4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2014 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
27 #include <sys/zfs_context.h>
28 #include <sys/fm/fs/zfs.h>
31 #include <sys/spa_impl.h>
32 #include <sys/vdev_impl.h>
33 #include <sys/zio_impl.h>
34 #include <sys/zio_compress.h>
35 #include <sys/zio_checksum.h>
36 #include <sys/dmu_objset.h>
39 #include <sys/trim_map.h>
40 #include <sys/blkptr.h>
41 #include <sys/zfeature.h>
43 SYSCTL_DECL(_vfs_zfs);
44 SYSCTL_NODE(_vfs_zfs, OID_AUTO, zio, CTLFLAG_RW, 0, "ZFS ZIO");
45 #if defined(__amd64__)
46 static int zio_use_uma = 1;
48 static int zio_use_uma = 0;
50 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, use_uma, CTLFLAG_RDTUN, &zio_use_uma, 0,
51 "Use uma(9) for ZIO allocations");
52 static int zio_exclude_metadata = 0;
53 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, exclude_metadata, CTLFLAG_RDTUN, &zio_exclude_metadata, 0,
54 "Exclude metadata buffers from dumps as well");
56 zio_trim_stats_t zio_trim_stats = {
57 { "bytes", KSTAT_DATA_UINT64,
58 "Number of bytes successfully TRIMmed" },
59 { "success", KSTAT_DATA_UINT64,
60 "Number of successful TRIM requests" },
61 { "unsupported", KSTAT_DATA_UINT64,
62 "Number of TRIM requests that failed because TRIM is not supported" },
63 { "failed", KSTAT_DATA_UINT64,
64 "Number of TRIM requests that failed for reasons other than not supported" },
67 static kstat_t *zio_trim_ksp;
70 * ==========================================================================
71 * I/O type descriptions
72 * ==========================================================================
74 const char *zio_type_name[ZIO_TYPES] = {
75 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
80 * ==========================================================================
82 * ==========================================================================
84 kmem_cache_t *zio_cache;
85 kmem_cache_t *zio_link_cache;
86 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
87 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
90 extern vmem_t *zio_alloc_arena;
94 * The following actions directly effect the spa's sync-to-convergence logic.
95 * The values below define the sync pass when we start performing the action.
96 * Care should be taken when changing these values as they directly impact
97 * spa_sync() performance. Tuning these values may introduce subtle performance
98 * pathologies and should only be done in the context of performance analysis.
99 * These tunables will eventually be removed and replaced with #defines once
100 * enough analysis has been done to determine optimal values.
102 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
103 * regular blocks are not deferred.
105 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
106 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_deferred_free, CTLFLAG_RDTUN,
107 &zfs_sync_pass_deferred_free, 0, "defer frees starting in this pass");
108 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
109 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_dont_compress, CTLFLAG_RDTUN,
110 &zfs_sync_pass_dont_compress, 0, "don't compress starting in this pass");
111 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
112 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_rewrite, CTLFLAG_RDTUN,
113 &zfs_sync_pass_rewrite, 0, "rewrite new bps starting in this pass");
116 * An allocating zio is one that either currently has the DVA allocate
117 * stage set or will have it later in its lifetime.
119 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
121 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
124 int zio_buf_debug_limit = 16384;
126 int zio_buf_debug_limit = 0;
133 zio_cache = kmem_cache_create("zio_cache",
134 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
135 zio_link_cache = kmem_cache_create("zio_link_cache",
136 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
141 * For small buffers, we want a cache for each multiple of
142 * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache
143 * for each quarter-power of 2. For large buffers, we want
144 * a cache for each multiple of PAGESIZE.
146 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
147 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
150 size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
152 while (p2 & (p2 - 1))
158 * If we are using watchpoints, put each buffer on its own page,
159 * to eliminate the performance overhead of trapping to the
160 * kernel when modifying a non-watched buffer that shares the
161 * page with a watched buffer.
163 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
167 if (size <= 4 * SPA_MINBLOCKSIZE) {
168 align = SPA_MINBLOCKSIZE;
169 } else if (IS_P2ALIGNED(size, PAGESIZE)) {
171 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
177 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
178 zio_buf_cache[c] = kmem_cache_create(name, size,
179 align, NULL, NULL, NULL, NULL, NULL, cflags);
182 * Since zio_data bufs do not appear in crash dumps, we
183 * pass KMC_NOTOUCH so that no allocator metadata is
184 * stored with the buffers.
186 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
187 zio_data_buf_cache[c] = kmem_cache_create(name, size,
188 align, NULL, NULL, NULL, NULL, NULL,
189 cflags | KMC_NOTOUCH | KMC_NODEBUG);
194 ASSERT(zio_buf_cache[c] != NULL);
195 if (zio_buf_cache[c - 1] == NULL)
196 zio_buf_cache[c - 1] = zio_buf_cache[c];
198 ASSERT(zio_data_buf_cache[c] != NULL);
199 if (zio_data_buf_cache[c - 1] == NULL)
200 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
206 zio_trim_ksp = kstat_create("zfs", 0, "zio_trim", "misc",
208 sizeof(zio_trim_stats) / sizeof(kstat_named_t),
211 if (zio_trim_ksp != NULL) {
212 zio_trim_ksp->ks_data = &zio_trim_stats;
213 kstat_install(zio_trim_ksp);
221 kmem_cache_t *last_cache = NULL;
222 kmem_cache_t *last_data_cache = NULL;
224 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
225 if (zio_buf_cache[c] != last_cache) {
226 last_cache = zio_buf_cache[c];
227 kmem_cache_destroy(zio_buf_cache[c]);
229 zio_buf_cache[c] = NULL;
231 if (zio_data_buf_cache[c] != last_data_cache) {
232 last_data_cache = zio_data_buf_cache[c];
233 kmem_cache_destroy(zio_data_buf_cache[c]);
235 zio_data_buf_cache[c] = NULL;
238 kmem_cache_destroy(zio_link_cache);
239 kmem_cache_destroy(zio_cache);
243 if (zio_trim_ksp != NULL) {
244 kstat_delete(zio_trim_ksp);
250 * ==========================================================================
251 * Allocate and free I/O buffers
252 * ==========================================================================
256 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
257 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
258 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
259 * excess / transient data in-core during a crashdump.
262 zio_buf_alloc(size_t size)
264 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
265 int flags = zio_exclude_metadata ? KM_NODEBUG : 0;
267 ASSERT3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
270 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
272 return (kmem_alloc(size, KM_SLEEP|flags));
276 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
277 * crashdump if the kernel panics. This exists so that we will limit the amount
278 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
279 * of kernel heap dumped to disk when the kernel panics)
282 zio_data_buf_alloc(size_t size)
284 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
286 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
289 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
291 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG));
295 zio_buf_free(void *buf, size_t size)
297 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
299 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
302 kmem_cache_free(zio_buf_cache[c], buf);
304 kmem_free(buf, size);
308 zio_data_buf_free(void *buf, size_t size)
310 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
312 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
315 kmem_cache_free(zio_data_buf_cache[c], buf);
317 kmem_free(buf, size);
321 * ==========================================================================
322 * Push and pop I/O transform buffers
323 * ==========================================================================
326 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
327 zio_transform_func_t *transform)
329 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
331 zt->zt_orig_data = zio->io_data;
332 zt->zt_orig_size = zio->io_size;
333 zt->zt_bufsize = bufsize;
334 zt->zt_transform = transform;
336 zt->zt_next = zio->io_transform_stack;
337 zio->io_transform_stack = zt;
344 zio_pop_transforms(zio_t *zio)
348 while ((zt = zio->io_transform_stack) != NULL) {
349 if (zt->zt_transform != NULL)
350 zt->zt_transform(zio,
351 zt->zt_orig_data, zt->zt_orig_size);
353 if (zt->zt_bufsize != 0)
354 zio_buf_free(zio->io_data, zt->zt_bufsize);
356 zio->io_data = zt->zt_orig_data;
357 zio->io_size = zt->zt_orig_size;
358 zio->io_transform_stack = zt->zt_next;
360 kmem_free(zt, sizeof (zio_transform_t));
365 * ==========================================================================
366 * I/O transform callbacks for subblocks and decompression
367 * ==========================================================================
370 zio_subblock(zio_t *zio, void *data, uint64_t size)
372 ASSERT(zio->io_size > size);
374 if (zio->io_type == ZIO_TYPE_READ)
375 bcopy(zio->io_data, data, size);
379 zio_decompress(zio_t *zio, void *data, uint64_t size)
381 if (zio->io_error == 0 &&
382 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
383 zio->io_data, data, zio->io_size, size) != 0)
384 zio->io_error = SET_ERROR(EIO);
388 * ==========================================================================
389 * I/O parent/child relationships and pipeline interlocks
390 * ==========================================================================
393 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
394 * continue calling these functions until they return NULL.
395 * Otherwise, the next caller will pick up the list walk in
396 * some indeterminate state. (Otherwise every caller would
397 * have to pass in a cookie to keep the state represented by
398 * io_walk_link, which gets annoying.)
401 zio_walk_parents(zio_t *cio)
403 zio_link_t *zl = cio->io_walk_link;
404 list_t *pl = &cio->io_parent_list;
406 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
407 cio->io_walk_link = zl;
412 ASSERT(zl->zl_child == cio);
413 return (zl->zl_parent);
417 zio_walk_children(zio_t *pio)
419 zio_link_t *zl = pio->io_walk_link;
420 list_t *cl = &pio->io_child_list;
422 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
423 pio->io_walk_link = zl;
428 ASSERT(zl->zl_parent == pio);
429 return (zl->zl_child);
433 zio_unique_parent(zio_t *cio)
435 zio_t *pio = zio_walk_parents(cio);
437 VERIFY(zio_walk_parents(cio) == NULL);
442 zio_add_child(zio_t *pio, zio_t *cio)
444 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
447 * Logical I/Os can have logical, gang, or vdev children.
448 * Gang I/Os can have gang or vdev children.
449 * Vdev I/Os can only have vdev children.
450 * The following ASSERT captures all of these constraints.
452 ASSERT(cio->io_child_type <= pio->io_child_type);
457 mutex_enter(&cio->io_lock);
458 mutex_enter(&pio->io_lock);
460 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
462 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
463 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
465 list_insert_head(&pio->io_child_list, zl);
466 list_insert_head(&cio->io_parent_list, zl);
468 pio->io_child_count++;
469 cio->io_parent_count++;
471 mutex_exit(&pio->io_lock);
472 mutex_exit(&cio->io_lock);
476 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
478 ASSERT(zl->zl_parent == pio);
479 ASSERT(zl->zl_child == cio);
481 mutex_enter(&cio->io_lock);
482 mutex_enter(&pio->io_lock);
484 list_remove(&pio->io_child_list, zl);
485 list_remove(&cio->io_parent_list, zl);
487 pio->io_child_count--;
488 cio->io_parent_count--;
490 mutex_exit(&pio->io_lock);
491 mutex_exit(&cio->io_lock);
493 kmem_cache_free(zio_link_cache, zl);
497 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
499 uint64_t *countp = &zio->io_children[child][wait];
500 boolean_t waiting = B_FALSE;
502 mutex_enter(&zio->io_lock);
503 ASSERT(zio->io_stall == NULL);
506 zio->io_stall = countp;
509 mutex_exit(&zio->io_lock);
515 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
517 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
518 int *errorp = &pio->io_child_error[zio->io_child_type];
520 mutex_enter(&pio->io_lock);
521 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
522 *errorp = zio_worst_error(*errorp, zio->io_error);
523 pio->io_reexecute |= zio->io_reexecute;
524 ASSERT3U(*countp, >, 0);
528 if (*countp == 0 && pio->io_stall == countp) {
529 pio->io_stall = NULL;
530 mutex_exit(&pio->io_lock);
533 mutex_exit(&pio->io_lock);
538 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
540 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
541 zio->io_error = zio->io_child_error[c];
545 * ==========================================================================
546 * Create the various types of I/O (read, write, free, etc)
547 * ==========================================================================
550 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
551 void *data, uint64_t size, zio_done_func_t *done, void *private,
552 zio_type_t type, zio_priority_t priority, enum zio_flag flags,
553 vdev_t *vd, uint64_t offset, const zbookmark_phys_t *zb,
554 enum zio_stage stage, enum zio_stage pipeline)
558 ASSERT3U(type == ZIO_TYPE_FREE || size, <=, SPA_MAXBLOCKSIZE);
559 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
560 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
562 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
563 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
564 ASSERT(vd || stage == ZIO_STAGE_OPEN);
566 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
567 bzero(zio, sizeof (zio_t));
569 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
570 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
572 list_create(&zio->io_parent_list, sizeof (zio_link_t),
573 offsetof(zio_link_t, zl_parent_node));
574 list_create(&zio->io_child_list, sizeof (zio_link_t),
575 offsetof(zio_link_t, zl_child_node));
578 zio->io_child_type = ZIO_CHILD_VDEV;
579 else if (flags & ZIO_FLAG_GANG_CHILD)
580 zio->io_child_type = ZIO_CHILD_GANG;
581 else if (flags & ZIO_FLAG_DDT_CHILD)
582 zio->io_child_type = ZIO_CHILD_DDT;
584 zio->io_child_type = ZIO_CHILD_LOGICAL;
587 zio->io_bp = (blkptr_t *)bp;
588 zio->io_bp_copy = *bp;
589 zio->io_bp_orig = *bp;
590 if (type != ZIO_TYPE_WRITE ||
591 zio->io_child_type == ZIO_CHILD_DDT)
592 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
593 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
594 zio->io_logical = zio;
595 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
596 pipeline |= ZIO_GANG_STAGES;
602 zio->io_private = private;
604 zio->io_priority = priority;
606 zio->io_offset = offset;
607 zio->io_orig_data = zio->io_data = data;
608 zio->io_orig_size = zio->io_size = size;
609 zio->io_orig_flags = zio->io_flags = flags;
610 zio->io_orig_stage = zio->io_stage = stage;
611 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
613 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
614 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
617 zio->io_bookmark = *zb;
620 if (zio->io_logical == NULL)
621 zio->io_logical = pio->io_logical;
622 if (zio->io_child_type == ZIO_CHILD_GANG)
623 zio->io_gang_leader = pio->io_gang_leader;
624 zio_add_child(pio, zio);
631 zio_destroy(zio_t *zio)
633 list_destroy(&zio->io_parent_list);
634 list_destroy(&zio->io_child_list);
635 mutex_destroy(&zio->io_lock);
636 cv_destroy(&zio->io_cv);
637 kmem_cache_free(zio_cache, zio);
641 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
642 void *private, enum zio_flag flags)
646 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
647 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
648 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
654 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
656 return (zio_null(NULL, spa, NULL, done, private, flags));
660 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
661 void *data, uint64_t size, zio_done_func_t *done, void *private,
662 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
666 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
667 data, size, done, private,
668 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
669 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
670 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
676 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
677 void *data, uint64_t size, const zio_prop_t *zp,
678 zio_done_func_t *ready, zio_done_func_t *physdone, zio_done_func_t *done,
680 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
684 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
685 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
686 zp->zp_compress >= ZIO_COMPRESS_OFF &&
687 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
688 DMU_OT_IS_VALID(zp->zp_type) &&
691 zp->zp_copies <= spa_max_replication(spa));
693 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
694 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
695 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
696 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
698 zio->io_ready = ready;
699 zio->io_physdone = physdone;
703 * Data can be NULL if we are going to call zio_write_override() to
704 * provide the already-allocated BP. But we may need the data to
705 * verify a dedup hit (if requested). In this case, don't try to
706 * dedup (just take the already-allocated BP verbatim).
708 if (data == NULL && zio->io_prop.zp_dedup_verify) {
709 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
716 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
717 uint64_t size, zio_done_func_t *done, void *private,
718 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
722 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
723 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
724 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
730 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
732 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
733 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
734 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
735 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
738 * We must reset the io_prop to match the values that existed
739 * when the bp was first written by dmu_sync() keeping in mind
740 * that nopwrite and dedup are mutually exclusive.
742 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
743 zio->io_prop.zp_nopwrite = nopwrite;
744 zio->io_prop.zp_copies = copies;
745 zio->io_bp_override = bp;
749 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
753 * The check for EMBEDDED is a performance optimization. We
754 * process the free here (by ignoring it) rather than
755 * putting it on the list and then processing it in zio_free_sync().
757 if (BP_IS_EMBEDDED(bp))
759 metaslab_check_free(spa, bp);
762 * Frees that are for the currently-syncing txg, are not going to be
763 * deferred, and which will not need to do a read (i.e. not GANG or
764 * DEDUP), can be processed immediately. Otherwise, put them on the
765 * in-memory list for later processing.
767 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
768 txg != spa->spa_syncing_txg ||
769 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
770 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
772 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp,
773 BP_GET_PSIZE(bp), 0)));
778 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
779 uint64_t size, enum zio_flag flags)
782 enum zio_stage stage = ZIO_FREE_PIPELINE;
784 ASSERT(!BP_IS_HOLE(bp));
785 ASSERT(spa_syncing_txg(spa) == txg);
786 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
788 if (BP_IS_EMBEDDED(bp))
789 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
791 metaslab_check_free(spa, bp);
794 if (zfs_trim_enabled)
795 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START |
796 ZIO_STAGE_VDEV_IO_ASSESS;
798 * GANG and DEDUP blocks can induce a read (for the gang block header,
799 * or the DDT), so issue them asynchronously so that this thread is
802 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
803 stage |= ZIO_STAGE_ISSUE_ASYNC;
805 flags |= ZIO_FLAG_DONT_QUEUE;
807 zio = zio_create(pio, spa, txg, bp, NULL, size,
808 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags,
809 NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
815 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
816 zio_done_func_t *done, void *private, enum zio_flag flags)
820 dprintf_bp(bp, "claiming in txg %llu", txg);
822 if (BP_IS_EMBEDDED(bp))
823 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
826 * A claim is an allocation of a specific block. Claims are needed
827 * to support immediate writes in the intent log. The issue is that
828 * immediate writes contain committed data, but in a txg that was
829 * *not* committed. Upon opening the pool after an unclean shutdown,
830 * the intent log claims all blocks that contain immediate write data
831 * so that the SPA knows they're in use.
833 * All claims *must* be resolved in the first txg -- before the SPA
834 * starts allocating blocks -- so that nothing is allocated twice.
835 * If txg == 0 we just verify that the block is claimable.
837 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
838 ASSERT(txg == spa_first_txg(spa) || txg == 0);
839 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
841 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
842 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
843 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
849 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset,
850 uint64_t size, zio_done_func_t *done, void *private,
851 zio_priority_t priority, enum zio_flag flags)
856 if (vd->vdev_children == 0) {
857 zio = zio_create(pio, spa, 0, NULL, NULL, size, done, private,
858 ZIO_TYPE_IOCTL, priority, flags, vd, offset, NULL,
859 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
863 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
865 for (c = 0; c < vd->vdev_children; c++)
866 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
867 offset, size, done, private, priority, flags));
874 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
875 void *data, int checksum, zio_done_func_t *done, void *private,
876 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
880 ASSERT(vd->vdev_children == 0);
881 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
882 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
883 ASSERT3U(offset + size, <=, vd->vdev_psize);
885 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
886 ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
887 NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
889 zio->io_prop.zp_checksum = checksum;
895 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
896 void *data, int checksum, zio_done_func_t *done, void *private,
897 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
901 ASSERT(vd->vdev_children == 0);
902 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
903 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
904 ASSERT3U(offset + size, <=, vd->vdev_psize);
906 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
907 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
908 NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
910 zio->io_prop.zp_checksum = checksum;
912 if (zio_checksum_table[checksum].ci_eck) {
914 * zec checksums are necessarily destructive -- they modify
915 * the end of the write buffer to hold the verifier/checksum.
916 * Therefore, we must make a local copy in case the data is
917 * being written to multiple places in parallel.
919 void *wbuf = zio_buf_alloc(size);
920 bcopy(data, wbuf, size);
921 zio_push_transform(zio, wbuf, size, size, NULL);
928 * Create a child I/O to do some work for us.
931 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
932 void *data, uint64_t size, int type, zio_priority_t priority,
933 enum zio_flag flags, zio_done_func_t *done, void *private)
935 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
938 ASSERT(vd->vdev_parent ==
939 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
941 if (type == ZIO_TYPE_READ && bp != NULL) {
943 * If we have the bp, then the child should perform the
944 * checksum and the parent need not. This pushes error
945 * detection as close to the leaves as possible and
946 * eliminates redundant checksums in the interior nodes.
948 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
949 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
952 /* Not all IO types require vdev io done stage e.g. free */
953 if (!(pio->io_pipeline & ZIO_STAGE_VDEV_IO_DONE))
954 pipeline &= ~ZIO_STAGE_VDEV_IO_DONE;
956 if (vd->vdev_children == 0)
957 offset += VDEV_LABEL_START_SIZE;
959 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
962 * If we've decided to do a repair, the write is not speculative --
963 * even if the original read was.
965 if (flags & ZIO_FLAG_IO_REPAIR)
966 flags &= ~ZIO_FLAG_SPECULATIVE;
968 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
969 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
970 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
972 zio->io_physdone = pio->io_physdone;
973 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
974 zio->io_logical->io_phys_children++;
980 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
981 int type, zio_priority_t priority, enum zio_flag flags,
982 zio_done_func_t *done, void *private)
986 ASSERT(vd->vdev_ops->vdev_op_leaf);
988 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
989 data, size, done, private, type, priority,
990 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
992 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
998 zio_flush(zio_t *zio, vdev_t *vd)
1000 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0,
1001 NULL, NULL, ZIO_PRIORITY_NOW,
1002 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1006 zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size)
1009 ASSERT(vd->vdev_ops->vdev_op_leaf);
1011 return (zio_create(zio, spa, 0, NULL, NULL, size, NULL, NULL,
1012 ZIO_TYPE_FREE, ZIO_PRIORITY_TRIM, ZIO_FLAG_DONT_AGGREGATE |
1013 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY,
1014 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PHYS_PIPELINE));
1018 zio_shrink(zio_t *zio, uint64_t size)
1020 ASSERT(zio->io_executor == NULL);
1021 ASSERT(zio->io_orig_size == zio->io_size);
1022 ASSERT(size <= zio->io_size);
1025 * We don't shrink for raidz because of problems with the
1026 * reconstruction when reading back less than the block size.
1027 * Note, BP_IS_RAIDZ() assumes no compression.
1029 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1030 if (!BP_IS_RAIDZ(zio->io_bp))
1031 zio->io_orig_size = zio->io_size = size;
1035 * ==========================================================================
1036 * Prepare to read and write logical blocks
1037 * ==========================================================================
1041 zio_read_bp_init(zio_t **ziop)
1044 blkptr_t *bp = zio->io_bp;
1046 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1047 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1048 !(zio->io_flags & ZIO_FLAG_RAW)) {
1050 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1051 void *cbuf = zio_buf_alloc(psize);
1053 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
1056 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1057 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1058 decode_embedded_bp_compressed(bp, zio->io_data);
1060 ASSERT(!BP_IS_EMBEDDED(bp));
1063 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1064 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1066 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1067 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1069 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1070 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1072 return (ZIO_PIPELINE_CONTINUE);
1076 zio_write_bp_init(zio_t **ziop)
1079 spa_t *spa = zio->io_spa;
1080 zio_prop_t *zp = &zio->io_prop;
1081 enum zio_compress compress = zp->zp_compress;
1082 blkptr_t *bp = zio->io_bp;
1083 uint64_t lsize = zio->io_size;
1084 uint64_t psize = lsize;
1088 * If our children haven't all reached the ready stage,
1089 * wait for them and then repeat this pipeline stage.
1091 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1092 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1093 return (ZIO_PIPELINE_STOP);
1095 if (!IO_IS_ALLOCATING(zio))
1096 return (ZIO_PIPELINE_CONTINUE);
1098 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1100 if (zio->io_bp_override) {
1101 ASSERT(bp->blk_birth != zio->io_txg);
1102 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1104 *bp = *zio->io_bp_override;
1105 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1107 if (BP_IS_EMBEDDED(bp))
1108 return (ZIO_PIPELINE_CONTINUE);
1111 * If we've been overridden and nopwrite is set then
1112 * set the flag accordingly to indicate that a nopwrite
1113 * has already occurred.
1115 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1116 ASSERT(!zp->zp_dedup);
1117 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1118 return (ZIO_PIPELINE_CONTINUE);
1121 ASSERT(!zp->zp_nopwrite);
1123 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1124 return (ZIO_PIPELINE_CONTINUE);
1126 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1127 zp->zp_dedup_verify);
1129 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1130 BP_SET_DEDUP(bp, 1);
1131 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1132 return (ZIO_PIPELINE_CONTINUE);
1134 zio->io_bp_override = NULL;
1138 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1140 * We're rewriting an existing block, which means we're
1141 * working on behalf of spa_sync(). For spa_sync() to
1142 * converge, it must eventually be the case that we don't
1143 * have to allocate new blocks. But compression changes
1144 * the blocksize, which forces a reallocate, and makes
1145 * convergence take longer. Therefore, after the first
1146 * few passes, stop compressing to ensure convergence.
1148 pass = spa_sync_pass(spa);
1150 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1151 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1152 ASSERT(!BP_GET_DEDUP(bp));
1154 if (pass >= zfs_sync_pass_dont_compress)
1155 compress = ZIO_COMPRESS_OFF;
1157 /* Make sure someone doesn't change their mind on overwrites */
1158 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1159 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1162 if (compress != ZIO_COMPRESS_OFF) {
1163 void *cbuf = zio_buf_alloc(lsize);
1164 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1165 if (psize == 0 || psize == lsize) {
1166 compress = ZIO_COMPRESS_OFF;
1167 zio_buf_free(cbuf, lsize);
1168 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1169 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1170 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1171 encode_embedded_bp_compressed(bp,
1172 cbuf, compress, lsize, psize);
1173 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1174 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1175 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1176 zio_buf_free(cbuf, lsize);
1177 bp->blk_birth = zio->io_txg;
1178 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1179 ASSERT(spa_feature_is_active(spa,
1180 SPA_FEATURE_EMBEDDED_DATA));
1181 return (ZIO_PIPELINE_CONTINUE);
1184 * Round up compressed size to MINBLOCKSIZE and
1188 P2ROUNDUP(psize, (size_t)SPA_MINBLOCKSIZE);
1189 if (rounded > psize) {
1190 bzero((char *)cbuf + psize, rounded - psize);
1193 if (psize == lsize) {
1194 compress = ZIO_COMPRESS_OFF;
1195 zio_buf_free(cbuf, lsize);
1197 zio_push_transform(zio, cbuf,
1198 psize, lsize, NULL);
1204 * The final pass of spa_sync() must be all rewrites, but the first
1205 * few passes offer a trade-off: allocating blocks defers convergence,
1206 * but newly allocated blocks are sequential, so they can be written
1207 * to disk faster. Therefore, we allow the first few passes of
1208 * spa_sync() to allocate new blocks, but force rewrites after that.
1209 * There should only be a handful of blocks after pass 1 in any case.
1211 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1212 BP_GET_PSIZE(bp) == psize &&
1213 pass >= zfs_sync_pass_rewrite) {
1215 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1216 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1217 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1220 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1224 if (zio->io_bp_orig.blk_birth != 0 &&
1225 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1226 BP_SET_LSIZE(bp, lsize);
1227 BP_SET_TYPE(bp, zp->zp_type);
1228 BP_SET_LEVEL(bp, zp->zp_level);
1229 BP_SET_BIRTH(bp, zio->io_txg, 0);
1231 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1233 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1234 BP_SET_LSIZE(bp, lsize);
1235 BP_SET_TYPE(bp, zp->zp_type);
1236 BP_SET_LEVEL(bp, zp->zp_level);
1237 BP_SET_PSIZE(bp, psize);
1238 BP_SET_COMPRESS(bp, compress);
1239 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1240 BP_SET_DEDUP(bp, zp->zp_dedup);
1241 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1243 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1244 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1245 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1247 if (zp->zp_nopwrite) {
1248 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1249 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1250 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1254 return (ZIO_PIPELINE_CONTINUE);
1258 zio_free_bp_init(zio_t **ziop)
1261 blkptr_t *bp = zio->io_bp;
1263 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1264 if (BP_GET_DEDUP(bp))
1265 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1268 return (ZIO_PIPELINE_CONTINUE);
1272 * ==========================================================================
1273 * Execute the I/O pipeline
1274 * ==========================================================================
1278 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1280 spa_t *spa = zio->io_spa;
1281 zio_type_t t = zio->io_type;
1282 int flags = (cutinline ? TQ_FRONT : 0);
1284 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1287 * If we're a config writer or a probe, the normal issue and
1288 * interrupt threads may all be blocked waiting for the config lock.
1289 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1291 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1295 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1297 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1301 * If this is a high priority I/O, then use the high priority taskq if
1304 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1305 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1308 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1311 * NB: We are assuming that the zio can only be dispatched
1312 * to a single taskq at a time. It would be a grievous error
1313 * to dispatch the zio to another taskq at the same time.
1315 #if defined(illumos) || !defined(_KERNEL)
1316 ASSERT(zio->io_tqent.tqent_next == NULL);
1318 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
1320 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1321 flags, &zio->io_tqent);
1325 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1327 kthread_t *executor = zio->io_executor;
1328 spa_t *spa = zio->io_spa;
1330 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1331 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1333 for (i = 0; i < tqs->stqs_count; i++) {
1334 if (taskq_member(tqs->stqs_taskq[i], executor))
1343 zio_issue_async(zio_t **ziop)
1347 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1349 return (ZIO_PIPELINE_STOP);
1353 zio_interrupt(zio_t *zio)
1355 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1359 * Execute the I/O pipeline until one of the following occurs:
1361 * (1) the I/O completes
1362 * (2) the pipeline stalls waiting for dependent child I/Os
1363 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1364 * (4) the I/O is delegated by vdev-level caching or aggregation
1365 * (5) the I/O is deferred due to vdev-level queueing
1366 * (6) the I/O is handed off to another thread.
1368 * In all cases, the pipeline stops whenever there's no CPU work; it never
1369 * burns a thread in cv_wait().
1371 * There's no locking on io_stage because there's no legitimate way
1372 * for multiple threads to be attempting to process the same I/O.
1374 static zio_pipe_stage_t *zio_pipeline[];
1377 zio_execute(zio_t *zio)
1379 zio->io_executor = curthread;
1381 while (zio->io_stage < ZIO_STAGE_DONE) {
1382 enum zio_stage pipeline = zio->io_pipeline;
1383 enum zio_stage stage = zio->io_stage;
1386 ASSERT(!MUTEX_HELD(&zio->io_lock));
1387 ASSERT(ISP2(stage));
1388 ASSERT(zio->io_stall == NULL);
1392 } while ((stage & pipeline) == 0);
1394 ASSERT(stage <= ZIO_STAGE_DONE);
1397 * If we are in interrupt context and this pipeline stage
1398 * will grab a config lock that is held across I/O,
1399 * or may wait for an I/O that needs an interrupt thread
1400 * to complete, issue async to avoid deadlock.
1402 * For VDEV_IO_START, we cut in line so that the io will
1403 * be sent to disk promptly.
1405 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1406 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1407 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1408 zio_requeue_io_start_cut_in_line : B_FALSE;
1409 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1413 zio->io_stage = stage;
1414 rv = zio_pipeline[highbit64(stage) - 1](&zio);
1416 if (rv == ZIO_PIPELINE_STOP)
1419 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1424 * ==========================================================================
1425 * Initiate I/O, either sync or async
1426 * ==========================================================================
1429 zio_wait(zio_t *zio)
1433 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1434 ASSERT(zio->io_executor == NULL);
1436 zio->io_waiter = curthread;
1440 mutex_enter(&zio->io_lock);
1441 while (zio->io_executor != NULL)
1442 cv_wait(&zio->io_cv, &zio->io_lock);
1443 mutex_exit(&zio->io_lock);
1445 error = zio->io_error;
1452 zio_nowait(zio_t *zio)
1454 ASSERT(zio->io_executor == NULL);
1456 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1457 zio_unique_parent(zio) == NULL) {
1459 * This is a logical async I/O with no parent to wait for it.
1460 * We add it to the spa_async_root_zio "Godfather" I/O which
1461 * will ensure they complete prior to unloading the pool.
1463 spa_t *spa = zio->io_spa;
1465 zio_add_child(spa->spa_async_zio_root, zio);
1472 * ==========================================================================
1473 * Reexecute or suspend/resume failed I/O
1474 * ==========================================================================
1478 zio_reexecute(zio_t *pio)
1480 zio_t *cio, *cio_next;
1482 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1483 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1484 ASSERT(pio->io_gang_leader == NULL);
1485 ASSERT(pio->io_gang_tree == NULL);
1487 pio->io_flags = pio->io_orig_flags;
1488 pio->io_stage = pio->io_orig_stage;
1489 pio->io_pipeline = pio->io_orig_pipeline;
1490 pio->io_reexecute = 0;
1491 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1493 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1494 pio->io_state[w] = 0;
1495 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1496 pio->io_child_error[c] = 0;
1498 if (IO_IS_ALLOCATING(pio))
1499 BP_ZERO(pio->io_bp);
1502 * As we reexecute pio's children, new children could be created.
1503 * New children go to the head of pio's io_child_list, however,
1504 * so we will (correctly) not reexecute them. The key is that
1505 * the remainder of pio's io_child_list, from 'cio_next' onward,
1506 * cannot be affected by any side effects of reexecuting 'cio'.
1508 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1509 cio_next = zio_walk_children(pio);
1510 mutex_enter(&pio->io_lock);
1511 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1512 pio->io_children[cio->io_child_type][w]++;
1513 mutex_exit(&pio->io_lock);
1518 * Now that all children have been reexecuted, execute the parent.
1519 * We don't reexecute "The Godfather" I/O here as it's the
1520 * responsibility of the caller to wait on him.
1522 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1527 zio_suspend(spa_t *spa, zio_t *zio)
1529 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1530 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1531 "failure and the failure mode property for this pool "
1532 "is set to panic.", spa_name(spa));
1534 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1536 mutex_enter(&spa->spa_suspend_lock);
1538 if (spa->spa_suspend_zio_root == NULL)
1539 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1540 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1541 ZIO_FLAG_GODFATHER);
1543 spa->spa_suspended = B_TRUE;
1546 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1547 ASSERT(zio != spa->spa_suspend_zio_root);
1548 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1549 ASSERT(zio_unique_parent(zio) == NULL);
1550 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1551 zio_add_child(spa->spa_suspend_zio_root, zio);
1554 mutex_exit(&spa->spa_suspend_lock);
1558 zio_resume(spa_t *spa)
1563 * Reexecute all previously suspended i/o.
1565 mutex_enter(&spa->spa_suspend_lock);
1566 spa->spa_suspended = B_FALSE;
1567 cv_broadcast(&spa->spa_suspend_cv);
1568 pio = spa->spa_suspend_zio_root;
1569 spa->spa_suspend_zio_root = NULL;
1570 mutex_exit(&spa->spa_suspend_lock);
1576 return (zio_wait(pio));
1580 zio_resume_wait(spa_t *spa)
1582 mutex_enter(&spa->spa_suspend_lock);
1583 while (spa_suspended(spa))
1584 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1585 mutex_exit(&spa->spa_suspend_lock);
1589 * ==========================================================================
1592 * A gang block is a collection of small blocks that looks to the DMU
1593 * like one large block. When zio_dva_allocate() cannot find a block
1594 * of the requested size, due to either severe fragmentation or the pool
1595 * being nearly full, it calls zio_write_gang_block() to construct the
1596 * block from smaller fragments.
1598 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1599 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1600 * an indirect block: it's an array of block pointers. It consumes
1601 * only one sector and hence is allocatable regardless of fragmentation.
1602 * The gang header's bps point to its gang members, which hold the data.
1604 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1605 * as the verifier to ensure uniqueness of the SHA256 checksum.
1606 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1607 * not the gang header. This ensures that data block signatures (needed for
1608 * deduplication) are independent of how the block is physically stored.
1610 * Gang blocks can be nested: a gang member may itself be a gang block.
1611 * Thus every gang block is a tree in which root and all interior nodes are
1612 * gang headers, and the leaves are normal blocks that contain user data.
1613 * The root of the gang tree is called the gang leader.
1615 * To perform any operation (read, rewrite, free, claim) on a gang block,
1616 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1617 * in the io_gang_tree field of the original logical i/o by recursively
1618 * reading the gang leader and all gang headers below it. This yields
1619 * an in-core tree containing the contents of every gang header and the
1620 * bps for every constituent of the gang block.
1622 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1623 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1624 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1625 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1626 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1627 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1628 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1629 * of the gang header plus zio_checksum_compute() of the data to update the
1630 * gang header's blk_cksum as described above.
1632 * The two-phase assemble/issue model solves the problem of partial failure --
1633 * what if you'd freed part of a gang block but then couldn't read the
1634 * gang header for another part? Assembling the entire gang tree first
1635 * ensures that all the necessary gang header I/O has succeeded before
1636 * starting the actual work of free, claim, or write. Once the gang tree
1637 * is assembled, free and claim are in-memory operations that cannot fail.
1639 * In the event that a gang write fails, zio_dva_unallocate() walks the
1640 * gang tree to immediately free (i.e. insert back into the space map)
1641 * everything we've allocated. This ensures that we don't get ENOSPC
1642 * errors during repeated suspend/resume cycles due to a flaky device.
1644 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1645 * the gang tree, we won't modify the block, so we can safely defer the free
1646 * (knowing that the block is still intact). If we *can* assemble the gang
1647 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1648 * each constituent bp and we can allocate a new block on the next sync pass.
1650 * In all cases, the gang tree allows complete recovery from partial failure.
1651 * ==========================================================================
1655 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1660 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1661 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1662 &pio->io_bookmark));
1666 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1671 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1672 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1673 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1675 * As we rewrite each gang header, the pipeline will compute
1676 * a new gang block header checksum for it; but no one will
1677 * compute a new data checksum, so we do that here. The one
1678 * exception is the gang leader: the pipeline already computed
1679 * its data checksum because that stage precedes gang assembly.
1680 * (Presently, nothing actually uses interior data checksums;
1681 * this is just good hygiene.)
1683 if (gn != pio->io_gang_leader->io_gang_tree) {
1684 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1685 data, BP_GET_PSIZE(bp));
1688 * If we are here to damage data for testing purposes,
1689 * leave the GBH alone so that we can detect the damage.
1691 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1692 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1694 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1695 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1696 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1704 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1706 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1707 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp),
1708 ZIO_GANG_CHILD_FLAGS(pio)));
1713 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1715 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1716 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1719 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1728 static void zio_gang_tree_assemble_done(zio_t *zio);
1730 static zio_gang_node_t *
1731 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1733 zio_gang_node_t *gn;
1735 ASSERT(*gnpp == NULL);
1737 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1738 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1745 zio_gang_node_free(zio_gang_node_t **gnpp)
1747 zio_gang_node_t *gn = *gnpp;
1749 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1750 ASSERT(gn->gn_child[g] == NULL);
1752 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1753 kmem_free(gn, sizeof (*gn));
1758 zio_gang_tree_free(zio_gang_node_t **gnpp)
1760 zio_gang_node_t *gn = *gnpp;
1765 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1766 zio_gang_tree_free(&gn->gn_child[g]);
1768 zio_gang_node_free(gnpp);
1772 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1774 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1776 ASSERT(gio->io_gang_leader == gio);
1777 ASSERT(BP_IS_GANG(bp));
1779 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1780 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1781 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1785 zio_gang_tree_assemble_done(zio_t *zio)
1787 zio_t *gio = zio->io_gang_leader;
1788 zio_gang_node_t *gn = zio->io_private;
1789 blkptr_t *bp = zio->io_bp;
1791 ASSERT(gio == zio_unique_parent(zio));
1792 ASSERT(zio->io_child_count == 0);
1797 if (BP_SHOULD_BYTESWAP(bp))
1798 byteswap_uint64_array(zio->io_data, zio->io_size);
1800 ASSERT(zio->io_data == gn->gn_gbh);
1801 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1802 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1804 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1805 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1806 if (!BP_IS_GANG(gbp))
1808 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1813 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1815 zio_t *gio = pio->io_gang_leader;
1818 ASSERT(BP_IS_GANG(bp) == !!gn);
1819 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1820 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1823 * If you're a gang header, your data is in gn->gn_gbh.
1824 * If you're a gang member, your data is in 'data' and gn == NULL.
1826 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1829 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1831 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1832 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1833 if (BP_IS_HOLE(gbp))
1835 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1836 data = (char *)data + BP_GET_PSIZE(gbp);
1840 if (gn == gio->io_gang_tree && gio->io_data != NULL)
1841 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1848 zio_gang_assemble(zio_t **ziop)
1851 blkptr_t *bp = zio->io_bp;
1853 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1854 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1856 zio->io_gang_leader = zio;
1858 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1860 return (ZIO_PIPELINE_CONTINUE);
1864 zio_gang_issue(zio_t **ziop)
1867 blkptr_t *bp = zio->io_bp;
1869 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1870 return (ZIO_PIPELINE_STOP);
1872 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1873 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1875 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1876 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1878 zio_gang_tree_free(&zio->io_gang_tree);
1880 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1882 return (ZIO_PIPELINE_CONTINUE);
1886 zio_write_gang_member_ready(zio_t *zio)
1888 zio_t *pio = zio_unique_parent(zio);
1889 zio_t *gio = zio->io_gang_leader;
1890 dva_t *cdva = zio->io_bp->blk_dva;
1891 dva_t *pdva = pio->io_bp->blk_dva;
1894 if (BP_IS_HOLE(zio->io_bp))
1897 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1899 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1900 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1901 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1902 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1903 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1905 mutex_enter(&pio->io_lock);
1906 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1907 ASSERT(DVA_GET_GANG(&pdva[d]));
1908 asize = DVA_GET_ASIZE(&pdva[d]);
1909 asize += DVA_GET_ASIZE(&cdva[d]);
1910 DVA_SET_ASIZE(&pdva[d], asize);
1912 mutex_exit(&pio->io_lock);
1916 zio_write_gang_block(zio_t *pio)
1918 spa_t *spa = pio->io_spa;
1919 blkptr_t *bp = pio->io_bp;
1920 zio_t *gio = pio->io_gang_leader;
1922 zio_gang_node_t *gn, **gnpp;
1923 zio_gbh_phys_t *gbh;
1924 uint64_t txg = pio->io_txg;
1925 uint64_t resid = pio->io_size;
1927 int copies = gio->io_prop.zp_copies;
1928 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1932 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1933 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1934 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1936 pio->io_error = error;
1937 return (ZIO_PIPELINE_CONTINUE);
1941 gnpp = &gio->io_gang_tree;
1943 gnpp = pio->io_private;
1944 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1947 gn = zio_gang_node_alloc(gnpp);
1949 bzero(gbh, SPA_GANGBLOCKSIZE);
1952 * Create the gang header.
1954 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1955 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1958 * Create and nowait the gang children.
1960 for (int g = 0; resid != 0; resid -= lsize, g++) {
1961 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1963 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1965 zp.zp_checksum = gio->io_prop.zp_checksum;
1966 zp.zp_compress = ZIO_COMPRESS_OFF;
1967 zp.zp_type = DMU_OT_NONE;
1969 zp.zp_copies = gio->io_prop.zp_copies;
1970 zp.zp_dedup = B_FALSE;
1971 zp.zp_dedup_verify = B_FALSE;
1972 zp.zp_nopwrite = B_FALSE;
1974 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1975 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1976 zio_write_gang_member_ready, NULL, NULL, &gn->gn_child[g],
1977 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1978 &pio->io_bookmark));
1982 * Set pio's pipeline to just wait for zio to finish.
1984 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1988 return (ZIO_PIPELINE_CONTINUE);
1992 * The zio_nop_write stage in the pipeline determines if allocating
1993 * a new bp is necessary. By leveraging a cryptographically secure checksum,
1994 * such as SHA256, we can compare the checksums of the new data and the old
1995 * to determine if allocating a new block is required. The nopwrite
1996 * feature can handle writes in either syncing or open context (i.e. zil
1997 * writes) and as a result is mutually exclusive with dedup.
2000 zio_nop_write(zio_t **ziop)
2003 blkptr_t *bp = zio->io_bp;
2004 blkptr_t *bp_orig = &zio->io_bp_orig;
2005 zio_prop_t *zp = &zio->io_prop;
2007 ASSERT(BP_GET_LEVEL(bp) == 0);
2008 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2009 ASSERT(zp->zp_nopwrite);
2010 ASSERT(!zp->zp_dedup);
2011 ASSERT(zio->io_bp_override == NULL);
2012 ASSERT(IO_IS_ALLOCATING(zio));
2015 * Check to see if the original bp and the new bp have matching
2016 * characteristics (i.e. same checksum, compression algorithms, etc).
2017 * If they don't then just continue with the pipeline which will
2018 * allocate a new bp.
2020 if (BP_IS_HOLE(bp_orig) ||
2021 !zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup ||
2022 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2023 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2024 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2025 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2026 return (ZIO_PIPELINE_CONTINUE);
2029 * If the checksums match then reset the pipeline so that we
2030 * avoid allocating a new bp and issuing any I/O.
2032 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2033 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup);
2034 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2035 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2036 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2037 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2038 sizeof (uint64_t)) == 0);
2041 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2042 zio->io_flags |= ZIO_FLAG_NOPWRITE;
2045 return (ZIO_PIPELINE_CONTINUE);
2049 * ==========================================================================
2051 * ==========================================================================
2054 zio_ddt_child_read_done(zio_t *zio)
2056 blkptr_t *bp = zio->io_bp;
2057 ddt_entry_t *dde = zio->io_private;
2059 zio_t *pio = zio_unique_parent(zio);
2061 mutex_enter(&pio->io_lock);
2062 ddp = ddt_phys_select(dde, bp);
2063 if (zio->io_error == 0)
2064 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2065 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
2066 dde->dde_repair_data = zio->io_data;
2068 zio_buf_free(zio->io_data, zio->io_size);
2069 mutex_exit(&pio->io_lock);
2073 zio_ddt_read_start(zio_t **ziop)
2076 blkptr_t *bp = zio->io_bp;
2078 ASSERT(BP_GET_DEDUP(bp));
2079 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2080 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2082 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2083 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2084 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2085 ddt_phys_t *ddp = dde->dde_phys;
2086 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2089 ASSERT(zio->io_vsd == NULL);
2092 if (ddp_self == NULL)
2093 return (ZIO_PIPELINE_CONTINUE);
2095 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2096 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2098 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2100 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2101 zio_buf_alloc(zio->io_size), zio->io_size,
2102 zio_ddt_child_read_done, dde, zio->io_priority,
2103 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
2104 &zio->io_bookmark));
2106 return (ZIO_PIPELINE_CONTINUE);
2109 zio_nowait(zio_read(zio, zio->io_spa, bp,
2110 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
2111 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2113 return (ZIO_PIPELINE_CONTINUE);
2117 zio_ddt_read_done(zio_t **ziop)
2120 blkptr_t *bp = zio->io_bp;
2122 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2123 return (ZIO_PIPELINE_STOP);
2125 ASSERT(BP_GET_DEDUP(bp));
2126 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2127 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2129 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2130 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2131 ddt_entry_t *dde = zio->io_vsd;
2133 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2134 return (ZIO_PIPELINE_CONTINUE);
2137 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2138 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2139 return (ZIO_PIPELINE_STOP);
2141 if (dde->dde_repair_data != NULL) {
2142 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2143 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2145 ddt_repair_done(ddt, dde);
2149 ASSERT(zio->io_vsd == NULL);
2151 return (ZIO_PIPELINE_CONTINUE);
2155 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2157 spa_t *spa = zio->io_spa;
2160 * Note: we compare the original data, not the transformed data,
2161 * because when zio->io_bp is an override bp, we will not have
2162 * pushed the I/O transforms. That's an important optimization
2163 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2165 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2166 zio_t *lio = dde->dde_lead_zio[p];
2169 return (lio->io_orig_size != zio->io_orig_size ||
2170 bcmp(zio->io_orig_data, lio->io_orig_data,
2171 zio->io_orig_size) != 0);
2175 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2176 ddt_phys_t *ddp = &dde->dde_phys[p];
2178 if (ddp->ddp_phys_birth != 0) {
2179 arc_buf_t *abuf = NULL;
2180 uint32_t aflags = ARC_WAIT;
2181 blkptr_t blk = *zio->io_bp;
2184 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2188 error = arc_read(NULL, spa, &blk,
2189 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2190 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2191 &aflags, &zio->io_bookmark);
2194 if (arc_buf_size(abuf) != zio->io_orig_size ||
2195 bcmp(abuf->b_data, zio->io_orig_data,
2196 zio->io_orig_size) != 0)
2197 error = SET_ERROR(EEXIST);
2198 VERIFY(arc_buf_remove_ref(abuf, &abuf));
2202 return (error != 0);
2210 zio_ddt_child_write_ready(zio_t *zio)
2212 int p = zio->io_prop.zp_copies;
2213 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2214 ddt_entry_t *dde = zio->io_private;
2215 ddt_phys_t *ddp = &dde->dde_phys[p];
2223 ASSERT(dde->dde_lead_zio[p] == zio);
2225 ddt_phys_fill(ddp, zio->io_bp);
2227 while ((pio = zio_walk_parents(zio)) != NULL)
2228 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2234 zio_ddt_child_write_done(zio_t *zio)
2236 int p = zio->io_prop.zp_copies;
2237 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2238 ddt_entry_t *dde = zio->io_private;
2239 ddt_phys_t *ddp = &dde->dde_phys[p];
2243 ASSERT(ddp->ddp_refcnt == 0);
2244 ASSERT(dde->dde_lead_zio[p] == zio);
2245 dde->dde_lead_zio[p] = NULL;
2247 if (zio->io_error == 0) {
2248 while (zio_walk_parents(zio) != NULL)
2249 ddt_phys_addref(ddp);
2251 ddt_phys_clear(ddp);
2258 zio_ddt_ditto_write_done(zio_t *zio)
2260 int p = DDT_PHYS_DITTO;
2261 zio_prop_t *zp = &zio->io_prop;
2262 blkptr_t *bp = zio->io_bp;
2263 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2264 ddt_entry_t *dde = zio->io_private;
2265 ddt_phys_t *ddp = &dde->dde_phys[p];
2266 ddt_key_t *ddk = &dde->dde_key;
2270 ASSERT(ddp->ddp_refcnt == 0);
2271 ASSERT(dde->dde_lead_zio[p] == zio);
2272 dde->dde_lead_zio[p] = NULL;
2274 if (zio->io_error == 0) {
2275 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2276 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2277 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2278 if (ddp->ddp_phys_birth != 0)
2279 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2280 ddt_phys_fill(ddp, bp);
2287 zio_ddt_write(zio_t **ziop)
2290 spa_t *spa = zio->io_spa;
2291 blkptr_t *bp = zio->io_bp;
2292 uint64_t txg = zio->io_txg;
2293 zio_prop_t *zp = &zio->io_prop;
2294 int p = zp->zp_copies;
2298 ddt_t *ddt = ddt_select(spa, bp);
2302 ASSERT(BP_GET_DEDUP(bp));
2303 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2304 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2307 dde = ddt_lookup(ddt, bp, B_TRUE);
2308 ddp = &dde->dde_phys[p];
2310 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2312 * If we're using a weak checksum, upgrade to a strong checksum
2313 * and try again. If we're already using a strong checksum,
2314 * we can't resolve it, so just convert to an ordinary write.
2315 * (And automatically e-mail a paper to Nature?)
2317 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2318 zp->zp_checksum = spa_dedup_checksum(spa);
2319 zio_pop_transforms(zio);
2320 zio->io_stage = ZIO_STAGE_OPEN;
2323 zp->zp_dedup = B_FALSE;
2325 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2327 return (ZIO_PIPELINE_CONTINUE);
2330 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2331 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2333 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2334 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2335 zio_prop_t czp = *zp;
2337 czp.zp_copies = ditto_copies;
2340 * If we arrived here with an override bp, we won't have run
2341 * the transform stack, so we won't have the data we need to
2342 * generate a child i/o. So, toss the override bp and restart.
2343 * This is safe, because using the override bp is just an
2344 * optimization; and it's rare, so the cost doesn't matter.
2346 if (zio->io_bp_override) {
2347 zio_pop_transforms(zio);
2348 zio->io_stage = ZIO_STAGE_OPEN;
2349 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2350 zio->io_bp_override = NULL;
2353 return (ZIO_PIPELINE_CONTINUE);
2356 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2357 zio->io_orig_size, &czp, NULL, NULL,
2358 zio_ddt_ditto_write_done, dde, zio->io_priority,
2359 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2361 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2362 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2365 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2366 if (ddp->ddp_phys_birth != 0)
2367 ddt_bp_fill(ddp, bp, txg);
2368 if (dde->dde_lead_zio[p] != NULL)
2369 zio_add_child(zio, dde->dde_lead_zio[p]);
2371 ddt_phys_addref(ddp);
2372 } else if (zio->io_bp_override) {
2373 ASSERT(bp->blk_birth == txg);
2374 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2375 ddt_phys_fill(ddp, bp);
2376 ddt_phys_addref(ddp);
2378 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2379 zio->io_orig_size, zp, zio_ddt_child_write_ready, NULL,
2380 zio_ddt_child_write_done, dde, zio->io_priority,
2381 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2383 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2384 dde->dde_lead_zio[p] = cio;
2394 return (ZIO_PIPELINE_CONTINUE);
2397 ddt_entry_t *freedde; /* for debugging */
2400 zio_ddt_free(zio_t **ziop)
2403 spa_t *spa = zio->io_spa;
2404 blkptr_t *bp = zio->io_bp;
2405 ddt_t *ddt = ddt_select(spa, bp);
2409 ASSERT(BP_GET_DEDUP(bp));
2410 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2413 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2414 ddp = ddt_phys_select(dde, bp);
2415 ddt_phys_decref(ddp);
2418 return (ZIO_PIPELINE_CONTINUE);
2422 * ==========================================================================
2423 * Allocate and free blocks
2424 * ==========================================================================
2427 zio_dva_allocate(zio_t **ziop)
2430 spa_t *spa = zio->io_spa;
2431 metaslab_class_t *mc = spa_normal_class(spa);
2432 blkptr_t *bp = zio->io_bp;
2436 if (zio->io_gang_leader == NULL) {
2437 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2438 zio->io_gang_leader = zio;
2441 ASSERT(BP_IS_HOLE(bp));
2442 ASSERT0(BP_GET_NDVAS(bp));
2443 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2444 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2445 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2448 * The dump device does not support gang blocks so allocation on
2449 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2450 * the "fast" gang feature.
2452 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2453 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2454 METASLAB_GANG_CHILD : 0;
2455 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2456 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2459 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2460 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2462 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2463 return (zio_write_gang_block(zio));
2464 zio->io_error = error;
2467 return (ZIO_PIPELINE_CONTINUE);
2471 zio_dva_free(zio_t **ziop)
2475 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2477 return (ZIO_PIPELINE_CONTINUE);
2481 zio_dva_claim(zio_t **ziop)
2486 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2488 zio->io_error = error;
2490 return (ZIO_PIPELINE_CONTINUE);
2494 * Undo an allocation. This is used by zio_done() when an I/O fails
2495 * and we want to give back the block we just allocated.
2496 * This handles both normal blocks and gang blocks.
2499 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2501 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2502 ASSERT(zio->io_bp_override == NULL);
2504 if (!BP_IS_HOLE(bp))
2505 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2508 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2509 zio_dva_unallocate(zio, gn->gn_child[g],
2510 &gn->gn_gbh->zg_blkptr[g]);
2516 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2519 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2520 uint64_t size, boolean_t use_slog)
2524 ASSERT(txg > spa_syncing_txg(spa));
2527 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2528 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2529 * when allocating them.
2532 error = metaslab_alloc(spa, spa_log_class(spa), size,
2533 new_bp, 1, txg, old_bp,
2534 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2538 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2539 new_bp, 1, txg, old_bp,
2540 METASLAB_HINTBP_AVOID);
2544 BP_SET_LSIZE(new_bp, size);
2545 BP_SET_PSIZE(new_bp, size);
2546 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2547 BP_SET_CHECKSUM(new_bp,
2548 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2549 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2550 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2551 BP_SET_LEVEL(new_bp, 0);
2552 BP_SET_DEDUP(new_bp, 0);
2553 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2560 * Free an intent log block.
2563 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2565 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2566 ASSERT(!BP_IS_GANG(bp));
2568 zio_free(spa, txg, bp);
2572 * ==========================================================================
2573 * Read, write and delete to physical devices
2574 * ==========================================================================
2577 zio_vdev_io_start(zio_t **ziop)
2580 vdev_t *vd = zio->io_vd;
2582 spa_t *spa = zio->io_spa;
2584 ASSERT(zio->io_error == 0);
2585 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2588 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2589 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2592 * The mirror_ops handle multiple DVAs in a single BP.
2594 return (vdev_mirror_ops.vdev_op_io_start(zio));
2597 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE &&
2598 zio->io_priority == ZIO_PRIORITY_NOW) {
2599 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg);
2600 return (ZIO_PIPELINE_CONTINUE);
2604 * We keep track of time-sensitive I/Os so that the scan thread
2605 * can quickly react to certain workloads. In particular, we care
2606 * about non-scrubbing, top-level reads and writes with the following
2608 * - synchronous writes of user data to non-slog devices
2609 * - any reads of user data
2610 * When these conditions are met, adjust the timestamp of spa_last_io
2611 * which allows the scan thread to adjust its workload accordingly.
2613 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2614 vd == vd->vdev_top && !vd->vdev_islog &&
2615 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2616 zio->io_txg != spa_syncing_txg(spa)) {
2617 uint64_t old = spa->spa_last_io;
2618 uint64_t new = ddi_get_lbolt64();
2620 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2623 align = 1ULL << vd->vdev_top->vdev_ashift;
2625 if ((!(zio->io_flags & ZIO_FLAG_PHYSICAL) ||
2626 (vd->vdev_top->vdev_physical_ashift > SPA_MINBLOCKSHIFT)) &&
2627 P2PHASE(zio->io_size, align) != 0) {
2628 /* Transform logical writes to be a full physical block size. */
2629 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2631 if (zio->io_type == ZIO_TYPE_READ ||
2632 zio->io_type == ZIO_TYPE_WRITE)
2633 abuf = zio_buf_alloc(asize);
2634 ASSERT(vd == vd->vdev_top);
2635 if (zio->io_type == ZIO_TYPE_WRITE) {
2636 bcopy(zio->io_data, abuf, zio->io_size);
2637 bzero(abuf + zio->io_size, asize - zio->io_size);
2639 zio_push_transform(zio, abuf, asize, abuf ? asize : 0,
2644 * If this is not a physical io, make sure that it is properly aligned
2645 * before proceeding.
2647 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
2648 ASSERT0(P2PHASE(zio->io_offset, align));
2649 ASSERT0(P2PHASE(zio->io_size, align));
2652 * For physical writes, we allow 512b aligned writes and assume
2653 * the device will perform a read-modify-write as necessary.
2655 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
2656 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
2659 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa));
2662 * If this is a repair I/O, and there's no self-healing involved --
2663 * that is, we're just resilvering what we expect to resilver --
2664 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2665 * This prevents spurious resilvering with nested replication.
2666 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2667 * A is out of date, we'll read from C+D, then use the data to
2668 * resilver A+B -- but we don't actually want to resilver B, just A.
2669 * The top-level mirror has no way to know this, so instead we just
2670 * discard unnecessary repairs as we work our way down the vdev tree.
2671 * The same logic applies to any form of nested replication:
2672 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2674 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2675 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2676 zio->io_txg != 0 && /* not a delegated i/o */
2677 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2678 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2679 zio_vdev_io_bypass(zio);
2680 return (ZIO_PIPELINE_CONTINUE);
2683 if (vd->vdev_ops->vdev_op_leaf) {
2684 switch (zio->io_type) {
2686 if (vdev_cache_read(zio))
2687 return (ZIO_PIPELINE_CONTINUE);
2689 case ZIO_TYPE_WRITE:
2691 if ((zio = vdev_queue_io(zio)) == NULL)
2692 return (ZIO_PIPELINE_STOP);
2695 if (!vdev_accessible(vd, zio)) {
2696 zio->io_error = SET_ERROR(ENXIO);
2698 return (ZIO_PIPELINE_STOP);
2703 * Note that we ignore repair writes for TRIM because they can
2704 * conflict with normal writes. This isn't an issue because, by
2705 * definition, we only repair blocks that aren't freed.
2707 if (zio->io_type == ZIO_TYPE_WRITE &&
2708 !(zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2709 !trim_map_write_start(zio))
2710 return (ZIO_PIPELINE_STOP);
2713 return (vd->vdev_ops->vdev_op_io_start(zio));
2717 zio_vdev_io_done(zio_t **ziop)
2720 vdev_t *vd = zio->io_vd;
2721 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2722 boolean_t unexpected_error = B_FALSE;
2724 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2725 return (ZIO_PIPELINE_STOP);
2727 ASSERT(zio->io_type == ZIO_TYPE_READ ||
2728 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE);
2730 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2731 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE ||
2732 zio->io_type == ZIO_TYPE_FREE)) {
2734 if (zio->io_type == ZIO_TYPE_WRITE &&
2735 !(zio->io_flags & ZIO_FLAG_IO_REPAIR))
2736 trim_map_write_done(zio);
2738 vdev_queue_io_done(zio);
2740 if (zio->io_type == ZIO_TYPE_WRITE)
2741 vdev_cache_write(zio);
2743 if (zio_injection_enabled && zio->io_error == 0)
2744 zio->io_error = zio_handle_device_injection(vd,
2747 if (zio_injection_enabled && zio->io_error == 0)
2748 zio->io_error = zio_handle_label_injection(zio, EIO);
2750 if (zio->io_error) {
2751 if (zio->io_error == ENOTSUP &&
2752 zio->io_type == ZIO_TYPE_FREE) {
2753 /* Not all devices support TRIM. */
2754 } else if (!vdev_accessible(vd, zio)) {
2755 zio->io_error = SET_ERROR(ENXIO);
2757 unexpected_error = B_TRUE;
2762 ops->vdev_op_io_done(zio);
2764 if (unexpected_error)
2765 VERIFY(vdev_probe(vd, zio) == NULL);
2767 return (ZIO_PIPELINE_CONTINUE);
2771 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2772 * disk, and use that to finish the checksum ereport later.
2775 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2776 const void *good_buf)
2778 /* no processing needed */
2779 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2784 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2786 void *buf = zio_buf_alloc(zio->io_size);
2788 bcopy(zio->io_data, buf, zio->io_size);
2790 zcr->zcr_cbinfo = zio->io_size;
2791 zcr->zcr_cbdata = buf;
2792 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2793 zcr->zcr_free = zio_buf_free;
2797 zio_vdev_io_assess(zio_t **ziop)
2800 vdev_t *vd = zio->io_vd;
2802 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2803 return (ZIO_PIPELINE_STOP);
2805 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2806 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2808 if (zio->io_vsd != NULL) {
2809 zio->io_vsd_ops->vsd_free(zio);
2813 if (zio_injection_enabled && zio->io_error == 0)
2814 zio->io_error = zio_handle_fault_injection(zio, EIO);
2816 if (zio->io_type == ZIO_TYPE_FREE &&
2817 zio->io_priority != ZIO_PRIORITY_NOW) {
2818 switch (zio->io_error) {
2820 ZIO_TRIM_STAT_INCR(bytes, zio->io_size);
2821 ZIO_TRIM_STAT_BUMP(success);
2824 ZIO_TRIM_STAT_BUMP(unsupported);
2827 ZIO_TRIM_STAT_BUMP(failed);
2833 * If the I/O failed, determine whether we should attempt to retry it.
2835 * On retry, we cut in line in the issue queue, since we don't want
2836 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2838 if (zio->io_error && vd == NULL &&
2839 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2840 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2841 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2843 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2844 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2845 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2846 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2847 zio_requeue_io_start_cut_in_line);
2848 return (ZIO_PIPELINE_STOP);
2852 * If we got an error on a leaf device, convert it to ENXIO
2853 * if the device is not accessible at all.
2855 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2856 !vdev_accessible(vd, zio))
2857 zio->io_error = SET_ERROR(ENXIO);
2860 * If we can't write to an interior vdev (mirror or RAID-Z),
2861 * set vdev_cant_write so that we stop trying to allocate from it.
2863 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2864 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
2865 vd->vdev_cant_write = B_TRUE;
2869 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2871 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2872 zio->io_physdone != NULL) {
2873 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
2874 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
2875 zio->io_physdone(zio->io_logical);
2878 return (ZIO_PIPELINE_CONTINUE);
2882 zio_vdev_io_reissue(zio_t *zio)
2884 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2885 ASSERT(zio->io_error == 0);
2887 zio->io_stage >>= 1;
2891 zio_vdev_io_redone(zio_t *zio)
2893 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2895 zio->io_stage >>= 1;
2899 zio_vdev_io_bypass(zio_t *zio)
2901 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2902 ASSERT(zio->io_error == 0);
2904 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2905 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2909 * ==========================================================================
2910 * Generate and verify checksums
2911 * ==========================================================================
2914 zio_checksum_generate(zio_t **ziop)
2917 blkptr_t *bp = zio->io_bp;
2918 enum zio_checksum checksum;
2922 * This is zio_write_phys().
2923 * We're either generating a label checksum, or none at all.
2925 checksum = zio->io_prop.zp_checksum;
2927 if (checksum == ZIO_CHECKSUM_OFF)
2928 return (ZIO_PIPELINE_CONTINUE);
2930 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2932 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2933 ASSERT(!IO_IS_ALLOCATING(zio));
2934 checksum = ZIO_CHECKSUM_GANG_HEADER;
2936 checksum = BP_GET_CHECKSUM(bp);
2940 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2942 return (ZIO_PIPELINE_CONTINUE);
2946 zio_checksum_verify(zio_t **ziop)
2949 zio_bad_cksum_t info;
2950 blkptr_t *bp = zio->io_bp;
2953 ASSERT(zio->io_vd != NULL);
2957 * This is zio_read_phys().
2958 * We're either verifying a label checksum, or nothing at all.
2960 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2961 return (ZIO_PIPELINE_CONTINUE);
2963 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2966 if ((error = zio_checksum_error(zio, &info)) != 0) {
2967 zio->io_error = error;
2968 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2969 zfs_ereport_start_checksum(zio->io_spa,
2970 zio->io_vd, zio, zio->io_offset,
2971 zio->io_size, NULL, &info);
2975 return (ZIO_PIPELINE_CONTINUE);
2979 * Called by RAID-Z to ensure we don't compute the checksum twice.
2982 zio_checksum_verified(zio_t *zio)
2984 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2988 * ==========================================================================
2989 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2990 * An error of 0 indicates success. ENXIO indicates whole-device failure,
2991 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2992 * indicate errors that are specific to one I/O, and most likely permanent.
2993 * Any other error is presumed to be worse because we weren't expecting it.
2994 * ==========================================================================
2997 zio_worst_error(int e1, int e2)
2999 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
3002 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
3003 if (e1 == zio_error_rank[r1])
3006 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
3007 if (e2 == zio_error_rank[r2])
3010 return (r1 > r2 ? e1 : e2);
3014 * ==========================================================================
3016 * ==========================================================================
3019 zio_ready(zio_t **ziop)
3022 blkptr_t *bp = zio->io_bp;
3023 zio_t *pio, *pio_next;
3025 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
3026 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
3027 return (ZIO_PIPELINE_STOP);
3029 if (zio->io_ready) {
3030 ASSERT(IO_IS_ALLOCATING(zio));
3031 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
3032 (zio->io_flags & ZIO_FLAG_NOPWRITE));
3033 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
3038 if (bp != NULL && bp != &zio->io_bp_copy)
3039 zio->io_bp_copy = *bp;
3042 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3044 mutex_enter(&zio->io_lock);
3045 zio->io_state[ZIO_WAIT_READY] = 1;
3046 pio = zio_walk_parents(zio);
3047 mutex_exit(&zio->io_lock);
3050 * As we notify zio's parents, new parents could be added.
3051 * New parents go to the head of zio's io_parent_list, however,
3052 * so we will (correctly) not notify them. The remainder of zio's
3053 * io_parent_list, from 'pio_next' onward, cannot change because
3054 * all parents must wait for us to be done before they can be done.
3056 for (; pio != NULL; pio = pio_next) {
3057 pio_next = zio_walk_parents(zio);
3058 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
3061 if (zio->io_flags & ZIO_FLAG_NODATA) {
3062 if (BP_IS_GANG(bp)) {
3063 zio->io_flags &= ~ZIO_FLAG_NODATA;
3065 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
3066 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
3070 if (zio_injection_enabled &&
3071 zio->io_spa->spa_syncing_txg == zio->io_txg)
3072 zio_handle_ignored_writes(zio);
3074 return (ZIO_PIPELINE_CONTINUE);
3078 zio_done(zio_t **ziop)
3081 spa_t *spa = zio->io_spa;
3082 zio_t *lio = zio->io_logical;
3083 blkptr_t *bp = zio->io_bp;
3084 vdev_t *vd = zio->io_vd;
3085 uint64_t psize = zio->io_size;
3086 zio_t *pio, *pio_next;
3089 * If our children haven't all completed,
3090 * wait for them and then repeat this pipeline stage.
3092 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
3093 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
3094 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
3095 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
3096 return (ZIO_PIPELINE_STOP);
3098 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3099 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3100 ASSERT(zio->io_children[c][w] == 0);
3102 if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
3103 ASSERT(bp->blk_pad[0] == 0);
3104 ASSERT(bp->blk_pad[1] == 0);
3105 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3106 (bp == zio_unique_parent(zio)->io_bp));
3107 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3108 zio->io_bp_override == NULL &&
3109 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3110 ASSERT(!BP_SHOULD_BYTESWAP(bp));
3111 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3112 ASSERT(BP_COUNT_GANG(bp) == 0 ||
3113 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3115 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3116 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3120 * If there were child vdev/gang/ddt errors, they apply to us now.
3122 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3123 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3124 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3127 * If the I/O on the transformed data was successful, generate any
3128 * checksum reports now while we still have the transformed data.
3130 if (zio->io_error == 0) {
3131 while (zio->io_cksum_report != NULL) {
3132 zio_cksum_report_t *zcr = zio->io_cksum_report;
3133 uint64_t align = zcr->zcr_align;
3134 uint64_t asize = P2ROUNDUP(psize, align);
3135 char *abuf = zio->io_data;
3137 if (asize != psize) {
3138 abuf = zio_buf_alloc(asize);
3139 bcopy(zio->io_data, abuf, psize);
3140 bzero(abuf + psize, asize - psize);
3143 zio->io_cksum_report = zcr->zcr_next;
3144 zcr->zcr_next = NULL;
3145 zcr->zcr_finish(zcr, abuf);
3146 zfs_ereport_free_checksum(zcr);
3149 zio_buf_free(abuf, asize);
3153 zio_pop_transforms(zio); /* note: may set zio->io_error */
3155 vdev_stat_update(zio, psize);
3157 if (zio->io_error) {
3159 * If this I/O is attached to a particular vdev,
3160 * generate an error message describing the I/O failure
3161 * at the block level. We ignore these errors if the
3162 * device is currently unavailable.
3164 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3165 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3167 if ((zio->io_error == EIO || !(zio->io_flags &
3168 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3171 * For logical I/O requests, tell the SPA to log the
3172 * error and generate a logical data ereport.
3174 spa_log_error(spa, zio);
3175 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3180 if (zio->io_error && zio == lio) {
3182 * Determine whether zio should be reexecuted. This will
3183 * propagate all the way to the root via zio_notify_parent().
3185 ASSERT(vd == NULL && bp != NULL);
3186 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3188 if (IO_IS_ALLOCATING(zio) &&
3189 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3190 if (zio->io_error != ENOSPC)
3191 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3193 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3196 if ((zio->io_type == ZIO_TYPE_READ ||
3197 zio->io_type == ZIO_TYPE_FREE) &&
3198 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3199 zio->io_error == ENXIO &&
3200 spa_load_state(spa) == SPA_LOAD_NONE &&
3201 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3202 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3204 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3205 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3208 * Here is a possibly good place to attempt to do
3209 * either combinatorial reconstruction or error correction
3210 * based on checksums. It also might be a good place
3211 * to send out preliminary ereports before we suspend
3217 * If there were logical child errors, they apply to us now.
3218 * We defer this until now to avoid conflating logical child
3219 * errors with errors that happened to the zio itself when
3220 * updating vdev stats and reporting FMA events above.
3222 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3224 if ((zio->io_error || zio->io_reexecute) &&
3225 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3226 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3227 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3229 zio_gang_tree_free(&zio->io_gang_tree);
3232 * Godfather I/Os should never suspend.
3234 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3235 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3236 zio->io_reexecute = 0;
3238 if (zio->io_reexecute) {
3240 * This is a logical I/O that wants to reexecute.
3242 * Reexecute is top-down. When an i/o fails, if it's not
3243 * the root, it simply notifies its parent and sticks around.
3244 * The parent, seeing that it still has children in zio_done(),
3245 * does the same. This percolates all the way up to the root.
3246 * The root i/o will reexecute or suspend the entire tree.
3248 * This approach ensures that zio_reexecute() honors
3249 * all the original i/o dependency relationships, e.g.
3250 * parents not executing until children are ready.
3252 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3254 zio->io_gang_leader = NULL;
3256 mutex_enter(&zio->io_lock);
3257 zio->io_state[ZIO_WAIT_DONE] = 1;
3258 mutex_exit(&zio->io_lock);
3261 * "The Godfather" I/O monitors its children but is
3262 * not a true parent to them. It will track them through
3263 * the pipeline but severs its ties whenever they get into
3264 * trouble (e.g. suspended). This allows "The Godfather"
3265 * I/O to return status without blocking.
3267 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3268 zio_link_t *zl = zio->io_walk_link;
3269 pio_next = zio_walk_parents(zio);
3271 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3272 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3273 zio_remove_child(pio, zio, zl);
3274 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3278 if ((pio = zio_unique_parent(zio)) != NULL) {
3280 * We're not a root i/o, so there's nothing to do
3281 * but notify our parent. Don't propagate errors
3282 * upward since we haven't permanently failed yet.
3284 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3285 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3286 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3287 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3289 * We'd fail again if we reexecuted now, so suspend
3290 * until conditions improve (e.g. device comes online).
3292 zio_suspend(spa, zio);
3295 * Reexecution is potentially a huge amount of work.
3296 * Hand it off to the otherwise-unused claim taskq.
3298 #if defined(illumos) || !defined(_KERNEL)
3299 ASSERT(zio->io_tqent.tqent_next == NULL);
3301 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
3303 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3304 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3307 return (ZIO_PIPELINE_STOP);
3310 ASSERT(zio->io_child_count == 0);
3311 ASSERT(zio->io_reexecute == 0);
3312 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3315 * Report any checksum errors, since the I/O is complete.
3317 while (zio->io_cksum_report != NULL) {
3318 zio_cksum_report_t *zcr = zio->io_cksum_report;
3319 zio->io_cksum_report = zcr->zcr_next;
3320 zcr->zcr_next = NULL;
3321 zcr->zcr_finish(zcr, NULL);
3322 zfs_ereport_free_checksum(zcr);
3326 * It is the responsibility of the done callback to ensure that this
3327 * particular zio is no longer discoverable for adoption, and as
3328 * such, cannot acquire any new parents.
3333 mutex_enter(&zio->io_lock);
3334 zio->io_state[ZIO_WAIT_DONE] = 1;
3335 mutex_exit(&zio->io_lock);
3337 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3338 zio_link_t *zl = zio->io_walk_link;
3339 pio_next = zio_walk_parents(zio);
3340 zio_remove_child(pio, zio, zl);
3341 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3344 if (zio->io_waiter != NULL) {
3345 mutex_enter(&zio->io_lock);
3346 zio->io_executor = NULL;
3347 cv_broadcast(&zio->io_cv);
3348 mutex_exit(&zio->io_lock);
3353 return (ZIO_PIPELINE_STOP);
3357 * ==========================================================================
3358 * I/O pipeline definition
3359 * ==========================================================================
3361 static zio_pipe_stage_t *zio_pipeline[] = {
3367 zio_checksum_generate,
3382 zio_checksum_verify,
3386 /* dnp is the dnode for zb1->zb_object */
3388 zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_phys_t *zb1,
3389 const zbookmark_phys_t *zb2)
3391 uint64_t zb1nextL0, zb2thisobj;
3393 ASSERT(zb1->zb_objset == zb2->zb_objset);
3394 ASSERT(zb2->zb_level == 0);
3396 /* The objset_phys_t isn't before anything. */
3400 zb1nextL0 = (zb1->zb_blkid + 1) <<
3401 ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
3403 zb2thisobj = zb2->zb_object ? zb2->zb_object :
3404 zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
3406 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3407 uint64_t nextobj = zb1nextL0 *
3408 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
3409 return (nextobj <= zb2thisobj);
3412 if (zb1->zb_object < zb2thisobj)
3414 if (zb1->zb_object > zb2thisobj)
3416 if (zb2->zb_object == DMU_META_DNODE_OBJECT)
3418 return (zb1nextL0 <= zb2->zb_blkid);