4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2012 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
27 #include <sys/zfs_context.h>
28 #include <sys/fm/fs/zfs.h>
31 #include <sys/spa_impl.h>
32 #include <sys/vdev_impl.h>
33 #include <sys/zio_impl.h>
34 #include <sys/zio_compress.h>
35 #include <sys/zio_checksum.h>
36 #include <sys/dmu_objset.h>
41 * ==========================================================================
43 * ==========================================================================
45 uint8_t zio_priority_table[ZIO_PRIORITY_TABLE_SIZE] = {
46 0, /* ZIO_PRIORITY_NOW */
47 0, /* ZIO_PRIORITY_SYNC_READ */
48 0, /* ZIO_PRIORITY_SYNC_WRITE */
49 0, /* ZIO_PRIORITY_LOG_WRITE */
50 1, /* ZIO_PRIORITY_CACHE_FILL */
51 1, /* ZIO_PRIORITY_AGG */
52 4, /* ZIO_PRIORITY_FREE */
53 4, /* ZIO_PRIORITY_ASYNC_WRITE */
54 6, /* ZIO_PRIORITY_ASYNC_READ */
55 10, /* ZIO_PRIORITY_RESILVER */
56 20, /* ZIO_PRIORITY_SCRUB */
57 2, /* ZIO_PRIORITY_DDT_PREFETCH */
61 * ==========================================================================
62 * I/O type descriptions
63 * ==========================================================================
65 char *zio_type_name[ZIO_TYPES] = {
66 "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl"
70 * ==========================================================================
72 * ==========================================================================
74 kmem_cache_t *zio_cache;
75 kmem_cache_t *zio_link_cache;
76 kmem_cache_t *zio_vdev_cache;
77 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
78 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
79 int zio_bulk_flags = 0;
80 int zio_delay_max = ZIO_DELAY_MAX;
83 extern vmem_t *zio_alloc_arena;
85 extern int zfs_mg_alloc_failures;
88 * An allocating zio is one that either currently has the DVA allocate
89 * stage set or will have it later in its lifetime.
91 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
93 int zio_requeue_io_start_cut_in_line = 1;
96 int zio_buf_debug_limit = 16384;
98 int zio_buf_debug_limit = 0;
101 static inline void __zio_execute(zio_t *zio);
104 zio_cons(void *arg, void *unused, int kmflag)
108 bzero(zio, sizeof (zio_t));
110 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
111 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
113 list_create(&zio->io_parent_list, sizeof (zio_link_t),
114 offsetof(zio_link_t, zl_parent_node));
115 list_create(&zio->io_child_list, sizeof (zio_link_t),
116 offsetof(zio_link_t, zl_child_node));
122 zio_dest(void *arg, void *unused)
126 mutex_destroy(&zio->io_lock);
127 cv_destroy(&zio->io_cv);
128 list_destroy(&zio->io_parent_list);
129 list_destroy(&zio->io_child_list);
136 vmem_t *data_alloc_arena = NULL;
139 data_alloc_arena = zio_alloc_arena;
141 zio_cache = kmem_cache_create("zio_cache", sizeof (zio_t), 0,
142 zio_cons, zio_dest, NULL, NULL, NULL, KMC_KMEM);
143 zio_link_cache = kmem_cache_create("zio_link_cache",
144 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, KMC_KMEM);
145 zio_vdev_cache = kmem_cache_create("zio_vdev_cache", sizeof(vdev_io_t),
146 PAGESIZE, NULL, NULL, NULL, NULL, NULL, KMC_VMEM);
149 * For small buffers, we want a cache for each multiple of
150 * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache
151 * for each quarter-power of 2. For large buffers, we want
152 * a cache for each multiple of PAGESIZE.
154 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
155 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
159 while (p2 & (p2 - 1))
162 if (size <= 4 * SPA_MINBLOCKSIZE) {
163 align = SPA_MINBLOCKSIZE;
164 } else if (P2PHASE(size, PAGESIZE) == 0) {
166 } else if (P2PHASE(size, p2 >> 2) == 0) {
172 int flags = zio_bulk_flags;
175 * The smallest buffers (512b) are heavily used and
176 * experience a lot of churn. The slabs allocated
177 * for them are also relatively small (32K). Thus
178 * in over to avoid expensive calls to vmalloc() we
179 * make an exception to the usual slab allocation
180 * policy and force these buffers to be kmem backed.
182 if (size == (1 << SPA_MINBLOCKSHIFT))
185 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
186 zio_buf_cache[c] = kmem_cache_create(name, size,
187 align, NULL, NULL, NULL, NULL, NULL, flags);
189 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
190 zio_data_buf_cache[c] = kmem_cache_create(name, size,
191 align, NULL, NULL, NULL, NULL,
192 data_alloc_arena, flags);
197 ASSERT(zio_buf_cache[c] != NULL);
198 if (zio_buf_cache[c - 1] == NULL)
199 zio_buf_cache[c - 1] = zio_buf_cache[c];
201 ASSERT(zio_data_buf_cache[c] != NULL);
202 if (zio_data_buf_cache[c - 1] == NULL)
203 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
207 * The zio write taskqs have 1 thread per cpu, allow 1/2 of the taskqs
208 * to fail 3 times per txg or 8 failures, whichever is greater.
210 zfs_mg_alloc_failures = MAX((3 * max_ncpus / 2), 8);
219 kmem_cache_t *last_cache = NULL;
220 kmem_cache_t *last_data_cache = NULL;
222 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
223 if (zio_buf_cache[c] != last_cache) {
224 last_cache = zio_buf_cache[c];
225 kmem_cache_destroy(zio_buf_cache[c]);
227 zio_buf_cache[c] = NULL;
229 if (zio_data_buf_cache[c] != last_data_cache) {
230 last_data_cache = zio_data_buf_cache[c];
231 kmem_cache_destroy(zio_data_buf_cache[c]);
233 zio_data_buf_cache[c] = NULL;
236 kmem_cache_destroy(zio_vdev_cache);
237 kmem_cache_destroy(zio_link_cache);
238 kmem_cache_destroy(zio_cache);
244 * ==========================================================================
245 * Allocate and free I/O buffers
246 * ==========================================================================
250 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
251 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
252 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
253 * excess / transient data in-core during a crashdump.
256 zio_buf_alloc(size_t size)
258 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
260 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
262 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
266 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
267 * crashdump if the kernel panics. This exists so that we will limit the amount
268 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
269 * of kernel heap dumped to disk when the kernel panics)
272 zio_data_buf_alloc(size_t size)
274 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
276 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
278 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
282 zio_buf_free(void *buf, size_t size)
284 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
286 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
288 kmem_cache_free(zio_buf_cache[c], buf);
292 zio_data_buf_free(void *buf, size_t size)
294 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
296 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
298 kmem_cache_free(zio_data_buf_cache[c], buf);
302 * Dedicated I/O buffers to ensure that memory fragmentation never prevents
303 * or significantly delays the issuing of a zio. These buffers are used
304 * to aggregate I/O and could be used for raidz stripes.
309 return (kmem_cache_alloc(zio_vdev_cache, KM_PUSHPAGE));
313 zio_vdev_free(void *buf)
315 kmem_cache_free(zio_vdev_cache, buf);
320 * ==========================================================================
321 * Push and pop I/O transform buffers
322 * ==========================================================================
325 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
326 zio_transform_func_t *transform)
328 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_PUSHPAGE);
330 zt->zt_orig_data = zio->io_data;
331 zt->zt_orig_size = zio->io_size;
332 zt->zt_bufsize = bufsize;
333 zt->zt_transform = transform;
335 zt->zt_next = zio->io_transform_stack;
336 zio->io_transform_stack = zt;
343 zio_pop_transforms(zio_t *zio)
347 while ((zt = zio->io_transform_stack) != NULL) {
348 if (zt->zt_transform != NULL)
349 zt->zt_transform(zio,
350 zt->zt_orig_data, zt->zt_orig_size);
352 if (zt->zt_bufsize != 0)
353 zio_buf_free(zio->io_data, zt->zt_bufsize);
355 zio->io_data = zt->zt_orig_data;
356 zio->io_size = zt->zt_orig_size;
357 zio->io_transform_stack = zt->zt_next;
359 kmem_free(zt, sizeof (zio_transform_t));
364 * ==========================================================================
365 * I/O transform callbacks for subblocks and decompression
366 * ==========================================================================
369 zio_subblock(zio_t *zio, void *data, uint64_t size)
371 ASSERT(zio->io_size > size);
373 if (zio->io_type == ZIO_TYPE_READ)
374 bcopy(zio->io_data, data, size);
378 zio_decompress(zio_t *zio, void *data, uint64_t size)
380 if (zio->io_error == 0 &&
381 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
382 zio->io_data, data, zio->io_size, size) != 0)
387 * ==========================================================================
388 * I/O parent/child relationships and pipeline interlocks
389 * ==========================================================================
392 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
393 * continue calling these functions until they return NULL.
394 * Otherwise, the next caller will pick up the list walk in
395 * some indeterminate state. (Otherwise every caller would
396 * have to pass in a cookie to keep the state represented by
397 * io_walk_link, which gets annoying.)
400 zio_walk_parents(zio_t *cio)
402 zio_link_t *zl = cio->io_walk_link;
403 list_t *pl = &cio->io_parent_list;
405 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
406 cio->io_walk_link = zl;
411 ASSERT(zl->zl_child == cio);
412 return (zl->zl_parent);
416 zio_walk_children(zio_t *pio)
418 zio_link_t *zl = pio->io_walk_link;
419 list_t *cl = &pio->io_child_list;
421 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
422 pio->io_walk_link = zl;
427 ASSERT(zl->zl_parent == pio);
428 return (zl->zl_child);
432 zio_unique_parent(zio_t *cio)
434 zio_t *pio = zio_walk_parents(cio);
436 VERIFY(zio_walk_parents(cio) == NULL);
441 zio_add_child(zio_t *pio, zio_t *cio)
443 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_PUSHPAGE);
447 * Logical I/Os can have logical, gang, or vdev children.
448 * Gang I/Os can have gang or vdev children.
449 * Vdev I/Os can only have vdev children.
450 * The following ASSERT captures all of these constraints.
452 ASSERT(cio->io_child_type <= pio->io_child_type);
457 mutex_enter(&cio->io_lock);
458 mutex_enter(&pio->io_lock);
460 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
462 for (w = 0; w < ZIO_WAIT_TYPES; w++)
463 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
465 list_insert_head(&pio->io_child_list, zl);
466 list_insert_head(&cio->io_parent_list, zl);
468 pio->io_child_count++;
469 cio->io_parent_count++;
471 mutex_exit(&pio->io_lock);
472 mutex_exit(&cio->io_lock);
476 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
478 ASSERT(zl->zl_parent == pio);
479 ASSERT(zl->zl_child == cio);
481 mutex_enter(&cio->io_lock);
482 mutex_enter(&pio->io_lock);
484 list_remove(&pio->io_child_list, zl);
485 list_remove(&cio->io_parent_list, zl);
487 pio->io_child_count--;
488 cio->io_parent_count--;
490 mutex_exit(&pio->io_lock);
491 mutex_exit(&cio->io_lock);
493 kmem_cache_free(zio_link_cache, zl);
497 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
499 uint64_t *countp = &zio->io_children[child][wait];
500 boolean_t waiting = B_FALSE;
502 mutex_enter(&zio->io_lock);
503 ASSERT(zio->io_stall == NULL);
506 zio->io_stall = countp;
509 mutex_exit(&zio->io_lock);
514 __attribute__((always_inline))
516 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
518 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
519 int *errorp = &pio->io_child_error[zio->io_child_type];
521 mutex_enter(&pio->io_lock);
522 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
523 *errorp = zio_worst_error(*errorp, zio->io_error);
524 pio->io_reexecute |= zio->io_reexecute;
525 ASSERT3U(*countp, >, 0);
526 if (--*countp == 0 && pio->io_stall == countp) {
527 pio->io_stall = NULL;
528 mutex_exit(&pio->io_lock);
531 mutex_exit(&pio->io_lock);
536 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
538 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
539 zio->io_error = zio->io_child_error[c];
543 * ==========================================================================
544 * Create the various types of I/O (read, write, free, etc)
545 * ==========================================================================
548 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
549 void *data, uint64_t size, zio_done_func_t *done, void *private,
550 zio_type_t type, int priority, enum zio_flag flags,
551 vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
552 enum zio_stage stage, enum zio_stage pipeline)
556 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
557 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
558 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
560 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
561 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
562 ASSERT(vd || stage == ZIO_STAGE_OPEN);
564 zio = kmem_cache_alloc(zio_cache, KM_PUSHPAGE);
567 zio->io_child_type = ZIO_CHILD_VDEV;
568 else if (flags & ZIO_FLAG_GANG_CHILD)
569 zio->io_child_type = ZIO_CHILD_GANG;
570 else if (flags & ZIO_FLAG_DDT_CHILD)
571 zio->io_child_type = ZIO_CHILD_DDT;
573 zio->io_child_type = ZIO_CHILD_LOGICAL;
576 zio->io_logical = NULL;
577 zio->io_bp = (blkptr_t *)bp;
578 zio->io_bp_copy = *bp;
579 zio->io_bp_orig = *bp;
580 if (type != ZIO_TYPE_WRITE ||
581 zio->io_child_type == ZIO_CHILD_DDT)
582 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
583 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
584 zio->io_logical = zio;
585 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
586 pipeline |= ZIO_GANG_STAGES;
588 zio->io_logical = NULL;
590 bzero(&zio->io_bp_copy, sizeof (blkptr_t));
591 bzero(&zio->io_bp_orig, sizeof (blkptr_t));
596 zio->io_ready = NULL;
598 zio->io_private = private;
599 zio->io_prev_space_delta = 0;
601 zio->io_priority = priority;
604 zio->io_vsd_ops = NULL;
605 zio->io_offset = offset;
606 zio->io_deadline = 0;
607 zio->io_orig_data = zio->io_data = data;
608 zio->io_orig_size = zio->io_size = size;
609 zio->io_orig_flags = zio->io_flags = flags;
610 zio->io_orig_stage = zio->io_stage = stage;
611 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
612 bzero(&zio->io_prop, sizeof (zio_prop_t));
614 zio->io_reexecute = 0;
615 zio->io_bp_override = NULL;
616 zio->io_walk_link = NULL;
617 zio->io_transform_stack = NULL;
620 zio->io_child_count = 0;
621 zio->io_parent_count = 0;
622 zio->io_stall = NULL;
623 zio->io_gang_leader = NULL;
624 zio->io_gang_tree = NULL;
625 zio->io_executor = NULL;
626 zio->io_waiter = NULL;
627 zio->io_cksum_report = NULL;
629 bzero(zio->io_child_error, sizeof (int) * ZIO_CHILD_TYPES);
630 bzero(zio->io_children,
631 sizeof (uint64_t) * ZIO_CHILD_TYPES * ZIO_WAIT_TYPES);
632 bzero(&zio->io_bookmark, sizeof (zbookmark_t));
634 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
635 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
638 zio->io_bookmark = *zb;
641 if (zio->io_logical == NULL)
642 zio->io_logical = pio->io_logical;
643 if (zio->io_child_type == ZIO_CHILD_GANG)
644 zio->io_gang_leader = pio->io_gang_leader;
645 zio_add_child(pio, zio);
648 taskq_init_ent(&zio->io_tqent);
654 zio_destroy(zio_t *zio)
656 kmem_cache_free(zio_cache, zio);
660 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
661 void *private, enum zio_flag flags)
665 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
666 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
667 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
673 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
675 return (zio_null(NULL, spa, NULL, done, private, flags));
679 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
680 void *data, uint64_t size, zio_done_func_t *done, void *private,
681 int priority, enum zio_flag flags, const zbookmark_t *zb)
685 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
686 data, size, done, private,
687 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
688 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
689 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
695 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
696 void *data, uint64_t size, const zio_prop_t *zp,
697 zio_done_func_t *ready, zio_done_func_t *done, void *private,
698 int priority, enum zio_flag flags, const zbookmark_t *zb)
702 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
703 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
704 zp->zp_compress >= ZIO_COMPRESS_OFF &&
705 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
706 zp->zp_type < DMU_OT_NUMTYPES &&
709 zp->zp_copies <= spa_max_replication(spa) &&
711 zp->zp_dedup_verify <= 1);
713 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
714 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
715 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
716 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
718 zio->io_ready = ready;
725 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
726 uint64_t size, zio_done_func_t *done, void *private, int priority,
727 enum zio_flag flags, zbookmark_t *zb)
731 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
732 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
733 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
739 zio_write_override(zio_t *zio, blkptr_t *bp, int copies)
741 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
742 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
743 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
744 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
746 zio->io_prop.zp_copies = copies;
747 zio->io_bp_override = bp;
751 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
753 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
757 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
762 dprintf_bp(bp, "freeing in txg %llu, pass %u",
763 (longlong_t)txg, spa->spa_sync_pass);
765 ASSERT(!BP_IS_HOLE(bp));
766 ASSERT(spa_syncing_txg(spa) == txg);
767 ASSERT(spa_sync_pass(spa) <= SYNC_PASS_DEFERRED_FREE);
769 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
770 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
771 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
777 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
778 zio_done_func_t *done, void *private, enum zio_flag flags)
783 * A claim is an allocation of a specific block. Claims are needed
784 * to support immediate writes in the intent log. The issue is that
785 * immediate writes contain committed data, but in a txg that was
786 * *not* committed. Upon opening the pool after an unclean shutdown,
787 * the intent log claims all blocks that contain immediate write data
788 * so that the SPA knows they're in use.
790 * All claims *must* be resolved in the first txg -- before the SPA
791 * starts allocating blocks -- so that nothing is allocated twice.
792 * If txg == 0 we just verify that the block is claimable.
794 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
795 ASSERT(txg == spa_first_txg(spa) || txg == 0);
796 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
798 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
799 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
800 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
806 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
807 zio_done_func_t *done, void *private, int priority, enum zio_flag flags)
812 if (vd->vdev_children == 0) {
813 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
814 ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
815 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
819 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
821 for (c = 0; c < vd->vdev_children; c++)
822 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
823 done, private, priority, flags));
830 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
831 void *data, int checksum, zio_done_func_t *done, void *private,
832 int priority, enum zio_flag flags, boolean_t labels)
836 ASSERT(vd->vdev_children == 0);
837 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
838 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
839 ASSERT3U(offset + size, <=, vd->vdev_psize);
841 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
842 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
843 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
845 zio->io_prop.zp_checksum = checksum;
851 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
852 void *data, int checksum, zio_done_func_t *done, void *private,
853 int priority, enum zio_flag flags, boolean_t labels)
857 ASSERT(vd->vdev_children == 0);
858 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
859 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
860 ASSERT3U(offset + size, <=, vd->vdev_psize);
862 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
863 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
864 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
866 zio->io_prop.zp_checksum = checksum;
868 if (zio_checksum_table[checksum].ci_eck) {
870 * zec checksums are necessarily destructive -- they modify
871 * the end of the write buffer to hold the verifier/checksum.
872 * Therefore, we must make a local copy in case the data is
873 * being written to multiple places in parallel.
875 void *wbuf = zio_buf_alloc(size);
876 bcopy(data, wbuf, size);
877 zio_push_transform(zio, wbuf, size, size, NULL);
884 * Create a child I/O to do some work for us.
887 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
888 void *data, uint64_t size, int type, int priority, enum zio_flag flags,
889 zio_done_func_t *done, void *private)
891 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
894 ASSERT(vd->vdev_parent ==
895 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
897 if (type == ZIO_TYPE_READ && bp != NULL) {
899 * If we have the bp, then the child should perform the
900 * checksum and the parent need not. This pushes error
901 * detection as close to the leaves as possible and
902 * eliminates redundant checksums in the interior nodes.
904 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
905 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
908 if (vd->vdev_children == 0)
909 offset += VDEV_LABEL_START_SIZE;
911 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
914 * If we've decided to do a repair, the write is not speculative --
915 * even if the original read was.
917 if (flags & ZIO_FLAG_IO_REPAIR)
918 flags &= ~ZIO_FLAG_SPECULATIVE;
920 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
921 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
922 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
928 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
929 int type, int priority, enum zio_flag flags,
930 zio_done_func_t *done, void *private)
934 ASSERT(vd->vdev_ops->vdev_op_leaf);
936 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
937 data, size, done, private, type, priority,
938 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
940 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
946 zio_flush(zio_t *zio, vdev_t *vd)
948 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
949 NULL, NULL, ZIO_PRIORITY_NOW,
950 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
954 zio_shrink(zio_t *zio, uint64_t size)
956 ASSERT(zio->io_executor == NULL);
957 ASSERT(zio->io_orig_size == zio->io_size);
958 ASSERT(size <= zio->io_size);
961 * We don't shrink for raidz because of problems with the
962 * reconstruction when reading back less than the block size.
963 * Note, BP_IS_RAIDZ() assumes no compression.
965 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
966 if (!BP_IS_RAIDZ(zio->io_bp))
967 zio->io_orig_size = zio->io_size = size;
971 * ==========================================================================
972 * Prepare to read and write logical blocks
973 * ==========================================================================
977 zio_read_bp_init(zio_t *zio)
979 blkptr_t *bp = zio->io_bp;
981 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
982 zio->io_child_type == ZIO_CHILD_LOGICAL &&
983 !(zio->io_flags & ZIO_FLAG_RAW)) {
984 uint64_t psize = BP_GET_PSIZE(bp);
985 void *cbuf = zio_buf_alloc(psize);
987 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
990 if (!dmu_ot[BP_GET_TYPE(bp)].ot_metadata && BP_GET_LEVEL(bp) == 0)
991 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
993 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
994 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
996 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
997 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
999 return (ZIO_PIPELINE_CONTINUE);
1003 zio_write_bp_init(zio_t *zio)
1005 spa_t *spa = zio->io_spa;
1006 zio_prop_t *zp = &zio->io_prop;
1007 enum zio_compress compress = zp->zp_compress;
1008 blkptr_t *bp = zio->io_bp;
1009 uint64_t lsize = zio->io_size;
1010 uint64_t psize = lsize;
1014 * If our children haven't all reached the ready stage,
1015 * wait for them and then repeat this pipeline stage.
1017 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1018 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1019 return (ZIO_PIPELINE_STOP);
1021 if (!IO_IS_ALLOCATING(zio))
1022 return (ZIO_PIPELINE_CONTINUE);
1024 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1026 if (zio->io_bp_override) {
1027 ASSERT(bp->blk_birth != zio->io_txg);
1028 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1030 *bp = *zio->io_bp_override;
1031 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1033 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1034 return (ZIO_PIPELINE_CONTINUE);
1036 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1037 zp->zp_dedup_verify);
1039 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1040 BP_SET_DEDUP(bp, 1);
1041 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1042 return (ZIO_PIPELINE_CONTINUE);
1044 zio->io_bp_override = NULL;
1048 if (bp->blk_birth == zio->io_txg) {
1050 * We're rewriting an existing block, which means we're
1051 * working on behalf of spa_sync(). For spa_sync() to
1052 * converge, it must eventually be the case that we don't
1053 * have to allocate new blocks. But compression changes
1054 * the blocksize, which forces a reallocate, and makes
1055 * convergence take longer. Therefore, after the first
1056 * few passes, stop compressing to ensure convergence.
1058 pass = spa_sync_pass(spa);
1060 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1061 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1062 ASSERT(!BP_GET_DEDUP(bp));
1064 if (pass > SYNC_PASS_DONT_COMPRESS)
1065 compress = ZIO_COMPRESS_OFF;
1067 /* Make sure someone doesn't change their mind on overwrites */
1068 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
1069 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1072 if (compress != ZIO_COMPRESS_OFF) {
1073 void *cbuf = zio_buf_alloc(lsize);
1074 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1075 if (psize == 0 || psize == lsize) {
1076 compress = ZIO_COMPRESS_OFF;
1077 zio_buf_free(cbuf, lsize);
1079 ASSERT(psize < lsize);
1080 zio_push_transform(zio, cbuf, psize, lsize, NULL);
1085 * The final pass of spa_sync() must be all rewrites, but the first
1086 * few passes offer a trade-off: allocating blocks defers convergence,
1087 * but newly allocated blocks are sequential, so they can be written
1088 * to disk faster. Therefore, we allow the first few passes of
1089 * spa_sync() to allocate new blocks, but force rewrites after that.
1090 * There should only be a handful of blocks after pass 1 in any case.
1092 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
1093 pass > SYNC_PASS_REWRITE) {
1094 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1096 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1097 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1100 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1104 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1106 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1107 BP_SET_LSIZE(bp, lsize);
1108 BP_SET_PSIZE(bp, psize);
1109 BP_SET_COMPRESS(bp, compress);
1110 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1111 BP_SET_TYPE(bp, zp->zp_type);
1112 BP_SET_LEVEL(bp, zp->zp_level);
1113 BP_SET_DEDUP(bp, zp->zp_dedup);
1114 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1116 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1117 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1118 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1122 return (ZIO_PIPELINE_CONTINUE);
1126 zio_free_bp_init(zio_t *zio)
1128 blkptr_t *bp = zio->io_bp;
1130 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1131 if (BP_GET_DEDUP(bp))
1132 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1135 return (ZIO_PIPELINE_CONTINUE);
1139 * ==========================================================================
1140 * Execute the I/O pipeline
1141 * ==========================================================================
1145 zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q, boolean_t cutinline)
1147 spa_t *spa = zio->io_spa;
1148 zio_type_t t = zio->io_type;
1149 int flags = (cutinline ? TQ_FRONT : 0);
1152 * If we're a config writer or a probe, the normal issue and
1153 * interrupt threads may all be blocked waiting for the config lock.
1154 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1156 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1160 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1162 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1166 * If this is a high priority I/O, then use the high priority taskq.
1168 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1169 spa->spa_zio_taskq[t][q + 1] != NULL)
1172 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1175 * NB: We are assuming that the zio can only be dispatched
1176 * to a single taskq at a time. It would be a grievous error
1177 * to dispatch the zio to another taskq at the same time.
1179 ASSERT(taskq_empty_ent(&zio->io_tqent));
1180 taskq_dispatch_ent(spa->spa_zio_taskq[t][q],
1181 (task_func_t *)zio_execute, zio, flags, &zio->io_tqent);
1185 zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
1187 kthread_t *executor = zio->io_executor;
1188 spa_t *spa = zio->io_spa;
1191 for (t = 0; t < ZIO_TYPES; t++)
1192 if (taskq_member(spa->spa_zio_taskq[t][q], executor))
1199 zio_issue_async(zio_t *zio)
1201 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1203 return (ZIO_PIPELINE_STOP);
1207 zio_interrupt(zio_t *zio)
1209 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1213 * Execute the I/O pipeline until one of the following occurs:
1214 * (1) the I/O completes; (2) the pipeline stalls waiting for
1215 * dependent child I/Os; (3) the I/O issues, so we're waiting
1216 * for an I/O completion interrupt; (4) the I/O is delegated by
1217 * vdev-level caching or aggregation; (5) the I/O is deferred
1218 * due to vdev-level queueing; (6) the I/O is handed off to
1219 * another thread. In all cases, the pipeline stops whenever
1220 * there's no CPU work; it never burns a thread in cv_wait().
1222 * There's no locking on io_stage because there's no legitimate way
1223 * for multiple threads to be attempting to process the same I/O.
1225 static zio_pipe_stage_t *zio_pipeline[];
1228 * zio_execute() is a wrapper around the static function
1229 * __zio_execute() so that we can force __zio_execute() to be
1230 * inlined. This reduces stack overhead which is important
1231 * because __zio_execute() is called recursively in several zio
1232 * code paths. zio_execute() itself cannot be inlined because
1233 * it is externally visible.
1236 zio_execute(zio_t *zio)
1241 __attribute__((always_inline))
1243 __zio_execute(zio_t *zio)
1245 zio->io_executor = curthread;
1247 while (zio->io_stage < ZIO_STAGE_DONE) {
1248 enum zio_stage pipeline = zio->io_pipeline;
1249 enum zio_stage stage = zio->io_stage;
1254 ASSERT(!MUTEX_HELD(&zio->io_lock));
1255 ASSERT(ISP2(stage));
1256 ASSERT(zio->io_stall == NULL);
1260 } while ((stage & pipeline) == 0);
1262 ASSERT(stage <= ZIO_STAGE_DONE);
1264 dsl = spa_get_dsl(zio->io_spa);
1265 cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1266 zio_requeue_io_start_cut_in_line : B_FALSE;
1269 * If we are in interrupt context and this pipeline stage
1270 * will grab a config lock that is held across I/O,
1271 * or may wait for an I/O that needs an interrupt thread
1272 * to complete, issue async to avoid deadlock.
1274 * If we are in the txg_sync_thread or being called
1275 * during pool init issue async to minimize stack depth.
1276 * Both of these call paths may be recursively called.
1278 * For VDEV_IO_START, we cut in line so that the io will
1279 * be sent to disk promptly.
1281 if (((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1282 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) ||
1283 (dsl != NULL && dsl_pool_sync_context(dsl))) {
1284 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1288 zio->io_stage = stage;
1289 rv = zio_pipeline[highbit(stage) - 1](zio);
1291 if (rv == ZIO_PIPELINE_STOP)
1294 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1300 * ==========================================================================
1301 * Initiate I/O, either sync or async
1302 * ==========================================================================
1305 zio_wait(zio_t *zio)
1309 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1310 ASSERT(zio->io_executor == NULL);
1312 zio->io_waiter = curthread;
1316 mutex_enter(&zio->io_lock);
1317 while (zio->io_executor != NULL)
1318 cv_wait(&zio->io_cv, &zio->io_lock);
1319 mutex_exit(&zio->io_lock);
1321 error = zio->io_error;
1328 zio_nowait(zio_t *zio)
1330 ASSERT(zio->io_executor == NULL);
1332 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1333 zio_unique_parent(zio) == NULL) {
1335 * This is a logical async I/O with no parent to wait for it.
1336 * We add it to the spa_async_root_zio "Godfather" I/O which
1337 * will ensure they complete prior to unloading the pool.
1339 spa_t *spa = zio->io_spa;
1341 zio_add_child(spa->spa_async_zio_root, zio);
1348 * ==========================================================================
1349 * Reexecute or suspend/resume failed I/O
1350 * ==========================================================================
1354 zio_reexecute(zio_t *pio)
1356 zio_t *cio, *cio_next;
1359 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1360 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1361 ASSERT(pio->io_gang_leader == NULL);
1362 ASSERT(pio->io_gang_tree == NULL);
1364 pio->io_flags = pio->io_orig_flags;
1365 pio->io_stage = pio->io_orig_stage;
1366 pio->io_pipeline = pio->io_orig_pipeline;
1367 pio->io_reexecute = 0;
1369 for (w = 0; w < ZIO_WAIT_TYPES; w++)
1370 pio->io_state[w] = 0;
1371 for (c = 0; c < ZIO_CHILD_TYPES; c++)
1372 pio->io_child_error[c] = 0;
1374 if (IO_IS_ALLOCATING(pio))
1375 BP_ZERO(pio->io_bp);
1378 * As we reexecute pio's children, new children could be created.
1379 * New children go to the head of pio's io_child_list, however,
1380 * so we will (correctly) not reexecute them. The key is that
1381 * the remainder of pio's io_child_list, from 'cio_next' onward,
1382 * cannot be affected by any side effects of reexecuting 'cio'.
1384 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1385 cio_next = zio_walk_children(pio);
1386 mutex_enter(&pio->io_lock);
1387 for (w = 0; w < ZIO_WAIT_TYPES; w++)
1388 pio->io_children[cio->io_child_type][w]++;
1389 mutex_exit(&pio->io_lock);
1394 * Now that all children have been reexecuted, execute the parent.
1395 * We don't reexecute "The Godfather" I/O here as it's the
1396 * responsibility of the caller to wait on him.
1398 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1403 zio_suspend(spa_t *spa, zio_t *zio)
1405 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1406 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1407 "failure and the failure mode property for this pool "
1408 "is set to panic.", spa_name(spa));
1410 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1412 mutex_enter(&spa->spa_suspend_lock);
1414 if (spa->spa_suspend_zio_root == NULL)
1415 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1416 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1417 ZIO_FLAG_GODFATHER);
1419 spa->spa_suspended = B_TRUE;
1422 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1423 ASSERT(zio != spa->spa_suspend_zio_root);
1424 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1425 ASSERT(zio_unique_parent(zio) == NULL);
1426 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1427 zio_add_child(spa->spa_suspend_zio_root, zio);
1430 mutex_exit(&spa->spa_suspend_lock);
1434 zio_resume(spa_t *spa)
1439 * Reexecute all previously suspended i/o.
1441 mutex_enter(&spa->spa_suspend_lock);
1442 spa->spa_suspended = B_FALSE;
1443 cv_broadcast(&spa->spa_suspend_cv);
1444 pio = spa->spa_suspend_zio_root;
1445 spa->spa_suspend_zio_root = NULL;
1446 mutex_exit(&spa->spa_suspend_lock);
1452 return (zio_wait(pio));
1456 zio_resume_wait(spa_t *spa)
1458 mutex_enter(&spa->spa_suspend_lock);
1459 while (spa_suspended(spa))
1460 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1461 mutex_exit(&spa->spa_suspend_lock);
1465 * ==========================================================================
1468 * A gang block is a collection of small blocks that looks to the DMU
1469 * like one large block. When zio_dva_allocate() cannot find a block
1470 * of the requested size, due to either severe fragmentation or the pool
1471 * being nearly full, it calls zio_write_gang_block() to construct the
1472 * block from smaller fragments.
1474 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1475 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1476 * an indirect block: it's an array of block pointers. It consumes
1477 * only one sector and hence is allocatable regardless of fragmentation.
1478 * The gang header's bps point to its gang members, which hold the data.
1480 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1481 * as the verifier to ensure uniqueness of the SHA256 checksum.
1482 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1483 * not the gang header. This ensures that data block signatures (needed for
1484 * deduplication) are independent of how the block is physically stored.
1486 * Gang blocks can be nested: a gang member may itself be a gang block.
1487 * Thus every gang block is a tree in which root and all interior nodes are
1488 * gang headers, and the leaves are normal blocks that contain user data.
1489 * The root of the gang tree is called the gang leader.
1491 * To perform any operation (read, rewrite, free, claim) on a gang block,
1492 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1493 * in the io_gang_tree field of the original logical i/o by recursively
1494 * reading the gang leader and all gang headers below it. This yields
1495 * an in-core tree containing the contents of every gang header and the
1496 * bps for every constituent of the gang block.
1498 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1499 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1500 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1501 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1502 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1503 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1504 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1505 * of the gang header plus zio_checksum_compute() of the data to update the
1506 * gang header's blk_cksum as described above.
1508 * The two-phase assemble/issue model solves the problem of partial failure --
1509 * what if you'd freed part of a gang block but then couldn't read the
1510 * gang header for another part? Assembling the entire gang tree first
1511 * ensures that all the necessary gang header I/O has succeeded before
1512 * starting the actual work of free, claim, or write. Once the gang tree
1513 * is assembled, free and claim are in-memory operations that cannot fail.
1515 * In the event that a gang write fails, zio_dva_unallocate() walks the
1516 * gang tree to immediately free (i.e. insert back into the space map)
1517 * everything we've allocated. This ensures that we don't get ENOSPC
1518 * errors during repeated suspend/resume cycles due to a flaky device.
1520 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1521 * the gang tree, we won't modify the block, so we can safely defer the free
1522 * (knowing that the block is still intact). If we *can* assemble the gang
1523 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1524 * each constituent bp and we can allocate a new block on the next sync pass.
1526 * In all cases, the gang tree allows complete recovery from partial failure.
1527 * ==========================================================================
1531 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1536 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1537 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1538 &pio->io_bookmark));
1542 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1547 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1548 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1549 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1551 * As we rewrite each gang header, the pipeline will compute
1552 * a new gang block header checksum for it; but no one will
1553 * compute a new data checksum, so we do that here. The one
1554 * exception is the gang leader: the pipeline already computed
1555 * its data checksum because that stage precedes gang assembly.
1556 * (Presently, nothing actually uses interior data checksums;
1557 * this is just good hygiene.)
1559 if (gn != pio->io_gang_leader->io_gang_tree) {
1560 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1561 data, BP_GET_PSIZE(bp));
1564 * If we are here to damage data for testing purposes,
1565 * leave the GBH alone so that we can detect the damage.
1567 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1568 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1570 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1571 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1572 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1580 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1582 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1583 ZIO_GANG_CHILD_FLAGS(pio)));
1588 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1590 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1591 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1594 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1603 static void zio_gang_tree_assemble_done(zio_t *zio);
1605 static zio_gang_node_t *
1606 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1608 zio_gang_node_t *gn;
1610 ASSERT(*gnpp == NULL);
1612 gn = kmem_zalloc(sizeof (*gn), KM_PUSHPAGE);
1613 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1620 zio_gang_node_free(zio_gang_node_t **gnpp)
1622 zio_gang_node_t *gn = *gnpp;
1625 for (g = 0; g < SPA_GBH_NBLKPTRS; g++)
1626 ASSERT(gn->gn_child[g] == NULL);
1628 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1629 kmem_free(gn, sizeof (*gn));
1634 zio_gang_tree_free(zio_gang_node_t **gnpp)
1636 zio_gang_node_t *gn = *gnpp;
1642 for (g = 0; g < SPA_GBH_NBLKPTRS; g++)
1643 zio_gang_tree_free(&gn->gn_child[g]);
1645 zio_gang_node_free(gnpp);
1649 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1651 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1653 ASSERT(gio->io_gang_leader == gio);
1654 ASSERT(BP_IS_GANG(bp));
1656 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1657 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1658 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1662 zio_gang_tree_assemble_done(zio_t *zio)
1664 zio_t *gio = zio->io_gang_leader;
1665 zio_gang_node_t *gn = zio->io_private;
1666 blkptr_t *bp = zio->io_bp;
1669 ASSERT(gio == zio_unique_parent(zio));
1670 ASSERT(zio->io_child_count == 0);
1675 if (BP_SHOULD_BYTESWAP(bp))
1676 byteswap_uint64_array(zio->io_data, zio->io_size);
1678 ASSERT(zio->io_data == gn->gn_gbh);
1679 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1680 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1682 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1683 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1684 if (!BP_IS_GANG(gbp))
1686 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1691 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1693 zio_t *gio = pio->io_gang_leader;
1697 ASSERT(BP_IS_GANG(bp) == !!gn);
1698 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1699 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1702 * If you're a gang header, your data is in gn->gn_gbh.
1703 * If you're a gang member, your data is in 'data' and gn == NULL.
1705 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1708 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1710 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1711 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1712 if (BP_IS_HOLE(gbp))
1714 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1715 data = (char *)data + BP_GET_PSIZE(gbp);
1719 if (gn == gio->io_gang_tree)
1720 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1727 zio_gang_assemble(zio_t *zio)
1729 blkptr_t *bp = zio->io_bp;
1731 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1732 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1734 zio->io_gang_leader = zio;
1736 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1738 return (ZIO_PIPELINE_CONTINUE);
1742 zio_gang_issue(zio_t *zio)
1744 blkptr_t *bp = zio->io_bp;
1746 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1747 return (ZIO_PIPELINE_STOP);
1749 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1750 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1752 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1753 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1755 zio_gang_tree_free(&zio->io_gang_tree);
1757 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1759 return (ZIO_PIPELINE_CONTINUE);
1763 zio_write_gang_member_ready(zio_t *zio)
1765 zio_t *pio = zio_unique_parent(zio);
1766 ASSERTV(zio_t *gio = zio->io_gang_leader;)
1767 dva_t *cdva = zio->io_bp->blk_dva;
1768 dva_t *pdva = pio->io_bp->blk_dva;
1772 if (BP_IS_HOLE(zio->io_bp))
1775 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1777 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1778 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1779 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1780 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1781 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1783 mutex_enter(&pio->io_lock);
1784 for (d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1785 ASSERT(DVA_GET_GANG(&pdva[d]));
1786 asize = DVA_GET_ASIZE(&pdva[d]);
1787 asize += DVA_GET_ASIZE(&cdva[d]);
1788 DVA_SET_ASIZE(&pdva[d], asize);
1790 mutex_exit(&pio->io_lock);
1794 zio_write_gang_block(zio_t *pio)
1796 spa_t *spa = pio->io_spa;
1797 blkptr_t *bp = pio->io_bp;
1798 zio_t *gio = pio->io_gang_leader;
1800 zio_gang_node_t *gn, **gnpp;
1801 zio_gbh_phys_t *gbh;
1802 uint64_t txg = pio->io_txg;
1803 uint64_t resid = pio->io_size;
1805 int copies = gio->io_prop.zp_copies;
1806 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1810 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1811 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1812 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1814 pio->io_error = error;
1815 return (ZIO_PIPELINE_CONTINUE);
1819 gnpp = &gio->io_gang_tree;
1821 gnpp = pio->io_private;
1822 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1825 gn = zio_gang_node_alloc(gnpp);
1827 bzero(gbh, SPA_GANGBLOCKSIZE);
1830 * Create the gang header.
1832 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1833 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1836 * Create and nowait the gang children.
1838 for (g = 0; resid != 0; resid -= lsize, g++) {
1839 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1841 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1843 zp.zp_checksum = gio->io_prop.zp_checksum;
1844 zp.zp_compress = ZIO_COMPRESS_OFF;
1845 zp.zp_type = DMU_OT_NONE;
1847 zp.zp_copies = gio->io_prop.zp_copies;
1849 zp.zp_dedup_verify = 0;
1851 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1852 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1853 zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1854 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1855 &pio->io_bookmark));
1859 * Set pio's pipeline to just wait for zio to finish.
1861 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1865 return (ZIO_PIPELINE_CONTINUE);
1869 * ==========================================================================
1871 * ==========================================================================
1874 zio_ddt_child_read_done(zio_t *zio)
1876 blkptr_t *bp = zio->io_bp;
1877 ddt_entry_t *dde = zio->io_private;
1879 zio_t *pio = zio_unique_parent(zio);
1881 mutex_enter(&pio->io_lock);
1882 ddp = ddt_phys_select(dde, bp);
1883 if (zio->io_error == 0)
1884 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1885 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1886 dde->dde_repair_data = zio->io_data;
1888 zio_buf_free(zio->io_data, zio->io_size);
1889 mutex_exit(&pio->io_lock);
1893 zio_ddt_read_start(zio_t *zio)
1895 blkptr_t *bp = zio->io_bp;
1898 ASSERT(BP_GET_DEDUP(bp));
1899 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1900 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1902 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1903 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1904 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1905 ddt_phys_t *ddp = dde->dde_phys;
1906 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1909 ASSERT(zio->io_vsd == NULL);
1912 if (ddp_self == NULL)
1913 return (ZIO_PIPELINE_CONTINUE);
1915 for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1916 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1918 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1920 zio_nowait(zio_read(zio, zio->io_spa, &blk,
1921 zio_buf_alloc(zio->io_size), zio->io_size,
1922 zio_ddt_child_read_done, dde, zio->io_priority,
1923 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1924 &zio->io_bookmark));
1926 return (ZIO_PIPELINE_CONTINUE);
1929 zio_nowait(zio_read(zio, zio->io_spa, bp,
1930 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1931 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1933 return (ZIO_PIPELINE_CONTINUE);
1937 zio_ddt_read_done(zio_t *zio)
1939 blkptr_t *bp = zio->io_bp;
1941 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1942 return (ZIO_PIPELINE_STOP);
1944 ASSERT(BP_GET_DEDUP(bp));
1945 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1946 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1948 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1949 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1950 ddt_entry_t *dde = zio->io_vsd;
1952 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
1953 return (ZIO_PIPELINE_CONTINUE);
1956 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
1957 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1958 return (ZIO_PIPELINE_STOP);
1960 if (dde->dde_repair_data != NULL) {
1961 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
1962 zio->io_child_error[ZIO_CHILD_DDT] = 0;
1964 ddt_repair_done(ddt, dde);
1968 ASSERT(zio->io_vsd == NULL);
1970 return (ZIO_PIPELINE_CONTINUE);
1974 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
1976 spa_t *spa = zio->io_spa;
1980 * Note: we compare the original data, not the transformed data,
1981 * because when zio->io_bp is an override bp, we will not have
1982 * pushed the I/O transforms. That's an important optimization
1983 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
1985 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1986 zio_t *lio = dde->dde_lead_zio[p];
1989 return (lio->io_orig_size != zio->io_orig_size ||
1990 bcmp(zio->io_orig_data, lio->io_orig_data,
1991 zio->io_orig_size) != 0);
1995 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1996 ddt_phys_t *ddp = &dde->dde_phys[p];
1998 if (ddp->ddp_phys_birth != 0) {
1999 arc_buf_t *abuf = NULL;
2000 uint32_t aflags = ARC_WAIT;
2001 blkptr_t blk = *zio->io_bp;
2004 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2008 error = arc_read_nolock(NULL, spa, &blk,
2009 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2010 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2011 &aflags, &zio->io_bookmark);
2014 if (arc_buf_size(abuf) != zio->io_orig_size ||
2015 bcmp(abuf->b_data, zio->io_orig_data,
2016 zio->io_orig_size) != 0)
2018 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
2022 return (error != 0);
2030 zio_ddt_child_write_ready(zio_t *zio)
2032 int p = zio->io_prop.zp_copies;
2033 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2034 ddt_entry_t *dde = zio->io_private;
2035 ddt_phys_t *ddp = &dde->dde_phys[p];
2043 ASSERT(dde->dde_lead_zio[p] == zio);
2045 ddt_phys_fill(ddp, zio->io_bp);
2047 while ((pio = zio_walk_parents(zio)) != NULL)
2048 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2054 zio_ddt_child_write_done(zio_t *zio)
2056 int p = zio->io_prop.zp_copies;
2057 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2058 ddt_entry_t *dde = zio->io_private;
2059 ddt_phys_t *ddp = &dde->dde_phys[p];
2063 ASSERT(ddp->ddp_refcnt == 0);
2064 ASSERT(dde->dde_lead_zio[p] == zio);
2065 dde->dde_lead_zio[p] = NULL;
2067 if (zio->io_error == 0) {
2068 while (zio_walk_parents(zio) != NULL)
2069 ddt_phys_addref(ddp);
2071 ddt_phys_clear(ddp);
2078 zio_ddt_ditto_write_done(zio_t *zio)
2080 int p = DDT_PHYS_DITTO;
2081 blkptr_t *bp = zio->io_bp;
2082 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2083 ddt_entry_t *dde = zio->io_private;
2084 ddt_phys_t *ddp = &dde->dde_phys[p];
2085 ddt_key_t *ddk = &dde->dde_key;
2086 ASSERTV(zio_prop_t *zp = &zio->io_prop);
2090 ASSERT(ddp->ddp_refcnt == 0);
2091 ASSERT(dde->dde_lead_zio[p] == zio);
2092 dde->dde_lead_zio[p] = NULL;
2094 if (zio->io_error == 0) {
2095 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2096 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2097 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2098 if (ddp->ddp_phys_birth != 0)
2099 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2100 ddt_phys_fill(ddp, bp);
2107 zio_ddt_write(zio_t *zio)
2109 spa_t *spa = zio->io_spa;
2110 blkptr_t *bp = zio->io_bp;
2111 uint64_t txg = zio->io_txg;
2112 zio_prop_t *zp = &zio->io_prop;
2113 int p = zp->zp_copies;
2117 ddt_t *ddt = ddt_select(spa, bp);
2121 ASSERT(BP_GET_DEDUP(bp));
2122 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2123 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2126 dde = ddt_lookup(ddt, bp, B_TRUE);
2127 ddp = &dde->dde_phys[p];
2129 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2131 * If we're using a weak checksum, upgrade to a strong checksum
2132 * and try again. If we're already using a strong checksum,
2133 * we can't resolve it, so just convert to an ordinary write.
2134 * (And automatically e-mail a paper to Nature?)
2136 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2137 zp->zp_checksum = spa_dedup_checksum(spa);
2138 zio_pop_transforms(zio);
2139 zio->io_stage = ZIO_STAGE_OPEN;
2144 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2146 return (ZIO_PIPELINE_CONTINUE);
2149 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2150 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2152 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2153 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2154 zio_prop_t czp = *zp;
2156 czp.zp_copies = ditto_copies;
2159 * If we arrived here with an override bp, we won't have run
2160 * the transform stack, so we won't have the data we need to
2161 * generate a child i/o. So, toss the override bp and restart.
2162 * This is safe, because using the override bp is just an
2163 * optimization; and it's rare, so the cost doesn't matter.
2165 if (zio->io_bp_override) {
2166 zio_pop_transforms(zio);
2167 zio->io_stage = ZIO_STAGE_OPEN;
2168 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2169 zio->io_bp_override = NULL;
2172 return (ZIO_PIPELINE_CONTINUE);
2175 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2176 zio->io_orig_size, &czp, NULL,
2177 zio_ddt_ditto_write_done, dde, zio->io_priority,
2178 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2180 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2181 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2184 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2185 if (ddp->ddp_phys_birth != 0)
2186 ddt_bp_fill(ddp, bp, txg);
2187 if (dde->dde_lead_zio[p] != NULL)
2188 zio_add_child(zio, dde->dde_lead_zio[p]);
2190 ddt_phys_addref(ddp);
2191 } else if (zio->io_bp_override) {
2192 ASSERT(bp->blk_birth == txg);
2193 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2194 ddt_phys_fill(ddp, bp);
2195 ddt_phys_addref(ddp);
2197 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2198 zio->io_orig_size, zp, zio_ddt_child_write_ready,
2199 zio_ddt_child_write_done, dde, zio->io_priority,
2200 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2202 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2203 dde->dde_lead_zio[p] = cio;
2213 return (ZIO_PIPELINE_CONTINUE);
2216 ddt_entry_t *freedde; /* for debugging */
2219 zio_ddt_free(zio_t *zio)
2221 spa_t *spa = zio->io_spa;
2222 blkptr_t *bp = zio->io_bp;
2223 ddt_t *ddt = ddt_select(spa, bp);
2227 ASSERT(BP_GET_DEDUP(bp));
2228 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2231 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2232 ddp = ddt_phys_select(dde, bp);
2233 ddt_phys_decref(ddp);
2236 return (ZIO_PIPELINE_CONTINUE);
2240 * ==========================================================================
2241 * Allocate and free blocks
2242 * ==========================================================================
2245 zio_dva_allocate(zio_t *zio)
2247 spa_t *spa = zio->io_spa;
2248 metaslab_class_t *mc = spa_normal_class(spa);
2249 blkptr_t *bp = zio->io_bp;
2253 if (zio->io_gang_leader == NULL) {
2254 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2255 zio->io_gang_leader = zio;
2258 ASSERT(BP_IS_HOLE(bp));
2259 ASSERT3U(BP_GET_NDVAS(bp), ==, 0);
2260 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2261 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2262 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2265 * The dump device does not support gang blocks so allocation on
2266 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2267 * the "fast" gang feature.
2269 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2270 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2271 METASLAB_GANG_CHILD : 0;
2272 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2273 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2276 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2277 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2279 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2280 return (zio_write_gang_block(zio));
2281 zio->io_error = error;
2284 return (ZIO_PIPELINE_CONTINUE);
2288 zio_dva_free(zio_t *zio)
2290 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2292 return (ZIO_PIPELINE_CONTINUE);
2296 zio_dva_claim(zio_t *zio)
2300 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2302 zio->io_error = error;
2304 return (ZIO_PIPELINE_CONTINUE);
2308 * Undo an allocation. This is used by zio_done() when an I/O fails
2309 * and we want to give back the block we just allocated.
2310 * This handles both normal blocks and gang blocks.
2313 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2317 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2318 ASSERT(zio->io_bp_override == NULL);
2320 if (!BP_IS_HOLE(bp))
2321 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2324 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2325 zio_dva_unallocate(zio, gn->gn_child[g],
2326 &gn->gn_gbh->zg_blkptr[g]);
2332 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2335 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2336 uint64_t size, boolean_t use_slog)
2340 ASSERT(txg > spa_syncing_txg(spa));
2343 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2344 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2345 * when allocating them.
2348 error = metaslab_alloc(spa, spa_log_class(spa), size,
2349 new_bp, 1, txg, old_bp,
2350 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2354 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2355 new_bp, 1, txg, old_bp,
2356 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2360 BP_SET_LSIZE(new_bp, size);
2361 BP_SET_PSIZE(new_bp, size);
2362 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2363 BP_SET_CHECKSUM(new_bp,
2364 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2365 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2366 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2367 BP_SET_LEVEL(new_bp, 0);
2368 BP_SET_DEDUP(new_bp, 0);
2369 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2376 * Free an intent log block.
2379 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2381 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2382 ASSERT(!BP_IS_GANG(bp));
2384 zio_free(spa, txg, bp);
2388 * ==========================================================================
2389 * Read and write to physical devices
2390 * ==========================================================================
2393 zio_vdev_io_start(zio_t *zio)
2395 vdev_t *vd = zio->io_vd;
2397 spa_t *spa = zio->io_spa;
2399 ASSERT(zio->io_error == 0);
2400 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2403 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2404 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2407 * The mirror_ops handle multiple DVAs in a single BP.
2409 return (vdev_mirror_ops.vdev_op_io_start(zio));
2413 * We keep track of time-sensitive I/Os so that the scan thread
2414 * can quickly react to certain workloads. In particular, we care
2415 * about non-scrubbing, top-level reads and writes with the following
2417 * - synchronous writes of user data to non-slog devices
2418 * - any reads of user data
2419 * When these conditions are met, adjust the timestamp of spa_last_io
2420 * which allows the scan thread to adjust its workload accordingly.
2422 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2423 vd == vd->vdev_top && !vd->vdev_islog &&
2424 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2425 zio->io_txg != spa_syncing_txg(spa)) {
2426 uint64_t old = spa->spa_last_io;
2427 uint64_t new = ddi_get_lbolt64();
2429 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2432 align = 1ULL << vd->vdev_top->vdev_ashift;
2434 if (P2PHASE(zio->io_size, align) != 0) {
2435 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2436 char *abuf = zio_buf_alloc(asize);
2437 ASSERT(vd == vd->vdev_top);
2438 if (zio->io_type == ZIO_TYPE_WRITE) {
2439 bcopy(zio->io_data, abuf, zio->io_size);
2440 bzero(abuf + zio->io_size, asize - zio->io_size);
2442 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2445 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2446 ASSERT(P2PHASE(zio->io_size, align) == 0);
2447 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2450 * If this is a repair I/O, and there's no self-healing involved --
2451 * that is, we're just resilvering what we expect to resilver --
2452 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2453 * This prevents spurious resilvering with nested replication.
2454 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2455 * A is out of date, we'll read from C+D, then use the data to
2456 * resilver A+B -- but we don't actually want to resilver B, just A.
2457 * The top-level mirror has no way to know this, so instead we just
2458 * discard unnecessary repairs as we work our way down the vdev tree.
2459 * The same logic applies to any form of nested replication:
2460 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2462 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2463 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2464 zio->io_txg != 0 && /* not a delegated i/o */
2465 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2466 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2467 zio_vdev_io_bypass(zio);
2468 return (ZIO_PIPELINE_CONTINUE);
2471 if (vd->vdev_ops->vdev_op_leaf &&
2472 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2474 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2475 return (ZIO_PIPELINE_CONTINUE);
2477 if ((zio = vdev_queue_io(zio)) == NULL)
2478 return (ZIO_PIPELINE_STOP);
2480 if (!vdev_accessible(vd, zio)) {
2481 zio->io_error = ENXIO;
2483 return (ZIO_PIPELINE_STOP);
2487 return (vd->vdev_ops->vdev_op_io_start(zio));
2491 zio_vdev_io_done(zio_t *zio)
2493 vdev_t *vd = zio->io_vd;
2494 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2495 boolean_t unexpected_error = B_FALSE;
2497 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2498 return (ZIO_PIPELINE_STOP);
2500 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2502 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2504 vdev_queue_io_done(zio);
2506 if (zio->io_type == ZIO_TYPE_WRITE)
2507 vdev_cache_write(zio);
2509 if (zio_injection_enabled && zio->io_error == 0)
2510 zio->io_error = zio_handle_device_injection(vd,
2513 if (zio_injection_enabled && zio->io_error == 0)
2514 zio->io_error = zio_handle_label_injection(zio, EIO);
2516 if (zio->io_error) {
2517 if (!vdev_accessible(vd, zio)) {
2518 zio->io_error = ENXIO;
2520 unexpected_error = B_TRUE;
2525 ops->vdev_op_io_done(zio);
2527 if (unexpected_error)
2528 VERIFY(vdev_probe(vd, zio) == NULL);
2530 return (ZIO_PIPELINE_CONTINUE);
2534 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2535 * disk, and use that to finish the checksum ereport later.
2538 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2539 const void *good_buf)
2541 /* no processing needed */
2542 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2547 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2549 void *buf = zio_buf_alloc(zio->io_size);
2551 bcopy(zio->io_data, buf, zio->io_size);
2553 zcr->zcr_cbinfo = zio->io_size;
2554 zcr->zcr_cbdata = buf;
2555 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2556 zcr->zcr_free = zio_buf_free;
2560 zio_vdev_io_assess(zio_t *zio)
2562 vdev_t *vd = zio->io_vd;
2564 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2565 return (ZIO_PIPELINE_STOP);
2567 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2568 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2570 if (zio->io_vsd != NULL) {
2571 zio->io_vsd_ops->vsd_free(zio);
2575 if (zio_injection_enabled && zio->io_error == 0)
2576 zio->io_error = zio_handle_fault_injection(zio, EIO);
2579 * If the I/O failed, determine whether we should attempt to retry it.
2581 * On retry, we cut in line in the issue queue, since we don't want
2582 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2584 if (zio->io_error && vd == NULL &&
2585 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2586 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2587 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2589 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2590 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2591 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2592 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2593 zio_requeue_io_start_cut_in_line);
2594 return (ZIO_PIPELINE_STOP);
2598 * If we got an error on a leaf device, convert it to ENXIO
2599 * if the device is not accessible at all.
2601 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2602 !vdev_accessible(vd, zio))
2603 zio->io_error = ENXIO;
2606 * If we can't write to an interior vdev (mirror or RAID-Z),
2607 * set vdev_cant_write so that we stop trying to allocate from it.
2609 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2610 vd != NULL && !vd->vdev_ops->vdev_op_leaf)
2611 vd->vdev_cant_write = B_TRUE;
2614 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2616 return (ZIO_PIPELINE_CONTINUE);
2620 zio_vdev_io_reissue(zio_t *zio)
2622 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2623 ASSERT(zio->io_error == 0);
2625 zio->io_stage >>= 1;
2629 zio_vdev_io_redone(zio_t *zio)
2631 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2633 zio->io_stage >>= 1;
2637 zio_vdev_io_bypass(zio_t *zio)
2639 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2640 ASSERT(zio->io_error == 0);
2642 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2643 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2647 * ==========================================================================
2648 * Generate and verify checksums
2649 * ==========================================================================
2652 zio_checksum_generate(zio_t *zio)
2654 blkptr_t *bp = zio->io_bp;
2655 enum zio_checksum checksum;
2659 * This is zio_write_phys().
2660 * We're either generating a label checksum, or none at all.
2662 checksum = zio->io_prop.zp_checksum;
2664 if (checksum == ZIO_CHECKSUM_OFF)
2665 return (ZIO_PIPELINE_CONTINUE);
2667 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2669 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2670 ASSERT(!IO_IS_ALLOCATING(zio));
2671 checksum = ZIO_CHECKSUM_GANG_HEADER;
2673 checksum = BP_GET_CHECKSUM(bp);
2677 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2679 return (ZIO_PIPELINE_CONTINUE);
2683 zio_checksum_verify(zio_t *zio)
2685 zio_bad_cksum_t info;
2686 blkptr_t *bp = zio->io_bp;
2689 ASSERT(zio->io_vd != NULL);
2693 * This is zio_read_phys().
2694 * We're either verifying a label checksum, or nothing at all.
2696 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2697 return (ZIO_PIPELINE_CONTINUE);
2699 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2702 if ((error = zio_checksum_error(zio, &info)) != 0) {
2703 zio->io_error = error;
2704 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2705 zfs_ereport_start_checksum(zio->io_spa,
2706 zio->io_vd, zio, zio->io_offset,
2707 zio->io_size, NULL, &info);
2711 return (ZIO_PIPELINE_CONTINUE);
2715 * Called by RAID-Z to ensure we don't compute the checksum twice.
2718 zio_checksum_verified(zio_t *zio)
2720 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2724 * ==========================================================================
2725 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2726 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2727 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2728 * indicate errors that are specific to one I/O, and most likely permanent.
2729 * Any other error is presumed to be worse because we weren't expecting it.
2730 * ==========================================================================
2733 zio_worst_error(int e1, int e2)
2735 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2738 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2739 if (e1 == zio_error_rank[r1])
2742 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2743 if (e2 == zio_error_rank[r2])
2746 return (r1 > r2 ? e1 : e2);
2750 * ==========================================================================
2752 * ==========================================================================
2755 zio_ready(zio_t *zio)
2757 blkptr_t *bp = zio->io_bp;
2758 zio_t *pio, *pio_next;
2760 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2761 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2762 return (ZIO_PIPELINE_STOP);
2764 if (zio->io_ready) {
2765 ASSERT(IO_IS_ALLOCATING(zio));
2766 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2767 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2772 if (bp != NULL && bp != &zio->io_bp_copy)
2773 zio->io_bp_copy = *bp;
2776 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2778 mutex_enter(&zio->io_lock);
2779 zio->io_state[ZIO_WAIT_READY] = 1;
2780 pio = zio_walk_parents(zio);
2781 mutex_exit(&zio->io_lock);
2784 * As we notify zio's parents, new parents could be added.
2785 * New parents go to the head of zio's io_parent_list, however,
2786 * so we will (correctly) not notify them. The remainder of zio's
2787 * io_parent_list, from 'pio_next' onward, cannot change because
2788 * all parents must wait for us to be done before they can be done.
2790 for (; pio != NULL; pio = pio_next) {
2791 pio_next = zio_walk_parents(zio);
2792 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2795 if (zio->io_flags & ZIO_FLAG_NODATA) {
2796 if (BP_IS_GANG(bp)) {
2797 zio->io_flags &= ~ZIO_FLAG_NODATA;
2799 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2800 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2804 if (zio_injection_enabled &&
2805 zio->io_spa->spa_syncing_txg == zio->io_txg)
2806 zio_handle_ignored_writes(zio);
2808 return (ZIO_PIPELINE_CONTINUE);
2812 zio_done(zio_t *zio)
2814 zio_t *pio, *pio_next;
2818 * If our children haven't all completed,
2819 * wait for them and then repeat this pipeline stage.
2821 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2822 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2823 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2824 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2825 return (ZIO_PIPELINE_STOP);
2827 for (c = 0; c < ZIO_CHILD_TYPES; c++)
2828 for (w = 0; w < ZIO_WAIT_TYPES; w++)
2829 ASSERT(zio->io_children[c][w] == 0);
2831 if (zio->io_bp != NULL) {
2832 ASSERT(zio->io_bp->blk_pad[0] == 0);
2833 ASSERT(zio->io_bp->blk_pad[1] == 0);
2834 ASSERT(bcmp(zio->io_bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2835 (zio->io_bp == zio_unique_parent(zio)->io_bp));
2836 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) &&
2837 zio->io_bp_override == NULL &&
2838 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2839 ASSERT(!BP_SHOULD_BYTESWAP(zio->io_bp));
2840 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2841 ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 ||
2842 (BP_COUNT_GANG(zio->io_bp) == BP_GET_NDVAS(zio->io_bp)));
2847 * If there were child vdev/gang/ddt errors, they apply to us now.
2849 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2850 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2851 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2854 * If the I/O on the transformed data was successful, generate any
2855 * checksum reports now while we still have the transformed data.
2857 if (zio->io_error == 0) {
2858 while (zio->io_cksum_report != NULL) {
2859 zio_cksum_report_t *zcr = zio->io_cksum_report;
2860 uint64_t align = zcr->zcr_align;
2861 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2862 char *abuf = zio->io_data;
2864 if (asize != zio->io_size) {
2865 abuf = zio_buf_alloc(asize);
2866 bcopy(zio->io_data, abuf, zio->io_size);
2867 bzero(abuf + zio->io_size, asize - zio->io_size);
2870 zio->io_cksum_report = zcr->zcr_next;
2871 zcr->zcr_next = NULL;
2872 zcr->zcr_finish(zcr, abuf);
2873 zfs_ereport_free_checksum(zcr);
2875 if (asize != zio->io_size)
2876 zio_buf_free(abuf, asize);
2880 zio_pop_transforms(zio); /* note: may set zio->io_error */
2882 vdev_stat_update(zio, zio->io_size);
2885 * If this I/O is attached to a particular vdev is slow, exeeding
2886 * 30 seconds to complete, post an error described the I/O delay.
2887 * We ignore these errors if the device is currently unavailable.
2889 if (zio->io_delay >= zio_delay_max) {
2890 if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd))
2891 zfs_ereport_post(FM_EREPORT_ZFS_DELAY, zio->io_spa,
2892 zio->io_vd, zio, 0, 0);
2895 if (zio->io_error) {
2897 * If this I/O is attached to a particular vdev,
2898 * generate an error message describing the I/O failure
2899 * at the block level. We ignore these errors if the
2900 * device is currently unavailable.
2902 if (zio->io_error != ECKSUM && zio->io_vd != NULL &&
2903 !vdev_is_dead(zio->io_vd))
2904 zfs_ereport_post(FM_EREPORT_ZFS_IO, zio->io_spa,
2905 zio->io_vd, zio, 0, 0);
2907 if ((zio->io_error == EIO || !(zio->io_flags &
2908 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2909 zio == zio->io_logical) {
2911 * For logical I/O requests, tell the SPA to log the
2912 * error and generate a logical data ereport.
2914 spa_log_error(zio->io_spa, zio);
2915 zfs_ereport_post(FM_EREPORT_ZFS_DATA, zio->io_spa, NULL, zio,
2920 if (zio->io_error && zio == zio->io_logical) {
2922 * Determine whether zio should be reexecuted. This will
2923 * propagate all the way to the root via zio_notify_parent().
2925 ASSERT(zio->io_vd == NULL && zio->io_bp != NULL);
2926 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2928 if (IO_IS_ALLOCATING(zio) &&
2929 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2930 if (zio->io_error != ENOSPC)
2931 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2933 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2936 if ((zio->io_type == ZIO_TYPE_READ ||
2937 zio->io_type == ZIO_TYPE_FREE) &&
2938 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
2939 zio->io_error == ENXIO &&
2940 spa_load_state(zio->io_spa) == SPA_LOAD_NONE &&
2941 spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE)
2942 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2944 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2945 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2948 * Here is a possibly good place to attempt to do
2949 * either combinatorial reconstruction or error correction
2950 * based on checksums. It also might be a good place
2951 * to send out preliminary ereports before we suspend
2957 * If there were logical child errors, they apply to us now.
2958 * We defer this until now to avoid conflating logical child
2959 * errors with errors that happened to the zio itself when
2960 * updating vdev stats and reporting FMA events above.
2962 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2964 if ((zio->io_error || zio->io_reexecute) &&
2965 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
2966 !(zio->io_flags & ZIO_FLAG_IO_REWRITE))
2967 zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp);
2969 zio_gang_tree_free(&zio->io_gang_tree);
2972 * Godfather I/Os should never suspend.
2974 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
2975 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
2976 zio->io_reexecute = 0;
2978 if (zio->io_reexecute) {
2980 * This is a logical I/O that wants to reexecute.
2982 * Reexecute is top-down. When an i/o fails, if it's not
2983 * the root, it simply notifies its parent and sticks around.
2984 * The parent, seeing that it still has children in zio_done(),
2985 * does the same. This percolates all the way up to the root.
2986 * The root i/o will reexecute or suspend the entire tree.
2988 * This approach ensures that zio_reexecute() honors
2989 * all the original i/o dependency relationships, e.g.
2990 * parents not executing until children are ready.
2992 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2994 zio->io_gang_leader = NULL;
2996 mutex_enter(&zio->io_lock);
2997 zio->io_state[ZIO_WAIT_DONE] = 1;
2998 mutex_exit(&zio->io_lock);
3001 * "The Godfather" I/O monitors its children but is
3002 * not a true parent to them. It will track them through
3003 * the pipeline but severs its ties whenever they get into
3004 * trouble (e.g. suspended). This allows "The Godfather"
3005 * I/O to return status without blocking.
3007 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3008 zio_link_t *zl = zio->io_walk_link;
3009 pio_next = zio_walk_parents(zio);
3011 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3012 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3013 zio_remove_child(pio, zio, zl);
3014 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3018 if ((pio = zio_unique_parent(zio)) != NULL) {
3020 * We're not a root i/o, so there's nothing to do
3021 * but notify our parent. Don't propagate errors
3022 * upward since we haven't permanently failed yet.
3024 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3025 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3026 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3027 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3029 * We'd fail again if we reexecuted now, so suspend
3030 * until conditions improve (e.g. device comes online).
3032 zio_suspend(zio->io_spa, zio);
3035 * Reexecution is potentially a huge amount of work.
3036 * Hand it off to the otherwise-unused claim taskq.
3038 ASSERT(taskq_empty_ent(&zio->io_tqent));
3039 (void) taskq_dispatch_ent(
3040 zio->io_spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
3041 (task_func_t *)zio_reexecute, zio, 0,
3044 return (ZIO_PIPELINE_STOP);
3047 ASSERT(zio->io_child_count == 0);
3048 ASSERT(zio->io_reexecute == 0);
3049 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3052 * Report any checksum errors, since the I/O is complete.
3054 while (zio->io_cksum_report != NULL) {
3055 zio_cksum_report_t *zcr = zio->io_cksum_report;
3056 zio->io_cksum_report = zcr->zcr_next;
3057 zcr->zcr_next = NULL;
3058 zcr->zcr_finish(zcr, NULL);
3059 zfs_ereport_free_checksum(zcr);
3063 * It is the responsibility of the done callback to ensure that this
3064 * particular zio is no longer discoverable for adoption, and as
3065 * such, cannot acquire any new parents.
3070 mutex_enter(&zio->io_lock);
3071 zio->io_state[ZIO_WAIT_DONE] = 1;
3072 mutex_exit(&zio->io_lock);
3074 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3075 zio_link_t *zl = zio->io_walk_link;
3076 pio_next = zio_walk_parents(zio);
3077 zio_remove_child(pio, zio, zl);
3078 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3081 if (zio->io_waiter != NULL) {
3082 mutex_enter(&zio->io_lock);
3083 zio->io_executor = NULL;
3084 cv_broadcast(&zio->io_cv);
3085 mutex_exit(&zio->io_lock);
3090 return (ZIO_PIPELINE_STOP);
3094 * ==========================================================================
3095 * I/O pipeline definition
3096 * ==========================================================================
3098 static zio_pipe_stage_t *zio_pipeline[] = {
3104 zio_checksum_generate,
3118 zio_checksum_verify,
3122 #if defined(_KERNEL) && defined(HAVE_SPL)
3123 /* Fault injection */
3124 EXPORT_SYMBOL(zio_injection_enabled);
3125 EXPORT_SYMBOL(zio_inject_fault);
3126 EXPORT_SYMBOL(zio_inject_list_next);
3127 EXPORT_SYMBOL(zio_clear_fault);
3128 EXPORT_SYMBOL(zio_handle_fault_injection);
3129 EXPORT_SYMBOL(zio_handle_device_injection);
3130 EXPORT_SYMBOL(zio_handle_label_injection);
3131 EXPORT_SYMBOL(zio_priority_table);
3132 EXPORT_SYMBOL(zio_type_name);
3134 module_param(zio_bulk_flags, int, 0644);
3135 MODULE_PARM_DESC(zio_bulk_flags, "Additional flags to pass to bulk buffers");
3137 module_param(zio_delay_max, int, 0644);
3138 MODULE_PARM_DESC(zio_delay_max, "Max zio millisec delay before posting event");
3140 module_param(zio_requeue_io_start_cut_in_line, int, 0644);
3141 MODULE_PARM_DESC(zio_requeue_io_start_cut_in_line, "Prioritize requeued I/O");