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/sysmacros.h>
28 #include <sys/zfs_context.h>
29 #include <sys/fm/fs/zfs.h>
32 #include <sys/spa_impl.h>
33 #include <sys/vdev_impl.h>
34 #include <sys/zio_impl.h>
35 #include <sys/zio_compress.h>
36 #include <sys/zio_checksum.h>
37 #include <sys/dmu_objset.h>
40 #include <sys/trim_map.h>
41 #include <sys/blkptr.h>
42 #include <sys/zfeature.h>
44 SYSCTL_DECL(_vfs_zfs);
45 SYSCTL_NODE(_vfs_zfs, OID_AUTO, zio, CTLFLAG_RW, 0, "ZFS ZIO");
46 #if defined(__amd64__)
47 static int zio_use_uma = 1;
49 static int zio_use_uma = 0;
51 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, use_uma, CTLFLAG_RDTUN, &zio_use_uma, 0,
52 "Use uma(9) for ZIO allocations");
53 static int zio_exclude_metadata = 0;
54 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, exclude_metadata, CTLFLAG_RDTUN, &zio_exclude_metadata, 0,
55 "Exclude metadata buffers from dumps as well");
57 zio_trim_stats_t zio_trim_stats = {
58 { "bytes", KSTAT_DATA_UINT64,
59 "Number of bytes successfully TRIMmed" },
60 { "success", KSTAT_DATA_UINT64,
61 "Number of successful TRIM requests" },
62 { "unsupported", KSTAT_DATA_UINT64,
63 "Number of TRIM requests that failed because TRIM is not supported" },
64 { "failed", KSTAT_DATA_UINT64,
65 "Number of TRIM requests that failed for reasons other than not supported" },
68 static kstat_t *zio_trim_ksp;
71 * ==========================================================================
72 * I/O type descriptions
73 * ==========================================================================
75 const char *zio_type_name[ZIO_TYPES] = {
76 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
81 * ==========================================================================
83 * ==========================================================================
85 kmem_cache_t *zio_cache;
86 kmem_cache_t *zio_link_cache;
87 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
88 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
91 extern vmem_t *zio_alloc_arena;
94 #define ZIO_PIPELINE_CONTINUE 0x100
95 #define ZIO_PIPELINE_STOP 0x101
98 * The following actions directly effect the spa's sync-to-convergence logic.
99 * The values below define the sync pass when we start performing the action.
100 * Care should be taken when changing these values as they directly impact
101 * spa_sync() performance. Tuning these values may introduce subtle performance
102 * pathologies and should only be done in the context of performance analysis.
103 * These tunables will eventually be removed and replaced with #defines once
104 * enough analysis has been done to determine optimal values.
106 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
107 * regular blocks are not deferred.
109 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
110 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_deferred_free, CTLFLAG_RDTUN,
111 &zfs_sync_pass_deferred_free, 0, "defer frees starting in this pass");
112 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
113 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_dont_compress, CTLFLAG_RDTUN,
114 &zfs_sync_pass_dont_compress, 0, "don't compress starting in this pass");
115 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
116 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_rewrite, CTLFLAG_RDTUN,
117 &zfs_sync_pass_rewrite, 0, "rewrite new bps starting in this pass");
120 * An allocating zio is one that either currently has the DVA allocate
121 * stage set or will have it later in its lifetime.
123 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
125 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
128 int zio_buf_debug_limit = 16384;
130 int zio_buf_debug_limit = 0;
137 zio_cache = kmem_cache_create("zio_cache",
138 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
139 zio_link_cache = kmem_cache_create("zio_link_cache",
140 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
145 * For small buffers, we want a cache for each multiple of
146 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
147 * for each quarter-power of 2.
149 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
150 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
153 size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
161 * If we are using watchpoints, put each buffer on its own page,
162 * to eliminate the performance overhead of trapping to the
163 * kernel when modifying a non-watched buffer that shares the
164 * page with a watched buffer.
166 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
170 if (size <= 4 * SPA_MINBLOCKSIZE) {
171 align = SPA_MINBLOCKSIZE;
172 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
173 align = MIN(p2 >> 2, PAGESIZE);
178 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
179 zio_buf_cache[c] = kmem_cache_create(name, size,
180 align, NULL, NULL, NULL, NULL, NULL, cflags);
183 * Since zio_data bufs do not appear in crash dumps, we
184 * pass KMC_NOTOUCH so that no allocator metadata is
185 * stored with the buffers.
187 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
188 zio_data_buf_cache[c] = kmem_cache_create(name, size,
189 align, NULL, NULL, NULL, NULL, NULL,
190 cflags | KMC_NOTOUCH | KMC_NODEBUG);
195 ASSERT(zio_buf_cache[c] != NULL);
196 if (zio_buf_cache[c - 1] == NULL)
197 zio_buf_cache[c - 1] = zio_buf_cache[c];
199 ASSERT(zio_data_buf_cache[c] != NULL);
200 if (zio_data_buf_cache[c - 1] == NULL)
201 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
207 zio_trim_ksp = kstat_create("zfs", 0, "zio_trim", "misc",
209 sizeof(zio_trim_stats) / sizeof(kstat_named_t),
212 if (zio_trim_ksp != NULL) {
213 zio_trim_ksp->ks_data = &zio_trim_stats;
214 kstat_install(zio_trim_ksp);
222 kmem_cache_t *last_cache = NULL;
223 kmem_cache_t *last_data_cache = NULL;
225 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
226 if (zio_buf_cache[c] != last_cache) {
227 last_cache = zio_buf_cache[c];
228 kmem_cache_destroy(zio_buf_cache[c]);
230 zio_buf_cache[c] = NULL;
232 if (zio_data_buf_cache[c] != last_data_cache) {
233 last_data_cache = zio_data_buf_cache[c];
234 kmem_cache_destroy(zio_data_buf_cache[c]);
236 zio_data_buf_cache[c] = NULL;
239 kmem_cache_destroy(zio_link_cache);
240 kmem_cache_destroy(zio_cache);
244 if (zio_trim_ksp != NULL) {
245 kstat_delete(zio_trim_ksp);
251 * ==========================================================================
252 * Allocate and free I/O buffers
253 * ==========================================================================
257 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
258 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
259 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
260 * excess / transient data in-core during a crashdump.
263 zio_buf_alloc(size_t size)
265 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
266 int flags = zio_exclude_metadata ? KM_NODEBUG : 0;
268 ASSERT3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
271 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
273 return (kmem_alloc(size, KM_SLEEP|flags));
277 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
278 * crashdump if the kernel panics. This exists so that we will limit the amount
279 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
280 * of kernel heap dumped to disk when the kernel panics)
283 zio_data_buf_alloc(size_t size)
285 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
287 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
290 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
292 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG));
296 zio_buf_free(void *buf, size_t size)
298 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
300 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
303 kmem_cache_free(zio_buf_cache[c], buf);
305 kmem_free(buf, size);
309 zio_data_buf_free(void *buf, size_t size)
311 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
313 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
316 kmem_cache_free(zio_data_buf_cache[c], buf);
318 kmem_free(buf, size);
322 * ==========================================================================
323 * Push and pop I/O transform buffers
324 * ==========================================================================
327 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
328 zio_transform_func_t *transform)
330 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
332 zt->zt_orig_data = zio->io_data;
333 zt->zt_orig_size = zio->io_size;
334 zt->zt_bufsize = bufsize;
335 zt->zt_transform = transform;
337 zt->zt_next = zio->io_transform_stack;
338 zio->io_transform_stack = zt;
345 zio_pop_transforms(zio_t *zio)
349 while ((zt = zio->io_transform_stack) != NULL) {
350 if (zt->zt_transform != NULL)
351 zt->zt_transform(zio,
352 zt->zt_orig_data, zt->zt_orig_size);
354 if (zt->zt_bufsize != 0)
355 zio_buf_free(zio->io_data, zt->zt_bufsize);
357 zio->io_data = zt->zt_orig_data;
358 zio->io_size = zt->zt_orig_size;
359 zio->io_transform_stack = zt->zt_next;
361 kmem_free(zt, sizeof (zio_transform_t));
366 * ==========================================================================
367 * I/O transform callbacks for subblocks and decompression
368 * ==========================================================================
371 zio_subblock(zio_t *zio, void *data, uint64_t size)
373 ASSERT(zio->io_size > size);
375 if (zio->io_type == ZIO_TYPE_READ)
376 bcopy(zio->io_data, data, size);
380 zio_decompress(zio_t *zio, void *data, uint64_t size)
382 if (zio->io_error == 0 &&
383 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
384 zio->io_data, data, zio->io_size, size) != 0)
385 zio->io_error = SET_ERROR(EIO);
389 * ==========================================================================
390 * I/O parent/child relationships and pipeline interlocks
391 * ==========================================================================
394 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
395 * continue calling these functions until they return NULL.
396 * Otherwise, the next caller will pick up the list walk in
397 * some indeterminate state. (Otherwise every caller would
398 * have to pass in a cookie to keep the state represented by
399 * io_walk_link, which gets annoying.)
402 zio_walk_parents(zio_t *cio)
404 zio_link_t *zl = cio->io_walk_link;
405 list_t *pl = &cio->io_parent_list;
407 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
408 cio->io_walk_link = zl;
413 ASSERT(zl->zl_child == cio);
414 return (zl->zl_parent);
418 zio_walk_children(zio_t *pio)
420 zio_link_t *zl = pio->io_walk_link;
421 list_t *cl = &pio->io_child_list;
423 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
424 pio->io_walk_link = zl;
429 ASSERT(zl->zl_parent == pio);
430 return (zl->zl_child);
434 zio_unique_parent(zio_t *cio)
436 zio_t *pio = zio_walk_parents(cio);
438 VERIFY(zio_walk_parents(cio) == NULL);
443 zio_add_child(zio_t *pio, zio_t *cio)
445 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
448 * Logical I/Os can have logical, gang, or vdev children.
449 * Gang I/Os can have gang or vdev children.
450 * Vdev I/Os can only have vdev children.
451 * The following ASSERT captures all of these constraints.
453 ASSERT(cio->io_child_type <= pio->io_child_type);
458 mutex_enter(&cio->io_lock);
459 mutex_enter(&pio->io_lock);
461 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
463 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
464 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
466 list_insert_head(&pio->io_child_list, zl);
467 list_insert_head(&cio->io_parent_list, zl);
469 pio->io_child_count++;
470 cio->io_parent_count++;
472 mutex_exit(&pio->io_lock);
473 mutex_exit(&cio->io_lock);
477 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
479 ASSERT(zl->zl_parent == pio);
480 ASSERT(zl->zl_child == cio);
482 mutex_enter(&cio->io_lock);
483 mutex_enter(&pio->io_lock);
485 list_remove(&pio->io_child_list, zl);
486 list_remove(&cio->io_parent_list, zl);
488 pio->io_child_count--;
489 cio->io_parent_count--;
491 mutex_exit(&pio->io_lock);
492 mutex_exit(&cio->io_lock);
494 kmem_cache_free(zio_link_cache, zl);
498 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
500 uint64_t *countp = &zio->io_children[child][wait];
501 boolean_t waiting = B_FALSE;
503 mutex_enter(&zio->io_lock);
504 ASSERT(zio->io_stall == NULL);
507 zio->io_stall = countp;
510 mutex_exit(&zio->io_lock);
516 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
518 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
519 int *errorp = &pio->io_child_error[zio->io_child_type];
521 mutex_enter(&pio->io_lock);
522 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
523 *errorp = zio_worst_error(*errorp, zio->io_error);
524 pio->io_reexecute |= zio->io_reexecute;
525 ASSERT3U(*countp, >, 0);
529 if (*countp == 0 && pio->io_stall == countp) {
530 pio->io_stall = NULL;
531 mutex_exit(&pio->io_lock);
534 mutex_exit(&pio->io_lock);
539 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
541 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
542 zio->io_error = zio->io_child_error[c];
546 * ==========================================================================
547 * Create the various types of I/O (read, write, free, etc)
548 * ==========================================================================
551 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
552 void *data, uint64_t size, zio_done_func_t *done, void *private,
553 zio_type_t type, zio_priority_t priority, enum zio_flag flags,
554 vdev_t *vd, uint64_t offset, const zbookmark_phys_t *zb,
555 enum zio_stage stage, enum zio_stage pipeline)
559 ASSERT3U(type == ZIO_TYPE_FREE || size, <=, SPA_MAXBLOCKSIZE);
560 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
561 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
563 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
564 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
565 ASSERT(vd || stage == ZIO_STAGE_OPEN);
567 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
568 bzero(zio, sizeof (zio_t));
570 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
571 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
573 list_create(&zio->io_parent_list, sizeof (zio_link_t),
574 offsetof(zio_link_t, zl_parent_node));
575 list_create(&zio->io_child_list, sizeof (zio_link_t),
576 offsetof(zio_link_t, zl_child_node));
579 zio->io_child_type = ZIO_CHILD_VDEV;
580 else if (flags & ZIO_FLAG_GANG_CHILD)
581 zio->io_child_type = ZIO_CHILD_GANG;
582 else if (flags & ZIO_FLAG_DDT_CHILD)
583 zio->io_child_type = ZIO_CHILD_DDT;
585 zio->io_child_type = ZIO_CHILD_LOGICAL;
588 zio->io_bp = (blkptr_t *)bp;
589 zio->io_bp_copy = *bp;
590 zio->io_bp_orig = *bp;
591 if (type != ZIO_TYPE_WRITE ||
592 zio->io_child_type == ZIO_CHILD_DDT)
593 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
594 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
595 zio->io_logical = zio;
596 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
597 pipeline |= ZIO_GANG_STAGES;
603 zio->io_private = private;
605 zio->io_priority = priority;
607 zio->io_offset = offset;
608 zio->io_orig_data = zio->io_data = data;
609 zio->io_orig_size = zio->io_size = size;
610 zio->io_orig_flags = zio->io_flags = flags;
611 zio->io_orig_stage = zio->io_stage = stage;
612 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
614 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
615 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
618 zio->io_bookmark = *zb;
621 if (zio->io_logical == NULL)
622 zio->io_logical = pio->io_logical;
623 if (zio->io_child_type == ZIO_CHILD_GANG)
624 zio->io_gang_leader = pio->io_gang_leader;
625 zio_add_child(pio, zio);
632 zio_destroy(zio_t *zio)
634 list_destroy(&zio->io_parent_list);
635 list_destroy(&zio->io_child_list);
636 mutex_destroy(&zio->io_lock);
637 cv_destroy(&zio->io_cv);
638 kmem_cache_free(zio_cache, zio);
642 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
643 void *private, enum zio_flag flags)
647 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
648 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
649 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
655 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
657 return (zio_null(NULL, spa, NULL, done, private, flags));
661 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
662 void *data, uint64_t size, zio_done_func_t *done, void *private,
663 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
667 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
668 data, size, done, private,
669 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
670 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
671 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
677 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
678 void *data, uint64_t size, const zio_prop_t *zp,
679 zio_done_func_t *ready, zio_done_func_t *physdone, zio_done_func_t *done,
681 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
685 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
686 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
687 zp->zp_compress >= ZIO_COMPRESS_OFF &&
688 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
689 DMU_OT_IS_VALID(zp->zp_type) &&
692 zp->zp_copies <= spa_max_replication(spa));
694 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
695 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
696 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
697 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
699 zio->io_ready = ready;
700 zio->io_physdone = physdone;
704 * Data can be NULL if we are going to call zio_write_override() to
705 * provide the already-allocated BP. But we may need the data to
706 * verify a dedup hit (if requested). In this case, don't try to
707 * dedup (just take the already-allocated BP verbatim).
709 if (data == NULL && zio->io_prop.zp_dedup_verify) {
710 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
717 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
718 uint64_t size, zio_done_func_t *done, void *private,
719 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
723 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
724 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
725 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
731 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
733 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
734 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
735 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
736 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
739 * We must reset the io_prop to match the values that existed
740 * when the bp was first written by dmu_sync() keeping in mind
741 * that nopwrite and dedup are mutually exclusive.
743 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
744 zio->io_prop.zp_nopwrite = nopwrite;
745 zio->io_prop.zp_copies = copies;
746 zio->io_bp_override = bp;
750 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
754 * The check for EMBEDDED is a performance optimization. We
755 * process the free here (by ignoring it) rather than
756 * putting it on the list and then processing it in zio_free_sync().
758 if (BP_IS_EMBEDDED(bp))
760 metaslab_check_free(spa, bp);
763 * Frees that are for the currently-syncing txg, are not going to be
764 * deferred, and which will not need to do a read (i.e. not GANG or
765 * DEDUP), can be processed immediately. Otherwise, put them on the
766 * in-memory list for later processing.
768 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
769 txg != spa->spa_syncing_txg ||
770 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
771 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
773 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp,
774 BP_GET_PSIZE(bp), 0)));
779 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
780 uint64_t size, enum zio_flag flags)
783 enum zio_stage stage = ZIO_FREE_PIPELINE;
785 ASSERT(!BP_IS_HOLE(bp));
786 ASSERT(spa_syncing_txg(spa) == txg);
787 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
789 if (BP_IS_EMBEDDED(bp))
790 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
792 metaslab_check_free(spa, bp);
795 if (zfs_trim_enabled)
796 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START |
797 ZIO_STAGE_VDEV_IO_ASSESS;
799 * GANG and DEDUP blocks can induce a read (for the gang block header,
800 * or the DDT), so issue them asynchronously so that this thread is
803 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
804 stage |= ZIO_STAGE_ISSUE_ASYNC;
806 flags |= ZIO_FLAG_DONT_QUEUE;
808 zio = zio_create(pio, spa, txg, bp, NULL, size,
809 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags,
810 NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
816 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
817 zio_done_func_t *done, void *private, enum zio_flag flags)
821 dprintf_bp(bp, "claiming in txg %llu", txg);
823 if (BP_IS_EMBEDDED(bp))
824 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
827 * A claim is an allocation of a specific block. Claims are needed
828 * to support immediate writes in the intent log. The issue is that
829 * immediate writes contain committed data, but in a txg that was
830 * *not* committed. Upon opening the pool after an unclean shutdown,
831 * the intent log claims all blocks that contain immediate write data
832 * so that the SPA knows they're in use.
834 * All claims *must* be resolved in the first txg -- before the SPA
835 * starts allocating blocks -- so that nothing is allocated twice.
836 * If txg == 0 we just verify that the block is claimable.
838 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
839 ASSERT(txg == spa_first_txg(spa) || txg == 0);
840 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
842 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
843 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
844 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
850 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset,
851 uint64_t size, zio_done_func_t *done, void *private,
852 zio_priority_t priority, enum zio_flag flags)
857 if (vd->vdev_children == 0) {
858 zio = zio_create(pio, spa, 0, NULL, NULL, size, done, private,
859 ZIO_TYPE_IOCTL, priority, flags, vd, offset, NULL,
860 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
864 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
866 for (c = 0; c < vd->vdev_children; c++)
867 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
868 offset, size, done, private, priority, flags));
875 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
876 void *data, int checksum, zio_done_func_t *done, void *private,
877 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
881 ASSERT(vd->vdev_children == 0);
882 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
883 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
884 ASSERT3U(offset + size, <=, vd->vdev_psize);
886 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
887 ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
888 NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
890 zio->io_prop.zp_checksum = checksum;
896 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
897 void *data, int checksum, zio_done_func_t *done, void *private,
898 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
902 ASSERT(vd->vdev_children == 0);
903 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
904 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
905 ASSERT3U(offset + size, <=, vd->vdev_psize);
907 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
908 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
909 NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
911 zio->io_prop.zp_checksum = checksum;
913 if (zio_checksum_table[checksum].ci_eck) {
915 * zec checksums are necessarily destructive -- they modify
916 * the end of the write buffer to hold the verifier/checksum.
917 * Therefore, we must make a local copy in case the data is
918 * being written to multiple places in parallel.
920 void *wbuf = zio_buf_alloc(size);
921 bcopy(data, wbuf, size);
922 zio_push_transform(zio, wbuf, size, size, NULL);
929 * Create a child I/O to do some work for us.
932 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
933 void *data, uint64_t size, int type, zio_priority_t priority,
934 enum zio_flag flags, zio_done_func_t *done, void *private)
936 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
939 ASSERT(vd->vdev_parent ==
940 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
942 if (type == ZIO_TYPE_READ && bp != NULL) {
944 * If we have the bp, then the child should perform the
945 * checksum and the parent need not. This pushes error
946 * detection as close to the leaves as possible and
947 * eliminates redundant checksums in the interior nodes.
949 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
950 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
953 /* Not all IO types require vdev io done stage e.g. free */
954 if (!(pio->io_pipeline & ZIO_STAGE_VDEV_IO_DONE))
955 pipeline &= ~ZIO_STAGE_VDEV_IO_DONE;
957 if (vd->vdev_children == 0)
958 offset += VDEV_LABEL_START_SIZE;
960 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
963 * If we've decided to do a repair, the write is not speculative --
964 * even if the original read was.
966 if (flags & ZIO_FLAG_IO_REPAIR)
967 flags &= ~ZIO_FLAG_SPECULATIVE;
969 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
970 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
971 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
973 zio->io_physdone = pio->io_physdone;
974 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
975 zio->io_logical->io_phys_children++;
981 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
982 int type, zio_priority_t priority, enum zio_flag flags,
983 zio_done_func_t *done, void *private)
987 ASSERT(vd->vdev_ops->vdev_op_leaf);
989 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
990 data, size, done, private, type, priority,
991 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
993 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
999 zio_flush(zio_t *zio, vdev_t *vd)
1001 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0,
1002 NULL, NULL, ZIO_PRIORITY_NOW,
1003 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1007 zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size)
1010 ASSERT(vd->vdev_ops->vdev_op_leaf);
1012 return (zio_create(zio, spa, 0, NULL, NULL, size, NULL, NULL,
1013 ZIO_TYPE_FREE, ZIO_PRIORITY_TRIM, ZIO_FLAG_DONT_AGGREGATE |
1014 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY,
1015 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PHYS_PIPELINE));
1019 zio_shrink(zio_t *zio, uint64_t size)
1021 ASSERT(zio->io_executor == NULL);
1022 ASSERT(zio->io_orig_size == zio->io_size);
1023 ASSERT(size <= zio->io_size);
1026 * We don't shrink for raidz because of problems with the
1027 * reconstruction when reading back less than the block size.
1028 * Note, BP_IS_RAIDZ() assumes no compression.
1030 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1031 if (!BP_IS_RAIDZ(zio->io_bp))
1032 zio->io_orig_size = zio->io_size = size;
1036 * ==========================================================================
1037 * Prepare to read and write logical blocks
1038 * ==========================================================================
1042 zio_read_bp_init(zio_t *zio)
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 *zio)
1078 spa_t *spa = zio->io_spa;
1079 zio_prop_t *zp = &zio->io_prop;
1080 enum zio_compress compress = zp->zp_compress;
1081 blkptr_t *bp = zio->io_bp;
1082 uint64_t lsize = zio->io_size;
1083 uint64_t psize = lsize;
1087 * If our children haven't all reached the ready stage,
1088 * wait for them and then repeat this pipeline stage.
1090 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1091 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1092 return (ZIO_PIPELINE_STOP);
1094 if (!IO_IS_ALLOCATING(zio))
1095 return (ZIO_PIPELINE_CONTINUE);
1097 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1099 if (zio->io_bp_override) {
1100 ASSERT(bp->blk_birth != zio->io_txg);
1101 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1103 *bp = *zio->io_bp_override;
1104 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1106 if (BP_IS_EMBEDDED(bp))
1107 return (ZIO_PIPELINE_CONTINUE);
1110 * If we've been overridden and nopwrite is set then
1111 * set the flag accordingly to indicate that a nopwrite
1112 * has already occurred.
1114 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1115 ASSERT(!zp->zp_dedup);
1116 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1117 return (ZIO_PIPELINE_CONTINUE);
1120 ASSERT(!zp->zp_nopwrite);
1122 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1123 return (ZIO_PIPELINE_CONTINUE);
1125 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1126 zp->zp_dedup_verify);
1128 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1129 BP_SET_DEDUP(bp, 1);
1130 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1131 return (ZIO_PIPELINE_CONTINUE);
1133 zio->io_bp_override = NULL;
1137 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1139 * We're rewriting an existing block, which means we're
1140 * working on behalf of spa_sync(). For spa_sync() to
1141 * converge, it must eventually be the case that we don't
1142 * have to allocate new blocks. But compression changes
1143 * the blocksize, which forces a reallocate, and makes
1144 * convergence take longer. Therefore, after the first
1145 * few passes, stop compressing to ensure convergence.
1147 pass = spa_sync_pass(spa);
1149 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1150 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1151 ASSERT(!BP_GET_DEDUP(bp));
1153 if (pass >= zfs_sync_pass_dont_compress)
1154 compress = ZIO_COMPRESS_OFF;
1156 /* Make sure someone doesn't change their mind on overwrites */
1157 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1158 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1161 if (compress != ZIO_COMPRESS_OFF) {
1162 void *cbuf = zio_buf_alloc(lsize);
1163 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1164 if (psize == 0 || psize == lsize) {
1165 compress = ZIO_COMPRESS_OFF;
1166 zio_buf_free(cbuf, lsize);
1167 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1168 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1169 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1170 encode_embedded_bp_compressed(bp,
1171 cbuf, compress, lsize, psize);
1172 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1173 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1174 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1175 zio_buf_free(cbuf, lsize);
1176 bp->blk_birth = zio->io_txg;
1177 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1178 ASSERT(spa_feature_is_active(spa,
1179 SPA_FEATURE_EMBEDDED_DATA));
1180 return (ZIO_PIPELINE_CONTINUE);
1183 * Round up compressed size to MINBLOCKSIZE and
1187 P2ROUNDUP(psize, (size_t)SPA_MINBLOCKSIZE);
1188 if (rounded > psize) {
1189 bzero((char *)cbuf + psize, rounded - psize);
1192 if (psize == lsize) {
1193 compress = ZIO_COMPRESS_OFF;
1194 zio_buf_free(cbuf, lsize);
1196 zio_push_transform(zio, cbuf,
1197 psize, lsize, NULL);
1203 * The final pass of spa_sync() must be all rewrites, but the first
1204 * few passes offer a trade-off: allocating blocks defers convergence,
1205 * but newly allocated blocks are sequential, so they can be written
1206 * to disk faster. Therefore, we allow the first few passes of
1207 * spa_sync() to allocate new blocks, but force rewrites after that.
1208 * There should only be a handful of blocks after pass 1 in any case.
1210 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1211 BP_GET_PSIZE(bp) == psize &&
1212 pass >= zfs_sync_pass_rewrite) {
1214 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1215 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1216 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1219 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1223 if (zio->io_bp_orig.blk_birth != 0 &&
1224 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1225 BP_SET_LSIZE(bp, lsize);
1226 BP_SET_TYPE(bp, zp->zp_type);
1227 BP_SET_LEVEL(bp, zp->zp_level);
1228 BP_SET_BIRTH(bp, zio->io_txg, 0);
1230 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1232 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1233 BP_SET_LSIZE(bp, lsize);
1234 BP_SET_TYPE(bp, zp->zp_type);
1235 BP_SET_LEVEL(bp, zp->zp_level);
1236 BP_SET_PSIZE(bp, psize);
1237 BP_SET_COMPRESS(bp, compress);
1238 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1239 BP_SET_DEDUP(bp, zp->zp_dedup);
1240 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1242 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1243 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1244 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1246 if (zp->zp_nopwrite) {
1247 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1248 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1249 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1253 return (ZIO_PIPELINE_CONTINUE);
1257 zio_free_bp_init(zio_t *zio)
1259 blkptr_t *bp = zio->io_bp;
1261 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1262 if (BP_GET_DEDUP(bp))
1263 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1266 return (ZIO_PIPELINE_CONTINUE);
1270 * ==========================================================================
1271 * Execute the I/O pipeline
1272 * ==========================================================================
1276 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1278 spa_t *spa = zio->io_spa;
1279 zio_type_t t = zio->io_type;
1280 int flags = (cutinline ? TQ_FRONT : 0);
1282 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1285 * If we're a config writer or a probe, the normal issue and
1286 * interrupt threads may all be blocked waiting for the config lock.
1287 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1289 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1293 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1295 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1299 * If this is a high priority I/O, then use the high priority taskq if
1302 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1303 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1306 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1309 * NB: We are assuming that the zio can only be dispatched
1310 * to a single taskq at a time. It would be a grievous error
1311 * to dispatch the zio to another taskq at the same time.
1313 #if defined(illumos) || !defined(_KERNEL)
1314 ASSERT(zio->io_tqent.tqent_next == NULL);
1316 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
1318 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1319 flags, &zio->io_tqent);
1323 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1325 kthread_t *executor = zio->io_executor;
1326 spa_t *spa = zio->io_spa;
1328 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1329 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1331 for (i = 0; i < tqs->stqs_count; i++) {
1332 if (taskq_member(tqs->stqs_taskq[i], executor))
1341 zio_issue_async(zio_t *zio)
1343 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1345 return (ZIO_PIPELINE_STOP);
1349 zio_interrupt(zio_t *zio)
1351 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1355 * Execute the I/O pipeline until one of the following occurs:
1357 * (1) the I/O completes
1358 * (2) the pipeline stalls waiting for dependent child I/Os
1359 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1360 * (4) the I/O is delegated by vdev-level caching or aggregation
1361 * (5) the I/O is deferred due to vdev-level queueing
1362 * (6) the I/O is handed off to another thread.
1364 * In all cases, the pipeline stops whenever there's no CPU work; it never
1365 * burns a thread in cv_wait().
1367 * There's no locking on io_stage because there's no legitimate way
1368 * for multiple threads to be attempting to process the same I/O.
1370 static zio_pipe_stage_t *zio_pipeline[];
1373 zio_execute(zio_t *zio)
1375 zio->io_executor = curthread;
1377 while (zio->io_stage < ZIO_STAGE_DONE) {
1378 enum zio_stage pipeline = zio->io_pipeline;
1379 enum zio_stage stage = zio->io_stage;
1382 ASSERT(!MUTEX_HELD(&zio->io_lock));
1383 ASSERT(ISP2(stage));
1384 ASSERT(zio->io_stall == NULL);
1388 } while ((stage & pipeline) == 0);
1390 ASSERT(stage <= ZIO_STAGE_DONE);
1393 * If we are in interrupt context and this pipeline stage
1394 * will grab a config lock that is held across I/O,
1395 * or may wait for an I/O that needs an interrupt thread
1396 * to complete, issue async to avoid deadlock.
1398 * For VDEV_IO_START, we cut in line so that the io will
1399 * be sent to disk promptly.
1401 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1402 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1403 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1404 zio_requeue_io_start_cut_in_line : B_FALSE;
1405 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1409 zio->io_stage = stage;
1410 rv = zio_pipeline[highbit64(stage) - 1](zio);
1412 if (rv == ZIO_PIPELINE_STOP)
1415 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1420 * ==========================================================================
1421 * Initiate I/O, either sync or async
1422 * ==========================================================================
1425 zio_wait(zio_t *zio)
1429 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1430 ASSERT(zio->io_executor == NULL);
1432 zio->io_waiter = curthread;
1436 mutex_enter(&zio->io_lock);
1437 while (zio->io_executor != NULL)
1438 cv_wait(&zio->io_cv, &zio->io_lock);
1439 mutex_exit(&zio->io_lock);
1441 error = zio->io_error;
1448 zio_nowait(zio_t *zio)
1450 ASSERT(zio->io_executor == NULL);
1452 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1453 zio_unique_parent(zio) == NULL) {
1455 * This is a logical async I/O with no parent to wait for it.
1456 * We add it to the spa_async_root_zio "Godfather" I/O which
1457 * will ensure they complete prior to unloading the pool.
1459 spa_t *spa = zio->io_spa;
1461 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1468 * ==========================================================================
1469 * Reexecute or suspend/resume failed I/O
1470 * ==========================================================================
1474 zio_reexecute(zio_t *pio)
1476 zio_t *cio, *cio_next;
1478 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1479 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1480 ASSERT(pio->io_gang_leader == NULL);
1481 ASSERT(pio->io_gang_tree == NULL);
1483 pio->io_flags = pio->io_orig_flags;
1484 pio->io_stage = pio->io_orig_stage;
1485 pio->io_pipeline = pio->io_orig_pipeline;
1486 pio->io_reexecute = 0;
1487 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1489 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1490 pio->io_state[w] = 0;
1491 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1492 pio->io_child_error[c] = 0;
1494 if (IO_IS_ALLOCATING(pio))
1495 BP_ZERO(pio->io_bp);
1498 * As we reexecute pio's children, new children could be created.
1499 * New children go to the head of pio's io_child_list, however,
1500 * so we will (correctly) not reexecute them. The key is that
1501 * the remainder of pio's io_child_list, from 'cio_next' onward,
1502 * cannot be affected by any side effects of reexecuting 'cio'.
1504 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1505 cio_next = zio_walk_children(pio);
1506 mutex_enter(&pio->io_lock);
1507 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1508 pio->io_children[cio->io_child_type][w]++;
1509 mutex_exit(&pio->io_lock);
1514 * Now that all children have been reexecuted, execute the parent.
1515 * We don't reexecute "The Godfather" I/O here as it's the
1516 * responsibility of the caller to wait on him.
1518 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1523 zio_suspend(spa_t *spa, zio_t *zio)
1525 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1526 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1527 "failure and the failure mode property for this pool "
1528 "is set to panic.", spa_name(spa));
1530 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1532 mutex_enter(&spa->spa_suspend_lock);
1534 if (spa->spa_suspend_zio_root == NULL)
1535 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1536 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1537 ZIO_FLAG_GODFATHER);
1539 spa->spa_suspended = B_TRUE;
1542 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1543 ASSERT(zio != spa->spa_suspend_zio_root);
1544 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1545 ASSERT(zio_unique_parent(zio) == NULL);
1546 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1547 zio_add_child(spa->spa_suspend_zio_root, zio);
1550 mutex_exit(&spa->spa_suspend_lock);
1554 zio_resume(spa_t *spa)
1559 * Reexecute all previously suspended i/o.
1561 mutex_enter(&spa->spa_suspend_lock);
1562 spa->spa_suspended = B_FALSE;
1563 cv_broadcast(&spa->spa_suspend_cv);
1564 pio = spa->spa_suspend_zio_root;
1565 spa->spa_suspend_zio_root = NULL;
1566 mutex_exit(&spa->spa_suspend_lock);
1572 return (zio_wait(pio));
1576 zio_resume_wait(spa_t *spa)
1578 mutex_enter(&spa->spa_suspend_lock);
1579 while (spa_suspended(spa))
1580 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1581 mutex_exit(&spa->spa_suspend_lock);
1585 * ==========================================================================
1588 * A gang block is a collection of small blocks that looks to the DMU
1589 * like one large block. When zio_dva_allocate() cannot find a block
1590 * of the requested size, due to either severe fragmentation or the pool
1591 * being nearly full, it calls zio_write_gang_block() to construct the
1592 * block from smaller fragments.
1594 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1595 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1596 * an indirect block: it's an array of block pointers. It consumes
1597 * only one sector and hence is allocatable regardless of fragmentation.
1598 * The gang header's bps point to its gang members, which hold the data.
1600 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1601 * as the verifier to ensure uniqueness of the SHA256 checksum.
1602 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1603 * not the gang header. This ensures that data block signatures (needed for
1604 * deduplication) are independent of how the block is physically stored.
1606 * Gang blocks can be nested: a gang member may itself be a gang block.
1607 * Thus every gang block is a tree in which root and all interior nodes are
1608 * gang headers, and the leaves are normal blocks that contain user data.
1609 * The root of the gang tree is called the gang leader.
1611 * To perform any operation (read, rewrite, free, claim) on a gang block,
1612 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1613 * in the io_gang_tree field of the original logical i/o by recursively
1614 * reading the gang leader and all gang headers below it. This yields
1615 * an in-core tree containing the contents of every gang header and the
1616 * bps for every constituent of the gang block.
1618 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1619 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1620 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1621 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1622 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1623 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1624 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1625 * of the gang header plus zio_checksum_compute() of the data to update the
1626 * gang header's blk_cksum as described above.
1628 * The two-phase assemble/issue model solves the problem of partial failure --
1629 * what if you'd freed part of a gang block but then couldn't read the
1630 * gang header for another part? Assembling the entire gang tree first
1631 * ensures that all the necessary gang header I/O has succeeded before
1632 * starting the actual work of free, claim, or write. Once the gang tree
1633 * is assembled, free and claim are in-memory operations that cannot fail.
1635 * In the event that a gang write fails, zio_dva_unallocate() walks the
1636 * gang tree to immediately free (i.e. insert back into the space map)
1637 * everything we've allocated. This ensures that we don't get ENOSPC
1638 * errors during repeated suspend/resume cycles due to a flaky device.
1640 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1641 * the gang tree, we won't modify the block, so we can safely defer the free
1642 * (knowing that the block is still intact). If we *can* assemble the gang
1643 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1644 * each constituent bp and we can allocate a new block on the next sync pass.
1646 * In all cases, the gang tree allows complete recovery from partial failure.
1647 * ==========================================================================
1651 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1656 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1657 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1658 &pio->io_bookmark));
1662 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1667 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1668 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1669 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1671 * As we rewrite each gang header, the pipeline will compute
1672 * a new gang block header checksum for it; but no one will
1673 * compute a new data checksum, so we do that here. The one
1674 * exception is the gang leader: the pipeline already computed
1675 * its data checksum because that stage precedes gang assembly.
1676 * (Presently, nothing actually uses interior data checksums;
1677 * this is just good hygiene.)
1679 if (gn != pio->io_gang_leader->io_gang_tree) {
1680 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1681 data, BP_GET_PSIZE(bp));
1684 * If we are here to damage data for testing purposes,
1685 * leave the GBH alone so that we can detect the damage.
1687 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1688 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1690 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1691 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1692 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1700 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1702 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1703 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp),
1704 ZIO_GANG_CHILD_FLAGS(pio)));
1709 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1711 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1712 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1715 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1724 static void zio_gang_tree_assemble_done(zio_t *zio);
1726 static zio_gang_node_t *
1727 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1729 zio_gang_node_t *gn;
1731 ASSERT(*gnpp == NULL);
1733 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1734 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1741 zio_gang_node_free(zio_gang_node_t **gnpp)
1743 zio_gang_node_t *gn = *gnpp;
1745 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1746 ASSERT(gn->gn_child[g] == NULL);
1748 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1749 kmem_free(gn, sizeof (*gn));
1754 zio_gang_tree_free(zio_gang_node_t **gnpp)
1756 zio_gang_node_t *gn = *gnpp;
1761 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1762 zio_gang_tree_free(&gn->gn_child[g]);
1764 zio_gang_node_free(gnpp);
1768 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1770 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1772 ASSERT(gio->io_gang_leader == gio);
1773 ASSERT(BP_IS_GANG(bp));
1775 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1776 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1777 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1781 zio_gang_tree_assemble_done(zio_t *zio)
1783 zio_t *gio = zio->io_gang_leader;
1784 zio_gang_node_t *gn = zio->io_private;
1785 blkptr_t *bp = zio->io_bp;
1787 ASSERT(gio == zio_unique_parent(zio));
1788 ASSERT(zio->io_child_count == 0);
1793 if (BP_SHOULD_BYTESWAP(bp))
1794 byteswap_uint64_array(zio->io_data, zio->io_size);
1796 ASSERT(zio->io_data == gn->gn_gbh);
1797 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1798 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1800 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1801 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1802 if (!BP_IS_GANG(gbp))
1804 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1809 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1811 zio_t *gio = pio->io_gang_leader;
1814 ASSERT(BP_IS_GANG(bp) == !!gn);
1815 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1816 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1819 * If you're a gang header, your data is in gn->gn_gbh.
1820 * If you're a gang member, your data is in 'data' and gn == NULL.
1822 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1825 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1827 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1828 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1829 if (BP_IS_HOLE(gbp))
1831 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1832 data = (char *)data + BP_GET_PSIZE(gbp);
1836 if (gn == gio->io_gang_tree && gio->io_data != NULL)
1837 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1844 zio_gang_assemble(zio_t *zio)
1846 blkptr_t *bp = zio->io_bp;
1848 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1849 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1851 zio->io_gang_leader = zio;
1853 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1855 return (ZIO_PIPELINE_CONTINUE);
1859 zio_gang_issue(zio_t *zio)
1861 blkptr_t *bp = zio->io_bp;
1863 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1864 return (ZIO_PIPELINE_STOP);
1866 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1867 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1869 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1870 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1872 zio_gang_tree_free(&zio->io_gang_tree);
1874 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1876 return (ZIO_PIPELINE_CONTINUE);
1880 zio_write_gang_member_ready(zio_t *zio)
1882 zio_t *pio = zio_unique_parent(zio);
1883 zio_t *gio = zio->io_gang_leader;
1884 dva_t *cdva = zio->io_bp->blk_dva;
1885 dva_t *pdva = pio->io_bp->blk_dva;
1888 if (BP_IS_HOLE(zio->io_bp))
1891 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1893 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1894 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1895 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1896 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1897 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1899 mutex_enter(&pio->io_lock);
1900 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1901 ASSERT(DVA_GET_GANG(&pdva[d]));
1902 asize = DVA_GET_ASIZE(&pdva[d]);
1903 asize += DVA_GET_ASIZE(&cdva[d]);
1904 DVA_SET_ASIZE(&pdva[d], asize);
1906 mutex_exit(&pio->io_lock);
1910 zio_write_gang_block(zio_t *pio)
1912 spa_t *spa = pio->io_spa;
1913 blkptr_t *bp = pio->io_bp;
1914 zio_t *gio = pio->io_gang_leader;
1916 zio_gang_node_t *gn, **gnpp;
1917 zio_gbh_phys_t *gbh;
1918 uint64_t txg = pio->io_txg;
1919 uint64_t resid = pio->io_size;
1921 int copies = gio->io_prop.zp_copies;
1922 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1926 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1927 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1928 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1930 pio->io_error = error;
1931 return (ZIO_PIPELINE_CONTINUE);
1935 gnpp = &gio->io_gang_tree;
1937 gnpp = pio->io_private;
1938 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1941 gn = zio_gang_node_alloc(gnpp);
1943 bzero(gbh, SPA_GANGBLOCKSIZE);
1946 * Create the gang header.
1948 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1949 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1952 * Create and nowait the gang children.
1954 for (int g = 0; resid != 0; resid -= lsize, g++) {
1955 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1957 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1959 zp.zp_checksum = gio->io_prop.zp_checksum;
1960 zp.zp_compress = ZIO_COMPRESS_OFF;
1961 zp.zp_type = DMU_OT_NONE;
1963 zp.zp_copies = gio->io_prop.zp_copies;
1964 zp.zp_dedup = B_FALSE;
1965 zp.zp_dedup_verify = B_FALSE;
1966 zp.zp_nopwrite = B_FALSE;
1968 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1969 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1970 zio_write_gang_member_ready, NULL, NULL, &gn->gn_child[g],
1971 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1972 &pio->io_bookmark));
1976 * Set pio's pipeline to just wait for zio to finish.
1978 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1982 return (ZIO_PIPELINE_CONTINUE);
1986 * The zio_nop_write stage in the pipeline determines if allocating
1987 * a new bp is necessary. By leveraging a cryptographically secure checksum,
1988 * such as SHA256, we can compare the checksums of the new data and the old
1989 * to determine if allocating a new block is required. The nopwrite
1990 * feature can handle writes in either syncing or open context (i.e. zil
1991 * writes) and as a result is mutually exclusive with dedup.
1994 zio_nop_write(zio_t *zio)
1996 blkptr_t *bp = zio->io_bp;
1997 blkptr_t *bp_orig = &zio->io_bp_orig;
1998 zio_prop_t *zp = &zio->io_prop;
2000 ASSERT(BP_GET_LEVEL(bp) == 0);
2001 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2002 ASSERT(zp->zp_nopwrite);
2003 ASSERT(!zp->zp_dedup);
2004 ASSERT(zio->io_bp_override == NULL);
2005 ASSERT(IO_IS_ALLOCATING(zio));
2008 * Check to see if the original bp and the new bp have matching
2009 * characteristics (i.e. same checksum, compression algorithms, etc).
2010 * If they don't then just continue with the pipeline which will
2011 * allocate a new bp.
2013 if (BP_IS_HOLE(bp_orig) ||
2014 !zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup ||
2015 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2016 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2017 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2018 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2019 return (ZIO_PIPELINE_CONTINUE);
2022 * If the checksums match then reset the pipeline so that we
2023 * avoid allocating a new bp and issuing any I/O.
2025 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2026 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup);
2027 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2028 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2029 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2030 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2031 sizeof (uint64_t)) == 0);
2034 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2035 zio->io_flags |= ZIO_FLAG_NOPWRITE;
2038 return (ZIO_PIPELINE_CONTINUE);
2042 * ==========================================================================
2044 * ==========================================================================
2047 zio_ddt_child_read_done(zio_t *zio)
2049 blkptr_t *bp = zio->io_bp;
2050 ddt_entry_t *dde = zio->io_private;
2052 zio_t *pio = zio_unique_parent(zio);
2054 mutex_enter(&pio->io_lock);
2055 ddp = ddt_phys_select(dde, bp);
2056 if (zio->io_error == 0)
2057 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2058 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
2059 dde->dde_repair_data = zio->io_data;
2061 zio_buf_free(zio->io_data, zio->io_size);
2062 mutex_exit(&pio->io_lock);
2066 zio_ddt_read_start(zio_t *zio)
2068 blkptr_t *bp = zio->io_bp;
2070 ASSERT(BP_GET_DEDUP(bp));
2071 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2072 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2074 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2075 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2076 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2077 ddt_phys_t *ddp = dde->dde_phys;
2078 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2081 ASSERT(zio->io_vsd == NULL);
2084 if (ddp_self == NULL)
2085 return (ZIO_PIPELINE_CONTINUE);
2087 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2088 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2090 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2092 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2093 zio_buf_alloc(zio->io_size), zio->io_size,
2094 zio_ddt_child_read_done, dde, zio->io_priority,
2095 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
2096 &zio->io_bookmark));
2098 return (ZIO_PIPELINE_CONTINUE);
2101 zio_nowait(zio_read(zio, zio->io_spa, bp,
2102 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
2103 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2105 return (ZIO_PIPELINE_CONTINUE);
2109 zio_ddt_read_done(zio_t *zio)
2111 blkptr_t *bp = zio->io_bp;
2113 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2114 return (ZIO_PIPELINE_STOP);
2116 ASSERT(BP_GET_DEDUP(bp));
2117 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2118 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2120 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2121 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2122 ddt_entry_t *dde = zio->io_vsd;
2124 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2125 return (ZIO_PIPELINE_CONTINUE);
2128 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2129 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2130 return (ZIO_PIPELINE_STOP);
2132 if (dde->dde_repair_data != NULL) {
2133 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2134 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2136 ddt_repair_done(ddt, dde);
2140 ASSERT(zio->io_vsd == NULL);
2142 return (ZIO_PIPELINE_CONTINUE);
2146 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2148 spa_t *spa = zio->io_spa;
2151 * Note: we compare the original data, not the transformed data,
2152 * because when zio->io_bp is an override bp, we will not have
2153 * pushed the I/O transforms. That's an important optimization
2154 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2156 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2157 zio_t *lio = dde->dde_lead_zio[p];
2160 return (lio->io_orig_size != zio->io_orig_size ||
2161 bcmp(zio->io_orig_data, lio->io_orig_data,
2162 zio->io_orig_size) != 0);
2166 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2167 ddt_phys_t *ddp = &dde->dde_phys[p];
2169 if (ddp->ddp_phys_birth != 0) {
2170 arc_buf_t *abuf = NULL;
2171 uint32_t aflags = ARC_WAIT;
2172 blkptr_t blk = *zio->io_bp;
2175 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2179 error = arc_read(NULL, spa, &blk,
2180 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2181 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2182 &aflags, &zio->io_bookmark);
2185 if (arc_buf_size(abuf) != zio->io_orig_size ||
2186 bcmp(abuf->b_data, zio->io_orig_data,
2187 zio->io_orig_size) != 0)
2188 error = SET_ERROR(EEXIST);
2189 VERIFY(arc_buf_remove_ref(abuf, &abuf));
2193 return (error != 0);
2201 zio_ddt_child_write_ready(zio_t *zio)
2203 int p = zio->io_prop.zp_copies;
2204 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2205 ddt_entry_t *dde = zio->io_private;
2206 ddt_phys_t *ddp = &dde->dde_phys[p];
2214 ASSERT(dde->dde_lead_zio[p] == zio);
2216 ddt_phys_fill(ddp, zio->io_bp);
2218 while ((pio = zio_walk_parents(zio)) != NULL)
2219 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2225 zio_ddt_child_write_done(zio_t *zio)
2227 int p = zio->io_prop.zp_copies;
2228 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2229 ddt_entry_t *dde = zio->io_private;
2230 ddt_phys_t *ddp = &dde->dde_phys[p];
2234 ASSERT(ddp->ddp_refcnt == 0);
2235 ASSERT(dde->dde_lead_zio[p] == zio);
2236 dde->dde_lead_zio[p] = NULL;
2238 if (zio->io_error == 0) {
2239 while (zio_walk_parents(zio) != NULL)
2240 ddt_phys_addref(ddp);
2242 ddt_phys_clear(ddp);
2249 zio_ddt_ditto_write_done(zio_t *zio)
2251 int p = DDT_PHYS_DITTO;
2252 zio_prop_t *zp = &zio->io_prop;
2253 blkptr_t *bp = zio->io_bp;
2254 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2255 ddt_entry_t *dde = zio->io_private;
2256 ddt_phys_t *ddp = &dde->dde_phys[p];
2257 ddt_key_t *ddk = &dde->dde_key;
2261 ASSERT(ddp->ddp_refcnt == 0);
2262 ASSERT(dde->dde_lead_zio[p] == zio);
2263 dde->dde_lead_zio[p] = NULL;
2265 if (zio->io_error == 0) {
2266 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2267 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2268 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2269 if (ddp->ddp_phys_birth != 0)
2270 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2271 ddt_phys_fill(ddp, bp);
2278 zio_ddt_write(zio_t *zio)
2280 spa_t *spa = zio->io_spa;
2281 blkptr_t *bp = zio->io_bp;
2282 uint64_t txg = zio->io_txg;
2283 zio_prop_t *zp = &zio->io_prop;
2284 int p = zp->zp_copies;
2288 ddt_t *ddt = ddt_select(spa, bp);
2292 ASSERT(BP_GET_DEDUP(bp));
2293 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2294 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2297 dde = ddt_lookup(ddt, bp, B_TRUE);
2298 ddp = &dde->dde_phys[p];
2300 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2302 * If we're using a weak checksum, upgrade to a strong checksum
2303 * and try again. If we're already using a strong checksum,
2304 * we can't resolve it, so just convert to an ordinary write.
2305 * (And automatically e-mail a paper to Nature?)
2307 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2308 zp->zp_checksum = spa_dedup_checksum(spa);
2309 zio_pop_transforms(zio);
2310 zio->io_stage = ZIO_STAGE_OPEN;
2313 zp->zp_dedup = B_FALSE;
2315 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2317 return (ZIO_PIPELINE_CONTINUE);
2320 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2321 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2323 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2324 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2325 zio_prop_t czp = *zp;
2327 czp.zp_copies = ditto_copies;
2330 * If we arrived here with an override bp, we won't have run
2331 * the transform stack, so we won't have the data we need to
2332 * generate a child i/o. So, toss the override bp and restart.
2333 * This is safe, because using the override bp is just an
2334 * optimization; and it's rare, so the cost doesn't matter.
2336 if (zio->io_bp_override) {
2337 zio_pop_transforms(zio);
2338 zio->io_stage = ZIO_STAGE_OPEN;
2339 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2340 zio->io_bp_override = NULL;
2343 return (ZIO_PIPELINE_CONTINUE);
2346 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2347 zio->io_orig_size, &czp, NULL, NULL,
2348 zio_ddt_ditto_write_done, dde, zio->io_priority,
2349 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2351 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2352 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2355 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2356 if (ddp->ddp_phys_birth != 0)
2357 ddt_bp_fill(ddp, bp, txg);
2358 if (dde->dde_lead_zio[p] != NULL)
2359 zio_add_child(zio, dde->dde_lead_zio[p]);
2361 ddt_phys_addref(ddp);
2362 } else if (zio->io_bp_override) {
2363 ASSERT(bp->blk_birth == txg);
2364 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2365 ddt_phys_fill(ddp, bp);
2366 ddt_phys_addref(ddp);
2368 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2369 zio->io_orig_size, zp, zio_ddt_child_write_ready, NULL,
2370 zio_ddt_child_write_done, dde, zio->io_priority,
2371 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2373 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2374 dde->dde_lead_zio[p] = cio;
2384 return (ZIO_PIPELINE_CONTINUE);
2387 ddt_entry_t *freedde; /* for debugging */
2390 zio_ddt_free(zio_t *zio)
2392 spa_t *spa = zio->io_spa;
2393 blkptr_t *bp = zio->io_bp;
2394 ddt_t *ddt = ddt_select(spa, bp);
2398 ASSERT(BP_GET_DEDUP(bp));
2399 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2402 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2403 ddp = ddt_phys_select(dde, bp);
2404 ddt_phys_decref(ddp);
2407 return (ZIO_PIPELINE_CONTINUE);
2411 * ==========================================================================
2412 * Allocate and free blocks
2413 * ==========================================================================
2416 zio_dva_allocate(zio_t *zio)
2418 spa_t *spa = zio->io_spa;
2419 metaslab_class_t *mc = spa_normal_class(spa);
2420 blkptr_t *bp = zio->io_bp;
2424 if (zio->io_gang_leader == NULL) {
2425 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2426 zio->io_gang_leader = zio;
2429 ASSERT(BP_IS_HOLE(bp));
2430 ASSERT0(BP_GET_NDVAS(bp));
2431 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2432 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2433 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2436 * The dump device does not support gang blocks so allocation on
2437 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2438 * the "fast" gang feature.
2440 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2441 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2442 METASLAB_GANG_CHILD : 0;
2443 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2444 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2447 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2448 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2450 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2451 return (zio_write_gang_block(zio));
2452 zio->io_error = error;
2455 return (ZIO_PIPELINE_CONTINUE);
2459 zio_dva_free(zio_t *zio)
2461 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2463 return (ZIO_PIPELINE_CONTINUE);
2467 zio_dva_claim(zio_t *zio)
2471 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2473 zio->io_error = error;
2475 return (ZIO_PIPELINE_CONTINUE);
2479 * Undo an allocation. This is used by zio_done() when an I/O fails
2480 * and we want to give back the block we just allocated.
2481 * This handles both normal blocks and gang blocks.
2484 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2486 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2487 ASSERT(zio->io_bp_override == NULL);
2489 if (!BP_IS_HOLE(bp))
2490 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2493 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2494 zio_dva_unallocate(zio, gn->gn_child[g],
2495 &gn->gn_gbh->zg_blkptr[g]);
2501 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2504 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2505 uint64_t size, boolean_t use_slog)
2509 ASSERT(txg > spa_syncing_txg(spa));
2512 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2513 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2514 * when allocating them.
2517 error = metaslab_alloc(spa, spa_log_class(spa), size,
2518 new_bp, 1, txg, old_bp,
2519 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2523 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2524 new_bp, 1, txg, old_bp,
2525 METASLAB_HINTBP_AVOID);
2529 BP_SET_LSIZE(new_bp, size);
2530 BP_SET_PSIZE(new_bp, size);
2531 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2532 BP_SET_CHECKSUM(new_bp,
2533 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2534 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2535 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2536 BP_SET_LEVEL(new_bp, 0);
2537 BP_SET_DEDUP(new_bp, 0);
2538 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2545 * Free an intent log block.
2548 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2550 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2551 ASSERT(!BP_IS_GANG(bp));
2553 zio_free(spa, txg, bp);
2557 * ==========================================================================
2558 * Read, write and delete to physical devices
2559 * ==========================================================================
2564 * Issue an I/O to the underlying vdev. Typically the issue pipeline
2565 * stops after this stage and will resume upon I/O completion.
2566 * However, there are instances where the vdev layer may need to
2567 * continue the pipeline when an I/O was not issued. Since the I/O
2568 * that was sent to the vdev layer might be different than the one
2569 * currently active in the pipeline (see vdev_queue_io()), we explicitly
2570 * force the underlying vdev layers to call either zio_execute() or
2571 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
2574 zio_vdev_io_start(zio_t *zio)
2576 vdev_t *vd = zio->io_vd;
2578 spa_t *spa = zio->io_spa;
2581 ASSERT(zio->io_error == 0);
2582 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2585 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2586 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2589 * The mirror_ops handle multiple DVAs in a single BP.
2591 vdev_mirror_ops.vdev_op_io_start(zio);
2592 return (ZIO_PIPELINE_STOP);
2595 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE &&
2596 zio->io_priority == ZIO_PRIORITY_NOW) {
2597 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg);
2598 return (ZIO_PIPELINE_CONTINUE);
2602 * We keep track of time-sensitive I/Os so that the scan thread
2603 * can quickly react to certain workloads. In particular, we care
2604 * about non-scrubbing, top-level reads and writes with the following
2606 * - synchronous writes of user data to non-slog devices
2607 * - any reads of user data
2608 * When these conditions are met, adjust the timestamp of spa_last_io
2609 * which allows the scan thread to adjust its workload accordingly.
2611 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2612 vd == vd->vdev_top && !vd->vdev_islog &&
2613 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2614 zio->io_txg != spa_syncing_txg(spa)) {
2615 uint64_t old = spa->spa_last_io;
2616 uint64_t new = ddi_get_lbolt64();
2618 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2621 align = 1ULL << vd->vdev_top->vdev_ashift;
2623 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
2624 P2PHASE(zio->io_size, align) != 0) {
2625 /* Transform logical writes to be a full physical block size. */
2626 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2628 if (zio->io_type == ZIO_TYPE_READ ||
2629 zio->io_type == ZIO_TYPE_WRITE)
2630 abuf = zio_buf_alloc(asize);
2631 ASSERT(vd == vd->vdev_top);
2632 if (zio->io_type == ZIO_TYPE_WRITE) {
2633 bcopy(zio->io_data, abuf, zio->io_size);
2634 bzero(abuf + zio->io_size, asize - zio->io_size);
2636 zio_push_transform(zio, abuf, asize, abuf ? asize : 0,
2641 * If this is not a physical io, make sure that it is properly aligned
2642 * before proceeding.
2644 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
2645 ASSERT0(P2PHASE(zio->io_offset, align));
2646 ASSERT0(P2PHASE(zio->io_size, align));
2649 * For physical writes, we allow 512b aligned writes and assume
2650 * the device will perform a read-modify-write as necessary.
2652 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
2653 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
2656 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa));
2659 * If this is a repair I/O, and there's no self-healing involved --
2660 * that is, we're just resilvering what we expect to resilver --
2661 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2662 * This prevents spurious resilvering with nested replication.
2663 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2664 * A is out of date, we'll read from C+D, then use the data to
2665 * resilver A+B -- but we don't actually want to resilver B, just A.
2666 * The top-level mirror has no way to know this, so instead we just
2667 * discard unnecessary repairs as we work our way down the vdev tree.
2668 * The same logic applies to any form of nested replication:
2669 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2671 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2672 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2673 zio->io_txg != 0 && /* not a delegated i/o */
2674 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2675 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2676 zio_vdev_io_bypass(zio);
2677 return (ZIO_PIPELINE_CONTINUE);
2680 if (vd->vdev_ops->vdev_op_leaf) {
2681 switch (zio->io_type) {
2683 if (vdev_cache_read(zio))
2684 return (ZIO_PIPELINE_CONTINUE);
2686 case ZIO_TYPE_WRITE:
2688 if ((zio = vdev_queue_io(zio)) == NULL)
2689 return (ZIO_PIPELINE_STOP);
2691 if (!vdev_accessible(vd, zio)) {
2692 zio->io_error = SET_ERROR(ENXIO);
2694 return (ZIO_PIPELINE_STOP);
2699 * Note that we ignore repair writes for TRIM because they can
2700 * conflict with normal writes. This isn't an issue because, by
2701 * definition, we only repair blocks that aren't freed.
2703 if (zio->io_type == ZIO_TYPE_WRITE &&
2704 !(zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2705 !trim_map_write_start(zio))
2706 return (ZIO_PIPELINE_STOP);
2709 vd->vdev_ops->vdev_op_io_start(zio);
2710 return (ZIO_PIPELINE_STOP);
2714 zio_vdev_io_done(zio_t *zio)
2716 vdev_t *vd = zio->io_vd;
2717 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2718 boolean_t unexpected_error = B_FALSE;
2720 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2721 return (ZIO_PIPELINE_STOP);
2723 ASSERT(zio->io_type == ZIO_TYPE_READ ||
2724 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE);
2726 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2727 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE ||
2728 zio->io_type == ZIO_TYPE_FREE)) {
2730 if (zio->io_type == ZIO_TYPE_WRITE &&
2731 !(zio->io_flags & ZIO_FLAG_IO_REPAIR))
2732 trim_map_write_done(zio);
2734 vdev_queue_io_done(zio);
2736 if (zio->io_type == ZIO_TYPE_WRITE)
2737 vdev_cache_write(zio);
2739 if (zio_injection_enabled && zio->io_error == 0)
2740 zio->io_error = zio_handle_device_injection(vd,
2743 if (zio_injection_enabled && zio->io_error == 0)
2744 zio->io_error = zio_handle_label_injection(zio, EIO);
2746 if (zio->io_error) {
2747 if (zio->io_error == ENOTSUP &&
2748 zio->io_type == ZIO_TYPE_FREE) {
2749 /* Not all devices support TRIM. */
2750 } else if (!vdev_accessible(vd, zio)) {
2751 zio->io_error = SET_ERROR(ENXIO);
2753 unexpected_error = B_TRUE;
2758 ops->vdev_op_io_done(zio);
2760 if (unexpected_error)
2761 VERIFY(vdev_probe(vd, zio) == NULL);
2763 return (ZIO_PIPELINE_CONTINUE);
2767 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2768 * disk, and use that to finish the checksum ereport later.
2771 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2772 const void *good_buf)
2774 /* no processing needed */
2775 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2780 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2782 void *buf = zio_buf_alloc(zio->io_size);
2784 bcopy(zio->io_data, buf, zio->io_size);
2786 zcr->zcr_cbinfo = zio->io_size;
2787 zcr->zcr_cbdata = buf;
2788 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2789 zcr->zcr_free = zio_buf_free;
2793 zio_vdev_io_assess(zio_t *zio)
2795 vdev_t *vd = zio->io_vd;
2797 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2798 return (ZIO_PIPELINE_STOP);
2800 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2801 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2803 if (zio->io_vsd != NULL) {
2804 zio->io_vsd_ops->vsd_free(zio);
2808 if (zio_injection_enabled && zio->io_error == 0)
2809 zio->io_error = zio_handle_fault_injection(zio, EIO);
2811 if (zio->io_type == ZIO_TYPE_FREE &&
2812 zio->io_priority != ZIO_PRIORITY_NOW) {
2813 switch (zio->io_error) {
2815 ZIO_TRIM_STAT_INCR(bytes, zio->io_size);
2816 ZIO_TRIM_STAT_BUMP(success);
2819 ZIO_TRIM_STAT_BUMP(unsupported);
2822 ZIO_TRIM_STAT_BUMP(failed);
2828 * If the I/O failed, determine whether we should attempt to retry it.
2830 * On retry, we cut in line in the issue queue, since we don't want
2831 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2833 if (zio->io_error && vd == NULL &&
2834 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2835 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2836 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2838 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2839 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2840 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2841 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2842 zio_requeue_io_start_cut_in_line);
2843 return (ZIO_PIPELINE_STOP);
2847 * If we got an error on a leaf device, convert it to ENXIO
2848 * if the device is not accessible at all.
2850 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2851 !vdev_accessible(vd, zio))
2852 zio->io_error = SET_ERROR(ENXIO);
2855 * If we can't write to an interior vdev (mirror or RAID-Z),
2856 * set vdev_cant_write so that we stop trying to allocate from it.
2858 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2859 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
2860 vd->vdev_cant_write = B_TRUE;
2864 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2866 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2867 zio->io_physdone != NULL) {
2868 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
2869 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
2870 zio->io_physdone(zio->io_logical);
2873 return (ZIO_PIPELINE_CONTINUE);
2877 zio_vdev_io_reissue(zio_t *zio)
2879 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2880 ASSERT(zio->io_error == 0);
2882 zio->io_stage >>= 1;
2886 zio_vdev_io_redone(zio_t *zio)
2888 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2890 zio->io_stage >>= 1;
2894 zio_vdev_io_bypass(zio_t *zio)
2896 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2897 ASSERT(zio->io_error == 0);
2899 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2900 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2904 * ==========================================================================
2905 * Generate and verify checksums
2906 * ==========================================================================
2909 zio_checksum_generate(zio_t *zio)
2911 blkptr_t *bp = zio->io_bp;
2912 enum zio_checksum checksum;
2916 * This is zio_write_phys().
2917 * We're either generating a label checksum, or none at all.
2919 checksum = zio->io_prop.zp_checksum;
2921 if (checksum == ZIO_CHECKSUM_OFF)
2922 return (ZIO_PIPELINE_CONTINUE);
2924 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2926 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2927 ASSERT(!IO_IS_ALLOCATING(zio));
2928 checksum = ZIO_CHECKSUM_GANG_HEADER;
2930 checksum = BP_GET_CHECKSUM(bp);
2934 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2936 return (ZIO_PIPELINE_CONTINUE);
2940 zio_checksum_verify(zio_t *zio)
2942 zio_bad_cksum_t info;
2943 blkptr_t *bp = zio->io_bp;
2946 ASSERT(zio->io_vd != NULL);
2950 * This is zio_read_phys().
2951 * We're either verifying a label checksum, or nothing at all.
2953 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2954 return (ZIO_PIPELINE_CONTINUE);
2956 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2959 if ((error = zio_checksum_error(zio, &info)) != 0) {
2960 zio->io_error = error;
2961 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2962 zfs_ereport_start_checksum(zio->io_spa,
2963 zio->io_vd, zio, zio->io_offset,
2964 zio->io_size, NULL, &info);
2968 return (ZIO_PIPELINE_CONTINUE);
2972 * Called by RAID-Z to ensure we don't compute the checksum twice.
2975 zio_checksum_verified(zio_t *zio)
2977 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2981 * ==========================================================================
2982 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2983 * An error of 0 indicates success. ENXIO indicates whole-device failure,
2984 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2985 * indicate errors that are specific to one I/O, and most likely permanent.
2986 * Any other error is presumed to be worse because we weren't expecting it.
2987 * ==========================================================================
2990 zio_worst_error(int e1, int e2)
2992 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2995 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2996 if (e1 == zio_error_rank[r1])
2999 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
3000 if (e2 == zio_error_rank[r2])
3003 return (r1 > r2 ? e1 : e2);
3007 * ==========================================================================
3009 * ==========================================================================
3012 zio_ready(zio_t *zio)
3014 blkptr_t *bp = zio->io_bp;
3015 zio_t *pio, *pio_next;
3017 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
3018 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
3019 return (ZIO_PIPELINE_STOP);
3021 if (zio->io_ready) {
3022 ASSERT(IO_IS_ALLOCATING(zio));
3023 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
3024 (zio->io_flags & ZIO_FLAG_NOPWRITE));
3025 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
3030 if (bp != NULL && bp != &zio->io_bp_copy)
3031 zio->io_bp_copy = *bp;
3034 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3036 mutex_enter(&zio->io_lock);
3037 zio->io_state[ZIO_WAIT_READY] = 1;
3038 pio = zio_walk_parents(zio);
3039 mutex_exit(&zio->io_lock);
3042 * As we notify zio's parents, new parents could be added.
3043 * New parents go to the head of zio's io_parent_list, however,
3044 * so we will (correctly) not notify them. The remainder of zio's
3045 * io_parent_list, from 'pio_next' onward, cannot change because
3046 * all parents must wait for us to be done before they can be done.
3048 for (; pio != NULL; pio = pio_next) {
3049 pio_next = zio_walk_parents(zio);
3050 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
3053 if (zio->io_flags & ZIO_FLAG_NODATA) {
3054 if (BP_IS_GANG(bp)) {
3055 zio->io_flags &= ~ZIO_FLAG_NODATA;
3057 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
3058 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
3062 if (zio_injection_enabled &&
3063 zio->io_spa->spa_syncing_txg == zio->io_txg)
3064 zio_handle_ignored_writes(zio);
3066 return (ZIO_PIPELINE_CONTINUE);
3070 zio_done(zio_t *zio)
3072 spa_t *spa = zio->io_spa;
3073 zio_t *lio = zio->io_logical;
3074 blkptr_t *bp = zio->io_bp;
3075 vdev_t *vd = zio->io_vd;
3076 uint64_t psize = zio->io_size;
3077 zio_t *pio, *pio_next;
3080 * If our children haven't all completed,
3081 * wait for them and then repeat this pipeline stage.
3083 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
3084 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
3085 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
3086 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
3087 return (ZIO_PIPELINE_STOP);
3089 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3090 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3091 ASSERT(zio->io_children[c][w] == 0);
3093 if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
3094 ASSERT(bp->blk_pad[0] == 0);
3095 ASSERT(bp->blk_pad[1] == 0);
3096 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3097 (bp == zio_unique_parent(zio)->io_bp));
3098 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3099 zio->io_bp_override == NULL &&
3100 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3101 ASSERT(!BP_SHOULD_BYTESWAP(bp));
3102 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3103 ASSERT(BP_COUNT_GANG(bp) == 0 ||
3104 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3106 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3107 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3111 * If there were child vdev/gang/ddt errors, they apply to us now.
3113 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3114 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3115 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3118 * If the I/O on the transformed data was successful, generate any
3119 * checksum reports now while we still have the transformed data.
3121 if (zio->io_error == 0) {
3122 while (zio->io_cksum_report != NULL) {
3123 zio_cksum_report_t *zcr = zio->io_cksum_report;
3124 uint64_t align = zcr->zcr_align;
3125 uint64_t asize = P2ROUNDUP(psize, align);
3126 char *abuf = zio->io_data;
3128 if (asize != psize) {
3129 abuf = zio_buf_alloc(asize);
3130 bcopy(zio->io_data, abuf, psize);
3131 bzero(abuf + psize, asize - psize);
3134 zio->io_cksum_report = zcr->zcr_next;
3135 zcr->zcr_next = NULL;
3136 zcr->zcr_finish(zcr, abuf);
3137 zfs_ereport_free_checksum(zcr);
3140 zio_buf_free(abuf, asize);
3144 zio_pop_transforms(zio); /* note: may set zio->io_error */
3146 vdev_stat_update(zio, psize);
3148 if (zio->io_error) {
3150 * If this I/O is attached to a particular vdev,
3151 * generate an error message describing the I/O failure
3152 * at the block level. We ignore these errors if the
3153 * device is currently unavailable.
3155 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3156 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3158 if ((zio->io_error == EIO || !(zio->io_flags &
3159 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3162 * For logical I/O requests, tell the SPA to log the
3163 * error and generate a logical data ereport.
3165 spa_log_error(spa, zio);
3166 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3171 if (zio->io_error && zio == lio) {
3173 * Determine whether zio should be reexecuted. This will
3174 * propagate all the way to the root via zio_notify_parent().
3176 ASSERT(vd == NULL && bp != NULL);
3177 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3179 if (IO_IS_ALLOCATING(zio) &&
3180 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3181 if (zio->io_error != ENOSPC)
3182 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3184 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3187 if ((zio->io_type == ZIO_TYPE_READ ||
3188 zio->io_type == ZIO_TYPE_FREE) &&
3189 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3190 zio->io_error == ENXIO &&
3191 spa_load_state(spa) == SPA_LOAD_NONE &&
3192 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3193 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3195 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3196 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3199 * Here is a possibly good place to attempt to do
3200 * either combinatorial reconstruction or error correction
3201 * based on checksums. It also might be a good place
3202 * to send out preliminary ereports before we suspend
3208 * If there were logical child errors, they apply to us now.
3209 * We defer this until now to avoid conflating logical child
3210 * errors with errors that happened to the zio itself when
3211 * updating vdev stats and reporting FMA events above.
3213 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3215 if ((zio->io_error || zio->io_reexecute) &&
3216 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3217 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3218 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3220 zio_gang_tree_free(&zio->io_gang_tree);
3223 * Godfather I/Os should never suspend.
3225 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3226 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3227 zio->io_reexecute = 0;
3229 if (zio->io_reexecute) {
3231 * This is a logical I/O that wants to reexecute.
3233 * Reexecute is top-down. When an i/o fails, if it's not
3234 * the root, it simply notifies its parent and sticks around.
3235 * The parent, seeing that it still has children in zio_done(),
3236 * does the same. This percolates all the way up to the root.
3237 * The root i/o will reexecute or suspend the entire tree.
3239 * This approach ensures that zio_reexecute() honors
3240 * all the original i/o dependency relationships, e.g.
3241 * parents not executing until children are ready.
3243 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3245 zio->io_gang_leader = NULL;
3247 mutex_enter(&zio->io_lock);
3248 zio->io_state[ZIO_WAIT_DONE] = 1;
3249 mutex_exit(&zio->io_lock);
3252 * "The Godfather" I/O monitors its children but is
3253 * not a true parent to them. It will track them through
3254 * the pipeline but severs its ties whenever they get into
3255 * trouble (e.g. suspended). This allows "The Godfather"
3256 * I/O to return status without blocking.
3258 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3259 zio_link_t *zl = zio->io_walk_link;
3260 pio_next = zio_walk_parents(zio);
3262 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3263 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3264 zio_remove_child(pio, zio, zl);
3265 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3269 if ((pio = zio_unique_parent(zio)) != NULL) {
3271 * We're not a root i/o, so there's nothing to do
3272 * but notify our parent. Don't propagate errors
3273 * upward since we haven't permanently failed yet.
3275 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3276 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3277 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3278 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3280 * We'd fail again if we reexecuted now, so suspend
3281 * until conditions improve (e.g. device comes online).
3283 zio_suspend(spa, zio);
3286 * Reexecution is potentially a huge amount of work.
3287 * Hand it off to the otherwise-unused claim taskq.
3289 #if defined(illumos) || !defined(_KERNEL)
3290 ASSERT(zio->io_tqent.tqent_next == NULL);
3292 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
3294 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3295 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3298 return (ZIO_PIPELINE_STOP);
3301 ASSERT(zio->io_child_count == 0);
3302 ASSERT(zio->io_reexecute == 0);
3303 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3306 * Report any checksum errors, since the I/O is complete.
3308 while (zio->io_cksum_report != NULL) {
3309 zio_cksum_report_t *zcr = zio->io_cksum_report;
3310 zio->io_cksum_report = zcr->zcr_next;
3311 zcr->zcr_next = NULL;
3312 zcr->zcr_finish(zcr, NULL);
3313 zfs_ereport_free_checksum(zcr);
3317 * It is the responsibility of the done callback to ensure that this
3318 * particular zio is no longer discoverable for adoption, and as
3319 * such, cannot acquire any new parents.
3324 mutex_enter(&zio->io_lock);
3325 zio->io_state[ZIO_WAIT_DONE] = 1;
3326 mutex_exit(&zio->io_lock);
3328 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3329 zio_link_t *zl = zio->io_walk_link;
3330 pio_next = zio_walk_parents(zio);
3331 zio_remove_child(pio, zio, zl);
3332 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3335 if (zio->io_waiter != NULL) {
3336 mutex_enter(&zio->io_lock);
3337 zio->io_executor = NULL;
3338 cv_broadcast(&zio->io_cv);
3339 mutex_exit(&zio->io_lock);
3344 return (ZIO_PIPELINE_STOP);
3348 * ==========================================================================
3349 * I/O pipeline definition
3350 * ==========================================================================
3352 static zio_pipe_stage_t *zio_pipeline[] = {
3358 zio_checksum_generate,
3373 zio_checksum_verify,
3377 /* dnp is the dnode for zb1->zb_object */
3379 zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_phys_t *zb1,
3380 const zbookmark_phys_t *zb2)
3382 uint64_t zb1nextL0, zb2thisobj;
3384 ASSERT(zb1->zb_objset == zb2->zb_objset);
3385 ASSERT(zb2->zb_level == 0);
3387 /* The objset_phys_t isn't before anything. */
3391 zb1nextL0 = (zb1->zb_blkid + 1) <<
3392 ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
3394 zb2thisobj = zb2->zb_object ? zb2->zb_object :
3395 zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
3397 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3398 uint64_t nextobj = zb1nextL0 *
3399 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
3400 return (nextobj <= zb2thisobj);
3403 if (zb1->zb_object < zb2thisobj)
3405 if (zb1->zb_object > zb2thisobj)
3407 if (zb2->zb_object == DMU_META_DNODE_OBJECT)
3409 return (zb1nextL0 <= zb2->zb_blkid);