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) 2013 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>
38 #include <sys/trim_map.h>
40 SYSCTL_DECL(_vfs_zfs);
41 SYSCTL_NODE(_vfs_zfs, OID_AUTO, zio, CTLFLAG_RW, 0, "ZFS ZIO");
42 static int zio_use_uma = 0;
43 TUNABLE_INT("vfs.zfs.zio.use_uma", &zio_use_uma);
44 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, use_uma, CTLFLAG_RDTUN, &zio_use_uma, 0,
45 "Use uma(9) for ZIO allocations");
46 static int zio_exclude_metadata = 0;
47 TUNABLE_INT("vfs.zfs.zio.exclude_metadata", &zio_exclude_metadata);
48 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, exclude_metadata, CTLFLAG_RDTUN, &zio_exclude_metadata, 0,
49 "Exclude metadata buffers from dumps as well");
51 zio_trim_stats_t zio_trim_stats = {
52 { "bytes", KSTAT_DATA_UINT64,
53 "Number of bytes successfully TRIMmed" },
54 { "success", KSTAT_DATA_UINT64,
55 "Number of successful TRIM requests" },
56 { "unsupported", KSTAT_DATA_UINT64,
57 "Number of TRIM requests that failed because TRIM is not supported" },
58 { "failed", KSTAT_DATA_UINT64,
59 "Number of TRIM requests that failed for reasons other than not supported" },
62 static kstat_t *zio_trim_ksp;
65 * ==========================================================================
67 * ==========================================================================
69 uint8_t zio_priority_table[ZIO_PRIORITY_TABLE_SIZE] = {
70 0, /* ZIO_PRIORITY_NOW */
71 0, /* ZIO_PRIORITY_SYNC_READ */
72 0, /* ZIO_PRIORITY_SYNC_WRITE */
73 0, /* ZIO_PRIORITY_LOG_WRITE */
74 1, /* ZIO_PRIORITY_CACHE_FILL */
75 1, /* ZIO_PRIORITY_AGG */
76 4, /* ZIO_PRIORITY_FREE */
77 4, /* ZIO_PRIORITY_ASYNC_WRITE */
78 6, /* ZIO_PRIORITY_ASYNC_READ */
79 10, /* ZIO_PRIORITY_RESILVER */
80 20, /* ZIO_PRIORITY_SCRUB */
81 2, /* ZIO_PRIORITY_DDT_PREFETCH */
82 30, /* ZIO_PRIORITY_TRIM */
86 * ==========================================================================
87 * I/O type descriptions
88 * ==========================================================================
90 char *zio_type_name[ZIO_TYPES] = {
91 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
96 * ==========================================================================
98 * ==========================================================================
100 kmem_cache_t *zio_cache;
101 kmem_cache_t *zio_link_cache;
102 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
103 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
106 extern vmem_t *zio_alloc_arena;
108 extern int zfs_mg_alloc_failures;
111 * The following actions directly effect the spa's sync-to-convergence logic.
112 * The values below define the sync pass when we start performing the action.
113 * Care should be taken when changing these values as they directly impact
114 * spa_sync() performance. Tuning these values may introduce subtle performance
115 * pathologies and should only be done in the context of performance analysis.
116 * These tunables will eventually be removed and replaced with #defines once
117 * enough analysis has been done to determine optimal values.
119 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
120 * regular blocks are not deferred.
122 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
123 TUNABLE_INT("vfs.zfs.sync_pass_deferred_free", &zfs_sync_pass_deferred_free);
124 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_deferred_free, CTLFLAG_RDTUN,
125 &zfs_sync_pass_deferred_free, 0, "defer frees starting in this pass");
126 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
127 TUNABLE_INT("vfs.zfs.sync_pass_dont_compress", &zfs_sync_pass_dont_compress);
128 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_dont_compress, CTLFLAG_RDTUN,
129 &zfs_sync_pass_dont_compress, 0, "don't compress starting in this pass");
130 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
131 TUNABLE_INT("vfs.zfs.sync_pass_rewrite", &zfs_sync_pass_rewrite);
132 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_rewrite, CTLFLAG_RDTUN,
133 &zfs_sync_pass_rewrite, 0, "rewrite new bps starting in this pass");
136 * An allocating zio is one that either currently has the DVA allocate
137 * stage set or will have it later in its lifetime.
139 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
141 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
144 int zio_buf_debug_limit = 16384;
146 int zio_buf_debug_limit = 0;
153 zio_cache = kmem_cache_create("zio_cache",
154 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
155 zio_link_cache = kmem_cache_create("zio_link_cache",
156 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
161 * For small buffers, we want a cache for each multiple of
162 * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache
163 * for each quarter-power of 2. For large buffers, we want
164 * a cache for each multiple of PAGESIZE.
166 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
167 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
170 size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
172 while (p2 & (p2 - 1))
178 * If we are using watchpoints, put each buffer on its own page,
179 * to eliminate the performance overhead of trapping to the
180 * kernel when modifying a non-watched buffer that shares the
181 * page with a watched buffer.
183 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
187 if (size <= 4 * SPA_MINBLOCKSIZE) {
188 align = SPA_MINBLOCKSIZE;
189 } else if (IS_P2ALIGNED(size, PAGESIZE)) {
191 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
197 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
198 zio_buf_cache[c] = kmem_cache_create(name, size,
199 align, NULL, NULL, NULL, NULL, NULL, cflags);
202 * Since zio_data bufs do not appear in crash dumps, we
203 * pass KMC_NOTOUCH so that no allocator metadata is
204 * stored with the buffers.
206 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
207 zio_data_buf_cache[c] = kmem_cache_create(name, size,
208 align, NULL, NULL, NULL, NULL, NULL,
209 cflags | KMC_NOTOUCH | KMC_NODEBUG);
214 ASSERT(zio_buf_cache[c] != NULL);
215 if (zio_buf_cache[c - 1] == NULL)
216 zio_buf_cache[c - 1] = zio_buf_cache[c];
218 ASSERT(zio_data_buf_cache[c] != NULL);
219 if (zio_data_buf_cache[c - 1] == NULL)
220 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
225 * The zio write taskqs have 1 thread per cpu, allow 1/2 of the taskqs
226 * to fail 3 times per txg or 8 failures, whichever is greater.
228 if (zfs_mg_alloc_failures == 0)
229 zfs_mg_alloc_failures = MAX((3 * max_ncpus / 2), 8);
230 else if (zfs_mg_alloc_failures < 8)
231 zfs_mg_alloc_failures = 8;
235 zio_trim_ksp = kstat_create("zfs", 0, "zio_trim", "misc",
237 sizeof(zio_trim_stats) / sizeof(kstat_named_t),
240 if (zio_trim_ksp != NULL) {
241 zio_trim_ksp->ks_data = &zio_trim_stats;
242 kstat_install(zio_trim_ksp);
250 kmem_cache_t *last_cache = NULL;
251 kmem_cache_t *last_data_cache = NULL;
253 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
254 if (zio_buf_cache[c] != last_cache) {
255 last_cache = zio_buf_cache[c];
256 kmem_cache_destroy(zio_buf_cache[c]);
258 zio_buf_cache[c] = NULL;
260 if (zio_data_buf_cache[c] != last_data_cache) {
261 last_data_cache = zio_data_buf_cache[c];
262 kmem_cache_destroy(zio_data_buf_cache[c]);
264 zio_data_buf_cache[c] = NULL;
267 kmem_cache_destroy(zio_link_cache);
268 kmem_cache_destroy(zio_cache);
272 if (zio_trim_ksp != NULL) {
273 kstat_delete(zio_trim_ksp);
279 * ==========================================================================
280 * Allocate and free I/O buffers
281 * ==========================================================================
285 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
286 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
287 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
288 * excess / transient data in-core during a crashdump.
291 zio_buf_alloc(size_t size)
293 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
294 int flags = zio_exclude_metadata ? KM_NODEBUG : 0;
296 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
299 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
301 return (kmem_alloc(size, KM_SLEEP|flags));
305 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
306 * crashdump if the kernel panics. This exists so that we will limit the amount
307 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
308 * of kernel heap dumped to disk when the kernel panics)
311 zio_data_buf_alloc(size_t size)
313 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
315 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
318 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
320 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG));
324 zio_buf_free(void *buf, size_t size)
326 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
328 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
331 kmem_cache_free(zio_buf_cache[c], buf);
333 kmem_free(buf, size);
337 zio_data_buf_free(void *buf, size_t size)
339 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
341 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
344 kmem_cache_free(zio_data_buf_cache[c], buf);
346 kmem_free(buf, size);
350 * ==========================================================================
351 * Push and pop I/O transform buffers
352 * ==========================================================================
355 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
356 zio_transform_func_t *transform)
358 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
360 zt->zt_orig_data = zio->io_data;
361 zt->zt_orig_size = zio->io_size;
362 zt->zt_bufsize = bufsize;
363 zt->zt_transform = transform;
365 zt->zt_next = zio->io_transform_stack;
366 zio->io_transform_stack = zt;
373 zio_pop_transforms(zio_t *zio)
377 while ((zt = zio->io_transform_stack) != NULL) {
378 if (zt->zt_transform != NULL)
379 zt->zt_transform(zio,
380 zt->zt_orig_data, zt->zt_orig_size);
382 if (zt->zt_bufsize != 0)
383 zio_buf_free(zio->io_data, zt->zt_bufsize);
385 zio->io_data = zt->zt_orig_data;
386 zio->io_size = zt->zt_orig_size;
387 zio->io_transform_stack = zt->zt_next;
389 kmem_free(zt, sizeof (zio_transform_t));
394 * ==========================================================================
395 * I/O transform callbacks for subblocks and decompression
396 * ==========================================================================
399 zio_subblock(zio_t *zio, void *data, uint64_t size)
401 ASSERT(zio->io_size > size);
403 if (zio->io_type == ZIO_TYPE_READ)
404 bcopy(zio->io_data, data, size);
408 zio_decompress(zio_t *zio, void *data, uint64_t size)
410 if (zio->io_error == 0 &&
411 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
412 zio->io_data, data, zio->io_size, size) != 0)
413 zio->io_error = SET_ERROR(EIO);
417 * ==========================================================================
418 * I/O parent/child relationships and pipeline interlocks
419 * ==========================================================================
422 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
423 * continue calling these functions until they return NULL.
424 * Otherwise, the next caller will pick up the list walk in
425 * some indeterminate state. (Otherwise every caller would
426 * have to pass in a cookie to keep the state represented by
427 * io_walk_link, which gets annoying.)
430 zio_walk_parents(zio_t *cio)
432 zio_link_t *zl = cio->io_walk_link;
433 list_t *pl = &cio->io_parent_list;
435 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
436 cio->io_walk_link = zl;
441 ASSERT(zl->zl_child == cio);
442 return (zl->zl_parent);
446 zio_walk_children(zio_t *pio)
448 zio_link_t *zl = pio->io_walk_link;
449 list_t *cl = &pio->io_child_list;
451 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
452 pio->io_walk_link = zl;
457 ASSERT(zl->zl_parent == pio);
458 return (zl->zl_child);
462 zio_unique_parent(zio_t *cio)
464 zio_t *pio = zio_walk_parents(cio);
466 VERIFY(zio_walk_parents(cio) == NULL);
471 zio_add_child(zio_t *pio, zio_t *cio)
473 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
476 * Logical I/Os can have logical, gang, or vdev children.
477 * Gang I/Os can have gang or vdev children.
478 * Vdev I/Os can only have vdev children.
479 * The following ASSERT captures all of these constraints.
481 ASSERT(cio->io_child_type <= pio->io_child_type);
486 mutex_enter(&cio->io_lock);
487 mutex_enter(&pio->io_lock);
489 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
491 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
492 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
494 list_insert_head(&pio->io_child_list, zl);
495 list_insert_head(&cio->io_parent_list, zl);
497 pio->io_child_count++;
498 cio->io_parent_count++;
500 mutex_exit(&pio->io_lock);
501 mutex_exit(&cio->io_lock);
505 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
507 ASSERT(zl->zl_parent == pio);
508 ASSERT(zl->zl_child == cio);
510 mutex_enter(&cio->io_lock);
511 mutex_enter(&pio->io_lock);
513 list_remove(&pio->io_child_list, zl);
514 list_remove(&cio->io_parent_list, zl);
516 pio->io_child_count--;
517 cio->io_parent_count--;
519 mutex_exit(&pio->io_lock);
520 mutex_exit(&cio->io_lock);
522 kmem_cache_free(zio_link_cache, zl);
526 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
528 uint64_t *countp = &zio->io_children[child][wait];
529 boolean_t waiting = B_FALSE;
531 mutex_enter(&zio->io_lock);
532 ASSERT(zio->io_stall == NULL);
535 zio->io_stall = countp;
538 mutex_exit(&zio->io_lock);
544 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
546 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
547 int *errorp = &pio->io_child_error[zio->io_child_type];
549 mutex_enter(&pio->io_lock);
550 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
551 *errorp = zio_worst_error(*errorp, zio->io_error);
552 pio->io_reexecute |= zio->io_reexecute;
553 ASSERT3U(*countp, >, 0);
554 if (--*countp == 0 && pio->io_stall == countp) {
555 pio->io_stall = NULL;
556 mutex_exit(&pio->io_lock);
559 mutex_exit(&pio->io_lock);
564 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
566 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
567 zio->io_error = zio->io_child_error[c];
571 * ==========================================================================
572 * Create the various types of I/O (read, write, free, etc)
573 * ==========================================================================
576 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
577 void *data, uint64_t size, zio_done_func_t *done, void *private,
578 zio_type_t type, int priority, enum zio_flag flags,
579 vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
580 enum zio_stage stage, enum zio_stage pipeline)
584 ASSERT3U(type == ZIO_TYPE_FREE || size, <=, SPA_MAXBLOCKSIZE);
585 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
586 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
588 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
589 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
590 ASSERT(vd || stage == ZIO_STAGE_OPEN);
592 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
593 bzero(zio, sizeof (zio_t));
595 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
596 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
598 list_create(&zio->io_parent_list, sizeof (zio_link_t),
599 offsetof(zio_link_t, zl_parent_node));
600 list_create(&zio->io_child_list, sizeof (zio_link_t),
601 offsetof(zio_link_t, zl_child_node));
604 zio->io_child_type = ZIO_CHILD_VDEV;
605 else if (flags & ZIO_FLAG_GANG_CHILD)
606 zio->io_child_type = ZIO_CHILD_GANG;
607 else if (flags & ZIO_FLAG_DDT_CHILD)
608 zio->io_child_type = ZIO_CHILD_DDT;
610 zio->io_child_type = ZIO_CHILD_LOGICAL;
613 zio->io_bp = (blkptr_t *)bp;
614 zio->io_bp_copy = *bp;
615 zio->io_bp_orig = *bp;
616 if (type != ZIO_TYPE_WRITE ||
617 zio->io_child_type == ZIO_CHILD_DDT)
618 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
619 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
620 zio->io_logical = zio;
621 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
622 pipeline |= ZIO_GANG_STAGES;
628 zio->io_private = private;
630 zio->io_priority = priority;
632 zio->io_offset = offset;
633 zio->io_orig_data = zio->io_data = data;
634 zio->io_orig_size = zio->io_size = size;
635 zio->io_orig_flags = zio->io_flags = flags;
636 zio->io_orig_stage = zio->io_stage = stage;
637 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
639 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
640 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
643 zio->io_bookmark = *zb;
646 if (zio->io_logical == NULL)
647 zio->io_logical = pio->io_logical;
648 if (zio->io_child_type == ZIO_CHILD_GANG)
649 zio->io_gang_leader = pio->io_gang_leader;
650 zio_add_child(pio, zio);
657 zio_destroy(zio_t *zio)
659 list_destroy(&zio->io_parent_list);
660 list_destroy(&zio->io_child_list);
661 mutex_destroy(&zio->io_lock);
662 cv_destroy(&zio->io_cv);
663 kmem_cache_free(zio_cache, zio);
667 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
668 void *private, enum zio_flag flags)
672 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
673 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
674 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
680 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
682 return (zio_null(NULL, spa, NULL, done, private, flags));
686 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
687 void *data, uint64_t size, zio_done_func_t *done, void *private,
688 int priority, enum zio_flag flags, const zbookmark_t *zb)
692 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
693 data, size, done, private,
694 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
695 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
696 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
702 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
703 void *data, uint64_t size, const zio_prop_t *zp,
704 zio_done_func_t *ready, zio_done_func_t *done, void *private,
705 int priority, enum zio_flag flags, const zbookmark_t *zb)
709 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
710 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
711 zp->zp_compress >= ZIO_COMPRESS_OFF &&
712 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
713 DMU_OT_IS_VALID(zp->zp_type) &&
716 zp->zp_copies <= spa_max_replication(spa));
718 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
719 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
720 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
721 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
723 zio->io_ready = ready;
730 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
731 uint64_t size, zio_done_func_t *done, void *private, int priority,
732 enum zio_flag flags, zbookmark_t *zb)
736 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
737 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
738 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
744 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
746 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
747 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
748 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
749 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
752 * We must reset the io_prop to match the values that existed
753 * when the bp was first written by dmu_sync() keeping in mind
754 * that nopwrite and dedup are mutually exclusive.
756 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
757 zio->io_prop.zp_nopwrite = nopwrite;
758 zio->io_prop.zp_copies = copies;
759 zio->io_bp_override = bp;
763 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
765 metaslab_check_free(spa, bp);
768 * Frees that are for the currently-syncing txg, are not going to be
769 * deferred, and which will not need to do a read (i.e. not GANG or
770 * DEDUP), can be processed immediately. Otherwise, put them on the
771 * in-memory list for later processing.
773 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
774 txg != spa->spa_syncing_txg ||
775 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
776 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
778 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp,
779 BP_GET_PSIZE(bp), 0)));
784 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
785 uint64_t size, enum zio_flag flags)
788 enum zio_stage stage = ZIO_FREE_PIPELINE;
790 dprintf_bp(bp, "freeing in txg %llu, pass %u",
791 (longlong_t)txg, spa->spa_sync_pass);
793 ASSERT(!BP_IS_HOLE(bp));
794 ASSERT(spa_syncing_txg(spa) == txg);
795 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
797 metaslab_check_free(spa, bp);
800 if (zfs_trim_enabled)
801 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START |
802 ZIO_STAGE_VDEV_IO_ASSESS;
804 * GANG and DEDUP blocks can induce a read (for the gang block header,
805 * or the DDT), so issue them asynchronously so that this thread is
808 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
809 stage |= ZIO_STAGE_ISSUE_ASYNC;
811 zio = zio_create(pio, spa, txg, bp, NULL, size,
812 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
813 NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
819 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
820 zio_done_func_t *done, void *private, enum zio_flag flags)
825 * A claim is an allocation of a specific block. Claims are needed
826 * to support immediate writes in the intent log. The issue is that
827 * immediate writes contain committed data, but in a txg that was
828 * *not* committed. Upon opening the pool after an unclean shutdown,
829 * the intent log claims all blocks that contain immediate write data
830 * so that the SPA knows they're in use.
832 * All claims *must* be resolved in the first txg -- before the SPA
833 * starts allocating blocks -- so that nothing is allocated twice.
834 * If txg == 0 we just verify that the block is claimable.
836 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
837 ASSERT(txg == spa_first_txg(spa) || txg == 0);
838 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
840 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
841 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
842 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
848 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset,
849 uint64_t size, zio_done_func_t *done, void *private, int priority,
855 if (vd->vdev_children == 0) {
856 zio = zio_create(pio, spa, 0, NULL, NULL, size, done, private,
857 ZIO_TYPE_IOCTL, priority, flags, vd, offset, NULL,
858 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
862 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
864 for (c = 0; c < vd->vdev_children; c++)
865 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
866 offset, size, done, private, priority, flags));
873 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
874 void *data, int checksum, zio_done_func_t *done, void *private,
875 int priority, enum zio_flag flags, boolean_t labels)
879 ASSERT(vd->vdev_children == 0);
880 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
881 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
882 ASSERT3U(offset + size, <=, vd->vdev_psize);
884 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
885 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
886 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
888 zio->io_prop.zp_checksum = checksum;
894 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
895 void *data, int checksum, zio_done_func_t *done, void *private,
896 int priority, enum zio_flag flags, boolean_t labels)
900 ASSERT(vd->vdev_children == 0);
901 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
902 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
903 ASSERT3U(offset + size, <=, vd->vdev_psize);
905 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
906 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
907 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
909 zio->io_prop.zp_checksum = checksum;
911 if (zio_checksum_table[checksum].ci_eck) {
913 * zec checksums are necessarily destructive -- they modify
914 * the end of the write buffer to hold the verifier/checksum.
915 * Therefore, we must make a local copy in case the data is
916 * being written to multiple places in parallel.
918 void *wbuf = zio_buf_alloc(size);
919 bcopy(data, wbuf, size);
920 zio_push_transform(zio, wbuf, size, size, NULL);
927 * Create a child I/O to do some work for us.
930 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
931 void *data, uint64_t size, int type, int priority, enum zio_flag flags,
932 zio_done_func_t *done, void *private)
934 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
937 ASSERT(vd->vdev_parent ==
938 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
940 if (type == ZIO_TYPE_READ && bp != NULL) {
942 * If we have the bp, then the child should perform the
943 * checksum and the parent need not. This pushes error
944 * detection as close to the leaves as possible and
945 * eliminates redundant checksums in the interior nodes.
947 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
948 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
951 if (vd->vdev_children == 0)
952 offset += VDEV_LABEL_START_SIZE;
954 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
957 * If we've decided to do a repair, the write is not speculative --
958 * even if the original read was.
960 if (flags & ZIO_FLAG_IO_REPAIR)
961 flags &= ~ZIO_FLAG_SPECULATIVE;
963 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
964 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
965 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
971 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
972 int type, int priority, enum zio_flag flags,
973 zio_done_func_t *done, void *private)
977 ASSERT(vd->vdev_ops->vdev_op_leaf);
979 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
980 data, size, done, private, type, priority,
981 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
983 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
989 zio_flush(zio_t *zio, vdev_t *vd)
991 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0,
992 NULL, NULL, ZIO_PRIORITY_NOW,
993 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
997 zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size)
1000 ASSERT(vd->vdev_ops->vdev_op_leaf);
1002 return zio_ioctl(zio, spa, vd, DKIOCTRIM, offset, size,
1003 NULL, NULL, ZIO_PRIORITY_TRIM,
1004 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY);
1008 zio_shrink(zio_t *zio, uint64_t size)
1010 ASSERT(zio->io_executor == NULL);
1011 ASSERT(zio->io_orig_size == zio->io_size);
1012 ASSERT(size <= zio->io_size);
1015 * We don't shrink for raidz because of problems with the
1016 * reconstruction when reading back less than the block size.
1017 * Note, BP_IS_RAIDZ() assumes no compression.
1019 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1020 if (!BP_IS_RAIDZ(zio->io_bp))
1021 zio->io_orig_size = zio->io_size = size;
1025 * ==========================================================================
1026 * Prepare to read and write logical blocks
1027 * ==========================================================================
1031 zio_read_bp_init(zio_t *zio)
1033 blkptr_t *bp = zio->io_bp;
1035 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1036 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1037 !(zio->io_flags & ZIO_FLAG_RAW)) {
1038 uint64_t psize = BP_GET_PSIZE(bp);
1039 void *cbuf = zio_buf_alloc(psize);
1041 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
1044 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1045 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1047 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1048 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1050 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1051 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1053 return (ZIO_PIPELINE_CONTINUE);
1057 zio_write_bp_init(zio_t *zio)
1059 spa_t *spa = zio->io_spa;
1060 zio_prop_t *zp = &zio->io_prop;
1061 enum zio_compress compress = zp->zp_compress;
1062 blkptr_t *bp = zio->io_bp;
1063 uint64_t lsize = zio->io_size;
1064 uint64_t psize = lsize;
1068 * If our children haven't all reached the ready stage,
1069 * wait for them and then repeat this pipeline stage.
1071 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1072 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1073 return (ZIO_PIPELINE_STOP);
1075 if (!IO_IS_ALLOCATING(zio))
1076 return (ZIO_PIPELINE_CONTINUE);
1078 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1080 if (zio->io_bp_override) {
1081 ASSERT(bp->blk_birth != zio->io_txg);
1082 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1084 *bp = *zio->io_bp_override;
1085 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1088 * If we've been overridden and nopwrite is set then
1089 * set the flag accordingly to indicate that a nopwrite
1090 * has already occurred.
1092 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1093 ASSERT(!zp->zp_dedup);
1094 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1095 return (ZIO_PIPELINE_CONTINUE);
1098 ASSERT(!zp->zp_nopwrite);
1100 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1101 return (ZIO_PIPELINE_CONTINUE);
1103 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1104 zp->zp_dedup_verify);
1106 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1107 BP_SET_DEDUP(bp, 1);
1108 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1109 return (ZIO_PIPELINE_CONTINUE);
1111 zio->io_bp_override = NULL;
1115 if (bp->blk_birth == zio->io_txg) {
1117 * We're rewriting an existing block, which means we're
1118 * working on behalf of spa_sync(). For spa_sync() to
1119 * converge, it must eventually be the case that we don't
1120 * have to allocate new blocks. But compression changes
1121 * the blocksize, which forces a reallocate, and makes
1122 * convergence take longer. Therefore, after the first
1123 * few passes, stop compressing to ensure convergence.
1125 pass = spa_sync_pass(spa);
1127 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1128 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1129 ASSERT(!BP_GET_DEDUP(bp));
1131 if (pass >= zfs_sync_pass_dont_compress)
1132 compress = ZIO_COMPRESS_OFF;
1134 /* Make sure someone doesn't change their mind on overwrites */
1135 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
1136 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1139 if (compress != ZIO_COMPRESS_OFF) {
1140 metaslab_class_t *mc = spa_normal_class(spa);
1141 void *cbuf = zio_buf_alloc(lsize);
1142 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize,
1143 (size_t)metaslab_class_get_minblocksize(mc));
1144 if (psize == 0 || psize == lsize) {
1145 compress = ZIO_COMPRESS_OFF;
1146 zio_buf_free(cbuf, lsize);
1148 ASSERT(psize < lsize);
1149 zio_push_transform(zio, cbuf, psize, lsize, NULL);
1154 * The final pass of spa_sync() must be all rewrites, but the first
1155 * few passes offer a trade-off: allocating blocks defers convergence,
1156 * but newly allocated blocks are sequential, so they can be written
1157 * to disk faster. Therefore, we allow the first few passes of
1158 * spa_sync() to allocate new blocks, but force rewrites after that.
1159 * There should only be a handful of blocks after pass 1 in any case.
1161 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
1162 pass >= zfs_sync_pass_rewrite) {
1164 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1165 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1166 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1169 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1173 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1175 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1176 BP_SET_LSIZE(bp, lsize);
1177 BP_SET_PSIZE(bp, psize);
1178 BP_SET_COMPRESS(bp, compress);
1179 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1180 BP_SET_TYPE(bp, zp->zp_type);
1181 BP_SET_LEVEL(bp, zp->zp_level);
1182 BP_SET_DEDUP(bp, zp->zp_dedup);
1183 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1185 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1186 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1187 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1189 if (zp->zp_nopwrite) {
1190 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1191 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1192 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1196 return (ZIO_PIPELINE_CONTINUE);
1200 zio_free_bp_init(zio_t *zio)
1202 blkptr_t *bp = zio->io_bp;
1204 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1205 if (BP_GET_DEDUP(bp))
1206 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1209 return (ZIO_PIPELINE_CONTINUE);
1213 * ==========================================================================
1214 * Execute the I/O pipeline
1215 * ==========================================================================
1219 zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q, boolean_t cutinline)
1221 spa_t *spa = zio->io_spa;
1222 zio_type_t t = zio->io_type;
1223 int flags = TQ_SLEEP | (cutinline ? TQ_FRONT : 0);
1225 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1228 * If we're a config writer or a probe, the normal issue and
1229 * interrupt threads may all be blocked waiting for the config lock.
1230 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1232 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1236 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1238 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1242 * If this is a high priority I/O, then use the high priority taskq.
1244 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1245 spa->spa_zio_taskq[t][q + 1] != NULL)
1248 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1250 (void) taskq_dispatch_safe(spa->spa_zio_taskq[t][q],
1251 (task_func_t *)zio_execute, zio, flags, &zio->io_task);
1253 (void) taskq_dispatch(spa->spa_zio_taskq[t][q],
1254 (task_func_t *)zio_execute, zio, flags);
1259 zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
1261 kthread_t *executor = zio->io_executor;
1262 spa_t *spa = zio->io_spa;
1264 for (zio_type_t t = 0; t < ZIO_TYPES; t++)
1265 if (taskq_member(spa->spa_zio_taskq[t][q], executor))
1272 zio_issue_async(zio_t *zio)
1274 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1276 return (ZIO_PIPELINE_STOP);
1280 zio_interrupt(zio_t *zio)
1282 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1286 * Execute the I/O pipeline until one of the following occurs:
1288 * (1) the I/O completes
1289 * (2) the pipeline stalls waiting for dependent child I/Os
1290 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1291 * (4) the I/O is delegated by vdev-level caching or aggregation
1292 * (5) the I/O is deferred due to vdev-level queueing
1293 * (6) the I/O is handed off to another thread.
1295 * In all cases, the pipeline stops whenever there's no CPU work; it never
1296 * burns a thread in cv_wait().
1298 * There's no locking on io_stage because there's no legitimate way
1299 * for multiple threads to be attempting to process the same I/O.
1301 static zio_pipe_stage_t *zio_pipeline[];
1304 zio_execute(zio_t *zio)
1306 zio->io_executor = curthread;
1308 while (zio->io_stage < ZIO_STAGE_DONE) {
1309 enum zio_stage pipeline = zio->io_pipeline;
1310 enum zio_stage stage = zio->io_stage;
1313 ASSERT(!MUTEX_HELD(&zio->io_lock));
1314 ASSERT(ISP2(stage));
1315 ASSERT(zio->io_stall == NULL);
1319 } while ((stage & pipeline) == 0);
1321 ASSERT(stage <= ZIO_STAGE_DONE);
1324 * If we are in interrupt context and this pipeline stage
1325 * will grab a config lock that is held across I/O,
1326 * or may wait for an I/O that needs an interrupt thread
1327 * to complete, issue async to avoid deadlock.
1329 * For VDEV_IO_START, we cut in line so that the io will
1330 * be sent to disk promptly.
1332 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1333 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1334 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1335 zio_requeue_io_start_cut_in_line : B_FALSE;
1336 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1340 zio->io_stage = stage;
1341 rv = zio_pipeline[highbit(stage) - 1](zio);
1343 if (rv == ZIO_PIPELINE_STOP)
1346 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1351 * ==========================================================================
1352 * Initiate I/O, either sync or async
1353 * ==========================================================================
1356 zio_wait(zio_t *zio)
1360 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1361 ASSERT(zio->io_executor == NULL);
1363 zio->io_waiter = curthread;
1367 mutex_enter(&zio->io_lock);
1368 while (zio->io_executor != NULL)
1369 cv_wait(&zio->io_cv, &zio->io_lock);
1370 mutex_exit(&zio->io_lock);
1372 error = zio->io_error;
1379 zio_nowait(zio_t *zio)
1381 ASSERT(zio->io_executor == NULL);
1383 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1384 zio_unique_parent(zio) == NULL) {
1386 * This is a logical async I/O with no parent to wait for it.
1387 * We add it to the spa_async_root_zio "Godfather" I/O which
1388 * will ensure they complete prior to unloading the pool.
1390 spa_t *spa = zio->io_spa;
1392 zio_add_child(spa->spa_async_zio_root, zio);
1399 * ==========================================================================
1400 * Reexecute or suspend/resume failed I/O
1401 * ==========================================================================
1405 zio_reexecute(zio_t *pio)
1407 zio_t *cio, *cio_next;
1409 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1410 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1411 ASSERT(pio->io_gang_leader == NULL);
1412 ASSERT(pio->io_gang_tree == NULL);
1414 pio->io_flags = pio->io_orig_flags;
1415 pio->io_stage = pio->io_orig_stage;
1416 pio->io_pipeline = pio->io_orig_pipeline;
1417 pio->io_reexecute = 0;
1418 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1420 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1421 pio->io_state[w] = 0;
1422 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1423 pio->io_child_error[c] = 0;
1425 if (IO_IS_ALLOCATING(pio))
1426 BP_ZERO(pio->io_bp);
1429 * As we reexecute pio's children, new children could be created.
1430 * New children go to the head of pio's io_child_list, however,
1431 * so we will (correctly) not reexecute them. The key is that
1432 * the remainder of pio's io_child_list, from 'cio_next' onward,
1433 * cannot be affected by any side effects of reexecuting 'cio'.
1435 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1436 cio_next = zio_walk_children(pio);
1437 mutex_enter(&pio->io_lock);
1438 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1439 pio->io_children[cio->io_child_type][w]++;
1440 mutex_exit(&pio->io_lock);
1445 * Now that all children have been reexecuted, execute the parent.
1446 * We don't reexecute "The Godfather" I/O here as it's the
1447 * responsibility of the caller to wait on him.
1449 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1454 zio_suspend(spa_t *spa, zio_t *zio)
1456 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1457 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1458 "failure and the failure mode property for this pool "
1459 "is set to panic.", spa_name(spa));
1461 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1463 mutex_enter(&spa->spa_suspend_lock);
1465 if (spa->spa_suspend_zio_root == NULL)
1466 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1467 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1468 ZIO_FLAG_GODFATHER);
1470 spa->spa_suspended = B_TRUE;
1473 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1474 ASSERT(zio != spa->spa_suspend_zio_root);
1475 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1476 ASSERT(zio_unique_parent(zio) == NULL);
1477 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1478 zio_add_child(spa->spa_suspend_zio_root, zio);
1481 mutex_exit(&spa->spa_suspend_lock);
1485 zio_resume(spa_t *spa)
1490 * Reexecute all previously suspended i/o.
1492 mutex_enter(&spa->spa_suspend_lock);
1493 spa->spa_suspended = B_FALSE;
1494 cv_broadcast(&spa->spa_suspend_cv);
1495 pio = spa->spa_suspend_zio_root;
1496 spa->spa_suspend_zio_root = NULL;
1497 mutex_exit(&spa->spa_suspend_lock);
1503 return (zio_wait(pio));
1507 zio_resume_wait(spa_t *spa)
1509 mutex_enter(&spa->spa_suspend_lock);
1510 while (spa_suspended(spa))
1511 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1512 mutex_exit(&spa->spa_suspend_lock);
1516 * ==========================================================================
1519 * A gang block is a collection of small blocks that looks to the DMU
1520 * like one large block. When zio_dva_allocate() cannot find a block
1521 * of the requested size, due to either severe fragmentation or the pool
1522 * being nearly full, it calls zio_write_gang_block() to construct the
1523 * block from smaller fragments.
1525 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1526 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1527 * an indirect block: it's an array of block pointers. It consumes
1528 * only one sector and hence is allocatable regardless of fragmentation.
1529 * The gang header's bps point to its gang members, which hold the data.
1531 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1532 * as the verifier to ensure uniqueness of the SHA256 checksum.
1533 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1534 * not the gang header. This ensures that data block signatures (needed for
1535 * deduplication) are independent of how the block is physically stored.
1537 * Gang blocks can be nested: a gang member may itself be a gang block.
1538 * Thus every gang block is a tree in which root and all interior nodes are
1539 * gang headers, and the leaves are normal blocks that contain user data.
1540 * The root of the gang tree is called the gang leader.
1542 * To perform any operation (read, rewrite, free, claim) on a gang block,
1543 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1544 * in the io_gang_tree field of the original logical i/o by recursively
1545 * reading the gang leader and all gang headers below it. This yields
1546 * an in-core tree containing the contents of every gang header and the
1547 * bps for every constituent of the gang block.
1549 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1550 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1551 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1552 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1553 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1554 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1555 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1556 * of the gang header plus zio_checksum_compute() of the data to update the
1557 * gang header's blk_cksum as described above.
1559 * The two-phase assemble/issue model solves the problem of partial failure --
1560 * what if you'd freed part of a gang block but then couldn't read the
1561 * gang header for another part? Assembling the entire gang tree first
1562 * ensures that all the necessary gang header I/O has succeeded before
1563 * starting the actual work of free, claim, or write. Once the gang tree
1564 * is assembled, free and claim are in-memory operations that cannot fail.
1566 * In the event that a gang write fails, zio_dva_unallocate() walks the
1567 * gang tree to immediately free (i.e. insert back into the space map)
1568 * everything we've allocated. This ensures that we don't get ENOSPC
1569 * errors during repeated suspend/resume cycles due to a flaky device.
1571 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1572 * the gang tree, we won't modify the block, so we can safely defer the free
1573 * (knowing that the block is still intact). If we *can* assemble the gang
1574 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1575 * each constituent bp and we can allocate a new block on the next sync pass.
1577 * In all cases, the gang tree allows complete recovery from partial failure.
1578 * ==========================================================================
1582 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1587 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1588 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1589 &pio->io_bookmark));
1593 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1598 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1599 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1600 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1602 * As we rewrite each gang header, the pipeline will compute
1603 * a new gang block header checksum for it; but no one will
1604 * compute a new data checksum, so we do that here. The one
1605 * exception is the gang leader: the pipeline already computed
1606 * its data checksum because that stage precedes gang assembly.
1607 * (Presently, nothing actually uses interior data checksums;
1608 * this is just good hygiene.)
1610 if (gn != pio->io_gang_leader->io_gang_tree) {
1611 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1612 data, BP_GET_PSIZE(bp));
1615 * If we are here to damage data for testing purposes,
1616 * leave the GBH alone so that we can detect the damage.
1618 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1619 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1621 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1622 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1623 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1631 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1633 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1634 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp),
1635 ZIO_GANG_CHILD_FLAGS(pio)));
1640 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1642 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1643 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1646 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1655 static void zio_gang_tree_assemble_done(zio_t *zio);
1657 static zio_gang_node_t *
1658 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1660 zio_gang_node_t *gn;
1662 ASSERT(*gnpp == NULL);
1664 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1665 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1672 zio_gang_node_free(zio_gang_node_t **gnpp)
1674 zio_gang_node_t *gn = *gnpp;
1676 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1677 ASSERT(gn->gn_child[g] == NULL);
1679 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1680 kmem_free(gn, sizeof (*gn));
1685 zio_gang_tree_free(zio_gang_node_t **gnpp)
1687 zio_gang_node_t *gn = *gnpp;
1692 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1693 zio_gang_tree_free(&gn->gn_child[g]);
1695 zio_gang_node_free(gnpp);
1699 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1701 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1703 ASSERT(gio->io_gang_leader == gio);
1704 ASSERT(BP_IS_GANG(bp));
1706 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1707 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1708 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1712 zio_gang_tree_assemble_done(zio_t *zio)
1714 zio_t *gio = zio->io_gang_leader;
1715 zio_gang_node_t *gn = zio->io_private;
1716 blkptr_t *bp = zio->io_bp;
1718 ASSERT(gio == zio_unique_parent(zio));
1719 ASSERT(zio->io_child_count == 0);
1724 if (BP_SHOULD_BYTESWAP(bp))
1725 byteswap_uint64_array(zio->io_data, zio->io_size);
1727 ASSERT(zio->io_data == gn->gn_gbh);
1728 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1729 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1731 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1732 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1733 if (!BP_IS_GANG(gbp))
1735 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1740 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1742 zio_t *gio = pio->io_gang_leader;
1745 ASSERT(BP_IS_GANG(bp) == !!gn);
1746 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1747 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1750 * If you're a gang header, your data is in gn->gn_gbh.
1751 * If you're a gang member, your data is in 'data' and gn == NULL.
1753 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1756 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1758 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1759 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1760 if (BP_IS_HOLE(gbp))
1762 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1763 data = (char *)data + BP_GET_PSIZE(gbp);
1767 if (gn == gio->io_gang_tree && gio->io_data != NULL)
1768 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1775 zio_gang_assemble(zio_t *zio)
1777 blkptr_t *bp = zio->io_bp;
1779 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1780 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1782 zio->io_gang_leader = zio;
1784 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1786 return (ZIO_PIPELINE_CONTINUE);
1790 zio_gang_issue(zio_t *zio)
1792 blkptr_t *bp = zio->io_bp;
1794 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1795 return (ZIO_PIPELINE_STOP);
1797 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1798 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1800 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1801 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1803 zio_gang_tree_free(&zio->io_gang_tree);
1805 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1807 return (ZIO_PIPELINE_CONTINUE);
1811 zio_write_gang_member_ready(zio_t *zio)
1813 zio_t *pio = zio_unique_parent(zio);
1814 zio_t *gio = zio->io_gang_leader;
1815 dva_t *cdva = zio->io_bp->blk_dva;
1816 dva_t *pdva = pio->io_bp->blk_dva;
1819 if (BP_IS_HOLE(zio->io_bp))
1822 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1824 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1825 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1826 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1827 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1828 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1830 mutex_enter(&pio->io_lock);
1831 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1832 ASSERT(DVA_GET_GANG(&pdva[d]));
1833 asize = DVA_GET_ASIZE(&pdva[d]);
1834 asize += DVA_GET_ASIZE(&cdva[d]);
1835 DVA_SET_ASIZE(&pdva[d], asize);
1837 mutex_exit(&pio->io_lock);
1841 zio_write_gang_block(zio_t *pio)
1843 spa_t *spa = pio->io_spa;
1844 blkptr_t *bp = pio->io_bp;
1845 zio_t *gio = pio->io_gang_leader;
1847 zio_gang_node_t *gn, **gnpp;
1848 zio_gbh_phys_t *gbh;
1849 uint64_t txg = pio->io_txg;
1850 uint64_t resid = pio->io_size;
1852 int copies = gio->io_prop.zp_copies;
1853 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1857 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1858 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1859 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1861 pio->io_error = error;
1862 return (ZIO_PIPELINE_CONTINUE);
1866 gnpp = &gio->io_gang_tree;
1868 gnpp = pio->io_private;
1869 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1872 gn = zio_gang_node_alloc(gnpp);
1874 bzero(gbh, SPA_GANGBLOCKSIZE);
1877 * Create the gang header.
1879 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1880 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1883 * Create and nowait the gang children.
1885 for (int g = 0; resid != 0; resid -= lsize, g++) {
1886 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1888 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1890 zp.zp_checksum = gio->io_prop.zp_checksum;
1891 zp.zp_compress = ZIO_COMPRESS_OFF;
1892 zp.zp_type = DMU_OT_NONE;
1894 zp.zp_copies = gio->io_prop.zp_copies;
1895 zp.zp_dedup = B_FALSE;
1896 zp.zp_dedup_verify = B_FALSE;
1897 zp.zp_nopwrite = B_FALSE;
1899 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1900 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1901 zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1902 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1903 &pio->io_bookmark));
1907 * Set pio's pipeline to just wait for zio to finish.
1909 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1913 return (ZIO_PIPELINE_CONTINUE);
1917 * The zio_nop_write stage in the pipeline determines if allocating
1918 * a new bp is necessary. By leveraging a cryptographically secure checksum,
1919 * such as SHA256, we can compare the checksums of the new data and the old
1920 * to determine if allocating a new block is required. The nopwrite
1921 * feature can handle writes in either syncing or open context (i.e. zil
1922 * writes) and as a result is mutually exclusive with dedup.
1925 zio_nop_write(zio_t *zio)
1927 blkptr_t *bp = zio->io_bp;
1928 blkptr_t *bp_orig = &zio->io_bp_orig;
1929 zio_prop_t *zp = &zio->io_prop;
1931 ASSERT(BP_GET_LEVEL(bp) == 0);
1932 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1933 ASSERT(zp->zp_nopwrite);
1934 ASSERT(!zp->zp_dedup);
1935 ASSERT(zio->io_bp_override == NULL);
1936 ASSERT(IO_IS_ALLOCATING(zio));
1939 * Check to see if the original bp and the new bp have matching
1940 * characteristics (i.e. same checksum, compression algorithms, etc).
1941 * If they don't then just continue with the pipeline which will
1942 * allocate a new bp.
1944 if (BP_IS_HOLE(bp_orig) ||
1945 !zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup ||
1946 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
1947 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
1948 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
1949 zp->zp_copies != BP_GET_NDVAS(bp_orig))
1950 return (ZIO_PIPELINE_CONTINUE);
1953 * If the checksums match then reset the pipeline so that we
1954 * avoid allocating a new bp and issuing any I/O.
1956 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
1957 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup);
1958 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
1959 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
1960 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
1961 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
1962 sizeof (uint64_t)) == 0);
1965 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1966 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1969 return (ZIO_PIPELINE_CONTINUE);
1973 * ==========================================================================
1975 * ==========================================================================
1978 zio_ddt_child_read_done(zio_t *zio)
1980 blkptr_t *bp = zio->io_bp;
1981 ddt_entry_t *dde = zio->io_private;
1983 zio_t *pio = zio_unique_parent(zio);
1985 mutex_enter(&pio->io_lock);
1986 ddp = ddt_phys_select(dde, bp);
1987 if (zio->io_error == 0)
1988 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1989 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1990 dde->dde_repair_data = zio->io_data;
1992 zio_buf_free(zio->io_data, zio->io_size);
1993 mutex_exit(&pio->io_lock);
1997 zio_ddt_read_start(zio_t *zio)
1999 blkptr_t *bp = zio->io_bp;
2001 ASSERT(BP_GET_DEDUP(bp));
2002 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2003 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2005 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2006 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2007 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2008 ddt_phys_t *ddp = dde->dde_phys;
2009 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2012 ASSERT(zio->io_vsd == NULL);
2015 if (ddp_self == NULL)
2016 return (ZIO_PIPELINE_CONTINUE);
2018 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2019 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2021 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2023 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2024 zio_buf_alloc(zio->io_size), zio->io_size,
2025 zio_ddt_child_read_done, dde, zio->io_priority,
2026 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
2027 &zio->io_bookmark));
2029 return (ZIO_PIPELINE_CONTINUE);
2032 zio_nowait(zio_read(zio, zio->io_spa, bp,
2033 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
2034 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2036 return (ZIO_PIPELINE_CONTINUE);
2040 zio_ddt_read_done(zio_t *zio)
2042 blkptr_t *bp = zio->io_bp;
2044 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2045 return (ZIO_PIPELINE_STOP);
2047 ASSERT(BP_GET_DEDUP(bp));
2048 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2049 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2051 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2052 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2053 ddt_entry_t *dde = zio->io_vsd;
2055 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2056 return (ZIO_PIPELINE_CONTINUE);
2059 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2060 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2061 return (ZIO_PIPELINE_STOP);
2063 if (dde->dde_repair_data != NULL) {
2064 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2065 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2067 ddt_repair_done(ddt, dde);
2071 ASSERT(zio->io_vsd == NULL);
2073 return (ZIO_PIPELINE_CONTINUE);
2077 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2079 spa_t *spa = zio->io_spa;
2082 * Note: we compare the original data, not the transformed data,
2083 * because when zio->io_bp is an override bp, we will not have
2084 * pushed the I/O transforms. That's an important optimization
2085 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2087 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2088 zio_t *lio = dde->dde_lead_zio[p];
2091 return (lio->io_orig_size != zio->io_orig_size ||
2092 bcmp(zio->io_orig_data, lio->io_orig_data,
2093 zio->io_orig_size) != 0);
2097 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2098 ddt_phys_t *ddp = &dde->dde_phys[p];
2100 if (ddp->ddp_phys_birth != 0) {
2101 arc_buf_t *abuf = NULL;
2102 uint32_t aflags = ARC_WAIT;
2103 blkptr_t blk = *zio->io_bp;
2106 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2110 error = arc_read(NULL, spa, &blk,
2111 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2112 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2113 &aflags, &zio->io_bookmark);
2116 if (arc_buf_size(abuf) != zio->io_orig_size ||
2117 bcmp(abuf->b_data, zio->io_orig_data,
2118 zio->io_orig_size) != 0)
2119 error = SET_ERROR(EEXIST);
2120 VERIFY(arc_buf_remove_ref(abuf, &abuf));
2124 return (error != 0);
2132 zio_ddt_child_write_ready(zio_t *zio)
2134 int p = zio->io_prop.zp_copies;
2135 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2136 ddt_entry_t *dde = zio->io_private;
2137 ddt_phys_t *ddp = &dde->dde_phys[p];
2145 ASSERT(dde->dde_lead_zio[p] == zio);
2147 ddt_phys_fill(ddp, zio->io_bp);
2149 while ((pio = zio_walk_parents(zio)) != NULL)
2150 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2156 zio_ddt_child_write_done(zio_t *zio)
2158 int p = zio->io_prop.zp_copies;
2159 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2160 ddt_entry_t *dde = zio->io_private;
2161 ddt_phys_t *ddp = &dde->dde_phys[p];
2165 ASSERT(ddp->ddp_refcnt == 0);
2166 ASSERT(dde->dde_lead_zio[p] == zio);
2167 dde->dde_lead_zio[p] = NULL;
2169 if (zio->io_error == 0) {
2170 while (zio_walk_parents(zio) != NULL)
2171 ddt_phys_addref(ddp);
2173 ddt_phys_clear(ddp);
2180 zio_ddt_ditto_write_done(zio_t *zio)
2182 int p = DDT_PHYS_DITTO;
2183 zio_prop_t *zp = &zio->io_prop;
2184 blkptr_t *bp = zio->io_bp;
2185 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2186 ddt_entry_t *dde = zio->io_private;
2187 ddt_phys_t *ddp = &dde->dde_phys[p];
2188 ddt_key_t *ddk = &dde->dde_key;
2192 ASSERT(ddp->ddp_refcnt == 0);
2193 ASSERT(dde->dde_lead_zio[p] == zio);
2194 dde->dde_lead_zio[p] = NULL;
2196 if (zio->io_error == 0) {
2197 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2198 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2199 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2200 if (ddp->ddp_phys_birth != 0)
2201 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2202 ddt_phys_fill(ddp, bp);
2209 zio_ddt_write(zio_t *zio)
2211 spa_t *spa = zio->io_spa;
2212 blkptr_t *bp = zio->io_bp;
2213 uint64_t txg = zio->io_txg;
2214 zio_prop_t *zp = &zio->io_prop;
2215 int p = zp->zp_copies;
2219 ddt_t *ddt = ddt_select(spa, bp);
2223 ASSERT(BP_GET_DEDUP(bp));
2224 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2225 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2228 dde = ddt_lookup(ddt, bp, B_TRUE);
2229 ddp = &dde->dde_phys[p];
2231 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2233 * If we're using a weak checksum, upgrade to a strong checksum
2234 * and try again. If we're already using a strong checksum,
2235 * we can't resolve it, so just convert to an ordinary write.
2236 * (And automatically e-mail a paper to Nature?)
2238 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2239 zp->zp_checksum = spa_dedup_checksum(spa);
2240 zio_pop_transforms(zio);
2241 zio->io_stage = ZIO_STAGE_OPEN;
2244 zp->zp_dedup = B_FALSE;
2246 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2248 return (ZIO_PIPELINE_CONTINUE);
2251 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2252 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2254 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2255 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2256 zio_prop_t czp = *zp;
2258 czp.zp_copies = ditto_copies;
2261 * If we arrived here with an override bp, we won't have run
2262 * the transform stack, so we won't have the data we need to
2263 * generate a child i/o. So, toss the override bp and restart.
2264 * This is safe, because using the override bp is just an
2265 * optimization; and it's rare, so the cost doesn't matter.
2267 if (zio->io_bp_override) {
2268 zio_pop_transforms(zio);
2269 zio->io_stage = ZIO_STAGE_OPEN;
2270 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2271 zio->io_bp_override = NULL;
2274 return (ZIO_PIPELINE_CONTINUE);
2277 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2278 zio->io_orig_size, &czp, NULL,
2279 zio_ddt_ditto_write_done, dde, zio->io_priority,
2280 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2282 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2283 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2286 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2287 if (ddp->ddp_phys_birth != 0)
2288 ddt_bp_fill(ddp, bp, txg);
2289 if (dde->dde_lead_zio[p] != NULL)
2290 zio_add_child(zio, dde->dde_lead_zio[p]);
2292 ddt_phys_addref(ddp);
2293 } else if (zio->io_bp_override) {
2294 ASSERT(bp->blk_birth == txg);
2295 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2296 ddt_phys_fill(ddp, bp);
2297 ddt_phys_addref(ddp);
2299 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2300 zio->io_orig_size, zp, zio_ddt_child_write_ready,
2301 zio_ddt_child_write_done, dde, zio->io_priority,
2302 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2304 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2305 dde->dde_lead_zio[p] = cio;
2315 return (ZIO_PIPELINE_CONTINUE);
2318 ddt_entry_t *freedde; /* for debugging */
2321 zio_ddt_free(zio_t *zio)
2323 spa_t *spa = zio->io_spa;
2324 blkptr_t *bp = zio->io_bp;
2325 ddt_t *ddt = ddt_select(spa, bp);
2329 ASSERT(BP_GET_DEDUP(bp));
2330 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2333 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2334 ddp = ddt_phys_select(dde, bp);
2335 ddt_phys_decref(ddp);
2338 return (ZIO_PIPELINE_CONTINUE);
2342 * ==========================================================================
2343 * Allocate and free blocks
2344 * ==========================================================================
2347 zio_dva_allocate(zio_t *zio)
2349 spa_t *spa = zio->io_spa;
2350 metaslab_class_t *mc = spa_normal_class(spa);
2351 blkptr_t *bp = zio->io_bp;
2355 if (zio->io_gang_leader == NULL) {
2356 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2357 zio->io_gang_leader = zio;
2360 ASSERT(BP_IS_HOLE(bp));
2361 ASSERT0(BP_GET_NDVAS(bp));
2362 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2363 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2364 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2367 * The dump device does not support gang blocks so allocation on
2368 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2369 * the "fast" gang feature.
2371 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2372 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2373 METASLAB_GANG_CHILD : 0;
2374 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2375 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2378 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2379 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2381 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2382 return (zio_write_gang_block(zio));
2383 zio->io_error = error;
2386 return (ZIO_PIPELINE_CONTINUE);
2390 zio_dva_free(zio_t *zio)
2392 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2394 return (ZIO_PIPELINE_CONTINUE);
2398 zio_dva_claim(zio_t *zio)
2402 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2404 zio->io_error = error;
2406 return (ZIO_PIPELINE_CONTINUE);
2410 * Undo an allocation. This is used by zio_done() when an I/O fails
2411 * and we want to give back the block we just allocated.
2412 * This handles both normal blocks and gang blocks.
2415 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2417 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2418 ASSERT(zio->io_bp_override == NULL);
2420 if (!BP_IS_HOLE(bp))
2421 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2424 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2425 zio_dva_unallocate(zio, gn->gn_child[g],
2426 &gn->gn_gbh->zg_blkptr[g]);
2432 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2435 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2436 uint64_t size, boolean_t use_slog)
2440 ASSERT(txg > spa_syncing_txg(spa));
2443 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2444 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2445 * when allocating them.
2448 error = metaslab_alloc(spa, spa_log_class(spa), size,
2449 new_bp, 1, txg, old_bp,
2450 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2454 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2455 new_bp, 1, txg, old_bp,
2456 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2460 BP_SET_LSIZE(new_bp, size);
2461 BP_SET_PSIZE(new_bp, size);
2462 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2463 BP_SET_CHECKSUM(new_bp,
2464 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2465 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2466 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2467 BP_SET_LEVEL(new_bp, 0);
2468 BP_SET_DEDUP(new_bp, 0);
2469 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2476 * Free an intent log block.
2479 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2481 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2482 ASSERT(!BP_IS_GANG(bp));
2484 zio_free(spa, txg, bp);
2488 * ==========================================================================
2489 * Read, write and delete to physical devices
2490 * ==========================================================================
2493 zio_vdev_io_start(zio_t *zio)
2495 vdev_t *vd = zio->io_vd;
2497 spa_t *spa = zio->io_spa;
2499 ASSERT(zio->io_error == 0);
2500 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2503 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2504 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2507 * The mirror_ops handle multiple DVAs in a single BP.
2509 return (vdev_mirror_ops.vdev_op_io_start(zio));
2512 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE) {
2513 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg);
2514 return (ZIO_PIPELINE_CONTINUE);
2518 * We keep track of time-sensitive I/Os so that the scan thread
2519 * can quickly react to certain workloads. In particular, we care
2520 * about non-scrubbing, top-level reads and writes with the following
2522 * - synchronous writes of user data to non-slog devices
2523 * - any reads of user data
2524 * When these conditions are met, adjust the timestamp of spa_last_io
2525 * which allows the scan thread to adjust its workload accordingly.
2527 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2528 vd == vd->vdev_top && !vd->vdev_islog &&
2529 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2530 zio->io_txg != spa_syncing_txg(spa)) {
2531 uint64_t old = spa->spa_last_io;
2532 uint64_t new = ddi_get_lbolt64();
2534 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2537 align = 1ULL << vd->vdev_top->vdev_ashift;
2539 if (P2PHASE(zio->io_size, align) != 0) {
2540 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2542 if (zio->io_type == ZIO_TYPE_READ ||
2543 zio->io_type == ZIO_TYPE_WRITE)
2544 abuf = zio_buf_alloc(asize);
2545 ASSERT(vd == vd->vdev_top);
2546 if (zio->io_type == ZIO_TYPE_WRITE) {
2547 bcopy(zio->io_data, abuf, zio->io_size);
2548 bzero(abuf + zio->io_size, asize - zio->io_size);
2550 zio_push_transform(zio, abuf, asize, abuf ? asize : 0,
2554 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2555 ASSERT(P2PHASE(zio->io_size, align) == 0);
2556 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa));
2559 * If this is a repair I/O, and there's no self-healing involved --
2560 * that is, we're just resilvering what we expect to resilver --
2561 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2562 * This prevents spurious resilvering with nested replication.
2563 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2564 * A is out of date, we'll read from C+D, then use the data to
2565 * resilver A+B -- but we don't actually want to resilver B, just A.
2566 * The top-level mirror has no way to know this, so instead we just
2567 * discard unnecessary repairs as we work our way down the vdev tree.
2568 * The same logic applies to any form of nested replication:
2569 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2571 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2572 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2573 zio->io_txg != 0 && /* not a delegated i/o */
2574 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2575 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2576 zio_vdev_io_bypass(zio);
2577 return (ZIO_PIPELINE_CONTINUE);
2580 if (vd->vdev_ops->vdev_op_leaf &&
2581 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2583 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2584 return (ZIO_PIPELINE_CONTINUE);
2586 if ((zio = vdev_queue_io(zio)) == NULL)
2587 return (ZIO_PIPELINE_STOP);
2589 if (!vdev_accessible(vd, zio)) {
2590 zio->io_error = SET_ERROR(ENXIO);
2592 return (ZIO_PIPELINE_STOP);
2597 * Note that we ignore repair writes for TRIM because they can conflict
2598 * with normal writes. This isn't an issue because, by definition, we
2599 * only repair blocks that aren't freed.
2601 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_WRITE &&
2602 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2603 if (!trim_map_write_start(zio))
2604 return (ZIO_PIPELINE_STOP);
2607 return (vd->vdev_ops->vdev_op_io_start(zio));
2611 zio_vdev_io_done(zio_t *zio)
2613 vdev_t *vd = zio->io_vd;
2614 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2615 boolean_t unexpected_error = B_FALSE;
2617 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2618 return (ZIO_PIPELINE_STOP);
2620 ASSERT(zio->io_type == ZIO_TYPE_READ ||
2621 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE);
2623 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2624 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2626 if (zio->io_type == ZIO_TYPE_WRITE &&
2627 !(zio->io_flags & ZIO_FLAG_IO_REPAIR))
2628 trim_map_write_done(zio);
2630 vdev_queue_io_done(zio);
2632 if (zio->io_type == ZIO_TYPE_WRITE)
2633 vdev_cache_write(zio);
2635 if (zio_injection_enabled && zio->io_error == 0)
2636 zio->io_error = zio_handle_device_injection(vd,
2639 if (zio_injection_enabled && zio->io_error == 0)
2640 zio->io_error = zio_handle_label_injection(zio, EIO);
2642 if (zio->io_error) {
2643 if (!vdev_accessible(vd, zio)) {
2644 zio->io_error = SET_ERROR(ENXIO);
2646 unexpected_error = B_TRUE;
2651 ops->vdev_op_io_done(zio);
2653 if (unexpected_error)
2654 VERIFY(vdev_probe(vd, zio) == NULL);
2656 return (ZIO_PIPELINE_CONTINUE);
2660 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2661 * disk, and use that to finish the checksum ereport later.
2664 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2665 const void *good_buf)
2667 /* no processing needed */
2668 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2673 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2675 void *buf = zio_buf_alloc(zio->io_size);
2677 bcopy(zio->io_data, buf, zio->io_size);
2679 zcr->zcr_cbinfo = zio->io_size;
2680 zcr->zcr_cbdata = buf;
2681 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2682 zcr->zcr_free = zio_buf_free;
2686 zio_vdev_io_assess(zio_t *zio)
2688 vdev_t *vd = zio->io_vd;
2690 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2691 return (ZIO_PIPELINE_STOP);
2693 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2694 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2696 if (zio->io_vsd != NULL) {
2697 zio->io_vsd_ops->vsd_free(zio);
2701 if (zio_injection_enabled && zio->io_error == 0)
2702 zio->io_error = zio_handle_fault_injection(zio, EIO);
2704 if (zio->io_type == ZIO_TYPE_IOCTL && zio->io_cmd == DKIOCTRIM)
2705 switch (zio->io_error) {
2707 ZIO_TRIM_STAT_INCR(bytes, zio->io_size);
2708 ZIO_TRIM_STAT_BUMP(success);
2711 ZIO_TRIM_STAT_BUMP(unsupported);
2714 ZIO_TRIM_STAT_BUMP(failed);
2719 * If the I/O failed, determine whether we should attempt to retry it.
2721 * On retry, we cut in line in the issue queue, since we don't want
2722 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2724 if (zio->io_error && vd == NULL &&
2725 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2726 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2727 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2729 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2730 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2731 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2732 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2733 zio_requeue_io_start_cut_in_line);
2734 return (ZIO_PIPELINE_STOP);
2738 * If we got an error on a leaf device, convert it to ENXIO
2739 * if the device is not accessible at all.
2741 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2742 !vdev_accessible(vd, zio))
2743 zio->io_error = SET_ERROR(ENXIO);
2746 * If we can't write to an interior vdev (mirror or RAID-Z),
2747 * set vdev_cant_write so that we stop trying to allocate from it.
2749 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2750 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
2751 vd->vdev_cant_write = B_TRUE;
2755 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2757 return (ZIO_PIPELINE_CONTINUE);
2761 zio_vdev_io_reissue(zio_t *zio)
2763 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2764 ASSERT(zio->io_error == 0);
2766 zio->io_stage >>= 1;
2770 zio_vdev_io_redone(zio_t *zio)
2772 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2774 zio->io_stage >>= 1;
2778 zio_vdev_io_bypass(zio_t *zio)
2780 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2781 ASSERT(zio->io_error == 0);
2783 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2784 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2788 * ==========================================================================
2789 * Generate and verify checksums
2790 * ==========================================================================
2793 zio_checksum_generate(zio_t *zio)
2795 blkptr_t *bp = zio->io_bp;
2796 enum zio_checksum checksum;
2800 * This is zio_write_phys().
2801 * We're either generating a label checksum, or none at all.
2803 checksum = zio->io_prop.zp_checksum;
2805 if (checksum == ZIO_CHECKSUM_OFF)
2806 return (ZIO_PIPELINE_CONTINUE);
2808 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2810 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2811 ASSERT(!IO_IS_ALLOCATING(zio));
2812 checksum = ZIO_CHECKSUM_GANG_HEADER;
2814 checksum = BP_GET_CHECKSUM(bp);
2818 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2820 return (ZIO_PIPELINE_CONTINUE);
2824 zio_checksum_verify(zio_t *zio)
2826 zio_bad_cksum_t info;
2827 blkptr_t *bp = zio->io_bp;
2830 ASSERT(zio->io_vd != NULL);
2834 * This is zio_read_phys().
2835 * We're either verifying a label checksum, or nothing at all.
2837 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2838 return (ZIO_PIPELINE_CONTINUE);
2840 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2843 if ((error = zio_checksum_error(zio, &info)) != 0) {
2844 zio->io_error = error;
2845 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2846 zfs_ereport_start_checksum(zio->io_spa,
2847 zio->io_vd, zio, zio->io_offset,
2848 zio->io_size, NULL, &info);
2852 return (ZIO_PIPELINE_CONTINUE);
2856 * Called by RAID-Z to ensure we don't compute the checksum twice.
2859 zio_checksum_verified(zio_t *zio)
2861 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2865 * ==========================================================================
2866 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2867 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2868 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2869 * indicate errors that are specific to one I/O, and most likely permanent.
2870 * Any other error is presumed to be worse because we weren't expecting it.
2871 * ==========================================================================
2874 zio_worst_error(int e1, int e2)
2876 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2879 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2880 if (e1 == zio_error_rank[r1])
2883 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2884 if (e2 == zio_error_rank[r2])
2887 return (r1 > r2 ? e1 : e2);
2891 * ==========================================================================
2893 * ==========================================================================
2896 zio_ready(zio_t *zio)
2898 blkptr_t *bp = zio->io_bp;
2899 zio_t *pio, *pio_next;
2901 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2902 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2903 return (ZIO_PIPELINE_STOP);
2905 if (zio->io_ready) {
2906 ASSERT(IO_IS_ALLOCATING(zio));
2907 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
2908 (zio->io_flags & ZIO_FLAG_NOPWRITE));
2909 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2914 if (bp != NULL && bp != &zio->io_bp_copy)
2915 zio->io_bp_copy = *bp;
2918 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2920 mutex_enter(&zio->io_lock);
2921 zio->io_state[ZIO_WAIT_READY] = 1;
2922 pio = zio_walk_parents(zio);
2923 mutex_exit(&zio->io_lock);
2926 * As we notify zio's parents, new parents could be added.
2927 * New parents go to the head of zio's io_parent_list, however,
2928 * so we will (correctly) not notify them. The remainder of zio's
2929 * io_parent_list, from 'pio_next' onward, cannot change because
2930 * all parents must wait for us to be done before they can be done.
2932 for (; pio != NULL; pio = pio_next) {
2933 pio_next = zio_walk_parents(zio);
2934 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2937 if (zio->io_flags & ZIO_FLAG_NODATA) {
2938 if (BP_IS_GANG(bp)) {
2939 zio->io_flags &= ~ZIO_FLAG_NODATA;
2941 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2942 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2946 if (zio_injection_enabled &&
2947 zio->io_spa->spa_syncing_txg == zio->io_txg)
2948 zio_handle_ignored_writes(zio);
2950 return (ZIO_PIPELINE_CONTINUE);
2954 zio_done(zio_t *zio)
2956 spa_t *spa = zio->io_spa;
2957 zio_t *lio = zio->io_logical;
2958 blkptr_t *bp = zio->io_bp;
2959 vdev_t *vd = zio->io_vd;
2960 uint64_t psize = zio->io_size;
2961 zio_t *pio, *pio_next;
2964 * If our children haven't all completed,
2965 * wait for them and then repeat this pipeline stage.
2967 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2968 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2969 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2970 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2971 return (ZIO_PIPELINE_STOP);
2973 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2974 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2975 ASSERT(zio->io_children[c][w] == 0);
2978 ASSERT(bp->blk_pad[0] == 0);
2979 ASSERT(bp->blk_pad[1] == 0);
2980 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2981 (bp == zio_unique_parent(zio)->io_bp));
2982 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
2983 zio->io_bp_override == NULL &&
2984 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2985 ASSERT(!BP_SHOULD_BYTESWAP(bp));
2986 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
2987 ASSERT(BP_COUNT_GANG(bp) == 0 ||
2988 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
2990 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
2991 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
2995 * If there were child vdev/gang/ddt errors, they apply to us now.
2997 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2998 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2999 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3002 * If the I/O on the transformed data was successful, generate any
3003 * checksum reports now while we still have the transformed data.
3005 if (zio->io_error == 0) {
3006 while (zio->io_cksum_report != NULL) {
3007 zio_cksum_report_t *zcr = zio->io_cksum_report;
3008 uint64_t align = zcr->zcr_align;
3009 uint64_t asize = P2ROUNDUP(psize, align);
3010 char *abuf = zio->io_data;
3012 if (asize != psize) {
3013 abuf = zio_buf_alloc(asize);
3014 bcopy(zio->io_data, abuf, psize);
3015 bzero(abuf + psize, asize - psize);
3018 zio->io_cksum_report = zcr->zcr_next;
3019 zcr->zcr_next = NULL;
3020 zcr->zcr_finish(zcr, abuf);
3021 zfs_ereport_free_checksum(zcr);
3024 zio_buf_free(abuf, asize);
3028 zio_pop_transforms(zio); /* note: may set zio->io_error */
3030 vdev_stat_update(zio, psize);
3032 if (zio->io_error) {
3034 * If this I/O is attached to a particular vdev,
3035 * generate an error message describing the I/O failure
3036 * at the block level. We ignore these errors if the
3037 * device is currently unavailable.
3039 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3040 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3042 if ((zio->io_error == EIO || !(zio->io_flags &
3043 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3046 * For logical I/O requests, tell the SPA to log the
3047 * error and generate a logical data ereport.
3049 spa_log_error(spa, zio);
3050 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3055 if (zio->io_error && zio == lio) {
3057 * Determine whether zio should be reexecuted. This will
3058 * propagate all the way to the root via zio_notify_parent().
3060 ASSERT(vd == NULL && bp != NULL);
3061 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3063 if (IO_IS_ALLOCATING(zio) &&
3064 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3065 if (zio->io_error != ENOSPC)
3066 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3068 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3071 if ((zio->io_type == ZIO_TYPE_READ ||
3072 zio->io_type == ZIO_TYPE_FREE) &&
3073 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3074 zio->io_error == ENXIO &&
3075 spa_load_state(spa) == SPA_LOAD_NONE &&
3076 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3077 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3079 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3080 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3083 * Here is a possibly good place to attempt to do
3084 * either combinatorial reconstruction or error correction
3085 * based on checksums. It also might be a good place
3086 * to send out preliminary ereports before we suspend
3092 * If there were logical child errors, they apply to us now.
3093 * We defer this until now to avoid conflating logical child
3094 * errors with errors that happened to the zio itself when
3095 * updating vdev stats and reporting FMA events above.
3097 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3099 if ((zio->io_error || zio->io_reexecute) &&
3100 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3101 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3102 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3104 zio_gang_tree_free(&zio->io_gang_tree);
3107 * Godfather I/Os should never suspend.
3109 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3110 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3111 zio->io_reexecute = 0;
3113 if (zio->io_reexecute) {
3115 * This is a logical I/O that wants to reexecute.
3117 * Reexecute is top-down. When an i/o fails, if it's not
3118 * the root, it simply notifies its parent and sticks around.
3119 * The parent, seeing that it still has children in zio_done(),
3120 * does the same. This percolates all the way up to the root.
3121 * The root i/o will reexecute or suspend the entire tree.
3123 * This approach ensures that zio_reexecute() honors
3124 * all the original i/o dependency relationships, e.g.
3125 * parents not executing until children are ready.
3127 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3129 zio->io_gang_leader = NULL;
3131 mutex_enter(&zio->io_lock);
3132 zio->io_state[ZIO_WAIT_DONE] = 1;
3133 mutex_exit(&zio->io_lock);
3136 * "The Godfather" I/O monitors its children but is
3137 * not a true parent to them. It will track them through
3138 * the pipeline but severs its ties whenever they get into
3139 * trouble (e.g. suspended). This allows "The Godfather"
3140 * I/O to return status without blocking.
3142 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3143 zio_link_t *zl = zio->io_walk_link;
3144 pio_next = zio_walk_parents(zio);
3146 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3147 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3148 zio_remove_child(pio, zio, zl);
3149 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3153 if ((pio = zio_unique_parent(zio)) != NULL) {
3155 * We're not a root i/o, so there's nothing to do
3156 * but notify our parent. Don't propagate errors
3157 * upward since we haven't permanently failed yet.
3159 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3160 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3161 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3162 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3164 * We'd fail again if we reexecuted now, so suspend
3165 * until conditions improve (e.g. device comes online).
3167 zio_suspend(spa, zio);
3170 * Reexecution is potentially a huge amount of work.
3171 * Hand it off to the otherwise-unused claim taskq.
3174 (void) taskq_dispatch_safe(
3175 spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
3176 (task_func_t *)zio_reexecute, zio, TQ_SLEEP,
3179 (void) taskq_dispatch(
3180 spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
3181 (task_func_t *)zio_reexecute, zio, TQ_SLEEP);
3184 return (ZIO_PIPELINE_STOP);
3187 ASSERT(zio->io_child_count == 0);
3188 ASSERT(zio->io_reexecute == 0);
3189 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3192 * Report any checksum errors, since the I/O is complete.
3194 while (zio->io_cksum_report != NULL) {
3195 zio_cksum_report_t *zcr = zio->io_cksum_report;
3196 zio->io_cksum_report = zcr->zcr_next;
3197 zcr->zcr_next = NULL;
3198 zcr->zcr_finish(zcr, NULL);
3199 zfs_ereport_free_checksum(zcr);
3203 * It is the responsibility of the done callback to ensure that this
3204 * particular zio is no longer discoverable for adoption, and as
3205 * such, cannot acquire any new parents.
3210 mutex_enter(&zio->io_lock);
3211 zio->io_state[ZIO_WAIT_DONE] = 1;
3212 mutex_exit(&zio->io_lock);
3214 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3215 zio_link_t *zl = zio->io_walk_link;
3216 pio_next = zio_walk_parents(zio);
3217 zio_remove_child(pio, zio, zl);
3218 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3221 if (zio->io_waiter != NULL) {
3222 mutex_enter(&zio->io_lock);
3223 zio->io_executor = NULL;
3224 cv_broadcast(&zio->io_cv);
3225 mutex_exit(&zio->io_lock);
3230 return (ZIO_PIPELINE_STOP);
3234 * ==========================================================================
3235 * I/O pipeline definition
3236 * ==========================================================================
3238 static zio_pipe_stage_t *zio_pipeline[] = {
3244 zio_checksum_generate,
3259 zio_checksum_verify,
3263 /* dnp is the dnode for zb1->zb_object */
3265 zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_t *zb1,
3266 const zbookmark_t *zb2)
3268 uint64_t zb1nextL0, zb2thisobj;
3270 ASSERT(zb1->zb_objset == zb2->zb_objset);
3271 ASSERT(zb2->zb_level == 0);
3274 * A bookmark in the deadlist is considered to be after
3277 if (zb2->zb_object == DMU_DEADLIST_OBJECT)
3280 /* The objset_phys_t isn't before anything. */
3284 zb1nextL0 = (zb1->zb_blkid + 1) <<
3285 ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
3287 zb2thisobj = zb2->zb_object ? zb2->zb_object :
3288 zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
3290 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3291 uint64_t nextobj = zb1nextL0 *
3292 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
3293 return (nextobj <= zb2thisobj);
3296 if (zb1->zb_object < zb2thisobj)
3298 if (zb1->zb_object > zb2thisobj)
3300 if (zb2->zb_object == DMU_META_DNODE_OBJECT)
3302 return (zb1nextL0 <= zb2->zb_blkid);