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.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
27 #include <sys/zfs_context.h>
28 #include <sys/fm/fs/zfs.h>
31 #include <sys/spa_impl.h>
32 #include <sys/vdev_impl.h>
33 #include <sys/zio_impl.h>
34 #include <sys/zio_compress.h>
35 #include <sys/zio_checksum.h>
36 #include <sys/dmu_objset.h>
41 * ==========================================================================
43 * ==========================================================================
45 uint8_t zio_priority_table[ZIO_PRIORITY_TABLE_SIZE] = {
46 0, /* ZIO_PRIORITY_NOW */
47 0, /* ZIO_PRIORITY_SYNC_READ */
48 0, /* ZIO_PRIORITY_SYNC_WRITE */
49 0, /* ZIO_PRIORITY_LOG_WRITE */
50 1, /* ZIO_PRIORITY_CACHE_FILL */
51 1, /* ZIO_PRIORITY_AGG */
52 4, /* ZIO_PRIORITY_FREE */
53 4, /* ZIO_PRIORITY_ASYNC_WRITE */
54 6, /* ZIO_PRIORITY_ASYNC_READ */
55 10, /* ZIO_PRIORITY_RESILVER */
56 20, /* ZIO_PRIORITY_SCRUB */
57 2, /* ZIO_PRIORITY_DDT_PREFETCH */
61 * ==========================================================================
62 * I/O type descriptions
63 * ==========================================================================
65 char *zio_type_name[ZIO_TYPES] = {
66 "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl"
70 * ==========================================================================
72 * ==========================================================================
74 kmem_cache_t *zio_cache;
75 kmem_cache_t *zio_link_cache;
76 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
77 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
78 int zio_bulk_flags = 0;
79 int zio_delay_max = ZIO_DELAY_MAX;
82 extern vmem_t *zio_alloc_arena;
84 extern int zfs_mg_alloc_failures;
87 * An allocating zio is one that either currently has the DVA allocate
88 * stage set or will have it later in its lifetime.
90 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
92 int zio_requeue_io_start_cut_in_line = 1;
95 int zio_buf_debug_limit = 16384;
97 int zio_buf_debug_limit = 0;
100 static inline void __zio_execute(zio_t *zio);
106 vmem_t *data_alloc_arena = NULL;
109 data_alloc_arena = zio_alloc_arena;
111 zio_cache = kmem_cache_create("zio_cache",
112 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, KMC_KMEM);
113 zio_link_cache = kmem_cache_create("zio_link_cache",
114 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, KMC_KMEM);
117 * For small buffers, we want a cache for each multiple of
118 * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache
119 * for each quarter-power of 2. For large buffers, we want
120 * a cache for each multiple of PAGESIZE.
122 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
123 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
127 while (p2 & (p2 - 1))
130 if (size <= 4 * SPA_MINBLOCKSIZE) {
131 align = SPA_MINBLOCKSIZE;
132 } else if (P2PHASE(size, PAGESIZE) == 0) {
134 } else if (P2PHASE(size, p2 >> 2) == 0) {
140 int flags = zio_bulk_flags;
143 * The smallest buffers (512b) are heavily used and
144 * experience a lot of churn. The slabs allocated
145 * for them are also relatively small (32K). Thus
146 * in over to avoid expensive calls to vmalloc() we
147 * make an exception to the usual slab allocation
148 * policy and force these buffers to be kmem backed.
150 if (size == (1 << SPA_MINBLOCKSHIFT))
153 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
154 zio_buf_cache[c] = kmem_cache_create(name, size,
155 align, NULL, NULL, NULL, NULL, NULL, flags);
157 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
158 zio_data_buf_cache[c] = kmem_cache_create(name, size,
159 align, NULL, NULL, NULL, NULL,
160 data_alloc_arena, flags);
165 ASSERT(zio_buf_cache[c] != NULL);
166 if (zio_buf_cache[c - 1] == NULL)
167 zio_buf_cache[c - 1] = zio_buf_cache[c];
169 ASSERT(zio_data_buf_cache[c] != NULL);
170 if (zio_data_buf_cache[c - 1] == NULL)
171 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
175 * The zio write taskqs have 1 thread per cpu, allow 1/2 of the taskqs
176 * to fail 3 times per txg or 8 failures, whichever is greater.
178 zfs_mg_alloc_failures = MAX((3 * max_ncpus / 2), 8);
187 kmem_cache_t *last_cache = NULL;
188 kmem_cache_t *last_data_cache = NULL;
190 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
191 if (zio_buf_cache[c] != last_cache) {
192 last_cache = zio_buf_cache[c];
193 kmem_cache_destroy(zio_buf_cache[c]);
195 zio_buf_cache[c] = NULL;
197 if (zio_data_buf_cache[c] != last_data_cache) {
198 last_data_cache = zio_data_buf_cache[c];
199 kmem_cache_destroy(zio_data_buf_cache[c]);
201 zio_data_buf_cache[c] = NULL;
204 kmem_cache_destroy(zio_link_cache);
205 kmem_cache_destroy(zio_cache);
211 * ==========================================================================
212 * Allocate and free I/O buffers
213 * ==========================================================================
217 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
218 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
219 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
220 * excess / transient data in-core during a crashdump.
223 zio_buf_alloc(size_t size)
225 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
227 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
229 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
233 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
234 * crashdump if the kernel panics. This exists so that we will limit the amount
235 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
236 * of kernel heap dumped to disk when the kernel panics)
239 zio_data_buf_alloc(size_t size)
241 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
243 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
245 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
249 zio_buf_free(void *buf, size_t size)
251 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
253 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
255 kmem_cache_free(zio_buf_cache[c], buf);
259 zio_data_buf_free(void *buf, size_t size)
261 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
263 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
265 kmem_cache_free(zio_data_buf_cache[c], buf);
269 * ==========================================================================
270 * Push and pop I/O transform buffers
271 * ==========================================================================
274 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
275 zio_transform_func_t *transform)
277 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_PUSHPAGE);
279 zt->zt_orig_data = zio->io_data;
280 zt->zt_orig_size = zio->io_size;
281 zt->zt_bufsize = bufsize;
282 zt->zt_transform = transform;
284 zt->zt_next = zio->io_transform_stack;
285 zio->io_transform_stack = zt;
292 zio_pop_transforms(zio_t *zio)
296 while ((zt = zio->io_transform_stack) != NULL) {
297 if (zt->zt_transform != NULL)
298 zt->zt_transform(zio,
299 zt->zt_orig_data, zt->zt_orig_size);
301 if (zt->zt_bufsize != 0)
302 zio_buf_free(zio->io_data, zt->zt_bufsize);
304 zio->io_data = zt->zt_orig_data;
305 zio->io_size = zt->zt_orig_size;
306 zio->io_transform_stack = zt->zt_next;
308 kmem_free(zt, sizeof (zio_transform_t));
313 * ==========================================================================
314 * I/O transform callbacks for subblocks and decompression
315 * ==========================================================================
318 zio_subblock(zio_t *zio, void *data, uint64_t size)
320 ASSERT(zio->io_size > size);
322 if (zio->io_type == ZIO_TYPE_READ)
323 bcopy(zio->io_data, data, size);
327 zio_decompress(zio_t *zio, void *data, uint64_t size)
329 if (zio->io_error == 0 &&
330 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
331 zio->io_data, data, zio->io_size, size) != 0)
336 * ==========================================================================
337 * I/O parent/child relationships and pipeline interlocks
338 * ==========================================================================
341 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
342 * continue calling these functions until they return NULL.
343 * Otherwise, the next caller will pick up the list walk in
344 * some indeterminate state. (Otherwise every caller would
345 * have to pass in a cookie to keep the state represented by
346 * io_walk_link, which gets annoying.)
349 zio_walk_parents(zio_t *cio)
351 zio_link_t *zl = cio->io_walk_link;
352 list_t *pl = &cio->io_parent_list;
354 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
355 cio->io_walk_link = zl;
360 ASSERT(zl->zl_child == cio);
361 return (zl->zl_parent);
365 zio_walk_children(zio_t *pio)
367 zio_link_t *zl = pio->io_walk_link;
368 list_t *cl = &pio->io_child_list;
370 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
371 pio->io_walk_link = zl;
376 ASSERT(zl->zl_parent == pio);
377 return (zl->zl_child);
381 zio_unique_parent(zio_t *cio)
383 zio_t *pio = zio_walk_parents(cio);
385 VERIFY(zio_walk_parents(cio) == NULL);
390 zio_add_child(zio_t *pio, zio_t *cio)
392 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_PUSHPAGE);
396 * Logical I/Os can have logical, gang, or vdev children.
397 * Gang I/Os can have gang or vdev children.
398 * Vdev I/Os can only have vdev children.
399 * The following ASSERT captures all of these constraints.
401 ASSERT(cio->io_child_type <= pio->io_child_type);
406 mutex_enter(&cio->io_lock);
407 mutex_enter(&pio->io_lock);
409 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
411 for (w = 0; w < ZIO_WAIT_TYPES; w++)
412 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
414 list_insert_head(&pio->io_child_list, zl);
415 list_insert_head(&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);
425 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
427 ASSERT(zl->zl_parent == pio);
428 ASSERT(zl->zl_child == cio);
430 mutex_enter(&cio->io_lock);
431 mutex_enter(&pio->io_lock);
433 list_remove(&pio->io_child_list, zl);
434 list_remove(&cio->io_parent_list, zl);
436 pio->io_child_count--;
437 cio->io_parent_count--;
439 mutex_exit(&pio->io_lock);
440 mutex_exit(&cio->io_lock);
442 kmem_cache_free(zio_link_cache, zl);
446 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
448 uint64_t *countp = &zio->io_children[child][wait];
449 boolean_t waiting = B_FALSE;
451 mutex_enter(&zio->io_lock);
452 ASSERT(zio->io_stall == NULL);
455 zio->io_stall = countp;
458 mutex_exit(&zio->io_lock);
463 __attribute__((always_inline))
465 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
467 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
468 int *errorp = &pio->io_child_error[zio->io_child_type];
470 mutex_enter(&pio->io_lock);
471 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
472 *errorp = zio_worst_error(*errorp, zio->io_error);
473 pio->io_reexecute |= zio->io_reexecute;
474 ASSERT3U(*countp, >, 0);
475 if (--*countp == 0 && pio->io_stall == countp) {
476 pio->io_stall = NULL;
477 mutex_exit(&pio->io_lock);
480 mutex_exit(&pio->io_lock);
485 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
487 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
488 zio->io_error = zio->io_child_error[c];
492 * ==========================================================================
493 * Create the various types of I/O (read, write, free, etc)
494 * ==========================================================================
497 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
498 void *data, uint64_t size, zio_done_func_t *done, void *private,
499 zio_type_t type, int priority, enum zio_flag flags,
500 vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
501 enum zio_stage stage, enum zio_stage pipeline)
505 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
506 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
507 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
509 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
510 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
511 ASSERT(vd || stage == ZIO_STAGE_OPEN);
513 zio = kmem_cache_alloc(zio_cache, KM_PUSHPAGE);
514 bzero(zio, sizeof (zio_t));
516 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
517 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
519 list_create(&zio->io_parent_list, sizeof (zio_link_t),
520 offsetof(zio_link_t, zl_parent_node));
521 list_create(&zio->io_child_list, sizeof (zio_link_t),
522 offsetof(zio_link_t, zl_child_node));
525 zio->io_child_type = ZIO_CHILD_VDEV;
526 else if (flags & ZIO_FLAG_GANG_CHILD)
527 zio->io_child_type = ZIO_CHILD_GANG;
528 else if (flags & ZIO_FLAG_DDT_CHILD)
529 zio->io_child_type = ZIO_CHILD_DDT;
531 zio->io_child_type = ZIO_CHILD_LOGICAL;
534 zio->io_bp = (blkptr_t *)bp;
535 zio->io_bp_copy = *bp;
536 zio->io_bp_orig = *bp;
537 if (type != ZIO_TYPE_WRITE ||
538 zio->io_child_type == ZIO_CHILD_DDT)
539 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
540 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
541 zio->io_logical = zio;
542 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
543 pipeline |= ZIO_GANG_STAGES;
549 zio->io_private = private;
551 zio->io_priority = priority;
553 zio->io_offset = offset;
554 zio->io_orig_data = zio->io_data = data;
555 zio->io_orig_size = zio->io_size = size;
556 zio->io_orig_flags = zio->io_flags = flags;
557 zio->io_orig_stage = zio->io_stage = stage;
558 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
560 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
561 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
564 zio->io_bookmark = *zb;
567 if (zio->io_logical == NULL)
568 zio->io_logical = pio->io_logical;
569 if (zio->io_child_type == ZIO_CHILD_GANG)
570 zio->io_gang_leader = pio->io_gang_leader;
571 zio_add_child(pio, zio);
574 taskq_init_ent(&zio->io_tqent);
580 zio_destroy(zio_t *zio)
582 list_destroy(&zio->io_parent_list);
583 list_destroy(&zio->io_child_list);
584 mutex_destroy(&zio->io_lock);
585 cv_destroy(&zio->io_cv);
586 kmem_cache_free(zio_cache, zio);
590 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
591 void *private, enum zio_flag flags)
595 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
596 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
597 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
603 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
605 return (zio_null(NULL, spa, NULL, done, private, flags));
609 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
610 void *data, uint64_t size, zio_done_func_t *done, void *private,
611 int priority, enum zio_flag flags, const zbookmark_t *zb)
615 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
616 data, size, done, private,
617 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
618 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
619 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
625 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
626 void *data, uint64_t size, const zio_prop_t *zp,
627 zio_done_func_t *ready, zio_done_func_t *done, void *private,
628 int priority, enum zio_flag flags, const zbookmark_t *zb)
632 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
633 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
634 zp->zp_compress >= ZIO_COMPRESS_OFF &&
635 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
636 zp->zp_type < DMU_OT_NUMTYPES &&
639 zp->zp_copies <= spa_max_replication(spa) &&
641 zp->zp_dedup_verify <= 1);
643 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
644 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
645 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
646 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
648 zio->io_ready = ready;
655 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
656 uint64_t size, zio_done_func_t *done, void *private, int priority,
657 enum zio_flag flags, zbookmark_t *zb)
661 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
662 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
663 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
669 zio_write_override(zio_t *zio, blkptr_t *bp, int copies)
671 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
672 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
673 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
674 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
676 zio->io_prop.zp_copies = copies;
677 zio->io_bp_override = bp;
681 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
683 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
687 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
692 dprintf_bp(bp, "freeing in txg %llu, pass %u",
693 (longlong_t)txg, spa->spa_sync_pass);
695 ASSERT(!BP_IS_HOLE(bp));
696 ASSERT(spa_syncing_txg(spa) == txg);
697 ASSERT(spa_sync_pass(spa) <= SYNC_PASS_DEFERRED_FREE);
699 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
700 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
701 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
707 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
708 zio_done_func_t *done, void *private, enum zio_flag flags)
713 * A claim is an allocation of a specific block. Claims are needed
714 * to support immediate writes in the intent log. The issue is that
715 * immediate writes contain committed data, but in a txg that was
716 * *not* committed. Upon opening the pool after an unclean shutdown,
717 * the intent log claims all blocks that contain immediate write data
718 * so that the SPA knows they're in use.
720 * All claims *must* be resolved in the first txg -- before the SPA
721 * starts allocating blocks -- so that nothing is allocated twice.
722 * If txg == 0 we just verify that the block is claimable.
724 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
725 ASSERT(txg == spa_first_txg(spa) || txg == 0);
726 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
728 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
729 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
730 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
736 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
737 zio_done_func_t *done, void *private, int priority, enum zio_flag flags)
742 if (vd->vdev_children == 0) {
743 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
744 ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
745 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
749 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
751 for (c = 0; c < vd->vdev_children; c++)
752 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
753 done, private, priority, flags));
760 zio_read_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_READ, priority, flags, vd, offset, NULL,
773 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
775 zio->io_prop.zp_checksum = checksum;
781 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
782 void *data, int checksum, zio_done_func_t *done, void *private,
783 int priority, enum zio_flag flags, boolean_t labels)
787 ASSERT(vd->vdev_children == 0);
788 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
789 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
790 ASSERT3U(offset + size, <=, vd->vdev_psize);
792 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
793 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
794 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
796 zio->io_prop.zp_checksum = checksum;
798 if (zio_checksum_table[checksum].ci_eck) {
800 * zec checksums are necessarily destructive -- they modify
801 * the end of the write buffer to hold the verifier/checksum.
802 * Therefore, we must make a local copy in case the data is
803 * being written to multiple places in parallel.
805 void *wbuf = zio_buf_alloc(size);
806 bcopy(data, wbuf, size);
807 zio_push_transform(zio, wbuf, size, size, NULL);
814 * Create a child I/O to do some work for us.
817 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
818 void *data, uint64_t size, int type, int priority, enum zio_flag flags,
819 zio_done_func_t *done, void *private)
821 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
824 ASSERT(vd->vdev_parent ==
825 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
827 if (type == ZIO_TYPE_READ && bp != NULL) {
829 * If we have the bp, then the child should perform the
830 * checksum and the parent need not. This pushes error
831 * detection as close to the leaves as possible and
832 * eliminates redundant checksums in the interior nodes.
834 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
835 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
838 if (vd->vdev_children == 0)
839 offset += VDEV_LABEL_START_SIZE;
841 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
844 * If we've decided to do a repair, the write is not speculative --
845 * even if the original read was.
847 if (flags & ZIO_FLAG_IO_REPAIR)
848 flags &= ~ZIO_FLAG_SPECULATIVE;
850 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
851 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
852 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
858 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
859 int type, int priority, enum zio_flag flags,
860 zio_done_func_t *done, void *private)
864 ASSERT(vd->vdev_ops->vdev_op_leaf);
866 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
867 data, size, done, private, type, priority,
868 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
870 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
876 zio_flush(zio_t *zio, vdev_t *vd)
878 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
879 NULL, NULL, ZIO_PRIORITY_NOW,
880 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
884 zio_shrink(zio_t *zio, uint64_t size)
886 ASSERT(zio->io_executor == NULL);
887 ASSERT(zio->io_orig_size == zio->io_size);
888 ASSERT(size <= zio->io_size);
891 * We don't shrink for raidz because of problems with the
892 * reconstruction when reading back less than the block size.
893 * Note, BP_IS_RAIDZ() assumes no compression.
895 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
896 if (!BP_IS_RAIDZ(zio->io_bp))
897 zio->io_orig_size = zio->io_size = size;
901 * ==========================================================================
902 * Prepare to read and write logical blocks
903 * ==========================================================================
907 zio_read_bp_init(zio_t *zio)
909 blkptr_t *bp = zio->io_bp;
911 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
912 zio->io_child_type == ZIO_CHILD_LOGICAL &&
913 !(zio->io_flags & ZIO_FLAG_RAW)) {
914 uint64_t psize = BP_GET_PSIZE(bp);
915 void *cbuf = zio_buf_alloc(psize);
917 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
920 if (!dmu_ot[BP_GET_TYPE(bp)].ot_metadata && BP_GET_LEVEL(bp) == 0)
921 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
923 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
924 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
926 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
927 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
929 return (ZIO_PIPELINE_CONTINUE);
933 zio_write_bp_init(zio_t *zio)
935 spa_t *spa = zio->io_spa;
936 zio_prop_t *zp = &zio->io_prop;
937 enum zio_compress compress = zp->zp_compress;
938 blkptr_t *bp = zio->io_bp;
939 uint64_t lsize = zio->io_size;
940 uint64_t psize = lsize;
944 * If our children haven't all reached the ready stage,
945 * wait for them and then repeat this pipeline stage.
947 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
948 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
949 return (ZIO_PIPELINE_STOP);
951 if (!IO_IS_ALLOCATING(zio))
952 return (ZIO_PIPELINE_CONTINUE);
954 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
956 if (zio->io_bp_override) {
957 ASSERT(bp->blk_birth != zio->io_txg);
958 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
960 *bp = *zio->io_bp_override;
961 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
963 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
964 return (ZIO_PIPELINE_CONTINUE);
966 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
967 zp->zp_dedup_verify);
969 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
971 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
972 return (ZIO_PIPELINE_CONTINUE);
974 zio->io_bp_override = NULL;
978 if (bp->blk_birth == zio->io_txg) {
980 * We're rewriting an existing block, which means we're
981 * working on behalf of spa_sync(). For spa_sync() to
982 * converge, it must eventually be the case that we don't
983 * have to allocate new blocks. But compression changes
984 * the blocksize, which forces a reallocate, and makes
985 * convergence take longer. Therefore, after the first
986 * few passes, stop compressing to ensure convergence.
988 pass = spa_sync_pass(spa);
990 ASSERT(zio->io_txg == spa_syncing_txg(spa));
991 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
992 ASSERT(!BP_GET_DEDUP(bp));
994 if (pass > SYNC_PASS_DONT_COMPRESS)
995 compress = ZIO_COMPRESS_OFF;
997 /* Make sure someone doesn't change their mind on overwrites */
998 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
999 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1002 if (compress != ZIO_COMPRESS_OFF) {
1003 void *cbuf = zio_buf_alloc(lsize);
1004 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1005 if (psize == 0 || psize == lsize) {
1006 compress = ZIO_COMPRESS_OFF;
1007 zio_buf_free(cbuf, lsize);
1009 ASSERT(psize < lsize);
1010 zio_push_transform(zio, cbuf, psize, lsize, NULL);
1015 * The final pass of spa_sync() must be all rewrites, but the first
1016 * few passes offer a trade-off: allocating blocks defers convergence,
1017 * but newly allocated blocks are sequential, so they can be written
1018 * to disk faster. Therefore, we allow the first few passes of
1019 * spa_sync() to allocate new blocks, but force rewrites after that.
1020 * There should only be a handful of blocks after pass 1 in any case.
1022 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
1023 pass > SYNC_PASS_REWRITE) {
1024 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1026 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1027 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1030 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1034 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1036 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1037 BP_SET_LSIZE(bp, lsize);
1038 BP_SET_PSIZE(bp, psize);
1039 BP_SET_COMPRESS(bp, compress);
1040 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1041 BP_SET_TYPE(bp, zp->zp_type);
1042 BP_SET_LEVEL(bp, zp->zp_level);
1043 BP_SET_DEDUP(bp, zp->zp_dedup);
1044 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1046 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1047 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1048 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1052 return (ZIO_PIPELINE_CONTINUE);
1056 zio_free_bp_init(zio_t *zio)
1058 blkptr_t *bp = zio->io_bp;
1060 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1061 if (BP_GET_DEDUP(bp))
1062 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1065 return (ZIO_PIPELINE_CONTINUE);
1069 * ==========================================================================
1070 * Execute the I/O pipeline
1071 * ==========================================================================
1075 zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q, boolean_t cutinline)
1077 spa_t *spa = zio->io_spa;
1078 zio_type_t t = zio->io_type;
1079 int flags = (cutinline ? TQ_FRONT : 0);
1082 * If we're a config writer or a probe, the normal issue and
1083 * interrupt threads may all be blocked waiting for the config lock.
1084 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1086 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1090 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1092 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1096 * If this is a high priority I/O, then use the high priority taskq.
1098 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1099 spa->spa_zio_taskq[t][q + 1] != NULL)
1102 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1105 * NB: We are assuming that the zio can only be dispatched
1106 * to a single taskq at a time. It would be a grievous error
1107 * to dispatch the zio to another taskq at the same time.
1109 ASSERT(taskq_empty_ent(&zio->io_tqent));
1110 taskq_dispatch_ent(spa->spa_zio_taskq[t][q],
1111 (task_func_t *)zio_execute, zio, flags, &zio->io_tqent);
1115 zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
1117 kthread_t *executor = zio->io_executor;
1118 spa_t *spa = zio->io_spa;
1121 for (t = 0; t < ZIO_TYPES; t++)
1122 if (taskq_member(spa->spa_zio_taskq[t][q], executor))
1129 zio_issue_async(zio_t *zio)
1131 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1133 return (ZIO_PIPELINE_STOP);
1137 zio_interrupt(zio_t *zio)
1139 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1143 * Execute the I/O pipeline until one of the following occurs:
1144 * (1) the I/O completes; (2) the pipeline stalls waiting for
1145 * dependent child I/Os; (3) the I/O issues, so we're waiting
1146 * for an I/O completion interrupt; (4) the I/O is delegated by
1147 * vdev-level caching or aggregation; (5) the I/O is deferred
1148 * due to vdev-level queueing; (6) the I/O is handed off to
1149 * another thread. In all cases, the pipeline stops whenever
1150 * there's no CPU work; it never burns a thread in cv_wait().
1152 * There's no locking on io_stage because there's no legitimate way
1153 * for multiple threads to be attempting to process the same I/O.
1155 static zio_pipe_stage_t *zio_pipeline[];
1158 * zio_execute() is a wrapper around the static function
1159 * __zio_execute() so that we can force __zio_execute() to be
1160 * inlined. This reduces stack overhead which is important
1161 * because __zio_execute() is called recursively in several zio
1162 * code paths. zio_execute() itself cannot be inlined because
1163 * it is externally visible.
1166 zio_execute(zio_t *zio)
1171 __attribute__((always_inline))
1173 __zio_execute(zio_t *zio)
1175 zio->io_executor = curthread;
1177 while (zio->io_stage < ZIO_STAGE_DONE) {
1178 enum zio_stage pipeline = zio->io_pipeline;
1179 enum zio_stage stage = zio->io_stage;
1184 ASSERT(!MUTEX_HELD(&zio->io_lock));
1185 ASSERT(ISP2(stage));
1186 ASSERT(zio->io_stall == NULL);
1190 } while ((stage & pipeline) == 0);
1192 ASSERT(stage <= ZIO_STAGE_DONE);
1194 dsl = spa_get_dsl(zio->io_spa);
1195 cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1196 zio_requeue_io_start_cut_in_line : B_FALSE;
1199 * If we are in interrupt context and this pipeline stage
1200 * will grab a config lock that is held across I/O,
1201 * or may wait for an I/O that needs an interrupt thread
1202 * to complete, issue async to avoid deadlock.
1204 * If we are in the txg_sync_thread or being called
1205 * during pool init issue async to minimize stack depth.
1206 * Both of these call paths may be recursively called.
1208 * For VDEV_IO_START, we cut in line so that the io will
1209 * be sent to disk promptly.
1211 if (((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1212 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) ||
1213 (dsl != NULL && dsl_pool_sync_context(dsl))) {
1214 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1218 zio->io_stage = stage;
1219 rv = zio_pipeline[highbit(stage) - 1](zio);
1221 if (rv == ZIO_PIPELINE_STOP)
1224 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1230 * ==========================================================================
1231 * Initiate I/O, either sync or async
1232 * ==========================================================================
1235 zio_wait(zio_t *zio)
1239 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1240 ASSERT(zio->io_executor == NULL);
1242 zio->io_waiter = curthread;
1246 mutex_enter(&zio->io_lock);
1247 while (zio->io_executor != NULL)
1248 cv_wait(&zio->io_cv, &zio->io_lock);
1249 mutex_exit(&zio->io_lock);
1251 error = zio->io_error;
1258 zio_nowait(zio_t *zio)
1260 ASSERT(zio->io_executor == NULL);
1262 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1263 zio_unique_parent(zio) == NULL) {
1265 * This is a logical async I/O with no parent to wait for it.
1266 * We add it to the spa_async_root_zio "Godfather" I/O which
1267 * will ensure they complete prior to unloading the pool.
1269 spa_t *spa = zio->io_spa;
1271 zio_add_child(spa->spa_async_zio_root, zio);
1278 * ==========================================================================
1279 * Reexecute or suspend/resume failed I/O
1280 * ==========================================================================
1284 zio_reexecute(zio_t *pio)
1286 zio_t *cio, *cio_next;
1289 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1290 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1291 ASSERT(pio->io_gang_leader == NULL);
1292 ASSERT(pio->io_gang_tree == NULL);
1294 pio->io_flags = pio->io_orig_flags;
1295 pio->io_stage = pio->io_orig_stage;
1296 pio->io_pipeline = pio->io_orig_pipeline;
1297 pio->io_reexecute = 0;
1299 for (w = 0; w < ZIO_WAIT_TYPES; w++)
1300 pio->io_state[w] = 0;
1301 for (c = 0; c < ZIO_CHILD_TYPES; c++)
1302 pio->io_child_error[c] = 0;
1304 if (IO_IS_ALLOCATING(pio))
1305 BP_ZERO(pio->io_bp);
1308 * As we reexecute pio's children, new children could be created.
1309 * New children go to the head of pio's io_child_list, however,
1310 * so we will (correctly) not reexecute them. The key is that
1311 * the remainder of pio's io_child_list, from 'cio_next' onward,
1312 * cannot be affected by any side effects of reexecuting 'cio'.
1314 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1315 cio_next = zio_walk_children(pio);
1316 mutex_enter(&pio->io_lock);
1317 for (w = 0; w < ZIO_WAIT_TYPES; w++)
1318 pio->io_children[cio->io_child_type][w]++;
1319 mutex_exit(&pio->io_lock);
1324 * Now that all children have been reexecuted, execute the parent.
1325 * We don't reexecute "The Godfather" I/O here as it's the
1326 * responsibility of the caller to wait on him.
1328 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1333 zio_suspend(spa_t *spa, zio_t *zio)
1335 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1336 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1337 "failure and the failure mode property for this pool "
1338 "is set to panic.", spa_name(spa));
1340 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1342 mutex_enter(&spa->spa_suspend_lock);
1344 if (spa->spa_suspend_zio_root == NULL)
1345 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1346 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1347 ZIO_FLAG_GODFATHER);
1349 spa->spa_suspended = B_TRUE;
1352 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1353 ASSERT(zio != spa->spa_suspend_zio_root);
1354 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1355 ASSERT(zio_unique_parent(zio) == NULL);
1356 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1357 zio_add_child(spa->spa_suspend_zio_root, zio);
1360 mutex_exit(&spa->spa_suspend_lock);
1364 zio_resume(spa_t *spa)
1369 * Reexecute all previously suspended i/o.
1371 mutex_enter(&spa->spa_suspend_lock);
1372 spa->spa_suspended = B_FALSE;
1373 cv_broadcast(&spa->spa_suspend_cv);
1374 pio = spa->spa_suspend_zio_root;
1375 spa->spa_suspend_zio_root = NULL;
1376 mutex_exit(&spa->spa_suspend_lock);
1382 return (zio_wait(pio));
1386 zio_resume_wait(spa_t *spa)
1388 mutex_enter(&spa->spa_suspend_lock);
1389 while (spa_suspended(spa))
1390 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1391 mutex_exit(&spa->spa_suspend_lock);
1395 * ==========================================================================
1398 * A gang block is a collection of small blocks that looks to the DMU
1399 * like one large block. When zio_dva_allocate() cannot find a block
1400 * of the requested size, due to either severe fragmentation or the pool
1401 * being nearly full, it calls zio_write_gang_block() to construct the
1402 * block from smaller fragments.
1404 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1405 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1406 * an indirect block: it's an array of block pointers. It consumes
1407 * only one sector and hence is allocatable regardless of fragmentation.
1408 * The gang header's bps point to its gang members, which hold the data.
1410 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1411 * as the verifier to ensure uniqueness of the SHA256 checksum.
1412 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1413 * not the gang header. This ensures that data block signatures (needed for
1414 * deduplication) are independent of how the block is physically stored.
1416 * Gang blocks can be nested: a gang member may itself be a gang block.
1417 * Thus every gang block is a tree in which root and all interior nodes are
1418 * gang headers, and the leaves are normal blocks that contain user data.
1419 * The root of the gang tree is called the gang leader.
1421 * To perform any operation (read, rewrite, free, claim) on a gang block,
1422 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1423 * in the io_gang_tree field of the original logical i/o by recursively
1424 * reading the gang leader and all gang headers below it. This yields
1425 * an in-core tree containing the contents of every gang header and the
1426 * bps for every constituent of the gang block.
1428 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1429 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1430 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1431 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1432 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1433 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1434 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1435 * of the gang header plus zio_checksum_compute() of the data to update the
1436 * gang header's blk_cksum as described above.
1438 * The two-phase assemble/issue model solves the problem of partial failure --
1439 * what if you'd freed part of a gang block but then couldn't read the
1440 * gang header for another part? Assembling the entire gang tree first
1441 * ensures that all the necessary gang header I/O has succeeded before
1442 * starting the actual work of free, claim, or write. Once the gang tree
1443 * is assembled, free and claim are in-memory operations that cannot fail.
1445 * In the event that a gang write fails, zio_dva_unallocate() walks the
1446 * gang tree to immediately free (i.e. insert back into the space map)
1447 * everything we've allocated. This ensures that we don't get ENOSPC
1448 * errors during repeated suspend/resume cycles due to a flaky device.
1450 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1451 * the gang tree, we won't modify the block, so we can safely defer the free
1452 * (knowing that the block is still intact). If we *can* assemble the gang
1453 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1454 * each constituent bp and we can allocate a new block on the next sync pass.
1456 * In all cases, the gang tree allows complete recovery from partial failure.
1457 * ==========================================================================
1461 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1466 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1467 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1468 &pio->io_bookmark));
1472 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1477 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1478 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1479 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1481 * As we rewrite each gang header, the pipeline will compute
1482 * a new gang block header checksum for it; but no one will
1483 * compute a new data checksum, so we do that here. The one
1484 * exception is the gang leader: the pipeline already computed
1485 * its data checksum because that stage precedes gang assembly.
1486 * (Presently, nothing actually uses interior data checksums;
1487 * this is just good hygiene.)
1489 if (gn != pio->io_gang_leader->io_gang_tree) {
1490 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1491 data, BP_GET_PSIZE(bp));
1494 * If we are here to damage data for testing purposes,
1495 * leave the GBH alone so that we can detect the damage.
1497 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1498 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1500 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1501 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1502 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1510 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1512 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1513 ZIO_GANG_CHILD_FLAGS(pio)));
1518 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1520 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1521 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1524 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1533 static void zio_gang_tree_assemble_done(zio_t *zio);
1535 static zio_gang_node_t *
1536 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1538 zio_gang_node_t *gn;
1540 ASSERT(*gnpp == NULL);
1542 gn = kmem_zalloc(sizeof (*gn), KM_PUSHPAGE);
1543 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1550 zio_gang_node_free(zio_gang_node_t **gnpp)
1552 zio_gang_node_t *gn = *gnpp;
1555 for (g = 0; g < SPA_GBH_NBLKPTRS; g++)
1556 ASSERT(gn->gn_child[g] == NULL);
1558 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1559 kmem_free(gn, sizeof (*gn));
1564 zio_gang_tree_free(zio_gang_node_t **gnpp)
1566 zio_gang_node_t *gn = *gnpp;
1572 for (g = 0; g < SPA_GBH_NBLKPTRS; g++)
1573 zio_gang_tree_free(&gn->gn_child[g]);
1575 zio_gang_node_free(gnpp);
1579 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1581 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1583 ASSERT(gio->io_gang_leader == gio);
1584 ASSERT(BP_IS_GANG(bp));
1586 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1587 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1588 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1592 zio_gang_tree_assemble_done(zio_t *zio)
1594 zio_t *gio = zio->io_gang_leader;
1595 zio_gang_node_t *gn = zio->io_private;
1596 blkptr_t *bp = zio->io_bp;
1599 ASSERT(gio == zio_unique_parent(zio));
1600 ASSERT(zio->io_child_count == 0);
1605 if (BP_SHOULD_BYTESWAP(bp))
1606 byteswap_uint64_array(zio->io_data, zio->io_size);
1608 ASSERT(zio->io_data == gn->gn_gbh);
1609 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1610 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1612 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1613 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1614 if (!BP_IS_GANG(gbp))
1616 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1621 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1623 zio_t *gio = pio->io_gang_leader;
1627 ASSERT(BP_IS_GANG(bp) == !!gn);
1628 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1629 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1632 * If you're a gang header, your data is in gn->gn_gbh.
1633 * If you're a gang member, your data is in 'data' and gn == NULL.
1635 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1638 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1640 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1641 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1642 if (BP_IS_HOLE(gbp))
1644 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1645 data = (char *)data + BP_GET_PSIZE(gbp);
1649 if (gn == gio->io_gang_tree)
1650 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1657 zio_gang_assemble(zio_t *zio)
1659 blkptr_t *bp = zio->io_bp;
1661 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1662 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1664 zio->io_gang_leader = zio;
1666 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1668 return (ZIO_PIPELINE_CONTINUE);
1672 zio_gang_issue(zio_t *zio)
1674 blkptr_t *bp = zio->io_bp;
1676 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1677 return (ZIO_PIPELINE_STOP);
1679 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1680 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1682 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1683 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1685 zio_gang_tree_free(&zio->io_gang_tree);
1687 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1689 return (ZIO_PIPELINE_CONTINUE);
1693 zio_write_gang_member_ready(zio_t *zio)
1695 zio_t *pio = zio_unique_parent(zio);
1696 ASSERTV(zio_t *gio = zio->io_gang_leader;)
1697 dva_t *cdva = zio->io_bp->blk_dva;
1698 dva_t *pdva = pio->io_bp->blk_dva;
1702 if (BP_IS_HOLE(zio->io_bp))
1705 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1707 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1708 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1709 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1710 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1711 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1713 mutex_enter(&pio->io_lock);
1714 for (d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1715 ASSERT(DVA_GET_GANG(&pdva[d]));
1716 asize = DVA_GET_ASIZE(&pdva[d]);
1717 asize += DVA_GET_ASIZE(&cdva[d]);
1718 DVA_SET_ASIZE(&pdva[d], asize);
1720 mutex_exit(&pio->io_lock);
1724 zio_write_gang_block(zio_t *pio)
1726 spa_t *spa = pio->io_spa;
1727 blkptr_t *bp = pio->io_bp;
1728 zio_t *gio = pio->io_gang_leader;
1730 zio_gang_node_t *gn, **gnpp;
1731 zio_gbh_phys_t *gbh;
1732 uint64_t txg = pio->io_txg;
1733 uint64_t resid = pio->io_size;
1735 int copies = gio->io_prop.zp_copies;
1736 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1740 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1741 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1742 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1744 pio->io_error = error;
1745 return (ZIO_PIPELINE_CONTINUE);
1749 gnpp = &gio->io_gang_tree;
1751 gnpp = pio->io_private;
1752 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1755 gn = zio_gang_node_alloc(gnpp);
1757 bzero(gbh, SPA_GANGBLOCKSIZE);
1760 * Create the gang header.
1762 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1763 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1766 * Create and nowait the gang children.
1768 for (g = 0; resid != 0; resid -= lsize, g++) {
1769 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1771 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1773 zp.zp_checksum = gio->io_prop.zp_checksum;
1774 zp.zp_compress = ZIO_COMPRESS_OFF;
1775 zp.zp_type = DMU_OT_NONE;
1777 zp.zp_copies = gio->io_prop.zp_copies;
1779 zp.zp_dedup_verify = 0;
1781 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1782 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1783 zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1784 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1785 &pio->io_bookmark));
1789 * Set pio's pipeline to just wait for zio to finish.
1791 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1795 return (ZIO_PIPELINE_CONTINUE);
1799 * ==========================================================================
1801 * ==========================================================================
1804 zio_ddt_child_read_done(zio_t *zio)
1806 blkptr_t *bp = zio->io_bp;
1807 ddt_entry_t *dde = zio->io_private;
1809 zio_t *pio = zio_unique_parent(zio);
1811 mutex_enter(&pio->io_lock);
1812 ddp = ddt_phys_select(dde, bp);
1813 if (zio->io_error == 0)
1814 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1815 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1816 dde->dde_repair_data = zio->io_data;
1818 zio_buf_free(zio->io_data, zio->io_size);
1819 mutex_exit(&pio->io_lock);
1823 zio_ddt_read_start(zio_t *zio)
1825 blkptr_t *bp = zio->io_bp;
1828 ASSERT(BP_GET_DEDUP(bp));
1829 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1830 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1832 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1833 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1834 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1835 ddt_phys_t *ddp = dde->dde_phys;
1836 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1839 ASSERT(zio->io_vsd == NULL);
1842 if (ddp_self == NULL)
1843 return (ZIO_PIPELINE_CONTINUE);
1845 for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1846 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1848 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1850 zio_nowait(zio_read(zio, zio->io_spa, &blk,
1851 zio_buf_alloc(zio->io_size), zio->io_size,
1852 zio_ddt_child_read_done, dde, zio->io_priority,
1853 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1854 &zio->io_bookmark));
1856 return (ZIO_PIPELINE_CONTINUE);
1859 zio_nowait(zio_read(zio, zio->io_spa, bp,
1860 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1861 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1863 return (ZIO_PIPELINE_CONTINUE);
1867 zio_ddt_read_done(zio_t *zio)
1869 blkptr_t *bp = zio->io_bp;
1871 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1872 return (ZIO_PIPELINE_STOP);
1874 ASSERT(BP_GET_DEDUP(bp));
1875 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1876 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1878 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1879 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1880 ddt_entry_t *dde = zio->io_vsd;
1882 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
1883 return (ZIO_PIPELINE_CONTINUE);
1886 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
1887 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1888 return (ZIO_PIPELINE_STOP);
1890 if (dde->dde_repair_data != NULL) {
1891 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
1892 zio->io_child_error[ZIO_CHILD_DDT] = 0;
1894 ddt_repair_done(ddt, dde);
1898 ASSERT(zio->io_vsd == NULL);
1900 return (ZIO_PIPELINE_CONTINUE);
1904 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
1906 spa_t *spa = zio->io_spa;
1910 * Note: we compare the original data, not the transformed data,
1911 * because when zio->io_bp is an override bp, we will not have
1912 * pushed the I/O transforms. That's an important optimization
1913 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
1915 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1916 zio_t *lio = dde->dde_lead_zio[p];
1919 return (lio->io_orig_size != zio->io_orig_size ||
1920 bcmp(zio->io_orig_data, lio->io_orig_data,
1921 zio->io_orig_size) != 0);
1925 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1926 ddt_phys_t *ddp = &dde->dde_phys[p];
1928 if (ddp->ddp_phys_birth != 0) {
1929 arc_buf_t *abuf = NULL;
1930 uint32_t aflags = ARC_WAIT;
1931 blkptr_t blk = *zio->io_bp;
1934 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
1938 error = arc_read_nolock(NULL, spa, &blk,
1939 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
1940 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1941 &aflags, &zio->io_bookmark);
1944 if (arc_buf_size(abuf) != zio->io_orig_size ||
1945 bcmp(abuf->b_data, zio->io_orig_data,
1946 zio->io_orig_size) != 0)
1948 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
1952 return (error != 0);
1960 zio_ddt_child_write_ready(zio_t *zio)
1962 int p = zio->io_prop.zp_copies;
1963 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1964 ddt_entry_t *dde = zio->io_private;
1965 ddt_phys_t *ddp = &dde->dde_phys[p];
1973 ASSERT(dde->dde_lead_zio[p] == zio);
1975 ddt_phys_fill(ddp, zio->io_bp);
1977 while ((pio = zio_walk_parents(zio)) != NULL)
1978 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
1984 zio_ddt_child_write_done(zio_t *zio)
1986 int p = zio->io_prop.zp_copies;
1987 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1988 ddt_entry_t *dde = zio->io_private;
1989 ddt_phys_t *ddp = &dde->dde_phys[p];
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 while (zio_walk_parents(zio) != NULL)
1999 ddt_phys_addref(ddp);
2001 ddt_phys_clear(ddp);
2008 zio_ddt_ditto_write_done(zio_t *zio)
2010 int p = DDT_PHYS_DITTO;
2011 blkptr_t *bp = zio->io_bp;
2012 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2013 ddt_entry_t *dde = zio->io_private;
2014 ddt_phys_t *ddp = &dde->dde_phys[p];
2015 ddt_key_t *ddk = &dde->dde_key;
2016 ASSERTV(zio_prop_t *zp = &zio->io_prop);
2020 ASSERT(ddp->ddp_refcnt == 0);
2021 ASSERT(dde->dde_lead_zio[p] == zio);
2022 dde->dde_lead_zio[p] = NULL;
2024 if (zio->io_error == 0) {
2025 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2026 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2027 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2028 if (ddp->ddp_phys_birth != 0)
2029 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2030 ddt_phys_fill(ddp, bp);
2037 zio_ddt_write(zio_t *zio)
2039 spa_t *spa = zio->io_spa;
2040 blkptr_t *bp = zio->io_bp;
2041 uint64_t txg = zio->io_txg;
2042 zio_prop_t *zp = &zio->io_prop;
2043 int p = zp->zp_copies;
2047 ddt_t *ddt = ddt_select(spa, bp);
2051 ASSERT(BP_GET_DEDUP(bp));
2052 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2053 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2056 dde = ddt_lookup(ddt, bp, B_TRUE);
2057 ddp = &dde->dde_phys[p];
2059 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2061 * If we're using a weak checksum, upgrade to a strong checksum
2062 * and try again. If we're already using a strong checksum,
2063 * we can't resolve it, so just convert to an ordinary write.
2064 * (And automatically e-mail a paper to Nature?)
2066 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2067 zp->zp_checksum = spa_dedup_checksum(spa);
2068 zio_pop_transforms(zio);
2069 zio->io_stage = ZIO_STAGE_OPEN;
2074 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2076 return (ZIO_PIPELINE_CONTINUE);
2079 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2080 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2082 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2083 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2084 zio_prop_t czp = *zp;
2086 czp.zp_copies = ditto_copies;
2089 * If we arrived here with an override bp, we won't have run
2090 * the transform stack, so we won't have the data we need to
2091 * generate a child i/o. So, toss the override bp and restart.
2092 * This is safe, because using the override bp is just an
2093 * optimization; and it's rare, so the cost doesn't matter.
2095 if (zio->io_bp_override) {
2096 zio_pop_transforms(zio);
2097 zio->io_stage = ZIO_STAGE_OPEN;
2098 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2099 zio->io_bp_override = NULL;
2102 return (ZIO_PIPELINE_CONTINUE);
2105 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2106 zio->io_orig_size, &czp, NULL,
2107 zio_ddt_ditto_write_done, dde, zio->io_priority,
2108 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2110 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2111 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2114 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2115 if (ddp->ddp_phys_birth != 0)
2116 ddt_bp_fill(ddp, bp, txg);
2117 if (dde->dde_lead_zio[p] != NULL)
2118 zio_add_child(zio, dde->dde_lead_zio[p]);
2120 ddt_phys_addref(ddp);
2121 } else if (zio->io_bp_override) {
2122 ASSERT(bp->blk_birth == txg);
2123 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2124 ddt_phys_fill(ddp, bp);
2125 ddt_phys_addref(ddp);
2127 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2128 zio->io_orig_size, zp, zio_ddt_child_write_ready,
2129 zio_ddt_child_write_done, dde, zio->io_priority,
2130 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2132 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2133 dde->dde_lead_zio[p] = cio;
2143 return (ZIO_PIPELINE_CONTINUE);
2146 ddt_entry_t *freedde; /* for debugging */
2149 zio_ddt_free(zio_t *zio)
2151 spa_t *spa = zio->io_spa;
2152 blkptr_t *bp = zio->io_bp;
2153 ddt_t *ddt = ddt_select(spa, bp);
2157 ASSERT(BP_GET_DEDUP(bp));
2158 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2161 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2162 ddp = ddt_phys_select(dde, bp);
2163 ddt_phys_decref(ddp);
2166 return (ZIO_PIPELINE_CONTINUE);
2170 * ==========================================================================
2171 * Allocate and free blocks
2172 * ==========================================================================
2175 zio_dva_allocate(zio_t *zio)
2177 spa_t *spa = zio->io_spa;
2178 metaslab_class_t *mc = spa_normal_class(spa);
2179 blkptr_t *bp = zio->io_bp;
2183 if (zio->io_gang_leader == NULL) {
2184 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2185 zio->io_gang_leader = zio;
2188 ASSERT(BP_IS_HOLE(bp));
2189 ASSERT3U(BP_GET_NDVAS(bp), ==, 0);
2190 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2191 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2192 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2195 * The dump device does not support gang blocks so allocation on
2196 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2197 * the "fast" gang feature.
2199 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2200 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2201 METASLAB_GANG_CHILD : 0;
2202 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2203 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2206 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2207 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2209 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2210 return (zio_write_gang_block(zio));
2211 zio->io_error = error;
2214 return (ZIO_PIPELINE_CONTINUE);
2218 zio_dva_free(zio_t *zio)
2220 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2222 return (ZIO_PIPELINE_CONTINUE);
2226 zio_dva_claim(zio_t *zio)
2230 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2232 zio->io_error = error;
2234 return (ZIO_PIPELINE_CONTINUE);
2238 * Undo an allocation. This is used by zio_done() when an I/O fails
2239 * and we want to give back the block we just allocated.
2240 * This handles both normal blocks and gang blocks.
2243 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2247 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2248 ASSERT(zio->io_bp_override == NULL);
2250 if (!BP_IS_HOLE(bp))
2251 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2254 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2255 zio_dva_unallocate(zio, gn->gn_child[g],
2256 &gn->gn_gbh->zg_blkptr[g]);
2262 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2265 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2266 uint64_t size, boolean_t use_slog)
2270 ASSERT(txg > spa_syncing_txg(spa));
2273 error = metaslab_alloc(spa, spa_log_class(spa), size,
2274 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2277 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2278 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2281 BP_SET_LSIZE(new_bp, size);
2282 BP_SET_PSIZE(new_bp, size);
2283 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2284 BP_SET_CHECKSUM(new_bp,
2285 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2286 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2287 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2288 BP_SET_LEVEL(new_bp, 0);
2289 BP_SET_DEDUP(new_bp, 0);
2290 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2297 * Free an intent log block.
2300 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2302 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2303 ASSERT(!BP_IS_GANG(bp));
2305 zio_free(spa, txg, bp);
2309 * ==========================================================================
2310 * Read and write to physical devices
2311 * ==========================================================================
2314 zio_vdev_io_start(zio_t *zio)
2316 vdev_t *vd = zio->io_vd;
2318 spa_t *spa = zio->io_spa;
2320 ASSERT(zio->io_error == 0);
2321 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2324 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2325 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2328 * The mirror_ops handle multiple DVAs in a single BP.
2330 return (vdev_mirror_ops.vdev_op_io_start(zio));
2334 * We keep track of time-sensitive I/Os so that the scan thread
2335 * can quickly react to certain workloads. In particular, we care
2336 * about non-scrubbing, top-level reads and writes with the following
2338 * - synchronous writes of user data to non-slog devices
2339 * - any reads of user data
2340 * When these conditions are met, adjust the timestamp of spa_last_io
2341 * which allows the scan thread to adjust its workload accordingly.
2343 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2344 vd == vd->vdev_top && !vd->vdev_islog &&
2345 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2346 zio->io_txg != spa_syncing_txg(spa)) {
2347 uint64_t old = spa->spa_last_io;
2348 uint64_t new = ddi_get_lbolt64();
2350 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2353 align = 1ULL << vd->vdev_top->vdev_ashift;
2355 if (P2PHASE(zio->io_size, align) != 0) {
2356 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2357 char *abuf = zio_buf_alloc(asize);
2358 ASSERT(vd == vd->vdev_top);
2359 if (zio->io_type == ZIO_TYPE_WRITE) {
2360 bcopy(zio->io_data, abuf, zio->io_size);
2361 bzero(abuf + zio->io_size, asize - zio->io_size);
2363 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2366 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2367 ASSERT(P2PHASE(zio->io_size, align) == 0);
2368 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2371 * If this is a repair I/O, and there's no self-healing involved --
2372 * that is, we're just resilvering what we expect to resilver --
2373 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2374 * This prevents spurious resilvering with nested replication.
2375 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2376 * A is out of date, we'll read from C+D, then use the data to
2377 * resilver A+B -- but we don't actually want to resilver B, just A.
2378 * The top-level mirror has no way to know this, so instead we just
2379 * discard unnecessary repairs as we work our way down the vdev tree.
2380 * The same logic applies to any form of nested replication:
2381 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2383 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2384 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2385 zio->io_txg != 0 && /* not a delegated i/o */
2386 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2387 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2388 zio_vdev_io_bypass(zio);
2389 return (ZIO_PIPELINE_CONTINUE);
2392 if (vd->vdev_ops->vdev_op_leaf &&
2393 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2395 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2396 return (ZIO_PIPELINE_CONTINUE);
2398 if ((zio = vdev_queue_io(zio)) == NULL)
2399 return (ZIO_PIPELINE_STOP);
2401 if (!vdev_accessible(vd, zio)) {
2402 zio->io_error = ENXIO;
2404 return (ZIO_PIPELINE_STOP);
2408 return (vd->vdev_ops->vdev_op_io_start(zio));
2412 zio_vdev_io_done(zio_t *zio)
2414 vdev_t *vd = zio->io_vd;
2415 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2416 boolean_t unexpected_error = B_FALSE;
2418 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2419 return (ZIO_PIPELINE_STOP);
2421 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2423 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2425 vdev_queue_io_done(zio);
2427 if (zio->io_type == ZIO_TYPE_WRITE)
2428 vdev_cache_write(zio);
2430 if (zio_injection_enabled && zio->io_error == 0)
2431 zio->io_error = zio_handle_device_injection(vd,
2434 if (zio_injection_enabled && zio->io_error == 0)
2435 zio->io_error = zio_handle_label_injection(zio, EIO);
2437 if (zio->io_error) {
2438 if (!vdev_accessible(vd, zio)) {
2439 zio->io_error = ENXIO;
2441 unexpected_error = B_TRUE;
2446 ops->vdev_op_io_done(zio);
2448 if (unexpected_error)
2449 VERIFY(vdev_probe(vd, zio) == NULL);
2451 return (ZIO_PIPELINE_CONTINUE);
2455 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2456 * disk, and use that to finish the checksum ereport later.
2459 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2460 const void *good_buf)
2462 /* no processing needed */
2463 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2468 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2470 void *buf = zio_buf_alloc(zio->io_size);
2472 bcopy(zio->io_data, buf, zio->io_size);
2474 zcr->zcr_cbinfo = zio->io_size;
2475 zcr->zcr_cbdata = buf;
2476 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2477 zcr->zcr_free = zio_buf_free;
2481 zio_vdev_io_assess(zio_t *zio)
2483 vdev_t *vd = zio->io_vd;
2485 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2486 return (ZIO_PIPELINE_STOP);
2488 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2489 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2491 if (zio->io_vsd != NULL) {
2492 zio->io_vsd_ops->vsd_free(zio);
2496 if (zio_injection_enabled && zio->io_error == 0)
2497 zio->io_error = zio_handle_fault_injection(zio, EIO);
2500 * If the I/O failed, determine whether we should attempt to retry it.
2502 * On retry, we cut in line in the issue queue, since we don't want
2503 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2505 if (zio->io_error && vd == NULL &&
2506 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2507 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2508 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2510 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2511 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2512 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2513 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2514 zio_requeue_io_start_cut_in_line);
2515 return (ZIO_PIPELINE_STOP);
2519 * If we got an error on a leaf device, convert it to ENXIO
2520 * if the device is not accessible at all.
2522 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2523 !vdev_accessible(vd, zio))
2524 zio->io_error = ENXIO;
2527 * If we can't write to an interior vdev (mirror or RAID-Z),
2528 * set vdev_cant_write so that we stop trying to allocate from it.
2530 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2531 vd != NULL && !vd->vdev_ops->vdev_op_leaf)
2532 vd->vdev_cant_write = B_TRUE;
2535 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2537 return (ZIO_PIPELINE_CONTINUE);
2541 zio_vdev_io_reissue(zio_t *zio)
2543 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2544 ASSERT(zio->io_error == 0);
2546 zio->io_stage >>= 1;
2550 zio_vdev_io_redone(zio_t *zio)
2552 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2554 zio->io_stage >>= 1;
2558 zio_vdev_io_bypass(zio_t *zio)
2560 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2561 ASSERT(zio->io_error == 0);
2563 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2564 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2568 * ==========================================================================
2569 * Generate and verify checksums
2570 * ==========================================================================
2573 zio_checksum_generate(zio_t *zio)
2575 blkptr_t *bp = zio->io_bp;
2576 enum zio_checksum checksum;
2580 * This is zio_write_phys().
2581 * We're either generating a label checksum, or none at all.
2583 checksum = zio->io_prop.zp_checksum;
2585 if (checksum == ZIO_CHECKSUM_OFF)
2586 return (ZIO_PIPELINE_CONTINUE);
2588 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2590 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2591 ASSERT(!IO_IS_ALLOCATING(zio));
2592 checksum = ZIO_CHECKSUM_GANG_HEADER;
2594 checksum = BP_GET_CHECKSUM(bp);
2598 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2600 return (ZIO_PIPELINE_CONTINUE);
2604 zio_checksum_verify(zio_t *zio)
2606 zio_bad_cksum_t info;
2607 blkptr_t *bp = zio->io_bp;
2610 ASSERT(zio->io_vd != NULL);
2614 * This is zio_read_phys().
2615 * We're either verifying a label checksum, or nothing at all.
2617 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2618 return (ZIO_PIPELINE_CONTINUE);
2620 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2623 if ((error = zio_checksum_error(zio, &info)) != 0) {
2624 zio->io_error = error;
2625 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2626 zfs_ereport_start_checksum(zio->io_spa,
2627 zio->io_vd, zio, zio->io_offset,
2628 zio->io_size, NULL, &info);
2632 return (ZIO_PIPELINE_CONTINUE);
2636 * Called by RAID-Z to ensure we don't compute the checksum twice.
2639 zio_checksum_verified(zio_t *zio)
2641 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2645 * ==========================================================================
2646 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2647 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2648 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2649 * indicate errors that are specific to one I/O, and most likely permanent.
2650 * Any other error is presumed to be worse because we weren't expecting it.
2651 * ==========================================================================
2654 zio_worst_error(int e1, int e2)
2656 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2659 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2660 if (e1 == zio_error_rank[r1])
2663 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2664 if (e2 == zio_error_rank[r2])
2667 return (r1 > r2 ? e1 : e2);
2671 * ==========================================================================
2673 * ==========================================================================
2676 zio_ready(zio_t *zio)
2678 blkptr_t *bp = zio->io_bp;
2679 zio_t *pio, *pio_next;
2681 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2682 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2683 return (ZIO_PIPELINE_STOP);
2685 if (zio->io_ready) {
2686 ASSERT(IO_IS_ALLOCATING(zio));
2687 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2688 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2693 if (bp != NULL && bp != &zio->io_bp_copy)
2694 zio->io_bp_copy = *bp;
2697 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2699 mutex_enter(&zio->io_lock);
2700 zio->io_state[ZIO_WAIT_READY] = 1;
2701 pio = zio_walk_parents(zio);
2702 mutex_exit(&zio->io_lock);
2705 * As we notify zio's parents, new parents could be added.
2706 * New parents go to the head of zio's io_parent_list, however,
2707 * so we will (correctly) not notify them. The remainder of zio's
2708 * io_parent_list, from 'pio_next' onward, cannot change because
2709 * all parents must wait for us to be done before they can be done.
2711 for (; pio != NULL; pio = pio_next) {
2712 pio_next = zio_walk_parents(zio);
2713 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2716 if (zio->io_flags & ZIO_FLAG_NODATA) {
2717 if (BP_IS_GANG(bp)) {
2718 zio->io_flags &= ~ZIO_FLAG_NODATA;
2720 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2721 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2725 if (zio_injection_enabled &&
2726 zio->io_spa->spa_syncing_txg == zio->io_txg)
2727 zio_handle_ignored_writes(zio);
2729 return (ZIO_PIPELINE_CONTINUE);
2733 zio_done(zio_t *zio)
2735 zio_t *pio, *pio_next;
2739 * If our children haven't all completed,
2740 * wait for them and then repeat this pipeline stage.
2742 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2743 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2744 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2745 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2746 return (ZIO_PIPELINE_STOP);
2748 for (c = 0; c < ZIO_CHILD_TYPES; c++)
2749 for (w = 0; w < ZIO_WAIT_TYPES; w++)
2750 ASSERT(zio->io_children[c][w] == 0);
2752 if (zio->io_bp != NULL) {
2753 ASSERT(zio->io_bp->blk_pad[0] == 0);
2754 ASSERT(zio->io_bp->blk_pad[1] == 0);
2755 ASSERT(bcmp(zio->io_bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2756 (zio->io_bp == zio_unique_parent(zio)->io_bp));
2757 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) &&
2758 zio->io_bp_override == NULL &&
2759 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2760 ASSERT(!BP_SHOULD_BYTESWAP(zio->io_bp));
2761 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2762 ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 ||
2763 (BP_COUNT_GANG(zio->io_bp) == BP_GET_NDVAS(zio->io_bp)));
2768 * If there were child vdev/gang/ddt errors, they apply to us now.
2770 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2771 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2772 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2775 * If the I/O on the transformed data was successful, generate any
2776 * checksum reports now while we still have the transformed data.
2778 if (zio->io_error == 0) {
2779 while (zio->io_cksum_report != NULL) {
2780 zio_cksum_report_t *zcr = zio->io_cksum_report;
2781 uint64_t align = zcr->zcr_align;
2782 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2783 char *abuf = zio->io_data;
2785 if (asize != zio->io_size) {
2786 abuf = zio_buf_alloc(asize);
2787 bcopy(zio->io_data, abuf, zio->io_size);
2788 bzero(abuf + zio->io_size, asize - zio->io_size);
2791 zio->io_cksum_report = zcr->zcr_next;
2792 zcr->zcr_next = NULL;
2793 zcr->zcr_finish(zcr, abuf);
2794 zfs_ereport_free_checksum(zcr);
2796 if (asize != zio->io_size)
2797 zio_buf_free(abuf, asize);
2801 zio_pop_transforms(zio); /* note: may set zio->io_error */
2803 vdev_stat_update(zio, zio->io_size);
2806 * If this I/O is attached to a particular vdev is slow, exeeding
2807 * 30 seconds to complete, post an error described the I/O delay.
2808 * We ignore these errors if the device is currently unavailable.
2810 if (zio->io_delay >= zio_delay_max) {
2811 if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd))
2812 zfs_ereport_post(FM_EREPORT_ZFS_DELAY, zio->io_spa,
2813 zio->io_vd, zio, 0, 0);
2816 if (zio->io_error) {
2818 * If this I/O is attached to a particular vdev,
2819 * generate an error message describing the I/O failure
2820 * at the block level. We ignore these errors if the
2821 * device is currently unavailable.
2823 if (zio->io_error != ECKSUM && zio->io_vd != NULL &&
2824 !vdev_is_dead(zio->io_vd))
2825 zfs_ereport_post(FM_EREPORT_ZFS_IO, zio->io_spa,
2826 zio->io_vd, zio, 0, 0);
2828 if ((zio->io_error == EIO || !(zio->io_flags &
2829 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2830 zio == zio->io_logical) {
2832 * For logical I/O requests, tell the SPA to log the
2833 * error and generate a logical data ereport.
2835 spa_log_error(zio->io_spa, zio);
2836 zfs_ereport_post(FM_EREPORT_ZFS_DATA, zio->io_spa, NULL, zio,
2841 if (zio->io_error && zio == zio->io_logical) {
2843 * Determine whether zio should be reexecuted. This will
2844 * propagate all the way to the root via zio_notify_parent().
2846 ASSERT(zio->io_vd == NULL && zio->io_bp != NULL);
2847 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2849 if (IO_IS_ALLOCATING(zio) &&
2850 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2851 if (zio->io_error != ENOSPC)
2852 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2854 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2857 if ((zio->io_type == ZIO_TYPE_READ ||
2858 zio->io_type == ZIO_TYPE_FREE) &&
2859 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
2860 zio->io_error == ENXIO &&
2861 spa_load_state(zio->io_spa) == SPA_LOAD_NONE &&
2862 spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE)
2863 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2865 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2866 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2869 * Here is a possibly good place to attempt to do
2870 * either combinatorial reconstruction or error correction
2871 * based on checksums. It also might be a good place
2872 * to send out preliminary ereports before we suspend
2878 * If there were logical child errors, they apply to us now.
2879 * We defer this until now to avoid conflating logical child
2880 * errors with errors that happened to the zio itself when
2881 * updating vdev stats and reporting FMA events above.
2883 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2885 if ((zio->io_error || zio->io_reexecute) &&
2886 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
2887 !(zio->io_flags & ZIO_FLAG_IO_REWRITE))
2888 zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp);
2890 zio_gang_tree_free(&zio->io_gang_tree);
2893 * Godfather I/Os should never suspend.
2895 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
2896 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
2897 zio->io_reexecute = 0;
2899 if (zio->io_reexecute) {
2901 * This is a logical I/O that wants to reexecute.
2903 * Reexecute is top-down. When an i/o fails, if it's not
2904 * the root, it simply notifies its parent and sticks around.
2905 * The parent, seeing that it still has children in zio_done(),
2906 * does the same. This percolates all the way up to the root.
2907 * The root i/o will reexecute or suspend the entire tree.
2909 * This approach ensures that zio_reexecute() honors
2910 * all the original i/o dependency relationships, e.g.
2911 * parents not executing until children are ready.
2913 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2915 zio->io_gang_leader = NULL;
2917 mutex_enter(&zio->io_lock);
2918 zio->io_state[ZIO_WAIT_DONE] = 1;
2919 mutex_exit(&zio->io_lock);
2922 * "The Godfather" I/O monitors its children but is
2923 * not a true parent to them. It will track them through
2924 * the pipeline but severs its ties whenever they get into
2925 * trouble (e.g. suspended). This allows "The Godfather"
2926 * I/O to return status without blocking.
2928 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2929 zio_link_t *zl = zio->io_walk_link;
2930 pio_next = zio_walk_parents(zio);
2932 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
2933 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
2934 zio_remove_child(pio, zio, zl);
2935 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2939 if ((pio = zio_unique_parent(zio)) != NULL) {
2941 * We're not a root i/o, so there's nothing to do
2942 * but notify our parent. Don't propagate errors
2943 * upward since we haven't permanently failed yet.
2945 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2946 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
2947 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2948 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
2950 * We'd fail again if we reexecuted now, so suspend
2951 * until conditions improve (e.g. device comes online).
2953 zio_suspend(zio->io_spa, zio);
2956 * Reexecution is potentially a huge amount of work.
2957 * Hand it off to the otherwise-unused claim taskq.
2959 ASSERT(taskq_empty_ent(&zio->io_tqent));
2960 (void) taskq_dispatch_ent(
2961 zio->io_spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
2962 (task_func_t *)zio_reexecute, zio, 0,
2965 return (ZIO_PIPELINE_STOP);
2968 ASSERT(zio->io_child_count == 0);
2969 ASSERT(zio->io_reexecute == 0);
2970 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
2973 * Report any checksum errors, since the I/O is complete.
2975 while (zio->io_cksum_report != NULL) {
2976 zio_cksum_report_t *zcr = zio->io_cksum_report;
2977 zio->io_cksum_report = zcr->zcr_next;
2978 zcr->zcr_next = NULL;
2979 zcr->zcr_finish(zcr, NULL);
2980 zfs_ereport_free_checksum(zcr);
2984 * It is the responsibility of the done callback to ensure that this
2985 * particular zio is no longer discoverable for adoption, and as
2986 * such, cannot acquire any new parents.
2991 mutex_enter(&zio->io_lock);
2992 zio->io_state[ZIO_WAIT_DONE] = 1;
2993 mutex_exit(&zio->io_lock);
2995 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2996 zio_link_t *zl = zio->io_walk_link;
2997 pio_next = zio_walk_parents(zio);
2998 zio_remove_child(pio, zio, zl);
2999 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3002 if (zio->io_waiter != NULL) {
3003 mutex_enter(&zio->io_lock);
3004 zio->io_executor = NULL;
3005 cv_broadcast(&zio->io_cv);
3006 mutex_exit(&zio->io_lock);
3011 return (ZIO_PIPELINE_STOP);
3015 * ==========================================================================
3016 * I/O pipeline definition
3017 * ==========================================================================
3019 static zio_pipe_stage_t *zio_pipeline[] = {
3025 zio_checksum_generate,
3039 zio_checksum_verify,
3043 #if defined(_KERNEL) && defined(HAVE_SPL)
3044 /* Fault injection */
3045 EXPORT_SYMBOL(zio_injection_enabled);
3046 EXPORT_SYMBOL(zio_inject_fault);
3047 EXPORT_SYMBOL(zio_inject_list_next);
3048 EXPORT_SYMBOL(zio_clear_fault);
3049 EXPORT_SYMBOL(zio_handle_fault_injection);
3050 EXPORT_SYMBOL(zio_handle_device_injection);
3051 EXPORT_SYMBOL(zio_handle_label_injection);
3052 EXPORT_SYMBOL(zio_priority_table);
3053 EXPORT_SYMBOL(zio_type_name);
3055 module_param(zio_bulk_flags, int, 0644);
3056 MODULE_PARM_DESC(zio_bulk_flags, "Additional flags to pass to bulk buffers");
3058 module_param(zio_delay_max, int, 0644);
3059 MODULE_PARM_DESC(zio_delay_max, "Max zio millisec delay before posting event");
3061 module_param(zio_requeue_io_start_cut_in_line, int, 0644);
3062 MODULE_PARM_DESC(zio_requeue_io_start_cut_in_line, "Prioritize requeued I/O");