4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2012 by Delphix. All rights reserved.
26 #include <sys/zfs_context.h>
27 #include <sys/fm/fs/zfs.h>
30 #include <sys/spa_impl.h>
31 #include <sys/vdev_impl.h>
32 #include <sys/zio_impl.h>
33 #include <sys/zio_compress.h>
34 #include <sys/zio_checksum.h>
35 #include <sys/dmu_objset.h>
39 SYSCTL_DECL(_vfs_zfs);
40 SYSCTL_NODE(_vfs_zfs, OID_AUTO, zio, CTLFLAG_RW, 0, "ZFS ZIO");
41 static int zio_use_uma = 0;
42 TUNABLE_INT("vfs.zfs.zio.use_uma", &zio_use_uma);
43 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, use_uma, CTLFLAG_RDTUN, &zio_use_uma, 0,
44 "Use uma(9) for ZIO allocations");
47 * ==========================================================================
49 * ==========================================================================
51 uint8_t zio_priority_table[ZIO_PRIORITY_TABLE_SIZE] = {
52 0, /* ZIO_PRIORITY_NOW */
53 0, /* ZIO_PRIORITY_SYNC_READ */
54 0, /* ZIO_PRIORITY_SYNC_WRITE */
55 0, /* ZIO_PRIORITY_LOG_WRITE */
56 1, /* ZIO_PRIORITY_CACHE_FILL */
57 1, /* ZIO_PRIORITY_AGG */
58 4, /* ZIO_PRIORITY_FREE */
59 4, /* ZIO_PRIORITY_ASYNC_WRITE */
60 6, /* ZIO_PRIORITY_ASYNC_READ */
61 10, /* ZIO_PRIORITY_RESILVER */
62 20, /* ZIO_PRIORITY_SCRUB */
63 2, /* ZIO_PRIORITY_DDT_PREFETCH */
67 * ==========================================================================
68 * I/O type descriptions
69 * ==========================================================================
71 char *zio_type_name[ZIO_TYPES] = {
72 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
77 * ==========================================================================
79 * ==========================================================================
81 kmem_cache_t *zio_cache;
82 kmem_cache_t *zio_link_cache;
83 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
84 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
87 extern vmem_t *zio_alloc_arena;
89 extern int zfs_mg_alloc_failures;
92 * An allocating zio is one that either currently has the DVA allocate
93 * stage set or will have it later in its lifetime.
95 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
97 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
100 int zio_buf_debug_limit = 16384;
102 int zio_buf_debug_limit = 0;
109 zio_cache = kmem_cache_create("zio_cache",
110 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
111 zio_link_cache = kmem_cache_create("zio_link_cache",
112 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
115 * For small buffers, we want a cache for each multiple of
116 * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache
117 * for each quarter-power of 2. For large buffers, we want
118 * a cache for each multiple of PAGESIZE.
120 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
121 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
124 size_t cflags = (size > zio_buf_debug_limit) ? (KMC_NODEBUG|KMC_NOTOUCH) : 0;
126 while (p2 & (p2 - 1))
132 * If we are using watchpoints, put each buffer on its own page,
133 * to eliminate the performance overhead of trapping to the
134 * kernel when modifying a non-watched buffer that shares the
135 * page with a watched buffer.
137 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
141 if (size <= 4 * SPA_MINBLOCKSIZE) {
142 align = SPA_MINBLOCKSIZE;
143 } else if (IS_P2ALIGNED(size, PAGESIZE)) {
145 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
151 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
152 zio_buf_cache[c] = kmem_cache_create(name, size,
153 align, NULL, NULL, NULL, NULL, NULL, cflags);
156 * Since zio_data bufs do not appear in crash dumps, we
157 * pass KMC_NOTOUCH so that no allocator metadata is
158 * stored with the buffers.
160 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
161 zio_data_buf_cache[c] = kmem_cache_create(name, size,
162 align, NULL, NULL, NULL, NULL, NULL,
163 cflags | KMC_NOTOUCH);
168 ASSERT(zio_buf_cache[c] != NULL);
169 if (zio_buf_cache[c - 1] == NULL)
170 zio_buf_cache[c - 1] = zio_buf_cache[c];
172 ASSERT(zio_data_buf_cache[c] != NULL);
173 if (zio_data_buf_cache[c - 1] == NULL)
174 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
178 * The zio write taskqs have 1 thread per cpu, allow 1/2 of the taskqs
179 * to fail 3 times per txg or 8 failures, whichever is greater.
181 if (zfs_mg_alloc_failures == 0)
182 zfs_mg_alloc_failures = MAX((3 * max_ncpus / 2), 8);
183 else if (zfs_mg_alloc_failures < 8)
184 zfs_mg_alloc_failures = 8;
193 kmem_cache_t *last_cache = NULL;
194 kmem_cache_t *last_data_cache = NULL;
196 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
197 if (zio_buf_cache[c] != last_cache) {
198 last_cache = zio_buf_cache[c];
199 kmem_cache_destroy(zio_buf_cache[c]);
201 zio_buf_cache[c] = NULL;
203 if (zio_data_buf_cache[c] != last_data_cache) {
204 last_data_cache = zio_data_buf_cache[c];
205 kmem_cache_destroy(zio_data_buf_cache[c]);
207 zio_data_buf_cache[c] = NULL;
210 kmem_cache_destroy(zio_link_cache);
211 kmem_cache_destroy(zio_cache);
217 * ==========================================================================
218 * Allocate and free I/O buffers
219 * ==========================================================================
223 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
224 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
225 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
226 * excess / transient data in-core during a crashdump.
229 zio_buf_alloc(size_t size)
231 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
233 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
236 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
238 return (kmem_alloc(size, KM_SLEEP));
242 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
243 * crashdump if the kernel panics. This exists so that we will limit the amount
244 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
245 * of kernel heap dumped to disk when the kernel panics)
248 zio_data_buf_alloc(size_t size)
250 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
252 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
255 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
257 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG));
261 zio_buf_free(void *buf, size_t size)
263 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
265 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
268 kmem_cache_free(zio_buf_cache[c], buf);
270 kmem_free(buf, size);
274 zio_data_buf_free(void *buf, size_t size)
276 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
278 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
281 kmem_cache_free(zio_data_buf_cache[c], buf);
283 kmem_free(buf, size);
287 * ==========================================================================
288 * Push and pop I/O transform buffers
289 * ==========================================================================
292 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
293 zio_transform_func_t *transform)
295 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
297 zt->zt_orig_data = zio->io_data;
298 zt->zt_orig_size = zio->io_size;
299 zt->zt_bufsize = bufsize;
300 zt->zt_transform = transform;
302 zt->zt_next = zio->io_transform_stack;
303 zio->io_transform_stack = zt;
310 zio_pop_transforms(zio_t *zio)
314 while ((zt = zio->io_transform_stack) != NULL) {
315 if (zt->zt_transform != NULL)
316 zt->zt_transform(zio,
317 zt->zt_orig_data, zt->zt_orig_size);
319 if (zt->zt_bufsize != 0)
320 zio_buf_free(zio->io_data, zt->zt_bufsize);
322 zio->io_data = zt->zt_orig_data;
323 zio->io_size = zt->zt_orig_size;
324 zio->io_transform_stack = zt->zt_next;
326 kmem_free(zt, sizeof (zio_transform_t));
331 * ==========================================================================
332 * I/O transform callbacks for subblocks and decompression
333 * ==========================================================================
336 zio_subblock(zio_t *zio, void *data, uint64_t size)
338 ASSERT(zio->io_size > size);
340 if (zio->io_type == ZIO_TYPE_READ)
341 bcopy(zio->io_data, data, size);
345 zio_decompress(zio_t *zio, void *data, uint64_t size)
347 if (zio->io_error == 0 &&
348 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
349 zio->io_data, data, zio->io_size, size) != 0)
354 * ==========================================================================
355 * I/O parent/child relationships and pipeline interlocks
356 * ==========================================================================
359 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
360 * continue calling these functions until they return NULL.
361 * Otherwise, the next caller will pick up the list walk in
362 * some indeterminate state. (Otherwise every caller would
363 * have to pass in a cookie to keep the state represented by
364 * io_walk_link, which gets annoying.)
367 zio_walk_parents(zio_t *cio)
369 zio_link_t *zl = cio->io_walk_link;
370 list_t *pl = &cio->io_parent_list;
372 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
373 cio->io_walk_link = zl;
378 ASSERT(zl->zl_child == cio);
379 return (zl->zl_parent);
383 zio_walk_children(zio_t *pio)
385 zio_link_t *zl = pio->io_walk_link;
386 list_t *cl = &pio->io_child_list;
388 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
389 pio->io_walk_link = zl;
394 ASSERT(zl->zl_parent == pio);
395 return (zl->zl_child);
399 zio_unique_parent(zio_t *cio)
401 zio_t *pio = zio_walk_parents(cio);
403 VERIFY(zio_walk_parents(cio) == NULL);
408 zio_add_child(zio_t *pio, zio_t *cio)
410 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
413 * Logical I/Os can have logical, gang, or vdev children.
414 * Gang I/Os can have gang or vdev children.
415 * Vdev I/Os can only have vdev children.
416 * The following ASSERT captures all of these constraints.
418 ASSERT(cio->io_child_type <= pio->io_child_type);
423 mutex_enter(&cio->io_lock);
424 mutex_enter(&pio->io_lock);
426 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
428 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
429 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
431 list_insert_head(&pio->io_child_list, zl);
432 list_insert_head(&cio->io_parent_list, zl);
434 pio->io_child_count++;
435 cio->io_parent_count++;
437 mutex_exit(&pio->io_lock);
438 mutex_exit(&cio->io_lock);
442 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
444 ASSERT(zl->zl_parent == pio);
445 ASSERT(zl->zl_child == cio);
447 mutex_enter(&cio->io_lock);
448 mutex_enter(&pio->io_lock);
450 list_remove(&pio->io_child_list, zl);
451 list_remove(&cio->io_parent_list, zl);
453 pio->io_child_count--;
454 cio->io_parent_count--;
456 mutex_exit(&pio->io_lock);
457 mutex_exit(&cio->io_lock);
459 kmem_cache_free(zio_link_cache, zl);
463 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
465 uint64_t *countp = &zio->io_children[child][wait];
466 boolean_t waiting = B_FALSE;
468 mutex_enter(&zio->io_lock);
469 ASSERT(zio->io_stall == NULL);
472 zio->io_stall = countp;
475 mutex_exit(&zio->io_lock);
481 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
483 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
484 int *errorp = &pio->io_child_error[zio->io_child_type];
486 mutex_enter(&pio->io_lock);
487 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
488 *errorp = zio_worst_error(*errorp, zio->io_error);
489 pio->io_reexecute |= zio->io_reexecute;
490 ASSERT3U(*countp, >, 0);
491 if (--*countp == 0 && pio->io_stall == countp) {
492 pio->io_stall = NULL;
493 mutex_exit(&pio->io_lock);
496 mutex_exit(&pio->io_lock);
501 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
503 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
504 zio->io_error = zio->io_child_error[c];
508 * ==========================================================================
509 * Create the various types of I/O (read, write, free, etc)
510 * ==========================================================================
513 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
514 void *data, uint64_t size, zio_done_func_t *done, void *private,
515 zio_type_t type, int priority, enum zio_flag flags,
516 vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
517 enum zio_stage stage, enum zio_stage pipeline)
521 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
522 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
523 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
525 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
526 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
527 ASSERT(vd || stage == ZIO_STAGE_OPEN);
529 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
530 bzero(zio, sizeof (zio_t));
532 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
533 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
535 list_create(&zio->io_parent_list, sizeof (zio_link_t),
536 offsetof(zio_link_t, zl_parent_node));
537 list_create(&zio->io_child_list, sizeof (zio_link_t),
538 offsetof(zio_link_t, zl_child_node));
541 zio->io_child_type = ZIO_CHILD_VDEV;
542 else if (flags & ZIO_FLAG_GANG_CHILD)
543 zio->io_child_type = ZIO_CHILD_GANG;
544 else if (flags & ZIO_FLAG_DDT_CHILD)
545 zio->io_child_type = ZIO_CHILD_DDT;
547 zio->io_child_type = ZIO_CHILD_LOGICAL;
550 zio->io_bp = (blkptr_t *)bp;
551 zio->io_bp_copy = *bp;
552 zio->io_bp_orig = *bp;
553 if (type != ZIO_TYPE_WRITE ||
554 zio->io_child_type == ZIO_CHILD_DDT)
555 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
556 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
557 zio->io_logical = zio;
558 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
559 pipeline |= ZIO_GANG_STAGES;
565 zio->io_private = private;
567 zio->io_priority = priority;
569 zio->io_offset = offset;
570 zio->io_orig_data = zio->io_data = data;
571 zio->io_orig_size = zio->io_size = size;
572 zio->io_orig_flags = zio->io_flags = flags;
573 zio->io_orig_stage = zio->io_stage = stage;
574 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
576 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
577 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
580 zio->io_bookmark = *zb;
583 if (zio->io_logical == NULL)
584 zio->io_logical = pio->io_logical;
585 if (zio->io_child_type == ZIO_CHILD_GANG)
586 zio->io_gang_leader = pio->io_gang_leader;
587 zio_add_child(pio, zio);
594 zio_destroy(zio_t *zio)
596 list_destroy(&zio->io_parent_list);
597 list_destroy(&zio->io_child_list);
598 mutex_destroy(&zio->io_lock);
599 cv_destroy(&zio->io_cv);
600 kmem_cache_free(zio_cache, zio);
604 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
605 void *private, enum zio_flag flags)
609 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
610 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
611 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
617 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
619 return (zio_null(NULL, spa, NULL, done, private, flags));
623 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
624 void *data, uint64_t size, zio_done_func_t *done, void *private,
625 int priority, enum zio_flag flags, const zbookmark_t *zb)
629 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
630 data, size, done, private,
631 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
632 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
633 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
639 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
640 void *data, uint64_t size, const zio_prop_t *zp,
641 zio_done_func_t *ready, zio_done_func_t *done, void *private,
642 int priority, enum zio_flag flags, const zbookmark_t *zb)
646 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
647 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
648 zp->zp_compress >= ZIO_COMPRESS_OFF &&
649 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
650 DMU_OT_IS_VALID(zp->zp_type) &&
653 zp->zp_copies <= spa_max_replication(spa) &&
655 zp->zp_dedup_verify <= 1);
657 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
658 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
659 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
660 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
662 zio->io_ready = ready;
669 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
670 uint64_t size, zio_done_func_t *done, void *private, int priority,
671 enum zio_flag flags, zbookmark_t *zb)
675 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
676 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
677 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
683 zio_write_override(zio_t *zio, blkptr_t *bp, int copies)
685 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
686 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
687 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
688 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
690 zio->io_prop.zp_copies = copies;
691 zio->io_bp_override = bp;
695 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
697 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
701 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
706 dprintf_bp(bp, "freeing in txg %llu, pass %u",
707 (longlong_t)txg, spa->spa_sync_pass);
709 ASSERT(!BP_IS_HOLE(bp));
710 ASSERT(spa_syncing_txg(spa) == txg);
711 ASSERT(spa_sync_pass(spa) <= SYNC_PASS_DEFERRED_FREE);
713 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
714 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
715 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
721 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
722 zio_done_func_t *done, void *private, enum zio_flag flags)
727 * A claim is an allocation of a specific block. Claims are needed
728 * to support immediate writes in the intent log. The issue is that
729 * immediate writes contain committed data, but in a txg that was
730 * *not* committed. Upon opening the pool after an unclean shutdown,
731 * the intent log claims all blocks that contain immediate write data
732 * so that the SPA knows they're in use.
734 * All claims *must* be resolved in the first txg -- before the SPA
735 * starts allocating blocks -- so that nothing is allocated twice.
736 * If txg == 0 we just verify that the block is claimable.
738 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
739 ASSERT(txg == spa_first_txg(spa) || txg == 0);
740 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
742 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
743 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
744 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
750 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
751 zio_done_func_t *done, void *private, int priority, enum zio_flag flags)
756 if (vd->vdev_children == 0) {
757 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
758 ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL,
759 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
763 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
765 for (c = 0; c < vd->vdev_children; c++)
766 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
767 done, private, priority, flags));
774 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
775 void *data, int checksum, zio_done_func_t *done, void *private,
776 int priority, enum zio_flag flags, boolean_t labels)
780 ASSERT(vd->vdev_children == 0);
781 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
782 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
783 ASSERT3U(offset + size, <=, vd->vdev_psize);
785 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
786 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
787 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
789 zio->io_prop.zp_checksum = checksum;
795 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
796 void *data, int checksum, zio_done_func_t *done, void *private,
797 int priority, enum zio_flag flags, boolean_t labels)
801 ASSERT(vd->vdev_children == 0);
802 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
803 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
804 ASSERT3U(offset + size, <=, vd->vdev_psize);
806 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
807 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
808 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
810 zio->io_prop.zp_checksum = checksum;
812 if (zio_checksum_table[checksum].ci_eck) {
814 * zec checksums are necessarily destructive -- they modify
815 * the end of the write buffer to hold the verifier/checksum.
816 * Therefore, we must make a local copy in case the data is
817 * being written to multiple places in parallel.
819 void *wbuf = zio_buf_alloc(size);
820 bcopy(data, wbuf, size);
821 zio_push_transform(zio, wbuf, size, size, NULL);
828 * Create a child I/O to do some work for us.
831 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
832 void *data, uint64_t size, int type, int priority, enum zio_flag flags,
833 zio_done_func_t *done, void *private)
835 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
838 ASSERT(vd->vdev_parent ==
839 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
841 if (type == ZIO_TYPE_READ && bp != NULL) {
843 * If we have the bp, then the child should perform the
844 * checksum and the parent need not. This pushes error
845 * detection as close to the leaves as possible and
846 * eliminates redundant checksums in the interior nodes.
848 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
849 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
852 if (vd->vdev_children == 0)
853 offset += VDEV_LABEL_START_SIZE;
855 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
858 * If we've decided to do a repair, the write is not speculative --
859 * even if the original read was.
861 if (flags & ZIO_FLAG_IO_REPAIR)
862 flags &= ~ZIO_FLAG_SPECULATIVE;
864 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
865 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
866 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
872 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
873 int type, int priority, enum zio_flag flags,
874 zio_done_func_t *done, void *private)
878 ASSERT(vd->vdev_ops->vdev_op_leaf);
880 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
881 data, size, done, private, type, priority,
882 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
884 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
890 zio_flush(zio_t *zio, vdev_t *vd)
892 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
893 NULL, NULL, ZIO_PRIORITY_NOW,
894 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
898 zio_shrink(zio_t *zio, uint64_t size)
900 ASSERT(zio->io_executor == NULL);
901 ASSERT(zio->io_orig_size == zio->io_size);
902 ASSERT(size <= zio->io_size);
905 * We don't shrink for raidz because of problems with the
906 * reconstruction when reading back less than the block size.
907 * Note, BP_IS_RAIDZ() assumes no compression.
909 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
910 if (!BP_IS_RAIDZ(zio->io_bp))
911 zio->io_orig_size = zio->io_size = size;
915 * ==========================================================================
916 * Prepare to read and write logical blocks
917 * ==========================================================================
921 zio_read_bp_init(zio_t *zio)
923 blkptr_t *bp = zio->io_bp;
925 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
926 zio->io_child_type == ZIO_CHILD_LOGICAL &&
927 !(zio->io_flags & ZIO_FLAG_RAW)) {
928 uint64_t psize = BP_GET_PSIZE(bp);
929 void *cbuf = zio_buf_alloc(psize);
931 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
934 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
935 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
937 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
938 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
940 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
941 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
943 return (ZIO_PIPELINE_CONTINUE);
947 zio_write_bp_init(zio_t *zio)
949 spa_t *spa = zio->io_spa;
950 zio_prop_t *zp = &zio->io_prop;
951 enum zio_compress compress = zp->zp_compress;
952 blkptr_t *bp = zio->io_bp;
953 uint64_t lsize = zio->io_size;
954 uint64_t psize = lsize;
958 * If our children haven't all reached the ready stage,
959 * wait for them and then repeat this pipeline stage.
961 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
962 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
963 return (ZIO_PIPELINE_STOP);
965 if (!IO_IS_ALLOCATING(zio))
966 return (ZIO_PIPELINE_CONTINUE);
968 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
970 if (zio->io_bp_override) {
971 ASSERT(bp->blk_birth != zio->io_txg);
972 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
974 *bp = *zio->io_bp_override;
975 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
977 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
978 return (ZIO_PIPELINE_CONTINUE);
980 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
981 zp->zp_dedup_verify);
983 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
985 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
986 return (ZIO_PIPELINE_CONTINUE);
988 zio->io_bp_override = NULL;
992 if (bp->blk_birth == zio->io_txg) {
994 * We're rewriting an existing block, which means we're
995 * working on behalf of spa_sync(). For spa_sync() to
996 * converge, it must eventually be the case that we don't
997 * have to allocate new blocks. But compression changes
998 * the blocksize, which forces a reallocate, and makes
999 * convergence take longer. Therefore, after the first
1000 * few passes, stop compressing to ensure convergence.
1002 pass = spa_sync_pass(spa);
1004 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1005 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1006 ASSERT(!BP_GET_DEDUP(bp));
1008 if (pass > SYNC_PASS_DONT_COMPRESS)
1009 compress = ZIO_COMPRESS_OFF;
1011 /* Make sure someone doesn't change their mind on overwrites */
1012 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
1013 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1016 if (compress != ZIO_COMPRESS_OFF) {
1017 void *cbuf = zio_buf_alloc(lsize);
1018 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1019 if (psize == 0 || psize == lsize) {
1020 compress = ZIO_COMPRESS_OFF;
1021 zio_buf_free(cbuf, lsize);
1023 ASSERT(psize < lsize);
1024 zio_push_transform(zio, cbuf, psize, lsize, NULL);
1029 * The final pass of spa_sync() must be all rewrites, but the first
1030 * few passes offer a trade-off: allocating blocks defers convergence,
1031 * but newly allocated blocks are sequential, so they can be written
1032 * to disk faster. Therefore, we allow the first few passes of
1033 * spa_sync() to allocate new blocks, but force rewrites after that.
1034 * There should only be a handful of blocks after pass 1 in any case.
1036 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
1037 pass > SYNC_PASS_REWRITE) {
1039 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1040 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1041 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1044 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1048 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1050 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1051 BP_SET_LSIZE(bp, lsize);
1052 BP_SET_PSIZE(bp, psize);
1053 BP_SET_COMPRESS(bp, compress);
1054 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1055 BP_SET_TYPE(bp, zp->zp_type);
1056 BP_SET_LEVEL(bp, zp->zp_level);
1057 BP_SET_DEDUP(bp, zp->zp_dedup);
1058 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1060 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1061 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1062 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1066 return (ZIO_PIPELINE_CONTINUE);
1070 zio_free_bp_init(zio_t *zio)
1072 blkptr_t *bp = zio->io_bp;
1074 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1075 if (BP_GET_DEDUP(bp))
1076 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1079 return (ZIO_PIPELINE_CONTINUE);
1083 * ==========================================================================
1084 * Execute the I/O pipeline
1085 * ==========================================================================
1089 zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q, boolean_t cutinline)
1091 spa_t *spa = zio->io_spa;
1092 zio_type_t t = zio->io_type;
1093 int flags = TQ_SLEEP | (cutinline ? TQ_FRONT : 0);
1095 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1098 * If we're a config writer or a probe, the normal issue and
1099 * interrupt threads may all be blocked waiting for the config lock.
1100 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1102 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1106 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1108 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1112 * If this is a high priority I/O, then use the high priority taskq.
1114 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1115 spa->spa_zio_taskq[t][q + 1] != NULL)
1118 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1120 (void) taskq_dispatch_safe(spa->spa_zio_taskq[t][q],
1121 (task_func_t *)zio_execute, zio, flags, &zio->io_task);
1123 (void) taskq_dispatch(spa->spa_zio_taskq[t][q],
1124 (task_func_t *)zio_execute, zio, flags);
1129 zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
1131 kthread_t *executor = zio->io_executor;
1132 spa_t *spa = zio->io_spa;
1134 for (zio_type_t t = 0; t < ZIO_TYPES; t++)
1135 if (taskq_member(spa->spa_zio_taskq[t][q], executor))
1142 zio_issue_async(zio_t *zio)
1144 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1146 return (ZIO_PIPELINE_STOP);
1150 zio_interrupt(zio_t *zio)
1152 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1156 * Execute the I/O pipeline until one of the following occurs:
1157 * (1) the I/O completes; (2) the pipeline stalls waiting for
1158 * dependent child I/Os; (3) the I/O issues, so we're waiting
1159 * for an I/O completion interrupt; (4) the I/O is delegated by
1160 * vdev-level caching or aggregation; (5) the I/O is deferred
1161 * due to vdev-level queueing; (6) the I/O is handed off to
1162 * another thread. In all cases, the pipeline stops whenever
1163 * there's no CPU work; it never burns a thread in cv_wait().
1165 * There's no locking on io_stage because there's no legitimate way
1166 * for multiple threads to be attempting to process the same I/O.
1168 static zio_pipe_stage_t *zio_pipeline[];
1171 zio_execute(zio_t *zio)
1173 zio->io_executor = curthread;
1175 while (zio->io_stage < ZIO_STAGE_DONE) {
1176 enum zio_stage pipeline = zio->io_pipeline;
1177 enum zio_stage stage = zio->io_stage;
1180 ASSERT(!MUTEX_HELD(&zio->io_lock));
1181 ASSERT(ISP2(stage));
1182 ASSERT(zio->io_stall == NULL);
1186 } while ((stage & pipeline) == 0);
1188 ASSERT(stage <= ZIO_STAGE_DONE);
1191 * If we are in interrupt context and this pipeline stage
1192 * will grab a config lock that is held across I/O,
1193 * or may wait for an I/O that needs an interrupt thread
1194 * to complete, issue async to avoid deadlock.
1196 * For VDEV_IO_START, we cut in line so that the io will
1197 * be sent to disk promptly.
1199 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1200 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1201 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1202 zio_requeue_io_start_cut_in_line : B_FALSE;
1203 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1207 zio->io_stage = stage;
1208 rv = zio_pipeline[highbit(stage) - 1](zio);
1210 if (rv == ZIO_PIPELINE_STOP)
1213 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1218 * ==========================================================================
1219 * Initiate I/O, either sync or async
1220 * ==========================================================================
1223 zio_wait(zio_t *zio)
1227 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1228 ASSERT(zio->io_executor == NULL);
1230 zio->io_waiter = curthread;
1234 mutex_enter(&zio->io_lock);
1235 while (zio->io_executor != NULL)
1236 cv_wait(&zio->io_cv, &zio->io_lock);
1237 mutex_exit(&zio->io_lock);
1239 error = zio->io_error;
1246 zio_nowait(zio_t *zio)
1248 ASSERT(zio->io_executor == NULL);
1250 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1251 zio_unique_parent(zio) == NULL) {
1253 * This is a logical async I/O with no parent to wait for it.
1254 * We add it to the spa_async_root_zio "Godfather" I/O which
1255 * will ensure they complete prior to unloading the pool.
1257 spa_t *spa = zio->io_spa;
1259 zio_add_child(spa->spa_async_zio_root, zio);
1266 * ==========================================================================
1267 * Reexecute or suspend/resume failed I/O
1268 * ==========================================================================
1272 zio_reexecute(zio_t *pio)
1274 zio_t *cio, *cio_next;
1276 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1277 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1278 ASSERT(pio->io_gang_leader == NULL);
1279 ASSERT(pio->io_gang_tree == NULL);
1281 pio->io_flags = pio->io_orig_flags;
1282 pio->io_stage = pio->io_orig_stage;
1283 pio->io_pipeline = pio->io_orig_pipeline;
1284 pio->io_reexecute = 0;
1286 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1287 pio->io_state[w] = 0;
1288 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1289 pio->io_child_error[c] = 0;
1291 if (IO_IS_ALLOCATING(pio))
1292 BP_ZERO(pio->io_bp);
1295 * As we reexecute pio's children, new children could be created.
1296 * New children go to the head of pio's io_child_list, however,
1297 * so we will (correctly) not reexecute them. The key is that
1298 * the remainder of pio's io_child_list, from 'cio_next' onward,
1299 * cannot be affected by any side effects of reexecuting 'cio'.
1301 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1302 cio_next = zio_walk_children(pio);
1303 mutex_enter(&pio->io_lock);
1304 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1305 pio->io_children[cio->io_child_type][w]++;
1306 mutex_exit(&pio->io_lock);
1311 * Now that all children have been reexecuted, execute the parent.
1312 * We don't reexecute "The Godfather" I/O here as it's the
1313 * responsibility of the caller to wait on him.
1315 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1320 zio_suspend(spa_t *spa, zio_t *zio)
1322 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1323 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1324 "failure and the failure mode property for this pool "
1325 "is set to panic.", spa_name(spa));
1327 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1329 mutex_enter(&spa->spa_suspend_lock);
1331 if (spa->spa_suspend_zio_root == NULL)
1332 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1333 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1334 ZIO_FLAG_GODFATHER);
1336 spa->spa_suspended = B_TRUE;
1339 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1340 ASSERT(zio != spa->spa_suspend_zio_root);
1341 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1342 ASSERT(zio_unique_parent(zio) == NULL);
1343 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1344 zio_add_child(spa->spa_suspend_zio_root, zio);
1347 mutex_exit(&spa->spa_suspend_lock);
1351 zio_resume(spa_t *spa)
1356 * Reexecute all previously suspended i/o.
1358 mutex_enter(&spa->spa_suspend_lock);
1359 spa->spa_suspended = B_FALSE;
1360 cv_broadcast(&spa->spa_suspend_cv);
1361 pio = spa->spa_suspend_zio_root;
1362 spa->spa_suspend_zio_root = NULL;
1363 mutex_exit(&spa->spa_suspend_lock);
1369 return (zio_wait(pio));
1373 zio_resume_wait(spa_t *spa)
1375 mutex_enter(&spa->spa_suspend_lock);
1376 while (spa_suspended(spa))
1377 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1378 mutex_exit(&spa->spa_suspend_lock);
1382 * ==========================================================================
1385 * A gang block is a collection of small blocks that looks to the DMU
1386 * like one large block. When zio_dva_allocate() cannot find a block
1387 * of the requested size, due to either severe fragmentation or the pool
1388 * being nearly full, it calls zio_write_gang_block() to construct the
1389 * block from smaller fragments.
1391 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1392 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1393 * an indirect block: it's an array of block pointers. It consumes
1394 * only one sector and hence is allocatable regardless of fragmentation.
1395 * The gang header's bps point to its gang members, which hold the data.
1397 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1398 * as the verifier to ensure uniqueness of the SHA256 checksum.
1399 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1400 * not the gang header. This ensures that data block signatures (needed for
1401 * deduplication) are independent of how the block is physically stored.
1403 * Gang blocks can be nested: a gang member may itself be a gang block.
1404 * Thus every gang block is a tree in which root and all interior nodes are
1405 * gang headers, and the leaves are normal blocks that contain user data.
1406 * The root of the gang tree is called the gang leader.
1408 * To perform any operation (read, rewrite, free, claim) on a gang block,
1409 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1410 * in the io_gang_tree field of the original logical i/o by recursively
1411 * reading the gang leader and all gang headers below it. This yields
1412 * an in-core tree containing the contents of every gang header and the
1413 * bps for every constituent of the gang block.
1415 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1416 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1417 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1418 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1419 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1420 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1421 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1422 * of the gang header plus zio_checksum_compute() of the data to update the
1423 * gang header's blk_cksum as described above.
1425 * The two-phase assemble/issue model solves the problem of partial failure --
1426 * what if you'd freed part of a gang block but then couldn't read the
1427 * gang header for another part? Assembling the entire gang tree first
1428 * ensures that all the necessary gang header I/O has succeeded before
1429 * starting the actual work of free, claim, or write. Once the gang tree
1430 * is assembled, free and claim are in-memory operations that cannot fail.
1432 * In the event that a gang write fails, zio_dva_unallocate() walks the
1433 * gang tree to immediately free (i.e. insert back into the space map)
1434 * everything we've allocated. This ensures that we don't get ENOSPC
1435 * errors during repeated suspend/resume cycles due to a flaky device.
1437 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1438 * the gang tree, we won't modify the block, so we can safely defer the free
1439 * (knowing that the block is still intact). If we *can* assemble the gang
1440 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1441 * each constituent bp and we can allocate a new block on the next sync pass.
1443 * In all cases, the gang tree allows complete recovery from partial failure.
1444 * ==========================================================================
1448 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1453 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1454 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1455 &pio->io_bookmark));
1459 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1464 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1465 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1466 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1468 * As we rewrite each gang header, the pipeline will compute
1469 * a new gang block header checksum for it; but no one will
1470 * compute a new data checksum, so we do that here. The one
1471 * exception is the gang leader: the pipeline already computed
1472 * its data checksum because that stage precedes gang assembly.
1473 * (Presently, nothing actually uses interior data checksums;
1474 * this is just good hygiene.)
1476 if (gn != pio->io_gang_leader->io_gang_tree) {
1477 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1478 data, BP_GET_PSIZE(bp));
1481 * If we are here to damage data for testing purposes,
1482 * leave the GBH alone so that we can detect the damage.
1484 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1485 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1487 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1488 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1489 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1497 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1499 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1500 ZIO_GANG_CHILD_FLAGS(pio)));
1505 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1507 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1508 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1511 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1520 static void zio_gang_tree_assemble_done(zio_t *zio);
1522 static zio_gang_node_t *
1523 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1525 zio_gang_node_t *gn;
1527 ASSERT(*gnpp == NULL);
1529 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1530 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1537 zio_gang_node_free(zio_gang_node_t **gnpp)
1539 zio_gang_node_t *gn = *gnpp;
1541 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1542 ASSERT(gn->gn_child[g] == NULL);
1544 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1545 kmem_free(gn, sizeof (*gn));
1550 zio_gang_tree_free(zio_gang_node_t **gnpp)
1552 zio_gang_node_t *gn = *gnpp;
1557 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1558 zio_gang_tree_free(&gn->gn_child[g]);
1560 zio_gang_node_free(gnpp);
1564 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1566 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1568 ASSERT(gio->io_gang_leader == gio);
1569 ASSERT(BP_IS_GANG(bp));
1571 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1572 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1573 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1577 zio_gang_tree_assemble_done(zio_t *zio)
1579 zio_t *gio = zio->io_gang_leader;
1580 zio_gang_node_t *gn = zio->io_private;
1581 blkptr_t *bp = zio->io_bp;
1583 ASSERT(gio == zio_unique_parent(zio));
1584 ASSERT(zio->io_child_count == 0);
1589 if (BP_SHOULD_BYTESWAP(bp))
1590 byteswap_uint64_array(zio->io_data, zio->io_size);
1592 ASSERT(zio->io_data == gn->gn_gbh);
1593 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1594 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1596 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1597 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1598 if (!BP_IS_GANG(gbp))
1600 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1605 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1607 zio_t *gio = pio->io_gang_leader;
1610 ASSERT(BP_IS_GANG(bp) == !!gn);
1611 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1612 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1615 * If you're a gang header, your data is in gn->gn_gbh.
1616 * If you're a gang member, your data is in 'data' and gn == NULL.
1618 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1621 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1623 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1624 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1625 if (BP_IS_HOLE(gbp))
1627 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1628 data = (char *)data + BP_GET_PSIZE(gbp);
1632 if (gn == gio->io_gang_tree)
1633 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1640 zio_gang_assemble(zio_t *zio)
1642 blkptr_t *bp = zio->io_bp;
1644 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1645 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1647 zio->io_gang_leader = zio;
1649 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1651 return (ZIO_PIPELINE_CONTINUE);
1655 zio_gang_issue(zio_t *zio)
1657 blkptr_t *bp = zio->io_bp;
1659 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1660 return (ZIO_PIPELINE_STOP);
1662 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1663 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1665 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1666 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1668 zio_gang_tree_free(&zio->io_gang_tree);
1670 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1672 return (ZIO_PIPELINE_CONTINUE);
1676 zio_write_gang_member_ready(zio_t *zio)
1678 zio_t *pio = zio_unique_parent(zio);
1679 zio_t *gio = zio->io_gang_leader;
1680 dva_t *cdva = zio->io_bp->blk_dva;
1681 dva_t *pdva = pio->io_bp->blk_dva;
1684 if (BP_IS_HOLE(zio->io_bp))
1687 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1689 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1690 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1691 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1692 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1693 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1695 mutex_enter(&pio->io_lock);
1696 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1697 ASSERT(DVA_GET_GANG(&pdva[d]));
1698 asize = DVA_GET_ASIZE(&pdva[d]);
1699 asize += DVA_GET_ASIZE(&cdva[d]);
1700 DVA_SET_ASIZE(&pdva[d], asize);
1702 mutex_exit(&pio->io_lock);
1706 zio_write_gang_block(zio_t *pio)
1708 spa_t *spa = pio->io_spa;
1709 blkptr_t *bp = pio->io_bp;
1710 zio_t *gio = pio->io_gang_leader;
1712 zio_gang_node_t *gn, **gnpp;
1713 zio_gbh_phys_t *gbh;
1714 uint64_t txg = pio->io_txg;
1715 uint64_t resid = pio->io_size;
1717 int copies = gio->io_prop.zp_copies;
1718 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1722 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1723 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1724 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1726 pio->io_error = error;
1727 return (ZIO_PIPELINE_CONTINUE);
1731 gnpp = &gio->io_gang_tree;
1733 gnpp = pio->io_private;
1734 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1737 gn = zio_gang_node_alloc(gnpp);
1739 bzero(gbh, SPA_GANGBLOCKSIZE);
1742 * Create the gang header.
1744 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1745 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1748 * Create and nowait the gang children.
1750 for (int g = 0; resid != 0; resid -= lsize, g++) {
1751 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1753 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1755 zp.zp_checksum = gio->io_prop.zp_checksum;
1756 zp.zp_compress = ZIO_COMPRESS_OFF;
1757 zp.zp_type = DMU_OT_NONE;
1759 zp.zp_copies = gio->io_prop.zp_copies;
1761 zp.zp_dedup_verify = 0;
1763 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1764 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1765 zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1766 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1767 &pio->io_bookmark));
1771 * Set pio's pipeline to just wait for zio to finish.
1773 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1777 return (ZIO_PIPELINE_CONTINUE);
1781 * ==========================================================================
1783 * ==========================================================================
1786 zio_ddt_child_read_done(zio_t *zio)
1788 blkptr_t *bp = zio->io_bp;
1789 ddt_entry_t *dde = zio->io_private;
1791 zio_t *pio = zio_unique_parent(zio);
1793 mutex_enter(&pio->io_lock);
1794 ddp = ddt_phys_select(dde, bp);
1795 if (zio->io_error == 0)
1796 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1797 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1798 dde->dde_repair_data = zio->io_data;
1800 zio_buf_free(zio->io_data, zio->io_size);
1801 mutex_exit(&pio->io_lock);
1805 zio_ddt_read_start(zio_t *zio)
1807 blkptr_t *bp = zio->io_bp;
1809 ASSERT(BP_GET_DEDUP(bp));
1810 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1811 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1813 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1814 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1815 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1816 ddt_phys_t *ddp = dde->dde_phys;
1817 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1820 ASSERT(zio->io_vsd == NULL);
1823 if (ddp_self == NULL)
1824 return (ZIO_PIPELINE_CONTINUE);
1826 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1827 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1829 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1831 zio_nowait(zio_read(zio, zio->io_spa, &blk,
1832 zio_buf_alloc(zio->io_size), zio->io_size,
1833 zio_ddt_child_read_done, dde, zio->io_priority,
1834 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1835 &zio->io_bookmark));
1837 return (ZIO_PIPELINE_CONTINUE);
1840 zio_nowait(zio_read(zio, zio->io_spa, bp,
1841 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1842 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1844 return (ZIO_PIPELINE_CONTINUE);
1848 zio_ddt_read_done(zio_t *zio)
1850 blkptr_t *bp = zio->io_bp;
1852 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
1853 return (ZIO_PIPELINE_STOP);
1855 ASSERT(BP_GET_DEDUP(bp));
1856 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1857 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1859 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1860 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1861 ddt_entry_t *dde = zio->io_vsd;
1863 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
1864 return (ZIO_PIPELINE_CONTINUE);
1867 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
1868 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1869 return (ZIO_PIPELINE_STOP);
1871 if (dde->dde_repair_data != NULL) {
1872 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
1873 zio->io_child_error[ZIO_CHILD_DDT] = 0;
1875 ddt_repair_done(ddt, dde);
1879 ASSERT(zio->io_vsd == NULL);
1881 return (ZIO_PIPELINE_CONTINUE);
1885 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
1887 spa_t *spa = zio->io_spa;
1890 * Note: we compare the original data, not the transformed data,
1891 * because when zio->io_bp is an override bp, we will not have
1892 * pushed the I/O transforms. That's an important optimization
1893 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
1895 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1896 zio_t *lio = dde->dde_lead_zio[p];
1899 return (lio->io_orig_size != zio->io_orig_size ||
1900 bcmp(zio->io_orig_data, lio->io_orig_data,
1901 zio->io_orig_size) != 0);
1905 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
1906 ddt_phys_t *ddp = &dde->dde_phys[p];
1908 if (ddp->ddp_phys_birth != 0) {
1909 arc_buf_t *abuf = NULL;
1910 uint32_t aflags = ARC_WAIT;
1911 blkptr_t blk = *zio->io_bp;
1914 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
1918 error = arc_read_nolock(NULL, spa, &blk,
1919 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
1920 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1921 &aflags, &zio->io_bookmark);
1924 if (arc_buf_size(abuf) != zio->io_orig_size ||
1925 bcmp(abuf->b_data, zio->io_orig_data,
1926 zio->io_orig_size) != 0)
1928 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
1932 return (error != 0);
1940 zio_ddt_child_write_ready(zio_t *zio)
1942 int p = zio->io_prop.zp_copies;
1943 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1944 ddt_entry_t *dde = zio->io_private;
1945 ddt_phys_t *ddp = &dde->dde_phys[p];
1953 ASSERT(dde->dde_lead_zio[p] == zio);
1955 ddt_phys_fill(ddp, zio->io_bp);
1957 while ((pio = zio_walk_parents(zio)) != NULL)
1958 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
1964 zio_ddt_child_write_done(zio_t *zio)
1966 int p = zio->io_prop.zp_copies;
1967 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
1968 ddt_entry_t *dde = zio->io_private;
1969 ddt_phys_t *ddp = &dde->dde_phys[p];
1973 ASSERT(ddp->ddp_refcnt == 0);
1974 ASSERT(dde->dde_lead_zio[p] == zio);
1975 dde->dde_lead_zio[p] = NULL;
1977 if (zio->io_error == 0) {
1978 while (zio_walk_parents(zio) != NULL)
1979 ddt_phys_addref(ddp);
1981 ddt_phys_clear(ddp);
1988 zio_ddt_ditto_write_done(zio_t *zio)
1990 int p = DDT_PHYS_DITTO;
1991 zio_prop_t *zp = &zio->io_prop;
1992 blkptr_t *bp = zio->io_bp;
1993 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1994 ddt_entry_t *dde = zio->io_private;
1995 ddt_phys_t *ddp = &dde->dde_phys[p];
1996 ddt_key_t *ddk = &dde->dde_key;
2000 ASSERT(ddp->ddp_refcnt == 0);
2001 ASSERT(dde->dde_lead_zio[p] == zio);
2002 dde->dde_lead_zio[p] = NULL;
2004 if (zio->io_error == 0) {
2005 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2006 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2007 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2008 if (ddp->ddp_phys_birth != 0)
2009 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2010 ddt_phys_fill(ddp, bp);
2017 zio_ddt_write(zio_t *zio)
2019 spa_t *spa = zio->io_spa;
2020 blkptr_t *bp = zio->io_bp;
2021 uint64_t txg = zio->io_txg;
2022 zio_prop_t *zp = &zio->io_prop;
2023 int p = zp->zp_copies;
2027 ddt_t *ddt = ddt_select(spa, bp);
2031 ASSERT(BP_GET_DEDUP(bp));
2032 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2033 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2036 dde = ddt_lookup(ddt, bp, B_TRUE);
2037 ddp = &dde->dde_phys[p];
2039 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2041 * If we're using a weak checksum, upgrade to a strong checksum
2042 * and try again. If we're already using a strong checksum,
2043 * we can't resolve it, so just convert to an ordinary write.
2044 * (And automatically e-mail a paper to Nature?)
2046 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2047 zp->zp_checksum = spa_dedup_checksum(spa);
2048 zio_pop_transforms(zio);
2049 zio->io_stage = ZIO_STAGE_OPEN;
2054 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2056 return (ZIO_PIPELINE_CONTINUE);
2059 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2060 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2062 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2063 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2064 zio_prop_t czp = *zp;
2066 czp.zp_copies = ditto_copies;
2069 * If we arrived here with an override bp, we won't have run
2070 * the transform stack, so we won't have the data we need to
2071 * generate a child i/o. So, toss the override bp and restart.
2072 * This is safe, because using the override bp is just an
2073 * optimization; and it's rare, so the cost doesn't matter.
2075 if (zio->io_bp_override) {
2076 zio_pop_transforms(zio);
2077 zio->io_stage = ZIO_STAGE_OPEN;
2078 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2079 zio->io_bp_override = NULL;
2082 return (ZIO_PIPELINE_CONTINUE);
2085 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2086 zio->io_orig_size, &czp, NULL,
2087 zio_ddt_ditto_write_done, dde, zio->io_priority,
2088 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2090 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2091 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2094 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2095 if (ddp->ddp_phys_birth != 0)
2096 ddt_bp_fill(ddp, bp, txg);
2097 if (dde->dde_lead_zio[p] != NULL)
2098 zio_add_child(zio, dde->dde_lead_zio[p]);
2100 ddt_phys_addref(ddp);
2101 } else if (zio->io_bp_override) {
2102 ASSERT(bp->blk_birth == txg);
2103 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2104 ddt_phys_fill(ddp, bp);
2105 ddt_phys_addref(ddp);
2107 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2108 zio->io_orig_size, zp, zio_ddt_child_write_ready,
2109 zio_ddt_child_write_done, dde, zio->io_priority,
2110 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2112 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2113 dde->dde_lead_zio[p] = cio;
2123 return (ZIO_PIPELINE_CONTINUE);
2126 ddt_entry_t *freedde; /* for debugging */
2129 zio_ddt_free(zio_t *zio)
2131 spa_t *spa = zio->io_spa;
2132 blkptr_t *bp = zio->io_bp;
2133 ddt_t *ddt = ddt_select(spa, bp);
2137 ASSERT(BP_GET_DEDUP(bp));
2138 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2141 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2142 ddp = ddt_phys_select(dde, bp);
2143 ddt_phys_decref(ddp);
2146 return (ZIO_PIPELINE_CONTINUE);
2150 * ==========================================================================
2151 * Allocate and free blocks
2152 * ==========================================================================
2155 zio_dva_allocate(zio_t *zio)
2157 spa_t *spa = zio->io_spa;
2158 metaslab_class_t *mc = spa_normal_class(spa);
2159 blkptr_t *bp = zio->io_bp;
2163 if (zio->io_gang_leader == NULL) {
2164 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2165 zio->io_gang_leader = zio;
2168 ASSERT(BP_IS_HOLE(bp));
2169 ASSERT0(BP_GET_NDVAS(bp));
2170 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2171 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2172 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2175 * The dump device does not support gang blocks so allocation on
2176 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2177 * the "fast" gang feature.
2179 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2180 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2181 METASLAB_GANG_CHILD : 0;
2182 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2183 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2186 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2187 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2189 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2190 return (zio_write_gang_block(zio));
2191 zio->io_error = error;
2194 return (ZIO_PIPELINE_CONTINUE);
2198 zio_dva_free(zio_t *zio)
2200 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2202 return (ZIO_PIPELINE_CONTINUE);
2206 zio_dva_claim(zio_t *zio)
2210 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2212 zio->io_error = error;
2214 return (ZIO_PIPELINE_CONTINUE);
2218 * Undo an allocation. This is used by zio_done() when an I/O fails
2219 * and we want to give back the block we just allocated.
2220 * This handles both normal blocks and gang blocks.
2223 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2225 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2226 ASSERT(zio->io_bp_override == NULL);
2228 if (!BP_IS_HOLE(bp))
2229 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2232 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2233 zio_dva_unallocate(zio, gn->gn_child[g],
2234 &gn->gn_gbh->zg_blkptr[g]);
2240 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2243 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2244 uint64_t size, boolean_t use_slog)
2248 ASSERT(txg > spa_syncing_txg(spa));
2251 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2252 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2253 * when allocating them.
2256 error = metaslab_alloc(spa, spa_log_class(spa), size,
2257 new_bp, 1, txg, old_bp,
2258 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2262 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2263 new_bp, 1, txg, old_bp,
2264 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2268 BP_SET_LSIZE(new_bp, size);
2269 BP_SET_PSIZE(new_bp, size);
2270 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2271 BP_SET_CHECKSUM(new_bp,
2272 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2273 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2274 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2275 BP_SET_LEVEL(new_bp, 0);
2276 BP_SET_DEDUP(new_bp, 0);
2277 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2284 * Free an intent log block.
2287 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2289 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2290 ASSERT(!BP_IS_GANG(bp));
2292 zio_free(spa, txg, bp);
2296 * ==========================================================================
2297 * Read and write to physical devices
2298 * ==========================================================================
2301 zio_vdev_io_start(zio_t *zio)
2303 vdev_t *vd = zio->io_vd;
2305 spa_t *spa = zio->io_spa;
2307 ASSERT(zio->io_error == 0);
2308 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2311 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2312 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2315 * The mirror_ops handle multiple DVAs in a single BP.
2317 return (vdev_mirror_ops.vdev_op_io_start(zio));
2321 * We keep track of time-sensitive I/Os so that the scan thread
2322 * can quickly react to certain workloads. In particular, we care
2323 * about non-scrubbing, top-level reads and writes with the following
2325 * - synchronous writes of user data to non-slog devices
2326 * - any reads of user data
2327 * When these conditions are met, adjust the timestamp of spa_last_io
2328 * which allows the scan thread to adjust its workload accordingly.
2330 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2331 vd == vd->vdev_top && !vd->vdev_islog &&
2332 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2333 zio->io_txg != spa_syncing_txg(spa)) {
2334 uint64_t old = spa->spa_last_io;
2335 uint64_t new = ddi_get_lbolt64();
2337 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2340 align = 1ULL << vd->vdev_top->vdev_ashift;
2342 if (P2PHASE(zio->io_size, align) != 0) {
2343 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2344 char *abuf = zio_buf_alloc(asize);
2345 ASSERT(vd == vd->vdev_top);
2346 if (zio->io_type == ZIO_TYPE_WRITE) {
2347 bcopy(zio->io_data, abuf, zio->io_size);
2348 bzero(abuf + zio->io_size, asize - zio->io_size);
2350 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2353 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2354 ASSERT(P2PHASE(zio->io_size, align) == 0);
2355 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2358 * If this is a repair I/O, and there's no self-healing involved --
2359 * that is, we're just resilvering what we expect to resilver --
2360 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2361 * This prevents spurious resilvering with nested replication.
2362 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2363 * A is out of date, we'll read from C+D, then use the data to
2364 * resilver A+B -- but we don't actually want to resilver B, just A.
2365 * The top-level mirror has no way to know this, so instead we just
2366 * discard unnecessary repairs as we work our way down the vdev tree.
2367 * The same logic applies to any form of nested replication:
2368 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2370 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2371 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2372 zio->io_txg != 0 && /* not a delegated i/o */
2373 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2374 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2375 zio_vdev_io_bypass(zio);
2376 return (ZIO_PIPELINE_CONTINUE);
2379 if (vd->vdev_ops->vdev_op_leaf &&
2380 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2382 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2383 return (ZIO_PIPELINE_CONTINUE);
2385 if ((zio = vdev_queue_io(zio)) == NULL)
2386 return (ZIO_PIPELINE_STOP);
2388 if (!vdev_accessible(vd, zio)) {
2389 zio->io_error = ENXIO;
2391 return (ZIO_PIPELINE_STOP);
2395 return (vd->vdev_ops->vdev_op_io_start(zio));
2399 zio_vdev_io_done(zio_t *zio)
2401 vdev_t *vd = zio->io_vd;
2402 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2403 boolean_t unexpected_error = B_FALSE;
2405 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2406 return (ZIO_PIPELINE_STOP);
2408 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
2410 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
2412 vdev_queue_io_done(zio);
2414 if (zio->io_type == ZIO_TYPE_WRITE)
2415 vdev_cache_write(zio);
2417 if (zio_injection_enabled && zio->io_error == 0)
2418 zio->io_error = zio_handle_device_injection(vd,
2421 if (zio_injection_enabled && zio->io_error == 0)
2422 zio->io_error = zio_handle_label_injection(zio, EIO);
2424 if (zio->io_error) {
2425 if (!vdev_accessible(vd, zio)) {
2426 zio->io_error = ENXIO;
2428 unexpected_error = B_TRUE;
2433 ops->vdev_op_io_done(zio);
2435 if (unexpected_error)
2436 VERIFY(vdev_probe(vd, zio) == NULL);
2438 return (ZIO_PIPELINE_CONTINUE);
2442 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2443 * disk, and use that to finish the checksum ereport later.
2446 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2447 const void *good_buf)
2449 /* no processing needed */
2450 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2455 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2457 void *buf = zio_buf_alloc(zio->io_size);
2459 bcopy(zio->io_data, buf, zio->io_size);
2461 zcr->zcr_cbinfo = zio->io_size;
2462 zcr->zcr_cbdata = buf;
2463 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2464 zcr->zcr_free = zio_buf_free;
2468 zio_vdev_io_assess(zio_t *zio)
2470 vdev_t *vd = zio->io_vd;
2472 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2473 return (ZIO_PIPELINE_STOP);
2475 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2476 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2478 if (zio->io_vsd != NULL) {
2479 zio->io_vsd_ops->vsd_free(zio);
2483 if (zio_injection_enabled && zio->io_error == 0)
2484 zio->io_error = zio_handle_fault_injection(zio, EIO);
2487 * If the I/O failed, determine whether we should attempt to retry it.
2489 * On retry, we cut in line in the issue queue, since we don't want
2490 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2492 if (zio->io_error && vd == NULL &&
2493 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2494 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2495 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2497 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2498 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2499 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2500 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2501 zio_requeue_io_start_cut_in_line);
2502 return (ZIO_PIPELINE_STOP);
2506 * If we got an error on a leaf device, convert it to ENXIO
2507 * if the device is not accessible at all.
2509 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2510 !vdev_accessible(vd, zio))
2511 zio->io_error = ENXIO;
2514 * If we can't write to an interior vdev (mirror or RAID-Z),
2515 * set vdev_cant_write so that we stop trying to allocate from it.
2517 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2518 vd != NULL && !vd->vdev_ops->vdev_op_leaf)
2519 vd->vdev_cant_write = B_TRUE;
2522 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2524 return (ZIO_PIPELINE_CONTINUE);
2528 zio_vdev_io_reissue(zio_t *zio)
2530 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2531 ASSERT(zio->io_error == 0);
2533 zio->io_stage >>= 1;
2537 zio_vdev_io_redone(zio_t *zio)
2539 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2541 zio->io_stage >>= 1;
2545 zio_vdev_io_bypass(zio_t *zio)
2547 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2548 ASSERT(zio->io_error == 0);
2550 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2551 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2555 * ==========================================================================
2556 * Generate and verify checksums
2557 * ==========================================================================
2560 zio_checksum_generate(zio_t *zio)
2562 blkptr_t *bp = zio->io_bp;
2563 enum zio_checksum checksum;
2567 * This is zio_write_phys().
2568 * We're either generating a label checksum, or none at all.
2570 checksum = zio->io_prop.zp_checksum;
2572 if (checksum == ZIO_CHECKSUM_OFF)
2573 return (ZIO_PIPELINE_CONTINUE);
2575 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2577 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2578 ASSERT(!IO_IS_ALLOCATING(zio));
2579 checksum = ZIO_CHECKSUM_GANG_HEADER;
2581 checksum = BP_GET_CHECKSUM(bp);
2585 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2587 return (ZIO_PIPELINE_CONTINUE);
2591 zio_checksum_verify(zio_t *zio)
2593 zio_bad_cksum_t info;
2594 blkptr_t *bp = zio->io_bp;
2597 ASSERT(zio->io_vd != NULL);
2601 * This is zio_read_phys().
2602 * We're either verifying a label checksum, or nothing at all.
2604 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2605 return (ZIO_PIPELINE_CONTINUE);
2607 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2610 if ((error = zio_checksum_error(zio, &info)) != 0) {
2611 zio->io_error = error;
2612 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2613 zfs_ereport_start_checksum(zio->io_spa,
2614 zio->io_vd, zio, zio->io_offset,
2615 zio->io_size, NULL, &info);
2619 return (ZIO_PIPELINE_CONTINUE);
2623 * Called by RAID-Z to ensure we don't compute the checksum twice.
2626 zio_checksum_verified(zio_t *zio)
2628 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2632 * ==========================================================================
2633 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2634 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2635 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2636 * indicate errors that are specific to one I/O, and most likely permanent.
2637 * Any other error is presumed to be worse because we weren't expecting it.
2638 * ==========================================================================
2641 zio_worst_error(int e1, int e2)
2643 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2646 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2647 if (e1 == zio_error_rank[r1])
2650 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2651 if (e2 == zio_error_rank[r2])
2654 return (r1 > r2 ? e1 : e2);
2658 * ==========================================================================
2660 * ==========================================================================
2663 zio_ready(zio_t *zio)
2665 blkptr_t *bp = zio->io_bp;
2666 zio_t *pio, *pio_next;
2668 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2669 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2670 return (ZIO_PIPELINE_STOP);
2672 if (zio->io_ready) {
2673 ASSERT(IO_IS_ALLOCATING(zio));
2674 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2675 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2680 if (bp != NULL && bp != &zio->io_bp_copy)
2681 zio->io_bp_copy = *bp;
2684 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2686 mutex_enter(&zio->io_lock);
2687 zio->io_state[ZIO_WAIT_READY] = 1;
2688 pio = zio_walk_parents(zio);
2689 mutex_exit(&zio->io_lock);
2692 * As we notify zio's parents, new parents could be added.
2693 * New parents go to the head of zio's io_parent_list, however,
2694 * so we will (correctly) not notify them. The remainder of zio's
2695 * io_parent_list, from 'pio_next' onward, cannot change because
2696 * all parents must wait for us to be done before they can be done.
2698 for (; pio != NULL; pio = pio_next) {
2699 pio_next = zio_walk_parents(zio);
2700 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2703 if (zio->io_flags & ZIO_FLAG_NODATA) {
2704 if (BP_IS_GANG(bp)) {
2705 zio->io_flags &= ~ZIO_FLAG_NODATA;
2707 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2708 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2712 if (zio_injection_enabled &&
2713 zio->io_spa->spa_syncing_txg == zio->io_txg)
2714 zio_handle_ignored_writes(zio);
2716 return (ZIO_PIPELINE_CONTINUE);
2720 zio_done(zio_t *zio)
2722 spa_t *spa = zio->io_spa;
2723 zio_t *lio = zio->io_logical;
2724 blkptr_t *bp = zio->io_bp;
2725 vdev_t *vd = zio->io_vd;
2726 uint64_t psize = zio->io_size;
2727 zio_t *pio, *pio_next;
2730 * If our children haven't all completed,
2731 * wait for them and then repeat this pipeline stage.
2733 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2734 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2735 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2736 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2737 return (ZIO_PIPELINE_STOP);
2739 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2740 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2741 ASSERT(zio->io_children[c][w] == 0);
2744 ASSERT(bp->blk_pad[0] == 0);
2745 ASSERT(bp->blk_pad[1] == 0);
2746 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2747 (bp == zio_unique_parent(zio)->io_bp));
2748 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
2749 zio->io_bp_override == NULL &&
2750 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2751 ASSERT(!BP_SHOULD_BYTESWAP(bp));
2752 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
2753 ASSERT(BP_COUNT_GANG(bp) == 0 ||
2754 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
2759 * If there were child vdev/gang/ddt errors, they apply to us now.
2761 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2762 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2763 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2766 * If the I/O on the transformed data was successful, generate any
2767 * checksum reports now while we still have the transformed data.
2769 if (zio->io_error == 0) {
2770 while (zio->io_cksum_report != NULL) {
2771 zio_cksum_report_t *zcr = zio->io_cksum_report;
2772 uint64_t align = zcr->zcr_align;
2773 uint64_t asize = P2ROUNDUP(psize, align);
2774 char *abuf = zio->io_data;
2776 if (asize != psize) {
2777 abuf = zio_buf_alloc(asize);
2778 bcopy(zio->io_data, abuf, psize);
2779 bzero(abuf + psize, asize - psize);
2782 zio->io_cksum_report = zcr->zcr_next;
2783 zcr->zcr_next = NULL;
2784 zcr->zcr_finish(zcr, abuf);
2785 zfs_ereport_free_checksum(zcr);
2788 zio_buf_free(abuf, asize);
2792 zio_pop_transforms(zio); /* note: may set zio->io_error */
2794 vdev_stat_update(zio, psize);
2796 if (zio->io_error) {
2798 * If this I/O is attached to a particular vdev,
2799 * generate an error message describing the I/O failure
2800 * at the block level. We ignore these errors if the
2801 * device is currently unavailable.
2803 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
2804 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
2806 if ((zio->io_error == EIO || !(zio->io_flags &
2807 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
2810 * For logical I/O requests, tell the SPA to log the
2811 * error and generate a logical data ereport.
2813 spa_log_error(spa, zio);
2814 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
2819 if (zio->io_error && zio == lio) {
2821 * Determine whether zio should be reexecuted. This will
2822 * propagate all the way to the root via zio_notify_parent().
2824 ASSERT(vd == NULL && bp != NULL);
2825 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2827 if (IO_IS_ALLOCATING(zio) &&
2828 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
2829 if (zio->io_error != ENOSPC)
2830 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
2832 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2835 if ((zio->io_type == ZIO_TYPE_READ ||
2836 zio->io_type == ZIO_TYPE_FREE) &&
2837 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
2838 zio->io_error == ENXIO &&
2839 spa_load_state(spa) == SPA_LOAD_NONE &&
2840 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
2841 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2843 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
2844 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
2847 * Here is a possibly good place to attempt to do
2848 * either combinatorial reconstruction or error correction
2849 * based on checksums. It also might be a good place
2850 * to send out preliminary ereports before we suspend
2856 * If there were logical child errors, they apply to us now.
2857 * We defer this until now to avoid conflating logical child
2858 * errors with errors that happened to the zio itself when
2859 * updating vdev stats and reporting FMA events above.
2861 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2863 if ((zio->io_error || zio->io_reexecute) &&
2864 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
2865 !(zio->io_flags & ZIO_FLAG_IO_REWRITE))
2866 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
2868 zio_gang_tree_free(&zio->io_gang_tree);
2871 * Godfather I/Os should never suspend.
2873 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
2874 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
2875 zio->io_reexecute = 0;
2877 if (zio->io_reexecute) {
2879 * This is a logical I/O that wants to reexecute.
2881 * Reexecute is top-down. When an i/o fails, if it's not
2882 * the root, it simply notifies its parent and sticks around.
2883 * The parent, seeing that it still has children in zio_done(),
2884 * does the same. This percolates all the way up to the root.
2885 * The root i/o will reexecute or suspend the entire tree.
2887 * This approach ensures that zio_reexecute() honors
2888 * all the original i/o dependency relationships, e.g.
2889 * parents not executing until children are ready.
2891 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2893 zio->io_gang_leader = NULL;
2895 mutex_enter(&zio->io_lock);
2896 zio->io_state[ZIO_WAIT_DONE] = 1;
2897 mutex_exit(&zio->io_lock);
2900 * "The Godfather" I/O monitors its children but is
2901 * not a true parent to them. It will track them through
2902 * the pipeline but severs its ties whenever they get into
2903 * trouble (e.g. suspended). This allows "The Godfather"
2904 * I/O to return status without blocking.
2906 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2907 zio_link_t *zl = zio->io_walk_link;
2908 pio_next = zio_walk_parents(zio);
2910 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
2911 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
2912 zio_remove_child(pio, zio, zl);
2913 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2917 if ((pio = zio_unique_parent(zio)) != NULL) {
2919 * We're not a root i/o, so there's nothing to do
2920 * but notify our parent. Don't propagate errors
2921 * upward since we haven't permanently failed yet.
2923 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2924 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
2925 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2926 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
2928 * We'd fail again if we reexecuted now, so suspend
2929 * until conditions improve (e.g. device comes online).
2931 zio_suspend(spa, zio);
2934 * Reexecution is potentially a huge amount of work.
2935 * Hand it off to the otherwise-unused claim taskq.
2938 (void) taskq_dispatch_safe(
2939 spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
2940 (task_func_t *)zio_reexecute, zio, TQ_SLEEP,
2943 (void) taskq_dispatch(
2944 spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
2945 (task_func_t *)zio_reexecute, zio, TQ_SLEEP);
2948 return (ZIO_PIPELINE_STOP);
2951 ASSERT(zio->io_child_count == 0);
2952 ASSERT(zio->io_reexecute == 0);
2953 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
2956 * Report any checksum errors, since the I/O is complete.
2958 while (zio->io_cksum_report != NULL) {
2959 zio_cksum_report_t *zcr = zio->io_cksum_report;
2960 zio->io_cksum_report = zcr->zcr_next;
2961 zcr->zcr_next = NULL;
2962 zcr->zcr_finish(zcr, NULL);
2963 zfs_ereport_free_checksum(zcr);
2967 * It is the responsibility of the done callback to ensure that this
2968 * particular zio is no longer discoverable for adoption, and as
2969 * such, cannot acquire any new parents.
2974 mutex_enter(&zio->io_lock);
2975 zio->io_state[ZIO_WAIT_DONE] = 1;
2976 mutex_exit(&zio->io_lock);
2978 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
2979 zio_link_t *zl = zio->io_walk_link;
2980 pio_next = zio_walk_parents(zio);
2981 zio_remove_child(pio, zio, zl);
2982 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
2985 if (zio->io_waiter != NULL) {
2986 mutex_enter(&zio->io_lock);
2987 zio->io_executor = NULL;
2988 cv_broadcast(&zio->io_cv);
2989 mutex_exit(&zio->io_lock);
2994 return (ZIO_PIPELINE_STOP);
2998 * ==========================================================================
2999 * I/O pipeline definition
3000 * ==========================================================================
3002 static zio_pipe_stage_t *zio_pipeline[] = {
3008 zio_checksum_generate,
3022 zio_checksum_verify,
3026 /* dnp is the dnode for zb1->zb_object */
3028 zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_t *zb1,
3029 const zbookmark_t *zb2)
3031 uint64_t zb1nextL0, zb2thisobj;
3033 ASSERT(zb1->zb_objset == zb2->zb_objset);
3034 ASSERT(zb2->zb_level == 0);
3037 * A bookmark in the deadlist is considered to be after
3040 if (zb2->zb_object == DMU_DEADLIST_OBJECT)
3043 /* The objset_phys_t isn't before anything. */
3047 zb1nextL0 = (zb1->zb_blkid + 1) <<
3048 ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
3050 zb2thisobj = zb2->zb_object ? zb2->zb_object :
3051 zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
3053 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3054 uint64_t nextobj = zb1nextL0 *
3055 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
3056 return (nextobj <= zb2thisobj);
3059 if (zb1->zb_object < zb2thisobj)
3061 if (zb1->zb_object > zb2thisobj)
3063 if (zb2->zb_object == DMU_META_DNODE_OBJECT)
3065 return (zb1nextL0 <= zb2->zb_blkid);