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) 2012 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
27 #include <sys/zfs_context.h>
28 #include <sys/fm/fs/zfs.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>
41 * ==========================================================================
43 * ==========================================================================
45 uint8_t zio_priority_table[ZIO_PRIORITY_TABLE_SIZE] = {
46 0, /* ZIO_PRIORITY_NOW */
47 0, /* ZIO_PRIORITY_SYNC_READ */
48 0, /* ZIO_PRIORITY_SYNC_WRITE */
49 0, /* ZIO_PRIORITY_LOG_WRITE */
50 1, /* ZIO_PRIORITY_CACHE_FILL */
51 1, /* ZIO_PRIORITY_AGG */
52 4, /* ZIO_PRIORITY_FREE */
53 4, /* ZIO_PRIORITY_ASYNC_WRITE */
54 6, /* ZIO_PRIORITY_ASYNC_READ */
55 10, /* ZIO_PRIORITY_RESILVER */
56 20, /* ZIO_PRIORITY_SCRUB */
57 2, /* ZIO_PRIORITY_DDT_PREFETCH */
61 * ==========================================================================
62 * I/O type descriptions
63 * ==========================================================================
65 char *zio_type_name[ZIO_TYPES] = {
66 "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl"
70 * ==========================================================================
72 * ==========================================================================
74 kmem_cache_t *zio_cache;
75 kmem_cache_t *zio_link_cache;
76 kmem_cache_t *zio_vdev_cache;
77 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
78 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
79 int zio_bulk_flags = 0;
80 int zio_delay_max = ZIO_DELAY_MAX;
83 extern vmem_t *zio_alloc_arena;
85 extern int zfs_mg_alloc_failures;
88 * An allocating zio is one that either currently has the DVA allocate
89 * stage set or will have it later in its lifetime.
91 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
93 int zio_requeue_io_start_cut_in_line = 1;
96 int zio_buf_debug_limit = 16384;
98 int zio_buf_debug_limit = 0;
101 static inline void __zio_execute(zio_t *zio);
104 zio_cons(void *arg, void *unused, int kmflag)
108 bzero(zio, sizeof (zio_t));
110 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
111 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
113 list_create(&zio->io_parent_list, sizeof (zio_link_t),
114 offsetof(zio_link_t, zl_parent_node));
115 list_create(&zio->io_child_list, sizeof (zio_link_t),
116 offsetof(zio_link_t, zl_child_node));
122 zio_dest(void *arg, void *unused)
126 mutex_destroy(&zio->io_lock);
127 cv_destroy(&zio->io_cv);
128 list_destroy(&zio->io_parent_list);
129 list_destroy(&zio->io_child_list);
136 vmem_t *data_alloc_arena = NULL;
139 data_alloc_arena = zio_alloc_arena;
141 zio_cache = kmem_cache_create("zio_cache", sizeof (zio_t), 0,
142 zio_cons, zio_dest, NULL, NULL, NULL, KMC_KMEM);
143 zio_link_cache = kmem_cache_create("zio_link_cache",
144 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, KMC_KMEM);
145 zio_vdev_cache = kmem_cache_create("zio_vdev_cache", sizeof(vdev_io_t),
146 PAGESIZE, NULL, NULL, NULL, NULL, NULL, KMC_VMEM);
149 * For small buffers, we want a cache for each multiple of
150 * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache
151 * for each quarter-power of 2. For large buffers, we want
152 * a cache for each multiple of PAGESIZE.
154 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
155 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
159 while (p2 & (p2 - 1))
162 if (size <= 4 * SPA_MINBLOCKSIZE) {
163 align = SPA_MINBLOCKSIZE;
164 } else if (P2PHASE(size, PAGESIZE) == 0) {
166 } else if (P2PHASE(size, p2 >> 2) == 0) {
172 int flags = zio_bulk_flags;
175 * The smallest buffers (512b) are heavily used and
176 * experience a lot of churn. The slabs allocated
177 * for them are also relatively small (32K). Thus
178 * in over to avoid expensive calls to vmalloc() we
179 * make an exception to the usual slab allocation
180 * policy and force these buffers to be kmem backed.
182 if (size == (1 << SPA_MINBLOCKSHIFT))
185 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
186 zio_buf_cache[c] = kmem_cache_create(name, size,
187 align, NULL, NULL, NULL, NULL, NULL, flags);
189 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
190 zio_data_buf_cache[c] = kmem_cache_create(name, size,
191 align, NULL, NULL, NULL, NULL,
192 data_alloc_arena, flags);
197 ASSERT(zio_buf_cache[c] != NULL);
198 if (zio_buf_cache[c - 1] == NULL)
199 zio_buf_cache[c - 1] = zio_buf_cache[c];
201 ASSERT(zio_data_buf_cache[c] != NULL);
202 if (zio_data_buf_cache[c - 1] == NULL)
203 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
207 * The zio write taskqs have 1 thread per cpu, allow 1/2 of the taskqs
208 * to fail 3 times per txg or 8 failures, whichever is greater.
210 zfs_mg_alloc_failures = MAX((3 * max_ncpus / 2), 8);
219 kmem_cache_t *last_cache = NULL;
220 kmem_cache_t *last_data_cache = NULL;
222 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
223 if (zio_buf_cache[c] != last_cache) {
224 last_cache = zio_buf_cache[c];
225 kmem_cache_destroy(zio_buf_cache[c]);
227 zio_buf_cache[c] = NULL;
229 if (zio_data_buf_cache[c] != last_data_cache) {
230 last_data_cache = zio_data_buf_cache[c];
231 kmem_cache_destroy(zio_data_buf_cache[c]);
233 zio_data_buf_cache[c] = NULL;
236 kmem_cache_destroy(zio_vdev_cache);
237 kmem_cache_destroy(zio_link_cache);
238 kmem_cache_destroy(zio_cache);
244 * ==========================================================================
245 * Allocate and free I/O buffers
246 * ==========================================================================
250 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
251 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
252 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
253 * excess / transient data in-core during a crashdump.
256 zio_buf_alloc(size_t size)
258 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
260 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
262 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE | KM_NODEBUG));
266 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
267 * crashdump if the kernel panics. This exists so that we will limit the amount
268 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
269 * of kernel heap dumped to disk when the kernel panics)
272 zio_data_buf_alloc(size_t size)
274 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
276 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
278 return (kmem_cache_alloc(zio_data_buf_cache[c],
279 KM_PUSHPAGE | KM_NODEBUG));
283 zio_buf_free(void *buf, size_t size)
285 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
287 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
289 kmem_cache_free(zio_buf_cache[c], buf);
293 zio_data_buf_free(void *buf, size_t size)
295 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
297 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
299 kmem_cache_free(zio_data_buf_cache[c], buf);
303 * Dedicated I/O buffers to ensure that memory fragmentation never prevents
304 * or significantly delays the issuing of a zio. These buffers are used
305 * to aggregate I/O and could be used for raidz stripes.
310 return (kmem_cache_alloc(zio_vdev_cache, KM_PUSHPAGE));
314 zio_vdev_free(void *buf)
316 kmem_cache_free(zio_vdev_cache, buf);
321 * ==========================================================================
322 * Push and pop I/O transform buffers
323 * ==========================================================================
326 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
327 zio_transform_func_t *transform)
329 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_PUSHPAGE);
331 zt->zt_orig_data = zio->io_data;
332 zt->zt_orig_size = zio->io_size;
333 zt->zt_bufsize = bufsize;
334 zt->zt_transform = transform;
336 zt->zt_next = zio->io_transform_stack;
337 zio->io_transform_stack = zt;
344 zio_pop_transforms(zio_t *zio)
348 while ((zt = zio->io_transform_stack) != NULL) {
349 if (zt->zt_transform != NULL)
350 zt->zt_transform(zio,
351 zt->zt_orig_data, zt->zt_orig_size);
353 if (zt->zt_bufsize != 0)
354 zio_buf_free(zio->io_data, zt->zt_bufsize);
356 zio->io_data = zt->zt_orig_data;
357 zio->io_size = zt->zt_orig_size;
358 zio->io_transform_stack = zt->zt_next;
360 kmem_free(zt, sizeof (zio_transform_t));
365 * ==========================================================================
366 * I/O transform callbacks for subblocks and decompression
367 * ==========================================================================
370 zio_subblock(zio_t *zio, void *data, uint64_t size)
372 ASSERT(zio->io_size > size);
374 if (zio->io_type == ZIO_TYPE_READ)
375 bcopy(zio->io_data, data, size);
379 zio_decompress(zio_t *zio, void *data, uint64_t size)
381 if (zio->io_error == 0 &&
382 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
383 zio->io_data, data, zio->io_size, size) != 0)
388 * ==========================================================================
389 * I/O parent/child relationships and pipeline interlocks
390 * ==========================================================================
393 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
394 * continue calling these functions until they return NULL.
395 * Otherwise, the next caller will pick up the list walk in
396 * some indeterminate state. (Otherwise every caller would
397 * have to pass in a cookie to keep the state represented by
398 * io_walk_link, which gets annoying.)
401 zio_walk_parents(zio_t *cio)
403 zio_link_t *zl = cio->io_walk_link;
404 list_t *pl = &cio->io_parent_list;
406 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
407 cio->io_walk_link = zl;
412 ASSERT(zl->zl_child == cio);
413 return (zl->zl_parent);
417 zio_walk_children(zio_t *pio)
419 zio_link_t *zl = pio->io_walk_link;
420 list_t *cl = &pio->io_child_list;
422 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
423 pio->io_walk_link = zl;
428 ASSERT(zl->zl_parent == pio);
429 return (zl->zl_child);
433 zio_unique_parent(zio_t *cio)
435 zio_t *pio = zio_walk_parents(cio);
437 VERIFY(zio_walk_parents(cio) == NULL);
442 zio_add_child(zio_t *pio, zio_t *cio)
444 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_PUSHPAGE);
448 * Logical I/Os can have logical, gang, or vdev children.
449 * Gang I/Os can have gang or vdev children.
450 * Vdev I/Os can only have vdev children.
451 * The following ASSERT captures all of these constraints.
453 ASSERT(cio->io_child_type <= pio->io_child_type);
458 mutex_enter(&cio->io_lock);
459 mutex_enter(&pio->io_lock);
461 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
463 for (w = 0; w < ZIO_WAIT_TYPES; w++)
464 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
466 list_insert_head(&pio->io_child_list, zl);
467 list_insert_head(&cio->io_parent_list, zl);
469 pio->io_child_count++;
470 cio->io_parent_count++;
472 mutex_exit(&pio->io_lock);
473 mutex_exit(&cio->io_lock);
477 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
479 ASSERT(zl->zl_parent == pio);
480 ASSERT(zl->zl_child == cio);
482 mutex_enter(&cio->io_lock);
483 mutex_enter(&pio->io_lock);
485 list_remove(&pio->io_child_list, zl);
486 list_remove(&cio->io_parent_list, zl);
488 pio->io_child_count--;
489 cio->io_parent_count--;
491 mutex_exit(&pio->io_lock);
492 mutex_exit(&cio->io_lock);
494 kmem_cache_free(zio_link_cache, zl);
498 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
500 uint64_t *countp = &zio->io_children[child][wait];
501 boolean_t waiting = B_FALSE;
503 mutex_enter(&zio->io_lock);
504 ASSERT(zio->io_stall == NULL);
507 zio->io_stall = countp;
510 mutex_exit(&zio->io_lock);
515 __attribute__((always_inline))
517 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
519 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
520 int *errorp = &pio->io_child_error[zio->io_child_type];
522 mutex_enter(&pio->io_lock);
523 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
524 *errorp = zio_worst_error(*errorp, zio->io_error);
525 pio->io_reexecute |= zio->io_reexecute;
526 ASSERT3U(*countp, >, 0);
527 if (--*countp == 0 && pio->io_stall == countp) {
528 pio->io_stall = NULL;
529 mutex_exit(&pio->io_lock);
532 mutex_exit(&pio->io_lock);
537 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
539 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
540 zio->io_error = zio->io_child_error[c];
544 * ==========================================================================
545 * Create the various types of I/O (read, write, free, etc)
546 * ==========================================================================
549 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
550 void *data, uint64_t size, zio_done_func_t *done, void *private,
551 zio_type_t type, int priority, enum zio_flag flags,
552 vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
553 enum zio_stage stage, enum zio_stage pipeline)
557 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
558 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
559 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
561 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
562 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
563 ASSERT(vd || stage == ZIO_STAGE_OPEN);
565 zio = kmem_cache_alloc(zio_cache, KM_PUSHPAGE);
568 zio->io_child_type = ZIO_CHILD_VDEV;
569 else if (flags & ZIO_FLAG_GANG_CHILD)
570 zio->io_child_type = ZIO_CHILD_GANG;
571 else if (flags & ZIO_FLAG_DDT_CHILD)
572 zio->io_child_type = ZIO_CHILD_DDT;
574 zio->io_child_type = ZIO_CHILD_LOGICAL;
577 zio->io_logical = NULL;
578 zio->io_bp = (blkptr_t *)bp;
579 zio->io_bp_copy = *bp;
580 zio->io_bp_orig = *bp;
581 if (type != ZIO_TYPE_WRITE ||
582 zio->io_child_type == ZIO_CHILD_DDT)
583 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
584 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
585 zio->io_logical = zio;
586 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
587 pipeline |= ZIO_GANG_STAGES;
589 zio->io_logical = NULL;
591 bzero(&zio->io_bp_copy, sizeof (blkptr_t));
592 bzero(&zio->io_bp_orig, sizeof (blkptr_t));
597 zio->io_ready = NULL;
599 zio->io_private = private;
600 zio->io_prev_space_delta = 0;
602 zio->io_priority = priority;
605 zio->io_vsd_ops = NULL;
606 zio->io_offset = offset;
607 zio->io_deadline = 0;
608 zio->io_orig_data = zio->io_data = data;
609 zio->io_orig_size = zio->io_size = size;
610 zio->io_orig_flags = zio->io_flags = flags;
611 zio->io_orig_stage = zio->io_stage = stage;
612 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
613 bzero(&zio->io_prop, sizeof (zio_prop_t));
615 zio->io_reexecute = 0;
616 zio->io_bp_override = NULL;
617 zio->io_walk_link = NULL;
618 zio->io_transform_stack = NULL;
621 zio->io_child_count = 0;
622 zio->io_parent_count = 0;
623 zio->io_stall = NULL;
624 zio->io_gang_leader = NULL;
625 zio->io_gang_tree = NULL;
626 zio->io_executor = NULL;
627 zio->io_waiter = NULL;
628 zio->io_cksum_report = NULL;
630 bzero(zio->io_child_error, sizeof (int) * ZIO_CHILD_TYPES);
631 bzero(zio->io_children,
632 sizeof (uint64_t) * ZIO_CHILD_TYPES * ZIO_WAIT_TYPES);
633 bzero(&zio->io_bookmark, sizeof (zbookmark_t));
635 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
636 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
639 zio->io_bookmark = *zb;
642 if (zio->io_logical == NULL)
643 zio->io_logical = pio->io_logical;
644 if (zio->io_child_type == ZIO_CHILD_GANG)
645 zio->io_gang_leader = pio->io_gang_leader;
646 zio_add_child(pio, zio);
649 taskq_init_ent(&zio->io_tqent);
655 zio_destroy(zio_t *zio)
657 kmem_cache_free(zio_cache, zio);
661 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
662 void *private, enum zio_flag flags)
666 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
667 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
668 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
674 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
676 return (zio_null(NULL, spa, NULL, done, private, flags));
680 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
681 void *data, uint64_t size, zio_done_func_t *done, void *private,
682 int priority, enum zio_flag flags, const zbookmark_t *zb)
686 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
687 data, size, done, private,
688 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
689 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
690 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
696 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
697 void *data, uint64_t size, const zio_prop_t *zp,
698 zio_done_func_t *ready, zio_done_func_t *done, void *private,
699 int priority, enum zio_flag flags, const zbookmark_t *zb)
703 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
704 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
705 zp->zp_compress >= ZIO_COMPRESS_OFF &&
706 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
707 zp->zp_type < DMU_OT_NUMTYPES &&
710 zp->zp_copies <= spa_max_replication(spa) &&
712 zp->zp_dedup_verify <= 1);
714 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
715 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
716 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
717 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
719 zio->io_ready = ready;
726 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
727 uint64_t size, zio_done_func_t *done, void *private, int priority,
728 enum zio_flag flags, zbookmark_t *zb)
732 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
733 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
734 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
740 zio_write_override(zio_t *zio, blkptr_t *bp, int copies)
742 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
743 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
744 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
745 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
747 zio->io_prop.zp_copies = copies;
748 zio->io_bp_override = bp;
752 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
754 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
758 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
763 dprintf_bp(bp, "freeing in txg %llu, pass %u",
764 (longlong_t)txg, spa->spa_sync_pass);
766 ASSERT(!BP_IS_HOLE(bp));
767 ASSERT(spa_syncing_txg(spa) == txg);
768 ASSERT(spa_sync_pass(spa) <= SYNC_PASS_DEFERRED_FREE);
770 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
771 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
772 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
778 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
779 zio_done_func_t *done, void *private, enum zio_flag flags)
784 * A claim is an allocation of a specific block. Claims are needed
785 * to support immediate writes in the intent log. The issue is that
786 * immediate writes contain committed data, but in a txg that was
787 * *not* committed. Upon opening the pool after an unclean shutdown,
788 * the intent log claims all blocks that contain immediate write data
789 * so that the SPA knows they're in use.
791 * All claims *must* be resolved in the first txg -- before the SPA
792 * starts allocating blocks -- so that nothing is allocated twice.
793 * If txg == 0 we just verify that the block is claimable.
795 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
796 ASSERT(txg == spa_first_txg(spa) || txg == 0);
797 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
799 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
800 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
801 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
807 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
808 zio_done_func_t *done, void *private, int priority, enum zio_flag flags)
813 if (vd->vdev_children == 0) {
814 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
815 ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
816 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
820 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
822 for (c = 0; c < vd->vdev_children; c++)
823 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
824 done, private, priority, flags));
831 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
832 void *data, int checksum, zio_done_func_t *done, void *private,
833 int priority, enum zio_flag flags, boolean_t labels)
837 ASSERT(vd->vdev_children == 0);
838 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
839 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
840 ASSERT3U(offset + size, <=, vd->vdev_psize);
842 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
843 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
844 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
846 zio->io_prop.zp_checksum = checksum;
852 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
853 void *data, int checksum, zio_done_func_t *done, void *private,
854 int priority, enum zio_flag flags, boolean_t labels)
858 ASSERT(vd->vdev_children == 0);
859 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
860 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
861 ASSERT3U(offset + size, <=, vd->vdev_psize);
863 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
864 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
865 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
867 zio->io_prop.zp_checksum = checksum;
869 if (zio_checksum_table[checksum].ci_eck) {
871 * zec checksums are necessarily destructive -- they modify
872 * the end of the write buffer to hold the verifier/checksum.
873 * Therefore, we must make a local copy in case the data is
874 * being written to multiple places in parallel.
876 void *wbuf = zio_buf_alloc(size);
877 bcopy(data, wbuf, size);
878 zio_push_transform(zio, wbuf, size, size, NULL);
885 * Create a child I/O to do some work for us.
888 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
889 void *data, uint64_t size, int type, int priority, enum zio_flag flags,
890 zio_done_func_t *done, void *private)
892 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
895 ASSERT(vd->vdev_parent ==
896 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
898 if (type == ZIO_TYPE_READ && bp != NULL) {
900 * If we have the bp, then the child should perform the
901 * checksum and the parent need not. This pushes error
902 * detection as close to the leaves as possible and
903 * eliminates redundant checksums in the interior nodes.
905 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
906 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
909 if (vd->vdev_children == 0)
910 offset += VDEV_LABEL_START_SIZE;
912 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
915 * If we've decided to do a repair, the write is not speculative --
916 * even if the original read was.
918 if (flags & ZIO_FLAG_IO_REPAIR)
919 flags &= ~ZIO_FLAG_SPECULATIVE;
921 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
922 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
923 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
929 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
930 int type, int priority, enum zio_flag flags,
931 zio_done_func_t *done, void *private)
935 ASSERT(vd->vdev_ops->vdev_op_leaf);
937 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
938 data, size, done, private, type, priority,
939 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
941 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
947 zio_flush(zio_t *zio, vdev_t *vd)
949 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
950 NULL, NULL, ZIO_PRIORITY_NOW,
951 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
955 zio_shrink(zio_t *zio, uint64_t size)
957 ASSERT(zio->io_executor == NULL);
958 ASSERT(zio->io_orig_size == zio->io_size);
959 ASSERT(size <= zio->io_size);
962 * We don't shrink for raidz because of problems with the
963 * reconstruction when reading back less than the block size.
964 * Note, BP_IS_RAIDZ() assumes no compression.
966 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
967 if (!BP_IS_RAIDZ(zio->io_bp))
968 zio->io_orig_size = zio->io_size = size;
972 * ==========================================================================
973 * Prepare to read and write logical blocks
974 * ==========================================================================
978 zio_read_bp_init(zio_t *zio)
980 blkptr_t *bp = zio->io_bp;
982 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
983 zio->io_child_type == ZIO_CHILD_LOGICAL &&
984 !(zio->io_flags & ZIO_FLAG_RAW)) {
985 uint64_t psize = BP_GET_PSIZE(bp);
986 void *cbuf = zio_buf_alloc(psize);
988 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
991 if (!dmu_ot[BP_GET_TYPE(bp)].ot_metadata && BP_GET_LEVEL(bp) == 0)
992 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
994 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
995 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
997 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
998 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1000 return (ZIO_PIPELINE_CONTINUE);
1004 zio_write_bp_init(zio_t *zio)
1006 spa_t *spa = zio->io_spa;
1007 zio_prop_t *zp = &zio->io_prop;
1008 enum zio_compress compress = zp->zp_compress;
1009 blkptr_t *bp = zio->io_bp;
1010 uint64_t lsize = zio->io_size;
1011 uint64_t psize = lsize;
1015 * If our children haven't all reached the ready stage,
1016 * wait for them and then repeat this pipeline stage.
1018 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1019 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1020 return (ZIO_PIPELINE_STOP);
1022 if (!IO_IS_ALLOCATING(zio))
1023 return (ZIO_PIPELINE_CONTINUE);
1025 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1027 if (zio->io_bp_override) {
1028 ASSERT(bp->blk_birth != zio->io_txg);
1029 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1031 *bp = *zio->io_bp_override;
1032 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1034 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1035 return (ZIO_PIPELINE_CONTINUE);
1037 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1038 zp->zp_dedup_verify);
1040 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1041 BP_SET_DEDUP(bp, 1);
1042 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1043 return (ZIO_PIPELINE_CONTINUE);
1045 zio->io_bp_override = NULL;
1049 if (bp->blk_birth == zio->io_txg) {
1051 * We're rewriting an existing block, which means we're
1052 * working on behalf of spa_sync(). For spa_sync() to
1053 * converge, it must eventually be the case that we don't
1054 * have to allocate new blocks. But compression changes
1055 * the blocksize, which forces a reallocate, and makes
1056 * convergence take longer. Therefore, after the first
1057 * few passes, stop compressing to ensure convergence.
1059 pass = spa_sync_pass(spa);
1061 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1062 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1063 ASSERT(!BP_GET_DEDUP(bp));
1065 if (pass > SYNC_PASS_DONT_COMPRESS)
1066 compress = ZIO_COMPRESS_OFF;
1068 /* Make sure someone doesn't change their mind on overwrites */
1069 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
1070 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1073 if (compress != ZIO_COMPRESS_OFF) {
1074 void *cbuf = zio_buf_alloc(lsize);
1075 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1076 if (psize == 0 || psize == lsize) {
1077 compress = ZIO_COMPRESS_OFF;
1078 zio_buf_free(cbuf, lsize);
1080 ASSERT(psize < lsize);
1081 zio_push_transform(zio, cbuf, psize, lsize, NULL);
1086 * The final pass of spa_sync() must be all rewrites, but the first
1087 * few passes offer a trade-off: allocating blocks defers convergence,
1088 * but newly allocated blocks are sequential, so they can be written
1089 * to disk faster. Therefore, we allow the first few passes of
1090 * spa_sync() to allocate new blocks, but force rewrites after that.
1091 * There should only be a handful of blocks after pass 1 in any case.
1093 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
1094 pass > SYNC_PASS_REWRITE) {
1095 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1097 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1098 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1101 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1105 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1107 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1108 BP_SET_LSIZE(bp, lsize);
1109 BP_SET_PSIZE(bp, psize);
1110 BP_SET_COMPRESS(bp, compress);
1111 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1112 BP_SET_TYPE(bp, zp->zp_type);
1113 BP_SET_LEVEL(bp, zp->zp_level);
1114 BP_SET_DEDUP(bp, zp->zp_dedup);
1115 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1117 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1118 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1119 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1123 return (ZIO_PIPELINE_CONTINUE);
1127 zio_free_bp_init(zio_t *zio)
1129 blkptr_t *bp = zio->io_bp;
1131 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1132 if (BP_GET_DEDUP(bp))
1133 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1136 return (ZIO_PIPELINE_CONTINUE);
1140 * ==========================================================================
1141 * Execute the I/O pipeline
1142 * ==========================================================================
1146 zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q, boolean_t cutinline)
1148 spa_t *spa = zio->io_spa;
1149 zio_type_t t = zio->io_type;
1150 int flags = (cutinline ? TQ_FRONT : 0);
1153 * If we're a config writer or a probe, the normal issue and
1154 * interrupt threads may all be blocked waiting for the config lock.
1155 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1157 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1161 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1163 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1167 * If this is a high priority I/O, then use the high priority taskq.
1169 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1170 spa->spa_zio_taskq[t][q + 1] != NULL)
1173 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1176 * NB: We are assuming that the zio can only be dispatched
1177 * to a single taskq at a time. It would be a grievous error
1178 * to dispatch the zio to another taskq at the same time.
1180 ASSERT(taskq_empty_ent(&zio->io_tqent));
1181 taskq_dispatch_ent(spa->spa_zio_taskq[t][q],
1182 (task_func_t *)zio_execute, zio, flags, &zio->io_tqent);
1186 zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
1188 kthread_t *executor = zio->io_executor;
1189 spa_t *spa = zio->io_spa;
1192 for (t = 0; t < ZIO_TYPES; t++)
1193 if (taskq_member(spa->spa_zio_taskq[t][q], executor))
1200 zio_issue_async(zio_t *zio)
1202 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1204 return (ZIO_PIPELINE_STOP);
1208 zio_interrupt(zio_t *zio)
1210 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1214 * Execute the I/O pipeline until one of the following occurs:
1215 * (1) the I/O completes; (2) the pipeline stalls waiting for
1216 * dependent child I/Os; (3) the I/O issues, so we're waiting
1217 * for an I/O completion interrupt; (4) the I/O is delegated by
1218 * vdev-level caching or aggregation; (5) the I/O is deferred
1219 * due to vdev-level queueing; (6) the I/O is handed off to
1220 * another thread. In all cases, the pipeline stops whenever
1221 * there's no CPU work; it never burns a thread in cv_wait().
1223 * There's no locking on io_stage because there's no legitimate way
1224 * for multiple threads to be attempting to process the same I/O.
1226 static zio_pipe_stage_t *zio_pipeline[];
1229 * zio_execute() is a wrapper around the static function
1230 * __zio_execute() so that we can force __zio_execute() to be
1231 * inlined. This reduces stack overhead which is important
1232 * because __zio_execute() is called recursively in several zio
1233 * code paths. zio_execute() itself cannot be inlined because
1234 * it is externally visible.
1237 zio_execute(zio_t *zio)
1242 __attribute__((always_inline))
1244 __zio_execute(zio_t *zio)
1246 zio->io_executor = curthread;
1248 while (zio->io_stage < ZIO_STAGE_DONE) {
1249 enum zio_stage pipeline = zio->io_pipeline;
1250 enum zio_stage stage = zio->io_stage;
1255 ASSERT(!MUTEX_HELD(&zio->io_lock));
1256 ASSERT(ISP2(stage));
1257 ASSERT(zio->io_stall == NULL);
1261 } while ((stage & pipeline) == 0);
1263 ASSERT(stage <= ZIO_STAGE_DONE);
1265 dsl = spa_get_dsl(zio->io_spa);
1266 cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1267 zio_requeue_io_start_cut_in_line : B_FALSE;
1270 * If we are in interrupt context and this pipeline stage
1271 * will grab a config lock that is held across I/O,
1272 * or may wait for an I/O that needs an interrupt thread
1273 * to complete, issue async to avoid deadlock.
1275 * If we are in the txg_sync_thread or being called
1276 * during pool init issue async to minimize stack depth.
1277 * Both of these call paths may be recursively called.
1279 * For VDEV_IO_START, we cut in line so that the io will
1280 * be sent to disk promptly.
1282 if (((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1283 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) ||
1284 (dsl != NULL && dsl_pool_sync_context(dsl))) {
1285 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1289 zio->io_stage = stage;
1290 rv = zio_pipeline[highbit(stage) - 1](zio);
1292 if (rv == ZIO_PIPELINE_STOP)
1295 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1301 * ==========================================================================
1302 * Initiate I/O, either sync or async
1303 * ==========================================================================
1306 zio_wait(zio_t *zio)
1311 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1312 ASSERT(zio->io_executor == NULL);
1314 zio->io_waiter = curthread;
1315 timeout = ddi_get_lbolt() + (zio_delay_max / MILLISEC * hz);
1319 mutex_enter(&zio->io_lock);
1320 while (zio->io_executor != NULL) {
1322 * Wake up periodically to prevent the kernel from complaining
1323 * about a blocked task. However, check zio_delay_max to see
1324 * if the I/O has exceeded the timeout and post an ereport.
1326 cv_timedwait_interruptible(&zio->io_cv, &zio->io_lock,
1327 ddi_get_lbolt() + hz);
1329 if (timeout && (ddi_get_lbolt() > timeout)) {
1330 zio->io_delay = zio_delay_max;
1331 zfs_ereport_post(FM_EREPORT_ZFS_DELAY,
1332 zio->io_spa, zio->io_vd, zio, 0, 0);
1336 mutex_exit(&zio->io_lock);
1338 error = zio->io_error;
1345 zio_nowait(zio_t *zio)
1347 ASSERT(zio->io_executor == NULL);
1349 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1350 zio_unique_parent(zio) == NULL) {
1352 * This is a logical async I/O with no parent to wait for it.
1353 * We add it to the spa_async_root_zio "Godfather" I/O which
1354 * will ensure they complete prior to unloading the pool.
1356 spa_t *spa = zio->io_spa;
1358 zio_add_child(spa->spa_async_zio_root, zio);
1365 * ==========================================================================
1366 * Reexecute or suspend/resume failed I/O
1367 * ==========================================================================
1371 zio_reexecute(zio_t *pio)
1373 zio_t *cio, *cio_next;
1376 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1377 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1378 ASSERT(pio->io_gang_leader == NULL);
1379 ASSERT(pio->io_gang_tree == NULL);
1381 pio->io_flags = pio->io_orig_flags;
1382 pio->io_stage = pio->io_orig_stage;
1383 pio->io_pipeline = pio->io_orig_pipeline;
1384 pio->io_reexecute = 0;
1386 for (w = 0; w < ZIO_WAIT_TYPES; w++)
1387 pio->io_state[w] = 0;
1388 for (c = 0; c < ZIO_CHILD_TYPES; c++)
1389 pio->io_child_error[c] = 0;
1391 if (IO_IS_ALLOCATING(pio))
1392 BP_ZERO(pio->io_bp);
1395 * As we reexecute pio's children, new children could be created.
1396 * New children go to the head of pio's io_child_list, however,
1397 * so we will (correctly) not reexecute them. The key is that
1398 * the remainder of pio's io_child_list, from 'cio_next' onward,
1399 * cannot be affected by any side effects of reexecuting 'cio'.
1401 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1402 cio_next = zio_walk_children(pio);
1403 mutex_enter(&pio->io_lock);
1404 for (w = 0; w < ZIO_WAIT_TYPES; w++)
1405 pio->io_children[cio->io_child_type][w]++;
1406 mutex_exit(&pio->io_lock);
1411 * Now that all children have been reexecuted, execute the parent.
1412 * We don't reexecute "The Godfather" I/O here as it's the
1413 * responsibility of the caller to wait on him.
1415 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1420 zio_suspend(spa_t *spa, zio_t *zio)
1422 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1423 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1424 "failure and the failure mode property for this pool "
1425 "is set to panic.", spa_name(spa));
1427 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1429 mutex_enter(&spa->spa_suspend_lock);
1431 if (spa->spa_suspend_zio_root == NULL)
1432 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1433 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1434 ZIO_FLAG_GODFATHER);
1436 spa->spa_suspended = B_TRUE;
1439 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1440 ASSERT(zio != spa->spa_suspend_zio_root);
1441 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1442 ASSERT(zio_unique_parent(zio) == NULL);
1443 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1444 zio_add_child(spa->spa_suspend_zio_root, zio);
1447 mutex_exit(&spa->spa_suspend_lock);
1451 zio_resume(spa_t *spa)
1456 * Reexecute all previously suspended i/o.
1458 mutex_enter(&spa->spa_suspend_lock);
1459 spa->spa_suspended = B_FALSE;
1460 cv_broadcast(&spa->spa_suspend_cv);
1461 pio = spa->spa_suspend_zio_root;
1462 spa->spa_suspend_zio_root = NULL;
1463 mutex_exit(&spa->spa_suspend_lock);
1469 return (zio_wait(pio));
1473 zio_resume_wait(spa_t *spa)
1475 mutex_enter(&spa->spa_suspend_lock);
1476 while (spa_suspended(spa))
1477 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1478 mutex_exit(&spa->spa_suspend_lock);
1482 * ==========================================================================
1485 * A gang block is a collection of small blocks that looks to the DMU
1486 * like one large block. When zio_dva_allocate() cannot find a block
1487 * of the requested size, due to either severe fragmentation or the pool
1488 * being nearly full, it calls zio_write_gang_block() to construct the
1489 * block from smaller fragments.
1491 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1492 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1493 * an indirect block: it's an array of block pointers. It consumes
1494 * only one sector and hence is allocatable regardless of fragmentation.
1495 * The gang header's bps point to its gang members, which hold the data.
1497 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1498 * as the verifier to ensure uniqueness of the SHA256 checksum.
1499 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1500 * not the gang header. This ensures that data block signatures (needed for
1501 * deduplication) are independent of how the block is physically stored.
1503 * Gang blocks can be nested: a gang member may itself be a gang block.
1504 * Thus every gang block is a tree in which root and all interior nodes are
1505 * gang headers, and the leaves are normal blocks that contain user data.
1506 * The root of the gang tree is called the gang leader.
1508 * To perform any operation (read, rewrite, free, claim) on a gang block,
1509 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1510 * in the io_gang_tree field of the original logical i/o by recursively
1511 * reading the gang leader and all gang headers below it. This yields
1512 * an in-core tree containing the contents of every gang header and the
1513 * bps for every constituent of the gang block.
1515 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1516 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1517 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1518 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1519 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1520 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1521 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1522 * of the gang header plus zio_checksum_compute() of the data to update the
1523 * gang header's blk_cksum as described above.
1525 * The two-phase assemble/issue model solves the problem of partial failure --
1526 * what if you'd freed part of a gang block but then couldn't read the
1527 * gang header for another part? Assembling the entire gang tree first
1528 * ensures that all the necessary gang header I/O has succeeded before
1529 * starting the actual work of free, claim, or write. Once the gang tree
1530 * is assembled, free and claim are in-memory operations that cannot fail.
1532 * In the event that a gang write fails, zio_dva_unallocate() walks the
1533 * gang tree to immediately free (i.e. insert back into the space map)
1534 * everything we've allocated. This ensures that we don't get ENOSPC
1535 * errors during repeated suspend/resume cycles due to a flaky device.
1537 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1538 * the gang tree, we won't modify the block, so we can safely defer the free
1539 * (knowing that the block is still intact). If we *can* assemble the gang
1540 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1541 * each constituent bp and we can allocate a new block on the next sync pass.
1543 * In all cases, the gang tree allows complete recovery from partial failure.
1544 * ==========================================================================
1548 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1553 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1554 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1555 &pio->io_bookmark));
1559 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1564 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1565 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1566 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1568 * As we rewrite each gang header, the pipeline will compute
1569 * a new gang block header checksum for it; but no one will
1570 * compute a new data checksum, so we do that here. The one
1571 * exception is the gang leader: the pipeline already computed
1572 * its data checksum because that stage precedes gang assembly.
1573 * (Presently, nothing actually uses interior data checksums;
1574 * this is just good hygiene.)
1576 if (gn != pio->io_gang_leader->io_gang_tree) {
1577 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1578 data, BP_GET_PSIZE(bp));
1581 * If we are here to damage data for testing purposes,
1582 * leave the GBH alone so that we can detect the damage.
1584 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1585 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1587 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1588 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1589 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1597 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1599 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1600 ZIO_GANG_CHILD_FLAGS(pio)));
1605 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1607 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1608 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1611 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1620 static void zio_gang_tree_assemble_done(zio_t *zio);
1622 static zio_gang_node_t *
1623 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1625 zio_gang_node_t *gn;
1627 ASSERT(*gnpp == NULL);
1629 gn = kmem_zalloc(sizeof (*gn), KM_PUSHPAGE);
1630 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1637 zio_gang_node_free(zio_gang_node_t **gnpp)
1639 zio_gang_node_t *gn = *gnpp;
1642 for (g = 0; g < SPA_GBH_NBLKPTRS; g++)
1643 ASSERT(gn->gn_child[g] == NULL);
1645 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1646 kmem_free(gn, sizeof (*gn));
1651 zio_gang_tree_free(zio_gang_node_t **gnpp)
1653 zio_gang_node_t *gn = *gnpp;
1659 for (g = 0; g < SPA_GBH_NBLKPTRS; g++)
1660 zio_gang_tree_free(&gn->gn_child[g]);
1662 zio_gang_node_free(gnpp);
1666 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1668 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1670 ASSERT(gio->io_gang_leader == gio);
1671 ASSERT(BP_IS_GANG(bp));
1673 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1674 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1675 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1679 zio_gang_tree_assemble_done(zio_t *zio)
1681 zio_t *gio = zio->io_gang_leader;
1682 zio_gang_node_t *gn = zio->io_private;
1683 blkptr_t *bp = zio->io_bp;
1686 ASSERT(gio == zio_unique_parent(zio));
1687 ASSERT(zio->io_child_count == 0);
1692 if (BP_SHOULD_BYTESWAP(bp))
1693 byteswap_uint64_array(zio->io_data, zio->io_size);
1695 ASSERT(zio->io_data == gn->gn_gbh);
1696 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1697 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1699 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1700 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1701 if (!BP_IS_GANG(gbp))
1703 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1708 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1710 zio_t *gio = pio->io_gang_leader;
1714 ASSERT(BP_IS_GANG(bp) == !!gn);
1715 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1716 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1719 * If you're a gang header, your data is in gn->gn_gbh.
1720 * If you're a gang member, your data is in 'data' and gn == NULL.
1722 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1725 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1727 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1728 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1729 if (BP_IS_HOLE(gbp))
1731 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1732 data = (char *)data + BP_GET_PSIZE(gbp);
1736 if (gn == gio->io_gang_tree)
1737 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1744 zio_gang_assemble(zio_t *zio)
1746 blkptr_t *bp = zio->io_bp;
1748 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1749 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1751 zio->io_gang_leader = zio;
1753 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1755 return (ZIO_PIPELINE_CONTINUE);
1759 zio_gang_issue(zio_t *zio)
1761 blkptr_t *bp = zio->io_bp;
1763 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1764 return (ZIO_PIPELINE_STOP);
1766 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1767 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1769 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1770 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1772 zio_gang_tree_free(&zio->io_gang_tree);
1774 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1776 return (ZIO_PIPELINE_CONTINUE);
1780 zio_write_gang_member_ready(zio_t *zio)
1782 zio_t *pio = zio_unique_parent(zio);
1783 ASSERTV(zio_t *gio = zio->io_gang_leader;)
1784 dva_t *cdva = zio->io_bp->blk_dva;
1785 dva_t *pdva = pio->io_bp->blk_dva;
1789 if (BP_IS_HOLE(zio->io_bp))
1792 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1794 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1795 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1796 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1797 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1798 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1800 mutex_enter(&pio->io_lock);
1801 for (d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1802 ASSERT(DVA_GET_GANG(&pdva[d]));
1803 asize = DVA_GET_ASIZE(&pdva[d]);
1804 asize += DVA_GET_ASIZE(&cdva[d]);
1805 DVA_SET_ASIZE(&pdva[d], asize);
1807 mutex_exit(&pio->io_lock);
1811 zio_write_gang_block(zio_t *pio)
1813 spa_t *spa = pio->io_spa;
1814 blkptr_t *bp = pio->io_bp;
1815 zio_t *gio = pio->io_gang_leader;
1817 zio_gang_node_t *gn, **gnpp;
1818 zio_gbh_phys_t *gbh;
1819 uint64_t txg = pio->io_txg;
1820 uint64_t resid = pio->io_size;
1822 int copies = gio->io_prop.zp_copies;
1823 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1827 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1828 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1829 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1831 pio->io_error = error;
1832 return (ZIO_PIPELINE_CONTINUE);
1836 gnpp = &gio->io_gang_tree;
1838 gnpp = pio->io_private;
1839 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1842 gn = zio_gang_node_alloc(gnpp);
1844 bzero(gbh, SPA_GANGBLOCKSIZE);
1847 * Create the gang header.
1849 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1850 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1853 * Create and nowait the gang children.
1855 for (g = 0; resid != 0; resid -= lsize, g++) {
1856 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1858 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1860 zp.zp_checksum = gio->io_prop.zp_checksum;
1861 zp.zp_compress = ZIO_COMPRESS_OFF;
1862 zp.zp_type = DMU_OT_NONE;
1864 zp.zp_copies = gio->io_prop.zp_copies;
1866 zp.zp_dedup_verify = 0;
1868 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1869 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1870 zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1871 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1872 &pio->io_bookmark));
1876 * Set pio's pipeline to just wait for zio to finish.
1878 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1881 * We didn't allocate this bp, so make sure it doesn't get unmarked.
1883 pio->io_flags &= ~ZIO_FLAG_FASTWRITE;
1887 return (ZIO_PIPELINE_CONTINUE);
1891 * ==========================================================================
1893 * ==========================================================================
1896 zio_ddt_child_read_done(zio_t *zio)
1898 blkptr_t *bp = zio->io_bp;
1899 ddt_entry_t *dde = zio->io_private;
1901 zio_t *pio = zio_unique_parent(zio);
1903 mutex_enter(&pio->io_lock);
1904 ddp = ddt_phys_select(dde, bp);
1905 if (zio->io_error == 0)
1906 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1907 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1908 dde->dde_repair_data = zio->io_data;
1910 zio_buf_free(zio->io_data, zio->io_size);
1911 mutex_exit(&pio->io_lock);
1915 zio_ddt_read_start(zio_t *zio)
1917 blkptr_t *bp = zio->io_bp;
1920 ASSERT(BP_GET_DEDUP(bp));
1921 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1922 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1924 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1925 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1926 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1927 ddt_phys_t *ddp = dde->dde_phys;
1928 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1931 ASSERT(zio->io_vsd == NULL);
1934 if (ddp_self == NULL)
1935 return (ZIO_PIPELINE_CONTINUE);
1937 for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1938 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1940 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1942 zio_nowait(zio_read(zio, zio->io_spa, &blk,
1943 zio_buf_alloc(zio->io_size), zio->io_size,
1944 zio_ddt_child_read_done, dde, zio->io_priority,
1945 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1946 &zio->io_bookmark));
1948 return (ZIO_PIPELINE_CONTINUE);
1951 zio_nowait(zio_read(zio, zio->io_spa, bp,
1952 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1953 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1955 return (ZIO_PIPELINE_CONTINUE);
1959 zio_ddt_read_done(zio_t *zio)
1961 blkptr_t *bp = zio->io_bp;
1963 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1964 return (ZIO_PIPELINE_STOP);
1966 ASSERT(BP_GET_DEDUP(bp));
1967 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1968 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1970 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1971 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1972 ddt_entry_t *dde = zio->io_vsd;
1974 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
1975 return (ZIO_PIPELINE_CONTINUE);
1978 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
1979 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1980 return (ZIO_PIPELINE_STOP);
1982 if (dde->dde_repair_data != NULL) {
1983 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
1984 zio->io_child_error[ZIO_CHILD_DDT] = 0;
1986 ddt_repair_done(ddt, dde);
1990 ASSERT(zio->io_vsd == NULL);
1992 return (ZIO_PIPELINE_CONTINUE);
1996 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
1998 spa_t *spa = zio->io_spa;
2002 * Note: we compare the original data, not the transformed data,
2003 * because when zio->io_bp is an override bp, we will not have
2004 * pushed the I/O transforms. That's an important optimization
2005 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2007 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2008 zio_t *lio = dde->dde_lead_zio[p];
2011 return (lio->io_orig_size != zio->io_orig_size ||
2012 bcmp(zio->io_orig_data, lio->io_orig_data,
2013 zio->io_orig_size) != 0);
2017 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2018 ddt_phys_t *ddp = &dde->dde_phys[p];
2020 if (ddp->ddp_phys_birth != 0) {
2021 arc_buf_t *abuf = NULL;
2022 uint32_t aflags = ARC_WAIT;
2023 blkptr_t blk = *zio->io_bp;
2026 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2030 error = arc_read_nolock(NULL, spa, &blk,
2031 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2032 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2033 &aflags, &zio->io_bookmark);
2036 if (arc_buf_size(abuf) != zio->io_orig_size ||
2037 bcmp(abuf->b_data, zio->io_orig_data,
2038 zio->io_orig_size) != 0)
2040 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
2044 return (error != 0);
2052 zio_ddt_child_write_ready(zio_t *zio)
2054 int p = zio->io_prop.zp_copies;
2055 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2056 ddt_entry_t *dde = zio->io_private;
2057 ddt_phys_t *ddp = &dde->dde_phys[p];
2065 ASSERT(dde->dde_lead_zio[p] == zio);
2067 ddt_phys_fill(ddp, zio->io_bp);
2069 while ((pio = zio_walk_parents(zio)) != NULL)
2070 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2076 zio_ddt_child_write_done(zio_t *zio)
2078 int p = zio->io_prop.zp_copies;
2079 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2080 ddt_entry_t *dde = zio->io_private;
2081 ddt_phys_t *ddp = &dde->dde_phys[p];
2085 ASSERT(ddp->ddp_refcnt == 0);
2086 ASSERT(dde->dde_lead_zio[p] == zio);
2087 dde->dde_lead_zio[p] = NULL;
2089 if (zio->io_error == 0) {
2090 while (zio_walk_parents(zio) != NULL)
2091 ddt_phys_addref(ddp);
2093 ddt_phys_clear(ddp);
2100 zio_ddt_ditto_write_done(zio_t *zio)
2102 int p = DDT_PHYS_DITTO;
2103 blkptr_t *bp = zio->io_bp;
2104 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2105 ddt_entry_t *dde = zio->io_private;
2106 ddt_phys_t *ddp = &dde->dde_phys[p];
2107 ddt_key_t *ddk = &dde->dde_key;
2108 ASSERTV(zio_prop_t *zp = &zio->io_prop);
2112 ASSERT(ddp->ddp_refcnt == 0);
2113 ASSERT(dde->dde_lead_zio[p] == zio);
2114 dde->dde_lead_zio[p] = NULL;
2116 if (zio->io_error == 0) {
2117 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2118 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2119 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2120 if (ddp->ddp_phys_birth != 0)
2121 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2122 ddt_phys_fill(ddp, bp);
2129 zio_ddt_write(zio_t *zio)
2131 spa_t *spa = zio->io_spa;
2132 blkptr_t *bp = zio->io_bp;
2133 uint64_t txg = zio->io_txg;
2134 zio_prop_t *zp = &zio->io_prop;
2135 int p = zp->zp_copies;
2139 ddt_t *ddt = ddt_select(spa, bp);
2143 ASSERT(BP_GET_DEDUP(bp));
2144 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2145 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2148 dde = ddt_lookup(ddt, bp, B_TRUE);
2149 ddp = &dde->dde_phys[p];
2151 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2153 * If we're using a weak checksum, upgrade to a strong checksum
2154 * and try again. If we're already using a strong checksum,
2155 * we can't resolve it, so just convert to an ordinary write.
2156 * (And automatically e-mail a paper to Nature?)
2158 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2159 zp->zp_checksum = spa_dedup_checksum(spa);
2160 zio_pop_transforms(zio);
2161 zio->io_stage = ZIO_STAGE_OPEN;
2166 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2168 return (ZIO_PIPELINE_CONTINUE);
2171 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2172 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2174 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2175 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2176 zio_prop_t czp = *zp;
2178 czp.zp_copies = ditto_copies;
2181 * If we arrived here with an override bp, we won't have run
2182 * the transform stack, so we won't have the data we need to
2183 * generate a child i/o. So, toss the override bp and restart.
2184 * This is safe, because using the override bp is just an
2185 * optimization; and it's rare, so the cost doesn't matter.
2187 if (zio->io_bp_override) {
2188 zio_pop_transforms(zio);
2189 zio->io_stage = ZIO_STAGE_OPEN;
2190 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2191 zio->io_bp_override = NULL;
2194 return (ZIO_PIPELINE_CONTINUE);
2197 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2198 zio->io_orig_size, &czp, NULL,
2199 zio_ddt_ditto_write_done, dde, zio->io_priority,
2200 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2202 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2203 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2206 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2207 if (ddp->ddp_phys_birth != 0)
2208 ddt_bp_fill(ddp, bp, txg);
2209 if (dde->dde_lead_zio[p] != NULL)
2210 zio_add_child(zio, dde->dde_lead_zio[p]);
2212 ddt_phys_addref(ddp);
2213 } else if (zio->io_bp_override) {
2214 ASSERT(bp->blk_birth == txg);
2215 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2216 ddt_phys_fill(ddp, bp);
2217 ddt_phys_addref(ddp);
2219 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2220 zio->io_orig_size, zp, zio_ddt_child_write_ready,
2221 zio_ddt_child_write_done, dde, zio->io_priority,
2222 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2224 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2225 dde->dde_lead_zio[p] = cio;
2235 return (ZIO_PIPELINE_CONTINUE);
2238 ddt_entry_t *freedde; /* for debugging */
2241 zio_ddt_free(zio_t *zio)
2243 spa_t *spa = zio->io_spa;
2244 blkptr_t *bp = zio->io_bp;
2245 ddt_t *ddt = ddt_select(spa, bp);
2249 ASSERT(BP_GET_DEDUP(bp));
2250 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2253 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2254 ddp = ddt_phys_select(dde, bp);
2255 ddt_phys_decref(ddp);
2258 return (ZIO_PIPELINE_CONTINUE);
2262 * ==========================================================================
2263 * Allocate and free blocks
2264 * ==========================================================================
2267 zio_dva_allocate(zio_t *zio)
2269 spa_t *spa = zio->io_spa;
2270 metaslab_class_t *mc = spa_normal_class(spa);
2271 blkptr_t *bp = zio->io_bp;
2275 if (zio->io_gang_leader == NULL) {
2276 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2277 zio->io_gang_leader = zio;
2280 ASSERT(BP_IS_HOLE(bp));
2281 ASSERT3U(BP_GET_NDVAS(bp), ==, 0);
2282 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2283 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2284 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2287 * The dump device does not support gang blocks so allocation on
2288 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2289 * the "fast" gang feature.
2291 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2292 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2293 METASLAB_GANG_CHILD : 0;
2294 flags |= (zio->io_flags & ZIO_FLAG_FASTWRITE) ? METASLAB_FASTWRITE : 0;
2295 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2296 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2299 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2300 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2302 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2303 return (zio_write_gang_block(zio));
2304 zio->io_error = error;
2307 return (ZIO_PIPELINE_CONTINUE);
2311 zio_dva_free(zio_t *zio)
2313 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2315 return (ZIO_PIPELINE_CONTINUE);
2319 zio_dva_claim(zio_t *zio)
2323 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2325 zio->io_error = error;
2327 return (ZIO_PIPELINE_CONTINUE);
2331 * Undo an allocation. This is used by zio_done() when an I/O fails
2332 * and we want to give back the block we just allocated.
2333 * This handles both normal blocks and gang blocks.
2336 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2340 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2341 ASSERT(zio->io_bp_override == NULL);
2343 if (!BP_IS_HOLE(bp))
2344 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2347 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2348 zio_dva_unallocate(zio, gn->gn_child[g],
2349 &gn->gn_gbh->zg_blkptr[g]);
2355 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2358 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, uint64_t size,
2363 ASSERT(txg > spa_syncing_txg(spa));
2366 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2367 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2368 * when allocating them.
2371 error = metaslab_alloc(spa, spa_log_class(spa), size,
2372 new_bp, 1, txg, NULL,
2373 METASLAB_FASTWRITE | METASLAB_GANG_AVOID);
2377 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2378 new_bp, 1, txg, NULL,
2379 METASLAB_FASTWRITE | METASLAB_GANG_AVOID);
2383 BP_SET_LSIZE(new_bp, size);
2384 BP_SET_PSIZE(new_bp, size);
2385 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2386 BP_SET_CHECKSUM(new_bp,
2387 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2388 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2389 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2390 BP_SET_LEVEL(new_bp, 0);
2391 BP_SET_DEDUP(new_bp, 0);
2392 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2399 * Free an intent log block.
2402 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2404 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2405 ASSERT(!BP_IS_GANG(bp));
2407 zio_free(spa, txg, bp);
2411 * ==========================================================================
2412 * Read and write to physical devices
2413 * ==========================================================================
2416 zio_vdev_io_start(zio_t *zio)
2418 vdev_t *vd = zio->io_vd;
2420 spa_t *spa = zio->io_spa;
2422 ASSERT(zio->io_error == 0);
2423 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2426 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2427 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2430 * The mirror_ops handle multiple DVAs in a single BP.
2432 return (vdev_mirror_ops.vdev_op_io_start(zio));
2436 * We keep track of time-sensitive I/Os so that the scan thread
2437 * can quickly react to certain workloads. In particular, we care
2438 * about non-scrubbing, top-level reads and writes with the following
2440 * - synchronous writes of user data to non-slog devices
2441 * - any reads of user data
2442 * When these conditions are met, adjust the timestamp of spa_last_io
2443 * which allows the scan thread to adjust its workload accordingly.
2445 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2446 vd == vd->vdev_top && !vd->vdev_islog &&
2447 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2448 zio->io_txg != spa_syncing_txg(spa)) {
2449 uint64_t old = spa->spa_last_io;
2450 uint64_t new = ddi_get_lbolt64();
2452 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2455 align = 1ULL << vd->vdev_top->vdev_ashift;
2457 if (P2PHASE(zio->io_size, align) != 0) {
2458 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2459 char *abuf = zio_buf_alloc(asize);
2460 ASSERT(vd == vd->vdev_top);
2461 if (zio->io_type == ZIO_TYPE_WRITE) {
2462 bcopy(zio->io_data, abuf, zio->io_size);
2463 bzero(abuf + zio->io_size, asize - zio->io_size);
2465 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2468 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2469 ASSERT(P2PHASE(zio->io_size, align) == 0);
2470 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2473 * If this is a repair I/O, and there's no self-healing involved --
2474 * that is, we're just resilvering what we expect to resilver --
2475 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2476 * This prevents spurious resilvering with nested replication.
2477 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2478 * A is out of date, we'll read from C+D, then use the data to
2479 * resilver A+B -- but we don't actually want to resilver B, just A.
2480 * The top-level mirror has no way to know this, so instead we just
2481 * discard unnecessary repairs as we work our way down the vdev tree.
2482 * The same logic applies to any form of nested replication:
2483 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2485 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2486 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2487 zio->io_txg != 0 && /* not a delegated i/o */
2488 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2489 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2490 zio_vdev_io_bypass(zio);
2491 return (ZIO_PIPELINE_CONTINUE);
2494 if (vd->vdev_ops->vdev_op_leaf &&
2495 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2497 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2498 return (ZIO_PIPELINE_CONTINUE);
2500 if ((zio = vdev_queue_io(zio)) == NULL)
2501 return (ZIO_PIPELINE_STOP);
2503 if (!vdev_accessible(vd, zio)) {
2504 zio->io_error = ENXIO;
2506 return (ZIO_PIPELINE_STOP);
2510 return (vd->vdev_ops->vdev_op_io_start(zio));
2514 zio_vdev_io_done(zio_t *zio)
2516 vdev_t *vd = zio->io_vd;
2517 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2518 boolean_t unexpected_error = B_FALSE;
2520 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2521 return (ZIO_PIPELINE_STOP);
2523 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2525 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2527 vdev_queue_io_done(zio);
2529 if (zio->io_type == ZIO_TYPE_WRITE)
2530 vdev_cache_write(zio);
2532 if (zio_injection_enabled && zio->io_error == 0)
2533 zio->io_error = zio_handle_device_injection(vd,
2536 if (zio_injection_enabled && zio->io_error == 0)
2537 zio->io_error = zio_handle_label_injection(zio, EIO);
2539 if (zio->io_error) {
2540 if (!vdev_accessible(vd, zio)) {
2541 zio->io_error = ENXIO;
2543 unexpected_error = B_TRUE;
2548 ops->vdev_op_io_done(zio);
2550 if (unexpected_error)
2551 VERIFY(vdev_probe(vd, zio) == NULL);
2553 return (ZIO_PIPELINE_CONTINUE);
2557 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2558 * disk, and use that to finish the checksum ereport later.
2561 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2562 const void *good_buf)
2564 /* no processing needed */
2565 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2570 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2572 void *buf = zio_buf_alloc(zio->io_size);
2574 bcopy(zio->io_data, buf, zio->io_size);
2576 zcr->zcr_cbinfo = zio->io_size;
2577 zcr->zcr_cbdata = buf;
2578 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2579 zcr->zcr_free = zio_buf_free;
2583 zio_vdev_io_assess(zio_t *zio)
2585 vdev_t *vd = zio->io_vd;
2587 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2588 return (ZIO_PIPELINE_STOP);
2590 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2591 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2593 if (zio->io_vsd != NULL) {
2594 zio->io_vsd_ops->vsd_free(zio);
2598 if (zio_injection_enabled && zio->io_error == 0)
2599 zio->io_error = zio_handle_fault_injection(zio, EIO);
2602 * If the I/O failed, determine whether we should attempt to retry it.
2604 * On retry, we cut in line in the issue queue, since we don't want
2605 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2607 if (zio->io_error && vd == NULL &&
2608 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2609 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2610 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2612 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2613 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2614 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2615 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2616 zio_requeue_io_start_cut_in_line);
2617 return (ZIO_PIPELINE_STOP);
2621 * If we got an error on a leaf device, convert it to ENXIO
2622 * if the device is not accessible at all.
2624 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2625 !vdev_accessible(vd, zio))
2626 zio->io_error = ENXIO;
2629 * If we can't write to an interior vdev (mirror or RAID-Z),
2630 * set vdev_cant_write so that we stop trying to allocate from it.
2632 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2633 vd != NULL && !vd->vdev_ops->vdev_op_leaf)
2634 vd->vdev_cant_write = B_TRUE;
2637 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2639 return (ZIO_PIPELINE_CONTINUE);
2643 zio_vdev_io_reissue(zio_t *zio)
2645 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2646 ASSERT(zio->io_error == 0);
2648 zio->io_stage >>= 1;
2652 zio_vdev_io_redone(zio_t *zio)
2654 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2656 zio->io_stage >>= 1;
2660 zio_vdev_io_bypass(zio_t *zio)
2662 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2663 ASSERT(zio->io_error == 0);
2665 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2666 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2670 * ==========================================================================
2671 * Generate and verify checksums
2672 * ==========================================================================
2675 zio_checksum_generate(zio_t *zio)
2677 blkptr_t *bp = zio->io_bp;
2678 enum zio_checksum checksum;
2682 * This is zio_write_phys().
2683 * We're either generating a label checksum, or none at all.
2685 checksum = zio->io_prop.zp_checksum;
2687 if (checksum == ZIO_CHECKSUM_OFF)
2688 return (ZIO_PIPELINE_CONTINUE);
2690 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2692 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2693 ASSERT(!IO_IS_ALLOCATING(zio));
2694 checksum = ZIO_CHECKSUM_GANG_HEADER;
2696 checksum = BP_GET_CHECKSUM(bp);
2700 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2702 return (ZIO_PIPELINE_CONTINUE);
2706 zio_checksum_verify(zio_t *zio)
2708 zio_bad_cksum_t info;
2709 blkptr_t *bp = zio->io_bp;
2712 ASSERT(zio->io_vd != NULL);
2716 * This is zio_read_phys().
2717 * We're either verifying a label checksum, or nothing at all.
2719 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2720 return (ZIO_PIPELINE_CONTINUE);
2722 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2725 if ((error = zio_checksum_error(zio, &info)) != 0) {
2726 zio->io_error = error;
2727 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2728 zfs_ereport_start_checksum(zio->io_spa,
2729 zio->io_vd, zio, zio->io_offset,
2730 zio->io_size, NULL, &info);
2734 return (ZIO_PIPELINE_CONTINUE);
2738 * Called by RAID-Z to ensure we don't compute the checksum twice.
2741 zio_checksum_verified(zio_t *zio)
2743 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2747 * ==========================================================================
2748 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2749 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2750 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2751 * indicate errors that are specific to one I/O, and most likely permanent.
2752 * Any other error is presumed to be worse because we weren't expecting it.
2753 * ==========================================================================
2756 zio_worst_error(int e1, int e2)
2758 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2761 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2762 if (e1 == zio_error_rank[r1])
2765 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2766 if (e2 == zio_error_rank[r2])
2769 return (r1 > r2 ? e1 : e2);
2773 * ==========================================================================
2775 * ==========================================================================
2778 zio_ready(zio_t *zio)
2780 blkptr_t *bp = zio->io_bp;
2781 zio_t *pio, *pio_next;
2783 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2784 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2785 return (ZIO_PIPELINE_STOP);
2787 if (zio->io_ready) {
2788 ASSERT(IO_IS_ALLOCATING(zio));
2789 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2790 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2795 if (bp != NULL && bp != &zio->io_bp_copy)
2796 zio->io_bp_copy = *bp;
2799 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2801 mutex_enter(&zio->io_lock);
2802 zio->io_state[ZIO_WAIT_READY] = 1;
2803 pio = zio_walk_parents(zio);
2804 mutex_exit(&zio->io_lock);
2807 * As we notify zio's parents, new parents could be added.
2808 * New parents go to the head of zio's io_parent_list, however,
2809 * so we will (correctly) not notify them. The remainder of zio's
2810 * io_parent_list, from 'pio_next' onward, cannot change because
2811 * all parents must wait for us to be done before they can be done.
2813 for (; pio != NULL; pio = pio_next) {
2814 pio_next = zio_walk_parents(zio);
2815 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2818 if (zio->io_flags & ZIO_FLAG_NODATA) {
2819 if (BP_IS_GANG(bp)) {
2820 zio->io_flags &= ~ZIO_FLAG_NODATA;
2822 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2823 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2827 if (zio_injection_enabled &&
2828 zio->io_spa->spa_syncing_txg == zio->io_txg)
2829 zio_handle_ignored_writes(zio);
2831 return (ZIO_PIPELINE_CONTINUE);
2835 zio_done(zio_t *zio)
2837 zio_t *pio, *pio_next;
2841 * If our children haven't all completed,
2842 * wait for them and then repeat this pipeline stage.
2844 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2845 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2846 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2847 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2848 return (ZIO_PIPELINE_STOP);
2850 for (c = 0; c < ZIO_CHILD_TYPES; c++)
2851 for (w = 0; w < ZIO_WAIT_TYPES; w++)
2852 ASSERT(zio->io_children[c][w] == 0);
2854 if (zio->io_bp != NULL) {
2855 ASSERT(zio->io_bp->blk_pad[0] == 0);
2856 ASSERT(zio->io_bp->blk_pad[1] == 0);
2857 ASSERT(bcmp(zio->io_bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2858 (zio->io_bp == zio_unique_parent(zio)->io_bp));
2859 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) &&
2860 zio->io_bp_override == NULL &&
2861 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2862 ASSERT(!BP_SHOULD_BYTESWAP(zio->io_bp));
2863 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2864 ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 ||
2865 (BP_COUNT_GANG(zio->io_bp) == BP_GET_NDVAS(zio->io_bp)));
2870 * If there were child vdev/gang/ddt errors, they apply to us now.
2872 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2873 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2874 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2877 * If the I/O on the transformed data was successful, generate any
2878 * checksum reports now while we still have the transformed data.
2880 if (zio->io_error == 0) {
2881 while (zio->io_cksum_report != NULL) {
2882 zio_cksum_report_t *zcr = zio->io_cksum_report;
2883 uint64_t align = zcr->zcr_align;
2884 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2885 char *abuf = zio->io_data;
2887 if (asize != zio->io_size) {
2888 abuf = zio_buf_alloc(asize);
2889 bcopy(zio->io_data, abuf, zio->io_size);
2890 bzero(abuf + zio->io_size, asize - zio->io_size);
2893 zio->io_cksum_report = zcr->zcr_next;
2894 zcr->zcr_next = NULL;
2895 zcr->zcr_finish(zcr, abuf);
2896 zfs_ereport_free_checksum(zcr);
2898 if (asize != zio->io_size)
2899 zio_buf_free(abuf, asize);
2903 zio_pop_transforms(zio); /* note: may set zio->io_error */
2905 vdev_stat_update(zio, zio->io_size);
2908 * When an I/O completes but was slow post an ereport.
2910 if (zio->io_delay >= zio_delay_max)
2911 zfs_ereport_post(FM_EREPORT_ZFS_DELAY, zio->io_spa,
2912 zio->io_vd, zio, 0, 0);
2914 if (zio->io_error) {
2916 * If this I/O is attached to a particular vdev,
2917 * generate an error message describing the I/O failure
2918 * at the block level. We ignore these errors if the
2919 * device is currently unavailable.
2921 if (zio->io_error != ECKSUM && zio->io_vd != NULL &&
2922 !vdev_is_dead(zio->io_vd))
2923 zfs_ereport_post(FM_EREPORT_ZFS_IO, zio->io_spa,
2924 zio->io_vd, zio, 0, 0);
2926 if ((zio->io_error == EIO || !(zio->io_flags &
2927 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2928 zio == zio->io_logical) {
2930 * For logical I/O requests, tell the SPA to log the
2931 * error and generate a logical data ereport.
2933 spa_log_error(zio->io_spa, zio);
2934 zfs_ereport_post(FM_EREPORT_ZFS_DATA, zio->io_spa, NULL, zio,
2939 if (zio->io_error && zio == zio->io_logical) {
2941 * Determine whether zio should be reexecuted. This will
2942 * propagate all the way to the root via zio_notify_parent().
2944 ASSERT(zio->io_vd == NULL && zio->io_bp != NULL);
2945 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2947 if (IO_IS_ALLOCATING(zio) &&
2948 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2949 if (zio->io_error != ENOSPC)
2950 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2952 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2955 if ((zio->io_type == ZIO_TYPE_READ ||
2956 zio->io_type == ZIO_TYPE_FREE) &&
2957 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
2958 zio->io_error == ENXIO &&
2959 spa_load_state(zio->io_spa) == SPA_LOAD_NONE &&
2960 spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE)
2961 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2963 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2964 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2967 * Here is a possibly good place to attempt to do
2968 * either combinatorial reconstruction or error correction
2969 * based on checksums. It also might be a good place
2970 * to send out preliminary ereports before we suspend
2976 * If there were logical child errors, they apply to us now.
2977 * We defer this until now to avoid conflating logical child
2978 * errors with errors that happened to the zio itself when
2979 * updating vdev stats and reporting FMA events above.
2981 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2983 if ((zio->io_error || zio->io_reexecute) &&
2984 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
2985 !(zio->io_flags & ZIO_FLAG_IO_REWRITE))
2986 zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp);
2988 zio_gang_tree_free(&zio->io_gang_tree);
2991 * Godfather I/Os should never suspend.
2993 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
2994 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
2995 zio->io_reexecute = 0;
2997 if (zio->io_reexecute) {
2999 * This is a logical I/O that wants to reexecute.
3001 * Reexecute is top-down. When an i/o fails, if it's not
3002 * the root, it simply notifies its parent and sticks around.
3003 * The parent, seeing that it still has children in zio_done(),
3004 * does the same. This percolates all the way up to the root.
3005 * The root i/o will reexecute or suspend the entire tree.
3007 * This approach ensures that zio_reexecute() honors
3008 * all the original i/o dependency relationships, e.g.
3009 * parents not executing until children are ready.
3011 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3013 zio->io_gang_leader = NULL;
3015 mutex_enter(&zio->io_lock);
3016 zio->io_state[ZIO_WAIT_DONE] = 1;
3017 mutex_exit(&zio->io_lock);
3020 * "The Godfather" I/O monitors its children but is
3021 * not a true parent to them. It will track them through
3022 * the pipeline but severs its ties whenever they get into
3023 * trouble (e.g. suspended). This allows "The Godfather"
3024 * I/O to return status without blocking.
3026 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3027 zio_link_t *zl = zio->io_walk_link;
3028 pio_next = zio_walk_parents(zio);
3030 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3031 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3032 zio_remove_child(pio, zio, zl);
3033 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3037 if ((pio = zio_unique_parent(zio)) != NULL) {
3039 * We're not a root i/o, so there's nothing to do
3040 * but notify our parent. Don't propagate errors
3041 * upward since we haven't permanently failed yet.
3043 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3044 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3045 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3046 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3048 * We'd fail again if we reexecuted now, so suspend
3049 * until conditions improve (e.g. device comes online).
3051 zio_suspend(zio->io_spa, zio);
3054 * Reexecution is potentially a huge amount of work.
3055 * Hand it off to the otherwise-unused claim taskq.
3057 ASSERT(taskq_empty_ent(&zio->io_tqent));
3058 (void) taskq_dispatch_ent(
3059 zio->io_spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
3060 (task_func_t *)zio_reexecute, zio, 0,
3063 return (ZIO_PIPELINE_STOP);
3066 ASSERT(zio->io_child_count == 0);
3067 ASSERT(zio->io_reexecute == 0);
3068 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3071 * Report any checksum errors, since the I/O is complete.
3073 while (zio->io_cksum_report != NULL) {
3074 zio_cksum_report_t *zcr = zio->io_cksum_report;
3075 zio->io_cksum_report = zcr->zcr_next;
3076 zcr->zcr_next = NULL;
3077 zcr->zcr_finish(zcr, NULL);
3078 zfs_ereport_free_checksum(zcr);
3081 if (zio->io_flags & ZIO_FLAG_FASTWRITE && zio->io_bp &&
3082 !BP_IS_HOLE(zio->io_bp)) {
3083 metaslab_fastwrite_unmark(zio->io_spa, zio->io_bp);
3087 * It is the responsibility of the done callback to ensure that this
3088 * particular zio is no longer discoverable for adoption, and as
3089 * such, cannot acquire any new parents.
3094 mutex_enter(&zio->io_lock);
3095 zio->io_state[ZIO_WAIT_DONE] = 1;
3096 mutex_exit(&zio->io_lock);
3098 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3099 zio_link_t *zl = zio->io_walk_link;
3100 pio_next = zio_walk_parents(zio);
3101 zio_remove_child(pio, zio, zl);
3102 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3105 if (zio->io_waiter != NULL) {
3106 mutex_enter(&zio->io_lock);
3107 zio->io_executor = NULL;
3108 cv_broadcast(&zio->io_cv);
3109 mutex_exit(&zio->io_lock);
3114 return (ZIO_PIPELINE_STOP);
3118 * ==========================================================================
3119 * I/O pipeline definition
3120 * ==========================================================================
3122 static zio_pipe_stage_t *zio_pipeline[] = {
3128 zio_checksum_generate,
3142 zio_checksum_verify,
3146 #if defined(_KERNEL) && defined(HAVE_SPL)
3147 /* Fault injection */
3148 EXPORT_SYMBOL(zio_injection_enabled);
3149 EXPORT_SYMBOL(zio_inject_fault);
3150 EXPORT_SYMBOL(zio_inject_list_next);
3151 EXPORT_SYMBOL(zio_clear_fault);
3152 EXPORT_SYMBOL(zio_handle_fault_injection);
3153 EXPORT_SYMBOL(zio_handle_device_injection);
3154 EXPORT_SYMBOL(zio_handle_label_injection);
3155 EXPORT_SYMBOL(zio_priority_table);
3156 EXPORT_SYMBOL(zio_type_name);
3158 module_param(zio_bulk_flags, int, 0644);
3159 MODULE_PARM_DESC(zio_bulk_flags, "Additional flags to pass to bulk buffers");
3161 module_param(zio_delay_max, int, 0644);
3162 MODULE_PARM_DESC(zio_delay_max, "Max zio millisec delay before posting event");
3164 module_param(zio_requeue_io_start_cut_in_line, int, 0644);
3165 MODULE_PARM_DESC(zio_requeue_io_start_cut_in_line, "Prioritize requeued I/O");