io_stage;
1168 ASSERT(!MUTEX_HELD(&zio->io_lock));
1169 ASSERT(ISP2(stage));
1170 ASSERT(zio->io_stall == NULL);
1174 } while ((stage & pipeline) == 0);
1176 ASSERT(stage <= ZIO_STAGE_DONE);
1179 * If we are in interrupt context and this pipeline stage
1180 * will grab a config lock that is held across I/O,
1181 * or may wait for an I/O that needs an interrupt thread
1182 * to complete, issue async to avoid deadlock.
1184 * For VDEV_IO_START, we cut in line so that the io will
1185 * be sent to disk promptly.
1187 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1188 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1189 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1190 zio_requeue_io_start_cut_in_line : B_FALSE;
1191 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1195 zio->io_stage = stage;
1196 rv = zio_pipeline[highbit(stage) - 1](zio);
1198 if (rv == ZIO_PIPELINE_STOP)
1201 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1206 * ==========================================================================
1207 * Initiate I/O, either sync or async
1208 * ==========================================================================
1211 zio_wait(zio_t *zio)
1215 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1216 ASSERT(zio->io_executor == NULL);
1218 zio->io_waiter = curthread;
1222 mutex_enter(&zio->io_lock);
1223 while (zio->io_executor != NULL)
1224 cv_wait(&zio->io_cv, &zio->io_lock);
1225 mutex_exit(&zio->io_lock);
1227 error = zio->io_error;
1234 zio_nowait(zio_t *zio)
1236 ASSERT(zio->io_executor == NULL);
1238 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1239 zio_unique_parent(zio) == NULL) {
1241 * This is a logical async I/O with no parent to wait for it.
1242 * We add it to the spa_async_root_zio "Godfather" I/O which
1243 * will ensure they complete prior to unloading the pool.
1245 spa_t *spa = zio->io_spa;
1247 zio_add_child(spa->spa_async_zio_root, zio);
1254 * ==========================================================================
1255 * Reexecute or suspend/resume failed I/O
1256 * ==========================================================================
1260 zio_reexecute(zio_t *pio)
1262 zio_t *cio, *cio_next;
1264 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1265 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1266 ASSERT(pio->io_gang_leader == NULL);
1267 ASSERT(pio->io_gang_tree == NULL);
1269 pio->io_flags = pio->io_orig_flags;
1270 pio->io_stage = pio->io_orig_stage;
1271 pio->io_pipeline = pio->io_orig_pipeline;
1272 pio->io_reexecute = 0;
1274 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1275 pio->io_state[w] = 0;
1276 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1277 pio->io_child_error[c] = 0;
1279 if (IO_IS_ALLOCATING(pio))
1280 BP_ZERO(pio->io_bp);
1283 * As we reexecute pio's children, new children could be created.
1284 * New children go to the head of pio's io_child_list, however,
1285 * so we will (correctly) not reexecute them. The key is that
1286 * the remainder of pio's io_child_list, from 'cio_next' onward,
1287 * cannot be affected by any side effects of reexecuting 'cio'.
1289 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1290 cio_next = zio_walk_children(pio);
1291 mutex_enter(&pio->io_lock);
1292 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1293 pio->io_children[cio->io_child_type][w]++;
1294 mutex_exit(&pio->io_lock);
1299 * Now that all children have been reexecuted, execute the parent.
1300 * We don't reexecute "The Godfather" I/O here as it's the
1301 * responsibility of the caller to wait on him.
1303 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1308 zio_suspend(spa_t *spa, zio_t *zio)
1310 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1311 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1312 "failure and the failure mode property for this pool "
1313 "is set to panic.", spa_name(spa));
1315 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1317 mutex_enter(&spa->spa_suspend_lock);
1319 if (spa->spa_suspend_zio_root == NULL)
1320 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1321 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1322 ZIO_FLAG_GODFATHER);
1324 spa->spa_suspended = B_TRUE;
1327 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1328 ASSERT(zio != spa->spa_suspend_zio_root);
1329 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1330 ASSERT(zio_unique_parent(zio) == NULL);
1331 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1332 zio_add_child(spa->spa_suspend_zio_root, zio);
1335 mutex_exit(&spa->spa_suspend_lock);
1339 zio_resume(spa_t *spa)
1344 * Reexecute all previously suspended i/o.
1346 mutex_enter(&spa->spa_suspend_lock);
1347 spa->spa_suspended = B_FALSE;
1348 cv_broadcast(&spa->spa_suspend_cv);
1349 pio = spa->spa_suspend_zio_root;
1350 spa->spa_suspend_zio_root = NULL;
1351 mutex_exit(&spa->spa_suspend_lock);
1357 return (zio_wait(pio));
1361 zio_resume_wait(spa_t *spa)
1363 mutex_enter(&spa->spa_suspend_lock);
1364 while (spa_suspended(spa))
1365 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1366 mutex_exit(&spa->spa_suspend_lock);
1370 * ==========================================================================
1373 * A gang block is a collection of small blocks that looks to the DMU
1374 * like one large block. When zio_dva_allocate() cannot find a block
1375 * of the requested size, due to either severe fragmentation or the pool
1376 * being nearly full, it calls zio_write_gang_block() to construct the
1377 * block from smaller fragments.
1379 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1380 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1381 * an indirect block: it's an array of block pointers. It consumes
1382 * only one sector and hence is allocatable regardless of fragmentation.
1383 * The gang header's bps point to its gang members, which hold the data.
1385 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1386 * as the verifier to ensure uniqueness of the SHA256 checksum.
1387 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1388 * not the gang header. This ensures that data block signatures (needed for
1389 * deduplication) are independent of how the block is physically stored.
1391 * Gang blocks can be nested: a gang member may itself be a gang block.
1392 * Thus every gang block is a tree in which root and all interior nodes are
1393 * gang headers, and the leaves are normal blocks that contain user data.
1394 * The root of the gang tree is called the gang leader.
1396 * To perform any operation (read, rewrite, free, claim) on a gang block,
1397 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1398 * in the io_gang_tree field of the original logical i/o by recursively
1399 * reading the gang leader and all gang headers below it. This yields
1400 * an in-core tree containing the contents of every gang header and the
1401 * bps for every constituent of the gang block.
1403 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1404 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1405 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1406 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1407 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1408 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1409 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1410 * of the gang header plus zio_checksum_compute() of the data to update the
1411 * gang header's blk_cksum as described above.
1413 * The two-phase assemble/issue model solves the problem of partial failure --
1414 * what if you'd freed part of a gang block but then couldn't read the
1415 * gang header for another part? Assembling the entire gang tree first
1416 * ensures that all the necessary gang header I/O has succeeded before
1417 * starting the actual work of free, claim, or write. Once the gang tree
1418 * is assembled, free and claim are in-memory operations that cannot fail.
1420 * In the event that a gang write fails, zio_dva_unallocate() walks the
1421 * gang tree to immediately free (i.e. insert back into the space map)
1422 * everything we've allocated. This ensures that we don't get ENOSPC
1423 * errors during repeated suspend/resume cycles due to a flaky device.
1425 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1426 * the gang tree, we won't modify the block, so we can safely defer the free
1427 * (knowing that the block is still intact). If we *can* assemble the gang
1428 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1429 * each constituent bp and we can allocate a new block on the next sync pass.
1431 * In all cases, the gang tree allows complete recovery from partial failure.
1432 * ==========================================================================
1436 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1441 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1442 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1443 &pio->io_bookmark));
1447 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1452 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1453 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1454 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1456 * As we rewrite each gang header, the pipeline will compute
1457 * a new gang block header checksum for it; but no one will
1458 * compute a new data checksum, so we do that here. The one
1459 * exception is the gang leader: the pipeline already computed
1460 * its data checksum because that stage precedes gang assembly.
1461 * (Presently, nothing actually uses interior data checksums;
1462 * this is just good hygiene.)
1464 if (gn != pio->io_gang_leader->io_gang_tree) {
1465 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1466 data, BP_GET_PSIZE(bp));
1469 * If we are here to damage data for testing purposes,
1470 * leave the GBH alone so that we can detect the damage.
1472 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1473 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1475 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1476 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1477 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1485 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1487 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1488 ZIO_GANG_CHILD_FLAGS(pio)));
1493 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1495 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1496 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1499 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1508 static void zio_gang_tree_assemble_done(zio_t *zio);
1510 static zio_gang_node_t *
1511 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1513 zio_gang_node_t *gn;
1515 ASSERT(*gnpp == NULL);
1517 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1518 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1525 zio_gang_node_free(zio_gang_node_t **gnpp)
1527 zio_gang_node_t *gn = *gnpp;
1529 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1530 ASSERT(gn->gn_child[g] == NULL);
1532 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1533 kmem_free(gn, sizeof (*gn));
1538 zio_gang_tree_free(zio_gang_node_t **gnpp)
1540 zio_gang_node_t *gn = *gnpp;
1545 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1546 zio_gang_tree_free(&gn->gn_child[g]);
1548 zio_gang_node_free(gnpp);
1552 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1554 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1556 ASSERT(gio->io_gang_leader == gio);
1557 ASSERT(BP_IS_GANG(bp));
1559 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1560 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1561 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1565 zio_gang_tree_assemble_done(zio_t *zio)
1567 zio_t *gio = zio->io_gang_leader;
1568 zio_gang_node_t *gn = zio->io_private;
1569 blkptr_t *bp = zio->io_bp;
1571 ASSERT(gio == zio_unique_parent(zio));
1572 ASSERT(zio->io_child_count == 0);
1577 if (BP_SHOULD_BYTESWAP(bp))
1578 byteswap_uint64_array(zio->io_data, zio->io_size);
1580 ASSERT(zio->io_data == gn->gn_gbh);
1581 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1582 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1584 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1585 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1586 if (!BP_IS_GANG(gbp))
1588 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1593 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1595 zio_t *gio = pio->io_gang_leader;
1598 ASSERT(BP_IS_GANG(bp) == !!gn);
1599 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1600 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1603 * If you're a gang header, your data is in gn->gn_gbh.
1604 * If you're a gang member, your data is in 'data' and gn == NULL.
1606 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1609 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1611 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1612 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1613 if (BP_IS_HOLE(gbp))
1615 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1616 data = (char *)data + BP_GET_PSIZE(gbp);
1620 if (gn == gio->io_gang_tree)
1621 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1628 zio_gang_assemble(zio_t *zio)
1630 blkptr_t *bp = zio->io_bp;
1632 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1633 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1635 zio->io_gang_leader = zio;
1637 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1639 return (ZIO_PIPELINE_CONTINUE);
1643 zio_gang_issue(zio_t *zio)
1645 blkptr_t *bp = zio->io_bp;
1647 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1648 return (ZIO_PIPELINE_STOP);
1650 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1651 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1653 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1654 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1656 zio_gang_tree_free(&zio->io_gang_tree);
1658 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1660 return (ZIO_PIPELINE_CONTINUE);
1664 zio_write_gang_member_ready(zio_t *zio)
1666 zio_t *pio = zio_unique_parent(zio);
1667 zio_t *gio = zio->io_gang_leader;
1668 dva_t *cdva = zio->io_bp->blk_dva;
1669 dva_t *pdva = pio->io_bp->blk_dva;
1672 if (BP_IS_HOLE(zio->io_bp))
1675 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1677 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1678 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1679 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1680 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1681 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1683 mutex_enter(&pio->io_lock);
1684 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1685 ASSERT(DVA_GET_GANG(&pdva[d]));
1686 asize = DVA_GET_ASIZE(&pdva[d]);
1687 asize += DVA_GET_ASIZE(&cdva[d]);
1688 DVA_SET_ASIZE(&pdva[d], asize);
1690 mutex_exit(&pio->io_lock);
1694 zio_write_gang_block(zio_t *pio)
1696 spa_t *spa = pio->io_spa;
1697 blkptr_t *bp = pio->io_bp;
1698 zio_t *gio = pio->io_gang_leader;
1700 zio_gang_node_t *gn, **gnpp;
1701 zio_gbh_phys_t *gbh;
1702 uint64_t txg = pio->io_txg;
1703 uint64_t resid = pio->io_size;
1705 int copies = gio->io_prop.zp_copies;
1706 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1710 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1711 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1712 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1714 pio->io_error = error;
1715 return (ZIO_PIPELINE_CONTINUE);
1719 gnpp = &gio->io_gang_tree;
1721 gnpp = pio->io_private;
1722 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1725 gn = zio_gang_node_alloc(gnpp);
1727 bzero(gbh, SPA_GANGBLOCKSIZE);
1730 * Create the gang header.
1732 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1733 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1736 * Create and nowait the gang children.
1738 for (int g = 0; resid != 0; resid -= lsize, g++) {
1739 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1741 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1743 zp.zp_checksum = gio->io_prop.zp_checksum;
1744 zp.zp_compress = ZIO_COMPRESS_OFF;
1745 zp.zp_type = DMU_OT_NONE;
1747 zp.zp_copies = gio->io_prop.zp_copies;
1749 zp.zp_dedup_verify = 0;
1751 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1752 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1753 zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1754 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1755 &pio->io_bookmark));
1759 * Set pio's pipeline to just wait for zio to finish.
1761 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1765 return (ZIO_PIPELINE_CONTINUE);
1769 * ==========================================================================
1771 * ==========================================================================
1774 zio_ddt_child_read_done(zio_t *zio)
1776 blkptr_t *bp = zio->io_bp;
1777 ddt_entry_t *dde = zio->io_private;
1779 zio_t *pio = zio_unique_parent(zio);
1781 mutex_enter(&pio->io_lock);
1782 ddp = ddt_phys_select(dde, bp);
1783 if (zio->io_error == 0)
1784 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1785 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1786 dde->dde_repair_data = zio->io_data;
1788 zio_buf_free(zio->io_data, zio->io_size);
1789 mutex_exit(&pio->io_lock);
1793 zio_ddt_read_start(zio_t *zio)
1795 blkptr_t *bp = zio->io_bp;
1797 ASSERT(BP_GET_DEDUP(bp));
1798 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1799 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1801 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1802 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1803 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1804 ddt_phys_t *ddp = dde->dde_phys;
1805 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1808 ASSERT(zio->io_vsd == NULL);
1811 if (ddp_self == NULL)
1812 return (ZIO_PIPELINE_CONTINUE);
1814 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1815 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1817 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1819 zio_nowait(zio_read(zio, zio->io_spa, &blk,
1820 zio_buf_alloc(zio->io_size), zio->io_size,
1821 zio_ddt_child_read_done, dde, zio->io_priority,
1822 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1823 &zio->io_bookmark));
1825 return (ZIO_PIPELINE_CONTINUE);
1828 zio_nowait(zio_read(zio, zio->io_spa, bp,
1829 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1830 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1832 return (ZIO_PIPELINE_CONTINUE);
1836 zio_ddt_read_done(zio_t *zio)
1838 blkptr_t *bp = zio->io_bp;
1840 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1841 return (ZIO_PIPELINE_STOP);
1843 ASSERT(BP_GET_DEDUP(bp));
1844 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1845 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1847 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1848 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1849 ddt_entry_t *dde = zio->io_vsd;
1851 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
1852 return (ZIO_PIPELINE_CONTINUE);
1855 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
1856 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1857 return (ZIO_PIPELINE_STOP);
1859 if (dde->dde_repair_data != NULL) {
1860 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
1861 zio->io_child_error[ZIO_CHILD_DDT] = 0;
1863 ddt_repair_done(ddt, dde);
1867 ASSERT(zio->io_vsd == NULL);
1869 return (ZIO_PIPELINE_CONTINUE);
1873 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
1875 spa_t *spa = zio->io_spa;
1878 * Note: we compare the original data, not the transformed data,
1879 * because when zio->io_bp is an override bp, we will not have
1880 * pushed the I/O transforms. That's an important optimization
1881 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
1883 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1884 zio_t *lio = dde->dde_lead_zio[p];
1887 return (lio->io_orig_size != zio->io_orig_size ||
1888 bcmp(zio->io_orig_data, lio->io_orig_data,
1889 zio->io_orig_size) != 0);
1893 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1894 ddt_phys_t *ddp = &dde->dde_phys[p];
1896 if (ddp->ddp_phys_birth != 0) {
1897 arc_buf_t *abuf = NULL;
1898 uint32_t aflags = ARC_WAIT;
1899 blkptr_t blk = *zio->io_bp;
1902 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
1906 error = arc_read_nolock(NULL, spa, &blk,
1907 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
1908 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1909 &aflags, &zio->io_bookmark);
1912 if (arc_buf_size(abuf) != zio->io_orig_size ||
1913 bcmp(abuf->b_data, zio->io_orig_data,
1914 zio->io_orig_size) != 0)
1916 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
1920 return (error != 0);
1928 zio_ddt_child_write_ready(zio_t *zio)
1930 int p = zio->io_prop.zp_copies;
1931 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1932 ddt_entry_t *dde = zio->io_private;
1933 ddt_phys_t *ddp = &dde->dde_phys[p];
1941 ASSERT(dde->dde_lead_zio[p] == zio);
1943 ddt_phys_fill(ddp, zio->io_bp);
1945 while ((pio = zio_walk_parents(zio)) != NULL)
1946 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
1952 zio_ddt_child_write_done(zio_t *zio)
1954 int p = zio->io_prop.zp_copies;
1955 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1956 ddt_entry_t *dde = zio->io_private;
1957 ddt_phys_t *ddp = &dde->dde_phys[p];
1961 ASSERT(ddp->ddp_refcnt == 0);
1962 ASSERT(dde->dde_lead_zio[p] == zio);
1963 dde->dde_lead_zio[p] = NULL;
1965 if (zio->io_error == 0) {
1966 while (zio_walk_parents(zio) != NULL)
1967 ddt_phys_addref(ddp);
1969 ddt_phys_clear(ddp);
1976 zio_ddt_ditto_write_done(zio_t *zio)
1978 int p = DDT_PHYS_DITTO;
1979 zio_prop_t *zp = &zio->io_prop;
1980 blkptr_t *bp = zio->io_bp;
1981 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1982 ddt_entry_t *dde = zio->io_private;
1983 ddt_phys_t *ddp = &dde->dde_phys[p];
1984 ddt_key_t *ddk = &dde->dde_key;
1988 ASSERT(ddp->ddp_refcnt == 0);
1989 ASSERT(dde->dde_lead_zio[p] == zio);
1990 dde->dde_lead_zio[p] = NULL;
1992 if (zio->io_error == 0) {
1993 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
1994 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
1995 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
1996 if (ddp->ddp_phys_birth != 0)
1997 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
1998 ddt_phys_fill(ddp, bp);
2005 zio_ddt_write(zio_t *zio)
2007 spa_t *spa = zio->io_spa;
2008 blkptr_t *bp = zio->io_bp;
2009 uint64_t txg = zio->io_txg;
2010 zio_prop_t *zp = &zio->io_prop;
2011 int p = zp->zp_copies;
2015 ddt_t *ddt = ddt_select(spa, bp);
2019 ASSERT(BP_GET_DEDUP(bp));
2020 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2021 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2024 dde = ddt_lookup(ddt, bp, B_TRUE);
2025 ddp = &dde->dde_phys[p];
2027 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2029 * If we're using a weak checksum, upgrade to a strong checksum
2030 * and try again. If we're already using a strong checksum,
2031 * we can't resolve it, so just convert to an ordinary write.
2032 * (And automatically e-mail a paper to Nature?)
2034 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2035 zp->zp_checksum = spa_dedup_checksum(spa);
2036 zio_pop_transforms(zio);
2037 zio->io_stage = ZIO_STAGE_OPEN;
2042 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2044 return (ZIO_PIPELINE_CONTINUE);
2047 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2048 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2050 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2051 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2052 zio_prop_t czp = *zp;
2054 czp.zp_copies = ditto_copies;
2057 * If we arrived here with an override bp, we won't have run
2058 * the transform stack, so we won't have the data we need to
2059 * generate a child i/o. So, toss the override bp and restart.
2060 * This is safe, because using the override bp is just an
2061 * optimization; and it's rare, so the cost doesn't matter.
2063 if (zio->io_bp_override) {
2064 zio_pop_transforms(zio);
2065 zio->io_stage = ZIO_STAGE_OPEN;
2066 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2067 zio->io_bp_override = NULL;
2070 return (ZIO_PIPELINE_CONTINUE);
2073 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2074 zio->io_orig_size, &czp, NULL,
2075 zio_ddt_ditto_write_done, dde, zio->io_priority,
2076 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2078 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2079 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2082 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2083 if (ddp->ddp_phys_birth != 0)
2084 ddt_bp_fill(ddp, bp, txg);
2085 if (dde->dde_lead_zio[p] != NULL)
2086 zio_add_child(zio, dde->dde_lead_zio[p]);
2088 ddt_phys_addref(ddp);
2089 } else if (zio->io_bp_override) {
2090 ASSERT(bp->blk_birth == txg);
2091 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2092 ddt_phys_fill(ddp, bp);
2093 ddt_phys_addref(ddp);
2095 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2096 zio->io_orig_size, zp, zio_ddt_child_write_ready,
2097 zio_ddt_child_write_done, dde, zio->io_priority,
2098 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2100 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2101 dde->dde_lead_zio[p] = cio;
2111 return (ZIO_PIPELINE_CONTINUE);
2114 ddt_entry_t *freedde; /* for debugging */
2117 zio_ddt_free(zio_t *zio)
2119 spa_t *spa = zio->io_spa;
2120 blkptr_t *bp = zio->io_bp;
2121 ddt_t *ddt = ddt_select(spa, bp);
2125 ASSERT(BP_GET_DEDUP(bp));
2126 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2129 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2130 ddp = ddt_phys_select(dde, bp);
2131 ddt_phys_decref(ddp);
2134 return (ZIO_PIPELINE_CONTINUE);
2138 * ==========================================================================
2139 * Allocate and free blocks
2140 * ==========================================================================
2143 zio_dva_allocate(zio_t *zio)
2145 spa_t *spa = zio->io_spa;
2146 metaslab_class_t *mc = spa_normal_class(spa);
2147 blkptr_t *bp = zio->io_bp;
2151 if (zio->io_gang_leader == NULL) {
2152 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2153 zio->io_gang_leader = zio;
2156 ASSERT(BP_IS_HOLE(bp));
2157 ASSERT3U(BP_GET_NDVAS(bp), ==, 0);
2158 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2159 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2160 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2163 * The dump device does not support gang blocks so allocation on
2164 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2165 * the "fast" gang feature.
2167 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2168 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2169 METASLAB_GANG_CHILD : 0;
2170 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2171 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2174 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2175 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2177 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2178 return (zio_write_gang_block(zio));
2179 zio->io_error = error;
2182 return (ZIO_PIPELINE_CONTINUE);
2186 zio_dva_free(zio_t *zio)
2188 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2190 return (ZIO_PIPELINE_CONTINUE);
2194 zio_dva_claim(zio_t *zio)
2198 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2200 zio->io_error = error;
2202 return (ZIO_PIPELINE_CONTINUE);
2206 * Undo an allocation. This is used by zio_done() when an I/O fails
2207 * and we want to give back the block we just allocated.
2208 * This handles both normal blocks and gang blocks.
2211 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2213 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2214 ASSERT(zio->io_bp_override == NULL);
2216 if (!BP_IS_HOLE(bp))
2217 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2220 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2221 zio_dva_unallocate(zio, gn->gn_child[g],
2222 &gn->gn_gbh->zg_blkptr[g]);
2228 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2231 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2232 uint64_t size, boolean_t use_slog)
2236 ASSERT(txg > spa_syncing_txg(spa));
2239 error = metaslab_alloc(spa, spa_log_class(spa), size,
2240 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2243 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2244 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2247 BP_SET_LSIZE(new_bp, size);
2248 BP_SET_PSIZE(new_bp, size);
2249 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2250 BP_SET_CHECKSUM(new_bp,
2251 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2252 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2253 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2254 BP_SET_LEVEL(new_bp, 0);
2255 BP_SET_DEDUP(new_bp, 0);
2256 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2263 * Free an intent log block.
2266 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2268 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2269 ASSERT(!BP_IS_GANG(bp));
2271 zio_free(spa, txg, bp);
2275 * ==========================================================================
2276 * Read and write to physical devices
2277 * ==========================================================================
2280 zio_vdev_io_start(zio_t *zio)
2282 vdev_t *vd = zio->io_vd;
2284 spa_t *spa = zio->io_spa;
2286 ASSERT(zio->io_error == 0);
2287 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2290 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2291 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2294 * The mirror_ops handle multiple DVAs in a single BP.
2296 return (vdev_mirror_ops.vdev_op_io_start(zio));
2300 * We keep track of time-sensitive I/Os so that the scan thread
2301 * can quickly react to certain workloads. In particular, we care
2302 * about non-scrubbing, top-level reads and writes with the following
2304 * - synchronous writes of user data to non-slog devices
2305 * - any reads of user data
2306 * When these conditions are met, adjust the timestamp of spa_last_io
2307 * which allows the scan thread to adjust its workload accordingly.
2309 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2310 vd == vd->vdev_top && !vd->vdev_islog &&
2311 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2312 zio->io_txg != spa_syncing_txg(spa)) {
2313 uint64_t old = spa->spa_last_io;
2314 uint64_t new = ddi_get_lbolt64();
2316 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2319 align = 1ULL << vd->vdev_top->vdev_ashift;
2321 if (P2PHASE(zio->io_size, align) != 0) {
2322 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2323 char *abuf = zio_buf_alloc(asize);
2324 ASSERT(vd == vd->vdev_top);
2325 if (zio->io_type == ZIO_TYPE_WRITE) {
2326 bcopy(zio->io_data, abuf, zio->io_size);
2327 bzero(abuf + zio->io_size, asize - zio->io_size);
2329 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2332 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2333 ASSERT(P2PHASE(zio->io_size, align) == 0);
2334 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2337 * If this is a repair I/O, and there's no self-healing involved --
2338 * that is, we're just resilvering what we expect to resilver --
2339 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2340 * This prevents spurious resilvering with nested replication.
2341 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2342 * A is out of date, we'll read from C+D, then use the data to
2343 * resilver A+B -- but we don't actually want to resilver B, just A.
2344 * The top-level mirror has no way to know this, so instead we just
2345 * discard unnecessary repairs as we work our way down the vdev tree.
2346 * The same logic applies to any form of nested replication:
2347 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2349 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2350 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2351 zio->io_txg != 0 && /* not a delegated i/o */
2352 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2353 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2354 zio_vdev_io_bypass(zio);
2355 return (ZIO_PIPELINE_CONTINUE);
2358 if (vd->vdev_ops->vdev_op_leaf &&
2359 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2361 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2362 return (ZIO_PIPELINE_CONTINUE);
2364 if ((zio = vdev_queue_io(zio)) == NULL)
2365 return (ZIO_PIPELINE_STOP);
2367 if (!vdev_accessible(vd, zio)) {
2368 zio->io_error = ENXIO;
2370 return (ZIO_PIPELINE_STOP);
2374 return (vd->vdev_ops->vdev_op_io_start(zio));
2378 zio_vdev_io_done(zio_t *zio)
2380 vdev_t *vd = zio->io_vd;
2381 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2382 boolean_t unexpected_error = B_FALSE;
2384 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2385 return (ZIO_PIPELINE_STOP);
2387 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2389 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2391 vdev_queue_io_done(zio);
2393 if (zio->io_type == ZIO_TYPE_WRITE)
2394 vdev_cache_write(zio);
2396 if (zio_injection_enabled && zio->io_error == 0)
2397 zio->io_error = zio_handle_device_injection(vd,
2400 if (zio_injection_enabled && zio->io_error == 0)
2401 zio->io_error = zio_handle_label_injection(zio, EIO);
2403 if (zio->io_error) {
2404 if (!vdev_accessible(vd, zio)) {
2405 zio->io_error = ENXIO;
2407 unexpected_error = B_TRUE;
2412 ops->vdev_op_io_done(zio);
2414 if (unexpected_error)
2415 VERIFY(vdev_probe(vd, zio) == NULL);
2417 return (ZIO_PIPELINE_CONTINUE);
2421 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2422 * disk, and use that to finish the checksum ereport later.
2425 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2426 const void *good_buf)
2428 /* no processing needed */
2429 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2434 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2436 void *buf = zio_buf_alloc(zio->io_size);
2438 bcopy(zio->io_data, buf, zio->io_size);
2440 zcr->zcr_cbinfo = zio->io_size;
2441 zcr->zcr_cbdata = buf;
2442 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2443 zcr->zcr_free = zio_buf_free;
2447 zio_vdev_io_assess(zio_t *zio)
2449 vdev_t *vd = zio->io_vd;
2451 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2452 return (ZIO_PIPELINE_STOP);
2454 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2455 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2457 if (zio->io_vsd != NULL) {
2458 zio->io_vsd_ops->vsd_free(zio);
2462 if (zio_injection_enabled && zio->io_error == 0)
2463 zio->io_error = zio_handle_fault_injection(zio, EIO);
2466 * If the I/O failed, determine whether we should attempt to retry it.
2468 * On retry, we cut in line in the issue queue, since we don't want
2469 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2471 if (zio->io_error && vd == NULL &&
2472 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2473 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2474 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2476 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2477 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2478 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2479 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2480 zio_requeue_io_start_cut_in_line);
2481 return (ZIO_PIPELINE_STOP);
2485 * If we got an error on a leaf device, convert it to ENXIO
2486 * if the device is not accessible at all.
2488 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2489 !vdev_accessible(vd, zio))
2490 zio->io_error = ENXIO;
2493 * If we can't write to an interior vdev (mirror or RAID-Z),
2494 * set vdev_cant_write so that we stop trying to allocate from it.
2496 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2497 vd != NULL && !vd->vdev_ops->vdev_op_leaf)
2498 vd->vdev_cant_write = B_TRUE;
2501 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2503 return (ZIO_PIPELINE_CONTINUE);
2507 zio_vdev_io_reissue(zio_t *zio)
2509 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2510 ASSERT(zio->io_error == 0);
2512 zio->io_stage >>= 1;
2516 zio_vdev_io_redone(zio_t *zio)
2518 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2520 zio->io_stage >>= 1;
2524 zio_vdev_io_bypass(zio_t *zio)
2526 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2527 ASSERT(zio->io_error == 0);
2529 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2530 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2534 * ==========================================================================
2535 * Generate and verify checksums
2536 * ==========================================================================
2539 zio_checksum_generate(zio_t *zio)
2541 blkptr_t *bp = zio->io_bp;
2542 enum zio_checksum checksum;
2546 * This is zio_write_phys().
2547 * We're either generating a label checksum, or none at all.
2549 checksum = zio->io_prop.zp_checksum;
2551 if (checksum == ZIO_CHECKSUM_OFF)
2552 return (ZIO_PIPELINE_CONTINUE);
2554 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2556 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2557 ASSERT(!IO_IS_ALLOCATING(zio));
2558 checksum = ZIO_CHECKSUM_GANG_HEADER;
2560 checksum = BP_GET_CHECKSUM(bp);
2564 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2566 return (ZIO_PIPELINE_CONTINUE);
2570 zio_checksum_verify(zio_t *zio)
2572 zio_bad_cksum_t info;
2573 blkptr_t *bp = zio->io_bp;
2576 ASSERT(zio->io_vd != NULL);
2580 * This is zio_read_phys().
2581 * We're either verifying a label checksum, or nothing at all.
2583 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2584 return (ZIO_PIPELINE_CONTINUE);
2586 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2589 if ((error = zio_checksum_error(zio, &info)) != 0) {
2590 zio->io_error = error;
2591 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2592 zfs_ereport_start_checksum(zio->io_spa,
2593 zio->io_vd, zio, zio->io_offset,
2594 zio->io_size, NULL, &info);
2598 return (ZIO_PIPELINE_CONTINUE);
2602 * Called by RAID-Z to ensure we don't compute the checksum twice.
2605 zio_checksum_verified(zio_t *zio)
2607 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2611 * ==========================================================================
2612 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2613 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2614 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2615 * indicate errors that are specific to one I/O, and most likely permanent.
2616 * Any other error is presumed to be worse because we weren't expecting it.
2617 * ==========================================================================
2620 zio_worst_error(int e1, int e2)
2622 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2625 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2626 if (e1 == zio_error_rank[r1])
2629 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2630 if (e2 == zio_error_rank[r2])
2633 return (r1 > r2 ? e1 : e2);
2637 * ==========================================================================
2639 * ==========================================================================
2642 zio_ready(zio_t *zio)
2644 blkptr_t *bp = zio->io_bp;
2645 zio_t *pio, *pio_next;
2647 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2648 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2649 return (ZIO_PIPELINE_STOP);
2651 if (zio->io_ready) {
2652 ASSERT(IO_IS_ALLOCATING(zio));
2653 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2654 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2659 if (bp != NULL && bp != &zio->io_bp_copy)
2660 zio->io_bp_copy = *bp;
2663 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2665 mutex_enter(&zio->io_lock);
2666 zio->io_state[ZIO_WAIT_READY] = 1;
2667 pio = zio_walk_parents(zio);
2668 mutex_exit(&zio->io_lock);
2671 * As we notify zio's parents, new parents could be added.
2672 * New parents go to the head of zio's io_parent_list, however,
2673 * so we will (correctly) not notify them. The remainder of zio's
2674 * io_parent_list, from 'pio_next' onward, cannot change because
2675 * all parents must wait for us to be done before they can be done.
2677 for (; pio != NULL; pio = pio_next) {
2678 pio_next = zio_walk_parents(zio);
2679 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2682 if (zio->io_flags & ZIO_FLAG_NODATA) {
2683 if (BP_IS_GANG(bp)) {
2684 zio->io_flags &= ~ZIO_FLAG_NODATA;
2686 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2687 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2691 if (zio_injection_enabled &&
2692 zio->io_spa->spa_syncing_txg == zio->io_txg)
2693 zio_handle_ignored_writes(zio);
2695 return (ZIO_PIPELINE_CONTINUE);
2699 zio_done(zio_t *zio)
2701 spa_t *spa = zio->io_spa;
2702 zio_t *lio = zio->io_logical;
2703 blkptr_t *bp = zio->io_bp;
2704 vdev_t *vd = zio->io_vd;
2705 uint64_t psize = zio->io_size;
2706 zio_t *pio, *pio_next;
2709 * If our children haven't all completed,
2710 * wait for them and then repeat this pipeline stage.
2712 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2713 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2714 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2715 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2716 return (ZIO_PIPELINE_STOP);
2718 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2719 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2720 ASSERT(zio->io_children[c][w] == 0);
2723 ASSERT(bp->blk_pad[0] == 0);
2724 ASSERT(bp->blk_pad[1] == 0);
2725 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2726 (bp == zio_unique_parent(zio)->io_bp));
2727 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
2728 zio->io_bp_override == NULL &&
2729 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2730 ASSERT(!BP_SHOULD_BYTESWAP(bp));
2731 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
2732 ASSERT(BP_COUNT_GANG(bp) == 0 ||
2733 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
2738 * If there were child vdev/gang/ddt errors, they apply to us now.
2740 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2741 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2742 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2745 * If the I/O on the transformed data was successful, generate any
2746 * checksum reports now while we still have the transformed data.
2748 if (zio->io_error == 0) {
2749 while (zio->io_cksum_report != NULL) {
2750 zio_cksum_report_t *zcr = zio->io_cksum_report;
2751 uint64_t align = zcr->zcr_align;
2752 uint64_t asize = P2ROUNDUP(psize, align);
2753 char *abuf = zio->io_data;
2755 if (asize != psize) {
2756 abuf = zio_buf_alloc(asize);
2757 bcopy(zio->io_data, abuf, psize);
2758 bzero(abuf + psize, asize - psize);
2761 zio->io_cksum_report = zcr->zcr_next;
2762 zcr->zcr_next = NULL;
2763 zcr->zcr_finish(zcr, abuf);
2764 zfs_ereport_free_checksum(zcr);
2767 zio_buf_free(abuf, asize);
2771 zio_pop_transforms(zio); /* note: may set zio->io_error */
2773 vdev_stat_update(zio, psize);
2775 if (zio->io_error) {
2777 * If this I/O is attached to a particular vdev,
2778 * generate an error message describing the I/O failure
2779 * at the block level. We ignore these errors if the
2780 * device is currently unavailable.
2782 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
2783 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
2785 if ((zio->io_error == EIO || !(zio->io_flags &
2786 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2789 * For logical I/O requests, tell the SPA to log the
2790 * error and generate a logical data ereport.
2792 spa_log_error(spa, zio);
2793 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
2798 if (zio->io_error && zio == lio) {
2800 * Determine whether zio should be reexecuted. This will
2801 * propagate all the way to the root via zio_notify_parent().
2803 ASSERT(vd == NULL && bp != NULL);
2804 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2806 if (IO_IS_ALLOCATING(zio) &&
2807 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2808 if (zio->io_error != ENOSPC)
2809 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2811 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2814 if ((zio->io_type == ZIO_TYPE_READ ||
2815 zio->io_type == ZIO_TYPE_FREE) &&
2816 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
2817 zio->io_error == ENXIO &&
2818 spa_load_state(spa) == SPA_LOAD_NONE &&
2819 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
2820 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2822 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2823 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2826 * Here is a possibly good place to attempt to do
2827 * either combinatorial reconstruction or error correction
2828 * based on checksums. It also might be a good place
2829 * to send out preliminary ereports before we suspend
2835 * If there were logical child errors, they apply to us now.
2836 * We defer this until now to avoid conflating logical child
2837 * errors with errors that happened to the zio itself when
2838 * updating vdev stats and reporting FMA events above.
2840 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2842 if ((zio->io_error || zio->io_reexecute) &&
2843 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
2844 !(zio->io_flags & ZIO_FLAG_IO_REWRITE))
2845 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
2847 zio_gang_tree_free(&zio->io_gang_tree);
2850 * Godfather I/Os should never suspend.
2852 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
2853 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
2854 zio->io_reexecute = 0;
2856 if (zio->io_reexecute) {
2858 * This is a logical I/O that wants to reexecute.
2860 * Reexecute is top-down. When an i/o fails, if it's not
2861 * the root, it simply notifies its parent and sticks around.
2862 * The parent, seeing that it still has children in zio_done(),
2863 * does the same. This percolates all the way up to the root.
2864 * The root i/o will reexecute or suspend the entire tree.
2866 * This approach ensures that zio_reexecute() honors
2867 * all the original i/o dependency relationships, e.g.
2868 * parents not executing until children are ready.
2870 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2872 zio->io_gang_leader = NULL;
2874 mutex_enter(&zio->io_lock);
2875 zio->io_state[ZIO_WAIT_DONE] = 1;
2876 mutex_exit(&zio->io_lock);
2879 * "The Godfather" I/O monitors its children but is
2880 * not a true parent to them. It will track them through
2881 * the pipeline but severs its ties whenever they get into
2882 * trouble (e.g. suspended). This allows "The Godfather"
2883 * I/O to return status without blocking.
2885 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2886 zio_link_t *zl = zio->io_walk_link;
2887 pio_next = zio_walk_parents(zio);
2889 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
2890 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
2891 zio_remove_child(pio, zio, zl);
2892 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2896 if ((pio = zio_unique_parent(zio)) != NULL) {
2898 * We're not a root i/o, so there's nothing to do
2899 * but notify our parent. Don't propagate errors
2900 * upward since we haven't permanently failed yet.
2902 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2903 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
2904 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2905 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
2907 * We'd fail again if we reexecuted now, so suspend
2908 * until conditions improve (e.g. device comes online).
2910 zio_suspend(spa, zio);
2913 * Reexecution is potentially a huge amount of work.
2914 * Hand it off to the otherwise-unused claim taskq.
2917 (void) taskq_dispatch_safe(
2918 spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
2919 (task_func_t *)zio_reexecute, zio, TQ_SLEEP,
2922 (void) taskq_dispatch(
2923 spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
2924 (task_func_t *)zio_reexecute, zio, TQ_SLEEP);
2927 return (ZIO_PIPELINE_STOP);
2930 ASSERT(zio->io_child_count == 0);
2931 ASSERT(zio->io_reexecute == 0);
2932 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
2935 * Report any checksum errors, since the I/O is complete.
2937 while (zio->io_cksum_report != NULL) {
2938 zio_cksum_report_t *zcr = zio->io_cksum_report;
2939 zio->io_cksum_report = zcr->zcr_next;
2940 zcr->zcr_next = NULL;
2941 zcr->zcr_finish(zcr, NULL);
2942 zfs_ereport_free_checksum(zcr);
2946 * It is the responsibility of the done callback to ensure that this
2947 * particular zio is no longer discoverable for adoption, and as
2948 * such, cannot acquire any new parents.
2953 mutex_enter(&zio->io_lock);
2954 zio->io_state[ZIO_WAIT_DONE] = 1;
2955 mutex_exit(&zio->io_lock);
2957 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2958 zio_link_t *zl = zio->io_walk_link;
2959 pio_next = zio_walk_parents(zio);
2960 zio_remove_child(pio, zio, zl);
2961 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2964 if (zio->io_waiter != NULL) {
2965 mutex_enter(&zio->io_lock);
2966 zio->io_executor = NULL;
2967 cv_broadcast(&zio->io_cv);
2968 mutex_exit(&zio->io_lock);
2973 return (ZIO_PIPELINE_STOP);
2977 * ==========================================================================
2978 * I/O pipeline definition
2979 * ==========================================================================
2981 static zio_pipe_stage_t *zio_pipeline[] = {
2987 zio_checksum_generate,
3001 zio_checksum_verify,