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_orig_data = zio->io_data = data;
613 zio->io_orig_size = zio->io_size = size;
614 zio->io_orig_flags = zio->io_flags = flags;
615 zio->io_orig_stage = zio->io_stage = stage;
616 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
617 bzero(&zio->io_prop, sizeof (zio_prop_t));
619 zio->io_reexecute = 0;
620 zio->io_bp_override = NULL;
621 zio->io_walk_link = NULL;
622 zio->io_transform_stack = NULL;
625 zio->io_child_count = 0;
626 zio->io_parent_count = 0;
627 zio->io_stall = NULL;
628 zio->io_gang_leader = NULL;
629 zio->io_gang_tree = NULL;
630 zio->io_executor = NULL;
631 zio->io_waiter = NULL;
632 zio->io_cksum_report = NULL;
634 bzero(zio->io_child_error, sizeof (int) * ZIO_CHILD_TYPES);
635 bzero(zio->io_children,
636 sizeof (uint64_t) * ZIO_CHILD_TYPES * ZIO_WAIT_TYPES);
637 bzero(&zio->io_bookmark, sizeof (zbookmark_t));
639 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
640 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
643 zio->io_bookmark = *zb;
646 if (zio->io_logical == NULL)
647 zio->io_logical = pio->io_logical;
648 if (zio->io_child_type == ZIO_CHILD_GANG)
649 zio->io_gang_leader = pio->io_gang_leader;
650 zio_add_child(pio, zio);
653 taskq_init_ent(&zio->io_tqent);
659 zio_destroy(zio_t *zio)
661 kmem_cache_free(zio_cache, zio);
665 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
666 void *private, enum zio_flag flags)
670 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
671 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
672 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
678 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
680 return (zio_null(NULL, spa, NULL, done, private, flags));
684 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
685 void *data, uint64_t size, zio_done_func_t *done, void *private,
686 int priority, enum zio_flag flags, const zbookmark_t *zb)
690 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
691 data, size, done, private,
692 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
693 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
694 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
700 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
701 void *data, uint64_t size, const zio_prop_t *zp,
702 zio_done_func_t *ready, zio_done_func_t *done, void *private,
703 int priority, enum zio_flag flags, const zbookmark_t *zb)
707 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
708 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
709 zp->zp_compress >= ZIO_COMPRESS_OFF &&
710 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
711 DMU_OT_IS_VALID(zp->zp_type) &&
714 zp->zp_copies <= spa_max_replication(spa) &&
716 zp->zp_dedup_verify <= 1);
718 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
719 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
720 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
721 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
723 zio->io_ready = ready;
730 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
731 uint64_t size, zio_done_func_t *done, void *private, int priority,
732 enum zio_flag flags, zbookmark_t *zb)
736 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
737 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
738 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
744 zio_write_override(zio_t *zio, blkptr_t *bp, int copies)
746 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
747 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
748 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
749 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
751 zio->io_prop.zp_copies = copies;
752 zio->io_bp_override = bp;
756 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
758 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
762 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
767 dprintf_bp(bp, "freeing in txg %llu, pass %u",
768 (longlong_t)txg, spa->spa_sync_pass);
770 ASSERT(!BP_IS_HOLE(bp));
771 ASSERT(spa_syncing_txg(spa) == txg);
772 ASSERT(spa_sync_pass(spa) <= SYNC_PASS_DEFERRED_FREE);
774 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
775 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
776 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
782 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
783 zio_done_func_t *done, void *private, enum zio_flag flags)
788 * A claim is an allocation of a specific block. Claims are needed
789 * to support immediate writes in the intent log. The issue is that
790 * immediate writes contain committed data, but in a txg that was
791 * *not* committed. Upon opening the pool after an unclean shutdown,
792 * the intent log claims all blocks that contain immediate write data
793 * so that the SPA knows they're in use.
795 * All claims *must* be resolved in the first txg -- before the SPA
796 * starts allocating blocks -- so that nothing is allocated twice.
797 * If txg == 0 we just verify that the block is claimable.
799 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
800 ASSERT(txg == spa_first_txg(spa) || txg == 0);
801 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
803 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
804 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
805 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
811 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
812 zio_done_func_t *done, void *private, int priority, enum zio_flag flags)
817 if (vd->vdev_children == 0) {
818 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
819 ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
820 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
824 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
826 for (c = 0; c < vd->vdev_children; c++)
827 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
828 done, private, priority, flags));
835 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
836 void *data, int checksum, zio_done_func_t *done, void *private,
837 int priority, enum zio_flag flags, boolean_t labels)
841 ASSERT(vd->vdev_children == 0);
842 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
843 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
844 ASSERT3U(offset + size, <=, vd->vdev_psize);
846 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
847 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
848 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
850 zio->io_prop.zp_checksum = checksum;
856 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
857 void *data, int checksum, zio_done_func_t *done, void *private,
858 int priority, enum zio_flag flags, boolean_t labels)
862 ASSERT(vd->vdev_children == 0);
863 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
864 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
865 ASSERT3U(offset + size, <=, vd->vdev_psize);
867 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
868 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
869 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
871 zio->io_prop.zp_checksum = checksum;
873 if (zio_checksum_table[checksum].ci_eck) {
875 * zec checksums are necessarily destructive -- they modify
876 * the end of the write buffer to hold the verifier/checksum.
877 * Therefore, we must make a local copy in case the data is
878 * being written to multiple places in parallel.
880 void *wbuf = zio_buf_alloc(size);
881 bcopy(data, wbuf, size);
882 zio_push_transform(zio, wbuf, size, size, NULL);
889 * Create a child I/O to do some work for us.
892 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
893 void *data, uint64_t size, int type, int priority, enum zio_flag flags,
894 zio_done_func_t *done, void *private)
896 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
899 ASSERT(vd->vdev_parent ==
900 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
902 if (type == ZIO_TYPE_READ && bp != NULL) {
904 * If we have the bp, then the child should perform the
905 * checksum and the parent need not. This pushes error
906 * detection as close to the leaves as possible and
907 * eliminates redundant checksums in the interior nodes.
909 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
910 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
913 if (vd->vdev_children == 0)
914 offset += VDEV_LABEL_START_SIZE;
916 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
919 * If we've decided to do a repair, the write is not speculative --
920 * even if the original read was.
922 if (flags & ZIO_FLAG_IO_REPAIR)
923 flags &= ~ZIO_FLAG_SPECULATIVE;
925 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
926 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
927 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
933 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
934 int type, int priority, enum zio_flag flags,
935 zio_done_func_t *done, void *private)
939 ASSERT(vd->vdev_ops->vdev_op_leaf);
941 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
942 data, size, done, private, type, priority,
943 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
945 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
951 zio_flush(zio_t *zio, vdev_t *vd)
953 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
954 NULL, NULL, ZIO_PRIORITY_NOW,
955 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
959 zio_shrink(zio_t *zio, uint64_t size)
961 ASSERT(zio->io_executor == NULL);
962 ASSERT(zio->io_orig_size == zio->io_size);
963 ASSERT(size <= zio->io_size);
966 * We don't shrink for raidz because of problems with the
967 * reconstruction when reading back less than the block size.
968 * Note, BP_IS_RAIDZ() assumes no compression.
970 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
971 if (!BP_IS_RAIDZ(zio->io_bp))
972 zio->io_orig_size = zio->io_size = size;
976 * ==========================================================================
977 * Prepare to read and write logical blocks
978 * ==========================================================================
982 zio_read_bp_init(zio_t *zio)
984 blkptr_t *bp = zio->io_bp;
986 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
987 zio->io_child_type == ZIO_CHILD_LOGICAL &&
988 !(zio->io_flags & ZIO_FLAG_RAW)) {
989 uint64_t psize = BP_GET_PSIZE(bp);
990 void *cbuf = zio_buf_alloc(psize);
992 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
995 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
996 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
998 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
999 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1001 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1002 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1004 return (ZIO_PIPELINE_CONTINUE);
1008 zio_write_bp_init(zio_t *zio)
1010 spa_t *spa = zio->io_spa;
1011 zio_prop_t *zp = &zio->io_prop;
1012 enum zio_compress compress = zp->zp_compress;
1013 blkptr_t *bp = zio->io_bp;
1014 uint64_t lsize = zio->io_size;
1015 uint64_t psize = lsize;
1019 * If our children haven't all reached the ready stage,
1020 * wait for them and then repeat this pipeline stage.
1022 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1023 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1024 return (ZIO_PIPELINE_STOP);
1026 if (!IO_IS_ALLOCATING(zio))
1027 return (ZIO_PIPELINE_CONTINUE);
1029 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1031 if (zio->io_bp_override) {
1032 ASSERT(bp->blk_birth != zio->io_txg);
1033 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1035 *bp = *zio->io_bp_override;
1036 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1038 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1039 return (ZIO_PIPELINE_CONTINUE);
1041 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1042 zp->zp_dedup_verify);
1044 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1045 BP_SET_DEDUP(bp, 1);
1046 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1047 return (ZIO_PIPELINE_CONTINUE);
1049 zio->io_bp_override = NULL;
1053 if (bp->blk_birth == zio->io_txg) {
1055 * We're rewriting an existing block, which means we're
1056 * working on behalf of spa_sync(). For spa_sync() to
1057 * converge, it must eventually be the case that we don't
1058 * have to allocate new blocks. But compression changes
1059 * the blocksize, which forces a reallocate, and makes
1060 * convergence take longer. Therefore, after the first
1061 * few passes, stop compressing to ensure convergence.
1063 pass = spa_sync_pass(spa);
1065 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1066 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1067 ASSERT(!BP_GET_DEDUP(bp));
1069 if (pass > SYNC_PASS_DONT_COMPRESS)
1070 compress = ZIO_COMPRESS_OFF;
1072 /* Make sure someone doesn't change their mind on overwrites */
1073 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
1074 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1077 if (compress != ZIO_COMPRESS_OFF) {
1078 void *cbuf = zio_buf_alloc(lsize);
1079 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1080 if (psize == 0 || psize == lsize) {
1081 compress = ZIO_COMPRESS_OFF;
1082 zio_buf_free(cbuf, lsize);
1084 ASSERT(psize < lsize);
1085 zio_push_transform(zio, cbuf, psize, lsize, NULL);
1090 * The final pass of spa_sync() must be all rewrites, but the first
1091 * few passes offer a trade-off: allocating blocks defers convergence,
1092 * but newly allocated blocks are sequential, so they can be written
1093 * to disk faster. Therefore, we allow the first few passes of
1094 * spa_sync() to allocate new blocks, but force rewrites after that.
1095 * There should only be a handful of blocks after pass 1 in any case.
1097 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
1098 pass > SYNC_PASS_REWRITE) {
1099 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1101 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1102 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1105 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1109 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1111 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1112 BP_SET_LSIZE(bp, lsize);
1113 BP_SET_PSIZE(bp, psize);
1114 BP_SET_COMPRESS(bp, compress);
1115 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1116 BP_SET_TYPE(bp, zp->zp_type);
1117 BP_SET_LEVEL(bp, zp->zp_level);
1118 BP_SET_DEDUP(bp, zp->zp_dedup);
1119 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1121 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1122 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1123 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1127 return (ZIO_PIPELINE_CONTINUE);
1131 zio_free_bp_init(zio_t *zio)
1133 blkptr_t *bp = zio->io_bp;
1135 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1136 if (BP_GET_DEDUP(bp))
1137 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1140 return (ZIO_PIPELINE_CONTINUE);
1144 * ==========================================================================
1145 * Execute the I/O pipeline
1146 * ==========================================================================
1150 zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q, boolean_t cutinline)
1152 spa_t *spa = zio->io_spa;
1153 zio_type_t t = zio->io_type;
1154 int flags = (cutinline ? TQ_FRONT : 0);
1157 * If we're a config writer or a probe, the normal issue and
1158 * interrupt threads may all be blocked waiting for the config lock.
1159 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1161 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1165 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1167 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1171 * If this is a high priority I/O, then use the high priority taskq.
1173 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1174 spa->spa_zio_taskq[t][q + 1] != NULL)
1177 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1180 * NB: We are assuming that the zio can only be dispatched
1181 * to a single taskq at a time. It would be a grievous error
1182 * to dispatch the zio to another taskq at the same time.
1184 ASSERT(taskq_empty_ent(&zio->io_tqent));
1185 taskq_dispatch_ent(spa->spa_zio_taskq[t][q],
1186 (task_func_t *)zio_execute, zio, flags, &zio->io_tqent);
1190 zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
1192 kthread_t *executor = zio->io_executor;
1193 spa_t *spa = zio->io_spa;
1196 for (t = 0; t < ZIO_TYPES; t++)
1197 if (taskq_member(spa->spa_zio_taskq[t][q], executor))
1204 zio_issue_async(zio_t *zio)
1206 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1208 return (ZIO_PIPELINE_STOP);
1212 zio_interrupt(zio_t *zio)
1214 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1218 * Execute the I/O pipeline until one of the following occurs:
1219 * (1) the I/O completes; (2) the pipeline stalls waiting for
1220 * dependent child I/Os; (3) the I/O issues, so we're waiting
1221 * for an I/O completion interrupt; (4) the I/O is delegated by
1222 * vdev-level caching or aggregation; (5) the I/O is deferred
1223 * due to vdev-level queueing; (6) the I/O is handed off to
1224 * another thread. In all cases, the pipeline stops whenever
1225 * there's no CPU work; it never burns a thread in cv_wait().
1227 * There's no locking on io_stage because there's no legitimate way
1228 * for multiple threads to be attempting to process the same I/O.
1230 static zio_pipe_stage_t *zio_pipeline[];
1233 * zio_execute() is a wrapper around the static function
1234 * __zio_execute() so that we can force __zio_execute() to be
1235 * inlined. This reduces stack overhead which is important
1236 * because __zio_execute() is called recursively in several zio
1237 * code paths. zio_execute() itself cannot be inlined because
1238 * it is externally visible.
1241 zio_execute(zio_t *zio)
1246 __attribute__((always_inline))
1248 __zio_execute(zio_t *zio)
1250 zio->io_executor = curthread;
1252 while (zio->io_stage < ZIO_STAGE_DONE) {
1253 enum zio_stage pipeline = zio->io_pipeline;
1254 enum zio_stage stage = zio->io_stage;
1259 ASSERT(!MUTEX_HELD(&zio->io_lock));
1260 ASSERT(ISP2(stage));
1261 ASSERT(zio->io_stall == NULL);
1265 } while ((stage & pipeline) == 0);
1267 ASSERT(stage <= ZIO_STAGE_DONE);
1269 dp = spa_get_dsl(zio->io_spa);
1270 cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1271 zio_requeue_io_start_cut_in_line : B_FALSE;
1274 * If we are in interrupt context and this pipeline stage
1275 * will grab a config lock that is held across I/O,
1276 * or may wait for an I/O that needs an interrupt thread
1277 * to complete, issue async to avoid deadlock.
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 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1290 * If we executing in the context of the tx_sync_thread,
1291 * or we are performing pool initialization outside of a
1292 * zio_taskq[ZIO_TASKQ_ISSUE] context. Then issue the zio
1293 * async to minimize stack usage for these deep call paths.
1295 if ((dp && curthread == dp->dp_tx.tx_sync_thread) ||
1296 (dp && spa_is_initializing(dp->dp_spa) &&
1297 !zio_taskq_member(zio, ZIO_TASKQ_ISSUE))) {
1298 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1303 zio->io_stage = stage;
1304 rv = zio_pipeline[highbit(stage) - 1](zio);
1306 if (rv == ZIO_PIPELINE_STOP)
1309 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1315 * ==========================================================================
1316 * Initiate I/O, either sync or async
1317 * ==========================================================================
1320 zio_wait(zio_t *zio)
1324 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1325 ASSERT(zio->io_executor == NULL);
1327 zio->io_waiter = curthread;
1331 mutex_enter(&zio->io_lock);
1332 while (zio->io_executor != NULL)
1333 cv_wait_io(&zio->io_cv, &zio->io_lock);
1334 mutex_exit(&zio->io_lock);
1336 error = zio->io_error;
1343 zio_nowait(zio_t *zio)
1345 ASSERT(zio->io_executor == NULL);
1347 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1348 zio_unique_parent(zio) == NULL) {
1350 * This is a logical async I/O with no parent to wait for it.
1351 * We add it to the spa_async_root_zio "Godfather" I/O which
1352 * will ensure they complete prior to unloading the pool.
1354 spa_t *spa = zio->io_spa;
1356 zio_add_child(spa->spa_async_zio_root, zio);
1363 * ==========================================================================
1364 * Reexecute or suspend/resume failed I/O
1365 * ==========================================================================
1369 zio_reexecute(zio_t *pio)
1371 zio_t *cio, *cio_next;
1374 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1375 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1376 ASSERT(pio->io_gang_leader == NULL);
1377 ASSERT(pio->io_gang_tree == NULL);
1379 pio->io_flags = pio->io_orig_flags;
1380 pio->io_stage = pio->io_orig_stage;
1381 pio->io_pipeline = pio->io_orig_pipeline;
1382 pio->io_reexecute = 0;
1384 for (w = 0; w < ZIO_WAIT_TYPES; w++)
1385 pio->io_state[w] = 0;
1386 for (c = 0; c < ZIO_CHILD_TYPES; c++)
1387 pio->io_child_error[c] = 0;
1389 if (IO_IS_ALLOCATING(pio))
1390 BP_ZERO(pio->io_bp);
1393 * As we reexecute pio's children, new children could be created.
1394 * New children go to the head of pio's io_child_list, however,
1395 * so we will (correctly) not reexecute them. The key is that
1396 * the remainder of pio's io_child_list, from 'cio_next' onward,
1397 * cannot be affected by any side effects of reexecuting 'cio'.
1399 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1400 cio_next = zio_walk_children(pio);
1401 mutex_enter(&pio->io_lock);
1402 for (w = 0; w < ZIO_WAIT_TYPES; w++)
1403 pio->io_children[cio->io_child_type][w]++;
1404 mutex_exit(&pio->io_lock);
1409 * Now that all children have been reexecuted, execute the parent.
1410 * We don't reexecute "The Godfather" I/O here as it's the
1411 * responsibility of the caller to wait on him.
1413 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1418 zio_suspend(spa_t *spa, zio_t *zio)
1420 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1421 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1422 "failure and the failure mode property for this pool "
1423 "is set to panic.", spa_name(spa));
1425 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1427 mutex_enter(&spa->spa_suspend_lock);
1429 if (spa->spa_suspend_zio_root == NULL)
1430 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1431 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1432 ZIO_FLAG_GODFATHER);
1434 spa->spa_suspended = B_TRUE;
1437 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1438 ASSERT(zio != spa->spa_suspend_zio_root);
1439 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1440 ASSERT(zio_unique_parent(zio) == NULL);
1441 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1442 zio_add_child(spa->spa_suspend_zio_root, zio);
1445 mutex_exit(&spa->spa_suspend_lock);
1449 zio_resume(spa_t *spa)
1454 * Reexecute all previously suspended i/o.
1456 mutex_enter(&spa->spa_suspend_lock);
1457 spa->spa_suspended = B_FALSE;
1458 cv_broadcast(&spa->spa_suspend_cv);
1459 pio = spa->spa_suspend_zio_root;
1460 spa->spa_suspend_zio_root = NULL;
1461 mutex_exit(&spa->spa_suspend_lock);
1467 return (zio_wait(pio));
1471 zio_resume_wait(spa_t *spa)
1473 mutex_enter(&spa->spa_suspend_lock);
1474 while (spa_suspended(spa))
1475 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1476 mutex_exit(&spa->spa_suspend_lock);
1480 * ==========================================================================
1483 * A gang block is a collection of small blocks that looks to the DMU
1484 * like one large block. When zio_dva_allocate() cannot find a block
1485 * of the requested size, due to either severe fragmentation or the pool
1486 * being nearly full, it calls zio_write_gang_block() to construct the
1487 * block from smaller fragments.
1489 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1490 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1491 * an indirect block: it's an array of block pointers. It consumes
1492 * only one sector and hence is allocatable regardless of fragmentation.
1493 * The gang header's bps point to its gang members, which hold the data.
1495 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1496 * as the verifier to ensure uniqueness of the SHA256 checksum.
1497 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1498 * not the gang header. This ensures that data block signatures (needed for
1499 * deduplication) are independent of how the block is physically stored.
1501 * Gang blocks can be nested: a gang member may itself be a gang block.
1502 * Thus every gang block is a tree in which root and all interior nodes are
1503 * gang headers, and the leaves are normal blocks that contain user data.
1504 * The root of the gang tree is called the gang leader.
1506 * To perform any operation (read, rewrite, free, claim) on a gang block,
1507 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1508 * in the io_gang_tree field of the original logical i/o by recursively
1509 * reading the gang leader and all gang headers below it. This yields
1510 * an in-core tree containing the contents of every gang header and the
1511 * bps for every constituent of the gang block.
1513 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1514 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1515 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1516 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1517 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1518 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1519 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1520 * of the gang header plus zio_checksum_compute() of the data to update the
1521 * gang header's blk_cksum as described above.
1523 * The two-phase assemble/issue model solves the problem of partial failure --
1524 * what if you'd freed part of a gang block but then couldn't read the
1525 * gang header for another part? Assembling the entire gang tree first
1526 * ensures that all the necessary gang header I/O has succeeded before
1527 * starting the actual work of free, claim, or write. Once the gang tree
1528 * is assembled, free and claim are in-memory operations that cannot fail.
1530 * In the event that a gang write fails, zio_dva_unallocate() walks the
1531 * gang tree to immediately free (i.e. insert back into the space map)
1532 * everything we've allocated. This ensures that we don't get ENOSPC
1533 * errors during repeated suspend/resume cycles due to a flaky device.
1535 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1536 * the gang tree, we won't modify the block, so we can safely defer the free
1537 * (knowing that the block is still intact). If we *can* assemble the gang
1538 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1539 * each constituent bp and we can allocate a new block on the next sync pass.
1541 * In all cases, the gang tree allows complete recovery from partial failure.
1542 * ==========================================================================
1546 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1551 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1552 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1553 &pio->io_bookmark));
1557 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1562 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1563 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1564 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1566 * As we rewrite each gang header, the pipeline will compute
1567 * a new gang block header checksum for it; but no one will
1568 * compute a new data checksum, so we do that here. The one
1569 * exception is the gang leader: the pipeline already computed
1570 * its data checksum because that stage precedes gang assembly.
1571 * (Presently, nothing actually uses interior data checksums;
1572 * this is just good hygiene.)
1574 if (gn != pio->io_gang_leader->io_gang_tree) {
1575 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1576 data, BP_GET_PSIZE(bp));
1579 * If we are here to damage data for testing purposes,
1580 * leave the GBH alone so that we can detect the damage.
1582 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1583 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1585 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1586 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1587 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1595 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1597 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1598 ZIO_GANG_CHILD_FLAGS(pio)));
1603 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1605 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1606 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1609 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1618 static void zio_gang_tree_assemble_done(zio_t *zio);
1620 static zio_gang_node_t *
1621 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1623 zio_gang_node_t *gn;
1625 ASSERT(*gnpp == NULL);
1627 gn = kmem_zalloc(sizeof (*gn), KM_PUSHPAGE);
1628 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1635 zio_gang_node_free(zio_gang_node_t **gnpp)
1637 zio_gang_node_t *gn = *gnpp;
1640 for (g = 0; g < SPA_GBH_NBLKPTRS; g++)
1641 ASSERT(gn->gn_child[g] == NULL);
1643 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1644 kmem_free(gn, sizeof (*gn));
1649 zio_gang_tree_free(zio_gang_node_t **gnpp)
1651 zio_gang_node_t *gn = *gnpp;
1657 for (g = 0; g < SPA_GBH_NBLKPTRS; g++)
1658 zio_gang_tree_free(&gn->gn_child[g]);
1660 zio_gang_node_free(gnpp);
1664 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1666 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1668 ASSERT(gio->io_gang_leader == gio);
1669 ASSERT(BP_IS_GANG(bp));
1671 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1672 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1673 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1677 zio_gang_tree_assemble_done(zio_t *zio)
1679 zio_t *gio = zio->io_gang_leader;
1680 zio_gang_node_t *gn = zio->io_private;
1681 blkptr_t *bp = zio->io_bp;
1684 ASSERT(gio == zio_unique_parent(zio));
1685 ASSERT(zio->io_child_count == 0);
1690 if (BP_SHOULD_BYTESWAP(bp))
1691 byteswap_uint64_array(zio->io_data, zio->io_size);
1693 ASSERT(zio->io_data == gn->gn_gbh);
1694 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1695 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1697 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1698 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1699 if (!BP_IS_GANG(gbp))
1701 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1706 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1708 zio_t *gio = pio->io_gang_leader;
1712 ASSERT(BP_IS_GANG(bp) == !!gn);
1713 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1714 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1717 * If you're a gang header, your data is in gn->gn_gbh.
1718 * If you're a gang member, your data is in 'data' and gn == NULL.
1720 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1723 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1725 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1726 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1727 if (BP_IS_HOLE(gbp))
1729 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1730 data = (char *)data + BP_GET_PSIZE(gbp);
1734 if (gn == gio->io_gang_tree)
1735 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1742 zio_gang_assemble(zio_t *zio)
1744 blkptr_t *bp = zio->io_bp;
1746 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1747 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1749 zio->io_gang_leader = zio;
1751 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1753 return (ZIO_PIPELINE_CONTINUE);
1757 zio_gang_issue(zio_t *zio)
1759 blkptr_t *bp = zio->io_bp;
1761 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1762 return (ZIO_PIPELINE_STOP);
1764 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1765 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1767 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1768 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1770 zio_gang_tree_free(&zio->io_gang_tree);
1772 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1774 return (ZIO_PIPELINE_CONTINUE);
1778 zio_write_gang_member_ready(zio_t *zio)
1780 zio_t *pio = zio_unique_parent(zio);
1781 ASSERTV(zio_t *gio = zio->io_gang_leader;)
1782 dva_t *cdva = zio->io_bp->blk_dva;
1783 dva_t *pdva = pio->io_bp->blk_dva;
1787 if (BP_IS_HOLE(zio->io_bp))
1790 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1792 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1793 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1794 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1795 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1796 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1798 mutex_enter(&pio->io_lock);
1799 for (d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1800 ASSERT(DVA_GET_GANG(&pdva[d]));
1801 asize = DVA_GET_ASIZE(&pdva[d]);
1802 asize += DVA_GET_ASIZE(&cdva[d]);
1803 DVA_SET_ASIZE(&pdva[d], asize);
1805 mutex_exit(&pio->io_lock);
1809 zio_write_gang_block(zio_t *pio)
1811 spa_t *spa = pio->io_spa;
1812 blkptr_t *bp = pio->io_bp;
1813 zio_t *gio = pio->io_gang_leader;
1815 zio_gang_node_t *gn, **gnpp;
1816 zio_gbh_phys_t *gbh;
1817 uint64_t txg = pio->io_txg;
1818 uint64_t resid = pio->io_size;
1820 int copies = gio->io_prop.zp_copies;
1821 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1825 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1826 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1827 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1829 pio->io_error = error;
1830 return (ZIO_PIPELINE_CONTINUE);
1834 gnpp = &gio->io_gang_tree;
1836 gnpp = pio->io_private;
1837 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1840 gn = zio_gang_node_alloc(gnpp);
1842 bzero(gbh, SPA_GANGBLOCKSIZE);
1845 * Create the gang header.
1847 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1848 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1851 * Create and nowait the gang children.
1853 for (g = 0; resid != 0; resid -= lsize, g++) {
1854 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1856 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1858 zp.zp_checksum = gio->io_prop.zp_checksum;
1859 zp.zp_compress = ZIO_COMPRESS_OFF;
1860 zp.zp_type = DMU_OT_NONE;
1862 zp.zp_copies = gio->io_prop.zp_copies;
1864 zp.zp_dedup_verify = 0;
1866 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1867 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1868 zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1869 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1870 &pio->io_bookmark));
1874 * Set pio's pipeline to just wait for zio to finish.
1876 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1879 * We didn't allocate this bp, so make sure it doesn't get unmarked.
1881 pio->io_flags &= ~ZIO_FLAG_FASTWRITE;
1885 return (ZIO_PIPELINE_CONTINUE);
1889 * ==========================================================================
1891 * ==========================================================================
1894 zio_ddt_child_read_done(zio_t *zio)
1896 blkptr_t *bp = zio->io_bp;
1897 ddt_entry_t *dde = zio->io_private;
1899 zio_t *pio = zio_unique_parent(zio);
1901 mutex_enter(&pio->io_lock);
1902 ddp = ddt_phys_select(dde, bp);
1903 if (zio->io_error == 0)
1904 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1905 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1906 dde->dde_repair_data = zio->io_data;
1908 zio_buf_free(zio->io_data, zio->io_size);
1909 mutex_exit(&pio->io_lock);
1913 zio_ddt_read_start(zio_t *zio)
1915 blkptr_t *bp = zio->io_bp;
1918 ASSERT(BP_GET_DEDUP(bp));
1919 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1920 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1922 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1923 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1924 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1925 ddt_phys_t *ddp = dde->dde_phys;
1926 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1929 ASSERT(zio->io_vsd == NULL);
1932 if (ddp_self == NULL)
1933 return (ZIO_PIPELINE_CONTINUE);
1935 for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1936 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1938 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1940 zio_nowait(zio_read(zio, zio->io_spa, &blk,
1941 zio_buf_alloc(zio->io_size), zio->io_size,
1942 zio_ddt_child_read_done, dde, zio->io_priority,
1943 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1944 &zio->io_bookmark));
1946 return (ZIO_PIPELINE_CONTINUE);
1949 zio_nowait(zio_read(zio, zio->io_spa, bp,
1950 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1951 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1953 return (ZIO_PIPELINE_CONTINUE);
1957 zio_ddt_read_done(zio_t *zio)
1959 blkptr_t *bp = zio->io_bp;
1961 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1962 return (ZIO_PIPELINE_STOP);
1964 ASSERT(BP_GET_DEDUP(bp));
1965 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1966 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1968 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1969 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1970 ddt_entry_t *dde = zio->io_vsd;
1972 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
1973 return (ZIO_PIPELINE_CONTINUE);
1976 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
1977 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1978 return (ZIO_PIPELINE_STOP);
1980 if (dde->dde_repair_data != NULL) {
1981 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
1982 zio->io_child_error[ZIO_CHILD_DDT] = 0;
1984 ddt_repair_done(ddt, dde);
1988 ASSERT(zio->io_vsd == NULL);
1990 return (ZIO_PIPELINE_CONTINUE);
1994 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
1996 spa_t *spa = zio->io_spa;
2000 * Note: we compare the original data, not the transformed data,
2001 * because when zio->io_bp is an override bp, we will not have
2002 * pushed the I/O transforms. That's an important optimization
2003 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2005 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2006 zio_t *lio = dde->dde_lead_zio[p];
2009 return (lio->io_orig_size != zio->io_orig_size ||
2010 bcmp(zio->io_orig_data, lio->io_orig_data,
2011 zio->io_orig_size) != 0);
2015 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2016 ddt_phys_t *ddp = &dde->dde_phys[p];
2018 if (ddp->ddp_phys_birth != 0) {
2019 arc_buf_t *abuf = NULL;
2020 uint32_t aflags = ARC_WAIT;
2021 blkptr_t blk = *zio->io_bp;
2024 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2028 error = arc_read_nolock(NULL, spa, &blk,
2029 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2030 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2031 &aflags, &zio->io_bookmark);
2034 if (arc_buf_size(abuf) != zio->io_orig_size ||
2035 bcmp(abuf->b_data, zio->io_orig_data,
2036 zio->io_orig_size) != 0)
2038 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
2042 return (error != 0);
2050 zio_ddt_child_write_ready(zio_t *zio)
2052 int p = zio->io_prop.zp_copies;
2053 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2054 ddt_entry_t *dde = zio->io_private;
2055 ddt_phys_t *ddp = &dde->dde_phys[p];
2063 ASSERT(dde->dde_lead_zio[p] == zio);
2065 ddt_phys_fill(ddp, zio->io_bp);
2067 while ((pio = zio_walk_parents(zio)) != NULL)
2068 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2074 zio_ddt_child_write_done(zio_t *zio)
2076 int p = zio->io_prop.zp_copies;
2077 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2078 ddt_entry_t *dde = zio->io_private;
2079 ddt_phys_t *ddp = &dde->dde_phys[p];
2083 ASSERT(ddp->ddp_refcnt == 0);
2084 ASSERT(dde->dde_lead_zio[p] == zio);
2085 dde->dde_lead_zio[p] = NULL;
2087 if (zio->io_error == 0) {
2088 while (zio_walk_parents(zio) != NULL)
2089 ddt_phys_addref(ddp);
2091 ddt_phys_clear(ddp);
2098 zio_ddt_ditto_write_done(zio_t *zio)
2100 int p = DDT_PHYS_DITTO;
2101 blkptr_t *bp = zio->io_bp;
2102 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2103 ddt_entry_t *dde = zio->io_private;
2104 ddt_phys_t *ddp = &dde->dde_phys[p];
2105 ddt_key_t *ddk = &dde->dde_key;
2106 ASSERTV(zio_prop_t *zp = &zio->io_prop);
2110 ASSERT(ddp->ddp_refcnt == 0);
2111 ASSERT(dde->dde_lead_zio[p] == zio);
2112 dde->dde_lead_zio[p] = NULL;
2114 if (zio->io_error == 0) {
2115 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2116 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2117 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2118 if (ddp->ddp_phys_birth != 0)
2119 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2120 ddt_phys_fill(ddp, bp);
2127 zio_ddt_write(zio_t *zio)
2129 spa_t *spa = zio->io_spa;
2130 blkptr_t *bp = zio->io_bp;
2131 uint64_t txg = zio->io_txg;
2132 zio_prop_t *zp = &zio->io_prop;
2133 int p = zp->zp_copies;
2137 ddt_t *ddt = ddt_select(spa, bp);
2141 ASSERT(BP_GET_DEDUP(bp));
2142 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2143 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2146 dde = ddt_lookup(ddt, bp, B_TRUE);
2147 ddp = &dde->dde_phys[p];
2149 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2151 * If we're using a weak checksum, upgrade to a strong checksum
2152 * and try again. If we're already using a strong checksum,
2153 * we can't resolve it, so just convert to an ordinary write.
2154 * (And automatically e-mail a paper to Nature?)
2156 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2157 zp->zp_checksum = spa_dedup_checksum(spa);
2158 zio_pop_transforms(zio);
2159 zio->io_stage = ZIO_STAGE_OPEN;
2164 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2166 return (ZIO_PIPELINE_CONTINUE);
2169 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2170 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2172 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2173 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2174 zio_prop_t czp = *zp;
2176 czp.zp_copies = ditto_copies;
2179 * If we arrived here with an override bp, we won't have run
2180 * the transform stack, so we won't have the data we need to
2181 * generate a child i/o. So, toss the override bp and restart.
2182 * This is safe, because using the override bp is just an
2183 * optimization; and it's rare, so the cost doesn't matter.
2185 if (zio->io_bp_override) {
2186 zio_pop_transforms(zio);
2187 zio->io_stage = ZIO_STAGE_OPEN;
2188 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2189 zio->io_bp_override = NULL;
2192 return (ZIO_PIPELINE_CONTINUE);
2195 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2196 zio->io_orig_size, &czp, NULL,
2197 zio_ddt_ditto_write_done, dde, zio->io_priority,
2198 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2200 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2201 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2204 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2205 if (ddp->ddp_phys_birth != 0)
2206 ddt_bp_fill(ddp, bp, txg);
2207 if (dde->dde_lead_zio[p] != NULL)
2208 zio_add_child(zio, dde->dde_lead_zio[p]);
2210 ddt_phys_addref(ddp);
2211 } else if (zio->io_bp_override) {
2212 ASSERT(bp->blk_birth == txg);
2213 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2214 ddt_phys_fill(ddp, bp);
2215 ddt_phys_addref(ddp);
2217 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2218 zio->io_orig_size, zp, zio_ddt_child_write_ready,
2219 zio_ddt_child_write_done, dde, zio->io_priority,
2220 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2222 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2223 dde->dde_lead_zio[p] = cio;
2233 return (ZIO_PIPELINE_CONTINUE);
2236 ddt_entry_t *freedde; /* for debugging */
2239 zio_ddt_free(zio_t *zio)
2241 spa_t *spa = zio->io_spa;
2242 blkptr_t *bp = zio->io_bp;
2243 ddt_t *ddt = ddt_select(spa, bp);
2247 ASSERT(BP_GET_DEDUP(bp));
2248 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2251 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2252 ddp = ddt_phys_select(dde, bp);
2253 ddt_phys_decref(ddp);
2256 return (ZIO_PIPELINE_CONTINUE);
2260 * ==========================================================================
2261 * Allocate and free blocks
2262 * ==========================================================================
2265 zio_dva_allocate(zio_t *zio)
2267 spa_t *spa = zio->io_spa;
2268 metaslab_class_t *mc = spa_normal_class(spa);
2269 blkptr_t *bp = zio->io_bp;
2273 if (zio->io_gang_leader == NULL) {
2274 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2275 zio->io_gang_leader = zio;
2278 ASSERT(BP_IS_HOLE(bp));
2279 ASSERT3U(BP_GET_NDVAS(bp), ==, 0);
2280 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2281 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2282 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2285 * The dump device does not support gang blocks so allocation on
2286 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2287 * the "fast" gang feature.
2289 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2290 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2291 METASLAB_GANG_CHILD : 0;
2292 flags |= (zio->io_flags & ZIO_FLAG_FASTWRITE) ? METASLAB_FASTWRITE : 0;
2293 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2294 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2297 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2298 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2300 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2301 return (zio_write_gang_block(zio));
2302 zio->io_error = error;
2305 return (ZIO_PIPELINE_CONTINUE);
2309 zio_dva_free(zio_t *zio)
2311 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2313 return (ZIO_PIPELINE_CONTINUE);
2317 zio_dva_claim(zio_t *zio)
2321 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2323 zio->io_error = error;
2325 return (ZIO_PIPELINE_CONTINUE);
2329 * Undo an allocation. This is used by zio_done() when an I/O fails
2330 * and we want to give back the block we just allocated.
2331 * This handles both normal blocks and gang blocks.
2334 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2338 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2339 ASSERT(zio->io_bp_override == NULL);
2341 if (!BP_IS_HOLE(bp))
2342 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2345 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2346 zio_dva_unallocate(zio, gn->gn_child[g],
2347 &gn->gn_gbh->zg_blkptr[g]);
2353 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2356 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, uint64_t size,
2361 ASSERT(txg > spa_syncing_txg(spa));
2364 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2365 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2366 * when allocating them.
2369 error = metaslab_alloc(spa, spa_log_class(spa), size,
2370 new_bp, 1, txg, NULL,
2371 METASLAB_FASTWRITE | METASLAB_GANG_AVOID);
2375 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2376 new_bp, 1, txg, NULL,
2377 METASLAB_FASTWRITE | METASLAB_GANG_AVOID);
2381 BP_SET_LSIZE(new_bp, size);
2382 BP_SET_PSIZE(new_bp, size);
2383 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2384 BP_SET_CHECKSUM(new_bp,
2385 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2386 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2387 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2388 BP_SET_LEVEL(new_bp, 0);
2389 BP_SET_DEDUP(new_bp, 0);
2390 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2397 * Free an intent log block.
2400 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2402 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2403 ASSERT(!BP_IS_GANG(bp));
2405 zio_free(spa, txg, bp);
2409 * ==========================================================================
2410 * Read and write to physical devices
2411 * ==========================================================================
2414 zio_vdev_io_start(zio_t *zio)
2416 vdev_t *vd = zio->io_vd;
2418 spa_t *spa = zio->io_spa;
2420 ASSERT(zio->io_error == 0);
2421 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2424 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2425 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2428 * The mirror_ops handle multiple DVAs in a single BP.
2430 return (vdev_mirror_ops.vdev_op_io_start(zio));
2434 * We keep track of time-sensitive I/Os so that the scan thread
2435 * can quickly react to certain workloads. In particular, we care
2436 * about non-scrubbing, top-level reads and writes with the following
2438 * - synchronous writes of user data to non-slog devices
2439 * - any reads of user data
2440 * When these conditions are met, adjust the timestamp of spa_last_io
2441 * which allows the scan thread to adjust its workload accordingly.
2443 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2444 vd == vd->vdev_top && !vd->vdev_islog &&
2445 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2446 zio->io_txg != spa_syncing_txg(spa)) {
2447 uint64_t old = spa->spa_last_io;
2448 uint64_t new = ddi_get_lbolt64();
2450 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2453 align = 1ULL << vd->vdev_top->vdev_ashift;
2455 if (P2PHASE(zio->io_size, align) != 0) {
2456 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2457 char *abuf = zio_buf_alloc(asize);
2458 ASSERT(vd == vd->vdev_top);
2459 if (zio->io_type == ZIO_TYPE_WRITE) {
2460 bcopy(zio->io_data, abuf, zio->io_size);
2461 bzero(abuf + zio->io_size, asize - zio->io_size);
2463 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2466 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2467 ASSERT(P2PHASE(zio->io_size, align) == 0);
2468 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2471 * If this is a repair I/O, and there's no self-healing involved --
2472 * that is, we're just resilvering what we expect to resilver --
2473 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2474 * This prevents spurious resilvering with nested replication.
2475 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2476 * A is out of date, we'll read from C+D, then use the data to
2477 * resilver A+B -- but we don't actually want to resilver B, just A.
2478 * The top-level mirror has no way to know this, so instead we just
2479 * discard unnecessary repairs as we work our way down the vdev tree.
2480 * The same logic applies to any form of nested replication:
2481 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2483 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2484 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2485 zio->io_txg != 0 && /* not a delegated i/o */
2486 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2487 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2488 zio_vdev_io_bypass(zio);
2489 return (ZIO_PIPELINE_CONTINUE);
2492 if (vd->vdev_ops->vdev_op_leaf &&
2493 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2495 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2496 return (ZIO_PIPELINE_CONTINUE);
2498 if ((zio = vdev_queue_io(zio)) == NULL)
2499 return (ZIO_PIPELINE_STOP);
2501 if (!vdev_accessible(vd, zio)) {
2502 zio->io_error = ENXIO;
2504 return (ZIO_PIPELINE_STOP);
2508 return (vd->vdev_ops->vdev_op_io_start(zio));
2512 zio_vdev_io_done(zio_t *zio)
2514 vdev_t *vd = zio->io_vd;
2515 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2516 boolean_t unexpected_error = B_FALSE;
2518 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2519 return (ZIO_PIPELINE_STOP);
2521 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2523 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2525 vdev_queue_io_done(zio);
2527 if (zio->io_type == ZIO_TYPE_WRITE)
2528 vdev_cache_write(zio);
2530 if (zio_injection_enabled && zio->io_error == 0)
2531 zio->io_error = zio_handle_device_injection(vd,
2534 if (zio_injection_enabled && zio->io_error == 0)
2535 zio->io_error = zio_handle_label_injection(zio, EIO);
2537 if (zio->io_error) {
2538 if (!vdev_accessible(vd, zio)) {
2539 zio->io_error = ENXIO;
2541 unexpected_error = B_TRUE;
2546 ops->vdev_op_io_done(zio);
2548 if (unexpected_error)
2549 VERIFY(vdev_probe(vd, zio) == NULL);
2551 return (ZIO_PIPELINE_CONTINUE);
2555 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2556 * disk, and use that to finish the checksum ereport later.
2559 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2560 const void *good_buf)
2562 /* no processing needed */
2563 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2568 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2570 void *buf = zio_buf_alloc(zio->io_size);
2572 bcopy(zio->io_data, buf, zio->io_size);
2574 zcr->zcr_cbinfo = zio->io_size;
2575 zcr->zcr_cbdata = buf;
2576 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2577 zcr->zcr_free = zio_buf_free;
2581 zio_vdev_io_assess(zio_t *zio)
2583 vdev_t *vd = zio->io_vd;
2585 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2586 return (ZIO_PIPELINE_STOP);
2588 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2589 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2591 if (zio->io_vsd != NULL) {
2592 zio->io_vsd_ops->vsd_free(zio);
2596 if (zio_injection_enabled && zio->io_error == 0)
2597 zio->io_error = zio_handle_fault_injection(zio, EIO);
2600 * If the I/O failed, determine whether we should attempt to retry it.
2602 * On retry, we cut in line in the issue queue, since we don't want
2603 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2605 if (zio->io_error && vd == NULL &&
2606 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2607 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2608 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2610 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2611 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2612 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2613 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2614 zio_requeue_io_start_cut_in_line);
2615 return (ZIO_PIPELINE_STOP);
2619 * If we got an error on a leaf device, convert it to ENXIO
2620 * if the device is not accessible at all.
2622 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2623 !vdev_accessible(vd, zio))
2624 zio->io_error = ENXIO;
2627 * If we can't write to an interior vdev (mirror or RAID-Z),
2628 * set vdev_cant_write so that we stop trying to allocate from it.
2630 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2631 vd != NULL && !vd->vdev_ops->vdev_op_leaf)
2632 vd->vdev_cant_write = B_TRUE;
2635 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2637 return (ZIO_PIPELINE_CONTINUE);
2641 zio_vdev_io_reissue(zio_t *zio)
2643 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2644 ASSERT(zio->io_error == 0);
2646 zio->io_stage >>= 1;
2650 zio_vdev_io_redone(zio_t *zio)
2652 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2654 zio->io_stage >>= 1;
2658 zio_vdev_io_bypass(zio_t *zio)
2660 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2661 ASSERT(zio->io_error == 0);
2663 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2664 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2668 * ==========================================================================
2669 * Generate and verify checksums
2670 * ==========================================================================
2673 zio_checksum_generate(zio_t *zio)
2675 blkptr_t *bp = zio->io_bp;
2676 enum zio_checksum checksum;
2680 * This is zio_write_phys().
2681 * We're either generating a label checksum, or none at all.
2683 checksum = zio->io_prop.zp_checksum;
2685 if (checksum == ZIO_CHECKSUM_OFF)
2686 return (ZIO_PIPELINE_CONTINUE);
2688 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2690 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2691 ASSERT(!IO_IS_ALLOCATING(zio));
2692 checksum = ZIO_CHECKSUM_GANG_HEADER;
2694 checksum = BP_GET_CHECKSUM(bp);
2698 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2700 return (ZIO_PIPELINE_CONTINUE);
2704 zio_checksum_verify(zio_t *zio)
2706 zio_bad_cksum_t info;
2707 blkptr_t *bp = zio->io_bp;
2710 ASSERT(zio->io_vd != NULL);
2714 * This is zio_read_phys().
2715 * We're either verifying a label checksum, or nothing at all.
2717 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2718 return (ZIO_PIPELINE_CONTINUE);
2720 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2723 if ((error = zio_checksum_error(zio, &info)) != 0) {
2724 zio->io_error = error;
2725 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2726 zfs_ereport_start_checksum(zio->io_spa,
2727 zio->io_vd, zio, zio->io_offset,
2728 zio->io_size, NULL, &info);
2732 return (ZIO_PIPELINE_CONTINUE);
2736 * Called by RAID-Z to ensure we don't compute the checksum twice.
2739 zio_checksum_verified(zio_t *zio)
2741 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2745 * ==========================================================================
2746 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2747 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2748 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2749 * indicate errors that are specific to one I/O, and most likely permanent.
2750 * Any other error is presumed to be worse because we weren't expecting it.
2751 * ==========================================================================
2754 zio_worst_error(int e1, int e2)
2756 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2759 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2760 if (e1 == zio_error_rank[r1])
2763 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2764 if (e2 == zio_error_rank[r2])
2767 return (r1 > r2 ? e1 : e2);
2771 * ==========================================================================
2773 * ==========================================================================
2776 zio_ready(zio_t *zio)
2778 blkptr_t *bp = zio->io_bp;
2779 zio_t *pio, *pio_next;
2781 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2782 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2783 return (ZIO_PIPELINE_STOP);
2785 if (zio->io_ready) {
2786 ASSERT(IO_IS_ALLOCATING(zio));
2787 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2788 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2793 if (bp != NULL && bp != &zio->io_bp_copy)
2794 zio->io_bp_copy = *bp;
2797 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2799 mutex_enter(&zio->io_lock);
2800 zio->io_state[ZIO_WAIT_READY] = 1;
2801 pio = zio_walk_parents(zio);
2802 mutex_exit(&zio->io_lock);
2805 * As we notify zio's parents, new parents could be added.
2806 * New parents go to the head of zio's io_parent_list, however,
2807 * so we will (correctly) not notify them. The remainder of zio's
2808 * io_parent_list, from 'pio_next' onward, cannot change because
2809 * all parents must wait for us to be done before they can be done.
2811 for (; pio != NULL; pio = pio_next) {
2812 pio_next = zio_walk_parents(zio);
2813 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2816 if (zio->io_flags & ZIO_FLAG_NODATA) {
2817 if (BP_IS_GANG(bp)) {
2818 zio->io_flags &= ~ZIO_FLAG_NODATA;
2820 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2821 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2825 if (zio_injection_enabled &&
2826 zio->io_spa->spa_syncing_txg == zio->io_txg)
2827 zio_handle_ignored_writes(zio);
2829 return (ZIO_PIPELINE_CONTINUE);
2833 zio_done(zio_t *zio)
2835 zio_t *pio, *pio_next;
2839 * If our children haven't all completed,
2840 * wait for them and then repeat this pipeline stage.
2842 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2843 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2844 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2845 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2846 return (ZIO_PIPELINE_STOP);
2848 for (c = 0; c < ZIO_CHILD_TYPES; c++)
2849 for (w = 0; w < ZIO_WAIT_TYPES; w++)
2850 ASSERT(zio->io_children[c][w] == 0);
2852 if (zio->io_bp != NULL) {
2853 ASSERT(zio->io_bp->blk_pad[0] == 0);
2854 ASSERT(zio->io_bp->blk_pad[1] == 0);
2855 ASSERT(bcmp(zio->io_bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2856 (zio->io_bp == zio_unique_parent(zio)->io_bp));
2857 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) &&
2858 zio->io_bp_override == NULL &&
2859 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2860 ASSERT(!BP_SHOULD_BYTESWAP(zio->io_bp));
2861 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2862 ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 ||
2863 (BP_COUNT_GANG(zio->io_bp) == BP_GET_NDVAS(zio->io_bp)));
2868 * If there were child vdev/gang/ddt errors, they apply to us now.
2870 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2871 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2872 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2875 * If the I/O on the transformed data was successful, generate any
2876 * checksum reports now while we still have the transformed data.
2878 if (zio->io_error == 0) {
2879 while (zio->io_cksum_report != NULL) {
2880 zio_cksum_report_t *zcr = zio->io_cksum_report;
2881 uint64_t align = zcr->zcr_align;
2882 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2883 char *abuf = zio->io_data;
2885 if (asize != zio->io_size) {
2886 abuf = zio_buf_alloc(asize);
2887 bcopy(zio->io_data, abuf, zio->io_size);
2888 bzero(abuf + zio->io_size, asize - zio->io_size);
2891 zio->io_cksum_report = zcr->zcr_next;
2892 zcr->zcr_next = NULL;
2893 zcr->zcr_finish(zcr, abuf);
2894 zfs_ereport_free_checksum(zcr);
2896 if (asize != zio->io_size)
2897 zio_buf_free(abuf, asize);
2901 zio_pop_transforms(zio); /* note: may set zio->io_error */
2903 vdev_stat_update(zio, zio->io_size);
2906 * If this I/O is attached to a particular vdev is slow, exeeding
2907 * 30 seconds to complete, post an error described the I/O delay.
2908 * We ignore these errors if the device is currently unavailable.
2910 if (zio->io_delay >= zio_delay_max) {
2911 if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd))
2912 zfs_ereport_post(FM_EREPORT_ZFS_DELAY, zio->io_spa,
2913 zio->io_vd, zio, 0, 0);
2916 if (zio->io_error) {
2918 * If this I/O is attached to a particular vdev,
2919 * generate an error message describing the I/O failure
2920 * at the block level. We ignore these errors if the
2921 * device is currently unavailable.
2923 if (zio->io_error != ECKSUM && zio->io_vd != NULL &&
2924 !vdev_is_dead(zio->io_vd))
2925 zfs_ereport_post(FM_EREPORT_ZFS_IO, zio->io_spa,
2926 zio->io_vd, zio, 0, 0);
2928 if ((zio->io_error == EIO || !(zio->io_flags &
2929 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2930 zio == zio->io_logical) {
2932 * For logical I/O requests, tell the SPA to log the
2933 * error and generate a logical data ereport.
2935 spa_log_error(zio->io_spa, zio);
2936 zfs_ereport_post(FM_EREPORT_ZFS_DATA, zio->io_spa, NULL, zio,
2941 if (zio->io_error && zio == zio->io_logical) {
2943 * Determine whether zio should be reexecuted. This will
2944 * propagate all the way to the root via zio_notify_parent().
2946 ASSERT(zio->io_vd == NULL && zio->io_bp != NULL);
2947 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2949 if (IO_IS_ALLOCATING(zio) &&
2950 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2951 if (zio->io_error != ENOSPC)
2952 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2954 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2957 if ((zio->io_type == ZIO_TYPE_READ ||
2958 zio->io_type == ZIO_TYPE_FREE) &&
2959 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
2960 zio->io_error == ENXIO &&
2961 spa_load_state(zio->io_spa) == SPA_LOAD_NONE &&
2962 spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE)
2963 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2965 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2966 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2969 * Here is a possibly good place to attempt to do
2970 * either combinatorial reconstruction or error correction
2971 * based on checksums. It also might be a good place
2972 * to send out preliminary ereports before we suspend
2978 * If there were logical child errors, they apply to us now.
2979 * We defer this until now to avoid conflating logical child
2980 * errors with errors that happened to the zio itself when
2981 * updating vdev stats and reporting FMA events above.
2983 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2985 if ((zio->io_error || zio->io_reexecute) &&
2986 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
2987 !(zio->io_flags & ZIO_FLAG_IO_REWRITE))
2988 zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp);
2990 zio_gang_tree_free(&zio->io_gang_tree);
2993 * Godfather I/Os should never suspend.
2995 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
2996 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
2997 zio->io_reexecute = 0;
2999 if (zio->io_reexecute) {
3001 * This is a logical I/O that wants to reexecute.
3003 * Reexecute is top-down. When an i/o fails, if it's not
3004 * the root, it simply notifies its parent and sticks around.
3005 * The parent, seeing that it still has children in zio_done(),
3006 * does the same. This percolates all the way up to the root.
3007 * The root i/o will reexecute or suspend the entire tree.
3009 * This approach ensures that zio_reexecute() honors
3010 * all the original i/o dependency relationships, e.g.
3011 * parents not executing until children are ready.
3013 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3015 zio->io_gang_leader = NULL;
3017 mutex_enter(&zio->io_lock);
3018 zio->io_state[ZIO_WAIT_DONE] = 1;
3019 mutex_exit(&zio->io_lock);
3022 * "The Godfather" I/O monitors its children but is
3023 * not a true parent to them. It will track them through
3024 * the pipeline but severs its ties whenever they get into
3025 * trouble (e.g. suspended). This allows "The Godfather"
3026 * I/O to return status without blocking.
3028 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3029 zio_link_t *zl = zio->io_walk_link;
3030 pio_next = zio_walk_parents(zio);
3032 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3033 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3034 zio_remove_child(pio, zio, zl);
3035 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3039 if ((pio = zio_unique_parent(zio)) != NULL) {
3041 * We're not a root i/o, so there's nothing to do
3042 * but notify our parent. Don't propagate errors
3043 * upward since we haven't permanently failed yet.
3045 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3046 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3047 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3048 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3050 * We'd fail again if we reexecuted now, so suspend
3051 * until conditions improve (e.g. device comes online).
3053 zio_suspend(zio->io_spa, zio);
3056 * Reexecution is potentially a huge amount of work.
3057 * Hand it off to the otherwise-unused claim taskq.
3059 ASSERT(taskq_empty_ent(&zio->io_tqent));
3060 (void) taskq_dispatch_ent(
3061 zio->io_spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
3062 (task_func_t *)zio_reexecute, zio, 0,
3065 return (ZIO_PIPELINE_STOP);
3068 ASSERT(zio->io_child_count == 0);
3069 ASSERT(zio->io_reexecute == 0);
3070 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3073 * Report any checksum errors, since the I/O is complete.
3075 while (zio->io_cksum_report != NULL) {
3076 zio_cksum_report_t *zcr = zio->io_cksum_report;
3077 zio->io_cksum_report = zcr->zcr_next;
3078 zcr->zcr_next = NULL;
3079 zcr->zcr_finish(zcr, NULL);
3080 zfs_ereport_free_checksum(zcr);
3083 if (zio->io_flags & ZIO_FLAG_FASTWRITE && zio->io_bp &&
3084 !BP_IS_HOLE(zio->io_bp)) {
3085 metaslab_fastwrite_unmark(zio->io_spa, zio->io_bp);
3089 * It is the responsibility of the done callback to ensure that this
3090 * particular zio is no longer discoverable for adoption, and as
3091 * such, cannot acquire any new parents.
3096 mutex_enter(&zio->io_lock);
3097 zio->io_state[ZIO_WAIT_DONE] = 1;
3098 mutex_exit(&zio->io_lock);
3100 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3101 zio_link_t *zl = zio->io_walk_link;
3102 pio_next = zio_walk_parents(zio);
3103 zio_remove_child(pio, zio, zl);
3104 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3107 if (zio->io_waiter != NULL) {
3108 mutex_enter(&zio->io_lock);
3109 zio->io_executor = NULL;
3110 cv_broadcast(&zio->io_cv);
3111 mutex_exit(&zio->io_lock);
3116 return (ZIO_PIPELINE_STOP);
3120 * ==========================================================================
3121 * I/O pipeline definition
3122 * ==========================================================================
3124 static zio_pipe_stage_t *zio_pipeline[] = {
3130 zio_checksum_generate,
3144 zio_checksum_verify,
3148 /* dnp is the dnode for zb1->zb_object */
3150 zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_t *zb1,
3151 const zbookmark_t *zb2)
3153 uint64_t zb1nextL0, zb2thisobj;
3155 ASSERT(zb1->zb_objset == zb2->zb_objset);
3156 ASSERT(zb2->zb_level == 0);
3159 * A bookmark in the deadlist is considered to be after
3162 if (zb2->zb_object == DMU_DEADLIST_OBJECT)
3165 /* The objset_phys_t isn't before anything. */
3169 zb1nextL0 = (zb1->zb_blkid + 1) <<
3170 ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
3172 zb2thisobj = zb2->zb_object ? zb2->zb_object :
3173 zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
3175 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3176 uint64_t nextobj = zb1nextL0 *
3177 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
3178 return (nextobj <= zb2thisobj);
3181 if (zb1->zb_object < zb2thisobj)
3183 if (zb1->zb_object > zb2thisobj)
3185 if (zb2->zb_object == DMU_META_DNODE_OBJECT)
3187 return (zb1nextL0 <= zb2->zb_blkid);
3190 #if defined(_KERNEL) && defined(HAVE_SPL)
3191 /* Fault injection */
3192 EXPORT_SYMBOL(zio_injection_enabled);
3193 EXPORT_SYMBOL(zio_inject_fault);
3194 EXPORT_SYMBOL(zio_inject_list_next);
3195 EXPORT_SYMBOL(zio_clear_fault);
3196 EXPORT_SYMBOL(zio_handle_fault_injection);
3197 EXPORT_SYMBOL(zio_handle_device_injection);
3198 EXPORT_SYMBOL(zio_handle_label_injection);
3199 EXPORT_SYMBOL(zio_priority_table);
3200 EXPORT_SYMBOL(zio_type_name);
3202 module_param(zio_bulk_flags, int, 0644);
3203 MODULE_PARM_DESC(zio_bulk_flags, "Additional flags to pass to bulk buffers");
3205 module_param(zio_delay_max, int, 0644);
3206 MODULE_PARM_DESC(zio_delay_max, "Max zio millisec delay before posting event");
3208 module_param(zio_requeue_io_start_cut_in_line, int, 0644);
3209 MODULE_PARM_DESC(zio_requeue_io_start_cut_in_line, "Prioritize requeued I/O");