4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011 by Delphix. All rights reserved.
26 #include <sys/zfs_context.h>
27 #include <sys/fm/fs/zfs.h>
30 #include <sys/spa_impl.h>
31 #include <sys/vdev_impl.h>
32 #include <sys/zio_impl.h>
33 #include <sys/zio_compress.h>
34 #include <sys/zio_checksum.h>
35 #include <sys/dmu_objset.h>
40 * ==========================================================================
42 * ==========================================================================
44 uint8_t zio_priority_table[ZIO_PRIORITY_TABLE_SIZE] = {
45 0, /* ZIO_PRIORITY_NOW */
46 0, /* ZIO_PRIORITY_SYNC_READ */
47 0, /* ZIO_PRIORITY_SYNC_WRITE */
48 0, /* ZIO_PRIORITY_LOG_WRITE */
49 1, /* ZIO_PRIORITY_CACHE_FILL */
50 1, /* ZIO_PRIORITY_AGG */
51 4, /* ZIO_PRIORITY_FREE */
52 4, /* ZIO_PRIORITY_ASYNC_WRITE */
53 6, /* ZIO_PRIORITY_ASYNC_READ */
54 10, /* ZIO_PRIORITY_RESILVER */
55 20, /* ZIO_PRIORITY_SCRUB */
56 2, /* ZIO_PRIORITY_DDT_PREFETCH */
60 * ==========================================================================
61 * I/O type descriptions
62 * ==========================================================================
64 char *zio_type_name[ZIO_TYPES] = {
65 "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl"
69 * ==========================================================================
71 * ==========================================================================
73 kmem_cache_t *zio_cache;
74 kmem_cache_t *zio_link_cache;
75 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
76 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
77 int zio_bulk_flags = 0;
78 int zio_delay_max = ZIO_DELAY_MAX;
81 extern vmem_t *zio_alloc_arena;
83 extern int zfs_mg_alloc_failures;
86 * An allocating zio is one that either currently has the DVA allocate
87 * stage set or will have it later in its lifetime.
89 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
91 int zio_requeue_io_start_cut_in_line = 1;
94 int zio_buf_debug_limit = 16384;
96 int zio_buf_debug_limit = 0;
99 static inline void __zio_execute(zio_t *zio);
105 vmem_t *data_alloc_arena = NULL;
108 data_alloc_arena = zio_alloc_arena;
110 zio_cache = kmem_cache_create("zio_cache",
111 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
112 zio_link_cache = kmem_cache_create("zio_link_cache",
113 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
116 * For small buffers, we want a cache for each multiple of
117 * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache
118 * for each quarter-power of 2. For large buffers, we want
119 * a cache for each multiple of PAGESIZE.
121 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
122 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
126 while (p2 & (p2 - 1))
129 if (size <= 4 * SPA_MINBLOCKSIZE) {
130 align = SPA_MINBLOCKSIZE;
131 } else if (P2PHASE(size, PAGESIZE) == 0) {
133 } else if (P2PHASE(size, p2 >> 2) == 0) {
139 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
140 zio_buf_cache[c] = kmem_cache_create(name, size,
141 align, NULL, NULL, NULL, NULL, NULL,
142 (size > zio_buf_debug_limit ? KMC_NODEBUG : 0) |
145 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
146 zio_data_buf_cache[c] = kmem_cache_create(name, size,
147 align, NULL, NULL, NULL, NULL, data_alloc_arena,
148 (size > zio_buf_debug_limit ? KMC_NODEBUG : 0) |
154 ASSERT(zio_buf_cache[c] != NULL);
155 if (zio_buf_cache[c - 1] == NULL)
156 zio_buf_cache[c - 1] = zio_buf_cache[c];
158 ASSERT(zio_data_buf_cache[c] != NULL);
159 if (zio_data_buf_cache[c - 1] == NULL)
160 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
164 * The zio write taskqs have 1 thread per cpu, allow 1/2 of the taskqs
165 * to fail 3 times per txg or 8 failures, whichever is greater.
167 zfs_mg_alloc_failures = MAX((3 * max_ncpus / 2), 8);
176 kmem_cache_t *last_cache = NULL;
177 kmem_cache_t *last_data_cache = NULL;
179 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
180 if (zio_buf_cache[c] != last_cache) {
181 last_cache = zio_buf_cache[c];
182 kmem_cache_destroy(zio_buf_cache[c]);
184 zio_buf_cache[c] = NULL;
186 if (zio_data_buf_cache[c] != last_data_cache) {
187 last_data_cache = zio_data_buf_cache[c];
188 kmem_cache_destroy(zio_data_buf_cache[c]);
190 zio_data_buf_cache[c] = NULL;
193 kmem_cache_destroy(zio_link_cache);
194 kmem_cache_destroy(zio_cache);
200 * ==========================================================================
201 * Allocate and free I/O buffers
202 * ==========================================================================
206 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
207 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
208 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
209 * excess / transient data in-core during a crashdump.
212 zio_buf_alloc(size_t size)
214 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
216 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
218 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
222 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
223 * crashdump if the kernel panics. This exists so that we will limit the amount
224 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
225 * of kernel heap dumped to disk when the kernel panics)
228 zio_data_buf_alloc(size_t size)
230 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
232 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
234 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
238 zio_buf_free(void *buf, size_t size)
240 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
242 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
244 kmem_cache_free(zio_buf_cache[c], buf);
248 zio_data_buf_free(void *buf, size_t size)
250 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
252 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
254 kmem_cache_free(zio_data_buf_cache[c], buf);
258 * ==========================================================================
259 * Push and pop I/O transform buffers
260 * ==========================================================================
263 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
264 zio_transform_func_t *transform)
266 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_PUSHPAGE);
268 zt->zt_orig_data = zio->io_data;
269 zt->zt_orig_size = zio->io_size;
270 zt->zt_bufsize = bufsize;
271 zt->zt_transform = transform;
273 zt->zt_next = zio->io_transform_stack;
274 zio->io_transform_stack = zt;
281 zio_pop_transforms(zio_t *zio)
285 while ((zt = zio->io_transform_stack) != NULL) {
286 if (zt->zt_transform != NULL)
287 zt->zt_transform(zio,
288 zt->zt_orig_data, zt->zt_orig_size);
290 if (zt->zt_bufsize != 0)
291 zio_buf_free(zio->io_data, zt->zt_bufsize);
293 zio->io_data = zt->zt_orig_data;
294 zio->io_size = zt->zt_orig_size;
295 zio->io_transform_stack = zt->zt_next;
297 kmem_free(zt, sizeof (zio_transform_t));
302 * ==========================================================================
303 * I/O transform callbacks for subblocks and decompression
304 * ==========================================================================
307 zio_subblock(zio_t *zio, void *data, uint64_t size)
309 ASSERT(zio->io_size > size);
311 if (zio->io_type == ZIO_TYPE_READ)
312 bcopy(zio->io_data, data, size);
316 zio_decompress(zio_t *zio, void *data, uint64_t size)
318 if (zio->io_error == 0 &&
319 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
320 zio->io_data, data, zio->io_size, size) != 0)
325 * ==========================================================================
326 * I/O parent/child relationships and pipeline interlocks
327 * ==========================================================================
330 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
331 * continue calling these functions until they return NULL.
332 * Otherwise, the next caller will pick up the list walk in
333 * some indeterminate state. (Otherwise every caller would
334 * have to pass in a cookie to keep the state represented by
335 * io_walk_link, which gets annoying.)
338 zio_walk_parents(zio_t *cio)
340 zio_link_t *zl = cio->io_walk_link;
341 list_t *pl = &cio->io_parent_list;
343 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
344 cio->io_walk_link = zl;
349 ASSERT(zl->zl_child == cio);
350 return (zl->zl_parent);
354 zio_walk_children(zio_t *pio)
356 zio_link_t *zl = pio->io_walk_link;
357 list_t *cl = &pio->io_child_list;
359 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
360 pio->io_walk_link = zl;
365 ASSERT(zl->zl_parent == pio);
366 return (zl->zl_child);
370 zio_unique_parent(zio_t *cio)
372 zio_t *pio = zio_walk_parents(cio);
374 VERIFY(zio_walk_parents(cio) == NULL);
379 zio_add_child(zio_t *pio, zio_t *cio)
381 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_PUSHPAGE);
385 * Logical I/Os can have logical, gang, or vdev children.
386 * Gang I/Os can have gang or vdev children.
387 * Vdev I/Os can only have vdev children.
388 * The following ASSERT captures all of these constraints.
390 ASSERT(cio->io_child_type <= pio->io_child_type);
395 mutex_enter(&cio->io_lock);
396 mutex_enter(&pio->io_lock);
398 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
400 for (w = 0; w < ZIO_WAIT_TYPES; w++)
401 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
403 list_insert_head(&pio->io_child_list, zl);
404 list_insert_head(&cio->io_parent_list, zl);
406 pio->io_child_count++;
407 cio->io_parent_count++;
409 mutex_exit(&pio->io_lock);
410 mutex_exit(&cio->io_lock);
414 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
416 ASSERT(zl->zl_parent == pio);
417 ASSERT(zl->zl_child == cio);
419 mutex_enter(&cio->io_lock);
420 mutex_enter(&pio->io_lock);
422 list_remove(&pio->io_child_list, zl);
423 list_remove(&cio->io_parent_list, zl);
425 pio->io_child_count--;
426 cio->io_parent_count--;
428 mutex_exit(&pio->io_lock);
429 mutex_exit(&cio->io_lock);
431 kmem_cache_free(zio_link_cache, zl);
435 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
437 uint64_t *countp = &zio->io_children[child][wait];
438 boolean_t waiting = B_FALSE;
440 mutex_enter(&zio->io_lock);
441 ASSERT(zio->io_stall == NULL);
444 zio->io_stall = countp;
447 mutex_exit(&zio->io_lock);
452 __attribute__((always_inline))
454 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
456 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
457 int *errorp = &pio->io_child_error[zio->io_child_type];
459 mutex_enter(&pio->io_lock);
460 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
461 *errorp = zio_worst_error(*errorp, zio->io_error);
462 pio->io_reexecute |= zio->io_reexecute;
463 ASSERT3U(*countp, >, 0);
464 if (--*countp == 0 && pio->io_stall == countp) {
465 pio->io_stall = NULL;
466 mutex_exit(&pio->io_lock);
469 mutex_exit(&pio->io_lock);
474 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
476 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
477 zio->io_error = zio->io_child_error[c];
481 * ==========================================================================
482 * Create the various types of I/O (read, write, free, etc)
483 * ==========================================================================
486 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
487 void *data, uint64_t size, zio_done_func_t *done, void *private,
488 zio_type_t type, int priority, enum zio_flag flags,
489 vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
490 enum zio_stage stage, enum zio_stage pipeline)
494 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
495 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
496 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
498 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
499 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
500 ASSERT(vd || stage == ZIO_STAGE_OPEN);
502 zio = kmem_cache_alloc(zio_cache, KM_PUSHPAGE);
503 bzero(zio, sizeof (zio_t));
505 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
506 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
508 list_create(&zio->io_parent_list, sizeof (zio_link_t),
509 offsetof(zio_link_t, zl_parent_node));
510 list_create(&zio->io_child_list, sizeof (zio_link_t),
511 offsetof(zio_link_t, zl_child_node));
514 zio->io_child_type = ZIO_CHILD_VDEV;
515 else if (flags & ZIO_FLAG_GANG_CHILD)
516 zio->io_child_type = ZIO_CHILD_GANG;
517 else if (flags & ZIO_FLAG_DDT_CHILD)
518 zio->io_child_type = ZIO_CHILD_DDT;
520 zio->io_child_type = ZIO_CHILD_LOGICAL;
523 zio->io_bp = (blkptr_t *)bp;
524 zio->io_bp_copy = *bp;
525 zio->io_bp_orig = *bp;
526 if (type != ZIO_TYPE_WRITE ||
527 zio->io_child_type == ZIO_CHILD_DDT)
528 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
529 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
530 zio->io_logical = zio;
531 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
532 pipeline |= ZIO_GANG_STAGES;
538 zio->io_private = private;
540 zio->io_priority = priority;
542 zio->io_offset = offset;
543 zio->io_orig_data = zio->io_data = data;
544 zio->io_orig_size = zio->io_size = size;
545 zio->io_orig_flags = zio->io_flags = flags;
546 zio->io_orig_stage = zio->io_stage = stage;
547 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
549 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
550 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
553 zio->io_bookmark = *zb;
556 if (zio->io_logical == NULL)
557 zio->io_logical = pio->io_logical;
558 if (zio->io_child_type == ZIO_CHILD_GANG)
559 zio->io_gang_leader = pio->io_gang_leader;
560 zio_add_child(pio, zio);
567 zio_destroy(zio_t *zio)
569 list_destroy(&zio->io_parent_list);
570 list_destroy(&zio->io_child_list);
571 mutex_destroy(&zio->io_lock);
572 cv_destroy(&zio->io_cv);
573 kmem_cache_free(zio_cache, zio);
577 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
578 void *private, enum zio_flag flags)
582 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
583 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
584 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
590 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
592 return (zio_null(NULL, spa, NULL, done, private, flags));
596 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
597 void *data, uint64_t size, zio_done_func_t *done, void *private,
598 int priority, enum zio_flag flags, const zbookmark_t *zb)
602 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
603 data, size, done, private,
604 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
605 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
606 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
612 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
613 void *data, uint64_t size, const zio_prop_t *zp,
614 zio_done_func_t *ready, zio_done_func_t *done, void *private,
615 int priority, enum zio_flag flags, const zbookmark_t *zb)
619 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
620 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
621 zp->zp_compress >= ZIO_COMPRESS_OFF &&
622 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
623 zp->zp_type < DMU_OT_NUMTYPES &&
626 zp->zp_copies <= spa_max_replication(spa) &&
628 zp->zp_dedup_verify <= 1);
630 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
631 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
632 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
633 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
635 zio->io_ready = ready;
642 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
643 uint64_t size, zio_done_func_t *done, void *private, int priority,
644 enum zio_flag flags, zbookmark_t *zb)
648 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
649 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
650 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
656 zio_write_override(zio_t *zio, blkptr_t *bp, int copies)
658 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
659 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
660 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
661 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
663 zio->io_prop.zp_copies = copies;
664 zio->io_bp_override = bp;
668 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
670 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
674 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
679 dprintf_bp(bp, "freeing in txg %llu, pass %u",
680 (longlong_t)txg, spa->spa_sync_pass);
682 ASSERT(!BP_IS_HOLE(bp));
683 ASSERT(spa_syncing_txg(spa) == txg);
684 ASSERT(spa_sync_pass(spa) <= SYNC_PASS_DEFERRED_FREE);
686 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
687 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
688 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
694 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
695 zio_done_func_t *done, void *private, enum zio_flag flags)
700 * A claim is an allocation of a specific block. Claims are needed
701 * to support immediate writes in the intent log. The issue is that
702 * immediate writes contain committed data, but in a txg that was
703 * *not* committed. Upon opening the pool after an unclean shutdown,
704 * the intent log claims all blocks that contain immediate write data
705 * so that the SPA knows they're in use.
707 * All claims *must* be resolved in the first txg -- before the SPA
708 * starts allocating blocks -- so that nothing is allocated twice.
709 * If txg == 0 we just verify that the block is claimable.
711 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
712 ASSERT(txg == spa_first_txg(spa) || txg == 0);
713 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
715 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
716 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
717 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
723 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
724 zio_done_func_t *done, void *private, int priority, enum zio_flag flags)
729 if (vd->vdev_children == 0) {
730 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
731 ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
732 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
736 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
738 for (c = 0; c < vd->vdev_children; c++)
739 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
740 done, private, priority, flags));
747 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
748 void *data, int checksum, zio_done_func_t *done, void *private,
749 int priority, enum zio_flag flags, boolean_t labels)
753 ASSERT(vd->vdev_children == 0);
754 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
755 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
756 ASSERT3U(offset + size, <=, vd->vdev_psize);
758 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
759 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
760 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
762 zio->io_prop.zp_checksum = checksum;
768 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
769 void *data, int checksum, zio_done_func_t *done, void *private,
770 int priority, enum zio_flag flags, boolean_t labels)
774 ASSERT(vd->vdev_children == 0);
775 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
776 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
777 ASSERT3U(offset + size, <=, vd->vdev_psize);
779 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
780 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
781 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
783 zio->io_prop.zp_checksum = checksum;
785 if (zio_checksum_table[checksum].ci_eck) {
787 * zec checksums are necessarily destructive -- they modify
788 * the end of the write buffer to hold the verifier/checksum.
789 * Therefore, we must make a local copy in case the data is
790 * being written to multiple places in parallel.
792 void *wbuf = zio_buf_alloc(size);
793 bcopy(data, wbuf, size);
794 zio_push_transform(zio, wbuf, size, size, NULL);
801 * Create a child I/O to do some work for us.
804 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
805 void *data, uint64_t size, int type, int priority, enum zio_flag flags,
806 zio_done_func_t *done, void *private)
808 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
811 ASSERT(vd->vdev_parent ==
812 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
814 if (type == ZIO_TYPE_READ && bp != NULL) {
816 * If we have the bp, then the child should perform the
817 * checksum and the parent need not. This pushes error
818 * detection as close to the leaves as possible and
819 * eliminates redundant checksums in the interior nodes.
821 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
822 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
825 if (vd->vdev_children == 0)
826 offset += VDEV_LABEL_START_SIZE;
828 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
831 * If we've decided to do a repair, the write is not speculative --
832 * even if the original read was.
834 if (flags & ZIO_FLAG_IO_REPAIR)
835 flags &= ~ZIO_FLAG_SPECULATIVE;
837 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
838 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
839 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
845 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
846 int type, int priority, enum zio_flag flags,
847 zio_done_func_t *done, void *private)
851 ASSERT(vd->vdev_ops->vdev_op_leaf);
853 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
854 data, size, done, private, type, priority,
855 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
857 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
863 zio_flush(zio_t *zio, vdev_t *vd)
865 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
866 NULL, NULL, ZIO_PRIORITY_NOW,
867 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
871 zio_shrink(zio_t *zio, uint64_t size)
873 ASSERT(zio->io_executor == NULL);
874 ASSERT(zio->io_orig_size == zio->io_size);
875 ASSERT(size <= zio->io_size);
878 * We don't shrink for raidz because of problems with the
879 * reconstruction when reading back less than the block size.
880 * Note, BP_IS_RAIDZ() assumes no compression.
882 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
883 if (!BP_IS_RAIDZ(zio->io_bp))
884 zio->io_orig_size = zio->io_size = size;
888 * ==========================================================================
889 * Prepare to read and write logical blocks
890 * ==========================================================================
894 zio_read_bp_init(zio_t *zio)
896 blkptr_t *bp = zio->io_bp;
898 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
899 zio->io_child_type == ZIO_CHILD_LOGICAL &&
900 !(zio->io_flags & ZIO_FLAG_RAW)) {
901 uint64_t psize = BP_GET_PSIZE(bp);
902 void *cbuf = zio_buf_alloc(psize);
904 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
907 if (!dmu_ot[BP_GET_TYPE(bp)].ot_metadata && BP_GET_LEVEL(bp) == 0)
908 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
910 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
911 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
913 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
914 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
916 return (ZIO_PIPELINE_CONTINUE);
920 zio_write_bp_init(zio_t *zio)
922 spa_t *spa = zio->io_spa;
923 zio_prop_t *zp = &zio->io_prop;
924 enum zio_compress compress = zp->zp_compress;
925 blkptr_t *bp = zio->io_bp;
926 uint64_t lsize = zio->io_size;
927 uint64_t psize = lsize;
931 * If our children haven't all reached the ready stage,
932 * wait for them and then repeat this pipeline stage.
934 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
935 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
936 return (ZIO_PIPELINE_STOP);
938 if (!IO_IS_ALLOCATING(zio))
939 return (ZIO_PIPELINE_CONTINUE);
941 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
943 if (zio->io_bp_override) {
944 ASSERT(bp->blk_birth != zio->io_txg);
945 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
947 *bp = *zio->io_bp_override;
948 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
950 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
951 return (ZIO_PIPELINE_CONTINUE);
953 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
954 zp->zp_dedup_verify);
956 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
958 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
959 return (ZIO_PIPELINE_CONTINUE);
961 zio->io_bp_override = NULL;
965 if (bp->blk_birth == zio->io_txg) {
967 * We're rewriting an existing block, which means we're
968 * working on behalf of spa_sync(). For spa_sync() to
969 * converge, it must eventually be the case that we don't
970 * have to allocate new blocks. But compression changes
971 * the blocksize, which forces a reallocate, and makes
972 * convergence take longer. Therefore, after the first
973 * few passes, stop compressing to ensure convergence.
975 pass = spa_sync_pass(spa);
977 ASSERT(zio->io_txg == spa_syncing_txg(spa));
978 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
979 ASSERT(!BP_GET_DEDUP(bp));
981 if (pass > SYNC_PASS_DONT_COMPRESS)
982 compress = ZIO_COMPRESS_OFF;
984 /* Make sure someone doesn't change their mind on overwrites */
985 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
986 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
989 if (compress != ZIO_COMPRESS_OFF) {
990 void *cbuf = zio_buf_alloc(lsize);
991 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
992 if (psize == 0 || psize == lsize) {
993 compress = ZIO_COMPRESS_OFF;
994 zio_buf_free(cbuf, lsize);
996 ASSERT(psize < lsize);
997 zio_push_transform(zio, cbuf, psize, lsize, NULL);
1002 * The final pass of spa_sync() must be all rewrites, but the first
1003 * few passes offer a trade-off: allocating blocks defers convergence,
1004 * but newly allocated blocks are sequential, so they can be written
1005 * to disk faster. Therefore, we allow the first few passes of
1006 * spa_sync() to allocate new blocks, but force rewrites after that.
1007 * There should only be a handful of blocks after pass 1 in any case.
1009 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
1010 pass > SYNC_PASS_REWRITE) {
1011 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1013 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1014 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1017 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1021 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1023 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1024 BP_SET_LSIZE(bp, lsize);
1025 BP_SET_PSIZE(bp, psize);
1026 BP_SET_COMPRESS(bp, compress);
1027 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1028 BP_SET_TYPE(bp, zp->zp_type);
1029 BP_SET_LEVEL(bp, zp->zp_level);
1030 BP_SET_DEDUP(bp, zp->zp_dedup);
1031 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1033 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1034 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1035 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1039 return (ZIO_PIPELINE_CONTINUE);
1043 zio_free_bp_init(zio_t *zio)
1045 blkptr_t *bp = zio->io_bp;
1047 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1048 if (BP_GET_DEDUP(bp))
1049 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1052 return (ZIO_PIPELINE_CONTINUE);
1056 * ==========================================================================
1057 * Execute the I/O pipeline
1058 * ==========================================================================
1062 zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q, boolean_t cutinline)
1064 spa_t *spa = zio->io_spa;
1065 zio_type_t t = zio->io_type;
1066 int flags = TQ_NOSLEEP | (cutinline ? TQ_FRONT : 0);
1069 * If we're a config writer or a probe, the normal issue and
1070 * interrupt threads may all be blocked waiting for the config lock.
1071 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1073 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1077 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1079 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1083 * If this is a high priority I/O, then use the high priority taskq.
1085 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1086 spa->spa_zio_taskq[t][q + 1] != NULL)
1089 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1091 while (taskq_dispatch(spa->spa_zio_taskq[t][q],
1092 (task_func_t *)zio_execute, zio, flags) == 0); /* do nothing */
1096 zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
1098 kthread_t *executor = zio->io_executor;
1099 spa_t *spa = zio->io_spa;
1102 for (t = 0; t < ZIO_TYPES; t++)
1103 if (taskq_member(spa->spa_zio_taskq[t][q], executor))
1110 zio_issue_async(zio_t *zio)
1112 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1114 return (ZIO_PIPELINE_STOP);
1118 zio_interrupt(zio_t *zio)
1120 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1124 * Execute the I/O pipeline until one of the following occurs:
1125 * (1) the I/O completes; (2) the pipeline stalls waiting for
1126 * dependent child I/Os; (3) the I/O issues, so we're waiting
1127 * for an I/O completion interrupt; (4) the I/O is delegated by
1128 * vdev-level caching or aggregation; (5) the I/O is deferred
1129 * due to vdev-level queueing; (6) the I/O is handed off to
1130 * another thread. In all cases, the pipeline stops whenever
1131 * there's no CPU work; it never burns a thread in cv_wait().
1133 * There's no locking on io_stage because there's no legitimate way
1134 * for multiple threads to be attempting to process the same I/O.
1136 static zio_pipe_stage_t *zio_pipeline[];
1139 * zio_execute() is a wrapper around the static function
1140 * __zio_execute() so that we can force __zio_execute() to be
1141 * inlined. This reduces stack overhead which is important
1142 * because __zio_execute() is called recursively in several zio
1143 * code paths. zio_execute() itself cannot be inlined because
1144 * it is externally visible.
1147 zio_execute(zio_t *zio)
1152 __attribute__((always_inline))
1154 __zio_execute(zio_t *zio)
1156 zio->io_executor = curthread;
1158 while (zio->io_stage < ZIO_STAGE_DONE) {
1159 enum zio_stage pipeline = zio->io_pipeline;
1160 enum zio_stage stage = zio->io_stage;
1165 ASSERT(!MUTEX_HELD(&zio->io_lock));
1166 ASSERT(ISP2(stage));
1167 ASSERT(zio->io_stall == NULL);
1171 } while ((stage & pipeline) == 0);
1173 ASSERT(stage <= ZIO_STAGE_DONE);
1175 dsl = spa_get_dsl(zio->io_spa);
1176 cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1177 zio_requeue_io_start_cut_in_line : B_FALSE;
1180 * If we are in interrupt context and this pipeline stage
1181 * will grab a config lock that is held across I/O,
1182 * or may wait for an I/O that needs an interrupt thread
1183 * to complete, issue async to avoid deadlock.
1185 * If we are in the txg_sync_thread or being called
1186 * during pool init issue async to minimize stack depth.
1187 * Both of these call paths may be recursively called.
1189 * For VDEV_IO_START, we cut in line so that the io will
1190 * be sent to disk promptly.
1192 if (((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1193 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) ||
1194 (dsl != NULL && dsl_pool_sync_context(dsl))) {
1195 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1199 zio->io_stage = stage;
1200 rv = zio_pipeline[highbit(stage) - 1](zio);
1202 if (rv == ZIO_PIPELINE_STOP)
1205 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1211 * ==========================================================================
1212 * Initiate I/O, either sync or async
1213 * ==========================================================================
1216 zio_wait(zio_t *zio)
1220 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1221 ASSERT(zio->io_executor == NULL);
1223 zio->io_waiter = curthread;
1227 mutex_enter(&zio->io_lock);
1228 while (zio->io_executor != NULL)
1229 cv_wait(&zio->io_cv, &zio->io_lock);
1230 mutex_exit(&zio->io_lock);
1232 error = zio->io_error;
1239 zio_nowait(zio_t *zio)
1241 ASSERT(zio->io_executor == NULL);
1243 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1244 zio_unique_parent(zio) == NULL) {
1246 * This is a logical async I/O with no parent to wait for it.
1247 * We add it to the spa_async_root_zio "Godfather" I/O which
1248 * will ensure they complete prior to unloading the pool.
1250 spa_t *spa = zio->io_spa;
1252 zio_add_child(spa->spa_async_zio_root, zio);
1259 * ==========================================================================
1260 * Reexecute or suspend/resume failed I/O
1261 * ==========================================================================
1265 zio_reexecute(zio_t *pio)
1267 zio_t *cio, *cio_next;
1270 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1271 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1272 ASSERT(pio->io_gang_leader == NULL);
1273 ASSERT(pio->io_gang_tree == NULL);
1275 pio->io_flags = pio->io_orig_flags;
1276 pio->io_stage = pio->io_orig_stage;
1277 pio->io_pipeline = pio->io_orig_pipeline;
1278 pio->io_reexecute = 0;
1280 for (w = 0; w < ZIO_WAIT_TYPES; w++)
1281 pio->io_state[w] = 0;
1282 for (c = 0; c < ZIO_CHILD_TYPES; c++)
1283 pio->io_child_error[c] = 0;
1285 if (IO_IS_ALLOCATING(pio))
1286 BP_ZERO(pio->io_bp);
1289 * As we reexecute pio's children, new children could be created.
1290 * New children go to the head of pio's io_child_list, however,
1291 * so we will (correctly) not reexecute them. The key is that
1292 * the remainder of pio's io_child_list, from 'cio_next' onward,
1293 * cannot be affected by any side effects of reexecuting 'cio'.
1295 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1296 cio_next = zio_walk_children(pio);
1297 mutex_enter(&pio->io_lock);
1298 for (w = 0; w < ZIO_WAIT_TYPES; w++)
1299 pio->io_children[cio->io_child_type][w]++;
1300 mutex_exit(&pio->io_lock);
1305 * Now that all children have been reexecuted, execute the parent.
1306 * We don't reexecute "The Godfather" I/O here as it's the
1307 * responsibility of the caller to wait on him.
1309 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1314 zio_suspend(spa_t *spa, zio_t *zio)
1316 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1317 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1318 "failure and the failure mode property for this pool "
1319 "is set to panic.", spa_name(spa));
1321 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1323 mutex_enter(&spa->spa_suspend_lock);
1325 if (spa->spa_suspend_zio_root == NULL)
1326 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1327 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1328 ZIO_FLAG_GODFATHER);
1330 spa->spa_suspended = B_TRUE;
1333 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1334 ASSERT(zio != spa->spa_suspend_zio_root);
1335 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1336 ASSERT(zio_unique_parent(zio) == NULL);
1337 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1338 zio_add_child(spa->spa_suspend_zio_root, zio);
1341 mutex_exit(&spa->spa_suspend_lock);
1345 zio_resume(spa_t *spa)
1350 * Reexecute all previously suspended i/o.
1352 mutex_enter(&spa->spa_suspend_lock);
1353 spa->spa_suspended = B_FALSE;
1354 cv_broadcast(&spa->spa_suspend_cv);
1355 pio = spa->spa_suspend_zio_root;
1356 spa->spa_suspend_zio_root = NULL;
1357 mutex_exit(&spa->spa_suspend_lock);
1363 return (zio_wait(pio));
1367 zio_resume_wait(spa_t *spa)
1369 mutex_enter(&spa->spa_suspend_lock);
1370 while (spa_suspended(spa))
1371 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1372 mutex_exit(&spa->spa_suspend_lock);
1376 * ==========================================================================
1379 * A gang block is a collection of small blocks that looks to the DMU
1380 * like one large block. When zio_dva_allocate() cannot find a block
1381 * of the requested size, due to either severe fragmentation or the pool
1382 * being nearly full, it calls zio_write_gang_block() to construct the
1383 * block from smaller fragments.
1385 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1386 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1387 * an indirect block: it's an array of block pointers. It consumes
1388 * only one sector and hence is allocatable regardless of fragmentation.
1389 * The gang header's bps point to its gang members, which hold the data.
1391 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1392 * as the verifier to ensure uniqueness of the SHA256 checksum.
1393 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1394 * not the gang header. This ensures that data block signatures (needed for
1395 * deduplication) are independent of how the block is physically stored.
1397 * Gang blocks can be nested: a gang member may itself be a gang block.
1398 * Thus every gang block is a tree in which root and all interior nodes are
1399 * gang headers, and the leaves are normal blocks that contain user data.
1400 * The root of the gang tree is called the gang leader.
1402 * To perform any operation (read, rewrite, free, claim) on a gang block,
1403 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1404 * in the io_gang_tree field of the original logical i/o by recursively
1405 * reading the gang leader and all gang headers below it. This yields
1406 * an in-core tree containing the contents of every gang header and the
1407 * bps for every constituent of the gang block.
1409 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1410 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1411 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1412 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1413 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1414 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1415 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1416 * of the gang header plus zio_checksum_compute() of the data to update the
1417 * gang header's blk_cksum as described above.
1419 * The two-phase assemble/issue model solves the problem of partial failure --
1420 * what if you'd freed part of a gang block but then couldn't read the
1421 * gang header for another part? Assembling the entire gang tree first
1422 * ensures that all the necessary gang header I/O has succeeded before
1423 * starting the actual work of free, claim, or write. Once the gang tree
1424 * is assembled, free and claim are in-memory operations that cannot fail.
1426 * In the event that a gang write fails, zio_dva_unallocate() walks the
1427 * gang tree to immediately free (i.e. insert back into the space map)
1428 * everything we've allocated. This ensures that we don't get ENOSPC
1429 * errors during repeated suspend/resume cycles due to a flaky device.
1431 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1432 * the gang tree, we won't modify the block, so we can safely defer the free
1433 * (knowing that the block is still intact). If we *can* assemble the gang
1434 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1435 * each constituent bp and we can allocate a new block on the next sync pass.
1437 * In all cases, the gang tree allows complete recovery from partial failure.
1438 * ==========================================================================
1442 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1447 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1448 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1449 &pio->io_bookmark));
1453 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1458 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1459 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1460 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1462 * As we rewrite each gang header, the pipeline will compute
1463 * a new gang block header checksum for it; but no one will
1464 * compute a new data checksum, so we do that here. The one
1465 * exception is the gang leader: the pipeline already computed
1466 * its data checksum because that stage precedes gang assembly.
1467 * (Presently, nothing actually uses interior data checksums;
1468 * this is just good hygiene.)
1470 if (gn != pio->io_gang_leader->io_gang_tree) {
1471 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1472 data, BP_GET_PSIZE(bp));
1475 * If we are here to damage data for testing purposes,
1476 * leave the GBH alone so that we can detect the damage.
1478 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1479 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1481 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1482 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1483 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1491 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1493 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1494 ZIO_GANG_CHILD_FLAGS(pio)));
1499 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1501 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1502 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1505 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1514 static void zio_gang_tree_assemble_done(zio_t *zio);
1516 static zio_gang_node_t *
1517 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1519 zio_gang_node_t *gn;
1521 ASSERT(*gnpp == NULL);
1523 gn = kmem_zalloc(sizeof (*gn), KM_PUSHPAGE);
1524 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1531 zio_gang_node_free(zio_gang_node_t **gnpp)
1533 zio_gang_node_t *gn = *gnpp;
1536 for (g = 0; g < SPA_GBH_NBLKPTRS; g++)
1537 ASSERT(gn->gn_child[g] == NULL);
1539 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1540 kmem_free(gn, sizeof (*gn));
1545 zio_gang_tree_free(zio_gang_node_t **gnpp)
1547 zio_gang_node_t *gn = *gnpp;
1553 for (g = 0; g < SPA_GBH_NBLKPTRS; g++)
1554 zio_gang_tree_free(&gn->gn_child[g]);
1556 zio_gang_node_free(gnpp);
1560 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1562 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1564 ASSERT(gio->io_gang_leader == gio);
1565 ASSERT(BP_IS_GANG(bp));
1567 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1568 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1569 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1573 zio_gang_tree_assemble_done(zio_t *zio)
1575 zio_t *gio = zio->io_gang_leader;
1576 zio_gang_node_t *gn = zio->io_private;
1577 blkptr_t *bp = zio->io_bp;
1580 ASSERT(gio == zio_unique_parent(zio));
1581 ASSERT(zio->io_child_count == 0);
1586 if (BP_SHOULD_BYTESWAP(bp))
1587 byteswap_uint64_array(zio->io_data, zio->io_size);
1589 ASSERT(zio->io_data == gn->gn_gbh);
1590 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1591 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1593 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1594 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1595 if (!BP_IS_GANG(gbp))
1597 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1602 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1604 zio_t *gio = pio->io_gang_leader;
1608 ASSERT(BP_IS_GANG(bp) == !!gn);
1609 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1610 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1613 * If you're a gang header, your data is in gn->gn_gbh.
1614 * If you're a gang member, your data is in 'data' and gn == NULL.
1616 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1619 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1621 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1622 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1623 if (BP_IS_HOLE(gbp))
1625 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1626 data = (char *)data + BP_GET_PSIZE(gbp);
1630 if (gn == gio->io_gang_tree)
1631 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1638 zio_gang_assemble(zio_t *zio)
1640 blkptr_t *bp = zio->io_bp;
1642 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1643 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1645 zio->io_gang_leader = zio;
1647 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1649 return (ZIO_PIPELINE_CONTINUE);
1653 zio_gang_issue(zio_t *zio)
1655 blkptr_t *bp = zio->io_bp;
1657 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1658 return (ZIO_PIPELINE_STOP);
1660 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1661 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1663 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1664 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1666 zio_gang_tree_free(&zio->io_gang_tree);
1668 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1670 return (ZIO_PIPELINE_CONTINUE);
1674 zio_write_gang_member_ready(zio_t *zio)
1676 zio_t *pio = zio_unique_parent(zio);
1677 ASSERTV(zio_t *gio = zio->io_gang_leader;)
1678 dva_t *cdva = zio->io_bp->blk_dva;
1679 dva_t *pdva = pio->io_bp->blk_dva;
1683 if (BP_IS_HOLE(zio->io_bp))
1686 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1688 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1689 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1690 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1691 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1692 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1694 mutex_enter(&pio->io_lock);
1695 for (d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1696 ASSERT(DVA_GET_GANG(&pdva[d]));
1697 asize = DVA_GET_ASIZE(&pdva[d]);
1698 asize += DVA_GET_ASIZE(&cdva[d]);
1699 DVA_SET_ASIZE(&pdva[d], asize);
1701 mutex_exit(&pio->io_lock);
1705 zio_write_gang_block(zio_t *pio)
1707 spa_t *spa = pio->io_spa;
1708 blkptr_t *bp = pio->io_bp;
1709 zio_t *gio = pio->io_gang_leader;
1711 zio_gang_node_t *gn, **gnpp;
1712 zio_gbh_phys_t *gbh;
1713 uint64_t txg = pio->io_txg;
1714 uint64_t resid = pio->io_size;
1716 int copies = gio->io_prop.zp_copies;
1717 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1721 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1722 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1723 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1725 pio->io_error = error;
1726 return (ZIO_PIPELINE_CONTINUE);
1730 gnpp = &gio->io_gang_tree;
1732 gnpp = pio->io_private;
1733 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1736 gn = zio_gang_node_alloc(gnpp);
1738 bzero(gbh, SPA_GANGBLOCKSIZE);
1741 * Create the gang header.
1743 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1744 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1747 * Create and nowait the gang children.
1749 for (g = 0; resid != 0; resid -= lsize, g++) {
1750 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1752 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1754 zp.zp_checksum = gio->io_prop.zp_checksum;
1755 zp.zp_compress = ZIO_COMPRESS_OFF;
1756 zp.zp_type = DMU_OT_NONE;
1758 zp.zp_copies = gio->io_prop.zp_copies;
1760 zp.zp_dedup_verify = 0;
1762 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1763 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1764 zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1765 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1766 &pio->io_bookmark));
1770 * Set pio's pipeline to just wait for zio to finish.
1772 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1776 return (ZIO_PIPELINE_CONTINUE);
1780 * ==========================================================================
1782 * ==========================================================================
1785 zio_ddt_child_read_done(zio_t *zio)
1787 blkptr_t *bp = zio->io_bp;
1788 ddt_entry_t *dde = zio->io_private;
1790 zio_t *pio = zio_unique_parent(zio);
1792 mutex_enter(&pio->io_lock);
1793 ddp = ddt_phys_select(dde, bp);
1794 if (zio->io_error == 0)
1795 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1796 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1797 dde->dde_repair_data = zio->io_data;
1799 zio_buf_free(zio->io_data, zio->io_size);
1800 mutex_exit(&pio->io_lock);
1804 zio_ddt_read_start(zio_t *zio)
1806 blkptr_t *bp = zio->io_bp;
1809 ASSERT(BP_GET_DEDUP(bp));
1810 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1811 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1813 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1814 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1815 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1816 ddt_phys_t *ddp = dde->dde_phys;
1817 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1820 ASSERT(zio->io_vsd == NULL);
1823 if (ddp_self == NULL)
1824 return (ZIO_PIPELINE_CONTINUE);
1826 for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1827 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1829 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1831 zio_nowait(zio_read(zio, zio->io_spa, &blk,
1832 zio_buf_alloc(zio->io_size), zio->io_size,
1833 zio_ddt_child_read_done, dde, zio->io_priority,
1834 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1835 &zio->io_bookmark));
1837 return (ZIO_PIPELINE_CONTINUE);
1840 zio_nowait(zio_read(zio, zio->io_spa, bp,
1841 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1842 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1844 return (ZIO_PIPELINE_CONTINUE);
1848 zio_ddt_read_done(zio_t *zio)
1850 blkptr_t *bp = zio->io_bp;
1852 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1853 return (ZIO_PIPELINE_STOP);
1855 ASSERT(BP_GET_DEDUP(bp));
1856 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1857 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1859 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1860 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1861 ddt_entry_t *dde = zio->io_vsd;
1863 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
1864 return (ZIO_PIPELINE_CONTINUE);
1867 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
1868 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1869 return (ZIO_PIPELINE_STOP);
1871 if (dde->dde_repair_data != NULL) {
1872 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
1873 zio->io_child_error[ZIO_CHILD_DDT] = 0;
1875 ddt_repair_done(ddt, dde);
1879 ASSERT(zio->io_vsd == NULL);
1881 return (ZIO_PIPELINE_CONTINUE);
1885 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
1887 spa_t *spa = zio->io_spa;
1891 * Note: we compare the original data, not the transformed data,
1892 * because when zio->io_bp is an override bp, we will not have
1893 * pushed the I/O transforms. That's an important optimization
1894 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
1896 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1897 zio_t *lio = dde->dde_lead_zio[p];
1900 return (lio->io_orig_size != zio->io_orig_size ||
1901 bcmp(zio->io_orig_data, lio->io_orig_data,
1902 zio->io_orig_size) != 0);
1906 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1907 ddt_phys_t *ddp = &dde->dde_phys[p];
1909 if (ddp->ddp_phys_birth != 0) {
1910 arc_buf_t *abuf = NULL;
1911 uint32_t aflags = ARC_WAIT;
1912 blkptr_t blk = *zio->io_bp;
1915 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
1919 error = arc_read_nolock(NULL, spa, &blk,
1920 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
1921 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1922 &aflags, &zio->io_bookmark);
1925 if (arc_buf_size(abuf) != zio->io_orig_size ||
1926 bcmp(abuf->b_data, zio->io_orig_data,
1927 zio->io_orig_size) != 0)
1929 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
1933 return (error != 0);
1941 zio_ddt_child_write_ready(zio_t *zio)
1943 int p = zio->io_prop.zp_copies;
1944 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1945 ddt_entry_t *dde = zio->io_private;
1946 ddt_phys_t *ddp = &dde->dde_phys[p];
1954 ASSERT(dde->dde_lead_zio[p] == zio);
1956 ddt_phys_fill(ddp, zio->io_bp);
1958 while ((pio = zio_walk_parents(zio)) != NULL)
1959 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
1965 zio_ddt_child_write_done(zio_t *zio)
1967 int p = zio->io_prop.zp_copies;
1968 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1969 ddt_entry_t *dde = zio->io_private;
1970 ddt_phys_t *ddp = &dde->dde_phys[p];
1974 ASSERT(ddp->ddp_refcnt == 0);
1975 ASSERT(dde->dde_lead_zio[p] == zio);
1976 dde->dde_lead_zio[p] = NULL;
1978 if (zio->io_error == 0) {
1979 while (zio_walk_parents(zio) != NULL)
1980 ddt_phys_addref(ddp);
1982 ddt_phys_clear(ddp);
1989 zio_ddt_ditto_write_done(zio_t *zio)
1991 int p = DDT_PHYS_DITTO;
1992 blkptr_t *bp = zio->io_bp;
1993 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1994 ddt_entry_t *dde = zio->io_private;
1995 ddt_phys_t *ddp = &dde->dde_phys[p];
1996 ddt_key_t *ddk = &dde->dde_key;
1997 ASSERTV(zio_prop_t *zp = &zio->io_prop);
2001 ASSERT(ddp->ddp_refcnt == 0);
2002 ASSERT(dde->dde_lead_zio[p] == zio);
2003 dde->dde_lead_zio[p] = NULL;
2005 if (zio->io_error == 0) {
2006 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2007 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2008 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2009 if (ddp->ddp_phys_birth != 0)
2010 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2011 ddt_phys_fill(ddp, bp);
2018 zio_ddt_write(zio_t *zio)
2020 spa_t *spa = zio->io_spa;
2021 blkptr_t *bp = zio->io_bp;
2022 uint64_t txg = zio->io_txg;
2023 zio_prop_t *zp = &zio->io_prop;
2024 int p = zp->zp_copies;
2028 ddt_t *ddt = ddt_select(spa, bp);
2032 ASSERT(BP_GET_DEDUP(bp));
2033 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2034 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2037 dde = ddt_lookup(ddt, bp, B_TRUE);
2038 ddp = &dde->dde_phys[p];
2040 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2042 * If we're using a weak checksum, upgrade to a strong checksum
2043 * and try again. If we're already using a strong checksum,
2044 * we can't resolve it, so just convert to an ordinary write.
2045 * (And automatically e-mail a paper to Nature?)
2047 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2048 zp->zp_checksum = spa_dedup_checksum(spa);
2049 zio_pop_transforms(zio);
2050 zio->io_stage = ZIO_STAGE_OPEN;
2055 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2057 return (ZIO_PIPELINE_CONTINUE);
2060 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2061 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2063 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2064 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2065 zio_prop_t czp = *zp;
2067 czp.zp_copies = ditto_copies;
2070 * If we arrived here with an override bp, we won't have run
2071 * the transform stack, so we won't have the data we need to
2072 * generate a child i/o. So, toss the override bp and restart.
2073 * This is safe, because using the override bp is just an
2074 * optimization; and it's rare, so the cost doesn't matter.
2076 if (zio->io_bp_override) {
2077 zio_pop_transforms(zio);
2078 zio->io_stage = ZIO_STAGE_OPEN;
2079 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2080 zio->io_bp_override = NULL;
2083 return (ZIO_PIPELINE_CONTINUE);
2086 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2087 zio->io_orig_size, &czp, NULL,
2088 zio_ddt_ditto_write_done, dde, zio->io_priority,
2089 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2091 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2092 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2095 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2096 if (ddp->ddp_phys_birth != 0)
2097 ddt_bp_fill(ddp, bp, txg);
2098 if (dde->dde_lead_zio[p] != NULL)
2099 zio_add_child(zio, dde->dde_lead_zio[p]);
2101 ddt_phys_addref(ddp);
2102 } else if (zio->io_bp_override) {
2103 ASSERT(bp->blk_birth == txg);
2104 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2105 ddt_phys_fill(ddp, bp);
2106 ddt_phys_addref(ddp);
2108 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2109 zio->io_orig_size, zp, zio_ddt_child_write_ready,
2110 zio_ddt_child_write_done, dde, zio->io_priority,
2111 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2113 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2114 dde->dde_lead_zio[p] = cio;
2124 return (ZIO_PIPELINE_CONTINUE);
2127 ddt_entry_t *freedde; /* for debugging */
2130 zio_ddt_free(zio_t *zio)
2132 spa_t *spa = zio->io_spa;
2133 blkptr_t *bp = zio->io_bp;
2134 ddt_t *ddt = ddt_select(spa, bp);
2138 ASSERT(BP_GET_DEDUP(bp));
2139 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2142 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2143 ddp = ddt_phys_select(dde, bp);
2144 ddt_phys_decref(ddp);
2147 return (ZIO_PIPELINE_CONTINUE);
2151 * ==========================================================================
2152 * Allocate and free blocks
2153 * ==========================================================================
2156 zio_dva_allocate(zio_t *zio)
2158 spa_t *spa = zio->io_spa;
2159 metaslab_class_t *mc = spa_normal_class(spa);
2160 blkptr_t *bp = zio->io_bp;
2164 if (zio->io_gang_leader == NULL) {
2165 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2166 zio->io_gang_leader = zio;
2169 ASSERT(BP_IS_HOLE(bp));
2170 ASSERT3U(BP_GET_NDVAS(bp), ==, 0);
2171 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2172 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2173 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2176 * The dump device does not support gang blocks so allocation on
2177 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2178 * the "fast" gang feature.
2180 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2181 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2182 METASLAB_GANG_CHILD : 0;
2183 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2184 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2187 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2188 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2190 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2191 return (zio_write_gang_block(zio));
2192 zio->io_error = error;
2195 return (ZIO_PIPELINE_CONTINUE);
2199 zio_dva_free(zio_t *zio)
2201 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2203 return (ZIO_PIPELINE_CONTINUE);
2207 zio_dva_claim(zio_t *zio)
2211 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2213 zio->io_error = error;
2215 return (ZIO_PIPELINE_CONTINUE);
2219 * Undo an allocation. This is used by zio_done() when an I/O fails
2220 * and we want to give back the block we just allocated.
2221 * This handles both normal blocks and gang blocks.
2224 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2228 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2229 ASSERT(zio->io_bp_override == NULL);
2231 if (!BP_IS_HOLE(bp))
2232 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2235 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2236 zio_dva_unallocate(zio, gn->gn_child[g],
2237 &gn->gn_gbh->zg_blkptr[g]);
2243 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2246 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2247 uint64_t size, boolean_t use_slog)
2251 ASSERT(txg > spa_syncing_txg(spa));
2254 error = metaslab_alloc(spa, spa_log_class(spa), size,
2255 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2258 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2259 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2262 BP_SET_LSIZE(new_bp, size);
2263 BP_SET_PSIZE(new_bp, size);
2264 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2265 BP_SET_CHECKSUM(new_bp,
2266 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2267 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2268 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2269 BP_SET_LEVEL(new_bp, 0);
2270 BP_SET_DEDUP(new_bp, 0);
2271 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2278 * Free an intent log block.
2281 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2283 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2284 ASSERT(!BP_IS_GANG(bp));
2286 zio_free(spa, txg, bp);
2290 * ==========================================================================
2291 * Read and write to physical devices
2292 * ==========================================================================
2295 zio_vdev_io_start(zio_t *zio)
2297 vdev_t *vd = zio->io_vd;
2299 spa_t *spa = zio->io_spa;
2301 ASSERT(zio->io_error == 0);
2302 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2305 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2306 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2309 * The mirror_ops handle multiple DVAs in a single BP.
2311 return (vdev_mirror_ops.vdev_op_io_start(zio));
2315 * We keep track of time-sensitive I/Os so that the scan thread
2316 * can quickly react to certain workloads. In particular, we care
2317 * about non-scrubbing, top-level reads and writes with the following
2319 * - synchronous writes of user data to non-slog devices
2320 * - any reads of user data
2321 * When these conditions are met, adjust the timestamp of spa_last_io
2322 * which allows the scan thread to adjust its workload accordingly.
2324 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2325 vd == vd->vdev_top && !vd->vdev_islog &&
2326 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2327 zio->io_txg != spa_syncing_txg(spa)) {
2328 uint64_t old = spa->spa_last_io;
2329 uint64_t new = ddi_get_lbolt64();
2331 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2334 align = 1ULL << vd->vdev_top->vdev_ashift;
2336 if (P2PHASE(zio->io_size, align) != 0) {
2337 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2338 char *abuf = zio_buf_alloc(asize);
2339 ASSERT(vd == vd->vdev_top);
2340 if (zio->io_type == ZIO_TYPE_WRITE) {
2341 bcopy(zio->io_data, abuf, zio->io_size);
2342 bzero(abuf + zio->io_size, asize - zio->io_size);
2344 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2347 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2348 ASSERT(P2PHASE(zio->io_size, align) == 0);
2349 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2352 * If this is a repair I/O, and there's no self-healing involved --
2353 * that is, we're just resilvering what we expect to resilver --
2354 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2355 * This prevents spurious resilvering with nested replication.
2356 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2357 * A is out of date, we'll read from C+D, then use the data to
2358 * resilver A+B -- but we don't actually want to resilver B, just A.
2359 * The top-level mirror has no way to know this, so instead we just
2360 * discard unnecessary repairs as we work our way down the vdev tree.
2361 * The same logic applies to any form of nested replication:
2362 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2364 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2365 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2366 zio->io_txg != 0 && /* not a delegated i/o */
2367 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2368 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2369 zio_vdev_io_bypass(zio);
2370 return (ZIO_PIPELINE_CONTINUE);
2373 if (vd->vdev_ops->vdev_op_leaf &&
2374 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2376 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2377 return (ZIO_PIPELINE_CONTINUE);
2379 if ((zio = vdev_queue_io(zio)) == NULL)
2380 return (ZIO_PIPELINE_STOP);
2382 if (!vdev_accessible(vd, zio)) {
2383 zio->io_error = ENXIO;
2385 return (ZIO_PIPELINE_STOP);
2389 return (vd->vdev_ops->vdev_op_io_start(zio));
2393 zio_vdev_io_done(zio_t *zio)
2395 vdev_t *vd = zio->io_vd;
2396 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2397 boolean_t unexpected_error = B_FALSE;
2399 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2400 return (ZIO_PIPELINE_STOP);
2402 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2404 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2406 vdev_queue_io_done(zio);
2408 if (zio->io_type == ZIO_TYPE_WRITE)
2409 vdev_cache_write(zio);
2411 if (zio_injection_enabled && zio->io_error == 0)
2412 zio->io_error = zio_handle_device_injection(vd,
2415 if (zio_injection_enabled && zio->io_error == 0)
2416 zio->io_error = zio_handle_label_injection(zio, EIO);
2418 if (zio->io_error) {
2419 if (!vdev_accessible(vd, zio)) {
2420 zio->io_error = ENXIO;
2422 unexpected_error = B_TRUE;
2427 ops->vdev_op_io_done(zio);
2429 if (unexpected_error)
2430 VERIFY(vdev_probe(vd, zio) == NULL);
2432 return (ZIO_PIPELINE_CONTINUE);
2436 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2437 * disk, and use that to finish the checksum ereport later.
2440 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2441 const void *good_buf)
2443 /* no processing needed */
2444 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2449 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2451 void *buf = zio_buf_alloc(zio->io_size);
2453 bcopy(zio->io_data, buf, zio->io_size);
2455 zcr->zcr_cbinfo = zio->io_size;
2456 zcr->zcr_cbdata = buf;
2457 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2458 zcr->zcr_free = zio_buf_free;
2462 zio_vdev_io_assess(zio_t *zio)
2464 vdev_t *vd = zio->io_vd;
2466 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2467 return (ZIO_PIPELINE_STOP);
2469 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2470 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2472 if (zio->io_vsd != NULL) {
2473 zio->io_vsd_ops->vsd_free(zio);
2477 if (zio_injection_enabled && zio->io_error == 0)
2478 zio->io_error = zio_handle_fault_injection(zio, EIO);
2481 * If the I/O failed, determine whether we should attempt to retry it.
2483 * On retry, we cut in line in the issue queue, since we don't want
2484 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2486 if (zio->io_error && vd == NULL &&
2487 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2488 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2489 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2491 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2492 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2493 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2494 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2495 zio_requeue_io_start_cut_in_line);
2496 return (ZIO_PIPELINE_STOP);
2500 * If we got an error on a leaf device, convert it to ENXIO
2501 * if the device is not accessible at all.
2503 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2504 !vdev_accessible(vd, zio))
2505 zio->io_error = ENXIO;
2508 * If we can't write to an interior vdev (mirror or RAID-Z),
2509 * set vdev_cant_write so that we stop trying to allocate from it.
2511 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2512 vd != NULL && !vd->vdev_ops->vdev_op_leaf)
2513 vd->vdev_cant_write = B_TRUE;
2516 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2518 return (ZIO_PIPELINE_CONTINUE);
2522 zio_vdev_io_reissue(zio_t *zio)
2524 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2525 ASSERT(zio->io_error == 0);
2527 zio->io_stage >>= 1;
2531 zio_vdev_io_redone(zio_t *zio)
2533 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2535 zio->io_stage >>= 1;
2539 zio_vdev_io_bypass(zio_t *zio)
2541 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2542 ASSERT(zio->io_error == 0);
2544 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2545 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2549 * ==========================================================================
2550 * Generate and verify checksums
2551 * ==========================================================================
2554 zio_checksum_generate(zio_t *zio)
2556 blkptr_t *bp = zio->io_bp;
2557 enum zio_checksum checksum;
2561 * This is zio_write_phys().
2562 * We're either generating a label checksum, or none at all.
2564 checksum = zio->io_prop.zp_checksum;
2566 if (checksum == ZIO_CHECKSUM_OFF)
2567 return (ZIO_PIPELINE_CONTINUE);
2569 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2571 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2572 ASSERT(!IO_IS_ALLOCATING(zio));
2573 checksum = ZIO_CHECKSUM_GANG_HEADER;
2575 checksum = BP_GET_CHECKSUM(bp);
2579 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2581 return (ZIO_PIPELINE_CONTINUE);
2585 zio_checksum_verify(zio_t *zio)
2587 zio_bad_cksum_t info;
2588 blkptr_t *bp = zio->io_bp;
2591 ASSERT(zio->io_vd != NULL);
2595 * This is zio_read_phys().
2596 * We're either verifying a label checksum, or nothing at all.
2598 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2599 return (ZIO_PIPELINE_CONTINUE);
2601 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2604 if ((error = zio_checksum_error(zio, &info)) != 0) {
2605 zio->io_error = error;
2606 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2607 zfs_ereport_start_checksum(zio->io_spa,
2608 zio->io_vd, zio, zio->io_offset,
2609 zio->io_size, NULL, &info);
2613 return (ZIO_PIPELINE_CONTINUE);
2617 * Called by RAID-Z to ensure we don't compute the checksum twice.
2620 zio_checksum_verified(zio_t *zio)
2622 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2626 * ==========================================================================
2627 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2628 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2629 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2630 * indicate errors that are specific to one I/O, and most likely permanent.
2631 * Any other error is presumed to be worse because we weren't expecting it.
2632 * ==========================================================================
2635 zio_worst_error(int e1, int e2)
2637 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2640 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2641 if (e1 == zio_error_rank[r1])
2644 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2645 if (e2 == zio_error_rank[r2])
2648 return (r1 > r2 ? e1 : e2);
2652 * ==========================================================================
2654 * ==========================================================================
2657 zio_ready(zio_t *zio)
2659 blkptr_t *bp = zio->io_bp;
2660 zio_t *pio, *pio_next;
2662 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2663 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2664 return (ZIO_PIPELINE_STOP);
2666 if (zio->io_ready) {
2667 ASSERT(IO_IS_ALLOCATING(zio));
2668 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2669 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2674 if (bp != NULL && bp != &zio->io_bp_copy)
2675 zio->io_bp_copy = *bp;
2678 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2680 mutex_enter(&zio->io_lock);
2681 zio->io_state[ZIO_WAIT_READY] = 1;
2682 pio = zio_walk_parents(zio);
2683 mutex_exit(&zio->io_lock);
2686 * As we notify zio's parents, new parents could be added.
2687 * New parents go to the head of zio's io_parent_list, however,
2688 * so we will (correctly) not notify them. The remainder of zio's
2689 * io_parent_list, from 'pio_next' onward, cannot change because
2690 * all parents must wait for us to be done before they can be done.
2692 for (; pio != NULL; pio = pio_next) {
2693 pio_next = zio_walk_parents(zio);
2694 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2697 if (zio->io_flags & ZIO_FLAG_NODATA) {
2698 if (BP_IS_GANG(bp)) {
2699 zio->io_flags &= ~ZIO_FLAG_NODATA;
2701 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2702 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2706 if (zio_injection_enabled &&
2707 zio->io_spa->spa_syncing_txg == zio->io_txg)
2708 zio_handle_ignored_writes(zio);
2710 return (ZIO_PIPELINE_CONTINUE);
2714 zio_done(zio_t *zio)
2716 zio_t *pio, *pio_next;
2720 * If our children haven't all completed,
2721 * wait for them and then repeat this pipeline stage.
2723 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2724 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2725 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2726 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2727 return (ZIO_PIPELINE_STOP);
2729 for (c = 0; c < ZIO_CHILD_TYPES; c++)
2730 for (w = 0; w < ZIO_WAIT_TYPES; w++)
2731 ASSERT(zio->io_children[c][w] == 0);
2733 if (zio->io_bp != NULL) {
2734 ASSERT(zio->io_bp->blk_pad[0] == 0);
2735 ASSERT(zio->io_bp->blk_pad[1] == 0);
2736 ASSERT(bcmp(zio->io_bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2737 (zio->io_bp == zio_unique_parent(zio)->io_bp));
2738 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) &&
2739 zio->io_bp_override == NULL &&
2740 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2741 ASSERT(!BP_SHOULD_BYTESWAP(zio->io_bp));
2742 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2743 ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 ||
2744 (BP_COUNT_GANG(zio->io_bp) == BP_GET_NDVAS(zio->io_bp)));
2749 * If there were child vdev/gang/ddt errors, they apply to us now.
2751 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2752 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2753 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2756 * If the I/O on the transformed data was successful, generate any
2757 * checksum reports now while we still have the transformed data.
2759 if (zio->io_error == 0) {
2760 while (zio->io_cksum_report != NULL) {
2761 zio_cksum_report_t *zcr = zio->io_cksum_report;
2762 uint64_t align = zcr->zcr_align;
2763 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2764 char *abuf = zio->io_data;
2766 if (asize != zio->io_size) {
2767 abuf = zio_buf_alloc(asize);
2768 bcopy(zio->io_data, abuf, zio->io_size);
2769 bzero(abuf + zio->io_size, asize - zio->io_size);
2772 zio->io_cksum_report = zcr->zcr_next;
2773 zcr->zcr_next = NULL;
2774 zcr->zcr_finish(zcr, abuf);
2775 zfs_ereport_free_checksum(zcr);
2777 if (asize != zio->io_size)
2778 zio_buf_free(abuf, asize);
2782 zio_pop_transforms(zio); /* note: may set zio->io_error */
2784 vdev_stat_update(zio, zio->io_size);
2787 * If this I/O is attached to a particular vdev is slow, exeeding
2788 * 30 seconds to complete, post an error described the I/O delay.
2789 * We ignore these errors if the device is currently unavailable.
2791 if (zio->io_delay >= zio_delay_max) {
2792 if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd))
2793 zfs_ereport_post(FM_EREPORT_ZFS_DELAY, zio->io_spa,
2794 zio->io_vd, zio, 0, 0);
2797 if (zio->io_error) {
2799 * If this I/O is attached to a particular vdev,
2800 * generate an error message describing the I/O failure
2801 * at the block level. We ignore these errors if the
2802 * device is currently unavailable.
2804 if (zio->io_error != ECKSUM && zio->io_vd != NULL &&
2805 !vdev_is_dead(zio->io_vd))
2806 zfs_ereport_post(FM_EREPORT_ZFS_IO, zio->io_spa,
2807 zio->io_vd, zio, 0, 0);
2809 if ((zio->io_error == EIO || !(zio->io_flags &
2810 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2811 zio == zio->io_logical) {
2813 * For logical I/O requests, tell the SPA to log the
2814 * error and generate a logical data ereport.
2816 spa_log_error(zio->io_spa, zio);
2817 zfs_ereport_post(FM_EREPORT_ZFS_DATA, zio->io_spa, NULL, zio,
2822 if (zio->io_error && zio == zio->io_logical) {
2824 * Determine whether zio should be reexecuted. This will
2825 * propagate all the way to the root via zio_notify_parent().
2827 ASSERT(zio->io_vd == NULL && zio->io_bp != NULL);
2828 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2830 if (IO_IS_ALLOCATING(zio) &&
2831 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2832 if (zio->io_error != ENOSPC)
2833 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2835 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2838 if ((zio->io_type == ZIO_TYPE_READ ||
2839 zio->io_type == ZIO_TYPE_FREE) &&
2840 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
2841 zio->io_error == ENXIO &&
2842 spa_load_state(zio->io_spa) == SPA_LOAD_NONE &&
2843 spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE)
2844 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2846 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2847 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2850 * Here is a possibly good place to attempt to do
2851 * either combinatorial reconstruction or error correction
2852 * based on checksums. It also might be a good place
2853 * to send out preliminary ereports before we suspend
2859 * If there were logical child errors, they apply to us now.
2860 * We defer this until now to avoid conflating logical child
2861 * errors with errors that happened to the zio itself when
2862 * updating vdev stats and reporting FMA events above.
2864 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2866 if ((zio->io_error || zio->io_reexecute) &&
2867 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
2868 !(zio->io_flags & ZIO_FLAG_IO_REWRITE))
2869 zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp);
2871 zio_gang_tree_free(&zio->io_gang_tree);
2874 * Godfather I/Os should never suspend.
2876 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
2877 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
2878 zio->io_reexecute = 0;
2880 if (zio->io_reexecute) {
2882 * This is a logical I/O that wants to reexecute.
2884 * Reexecute is top-down. When an i/o fails, if it's not
2885 * the root, it simply notifies its parent and sticks around.
2886 * The parent, seeing that it still has children in zio_done(),
2887 * does the same. This percolates all the way up to the root.
2888 * The root i/o will reexecute or suspend the entire tree.
2890 * This approach ensures that zio_reexecute() honors
2891 * all the original i/o dependency relationships, e.g.
2892 * parents not executing until children are ready.
2894 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2896 zio->io_gang_leader = NULL;
2898 mutex_enter(&zio->io_lock);
2899 zio->io_state[ZIO_WAIT_DONE] = 1;
2900 mutex_exit(&zio->io_lock);
2903 * "The Godfather" I/O monitors its children but is
2904 * not a true parent to them. It will track them through
2905 * the pipeline but severs its ties whenever they get into
2906 * trouble (e.g. suspended). This allows "The Godfather"
2907 * I/O to return status without blocking.
2909 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2910 zio_link_t *zl = zio->io_walk_link;
2911 pio_next = zio_walk_parents(zio);
2913 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
2914 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
2915 zio_remove_child(pio, zio, zl);
2916 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2920 if ((pio = zio_unique_parent(zio)) != NULL) {
2922 * We're not a root i/o, so there's nothing to do
2923 * but notify our parent. Don't propagate errors
2924 * upward since we haven't permanently failed yet.
2926 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2927 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
2928 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2929 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
2931 * We'd fail again if we reexecuted now, so suspend
2932 * until conditions improve (e.g. device comes online).
2934 zio_suspend(zio->io_spa, zio);
2937 * Reexecution is potentially a huge amount of work.
2938 * Hand it off to the otherwise-unused claim taskq.
2940 (void) taskq_dispatch(
2941 zio->io_spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
2942 (task_func_t *)zio_reexecute, zio, TQ_SLEEP);
2944 return (ZIO_PIPELINE_STOP);
2947 ASSERT(zio->io_child_count == 0);
2948 ASSERT(zio->io_reexecute == 0);
2949 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
2952 * Report any checksum errors, since the I/O is complete.
2954 while (zio->io_cksum_report != NULL) {
2955 zio_cksum_report_t *zcr = zio->io_cksum_report;
2956 zio->io_cksum_report = zcr->zcr_next;
2957 zcr->zcr_next = NULL;
2958 zcr->zcr_finish(zcr, NULL);
2959 zfs_ereport_free_checksum(zcr);
2963 * It is the responsibility of the done callback to ensure that this
2964 * particular zio is no longer discoverable for adoption, and as
2965 * such, cannot acquire any new parents.
2970 mutex_enter(&zio->io_lock);
2971 zio->io_state[ZIO_WAIT_DONE] = 1;
2972 mutex_exit(&zio->io_lock);
2974 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2975 zio_link_t *zl = zio->io_walk_link;
2976 pio_next = zio_walk_parents(zio);
2977 zio_remove_child(pio, zio, zl);
2978 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2981 if (zio->io_waiter != NULL) {
2982 mutex_enter(&zio->io_lock);
2983 zio->io_executor = NULL;
2984 cv_broadcast(&zio->io_cv);
2985 mutex_exit(&zio->io_lock);
2990 return (ZIO_PIPELINE_STOP);
2994 * ==========================================================================
2995 * I/O pipeline definition
2996 * ==========================================================================
2998 static zio_pipe_stage_t *zio_pipeline[] = {
3004 zio_checksum_generate,
3018 zio_checksum_verify,
3022 #if defined(_KERNEL) && defined(HAVE_SPL)
3023 /* Fault injection */
3024 EXPORT_SYMBOL(zio_injection_enabled);
3025 EXPORT_SYMBOL(zio_inject_fault);
3026 EXPORT_SYMBOL(zio_inject_list_next);
3027 EXPORT_SYMBOL(zio_clear_fault);
3028 EXPORT_SYMBOL(zio_handle_fault_injection);
3029 EXPORT_SYMBOL(zio_handle_device_injection);
3030 EXPORT_SYMBOL(zio_handle_label_injection);
3031 EXPORT_SYMBOL(zio_priority_table);
3032 EXPORT_SYMBOL(zio_type_name);
3034 module_param(zio_bulk_flags, int, 0644);
3035 MODULE_PARM_DESC(zio_bulk_flags, "Additional flags to pass to bulk buffers");
3037 module_param(zio_delay_max, int, 0644);
3038 MODULE_PARM_DESC(zio_delay_max, "Max zio millisec delay before posting event");
3040 module_param(zio_requeue_io_start_cut_in_line, int, 0644);
3041 MODULE_PARM_DESC(zio_requeue_io_start_cut_in_line, "Prioritize requeued I/O");