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, 2018 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
25 * Copyright (c) 2014 Integros [integros.com]
26 * Copyright (c) 2017, Intel Corporation.
29 #include <sys/sysmacros.h>
30 #include <sys/zfs_context.h>
31 #include <sys/fm/fs/zfs.h>
34 #include <sys/spa_impl.h>
35 #include <sys/vdev_impl.h>
36 #include <sys/zio_impl.h>
37 #include <sys/zio_compress.h>
38 #include <sys/zio_checksum.h>
39 #include <sys/dmu_objset.h>
42 #include <sys/trim_map.h>
43 #include <sys/blkptr.h>
44 #include <sys/zfeature.h>
45 #include <sys/dsl_scan.h>
46 #include <sys/metaslab_impl.h>
48 #include <sys/cityhash.h>
50 SYSCTL_DECL(_vfs_zfs);
51 SYSCTL_NODE(_vfs_zfs, OID_AUTO, zio, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
53 #if defined(__amd64__)
54 static int zio_use_uma = 1;
56 static int zio_use_uma = 0;
58 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, use_uma, CTLFLAG_RDTUN, &zio_use_uma, 0,
59 "Use uma(9) for ZIO allocations");
60 static int zio_exclude_metadata = 0;
61 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, exclude_metadata, CTLFLAG_RDTUN, &zio_exclude_metadata, 0,
62 "Exclude metadata buffers from dumps as well");
64 zio_trim_stats_t zio_trim_stats = {
65 { "bytes", KSTAT_DATA_UINT64,
66 "Number of bytes successfully TRIMmed" },
67 { "success", KSTAT_DATA_UINT64,
68 "Number of successful TRIM requests" },
69 { "unsupported", KSTAT_DATA_UINT64,
70 "Number of TRIM requests that failed because TRIM is not supported" },
71 { "failed", KSTAT_DATA_UINT64,
72 "Number of TRIM requests that failed for reasons other than not supported" },
75 static kstat_t *zio_trim_ksp;
78 * ==========================================================================
79 * I/O type descriptions
80 * ==========================================================================
82 const char *zio_type_name[ZIO_TYPES] = {
83 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
87 boolean_t zio_dva_throttle_enabled = B_TRUE;
88 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, dva_throttle_enabled, CTLFLAG_RWTUN,
89 &zio_dva_throttle_enabled, 0, "Enable allocation throttling");
92 * ==========================================================================
94 * ==========================================================================
96 kmem_cache_t *zio_cache;
97 kmem_cache_t *zio_link_cache;
98 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
99 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
102 extern vmem_t *zio_alloc_arena;
105 #define BP_SPANB(indblkshift, level) \
106 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
107 #define COMPARE_META_LEVEL 0x80000000ul
109 * The following actions directly effect the spa's sync-to-convergence logic.
110 * The values below define the sync pass when we start performing the action.
111 * Care should be taken when changing these values as they directly impact
112 * spa_sync() performance. Tuning these values may introduce subtle performance
113 * pathologies and should only be done in the context of performance analysis.
114 * These tunables will eventually be removed and replaced with #defines once
115 * enough analysis has been done to determine optimal values.
117 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
118 * regular blocks are not deferred.
120 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
121 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_deferred_free, CTLFLAG_RDTUN,
122 &zfs_sync_pass_deferred_free, 0, "defer frees starting in this pass");
123 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
124 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_dont_compress, CTLFLAG_RDTUN,
125 &zfs_sync_pass_dont_compress, 0, "don't compress starting in this pass");
126 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
127 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_rewrite, CTLFLAG_RDTUN,
128 &zfs_sync_pass_rewrite, 0, "rewrite new bps starting in this pass");
131 * An allocating zio is one that either currently has the DVA allocate
132 * stage set or will have it later in its lifetime.
134 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
136 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
140 int zio_buf_debug_limit = 16384;
142 int zio_buf_debug_limit = 0;
146 static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t);
152 zio_cache = kmem_cache_create("zio_cache",
153 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
154 zio_link_cache = kmem_cache_create("zio_link_cache",
155 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
160 * For small buffers, we want a cache for each multiple of
161 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
162 * for each quarter-power of 2.
164 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
165 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
168 int cflags = zio_exclude_metadata ? KMC_NODEBUG : 0;
176 * If we are using watchpoints, put each buffer on its own page,
177 * to eliminate the performance overhead of trapping to the
178 * kernel when modifying a non-watched buffer that shares the
179 * page with a watched buffer.
181 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
185 if (size <= 4 * SPA_MINBLOCKSIZE) {
186 align = SPA_MINBLOCKSIZE;
187 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
188 align = MIN(p2 >> 2, PAGESIZE);
193 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
194 zio_buf_cache[c] = kmem_cache_create(name, size,
195 align, NULL, NULL, NULL, NULL, NULL, cflags);
198 * Since zio_data bufs do not appear in crash dumps, we
199 * pass KMC_NOTOUCH so that no allocator metadata is
200 * stored with the buffers.
202 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
203 zio_data_buf_cache[c] = kmem_cache_create(name, size,
204 align, NULL, NULL, NULL, NULL, NULL,
205 cflags | KMC_NOTOUCH | KMC_NODEBUG);
210 ASSERT(zio_buf_cache[c] != NULL);
211 if (zio_buf_cache[c - 1] == NULL)
212 zio_buf_cache[c - 1] = zio_buf_cache[c];
214 ASSERT(zio_data_buf_cache[c] != NULL);
215 if (zio_data_buf_cache[c - 1] == NULL)
216 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
222 zio_trim_ksp = kstat_create("zfs", 0, "zio_trim", "misc",
224 sizeof(zio_trim_stats) / sizeof(kstat_named_t),
227 if (zio_trim_ksp != NULL) {
228 zio_trim_ksp->ks_data = &zio_trim_stats;
229 kstat_install(zio_trim_ksp);
237 kmem_cache_t *last_cache = NULL;
238 kmem_cache_t *last_data_cache = NULL;
240 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
241 if (zio_buf_cache[c] != last_cache) {
242 last_cache = zio_buf_cache[c];
243 kmem_cache_destroy(zio_buf_cache[c]);
245 zio_buf_cache[c] = NULL;
247 if (zio_data_buf_cache[c] != last_data_cache) {
248 last_data_cache = zio_data_buf_cache[c];
249 kmem_cache_destroy(zio_data_buf_cache[c]);
251 zio_data_buf_cache[c] = NULL;
254 kmem_cache_destroy(zio_link_cache);
255 kmem_cache_destroy(zio_cache);
259 if (zio_trim_ksp != NULL) {
260 kstat_delete(zio_trim_ksp);
266 * ==========================================================================
267 * Allocate and free I/O buffers
268 * ==========================================================================
272 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
273 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
274 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
275 * excess / transient data in-core during a crashdump.
278 zio_buf_alloc(size_t size)
280 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
281 int flags = zio_exclude_metadata ? KM_NODEBUG : 0;
283 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
286 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
288 return (kmem_alloc(size, KM_SLEEP|flags));
292 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
293 * crashdump if the kernel panics. This exists so that we will limit the amount
294 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
295 * of kernel heap dumped to disk when the kernel panics)
298 zio_data_buf_alloc(size_t size)
300 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
302 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
305 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
307 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG));
311 zio_buf_free(void *buf, size_t size)
313 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
315 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
318 kmem_cache_free(zio_buf_cache[c], buf);
320 kmem_free(buf, size);
324 zio_data_buf_free(void *buf, size_t size)
326 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
328 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
331 kmem_cache_free(zio_data_buf_cache[c], buf);
333 kmem_free(buf, size);
337 * ==========================================================================
338 * Push and pop I/O transform buffers
339 * ==========================================================================
342 zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize,
343 zio_transform_func_t *transform)
345 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
348 * Ensure that anyone expecting this zio to contain a linear ABD isn't
349 * going to get a nasty surprise when they try to access the data.
352 IMPLY(abd_is_linear(zio->io_abd), abd_is_linear(data));
354 IMPLY(zio->io_abd != NULL && abd_is_linear(zio->io_abd),
355 abd_is_linear(data));
358 zt->zt_orig_abd = zio->io_abd;
359 zt->zt_orig_size = zio->io_size;
360 zt->zt_bufsize = bufsize;
361 zt->zt_transform = transform;
363 zt->zt_next = zio->io_transform_stack;
364 zio->io_transform_stack = zt;
371 zio_pop_transforms(zio_t *zio)
375 while ((zt = zio->io_transform_stack) != NULL) {
376 if (zt->zt_transform != NULL)
377 zt->zt_transform(zio,
378 zt->zt_orig_abd, zt->zt_orig_size);
380 if (zt->zt_bufsize != 0)
381 abd_free(zio->io_abd);
383 zio->io_abd = zt->zt_orig_abd;
384 zio->io_size = zt->zt_orig_size;
385 zio->io_transform_stack = zt->zt_next;
387 kmem_free(zt, sizeof (zio_transform_t));
392 * ==========================================================================
393 * I/O transform callbacks for subblocks and decompression
394 * ==========================================================================
397 zio_subblock(zio_t *zio, abd_t *data, uint64_t size)
399 ASSERT(zio->io_size > size);
401 if (zio->io_type == ZIO_TYPE_READ)
402 abd_copy(data, zio->io_abd, size);
406 zio_decompress(zio_t *zio, abd_t *data, uint64_t size)
408 if (zio->io_error == 0) {
409 void *tmp = abd_borrow_buf(data, size);
410 int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
411 zio->io_abd, tmp, zio->io_size, size);
412 abd_return_buf_copy(data, tmp, size);
415 zio->io_error = SET_ERROR(EIO);
420 * ==========================================================================
421 * I/O parent/child relationships and pipeline interlocks
422 * ==========================================================================
425 zio_walk_parents(zio_t *cio, zio_link_t **zl)
427 list_t *pl = &cio->io_parent_list;
429 *zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl);
433 ASSERT((*zl)->zl_child == cio);
434 return ((*zl)->zl_parent);
438 zio_walk_children(zio_t *pio, zio_link_t **zl)
440 list_t *cl = &pio->io_child_list;
442 ASSERT(MUTEX_HELD(&pio->io_lock));
444 *zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl);
448 ASSERT((*zl)->zl_parent == pio);
449 return ((*zl)->zl_child);
453 zio_unique_parent(zio_t *cio)
455 zio_link_t *zl = NULL;
456 zio_t *pio = zio_walk_parents(cio, &zl);
458 VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL);
463 zio_add_child(zio_t *pio, zio_t *cio)
465 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
468 * Logical I/Os can have logical, gang, or vdev children.
469 * Gang I/Os can have gang or vdev children.
470 * Vdev I/Os can only have vdev children.
471 * The following ASSERT captures all of these constraints.
473 ASSERT3S(cio->io_child_type, <=, pio->io_child_type);
478 mutex_enter(&pio->io_lock);
479 mutex_enter(&cio->io_lock);
481 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
483 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
484 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
486 list_insert_head(&pio->io_child_list, zl);
487 list_insert_head(&cio->io_parent_list, zl);
489 pio->io_child_count++;
490 cio->io_parent_count++;
492 mutex_exit(&cio->io_lock);
493 mutex_exit(&pio->io_lock);
497 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
499 ASSERT(zl->zl_parent == pio);
500 ASSERT(zl->zl_child == cio);
502 mutex_enter(&pio->io_lock);
503 mutex_enter(&cio->io_lock);
505 list_remove(&pio->io_child_list, zl);
506 list_remove(&cio->io_parent_list, zl);
508 pio->io_child_count--;
509 cio->io_parent_count--;
511 mutex_exit(&cio->io_lock);
512 mutex_exit(&pio->io_lock);
513 kmem_cache_free(zio_link_cache, zl);
517 zio_wait_for_children(zio_t *zio, uint8_t childbits, enum zio_wait_type wait)
519 boolean_t waiting = B_FALSE;
521 mutex_enter(&zio->io_lock);
522 ASSERT(zio->io_stall == NULL);
523 for (int c = 0; c < ZIO_CHILD_TYPES; c++) {
524 if (!(ZIO_CHILD_BIT_IS_SET(childbits, c)))
527 uint64_t *countp = &zio->io_children[c][wait];
530 ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN);
531 zio->io_stall = countp;
536 mutex_exit(&zio->io_lock);
541 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait,
542 zio_t **next_to_executep)
544 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
545 int *errorp = &pio->io_child_error[zio->io_child_type];
547 mutex_enter(&pio->io_lock);
548 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
549 *errorp = zio_worst_error(*errorp, zio->io_error);
550 pio->io_reexecute |= zio->io_reexecute;
551 ASSERT3U(*countp, >, 0);
555 if (*countp == 0 && pio->io_stall == countp) {
556 zio_taskq_type_t type =
557 pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE :
559 pio->io_stall = NULL;
560 mutex_exit(&pio->io_lock);
563 * If we can tell the caller to execute this parent next, do
564 * so. Otherwise dispatch the parent zio as its own task.
566 * Having the caller execute the parent when possible reduces
567 * locking on the zio taskq's, reduces context switch
568 * overhead, and has no recursion penalty. Note that one
569 * read from disk typically causes at least 3 zio's: a
570 * zio_null(), the logical zio_read(), and then a physical
571 * zio. When the physical ZIO completes, we are able to call
572 * zio_done() on all 3 of these zio's from one invocation of
573 * zio_execute() by returning the parent back to
574 * zio_execute(). Since the parent isn't executed until this
575 * thread returns back to zio_execute(), the caller should do
578 * In other cases, dispatching the parent prevents
579 * overflowing the stack when we have deeply nested
580 * parent-child relationships, as we do with the "mega zio"
581 * of writes for spa_sync(), and the chain of ZIL blocks.
583 if (next_to_executep != NULL && *next_to_executep == NULL) {
584 *next_to_executep = pio;
586 zio_taskq_dispatch(pio, type, B_FALSE);
589 mutex_exit(&pio->io_lock);
594 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
596 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
597 zio->io_error = zio->io_child_error[c];
601 zio_bookmark_compare(const void *x1, const void *x2)
603 const zio_t *z1 = x1;
604 const zio_t *z2 = x2;
606 if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset)
608 if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset)
611 if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object)
613 if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object)
616 if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level)
618 if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level)
621 if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid)
623 if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid)
635 * ==========================================================================
636 * Create the various types of I/O (read, write, free, etc)
637 * ==========================================================================
640 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
641 abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done,
642 void *private, zio_type_t type, zio_priority_t priority,
643 enum zio_flag flags, vdev_t *vd, uint64_t offset,
644 const zbookmark_phys_t *zb, enum zio_stage stage, enum zio_stage pipeline)
648 IMPLY(type != ZIO_TYPE_FREE, psize <= SPA_MAXBLOCKSIZE);
649 ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0);
650 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
652 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
653 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
654 ASSERT(vd || stage == ZIO_STAGE_OPEN);
656 IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW) != 0);
658 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
659 bzero(zio, sizeof (zio_t));
661 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
662 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
663 #if defined(__FreeBSD__) && defined(_KERNEL)
664 callout_init(&zio->io_timer, 1);
667 list_create(&zio->io_parent_list, sizeof (zio_link_t),
668 offsetof(zio_link_t, zl_parent_node));
669 list_create(&zio->io_child_list, sizeof (zio_link_t),
670 offsetof(zio_link_t, zl_child_node));
671 metaslab_trace_init(&zio->io_alloc_list);
674 zio->io_child_type = ZIO_CHILD_VDEV;
675 else if (flags & ZIO_FLAG_GANG_CHILD)
676 zio->io_child_type = ZIO_CHILD_GANG;
677 else if (flags & ZIO_FLAG_DDT_CHILD)
678 zio->io_child_type = ZIO_CHILD_DDT;
680 zio->io_child_type = ZIO_CHILD_LOGICAL;
683 zio->io_bp = (blkptr_t *)bp;
684 zio->io_bp_copy = *bp;
685 zio->io_bp_orig = *bp;
686 if (type != ZIO_TYPE_WRITE ||
687 zio->io_child_type == ZIO_CHILD_DDT)
688 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
689 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
690 zio->io_logical = zio;
691 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
692 pipeline |= ZIO_GANG_STAGES;
698 zio->io_private = private;
700 zio->io_priority = priority;
702 zio->io_offset = offset;
703 zio->io_orig_abd = zio->io_abd = data;
704 zio->io_orig_size = zio->io_size = psize;
705 zio->io_lsize = lsize;
706 zio->io_orig_flags = zio->io_flags = flags;
707 zio->io_orig_stage = zio->io_stage = stage;
708 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
709 zio->io_pipeline_trace = ZIO_STAGE_OPEN;
711 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
712 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
715 zio->io_bookmark = *zb;
718 if (zio->io_metaslab_class == NULL)
719 zio->io_metaslab_class = pio->io_metaslab_class;
720 if (zio->io_logical == NULL)
721 zio->io_logical = pio->io_logical;
722 if (zio->io_child_type == ZIO_CHILD_GANG)
723 zio->io_gang_leader = pio->io_gang_leader;
724 zio_add_child(pio, zio);
731 zio_destroy(zio_t *zio)
734 KASSERT(!(callout_active(&zio->io_timer) ||
735 callout_pending(&zio->io_timer)), ("zio_destroy: timer active"));
737 metaslab_trace_fini(&zio->io_alloc_list);
738 list_destroy(&zio->io_parent_list);
739 list_destroy(&zio->io_child_list);
740 mutex_destroy(&zio->io_lock);
741 cv_destroy(&zio->io_cv);
742 kmem_cache_free(zio_cache, zio);
746 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
747 void *private, enum zio_flag flags)
751 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
752 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
753 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
759 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
761 return (zio_null(NULL, spa, NULL, done, private, flags));
765 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
767 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
768 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
769 bp, (longlong_t)BP_GET_TYPE(bp));
771 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
772 BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
773 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
774 bp, (longlong_t)BP_GET_CHECKSUM(bp));
776 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
777 BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
778 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
779 bp, (longlong_t)BP_GET_COMPRESS(bp));
781 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
782 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
783 bp, (longlong_t)BP_GET_LSIZE(bp));
785 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
786 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
787 bp, (longlong_t)BP_GET_PSIZE(bp));
790 if (BP_IS_EMBEDDED(bp)) {
791 if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
792 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
793 bp, (longlong_t)BPE_GET_ETYPE(bp));
798 * Do not verify individual DVAs if the config is not trusted. This
799 * will be done once the zio is executed in vdev_mirror_map_alloc.
801 if (!spa->spa_trust_config)
805 * Pool-specific checks.
807 * Note: it would be nice to verify that the blk_birth and
808 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
809 * allows the birth time of log blocks (and dmu_sync()-ed blocks
810 * that are in the log) to be arbitrarily large.
812 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
813 uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
814 if (vdevid >= spa->spa_root_vdev->vdev_children) {
815 zfs_panic_recover("blkptr at %p DVA %u has invalid "
817 bp, i, (longlong_t)vdevid);
820 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
822 zfs_panic_recover("blkptr at %p DVA %u has invalid "
824 bp, i, (longlong_t)vdevid);
827 if (vd->vdev_ops == &vdev_hole_ops) {
828 zfs_panic_recover("blkptr at %p DVA %u has hole "
830 bp, i, (longlong_t)vdevid);
833 if (vd->vdev_ops == &vdev_missing_ops) {
835 * "missing" vdevs are valid during import, but we
836 * don't have their detailed info (e.g. asize), so
837 * we can't perform any more checks on them.
841 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
842 uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
844 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
845 if (offset + asize > vd->vdev_asize) {
846 zfs_panic_recover("blkptr at %p DVA %u has invalid "
848 bp, i, (longlong_t)offset);
854 zfs_dva_valid(spa_t *spa, const dva_t *dva, const blkptr_t *bp)
856 uint64_t vdevid = DVA_GET_VDEV(dva);
858 if (vdevid >= spa->spa_root_vdev->vdev_children)
861 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
865 if (vd->vdev_ops == &vdev_hole_ops)
868 if (vd->vdev_ops == &vdev_missing_ops) {
872 uint64_t offset = DVA_GET_OFFSET(dva);
873 uint64_t asize = DVA_GET_ASIZE(dva);
876 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
877 if (offset + asize > vd->vdev_asize)
884 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
885 abd_t *data, uint64_t size, zio_done_func_t *done, void *private,
886 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
890 zfs_blkptr_verify(spa, bp);
892 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
893 data, size, size, done, private,
894 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
895 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
896 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
902 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
903 abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp,
904 zio_done_func_t *ready, zio_done_func_t *children_ready,
905 zio_done_func_t *physdone, zio_done_func_t *done,
906 void *private, zio_priority_t priority, enum zio_flag flags,
907 const zbookmark_phys_t *zb)
911 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
912 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
913 zp->zp_compress >= ZIO_COMPRESS_OFF &&
914 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
915 DMU_OT_IS_VALID(zp->zp_type) &&
918 zp->zp_copies <= spa_max_replication(spa));
920 zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private,
921 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
922 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
923 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
925 zio->io_ready = ready;
926 zio->io_children_ready = children_ready;
927 zio->io_physdone = physdone;
931 * Data can be NULL if we are going to call zio_write_override() to
932 * provide the already-allocated BP. But we may need the data to
933 * verify a dedup hit (if requested). In this case, don't try to
934 * dedup (just take the already-allocated BP verbatim).
936 if (data == NULL && zio->io_prop.zp_dedup_verify) {
937 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
944 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data,
945 uint64_t size, zio_done_func_t *done, void *private,
946 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
950 zio = zio_create(pio, spa, txg, bp, data, size, size, done, private,
951 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb,
952 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
958 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
960 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
961 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
962 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
963 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
966 * We must reset the io_prop to match the values that existed
967 * when the bp was first written by dmu_sync() keeping in mind
968 * that nopwrite and dedup are mutually exclusive.
970 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
971 zio->io_prop.zp_nopwrite = nopwrite;
972 zio->io_prop.zp_copies = copies;
973 zio->io_bp_override = bp;
977 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
980 zfs_blkptr_verify(spa, bp);
983 * The check for EMBEDDED is a performance optimization. We
984 * process the free here (by ignoring it) rather than
985 * putting it on the list and then processing it in zio_free_sync().
987 if (BP_IS_EMBEDDED(bp))
989 metaslab_check_free(spa, bp);
992 * Frees that are for the currently-syncing txg, are not going to be
993 * deferred, and which will not need to do a read (i.e. not GANG or
994 * DEDUP), can be processed immediately. Otherwise, put them on the
995 * in-memory list for later processing.
997 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
998 txg != spa->spa_syncing_txg ||
999 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
1000 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
1002 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp,
1003 BP_GET_PSIZE(bp), 0)));
1008 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1009 uint64_t size, enum zio_flag flags)
1012 enum zio_stage stage = ZIO_FREE_PIPELINE;
1014 ASSERT(!BP_IS_HOLE(bp));
1015 ASSERT(spa_syncing_txg(spa) == txg);
1016 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
1018 if (BP_IS_EMBEDDED(bp))
1019 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
1021 metaslab_check_free(spa, bp);
1023 dsl_scan_freed(spa, bp);
1025 if (zfs_trim_enabled)
1026 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START |
1027 ZIO_STAGE_VDEV_IO_ASSESS;
1029 * GANG and DEDUP blocks can induce a read (for the gang block header,
1030 * or the DDT), so issue them asynchronously so that this thread is
1033 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
1034 stage |= ZIO_STAGE_ISSUE_ASYNC;
1036 flags |= ZIO_FLAG_DONT_QUEUE;
1038 zio = zio_create(pio, spa, txg, bp, NULL, size,
1039 size, NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW,
1040 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
1046 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1047 zio_done_func_t *done, void *private, enum zio_flag flags)
1051 zfs_blkptr_verify(spa, bp);
1053 if (BP_IS_EMBEDDED(bp))
1054 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
1057 * A claim is an allocation of a specific block. Claims are needed
1058 * to support immediate writes in the intent log. The issue is that
1059 * immediate writes contain committed data, but in a txg that was
1060 * *not* committed. Upon opening the pool after an unclean shutdown,
1061 * the intent log claims all blocks that contain immediate write data
1062 * so that the SPA knows they're in use.
1064 * All claims *must* be resolved in the first txg -- before the SPA
1065 * starts allocating blocks -- so that nothing is allocated twice.
1066 * If txg == 0 we just verify that the block is claimable.
1068 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <,
1069 spa_min_claim_txg(spa));
1070 ASSERT(txg == spa_min_claim_txg(spa) || txg == 0);
1071 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
1073 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1074 BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW,
1075 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
1076 ASSERT0(zio->io_queued_timestamp);
1082 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset,
1083 uint64_t size, zio_done_func_t *done, void *private,
1084 zio_priority_t priority, enum zio_flag flags)
1089 if (vd->vdev_children == 0) {
1090 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
1091 ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
1092 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
1096 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
1098 for (c = 0; c < vd->vdev_children; c++)
1099 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
1100 offset, size, done, private, priority, flags));
1107 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1108 abd_t *data, int checksum, zio_done_func_t *done, void *private,
1109 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1113 ASSERT(vd->vdev_children == 0);
1114 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1115 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1116 ASSERT3U(offset + size, <=, vd->vdev_psize);
1118 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1119 private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1120 offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
1122 zio->io_prop.zp_checksum = checksum;
1128 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1129 abd_t *data, int checksum, zio_done_func_t *done, void *private,
1130 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1134 ASSERT(vd->vdev_children == 0);
1135 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1136 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1137 ASSERT3U(offset + size, <=, vd->vdev_psize);
1139 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1140 private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1141 offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
1143 zio->io_prop.zp_checksum = checksum;
1145 if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
1147 * zec checksums are necessarily destructive -- they modify
1148 * the end of the write buffer to hold the verifier/checksum.
1149 * Therefore, we must make a local copy in case the data is
1150 * being written to multiple places in parallel.
1152 abd_t *wbuf = abd_alloc_sametype(data, size);
1153 abd_copy(wbuf, data, size);
1155 zio_push_transform(zio, wbuf, size, size, NULL);
1162 * Create a child I/O to do some work for us.
1165 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1166 abd_t *data, uint64_t size, int type, zio_priority_t priority,
1167 enum zio_flag flags, zio_done_func_t *done, void *private)
1169 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1173 * vdev child I/Os do not propagate their error to the parent.
1174 * Therefore, for correct operation the caller *must* check for
1175 * and handle the error in the child i/o's done callback.
1176 * The only exceptions are i/os that we don't care about
1177 * (OPTIONAL or REPAIR).
1179 ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) ||
1182 if (type == ZIO_TYPE_READ && bp != NULL) {
1184 * If we have the bp, then the child should perform the
1185 * checksum and the parent need not. This pushes error
1186 * detection as close to the leaves as possible and
1187 * eliminates redundant checksums in the interior nodes.
1189 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1190 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1193 /* Not all IO types require vdev io done stage e.g. free */
1194 if (type == ZIO_TYPE_FREE &&
1195 !(pio->io_pipeline & ZIO_STAGE_VDEV_IO_DONE))
1196 pipeline &= ~ZIO_STAGE_VDEV_IO_DONE;
1198 if (vd->vdev_ops->vdev_op_leaf) {
1199 ASSERT0(vd->vdev_children);
1200 offset += VDEV_LABEL_START_SIZE;
1203 flags |= ZIO_VDEV_CHILD_FLAGS(pio);
1206 * If we've decided to do a repair, the write is not speculative --
1207 * even if the original read was.
1209 if (flags & ZIO_FLAG_IO_REPAIR)
1210 flags &= ~ZIO_FLAG_SPECULATIVE;
1213 * If we're creating a child I/O that is not associated with a
1214 * top-level vdev, then the child zio is not an allocating I/O.
1215 * If this is a retried I/O then we ignore it since we will
1216 * have already processed the original allocating I/O.
1218 if (flags & ZIO_FLAG_IO_ALLOCATING &&
1219 (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) {
1220 ASSERT(pio->io_metaslab_class != NULL);
1221 ASSERT(pio->io_metaslab_class->mc_alloc_throttle_enabled);
1222 ASSERT(type == ZIO_TYPE_WRITE);
1223 ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE);
1224 ASSERT(!(flags & ZIO_FLAG_IO_REPAIR));
1225 ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) ||
1226 pio->io_child_type == ZIO_CHILD_GANG);
1228 flags &= ~ZIO_FLAG_IO_ALLOCATING;
1231 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size,
1232 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1233 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1234 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
1236 zio->io_physdone = pio->io_physdone;
1237 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1238 zio->io_logical->io_phys_children++;
1244 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size,
1245 zio_type_t type, zio_priority_t priority, enum zio_flag flags,
1246 zio_done_func_t *done, void *private)
1250 ASSERT(vd->vdev_ops->vdev_op_leaf);
1252 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1253 data, size, size, done, private, type, priority,
1254 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1256 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1262 zio_flush(zio_t *zio, vdev_t *vd)
1264 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0,
1265 NULL, NULL, ZIO_PRIORITY_NOW,
1266 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1270 zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size)
1273 ASSERT(vd->vdev_ops->vdev_op_leaf);
1275 return (zio_create(zio, spa, 0, NULL, NULL, size, size, NULL, NULL,
1276 ZIO_TYPE_FREE, ZIO_PRIORITY_TRIM, ZIO_FLAG_DONT_AGGREGATE |
1277 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY,
1278 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PHYS_PIPELINE));
1282 zio_shrink(zio_t *zio, uint64_t size)
1284 ASSERT3P(zio->io_executor, ==, NULL);
1285 ASSERT3P(zio->io_orig_size, ==, zio->io_size);
1286 ASSERT3U(size, <=, zio->io_size);
1289 * We don't shrink for raidz because of problems with the
1290 * reconstruction when reading back less than the block size.
1291 * Note, BP_IS_RAIDZ() assumes no compression.
1293 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1294 if (!BP_IS_RAIDZ(zio->io_bp)) {
1295 /* we are not doing a raw write */
1296 ASSERT3U(zio->io_size, ==, zio->io_lsize);
1297 zio->io_orig_size = zio->io_size = zio->io_lsize = size;
1302 * ==========================================================================
1303 * Prepare to read and write logical blocks
1304 * ==========================================================================
1308 zio_read_bp_init(zio_t *zio)
1310 blkptr_t *bp = zio->io_bp;
1312 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1314 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1315 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1316 !(zio->io_flags & ZIO_FLAG_RAW)) {
1318 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1319 zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1320 psize, psize, zio_decompress);
1323 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1324 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1326 int psize = BPE_GET_PSIZE(bp);
1327 void *data = abd_borrow_buf(zio->io_abd, psize);
1328 decode_embedded_bp_compressed(bp, data);
1329 abd_return_buf_copy(zio->io_abd, data, psize);
1331 ASSERT(!BP_IS_EMBEDDED(bp));
1332 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1335 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1336 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1338 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1339 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1341 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1342 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1348 zio_write_bp_init(zio_t *zio)
1350 if (!IO_IS_ALLOCATING(zio))
1353 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1355 if (zio->io_bp_override) {
1356 blkptr_t *bp = zio->io_bp;
1357 zio_prop_t *zp = &zio->io_prop;
1359 ASSERT(bp->blk_birth != zio->io_txg);
1360 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1362 *bp = *zio->io_bp_override;
1363 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1365 if (BP_IS_EMBEDDED(bp))
1369 * If we've been overridden and nopwrite is set then
1370 * set the flag accordingly to indicate that a nopwrite
1371 * has already occurred.
1373 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1374 ASSERT(!zp->zp_dedup);
1375 ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum);
1376 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1380 ASSERT(!zp->zp_nopwrite);
1382 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1385 ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags &
1386 ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify);
1388 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1389 BP_SET_DEDUP(bp, 1);
1390 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1395 * We were unable to handle this as an override bp, treat
1396 * it as a regular write I/O.
1398 zio->io_bp_override = NULL;
1399 *bp = zio->io_bp_orig;
1400 zio->io_pipeline = zio->io_orig_pipeline;
1407 zio_write_compress(zio_t *zio)
1409 spa_t *spa = zio->io_spa;
1410 zio_prop_t *zp = &zio->io_prop;
1411 enum zio_compress compress = zp->zp_compress;
1412 blkptr_t *bp = zio->io_bp;
1413 uint64_t lsize = zio->io_lsize;
1414 uint64_t psize = zio->io_size;
1417 EQUIV(lsize != psize, (zio->io_flags & ZIO_FLAG_RAW) != 0);
1420 * If our children haven't all reached the ready stage,
1421 * wait for them and then repeat this pipeline stage.
1423 if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT |
1424 ZIO_CHILD_GANG_BIT, ZIO_WAIT_READY)) {
1428 if (!IO_IS_ALLOCATING(zio))
1431 if (zio->io_children_ready != NULL) {
1433 * Now that all our children are ready, run the callback
1434 * associated with this zio in case it wants to modify the
1435 * data to be written.
1437 ASSERT3U(zp->zp_level, >, 0);
1438 zio->io_children_ready(zio);
1441 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1442 ASSERT(zio->io_bp_override == NULL);
1444 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1446 * We're rewriting an existing block, which means we're
1447 * working on behalf of spa_sync(). For spa_sync() to
1448 * converge, it must eventually be the case that we don't
1449 * have to allocate new blocks. But compression changes
1450 * the blocksize, which forces a reallocate, and makes
1451 * convergence take longer. Therefore, after the first
1452 * few passes, stop compressing to ensure convergence.
1454 pass = spa_sync_pass(spa);
1456 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1457 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1458 ASSERT(!BP_GET_DEDUP(bp));
1460 if (pass >= zfs_sync_pass_dont_compress)
1461 compress = ZIO_COMPRESS_OFF;
1463 /* Make sure someone doesn't change their mind on overwrites */
1464 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1465 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1468 /* If it's a compressed write that is not raw, compress the buffer. */
1469 if (compress != ZIO_COMPRESS_OFF && psize == lsize) {
1470 void *cbuf = zio_buf_alloc(lsize);
1471 psize = zio_compress_data(compress, zio->io_abd, cbuf, lsize);
1472 if (psize == 0 || psize == lsize) {
1473 compress = ZIO_COMPRESS_OFF;
1474 zio_buf_free(cbuf, lsize);
1475 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1476 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1477 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1478 encode_embedded_bp_compressed(bp,
1479 cbuf, compress, lsize, psize);
1480 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1481 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1482 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1483 zio_buf_free(cbuf, lsize);
1484 bp->blk_birth = zio->io_txg;
1485 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1486 ASSERT(spa_feature_is_active(spa,
1487 SPA_FEATURE_EMBEDDED_DATA));
1491 * Round up compressed size up to the ashift
1492 * of the smallest-ashift device, and zero the tail.
1493 * This ensures that the compressed size of the BP
1494 * (and thus compressratio property) are correct,
1495 * in that we charge for the padding used to fill out
1498 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
1499 size_t rounded = (size_t)P2ROUNDUP(psize,
1500 1ULL << spa->spa_min_ashift);
1501 if (rounded >= lsize) {
1502 compress = ZIO_COMPRESS_OFF;
1503 zio_buf_free(cbuf, lsize);
1506 abd_t *cdata = abd_get_from_buf(cbuf, lsize);
1507 abd_take_ownership_of_buf(cdata, B_TRUE);
1508 abd_zero_off(cdata, psize, rounded - psize);
1510 zio_push_transform(zio, cdata,
1511 psize, lsize, NULL);
1516 * We were unable to handle this as an override bp, treat
1517 * it as a regular write I/O.
1519 zio->io_bp_override = NULL;
1520 *bp = zio->io_bp_orig;
1521 zio->io_pipeline = zio->io_orig_pipeline;
1523 ASSERT3U(psize, !=, 0);
1527 * The final pass of spa_sync() must be all rewrites, but the first
1528 * few passes offer a trade-off: allocating blocks defers convergence,
1529 * but newly allocated blocks are sequential, so they can be written
1530 * to disk faster. Therefore, we allow the first few passes of
1531 * spa_sync() to allocate new blocks, but force rewrites after that.
1532 * There should only be a handful of blocks after pass 1 in any case.
1534 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1535 BP_GET_PSIZE(bp) == psize &&
1536 pass >= zfs_sync_pass_rewrite) {
1537 VERIFY3U(psize, !=, 0);
1538 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1540 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1541 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1544 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1548 if (zio->io_bp_orig.blk_birth != 0 &&
1549 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1550 BP_SET_LSIZE(bp, lsize);
1551 BP_SET_TYPE(bp, zp->zp_type);
1552 BP_SET_LEVEL(bp, zp->zp_level);
1553 BP_SET_BIRTH(bp, zio->io_txg, 0);
1555 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1557 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1558 BP_SET_LSIZE(bp, lsize);
1559 BP_SET_TYPE(bp, zp->zp_type);
1560 BP_SET_LEVEL(bp, zp->zp_level);
1561 BP_SET_PSIZE(bp, psize);
1562 BP_SET_COMPRESS(bp, compress);
1563 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1564 BP_SET_DEDUP(bp, zp->zp_dedup);
1565 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1567 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1568 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1569 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1571 if (zp->zp_nopwrite) {
1572 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1573 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1574 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1581 zio_free_bp_init(zio_t *zio)
1583 blkptr_t *bp = zio->io_bp;
1585 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1586 if (BP_GET_DEDUP(bp))
1587 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1590 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1596 * ==========================================================================
1597 * Execute the I/O pipeline
1598 * ==========================================================================
1602 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1604 spa_t *spa = zio->io_spa;
1605 zio_type_t t = zio->io_type;
1606 int flags = (cutinline ? TQ_FRONT : 0);
1608 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1611 * If we're a config writer or a probe, the normal issue and
1612 * interrupt threads may all be blocked waiting for the config lock.
1613 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1615 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1619 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1621 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1625 * If this is a high priority I/O, then use the high priority taskq if
1628 if ((zio->io_priority == ZIO_PRIORITY_NOW ||
1629 zio->io_priority == ZIO_PRIORITY_SYNC_WRITE) &&
1630 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1633 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1636 * NB: We are assuming that the zio can only be dispatched
1637 * to a single taskq at a time. It would be a grievous error
1638 * to dispatch the zio to another taskq at the same time.
1640 #if defined(illumos) || !defined(_KERNEL)
1641 ASSERT(zio->io_tqent.tqent_next == NULL);
1643 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
1645 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1646 flags, &zio->io_tqent);
1650 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1652 kthread_t *executor = zio->io_executor;
1653 spa_t *spa = zio->io_spa;
1655 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1656 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1658 for (i = 0; i < tqs->stqs_count; i++) {
1659 if (taskq_member(tqs->stqs_taskq[i], executor))
1668 zio_issue_async(zio_t *zio)
1670 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1676 zio_interrupt(zio_t *zio)
1678 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1682 zio_delay_interrupt(zio_t *zio)
1685 * The timeout_generic() function isn't defined in userspace, so
1686 * rather than trying to implement the function, the zio delay
1687 * functionality has been disabled for userspace builds.
1692 * If io_target_timestamp is zero, then no delay has been registered
1693 * for this IO, thus jump to the end of this function and "skip" the
1694 * delay; issuing it directly to the zio layer.
1696 if (zio->io_target_timestamp != 0) {
1697 hrtime_t now = gethrtime();
1699 if (now >= zio->io_target_timestamp) {
1701 * This IO has already taken longer than the target
1702 * delay to complete, so we don't want to delay it
1703 * any longer; we "miss" the delay and issue it
1704 * directly to the zio layer. This is likely due to
1705 * the target latency being set to a value less than
1706 * the underlying hardware can satisfy (e.g. delay
1707 * set to 1ms, but the disks take 10ms to complete an
1711 DTRACE_PROBE2(zio__delay__miss, zio_t *, zio,
1716 hrtime_t diff = zio->io_target_timestamp - now;
1718 DTRACE_PROBE3(zio__delay__hit, zio_t *, zio,
1719 hrtime_t, now, hrtime_t, diff);
1722 callout_reset_sbt(&zio->io_timer, nstosbt(diff), 0,
1723 (void (*)(void *))zio_interrupt, zio, C_HARDCLOCK);
1725 (void) timeout_generic(CALLOUT_NORMAL,
1726 (void (*)(void *))zio_interrupt, zio, diff, 1, 0);
1734 DTRACE_PROBE1(zio__delay__skip, zio_t *, zio);
1739 * Execute the I/O pipeline until one of the following occurs:
1741 * (1) the I/O completes
1742 * (2) the pipeline stalls waiting for dependent child I/Os
1743 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1744 * (4) the I/O is delegated by vdev-level caching or aggregation
1745 * (5) the I/O is deferred due to vdev-level queueing
1746 * (6) the I/O is handed off to another thread.
1748 * In all cases, the pipeline stops whenever there's no CPU work; it never
1749 * burns a thread in cv_wait().
1751 * There's no locking on io_stage because there's no legitimate way
1752 * for multiple threads to be attempting to process the same I/O.
1754 static zio_pipe_stage_t *zio_pipeline[];
1757 zio_execute(zio_t *zio)
1759 ASSERT3U(zio->io_queued_timestamp, >, 0);
1761 while (zio->io_stage < ZIO_STAGE_DONE) {
1762 enum zio_stage pipeline = zio->io_pipeline;
1763 enum zio_stage stage = zio->io_stage;
1765 zio->io_executor = curthread;
1767 ASSERT(!MUTEX_HELD(&zio->io_lock));
1768 ASSERT(ISP2(stage));
1769 ASSERT(zio->io_stall == NULL);
1773 } while ((stage & pipeline) == 0);
1775 ASSERT(stage <= ZIO_STAGE_DONE);
1778 * If we are in interrupt context and this pipeline stage
1779 * will grab a config lock that is held across I/O,
1780 * or may wait for an I/O that needs an interrupt thread
1781 * to complete, issue async to avoid deadlock.
1783 * For VDEV_IO_START, we cut in line so that the io will
1784 * be sent to disk promptly.
1786 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1787 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1788 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1789 zio_requeue_io_start_cut_in_line : B_FALSE;
1790 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1794 zio->io_stage = stage;
1795 zio->io_pipeline_trace |= zio->io_stage;
1798 * The zio pipeline stage returns the next zio to execute
1799 * (typically the same as this one), or NULL if we should
1802 zio = zio_pipeline[highbit64(stage) - 1](zio);
1810 * ==========================================================================
1811 * Initiate I/O, either sync or async
1812 * ==========================================================================
1815 zio_wait(zio_t *zio)
1819 ASSERT3P(zio->io_stage, ==, ZIO_STAGE_OPEN);
1820 ASSERT3P(zio->io_executor, ==, NULL);
1822 zio->io_waiter = curthread;
1823 ASSERT0(zio->io_queued_timestamp);
1824 zio->io_queued_timestamp = gethrtime();
1828 mutex_enter(&zio->io_lock);
1829 while (zio->io_executor != NULL)
1830 cv_wait(&zio->io_cv, &zio->io_lock);
1831 mutex_exit(&zio->io_lock);
1833 error = zio->io_error;
1840 zio_nowait(zio_t *zio)
1842 ASSERT3P(zio->io_executor, ==, NULL);
1844 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1845 zio_unique_parent(zio) == NULL) {
1847 * This is a logical async I/O with no parent to wait for it.
1848 * We add it to the spa_async_root_zio "Godfather" I/O which
1849 * will ensure they complete prior to unloading the pool.
1851 spa_t *spa = zio->io_spa;
1853 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1856 ASSERT0(zio->io_queued_timestamp);
1857 zio->io_queued_timestamp = gethrtime();
1862 * ==========================================================================
1863 * Reexecute, cancel, or suspend/resume failed I/O
1864 * ==========================================================================
1868 zio_reexecute(zio_t *pio)
1870 zio_t *cio, *cio_next;
1872 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1873 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1874 ASSERT(pio->io_gang_leader == NULL);
1875 ASSERT(pio->io_gang_tree == NULL);
1877 pio->io_flags = pio->io_orig_flags;
1878 pio->io_stage = pio->io_orig_stage;
1879 pio->io_pipeline = pio->io_orig_pipeline;
1880 pio->io_reexecute = 0;
1881 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1882 pio->io_pipeline_trace = 0;
1884 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1885 pio->io_state[w] = 0;
1886 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1887 pio->io_child_error[c] = 0;
1889 if (IO_IS_ALLOCATING(pio))
1890 BP_ZERO(pio->io_bp);
1893 * As we reexecute pio's children, new children could be created.
1894 * New children go to the head of pio's io_child_list, however,
1895 * so we will (correctly) not reexecute them. The key is that
1896 * the remainder of pio's io_child_list, from 'cio_next' onward,
1897 * cannot be affected by any side effects of reexecuting 'cio'.
1899 zio_link_t *zl = NULL;
1900 mutex_enter(&pio->io_lock);
1901 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
1902 cio_next = zio_walk_children(pio, &zl);
1903 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1904 pio->io_children[cio->io_child_type][w]++;
1905 mutex_exit(&pio->io_lock);
1907 mutex_enter(&pio->io_lock);
1909 mutex_exit(&pio->io_lock);
1912 * Now that all children have been reexecuted, execute the parent.
1913 * We don't reexecute "The Godfather" I/O here as it's the
1914 * responsibility of the caller to wait on it.
1916 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) {
1917 pio->io_queued_timestamp = gethrtime();
1923 zio_suspend(spa_t *spa, zio_t *zio, zio_suspend_reason_t reason)
1925 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1926 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1927 "failure and the failure mode property for this pool "
1928 "is set to panic.", spa_name(spa));
1930 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1932 mutex_enter(&spa->spa_suspend_lock);
1934 if (spa->spa_suspend_zio_root == NULL)
1935 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1936 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1937 ZIO_FLAG_GODFATHER);
1939 spa->spa_suspended = reason;
1942 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1943 ASSERT(zio != spa->spa_suspend_zio_root);
1944 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1945 ASSERT(zio_unique_parent(zio) == NULL);
1946 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1947 zio_add_child(spa->spa_suspend_zio_root, zio);
1950 mutex_exit(&spa->spa_suspend_lock);
1954 zio_resume(spa_t *spa)
1959 * Reexecute all previously suspended i/o.
1961 mutex_enter(&spa->spa_suspend_lock);
1962 spa->spa_suspended = ZIO_SUSPEND_NONE;
1963 cv_broadcast(&spa->spa_suspend_cv);
1964 pio = spa->spa_suspend_zio_root;
1965 spa->spa_suspend_zio_root = NULL;
1966 mutex_exit(&spa->spa_suspend_lock);
1972 return (zio_wait(pio));
1976 zio_resume_wait(spa_t *spa)
1978 mutex_enter(&spa->spa_suspend_lock);
1979 while (spa_suspended(spa))
1980 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1981 mutex_exit(&spa->spa_suspend_lock);
1985 * ==========================================================================
1988 * A gang block is a collection of small blocks that looks to the DMU
1989 * like one large block. When zio_dva_allocate() cannot find a block
1990 * of the requested size, due to either severe fragmentation or the pool
1991 * being nearly full, it calls zio_write_gang_block() to construct the
1992 * block from smaller fragments.
1994 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1995 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1996 * an indirect block: it's an array of block pointers. It consumes
1997 * only one sector and hence is allocatable regardless of fragmentation.
1998 * The gang header's bps point to its gang members, which hold the data.
2000 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
2001 * as the verifier to ensure uniqueness of the SHA256 checksum.
2002 * Critically, the gang block bp's blk_cksum is the checksum of the data,
2003 * not the gang header. This ensures that data block signatures (needed for
2004 * deduplication) are independent of how the block is physically stored.
2006 * Gang blocks can be nested: a gang member may itself be a gang block.
2007 * Thus every gang block is a tree in which root and all interior nodes are
2008 * gang headers, and the leaves are normal blocks that contain user data.
2009 * The root of the gang tree is called the gang leader.
2011 * To perform any operation (read, rewrite, free, claim) on a gang block,
2012 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
2013 * in the io_gang_tree field of the original logical i/o by recursively
2014 * reading the gang leader and all gang headers below it. This yields
2015 * an in-core tree containing the contents of every gang header and the
2016 * bps for every constituent of the gang block.
2018 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
2019 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
2020 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
2021 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
2022 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
2023 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
2024 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
2025 * of the gang header plus zio_checksum_compute() of the data to update the
2026 * gang header's blk_cksum as described above.
2028 * The two-phase assemble/issue model solves the problem of partial failure --
2029 * what if you'd freed part of a gang block but then couldn't read the
2030 * gang header for another part? Assembling the entire gang tree first
2031 * ensures that all the necessary gang header I/O has succeeded before
2032 * starting the actual work of free, claim, or write. Once the gang tree
2033 * is assembled, free and claim are in-memory operations that cannot fail.
2035 * In the event that a gang write fails, zio_dva_unallocate() walks the
2036 * gang tree to immediately free (i.e. insert back into the space map)
2037 * everything we've allocated. This ensures that we don't get ENOSPC
2038 * errors during repeated suspend/resume cycles due to a flaky device.
2040 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
2041 * the gang tree, we won't modify the block, so we can safely defer the free
2042 * (knowing that the block is still intact). If we *can* assemble the gang
2043 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
2044 * each constituent bp and we can allocate a new block on the next sync pass.
2046 * In all cases, the gang tree allows complete recovery from partial failure.
2047 * ==========================================================================
2051 zio_gang_issue_func_done(zio_t *zio)
2053 abd_put(zio->io_abd);
2057 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2063 return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset),
2064 BP_GET_PSIZE(bp), zio_gang_issue_func_done,
2065 NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2066 &pio->io_bookmark));
2070 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2077 abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2078 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2079 gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL,
2080 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2083 * As we rewrite each gang header, the pipeline will compute
2084 * a new gang block header checksum for it; but no one will
2085 * compute a new data checksum, so we do that here. The one
2086 * exception is the gang leader: the pipeline already computed
2087 * its data checksum because that stage precedes gang assembly.
2088 * (Presently, nothing actually uses interior data checksums;
2089 * this is just good hygiene.)
2091 if (gn != pio->io_gang_leader->io_gang_tree) {
2092 abd_t *buf = abd_get_offset(data, offset);
2094 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
2095 buf, BP_GET_PSIZE(bp));
2100 * If we are here to damage data for testing purposes,
2101 * leave the GBH alone so that we can detect the damage.
2103 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
2104 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2106 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2107 abd_get_offset(data, offset), BP_GET_PSIZE(bp),
2108 zio_gang_issue_func_done, NULL, pio->io_priority,
2109 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2117 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2120 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
2121 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp),
2122 ZIO_GANG_CHILD_FLAGS(pio)));
2127 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2130 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
2131 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
2134 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
2143 static void zio_gang_tree_assemble_done(zio_t *zio);
2145 static zio_gang_node_t *
2146 zio_gang_node_alloc(zio_gang_node_t **gnpp)
2148 zio_gang_node_t *gn;
2150 ASSERT(*gnpp == NULL);
2152 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
2153 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
2160 zio_gang_node_free(zio_gang_node_t **gnpp)
2162 zio_gang_node_t *gn = *gnpp;
2164 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2165 ASSERT(gn->gn_child[g] == NULL);
2167 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2168 kmem_free(gn, sizeof (*gn));
2173 zio_gang_tree_free(zio_gang_node_t **gnpp)
2175 zio_gang_node_t *gn = *gnpp;
2180 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2181 zio_gang_tree_free(&gn->gn_child[g]);
2183 zio_gang_node_free(gnpp);
2187 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
2189 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
2190 abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2192 ASSERT(gio->io_gang_leader == gio);
2193 ASSERT(BP_IS_GANG(bp));
2195 zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2196 zio_gang_tree_assemble_done, gn, gio->io_priority,
2197 ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
2201 zio_gang_tree_assemble_done(zio_t *zio)
2203 zio_t *gio = zio->io_gang_leader;
2204 zio_gang_node_t *gn = zio->io_private;
2205 blkptr_t *bp = zio->io_bp;
2207 ASSERT(gio == zio_unique_parent(zio));
2208 ASSERT(zio->io_child_count == 0);
2213 /* this ABD was created from a linear buf in zio_gang_tree_assemble */
2214 if (BP_SHOULD_BYTESWAP(bp))
2215 byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size);
2217 ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh);
2218 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
2219 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2221 abd_put(zio->io_abd);
2223 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2224 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2225 if (!BP_IS_GANG(gbp))
2227 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
2232 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data,
2235 zio_t *gio = pio->io_gang_leader;
2238 ASSERT(BP_IS_GANG(bp) == !!gn);
2239 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
2240 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
2243 * If you're a gang header, your data is in gn->gn_gbh.
2244 * If you're a gang member, your data is in 'data' and gn == NULL.
2246 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset);
2249 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2251 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2252 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2253 if (BP_IS_HOLE(gbp))
2255 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data,
2257 offset += BP_GET_PSIZE(gbp);
2261 if (gn == gio->io_gang_tree && gio->io_abd != NULL)
2262 ASSERT3U(gio->io_size, ==, offset);
2269 zio_gang_assemble(zio_t *zio)
2271 blkptr_t *bp = zio->io_bp;
2273 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
2274 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2276 zio->io_gang_leader = zio;
2278 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
2284 zio_gang_issue(zio_t *zio)
2286 blkptr_t *bp = zio->io_bp;
2288 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT, ZIO_WAIT_DONE)) {
2292 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
2293 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2295 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
2296 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd,
2299 zio_gang_tree_free(&zio->io_gang_tree);
2301 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2307 zio_write_gang_member_ready(zio_t *zio)
2309 zio_t *pio = zio_unique_parent(zio);
2310 zio_t *gio = zio->io_gang_leader;
2311 dva_t *cdva = zio->io_bp->blk_dva;
2312 dva_t *pdva = pio->io_bp->blk_dva;
2315 if (BP_IS_HOLE(zio->io_bp))
2318 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
2320 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
2321 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
2322 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2323 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
2324 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
2326 mutex_enter(&pio->io_lock);
2327 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
2328 ASSERT(DVA_GET_GANG(&pdva[d]));
2329 asize = DVA_GET_ASIZE(&pdva[d]);
2330 asize += DVA_GET_ASIZE(&cdva[d]);
2331 DVA_SET_ASIZE(&pdva[d], asize);
2333 mutex_exit(&pio->io_lock);
2337 zio_write_gang_done(zio_t *zio)
2340 * The io_abd field will be NULL for a zio with no data. The io_flags
2341 * will initially have the ZIO_FLAG_NODATA bit flag set, but we can't
2342 * check for it here as it is cleared in zio_ready.
2344 if (zio->io_abd != NULL)
2345 abd_put(zio->io_abd);
2349 zio_write_gang_block(zio_t *pio)
2351 spa_t *spa = pio->io_spa;
2352 metaslab_class_t *mc = spa_normal_class(spa);
2353 blkptr_t *bp = pio->io_bp;
2354 zio_t *gio = pio->io_gang_leader;
2356 zio_gang_node_t *gn, **gnpp;
2357 zio_gbh_phys_t *gbh;
2359 uint64_t txg = pio->io_txg;
2360 uint64_t resid = pio->io_size;
2362 int copies = gio->io_prop.zp_copies;
2363 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
2366 boolean_t has_data = !(pio->io_flags & ZIO_FLAG_NODATA);
2368 int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER;
2369 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2370 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2373 flags |= METASLAB_ASYNC_ALLOC;
2374 VERIFY(zfs_refcount_held(&mc->mc_alloc_slots[pio->io_allocator],
2378 * The logical zio has already placed a reservation for
2379 * 'copies' allocation slots but gang blocks may require
2380 * additional copies. These additional copies
2381 * (i.e. gbh_copies - copies) are guaranteed to succeed
2382 * since metaslab_class_throttle_reserve() always allows
2383 * additional reservations for gang blocks.
2385 VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies,
2386 pio->io_allocator, pio, flags));
2389 error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE,
2390 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags,
2391 &pio->io_alloc_list, pio, pio->io_allocator);
2393 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2394 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2398 * If we failed to allocate the gang block header then
2399 * we remove any additional allocation reservations that
2400 * we placed here. The original reservation will
2401 * be removed when the logical I/O goes to the ready
2404 metaslab_class_throttle_unreserve(mc,
2405 gbh_copies - copies, pio->io_allocator, pio);
2407 pio->io_error = error;
2412 gnpp = &gio->io_gang_tree;
2414 gnpp = pio->io_private;
2415 ASSERT(pio->io_ready == zio_write_gang_member_ready);
2418 gn = zio_gang_node_alloc(gnpp);
2420 bzero(gbh, SPA_GANGBLOCKSIZE);
2421 gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE);
2424 * Create the gang header.
2426 zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2427 zio_write_gang_done, NULL, pio->io_priority,
2428 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2431 * Create and nowait the gang children.
2433 for (int g = 0; resid != 0; resid -= lsize, g++) {
2434 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2436 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2438 zp.zp_checksum = gio->io_prop.zp_checksum;
2439 zp.zp_compress = ZIO_COMPRESS_OFF;
2440 zp.zp_type = DMU_OT_NONE;
2442 zp.zp_copies = gio->io_prop.zp_copies;
2443 zp.zp_dedup = B_FALSE;
2444 zp.zp_dedup_verify = B_FALSE;
2445 zp.zp_nopwrite = B_FALSE;
2447 zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2448 has_data ? abd_get_offset(pio->io_abd, pio->io_size -
2449 resid) : NULL, lsize, lsize, &zp,
2450 zio_write_gang_member_ready, NULL, NULL,
2451 zio_write_gang_done, &gn->gn_child[g], pio->io_priority,
2452 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2454 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2455 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2459 * Gang children won't throttle but we should
2460 * account for their work, so reserve an allocation
2461 * slot for them here.
2463 VERIFY(metaslab_class_throttle_reserve(mc,
2464 zp.zp_copies, cio->io_allocator, cio, flags));
2470 * Set pio's pipeline to just wait for zio to finish.
2472 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2480 * The zio_nop_write stage in the pipeline determines if allocating a
2481 * new bp is necessary. The nopwrite feature can handle writes in
2482 * either syncing or open context (i.e. zil writes) and as a result is
2483 * mutually exclusive with dedup.
2485 * By leveraging a cryptographically secure checksum, such as SHA256, we
2486 * can compare the checksums of the new data and the old to determine if
2487 * allocating a new block is required. Note that our requirements for
2488 * cryptographic strength are fairly weak: there can't be any accidental
2489 * hash collisions, but we don't need to be secure against intentional
2490 * (malicious) collisions. To trigger a nopwrite, you have to be able
2491 * to write the file to begin with, and triggering an incorrect (hash
2492 * collision) nopwrite is no worse than simply writing to the file.
2493 * That said, there are no known attacks against the checksum algorithms
2494 * used for nopwrite, assuming that the salt and the checksums
2495 * themselves remain secret.
2498 zio_nop_write(zio_t *zio)
2500 blkptr_t *bp = zio->io_bp;
2501 blkptr_t *bp_orig = &zio->io_bp_orig;
2502 zio_prop_t *zp = &zio->io_prop;
2504 ASSERT(BP_GET_LEVEL(bp) == 0);
2505 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2506 ASSERT(zp->zp_nopwrite);
2507 ASSERT(!zp->zp_dedup);
2508 ASSERT(zio->io_bp_override == NULL);
2509 ASSERT(IO_IS_ALLOCATING(zio));
2512 * Check to see if the original bp and the new bp have matching
2513 * characteristics (i.e. same checksum, compression algorithms, etc).
2514 * If they don't then just continue with the pipeline which will
2515 * allocate a new bp.
2517 if (BP_IS_HOLE(bp_orig) ||
2518 !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
2519 ZCHECKSUM_FLAG_NOPWRITE) ||
2520 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2521 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2522 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2523 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2527 * If the checksums match then reset the pipeline so that we
2528 * avoid allocating a new bp and issuing any I/O.
2530 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2531 ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
2532 ZCHECKSUM_FLAG_NOPWRITE);
2533 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2534 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2535 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2536 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2537 sizeof (uint64_t)) == 0);
2540 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2541 zio->io_flags |= ZIO_FLAG_NOPWRITE;
2548 * ==========================================================================
2550 * ==========================================================================
2553 zio_ddt_child_read_done(zio_t *zio)
2555 blkptr_t *bp = zio->io_bp;
2556 ddt_entry_t *dde = zio->io_private;
2558 zio_t *pio = zio_unique_parent(zio);
2560 mutex_enter(&pio->io_lock);
2561 ddp = ddt_phys_select(dde, bp);
2562 if (zio->io_error == 0)
2563 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2565 if (zio->io_error == 0 && dde->dde_repair_abd == NULL)
2566 dde->dde_repair_abd = zio->io_abd;
2568 abd_free(zio->io_abd);
2569 mutex_exit(&pio->io_lock);
2573 zio_ddt_read_start(zio_t *zio)
2575 blkptr_t *bp = zio->io_bp;
2577 ASSERT(BP_GET_DEDUP(bp));
2578 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2579 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2581 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2582 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2583 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2584 ddt_phys_t *ddp = dde->dde_phys;
2585 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2588 ASSERT(zio->io_vsd == NULL);
2591 if (ddp_self == NULL)
2594 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2595 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2597 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2599 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2600 abd_alloc_for_io(zio->io_size, B_TRUE),
2601 zio->io_size, zio_ddt_child_read_done, dde,
2602 zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) |
2603 ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark));
2608 zio_nowait(zio_read(zio, zio->io_spa, bp,
2609 zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority,
2610 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2616 zio_ddt_read_done(zio_t *zio)
2618 blkptr_t *bp = zio->io_bp;
2620 if (zio_wait_for_children(zio, ZIO_CHILD_DDT_BIT, ZIO_WAIT_DONE)) {
2624 ASSERT(BP_GET_DEDUP(bp));
2625 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2626 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2628 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2629 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2630 ddt_entry_t *dde = zio->io_vsd;
2632 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2636 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2637 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2640 if (dde->dde_repair_abd != NULL) {
2641 abd_copy(zio->io_abd, dde->dde_repair_abd,
2643 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2645 ddt_repair_done(ddt, dde);
2649 ASSERT(zio->io_vsd == NULL);
2655 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2657 spa_t *spa = zio->io_spa;
2658 boolean_t do_raw = (zio->io_flags & ZIO_FLAG_RAW);
2660 /* We should never get a raw, override zio */
2661 ASSERT(!(zio->io_bp_override && do_raw));
2664 * Note: we compare the original data, not the transformed data,
2665 * because when zio->io_bp is an override bp, we will not have
2666 * pushed the I/O transforms. That's an important optimization
2667 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2669 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2670 zio_t *lio = dde->dde_lead_zio[p];
2673 return (lio->io_orig_size != zio->io_orig_size ||
2674 abd_cmp(zio->io_orig_abd, lio->io_orig_abd,
2675 zio->io_orig_size) != 0);
2679 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2680 ddt_phys_t *ddp = &dde->dde_phys[p];
2682 if (ddp->ddp_phys_birth != 0) {
2683 arc_buf_t *abuf = NULL;
2684 arc_flags_t aflags = ARC_FLAG_WAIT;
2685 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE;
2686 blkptr_t blk = *zio->io_bp;
2689 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2694 * Intuitively, it would make more sense to compare
2695 * io_abd than io_orig_abd in the raw case since you
2696 * don't want to look at any transformations that have
2697 * happened to the data. However, for raw I/Os the
2698 * data will actually be the same in io_abd and
2699 * io_orig_abd, so all we have to do is issue this as
2703 zio_flags |= ZIO_FLAG_RAW;
2704 ASSERT3U(zio->io_size, ==, zio->io_orig_size);
2705 ASSERT0(abd_cmp(zio->io_abd, zio->io_orig_abd,
2707 ASSERT3P(zio->io_transform_stack, ==, NULL);
2710 error = arc_read(NULL, spa, &blk,
2711 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2712 zio_flags, &aflags, &zio->io_bookmark);
2715 if (arc_buf_size(abuf) != zio->io_orig_size ||
2716 abd_cmp_buf(zio->io_orig_abd, abuf->b_data,
2717 zio->io_orig_size) != 0)
2718 error = SET_ERROR(EEXIST);
2719 arc_buf_destroy(abuf, &abuf);
2723 return (error != 0);
2731 zio_ddt_child_write_ready(zio_t *zio)
2733 int p = zio->io_prop.zp_copies;
2734 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2735 ddt_entry_t *dde = zio->io_private;
2736 ddt_phys_t *ddp = &dde->dde_phys[p];
2744 ASSERT(dde->dde_lead_zio[p] == zio);
2746 ddt_phys_fill(ddp, zio->io_bp);
2748 zio_link_t *zl = NULL;
2749 while ((pio = zio_walk_parents(zio, &zl)) != NULL)
2750 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2756 zio_ddt_child_write_done(zio_t *zio)
2758 int p = zio->io_prop.zp_copies;
2759 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2760 ddt_entry_t *dde = zio->io_private;
2761 ddt_phys_t *ddp = &dde->dde_phys[p];
2765 ASSERT(ddp->ddp_refcnt == 0);
2766 ASSERT(dde->dde_lead_zio[p] == zio);
2767 dde->dde_lead_zio[p] = NULL;
2769 if (zio->io_error == 0) {
2770 zio_link_t *zl = NULL;
2771 while (zio_walk_parents(zio, &zl) != NULL)
2772 ddt_phys_addref(ddp);
2774 ddt_phys_clear(ddp);
2781 zio_ddt_ditto_write_done(zio_t *zio)
2783 int p = DDT_PHYS_DITTO;
2784 zio_prop_t *zp = &zio->io_prop;
2785 blkptr_t *bp = zio->io_bp;
2786 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2787 ddt_entry_t *dde = zio->io_private;
2788 ddt_phys_t *ddp = &dde->dde_phys[p];
2789 ddt_key_t *ddk = &dde->dde_key;
2793 ASSERT(ddp->ddp_refcnt == 0);
2794 ASSERT(dde->dde_lead_zio[p] == zio);
2795 dde->dde_lead_zio[p] = NULL;
2797 if (zio->io_error == 0) {
2798 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2799 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2800 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2801 if (ddp->ddp_phys_birth != 0)
2802 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2803 ddt_phys_fill(ddp, bp);
2810 zio_ddt_write(zio_t *zio)
2812 spa_t *spa = zio->io_spa;
2813 blkptr_t *bp = zio->io_bp;
2814 uint64_t txg = zio->io_txg;
2815 zio_prop_t *zp = &zio->io_prop;
2816 int p = zp->zp_copies;
2820 ddt_t *ddt = ddt_select(spa, bp);
2824 ASSERT(BP_GET_DEDUP(bp));
2825 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2826 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2827 ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW)));
2830 dde = ddt_lookup(ddt, bp, B_TRUE);
2831 ddp = &dde->dde_phys[p];
2833 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2835 * If we're using a weak checksum, upgrade to a strong checksum
2836 * and try again. If we're already using a strong checksum,
2837 * we can't resolve it, so just convert to an ordinary write.
2838 * (And automatically e-mail a paper to Nature?)
2840 if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
2841 ZCHECKSUM_FLAG_DEDUP)) {
2842 zp->zp_checksum = spa_dedup_checksum(spa);
2843 zio_pop_transforms(zio);
2844 zio->io_stage = ZIO_STAGE_OPEN;
2847 zp->zp_dedup = B_FALSE;
2848 BP_SET_DEDUP(bp, B_FALSE);
2850 ASSERT(!BP_GET_DEDUP(bp));
2851 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2856 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2857 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2859 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2860 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2861 zio_prop_t czp = *zp;
2863 czp.zp_copies = ditto_copies;
2866 * If we arrived here with an override bp, we won't have run
2867 * the transform stack, so we won't have the data we need to
2868 * generate a child i/o. So, toss the override bp and restart.
2869 * This is safe, because using the override bp is just an
2870 * optimization; and it's rare, so the cost doesn't matter.
2872 if (zio->io_bp_override) {
2873 zio_pop_transforms(zio);
2874 zio->io_stage = ZIO_STAGE_OPEN;
2875 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2876 zio->io_bp_override = NULL;
2882 dio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
2883 zio->io_orig_size, zio->io_orig_size, &czp, NULL, NULL,
2884 NULL, zio_ddt_ditto_write_done, dde, zio->io_priority,
2885 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2887 zio_push_transform(dio, zio->io_abd, zio->io_size, 0, NULL);
2888 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2891 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2892 if (ddp->ddp_phys_birth != 0)
2893 ddt_bp_fill(ddp, bp, txg);
2894 if (dde->dde_lead_zio[p] != NULL)
2895 zio_add_child(zio, dde->dde_lead_zio[p]);
2897 ddt_phys_addref(ddp);
2898 } else if (zio->io_bp_override) {
2899 ASSERT(bp->blk_birth == txg);
2900 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2901 ddt_phys_fill(ddp, bp);
2902 ddt_phys_addref(ddp);
2904 cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
2905 zio->io_orig_size, zio->io_orig_size, zp,
2906 zio_ddt_child_write_ready, NULL, NULL,
2907 zio_ddt_child_write_done, dde, zio->io_priority,
2908 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2910 zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL);
2911 dde->dde_lead_zio[p] = cio;
2924 ddt_entry_t *freedde; /* for debugging */
2927 zio_ddt_free(zio_t *zio)
2929 spa_t *spa = zio->io_spa;
2930 blkptr_t *bp = zio->io_bp;
2931 ddt_t *ddt = ddt_select(spa, bp);
2935 ASSERT(BP_GET_DEDUP(bp));
2936 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2939 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2941 ddp = ddt_phys_select(dde, bp);
2943 ddt_phys_decref(ddp);
2951 * ==========================================================================
2952 * Allocate and free blocks
2953 * ==========================================================================
2957 zio_io_to_allocate(spa_t *spa, int allocator)
2961 ASSERT(MUTEX_HELD(&spa->spa_alloc_locks[allocator]));
2963 zio = avl_first(&spa->spa_alloc_trees[allocator]);
2967 ASSERT(IO_IS_ALLOCATING(zio));
2970 * Try to place a reservation for this zio. If we're unable to
2971 * reserve then we throttle.
2973 ASSERT3U(zio->io_allocator, ==, allocator);
2974 if (!metaslab_class_throttle_reserve(zio->io_metaslab_class,
2975 zio->io_prop.zp_copies, zio->io_allocator, zio, 0)) {
2979 avl_remove(&spa->spa_alloc_trees[allocator], zio);
2980 ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE);
2986 zio_dva_throttle(zio_t *zio)
2988 spa_t *spa = zio->io_spa;
2990 metaslab_class_t *mc;
2992 /* locate an appropriate allocation class */
2993 mc = spa_preferred_class(spa, zio->io_size, zio->io_prop.zp_type,
2994 zio->io_prop.zp_level, zio->io_prop.zp_zpl_smallblk);
2996 if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE ||
2997 !mc->mc_alloc_throttle_enabled ||
2998 zio->io_child_type == ZIO_CHILD_GANG ||
2999 zio->io_flags & ZIO_FLAG_NODATA) {
3003 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3005 ASSERT3U(zio->io_queued_timestamp, >, 0);
3006 ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE);
3008 zbookmark_phys_t *bm = &zio->io_bookmark;
3010 * We want to try to use as many allocators as possible to help improve
3011 * performance, but we also want logically adjacent IOs to be physically
3012 * adjacent to improve sequential read performance. We chunk each object
3013 * into 2^20 block regions, and then hash based on the objset, object,
3014 * level, and region to accomplish both of these goals.
3016 zio->io_allocator = cityhash4(bm->zb_objset, bm->zb_object,
3017 bm->zb_level, bm->zb_blkid >> 20) % spa->spa_alloc_count;
3018 mutex_enter(&spa->spa_alloc_locks[zio->io_allocator]);
3019 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3020 zio->io_metaslab_class = mc;
3021 avl_add(&spa->spa_alloc_trees[zio->io_allocator], zio);
3022 nio = zio_io_to_allocate(spa, zio->io_allocator);
3023 mutex_exit(&spa->spa_alloc_locks[zio->io_allocator]);
3029 zio_allocate_dispatch(spa_t *spa, int allocator)
3033 mutex_enter(&spa->spa_alloc_locks[allocator]);
3034 zio = zio_io_to_allocate(spa, allocator);
3035 mutex_exit(&spa->spa_alloc_locks[allocator]);
3039 ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE);
3040 ASSERT0(zio->io_error);
3041 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE);
3045 zio_dva_allocate(zio_t *zio)
3047 spa_t *spa = zio->io_spa;
3048 metaslab_class_t *mc;
3049 blkptr_t *bp = zio->io_bp;
3053 if (zio->io_gang_leader == NULL) {
3054 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3055 zio->io_gang_leader = zio;
3058 ASSERT(BP_IS_HOLE(bp));
3059 ASSERT0(BP_GET_NDVAS(bp));
3060 ASSERT3U(zio->io_prop.zp_copies, >, 0);
3061 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
3062 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
3064 if (zio->io_flags & ZIO_FLAG_NODATA)
3065 flags |= METASLAB_DONT_THROTTLE;
3066 if (zio->io_flags & ZIO_FLAG_GANG_CHILD)
3067 flags |= METASLAB_GANG_CHILD;
3068 if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE)
3069 flags |= METASLAB_ASYNC_ALLOC;
3072 * if not already chosen, locate an appropriate allocation class
3074 mc = zio->io_metaslab_class;
3076 mc = spa_preferred_class(spa, zio->io_size,
3077 zio->io_prop.zp_type, zio->io_prop.zp_level,
3078 zio->io_prop.zp_zpl_smallblk);
3079 zio->io_metaslab_class = mc;
3082 error = metaslab_alloc(spa, mc, zio->io_size, bp,
3083 zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
3084 &zio->io_alloc_list, zio, zio->io_allocator);
3087 * Fallback to normal class when an alloc class is full
3089 if (error == ENOSPC && mc != spa_normal_class(spa)) {
3091 * If throttling, transfer reservation over to normal class.
3092 * The io_allocator slot can remain the same even though we
3093 * are switching classes.
3095 if (mc->mc_alloc_throttle_enabled &&
3096 (zio->io_flags & ZIO_FLAG_IO_ALLOCATING)) {
3097 metaslab_class_throttle_unreserve(mc,
3098 zio->io_prop.zp_copies, zio->io_allocator, zio);
3099 zio->io_flags &= ~ZIO_FLAG_IO_ALLOCATING;
3101 mc = spa_normal_class(spa);
3102 VERIFY(metaslab_class_throttle_reserve(mc,
3103 zio->io_prop.zp_copies, zio->io_allocator, zio,
3104 flags | METASLAB_MUST_RESERVE));
3106 mc = spa_normal_class(spa);
3108 zio->io_metaslab_class = mc;
3110 error = metaslab_alloc(spa, mc, zio->io_size, bp,
3111 zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
3112 &zio->io_alloc_list, zio, zio->io_allocator);
3116 zfs_dbgmsg("%s: metaslab allocation failure: zio %p, "
3117 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
3119 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
3120 return (zio_write_gang_block(zio));
3121 zio->io_error = error;
3128 zio_dva_free(zio_t *zio)
3130 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
3136 zio_dva_claim(zio_t *zio)
3140 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
3142 zio->io_error = error;
3148 * Undo an allocation. This is used by zio_done() when an I/O fails
3149 * and we want to give back the block we just allocated.
3150 * This handles both normal blocks and gang blocks.
3153 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
3155 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
3156 ASSERT(zio->io_bp_override == NULL);
3158 if (!BP_IS_HOLE(bp))
3159 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
3162 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
3163 zio_dva_unallocate(zio, gn->gn_child[g],
3164 &gn->gn_gbh->zg_blkptr[g]);
3170 * Try to allocate an intent log block. Return 0 on success, errno on failure.
3173 zio_alloc_zil(spa_t *spa, uint64_t objset, uint64_t txg, blkptr_t *new_bp,
3174 blkptr_t *old_bp, uint64_t size, boolean_t *slog)
3177 zio_alloc_list_t io_alloc_list;
3179 ASSERT(txg > spa_syncing_txg(spa));
3181 metaslab_trace_init(&io_alloc_list);
3184 * Block pointer fields are useful to metaslabs for stats and debugging.
3185 * Fill in the obvious ones before calling into metaslab_alloc().
3187 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3188 BP_SET_PSIZE(new_bp, size);
3189 BP_SET_LEVEL(new_bp, 0);
3192 * When allocating a zil block, we don't have information about
3193 * the final destination of the block except the objset it's part
3194 * of, so we just hash the objset ID to pick the allocator to get
3197 error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1,
3198 txg, old_bp, METASLAB_HINTBP_AVOID, &io_alloc_list, NULL,
3199 cityhash4(0, 0, 0, objset) % spa->spa_alloc_count);
3203 error = metaslab_alloc(spa, spa_normal_class(spa), size,
3204 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID,
3205 &io_alloc_list, NULL, cityhash4(0, 0, 0, objset) %
3206 spa->spa_alloc_count);
3210 metaslab_trace_fini(&io_alloc_list);
3213 BP_SET_LSIZE(new_bp, size);
3214 BP_SET_PSIZE(new_bp, size);
3215 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
3216 BP_SET_CHECKSUM(new_bp,
3217 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
3218 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
3219 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3220 BP_SET_LEVEL(new_bp, 0);
3221 BP_SET_DEDUP(new_bp, 0);
3222 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
3224 zfs_dbgmsg("%s: zil block allocation failure: "
3225 "size %llu, error %d", spa_name(spa), size, error);
3232 * ==========================================================================
3233 * Read, write and delete to physical devices
3234 * ==========================================================================
3239 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3240 * stops after this stage and will resume upon I/O completion.
3241 * However, there are instances where the vdev layer may need to
3242 * continue the pipeline when an I/O was not issued. Since the I/O
3243 * that was sent to the vdev layer might be different than the one
3244 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3245 * force the underlying vdev layers to call either zio_execute() or
3246 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3249 zio_vdev_io_start(zio_t *zio)
3251 vdev_t *vd = zio->io_vd;
3253 spa_t *spa = zio->io_spa;
3256 ASSERT(zio->io_error == 0);
3257 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
3260 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3261 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
3264 * The mirror_ops handle multiple DVAs in a single BP.
3266 vdev_mirror_ops.vdev_op_io_start(zio);
3270 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE &&
3271 zio->io_priority == ZIO_PRIORITY_NOW) {
3272 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg);
3276 ASSERT3P(zio->io_logical, !=, zio);
3277 if (zio->io_type == ZIO_TYPE_WRITE) {
3278 ASSERT(spa->spa_trust_config);
3280 if (zio->io_vd->vdev_removing) {
3282 * Note: the code can handle other kinds of writes,
3283 * but we don't expect them.
3285 ASSERT(zio->io_flags &
3286 (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL |
3287 ZIO_FLAG_RESILVER | ZIO_FLAG_INDUCE_DAMAGE));
3292 * We keep track of time-sensitive I/Os so that the scan thread
3293 * can quickly react to certain workloads. In particular, we care
3294 * about non-scrubbing, top-level reads and writes with the following
3296 * - synchronous writes of user data to non-slog devices
3297 * - any reads of user data
3298 * When these conditions are met, adjust the timestamp of spa_last_io
3299 * which allows the scan thread to adjust its workload accordingly.
3301 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
3302 vd == vd->vdev_top && !vd->vdev_islog &&
3303 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
3304 zio->io_txg != spa_syncing_txg(spa)) {
3305 uint64_t old = spa->spa_last_io;
3306 uint64_t new = ddi_get_lbolt64();
3308 (void) atomic_cas_64(&spa->spa_last_io, old, new);
3310 align = 1ULL << vd->vdev_top->vdev_ashift;
3312 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
3313 P2PHASE(zio->io_size, align) != 0) {
3314 /* Transform logical writes to be a full physical block size. */
3315 uint64_t asize = P2ROUNDUP(zio->io_size, align);
3317 if (zio->io_type == ZIO_TYPE_READ ||
3318 zio->io_type == ZIO_TYPE_WRITE)
3319 abuf = abd_alloc_sametype(zio->io_abd, asize);
3320 ASSERT(vd == vd->vdev_top);
3321 if (zio->io_type == ZIO_TYPE_WRITE) {
3322 abd_copy(abuf, zio->io_abd, zio->io_size);
3323 abd_zero_off(abuf, zio->io_size, asize - zio->io_size);
3325 zio_push_transform(zio, abuf, asize, abuf ? asize : 0,
3330 * If this is not a physical io, make sure that it is properly aligned
3331 * before proceeding.
3333 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
3334 ASSERT0(P2PHASE(zio->io_offset, align));
3335 ASSERT0(P2PHASE(zio->io_size, align));
3338 * For the physical io we allow alignment
3339 * to a logical block size.
3341 uint64_t log_align =
3342 1ULL << vd->vdev_top->vdev_logical_ashift;
3343 ASSERT0(P2PHASE(zio->io_offset, log_align));
3344 ASSERT0(P2PHASE(zio->io_size, log_align));
3347 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa));
3350 * If this is a repair I/O, and there's no self-healing involved --
3351 * that is, we're just resilvering what we expect to resilver --
3352 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3353 * This prevents spurious resilvering.
3355 * There are a few ways that we can end up creating these spurious
3358 * 1. A resilver i/o will be issued if any DVA in the BP has a
3359 * dirty DTL. The mirror code will issue resilver writes to
3360 * each DVA, including the one(s) that are not on vdevs with dirty
3363 * 2. With nested replication, which happens when we have a
3364 * "replacing" or "spare" vdev that's a child of a mirror or raidz.
3365 * For example, given mirror(replacing(A+B), C), it's likely that
3366 * only A is out of date (it's the new device). In this case, we'll
3367 * read from C, then use the data to resilver A+B -- but we don't
3368 * actually want to resilver B, just A. The top-level mirror has no
3369 * way to know this, so instead we just discard unnecessary repairs
3370 * as we work our way down the vdev tree.
3372 * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc.
3373 * The same logic applies to any form of nested replication: ditto
3374 * + mirror, RAID-Z + replacing, etc.
3376 * However, indirect vdevs point off to other vdevs which may have
3377 * DTL's, so we never bypass them. The child i/os on concrete vdevs
3378 * will be properly bypassed instead.
3380 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3381 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
3382 zio->io_txg != 0 && /* not a delegated i/o */
3383 vd->vdev_ops != &vdev_indirect_ops &&
3384 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
3385 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3386 zio_vdev_io_bypass(zio);
3390 if (vd->vdev_ops->vdev_op_leaf) {
3391 switch (zio->io_type) {
3393 if (vdev_cache_read(zio))
3396 case ZIO_TYPE_WRITE:
3398 if ((zio = vdev_queue_io(zio)) == NULL)
3401 if (!vdev_accessible(vd, zio)) {
3402 zio->io_error = SET_ERROR(ENXIO);
3409 * Note that we ignore repair writes for TRIM because they can
3410 * conflict with normal writes. This isn't an issue because, by
3411 * definition, we only repair blocks that aren't freed.
3413 if (zio->io_type == ZIO_TYPE_WRITE &&
3414 !(zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3415 !trim_map_write_start(zio))
3419 vd->vdev_ops->vdev_op_io_start(zio);
3424 zio_vdev_io_done(zio_t *zio)
3426 vdev_t *vd = zio->io_vd;
3427 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
3428 boolean_t unexpected_error = B_FALSE;
3430 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3434 ASSERT(zio->io_type == ZIO_TYPE_READ ||
3435 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE);
3437 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3438 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE ||
3439 zio->io_type == ZIO_TYPE_FREE)) {
3441 if (zio->io_type == ZIO_TYPE_WRITE &&
3442 !(zio->io_flags & ZIO_FLAG_IO_REPAIR))
3443 trim_map_write_done(zio);
3445 vdev_queue_io_done(zio);
3447 if (zio->io_type == ZIO_TYPE_WRITE)
3448 vdev_cache_write(zio);
3450 if (zio_injection_enabled && zio->io_error == 0)
3451 zio->io_error = zio_handle_device_injection(vd,
3454 if (zio_injection_enabled && zio->io_error == 0)
3455 zio->io_error = zio_handle_label_injection(zio, EIO);
3457 if (zio->io_error) {
3458 if (zio->io_error == ENOTSUP &&
3459 zio->io_type == ZIO_TYPE_FREE) {
3460 /* Not all devices support TRIM. */
3461 } else if (!vdev_accessible(vd, zio)) {
3462 zio->io_error = SET_ERROR(ENXIO);
3464 unexpected_error = B_TRUE;
3469 ops->vdev_op_io_done(zio);
3471 if (unexpected_error)
3472 VERIFY(vdev_probe(vd, zio) == NULL);
3478 * This function is used to change the priority of an existing zio that is
3479 * currently in-flight. This is used by the arc to upgrade priority in the
3480 * event that a demand read is made for a block that is currently queued
3481 * as a scrub or async read IO. Otherwise, the high priority read request
3482 * would end up having to wait for the lower priority IO.
3485 zio_change_priority(zio_t *pio, zio_priority_t priority)
3487 zio_t *cio, *cio_next;
3488 zio_link_t *zl = NULL;
3490 ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
3492 if (pio->io_vd != NULL && pio->io_vd->vdev_ops->vdev_op_leaf) {
3493 vdev_queue_change_io_priority(pio, priority);
3495 pio->io_priority = priority;
3498 mutex_enter(&pio->io_lock);
3499 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
3500 cio_next = zio_walk_children(pio, &zl);
3501 zio_change_priority(cio, priority);
3503 mutex_exit(&pio->io_lock);
3507 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3508 * disk, and use that to finish the checksum ereport later.
3511 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
3512 const void *good_buf)
3514 /* no processing needed */
3515 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
3520 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
3522 void *buf = zio_buf_alloc(zio->io_size);
3524 abd_copy_to_buf(buf, zio->io_abd, zio->io_size);
3526 zcr->zcr_cbinfo = zio->io_size;
3527 zcr->zcr_cbdata = buf;
3528 zcr->zcr_finish = zio_vsd_default_cksum_finish;
3529 zcr->zcr_free = zio_buf_free;
3533 zio_vdev_io_assess(zio_t *zio)
3535 vdev_t *vd = zio->io_vd;
3537 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3541 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3542 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
3544 if (zio->io_vsd != NULL) {
3545 zio->io_vsd_ops->vsd_free(zio);
3549 if (zio_injection_enabled && zio->io_error == 0)
3550 zio->io_error = zio_handle_fault_injection(zio, EIO);
3552 if (zio->io_type == ZIO_TYPE_FREE &&
3553 zio->io_priority != ZIO_PRIORITY_NOW) {
3554 switch (zio->io_error) {
3556 ZIO_TRIM_STAT_INCR(bytes, zio->io_size);
3557 ZIO_TRIM_STAT_BUMP(success);
3560 ZIO_TRIM_STAT_BUMP(unsupported);
3563 ZIO_TRIM_STAT_BUMP(failed);
3569 * If the I/O failed, determine whether we should attempt to retry it.
3571 * On retry, we cut in line in the issue queue, since we don't want
3572 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3574 if (zio->io_error && vd == NULL &&
3575 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
3576 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
3577 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
3579 zio->io_flags |= ZIO_FLAG_IO_RETRY |
3580 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
3581 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
3582 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
3583 zio_requeue_io_start_cut_in_line);
3588 * If we got an error on a leaf device, convert it to ENXIO
3589 * if the device is not accessible at all.
3591 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3592 !vdev_accessible(vd, zio))
3593 zio->io_error = SET_ERROR(ENXIO);
3596 * If we can't write to an interior vdev (mirror or RAID-Z),
3597 * set vdev_cant_write so that we stop trying to allocate from it.
3599 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
3600 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
3601 vd->vdev_cant_write = B_TRUE;
3605 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
3606 * attempts will ever succeed. In this case we set a persistent bit so
3607 * that we don't bother with it in the future.
3609 if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) &&
3610 zio->io_type == ZIO_TYPE_IOCTL &&
3611 zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL)
3612 vd->vdev_nowritecache = B_TRUE;
3615 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3617 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3618 zio->io_physdone != NULL) {
3619 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
3620 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
3621 zio->io_physdone(zio->io_logical);
3628 zio_vdev_io_reissue(zio_t *zio)
3630 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3631 ASSERT(zio->io_error == 0);
3633 zio->io_stage >>= 1;
3637 zio_vdev_io_redone(zio_t *zio)
3639 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
3641 zio->io_stage >>= 1;
3645 zio_vdev_io_bypass(zio_t *zio)
3647 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3648 ASSERT(zio->io_error == 0);
3650 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
3651 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
3655 * ==========================================================================
3656 * Generate and verify checksums
3657 * ==========================================================================
3660 zio_checksum_generate(zio_t *zio)
3662 blkptr_t *bp = zio->io_bp;
3663 enum zio_checksum checksum;
3667 * This is zio_write_phys().
3668 * We're either generating a label checksum, or none at all.
3670 checksum = zio->io_prop.zp_checksum;
3672 if (checksum == ZIO_CHECKSUM_OFF)
3675 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
3677 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
3678 ASSERT(!IO_IS_ALLOCATING(zio));
3679 checksum = ZIO_CHECKSUM_GANG_HEADER;
3681 checksum = BP_GET_CHECKSUM(bp);
3685 zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size);
3691 zio_checksum_verify(zio_t *zio)
3693 zio_bad_cksum_t info;
3694 blkptr_t *bp = zio->io_bp;
3697 ASSERT(zio->io_vd != NULL);
3701 * This is zio_read_phys().
3702 * We're either verifying a label checksum, or nothing at all.
3704 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
3707 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
3710 if ((error = zio_checksum_error(zio, &info)) != 0) {
3711 zio->io_error = error;
3712 if (error == ECKSUM &&
3713 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
3714 zfs_ereport_start_checksum(zio->io_spa,
3715 zio->io_vd, zio, zio->io_offset,
3716 zio->io_size, NULL, &info);
3724 * Called by RAID-Z to ensure we don't compute the checksum twice.
3727 zio_checksum_verified(zio_t *zio)
3729 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
3733 * ==========================================================================
3734 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
3735 * An error of 0 indicates success. ENXIO indicates whole-device failure,
3736 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
3737 * indicate errors that are specific to one I/O, and most likely permanent.
3738 * Any other error is presumed to be worse because we weren't expecting it.
3739 * ==========================================================================
3742 zio_worst_error(int e1, int e2)
3744 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
3747 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
3748 if (e1 == zio_error_rank[r1])
3751 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
3752 if (e2 == zio_error_rank[r2])
3755 return (r1 > r2 ? e1 : e2);
3759 * ==========================================================================
3761 * ==========================================================================
3764 zio_ready(zio_t *zio)
3766 blkptr_t *bp = zio->io_bp;
3767 zio_t *pio, *pio_next;
3768 zio_link_t *zl = NULL;
3770 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT | ZIO_CHILD_DDT_BIT,
3775 if (zio->io_ready) {
3776 ASSERT(IO_IS_ALLOCATING(zio));
3777 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
3778 (zio->io_flags & ZIO_FLAG_NOPWRITE));
3779 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
3784 if (bp != NULL && bp != &zio->io_bp_copy)
3785 zio->io_bp_copy = *bp;
3787 if (zio->io_error != 0) {
3788 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3790 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3791 ASSERT(IO_IS_ALLOCATING(zio));
3792 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3793 ASSERT(zio->io_metaslab_class != NULL);
3796 * We were unable to allocate anything, unreserve and
3797 * issue the next I/O to allocate.
3799 metaslab_class_throttle_unreserve(
3800 zio->io_metaslab_class, zio->io_prop.zp_copies,
3801 zio->io_allocator, zio);
3802 zio_allocate_dispatch(zio->io_spa, zio->io_allocator);
3806 mutex_enter(&zio->io_lock);
3807 zio->io_state[ZIO_WAIT_READY] = 1;
3808 pio = zio_walk_parents(zio, &zl);
3809 mutex_exit(&zio->io_lock);
3812 * As we notify zio's parents, new parents could be added.
3813 * New parents go to the head of zio's io_parent_list, however,
3814 * so we will (correctly) not notify them. The remainder of zio's
3815 * io_parent_list, from 'pio_next' onward, cannot change because
3816 * all parents must wait for us to be done before they can be done.
3818 for (; pio != NULL; pio = pio_next) {
3819 pio_next = zio_walk_parents(zio, &zl);
3820 zio_notify_parent(pio, zio, ZIO_WAIT_READY, NULL);
3823 if (zio->io_flags & ZIO_FLAG_NODATA) {
3824 if (BP_IS_GANG(bp)) {
3825 zio->io_flags &= ~ZIO_FLAG_NODATA;
3827 ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE);
3828 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
3832 if (zio_injection_enabled &&
3833 zio->io_spa->spa_syncing_txg == zio->io_txg)
3834 zio_handle_ignored_writes(zio);
3840 * Update the allocation throttle accounting.
3843 zio_dva_throttle_done(zio_t *zio)
3845 zio_t *lio = zio->io_logical;
3846 zio_t *pio = zio_unique_parent(zio);
3847 vdev_t *vd = zio->io_vd;
3848 int flags = METASLAB_ASYNC_ALLOC;
3850 ASSERT3P(zio->io_bp, !=, NULL);
3851 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
3852 ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE);
3853 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
3855 ASSERT3P(vd, ==, vd->vdev_top);
3856 ASSERT(!(zio->io_flags & (ZIO_FLAG_IO_REPAIR | ZIO_FLAG_IO_RETRY)));
3857 ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING);
3858 ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE));
3859 ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA));
3862 * Parents of gang children can have two flavors -- ones that
3863 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
3864 * and ones that allocated the constituent blocks. The allocation
3865 * throttle needs to know the allocating parent zio so we must find
3868 if (pio->io_child_type == ZIO_CHILD_GANG) {
3870 * If our parent is a rewrite gang child then our grandparent
3871 * would have been the one that performed the allocation.
3873 if (pio->io_flags & ZIO_FLAG_IO_REWRITE)
3874 pio = zio_unique_parent(pio);
3875 flags |= METASLAB_GANG_CHILD;
3878 ASSERT(IO_IS_ALLOCATING(pio));
3879 ASSERT3P(zio, !=, zio->io_logical);
3880 ASSERT(zio->io_logical != NULL);
3881 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
3882 ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE);
3883 ASSERT(zio->io_metaslab_class != NULL);
3885 mutex_enter(&pio->io_lock);
3886 metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags,
3887 pio->io_allocator, B_TRUE);
3888 mutex_exit(&pio->io_lock);
3890 metaslab_class_throttle_unreserve(zio->io_metaslab_class, 1,
3891 pio->io_allocator, pio);
3894 * Call into the pipeline to see if there is more work that
3895 * needs to be done. If there is work to be done it will be
3896 * dispatched to another taskq thread.
3898 zio_allocate_dispatch(zio->io_spa, pio->io_allocator);
3902 zio_done(zio_t *zio)
3904 spa_t *spa = zio->io_spa;
3905 zio_t *lio = zio->io_logical;
3906 blkptr_t *bp = zio->io_bp;
3907 vdev_t *vd = zio->io_vd;
3908 uint64_t psize = zio->io_size;
3909 zio_t *pio, *pio_next;
3910 zio_link_t *zl = NULL;
3913 * If our children haven't all completed,
3914 * wait for them and then repeat this pipeline stage.
3916 if (zio_wait_for_children(zio, ZIO_CHILD_ALL_BITS, ZIO_WAIT_DONE)) {
3921 * If the allocation throttle is enabled, then update the accounting.
3922 * We only track child I/Os that are part of an allocating async
3923 * write. We must do this since the allocation is performed
3924 * by the logical I/O but the actual write is done by child I/Os.
3926 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING &&
3927 zio->io_child_type == ZIO_CHILD_VDEV) {
3928 ASSERT(zio->io_metaslab_class != NULL);
3929 ASSERT(zio->io_metaslab_class->mc_alloc_throttle_enabled);
3930 zio_dva_throttle_done(zio);
3934 * If the allocation throttle is enabled, verify that
3935 * we have decremented the refcounts for every I/O that was throttled.
3937 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3938 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3939 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3942 metaslab_group_alloc_verify(spa, zio->io_bp, zio,
3944 VERIFY(zfs_refcount_not_held(
3945 &zio->io_metaslab_class->mc_alloc_slots[zio->io_allocator],
3949 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3950 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3951 ASSERT(zio->io_children[c][w] == 0);
3953 if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
3954 ASSERT(bp->blk_pad[0] == 0);
3955 ASSERT(bp->blk_pad[1] == 0);
3956 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3957 (bp == zio_unique_parent(zio)->io_bp));
3958 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3959 zio->io_bp_override == NULL &&
3960 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3961 ASSERT(!BP_SHOULD_BYTESWAP(bp));
3962 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3963 ASSERT(BP_COUNT_GANG(bp) == 0 ||
3964 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3966 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3967 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3971 * If there were child vdev/gang/ddt errors, they apply to us now.
3973 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3974 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3975 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3978 * If the I/O on the transformed data was successful, generate any
3979 * checksum reports now while we still have the transformed data.
3981 if (zio->io_error == 0) {
3982 while (zio->io_cksum_report != NULL) {
3983 zio_cksum_report_t *zcr = zio->io_cksum_report;
3984 uint64_t align = zcr->zcr_align;
3985 uint64_t asize = P2ROUNDUP(psize, align);
3987 abd_t *adata = zio->io_abd;
3989 if (asize != psize) {
3990 adata = abd_alloc_linear(asize, B_TRUE);
3991 abd_copy(adata, zio->io_abd, psize);
3992 abd_zero_off(adata, psize, asize - psize);
3996 abuf = abd_borrow_buf_copy(adata, asize);
3998 zio->io_cksum_report = zcr->zcr_next;
3999 zcr->zcr_next = NULL;
4000 zcr->zcr_finish(zcr, abuf);
4001 zfs_ereport_free_checksum(zcr);
4004 abd_return_buf(adata, abuf, asize);
4011 zio_pop_transforms(zio); /* note: may set zio->io_error */
4013 vdev_stat_update(zio, psize);
4015 if (zio->io_error) {
4017 * If this I/O is attached to a particular vdev,
4018 * generate an error message describing the I/O failure
4019 * at the block level. We ignore these errors if the
4020 * device is currently unavailable.
4022 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
4023 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
4025 if ((zio->io_error == EIO || !(zio->io_flags &
4026 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
4029 * For logical I/O requests, tell the SPA to log the
4030 * error and generate a logical data ereport.
4032 spa_log_error(spa, zio);
4033 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
4038 if (zio->io_error && zio == lio) {
4040 * Determine whether zio should be reexecuted. This will
4041 * propagate all the way to the root via zio_notify_parent().
4043 ASSERT(vd == NULL && bp != NULL);
4044 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
4046 if (IO_IS_ALLOCATING(zio) &&
4047 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
4048 if (zio->io_error != ENOSPC)
4049 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
4051 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4054 if ((zio->io_type == ZIO_TYPE_READ ||
4055 zio->io_type == ZIO_TYPE_FREE) &&
4056 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
4057 zio->io_error == ENXIO &&
4058 spa_load_state(spa) == SPA_LOAD_NONE &&
4059 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
4060 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4062 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
4063 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4066 * Here is a possibly good place to attempt to do
4067 * either combinatorial reconstruction or error correction
4068 * based on checksums. It also might be a good place
4069 * to send out preliminary ereports before we suspend
4075 * If there were logical child errors, they apply to us now.
4076 * We defer this until now to avoid conflating logical child
4077 * errors with errors that happened to the zio itself when
4078 * updating vdev stats and reporting FMA events above.
4080 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
4082 if ((zio->io_error || zio->io_reexecute) &&
4083 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
4084 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
4085 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
4087 zio_gang_tree_free(&zio->io_gang_tree);
4090 * Godfather I/Os should never suspend.
4092 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
4093 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
4094 zio->io_reexecute = 0;
4096 if (zio->io_reexecute) {
4098 * This is a logical I/O that wants to reexecute.
4100 * Reexecute is top-down. When an i/o fails, if it's not
4101 * the root, it simply notifies its parent and sticks around.
4102 * The parent, seeing that it still has children in zio_done(),
4103 * does the same. This percolates all the way up to the root.
4104 * The root i/o will reexecute or suspend the entire tree.
4106 * This approach ensures that zio_reexecute() honors
4107 * all the original i/o dependency relationships, e.g.
4108 * parents not executing until children are ready.
4110 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
4112 zio->io_gang_leader = NULL;
4114 mutex_enter(&zio->io_lock);
4115 zio->io_state[ZIO_WAIT_DONE] = 1;
4116 mutex_exit(&zio->io_lock);
4119 * "The Godfather" I/O monitors its children but is
4120 * not a true parent to them. It will track them through
4121 * the pipeline but severs its ties whenever they get into
4122 * trouble (e.g. suspended). This allows "The Godfather"
4123 * I/O to return status without blocking.
4126 for (pio = zio_walk_parents(zio, &zl); pio != NULL;
4128 zio_link_t *remove_zl = zl;
4129 pio_next = zio_walk_parents(zio, &zl);
4131 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
4132 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
4133 zio_remove_child(pio, zio, remove_zl);
4135 * This is a rare code path, so we don't
4136 * bother with "next_to_execute".
4138 zio_notify_parent(pio, zio, ZIO_WAIT_DONE,
4143 if ((pio = zio_unique_parent(zio)) != NULL) {
4145 * We're not a root i/o, so there's nothing to do
4146 * but notify our parent. Don't propagate errors
4147 * upward since we haven't permanently failed yet.
4149 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
4150 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
4152 * This is a rare code path, so we don't bother with
4153 * "next_to_execute".
4155 zio_notify_parent(pio, zio, ZIO_WAIT_DONE, NULL);
4156 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
4158 * We'd fail again if we reexecuted now, so suspend
4159 * until conditions improve (e.g. device comes online).
4161 zio_suspend(zio->io_spa, zio, ZIO_SUSPEND_IOERR);
4164 * Reexecution is potentially a huge amount of work.
4165 * Hand it off to the otherwise-unused claim taskq.
4167 #if defined(illumos) || !defined(_KERNEL)
4168 ASSERT(zio->io_tqent.tqent_next == NULL);
4170 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
4172 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
4173 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
4179 ASSERT(zio->io_child_count == 0);
4180 ASSERT(zio->io_reexecute == 0);
4181 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
4184 * Report any checksum errors, since the I/O is complete.
4186 while (zio->io_cksum_report != NULL) {
4187 zio_cksum_report_t *zcr = zio->io_cksum_report;
4188 zio->io_cksum_report = zcr->zcr_next;
4189 zcr->zcr_next = NULL;
4190 zcr->zcr_finish(zcr, NULL);
4191 zfs_ereport_free_checksum(zcr);
4195 * It is the responsibility of the done callback to ensure that this
4196 * particular zio is no longer discoverable for adoption, and as
4197 * such, cannot acquire any new parents.
4202 mutex_enter(&zio->io_lock);
4203 zio->io_state[ZIO_WAIT_DONE] = 1;
4204 mutex_exit(&zio->io_lock);
4207 * We are done executing this zio. We may want to execute a parent
4208 * next. See the comment in zio_notify_parent().
4210 zio_t *next_to_execute = NULL;
4212 for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) {
4213 zio_link_t *remove_zl = zl;
4214 pio_next = zio_walk_parents(zio, &zl);
4215 zio_remove_child(pio, zio, remove_zl);
4216 zio_notify_parent(pio, zio, ZIO_WAIT_DONE, &next_to_execute);
4219 if (zio->io_waiter != NULL) {
4220 mutex_enter(&zio->io_lock);
4221 zio->io_executor = NULL;
4222 cv_broadcast(&zio->io_cv);
4223 mutex_exit(&zio->io_lock);
4228 return (next_to_execute);
4232 * ==========================================================================
4233 * I/O pipeline definition
4234 * ==========================================================================
4236 static zio_pipe_stage_t *zio_pipeline[] = {
4243 zio_checksum_generate,
4259 zio_checksum_verify,
4267 * Compare two zbookmark_phys_t's to see which we would reach first in a
4268 * pre-order traversal of the object tree.
4270 * This is simple in every case aside from the meta-dnode object. For all other
4271 * objects, we traverse them in order (object 1 before object 2, and so on).
4272 * However, all of these objects are traversed while traversing object 0, since
4273 * the data it points to is the list of objects. Thus, we need to convert to a
4274 * canonical representation so we can compare meta-dnode bookmarks to
4275 * non-meta-dnode bookmarks.
4277 * We do this by calculating "equivalents" for each field of the zbookmark.
4278 * zbookmarks outside of the meta-dnode use their own object and level, and
4279 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
4280 * blocks this bookmark refers to) by multiplying their blkid by their span
4281 * (the number of L0 blocks contained within one block at their level).
4282 * zbookmarks inside the meta-dnode calculate their object equivalent
4283 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
4284 * level + 1<<31 (any value larger than a level could ever be) for their level.
4285 * This causes them to always compare before a bookmark in their object
4286 * equivalent, compare appropriately to bookmarks in other objects, and to
4287 * compare appropriately to other bookmarks in the meta-dnode.
4290 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
4291 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
4294 * These variables represent the "equivalent" values for the zbookmark,
4295 * after converting zbookmarks inside the meta dnode to their
4296 * normal-object equivalents.
4298 uint64_t zb1obj, zb2obj;
4299 uint64_t zb1L0, zb2L0;
4300 uint64_t zb1level, zb2level;
4302 if (zb1->zb_object == zb2->zb_object &&
4303 zb1->zb_level == zb2->zb_level &&
4304 zb1->zb_blkid == zb2->zb_blkid)
4308 * BP_SPANB calculates the span in blocks.
4310 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
4311 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
4313 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
4314 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4316 zb1level = zb1->zb_level + COMPARE_META_LEVEL;
4318 zb1obj = zb1->zb_object;
4319 zb1level = zb1->zb_level;
4322 if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
4323 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4325 zb2level = zb2->zb_level + COMPARE_META_LEVEL;
4327 zb2obj = zb2->zb_object;
4328 zb2level = zb2->zb_level;
4331 /* Now that we have a canonical representation, do the comparison. */
4332 if (zb1obj != zb2obj)
4333 return (zb1obj < zb2obj ? -1 : 1);
4334 else if (zb1L0 != zb2L0)
4335 return (zb1L0 < zb2L0 ? -1 : 1);
4336 else if (zb1level != zb2level)
4337 return (zb1level > zb2level ? -1 : 1);
4339 * This can (theoretically) happen if the bookmarks have the same object
4340 * and level, but different blkids, if the block sizes are not the same.
4341 * There is presently no way to change the indirect block sizes
4347 * This function checks the following: given that last_block is the place that
4348 * our traversal stopped last time, does that guarantee that we've visited
4349 * every node under subtree_root? Therefore, we can't just use the raw output
4350 * of zbookmark_compare. We have to pass in a modified version of
4351 * subtree_root; by incrementing the block id, and then checking whether
4352 * last_block is before or equal to that, we can tell whether or not having
4353 * visited last_block implies that all of subtree_root's children have been
4357 zbookmark_subtree_completed(const dnode_phys_t *dnp,
4358 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
4360 zbookmark_phys_t mod_zb = *subtree_root;
4362 ASSERT(last_block->zb_level == 0);
4364 /* The objset_phys_t isn't before anything. */
4369 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
4370 * data block size in sectors, because that variable is only used if
4371 * the bookmark refers to a block in the meta-dnode. Since we don't
4372 * know without examining it what object it refers to, and there's no
4373 * harm in passing in this value in other cases, we always pass it in.
4375 * We pass in 0 for the indirect block size shift because zb2 must be
4376 * level 0. The indirect block size is only used to calculate the span
4377 * of the bookmark, but since the bookmark must be level 0, the span is
4378 * always 1, so the math works out.
4380 * If you make changes to how the zbookmark_compare code works, be sure
4381 * to make sure that this code still works afterwards.
4383 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
4384 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,