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);
221 kmem_cache_t *last_cache = NULL;
222 kmem_cache_t *last_data_cache = NULL;
224 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
225 if (zio_buf_cache[c] != last_cache) {
226 last_cache = zio_buf_cache[c];
227 kmem_cache_destroy(zio_buf_cache[c]);
229 zio_buf_cache[c] = NULL;
231 if (zio_data_buf_cache[c] != last_data_cache) {
232 last_data_cache = zio_data_buf_cache[c];
233 kmem_cache_destroy(zio_data_buf_cache[c]);
235 zio_data_buf_cache[c] = NULL;
238 kmem_cache_destroy(zio_vdev_cache);
239 kmem_cache_destroy(zio_link_cache);
240 kmem_cache_destroy(zio_cache);
248 * ==========================================================================
249 * Allocate and free I/O buffers
250 * ==========================================================================
254 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
255 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
256 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
257 * excess / transient data in-core during a crashdump.
260 zio_buf_alloc(size_t size)
262 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
264 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
266 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE | KM_NODEBUG));
270 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
271 * crashdump if the kernel panics. This exists so that we will limit the amount
272 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
273 * of kernel heap dumped to disk when the kernel panics)
276 zio_data_buf_alloc(size_t size)
278 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
280 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
282 return (kmem_cache_alloc(zio_data_buf_cache[c],
283 KM_PUSHPAGE | KM_NODEBUG));
287 zio_buf_free(void *buf, size_t size)
289 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
291 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
293 kmem_cache_free(zio_buf_cache[c], buf);
297 zio_data_buf_free(void *buf, size_t size)
299 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
301 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
303 kmem_cache_free(zio_data_buf_cache[c], buf);
307 * Dedicated I/O buffers to ensure that memory fragmentation never prevents
308 * or significantly delays the issuing of a zio. These buffers are used
309 * to aggregate I/O and could be used for raidz stripes.
314 return (kmem_cache_alloc(zio_vdev_cache, KM_PUSHPAGE));
318 zio_vdev_free(void *buf)
320 kmem_cache_free(zio_vdev_cache, buf);
325 * ==========================================================================
326 * Push and pop I/O transform buffers
327 * ==========================================================================
330 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
331 zio_transform_func_t *transform)
333 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_PUSHPAGE);
335 zt->zt_orig_data = zio->io_data;
336 zt->zt_orig_size = zio->io_size;
337 zt->zt_bufsize = bufsize;
338 zt->zt_transform = transform;
340 zt->zt_next = zio->io_transform_stack;
341 zio->io_transform_stack = zt;
348 zio_pop_transforms(zio_t *zio)
352 while ((zt = zio->io_transform_stack) != NULL) {
353 if (zt->zt_transform != NULL)
354 zt->zt_transform(zio,
355 zt->zt_orig_data, zt->zt_orig_size);
357 if (zt->zt_bufsize != 0)
358 zio_buf_free(zio->io_data, zt->zt_bufsize);
360 zio->io_data = zt->zt_orig_data;
361 zio->io_size = zt->zt_orig_size;
362 zio->io_transform_stack = zt->zt_next;
364 kmem_free(zt, sizeof (zio_transform_t));
369 * ==========================================================================
370 * I/O transform callbacks for subblocks and decompression
371 * ==========================================================================
374 zio_subblock(zio_t *zio, void *data, uint64_t size)
376 ASSERT(zio->io_size > size);
378 if (zio->io_type == ZIO_TYPE_READ)
379 bcopy(zio->io_data, data, size);
383 zio_decompress(zio_t *zio, void *data, uint64_t size)
385 if (zio->io_error == 0 &&
386 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
387 zio->io_data, data, zio->io_size, size) != 0)
392 * ==========================================================================
393 * I/O parent/child relationships and pipeline interlocks
394 * ==========================================================================
397 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
398 * continue calling these functions until they return NULL.
399 * Otherwise, the next caller will pick up the list walk in
400 * some indeterminate state. (Otherwise every caller would
401 * have to pass in a cookie to keep the state represented by
402 * io_walk_link, which gets annoying.)
405 zio_walk_parents(zio_t *cio)
407 zio_link_t *zl = cio->io_walk_link;
408 list_t *pl = &cio->io_parent_list;
410 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
411 cio->io_walk_link = zl;
416 ASSERT(zl->zl_child == cio);
417 return (zl->zl_parent);
421 zio_walk_children(zio_t *pio)
423 zio_link_t *zl = pio->io_walk_link;
424 list_t *cl = &pio->io_child_list;
426 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
427 pio->io_walk_link = zl;
432 ASSERT(zl->zl_parent == pio);
433 return (zl->zl_child);
437 zio_unique_parent(zio_t *cio)
439 zio_t *pio = zio_walk_parents(cio);
441 VERIFY(zio_walk_parents(cio) == NULL);
446 zio_add_child(zio_t *pio, zio_t *cio)
448 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_PUSHPAGE);
452 * Logical I/Os can have logical, gang, or vdev children.
453 * Gang I/Os can have gang or vdev children.
454 * Vdev I/Os can only have vdev children.
455 * The following ASSERT captures all of these constraints.
457 ASSERT(cio->io_child_type <= pio->io_child_type);
462 mutex_enter(&cio->io_lock);
463 mutex_enter(&pio->io_lock);
465 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
467 for (w = 0; w < ZIO_WAIT_TYPES; w++)
468 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
470 list_insert_head(&pio->io_child_list, zl);
471 list_insert_head(&cio->io_parent_list, zl);
473 pio->io_child_count++;
474 cio->io_parent_count++;
476 mutex_exit(&pio->io_lock);
477 mutex_exit(&cio->io_lock);
481 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
483 ASSERT(zl->zl_parent == pio);
484 ASSERT(zl->zl_child == cio);
486 mutex_enter(&cio->io_lock);
487 mutex_enter(&pio->io_lock);
489 list_remove(&pio->io_child_list, zl);
490 list_remove(&cio->io_parent_list, zl);
492 pio->io_child_count--;
493 cio->io_parent_count--;
495 mutex_exit(&pio->io_lock);
496 mutex_exit(&cio->io_lock);
498 kmem_cache_free(zio_link_cache, zl);
502 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
504 uint64_t *countp = &zio->io_children[child][wait];
505 boolean_t waiting = B_FALSE;
507 mutex_enter(&zio->io_lock);
508 ASSERT(zio->io_stall == NULL);
511 zio->io_stall = countp;
514 mutex_exit(&zio->io_lock);
519 __attribute__((always_inline))
521 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
523 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
524 int *errorp = &pio->io_child_error[zio->io_child_type];
526 mutex_enter(&pio->io_lock);
527 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
528 *errorp = zio_worst_error(*errorp, zio->io_error);
529 pio->io_reexecute |= zio->io_reexecute;
530 ASSERT3U(*countp, >, 0);
531 if (--*countp == 0 && pio->io_stall == countp) {
532 pio->io_stall = NULL;
533 mutex_exit(&pio->io_lock);
536 mutex_exit(&pio->io_lock);
541 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
543 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
544 zio->io_error = zio->io_child_error[c];
548 * ==========================================================================
549 * Create the various types of I/O (read, write, free, etc)
550 * ==========================================================================
553 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
554 void *data, uint64_t size, zio_done_func_t *done, void *private,
555 zio_type_t type, int priority, enum zio_flag flags,
556 vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
557 enum zio_stage stage, enum zio_stage pipeline)
561 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
562 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
563 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
565 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
566 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
567 ASSERT(vd || stage == ZIO_STAGE_OPEN);
569 zio = kmem_cache_alloc(zio_cache, KM_PUSHPAGE);
572 zio->io_child_type = ZIO_CHILD_VDEV;
573 else if (flags & ZIO_FLAG_GANG_CHILD)
574 zio->io_child_type = ZIO_CHILD_GANG;
575 else if (flags & ZIO_FLAG_DDT_CHILD)
576 zio->io_child_type = ZIO_CHILD_DDT;
578 zio->io_child_type = ZIO_CHILD_LOGICAL;
581 zio->io_logical = NULL;
582 zio->io_bp = (blkptr_t *)bp;
583 zio->io_bp_copy = *bp;
584 zio->io_bp_orig = *bp;
585 if (type != ZIO_TYPE_WRITE ||
586 zio->io_child_type == ZIO_CHILD_DDT)
587 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
588 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
589 zio->io_logical = zio;
590 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
591 pipeline |= ZIO_GANG_STAGES;
593 zio->io_logical = NULL;
595 bzero(&zio->io_bp_copy, sizeof (blkptr_t));
596 bzero(&zio->io_bp_orig, sizeof (blkptr_t));
601 zio->io_ready = NULL;
603 zio->io_private = private;
604 zio->io_prev_space_delta = 0;
606 zio->io_priority = priority;
609 zio->io_vsd_ops = NULL;
610 zio->io_offset = offset;
611 zio->io_deadline = 0;
612 zio->io_timestamp = 0;
615 zio->io_orig_data = zio->io_data = data;
616 zio->io_orig_size = zio->io_size = size;
617 zio->io_orig_flags = zio->io_flags = flags;
618 zio->io_orig_stage = zio->io_stage = stage;
619 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
620 bzero(&zio->io_prop, sizeof (zio_prop_t));
622 zio->io_reexecute = 0;
623 zio->io_bp_override = NULL;
624 zio->io_walk_link = NULL;
625 zio->io_transform_stack = NULL;
627 zio->io_child_count = 0;
628 zio->io_parent_count = 0;
629 zio->io_stall = NULL;
630 zio->io_gang_leader = NULL;
631 zio->io_gang_tree = NULL;
632 zio->io_executor = NULL;
633 zio->io_waiter = NULL;
634 zio->io_cksum_report = NULL;
636 bzero(zio->io_child_error, sizeof (int) * ZIO_CHILD_TYPES);
637 bzero(zio->io_children,
638 sizeof (uint64_t) * ZIO_CHILD_TYPES * ZIO_WAIT_TYPES);
639 bzero(&zio->io_bookmark, sizeof (zbookmark_t));
641 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
642 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
645 zio->io_bookmark = *zb;
648 if (zio->io_logical == NULL)
649 zio->io_logical = pio->io_logical;
650 if (zio->io_child_type == ZIO_CHILD_GANG)
651 zio->io_gang_leader = pio->io_gang_leader;
652 zio_add_child(pio, zio);
655 taskq_init_ent(&zio->io_tqent);
661 zio_destroy(zio_t *zio)
663 kmem_cache_free(zio_cache, zio);
667 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
668 void *private, enum zio_flag flags)
672 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
673 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
674 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
680 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
682 return (zio_null(NULL, spa, NULL, done, private, flags));
686 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
687 void *data, uint64_t size, zio_done_func_t *done, void *private,
688 int priority, enum zio_flag flags, const zbookmark_t *zb)
692 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
693 data, size, done, private,
694 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
695 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
696 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
702 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
703 void *data, uint64_t size, const zio_prop_t *zp,
704 zio_done_func_t *ready, zio_done_func_t *done, void *private,
705 int priority, enum zio_flag flags, const zbookmark_t *zb)
709 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
710 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
711 zp->zp_compress >= ZIO_COMPRESS_OFF &&
712 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
713 DMU_OT_IS_VALID(zp->zp_type) &&
716 zp->zp_copies <= spa_max_replication(spa) &&
718 zp->zp_dedup_verify <= 1);
720 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
721 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
722 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
723 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
725 zio->io_ready = ready;
732 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
733 uint64_t size, zio_done_func_t *done, void *private, int priority,
734 enum zio_flag flags, zbookmark_t *zb)
738 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
739 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
740 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
746 zio_write_override(zio_t *zio, blkptr_t *bp, int copies)
748 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
749 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
750 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
751 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
753 zio->io_prop.zp_copies = copies;
754 zio->io_bp_override = bp;
758 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
760 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
764 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
769 dprintf_bp(bp, "freeing in txg %llu, pass %u",
770 (longlong_t)txg, spa->spa_sync_pass);
772 ASSERT(!BP_IS_HOLE(bp));
773 ASSERT(spa_syncing_txg(spa) == txg);
774 ASSERT(spa_sync_pass(spa) <= SYNC_PASS_DEFERRED_FREE);
776 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
777 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
778 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
784 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
785 zio_done_func_t *done, void *private, enum zio_flag flags)
790 * A claim is an allocation of a specific block. Claims are needed
791 * to support immediate writes in the intent log. The issue is that
792 * immediate writes contain committed data, but in a txg that was
793 * *not* committed. Upon opening the pool after an unclean shutdown,
794 * the intent log claims all blocks that contain immediate write data
795 * so that the SPA knows they're in use.
797 * All claims *must* be resolved in the first txg -- before the SPA
798 * starts allocating blocks -- so that nothing is allocated twice.
799 * If txg == 0 we just verify that the block is claimable.
801 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
802 ASSERT(txg == spa_first_txg(spa) || txg == 0);
803 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
805 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
806 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
807 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
813 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
814 zio_done_func_t *done, void *private, int priority, enum zio_flag flags)
819 if (vd->vdev_children == 0) {
820 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
821 ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
822 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
826 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
828 for (c = 0; c < vd->vdev_children; c++)
829 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
830 done, private, priority, flags));
837 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
838 void *data, int checksum, zio_done_func_t *done, void *private,
839 int priority, enum zio_flag flags, boolean_t labels)
843 ASSERT(vd->vdev_children == 0);
844 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
845 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
846 ASSERT3U(offset + size, <=, vd->vdev_psize);
848 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
849 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
850 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
852 zio->io_prop.zp_checksum = checksum;
858 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
859 void *data, int checksum, zio_done_func_t *done, void *private,
860 int priority, enum zio_flag flags, boolean_t labels)
864 ASSERT(vd->vdev_children == 0);
865 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
866 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
867 ASSERT3U(offset + size, <=, vd->vdev_psize);
869 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
870 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
871 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
873 zio->io_prop.zp_checksum = checksum;
875 if (zio_checksum_table[checksum].ci_eck) {
877 * zec checksums are necessarily destructive -- they modify
878 * the end of the write buffer to hold the verifier/checksum.
879 * Therefore, we must make a local copy in case the data is
880 * being written to multiple places in parallel.
882 void *wbuf = zio_buf_alloc(size);
883 bcopy(data, wbuf, size);
884 zio_push_transform(zio, wbuf, size, size, NULL);
891 * Create a child I/O to do some work for us.
894 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
895 void *data, uint64_t size, int type, int priority, enum zio_flag flags,
896 zio_done_func_t *done, void *private)
898 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
901 ASSERT(vd->vdev_parent ==
902 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
904 if (type == ZIO_TYPE_READ && bp != NULL) {
906 * If we have the bp, then the child should perform the
907 * checksum and the parent need not. This pushes error
908 * detection as close to the leaves as possible and
909 * eliminates redundant checksums in the interior nodes.
911 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
912 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
915 if (vd->vdev_children == 0)
916 offset += VDEV_LABEL_START_SIZE;
918 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
921 * If we've decided to do a repair, the write is not speculative --
922 * even if the original read was.
924 if (flags & ZIO_FLAG_IO_REPAIR)
925 flags &= ~ZIO_FLAG_SPECULATIVE;
927 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
928 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
929 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
935 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
936 int type, int priority, enum zio_flag flags,
937 zio_done_func_t *done, void *private)
941 ASSERT(vd->vdev_ops->vdev_op_leaf);
943 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
944 data, size, done, private, type, priority,
945 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
947 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
953 zio_flush(zio_t *zio, vdev_t *vd)
955 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
956 NULL, NULL, ZIO_PRIORITY_NOW,
957 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
961 zio_shrink(zio_t *zio, uint64_t size)
963 ASSERT(zio->io_executor == NULL);
964 ASSERT(zio->io_orig_size == zio->io_size);
965 ASSERT(size <= zio->io_size);
968 * We don't shrink for raidz because of problems with the
969 * reconstruction when reading back less than the block size.
970 * Note, BP_IS_RAIDZ() assumes no compression.
972 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
973 if (!BP_IS_RAIDZ(zio->io_bp))
974 zio->io_orig_size = zio->io_size = size;
978 * ==========================================================================
979 * Prepare to read and write logical blocks
980 * ==========================================================================
984 zio_read_bp_init(zio_t *zio)
986 blkptr_t *bp = zio->io_bp;
988 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
989 zio->io_child_type == ZIO_CHILD_LOGICAL &&
990 !(zio->io_flags & ZIO_FLAG_RAW)) {
991 uint64_t psize = BP_GET_PSIZE(bp);
992 void *cbuf = zio_buf_alloc(psize);
994 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
997 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
998 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1000 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1001 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1003 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1004 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1006 return (ZIO_PIPELINE_CONTINUE);
1010 zio_write_bp_init(zio_t *zio)
1012 spa_t *spa = zio->io_spa;
1013 zio_prop_t *zp = &zio->io_prop;
1014 enum zio_compress compress = zp->zp_compress;
1015 blkptr_t *bp = zio->io_bp;
1016 uint64_t lsize = zio->io_size;
1017 uint64_t psize = lsize;
1021 * If our children haven't all reached the ready stage,
1022 * wait for them and then repeat this pipeline stage.
1024 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1025 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1026 return (ZIO_PIPELINE_STOP);
1028 if (!IO_IS_ALLOCATING(zio))
1029 return (ZIO_PIPELINE_CONTINUE);
1031 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1033 if (zio->io_bp_override) {
1034 ASSERT(bp->blk_birth != zio->io_txg);
1035 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1037 *bp = *zio->io_bp_override;
1038 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1040 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1041 return (ZIO_PIPELINE_CONTINUE);
1043 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1044 zp->zp_dedup_verify);
1046 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1047 BP_SET_DEDUP(bp, 1);
1048 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1049 return (ZIO_PIPELINE_CONTINUE);
1051 zio->io_bp_override = NULL;
1055 if (bp->blk_birth == zio->io_txg) {
1057 * We're rewriting an existing block, which means we're
1058 * working on behalf of spa_sync(). For spa_sync() to
1059 * converge, it must eventually be the case that we don't
1060 * have to allocate new blocks. But compression changes
1061 * the blocksize, which forces a reallocate, and makes
1062 * convergence take longer. Therefore, after the first
1063 * few passes, stop compressing to ensure convergence.
1065 pass = spa_sync_pass(spa);
1067 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1068 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1069 ASSERT(!BP_GET_DEDUP(bp));
1071 if (pass > SYNC_PASS_DONT_COMPRESS)
1072 compress = ZIO_COMPRESS_OFF;
1074 /* Make sure someone doesn't change their mind on overwrites */
1075 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
1076 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1079 if (compress != ZIO_COMPRESS_OFF) {
1080 void *cbuf = zio_buf_alloc(lsize);
1081 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1082 if (psize == 0 || psize == lsize) {
1083 compress = ZIO_COMPRESS_OFF;
1084 zio_buf_free(cbuf, lsize);
1086 ASSERT(psize < lsize);
1087 zio_push_transform(zio, cbuf, psize, lsize, NULL);
1092 * The final pass of spa_sync() must be all rewrites, but the first
1093 * few passes offer a trade-off: allocating blocks defers convergence,
1094 * but newly allocated blocks are sequential, so they can be written
1095 * to disk faster. Therefore, we allow the first few passes of
1096 * spa_sync() to allocate new blocks, but force rewrites after that.
1097 * There should only be a handful of blocks after pass 1 in any case.
1099 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
1100 pass > SYNC_PASS_REWRITE) {
1101 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1103 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1104 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1107 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1111 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1113 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1114 BP_SET_LSIZE(bp, lsize);
1115 BP_SET_PSIZE(bp, psize);
1116 BP_SET_COMPRESS(bp, compress);
1117 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1118 BP_SET_TYPE(bp, zp->zp_type);
1119 BP_SET_LEVEL(bp, zp->zp_level);
1120 BP_SET_DEDUP(bp, zp->zp_dedup);
1121 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1123 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1124 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1125 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1129 return (ZIO_PIPELINE_CONTINUE);
1133 zio_free_bp_init(zio_t *zio)
1135 blkptr_t *bp = zio->io_bp;
1137 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1138 if (BP_GET_DEDUP(bp))
1139 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1142 return (ZIO_PIPELINE_CONTINUE);
1146 * ==========================================================================
1147 * Execute the I/O pipeline
1148 * ==========================================================================
1152 zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q, boolean_t cutinline)
1154 spa_t *spa = zio->io_spa;
1155 zio_type_t t = zio->io_type;
1156 int flags = (cutinline ? TQ_FRONT : 0);
1159 * If we're a config writer or a probe, the normal issue and
1160 * interrupt threads may all be blocked waiting for the config lock.
1161 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1163 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1167 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1169 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1173 * If this is a high priority I/O, then use the high priority taskq.
1175 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1176 spa->spa_zio_taskq[t][q + 1] != NULL)
1179 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1182 * NB: We are assuming that the zio can only be dispatched
1183 * to a single taskq at a time. It would be a grievous error
1184 * to dispatch the zio to another taskq at the same time.
1186 ASSERT(taskq_empty_ent(&zio->io_tqent));
1187 taskq_dispatch_ent(spa->spa_zio_taskq[t][q],
1188 (task_func_t *)zio_execute, zio, flags, &zio->io_tqent);
1192 zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
1194 kthread_t *executor = zio->io_executor;
1195 spa_t *spa = zio->io_spa;
1198 for (t = 0; t < ZIO_TYPES; t++)
1199 if (taskq_member(spa->spa_zio_taskq[t][q], executor))
1206 zio_issue_async(zio_t *zio)
1208 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1210 return (ZIO_PIPELINE_STOP);
1214 zio_interrupt(zio_t *zio)
1216 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1220 * Execute the I/O pipeline until one of the following occurs:
1221 * (1) the I/O completes; (2) the pipeline stalls waiting for
1222 * dependent child I/Os; (3) the I/O issues, so we're waiting
1223 * for an I/O completion interrupt; (4) the I/O is delegated by
1224 * vdev-level caching or aggregation; (5) the I/O is deferred
1225 * due to vdev-level queueing; (6) the I/O is handed off to
1226 * another thread. In all cases, the pipeline stops whenever
1227 * there's no CPU work; it never burns a thread in cv_wait().
1229 * There's no locking on io_stage because there's no legitimate way
1230 * for multiple threads to be attempting to process the same I/O.
1232 static zio_pipe_stage_t *zio_pipeline[];
1235 * zio_execute() is a wrapper around the static function
1236 * __zio_execute() so that we can force __zio_execute() to be
1237 * inlined. This reduces stack overhead which is important
1238 * because __zio_execute() is called recursively in several zio
1239 * code paths. zio_execute() itself cannot be inlined because
1240 * it is externally visible.
1243 zio_execute(zio_t *zio)
1248 __attribute__((always_inline))
1250 __zio_execute(zio_t *zio)
1252 zio->io_executor = curthread;
1254 while (zio->io_stage < ZIO_STAGE_DONE) {
1255 enum zio_stage pipeline = zio->io_pipeline;
1256 enum zio_stage stage = zio->io_stage;
1261 ASSERT(!MUTEX_HELD(&zio->io_lock));
1262 ASSERT(ISP2(stage));
1263 ASSERT(zio->io_stall == NULL);
1267 } while ((stage & pipeline) == 0);
1269 ASSERT(stage <= ZIO_STAGE_DONE);
1271 dp = spa_get_dsl(zio->io_spa);
1272 cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1273 zio_requeue_io_start_cut_in_line : B_FALSE;
1276 * If we are in interrupt context and this pipeline stage
1277 * will grab a config lock that is held across I/O,
1278 * or may wait for an I/O that needs an interrupt thread
1279 * to complete, issue async to avoid deadlock.
1281 * For VDEV_IO_START, we cut in line so that the io will
1282 * be sent to disk promptly.
1284 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1285 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1286 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1292 * If we executing in the context of the tx_sync_thread,
1293 * or we are performing pool initialization outside of a
1294 * zio_taskq[ZIO_TASKQ_ISSUE] context. Then issue the zio
1295 * async to minimize stack usage for these deep call paths.
1297 if ((dp && curthread == dp->dp_tx.tx_sync_thread) ||
1298 (dp && spa_is_initializing(dp->dp_spa) &&
1299 !zio_taskq_member(zio, ZIO_TASKQ_ISSUE))) {
1300 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1305 zio->io_stage = stage;
1306 rv = zio_pipeline[highbit(stage) - 1](zio);
1308 if (rv == ZIO_PIPELINE_STOP)
1311 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1317 * ==========================================================================
1318 * Initiate I/O, either sync or async
1319 * ==========================================================================
1322 zio_wait(zio_t *zio)
1326 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1327 ASSERT(zio->io_executor == NULL);
1329 zio->io_waiter = curthread;
1333 mutex_enter(&zio->io_lock);
1334 while (zio->io_executor != NULL)
1335 cv_wait_io(&zio->io_cv, &zio->io_lock);
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);
2255 ddp = ddt_phys_select(dde, bp);
2257 ddt_phys_decref(ddp);
2261 return (ZIO_PIPELINE_CONTINUE);
2265 * ==========================================================================
2266 * Allocate and free blocks
2267 * ==========================================================================
2270 zio_dva_allocate(zio_t *zio)
2272 spa_t *spa = zio->io_spa;
2273 metaslab_class_t *mc = spa_normal_class(spa);
2274 blkptr_t *bp = zio->io_bp;
2278 if (zio->io_gang_leader == NULL) {
2279 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2280 zio->io_gang_leader = zio;
2283 ASSERT(BP_IS_HOLE(bp));
2284 ASSERT3U(BP_GET_NDVAS(bp), ==, 0);
2285 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2286 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2287 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2290 * The dump device does not support gang blocks so allocation on
2291 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2292 * the "fast" gang feature.
2294 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2295 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2296 METASLAB_GANG_CHILD : 0;
2297 flags |= (zio->io_flags & ZIO_FLAG_FASTWRITE) ? METASLAB_FASTWRITE : 0;
2298 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2299 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2302 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2303 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2305 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2306 return (zio_write_gang_block(zio));
2307 zio->io_error = error;
2310 return (ZIO_PIPELINE_CONTINUE);
2314 zio_dva_free(zio_t *zio)
2316 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2318 return (ZIO_PIPELINE_CONTINUE);
2322 zio_dva_claim(zio_t *zio)
2326 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2328 zio->io_error = error;
2330 return (ZIO_PIPELINE_CONTINUE);
2334 * Undo an allocation. This is used by zio_done() when an I/O fails
2335 * and we want to give back the block we just allocated.
2336 * This handles both normal blocks and gang blocks.
2339 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2343 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2344 ASSERT(zio->io_bp_override == NULL);
2346 if (!BP_IS_HOLE(bp))
2347 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2350 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2351 zio_dva_unallocate(zio, gn->gn_child[g],
2352 &gn->gn_gbh->zg_blkptr[g]);
2358 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2361 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, uint64_t size,
2366 ASSERT(txg > spa_syncing_txg(spa));
2369 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2370 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2371 * when allocating them.
2374 error = metaslab_alloc(spa, spa_log_class(spa), size,
2375 new_bp, 1, txg, NULL,
2376 METASLAB_FASTWRITE | METASLAB_GANG_AVOID);
2380 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2381 new_bp, 1, txg, NULL,
2382 METASLAB_FASTWRITE | METASLAB_GANG_AVOID);
2386 BP_SET_LSIZE(new_bp, size);
2387 BP_SET_PSIZE(new_bp, size);
2388 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2389 BP_SET_CHECKSUM(new_bp,
2390 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2391 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2392 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2393 BP_SET_LEVEL(new_bp, 0);
2394 BP_SET_DEDUP(new_bp, 0);
2395 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2402 * Free an intent log block.
2405 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2407 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2408 ASSERT(!BP_IS_GANG(bp));
2410 zio_free(spa, txg, bp);
2414 * ==========================================================================
2415 * Read and write to physical devices
2416 * ==========================================================================
2419 zio_vdev_io_start(zio_t *zio)
2421 vdev_t *vd = zio->io_vd;
2423 spa_t *spa = zio->io_spa;
2425 ASSERT(zio->io_error == 0);
2426 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2429 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2430 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2433 * The mirror_ops handle multiple DVAs in a single BP.
2435 return (vdev_mirror_ops.vdev_op_io_start(zio));
2439 * We keep track of time-sensitive I/Os so that the scan thread
2440 * can quickly react to certain workloads. In particular, we care
2441 * about non-scrubbing, top-level reads and writes with the following
2443 * - synchronous writes of user data to non-slog devices
2444 * - any reads of user data
2445 * When these conditions are met, adjust the timestamp of spa_last_io
2446 * which allows the scan thread to adjust its workload accordingly.
2448 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2449 vd == vd->vdev_top && !vd->vdev_islog &&
2450 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2451 zio->io_txg != spa_syncing_txg(spa)) {
2452 uint64_t old = spa->spa_last_io;
2453 uint64_t new = ddi_get_lbolt64();
2455 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2458 align = 1ULL << vd->vdev_top->vdev_ashift;
2460 if (P2PHASE(zio->io_size, align) != 0) {
2461 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2462 char *abuf = zio_buf_alloc(asize);
2463 ASSERT(vd == vd->vdev_top);
2464 if (zio->io_type == ZIO_TYPE_WRITE) {
2465 bcopy(zio->io_data, abuf, zio->io_size);
2466 bzero(abuf + zio->io_size, asize - zio->io_size);
2468 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2471 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2472 ASSERT(P2PHASE(zio->io_size, align) == 0);
2473 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2476 * If this is a repair I/O, and there's no self-healing involved --
2477 * that is, we're just resilvering what we expect to resilver --
2478 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2479 * This prevents spurious resilvering with nested replication.
2480 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2481 * A is out of date, we'll read from C+D, then use the data to
2482 * resilver A+B -- but we don't actually want to resilver B, just A.
2483 * The top-level mirror has no way to know this, so instead we just
2484 * discard unnecessary repairs as we work our way down the vdev tree.
2485 * The same logic applies to any form of nested replication:
2486 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2488 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2489 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2490 zio->io_txg != 0 && /* not a delegated i/o */
2491 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2492 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2493 zio_vdev_io_bypass(zio);
2494 return (ZIO_PIPELINE_CONTINUE);
2497 if (vd->vdev_ops->vdev_op_leaf &&
2498 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2500 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2501 return (ZIO_PIPELINE_CONTINUE);
2503 if ((zio = vdev_queue_io(zio)) == NULL)
2504 return (ZIO_PIPELINE_STOP);
2506 if (!vdev_accessible(vd, zio)) {
2507 zio->io_error = ENXIO;
2509 return (ZIO_PIPELINE_STOP);
2513 return (vd->vdev_ops->vdev_op_io_start(zio));
2517 zio_vdev_io_done(zio_t *zio)
2519 vdev_t *vd = zio->io_vd;
2520 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2521 boolean_t unexpected_error = B_FALSE;
2523 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2524 return (ZIO_PIPELINE_STOP);
2526 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2528 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2530 vdev_queue_io_done(zio);
2532 if (zio->io_type == ZIO_TYPE_WRITE)
2533 vdev_cache_write(zio);
2535 if (zio_injection_enabled && zio->io_error == 0)
2536 zio->io_error = zio_handle_device_injection(vd,
2539 if (zio_injection_enabled && zio->io_error == 0)
2540 zio->io_error = zio_handle_label_injection(zio, EIO);
2542 if (zio->io_error) {
2543 if (!vdev_accessible(vd, zio)) {
2544 zio->io_error = ENXIO;
2546 unexpected_error = B_TRUE;
2551 ops->vdev_op_io_done(zio);
2553 if (unexpected_error)
2554 VERIFY(vdev_probe(vd, zio) == NULL);
2556 return (ZIO_PIPELINE_CONTINUE);
2560 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2561 * disk, and use that to finish the checksum ereport later.
2564 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2565 const void *good_buf)
2567 /* no processing needed */
2568 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2573 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2575 void *buf = zio_buf_alloc(zio->io_size);
2577 bcopy(zio->io_data, buf, zio->io_size);
2579 zcr->zcr_cbinfo = zio->io_size;
2580 zcr->zcr_cbdata = buf;
2581 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2582 zcr->zcr_free = zio_buf_free;
2586 zio_vdev_io_assess(zio_t *zio)
2588 vdev_t *vd = zio->io_vd;
2590 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2591 return (ZIO_PIPELINE_STOP);
2593 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2594 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2596 if (zio->io_vsd != NULL) {
2597 zio->io_vsd_ops->vsd_free(zio);
2601 if (zio_injection_enabled && zio->io_error == 0)
2602 zio->io_error = zio_handle_fault_injection(zio, EIO);
2605 * If the I/O failed, determine whether we should attempt to retry it.
2607 * On retry, we cut in line in the issue queue, since we don't want
2608 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2610 if (zio->io_error && vd == NULL &&
2611 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2612 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2613 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2615 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2616 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2617 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2618 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2619 zio_requeue_io_start_cut_in_line);
2620 return (ZIO_PIPELINE_STOP);
2624 * If we got an error on a leaf device, convert it to ENXIO
2625 * if the device is not accessible at all.
2627 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2628 !vdev_accessible(vd, zio))
2629 zio->io_error = ENXIO;
2632 * If we can't write to an interior vdev (mirror or RAID-Z),
2633 * set vdev_cant_write so that we stop trying to allocate from it.
2635 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2636 vd != NULL && !vd->vdev_ops->vdev_op_leaf)
2637 vd->vdev_cant_write = B_TRUE;
2640 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2642 return (ZIO_PIPELINE_CONTINUE);
2646 zio_vdev_io_reissue(zio_t *zio)
2648 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2649 ASSERT(zio->io_error == 0);
2651 zio->io_stage >>= 1;
2655 zio_vdev_io_redone(zio_t *zio)
2657 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2659 zio->io_stage >>= 1;
2663 zio_vdev_io_bypass(zio_t *zio)
2665 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2666 ASSERT(zio->io_error == 0);
2668 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2669 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2673 * ==========================================================================
2674 * Generate and verify checksums
2675 * ==========================================================================
2678 zio_checksum_generate(zio_t *zio)
2680 blkptr_t *bp = zio->io_bp;
2681 enum zio_checksum checksum;
2685 * This is zio_write_phys().
2686 * We're either generating a label checksum, or none at all.
2688 checksum = zio->io_prop.zp_checksum;
2690 if (checksum == ZIO_CHECKSUM_OFF)
2691 return (ZIO_PIPELINE_CONTINUE);
2693 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2695 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2696 ASSERT(!IO_IS_ALLOCATING(zio));
2697 checksum = ZIO_CHECKSUM_GANG_HEADER;
2699 checksum = BP_GET_CHECKSUM(bp);
2703 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2705 return (ZIO_PIPELINE_CONTINUE);
2709 zio_checksum_verify(zio_t *zio)
2711 zio_bad_cksum_t info;
2712 blkptr_t *bp = zio->io_bp;
2715 ASSERT(zio->io_vd != NULL);
2719 * This is zio_read_phys().
2720 * We're either verifying a label checksum, or nothing at all.
2722 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2723 return (ZIO_PIPELINE_CONTINUE);
2725 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2728 if ((error = zio_checksum_error(zio, &info)) != 0) {
2729 zio->io_error = error;
2730 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2731 zfs_ereport_start_checksum(zio->io_spa,
2732 zio->io_vd, zio, zio->io_offset,
2733 zio->io_size, NULL, &info);
2737 return (ZIO_PIPELINE_CONTINUE);
2741 * Called by RAID-Z to ensure we don't compute the checksum twice.
2744 zio_checksum_verified(zio_t *zio)
2746 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2750 * ==========================================================================
2751 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2752 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2753 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2754 * indicate errors that are specific to one I/O, and most likely permanent.
2755 * Any other error is presumed to be worse because we weren't expecting it.
2756 * ==========================================================================
2759 zio_worst_error(int e1, int e2)
2761 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2764 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2765 if (e1 == zio_error_rank[r1])
2768 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2769 if (e2 == zio_error_rank[r2])
2772 return (r1 > r2 ? e1 : e2);
2776 * ==========================================================================
2778 * ==========================================================================
2781 zio_ready(zio_t *zio)
2783 blkptr_t *bp = zio->io_bp;
2784 zio_t *pio, *pio_next;
2786 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2787 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2788 return (ZIO_PIPELINE_STOP);
2790 if (zio->io_ready) {
2791 ASSERT(IO_IS_ALLOCATING(zio));
2792 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2793 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2798 if (bp != NULL && bp != &zio->io_bp_copy)
2799 zio->io_bp_copy = *bp;
2802 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2804 mutex_enter(&zio->io_lock);
2805 zio->io_state[ZIO_WAIT_READY] = 1;
2806 pio = zio_walk_parents(zio);
2807 mutex_exit(&zio->io_lock);
2810 * As we notify zio's parents, new parents could be added.
2811 * New parents go to the head of zio's io_parent_list, however,
2812 * so we will (correctly) not notify them. The remainder of zio's
2813 * io_parent_list, from 'pio_next' onward, cannot change because
2814 * all parents must wait for us to be done before they can be done.
2816 for (; pio != NULL; pio = pio_next) {
2817 pio_next = zio_walk_parents(zio);
2818 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2821 if (zio->io_flags & ZIO_FLAG_NODATA) {
2822 if (BP_IS_GANG(bp)) {
2823 zio->io_flags &= ~ZIO_FLAG_NODATA;
2825 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2826 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2830 if (zio_injection_enabled &&
2831 zio->io_spa->spa_syncing_txg == zio->io_txg)
2832 zio_handle_ignored_writes(zio);
2834 return (ZIO_PIPELINE_CONTINUE);
2838 zio_done(zio_t *zio)
2840 zio_t *pio, *pio_next;
2844 * If our children haven't all completed,
2845 * wait for them and then repeat this pipeline stage.
2847 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2848 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2849 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2850 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2851 return (ZIO_PIPELINE_STOP);
2853 for (c = 0; c < ZIO_CHILD_TYPES; c++)
2854 for (w = 0; w < ZIO_WAIT_TYPES; w++)
2855 ASSERT(zio->io_children[c][w] == 0);
2857 if (zio->io_bp != NULL) {
2858 ASSERT(zio->io_bp->blk_pad[0] == 0);
2859 ASSERT(zio->io_bp->blk_pad[1] == 0);
2860 ASSERT(bcmp(zio->io_bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2861 (zio->io_bp == zio_unique_parent(zio)->io_bp));
2862 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) &&
2863 zio->io_bp_override == NULL &&
2864 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2865 ASSERT(!BP_SHOULD_BYTESWAP(zio->io_bp));
2866 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2867 ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 ||
2868 (BP_COUNT_GANG(zio->io_bp) == BP_GET_NDVAS(zio->io_bp)));
2873 * If there were child vdev/gang/ddt errors, they apply to us now.
2875 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2876 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2877 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2880 * If the I/O on the transformed data was successful, generate any
2881 * checksum reports now while we still have the transformed data.
2883 if (zio->io_error == 0) {
2884 while (zio->io_cksum_report != NULL) {
2885 zio_cksum_report_t *zcr = zio->io_cksum_report;
2886 uint64_t align = zcr->zcr_align;
2887 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2888 char *abuf = zio->io_data;
2890 if (asize != zio->io_size) {
2891 abuf = zio_buf_alloc(asize);
2892 bcopy(zio->io_data, abuf, zio->io_size);
2893 bzero(abuf + zio->io_size, asize - zio->io_size);
2896 zio->io_cksum_report = zcr->zcr_next;
2897 zcr->zcr_next = NULL;
2898 zcr->zcr_finish(zcr, abuf);
2899 zfs_ereport_free_checksum(zcr);
2901 if (asize != zio->io_size)
2902 zio_buf_free(abuf, asize);
2906 zio_pop_transforms(zio); /* note: may set zio->io_error */
2908 vdev_stat_update(zio, zio->io_size);
2911 * If this I/O is attached to a particular vdev is slow, exceeding
2912 * 30 seconds to complete, post an error described the I/O delay.
2913 * We ignore these errors if the device is currently unavailable.
2915 if (zio->io_delay >= MSEC_TO_TICK(zio_delay_max)) {
2916 if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd))
2917 zfs_ereport_post(FM_EREPORT_ZFS_DELAY, zio->io_spa,
2918 zio->io_vd, zio, 0, 0);
2921 if (zio->io_error) {
2923 * If this I/O is attached to a particular vdev,
2924 * generate an error message describing the I/O failure
2925 * at the block level. We ignore these errors if the
2926 * device is currently unavailable.
2928 if (zio->io_error != ECKSUM && zio->io_vd != NULL &&
2929 !vdev_is_dead(zio->io_vd))
2930 zfs_ereport_post(FM_EREPORT_ZFS_IO, zio->io_spa,
2931 zio->io_vd, zio, 0, 0);
2933 if ((zio->io_error == EIO || !(zio->io_flags &
2934 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2935 zio == zio->io_logical) {
2937 * For logical I/O requests, tell the SPA to log the
2938 * error and generate a logical data ereport.
2940 spa_log_error(zio->io_spa, zio);
2941 zfs_ereport_post(FM_EREPORT_ZFS_DATA, zio->io_spa, NULL, zio,
2946 if (zio->io_error && zio == zio->io_logical) {
2948 * Determine whether zio should be reexecuted. This will
2949 * propagate all the way to the root via zio_notify_parent().
2951 ASSERT(zio->io_vd == NULL && zio->io_bp != NULL);
2952 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2954 if (IO_IS_ALLOCATING(zio) &&
2955 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2956 if (zio->io_error != ENOSPC)
2957 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2959 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2962 if ((zio->io_type == ZIO_TYPE_READ ||
2963 zio->io_type == ZIO_TYPE_FREE) &&
2964 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
2965 zio->io_error == ENXIO &&
2966 spa_load_state(zio->io_spa) == SPA_LOAD_NONE &&
2967 spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE)
2968 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2970 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2971 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2974 * Here is a possibly good place to attempt to do
2975 * either combinatorial reconstruction or error correction
2976 * based on checksums. It also might be a good place
2977 * to send out preliminary ereports before we suspend
2983 * If there were logical child errors, they apply to us now.
2984 * We defer this until now to avoid conflating logical child
2985 * errors with errors that happened to the zio itself when
2986 * updating vdev stats and reporting FMA events above.
2988 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2990 if ((zio->io_error || zio->io_reexecute) &&
2991 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
2992 !(zio->io_flags & ZIO_FLAG_IO_REWRITE))
2993 zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp);
2995 zio_gang_tree_free(&zio->io_gang_tree);
2998 * Godfather I/Os should never suspend.
3000 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3001 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3002 zio->io_reexecute = 0;
3004 if (zio->io_reexecute) {
3006 * This is a logical I/O that wants to reexecute.
3008 * Reexecute is top-down. When an i/o fails, if it's not
3009 * the root, it simply notifies its parent and sticks around.
3010 * The parent, seeing that it still has children in zio_done(),
3011 * does the same. This percolates all the way up to the root.
3012 * The root i/o will reexecute or suspend the entire tree.
3014 * This approach ensures that zio_reexecute() honors
3015 * all the original i/o dependency relationships, e.g.
3016 * parents not executing until children are ready.
3018 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3020 zio->io_gang_leader = NULL;
3022 mutex_enter(&zio->io_lock);
3023 zio->io_state[ZIO_WAIT_DONE] = 1;
3024 mutex_exit(&zio->io_lock);
3027 * "The Godfather" I/O monitors its children but is
3028 * not a true parent to them. It will track them through
3029 * the pipeline but severs its ties whenever they get into
3030 * trouble (e.g. suspended). This allows "The Godfather"
3031 * I/O to return status without blocking.
3033 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3034 zio_link_t *zl = zio->io_walk_link;
3035 pio_next = zio_walk_parents(zio);
3037 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3038 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3039 zio_remove_child(pio, zio, zl);
3040 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3044 if ((pio = zio_unique_parent(zio)) != NULL) {
3046 * We're not a root i/o, so there's nothing to do
3047 * but notify our parent. Don't propagate errors
3048 * upward since we haven't permanently failed yet.
3050 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3051 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3052 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3053 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3055 * We'd fail again if we reexecuted now, so suspend
3056 * until conditions improve (e.g. device comes online).
3058 zio_suspend(zio->io_spa, zio);
3061 * Reexecution is potentially a huge amount of work.
3062 * Hand it off to the otherwise-unused claim taskq.
3064 ASSERT(taskq_empty_ent(&zio->io_tqent));
3065 (void) taskq_dispatch_ent(
3066 zio->io_spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
3067 (task_func_t *)zio_reexecute, zio, 0,
3070 return (ZIO_PIPELINE_STOP);
3073 ASSERT(zio->io_child_count == 0);
3074 ASSERT(zio->io_reexecute == 0);
3075 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3078 * Report any checksum errors, since the I/O is complete.
3080 while (zio->io_cksum_report != NULL) {
3081 zio_cksum_report_t *zcr = zio->io_cksum_report;
3082 zio->io_cksum_report = zcr->zcr_next;
3083 zcr->zcr_next = NULL;
3084 zcr->zcr_finish(zcr, NULL);
3085 zfs_ereport_free_checksum(zcr);
3088 if (zio->io_flags & ZIO_FLAG_FASTWRITE && zio->io_bp &&
3089 !BP_IS_HOLE(zio->io_bp)) {
3090 metaslab_fastwrite_unmark(zio->io_spa, zio->io_bp);
3094 * It is the responsibility of the done callback to ensure that this
3095 * particular zio is no longer discoverable for adoption, and as
3096 * such, cannot acquire any new parents.
3101 mutex_enter(&zio->io_lock);
3102 zio->io_state[ZIO_WAIT_DONE] = 1;
3103 mutex_exit(&zio->io_lock);
3105 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3106 zio_link_t *zl = zio->io_walk_link;
3107 pio_next = zio_walk_parents(zio);
3108 zio_remove_child(pio, zio, zl);
3109 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3112 if (zio->io_waiter != NULL) {
3113 mutex_enter(&zio->io_lock);
3114 zio->io_executor = NULL;
3115 cv_broadcast(&zio->io_cv);
3116 mutex_exit(&zio->io_lock);
3121 return (ZIO_PIPELINE_STOP);
3125 * ==========================================================================
3126 * I/O pipeline definition
3127 * ==========================================================================
3129 static zio_pipe_stage_t *zio_pipeline[] = {
3135 zio_checksum_generate,
3149 zio_checksum_verify,
3153 /* dnp is the dnode for zb1->zb_object */
3155 zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_t *zb1,
3156 const zbookmark_t *zb2)
3158 uint64_t zb1nextL0, zb2thisobj;
3160 ASSERT(zb1->zb_objset == zb2->zb_objset);
3161 ASSERT(zb2->zb_level == 0);
3164 * A bookmark in the deadlist is considered to be after
3167 if (zb2->zb_object == DMU_DEADLIST_OBJECT)
3170 /* The objset_phys_t isn't before anything. */
3174 zb1nextL0 = (zb1->zb_blkid + 1) <<
3175 ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
3177 zb2thisobj = zb2->zb_object ? zb2->zb_object :
3178 zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
3180 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3181 uint64_t nextobj = zb1nextL0 *
3182 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
3183 return (nextobj <= zb2thisobj);
3186 if (zb1->zb_object < zb2thisobj)
3188 if (zb1->zb_object > zb2thisobj)
3190 if (zb2->zb_object == DMU_META_DNODE_OBJECT)
3192 return (zb1nextL0 <= zb2->zb_blkid);
3195 #if defined(_KERNEL) && defined(HAVE_SPL)
3196 /* Fault injection */
3197 EXPORT_SYMBOL(zio_injection_enabled);
3198 EXPORT_SYMBOL(zio_inject_fault);
3199 EXPORT_SYMBOL(zio_inject_list_next);
3200 EXPORT_SYMBOL(zio_clear_fault);
3201 EXPORT_SYMBOL(zio_handle_fault_injection);
3202 EXPORT_SYMBOL(zio_handle_device_injection);
3203 EXPORT_SYMBOL(zio_handle_label_injection);
3204 EXPORT_SYMBOL(zio_priority_table);
3205 EXPORT_SYMBOL(zio_type_name);
3207 module_param(zio_bulk_flags, int, 0644);
3208 MODULE_PARM_DESC(zio_bulk_flags, "Additional flags to pass to bulk buffers");
3210 module_param(zio_delay_max, int, 0644);
3211 MODULE_PARM_DESC(zio_delay_max, "Max zio millisec delay before posting event");
3213 module_param(zio_requeue_io_start_cut_in_line, int, 0644);
3214 MODULE_PARM_DESC(zio_requeue_io_start_cut_in_line, "Prioritize requeued I/O");