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);
103 zio_cons(void *arg, void *unused, int kmflag)
107 bzero(zio, sizeof (zio_t));
109 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
110 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
112 list_create(&zio->io_parent_list, sizeof (zio_link_t),
113 offsetof(zio_link_t, zl_parent_node));
114 list_create(&zio->io_child_list, sizeof (zio_link_t),
115 offsetof(zio_link_t, zl_child_node));
121 zio_dest(void *arg, void *unused)
125 mutex_destroy(&zio->io_lock);
126 cv_destroy(&zio->io_cv);
127 list_destroy(&zio->io_parent_list);
128 list_destroy(&zio->io_child_list);
135 vmem_t *data_alloc_arena = NULL;
138 data_alloc_arena = zio_alloc_arena;
140 zio_cache = kmem_cache_create("zio_cache", sizeof (zio_t), 0,
141 zio_cons, zio_dest, NULL, NULL, NULL, KMC_KMEM);
142 zio_link_cache = kmem_cache_create("zio_link_cache",
143 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, KMC_KMEM);
146 * For small buffers, we want a cache for each multiple of
147 * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache
148 * for each quarter-power of 2. For large buffers, we want
149 * a cache for each multiple of PAGESIZE.
151 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
152 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
156 while (p2 & (p2 - 1))
159 if (size <= 4 * SPA_MINBLOCKSIZE) {
160 align = SPA_MINBLOCKSIZE;
161 } else if (P2PHASE(size, PAGESIZE) == 0) {
163 } else if (P2PHASE(size, p2 >> 2) == 0) {
169 int flags = zio_bulk_flags;
172 * The smallest buffers (512b) are heavily used and
173 * experience a lot of churn. The slabs allocated
174 * for them are also relatively small (32K). Thus
175 * in over to avoid expensive calls to vmalloc() we
176 * make an exception to the usual slab allocation
177 * policy and force these buffers to be kmem backed.
179 if (size == (1 << SPA_MINBLOCKSHIFT))
182 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
183 zio_buf_cache[c] = kmem_cache_create(name, size,
184 align, NULL, NULL, NULL, NULL, NULL, flags);
186 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
187 zio_data_buf_cache[c] = kmem_cache_create(name, size,
188 align, NULL, NULL, NULL, NULL,
189 data_alloc_arena, flags);
194 ASSERT(zio_buf_cache[c] != NULL);
195 if (zio_buf_cache[c - 1] == NULL)
196 zio_buf_cache[c - 1] = zio_buf_cache[c];
198 ASSERT(zio_data_buf_cache[c] != NULL);
199 if (zio_data_buf_cache[c - 1] == NULL)
200 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
204 * The zio write taskqs have 1 thread per cpu, allow 1/2 of the taskqs
205 * to fail 3 times per txg or 8 failures, whichever is greater.
207 zfs_mg_alloc_failures = MAX((3 * max_ncpus / 2), 8);
216 kmem_cache_t *last_cache = NULL;
217 kmem_cache_t *last_data_cache = NULL;
219 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
220 if (zio_buf_cache[c] != last_cache) {
221 last_cache = zio_buf_cache[c];
222 kmem_cache_destroy(zio_buf_cache[c]);
224 zio_buf_cache[c] = NULL;
226 if (zio_data_buf_cache[c] != last_data_cache) {
227 last_data_cache = zio_data_buf_cache[c];
228 kmem_cache_destroy(zio_data_buf_cache[c]);
230 zio_data_buf_cache[c] = NULL;
233 kmem_cache_destroy(zio_link_cache);
234 kmem_cache_destroy(zio_cache);
240 * ==========================================================================
241 * Allocate and free I/O buffers
242 * ==========================================================================
246 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
247 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
248 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
249 * excess / transient data in-core during a crashdump.
252 zio_buf_alloc(size_t size)
254 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
256 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
258 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
262 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
263 * crashdump if the kernel panics. This exists so that we will limit the amount
264 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
265 * of kernel heap dumped to disk when the kernel panics)
268 zio_data_buf_alloc(size_t size)
270 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
272 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
274 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
278 zio_buf_free(void *buf, size_t size)
280 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
282 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
284 kmem_cache_free(zio_buf_cache[c], buf);
288 zio_data_buf_free(void *buf, size_t size)
290 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
292 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
294 kmem_cache_free(zio_data_buf_cache[c], buf);
298 * ==========================================================================
299 * Push and pop I/O transform buffers
300 * ==========================================================================
303 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
304 zio_transform_func_t *transform)
306 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_PUSHPAGE);
308 zt->zt_orig_data = zio->io_data;
309 zt->zt_orig_size = zio->io_size;
310 zt->zt_bufsize = bufsize;
311 zt->zt_transform = transform;
313 zt->zt_next = zio->io_transform_stack;
314 zio->io_transform_stack = zt;
321 zio_pop_transforms(zio_t *zio)
325 while ((zt = zio->io_transform_stack) != NULL) {
326 if (zt->zt_transform != NULL)
327 zt->zt_transform(zio,
328 zt->zt_orig_data, zt->zt_orig_size);
330 if (zt->zt_bufsize != 0)
331 zio_buf_free(zio->io_data, zt->zt_bufsize);
333 zio->io_data = zt->zt_orig_data;
334 zio->io_size = zt->zt_orig_size;
335 zio->io_transform_stack = zt->zt_next;
337 kmem_free(zt, sizeof (zio_transform_t));
342 * ==========================================================================
343 * I/O transform callbacks for subblocks and decompression
344 * ==========================================================================
347 zio_subblock(zio_t *zio, void *data, uint64_t size)
349 ASSERT(zio->io_size > size);
351 if (zio->io_type == ZIO_TYPE_READ)
352 bcopy(zio->io_data, data, size);
356 zio_decompress(zio_t *zio, void *data, uint64_t size)
358 if (zio->io_error == 0 &&
359 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
360 zio->io_data, data, zio->io_size, size) != 0)
365 * ==========================================================================
366 * I/O parent/child relationships and pipeline interlocks
367 * ==========================================================================
370 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
371 * continue calling these functions until they return NULL.
372 * Otherwise, the next caller will pick up the list walk in
373 * some indeterminate state. (Otherwise every caller would
374 * have to pass in a cookie to keep the state represented by
375 * io_walk_link, which gets annoying.)
378 zio_walk_parents(zio_t *cio)
380 zio_link_t *zl = cio->io_walk_link;
381 list_t *pl = &cio->io_parent_list;
383 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
384 cio->io_walk_link = zl;
389 ASSERT(zl->zl_child == cio);
390 return (zl->zl_parent);
394 zio_walk_children(zio_t *pio)
396 zio_link_t *zl = pio->io_walk_link;
397 list_t *cl = &pio->io_child_list;
399 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
400 pio->io_walk_link = zl;
405 ASSERT(zl->zl_parent == pio);
406 return (zl->zl_child);
410 zio_unique_parent(zio_t *cio)
412 zio_t *pio = zio_walk_parents(cio);
414 VERIFY(zio_walk_parents(cio) == NULL);
419 zio_add_child(zio_t *pio, zio_t *cio)
421 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_PUSHPAGE);
425 * Logical I/Os can have logical, gang, or vdev children.
426 * Gang I/Os can have gang or vdev children.
427 * Vdev I/Os can only have vdev children.
428 * The following ASSERT captures all of these constraints.
430 ASSERT(cio->io_child_type <= pio->io_child_type);
435 mutex_enter(&cio->io_lock);
436 mutex_enter(&pio->io_lock);
438 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
440 for (w = 0; w < ZIO_WAIT_TYPES; w++)
441 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
443 list_insert_head(&pio->io_child_list, zl);
444 list_insert_head(&cio->io_parent_list, zl);
446 pio->io_child_count++;
447 cio->io_parent_count++;
449 mutex_exit(&pio->io_lock);
450 mutex_exit(&cio->io_lock);
454 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
456 ASSERT(zl->zl_parent == pio);
457 ASSERT(zl->zl_child == cio);
459 mutex_enter(&cio->io_lock);
460 mutex_enter(&pio->io_lock);
462 list_remove(&pio->io_child_list, zl);
463 list_remove(&cio->io_parent_list, zl);
465 pio->io_child_count--;
466 cio->io_parent_count--;
468 mutex_exit(&pio->io_lock);
469 mutex_exit(&cio->io_lock);
471 kmem_cache_free(zio_link_cache, zl);
475 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
477 uint64_t *countp = &zio->io_children[child][wait];
478 boolean_t waiting = B_FALSE;
480 mutex_enter(&zio->io_lock);
481 ASSERT(zio->io_stall == NULL);
484 zio->io_stall = countp;
487 mutex_exit(&zio->io_lock);
492 __attribute__((always_inline))
494 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
496 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
497 int *errorp = &pio->io_child_error[zio->io_child_type];
499 mutex_enter(&pio->io_lock);
500 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
501 *errorp = zio_worst_error(*errorp, zio->io_error);
502 pio->io_reexecute |= zio->io_reexecute;
503 ASSERT3U(*countp, >, 0);
504 if (--*countp == 0 && pio->io_stall == countp) {
505 pio->io_stall = NULL;
506 mutex_exit(&pio->io_lock);
509 mutex_exit(&pio->io_lock);
514 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
516 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
517 zio->io_error = zio->io_child_error[c];
521 * ==========================================================================
522 * Create the various types of I/O (read, write, free, etc)
523 * ==========================================================================
526 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
527 void *data, uint64_t size, zio_done_func_t *done, void *private,
528 zio_type_t type, int priority, enum zio_flag flags,
529 vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
530 enum zio_stage stage, enum zio_stage pipeline)
534 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
535 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
536 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
538 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
539 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
540 ASSERT(vd || stage == ZIO_STAGE_OPEN);
542 zio = kmem_cache_alloc(zio_cache, KM_PUSHPAGE);
545 zio->io_child_type = ZIO_CHILD_VDEV;
546 else if (flags & ZIO_FLAG_GANG_CHILD)
547 zio->io_child_type = ZIO_CHILD_GANG;
548 else if (flags & ZIO_FLAG_DDT_CHILD)
549 zio->io_child_type = ZIO_CHILD_DDT;
551 zio->io_child_type = ZIO_CHILD_LOGICAL;
554 zio->io_bp = (blkptr_t *)bp;
555 zio->io_bp_copy = *bp;
556 zio->io_bp_orig = *bp;
557 if (type != ZIO_TYPE_WRITE ||
558 zio->io_child_type == ZIO_CHILD_DDT)
559 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
560 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
561 zio->io_logical = zio;
562 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
563 pipeline |= ZIO_GANG_STAGES;
565 zio->io_logical = NULL;
567 bzero(&zio->io_bp_copy, sizeof (blkptr_t));
568 bzero(&zio->io_bp_orig, sizeof (blkptr_t));
573 zio->io_ready = NULL;
575 zio->io_private = private;
576 zio->io_prev_space_delta = 0;
578 zio->io_priority = priority;
581 zio->io_vsd_ops = NULL;
582 zio->io_offset = offset;
583 zio->io_deadline = 0;
584 zio->io_orig_data = zio->io_data = data;
585 zio->io_orig_size = zio->io_size = size;
586 zio->io_orig_flags = zio->io_flags = flags;
587 zio->io_orig_stage = zio->io_stage = stage;
588 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
589 bzero(&zio->io_prop, sizeof (zio_prop_t));
591 zio->io_reexecute = 0;
592 zio->io_bp_override = NULL;
593 zio->io_walk_link = NULL;
594 zio->io_transform_stack = NULL;
597 zio->io_child_count = 0;
598 zio->io_parent_count = 0;
599 zio->io_stall = NULL;
600 zio->io_gang_leader = NULL;
601 zio->io_gang_tree = NULL;
602 zio->io_executor = NULL;
603 zio->io_waiter = NULL;
604 zio->io_cksum_report = NULL;
606 bzero(zio->io_child_error, sizeof (int) * ZIO_CHILD_TYPES);
607 bzero(zio->io_children,
608 sizeof (uint64_t) * ZIO_CHILD_TYPES * ZIO_WAIT_TYPES);
609 bzero(&zio->io_bookmark, sizeof (zbookmark_t));
611 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
612 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
615 zio->io_bookmark = *zb;
618 if (zio->io_logical == NULL)
619 zio->io_logical = pio->io_logical;
620 if (zio->io_child_type == ZIO_CHILD_GANG)
621 zio->io_gang_leader = pio->io_gang_leader;
622 zio_add_child(pio, zio);
625 taskq_init_ent(&zio->io_tqent);
631 zio_destroy(zio_t *zio)
633 kmem_cache_free(zio_cache, zio);
637 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
638 void *private, enum zio_flag flags)
642 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
643 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
644 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
650 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
652 return (zio_null(NULL, spa, NULL, done, private, flags));
656 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
657 void *data, uint64_t size, zio_done_func_t *done, void *private,
658 int priority, enum zio_flag flags, const zbookmark_t *zb)
662 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
663 data, size, done, private,
664 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
665 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
666 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
672 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
673 void *data, uint64_t size, const zio_prop_t *zp,
674 zio_done_func_t *ready, zio_done_func_t *done, void *private,
675 int priority, enum zio_flag flags, const zbookmark_t *zb)
679 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
680 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
681 zp->zp_compress >= ZIO_COMPRESS_OFF &&
682 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
683 zp->zp_type < DMU_OT_NUMTYPES &&
686 zp->zp_copies <= spa_max_replication(spa) &&
688 zp->zp_dedup_verify <= 1);
690 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
691 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
692 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
693 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
695 zio->io_ready = ready;
702 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
703 uint64_t size, zio_done_func_t *done, void *private, int priority,
704 enum zio_flag flags, zbookmark_t *zb)
708 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
709 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
710 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
716 zio_write_override(zio_t *zio, blkptr_t *bp, int copies)
718 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
719 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
720 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
721 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
723 zio->io_prop.zp_copies = copies;
724 zio->io_bp_override = bp;
728 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
730 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
734 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
739 dprintf_bp(bp, "freeing in txg %llu, pass %u",
740 (longlong_t)txg, spa->spa_sync_pass);
742 ASSERT(!BP_IS_HOLE(bp));
743 ASSERT(spa_syncing_txg(spa) == txg);
744 ASSERT(spa_sync_pass(spa) <= SYNC_PASS_DEFERRED_FREE);
746 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
747 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
748 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
754 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
755 zio_done_func_t *done, void *private, enum zio_flag flags)
760 * A claim is an allocation of a specific block. Claims are needed
761 * to support immediate writes in the intent log. The issue is that
762 * immediate writes contain committed data, but in a txg that was
763 * *not* committed. Upon opening the pool after an unclean shutdown,
764 * the intent log claims all blocks that contain immediate write data
765 * so that the SPA knows they're in use.
767 * All claims *must* be resolved in the first txg -- before the SPA
768 * starts allocating blocks -- so that nothing is allocated twice.
769 * If txg == 0 we just verify that the block is claimable.
771 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
772 ASSERT(txg == spa_first_txg(spa) || txg == 0);
773 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
775 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
776 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
777 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
783 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
784 zio_done_func_t *done, void *private, int priority, enum zio_flag flags)
789 if (vd->vdev_children == 0) {
790 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
791 ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
792 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
796 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
798 for (c = 0; c < vd->vdev_children; c++)
799 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
800 done, private, priority, flags));
807 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
808 void *data, int checksum, zio_done_func_t *done, void *private,
809 int priority, enum zio_flag flags, boolean_t labels)
813 ASSERT(vd->vdev_children == 0);
814 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
815 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
816 ASSERT3U(offset + size, <=, vd->vdev_psize);
818 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
819 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
820 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
822 zio->io_prop.zp_checksum = checksum;
828 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
829 void *data, int checksum, zio_done_func_t *done, void *private,
830 int priority, enum zio_flag flags, boolean_t labels)
834 ASSERT(vd->vdev_children == 0);
835 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
836 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
837 ASSERT3U(offset + size, <=, vd->vdev_psize);
839 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
840 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
841 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
843 zio->io_prop.zp_checksum = checksum;
845 if (zio_checksum_table[checksum].ci_eck) {
847 * zec checksums are necessarily destructive -- they modify
848 * the end of the write buffer to hold the verifier/checksum.
849 * Therefore, we must make a local copy in case the data is
850 * being written to multiple places in parallel.
852 void *wbuf = zio_buf_alloc(size);
853 bcopy(data, wbuf, size);
854 zio_push_transform(zio, wbuf, size, size, NULL);
861 * Create a child I/O to do some work for us.
864 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
865 void *data, uint64_t size, int type, int priority, enum zio_flag flags,
866 zio_done_func_t *done, void *private)
868 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
871 ASSERT(vd->vdev_parent ==
872 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
874 if (type == ZIO_TYPE_READ && bp != NULL) {
876 * If we have the bp, then the child should perform the
877 * checksum and the parent need not. This pushes error
878 * detection as close to the leaves as possible and
879 * eliminates redundant checksums in the interior nodes.
881 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
882 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
885 if (vd->vdev_children == 0)
886 offset += VDEV_LABEL_START_SIZE;
888 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
891 * If we've decided to do a repair, the write is not speculative --
892 * even if the original read was.
894 if (flags & ZIO_FLAG_IO_REPAIR)
895 flags &= ~ZIO_FLAG_SPECULATIVE;
897 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
898 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
899 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
905 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
906 int type, int priority, enum zio_flag flags,
907 zio_done_func_t *done, void *private)
911 ASSERT(vd->vdev_ops->vdev_op_leaf);
913 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
914 data, size, done, private, type, priority,
915 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
917 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
923 zio_flush(zio_t *zio, vdev_t *vd)
925 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
926 NULL, NULL, ZIO_PRIORITY_NOW,
927 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
931 zio_shrink(zio_t *zio, uint64_t size)
933 ASSERT(zio->io_executor == NULL);
934 ASSERT(zio->io_orig_size == zio->io_size);
935 ASSERT(size <= zio->io_size);
938 * We don't shrink for raidz because of problems with the
939 * reconstruction when reading back less than the block size.
940 * Note, BP_IS_RAIDZ() assumes no compression.
942 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
943 if (!BP_IS_RAIDZ(zio->io_bp))
944 zio->io_orig_size = zio->io_size = size;
948 * ==========================================================================
949 * Prepare to read and write logical blocks
950 * ==========================================================================
954 zio_read_bp_init(zio_t *zio)
956 blkptr_t *bp = zio->io_bp;
958 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
959 zio->io_child_type == ZIO_CHILD_LOGICAL &&
960 !(zio->io_flags & ZIO_FLAG_RAW)) {
961 uint64_t psize = BP_GET_PSIZE(bp);
962 void *cbuf = zio_buf_alloc(psize);
964 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
967 if (!dmu_ot[BP_GET_TYPE(bp)].ot_metadata && BP_GET_LEVEL(bp) == 0)
968 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
970 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
971 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
973 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
974 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
976 return (ZIO_PIPELINE_CONTINUE);
980 zio_write_bp_init(zio_t *zio)
982 spa_t *spa = zio->io_spa;
983 zio_prop_t *zp = &zio->io_prop;
984 enum zio_compress compress = zp->zp_compress;
985 blkptr_t *bp = zio->io_bp;
986 uint64_t lsize = zio->io_size;
987 uint64_t psize = lsize;
991 * If our children haven't all reached the ready stage,
992 * wait for them and then repeat this pipeline stage.
994 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
995 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
996 return (ZIO_PIPELINE_STOP);
998 if (!IO_IS_ALLOCATING(zio))
999 return (ZIO_PIPELINE_CONTINUE);
1001 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1003 if (zio->io_bp_override) {
1004 ASSERT(bp->blk_birth != zio->io_txg);
1005 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1007 *bp = *zio->io_bp_override;
1008 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1010 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1011 return (ZIO_PIPELINE_CONTINUE);
1013 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1014 zp->zp_dedup_verify);
1016 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1017 BP_SET_DEDUP(bp, 1);
1018 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1019 return (ZIO_PIPELINE_CONTINUE);
1021 zio->io_bp_override = NULL;
1025 if (bp->blk_birth == zio->io_txg) {
1027 * We're rewriting an existing block, which means we're
1028 * working on behalf of spa_sync(). For spa_sync() to
1029 * converge, it must eventually be the case that we don't
1030 * have to allocate new blocks. But compression changes
1031 * the blocksize, which forces a reallocate, and makes
1032 * convergence take longer. Therefore, after the first
1033 * few passes, stop compressing to ensure convergence.
1035 pass = spa_sync_pass(spa);
1037 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1038 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1039 ASSERT(!BP_GET_DEDUP(bp));
1041 if (pass > SYNC_PASS_DONT_COMPRESS)
1042 compress = ZIO_COMPRESS_OFF;
1044 /* Make sure someone doesn't change their mind on overwrites */
1045 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
1046 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1049 if (compress != ZIO_COMPRESS_OFF) {
1050 void *cbuf = zio_buf_alloc(lsize);
1051 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1052 if (psize == 0 || psize == lsize) {
1053 compress = ZIO_COMPRESS_OFF;
1054 zio_buf_free(cbuf, lsize);
1056 ASSERT(psize < lsize);
1057 zio_push_transform(zio, cbuf, psize, lsize, NULL);
1062 * The final pass of spa_sync() must be all rewrites, but the first
1063 * few passes offer a trade-off: allocating blocks defers convergence,
1064 * but newly allocated blocks are sequential, so they can be written
1065 * to disk faster. Therefore, we allow the first few passes of
1066 * spa_sync() to allocate new blocks, but force rewrites after that.
1067 * There should only be a handful of blocks after pass 1 in any case.
1069 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
1070 pass > SYNC_PASS_REWRITE) {
1071 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1073 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1074 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1077 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1081 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1083 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1084 BP_SET_LSIZE(bp, lsize);
1085 BP_SET_PSIZE(bp, psize);
1086 BP_SET_COMPRESS(bp, compress);
1087 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1088 BP_SET_TYPE(bp, zp->zp_type);
1089 BP_SET_LEVEL(bp, zp->zp_level);
1090 BP_SET_DEDUP(bp, zp->zp_dedup);
1091 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1093 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1094 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1095 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1099 return (ZIO_PIPELINE_CONTINUE);
1103 zio_free_bp_init(zio_t *zio)
1105 blkptr_t *bp = zio->io_bp;
1107 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1108 if (BP_GET_DEDUP(bp))
1109 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1112 return (ZIO_PIPELINE_CONTINUE);
1116 * ==========================================================================
1117 * Execute the I/O pipeline
1118 * ==========================================================================
1122 zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q, boolean_t cutinline)
1124 spa_t *spa = zio->io_spa;
1125 zio_type_t t = zio->io_type;
1126 int flags = (cutinline ? TQ_FRONT : 0);
1129 * If we're a config writer or a probe, the normal issue and
1130 * interrupt threads may all be blocked waiting for the config lock.
1131 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1133 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1137 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1139 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1143 * If this is a high priority I/O, then use the high priority taskq.
1145 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1146 spa->spa_zio_taskq[t][q + 1] != NULL)
1149 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1152 * NB: We are assuming that the zio can only be dispatched
1153 * to a single taskq at a time. It would be a grievous error
1154 * to dispatch the zio to another taskq at the same time.
1156 ASSERT(taskq_empty_ent(&zio->io_tqent));
1157 taskq_dispatch_ent(spa->spa_zio_taskq[t][q],
1158 (task_func_t *)zio_execute, zio, flags, &zio->io_tqent);
1162 zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
1164 kthread_t *executor = zio->io_executor;
1165 spa_t *spa = zio->io_spa;
1168 for (t = 0; t < ZIO_TYPES; t++)
1169 if (taskq_member(spa->spa_zio_taskq[t][q], executor))
1176 zio_issue_async(zio_t *zio)
1178 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1180 return (ZIO_PIPELINE_STOP);
1184 zio_interrupt(zio_t *zio)
1186 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1190 * Execute the I/O pipeline until one of the following occurs:
1191 * (1) the I/O completes; (2) the pipeline stalls waiting for
1192 * dependent child I/Os; (3) the I/O issues, so we're waiting
1193 * for an I/O completion interrupt; (4) the I/O is delegated by
1194 * vdev-level caching or aggregation; (5) the I/O is deferred
1195 * due to vdev-level queueing; (6) the I/O is handed off to
1196 * another thread. In all cases, the pipeline stops whenever
1197 * there's no CPU work; it never burns a thread in cv_wait().
1199 * There's no locking on io_stage because there's no legitimate way
1200 * for multiple threads to be attempting to process the same I/O.
1202 static zio_pipe_stage_t *zio_pipeline[];
1205 * zio_execute() is a wrapper around the static function
1206 * __zio_execute() so that we can force __zio_execute() to be
1207 * inlined. This reduces stack overhead which is important
1208 * because __zio_execute() is called recursively in several zio
1209 * code paths. zio_execute() itself cannot be inlined because
1210 * it is externally visible.
1213 zio_execute(zio_t *zio)
1218 __attribute__((always_inline))
1220 __zio_execute(zio_t *zio)
1222 zio->io_executor = curthread;
1224 while (zio->io_stage < ZIO_STAGE_DONE) {
1225 enum zio_stage pipeline = zio->io_pipeline;
1226 enum zio_stage stage = zio->io_stage;
1231 ASSERT(!MUTEX_HELD(&zio->io_lock));
1232 ASSERT(ISP2(stage));
1233 ASSERT(zio->io_stall == NULL);
1237 } while ((stage & pipeline) == 0);
1239 ASSERT(stage <= ZIO_STAGE_DONE);
1241 dsl = spa_get_dsl(zio->io_spa);
1242 cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1243 zio_requeue_io_start_cut_in_line : B_FALSE;
1246 * If we are in interrupt context and this pipeline stage
1247 * will grab a config lock that is held across I/O,
1248 * or may wait for an I/O that needs an interrupt thread
1249 * to complete, issue async to avoid deadlock.
1251 * If we are in the txg_sync_thread or being called
1252 * during pool init issue async to minimize stack depth.
1253 * Both of these call paths may be recursively called.
1255 * For VDEV_IO_START, we cut in line so that the io will
1256 * be sent to disk promptly.
1258 if (((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1259 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) ||
1260 (dsl != NULL && dsl_pool_sync_context(dsl))) {
1261 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1265 zio->io_stage = stage;
1266 rv = zio_pipeline[highbit(stage) - 1](zio);
1268 if (rv == ZIO_PIPELINE_STOP)
1271 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1277 * ==========================================================================
1278 * Initiate I/O, either sync or async
1279 * ==========================================================================
1282 zio_wait(zio_t *zio)
1286 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1287 ASSERT(zio->io_executor == NULL);
1289 zio->io_waiter = curthread;
1293 mutex_enter(&zio->io_lock);
1294 while (zio->io_executor != NULL)
1295 cv_wait(&zio->io_cv, &zio->io_lock);
1296 mutex_exit(&zio->io_lock);
1298 error = zio->io_error;
1305 zio_nowait(zio_t *zio)
1307 ASSERT(zio->io_executor == NULL);
1309 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1310 zio_unique_parent(zio) == NULL) {
1312 * This is a logical async I/O with no parent to wait for it.
1313 * We add it to the spa_async_root_zio "Godfather" I/O which
1314 * will ensure they complete prior to unloading the pool.
1316 spa_t *spa = zio->io_spa;
1318 zio_add_child(spa->spa_async_zio_root, zio);
1325 * ==========================================================================
1326 * Reexecute or suspend/resume failed I/O
1327 * ==========================================================================
1331 zio_reexecute(zio_t *pio)
1333 zio_t *cio, *cio_next;
1336 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1337 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1338 ASSERT(pio->io_gang_leader == NULL);
1339 ASSERT(pio->io_gang_tree == NULL);
1341 pio->io_flags = pio->io_orig_flags;
1342 pio->io_stage = pio->io_orig_stage;
1343 pio->io_pipeline = pio->io_orig_pipeline;
1344 pio->io_reexecute = 0;
1346 for (w = 0; w < ZIO_WAIT_TYPES; w++)
1347 pio->io_state[w] = 0;
1348 for (c = 0; c < ZIO_CHILD_TYPES; c++)
1349 pio->io_child_error[c] = 0;
1351 if (IO_IS_ALLOCATING(pio))
1352 BP_ZERO(pio->io_bp);
1355 * As we reexecute pio's children, new children could be created.
1356 * New children go to the head of pio's io_child_list, however,
1357 * so we will (correctly) not reexecute them. The key is that
1358 * the remainder of pio's io_child_list, from 'cio_next' onward,
1359 * cannot be affected by any side effects of reexecuting 'cio'.
1361 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1362 cio_next = zio_walk_children(pio);
1363 mutex_enter(&pio->io_lock);
1364 for (w = 0; w < ZIO_WAIT_TYPES; w++)
1365 pio->io_children[cio->io_child_type][w]++;
1366 mutex_exit(&pio->io_lock);
1371 * Now that all children have been reexecuted, execute the parent.
1372 * We don't reexecute "The Godfather" I/O here as it's the
1373 * responsibility of the caller to wait on him.
1375 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1380 zio_suspend(spa_t *spa, zio_t *zio)
1382 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1383 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1384 "failure and the failure mode property for this pool "
1385 "is set to panic.", spa_name(spa));
1387 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1389 mutex_enter(&spa->spa_suspend_lock);
1391 if (spa->spa_suspend_zio_root == NULL)
1392 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1393 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1394 ZIO_FLAG_GODFATHER);
1396 spa->spa_suspended = B_TRUE;
1399 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1400 ASSERT(zio != spa->spa_suspend_zio_root);
1401 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1402 ASSERT(zio_unique_parent(zio) == NULL);
1403 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1404 zio_add_child(spa->spa_suspend_zio_root, zio);
1407 mutex_exit(&spa->spa_suspend_lock);
1411 zio_resume(spa_t *spa)
1416 * Reexecute all previously suspended i/o.
1418 mutex_enter(&spa->spa_suspend_lock);
1419 spa->spa_suspended = B_FALSE;
1420 cv_broadcast(&spa->spa_suspend_cv);
1421 pio = spa->spa_suspend_zio_root;
1422 spa->spa_suspend_zio_root = NULL;
1423 mutex_exit(&spa->spa_suspend_lock);
1429 return (zio_wait(pio));
1433 zio_resume_wait(spa_t *spa)
1435 mutex_enter(&spa->spa_suspend_lock);
1436 while (spa_suspended(spa))
1437 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1438 mutex_exit(&spa->spa_suspend_lock);
1442 * ==========================================================================
1445 * A gang block is a collection of small blocks that looks to the DMU
1446 * like one large block. When zio_dva_allocate() cannot find a block
1447 * of the requested size, due to either severe fragmentation or the pool
1448 * being nearly full, it calls zio_write_gang_block() to construct the
1449 * block from smaller fragments.
1451 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1452 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1453 * an indirect block: it's an array of block pointers. It consumes
1454 * only one sector and hence is allocatable regardless of fragmentation.
1455 * The gang header's bps point to its gang members, which hold the data.
1457 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1458 * as the verifier to ensure uniqueness of the SHA256 checksum.
1459 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1460 * not the gang header. This ensures that data block signatures (needed for
1461 * deduplication) are independent of how the block is physically stored.
1463 * Gang blocks can be nested: a gang member may itself be a gang block.
1464 * Thus every gang block is a tree in which root and all interior nodes are
1465 * gang headers, and the leaves are normal blocks that contain user data.
1466 * The root of the gang tree is called the gang leader.
1468 * To perform any operation (read, rewrite, free, claim) on a gang block,
1469 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1470 * in the io_gang_tree field of the original logical i/o by recursively
1471 * reading the gang leader and all gang headers below it. This yields
1472 * an in-core tree containing the contents of every gang header and the
1473 * bps for every constituent of the gang block.
1475 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1476 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1477 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1478 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1479 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1480 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1481 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1482 * of the gang header plus zio_checksum_compute() of the data to update the
1483 * gang header's blk_cksum as described above.
1485 * The two-phase assemble/issue model solves the problem of partial failure --
1486 * what if you'd freed part of a gang block but then couldn't read the
1487 * gang header for another part? Assembling the entire gang tree first
1488 * ensures that all the necessary gang header I/O has succeeded before
1489 * starting the actual work of free, claim, or write. Once the gang tree
1490 * is assembled, free and claim are in-memory operations that cannot fail.
1492 * In the event that a gang write fails, zio_dva_unallocate() walks the
1493 * gang tree to immediately free (i.e. insert back into the space map)
1494 * everything we've allocated. This ensures that we don't get ENOSPC
1495 * errors during repeated suspend/resume cycles due to a flaky device.
1497 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1498 * the gang tree, we won't modify the block, so we can safely defer the free
1499 * (knowing that the block is still intact). If we *can* assemble the gang
1500 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1501 * each constituent bp and we can allocate a new block on the next sync pass.
1503 * In all cases, the gang tree allows complete recovery from partial failure.
1504 * ==========================================================================
1508 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1513 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1514 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1515 &pio->io_bookmark));
1519 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1524 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1525 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1526 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1528 * As we rewrite each gang header, the pipeline will compute
1529 * a new gang block header checksum for it; but no one will
1530 * compute a new data checksum, so we do that here. The one
1531 * exception is the gang leader: the pipeline already computed
1532 * its data checksum because that stage precedes gang assembly.
1533 * (Presently, nothing actually uses interior data checksums;
1534 * this is just good hygiene.)
1536 if (gn != pio->io_gang_leader->io_gang_tree) {
1537 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1538 data, BP_GET_PSIZE(bp));
1541 * If we are here to damage data for testing purposes,
1542 * leave the GBH alone so that we can detect the damage.
1544 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1545 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1547 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1548 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1549 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1557 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1559 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1560 ZIO_GANG_CHILD_FLAGS(pio)));
1565 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1567 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1568 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1571 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1580 static void zio_gang_tree_assemble_done(zio_t *zio);
1582 static zio_gang_node_t *
1583 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1585 zio_gang_node_t *gn;
1587 ASSERT(*gnpp == NULL);
1589 gn = kmem_zalloc(sizeof (*gn), KM_PUSHPAGE);
1590 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1597 zio_gang_node_free(zio_gang_node_t **gnpp)
1599 zio_gang_node_t *gn = *gnpp;
1602 for (g = 0; g < SPA_GBH_NBLKPTRS; g++)
1603 ASSERT(gn->gn_child[g] == NULL);
1605 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1606 kmem_free(gn, sizeof (*gn));
1611 zio_gang_tree_free(zio_gang_node_t **gnpp)
1613 zio_gang_node_t *gn = *gnpp;
1619 for (g = 0; g < SPA_GBH_NBLKPTRS; g++)
1620 zio_gang_tree_free(&gn->gn_child[g]);
1622 zio_gang_node_free(gnpp);
1626 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1628 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1630 ASSERT(gio->io_gang_leader == gio);
1631 ASSERT(BP_IS_GANG(bp));
1633 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1634 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1635 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1639 zio_gang_tree_assemble_done(zio_t *zio)
1641 zio_t *gio = zio->io_gang_leader;
1642 zio_gang_node_t *gn = zio->io_private;
1643 blkptr_t *bp = zio->io_bp;
1646 ASSERT(gio == zio_unique_parent(zio));
1647 ASSERT(zio->io_child_count == 0);
1652 if (BP_SHOULD_BYTESWAP(bp))
1653 byteswap_uint64_array(zio->io_data, zio->io_size);
1655 ASSERT(zio->io_data == gn->gn_gbh);
1656 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1657 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1659 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1660 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1661 if (!BP_IS_GANG(gbp))
1663 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1668 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1670 zio_t *gio = pio->io_gang_leader;
1674 ASSERT(BP_IS_GANG(bp) == !!gn);
1675 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1676 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1679 * If you're a gang header, your data is in gn->gn_gbh.
1680 * If you're a gang member, your data is in 'data' and gn == NULL.
1682 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1685 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1687 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1688 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1689 if (BP_IS_HOLE(gbp))
1691 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1692 data = (char *)data + BP_GET_PSIZE(gbp);
1696 if (gn == gio->io_gang_tree)
1697 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1704 zio_gang_assemble(zio_t *zio)
1706 blkptr_t *bp = zio->io_bp;
1708 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1709 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1711 zio->io_gang_leader = zio;
1713 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1715 return (ZIO_PIPELINE_CONTINUE);
1719 zio_gang_issue(zio_t *zio)
1721 blkptr_t *bp = zio->io_bp;
1723 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1724 return (ZIO_PIPELINE_STOP);
1726 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1727 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1729 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1730 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1732 zio_gang_tree_free(&zio->io_gang_tree);
1734 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1736 return (ZIO_PIPELINE_CONTINUE);
1740 zio_write_gang_member_ready(zio_t *zio)
1742 zio_t *pio = zio_unique_parent(zio);
1743 ASSERTV(zio_t *gio = zio->io_gang_leader;)
1744 dva_t *cdva = zio->io_bp->blk_dva;
1745 dva_t *pdva = pio->io_bp->blk_dva;
1749 if (BP_IS_HOLE(zio->io_bp))
1752 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1754 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1755 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1756 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1757 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1758 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1760 mutex_enter(&pio->io_lock);
1761 for (d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1762 ASSERT(DVA_GET_GANG(&pdva[d]));
1763 asize = DVA_GET_ASIZE(&pdva[d]);
1764 asize += DVA_GET_ASIZE(&cdva[d]);
1765 DVA_SET_ASIZE(&pdva[d], asize);
1767 mutex_exit(&pio->io_lock);
1771 zio_write_gang_block(zio_t *pio)
1773 spa_t *spa = pio->io_spa;
1774 blkptr_t *bp = pio->io_bp;
1775 zio_t *gio = pio->io_gang_leader;
1777 zio_gang_node_t *gn, **gnpp;
1778 zio_gbh_phys_t *gbh;
1779 uint64_t txg = pio->io_txg;
1780 uint64_t resid = pio->io_size;
1782 int copies = gio->io_prop.zp_copies;
1783 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1787 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1788 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1789 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1791 pio->io_error = error;
1792 return (ZIO_PIPELINE_CONTINUE);
1796 gnpp = &gio->io_gang_tree;
1798 gnpp = pio->io_private;
1799 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1802 gn = zio_gang_node_alloc(gnpp);
1804 bzero(gbh, SPA_GANGBLOCKSIZE);
1807 * Create the gang header.
1809 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1810 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1813 * Create and nowait the gang children.
1815 for (g = 0; resid != 0; resid -= lsize, g++) {
1816 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1818 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1820 zp.zp_checksum = gio->io_prop.zp_checksum;
1821 zp.zp_compress = ZIO_COMPRESS_OFF;
1822 zp.zp_type = DMU_OT_NONE;
1824 zp.zp_copies = gio->io_prop.zp_copies;
1826 zp.zp_dedup_verify = 0;
1828 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1829 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1830 zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1831 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1832 &pio->io_bookmark));
1836 * Set pio's pipeline to just wait for zio to finish.
1838 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1842 return (ZIO_PIPELINE_CONTINUE);
1846 * ==========================================================================
1848 * ==========================================================================
1851 zio_ddt_child_read_done(zio_t *zio)
1853 blkptr_t *bp = zio->io_bp;
1854 ddt_entry_t *dde = zio->io_private;
1856 zio_t *pio = zio_unique_parent(zio);
1858 mutex_enter(&pio->io_lock);
1859 ddp = ddt_phys_select(dde, bp);
1860 if (zio->io_error == 0)
1861 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1862 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1863 dde->dde_repair_data = zio->io_data;
1865 zio_buf_free(zio->io_data, zio->io_size);
1866 mutex_exit(&pio->io_lock);
1870 zio_ddt_read_start(zio_t *zio)
1872 blkptr_t *bp = zio->io_bp;
1875 ASSERT(BP_GET_DEDUP(bp));
1876 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1877 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1879 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1880 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1881 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1882 ddt_phys_t *ddp = dde->dde_phys;
1883 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1886 ASSERT(zio->io_vsd == NULL);
1889 if (ddp_self == NULL)
1890 return (ZIO_PIPELINE_CONTINUE);
1892 for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1893 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1895 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1897 zio_nowait(zio_read(zio, zio->io_spa, &blk,
1898 zio_buf_alloc(zio->io_size), zio->io_size,
1899 zio_ddt_child_read_done, dde, zio->io_priority,
1900 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1901 &zio->io_bookmark));
1903 return (ZIO_PIPELINE_CONTINUE);
1906 zio_nowait(zio_read(zio, zio->io_spa, bp,
1907 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1908 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1910 return (ZIO_PIPELINE_CONTINUE);
1914 zio_ddt_read_done(zio_t *zio)
1916 blkptr_t *bp = zio->io_bp;
1918 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1919 return (ZIO_PIPELINE_STOP);
1921 ASSERT(BP_GET_DEDUP(bp));
1922 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1923 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1925 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1926 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1927 ddt_entry_t *dde = zio->io_vsd;
1929 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
1930 return (ZIO_PIPELINE_CONTINUE);
1933 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
1934 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1935 return (ZIO_PIPELINE_STOP);
1937 if (dde->dde_repair_data != NULL) {
1938 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
1939 zio->io_child_error[ZIO_CHILD_DDT] = 0;
1941 ddt_repair_done(ddt, dde);
1945 ASSERT(zio->io_vsd == NULL);
1947 return (ZIO_PIPELINE_CONTINUE);
1951 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
1953 spa_t *spa = zio->io_spa;
1957 * Note: we compare the original data, not the transformed data,
1958 * because when zio->io_bp is an override bp, we will not have
1959 * pushed the I/O transforms. That's an important optimization
1960 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
1962 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1963 zio_t *lio = dde->dde_lead_zio[p];
1966 return (lio->io_orig_size != zio->io_orig_size ||
1967 bcmp(zio->io_orig_data, lio->io_orig_data,
1968 zio->io_orig_size) != 0);
1972 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1973 ddt_phys_t *ddp = &dde->dde_phys[p];
1975 if (ddp->ddp_phys_birth != 0) {
1976 arc_buf_t *abuf = NULL;
1977 uint32_t aflags = ARC_WAIT;
1978 blkptr_t blk = *zio->io_bp;
1981 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
1985 error = arc_read_nolock(NULL, spa, &blk,
1986 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
1987 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1988 &aflags, &zio->io_bookmark);
1991 if (arc_buf_size(abuf) != zio->io_orig_size ||
1992 bcmp(abuf->b_data, zio->io_orig_data,
1993 zio->io_orig_size) != 0)
1995 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
1999 return (error != 0);
2007 zio_ddt_child_write_ready(zio_t *zio)
2009 int p = zio->io_prop.zp_copies;
2010 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2011 ddt_entry_t *dde = zio->io_private;
2012 ddt_phys_t *ddp = &dde->dde_phys[p];
2020 ASSERT(dde->dde_lead_zio[p] == zio);
2022 ddt_phys_fill(ddp, zio->io_bp);
2024 while ((pio = zio_walk_parents(zio)) != NULL)
2025 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2031 zio_ddt_child_write_done(zio_t *zio)
2033 int p = zio->io_prop.zp_copies;
2034 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2035 ddt_entry_t *dde = zio->io_private;
2036 ddt_phys_t *ddp = &dde->dde_phys[p];
2040 ASSERT(ddp->ddp_refcnt == 0);
2041 ASSERT(dde->dde_lead_zio[p] == zio);
2042 dde->dde_lead_zio[p] = NULL;
2044 if (zio->io_error == 0) {
2045 while (zio_walk_parents(zio) != NULL)
2046 ddt_phys_addref(ddp);
2048 ddt_phys_clear(ddp);
2055 zio_ddt_ditto_write_done(zio_t *zio)
2057 int p = DDT_PHYS_DITTO;
2058 blkptr_t *bp = zio->io_bp;
2059 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2060 ddt_entry_t *dde = zio->io_private;
2061 ddt_phys_t *ddp = &dde->dde_phys[p];
2062 ddt_key_t *ddk = &dde->dde_key;
2063 ASSERTV(zio_prop_t *zp = &zio->io_prop);
2067 ASSERT(ddp->ddp_refcnt == 0);
2068 ASSERT(dde->dde_lead_zio[p] == zio);
2069 dde->dde_lead_zio[p] = NULL;
2071 if (zio->io_error == 0) {
2072 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2073 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2074 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2075 if (ddp->ddp_phys_birth != 0)
2076 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2077 ddt_phys_fill(ddp, bp);
2084 zio_ddt_write(zio_t *zio)
2086 spa_t *spa = zio->io_spa;
2087 blkptr_t *bp = zio->io_bp;
2088 uint64_t txg = zio->io_txg;
2089 zio_prop_t *zp = &zio->io_prop;
2090 int p = zp->zp_copies;
2094 ddt_t *ddt = ddt_select(spa, bp);
2098 ASSERT(BP_GET_DEDUP(bp));
2099 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2100 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2103 dde = ddt_lookup(ddt, bp, B_TRUE);
2104 ddp = &dde->dde_phys[p];
2106 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2108 * If we're using a weak checksum, upgrade to a strong checksum
2109 * and try again. If we're already using a strong checksum,
2110 * we can't resolve it, so just convert to an ordinary write.
2111 * (And automatically e-mail a paper to Nature?)
2113 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2114 zp->zp_checksum = spa_dedup_checksum(spa);
2115 zio_pop_transforms(zio);
2116 zio->io_stage = ZIO_STAGE_OPEN;
2121 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2123 return (ZIO_PIPELINE_CONTINUE);
2126 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2127 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2129 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2130 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2131 zio_prop_t czp = *zp;
2133 czp.zp_copies = ditto_copies;
2136 * If we arrived here with an override bp, we won't have run
2137 * the transform stack, so we won't have the data we need to
2138 * generate a child i/o. So, toss the override bp and restart.
2139 * This is safe, because using the override bp is just an
2140 * optimization; and it's rare, so the cost doesn't matter.
2142 if (zio->io_bp_override) {
2143 zio_pop_transforms(zio);
2144 zio->io_stage = ZIO_STAGE_OPEN;
2145 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2146 zio->io_bp_override = NULL;
2149 return (ZIO_PIPELINE_CONTINUE);
2152 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2153 zio->io_orig_size, &czp, NULL,
2154 zio_ddt_ditto_write_done, dde, zio->io_priority,
2155 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2157 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2158 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2161 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2162 if (ddp->ddp_phys_birth != 0)
2163 ddt_bp_fill(ddp, bp, txg);
2164 if (dde->dde_lead_zio[p] != NULL)
2165 zio_add_child(zio, dde->dde_lead_zio[p]);
2167 ddt_phys_addref(ddp);
2168 } else if (zio->io_bp_override) {
2169 ASSERT(bp->blk_birth == txg);
2170 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2171 ddt_phys_fill(ddp, bp);
2172 ddt_phys_addref(ddp);
2174 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2175 zio->io_orig_size, zp, zio_ddt_child_write_ready,
2176 zio_ddt_child_write_done, dde, zio->io_priority,
2177 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2179 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2180 dde->dde_lead_zio[p] = cio;
2190 return (ZIO_PIPELINE_CONTINUE);
2193 ddt_entry_t *freedde; /* for debugging */
2196 zio_ddt_free(zio_t *zio)
2198 spa_t *spa = zio->io_spa;
2199 blkptr_t *bp = zio->io_bp;
2200 ddt_t *ddt = ddt_select(spa, bp);
2204 ASSERT(BP_GET_DEDUP(bp));
2205 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2208 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2209 ddp = ddt_phys_select(dde, bp);
2210 ddt_phys_decref(ddp);
2213 return (ZIO_PIPELINE_CONTINUE);
2217 * ==========================================================================
2218 * Allocate and free blocks
2219 * ==========================================================================
2222 zio_dva_allocate(zio_t *zio)
2224 spa_t *spa = zio->io_spa;
2225 metaslab_class_t *mc = spa_normal_class(spa);
2226 blkptr_t *bp = zio->io_bp;
2230 if (zio->io_gang_leader == NULL) {
2231 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2232 zio->io_gang_leader = zio;
2235 ASSERT(BP_IS_HOLE(bp));
2236 ASSERT3U(BP_GET_NDVAS(bp), ==, 0);
2237 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2238 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2239 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2242 * The dump device does not support gang blocks so allocation on
2243 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2244 * the "fast" gang feature.
2246 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2247 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2248 METASLAB_GANG_CHILD : 0;
2249 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2250 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2253 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2254 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2256 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2257 return (zio_write_gang_block(zio));
2258 zio->io_error = error;
2261 return (ZIO_PIPELINE_CONTINUE);
2265 zio_dva_free(zio_t *zio)
2267 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2269 return (ZIO_PIPELINE_CONTINUE);
2273 zio_dva_claim(zio_t *zio)
2277 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2279 zio->io_error = error;
2281 return (ZIO_PIPELINE_CONTINUE);
2285 * Undo an allocation. This is used by zio_done() when an I/O fails
2286 * and we want to give back the block we just allocated.
2287 * This handles both normal blocks and gang blocks.
2290 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2294 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2295 ASSERT(zio->io_bp_override == NULL);
2297 if (!BP_IS_HOLE(bp))
2298 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2301 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2302 zio_dva_unallocate(zio, gn->gn_child[g],
2303 &gn->gn_gbh->zg_blkptr[g]);
2309 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2312 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2313 uint64_t size, boolean_t use_slog)
2317 ASSERT(txg > spa_syncing_txg(spa));
2320 error = metaslab_alloc(spa, spa_log_class(spa), size,
2321 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2324 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2325 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2328 BP_SET_LSIZE(new_bp, size);
2329 BP_SET_PSIZE(new_bp, size);
2330 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2331 BP_SET_CHECKSUM(new_bp,
2332 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2333 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2334 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2335 BP_SET_LEVEL(new_bp, 0);
2336 BP_SET_DEDUP(new_bp, 0);
2337 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2344 * Free an intent log block.
2347 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2349 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2350 ASSERT(!BP_IS_GANG(bp));
2352 zio_free(spa, txg, bp);
2356 * ==========================================================================
2357 * Read and write to physical devices
2358 * ==========================================================================
2361 zio_vdev_io_start(zio_t *zio)
2363 vdev_t *vd = zio->io_vd;
2365 spa_t *spa = zio->io_spa;
2367 ASSERT(zio->io_error == 0);
2368 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2371 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2372 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2375 * The mirror_ops handle multiple DVAs in a single BP.
2377 return (vdev_mirror_ops.vdev_op_io_start(zio));
2381 * We keep track of time-sensitive I/Os so that the scan thread
2382 * can quickly react to certain workloads. In particular, we care
2383 * about non-scrubbing, top-level reads and writes with the following
2385 * - synchronous writes of user data to non-slog devices
2386 * - any reads of user data
2387 * When these conditions are met, adjust the timestamp of spa_last_io
2388 * which allows the scan thread to adjust its workload accordingly.
2390 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2391 vd == vd->vdev_top && !vd->vdev_islog &&
2392 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2393 zio->io_txg != spa_syncing_txg(spa)) {
2394 uint64_t old = spa->spa_last_io;
2395 uint64_t new = ddi_get_lbolt64();
2397 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2400 align = 1ULL << vd->vdev_top->vdev_ashift;
2402 if (P2PHASE(zio->io_size, align) != 0) {
2403 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2404 char *abuf = zio_buf_alloc(asize);
2405 ASSERT(vd == vd->vdev_top);
2406 if (zio->io_type == ZIO_TYPE_WRITE) {
2407 bcopy(zio->io_data, abuf, zio->io_size);
2408 bzero(abuf + zio->io_size, asize - zio->io_size);
2410 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2413 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2414 ASSERT(P2PHASE(zio->io_size, align) == 0);
2415 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2418 * If this is a repair I/O, and there's no self-healing involved --
2419 * that is, we're just resilvering what we expect to resilver --
2420 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2421 * This prevents spurious resilvering with nested replication.
2422 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2423 * A is out of date, we'll read from C+D, then use the data to
2424 * resilver A+B -- but we don't actually want to resilver B, just A.
2425 * The top-level mirror has no way to know this, so instead we just
2426 * discard unnecessary repairs as we work our way down the vdev tree.
2427 * The same logic applies to any form of nested replication:
2428 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2430 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2431 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2432 zio->io_txg != 0 && /* not a delegated i/o */
2433 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2434 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2435 zio_vdev_io_bypass(zio);
2436 return (ZIO_PIPELINE_CONTINUE);
2439 if (vd->vdev_ops->vdev_op_leaf &&
2440 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2442 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2443 return (ZIO_PIPELINE_CONTINUE);
2445 if ((zio = vdev_queue_io(zio)) == NULL)
2446 return (ZIO_PIPELINE_STOP);
2448 if (!vdev_accessible(vd, zio)) {
2449 zio->io_error = ENXIO;
2451 return (ZIO_PIPELINE_STOP);
2455 return (vd->vdev_ops->vdev_op_io_start(zio));
2459 zio_vdev_io_done(zio_t *zio)
2461 vdev_t *vd = zio->io_vd;
2462 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2463 boolean_t unexpected_error = B_FALSE;
2465 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2466 return (ZIO_PIPELINE_STOP);
2468 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2470 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2472 vdev_queue_io_done(zio);
2474 if (zio->io_type == ZIO_TYPE_WRITE)
2475 vdev_cache_write(zio);
2477 if (zio_injection_enabled && zio->io_error == 0)
2478 zio->io_error = zio_handle_device_injection(vd,
2481 if (zio_injection_enabled && zio->io_error == 0)
2482 zio->io_error = zio_handle_label_injection(zio, EIO);
2484 if (zio->io_error) {
2485 if (!vdev_accessible(vd, zio)) {
2486 zio->io_error = ENXIO;
2488 unexpected_error = B_TRUE;
2493 ops->vdev_op_io_done(zio);
2495 if (unexpected_error)
2496 VERIFY(vdev_probe(vd, zio) == NULL);
2498 return (ZIO_PIPELINE_CONTINUE);
2502 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2503 * disk, and use that to finish the checksum ereport later.
2506 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2507 const void *good_buf)
2509 /* no processing needed */
2510 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2515 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2517 void *buf = zio_buf_alloc(zio->io_size);
2519 bcopy(zio->io_data, buf, zio->io_size);
2521 zcr->zcr_cbinfo = zio->io_size;
2522 zcr->zcr_cbdata = buf;
2523 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2524 zcr->zcr_free = zio_buf_free;
2528 zio_vdev_io_assess(zio_t *zio)
2530 vdev_t *vd = zio->io_vd;
2532 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2533 return (ZIO_PIPELINE_STOP);
2535 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2536 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2538 if (zio->io_vsd != NULL) {
2539 zio->io_vsd_ops->vsd_free(zio);
2543 if (zio_injection_enabled && zio->io_error == 0)
2544 zio->io_error = zio_handle_fault_injection(zio, EIO);
2547 * If the I/O failed, determine whether we should attempt to retry it.
2549 * On retry, we cut in line in the issue queue, since we don't want
2550 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2552 if (zio->io_error && vd == NULL &&
2553 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2554 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2555 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2557 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2558 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2559 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2560 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2561 zio_requeue_io_start_cut_in_line);
2562 return (ZIO_PIPELINE_STOP);
2566 * If we got an error on a leaf device, convert it to ENXIO
2567 * if the device is not accessible at all.
2569 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2570 !vdev_accessible(vd, zio))
2571 zio->io_error = ENXIO;
2574 * If we can't write to an interior vdev (mirror or RAID-Z),
2575 * set vdev_cant_write so that we stop trying to allocate from it.
2577 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2578 vd != NULL && !vd->vdev_ops->vdev_op_leaf)
2579 vd->vdev_cant_write = B_TRUE;
2582 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2584 return (ZIO_PIPELINE_CONTINUE);
2588 zio_vdev_io_reissue(zio_t *zio)
2590 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2591 ASSERT(zio->io_error == 0);
2593 zio->io_stage >>= 1;
2597 zio_vdev_io_redone(zio_t *zio)
2599 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2601 zio->io_stage >>= 1;
2605 zio_vdev_io_bypass(zio_t *zio)
2607 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2608 ASSERT(zio->io_error == 0);
2610 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2611 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2615 * ==========================================================================
2616 * Generate and verify checksums
2617 * ==========================================================================
2620 zio_checksum_generate(zio_t *zio)
2622 blkptr_t *bp = zio->io_bp;
2623 enum zio_checksum checksum;
2627 * This is zio_write_phys().
2628 * We're either generating a label checksum, or none at all.
2630 checksum = zio->io_prop.zp_checksum;
2632 if (checksum == ZIO_CHECKSUM_OFF)
2633 return (ZIO_PIPELINE_CONTINUE);
2635 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2637 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2638 ASSERT(!IO_IS_ALLOCATING(zio));
2639 checksum = ZIO_CHECKSUM_GANG_HEADER;
2641 checksum = BP_GET_CHECKSUM(bp);
2645 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2647 return (ZIO_PIPELINE_CONTINUE);
2651 zio_checksum_verify(zio_t *zio)
2653 zio_bad_cksum_t info;
2654 blkptr_t *bp = zio->io_bp;
2657 ASSERT(zio->io_vd != NULL);
2661 * This is zio_read_phys().
2662 * We're either verifying a label checksum, or nothing at all.
2664 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2665 return (ZIO_PIPELINE_CONTINUE);
2667 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2670 if ((error = zio_checksum_error(zio, &info)) != 0) {
2671 zio->io_error = error;
2672 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2673 zfs_ereport_start_checksum(zio->io_spa,
2674 zio->io_vd, zio, zio->io_offset,
2675 zio->io_size, NULL, &info);
2679 return (ZIO_PIPELINE_CONTINUE);
2683 * Called by RAID-Z to ensure we don't compute the checksum twice.
2686 zio_checksum_verified(zio_t *zio)
2688 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2692 * ==========================================================================
2693 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2694 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2695 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2696 * indicate errors that are specific to one I/O, and most likely permanent.
2697 * Any other error is presumed to be worse because we weren't expecting it.
2698 * ==========================================================================
2701 zio_worst_error(int e1, int e2)
2703 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2706 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2707 if (e1 == zio_error_rank[r1])
2710 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2711 if (e2 == zio_error_rank[r2])
2714 return (r1 > r2 ? e1 : e2);
2718 * ==========================================================================
2720 * ==========================================================================
2723 zio_ready(zio_t *zio)
2725 blkptr_t *bp = zio->io_bp;
2726 zio_t *pio, *pio_next;
2728 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2729 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2730 return (ZIO_PIPELINE_STOP);
2732 if (zio->io_ready) {
2733 ASSERT(IO_IS_ALLOCATING(zio));
2734 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2735 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2740 if (bp != NULL && bp != &zio->io_bp_copy)
2741 zio->io_bp_copy = *bp;
2744 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2746 mutex_enter(&zio->io_lock);
2747 zio->io_state[ZIO_WAIT_READY] = 1;
2748 pio = zio_walk_parents(zio);
2749 mutex_exit(&zio->io_lock);
2752 * As we notify zio's parents, new parents could be added.
2753 * New parents go to the head of zio's io_parent_list, however,
2754 * so we will (correctly) not notify them. The remainder of zio's
2755 * io_parent_list, from 'pio_next' onward, cannot change because
2756 * all parents must wait for us to be done before they can be done.
2758 for (; pio != NULL; pio = pio_next) {
2759 pio_next = zio_walk_parents(zio);
2760 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2763 if (zio->io_flags & ZIO_FLAG_NODATA) {
2764 if (BP_IS_GANG(bp)) {
2765 zio->io_flags &= ~ZIO_FLAG_NODATA;
2767 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2768 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2772 if (zio_injection_enabled &&
2773 zio->io_spa->spa_syncing_txg == zio->io_txg)
2774 zio_handle_ignored_writes(zio);
2776 return (ZIO_PIPELINE_CONTINUE);
2780 zio_done(zio_t *zio)
2782 zio_t *pio, *pio_next;
2786 * If our children haven't all completed,
2787 * wait for them and then repeat this pipeline stage.
2789 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2790 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2791 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2792 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2793 return (ZIO_PIPELINE_STOP);
2795 for (c = 0; c < ZIO_CHILD_TYPES; c++)
2796 for (w = 0; w < ZIO_WAIT_TYPES; w++)
2797 ASSERT(zio->io_children[c][w] == 0);
2799 if (zio->io_bp != NULL) {
2800 ASSERT(zio->io_bp->blk_pad[0] == 0);
2801 ASSERT(zio->io_bp->blk_pad[1] == 0);
2802 ASSERT(bcmp(zio->io_bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2803 (zio->io_bp == zio_unique_parent(zio)->io_bp));
2804 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) &&
2805 zio->io_bp_override == NULL &&
2806 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2807 ASSERT(!BP_SHOULD_BYTESWAP(zio->io_bp));
2808 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2809 ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 ||
2810 (BP_COUNT_GANG(zio->io_bp) == BP_GET_NDVAS(zio->io_bp)));
2815 * If there were child vdev/gang/ddt errors, they apply to us now.
2817 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2818 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2819 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2822 * If the I/O on the transformed data was successful, generate any
2823 * checksum reports now while we still have the transformed data.
2825 if (zio->io_error == 0) {
2826 while (zio->io_cksum_report != NULL) {
2827 zio_cksum_report_t *zcr = zio->io_cksum_report;
2828 uint64_t align = zcr->zcr_align;
2829 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2830 char *abuf = zio->io_data;
2832 if (asize != zio->io_size) {
2833 abuf = zio_buf_alloc(asize);
2834 bcopy(zio->io_data, abuf, zio->io_size);
2835 bzero(abuf + zio->io_size, asize - zio->io_size);
2838 zio->io_cksum_report = zcr->zcr_next;
2839 zcr->zcr_next = NULL;
2840 zcr->zcr_finish(zcr, abuf);
2841 zfs_ereport_free_checksum(zcr);
2843 if (asize != zio->io_size)
2844 zio_buf_free(abuf, asize);
2848 zio_pop_transforms(zio); /* note: may set zio->io_error */
2850 vdev_stat_update(zio, zio->io_size);
2853 * If this I/O is attached to a particular vdev is slow, exeeding
2854 * 30 seconds to complete, post an error described the I/O delay.
2855 * We ignore these errors if the device is currently unavailable.
2857 if (zio->io_delay >= zio_delay_max) {
2858 if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd))
2859 zfs_ereport_post(FM_EREPORT_ZFS_DELAY, zio->io_spa,
2860 zio->io_vd, zio, 0, 0);
2863 if (zio->io_error) {
2865 * If this I/O is attached to a particular vdev,
2866 * generate an error message describing the I/O failure
2867 * at the block level. We ignore these errors if the
2868 * device is currently unavailable.
2870 if (zio->io_error != ECKSUM && zio->io_vd != NULL &&
2871 !vdev_is_dead(zio->io_vd))
2872 zfs_ereport_post(FM_EREPORT_ZFS_IO, zio->io_spa,
2873 zio->io_vd, zio, 0, 0);
2875 if ((zio->io_error == EIO || !(zio->io_flags &
2876 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2877 zio == zio->io_logical) {
2879 * For logical I/O requests, tell the SPA to log the
2880 * error and generate a logical data ereport.
2882 spa_log_error(zio->io_spa, zio);
2883 zfs_ereport_post(FM_EREPORT_ZFS_DATA, zio->io_spa, NULL, zio,
2888 if (zio->io_error && zio == zio->io_logical) {
2890 * Determine whether zio should be reexecuted. This will
2891 * propagate all the way to the root via zio_notify_parent().
2893 ASSERT(zio->io_vd == NULL && zio->io_bp != NULL);
2894 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2896 if (IO_IS_ALLOCATING(zio) &&
2897 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2898 if (zio->io_error != ENOSPC)
2899 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2901 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2904 if ((zio->io_type == ZIO_TYPE_READ ||
2905 zio->io_type == ZIO_TYPE_FREE) &&
2906 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
2907 zio->io_error == ENXIO &&
2908 spa_load_state(zio->io_spa) == SPA_LOAD_NONE &&
2909 spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE)
2910 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2912 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2913 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2916 * Here is a possibly good place to attempt to do
2917 * either combinatorial reconstruction or error correction
2918 * based on checksums. It also might be a good place
2919 * to send out preliminary ereports before we suspend
2925 * If there were logical child errors, they apply to us now.
2926 * We defer this until now to avoid conflating logical child
2927 * errors with errors that happened to the zio itself when
2928 * updating vdev stats and reporting FMA events above.
2930 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2932 if ((zio->io_error || zio->io_reexecute) &&
2933 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
2934 !(zio->io_flags & ZIO_FLAG_IO_REWRITE))
2935 zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp);
2937 zio_gang_tree_free(&zio->io_gang_tree);
2940 * Godfather I/Os should never suspend.
2942 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
2943 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
2944 zio->io_reexecute = 0;
2946 if (zio->io_reexecute) {
2948 * This is a logical I/O that wants to reexecute.
2950 * Reexecute is top-down. When an i/o fails, if it's not
2951 * the root, it simply notifies its parent and sticks around.
2952 * The parent, seeing that it still has children in zio_done(),
2953 * does the same. This percolates all the way up to the root.
2954 * The root i/o will reexecute or suspend the entire tree.
2956 * This approach ensures that zio_reexecute() honors
2957 * all the original i/o dependency relationships, e.g.
2958 * parents not executing until children are ready.
2960 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2962 zio->io_gang_leader = NULL;
2964 mutex_enter(&zio->io_lock);
2965 zio->io_state[ZIO_WAIT_DONE] = 1;
2966 mutex_exit(&zio->io_lock);
2969 * "The Godfather" I/O monitors its children but is
2970 * not a true parent to them. It will track them through
2971 * the pipeline but severs its ties whenever they get into
2972 * trouble (e.g. suspended). This allows "The Godfather"
2973 * I/O to return status without blocking.
2975 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2976 zio_link_t *zl = zio->io_walk_link;
2977 pio_next = zio_walk_parents(zio);
2979 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
2980 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
2981 zio_remove_child(pio, zio, zl);
2982 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2986 if ((pio = zio_unique_parent(zio)) != NULL) {
2988 * We're not a root i/o, so there's nothing to do
2989 * but notify our parent. Don't propagate errors
2990 * upward since we haven't permanently failed yet.
2992 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2993 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
2994 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2995 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
2997 * We'd fail again if we reexecuted now, so suspend
2998 * until conditions improve (e.g. device comes online).
3000 zio_suspend(zio->io_spa, zio);
3003 * Reexecution is potentially a huge amount of work.
3004 * Hand it off to the otherwise-unused claim taskq.
3006 ASSERT(taskq_empty_ent(&zio->io_tqent));
3007 (void) taskq_dispatch_ent(
3008 zio->io_spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
3009 (task_func_t *)zio_reexecute, zio, 0,
3012 return (ZIO_PIPELINE_STOP);
3015 ASSERT(zio->io_child_count == 0);
3016 ASSERT(zio->io_reexecute == 0);
3017 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3020 * Report any checksum errors, since the I/O is complete.
3022 while (zio->io_cksum_report != NULL) {
3023 zio_cksum_report_t *zcr = zio->io_cksum_report;
3024 zio->io_cksum_report = zcr->zcr_next;
3025 zcr->zcr_next = NULL;
3026 zcr->zcr_finish(zcr, NULL);
3027 zfs_ereport_free_checksum(zcr);
3031 * It is the responsibility of the done callback to ensure that this
3032 * particular zio is no longer discoverable for adoption, and as
3033 * such, cannot acquire any new parents.
3038 mutex_enter(&zio->io_lock);
3039 zio->io_state[ZIO_WAIT_DONE] = 1;
3040 mutex_exit(&zio->io_lock);
3042 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3043 zio_link_t *zl = zio->io_walk_link;
3044 pio_next = zio_walk_parents(zio);
3045 zio_remove_child(pio, zio, zl);
3046 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3049 if (zio->io_waiter != NULL) {
3050 mutex_enter(&zio->io_lock);
3051 zio->io_executor = NULL;
3052 cv_broadcast(&zio->io_cv);
3053 mutex_exit(&zio->io_lock);
3058 return (ZIO_PIPELINE_STOP);
3062 * ==========================================================================
3063 * I/O pipeline definition
3064 * ==========================================================================
3066 static zio_pipe_stage_t *zio_pipeline[] = {
3072 zio_checksum_generate,
3086 zio_checksum_verify,
3090 #if defined(_KERNEL) && defined(HAVE_SPL)
3091 /* Fault injection */
3092 EXPORT_SYMBOL(zio_injection_enabled);
3093 EXPORT_SYMBOL(zio_inject_fault);
3094 EXPORT_SYMBOL(zio_inject_list_next);
3095 EXPORT_SYMBOL(zio_clear_fault);
3096 EXPORT_SYMBOL(zio_handle_fault_injection);
3097 EXPORT_SYMBOL(zio_handle_device_injection);
3098 EXPORT_SYMBOL(zio_handle_label_injection);
3099 EXPORT_SYMBOL(zio_priority_table);
3100 EXPORT_SYMBOL(zio_type_name);
3102 module_param(zio_bulk_flags, int, 0644);
3103 MODULE_PARM_DESC(zio_bulk_flags, "Additional flags to pass to bulk buffers");
3105 module_param(zio_delay_max, int, 0644);
3106 MODULE_PARM_DESC(zio_delay_max, "Max zio millisec delay before posting event");
3108 module_param(zio_requeue_io_start_cut_in_line, int, 0644);
3109 MODULE_PARM_DESC(zio_requeue_io_start_cut_in_line, "Prioritize requeued I/O");