4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2012 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
27 #include <sys/zfs_context.h>
28 #include <sys/fm/fs/zfs.h>
31 #include <sys/spa_impl.h>
32 #include <sys/vdev_impl.h>
33 #include <sys/zio_impl.h>
34 #include <sys/zio_compress.h>
35 #include <sys/zio_checksum.h>
36 #include <sys/dmu_objset.h>
41 * ==========================================================================
43 * ==========================================================================
45 uint8_t zio_priority_table[ZIO_PRIORITY_TABLE_SIZE] = {
46 0, /* ZIO_PRIORITY_NOW */
47 0, /* ZIO_PRIORITY_SYNC_READ */
48 0, /* ZIO_PRIORITY_SYNC_WRITE */
49 0, /* ZIO_PRIORITY_LOG_WRITE */
50 1, /* ZIO_PRIORITY_CACHE_FILL */
51 1, /* ZIO_PRIORITY_AGG */
52 4, /* ZIO_PRIORITY_FREE */
53 4, /* ZIO_PRIORITY_ASYNC_WRITE */
54 6, /* ZIO_PRIORITY_ASYNC_READ */
55 10, /* ZIO_PRIORITY_RESILVER */
56 20, /* ZIO_PRIORITY_SCRUB */
57 2, /* ZIO_PRIORITY_DDT_PREFETCH */
61 * ==========================================================================
62 * I/O type descriptions
63 * ==========================================================================
65 char *zio_type_name[ZIO_TYPES] = {
66 "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl"
70 * ==========================================================================
72 * ==========================================================================
74 kmem_cache_t *zio_cache;
75 kmem_cache_t *zio_link_cache;
76 kmem_cache_t *zio_vdev_cache;
77 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
78 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
79 int zio_bulk_flags = 0;
80 int zio_delay_max = ZIO_DELAY_MAX;
83 extern vmem_t *zio_alloc_arena;
85 extern int zfs_mg_alloc_failures;
88 * The following actions directly effect the spa's sync-to-convergence logic.
89 * The values below define the sync pass when we start performing the action.
90 * Care should be taken when changing these values as they directly impact
91 * spa_sync() performance. Tuning these values may introduce subtle performance
92 * pathologies and should only be done in the context of performance analysis.
93 * These tunables will eventually be removed and replaced with #defines once
94 * enough analysis has been done to determine optimal values.
96 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
97 * regular blocks are not deferred.
99 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
100 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
101 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
104 * An allocating zio is one that either currently has the DVA allocate
105 * stage set or will have it later in its lifetime.
107 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
109 int zio_requeue_io_start_cut_in_line = 1;
112 int zio_buf_debug_limit = 16384;
114 int zio_buf_debug_limit = 0;
117 static inline void __zio_execute(zio_t *zio);
120 zio_cons(void *arg, void *unused, int kmflag)
124 bzero(zio, sizeof (zio_t));
126 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
127 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
129 list_create(&zio->io_parent_list, sizeof (zio_link_t),
130 offsetof(zio_link_t, zl_parent_node));
131 list_create(&zio->io_child_list, sizeof (zio_link_t),
132 offsetof(zio_link_t, zl_child_node));
138 zio_dest(void *arg, void *unused)
142 mutex_destroy(&zio->io_lock);
143 cv_destroy(&zio->io_cv);
144 list_destroy(&zio->io_parent_list);
145 list_destroy(&zio->io_child_list);
152 vmem_t *data_alloc_arena = NULL;
155 data_alloc_arena = zio_alloc_arena;
157 zio_cache = kmem_cache_create("zio_cache", sizeof (zio_t), 0,
158 zio_cons, zio_dest, NULL, NULL, NULL, KMC_KMEM);
159 zio_link_cache = kmem_cache_create("zio_link_cache",
160 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, KMC_KMEM);
161 zio_vdev_cache = kmem_cache_create("zio_vdev_cache", sizeof(vdev_io_t),
162 PAGESIZE, NULL, NULL, NULL, NULL, NULL, KMC_VMEM);
165 * For small buffers, we want a cache for each multiple of
166 * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache
167 * for each quarter-power of 2. For large buffers, we want
168 * a cache for each multiple of PAGESIZE.
170 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
171 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
175 while (p2 & (p2 - 1))
178 if (size <= 4 * SPA_MINBLOCKSIZE) {
179 align = SPA_MINBLOCKSIZE;
180 } else if (P2PHASE(size, PAGESIZE) == 0) {
182 } else if (P2PHASE(size, p2 >> 2) == 0) {
188 int flags = zio_bulk_flags;
191 * The smallest buffers (512b) are heavily used and
192 * experience a lot of churn. The slabs allocated
193 * for them are also relatively small (32K). Thus
194 * in over to avoid expensive calls to vmalloc() we
195 * make an exception to the usual slab allocation
196 * policy and force these buffers to be kmem backed.
198 if (size == (1 << SPA_MINBLOCKSHIFT))
201 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
202 zio_buf_cache[c] = kmem_cache_create(name, size,
203 align, NULL, NULL, NULL, NULL, NULL, flags);
205 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
206 zio_data_buf_cache[c] = kmem_cache_create(name, size,
207 align, NULL, NULL, NULL, NULL,
208 data_alloc_arena, flags);
213 ASSERT(zio_buf_cache[c] != NULL);
214 if (zio_buf_cache[c - 1] == NULL)
215 zio_buf_cache[c - 1] = zio_buf_cache[c];
217 ASSERT(zio_data_buf_cache[c] != NULL);
218 if (zio_data_buf_cache[c - 1] == NULL)
219 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
223 * The zio write taskqs have 1 thread per cpu, allow 1/2 of the taskqs
224 * to fail 3 times per txg or 8 failures, whichever is greater.
226 zfs_mg_alloc_failures = MAX((3 * max_ncpus / 2), 8);
237 kmem_cache_t *last_cache = NULL;
238 kmem_cache_t *last_data_cache = NULL;
240 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
241 if (zio_buf_cache[c] != last_cache) {
242 last_cache = zio_buf_cache[c];
243 kmem_cache_destroy(zio_buf_cache[c]);
245 zio_buf_cache[c] = NULL;
247 if (zio_data_buf_cache[c] != last_data_cache) {
248 last_data_cache = zio_data_buf_cache[c];
249 kmem_cache_destroy(zio_data_buf_cache[c]);
251 zio_data_buf_cache[c] = NULL;
254 kmem_cache_destroy(zio_vdev_cache);
255 kmem_cache_destroy(zio_link_cache);
256 kmem_cache_destroy(zio_cache);
264 * ==========================================================================
265 * Allocate and free I/O buffers
266 * ==========================================================================
270 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
271 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
272 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
273 * excess / transient data in-core during a crashdump.
276 zio_buf_alloc(size_t size)
278 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
280 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
282 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE | KM_NODEBUG));
286 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
287 * crashdump if the kernel panics. This exists so that we will limit the amount
288 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
289 * of kernel heap dumped to disk when the kernel panics)
292 zio_data_buf_alloc(size_t size)
294 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
296 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
298 return (kmem_cache_alloc(zio_data_buf_cache[c],
299 KM_PUSHPAGE | KM_NODEBUG));
303 zio_buf_free(void *buf, size_t size)
305 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
307 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
309 kmem_cache_free(zio_buf_cache[c], buf);
313 zio_data_buf_free(void *buf, size_t size)
315 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
317 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
319 kmem_cache_free(zio_data_buf_cache[c], buf);
323 * Dedicated I/O buffers to ensure that memory fragmentation never prevents
324 * or significantly delays the issuing of a zio. These buffers are used
325 * to aggregate I/O and could be used for raidz stripes.
330 return (kmem_cache_alloc(zio_vdev_cache, KM_PUSHPAGE));
334 zio_vdev_free(void *buf)
336 kmem_cache_free(zio_vdev_cache, buf);
341 * ==========================================================================
342 * Push and pop I/O transform buffers
343 * ==========================================================================
346 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
347 zio_transform_func_t *transform)
349 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_PUSHPAGE);
351 zt->zt_orig_data = zio->io_data;
352 zt->zt_orig_size = zio->io_size;
353 zt->zt_bufsize = bufsize;
354 zt->zt_transform = transform;
356 zt->zt_next = zio->io_transform_stack;
357 zio->io_transform_stack = zt;
364 zio_pop_transforms(zio_t *zio)
368 while ((zt = zio->io_transform_stack) != NULL) {
369 if (zt->zt_transform != NULL)
370 zt->zt_transform(zio,
371 zt->zt_orig_data, zt->zt_orig_size);
373 if (zt->zt_bufsize != 0)
374 zio_buf_free(zio->io_data, zt->zt_bufsize);
376 zio->io_data = zt->zt_orig_data;
377 zio->io_size = zt->zt_orig_size;
378 zio->io_transform_stack = zt->zt_next;
380 kmem_free(zt, sizeof (zio_transform_t));
385 * ==========================================================================
386 * I/O transform callbacks for subblocks and decompression
387 * ==========================================================================
390 zio_subblock(zio_t *zio, void *data, uint64_t size)
392 ASSERT(zio->io_size > size);
394 if (zio->io_type == ZIO_TYPE_READ)
395 bcopy(zio->io_data, data, size);
399 zio_decompress(zio_t *zio, void *data, uint64_t size)
401 if (zio->io_error == 0 &&
402 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
403 zio->io_data, data, zio->io_size, size) != 0)
408 * ==========================================================================
409 * I/O parent/child relationships and pipeline interlocks
410 * ==========================================================================
413 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
414 * continue calling these functions until they return NULL.
415 * Otherwise, the next caller will pick up the list walk in
416 * some indeterminate state. (Otherwise every caller would
417 * have to pass in a cookie to keep the state represented by
418 * io_walk_link, which gets annoying.)
421 zio_walk_parents(zio_t *cio)
423 zio_link_t *zl = cio->io_walk_link;
424 list_t *pl = &cio->io_parent_list;
426 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
427 cio->io_walk_link = zl;
432 ASSERT(zl->zl_child == cio);
433 return (zl->zl_parent);
437 zio_walk_children(zio_t *pio)
439 zio_link_t *zl = pio->io_walk_link;
440 list_t *cl = &pio->io_child_list;
442 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
443 pio->io_walk_link = zl;
448 ASSERT(zl->zl_parent == pio);
449 return (zl->zl_child);
453 zio_unique_parent(zio_t *cio)
455 zio_t *pio = zio_walk_parents(cio);
457 VERIFY(zio_walk_parents(cio) == NULL);
462 zio_add_child(zio_t *pio, zio_t *cio)
464 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_PUSHPAGE);
468 * Logical I/Os can have logical, gang, or vdev children.
469 * Gang I/Os can have gang or vdev children.
470 * Vdev I/Os can only have vdev children.
471 * The following ASSERT captures all of these constraints.
473 ASSERT(cio->io_child_type <= pio->io_child_type);
478 mutex_enter(&cio->io_lock);
479 mutex_enter(&pio->io_lock);
481 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
483 for (w = 0; w < ZIO_WAIT_TYPES; w++)
484 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
486 list_insert_head(&pio->io_child_list, zl);
487 list_insert_head(&cio->io_parent_list, zl);
489 pio->io_child_count++;
490 cio->io_parent_count++;
492 mutex_exit(&pio->io_lock);
493 mutex_exit(&cio->io_lock);
497 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
499 ASSERT(zl->zl_parent == pio);
500 ASSERT(zl->zl_child == cio);
502 mutex_enter(&cio->io_lock);
503 mutex_enter(&pio->io_lock);
505 list_remove(&pio->io_child_list, zl);
506 list_remove(&cio->io_parent_list, zl);
508 pio->io_child_count--;
509 cio->io_parent_count--;
511 mutex_exit(&pio->io_lock);
512 mutex_exit(&cio->io_lock);
514 kmem_cache_free(zio_link_cache, zl);
518 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
520 uint64_t *countp = &zio->io_children[child][wait];
521 boolean_t waiting = B_FALSE;
523 mutex_enter(&zio->io_lock);
524 ASSERT(zio->io_stall == NULL);
527 zio->io_stall = countp;
530 mutex_exit(&zio->io_lock);
535 __attribute__((always_inline))
537 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
539 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
540 int *errorp = &pio->io_child_error[zio->io_child_type];
542 mutex_enter(&pio->io_lock);
543 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
544 *errorp = zio_worst_error(*errorp, zio->io_error);
545 pio->io_reexecute |= zio->io_reexecute;
546 ASSERT3U(*countp, >, 0);
547 if (--*countp == 0 && pio->io_stall == countp) {
548 pio->io_stall = NULL;
549 mutex_exit(&pio->io_lock);
552 mutex_exit(&pio->io_lock);
557 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
559 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
560 zio->io_error = zio->io_child_error[c];
564 * ==========================================================================
565 * Create the various types of I/O (read, write, free, etc)
566 * ==========================================================================
569 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
570 void *data, uint64_t size, zio_done_func_t *done, void *private,
571 zio_type_t type, int priority, enum zio_flag flags,
572 vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
573 enum zio_stage stage, enum zio_stage pipeline)
577 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
578 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
579 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
581 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
582 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
583 ASSERT(vd || stage == ZIO_STAGE_OPEN);
585 zio = kmem_cache_alloc(zio_cache, KM_PUSHPAGE);
588 zio->io_child_type = ZIO_CHILD_VDEV;
589 else if (flags & ZIO_FLAG_GANG_CHILD)
590 zio->io_child_type = ZIO_CHILD_GANG;
591 else if (flags & ZIO_FLAG_DDT_CHILD)
592 zio->io_child_type = ZIO_CHILD_DDT;
594 zio->io_child_type = ZIO_CHILD_LOGICAL;
597 zio->io_logical = NULL;
598 zio->io_bp = (blkptr_t *)bp;
599 zio->io_bp_copy = *bp;
600 zio->io_bp_orig = *bp;
601 if (type != ZIO_TYPE_WRITE ||
602 zio->io_child_type == ZIO_CHILD_DDT)
603 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
604 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
605 zio->io_logical = zio;
606 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
607 pipeline |= ZIO_GANG_STAGES;
609 zio->io_logical = NULL;
611 bzero(&zio->io_bp_copy, sizeof (blkptr_t));
612 bzero(&zio->io_bp_orig, sizeof (blkptr_t));
617 zio->io_ready = NULL;
619 zio->io_private = private;
620 zio->io_prev_space_delta = 0;
622 zio->io_priority = priority;
625 zio->io_vsd_ops = NULL;
626 zio->io_offset = offset;
627 zio->io_deadline = 0;
628 zio->io_timestamp = 0;
631 zio->io_orig_data = zio->io_data = data;
632 zio->io_orig_size = zio->io_size = size;
633 zio->io_orig_flags = zio->io_flags = flags;
634 zio->io_orig_stage = zio->io_stage = stage;
635 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
636 bzero(&zio->io_prop, sizeof (zio_prop_t));
638 zio->io_reexecute = 0;
639 zio->io_bp_override = NULL;
640 zio->io_walk_link = NULL;
641 zio->io_transform_stack = NULL;
643 zio->io_child_count = 0;
644 zio->io_parent_count = 0;
645 zio->io_stall = NULL;
646 zio->io_gang_leader = NULL;
647 zio->io_gang_tree = NULL;
648 zio->io_executor = NULL;
649 zio->io_waiter = NULL;
650 zio->io_cksum_report = NULL;
652 bzero(zio->io_child_error, sizeof (int) * ZIO_CHILD_TYPES);
653 bzero(zio->io_children,
654 sizeof (uint64_t) * ZIO_CHILD_TYPES * ZIO_WAIT_TYPES);
655 bzero(&zio->io_bookmark, sizeof (zbookmark_t));
657 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
658 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
661 zio->io_bookmark = *zb;
664 if (zio->io_logical == NULL)
665 zio->io_logical = pio->io_logical;
666 if (zio->io_child_type == ZIO_CHILD_GANG)
667 zio->io_gang_leader = pio->io_gang_leader;
668 zio_add_child(pio, zio);
671 taskq_init_ent(&zio->io_tqent);
677 zio_destroy(zio_t *zio)
679 kmem_cache_free(zio_cache, zio);
683 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
684 void *private, enum zio_flag flags)
688 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
689 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
690 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
696 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
698 return (zio_null(NULL, spa, NULL, done, private, flags));
702 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
703 void *data, uint64_t size, zio_done_func_t *done, void *private,
704 int priority, enum zio_flag flags, const zbookmark_t *zb)
708 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
709 data, size, done, private,
710 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
711 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
712 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
718 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
719 void *data, uint64_t size, const zio_prop_t *zp,
720 zio_done_func_t *ready, zio_done_func_t *done, void *private,
721 int priority, enum zio_flag flags, const zbookmark_t *zb)
725 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
726 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
727 zp->zp_compress >= ZIO_COMPRESS_OFF &&
728 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
729 DMU_OT_IS_VALID(zp->zp_type) &&
732 zp->zp_copies <= spa_max_replication(spa) &&
734 zp->zp_dedup_verify <= 1);
736 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
737 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
738 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
739 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
741 zio->io_ready = ready;
748 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
749 uint64_t size, zio_done_func_t *done, void *private, int priority,
750 enum zio_flag flags, zbookmark_t *zb)
754 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
755 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
756 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
762 zio_write_override(zio_t *zio, blkptr_t *bp, int copies)
764 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
765 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
766 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
767 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
769 zio->io_prop.zp_copies = copies;
770 zio->io_bp_override = bp;
774 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
776 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
780 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
785 dprintf_bp(bp, "freeing in txg %llu, pass %u",
786 (longlong_t)txg, spa->spa_sync_pass);
788 ASSERT(!BP_IS_HOLE(bp));
789 ASSERT(spa_syncing_txg(spa) == txg);
790 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
794 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
795 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
796 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
802 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
803 zio_done_func_t *done, void *private, enum zio_flag flags)
808 * A claim is an allocation of a specific block. Claims are needed
809 * to support immediate writes in the intent log. The issue is that
810 * immediate writes contain committed data, but in a txg that was
811 * *not* committed. Upon opening the pool after an unclean shutdown,
812 * the intent log claims all blocks that contain immediate write data
813 * so that the SPA knows they're in use.
815 * All claims *must* be resolved in the first txg -- before the SPA
816 * starts allocating blocks -- so that nothing is allocated twice.
817 * If txg == 0 we just verify that the block is claimable.
819 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
820 ASSERT(txg == spa_first_txg(spa) || txg == 0);
821 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
823 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
824 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
825 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
831 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
832 zio_done_func_t *done, void *private, int priority, enum zio_flag flags)
837 if (vd->vdev_children == 0) {
838 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
839 ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
840 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
844 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
846 for (c = 0; c < vd->vdev_children; c++)
847 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
848 done, private, priority, flags));
855 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
856 void *data, int checksum, zio_done_func_t *done, void *private,
857 int priority, enum zio_flag flags, boolean_t labels)
861 ASSERT(vd->vdev_children == 0);
862 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
863 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
864 ASSERT3U(offset + size, <=, vd->vdev_psize);
866 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
867 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
868 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
870 zio->io_prop.zp_checksum = checksum;
876 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
877 void *data, int checksum, zio_done_func_t *done, void *private,
878 int priority, enum zio_flag flags, boolean_t labels)
882 ASSERT(vd->vdev_children == 0);
883 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
884 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
885 ASSERT3U(offset + size, <=, vd->vdev_psize);
887 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
888 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
889 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
891 zio->io_prop.zp_checksum = checksum;
893 if (zio_checksum_table[checksum].ci_eck) {
895 * zec checksums are necessarily destructive -- they modify
896 * the end of the write buffer to hold the verifier/checksum.
897 * Therefore, we must make a local copy in case the data is
898 * being written to multiple places in parallel.
900 void *wbuf = zio_buf_alloc(size);
901 bcopy(data, wbuf, size);
902 zio_push_transform(zio, wbuf, size, size, NULL);
909 * Create a child I/O to do some work for us.
912 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
913 void *data, uint64_t size, int type, int priority, enum zio_flag flags,
914 zio_done_func_t *done, void *private)
916 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
919 ASSERT(vd->vdev_parent ==
920 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
922 if (type == ZIO_TYPE_READ && bp != NULL) {
924 * If we have the bp, then the child should perform the
925 * checksum and the parent need not. This pushes error
926 * detection as close to the leaves as possible and
927 * eliminates redundant checksums in the interior nodes.
929 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
930 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
933 if (vd->vdev_children == 0)
934 offset += VDEV_LABEL_START_SIZE;
936 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
939 * If we've decided to do a repair, the write is not speculative --
940 * even if the original read was.
942 if (flags & ZIO_FLAG_IO_REPAIR)
943 flags &= ~ZIO_FLAG_SPECULATIVE;
945 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
946 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
947 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
953 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
954 int type, int priority, enum zio_flag flags,
955 zio_done_func_t *done, void *private)
959 ASSERT(vd->vdev_ops->vdev_op_leaf);
961 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
962 data, size, done, private, type, priority,
963 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
965 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
971 zio_flush(zio_t *zio, vdev_t *vd)
973 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
974 NULL, NULL, ZIO_PRIORITY_NOW,
975 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
979 zio_shrink(zio_t *zio, uint64_t size)
981 ASSERT(zio->io_executor == NULL);
982 ASSERT(zio->io_orig_size == zio->io_size);
983 ASSERT(size <= zio->io_size);
986 * We don't shrink for raidz because of problems with the
987 * reconstruction when reading back less than the block size.
988 * Note, BP_IS_RAIDZ() assumes no compression.
990 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
991 if (!BP_IS_RAIDZ(zio->io_bp))
992 zio->io_orig_size = zio->io_size = size;
996 * ==========================================================================
997 * Prepare to read and write logical blocks
998 * ==========================================================================
1002 zio_read_bp_init(zio_t *zio)
1004 blkptr_t *bp = zio->io_bp;
1006 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1007 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1008 !(zio->io_flags & ZIO_FLAG_RAW)) {
1009 uint64_t psize = BP_GET_PSIZE(bp);
1010 void *cbuf = zio_buf_alloc(psize);
1012 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
1015 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1016 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1018 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1019 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1021 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1022 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1024 return (ZIO_PIPELINE_CONTINUE);
1028 zio_write_bp_init(zio_t *zio)
1030 spa_t *spa = zio->io_spa;
1031 zio_prop_t *zp = &zio->io_prop;
1032 enum zio_compress compress = zp->zp_compress;
1033 blkptr_t *bp = zio->io_bp;
1034 uint64_t lsize = zio->io_size;
1035 uint64_t psize = lsize;
1039 * If our children haven't all reached the ready stage,
1040 * wait for them and then repeat this pipeline stage.
1042 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1043 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1044 return (ZIO_PIPELINE_STOP);
1046 if (!IO_IS_ALLOCATING(zio))
1047 return (ZIO_PIPELINE_CONTINUE);
1049 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1051 if (zio->io_bp_override) {
1052 ASSERT(bp->blk_birth != zio->io_txg);
1053 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1055 *bp = *zio->io_bp_override;
1056 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1058 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1059 return (ZIO_PIPELINE_CONTINUE);
1061 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1062 zp->zp_dedup_verify);
1064 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1065 BP_SET_DEDUP(bp, 1);
1066 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1067 return (ZIO_PIPELINE_CONTINUE);
1069 zio->io_bp_override = NULL;
1073 if (bp->blk_birth == zio->io_txg) {
1075 * We're rewriting an existing block, which means we're
1076 * working on behalf of spa_sync(). For spa_sync() to
1077 * converge, it must eventually be the case that we don't
1078 * have to allocate new blocks. But compression changes
1079 * the blocksize, which forces a reallocate, and makes
1080 * convergence take longer. Therefore, after the first
1081 * few passes, stop compressing to ensure convergence.
1083 pass = spa_sync_pass(spa);
1085 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1086 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1087 ASSERT(!BP_GET_DEDUP(bp));
1089 if (pass >= zfs_sync_pass_dont_compress)
1090 compress = ZIO_COMPRESS_OFF;
1092 /* Make sure someone doesn't change their mind on overwrites */
1093 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
1094 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1097 if (compress != ZIO_COMPRESS_OFF) {
1098 void *cbuf = zio_buf_alloc(lsize);
1099 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1100 if (psize == 0 || psize == lsize) {
1101 compress = ZIO_COMPRESS_OFF;
1102 zio_buf_free(cbuf, lsize);
1104 ASSERT(psize < lsize);
1105 zio_push_transform(zio, cbuf, psize, lsize, NULL);
1110 * The final pass of spa_sync() must be all rewrites, but the first
1111 * few passes offer a trade-off: allocating blocks defers convergence,
1112 * but newly allocated blocks are sequential, so they can be written
1113 * to disk faster. Therefore, we allow the first few passes of
1114 * spa_sync() to allocate new blocks, but force rewrites after that.
1115 * There should only be a handful of blocks after pass 1 in any case.
1117 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
1118 pass >= zfs_sync_pass_rewrite) {
1119 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1121 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1122 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1125 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1129 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1131 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1132 BP_SET_LSIZE(bp, lsize);
1133 BP_SET_PSIZE(bp, psize);
1134 BP_SET_COMPRESS(bp, compress);
1135 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1136 BP_SET_TYPE(bp, zp->zp_type);
1137 BP_SET_LEVEL(bp, zp->zp_level);
1138 BP_SET_DEDUP(bp, zp->zp_dedup);
1139 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1141 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1142 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1143 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1147 return (ZIO_PIPELINE_CONTINUE);
1151 zio_free_bp_init(zio_t *zio)
1153 blkptr_t *bp = zio->io_bp;
1155 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1156 if (BP_GET_DEDUP(bp))
1157 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1160 return (ZIO_PIPELINE_CONTINUE);
1164 * ==========================================================================
1165 * Execute the I/O pipeline
1166 * ==========================================================================
1170 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1172 spa_t *spa = zio->io_spa;
1173 zio_type_t t = zio->io_type;
1174 int flags = (cutinline ? TQ_FRONT : 0);
1177 * If we're a config writer or a probe, the normal issue and
1178 * interrupt threads may all be blocked waiting for the config lock.
1179 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1181 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1185 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1187 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1191 * If this is a high priority I/O, then use the high priority taskq if
1194 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1195 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1198 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1201 * NB: We are assuming that the zio can only be dispatched
1202 * to a single taskq at a time. It would be a grievous error
1203 * to dispatch the zio to another taskq at the same time.
1205 ASSERT(taskq_empty_ent(&zio->io_tqent));
1206 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1207 flags, &zio->io_tqent);
1211 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1213 kthread_t *executor = zio->io_executor;
1214 spa_t *spa = zio->io_spa;
1217 for (t = 0; t < ZIO_TYPES; t++) {
1218 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1220 for (i = 0; i < tqs->stqs_count; i++) {
1221 if (taskq_member(tqs->stqs_taskq[i], executor))
1230 zio_issue_async(zio_t *zio)
1232 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1234 return (ZIO_PIPELINE_STOP);
1238 zio_interrupt(zio_t *zio)
1240 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1244 * Execute the I/O pipeline until one of the following occurs:
1245 * (1) the I/O completes; (2) the pipeline stalls waiting for
1246 * dependent child I/Os; (3) the I/O issues, so we're waiting
1247 * for an I/O completion interrupt; (4) the I/O is delegated by
1248 * vdev-level caching or aggregation; (5) the I/O is deferred
1249 * due to vdev-level queueing; (6) the I/O is handed off to
1250 * another thread. In all cases, the pipeline stops whenever
1251 * there's no CPU work; it never burns a thread in cv_wait_io().
1253 * There's no locking on io_stage because there's no legitimate way
1254 * for multiple threads to be attempting to process the same I/O.
1256 static zio_pipe_stage_t *zio_pipeline[];
1259 * zio_execute() is a wrapper around the static function
1260 * __zio_execute() so that we can force __zio_execute() to be
1261 * inlined. This reduces stack overhead which is important
1262 * because __zio_execute() is called recursively in several zio
1263 * code paths. zio_execute() itself cannot be inlined because
1264 * it is externally visible.
1267 zio_execute(zio_t *zio)
1272 __attribute__((always_inline))
1274 __zio_execute(zio_t *zio)
1276 zio->io_executor = curthread;
1278 while (zio->io_stage < ZIO_STAGE_DONE) {
1279 enum zio_stage pipeline = zio->io_pipeline;
1280 enum zio_stage stage = zio->io_stage;
1285 ASSERT(!MUTEX_HELD(&zio->io_lock));
1286 ASSERT(ISP2(stage));
1287 ASSERT(zio->io_stall == NULL);
1291 } while ((stage & pipeline) == 0);
1293 ASSERT(stage <= ZIO_STAGE_DONE);
1295 dp = spa_get_dsl(zio->io_spa);
1296 cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1297 zio_requeue_io_start_cut_in_line : B_FALSE;
1300 * If we are in interrupt context and this pipeline stage
1301 * will grab a config lock that is held across I/O,
1302 * or may wait for an I/O that needs an interrupt thread
1303 * to complete, issue async to avoid deadlock.
1305 * For VDEV_IO_START, we cut in line so that the io will
1306 * be sent to disk promptly.
1308 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1309 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1310 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1316 * If we executing in the context of the tx_sync_thread,
1317 * or we are performing pool initialization outside of a
1318 * zio_taskq[ZIO_TASKQ_ISSUE] context. Then issue the zio
1319 * async to minimize stack usage for these deep call paths.
1321 if ((dp && curthread == dp->dp_tx.tx_sync_thread) ||
1322 (dp && spa_is_initializing(dp->dp_spa) &&
1323 !zio_taskq_member(zio, ZIO_TASKQ_ISSUE))) {
1324 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1329 zio->io_stage = stage;
1330 rv = zio_pipeline[highbit(stage) - 1](zio);
1332 if (rv == ZIO_PIPELINE_STOP)
1335 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1341 * ==========================================================================
1342 * Initiate I/O, either sync or async
1343 * ==========================================================================
1346 zio_wait(zio_t *zio)
1350 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1351 ASSERT(zio->io_executor == NULL);
1353 zio->io_waiter = curthread;
1357 mutex_enter(&zio->io_lock);
1358 while (zio->io_executor != NULL)
1359 cv_wait_io(&zio->io_cv, &zio->io_lock);
1360 mutex_exit(&zio->io_lock);
1362 error = zio->io_error;
1369 zio_nowait(zio_t *zio)
1371 ASSERT(zio->io_executor == NULL);
1373 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1374 zio_unique_parent(zio) == NULL) {
1376 * This is a logical async I/O with no parent to wait for it.
1377 * We add it to the spa_async_root_zio "Godfather" I/O which
1378 * will ensure they complete prior to unloading the pool.
1380 spa_t *spa = zio->io_spa;
1382 zio_add_child(spa->spa_async_zio_root, zio);
1389 * ==========================================================================
1390 * Reexecute or suspend/resume failed I/O
1391 * ==========================================================================
1395 zio_reexecute(zio_t *pio)
1397 zio_t *cio, *cio_next;
1400 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1401 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1402 ASSERT(pio->io_gang_leader == NULL);
1403 ASSERT(pio->io_gang_tree == NULL);
1405 pio->io_flags = pio->io_orig_flags;
1406 pio->io_stage = pio->io_orig_stage;
1407 pio->io_pipeline = pio->io_orig_pipeline;
1408 pio->io_reexecute = 0;
1410 for (w = 0; w < ZIO_WAIT_TYPES; w++)
1411 pio->io_state[w] = 0;
1412 for (c = 0; c < ZIO_CHILD_TYPES; c++)
1413 pio->io_child_error[c] = 0;
1415 if (IO_IS_ALLOCATING(pio))
1416 BP_ZERO(pio->io_bp);
1419 * As we reexecute pio's children, new children could be created.
1420 * New children go to the head of pio's io_child_list, however,
1421 * so we will (correctly) not reexecute them. The key is that
1422 * the remainder of pio's io_child_list, from 'cio_next' onward,
1423 * cannot be affected by any side effects of reexecuting 'cio'.
1425 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1426 cio_next = zio_walk_children(pio);
1427 mutex_enter(&pio->io_lock);
1428 for (w = 0; w < ZIO_WAIT_TYPES; w++)
1429 pio->io_children[cio->io_child_type][w]++;
1430 mutex_exit(&pio->io_lock);
1435 * Now that all children have been reexecuted, execute the parent.
1436 * We don't reexecute "The Godfather" I/O here as it's the
1437 * responsibility of the caller to wait on him.
1439 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1444 zio_suspend(spa_t *spa, zio_t *zio)
1446 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1447 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1448 "failure and the failure mode property for this pool "
1449 "is set to panic.", spa_name(spa));
1451 cmn_err(CE_WARN, "Pool '%s' has encountered an uncorrectable I/O "
1452 "failure and has been suspended.\n", spa_name(spa));
1454 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1456 mutex_enter(&spa->spa_suspend_lock);
1458 if (spa->spa_suspend_zio_root == NULL)
1459 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1460 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1461 ZIO_FLAG_GODFATHER);
1463 spa->spa_suspended = B_TRUE;
1466 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1467 ASSERT(zio != spa->spa_suspend_zio_root);
1468 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1469 ASSERT(zio_unique_parent(zio) == NULL);
1470 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1471 zio_add_child(spa->spa_suspend_zio_root, zio);
1474 mutex_exit(&spa->spa_suspend_lock);
1478 zio_resume(spa_t *spa)
1483 * Reexecute all previously suspended i/o.
1485 mutex_enter(&spa->spa_suspend_lock);
1486 spa->spa_suspended = B_FALSE;
1487 cv_broadcast(&spa->spa_suspend_cv);
1488 pio = spa->spa_suspend_zio_root;
1489 spa->spa_suspend_zio_root = NULL;
1490 mutex_exit(&spa->spa_suspend_lock);
1496 return (zio_wait(pio));
1500 zio_resume_wait(spa_t *spa)
1502 mutex_enter(&spa->spa_suspend_lock);
1503 while (spa_suspended(spa))
1504 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1505 mutex_exit(&spa->spa_suspend_lock);
1509 * ==========================================================================
1512 * A gang block is a collection of small blocks that looks to the DMU
1513 * like one large block. When zio_dva_allocate() cannot find a block
1514 * of the requested size, due to either severe fragmentation or the pool
1515 * being nearly full, it calls zio_write_gang_block() to construct the
1516 * block from smaller fragments.
1518 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1519 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1520 * an indirect block: it's an array of block pointers. It consumes
1521 * only one sector and hence is allocatable regardless of fragmentation.
1522 * The gang header's bps point to its gang members, which hold the data.
1524 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1525 * as the verifier to ensure uniqueness of the SHA256 checksum.
1526 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1527 * not the gang header. This ensures that data block signatures (needed for
1528 * deduplication) are independent of how the block is physically stored.
1530 * Gang blocks can be nested: a gang member may itself be a gang block.
1531 * Thus every gang block is a tree in which root and all interior nodes are
1532 * gang headers, and the leaves are normal blocks that contain user data.
1533 * The root of the gang tree is called the gang leader.
1535 * To perform any operation (read, rewrite, free, claim) on a gang block,
1536 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1537 * in the io_gang_tree field of the original logical i/o by recursively
1538 * reading the gang leader and all gang headers below it. This yields
1539 * an in-core tree containing the contents of every gang header and the
1540 * bps for every constituent of the gang block.
1542 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1543 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1544 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1545 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1546 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1547 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1548 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1549 * of the gang header plus zio_checksum_compute() of the data to update the
1550 * gang header's blk_cksum as described above.
1552 * The two-phase assemble/issue model solves the problem of partial failure --
1553 * what if you'd freed part of a gang block but then couldn't read the
1554 * gang header for another part? Assembling the entire gang tree first
1555 * ensures that all the necessary gang header I/O has succeeded before
1556 * starting the actual work of free, claim, or write. Once the gang tree
1557 * is assembled, free and claim are in-memory operations that cannot fail.
1559 * In the event that a gang write fails, zio_dva_unallocate() walks the
1560 * gang tree to immediately free (i.e. insert back into the space map)
1561 * everything we've allocated. This ensures that we don't get ENOSPC
1562 * errors during repeated suspend/resume cycles due to a flaky device.
1564 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1565 * the gang tree, we won't modify the block, so we can safely defer the free
1566 * (knowing that the block is still intact). If we *can* assemble the gang
1567 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1568 * each constituent bp and we can allocate a new block on the next sync pass.
1570 * In all cases, the gang tree allows complete recovery from partial failure.
1571 * ==========================================================================
1575 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1580 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1581 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1582 &pio->io_bookmark));
1586 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1591 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1592 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1593 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1595 * As we rewrite each gang header, the pipeline will compute
1596 * a new gang block header checksum for it; but no one will
1597 * compute a new data checksum, so we do that here. The one
1598 * exception is the gang leader: the pipeline already computed
1599 * its data checksum because that stage precedes gang assembly.
1600 * (Presently, nothing actually uses interior data checksums;
1601 * this is just good hygiene.)
1603 if (gn != pio->io_gang_leader->io_gang_tree) {
1604 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1605 data, BP_GET_PSIZE(bp));
1608 * If we are here to damage data for testing purposes,
1609 * leave the GBH alone so that we can detect the damage.
1611 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1612 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1614 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1615 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1616 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1624 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1626 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1627 ZIO_GANG_CHILD_FLAGS(pio)));
1632 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1634 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1635 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1638 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1647 static void zio_gang_tree_assemble_done(zio_t *zio);
1649 static zio_gang_node_t *
1650 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1652 zio_gang_node_t *gn;
1654 ASSERT(*gnpp == NULL);
1656 gn = kmem_zalloc(sizeof (*gn), KM_PUSHPAGE);
1657 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1664 zio_gang_node_free(zio_gang_node_t **gnpp)
1666 zio_gang_node_t *gn = *gnpp;
1669 for (g = 0; g < SPA_GBH_NBLKPTRS; g++)
1670 ASSERT(gn->gn_child[g] == NULL);
1672 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1673 kmem_free(gn, sizeof (*gn));
1678 zio_gang_tree_free(zio_gang_node_t **gnpp)
1680 zio_gang_node_t *gn = *gnpp;
1686 for (g = 0; g < SPA_GBH_NBLKPTRS; g++)
1687 zio_gang_tree_free(&gn->gn_child[g]);
1689 zio_gang_node_free(gnpp);
1693 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1695 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1697 ASSERT(gio->io_gang_leader == gio);
1698 ASSERT(BP_IS_GANG(bp));
1700 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1701 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1702 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1706 zio_gang_tree_assemble_done(zio_t *zio)
1708 zio_t *gio = zio->io_gang_leader;
1709 zio_gang_node_t *gn = zio->io_private;
1710 blkptr_t *bp = zio->io_bp;
1713 ASSERT(gio == zio_unique_parent(zio));
1714 ASSERT(zio->io_child_count == 0);
1719 if (BP_SHOULD_BYTESWAP(bp))
1720 byteswap_uint64_array(zio->io_data, zio->io_size);
1722 ASSERT(zio->io_data == gn->gn_gbh);
1723 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1724 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1726 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1727 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1728 if (!BP_IS_GANG(gbp))
1730 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1735 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1737 zio_t *gio = pio->io_gang_leader;
1741 ASSERT(BP_IS_GANG(bp) == !!gn);
1742 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1743 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1746 * If you're a gang header, your data is in gn->gn_gbh.
1747 * If you're a gang member, your data is in 'data' and gn == NULL.
1749 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1752 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1754 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1755 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1756 if (BP_IS_HOLE(gbp))
1758 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1759 data = (char *)data + BP_GET_PSIZE(gbp);
1763 if (gn == gio->io_gang_tree)
1764 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1771 zio_gang_assemble(zio_t *zio)
1773 blkptr_t *bp = zio->io_bp;
1775 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1776 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1778 zio->io_gang_leader = zio;
1780 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1782 return (ZIO_PIPELINE_CONTINUE);
1786 zio_gang_issue(zio_t *zio)
1788 blkptr_t *bp = zio->io_bp;
1790 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1791 return (ZIO_PIPELINE_STOP);
1793 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1794 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1796 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1797 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1799 zio_gang_tree_free(&zio->io_gang_tree);
1801 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1803 return (ZIO_PIPELINE_CONTINUE);
1807 zio_write_gang_member_ready(zio_t *zio)
1809 zio_t *pio = zio_unique_parent(zio);
1810 ASSERTV(zio_t *gio = zio->io_gang_leader;)
1811 dva_t *cdva = zio->io_bp->blk_dva;
1812 dva_t *pdva = pio->io_bp->blk_dva;
1816 if (BP_IS_HOLE(zio->io_bp))
1819 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1821 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1822 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1823 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1824 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1825 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1827 mutex_enter(&pio->io_lock);
1828 for (d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1829 ASSERT(DVA_GET_GANG(&pdva[d]));
1830 asize = DVA_GET_ASIZE(&pdva[d]);
1831 asize += DVA_GET_ASIZE(&cdva[d]);
1832 DVA_SET_ASIZE(&pdva[d], asize);
1834 mutex_exit(&pio->io_lock);
1838 zio_write_gang_block(zio_t *pio)
1840 spa_t *spa = pio->io_spa;
1841 blkptr_t *bp = pio->io_bp;
1842 zio_t *gio = pio->io_gang_leader;
1844 zio_gang_node_t *gn, **gnpp;
1845 zio_gbh_phys_t *gbh;
1846 uint64_t txg = pio->io_txg;
1847 uint64_t resid = pio->io_size;
1849 int copies = gio->io_prop.zp_copies;
1850 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1854 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1855 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1856 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1858 pio->io_error = error;
1859 return (ZIO_PIPELINE_CONTINUE);
1863 gnpp = &gio->io_gang_tree;
1865 gnpp = pio->io_private;
1866 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1869 gn = zio_gang_node_alloc(gnpp);
1871 bzero(gbh, SPA_GANGBLOCKSIZE);
1874 * Create the gang header.
1876 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1877 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1880 * Create and nowait the gang children.
1882 for (g = 0; resid != 0; resid -= lsize, g++) {
1883 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1885 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1887 zp.zp_checksum = gio->io_prop.zp_checksum;
1888 zp.zp_compress = ZIO_COMPRESS_OFF;
1889 zp.zp_type = DMU_OT_NONE;
1891 zp.zp_copies = gio->io_prop.zp_copies;
1893 zp.zp_dedup_verify = 0;
1895 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1896 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1897 zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1898 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1899 &pio->io_bookmark));
1903 * Set pio's pipeline to just wait for zio to finish.
1905 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1908 * We didn't allocate this bp, so make sure it doesn't get unmarked.
1910 pio->io_flags &= ~ZIO_FLAG_FASTWRITE;
1914 return (ZIO_PIPELINE_CONTINUE);
1918 * ==========================================================================
1920 * ==========================================================================
1923 zio_ddt_child_read_done(zio_t *zio)
1925 blkptr_t *bp = zio->io_bp;
1926 ddt_entry_t *dde = zio->io_private;
1928 zio_t *pio = zio_unique_parent(zio);
1930 mutex_enter(&pio->io_lock);
1931 ddp = ddt_phys_select(dde, bp);
1932 if (zio->io_error == 0)
1933 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1934 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1935 dde->dde_repair_data = zio->io_data;
1937 zio_buf_free(zio->io_data, zio->io_size);
1938 mutex_exit(&pio->io_lock);
1942 zio_ddt_read_start(zio_t *zio)
1944 blkptr_t *bp = zio->io_bp;
1947 ASSERT(BP_GET_DEDUP(bp));
1948 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1949 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1951 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1952 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1953 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1954 ddt_phys_t *ddp = dde->dde_phys;
1955 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1958 ASSERT(zio->io_vsd == NULL);
1961 if (ddp_self == NULL)
1962 return (ZIO_PIPELINE_CONTINUE);
1964 for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1965 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1967 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1969 zio_nowait(zio_read(zio, zio->io_spa, &blk,
1970 zio_buf_alloc(zio->io_size), zio->io_size,
1971 zio_ddt_child_read_done, dde, zio->io_priority,
1972 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1973 &zio->io_bookmark));
1975 return (ZIO_PIPELINE_CONTINUE);
1978 zio_nowait(zio_read(zio, zio->io_spa, bp,
1979 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1980 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1982 return (ZIO_PIPELINE_CONTINUE);
1986 zio_ddt_read_done(zio_t *zio)
1988 blkptr_t *bp = zio->io_bp;
1990 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1991 return (ZIO_PIPELINE_STOP);
1993 ASSERT(BP_GET_DEDUP(bp));
1994 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1995 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1997 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1998 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1999 ddt_entry_t *dde = zio->io_vsd;
2001 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2002 return (ZIO_PIPELINE_CONTINUE);
2005 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2006 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2007 return (ZIO_PIPELINE_STOP);
2009 if (dde->dde_repair_data != NULL) {
2010 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2011 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2013 ddt_repair_done(ddt, dde);
2017 ASSERT(zio->io_vsd == NULL);
2019 return (ZIO_PIPELINE_CONTINUE);
2023 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2025 spa_t *spa = zio->io_spa;
2029 * Note: we compare the original data, not the transformed data,
2030 * because when zio->io_bp is an override bp, we will not have
2031 * pushed the I/O transforms. That's an important optimization
2032 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2034 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2035 zio_t *lio = dde->dde_lead_zio[p];
2038 return (lio->io_orig_size != zio->io_orig_size ||
2039 bcmp(zio->io_orig_data, lio->io_orig_data,
2040 zio->io_orig_size) != 0);
2044 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2045 ddt_phys_t *ddp = &dde->dde_phys[p];
2047 if (ddp->ddp_phys_birth != 0) {
2048 arc_buf_t *abuf = NULL;
2049 uint32_t aflags = ARC_WAIT;
2050 blkptr_t blk = *zio->io_bp;
2053 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2057 error = arc_read(NULL, spa, &blk,
2058 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2059 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2060 &aflags, &zio->io_bookmark);
2063 if (arc_buf_size(abuf) != zio->io_orig_size ||
2064 bcmp(abuf->b_data, zio->io_orig_data,
2065 zio->io_orig_size) != 0)
2067 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
2071 return (error != 0);
2079 zio_ddt_child_write_ready(zio_t *zio)
2081 int p = zio->io_prop.zp_copies;
2082 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2083 ddt_entry_t *dde = zio->io_private;
2084 ddt_phys_t *ddp = &dde->dde_phys[p];
2092 ASSERT(dde->dde_lead_zio[p] == zio);
2094 ddt_phys_fill(ddp, zio->io_bp);
2096 while ((pio = zio_walk_parents(zio)) != NULL)
2097 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2103 zio_ddt_child_write_done(zio_t *zio)
2105 int p = zio->io_prop.zp_copies;
2106 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2107 ddt_entry_t *dde = zio->io_private;
2108 ddt_phys_t *ddp = &dde->dde_phys[p];
2112 ASSERT(ddp->ddp_refcnt == 0);
2113 ASSERT(dde->dde_lead_zio[p] == zio);
2114 dde->dde_lead_zio[p] = NULL;
2116 if (zio->io_error == 0) {
2117 while (zio_walk_parents(zio) != NULL)
2118 ddt_phys_addref(ddp);
2120 ddt_phys_clear(ddp);
2127 zio_ddt_ditto_write_done(zio_t *zio)
2129 int p = DDT_PHYS_DITTO;
2130 blkptr_t *bp = zio->io_bp;
2131 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2132 ddt_entry_t *dde = zio->io_private;
2133 ddt_phys_t *ddp = &dde->dde_phys[p];
2134 ddt_key_t *ddk = &dde->dde_key;
2135 ASSERTV(zio_prop_t *zp = &zio->io_prop);
2139 ASSERT(ddp->ddp_refcnt == 0);
2140 ASSERT(dde->dde_lead_zio[p] == zio);
2141 dde->dde_lead_zio[p] = NULL;
2143 if (zio->io_error == 0) {
2144 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2145 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2146 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2147 if (ddp->ddp_phys_birth != 0)
2148 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2149 ddt_phys_fill(ddp, bp);
2156 zio_ddt_write(zio_t *zio)
2158 spa_t *spa = zio->io_spa;
2159 blkptr_t *bp = zio->io_bp;
2160 uint64_t txg = zio->io_txg;
2161 zio_prop_t *zp = &zio->io_prop;
2162 int p = zp->zp_copies;
2166 ddt_t *ddt = ddt_select(spa, bp);
2170 ASSERT(BP_GET_DEDUP(bp));
2171 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2172 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2175 dde = ddt_lookup(ddt, bp, B_TRUE);
2176 ddp = &dde->dde_phys[p];
2178 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2180 * If we're using a weak checksum, upgrade to a strong checksum
2181 * and try again. If we're already using a strong checksum,
2182 * we can't resolve it, so just convert to an ordinary write.
2183 * (And automatically e-mail a paper to Nature?)
2185 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2186 zp->zp_checksum = spa_dedup_checksum(spa);
2187 zio_pop_transforms(zio);
2188 zio->io_stage = ZIO_STAGE_OPEN;
2193 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2195 return (ZIO_PIPELINE_CONTINUE);
2198 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2199 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2201 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2202 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2203 zio_prop_t czp = *zp;
2205 czp.zp_copies = ditto_copies;
2208 * If we arrived here with an override bp, we won't have run
2209 * the transform stack, so we won't have the data we need to
2210 * generate a child i/o. So, toss the override bp and restart.
2211 * This is safe, because using the override bp is just an
2212 * optimization; and it's rare, so the cost doesn't matter.
2214 if (zio->io_bp_override) {
2215 zio_pop_transforms(zio);
2216 zio->io_stage = ZIO_STAGE_OPEN;
2217 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2218 zio->io_bp_override = NULL;
2221 return (ZIO_PIPELINE_CONTINUE);
2224 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2225 zio->io_orig_size, &czp, NULL,
2226 zio_ddt_ditto_write_done, dde, zio->io_priority,
2227 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2229 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2230 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2233 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2234 if (ddp->ddp_phys_birth != 0)
2235 ddt_bp_fill(ddp, bp, txg);
2236 if (dde->dde_lead_zio[p] != NULL)
2237 zio_add_child(zio, dde->dde_lead_zio[p]);
2239 ddt_phys_addref(ddp);
2240 } else if (zio->io_bp_override) {
2241 ASSERT(bp->blk_birth == txg);
2242 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2243 ddt_phys_fill(ddp, bp);
2244 ddt_phys_addref(ddp);
2246 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2247 zio->io_orig_size, zp, zio_ddt_child_write_ready,
2248 zio_ddt_child_write_done, dde, zio->io_priority,
2249 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2251 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2252 dde->dde_lead_zio[p] = cio;
2262 return (ZIO_PIPELINE_CONTINUE);
2265 ddt_entry_t *freedde; /* for debugging */
2268 zio_ddt_free(zio_t *zio)
2270 spa_t *spa = zio->io_spa;
2271 blkptr_t *bp = zio->io_bp;
2272 ddt_t *ddt = ddt_select(spa, bp);
2276 ASSERT(BP_GET_DEDUP(bp));
2277 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2280 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2282 ddp = ddt_phys_select(dde, bp);
2284 ddt_phys_decref(ddp);
2288 return (ZIO_PIPELINE_CONTINUE);
2292 * ==========================================================================
2293 * Allocate and free blocks
2294 * ==========================================================================
2297 zio_dva_allocate(zio_t *zio)
2299 spa_t *spa = zio->io_spa;
2300 metaslab_class_t *mc = spa_normal_class(spa);
2301 blkptr_t *bp = zio->io_bp;
2305 if (zio->io_gang_leader == NULL) {
2306 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2307 zio->io_gang_leader = zio;
2310 ASSERT(BP_IS_HOLE(bp));
2311 ASSERT0(BP_GET_NDVAS(bp));
2312 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2313 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2314 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2317 * The dump device does not support gang blocks so allocation on
2318 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2319 * the "fast" gang feature.
2321 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2322 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2323 METASLAB_GANG_CHILD : 0;
2324 flags |= (zio->io_flags & ZIO_FLAG_FASTWRITE) ? METASLAB_FASTWRITE : 0;
2325 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2326 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2329 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2330 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2332 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2333 return (zio_write_gang_block(zio));
2334 zio->io_error = error;
2337 return (ZIO_PIPELINE_CONTINUE);
2341 zio_dva_free(zio_t *zio)
2343 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2345 return (ZIO_PIPELINE_CONTINUE);
2349 zio_dva_claim(zio_t *zio)
2353 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2355 zio->io_error = error;
2357 return (ZIO_PIPELINE_CONTINUE);
2361 * Undo an allocation. This is used by zio_done() when an I/O fails
2362 * and we want to give back the block we just allocated.
2363 * This handles both normal blocks and gang blocks.
2366 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2370 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2371 ASSERT(zio->io_bp_override == NULL);
2373 if (!BP_IS_HOLE(bp))
2374 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2377 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2378 zio_dva_unallocate(zio, gn->gn_child[g],
2379 &gn->gn_gbh->zg_blkptr[g]);
2385 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2388 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, uint64_t size,
2393 ASSERT(txg > spa_syncing_txg(spa));
2396 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2397 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2398 * when allocating them.
2401 error = metaslab_alloc(spa, spa_log_class(spa), size,
2402 new_bp, 1, txg, NULL,
2403 METASLAB_FASTWRITE | METASLAB_GANG_AVOID);
2407 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2408 new_bp, 1, txg, NULL,
2409 METASLAB_FASTWRITE | METASLAB_GANG_AVOID);
2413 BP_SET_LSIZE(new_bp, size);
2414 BP_SET_PSIZE(new_bp, size);
2415 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2416 BP_SET_CHECKSUM(new_bp,
2417 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2418 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2419 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2420 BP_SET_LEVEL(new_bp, 0);
2421 BP_SET_DEDUP(new_bp, 0);
2422 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2429 * Free an intent log block.
2432 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2434 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2435 ASSERT(!BP_IS_GANG(bp));
2437 zio_free(spa, txg, bp);
2441 * ==========================================================================
2442 * Read and write to physical devices
2443 * ==========================================================================
2446 zio_vdev_io_start(zio_t *zio)
2448 vdev_t *vd = zio->io_vd;
2450 spa_t *spa = zio->io_spa;
2452 ASSERT(zio->io_error == 0);
2453 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2456 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2457 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2460 * The mirror_ops handle multiple DVAs in a single BP.
2462 return (vdev_mirror_ops.vdev_op_io_start(zio));
2466 * We keep track of time-sensitive I/Os so that the scan thread
2467 * can quickly react to certain workloads. In particular, we care
2468 * about non-scrubbing, top-level reads and writes with the following
2470 * - synchronous writes of user data to non-slog devices
2471 * - any reads of user data
2472 * When these conditions are met, adjust the timestamp of spa_last_io
2473 * which allows the scan thread to adjust its workload accordingly.
2475 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2476 vd == vd->vdev_top && !vd->vdev_islog &&
2477 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2478 zio->io_txg != spa_syncing_txg(spa)) {
2479 uint64_t old = spa->spa_last_io;
2480 uint64_t new = ddi_get_lbolt64();
2482 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2485 align = 1ULL << vd->vdev_top->vdev_ashift;
2487 if (P2PHASE(zio->io_size, align) != 0) {
2488 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2489 char *abuf = zio_buf_alloc(asize);
2490 ASSERT(vd == vd->vdev_top);
2491 if (zio->io_type == ZIO_TYPE_WRITE) {
2492 bcopy(zio->io_data, abuf, zio->io_size);
2493 bzero(abuf + zio->io_size, asize - zio->io_size);
2495 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2498 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2499 ASSERT(P2PHASE(zio->io_size, align) == 0);
2500 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2503 * If this is a repair I/O, and there's no self-healing involved --
2504 * that is, we're just resilvering what we expect to resilver --
2505 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2506 * This prevents spurious resilvering with nested replication.
2507 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2508 * A is out of date, we'll read from C+D, then use the data to
2509 * resilver A+B -- but we don't actually want to resilver B, just A.
2510 * The top-level mirror has no way to know this, so instead we just
2511 * discard unnecessary repairs as we work our way down the vdev tree.
2512 * The same logic applies to any form of nested replication:
2513 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2515 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2516 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2517 zio->io_txg != 0 && /* not a delegated i/o */
2518 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2519 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2520 zio_vdev_io_bypass(zio);
2521 return (ZIO_PIPELINE_CONTINUE);
2524 if (vd->vdev_ops->vdev_op_leaf &&
2525 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2527 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2528 return (ZIO_PIPELINE_CONTINUE);
2530 if ((zio = vdev_queue_io(zio)) == NULL)
2531 return (ZIO_PIPELINE_STOP);
2533 if (!vdev_accessible(vd, zio)) {
2534 zio->io_error = ENXIO;
2536 return (ZIO_PIPELINE_STOP);
2540 return (vd->vdev_ops->vdev_op_io_start(zio));
2544 zio_vdev_io_done(zio_t *zio)
2546 vdev_t *vd = zio->io_vd;
2547 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2548 boolean_t unexpected_error = B_FALSE;
2550 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2551 return (ZIO_PIPELINE_STOP);
2553 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2555 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2557 vdev_queue_io_done(zio);
2559 if (zio->io_type == ZIO_TYPE_WRITE)
2560 vdev_cache_write(zio);
2562 if (zio_injection_enabled && zio->io_error == 0)
2563 zio->io_error = zio_handle_device_injection(vd,
2566 if (zio_injection_enabled && zio->io_error == 0)
2567 zio->io_error = zio_handle_label_injection(zio, EIO);
2569 if (zio->io_error) {
2570 if (!vdev_accessible(vd, zio)) {
2571 zio->io_error = ENXIO;
2573 unexpected_error = B_TRUE;
2578 ops->vdev_op_io_done(zio);
2580 if (unexpected_error)
2581 VERIFY(vdev_probe(vd, zio) == NULL);
2583 return (ZIO_PIPELINE_CONTINUE);
2587 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2588 * disk, and use that to finish the checksum ereport later.
2591 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2592 const void *good_buf)
2594 /* no processing needed */
2595 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2600 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2602 void *buf = zio_buf_alloc(zio->io_size);
2604 bcopy(zio->io_data, buf, zio->io_size);
2606 zcr->zcr_cbinfo = zio->io_size;
2607 zcr->zcr_cbdata = buf;
2608 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2609 zcr->zcr_free = zio_buf_free;
2613 zio_vdev_io_assess(zio_t *zio)
2615 vdev_t *vd = zio->io_vd;
2617 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2618 return (ZIO_PIPELINE_STOP);
2620 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2621 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2623 if (zio->io_vsd != NULL) {
2624 zio->io_vsd_ops->vsd_free(zio);
2628 if (zio_injection_enabled && zio->io_error == 0)
2629 zio->io_error = zio_handle_fault_injection(zio, EIO);
2632 * If the I/O failed, determine whether we should attempt to retry it.
2634 * On retry, we cut in line in the issue queue, since we don't want
2635 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2637 if (zio->io_error && vd == NULL &&
2638 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2639 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2640 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2642 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2643 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2644 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2645 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2646 zio_requeue_io_start_cut_in_line);
2647 return (ZIO_PIPELINE_STOP);
2651 * If we got an error on a leaf device, convert it to ENXIO
2652 * if the device is not accessible at all.
2654 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2655 !vdev_accessible(vd, zio))
2656 zio->io_error = ENXIO;
2659 * If we can't write to an interior vdev (mirror or RAID-Z),
2660 * set vdev_cant_write so that we stop trying to allocate from it.
2662 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2663 vd != NULL && !vd->vdev_ops->vdev_op_leaf)
2664 vd->vdev_cant_write = B_TRUE;
2667 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2669 return (ZIO_PIPELINE_CONTINUE);
2673 zio_vdev_io_reissue(zio_t *zio)
2675 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2676 ASSERT(zio->io_error == 0);
2678 zio->io_stage >>= 1;
2682 zio_vdev_io_redone(zio_t *zio)
2684 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2686 zio->io_stage >>= 1;
2690 zio_vdev_io_bypass(zio_t *zio)
2692 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2693 ASSERT(zio->io_error == 0);
2695 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2696 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2700 * ==========================================================================
2701 * Generate and verify checksums
2702 * ==========================================================================
2705 zio_checksum_generate(zio_t *zio)
2707 blkptr_t *bp = zio->io_bp;
2708 enum zio_checksum checksum;
2712 * This is zio_write_phys().
2713 * We're either generating a label checksum, or none at all.
2715 checksum = zio->io_prop.zp_checksum;
2717 if (checksum == ZIO_CHECKSUM_OFF)
2718 return (ZIO_PIPELINE_CONTINUE);
2720 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2722 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2723 ASSERT(!IO_IS_ALLOCATING(zio));
2724 checksum = ZIO_CHECKSUM_GANG_HEADER;
2726 checksum = BP_GET_CHECKSUM(bp);
2730 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2732 return (ZIO_PIPELINE_CONTINUE);
2736 zio_checksum_verify(zio_t *zio)
2738 zio_bad_cksum_t info;
2739 blkptr_t *bp = zio->io_bp;
2742 ASSERT(zio->io_vd != NULL);
2746 * This is zio_read_phys().
2747 * We're either verifying a label checksum, or nothing at all.
2749 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2750 return (ZIO_PIPELINE_CONTINUE);
2752 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2755 if ((error = zio_checksum_error(zio, &info)) != 0) {
2756 zio->io_error = error;
2757 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2758 zfs_ereport_start_checksum(zio->io_spa,
2759 zio->io_vd, zio, zio->io_offset,
2760 zio->io_size, NULL, &info);
2764 return (ZIO_PIPELINE_CONTINUE);
2768 * Called by RAID-Z to ensure we don't compute the checksum twice.
2771 zio_checksum_verified(zio_t *zio)
2773 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2777 * ==========================================================================
2778 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2779 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2780 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2781 * indicate errors that are specific to one I/O, and most likely permanent.
2782 * Any other error is presumed to be worse because we weren't expecting it.
2783 * ==========================================================================
2786 zio_worst_error(int e1, int e2)
2788 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2791 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2792 if (e1 == zio_error_rank[r1])
2795 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2796 if (e2 == zio_error_rank[r2])
2799 return (r1 > r2 ? e1 : e2);
2803 * ==========================================================================
2805 * ==========================================================================
2808 zio_ready(zio_t *zio)
2810 blkptr_t *bp = zio->io_bp;
2811 zio_t *pio, *pio_next;
2813 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2814 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2815 return (ZIO_PIPELINE_STOP);
2817 if (zio->io_ready) {
2818 ASSERT(IO_IS_ALLOCATING(zio));
2819 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2820 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2825 if (bp != NULL && bp != &zio->io_bp_copy)
2826 zio->io_bp_copy = *bp;
2829 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2831 mutex_enter(&zio->io_lock);
2832 zio->io_state[ZIO_WAIT_READY] = 1;
2833 pio = zio_walk_parents(zio);
2834 mutex_exit(&zio->io_lock);
2837 * As we notify zio's parents, new parents could be added.
2838 * New parents go to the head of zio's io_parent_list, however,
2839 * so we will (correctly) not notify them. The remainder of zio's
2840 * io_parent_list, from 'pio_next' onward, cannot change because
2841 * all parents must wait for us to be done before they can be done.
2843 for (; pio != NULL; pio = pio_next) {
2844 pio_next = zio_walk_parents(zio);
2845 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2848 if (zio->io_flags & ZIO_FLAG_NODATA) {
2849 if (BP_IS_GANG(bp)) {
2850 zio->io_flags &= ~ZIO_FLAG_NODATA;
2852 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2853 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2857 if (zio_injection_enabled &&
2858 zio->io_spa->spa_syncing_txg == zio->io_txg)
2859 zio_handle_ignored_writes(zio);
2861 return (ZIO_PIPELINE_CONTINUE);
2865 zio_done(zio_t *zio)
2867 zio_t *pio, *pio_next;
2871 * If our children haven't all completed,
2872 * wait for them and then repeat this pipeline stage.
2874 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2875 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2876 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2877 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2878 return (ZIO_PIPELINE_STOP);
2880 for (c = 0; c < ZIO_CHILD_TYPES; c++)
2881 for (w = 0; w < ZIO_WAIT_TYPES; w++)
2882 ASSERT(zio->io_children[c][w] == 0);
2884 if (zio->io_bp != NULL) {
2885 ASSERT(zio->io_bp->blk_pad[0] == 0);
2886 ASSERT(zio->io_bp->blk_pad[1] == 0);
2887 ASSERT(bcmp(zio->io_bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2888 (zio->io_bp == zio_unique_parent(zio)->io_bp));
2889 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) &&
2890 zio->io_bp_override == NULL &&
2891 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2892 ASSERT(!BP_SHOULD_BYTESWAP(zio->io_bp));
2893 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2894 ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 ||
2895 (BP_COUNT_GANG(zio->io_bp) == BP_GET_NDVAS(zio->io_bp)));
2900 * If there were child vdev/gang/ddt errors, they apply to us now.
2902 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2903 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2904 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2907 * If the I/O on the transformed data was successful, generate any
2908 * checksum reports now while we still have the transformed data.
2910 if (zio->io_error == 0) {
2911 while (zio->io_cksum_report != NULL) {
2912 zio_cksum_report_t *zcr = zio->io_cksum_report;
2913 uint64_t align = zcr->zcr_align;
2914 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2915 char *abuf = zio->io_data;
2917 if (asize != zio->io_size) {
2918 abuf = zio_buf_alloc(asize);
2919 bcopy(zio->io_data, abuf, zio->io_size);
2920 bzero(abuf + zio->io_size, asize - zio->io_size);
2923 zio->io_cksum_report = zcr->zcr_next;
2924 zcr->zcr_next = NULL;
2925 zcr->zcr_finish(zcr, abuf);
2926 zfs_ereport_free_checksum(zcr);
2928 if (asize != zio->io_size)
2929 zio_buf_free(abuf, asize);
2933 zio_pop_transforms(zio); /* note: may set zio->io_error */
2935 vdev_stat_update(zio, zio->io_size);
2938 * If this I/O is attached to a particular vdev is slow, exceeding
2939 * 30 seconds to complete, post an error described the I/O delay.
2940 * We ignore these errors if the device is currently unavailable.
2942 if (zio->io_delay >= MSEC_TO_TICK(zio_delay_max)) {
2943 if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd))
2944 zfs_ereport_post(FM_EREPORT_ZFS_DELAY, zio->io_spa,
2945 zio->io_vd, zio, 0, 0);
2948 if (zio->io_error) {
2950 * If this I/O is attached to a particular vdev,
2951 * generate an error message describing the I/O failure
2952 * at the block level. We ignore these errors if the
2953 * device is currently unavailable.
2955 if (zio->io_error != ECKSUM && zio->io_vd != NULL &&
2956 !vdev_is_dead(zio->io_vd))
2957 zfs_ereport_post(FM_EREPORT_ZFS_IO, zio->io_spa,
2958 zio->io_vd, zio, 0, 0);
2960 if ((zio->io_error == EIO || !(zio->io_flags &
2961 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2962 zio == zio->io_logical) {
2964 * For logical I/O requests, tell the SPA to log the
2965 * error and generate a logical data ereport.
2967 spa_log_error(zio->io_spa, zio);
2968 zfs_ereport_post(FM_EREPORT_ZFS_DATA, zio->io_spa, NULL, zio,
2973 if (zio->io_error && zio == zio->io_logical) {
2975 * Determine whether zio should be reexecuted. This will
2976 * propagate all the way to the root via zio_notify_parent().
2978 ASSERT(zio->io_vd == NULL && zio->io_bp != NULL);
2979 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2981 if (IO_IS_ALLOCATING(zio) &&
2982 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2983 if (zio->io_error != ENOSPC)
2984 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2986 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2989 if ((zio->io_type == ZIO_TYPE_READ ||
2990 zio->io_type == ZIO_TYPE_FREE) &&
2991 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
2992 zio->io_error == ENXIO &&
2993 spa_load_state(zio->io_spa) == SPA_LOAD_NONE &&
2994 spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE)
2995 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2997 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2998 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3001 * Here is a possibly good place to attempt to do
3002 * either combinatorial reconstruction or error correction
3003 * based on checksums. It also might be a good place
3004 * to send out preliminary ereports before we suspend
3010 * If there were logical child errors, they apply to us now.
3011 * We defer this until now to avoid conflating logical child
3012 * errors with errors that happened to the zio itself when
3013 * updating vdev stats and reporting FMA events above.
3015 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3017 if ((zio->io_error || zio->io_reexecute) &&
3018 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3019 !(zio->io_flags & ZIO_FLAG_IO_REWRITE))
3020 zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp);
3022 zio_gang_tree_free(&zio->io_gang_tree);
3025 * Godfather I/Os should never suspend.
3027 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3028 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3029 zio->io_reexecute = 0;
3031 if (zio->io_reexecute) {
3033 * This is a logical I/O that wants to reexecute.
3035 * Reexecute is top-down. When an i/o fails, if it's not
3036 * the root, it simply notifies its parent and sticks around.
3037 * The parent, seeing that it still has children in zio_done(),
3038 * does the same. This percolates all the way up to the root.
3039 * The root i/o will reexecute or suspend the entire tree.
3041 * This approach ensures that zio_reexecute() honors
3042 * all the original i/o dependency relationships, e.g.
3043 * parents not executing until children are ready.
3045 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3047 zio->io_gang_leader = NULL;
3049 mutex_enter(&zio->io_lock);
3050 zio->io_state[ZIO_WAIT_DONE] = 1;
3051 mutex_exit(&zio->io_lock);
3054 * "The Godfather" I/O monitors its children but is
3055 * not a true parent to them. It will track them through
3056 * the pipeline but severs its ties whenever they get into
3057 * trouble (e.g. suspended). This allows "The Godfather"
3058 * I/O to return status without blocking.
3060 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3061 zio_link_t *zl = zio->io_walk_link;
3062 pio_next = zio_walk_parents(zio);
3064 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3065 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3066 zio_remove_child(pio, zio, zl);
3067 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3071 if ((pio = zio_unique_parent(zio)) != NULL) {
3073 * We're not a root i/o, so there's nothing to do
3074 * but notify our parent. Don't propagate errors
3075 * upward since we haven't permanently failed yet.
3077 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3078 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3079 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3080 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3082 * We'd fail again if we reexecuted now, so suspend
3083 * until conditions improve (e.g. device comes online).
3085 zio_suspend(zio->io_spa, zio);
3088 * Reexecution is potentially a huge amount of work.
3089 * Hand it off to the otherwise-unused claim taskq.
3091 ASSERT(taskq_empty_ent(&zio->io_tqent));
3092 spa_taskq_dispatch_ent(zio->io_spa,
3093 ZIO_TYPE_CLAIM, ZIO_TASKQ_ISSUE,
3094 (task_func_t *)zio_reexecute, zio, 0,
3097 return (ZIO_PIPELINE_STOP);
3100 ASSERT(zio->io_child_count == 0);
3101 ASSERT(zio->io_reexecute == 0);
3102 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3105 * Report any checksum errors, since the I/O is complete.
3107 while (zio->io_cksum_report != NULL) {
3108 zio_cksum_report_t *zcr = zio->io_cksum_report;
3109 zio->io_cksum_report = zcr->zcr_next;
3110 zcr->zcr_next = NULL;
3111 zcr->zcr_finish(zcr, NULL);
3112 zfs_ereport_free_checksum(zcr);
3115 if (zio->io_flags & ZIO_FLAG_FASTWRITE && zio->io_bp &&
3116 !BP_IS_HOLE(zio->io_bp)) {
3117 metaslab_fastwrite_unmark(zio->io_spa, zio->io_bp);
3121 * It is the responsibility of the done callback to ensure that this
3122 * particular zio is no longer discoverable for adoption, and as
3123 * such, cannot acquire any new parents.
3128 mutex_enter(&zio->io_lock);
3129 zio->io_state[ZIO_WAIT_DONE] = 1;
3130 mutex_exit(&zio->io_lock);
3132 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3133 zio_link_t *zl = zio->io_walk_link;
3134 pio_next = zio_walk_parents(zio);
3135 zio_remove_child(pio, zio, zl);
3136 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3139 if (zio->io_waiter != NULL) {
3140 mutex_enter(&zio->io_lock);
3141 zio->io_executor = NULL;
3142 cv_broadcast(&zio->io_cv);
3143 mutex_exit(&zio->io_lock);
3148 return (ZIO_PIPELINE_STOP);
3152 * ==========================================================================
3153 * I/O pipeline definition
3154 * ==========================================================================
3156 static zio_pipe_stage_t *zio_pipeline[] = {
3162 zio_checksum_generate,
3176 zio_checksum_verify,
3180 /* dnp is the dnode for zb1->zb_object */
3182 zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_t *zb1,
3183 const zbookmark_t *zb2)
3185 uint64_t zb1nextL0, zb2thisobj;
3187 ASSERT(zb1->zb_objset == zb2->zb_objset);
3188 ASSERT(zb2->zb_level == 0);
3191 * A bookmark in the deadlist is considered to be after
3194 if (zb2->zb_object == DMU_DEADLIST_OBJECT)
3197 /* The objset_phys_t isn't before anything. */
3201 zb1nextL0 = (zb1->zb_blkid + 1) <<
3202 ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
3204 zb2thisobj = zb2->zb_object ? zb2->zb_object :
3205 zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
3207 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3208 uint64_t nextobj = zb1nextL0 *
3209 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
3210 return (nextobj <= zb2thisobj);
3213 if (zb1->zb_object < zb2thisobj)
3215 if (zb1->zb_object > zb2thisobj)
3217 if (zb2->zb_object == DMU_META_DNODE_OBJECT)
3219 return (zb1nextL0 <= zb2->zb_blkid);
3222 #if defined(_KERNEL) && defined(HAVE_SPL)
3223 /* Fault injection */
3224 EXPORT_SYMBOL(zio_injection_enabled);
3225 EXPORT_SYMBOL(zio_inject_fault);
3226 EXPORT_SYMBOL(zio_inject_list_next);
3227 EXPORT_SYMBOL(zio_clear_fault);
3228 EXPORT_SYMBOL(zio_handle_fault_injection);
3229 EXPORT_SYMBOL(zio_handle_device_injection);
3230 EXPORT_SYMBOL(zio_handle_label_injection);
3231 EXPORT_SYMBOL(zio_priority_table);
3232 EXPORT_SYMBOL(zio_type_name);
3234 module_param(zio_bulk_flags, int, 0644);
3235 MODULE_PARM_DESC(zio_bulk_flags, "Additional flags to pass to bulk buffers");
3237 module_param(zio_delay_max, int, 0644);
3238 MODULE_PARM_DESC(zio_delay_max, "Max zio millisec delay before posting event");
3240 module_param(zio_requeue_io_start_cut_in_line, int, 0644);
3241 MODULE_PARM_DESC(zio_requeue_io_start_cut_in_line, "Prioritize requeued I/O");
3243 module_param(zfs_sync_pass_deferred_free, int, 0644);
3244 MODULE_PARM_DESC(zfs_sync_pass_deferred_free,
3245 "defer frees starting in this pass");
3247 module_param(zfs_sync_pass_dont_compress, int, 0644);
3248 MODULE_PARM_DESC(zfs_sync_pass_dont_compress,
3249 "don't compress starting in this pass");
3251 module_param(zfs_sync_pass_rewrite, int, 0644);
3252 MODULE_PARM_DESC(zfs_sync_pass_rewrite,
3253 "rewrite new bps starting in this pass");