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_logical = NULL;
555 zio->io_bp = (blkptr_t *)bp;
556 zio->io_bp_copy = *bp;
557 zio->io_bp_orig = *bp;
558 if (type != ZIO_TYPE_WRITE ||
559 zio->io_child_type == ZIO_CHILD_DDT)
560 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
561 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
562 zio->io_logical = zio;
563 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
564 pipeline |= ZIO_GANG_STAGES;
566 zio->io_logical = NULL;
568 bzero(&zio->io_bp_copy, sizeof (blkptr_t));
569 bzero(&zio->io_bp_orig, sizeof (blkptr_t));
574 zio->io_ready = NULL;
576 zio->io_private = private;
577 zio->io_prev_space_delta = 0;
579 zio->io_priority = priority;
582 zio->io_vsd_ops = NULL;
583 zio->io_offset = offset;
584 zio->io_deadline = 0;
585 zio->io_orig_data = zio->io_data = data;
586 zio->io_orig_size = zio->io_size = size;
587 zio->io_orig_flags = zio->io_flags = flags;
588 zio->io_orig_stage = zio->io_stage = stage;
589 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
590 bzero(&zio->io_prop, sizeof (zio_prop_t));
592 zio->io_reexecute = 0;
593 zio->io_bp_override = NULL;
594 zio->io_walk_link = NULL;
595 zio->io_transform_stack = NULL;
598 zio->io_child_count = 0;
599 zio->io_parent_count = 0;
600 zio->io_stall = NULL;
601 zio->io_gang_leader = NULL;
602 zio->io_gang_tree = NULL;
603 zio->io_executor = NULL;
604 zio->io_waiter = NULL;
605 zio->io_cksum_report = NULL;
607 bzero(zio->io_child_error, sizeof (int) * ZIO_CHILD_TYPES);
608 bzero(zio->io_children,
609 sizeof (uint64_t) * ZIO_CHILD_TYPES * ZIO_WAIT_TYPES);
610 bzero(&zio->io_bookmark, sizeof (zbookmark_t));
612 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
613 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
616 zio->io_bookmark = *zb;
619 if (zio->io_logical == NULL)
620 zio->io_logical = pio->io_logical;
621 if (zio->io_child_type == ZIO_CHILD_GANG)
622 zio->io_gang_leader = pio->io_gang_leader;
623 zio_add_child(pio, zio);
626 taskq_init_ent(&zio->io_tqent);
632 zio_destroy(zio_t *zio)
634 kmem_cache_free(zio_cache, zio);
638 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
639 void *private, enum zio_flag flags)
643 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
644 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
645 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
651 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
653 return (zio_null(NULL, spa, NULL, done, private, flags));
657 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
658 void *data, uint64_t size, zio_done_func_t *done, void *private,
659 int priority, enum zio_flag flags, const zbookmark_t *zb)
663 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
664 data, size, done, private,
665 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
666 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
667 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
673 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
674 void *data, uint64_t size, const zio_prop_t *zp,
675 zio_done_func_t *ready, zio_done_func_t *done, void *private,
676 int priority, enum zio_flag flags, const zbookmark_t *zb)
680 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
681 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
682 zp->zp_compress >= ZIO_COMPRESS_OFF &&
683 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
684 zp->zp_type < DMU_OT_NUMTYPES &&
687 zp->zp_copies <= spa_max_replication(spa) &&
689 zp->zp_dedup_verify <= 1);
691 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
692 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
693 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
694 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
696 zio->io_ready = ready;
703 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
704 uint64_t size, zio_done_func_t *done, void *private, int priority,
705 enum zio_flag flags, zbookmark_t *zb)
709 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
710 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
711 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
717 zio_write_override(zio_t *zio, blkptr_t *bp, int copies)
719 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
720 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
721 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
722 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
724 zio->io_prop.zp_copies = copies;
725 zio->io_bp_override = bp;
729 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
731 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
735 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
740 dprintf_bp(bp, "freeing in txg %llu, pass %u",
741 (longlong_t)txg, spa->spa_sync_pass);
743 ASSERT(!BP_IS_HOLE(bp));
744 ASSERT(spa_syncing_txg(spa) == txg);
745 ASSERT(spa_sync_pass(spa) <= SYNC_PASS_DEFERRED_FREE);
747 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
748 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
749 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
755 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
756 zio_done_func_t *done, void *private, enum zio_flag flags)
761 * A claim is an allocation of a specific block. Claims are needed
762 * to support immediate writes in the intent log. The issue is that
763 * immediate writes contain committed data, but in a txg that was
764 * *not* committed. Upon opening the pool after an unclean shutdown,
765 * the intent log claims all blocks that contain immediate write data
766 * so that the SPA knows they're in use.
768 * All claims *must* be resolved in the first txg -- before the SPA
769 * starts allocating blocks -- so that nothing is allocated twice.
770 * If txg == 0 we just verify that the block is claimable.
772 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
773 ASSERT(txg == spa_first_txg(spa) || txg == 0);
774 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
776 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
777 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
778 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
784 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
785 zio_done_func_t *done, void *private, int priority, enum zio_flag flags)
790 if (vd->vdev_children == 0) {
791 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
792 ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
793 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
797 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
799 for (c = 0; c < vd->vdev_children; c++)
800 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
801 done, private, priority, flags));
808 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
809 void *data, int checksum, zio_done_func_t *done, void *private,
810 int priority, enum zio_flag flags, boolean_t labels)
814 ASSERT(vd->vdev_children == 0);
815 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
816 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
817 ASSERT3U(offset + size, <=, vd->vdev_psize);
819 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
820 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
821 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
823 zio->io_prop.zp_checksum = checksum;
829 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
830 void *data, int checksum, zio_done_func_t *done, void *private,
831 int priority, enum zio_flag flags, boolean_t labels)
835 ASSERT(vd->vdev_children == 0);
836 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
837 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
838 ASSERT3U(offset + size, <=, vd->vdev_psize);
840 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
841 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
842 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
844 zio->io_prop.zp_checksum = checksum;
846 if (zio_checksum_table[checksum].ci_eck) {
848 * zec checksums are necessarily destructive -- they modify
849 * the end of the write buffer to hold the verifier/checksum.
850 * Therefore, we must make a local copy in case the data is
851 * being written to multiple places in parallel.
853 void *wbuf = zio_buf_alloc(size);
854 bcopy(data, wbuf, size);
855 zio_push_transform(zio, wbuf, size, size, NULL);
862 * Create a child I/O to do some work for us.
865 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
866 void *data, uint64_t size, int type, int priority, enum zio_flag flags,
867 zio_done_func_t *done, void *private)
869 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
872 ASSERT(vd->vdev_parent ==
873 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
875 if (type == ZIO_TYPE_READ && bp != NULL) {
877 * If we have the bp, then the child should perform the
878 * checksum and the parent need not. This pushes error
879 * detection as close to the leaves as possible and
880 * eliminates redundant checksums in the interior nodes.
882 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
883 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
886 if (vd->vdev_children == 0)
887 offset += VDEV_LABEL_START_SIZE;
889 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
892 * If we've decided to do a repair, the write is not speculative --
893 * even if the original read was.
895 if (flags & ZIO_FLAG_IO_REPAIR)
896 flags &= ~ZIO_FLAG_SPECULATIVE;
898 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
899 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
900 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
906 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
907 int type, int priority, enum zio_flag flags,
908 zio_done_func_t *done, void *private)
912 ASSERT(vd->vdev_ops->vdev_op_leaf);
914 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
915 data, size, done, private, type, priority,
916 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
918 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
924 zio_flush(zio_t *zio, vdev_t *vd)
926 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
927 NULL, NULL, ZIO_PRIORITY_NOW,
928 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
932 zio_shrink(zio_t *zio, uint64_t size)
934 ASSERT(zio->io_executor == NULL);
935 ASSERT(zio->io_orig_size == zio->io_size);
936 ASSERT(size <= zio->io_size);
939 * We don't shrink for raidz because of problems with the
940 * reconstruction when reading back less than the block size.
941 * Note, BP_IS_RAIDZ() assumes no compression.
943 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
944 if (!BP_IS_RAIDZ(zio->io_bp))
945 zio->io_orig_size = zio->io_size = size;
949 * ==========================================================================
950 * Prepare to read and write logical blocks
951 * ==========================================================================
955 zio_read_bp_init(zio_t *zio)
957 blkptr_t *bp = zio->io_bp;
959 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
960 zio->io_child_type == ZIO_CHILD_LOGICAL &&
961 !(zio->io_flags & ZIO_FLAG_RAW)) {
962 uint64_t psize = BP_GET_PSIZE(bp);
963 void *cbuf = zio_buf_alloc(psize);
965 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
968 if (!dmu_ot[BP_GET_TYPE(bp)].ot_metadata && BP_GET_LEVEL(bp) == 0)
969 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
971 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
972 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
974 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
975 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
977 return (ZIO_PIPELINE_CONTINUE);
981 zio_write_bp_init(zio_t *zio)
983 spa_t *spa = zio->io_spa;
984 zio_prop_t *zp = &zio->io_prop;
985 enum zio_compress compress = zp->zp_compress;
986 blkptr_t *bp = zio->io_bp;
987 uint64_t lsize = zio->io_size;
988 uint64_t psize = lsize;
992 * If our children haven't all reached the ready stage,
993 * wait for them and then repeat this pipeline stage.
995 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
996 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
997 return (ZIO_PIPELINE_STOP);
999 if (!IO_IS_ALLOCATING(zio))
1000 return (ZIO_PIPELINE_CONTINUE);
1002 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1004 if (zio->io_bp_override) {
1005 ASSERT(bp->blk_birth != zio->io_txg);
1006 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1008 *bp = *zio->io_bp_override;
1009 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1011 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1012 return (ZIO_PIPELINE_CONTINUE);
1014 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1015 zp->zp_dedup_verify);
1017 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1018 BP_SET_DEDUP(bp, 1);
1019 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1020 return (ZIO_PIPELINE_CONTINUE);
1022 zio->io_bp_override = NULL;
1026 if (bp->blk_birth == zio->io_txg) {
1028 * We're rewriting an existing block, which means we're
1029 * working on behalf of spa_sync(). For spa_sync() to
1030 * converge, it must eventually be the case that we don't
1031 * have to allocate new blocks. But compression changes
1032 * the blocksize, which forces a reallocate, and makes
1033 * convergence take longer. Therefore, after the first
1034 * few passes, stop compressing to ensure convergence.
1036 pass = spa_sync_pass(spa);
1038 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1039 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1040 ASSERT(!BP_GET_DEDUP(bp));
1042 if (pass > SYNC_PASS_DONT_COMPRESS)
1043 compress = ZIO_COMPRESS_OFF;
1045 /* Make sure someone doesn't change their mind on overwrites */
1046 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
1047 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1050 if (compress != ZIO_COMPRESS_OFF) {
1051 void *cbuf = zio_buf_alloc(lsize);
1052 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1053 if (psize == 0 || psize == lsize) {
1054 compress = ZIO_COMPRESS_OFF;
1055 zio_buf_free(cbuf, lsize);
1057 ASSERT(psize < lsize);
1058 zio_push_transform(zio, cbuf, psize, lsize, NULL);
1063 * The final pass of spa_sync() must be all rewrites, but the first
1064 * few passes offer a trade-off: allocating blocks defers convergence,
1065 * but newly allocated blocks are sequential, so they can be written
1066 * to disk faster. Therefore, we allow the first few passes of
1067 * spa_sync() to allocate new blocks, but force rewrites after that.
1068 * There should only be a handful of blocks after pass 1 in any case.
1070 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
1071 pass > SYNC_PASS_REWRITE) {
1072 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1074 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1075 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1078 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1082 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1084 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1085 BP_SET_LSIZE(bp, lsize);
1086 BP_SET_PSIZE(bp, psize);
1087 BP_SET_COMPRESS(bp, compress);
1088 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1089 BP_SET_TYPE(bp, zp->zp_type);
1090 BP_SET_LEVEL(bp, zp->zp_level);
1091 BP_SET_DEDUP(bp, zp->zp_dedup);
1092 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1094 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1095 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1096 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1100 return (ZIO_PIPELINE_CONTINUE);
1104 zio_free_bp_init(zio_t *zio)
1106 blkptr_t *bp = zio->io_bp;
1108 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1109 if (BP_GET_DEDUP(bp))
1110 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1113 return (ZIO_PIPELINE_CONTINUE);
1117 * ==========================================================================
1118 * Execute the I/O pipeline
1119 * ==========================================================================
1123 zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q, boolean_t cutinline)
1125 spa_t *spa = zio->io_spa;
1126 zio_type_t t = zio->io_type;
1127 int flags = (cutinline ? TQ_FRONT : 0);
1130 * If we're a config writer or a probe, the normal issue and
1131 * interrupt threads may all be blocked waiting for the config lock.
1132 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1134 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1138 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1140 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1144 * If this is a high priority I/O, then use the high priority taskq.
1146 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1147 spa->spa_zio_taskq[t][q + 1] != NULL)
1150 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1153 * NB: We are assuming that the zio can only be dispatched
1154 * to a single taskq at a time. It would be a grievous error
1155 * to dispatch the zio to another taskq at the same time.
1157 ASSERT(taskq_empty_ent(&zio->io_tqent));
1158 taskq_dispatch_ent(spa->spa_zio_taskq[t][q],
1159 (task_func_t *)zio_execute, zio, flags, &zio->io_tqent);
1163 zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
1165 kthread_t *executor = zio->io_executor;
1166 spa_t *spa = zio->io_spa;
1169 for (t = 0; t < ZIO_TYPES; t++)
1170 if (taskq_member(spa->spa_zio_taskq[t][q], executor))
1177 zio_issue_async(zio_t *zio)
1179 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1181 return (ZIO_PIPELINE_STOP);
1185 zio_interrupt(zio_t *zio)
1187 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1191 * Execute the I/O pipeline until one of the following occurs:
1192 * (1) the I/O completes; (2) the pipeline stalls waiting for
1193 * dependent child I/Os; (3) the I/O issues, so we're waiting
1194 * for an I/O completion interrupt; (4) the I/O is delegated by
1195 * vdev-level caching or aggregation; (5) the I/O is deferred
1196 * due to vdev-level queueing; (6) the I/O is handed off to
1197 * another thread. In all cases, the pipeline stops whenever
1198 * there's no CPU work; it never burns a thread in cv_wait().
1200 * There's no locking on io_stage because there's no legitimate way
1201 * for multiple threads to be attempting to process the same I/O.
1203 static zio_pipe_stage_t *zio_pipeline[];
1206 * zio_execute() is a wrapper around the static function
1207 * __zio_execute() so that we can force __zio_execute() to be
1208 * inlined. This reduces stack overhead which is important
1209 * because __zio_execute() is called recursively in several zio
1210 * code paths. zio_execute() itself cannot be inlined because
1211 * it is externally visible.
1214 zio_execute(zio_t *zio)
1219 __attribute__((always_inline))
1221 __zio_execute(zio_t *zio)
1223 zio->io_executor = curthread;
1225 while (zio->io_stage < ZIO_STAGE_DONE) {
1226 enum zio_stage pipeline = zio->io_pipeline;
1227 enum zio_stage stage = zio->io_stage;
1232 ASSERT(!MUTEX_HELD(&zio->io_lock));
1233 ASSERT(ISP2(stage));
1234 ASSERT(zio->io_stall == NULL);
1238 } while ((stage & pipeline) == 0);
1240 ASSERT(stage <= ZIO_STAGE_DONE);
1242 dsl = spa_get_dsl(zio->io_spa);
1243 cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1244 zio_requeue_io_start_cut_in_line : B_FALSE;
1247 * If we are in interrupt context and this pipeline stage
1248 * will grab a config lock that is held across I/O,
1249 * or may wait for an I/O that needs an interrupt thread
1250 * to complete, issue async to avoid deadlock.
1252 * If we are in the txg_sync_thread or being called
1253 * during pool init issue async to minimize stack depth.
1254 * Both of these call paths may be recursively called.
1256 * For VDEV_IO_START, we cut in line so that the io will
1257 * be sent to disk promptly.
1259 if (((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1260 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) ||
1261 (dsl != NULL && dsl_pool_sync_context(dsl))) {
1262 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1266 zio->io_stage = stage;
1267 rv = zio_pipeline[highbit(stage) - 1](zio);
1269 if (rv == ZIO_PIPELINE_STOP)
1272 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1278 * ==========================================================================
1279 * Initiate I/O, either sync or async
1280 * ==========================================================================
1283 zio_wait(zio_t *zio)
1287 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1288 ASSERT(zio->io_executor == NULL);
1290 zio->io_waiter = curthread;
1294 mutex_enter(&zio->io_lock);
1295 while (zio->io_executor != NULL)
1296 cv_wait(&zio->io_cv, &zio->io_lock);
1297 mutex_exit(&zio->io_lock);
1299 error = zio->io_error;
1306 zio_nowait(zio_t *zio)
1308 ASSERT(zio->io_executor == NULL);
1310 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1311 zio_unique_parent(zio) == NULL) {
1313 * This is a logical async I/O with no parent to wait for it.
1314 * We add it to the spa_async_root_zio "Godfather" I/O which
1315 * will ensure they complete prior to unloading the pool.
1317 spa_t *spa = zio->io_spa;
1319 zio_add_child(spa->spa_async_zio_root, zio);
1326 * ==========================================================================
1327 * Reexecute or suspend/resume failed I/O
1328 * ==========================================================================
1332 zio_reexecute(zio_t *pio)
1334 zio_t *cio, *cio_next;
1337 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1338 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1339 ASSERT(pio->io_gang_leader == NULL);
1340 ASSERT(pio->io_gang_tree == NULL);
1342 pio->io_flags = pio->io_orig_flags;
1343 pio->io_stage = pio->io_orig_stage;
1344 pio->io_pipeline = pio->io_orig_pipeline;
1345 pio->io_reexecute = 0;
1347 for (w = 0; w < ZIO_WAIT_TYPES; w++)
1348 pio->io_state[w] = 0;
1349 for (c = 0; c < ZIO_CHILD_TYPES; c++)
1350 pio->io_child_error[c] = 0;
1352 if (IO_IS_ALLOCATING(pio))
1353 BP_ZERO(pio->io_bp);
1356 * As we reexecute pio's children, new children could be created.
1357 * New children go to the head of pio's io_child_list, however,
1358 * so we will (correctly) not reexecute them. The key is that
1359 * the remainder of pio's io_child_list, from 'cio_next' onward,
1360 * cannot be affected by any side effects of reexecuting 'cio'.
1362 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1363 cio_next = zio_walk_children(pio);
1364 mutex_enter(&pio->io_lock);
1365 for (w = 0; w < ZIO_WAIT_TYPES; w++)
1366 pio->io_children[cio->io_child_type][w]++;
1367 mutex_exit(&pio->io_lock);
1372 * Now that all children have been reexecuted, execute the parent.
1373 * We don't reexecute "The Godfather" I/O here as it's the
1374 * responsibility of the caller to wait on him.
1376 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1381 zio_suspend(spa_t *spa, zio_t *zio)
1383 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1384 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1385 "failure and the failure mode property for this pool "
1386 "is set to panic.", spa_name(spa));
1388 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1390 mutex_enter(&spa->spa_suspend_lock);
1392 if (spa->spa_suspend_zio_root == NULL)
1393 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1394 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1395 ZIO_FLAG_GODFATHER);
1397 spa->spa_suspended = B_TRUE;
1400 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1401 ASSERT(zio != spa->spa_suspend_zio_root);
1402 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1403 ASSERT(zio_unique_parent(zio) == NULL);
1404 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1405 zio_add_child(spa->spa_suspend_zio_root, zio);
1408 mutex_exit(&spa->spa_suspend_lock);
1412 zio_resume(spa_t *spa)
1417 * Reexecute all previously suspended i/o.
1419 mutex_enter(&spa->spa_suspend_lock);
1420 spa->spa_suspended = B_FALSE;
1421 cv_broadcast(&spa->spa_suspend_cv);
1422 pio = spa->spa_suspend_zio_root;
1423 spa->spa_suspend_zio_root = NULL;
1424 mutex_exit(&spa->spa_suspend_lock);
1430 return (zio_wait(pio));
1434 zio_resume_wait(spa_t *spa)
1436 mutex_enter(&spa->spa_suspend_lock);
1437 while (spa_suspended(spa))
1438 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1439 mutex_exit(&spa->spa_suspend_lock);
1443 * ==========================================================================
1446 * A gang block is a collection of small blocks that looks to the DMU
1447 * like one large block. When zio_dva_allocate() cannot find a block
1448 * of the requested size, due to either severe fragmentation or the pool
1449 * being nearly full, it calls zio_write_gang_block() to construct the
1450 * block from smaller fragments.
1452 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1453 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1454 * an indirect block: it's an array of block pointers. It consumes
1455 * only one sector and hence is allocatable regardless of fragmentation.
1456 * The gang header's bps point to its gang members, which hold the data.
1458 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1459 * as the verifier to ensure uniqueness of the SHA256 checksum.
1460 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1461 * not the gang header. This ensures that data block signatures (needed for
1462 * deduplication) are independent of how the block is physically stored.
1464 * Gang blocks can be nested: a gang member may itself be a gang block.
1465 * Thus every gang block is a tree in which root and all interior nodes are
1466 * gang headers, and the leaves are normal blocks that contain user data.
1467 * The root of the gang tree is called the gang leader.
1469 * To perform any operation (read, rewrite, free, claim) on a gang block,
1470 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1471 * in the io_gang_tree field of the original logical i/o by recursively
1472 * reading the gang leader and all gang headers below it. This yields
1473 * an in-core tree containing the contents of every gang header and the
1474 * bps for every constituent of the gang block.
1476 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1477 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1478 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1479 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1480 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1481 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1482 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1483 * of the gang header plus zio_checksum_compute() of the data to update the
1484 * gang header's blk_cksum as described above.
1486 * The two-phase assemble/issue model solves the problem of partial failure --
1487 * what if you'd freed part of a gang block but then couldn't read the
1488 * gang header for another part? Assembling the entire gang tree first
1489 * ensures that all the necessary gang header I/O has succeeded before
1490 * starting the actual work of free, claim, or write. Once the gang tree
1491 * is assembled, free and claim are in-memory operations that cannot fail.
1493 * In the event that a gang write fails, zio_dva_unallocate() walks the
1494 * gang tree to immediately free (i.e. insert back into the space map)
1495 * everything we've allocated. This ensures that we don't get ENOSPC
1496 * errors during repeated suspend/resume cycles due to a flaky device.
1498 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1499 * the gang tree, we won't modify the block, so we can safely defer the free
1500 * (knowing that the block is still intact). If we *can* assemble the gang
1501 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1502 * each constituent bp and we can allocate a new block on the next sync pass.
1504 * In all cases, the gang tree allows complete recovery from partial failure.
1505 * ==========================================================================
1509 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1514 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1515 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1516 &pio->io_bookmark));
1520 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1525 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1526 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1527 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1529 * As we rewrite each gang header, the pipeline will compute
1530 * a new gang block header checksum for it; but no one will
1531 * compute a new data checksum, so we do that here. The one
1532 * exception is the gang leader: the pipeline already computed
1533 * its data checksum because that stage precedes gang assembly.
1534 * (Presently, nothing actually uses interior data checksums;
1535 * this is just good hygiene.)
1537 if (gn != pio->io_gang_leader->io_gang_tree) {
1538 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1539 data, BP_GET_PSIZE(bp));
1542 * If we are here to damage data for testing purposes,
1543 * leave the GBH alone so that we can detect the damage.
1545 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1546 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1548 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1549 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1550 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1558 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1560 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1561 ZIO_GANG_CHILD_FLAGS(pio)));
1566 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1568 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1569 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1572 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1581 static void zio_gang_tree_assemble_done(zio_t *zio);
1583 static zio_gang_node_t *
1584 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1586 zio_gang_node_t *gn;
1588 ASSERT(*gnpp == NULL);
1590 gn = kmem_zalloc(sizeof (*gn), KM_PUSHPAGE);
1591 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1598 zio_gang_node_free(zio_gang_node_t **gnpp)
1600 zio_gang_node_t *gn = *gnpp;
1603 for (g = 0; g < SPA_GBH_NBLKPTRS; g++)
1604 ASSERT(gn->gn_child[g] == NULL);
1606 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1607 kmem_free(gn, sizeof (*gn));
1612 zio_gang_tree_free(zio_gang_node_t **gnpp)
1614 zio_gang_node_t *gn = *gnpp;
1620 for (g = 0; g < SPA_GBH_NBLKPTRS; g++)
1621 zio_gang_tree_free(&gn->gn_child[g]);
1623 zio_gang_node_free(gnpp);
1627 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1629 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1631 ASSERT(gio->io_gang_leader == gio);
1632 ASSERT(BP_IS_GANG(bp));
1634 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1635 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1636 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1640 zio_gang_tree_assemble_done(zio_t *zio)
1642 zio_t *gio = zio->io_gang_leader;
1643 zio_gang_node_t *gn = zio->io_private;
1644 blkptr_t *bp = zio->io_bp;
1647 ASSERT(gio == zio_unique_parent(zio));
1648 ASSERT(zio->io_child_count == 0);
1653 if (BP_SHOULD_BYTESWAP(bp))
1654 byteswap_uint64_array(zio->io_data, zio->io_size);
1656 ASSERT(zio->io_data == gn->gn_gbh);
1657 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1658 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1660 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1661 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1662 if (!BP_IS_GANG(gbp))
1664 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1669 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1671 zio_t *gio = pio->io_gang_leader;
1675 ASSERT(BP_IS_GANG(bp) == !!gn);
1676 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1677 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1680 * If you're a gang header, your data is in gn->gn_gbh.
1681 * If you're a gang member, your data is in 'data' and gn == NULL.
1683 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1686 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1688 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1689 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1690 if (BP_IS_HOLE(gbp))
1692 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1693 data = (char *)data + BP_GET_PSIZE(gbp);
1697 if (gn == gio->io_gang_tree)
1698 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1705 zio_gang_assemble(zio_t *zio)
1707 blkptr_t *bp = zio->io_bp;
1709 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1710 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1712 zio->io_gang_leader = zio;
1714 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1716 return (ZIO_PIPELINE_CONTINUE);
1720 zio_gang_issue(zio_t *zio)
1722 blkptr_t *bp = zio->io_bp;
1724 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1725 return (ZIO_PIPELINE_STOP);
1727 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1728 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1730 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1731 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1733 zio_gang_tree_free(&zio->io_gang_tree);
1735 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1737 return (ZIO_PIPELINE_CONTINUE);
1741 zio_write_gang_member_ready(zio_t *zio)
1743 zio_t *pio = zio_unique_parent(zio);
1744 ASSERTV(zio_t *gio = zio->io_gang_leader;)
1745 dva_t *cdva = zio->io_bp->blk_dva;
1746 dva_t *pdva = pio->io_bp->blk_dva;
1750 if (BP_IS_HOLE(zio->io_bp))
1753 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1755 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1756 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1757 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1758 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1759 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1761 mutex_enter(&pio->io_lock);
1762 for (d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1763 ASSERT(DVA_GET_GANG(&pdva[d]));
1764 asize = DVA_GET_ASIZE(&pdva[d]);
1765 asize += DVA_GET_ASIZE(&cdva[d]);
1766 DVA_SET_ASIZE(&pdva[d], asize);
1768 mutex_exit(&pio->io_lock);
1772 zio_write_gang_block(zio_t *pio)
1774 spa_t *spa = pio->io_spa;
1775 blkptr_t *bp = pio->io_bp;
1776 zio_t *gio = pio->io_gang_leader;
1778 zio_gang_node_t *gn, **gnpp;
1779 zio_gbh_phys_t *gbh;
1780 uint64_t txg = pio->io_txg;
1781 uint64_t resid = pio->io_size;
1783 int copies = gio->io_prop.zp_copies;
1784 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1788 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1789 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1790 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1792 pio->io_error = error;
1793 return (ZIO_PIPELINE_CONTINUE);
1797 gnpp = &gio->io_gang_tree;
1799 gnpp = pio->io_private;
1800 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1803 gn = zio_gang_node_alloc(gnpp);
1805 bzero(gbh, SPA_GANGBLOCKSIZE);
1808 * Create the gang header.
1810 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1811 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1814 * Create and nowait the gang children.
1816 for (g = 0; resid != 0; resid -= lsize, g++) {
1817 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1819 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1821 zp.zp_checksum = gio->io_prop.zp_checksum;
1822 zp.zp_compress = ZIO_COMPRESS_OFF;
1823 zp.zp_type = DMU_OT_NONE;
1825 zp.zp_copies = gio->io_prop.zp_copies;
1827 zp.zp_dedup_verify = 0;
1829 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1830 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1831 zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1832 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1833 &pio->io_bookmark));
1837 * Set pio's pipeline to just wait for zio to finish.
1839 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1843 return (ZIO_PIPELINE_CONTINUE);
1847 * ==========================================================================
1849 * ==========================================================================
1852 zio_ddt_child_read_done(zio_t *zio)
1854 blkptr_t *bp = zio->io_bp;
1855 ddt_entry_t *dde = zio->io_private;
1857 zio_t *pio = zio_unique_parent(zio);
1859 mutex_enter(&pio->io_lock);
1860 ddp = ddt_phys_select(dde, bp);
1861 if (zio->io_error == 0)
1862 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1863 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1864 dde->dde_repair_data = zio->io_data;
1866 zio_buf_free(zio->io_data, zio->io_size);
1867 mutex_exit(&pio->io_lock);
1871 zio_ddt_read_start(zio_t *zio)
1873 blkptr_t *bp = zio->io_bp;
1876 ASSERT(BP_GET_DEDUP(bp));
1877 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1878 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1880 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1881 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1882 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1883 ddt_phys_t *ddp = dde->dde_phys;
1884 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1887 ASSERT(zio->io_vsd == NULL);
1890 if (ddp_self == NULL)
1891 return (ZIO_PIPELINE_CONTINUE);
1893 for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1894 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1896 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1898 zio_nowait(zio_read(zio, zio->io_spa, &blk,
1899 zio_buf_alloc(zio->io_size), zio->io_size,
1900 zio_ddt_child_read_done, dde, zio->io_priority,
1901 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1902 &zio->io_bookmark));
1904 return (ZIO_PIPELINE_CONTINUE);
1907 zio_nowait(zio_read(zio, zio->io_spa, bp,
1908 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1909 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1911 return (ZIO_PIPELINE_CONTINUE);
1915 zio_ddt_read_done(zio_t *zio)
1917 blkptr_t *bp = zio->io_bp;
1919 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1920 return (ZIO_PIPELINE_STOP);
1922 ASSERT(BP_GET_DEDUP(bp));
1923 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1924 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1926 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1927 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1928 ddt_entry_t *dde = zio->io_vsd;
1930 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
1931 return (ZIO_PIPELINE_CONTINUE);
1934 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
1935 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1936 return (ZIO_PIPELINE_STOP);
1938 if (dde->dde_repair_data != NULL) {
1939 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
1940 zio->io_child_error[ZIO_CHILD_DDT] = 0;
1942 ddt_repair_done(ddt, dde);
1946 ASSERT(zio->io_vsd == NULL);
1948 return (ZIO_PIPELINE_CONTINUE);
1952 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
1954 spa_t *spa = zio->io_spa;
1958 * Note: we compare the original data, not the transformed data,
1959 * because when zio->io_bp is an override bp, we will not have
1960 * pushed the I/O transforms. That's an important optimization
1961 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
1963 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1964 zio_t *lio = dde->dde_lead_zio[p];
1967 return (lio->io_orig_size != zio->io_orig_size ||
1968 bcmp(zio->io_orig_data, lio->io_orig_data,
1969 zio->io_orig_size) != 0);
1973 for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1974 ddt_phys_t *ddp = &dde->dde_phys[p];
1976 if (ddp->ddp_phys_birth != 0) {
1977 arc_buf_t *abuf = NULL;
1978 uint32_t aflags = ARC_WAIT;
1979 blkptr_t blk = *zio->io_bp;
1982 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
1986 error = arc_read_nolock(NULL, spa, &blk,
1987 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
1988 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1989 &aflags, &zio->io_bookmark);
1992 if (arc_buf_size(abuf) != zio->io_orig_size ||
1993 bcmp(abuf->b_data, zio->io_orig_data,
1994 zio->io_orig_size) != 0)
1996 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
2000 return (error != 0);
2008 zio_ddt_child_write_ready(zio_t *zio)
2010 int p = zio->io_prop.zp_copies;
2011 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2012 ddt_entry_t *dde = zio->io_private;
2013 ddt_phys_t *ddp = &dde->dde_phys[p];
2021 ASSERT(dde->dde_lead_zio[p] == zio);
2023 ddt_phys_fill(ddp, zio->io_bp);
2025 while ((pio = zio_walk_parents(zio)) != NULL)
2026 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2032 zio_ddt_child_write_done(zio_t *zio)
2034 int p = zio->io_prop.zp_copies;
2035 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2036 ddt_entry_t *dde = zio->io_private;
2037 ddt_phys_t *ddp = &dde->dde_phys[p];
2041 ASSERT(ddp->ddp_refcnt == 0);
2042 ASSERT(dde->dde_lead_zio[p] == zio);
2043 dde->dde_lead_zio[p] = NULL;
2045 if (zio->io_error == 0) {
2046 while (zio_walk_parents(zio) != NULL)
2047 ddt_phys_addref(ddp);
2049 ddt_phys_clear(ddp);
2056 zio_ddt_ditto_write_done(zio_t *zio)
2058 int p = DDT_PHYS_DITTO;
2059 blkptr_t *bp = zio->io_bp;
2060 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2061 ddt_entry_t *dde = zio->io_private;
2062 ddt_phys_t *ddp = &dde->dde_phys[p];
2063 ddt_key_t *ddk = &dde->dde_key;
2064 ASSERTV(zio_prop_t *zp = &zio->io_prop);
2068 ASSERT(ddp->ddp_refcnt == 0);
2069 ASSERT(dde->dde_lead_zio[p] == zio);
2070 dde->dde_lead_zio[p] = NULL;
2072 if (zio->io_error == 0) {
2073 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2074 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2075 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2076 if (ddp->ddp_phys_birth != 0)
2077 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2078 ddt_phys_fill(ddp, bp);
2085 zio_ddt_write(zio_t *zio)
2087 spa_t *spa = zio->io_spa;
2088 blkptr_t *bp = zio->io_bp;
2089 uint64_t txg = zio->io_txg;
2090 zio_prop_t *zp = &zio->io_prop;
2091 int p = zp->zp_copies;
2095 ddt_t *ddt = ddt_select(spa, bp);
2099 ASSERT(BP_GET_DEDUP(bp));
2100 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2101 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2104 dde = ddt_lookup(ddt, bp, B_TRUE);
2105 ddp = &dde->dde_phys[p];
2107 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2109 * If we're using a weak checksum, upgrade to a strong checksum
2110 * and try again. If we're already using a strong checksum,
2111 * we can't resolve it, so just convert to an ordinary write.
2112 * (And automatically e-mail a paper to Nature?)
2114 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2115 zp->zp_checksum = spa_dedup_checksum(spa);
2116 zio_pop_transforms(zio);
2117 zio->io_stage = ZIO_STAGE_OPEN;
2122 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2124 return (ZIO_PIPELINE_CONTINUE);
2127 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2128 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2130 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2131 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2132 zio_prop_t czp = *zp;
2134 czp.zp_copies = ditto_copies;
2137 * If we arrived here with an override bp, we won't have run
2138 * the transform stack, so we won't have the data we need to
2139 * generate a child i/o. So, toss the override bp and restart.
2140 * This is safe, because using the override bp is just an
2141 * optimization; and it's rare, so the cost doesn't matter.
2143 if (zio->io_bp_override) {
2144 zio_pop_transforms(zio);
2145 zio->io_stage = ZIO_STAGE_OPEN;
2146 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2147 zio->io_bp_override = NULL;
2150 return (ZIO_PIPELINE_CONTINUE);
2153 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2154 zio->io_orig_size, &czp, NULL,
2155 zio_ddt_ditto_write_done, dde, zio->io_priority,
2156 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2158 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2159 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2162 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2163 if (ddp->ddp_phys_birth != 0)
2164 ddt_bp_fill(ddp, bp, txg);
2165 if (dde->dde_lead_zio[p] != NULL)
2166 zio_add_child(zio, dde->dde_lead_zio[p]);
2168 ddt_phys_addref(ddp);
2169 } else if (zio->io_bp_override) {
2170 ASSERT(bp->blk_birth == txg);
2171 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2172 ddt_phys_fill(ddp, bp);
2173 ddt_phys_addref(ddp);
2175 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2176 zio->io_orig_size, zp, zio_ddt_child_write_ready,
2177 zio_ddt_child_write_done, dde, zio->io_priority,
2178 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2180 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2181 dde->dde_lead_zio[p] = cio;
2191 return (ZIO_PIPELINE_CONTINUE);
2194 ddt_entry_t *freedde; /* for debugging */
2197 zio_ddt_free(zio_t *zio)
2199 spa_t *spa = zio->io_spa;
2200 blkptr_t *bp = zio->io_bp;
2201 ddt_t *ddt = ddt_select(spa, bp);
2205 ASSERT(BP_GET_DEDUP(bp));
2206 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2209 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2210 ddp = ddt_phys_select(dde, bp);
2211 ddt_phys_decref(ddp);
2214 return (ZIO_PIPELINE_CONTINUE);
2218 * ==========================================================================
2219 * Allocate and free blocks
2220 * ==========================================================================
2223 zio_dva_allocate(zio_t *zio)
2225 spa_t *spa = zio->io_spa;
2226 metaslab_class_t *mc = spa_normal_class(spa);
2227 blkptr_t *bp = zio->io_bp;
2231 if (zio->io_gang_leader == NULL) {
2232 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2233 zio->io_gang_leader = zio;
2236 ASSERT(BP_IS_HOLE(bp));
2237 ASSERT3U(BP_GET_NDVAS(bp), ==, 0);
2238 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2239 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2240 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2243 * The dump device does not support gang blocks so allocation on
2244 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2245 * the "fast" gang feature.
2247 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2248 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2249 METASLAB_GANG_CHILD : 0;
2250 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2251 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2254 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2255 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2257 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2258 return (zio_write_gang_block(zio));
2259 zio->io_error = error;
2262 return (ZIO_PIPELINE_CONTINUE);
2266 zio_dva_free(zio_t *zio)
2268 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2270 return (ZIO_PIPELINE_CONTINUE);
2274 zio_dva_claim(zio_t *zio)
2278 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2280 zio->io_error = error;
2282 return (ZIO_PIPELINE_CONTINUE);
2286 * Undo an allocation. This is used by zio_done() when an I/O fails
2287 * and we want to give back the block we just allocated.
2288 * This handles both normal blocks and gang blocks.
2291 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2295 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2296 ASSERT(zio->io_bp_override == NULL);
2298 if (!BP_IS_HOLE(bp))
2299 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2302 for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2303 zio_dva_unallocate(zio, gn->gn_child[g],
2304 &gn->gn_gbh->zg_blkptr[g]);
2310 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2313 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2314 uint64_t size, boolean_t use_slog)
2318 ASSERT(txg > spa_syncing_txg(spa));
2321 error = metaslab_alloc(spa, spa_log_class(spa), size,
2322 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2325 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2326 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID);
2329 BP_SET_LSIZE(new_bp, size);
2330 BP_SET_PSIZE(new_bp, size);
2331 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2332 BP_SET_CHECKSUM(new_bp,
2333 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2334 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2335 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2336 BP_SET_LEVEL(new_bp, 0);
2337 BP_SET_DEDUP(new_bp, 0);
2338 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2345 * Free an intent log block.
2348 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2350 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2351 ASSERT(!BP_IS_GANG(bp));
2353 zio_free(spa, txg, bp);
2357 * ==========================================================================
2358 * Read and write to physical devices
2359 * ==========================================================================
2362 zio_vdev_io_start(zio_t *zio)
2364 vdev_t *vd = zio->io_vd;
2366 spa_t *spa = zio->io_spa;
2368 ASSERT(zio->io_error == 0);
2369 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2372 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2373 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2376 * The mirror_ops handle multiple DVAs in a single BP.
2378 return (vdev_mirror_ops.vdev_op_io_start(zio));
2382 * We keep track of time-sensitive I/Os so that the scan thread
2383 * can quickly react to certain workloads. In particular, we care
2384 * about non-scrubbing, top-level reads and writes with the following
2386 * - synchronous writes of user data to non-slog devices
2387 * - any reads of user data
2388 * When these conditions are met, adjust the timestamp of spa_last_io
2389 * which allows the scan thread to adjust its workload accordingly.
2391 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2392 vd == vd->vdev_top && !vd->vdev_islog &&
2393 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2394 zio->io_txg != spa_syncing_txg(spa)) {
2395 uint64_t old = spa->spa_last_io;
2396 uint64_t new = ddi_get_lbolt64();
2398 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2401 align = 1ULL << vd->vdev_top->vdev_ashift;
2403 if (P2PHASE(zio->io_size, align) != 0) {
2404 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2405 char *abuf = zio_buf_alloc(asize);
2406 ASSERT(vd == vd->vdev_top);
2407 if (zio->io_type == ZIO_TYPE_WRITE) {
2408 bcopy(zio->io_data, abuf, zio->io_size);
2409 bzero(abuf + zio->io_size, asize - zio->io_size);
2411 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2414 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2415 ASSERT(P2PHASE(zio->io_size, align) == 0);
2416 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2419 * If this is a repair I/O, and there's no self-healing involved --
2420 * that is, we're just resilvering what we expect to resilver --
2421 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2422 * This prevents spurious resilvering with nested replication.
2423 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2424 * A is out of date, we'll read from C+D, then use the data to
2425 * resilver A+B -- but we don't actually want to resilver B, just A.
2426 * The top-level mirror has no way to know this, so instead we just
2427 * discard unnecessary repairs as we work our way down the vdev tree.
2428 * The same logic applies to any form of nested replication:
2429 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2431 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2432 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2433 zio->io_txg != 0 && /* not a delegated i/o */
2434 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2435 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2436 zio_vdev_io_bypass(zio);
2437 return (ZIO_PIPELINE_CONTINUE);
2440 if (vd->vdev_ops->vdev_op_leaf &&
2441 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2443 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2444 return (ZIO_PIPELINE_CONTINUE);
2446 if ((zio = vdev_queue_io(zio)) == NULL)
2447 return (ZIO_PIPELINE_STOP);
2449 if (!vdev_accessible(vd, zio)) {
2450 zio->io_error = ENXIO;
2452 return (ZIO_PIPELINE_STOP);
2456 return (vd->vdev_ops->vdev_op_io_start(zio));
2460 zio_vdev_io_done(zio_t *zio)
2462 vdev_t *vd = zio->io_vd;
2463 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2464 boolean_t unexpected_error = B_FALSE;
2466 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2467 return (ZIO_PIPELINE_STOP);
2469 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2471 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2473 vdev_queue_io_done(zio);
2475 if (zio->io_type == ZIO_TYPE_WRITE)
2476 vdev_cache_write(zio);
2478 if (zio_injection_enabled && zio->io_error == 0)
2479 zio->io_error = zio_handle_device_injection(vd,
2482 if (zio_injection_enabled && zio->io_error == 0)
2483 zio->io_error = zio_handle_label_injection(zio, EIO);
2485 if (zio->io_error) {
2486 if (!vdev_accessible(vd, zio)) {
2487 zio->io_error = ENXIO;
2489 unexpected_error = B_TRUE;
2494 ops->vdev_op_io_done(zio);
2496 if (unexpected_error)
2497 VERIFY(vdev_probe(vd, zio) == NULL);
2499 return (ZIO_PIPELINE_CONTINUE);
2503 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2504 * disk, and use that to finish the checksum ereport later.
2507 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2508 const void *good_buf)
2510 /* no processing needed */
2511 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2516 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2518 void *buf = zio_buf_alloc(zio->io_size);
2520 bcopy(zio->io_data, buf, zio->io_size);
2522 zcr->zcr_cbinfo = zio->io_size;
2523 zcr->zcr_cbdata = buf;
2524 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2525 zcr->zcr_free = zio_buf_free;
2529 zio_vdev_io_assess(zio_t *zio)
2531 vdev_t *vd = zio->io_vd;
2533 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2534 return (ZIO_PIPELINE_STOP);
2536 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2537 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2539 if (zio->io_vsd != NULL) {
2540 zio->io_vsd_ops->vsd_free(zio);
2544 if (zio_injection_enabled && zio->io_error == 0)
2545 zio->io_error = zio_handle_fault_injection(zio, EIO);
2548 * If the I/O failed, determine whether we should attempt to retry it.
2550 * On retry, we cut in line in the issue queue, since we don't want
2551 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2553 if (zio->io_error && vd == NULL &&
2554 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2555 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2556 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2558 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2559 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2560 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2561 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2562 zio_requeue_io_start_cut_in_line);
2563 return (ZIO_PIPELINE_STOP);
2567 * If we got an error on a leaf device, convert it to ENXIO
2568 * if the device is not accessible at all.
2570 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2571 !vdev_accessible(vd, zio))
2572 zio->io_error = ENXIO;
2575 * If we can't write to an interior vdev (mirror or RAID-Z),
2576 * set vdev_cant_write so that we stop trying to allocate from it.
2578 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2579 vd != NULL && !vd->vdev_ops->vdev_op_leaf)
2580 vd->vdev_cant_write = B_TRUE;
2583 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2585 return (ZIO_PIPELINE_CONTINUE);
2589 zio_vdev_io_reissue(zio_t *zio)
2591 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2592 ASSERT(zio->io_error == 0);
2594 zio->io_stage >>= 1;
2598 zio_vdev_io_redone(zio_t *zio)
2600 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2602 zio->io_stage >>= 1;
2606 zio_vdev_io_bypass(zio_t *zio)
2608 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2609 ASSERT(zio->io_error == 0);
2611 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2612 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2616 * ==========================================================================
2617 * Generate and verify checksums
2618 * ==========================================================================
2621 zio_checksum_generate(zio_t *zio)
2623 blkptr_t *bp = zio->io_bp;
2624 enum zio_checksum checksum;
2628 * This is zio_write_phys().
2629 * We're either generating a label checksum, or none at all.
2631 checksum = zio->io_prop.zp_checksum;
2633 if (checksum == ZIO_CHECKSUM_OFF)
2634 return (ZIO_PIPELINE_CONTINUE);
2636 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2638 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2639 ASSERT(!IO_IS_ALLOCATING(zio));
2640 checksum = ZIO_CHECKSUM_GANG_HEADER;
2642 checksum = BP_GET_CHECKSUM(bp);
2646 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2648 return (ZIO_PIPELINE_CONTINUE);
2652 zio_checksum_verify(zio_t *zio)
2654 zio_bad_cksum_t info;
2655 blkptr_t *bp = zio->io_bp;
2658 ASSERT(zio->io_vd != NULL);
2662 * This is zio_read_phys().
2663 * We're either verifying a label checksum, or nothing at all.
2665 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2666 return (ZIO_PIPELINE_CONTINUE);
2668 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2671 if ((error = zio_checksum_error(zio, &info)) != 0) {
2672 zio->io_error = error;
2673 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2674 zfs_ereport_start_checksum(zio->io_spa,
2675 zio->io_vd, zio, zio->io_offset,
2676 zio->io_size, NULL, &info);
2680 return (ZIO_PIPELINE_CONTINUE);
2684 * Called by RAID-Z to ensure we don't compute the checksum twice.
2687 zio_checksum_verified(zio_t *zio)
2689 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2693 * ==========================================================================
2694 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2695 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2696 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2697 * indicate errors that are specific to one I/O, and most likely permanent.
2698 * Any other error is presumed to be worse because we weren't expecting it.
2699 * ==========================================================================
2702 zio_worst_error(int e1, int e2)
2704 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2707 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2708 if (e1 == zio_error_rank[r1])
2711 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2712 if (e2 == zio_error_rank[r2])
2715 return (r1 > r2 ? e1 : e2);
2719 * ==========================================================================
2721 * ==========================================================================
2724 zio_ready(zio_t *zio)
2726 blkptr_t *bp = zio->io_bp;
2727 zio_t *pio, *pio_next;
2729 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2730 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2731 return (ZIO_PIPELINE_STOP);
2733 if (zio->io_ready) {
2734 ASSERT(IO_IS_ALLOCATING(zio));
2735 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2736 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2741 if (bp != NULL && bp != &zio->io_bp_copy)
2742 zio->io_bp_copy = *bp;
2745 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2747 mutex_enter(&zio->io_lock);
2748 zio->io_state[ZIO_WAIT_READY] = 1;
2749 pio = zio_walk_parents(zio);
2750 mutex_exit(&zio->io_lock);
2753 * As we notify zio's parents, new parents could be added.
2754 * New parents go to the head of zio's io_parent_list, however,
2755 * so we will (correctly) not notify them. The remainder of zio's
2756 * io_parent_list, from 'pio_next' onward, cannot change because
2757 * all parents must wait for us to be done before they can be done.
2759 for (; pio != NULL; pio = pio_next) {
2760 pio_next = zio_walk_parents(zio);
2761 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2764 if (zio->io_flags & ZIO_FLAG_NODATA) {
2765 if (BP_IS_GANG(bp)) {
2766 zio->io_flags &= ~ZIO_FLAG_NODATA;
2768 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2769 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2773 if (zio_injection_enabled &&
2774 zio->io_spa->spa_syncing_txg == zio->io_txg)
2775 zio_handle_ignored_writes(zio);
2777 return (ZIO_PIPELINE_CONTINUE);
2781 zio_done(zio_t *zio)
2783 zio_t *pio, *pio_next;
2787 * If our children haven't all completed,
2788 * wait for them and then repeat this pipeline stage.
2790 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2791 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2792 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2793 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2794 return (ZIO_PIPELINE_STOP);
2796 for (c = 0; c < ZIO_CHILD_TYPES; c++)
2797 for (w = 0; w < ZIO_WAIT_TYPES; w++)
2798 ASSERT(zio->io_children[c][w] == 0);
2800 if (zio->io_bp != NULL) {
2801 ASSERT(zio->io_bp->blk_pad[0] == 0);
2802 ASSERT(zio->io_bp->blk_pad[1] == 0);
2803 ASSERT(bcmp(zio->io_bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2804 (zio->io_bp == zio_unique_parent(zio)->io_bp));
2805 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) &&
2806 zio->io_bp_override == NULL &&
2807 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2808 ASSERT(!BP_SHOULD_BYTESWAP(zio->io_bp));
2809 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2810 ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 ||
2811 (BP_COUNT_GANG(zio->io_bp) == BP_GET_NDVAS(zio->io_bp)));
2816 * If there were child vdev/gang/ddt errors, they apply to us now.
2818 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2819 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2820 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2823 * If the I/O on the transformed data was successful, generate any
2824 * checksum reports now while we still have the transformed data.
2826 if (zio->io_error == 0) {
2827 while (zio->io_cksum_report != NULL) {
2828 zio_cksum_report_t *zcr = zio->io_cksum_report;
2829 uint64_t align = zcr->zcr_align;
2830 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2831 char *abuf = zio->io_data;
2833 if (asize != zio->io_size) {
2834 abuf = zio_buf_alloc(asize);
2835 bcopy(zio->io_data, abuf, zio->io_size);
2836 bzero(abuf + zio->io_size, asize - zio->io_size);
2839 zio->io_cksum_report = zcr->zcr_next;
2840 zcr->zcr_next = NULL;
2841 zcr->zcr_finish(zcr, abuf);
2842 zfs_ereport_free_checksum(zcr);
2844 if (asize != zio->io_size)
2845 zio_buf_free(abuf, asize);
2849 zio_pop_transforms(zio); /* note: may set zio->io_error */
2851 vdev_stat_update(zio, zio->io_size);
2854 * If this I/O is attached to a particular vdev is slow, exeeding
2855 * 30 seconds to complete, post an error described the I/O delay.
2856 * We ignore these errors if the device is currently unavailable.
2858 if (zio->io_delay >= zio_delay_max) {
2859 if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd))
2860 zfs_ereport_post(FM_EREPORT_ZFS_DELAY, zio->io_spa,
2861 zio->io_vd, zio, 0, 0);
2864 if (zio->io_error) {
2866 * If this I/O is attached to a particular vdev,
2867 * generate an error message describing the I/O failure
2868 * at the block level. We ignore these errors if the
2869 * device is currently unavailable.
2871 if (zio->io_error != ECKSUM && zio->io_vd != NULL &&
2872 !vdev_is_dead(zio->io_vd))
2873 zfs_ereport_post(FM_EREPORT_ZFS_IO, zio->io_spa,
2874 zio->io_vd, zio, 0, 0);
2876 if ((zio->io_error == EIO || !(zio->io_flags &
2877 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2878 zio == zio->io_logical) {
2880 * For logical I/O requests, tell the SPA to log the
2881 * error and generate a logical data ereport.
2883 spa_log_error(zio->io_spa, zio);
2884 zfs_ereport_post(FM_EREPORT_ZFS_DATA, zio->io_spa, NULL, zio,
2889 if (zio->io_error && zio == zio->io_logical) {
2891 * Determine whether zio should be reexecuted. This will
2892 * propagate all the way to the root via zio_notify_parent().
2894 ASSERT(zio->io_vd == NULL && zio->io_bp != NULL);
2895 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2897 if (IO_IS_ALLOCATING(zio) &&
2898 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2899 if (zio->io_error != ENOSPC)
2900 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2902 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2905 if ((zio->io_type == ZIO_TYPE_READ ||
2906 zio->io_type == ZIO_TYPE_FREE) &&
2907 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
2908 zio->io_error == ENXIO &&
2909 spa_load_state(zio->io_spa) == SPA_LOAD_NONE &&
2910 spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE)
2911 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2913 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2914 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2917 * Here is a possibly good place to attempt to do
2918 * either combinatorial reconstruction or error correction
2919 * based on checksums. It also might be a good place
2920 * to send out preliminary ereports before we suspend
2926 * If there were logical child errors, they apply to us now.
2927 * We defer this until now to avoid conflating logical child
2928 * errors with errors that happened to the zio itself when
2929 * updating vdev stats and reporting FMA events above.
2931 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2933 if ((zio->io_error || zio->io_reexecute) &&
2934 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
2935 !(zio->io_flags & ZIO_FLAG_IO_REWRITE))
2936 zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp);
2938 zio_gang_tree_free(&zio->io_gang_tree);
2941 * Godfather I/Os should never suspend.
2943 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
2944 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
2945 zio->io_reexecute = 0;
2947 if (zio->io_reexecute) {
2949 * This is a logical I/O that wants to reexecute.
2951 * Reexecute is top-down. When an i/o fails, if it's not
2952 * the root, it simply notifies its parent and sticks around.
2953 * The parent, seeing that it still has children in zio_done(),
2954 * does the same. This percolates all the way up to the root.
2955 * The root i/o will reexecute or suspend the entire tree.
2957 * This approach ensures that zio_reexecute() honors
2958 * all the original i/o dependency relationships, e.g.
2959 * parents not executing until children are ready.
2961 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2963 zio->io_gang_leader = NULL;
2965 mutex_enter(&zio->io_lock);
2966 zio->io_state[ZIO_WAIT_DONE] = 1;
2967 mutex_exit(&zio->io_lock);
2970 * "The Godfather" I/O monitors its children but is
2971 * not a true parent to them. It will track them through
2972 * the pipeline but severs its ties whenever they get into
2973 * trouble (e.g. suspended). This allows "The Godfather"
2974 * I/O to return status without blocking.
2976 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2977 zio_link_t *zl = zio->io_walk_link;
2978 pio_next = zio_walk_parents(zio);
2980 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
2981 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
2982 zio_remove_child(pio, zio, zl);
2983 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2987 if ((pio = zio_unique_parent(zio)) != NULL) {
2989 * We're not a root i/o, so there's nothing to do
2990 * but notify our parent. Don't propagate errors
2991 * upward since we haven't permanently failed yet.
2993 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2994 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
2995 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2996 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
2998 * We'd fail again if we reexecuted now, so suspend
2999 * until conditions improve (e.g. device comes online).
3001 zio_suspend(zio->io_spa, zio);
3004 * Reexecution is potentially a huge amount of work.
3005 * Hand it off to the otherwise-unused claim taskq.
3007 ASSERT(taskq_empty_ent(&zio->io_tqent));
3008 (void) taskq_dispatch_ent(
3009 zio->io_spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
3010 (task_func_t *)zio_reexecute, zio, 0,
3013 return (ZIO_PIPELINE_STOP);
3016 ASSERT(zio->io_child_count == 0);
3017 ASSERT(zio->io_reexecute == 0);
3018 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3021 * Report any checksum errors, since the I/O is complete.
3023 while (zio->io_cksum_report != NULL) {
3024 zio_cksum_report_t *zcr = zio->io_cksum_report;
3025 zio->io_cksum_report = zcr->zcr_next;
3026 zcr->zcr_next = NULL;
3027 zcr->zcr_finish(zcr, NULL);
3028 zfs_ereport_free_checksum(zcr);
3032 * It is the responsibility of the done callback to ensure that this
3033 * particular zio is no longer discoverable for adoption, and as
3034 * such, cannot acquire any new parents.
3039 mutex_enter(&zio->io_lock);
3040 zio->io_state[ZIO_WAIT_DONE] = 1;
3041 mutex_exit(&zio->io_lock);
3043 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3044 zio_link_t *zl = zio->io_walk_link;
3045 pio_next = zio_walk_parents(zio);
3046 zio_remove_child(pio, zio, zl);
3047 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3050 if (zio->io_waiter != NULL) {
3051 mutex_enter(&zio->io_lock);
3052 zio->io_executor = NULL;
3053 cv_broadcast(&zio->io_cv);
3054 mutex_exit(&zio->io_lock);
3059 return (ZIO_PIPELINE_STOP);
3063 * ==========================================================================
3064 * I/O pipeline definition
3065 * ==========================================================================
3067 static zio_pipe_stage_t *zio_pipeline[] = {
3073 zio_checksum_generate,
3087 zio_checksum_verify,
3091 #if defined(_KERNEL) && defined(HAVE_SPL)
3092 /* Fault injection */
3093 EXPORT_SYMBOL(zio_injection_enabled);
3094 EXPORT_SYMBOL(zio_inject_fault);
3095 EXPORT_SYMBOL(zio_inject_list_next);
3096 EXPORT_SYMBOL(zio_clear_fault);
3097 EXPORT_SYMBOL(zio_handle_fault_injection);
3098 EXPORT_SYMBOL(zio_handle_device_injection);
3099 EXPORT_SYMBOL(zio_handle_label_injection);
3100 EXPORT_SYMBOL(zio_priority_table);
3101 EXPORT_SYMBOL(zio_type_name);
3103 module_param(zio_bulk_flags, int, 0644);
3104 MODULE_PARM_DESC(zio_bulk_flags, "Additional flags to pass to bulk buffers");
3106 module_param(zio_delay_max, int, 0644);
3107 MODULE_PARM_DESC(zio_delay_max, "Max zio millisec delay before posting event");
3109 module_param(zio_requeue_io_start_cut_in_line, int, 0644);
3110 MODULE_PARM_DESC(zio_requeue_io_start_cut_in_line, "Prioritize requeued I/O");