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.
25 #include <sys/zfs_context.h>
26 #include <sys/fm/fs/zfs.h>
29 #include <sys/spa_impl.h>
30 #include <sys/vdev_impl.h>
31 #include <sys/zio_impl.h>
32 #include <sys/zio_compress.h>
33 #include <sys/zio_checksum.h>
34 #include <sys/dmu_objset.h>
39 * ==========================================================================
41 * ==========================================================================
43 uint8_t zio_priority_table[ZIO_PRIORITY_TABLE_SIZE] = {
44 0, /* ZIO_PRIORITY_NOW */
45 0, /* ZIO_PRIORITY_SYNC_READ */
46 0, /* ZIO_PRIORITY_SYNC_WRITE */
47 0, /* ZIO_PRIORITY_LOG_WRITE */
48 1, /* ZIO_PRIORITY_CACHE_FILL */
49 1, /* ZIO_PRIORITY_AGG */
50 4, /* ZIO_PRIORITY_FREE */
51 4, /* ZIO_PRIORITY_ASYNC_WRITE */
52 6, /* ZIO_PRIORITY_ASYNC_READ */
53 10, /* ZIO_PRIORITY_RESILVER */
54 20, /* ZIO_PRIORITY_SCRUB */
55 2, /* ZIO_PRIORITY_DDT_PREFETCH */
59 * ==========================================================================
60 * I/O type descriptions
61 * ==========================================================================
63 char *zio_type_name[ZIO_TYPES] = {
64 "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl"
68 * ==========================================================================
70 * ==========================================================================
72 kmem_cache_t *zio_cache;
73 kmem_cache_t *zio_link_cache;
74 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
75 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
76 int zio_bulk_flags = 0;
77 int zio_delay_max = ZIO_DELAY_MAX;
80 extern vmem_t *zio_alloc_arena;
84 * An allocating zio is one that either currently has the DVA allocate
85 * stage set or will have it later in its lifetime.
87 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
89 int zio_requeue_io_start_cut_in_line = 1;
92 int zio_buf_debug_limit = 16384;
94 int zio_buf_debug_limit = 0;
97 static inline void __zio_execute(zio_t *zio);
103 vmem_t *data_alloc_arena = NULL;
106 data_alloc_arena = zio_alloc_arena;
108 zio_cache = kmem_cache_create("zio_cache",
109 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
110 zio_link_cache = kmem_cache_create("zio_link_cache",
111 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
114 * For small buffers, we want a cache for each multiple of
115 * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache
116 * for each quarter-power of 2. For large buffers, we want
117 * a cache for each multiple of PAGESIZE.
119 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
120 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
124 while (p2 & (p2 - 1))
127 if (size <= 4 * SPA_MINBLOCKSIZE) {
128 align = SPA_MINBLOCKSIZE;
129 } else if (P2PHASE(size, PAGESIZE) == 0) {
131 } else if (P2PHASE(size, p2 >> 2) == 0) {
137 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
138 zio_buf_cache[c] = kmem_cache_create(name, size,
139 align, NULL, NULL, NULL, NULL, NULL,
140 (size > zio_buf_debug_limit ? KMC_NODEBUG : 0) |
143 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
144 zio_data_buf_cache[c] = kmem_cache_create(name, size,
145 align, NULL, NULL, NULL, NULL, data_alloc_arena,
146 (size > zio_buf_debug_limit ? KMC_NODEBUG : 0) |
152 ASSERT(zio_buf_cache[c] != NULL);
153 if (zio_buf_cache[c - 1] == NULL)
154 zio_buf_cache[c - 1] = zio_buf_cache[c];
156 ASSERT(zio_data_buf_cache[c] != NULL);
157 if (zio_data_buf_cache[c - 1] == NULL)
158 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
168 kmem_cache_t *last_cache = NULL;
169 kmem_cache_t *last_data_cache = NULL;
171 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
172 if (zio_buf_cache[c] != last_cache) {
173 last_cache = zio_buf_cache[c];
174 kmem_cache_destroy(zio_buf_cache[c]);
176 zio_buf_cache[c] = NULL;
178 if (zio_data_buf_cache[c] != last_data_cache) {
179 last_data_cache = zio_data_buf_cache[c];
180 kmem_cache_destroy(zio_data_buf_cache[c]);
182 zio_data_buf_cache[c] = NULL;
185 kmem_cache_destroy(zio_link_cache);
186 kmem_cache_destroy(zio_cache);
192 * ==========================================================================
193 * Allocate and free I/O buffers
194 * ==========================================================================
198 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
199 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
200 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
201 * excess / transient data in-core during a crashdump.
204 zio_buf_alloc(size_t size)
206 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
208 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
210 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
214 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
215 * crashdump if the kernel panics. This exists so that we will limit the amount
216 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
217 * of kernel heap dumped to disk when the kernel panics)
220 zio_data_buf_alloc(size_t size)
222 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
224 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
226 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
230 zio_buf_free(void *buf, size_t size)
232 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
234 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
236 kmem_cache_free(zio_buf_cache[c], buf);
240 zio_data_buf_free(void *buf, size_t size)
242 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
244 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
246 kmem_cache_free(zio_data_buf_cache[c], buf);
250 * ==========================================================================
251 * Push and pop I/O transform buffers
252 * ==========================================================================
255 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
256 zio_transform_func_t *transform)
258 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_PUSHPAGE);
260 zt->zt_orig_data = zio->io_data;
261 zt->zt_orig_size = zio->io_size;
262 zt->zt_bufsize = bufsize;
263 zt->zt_transform = transform;
265 zt->zt_next = zio->io_transform_stack;
266 zio->io_transform_stack = zt;
273 zio_pop_transforms(zio_t *zio)
277 while ((zt = zio->io_transform_stack) != NULL) {
278 if (zt->zt_transform != NULL)
279 zt->zt_transform(zio,
280 zt->zt_orig_data, zt->zt_orig_size);
282 if (zt->zt_bufsize != 0)
283 zio_buf_free(zio->io_data, zt->zt_bufsize);
285 zio->io_data = zt->zt_orig_data;
286 zio->io_size = zt->zt_orig_size;
287 zio->io_transform_stack = zt->zt_next;
289 kmem_free(zt, sizeof (zio_transform_t));
294 * ==========================================================================
295 * I/O transform callbacks for subblocks and decompression
296 * ==========================================================================
299 zio_subblock(zio_t *zio, void *data, uint64_t size)
301 ASSERT(zio->io_size > size);
303 if (zio->io_type == ZIO_TYPE_READ)
304 bcopy(zio->io_data, data, size);
308 zio_decompress(zio_t *zio, void *data, uint64_t size)
310 if (zio->io_error == 0 &&
311 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
312 zio->io_data, data, zio->io_size, size) != 0)
317 * ==========================================================================
318 * I/O parent/child relationships and pipeline interlocks
319 * ==========================================================================
322 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
323 * continue calling these functions until they return NULL.
324 * Otherwise, the next caller will pick up the list walk in
325 * some indeterminate state. (Otherwise every caller would
326 * have to pass in a cookie to keep the state represented by
327 * io_walk_link, which gets annoying.)
330 zio_walk_parents(zio_t *cio)
332 zio_link_t *zl = cio->io_walk_link;
333 list_t *pl = &cio->io_parent_list;
335 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
336 cio->io_walk_link = zl;
341 ASSERT(zl->zl_child == cio);
342 return (zl->zl_parent);
346 zio_walk_children(zio_t *pio)
348 zio_link_t *zl = pio->io_walk_link;
349 list_t *cl = &pio->io_child_list;
351 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
352 pio->io_walk_link = zl;
357 ASSERT(zl->zl_parent == pio);
358 return (zl->zl_child);
362 zio_unique_parent(zio_t *cio)
364 zio_t *pio = zio_walk_parents(cio);
366 VERIFY(zio_walk_parents(cio) == NULL);
371 zio_add_child(zio_t *pio, zio_t *cio)
373 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_PUSHPAGE);
377 * Logical I/Os can have logical, gang, or vdev children.
378 * Gang I/Os can have gang or vdev children.
379 * Vdev I/Os can only have vdev children.
380 * The following ASSERT captures all of these constraints.
382 ASSERT(cio->io_child_type <= pio->io_child_type);
387 mutex_enter(&cio->io_lock);
388 mutex_enter(&pio->io_lock);
390 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
392 for (w = 0; w < ZIO_WAIT_TYPES; w++)
393 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
395 list_insert_head(&pio->io_child_list, zl);
396 list_insert_head(&cio->io_parent_list, zl);
398 pio->io_child_count++;
399 cio->io_parent_count++;
401 mutex_exit(&pio->io_lock);
402 mutex_exit(&cio->io_lock);
406 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
408 ASSERT(zl->zl_parent == pio);
409 ASSERT(zl->zl_child == cio);
411 mutex_enter(&cio->io_lock);
412 mutex_enter(&pio->io_lock);
414 list_remove(&pio->io_child_list, zl);
415 list_remove(&cio->io_parent_list, zl);
417 pio->io_child_count--;
418 cio->io_parent_count--;
420 mutex_exit(&pio->io_lock);
421 mutex_exit(&cio->io_lock);
423 kmem_cache_free(zio_link_cache, zl);
427 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
429 uint64_t *countp = &zio->io_children[child][wait];
430 boolean_t waiting = B_FALSE;
432 mutex_enter(&zio->io_lock);
433 ASSERT(zio->io_stall == NULL);
436 zio->io_stall = countp;
439 mutex_exit(&zio->io_lock);
444 __attribute__((always_inline))
446 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
448 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
449 int *errorp = &pio->io_child_error[zio->io_child_type];
451 mutex_enter(&pio->io_lock);
452 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
453 *errorp = zio_worst_error(*errorp, zio->io_error);
454 pio->io_reexecute |= zio->io_reexecute;
455 ASSERT3U(*countp, >, 0);
456 if (--*countp == 0 && pio->io_stall == countp) {
457 pio->io_stall = NULL;
458 mutex_exit(&pio->io_lock);
461 mutex_exit(&pio->io_lock);
466 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
468 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
469 zio->io_error = zio->io_child_error[c];
473 * ==========================================================================
474 * Create the various types of I/O (read, write, free, etc)
475 * ==========================================================================
478 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
479 void *data, uint64_t size, zio_done_func_t *done, void *private,
480 zio_type_t type, int priority, enum zio_flag flags,
481 vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
482 enum zio_stage stage, enum zio_stage pipeline)
486 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
487 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
488 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
490 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
491 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
492 ASSERT(vd || stage == ZIO_STAGE_OPEN);
494 zio = kmem_cache_alloc(zio_cache, KM_PUSHPAGE);
495 bzero(zio, sizeof (zio_t));
497 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
498 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
500 list_create(&zio->io_parent_list, sizeof (zio_link_t),
501 offsetof(zio_link_t, zl_parent_node));
502 list_create(&zio->io_child_list, sizeof (zio_link_t),
503 offsetof(zio_link_t, zl_child_node));
506 zio->io_child_type = ZIO_CHILD_VDEV;
507 else if (flags & ZIO_FLAG_GANG_CHILD)
508 zio->io_child_type = ZIO_CHILD_GANG;
509 else if (flags & ZIO_FLAG_DDT_CHILD)
510 zio->io_child_type = ZIO_CHILD_DDT;
512 zio->io_child_type = ZIO_CHILD_LOGICAL;
515 zio->io_bp = (blkptr_t *)bp;
516 zio->io_bp_copy = *bp;
517 zio->io_bp_orig = *bp;
518 if (type != ZIO_TYPE_WRITE ||
519 zio->io_child_type == ZIO_CHILD_DDT)
520 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
521 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
522 zio->io_logical = zio;
523 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
524 pipeline |= ZIO_GANG_STAGES;
530 zio->io_private = private;
532 zio->io_priority = priority;
534 zio->io_offset = offset;
535 zio->io_orig_data = zio->io_data = data;
536 zio->io_orig_size = zio->io_size = size;
537 zio->io_orig_flags = zio->io_flags = flags;
538 zio->io_orig_stage = zio->io_stage = stage;
539 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
541 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
542 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
545 zio->io_bookmark = *zb;
548 if (zio->io_logical == NULL)
549 zio->io_logical = pio->io_logical;
550 if (zio->io_child_type == ZIO_CHILD_GANG)
551 zio->io_gang_leader = pio->io_gang_leader;
552 zio_add_child(pio, zio);
559 zio_destroy(zio_t *zio)
561 list_destroy(&zio->io_parent_list);
562 list_destroy(&zio->io_child_list);
563 mutex_destroy(&zio->io_lock);
564 cv_destroy(&zio->io_cv);
565 kmem_cache_free(zio_cache, zio);
569 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
570 void *private, enum zio_flag flags)
574 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
575 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
576 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
582 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
584 return (zio_null(NULL, spa, NULL, done, private, flags));
588 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
589 void *data, uint64_t size, zio_done_func_t *done, void *private,
590 int priority, enum zio_flag flags, const zbookmark_t *zb)
594 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
595 data, size, done, private,
596 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
597 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
598 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
604 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
605 void *data, uint64_t size, const zio_prop_t *zp,
606 zio_done_func_t *ready, zio_done_func_t *done, void *private,
607 int priority, enum zio_flag flags, const zbookmark_t *zb)
611 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
612 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
613 zp->zp_compress >= ZIO_COMPRESS_OFF &&
614 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
615 zp->zp_type < DMU_OT_NUMTYPES &&
618 zp->zp_copies <= spa_max_replication(spa) &&
620 zp->zp_dedup_verify <= 1);
622 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
623 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
624 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
625 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
627 zio->io_ready = ready;
634 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
635 uint64_t size, zio_done_func_t *done, void *private, int priority,
636 enum zio_flag flags, zbookmark_t *zb)
640 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
641 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
642 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
648 zio_write_override(zio_t *zio, blkptr_t *bp, int copies)
650 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
651 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
652 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
653 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
655 zio->io_prop.zp_copies = copies;
656 zio->io_bp_override = bp;
660 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
662 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
666 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
671 dprintf_bp(bp, "freeing in txg %llu, pass %u",
672 (longlong_t)txg, spa->spa_sync_pass);
674 ASSERT(!BP_IS_HOLE(bp));
675 ASSERT(spa_syncing_txg(spa) == txg);
676 ASSERT(spa_sync_pass(spa) <= SYNC_PASS_DEFERRED_FREE);
678 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
679 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
680 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
686 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
687 zio_done_func_t *done, void *private, enum zio_flag flags)
692 * A claim is an allocation of a specific block. Claims are needed
693 * to support immediate writes in the intent log. The issue is that
694 * immediate writes contain committed data, but in a txg that was
695 * *not* committed. Upon opening the pool after an unclean shutdown,
696 * the intent log claims all blocks that contain immediate write data
697 * so that the SPA knows they're in use.
699 * All claims *must* be resolved in the first txg -- before the SPA
700 * starts allocating blocks -- so that nothing is allocated twice.
701 * If txg == 0 we just verify that the block is claimable.
703 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
704 ASSERT(txg == spa_first_txg(spa) || txg == 0);
705 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
707 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
708 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
709 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
715 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
716 zio_done_func_t *done, void *private, int priority, enum zio_flag flags)
721 if (vd->vdev_children == 0) {
722 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
723 ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
724 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
728 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
730 for (c = 0; c < vd->vdev_children; c++)
731 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
732 done, private, priority, flags));
739 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
740 void *data, int checksum, zio_done_func_t *done, void *private,
741 int priority, enum zio_flag flags, boolean_t labels)
745 ASSERT(vd->vdev_children == 0);
746 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
747 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
748 ASSERT3U(offset + size, <=, vd->vdev_psize);
750 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
751 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
752 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
754 zio->io_prop.zp_checksum = checksum;
760 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
761 void *data, int checksum, zio_done_func_t *done, void *private,
762 int priority, enum zio_flag flags, boolean_t labels)
766 ASSERT(vd->vdev_children == 0);
767 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
768 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
769 ASSERT3U(offset + size, <=, vd->vdev_psize);
771 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
772 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
773 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
775 zio->io_prop.zp_checksum = checksum;
777 if (zio_checksum_table[checksum].ci_eck) {
779 * zec checksums are necessarily destructive -- they modify
780 * the end of the write buffer to hold the verifier/checksum.
781 * Therefore, we must make a local copy in case the data is
782 * being written to multiple places in parallel.
784 void *wbuf = zio_buf_alloc(size);
785 bcopy(data, wbuf, size);
786 zio_push_transform(zio, wbuf, size, size, NULL);
793 * Create a child I/O to do some work for us.
796 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
797 void *data, uint64_t size, int type, int priority, enum zio_flag flags,
798 zio_done_func_t *done, void *private)
800 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
803 ASSERT(vd->vdev_parent ==
804 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
806 if (type == ZIO_TYPE_READ && bp != NULL) {
808 * If we have the bp, then the child should perform the
809 * checksum and the parent need not. This pushes error
810 * detection as close to the leaves as possible and
811 * eliminates redundant checksums in the interior nodes.
813 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
814 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
817 if (vd->vdev_children == 0)
818 offset += VDEV_LABEL_START_SIZE;
820 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
823 * If we've decided to do a repair, the write is not speculative --
824 * even if the original read was.
826 if (flags & ZIO_FLAG_IO_REPAIR)
827 flags &= ~ZIO_FLAG_SPECULATIVE;
829 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
830 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
831 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
837 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
838 int type, int priority, enum zio_flag flags,
839 zio_done_func_t *done, void *private)
843 ASSERT(vd->vdev_ops->vdev_op_leaf);
845 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
846 data, size, done, private, type, priority,
847 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
849 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
855 zio_flush(zio_t *zio, vdev_t *vd)
857 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
858 NULL, NULL, ZIO_PRIORITY_NOW,
859 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
863 zio_shrink(zio_t *zio, uint64_t size)
865 ASSERT(zio->io_executor == NULL);
866 ASSERT(zio->io_orig_size == zio->io_size);
867 ASSERT(size <= zio->io_size);
870 * We don't shrink for raidz because of problems with the
871 * reconstruction when reading back less than the block size.
872 * Note, BP_IS_RAIDZ() assumes no compression.
874 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
875 if (!BP_IS_RAIDZ(zio->io_bp))
876 zio->io_orig_size = zio->io_size = size;
880 * ==========================================================================
881 * Prepare to read and write logical blocks
882 * ==========================================================================
886 zio_read_bp_init(zio_t *zio)
888 blkptr_t *bp = zio->io_bp;
890 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
891 zio->io_child_type == ZIO_CHILD_LOGICAL &&
892 !(zio->io_flags & ZIO_FLAG_RAW)) {
893 uint64_t psize = BP_GET_PSIZE(bp);
894 void *cbuf = zio_buf_alloc(psize);
896 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
899 if (!dmu_ot[BP_GET_TYPE(bp)].ot_metadata && BP_GET_LEVEL(bp) == 0)
900 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
902 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
903 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
905 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
906 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
908 return (ZIO_PIPELINE_CONTINUE);
912 zio_write_bp_init(zio_t *zio)
914 spa_t *spa = zio->io_spa;
915 zio_prop_t *zp = &zio->io_prop;
916 enum zio_compress compress = zp->zp_compress;
917 blkptr_t *bp = zio->io_bp;
918 uint64_t lsize = zio->io_size;
919 uint64_t psize = lsize;
923 * If our children haven't all reached the ready stage,
924 * wait for them and then repeat this pipeline stage.
926 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
927 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
928 return (ZIO_PIPELINE_STOP);
930 if (!IO_IS_ALLOCATING(zio))
931 return (ZIO_PIPELINE_CONTINUE);
933 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
935 if (zio->io_bp_override) {
936 ASSERT(bp->blk_birth != zio->io_txg);
937 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
939 *bp = *zio->io_bp_override;
940 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
942 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
943 return (ZIO_PIPELINE_CONTINUE);
945 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
946 zp->zp_dedup_verify);
948 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
950 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
951 return (ZIO_PIPELINE_CONTINUE);
953 zio->io_bp_override = NULL;
957 if (bp->blk_birth == zio->io_txg) {
959 * We're rewriting an existing block, which means we're
960 * working on behalf of spa_sync(). For spa_sync() to
961 * converge, it must eventually be the case that we don't
962 * have to allocate new blocks. But compression changes
963 * the blocksize, which forces a reallocate, and makes
964 * convergence take longer. Therefore, after the first
965 * few passes, stop compressing to ensure convergence.
967 pass = spa_sync_pass(spa);
969 ASSERT(zio->io_txg == spa_syncing_txg(spa));
970 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
971 ASSERT(!BP_GET_DEDUP(bp));
973 if (pass > SYNC_PASS_DONT_COMPRESS)
974 compress = ZIO_COMPRESS_OFF;
976 /* Make sure someone doesn't change their mind on overwrites */
977 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
978 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
981 if (compress != ZIO_COMPRESS_OFF) {
982 void *cbuf = zio_buf_alloc(lsize);
983 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
984 if (psize == 0 || psize == lsize) {
985 compress = ZIO_COMPRESS_OFF;
986 zio_buf_free(cbuf, lsize);
988 ASSERT(psize < lsize);
989 zio_push_transform(zio, cbuf, psize, lsize, NULL);
994 * The final pass of spa_sync() must be all rewrites, but the first
995 * few passes offer a trade-off: allocating blocks defers convergence,
996 * but newly allocated blocks are sequential, so they can be written
997 * to disk faster. Therefore, we allow the first few passes of
998 * spa_sync() to allocate new blocks, but force rewrites after that.
999 * There should only be a handful of blocks after pass 1 in any case.
1001 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
1002 pass > SYNC_PASS_REWRITE) {
1003 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1005 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1006 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1009 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1013 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1015 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1016 BP_SET_LSIZE(bp, lsize);
1017 BP_SET_PSIZE(bp, psize);
1018 BP_SET_COMPRESS(bp, compress);
1019 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1020 BP_SET_TYPE(bp, zp->zp_type);
1021 BP_SET_LEVEL(bp, zp->zp_level);
1022 BP_SET_DEDUP(bp, zp->zp_dedup);
1023 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1025 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1026 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1027 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1031 return (ZIO_PIPELINE_CONTINUE);
1035 zio_free_bp_init(zio_t *zio)
1037 blkptr_t *bp = zio->io_bp;
1039 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1040 if (BP_GET_DEDUP(bp))
1041 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1044 return (ZIO_PIPELINE_CONTINUE);
1048 * ==========================================================================
1049 * Execute the I/O pipeline
1050 * ==========================================================================
1054 zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q, boolean_t cutinline)
1056 spa_t *spa = zio->io_spa;
1057 zio_type_t t = zio->io_type;
1058 int flags = TQ_NOSLEEP | (cutinline ? TQ_FRONT : 0);
1061 * If we're a config writer or a probe, the normal issue and
1062 * interrupt threads may all be blocked waiting for the config lock.
1063 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1065 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1069 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1071 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1075 * If this is a high priority I/O, then use the high priority taskq.
1077 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1078 spa->spa_zio_taskq[t][q + 1] != NULL)
1081 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1083 while (taskq_dispatch(spa->spa_zio_taskq[t][q],
1084 (task_func_t *)zio_execute, zio, flags) == 0); /* do nothing */
1088 zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
1090 kthread_t *executor = zio->io_executor;
1091 spa_t *spa = zio->io_spa;
1094 for (t = 0; t < ZIO_TYPES; t++)
1095 if (taskq_member(spa->spa_zio_taskq[t][q], executor))
1102 zio_issue_async(zio_t *zio)
1104 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1106 return (ZIO_PIPELINE_STOP);
1110 zio_interrupt(zio_t *zio)
1112 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1116 * Execute the I/O pipeline until one of the following occurs:
1117 * (1) the I/O completes; (2) the pipeline stalls waiting for
1118 * dependent child I/Os; (3) the I/O issues, so we're waiting
1119 * for an I/O completion interrupt; (4) the I/O is delegated by
1120 * vdev-level caching or aggregation; (5) the I/O is deferred
1121 * due to vdev-level queueing; (6) the I/O is handed off to
1122 * another thread. In all cases, the pipeline stops whenever
1123 * there's no CPU work; it never burns a thread in cv_wait().
1125 * There's no locking on io_stage because there's no legitimate way
1126 * for multiple threads to be attempting to process the same I/O.
1128 static zio_pipe_stage_t *zio_pipeline[];
1131 * zio_execute() is a wrapper around the static function
1132 * __zio_execute() so that we can force __zio_execute() to be
1133 * inlined. This reduces stack overhead which is important
1134 * because __zio_execute() is called recursively in several zio
1135 * code paths. zio_execute() itself cannot be inlined because
1136 * it is externally visible.
1139 zio_execute(zio_t *zio)
1144 __attribute__((always_inline))
1146 __zio_execute(zio_t *zio)
1148 zio->io_executor = curthread;
1150 while (zio->io_stage < ZIO_STAGE_DONE) {
1151 enum zio_stage pipeline = zio->io_pipeline;
1152 enum zio_stage stage = zio->io_stage;
1157 ASSERT(!MUTEX_HELD(&zio->io_lock));
1158 ASSERT(ISP2(stage));
1159 ASSERT(zio->io_stall == NULL);
1163 } while ((stage & pipeline) == 0);
1165 ASSERT(stage <= ZIO_STAGE_DONE);
1167 dsl = spa_get_dsl(zio->io_spa);
1168 cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1169 zio_requeue_io_start_cut_in_line : B_FALSE;
1172 * If we are in interrupt context and this pipeline stage
1173 * will grab a config lock that is held across I/O,
1174 * or may wait for an I/O that needs an interrupt thread
1175 * to complete, issue async to avoid deadlock.
1177 * If we are in the txg_sync_thread or being called
1178 * during pool init issue async to minimize stack depth.
1179 * Both of these call paths may be recursively called.
1181 * For VDEV_IO_START, we cut in line so that the io will
1182 * be sent to disk promptly.
1184 if (((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1185 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) ||
1186 (dsl != NULL && dsl_pool_sync_context(dsl))) {
1187 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1191 zio->io_stage = stage;
1192 rv = zio_pipeline[highbit(stage) - 1](zio);
1194 if (rv == ZIO_PIPELINE_STOP)
1197 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1203 * ==========================================================================
1204 * Initiate I/O, either sync or async
1205 * ==========================================================================
1208 zio_wait(zio_t *zio)
1212 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1213 ASSERT(zio->io_executor == NULL);
1215 zio->io_waiter = curthread;
1219 mutex_enter(&zio->io_lock);
1220 while (zio->io_executor != NULL)
1221 cv_wait(&zio->io_cv, &zio->io_lock);
1222 mutex_exit(&zio->io_lock);
1224 error = zio->io_error;
1231 zio_nowait(zio_t *zio)
1233 ASSERT(zio->io_executor == NULL);
1235 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1236 zio_unique_parent(zio) == NULL) {
1238 * This is a logical async I/O with no parent to wait for it.
1239 * We add it to the spa_async_root_zio "Godfather" I/O which
1240 * will ensure they complete prior to unloading the pool.
1242 spa_t *spa = zio->io_spa;
1244 zio_add_child(spa->spa_async_zio_root, zio);
1251 * ==========================================================================
1252 * Reexecute or suspend/resume failed I/O
1253 * ==========================================================================
1257 zio_reexecute(zio_t *pio)
1259 zio_t *cio, *cio_next;
1262 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1263 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1264 ASSERT(pio->io_gang_leader == NULL);
1265 ASSERT(pio->io_gang_tree == NULL);
1267 pio->io_flags = pio->io_orig_flags;
1268 pio->io_stage = pio->io_orig_stage;
1269 pio->io_pipeline = pio->io_orig_pipeline;
1270 pio->io_reexecute = 0;
1272 for (w = 0; w < ZIO_WAIT_TYPES; w++)
1273 pio->io_state[w] = 0;
1274 for (c = 0; c < ZIO_CHILD_TYPES; c++)
1275 pio->io_child_error[c] = 0;
1277 if (IO_IS_ALLOCATING(pio))
1278 BP_ZERO(pio->io_bp);
1281 * As we reexecute pio's children, new children could be created.
1282 * New children go to the head of pio's io_child_list, however,
1283 * so we will (correctly) not reexecute them. The key is that
1284 * the remainder of pio's io_child_list, from 'cio_next' onward,
1285 * cannot be affected by any side effects of reexecuting 'cio'.
1287 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1288 cio_next = zio_walk_children(pio);
1289 mutex_enter(&pio->io_lock);
1290 for (w = 0; w < ZIO_WAIT_TYPES; w++)
1291 pio->io_children[cio->io_child_type][w]++;
1292 mutex_exit(&pio->io_lock);
1297 * Now that all children have been reexecuted, execute the parent.
1298 * We don't reexecute "The Godfather" I/O here as it's the
1299 * responsibility of the caller to wait on him.
1301 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1306 zio_suspend(spa_t *spa, zio_t *zio)
1308 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1309 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1310 "failure and the failure mode property for this pool "
1311 "is set to panic.", spa_name(spa));
1313 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1315 mutex_enter(&spa->spa_suspend_lock);
1317 if (spa->spa_suspend_zio_root == NULL)
1318 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1319 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1320 ZIO_FLAG_GODFATHER);
1322 spa->spa_suspended = B_TRUE;
1325 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1326 ASSERT(zio != spa->spa_suspend_zio_root);
1327 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1328 ASSERT(zio_unique_parent(zio) == NULL);
1329 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1330 zio_add_child(spa->spa_suspend_zio_root, zio);
1333 mutex_exit(&spa->spa_suspend_lock);
1337 zio_resume(spa_t *spa)
1342 * Reexecute all previously suspended i/o.
1344 mutex_enter(&spa->spa_suspend_lock);
1345 spa->spa_suspended = B_FALSE;
1346 cv_broadcast(&spa->spa_suspend_cv);
1347 pio = spa->spa_suspend_zio_root;
1348 spa->spa_suspend_zio_root = NULL;
1349 mutex_exit(&spa->spa_suspend_lock);
1355 return (zio_wait(pio));
1359 zio_resume_wait(spa_t *spa)
1361 mutex_enter(&spa->spa_suspend_lock);
1362 while (spa_suspended(spa))
1363 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1364 mutex_exit(&spa->spa_suspend_lock);
1368 * ==========================================================================
1371 * A gang block is a collection of small blocks that looks to the DMU
1372 * like one large block. When zio_dva_allocate() cannot find a block
1373 * of the requested size, due to either severe fragmentation or the pool
1374 * being nearly full, it calls zio_write_gang_block() to construct the
1375 * block from smaller fragments.
1377 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1378 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1379 * an indirect block: it's an array of block pointers. It consumes
1380 * only one sector and hence is allocatable regardless of fragmentation.
1381 * The gang header's bps point to its gang members, which hold the data.
1383 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1384 * as the verifier to ensure uniqueness of the SHA256 checksum.
1385 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1386 * not the gang header. This ensures that data block signatures (needed for
1387 * deduplication) are independent of how the block is physically stored.
1389 * Gang blocks can be nested: a gang member may itself be a gang block.
1390 * Thus every gang block is a tree in which root and all interior nodes are
1391 * gang headers, and the leaves are normal blocks that contain user data.
1392 * The root of the gang tree is called the gang leader.
1394 * To perform any operation (read, rewrite, free, claim) on a gang block,
1395 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1396 * in the io_gang_tree field of the original logical i/o by recursively
1397 * reading the gang leader and all gang headers below it. This yields
1398 * an in-core tree containing the contents of every gang header and the
1399 * bps for every constituent of the gang block.
1401 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1402 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1403 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1404 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1405 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1406 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1407 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1408 * of the gang header plus zio_checksum_compute() of the data to update the
1409 * gang header's blk_cksum as described above.
1411 * The two-phase assemble/issue model solves the problem of partial failure --
1412 * what if you'd freed part of a gang block but then couldn't read the
1413 * gang header for another part? Assembling the entire gang tree first
1414 * ensures that all the necessary gang header I/O has succeeded before
1415 * starting the actual work of free, claim, or write. Once the gang tree
1416 * is assembled, free and claim are in-memory operations that cannot fail.
1418 * In the event that a gang write fails, zio_dva_unallocate() walks the
1419 * gang tree to immediately free (i.e. insert back into the space map)
1420 * everything we've allocated. This ensures that we don't get ENOSPC
1421 * errors during repeated suspend/resume cycles due to a flaky device.
1423 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1424 * the gang tree, we won't modify the block, so we can safely defer the free
1425 * (knowing that the block is still intact). If we *can* assemble the gang
1426 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1427 * each constituent bp and we can allocate a new block on the next sync pass.
1429 * In all cases, the gang tree allows complete recovery from partial failure.
1430 * ==========================================================================
1434 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1439 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1440 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1441 &pio->io_bookmark));
1445 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1450 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1451 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1452 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1454 * As we rewrite each gang header, the pipeline will compute
1455 * a new gang block header checksum for it; but no one will
1456 * compute a new data checksum, so we do that here. The one
1457 * exception is the gang leader: the pipeline already computed
1458 * its data checksum because that stage precedes gang assembly.
1459 * (Presently, nothing actually uses interior data checksums;
1460 * this is just good hygiene.)
1462 if (gn != pio->io_gang_leader->io_gang_tree) {
1463 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1464 data, BP_GET_PSIZE(bp));
1467 * If we are here to damage data for testing purposes,
1468 * leave the GBH alone so that we can detect the damage.
1470 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1471 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1473 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1474 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1475 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1483 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1485 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1486 ZIO_GANG_CHILD_FLAGS(pio)));
1491 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1493 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1494 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1497 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1506 static void zio_gang_tree_assemble_done(zio_t *zio);
1508 static zio_gang_node_t *
1509 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1511 zio_gang_node_t *gn;
1513 ASSERT(*gnpp == NULL);
1515 gn = kmem_zalloc(sizeof (*gn), KM_PUSHPAGE);
1516 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1523 zio_gang_node_free(zio_gang_node_t **gnpp)
1525 zio_gang_node_t *gn = *gnpp;
1528 for (g = 0; g < SPA_GBH_NBLKPTRS; g++)
1529 ASSERT(gn->gn_child[g] == NULL);
1531 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1532 kmem_free(gn, sizeof (*gn));
1537 zio_gang_tree_free(zio_gang_node_t **gnpp)
1539 zio_gang_node_t *gn = *gnpp;
1545 for (g = 0; g < SPA_GBH_NBLKPTRS; g++)
1546 zio_gang_tree_free(&gn->gn_child[g]);
1548 zio_gang_node_free(gnpp);
1552 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1554 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1556 ASSERT(gio->io_gang_leader == gio);
1557 ASSERT(BP_IS_GANG(bp));
1559 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1560 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1561 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1565 zio_gang_tree_assemble_done(zio_t *zio)
1567 zio_t *gio = zio->io_gang_leader;
1568 zio_gang_node_t *gn = zio->io_private;
1569 blkptr_t *bp = zio->io_bp;
1572 ASSERT(gio == zio_unique_parent(zio));
1573 ASSERT(zio->io_child_count == 0);
1578 if (BP_SHOULD_BYTESWAP(bp))
1579 byteswap_uint64_array(zio->io_data, zio->io_size);
1581 ASSERT(zio->io_data == gn->gn_gbh);
1582 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1583 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1585 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1586 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1587 if (!BP_IS_GANG(gbp))
1589 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1594 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1596 zio_t *gio = pio->io_gang_leader;
1600 ASSERT(BP_IS_GANG(bp) == !!gn);
1601 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1602 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1605 * If you're a gang header, your data is in gn->gn_gbh.
1606 * If you're a gang member, your data is in 'data' and gn == NULL.
1608 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1611 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1613 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1614 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1615 if (BP_IS_HOLE(gbp))
1617 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1618 data = (char *)data + BP_GET_PSIZE(gbp);
1622 if (gn == gio->io_gang_tree)
1623 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1630 zio_gang_assemble(zio_t *zio)
1632 blkptr_t *bp = zio->io_bp;
1634 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1635 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1637 zio->io_gang_leader = zio;
1639 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1641 return (ZIO_PIPELINE_CONTINUE);
1645 zio_gang_issue(zio_t *zio)
1647 blkptr_t *bp = zio->io_bp;
1649 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1650 return (ZIO_PIPELINE_STOP);
1652 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1653 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1655 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1656 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1658 zio_gang_tree_free(&zio->io_gang_tree);
1660 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1662 return (ZIO_PIPELINE_CONTINUE);
1666 zio_write_gang_member_ready(zio_t *zio)
1668 zio_t *pio = zio_unique_parent(zio);
1669 ASSERTV(zio_t *gio = zio->io_gang_leader;)
1670 dva_t *cdva = zio->io_bp->blk_dva;
1671 dva_t *pdva = pio->io_bp->blk_dva;
1675 if (BP_IS_HOLE(zio->io_bp))
1678 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1680 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1681 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1682 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1683 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1684 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1686 mutex_enter(&pio->io_lock);
1687 for (d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1688 ASSERT(DVA_GET_GANG(&pdva[d]));
1689 asize = DVA_GET_ASIZE(&pdva[d]);
1690 asize += DVA_GET_ASIZE(&cdva[d]);
1691 DVA_SET_ASIZE(&pdva[d], asize);
1693 mutex_exit(&pio->io_lock);
1697 zio_write_gang_block(zio_t *pio)
1699 spa_t *spa = pio->io_spa;
1700 blkptr_t *bp = pio->io_bp;
1701 zio_t *gio = pio->io_gang_leader;
1703 zio_gang_node_t *gn, **gnpp;
1704 zio_gbh_phys_t *gbh;
1705 uint64_t txg = pio->io_txg;
1706 uint64_t resid = pio->io_size;
1708 int copies = gio->io_prop.zp_copies;
1709 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1713 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1714 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1715 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1717 pio->io_error = error;
1718 return (ZIO_PIPELINE_CONTINUE);
1722 gnpp = &gio->io_gang_tree;
1724 gnpp = pio->io_private;
1725 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1728 gn = zio_gang_node_alloc(gnpp);
1730 bzero(gbh, SPA_GANGBLOCKSIZE);
1733 * Create the gang header.
1735 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1736 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1739 * Create and nowait the gang children.
1741 for (g = 0; resid != 0; resid -= lsize, g++) {
1742 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1744 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1746 zp.zp_checksum = gio->io_prop.zp_checksum;
1747 zp.zp_compress = ZIO_COMPRESS_OFF;
1748 zp.zp_type = DMU_OT_NONE;
1750 zp.zp_copies = gio->io_prop.zp_copies;
1752 zp.zp_dedup_verify = 0;
1754 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1755 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1756 zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1757 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1758 &pio->io_bookmark));
1762 * Set pio's pipeline to just wait for zio to finish.
1764 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1768 return (ZIO_PIPELINE_CONTINUE);
1772 * ==========================================================================
1774 * ==========================================================================
1777 zio_ddt_child_read_done(zio_t *zio)
1779 blkptr_t *bp = zio->io_bp;
1780 ddt_entry_t *dde = zio->io_private;
1782 zio_t *pio = zio_unique_parent(zio);
1784 mutex_enter(&pio->io_lock);
1785 ddp = ddt_phys_select(dde, bp);
1786 if (zio->io_error == 0)
1787 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1788 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1789 dde->dde_repair_data = zio->io_data;
1791 zio_buf_free(zio->io_data, zio->io_size);
1792 mutex_exit(&pio->io_lock);
1796 zio_ddt_read_start(zio_t *zio)
1798 blkptr_t *bp = zio->io_bp;
1801 ASSERT(BP_GET_DEDUP(bp));
1802 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1803 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1805 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1806 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1807 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1808 ddt_phys_t *ddp = dde->dde_phys;
1809 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1812 ASSERT(zio->io_vsd == NULL);
1815 if (ddp_self == NULL)
1816 return (ZIO_PIPELINE_CONTINUE);
1818 for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1819 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1821 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1823 zio_nowait(zio_read(zio, zio->io_spa, &blk,
1824 zio_buf_alloc(zio->io_size), zio->io_size,
1825 zio_ddt_child_read_done, dde, zio->io_priority,
1826 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1827 &zio->io_bookmark));
1829 return (ZIO_PIPELINE_CONTINUE);
1832 zio_nowait(zio_read(zio, zio->io_spa, bp,
1833 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1834 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1836 return (ZIO_PIPELINE_CONTINUE);
1840 zio_ddt_read_done(zio_t *zio)
1842 blkptr_t *bp = zio->io_bp;
1844 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1845 return (ZIO_PIPELINE_STOP);
1847 ASSERT(BP_GET_DEDUP(bp));
1848 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1849 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1851 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1852 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1853 ddt_entry_t *dde = zio->io_vsd;
1855 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
1856 return (ZIO_PIPELINE_CONTINUE);
1859 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
1860 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1861 return (ZIO_PIPELINE_STOP);
1863 if (dde->dde_repair_data != NULL) {
1864 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
1865 zio->io_child_error[ZIO_CHILD_DDT] = 0;
1867 ddt_repair_done(ddt, dde);
1871 ASSERT(zio->io_vsd == NULL);
1873 return (ZIO_PIPELINE_CONTINUE);
1877 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
1879 spa_t *spa = zio->io_spa;
1883 * Note: we compare the original data, not the transformed data,
1884 * because when zio->io_bp is an override bp, we will not have
1885 * pushed the I/O transforms. That's an important optimization
1886 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
1888 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1889 zio_t *lio = dde->dde_lead_zio[p];
1892 return (lio->io_orig_size != zio->io_orig_size ||
1893 bcmp(zio->io_orig_data, lio->io_orig_data,
1894 zio->io_orig_size) != 0);
1898 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1899 ddt_phys_t *ddp = &dde->dde_phys[p];
1901 if (ddp->ddp_phys_birth != 0) {
1902 arc_buf_t *abuf = NULL;
1903 uint32_t aflags = ARC_WAIT;
1904 blkptr_t blk = *zio->io_bp;
1907 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
1911 error = arc_read_nolock(NULL, spa, &blk,
1912 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
1913 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1914 &aflags, &zio->io_bookmark);
1917 if (arc_buf_size(abuf) != zio->io_orig_size ||
1918 bcmp(abuf->b_data, zio->io_orig_data,
1919 zio->io_orig_size) != 0)
1921 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
1925 return (error != 0);
1933 zio_ddt_child_write_ready(zio_t *zio)
1935 int p = zio->io_prop.zp_copies;
1936 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1937 ddt_entry_t *dde = zio->io_private;
1938 ddt_phys_t *ddp = &dde->dde_phys[p];
1946 ASSERT(dde->dde_lead_zio[p] == zio);
1948 ddt_phys_fill(ddp, zio->io_bp);
1950 while ((pio = zio_walk_parents(zio)) != NULL)
1951 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
1957 zio_ddt_child_write_done(zio_t *zio)
1959 int p = zio->io_prop.zp_copies;
1960 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1961 ddt_entry_t *dde = zio->io_private;
1962 ddt_phys_t *ddp = &dde->dde_phys[p];
1966 ASSERT(ddp->ddp_refcnt == 0);
1967 ASSERT(dde->dde_lead_zio[p] == zio);
1968 dde->dde_lead_zio[p] = NULL;
1970 if (zio->io_error == 0) {
1971 while (zio_walk_parents(zio) != NULL)
1972 ddt_phys_addref(ddp);
1974 ddt_phys_clear(ddp);
1981 zio_ddt_ditto_write_done(zio_t *zio)
1983 int p = DDT_PHYS_DITTO;
1984 blkptr_t *bp = zio->io_bp;
1985 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1986 ddt_entry_t *dde = zio->io_private;
1987 ddt_phys_t *ddp = &dde->dde_phys[p];
1988 ddt_key_t *ddk = &dde->dde_key;
1989 ASSERTV(zio_prop_t *zp = &zio->io_prop);
1993 ASSERT(ddp->ddp_refcnt == 0);
1994 ASSERT(dde->dde_lead_zio[p] == zio);
1995 dde->dde_lead_zio[p] = NULL;
1997 if (zio->io_error == 0) {
1998 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
1999 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2000 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2001 if (ddp->ddp_phys_birth != 0)
2002 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2003 ddt_phys_fill(ddp, bp);
2010 zio_ddt_write(zio_t *zio)
2012 spa_t *spa = zio->io_spa;
2013 blkptr_t *bp = zio->io_bp;
2014 uint64_t txg = zio->io_txg;
2015 zio_prop_t *zp = &zio->io_prop;
2016 int p = zp->zp_copies;
2020 ddt_t *ddt = ddt_select(spa, bp);
2024 ASSERT(BP_GET_DEDUP(bp));
2025 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2026 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2029 dde = ddt_lookup(ddt, bp, B_TRUE);
2030 ddp = &dde->dde_phys[p];
2032 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2034 * If we're using a weak checksum, upgrade to a strong checksum
2035 * and try again. If we're already using a strong checksum,
2036 * we can't resolve it, so just convert to an ordinary write.
2037 * (And automatically e-mail a paper to Nature?)
2039 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2040 zp->zp_checksum = spa_dedup_checksum(spa);
2041 zio_pop_transforms(zio);
2042 zio->io_stage = ZIO_STAGE_OPEN;
2047 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2049 return (ZIO_PIPELINE_CONTINUE);
2052 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2053 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2055 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2056 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2057 zio_prop_t czp = *zp;
2059 czp.zp_copies = ditto_copies;
2062 * If we arrived here with an override bp, we won't have run
2063 * the transform stack, so we won't have the data we need to
2064 * generate a child i/o. So, toss the override bp and restart.
2065 * This is safe, because using the override bp is just an
2066 * optimization; and it's rare, so the cost doesn't matter.
2068 if (zio->io_bp_override) {
2069 zio_pop_transforms(zio);
2070 zio->io_stage = ZIO_STAGE_OPEN;
2071 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2072 zio->io_bp_override = NULL;
2075 return (ZIO_PIPELINE_CONTINUE);
2078 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2079 zio->io_orig_size, &czp, NULL,
2080 zio_ddt_ditto_write_done, dde, zio->io_priority,
2081 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2083 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2084 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2087 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2088 if (ddp->ddp_phys_birth != 0)
2089 ddt_bp_fill(ddp, bp, txg);
2090 if (dde->dde_lead_zio[p] != NULL)
2091 zio_add_child(zio, dde->dde_lead_zio[p]);
2093 ddt_phys_addref(ddp);
2094 } else if (zio->io_bp_override) {
2095 ASSERT(bp->blk_birth == txg);
2096 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2097 ddt_phys_fill(ddp, bp);
2098 ddt_phys_addref(ddp);
2100 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2101 zio->io_orig_size, zp, zio_ddt_child_write_ready,
2102 zio_ddt_child_write_done, dde, zio->io_priority,
2103 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2105 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2106 dde->dde_lead_zio[p] = cio;
2116 return (ZIO_PIPELINE_CONTINUE);
2119 ddt_entry_t *freedde; /* for debugging */
2122 zio_ddt_free(zio_t *zio)
2124 spa_t *spa = zio->io_spa;
2125 blkptr_t *bp = zio->io_bp;
2126 ddt_t *ddt = ddt_select(spa, bp);
2130 ASSERT(BP_GET_DEDUP(bp));
2131 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2134 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2135 ddp = ddt_phys_select(dde, bp);
2136 ddt_phys_decref(ddp);
2139 return (ZIO_PIPELINE_CONTINUE);
2143 * ==========================================================================
2144 * Allocate and free blocks
2145 * ==========================================================================
2148 zio_dva_allocate(zio_t *zio)
2150 spa_t *spa = zio->io_spa;
2151 metaslab_class_t *mc = spa_normal_class(spa);
2152 blkptr_t *bp = zio->io_bp;
2155 if (zio->io_gang_leader == NULL) {
2156 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2157 zio->io_gang_leader = zio;
2160 ASSERT(BP_IS_HOLE(bp));
2161 ASSERT3U(BP_GET_NDVAS(bp), ==, 0);
2162 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2163 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2164 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2166 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2167 zio->io_prop.zp_copies, zio->io_txg, NULL, 0);
2170 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2171 return (zio_write_gang_block(zio));
2172 zio->io_error = error;
2175 return (ZIO_PIPELINE_CONTINUE);
2179 zio_dva_free(zio_t *zio)
2181 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2183 return (ZIO_PIPELINE_CONTINUE);
2187 zio_dva_claim(zio_t *zio)
2191 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2193 zio->io_error = error;
2195 return (ZIO_PIPELINE_CONTINUE);
2199 * Undo an allocation. This is used by zio_done() when an I/O fails
2200 * and we want to give back the block we just allocated.
2201 * This handles both normal blocks and gang blocks.
2204 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2208 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2209 ASSERT(zio->io_bp_override == NULL);
2211 if (!BP_IS_HOLE(bp))
2212 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2215 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2216 zio_dva_unallocate(zio, gn->gn_child[g],
2217 &gn->gn_gbh->zg_blkptr[g]);
2223 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2226 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2227 uint64_t size, boolean_t use_slog)
2231 ASSERT(txg > spa_syncing_txg(spa));
2234 error = metaslab_alloc(spa, spa_log_class(spa), size,
2235 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2238 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2239 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2242 BP_SET_LSIZE(new_bp, size);
2243 BP_SET_PSIZE(new_bp, size);
2244 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2245 BP_SET_CHECKSUM(new_bp,
2246 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2247 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2248 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2249 BP_SET_LEVEL(new_bp, 0);
2250 BP_SET_DEDUP(new_bp, 0);
2251 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2258 * Free an intent log block.
2261 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2263 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2264 ASSERT(!BP_IS_GANG(bp));
2266 zio_free(spa, txg, bp);
2270 * ==========================================================================
2271 * Read and write to physical devices
2272 * ==========================================================================
2275 zio_vdev_io_start(zio_t *zio)
2277 vdev_t *vd = zio->io_vd;
2279 spa_t *spa = zio->io_spa;
2281 ASSERT(zio->io_error == 0);
2282 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2285 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2286 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2289 * The mirror_ops handle multiple DVAs in a single BP.
2291 return (vdev_mirror_ops.vdev_op_io_start(zio));
2295 * We keep track of time-sensitive I/Os so that the scan thread
2296 * can quickly react to certain workloads. In particular, we care
2297 * about non-scrubbing, top-level reads and writes with the following
2299 * - synchronous writes of user data to non-slog devices
2300 * - any reads of user data
2301 * When these conditions are met, adjust the timestamp of spa_last_io
2302 * which allows the scan thread to adjust its workload accordingly.
2304 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2305 vd == vd->vdev_top && !vd->vdev_islog &&
2306 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2307 zio->io_txg != spa_syncing_txg(spa)) {
2308 uint64_t old = spa->spa_last_io;
2309 uint64_t new = ddi_get_lbolt64();
2311 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2314 align = 1ULL << vd->vdev_top->vdev_ashift;
2316 if (P2PHASE(zio->io_size, align) != 0) {
2317 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2318 char *abuf = zio_buf_alloc(asize);
2319 ASSERT(vd == vd->vdev_top);
2320 if (zio->io_type == ZIO_TYPE_WRITE) {
2321 bcopy(zio->io_data, abuf, zio->io_size);
2322 bzero(abuf + zio->io_size, asize - zio->io_size);
2324 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2327 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2328 ASSERT(P2PHASE(zio->io_size, align) == 0);
2329 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2332 * If this is a repair I/O, and there's no self-healing involved --
2333 * that is, we're just resilvering what we expect to resilver --
2334 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2335 * This prevents spurious resilvering with nested replication.
2336 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2337 * A is out of date, we'll read from C+D, then use the data to
2338 * resilver A+B -- but we don't actually want to resilver B, just A.
2339 * The top-level mirror has no way to know this, so instead we just
2340 * discard unnecessary repairs as we work our way down the vdev tree.
2341 * The same logic applies to any form of nested replication:
2342 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2344 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2345 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2346 zio->io_txg != 0 && /* not a delegated i/o */
2347 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2348 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2349 zio_vdev_io_bypass(zio);
2350 return (ZIO_PIPELINE_CONTINUE);
2353 if (vd->vdev_ops->vdev_op_leaf &&
2354 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2356 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2357 return (ZIO_PIPELINE_CONTINUE);
2359 if ((zio = vdev_queue_io(zio)) == NULL)
2360 return (ZIO_PIPELINE_STOP);
2362 if (!vdev_accessible(vd, zio)) {
2363 zio->io_error = ENXIO;
2365 return (ZIO_PIPELINE_STOP);
2369 return (vd->vdev_ops->vdev_op_io_start(zio));
2373 zio_vdev_io_done(zio_t *zio)
2375 vdev_t *vd = zio->io_vd;
2376 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2377 boolean_t unexpected_error = B_FALSE;
2379 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2380 return (ZIO_PIPELINE_STOP);
2382 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2384 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2386 vdev_queue_io_done(zio);
2388 if (zio->io_type == ZIO_TYPE_WRITE)
2389 vdev_cache_write(zio);
2391 if (zio_injection_enabled && zio->io_error == 0)
2392 zio->io_error = zio_handle_device_injection(vd,
2395 if (zio_injection_enabled && zio->io_error == 0)
2396 zio->io_error = zio_handle_label_injection(zio, EIO);
2398 if (zio->io_error) {
2399 if (!vdev_accessible(vd, zio)) {
2400 zio->io_error = ENXIO;
2402 unexpected_error = B_TRUE;
2407 ops->vdev_op_io_done(zio);
2409 if (unexpected_error)
2410 VERIFY(vdev_probe(vd, zio) == NULL);
2412 return (ZIO_PIPELINE_CONTINUE);
2416 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2417 * disk, and use that to finish the checksum ereport later.
2420 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2421 const void *good_buf)
2423 /* no processing needed */
2424 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2429 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2431 void *buf = zio_buf_alloc(zio->io_size);
2433 bcopy(zio->io_data, buf, zio->io_size);
2435 zcr->zcr_cbinfo = zio->io_size;
2436 zcr->zcr_cbdata = buf;
2437 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2438 zcr->zcr_free = zio_buf_free;
2442 zio_vdev_io_assess(zio_t *zio)
2444 vdev_t *vd = zio->io_vd;
2446 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2447 return (ZIO_PIPELINE_STOP);
2449 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2450 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2452 if (zio->io_vsd != NULL) {
2453 zio->io_vsd_ops->vsd_free(zio);
2457 if (zio_injection_enabled && zio->io_error == 0)
2458 zio->io_error = zio_handle_fault_injection(zio, EIO);
2461 * If the I/O failed, determine whether we should attempt to retry it.
2463 * On retry, we cut in line in the issue queue, since we don't want
2464 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2466 if (zio->io_error && vd == NULL &&
2467 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2468 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2469 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2471 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2472 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2473 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2474 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2475 zio_requeue_io_start_cut_in_line);
2476 return (ZIO_PIPELINE_STOP);
2480 * If we got an error on a leaf device, convert it to ENXIO
2481 * if the device is not accessible at all.
2483 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2484 !vdev_accessible(vd, zio))
2485 zio->io_error = ENXIO;
2488 * If we can't write to an interior vdev (mirror or RAID-Z),
2489 * set vdev_cant_write so that we stop trying to allocate from it.
2491 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2492 vd != NULL && !vd->vdev_ops->vdev_op_leaf)
2493 vd->vdev_cant_write = B_TRUE;
2496 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2498 return (ZIO_PIPELINE_CONTINUE);
2502 zio_vdev_io_reissue(zio_t *zio)
2504 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2505 ASSERT(zio->io_error == 0);
2507 zio->io_stage >>= 1;
2511 zio_vdev_io_redone(zio_t *zio)
2513 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2515 zio->io_stage >>= 1;
2519 zio_vdev_io_bypass(zio_t *zio)
2521 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2522 ASSERT(zio->io_error == 0);
2524 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2525 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2529 * ==========================================================================
2530 * Generate and verify checksums
2531 * ==========================================================================
2534 zio_checksum_generate(zio_t *zio)
2536 blkptr_t *bp = zio->io_bp;
2537 enum zio_checksum checksum;
2541 * This is zio_write_phys().
2542 * We're either generating a label checksum, or none at all.
2544 checksum = zio->io_prop.zp_checksum;
2546 if (checksum == ZIO_CHECKSUM_OFF)
2547 return (ZIO_PIPELINE_CONTINUE);
2549 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2551 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2552 ASSERT(!IO_IS_ALLOCATING(zio));
2553 checksum = ZIO_CHECKSUM_GANG_HEADER;
2555 checksum = BP_GET_CHECKSUM(bp);
2559 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2561 return (ZIO_PIPELINE_CONTINUE);
2565 zio_checksum_verify(zio_t *zio)
2567 zio_bad_cksum_t info;
2568 blkptr_t *bp = zio->io_bp;
2571 ASSERT(zio->io_vd != NULL);
2575 * This is zio_read_phys().
2576 * We're either verifying a label checksum, or nothing at all.
2578 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2579 return (ZIO_PIPELINE_CONTINUE);
2581 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2584 if ((error = zio_checksum_error(zio, &info)) != 0) {
2585 zio->io_error = error;
2586 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2587 zfs_ereport_start_checksum(zio->io_spa,
2588 zio->io_vd, zio, zio->io_offset,
2589 zio->io_size, NULL, &info);
2593 return (ZIO_PIPELINE_CONTINUE);
2597 * Called by RAID-Z to ensure we don't compute the checksum twice.
2600 zio_checksum_verified(zio_t *zio)
2602 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2606 * ==========================================================================
2607 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2608 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2609 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2610 * indicate errors that are specific to one I/O, and most likely permanent.
2611 * Any other error is presumed to be worse because we weren't expecting it.
2612 * ==========================================================================
2615 zio_worst_error(int e1, int e2)
2617 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2620 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2621 if (e1 == zio_error_rank[r1])
2624 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2625 if (e2 == zio_error_rank[r2])
2628 return (r1 > r2 ? e1 : e2);
2632 * ==========================================================================
2634 * ==========================================================================
2637 zio_ready(zio_t *zio)
2639 blkptr_t *bp = zio->io_bp;
2640 zio_t *pio, *pio_next;
2642 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2643 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2644 return (ZIO_PIPELINE_STOP);
2646 if (zio->io_ready) {
2647 ASSERT(IO_IS_ALLOCATING(zio));
2648 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2649 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2654 if (bp != NULL && bp != &zio->io_bp_copy)
2655 zio->io_bp_copy = *bp;
2658 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2660 mutex_enter(&zio->io_lock);
2661 zio->io_state[ZIO_WAIT_READY] = 1;
2662 pio = zio_walk_parents(zio);
2663 mutex_exit(&zio->io_lock);
2666 * As we notify zio's parents, new parents could be added.
2667 * New parents go to the head of zio's io_parent_list, however,
2668 * so we will (correctly) not notify them. The remainder of zio's
2669 * io_parent_list, from 'pio_next' onward, cannot change because
2670 * all parents must wait for us to be done before they can be done.
2672 for (; pio != NULL; pio = pio_next) {
2673 pio_next = zio_walk_parents(zio);
2674 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2677 if (zio->io_flags & ZIO_FLAG_NODATA) {
2678 if (BP_IS_GANG(bp)) {
2679 zio->io_flags &= ~ZIO_FLAG_NODATA;
2681 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2682 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2686 if (zio_injection_enabled &&
2687 zio->io_spa->spa_syncing_txg == zio->io_txg)
2688 zio_handle_ignored_writes(zio);
2690 return (ZIO_PIPELINE_CONTINUE);
2694 zio_done(zio_t *zio)
2696 zio_t *pio, *pio_next;
2700 * If our children haven't all completed,
2701 * wait for them and then repeat this pipeline stage.
2703 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2704 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2705 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2706 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2707 return (ZIO_PIPELINE_STOP);
2709 for (c = 0; c < ZIO_CHILD_TYPES; c++)
2710 for (w = 0; w < ZIO_WAIT_TYPES; w++)
2711 ASSERT(zio->io_children[c][w] == 0);
2713 if (zio->io_bp != NULL) {
2714 ASSERT(zio->io_bp->blk_pad[0] == 0);
2715 ASSERT(zio->io_bp->blk_pad[1] == 0);
2716 ASSERT(bcmp(zio->io_bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2717 (zio->io_bp == zio_unique_parent(zio)->io_bp));
2718 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) &&
2719 zio->io_bp_override == NULL &&
2720 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2721 ASSERT(!BP_SHOULD_BYTESWAP(zio->io_bp));
2722 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2723 ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 ||
2724 (BP_COUNT_GANG(zio->io_bp) == BP_GET_NDVAS(zio->io_bp)));
2729 * If there were child vdev/gang/ddt errors, they apply to us now.
2731 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2732 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2733 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2736 * If the I/O on the transformed data was successful, generate any
2737 * checksum reports now while we still have the transformed data.
2739 if (zio->io_error == 0) {
2740 while (zio->io_cksum_report != NULL) {
2741 zio_cksum_report_t *zcr = zio->io_cksum_report;
2742 uint64_t align = zcr->zcr_align;
2743 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2744 char *abuf = zio->io_data;
2746 if (asize != zio->io_size) {
2747 abuf = zio_buf_alloc(asize);
2748 bcopy(zio->io_data, abuf, zio->io_size);
2749 bzero(abuf + zio->io_size, asize - zio->io_size);
2752 zio->io_cksum_report = zcr->zcr_next;
2753 zcr->zcr_next = NULL;
2754 zcr->zcr_finish(zcr, abuf);
2755 zfs_ereport_free_checksum(zcr);
2757 if (asize != zio->io_size)
2758 zio_buf_free(abuf, asize);
2762 zio_pop_transforms(zio); /* note: may set zio->io_error */
2764 vdev_stat_update(zio, zio->io_size);
2767 * If this I/O is attached to a particular vdev is slow, exeeding
2768 * 30 seconds to complete, post an error described the I/O delay.
2769 * We ignore these errors if the device is currently unavailable.
2771 if (zio->io_delay >= zio_delay_max) {
2772 if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd))
2773 zfs_ereport_post(FM_EREPORT_ZFS_DELAY, zio->io_spa,
2774 zio->io_vd, zio, 0, 0);
2777 if (zio->io_error) {
2779 * If this I/O is attached to a particular vdev,
2780 * generate an error message describing the I/O failure
2781 * at the block level. We ignore these errors if the
2782 * device is currently unavailable.
2784 if (zio->io_error != ECKSUM && zio->io_vd != NULL &&
2785 !vdev_is_dead(zio->io_vd))
2786 zfs_ereport_post(FM_EREPORT_ZFS_IO, zio->io_spa,
2787 zio->io_vd, zio, 0, 0);
2789 if ((zio->io_error == EIO || !(zio->io_flags &
2790 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2791 zio == zio->io_logical) {
2793 * For logical I/O requests, tell the SPA to log the
2794 * error and generate a logical data ereport.
2796 spa_log_error(zio->io_spa, zio);
2797 zfs_ereport_post(FM_EREPORT_ZFS_DATA, zio->io_spa, NULL, zio,
2802 if (zio->io_error && zio == zio->io_logical) {
2804 * Determine whether zio should be reexecuted. This will
2805 * propagate all the way to the root via zio_notify_parent().
2807 ASSERT(zio->io_vd == NULL && zio->io_bp != NULL);
2808 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2810 if (IO_IS_ALLOCATING(zio) &&
2811 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2812 if (zio->io_error != ENOSPC)
2813 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2815 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2818 if ((zio->io_type == ZIO_TYPE_READ ||
2819 zio->io_type == ZIO_TYPE_FREE) &&
2820 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
2821 zio->io_error == ENXIO &&
2822 spa_load_state(zio->io_spa) == SPA_LOAD_NONE &&
2823 spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE)
2824 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2826 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2827 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2830 * Here is a possibly good place to attempt to do
2831 * either combinatorial reconstruction or error correction
2832 * based on checksums. It also might be a good place
2833 * to send out preliminary ereports before we suspend
2839 * If there were logical child errors, they apply to us now.
2840 * We defer this until now to avoid conflating logical child
2841 * errors with errors that happened to the zio itself when
2842 * updating vdev stats and reporting FMA events above.
2844 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2846 if ((zio->io_error || zio->io_reexecute) &&
2847 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
2848 !(zio->io_flags & ZIO_FLAG_IO_REWRITE))
2849 zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp);
2851 zio_gang_tree_free(&zio->io_gang_tree);
2854 * Godfather I/Os should never suspend.
2856 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
2857 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
2858 zio->io_reexecute = 0;
2860 if (zio->io_reexecute) {
2862 * This is a logical I/O that wants to reexecute.
2864 * Reexecute is top-down. When an i/o fails, if it's not
2865 * the root, it simply notifies its parent and sticks around.
2866 * The parent, seeing that it still has children in zio_done(),
2867 * does the same. This percolates all the way up to the root.
2868 * The root i/o will reexecute or suspend the entire tree.
2870 * This approach ensures that zio_reexecute() honors
2871 * all the original i/o dependency relationships, e.g.
2872 * parents not executing until children are ready.
2874 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2876 zio->io_gang_leader = NULL;
2878 mutex_enter(&zio->io_lock);
2879 zio->io_state[ZIO_WAIT_DONE] = 1;
2880 mutex_exit(&zio->io_lock);
2883 * "The Godfather" I/O monitors its children but is
2884 * not a true parent to them. It will track them through
2885 * the pipeline but severs its ties whenever they get into
2886 * trouble (e.g. suspended). This allows "The Godfather"
2887 * I/O to return status without blocking.
2889 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2890 zio_link_t *zl = zio->io_walk_link;
2891 pio_next = zio_walk_parents(zio);
2893 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
2894 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
2895 zio_remove_child(pio, zio, zl);
2896 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2900 if ((pio = zio_unique_parent(zio)) != NULL) {
2902 * We're not a root i/o, so there's nothing to do
2903 * but notify our parent. Don't propagate errors
2904 * upward since we haven't permanently failed yet.
2906 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2907 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
2908 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2909 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
2911 * We'd fail again if we reexecuted now, so suspend
2912 * until conditions improve (e.g. device comes online).
2914 zio_suspend(zio->io_spa, zio);
2917 * Reexecution is potentially a huge amount of work.
2918 * Hand it off to the otherwise-unused claim taskq.
2920 (void) taskq_dispatch(
2921 zio->io_spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
2922 (task_func_t *)zio_reexecute, zio, TQ_SLEEP);
2924 return (ZIO_PIPELINE_STOP);
2927 ASSERT(zio->io_child_count == 0);
2928 ASSERT(zio->io_reexecute == 0);
2929 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
2932 * Report any checksum errors, since the I/O is complete.
2934 while (zio->io_cksum_report != NULL) {
2935 zio_cksum_report_t *zcr = zio->io_cksum_report;
2936 zio->io_cksum_report = zcr->zcr_next;
2937 zcr->zcr_next = NULL;
2938 zcr->zcr_finish(zcr, NULL);
2939 zfs_ereport_free_checksum(zcr);
2943 * It is the responsibility of the done callback to ensure that this
2944 * particular zio is no longer discoverable for adoption, and as
2945 * such, cannot acquire any new parents.
2950 mutex_enter(&zio->io_lock);
2951 zio->io_state[ZIO_WAIT_DONE] = 1;
2952 mutex_exit(&zio->io_lock);
2954 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2955 zio_link_t *zl = zio->io_walk_link;
2956 pio_next = zio_walk_parents(zio);
2957 zio_remove_child(pio, zio, zl);
2958 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2961 if (zio->io_waiter != NULL) {
2962 mutex_enter(&zio->io_lock);
2963 zio->io_executor = NULL;
2964 cv_broadcast(&zio->io_cv);
2965 mutex_exit(&zio->io_lock);
2970 return (ZIO_PIPELINE_STOP);
2974 * ==========================================================================
2975 * I/O pipeline definition
2976 * ==========================================================================
2978 static zio_pipe_stage_t *zio_pipeline[] = {
2984 zio_checksum_generate,
2998 zio_checksum_verify,
3002 #if defined(_KERNEL) && defined(HAVE_SPL)
3003 /* Fault injection */
3004 EXPORT_SYMBOL(zio_injection_enabled);
3005 EXPORT_SYMBOL(zio_inject_fault);
3006 EXPORT_SYMBOL(zio_inject_list_next);
3007 EXPORT_SYMBOL(zio_clear_fault);
3008 EXPORT_SYMBOL(zio_handle_fault_injection);
3009 EXPORT_SYMBOL(zio_handle_device_injection);
3010 EXPORT_SYMBOL(zio_handle_label_injection);
3011 EXPORT_SYMBOL(zio_priority_table);
3012 EXPORT_SYMBOL(zio_type_name);
3014 module_param(zio_bulk_flags, int, 0644);
3015 MODULE_PARM_DESC(zio_bulk_flags, "Additional flags to pass to bulk buffers");
3017 module_param(zio_delay_max, int, 0644);
3018 MODULE_PARM_DESC(zio_delay_max, "Max zio millisec delay before posting event");
3020 module_param(zio_requeue_io_start_cut_in_line, int, 0644);
3021 MODULE_PARM_DESC(zio_requeue_io_start_cut_in_line, "Prioritize requeued I/O");