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, 2017 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
25 * Copyright (c) 2014 Integros [integros.com]
28 #include <sys/sysmacros.h>
29 #include <sys/zfs_context.h>
30 #include <sys/fm/fs/zfs.h>
33 #include <sys/spa_impl.h>
34 #include <sys/vdev_impl.h>
35 #include <sys/zio_impl.h>
36 #include <sys/zio_compress.h>
37 #include <sys/zio_checksum.h>
38 #include <sys/dmu_objset.h>
41 #include <sys/trim_map.h>
42 #include <sys/blkptr.h>
43 #include <sys/zfeature.h>
44 #include <sys/metaslab_impl.h>
47 SYSCTL_DECL(_vfs_zfs);
48 SYSCTL_NODE(_vfs_zfs, OID_AUTO, zio, CTLFLAG_RW, 0, "ZFS ZIO");
49 #if defined(__amd64__)
50 static int zio_use_uma = 1;
52 static int zio_use_uma = 0;
54 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, use_uma, CTLFLAG_RDTUN, &zio_use_uma, 0,
55 "Use uma(9) for ZIO allocations");
56 static int zio_exclude_metadata = 0;
57 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, exclude_metadata, CTLFLAG_RDTUN, &zio_exclude_metadata, 0,
58 "Exclude metadata buffers from dumps as well");
60 zio_trim_stats_t zio_trim_stats = {
61 { "bytes", KSTAT_DATA_UINT64,
62 "Number of bytes successfully TRIMmed" },
63 { "success", KSTAT_DATA_UINT64,
64 "Number of successful TRIM requests" },
65 { "unsupported", KSTAT_DATA_UINT64,
66 "Number of TRIM requests that failed because TRIM is not supported" },
67 { "failed", KSTAT_DATA_UINT64,
68 "Number of TRIM requests that failed for reasons other than not supported" },
71 static kstat_t *zio_trim_ksp;
74 * ==========================================================================
75 * I/O type descriptions
76 * ==========================================================================
78 const char *zio_type_name[ZIO_TYPES] = {
79 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
83 boolean_t zio_dva_throttle_enabled = B_TRUE;
84 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, dva_throttle_enabled, CTLFLAG_RDTUN,
85 &zio_dva_throttle_enabled, 0, "");
88 * ==========================================================================
90 * ==========================================================================
92 kmem_cache_t *zio_cache;
93 kmem_cache_t *zio_link_cache;
94 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
95 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
98 extern vmem_t *zio_alloc_arena;
101 #define ZIO_PIPELINE_CONTINUE 0x100
102 #define ZIO_PIPELINE_STOP 0x101
104 #define BP_SPANB(indblkshift, level) \
105 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
106 #define COMPARE_META_LEVEL 0x80000000ul
108 * The following actions directly effect the spa's sync-to-convergence logic.
109 * The values below define the sync pass when we start performing the action.
110 * Care should be taken when changing these values as they directly impact
111 * spa_sync() performance. Tuning these values may introduce subtle performance
112 * pathologies and should only be done in the context of performance analysis.
113 * These tunables will eventually be removed and replaced with #defines once
114 * enough analysis has been done to determine optimal values.
116 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
117 * regular blocks are not deferred.
119 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
120 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_deferred_free, CTLFLAG_RDTUN,
121 &zfs_sync_pass_deferred_free, 0, "defer frees starting in this pass");
122 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
123 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_dont_compress, CTLFLAG_RDTUN,
124 &zfs_sync_pass_dont_compress, 0, "don't compress starting in this pass");
125 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
126 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_rewrite, CTLFLAG_RDTUN,
127 &zfs_sync_pass_rewrite, 0, "rewrite new bps starting in this pass");
130 * An allocating zio is one that either currently has the DVA allocate
131 * stage set or will have it later in its lifetime.
133 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
135 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
139 int zio_buf_debug_limit = 16384;
141 int zio_buf_debug_limit = 0;
145 static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t);
151 zio_cache = kmem_cache_create("zio_cache",
152 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
153 zio_link_cache = kmem_cache_create("zio_link_cache",
154 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
159 * For small buffers, we want a cache for each multiple of
160 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
161 * for each quarter-power of 2.
163 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
164 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
167 int cflags = zio_exclude_metadata ? KMC_NODEBUG : 0;
175 * If we are using watchpoints, put each buffer on its own page,
176 * to eliminate the performance overhead of trapping to the
177 * kernel when modifying a non-watched buffer that shares the
178 * page with a watched buffer.
180 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
184 if (size <= 4 * SPA_MINBLOCKSIZE) {
185 align = SPA_MINBLOCKSIZE;
186 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
187 align = MIN(p2 >> 2, PAGESIZE);
192 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
193 zio_buf_cache[c] = kmem_cache_create(name, size,
194 align, NULL, NULL, NULL, NULL, NULL, cflags);
197 * Since zio_data bufs do not appear in crash dumps, we
198 * pass KMC_NOTOUCH so that no allocator metadata is
199 * stored with the buffers.
201 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
202 zio_data_buf_cache[c] = kmem_cache_create(name, size,
203 align, NULL, NULL, NULL, NULL, NULL,
204 cflags | KMC_NOTOUCH | KMC_NODEBUG);
209 ASSERT(zio_buf_cache[c] != NULL);
210 if (zio_buf_cache[c - 1] == NULL)
211 zio_buf_cache[c - 1] = zio_buf_cache[c];
213 ASSERT(zio_data_buf_cache[c] != NULL);
214 if (zio_data_buf_cache[c - 1] == NULL)
215 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
221 zio_trim_ksp = kstat_create("zfs", 0, "zio_trim", "misc",
223 sizeof(zio_trim_stats) / sizeof(kstat_named_t),
226 if (zio_trim_ksp != NULL) {
227 zio_trim_ksp->ks_data = &zio_trim_stats;
228 kstat_install(zio_trim_ksp);
236 kmem_cache_t *last_cache = NULL;
237 kmem_cache_t *last_data_cache = NULL;
239 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
240 if (zio_buf_cache[c] != last_cache) {
241 last_cache = zio_buf_cache[c];
242 kmem_cache_destroy(zio_buf_cache[c]);
244 zio_buf_cache[c] = NULL;
246 if (zio_data_buf_cache[c] != last_data_cache) {
247 last_data_cache = zio_data_buf_cache[c];
248 kmem_cache_destroy(zio_data_buf_cache[c]);
250 zio_data_buf_cache[c] = NULL;
253 kmem_cache_destroy(zio_link_cache);
254 kmem_cache_destroy(zio_cache);
258 if (zio_trim_ksp != NULL) {
259 kstat_delete(zio_trim_ksp);
265 * ==========================================================================
266 * Allocate and free I/O buffers
267 * ==========================================================================
271 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
272 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
273 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
274 * excess / transient data in-core during a crashdump.
277 zio_buf_alloc(size_t size)
279 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
280 int flags = zio_exclude_metadata ? KM_NODEBUG : 0;
282 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
285 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
287 return (kmem_alloc(size, KM_SLEEP|flags));
291 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
292 * crashdump if the kernel panics. This exists so that we will limit the amount
293 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
294 * of kernel heap dumped to disk when the kernel panics)
297 zio_data_buf_alloc(size_t size)
299 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
301 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
304 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
306 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG));
310 zio_buf_free(void *buf, size_t size)
312 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
314 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
317 kmem_cache_free(zio_buf_cache[c], buf);
319 kmem_free(buf, size);
323 zio_data_buf_free(void *buf, size_t size)
325 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
327 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
330 kmem_cache_free(zio_data_buf_cache[c], buf);
332 kmem_free(buf, size);
336 * ==========================================================================
337 * Push and pop I/O transform buffers
338 * ==========================================================================
341 zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize,
342 zio_transform_func_t *transform)
344 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
347 * Ensure that anyone expecting this zio to contain a linear ABD isn't
348 * going to get a nasty surprise when they try to access the data.
350 IMPLY(abd_is_linear(zio->io_abd), abd_is_linear(data));
352 zt->zt_orig_abd = zio->io_abd;
353 zt->zt_orig_size = zio->io_size;
354 zt->zt_bufsize = bufsize;
355 zt->zt_transform = transform;
357 zt->zt_next = zio->io_transform_stack;
358 zio->io_transform_stack = zt;
365 zio_pop_transforms(zio_t *zio)
369 while ((zt = zio->io_transform_stack) != NULL) {
370 if (zt->zt_transform != NULL)
371 zt->zt_transform(zio,
372 zt->zt_orig_abd, zt->zt_orig_size);
374 if (zt->zt_bufsize != 0)
375 abd_free(zio->io_abd);
377 zio->io_abd = zt->zt_orig_abd;
378 zio->io_size = zt->zt_orig_size;
379 zio->io_transform_stack = zt->zt_next;
381 kmem_free(zt, sizeof (zio_transform_t));
386 * ==========================================================================
387 * I/O transform callbacks for subblocks and decompression
388 * ==========================================================================
391 zio_subblock(zio_t *zio, abd_t *data, uint64_t size)
393 ASSERT(zio->io_size > size);
395 if (zio->io_type == ZIO_TYPE_READ)
396 abd_copy(data, zio->io_abd, size);
400 zio_decompress(zio_t *zio, abd_t *data, uint64_t size)
402 if (zio->io_error == 0) {
403 void *tmp = abd_borrow_buf(data, size);
404 int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
405 zio->io_abd, tmp, zio->io_size, size);
406 abd_return_buf_copy(data, tmp, size);
409 zio->io_error = SET_ERROR(EIO);
414 * ==========================================================================
415 * I/O parent/child relationships and pipeline interlocks
416 * ==========================================================================
419 zio_walk_parents(zio_t *cio, zio_link_t **zl)
421 list_t *pl = &cio->io_parent_list;
423 *zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl);
427 ASSERT((*zl)->zl_child == cio);
428 return ((*zl)->zl_parent);
432 zio_walk_children(zio_t *pio, zio_link_t **zl)
434 list_t *cl = &pio->io_child_list;
436 *zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl);
440 ASSERT((*zl)->zl_parent == pio);
441 return ((*zl)->zl_child);
445 zio_unique_parent(zio_t *cio)
447 zio_link_t *zl = NULL;
448 zio_t *pio = zio_walk_parents(cio, &zl);
450 VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL);
455 zio_add_child(zio_t *pio, zio_t *cio)
457 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
460 * Logical I/Os can have logical, gang, or vdev children.
461 * Gang I/Os can have gang or vdev children.
462 * Vdev I/Os can only have vdev children.
463 * The following ASSERT captures all of these constraints.
465 ASSERT(cio->io_child_type <= pio->io_child_type);
470 mutex_enter(&cio->io_lock);
471 mutex_enter(&pio->io_lock);
473 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
475 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
476 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
478 list_insert_head(&pio->io_child_list, zl);
479 list_insert_head(&cio->io_parent_list, zl);
481 pio->io_child_count++;
482 cio->io_parent_count++;
484 mutex_exit(&pio->io_lock);
485 mutex_exit(&cio->io_lock);
489 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
491 ASSERT(zl->zl_parent == pio);
492 ASSERT(zl->zl_child == cio);
494 mutex_enter(&cio->io_lock);
495 mutex_enter(&pio->io_lock);
497 list_remove(&pio->io_child_list, zl);
498 list_remove(&cio->io_parent_list, zl);
500 pio->io_child_count--;
501 cio->io_parent_count--;
503 mutex_exit(&pio->io_lock);
504 mutex_exit(&cio->io_lock);
506 kmem_cache_free(zio_link_cache, zl);
510 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
512 uint64_t *countp = &zio->io_children[child][wait];
513 boolean_t waiting = B_FALSE;
515 mutex_enter(&zio->io_lock);
516 ASSERT(zio->io_stall == NULL);
519 ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN);
520 zio->io_stall = countp;
523 mutex_exit(&zio->io_lock);
529 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
531 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
532 int *errorp = &pio->io_child_error[zio->io_child_type];
534 mutex_enter(&pio->io_lock);
535 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
536 *errorp = zio_worst_error(*errorp, zio->io_error);
537 pio->io_reexecute |= zio->io_reexecute;
538 ASSERT3U(*countp, >, 0);
542 if (*countp == 0 && pio->io_stall == countp) {
543 zio_taskq_type_t type =
544 pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE :
546 pio->io_stall = NULL;
547 mutex_exit(&pio->io_lock);
549 * Dispatch the parent zio in its own taskq so that
550 * the child can continue to make progress. This also
551 * prevents overflowing the stack when we have deeply nested
552 * parent-child relationships.
554 zio_taskq_dispatch(pio, type, B_FALSE);
556 mutex_exit(&pio->io_lock);
561 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
563 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
564 zio->io_error = zio->io_child_error[c];
568 zio_bookmark_compare(const void *x1, const void *x2)
570 const zio_t *z1 = x1;
571 const zio_t *z2 = x2;
573 if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset)
575 if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset)
578 if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object)
580 if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object)
583 if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level)
585 if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level)
588 if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid)
590 if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid)
602 * ==========================================================================
603 * Create the various types of I/O (read, write, free, etc)
604 * ==========================================================================
607 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
608 abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done,
609 void *private, zio_type_t type, zio_priority_t priority,
610 enum zio_flag flags, vdev_t *vd, uint64_t offset,
611 const zbookmark_phys_t *zb, enum zio_stage stage, enum zio_stage pipeline)
615 ASSERT3U(type == ZIO_TYPE_FREE || psize, <=, SPA_MAXBLOCKSIZE);
616 ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0);
617 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
619 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
620 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
621 ASSERT(vd || stage == ZIO_STAGE_OPEN);
623 IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW) != 0);
625 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
626 bzero(zio, sizeof (zio_t));
628 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
629 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
631 list_create(&zio->io_parent_list, sizeof (zio_link_t),
632 offsetof(zio_link_t, zl_parent_node));
633 list_create(&zio->io_child_list, sizeof (zio_link_t),
634 offsetof(zio_link_t, zl_child_node));
635 metaslab_trace_init(&zio->io_alloc_list);
638 zio->io_child_type = ZIO_CHILD_VDEV;
639 else if (flags & ZIO_FLAG_GANG_CHILD)
640 zio->io_child_type = ZIO_CHILD_GANG;
641 else if (flags & ZIO_FLAG_DDT_CHILD)
642 zio->io_child_type = ZIO_CHILD_DDT;
644 zio->io_child_type = ZIO_CHILD_LOGICAL;
647 zio->io_bp = (blkptr_t *)bp;
648 zio->io_bp_copy = *bp;
649 zio->io_bp_orig = *bp;
650 if (type != ZIO_TYPE_WRITE ||
651 zio->io_child_type == ZIO_CHILD_DDT)
652 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
653 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
654 zio->io_logical = zio;
655 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
656 pipeline |= ZIO_GANG_STAGES;
662 zio->io_private = private;
664 zio->io_priority = priority;
666 zio->io_offset = offset;
667 zio->io_orig_abd = zio->io_abd = data;
668 zio->io_orig_size = zio->io_size = psize;
669 zio->io_lsize = lsize;
670 zio->io_orig_flags = zio->io_flags = flags;
671 zio->io_orig_stage = zio->io_stage = stage;
672 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
673 zio->io_pipeline_trace = ZIO_STAGE_OPEN;
675 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
676 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
679 zio->io_bookmark = *zb;
682 if (zio->io_logical == NULL)
683 zio->io_logical = pio->io_logical;
684 if (zio->io_child_type == ZIO_CHILD_GANG)
685 zio->io_gang_leader = pio->io_gang_leader;
686 zio_add_child(pio, zio);
693 zio_destroy(zio_t *zio)
695 metaslab_trace_fini(&zio->io_alloc_list);
696 list_destroy(&zio->io_parent_list);
697 list_destroy(&zio->io_child_list);
698 mutex_destroy(&zio->io_lock);
699 cv_destroy(&zio->io_cv);
700 kmem_cache_free(zio_cache, zio);
704 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
705 void *private, enum zio_flag flags)
709 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
710 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
711 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
717 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
719 return (zio_null(NULL, spa, NULL, done, private, flags));
723 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
725 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
726 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
727 bp, (longlong_t)BP_GET_TYPE(bp));
729 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
730 BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
731 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
732 bp, (longlong_t)BP_GET_CHECKSUM(bp));
734 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
735 BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
736 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
737 bp, (longlong_t)BP_GET_COMPRESS(bp));
739 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
740 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
741 bp, (longlong_t)BP_GET_LSIZE(bp));
743 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
744 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
745 bp, (longlong_t)BP_GET_PSIZE(bp));
748 if (BP_IS_EMBEDDED(bp)) {
749 if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
750 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
751 bp, (longlong_t)BPE_GET_ETYPE(bp));
756 * Pool-specific checks.
758 * Note: it would be nice to verify that the blk_birth and
759 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
760 * allows the birth time of log blocks (and dmu_sync()-ed blocks
761 * that are in the log) to be arbitrarily large.
763 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
764 uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
765 if (vdevid >= spa->spa_root_vdev->vdev_children) {
766 zfs_panic_recover("blkptr at %p DVA %u has invalid "
768 bp, i, (longlong_t)vdevid);
771 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
773 zfs_panic_recover("blkptr at %p DVA %u has invalid "
775 bp, i, (longlong_t)vdevid);
778 if (vd->vdev_ops == &vdev_hole_ops) {
779 zfs_panic_recover("blkptr at %p DVA %u has hole "
781 bp, i, (longlong_t)vdevid);
784 if (vd->vdev_ops == &vdev_missing_ops) {
786 * "missing" vdevs are valid during import, but we
787 * don't have their detailed info (e.g. asize), so
788 * we can't perform any more checks on them.
792 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
793 uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
795 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
796 if (offset + asize > vd->vdev_asize) {
797 zfs_panic_recover("blkptr at %p DVA %u has invalid "
799 bp, i, (longlong_t)offset);
805 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
806 abd_t *data, uint64_t size, zio_done_func_t *done, void *private,
807 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
811 zfs_blkptr_verify(spa, bp);
813 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
814 data, size, size, done, private,
815 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
816 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
817 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
823 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
824 abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp,
825 zio_done_func_t *ready, zio_done_func_t *children_ready,
826 zio_done_func_t *physdone, zio_done_func_t *done,
827 void *private, zio_priority_t priority, enum zio_flag flags,
828 const zbookmark_phys_t *zb)
832 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
833 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
834 zp->zp_compress >= ZIO_COMPRESS_OFF &&
835 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
836 DMU_OT_IS_VALID(zp->zp_type) &&
839 zp->zp_copies <= spa_max_replication(spa));
841 zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private,
842 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
843 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
844 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
846 zio->io_ready = ready;
847 zio->io_children_ready = children_ready;
848 zio->io_physdone = physdone;
852 * Data can be NULL if we are going to call zio_write_override() to
853 * provide the already-allocated BP. But we may need the data to
854 * verify a dedup hit (if requested). In this case, don't try to
855 * dedup (just take the already-allocated BP verbatim).
857 if (data == NULL && zio->io_prop.zp_dedup_verify) {
858 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
865 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data,
866 uint64_t size, zio_done_func_t *done, void *private,
867 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
871 zio = zio_create(pio, spa, txg, bp, data, size, size, done, private,
872 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb,
873 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
879 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
881 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
882 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
883 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
884 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
887 * We must reset the io_prop to match the values that existed
888 * when the bp was first written by dmu_sync() keeping in mind
889 * that nopwrite and dedup are mutually exclusive.
891 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
892 zio->io_prop.zp_nopwrite = nopwrite;
893 zio->io_prop.zp_copies = copies;
894 zio->io_bp_override = bp;
898 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
902 * The check for EMBEDDED is a performance optimization. We
903 * process the free here (by ignoring it) rather than
904 * putting it on the list and then processing it in zio_free_sync().
906 if (BP_IS_EMBEDDED(bp))
908 metaslab_check_free(spa, bp);
911 * Frees that are for the currently-syncing txg, are not going to be
912 * deferred, and which will not need to do a read (i.e. not GANG or
913 * DEDUP), can be processed immediately. Otherwise, put them on the
914 * in-memory list for later processing.
916 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
917 txg != spa->spa_syncing_txg ||
918 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
919 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
921 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp,
922 BP_GET_PSIZE(bp), 0)));
927 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
928 uint64_t size, enum zio_flag flags)
931 enum zio_stage stage = ZIO_FREE_PIPELINE;
933 ASSERT(!BP_IS_HOLE(bp));
934 ASSERT(spa_syncing_txg(spa) == txg);
935 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
937 if (BP_IS_EMBEDDED(bp))
938 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
940 metaslab_check_free(spa, bp);
943 if (zfs_trim_enabled)
944 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START |
945 ZIO_STAGE_VDEV_IO_ASSESS;
947 * GANG and DEDUP blocks can induce a read (for the gang block header,
948 * or the DDT), so issue them asynchronously so that this thread is
951 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
952 stage |= ZIO_STAGE_ISSUE_ASYNC;
954 flags |= ZIO_FLAG_DONT_QUEUE;
956 zio = zio_create(pio, spa, txg, bp, NULL, size,
957 size, NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW,
958 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
964 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
965 zio_done_func_t *done, void *private, enum zio_flag flags)
969 dprintf_bp(bp, "claiming in txg %llu", txg);
971 if (BP_IS_EMBEDDED(bp))
972 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
975 * A claim is an allocation of a specific block. Claims are needed
976 * to support immediate writes in the intent log. The issue is that
977 * immediate writes contain committed data, but in a txg that was
978 * *not* committed. Upon opening the pool after an unclean shutdown,
979 * the intent log claims all blocks that contain immediate write data
980 * so that the SPA knows they're in use.
982 * All claims *must* be resolved in the first txg -- before the SPA
983 * starts allocating blocks -- so that nothing is allocated twice.
984 * If txg == 0 we just verify that the block is claimable.
986 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
987 ASSERT(txg == spa_first_txg(spa) || txg == 0);
988 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
990 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
991 BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW,
992 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
993 ASSERT0(zio->io_queued_timestamp);
999 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset,
1000 uint64_t size, zio_done_func_t *done, void *private,
1001 zio_priority_t priority, enum zio_flag flags)
1006 if (vd->vdev_children == 0) {
1007 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
1008 ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
1009 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
1013 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
1015 for (c = 0; c < vd->vdev_children; c++)
1016 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
1017 offset, size, done, private, priority, flags));
1024 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1025 abd_t *data, int checksum, zio_done_func_t *done, void *private,
1026 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1030 ASSERT(vd->vdev_children == 0);
1031 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1032 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1033 ASSERT3U(offset + size, <=, vd->vdev_psize);
1035 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1036 private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1037 offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
1039 zio->io_prop.zp_checksum = checksum;
1045 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1046 abd_t *data, int checksum, zio_done_func_t *done, void *private,
1047 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1051 ASSERT(vd->vdev_children == 0);
1052 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1053 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1054 ASSERT3U(offset + size, <=, vd->vdev_psize);
1056 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1057 private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1058 offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
1060 zio->io_prop.zp_checksum = checksum;
1062 if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
1064 * zec checksums are necessarily destructive -- they modify
1065 * the end of the write buffer to hold the verifier/checksum.
1066 * Therefore, we must make a local copy in case the data is
1067 * being written to multiple places in parallel.
1069 abd_t *wbuf = abd_alloc_sametype(data, size);
1070 abd_copy(wbuf, data, size);
1072 zio_push_transform(zio, wbuf, size, size, NULL);
1079 * Create a child I/O to do some work for us.
1082 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1083 abd_t *data, uint64_t size, int type, zio_priority_t priority,
1084 enum zio_flag flags, zio_done_func_t *done, void *private)
1086 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1089 ASSERT(vd->vdev_parent ==
1090 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
1092 if (type == ZIO_TYPE_READ && bp != NULL) {
1094 * If we have the bp, then the child should perform the
1095 * checksum and the parent need not. This pushes error
1096 * detection as close to the leaves as possible and
1097 * eliminates redundant checksums in the interior nodes.
1099 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1100 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1103 /* Not all IO types require vdev io done stage e.g. free */
1104 if (!(pio->io_pipeline & ZIO_STAGE_VDEV_IO_DONE))
1105 pipeline &= ~ZIO_STAGE_VDEV_IO_DONE;
1107 if (vd->vdev_children == 0)
1108 offset += VDEV_LABEL_START_SIZE;
1110 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
1113 * If we've decided to do a repair, the write is not speculative --
1114 * even if the original read was.
1116 if (flags & ZIO_FLAG_IO_REPAIR)
1117 flags &= ~ZIO_FLAG_SPECULATIVE;
1120 * If we're creating a child I/O that is not associated with a
1121 * top-level vdev, then the child zio is not an allocating I/O.
1122 * If this is a retried I/O then we ignore it since we will
1123 * have already processed the original allocating I/O.
1125 if (flags & ZIO_FLAG_IO_ALLOCATING &&
1126 (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) {
1127 metaslab_class_t *mc = spa_normal_class(pio->io_spa);
1129 ASSERT(mc->mc_alloc_throttle_enabled);
1130 ASSERT(type == ZIO_TYPE_WRITE);
1131 ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE);
1132 ASSERT(!(flags & ZIO_FLAG_IO_REPAIR));
1133 ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) ||
1134 pio->io_child_type == ZIO_CHILD_GANG);
1136 flags &= ~ZIO_FLAG_IO_ALLOCATING;
1139 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size,
1140 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1141 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1142 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
1144 zio->io_physdone = pio->io_physdone;
1145 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1146 zio->io_logical->io_phys_children++;
1152 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size,
1153 int type, zio_priority_t priority, enum zio_flag flags,
1154 zio_done_func_t *done, void *private)
1158 ASSERT(vd->vdev_ops->vdev_op_leaf);
1160 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1161 data, size, size, done, private, type, priority,
1162 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1164 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1170 zio_flush(zio_t *zio, vdev_t *vd)
1172 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0,
1173 NULL, NULL, ZIO_PRIORITY_NOW,
1174 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1178 zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size)
1181 ASSERT(vd->vdev_ops->vdev_op_leaf);
1183 return (zio_create(zio, spa, 0, NULL, NULL, size, size, NULL, NULL,
1184 ZIO_TYPE_FREE, ZIO_PRIORITY_TRIM, ZIO_FLAG_DONT_AGGREGATE |
1185 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY,
1186 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PHYS_PIPELINE));
1190 zio_shrink(zio_t *zio, uint64_t size)
1192 ASSERT(zio->io_executor == NULL);
1193 ASSERT(zio->io_orig_size == zio->io_size);
1194 ASSERT(size <= zio->io_size);
1197 * We don't shrink for raidz because of problems with the
1198 * reconstruction when reading back less than the block size.
1199 * Note, BP_IS_RAIDZ() assumes no compression.
1201 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1202 if (!BP_IS_RAIDZ(zio->io_bp)) {
1203 /* we are not doing a raw write */
1204 ASSERT3U(zio->io_size, ==, zio->io_lsize);
1205 zio->io_orig_size = zio->io_size = zio->io_lsize = size;
1210 * ==========================================================================
1211 * Prepare to read and write logical blocks
1212 * ==========================================================================
1216 zio_read_bp_init(zio_t *zio)
1218 blkptr_t *bp = zio->io_bp;
1220 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1221 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1222 !(zio->io_flags & ZIO_FLAG_RAW)) {
1224 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1225 zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1226 psize, psize, zio_decompress);
1229 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1230 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1232 int psize = BPE_GET_PSIZE(bp);
1233 void *data = abd_borrow_buf(zio->io_abd, psize);
1234 decode_embedded_bp_compressed(bp, data);
1235 abd_return_buf_copy(zio->io_abd, data, psize);
1237 ASSERT(!BP_IS_EMBEDDED(bp));
1240 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1241 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1243 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1244 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1246 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1247 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1249 return (ZIO_PIPELINE_CONTINUE);
1253 zio_write_bp_init(zio_t *zio)
1255 if (!IO_IS_ALLOCATING(zio))
1256 return (ZIO_PIPELINE_CONTINUE);
1258 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1260 if (zio->io_bp_override) {
1261 blkptr_t *bp = zio->io_bp;
1262 zio_prop_t *zp = &zio->io_prop;
1264 ASSERT(bp->blk_birth != zio->io_txg);
1265 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1267 *bp = *zio->io_bp_override;
1268 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1270 if (BP_IS_EMBEDDED(bp))
1271 return (ZIO_PIPELINE_CONTINUE);
1274 * If we've been overridden and nopwrite is set then
1275 * set the flag accordingly to indicate that a nopwrite
1276 * has already occurred.
1278 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1279 ASSERT(!zp->zp_dedup);
1280 ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum);
1281 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1282 return (ZIO_PIPELINE_CONTINUE);
1285 ASSERT(!zp->zp_nopwrite);
1287 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1288 return (ZIO_PIPELINE_CONTINUE);
1290 ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags &
1291 ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify);
1293 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1294 BP_SET_DEDUP(bp, 1);
1295 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1296 return (ZIO_PIPELINE_CONTINUE);
1300 * We were unable to handle this as an override bp, treat
1301 * it as a regular write I/O.
1303 zio->io_bp_override = NULL;
1304 *bp = zio->io_bp_orig;
1305 zio->io_pipeline = zio->io_orig_pipeline;
1308 return (ZIO_PIPELINE_CONTINUE);
1312 zio_write_compress(zio_t *zio)
1314 spa_t *spa = zio->io_spa;
1315 zio_prop_t *zp = &zio->io_prop;
1316 enum zio_compress compress = zp->zp_compress;
1317 blkptr_t *bp = zio->io_bp;
1318 uint64_t lsize = zio->io_lsize;
1319 uint64_t psize = zio->io_size;
1322 EQUIV(lsize != psize, (zio->io_flags & ZIO_FLAG_RAW) != 0);
1325 * If our children haven't all reached the ready stage,
1326 * wait for them and then repeat this pipeline stage.
1328 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1329 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1330 return (ZIO_PIPELINE_STOP);
1332 if (!IO_IS_ALLOCATING(zio))
1333 return (ZIO_PIPELINE_CONTINUE);
1335 if (zio->io_children_ready != NULL) {
1337 * Now that all our children are ready, run the callback
1338 * associated with this zio in case it wants to modify the
1339 * data to be written.
1341 ASSERT3U(zp->zp_level, >, 0);
1342 zio->io_children_ready(zio);
1345 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1346 ASSERT(zio->io_bp_override == NULL);
1348 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1350 * We're rewriting an existing block, which means we're
1351 * working on behalf of spa_sync(). For spa_sync() to
1352 * converge, it must eventually be the case that we don't
1353 * have to allocate new blocks. But compression changes
1354 * the blocksize, which forces a reallocate, and makes
1355 * convergence take longer. Therefore, after the first
1356 * few passes, stop compressing to ensure convergence.
1358 pass = spa_sync_pass(spa);
1360 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1361 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1362 ASSERT(!BP_GET_DEDUP(bp));
1364 if (pass >= zfs_sync_pass_dont_compress)
1365 compress = ZIO_COMPRESS_OFF;
1367 /* Make sure someone doesn't change their mind on overwrites */
1368 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1369 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1372 /* If it's a compressed write that is not raw, compress the buffer. */
1373 if (compress != ZIO_COMPRESS_OFF && psize == lsize) {
1374 void *cbuf = zio_buf_alloc(lsize);
1375 psize = zio_compress_data(compress, zio->io_abd, cbuf, lsize);
1376 if (psize == 0 || psize == lsize) {
1377 compress = ZIO_COMPRESS_OFF;
1378 zio_buf_free(cbuf, lsize);
1379 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1380 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1381 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1382 encode_embedded_bp_compressed(bp,
1383 cbuf, compress, lsize, psize);
1384 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1385 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1386 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1387 zio_buf_free(cbuf, lsize);
1388 bp->blk_birth = zio->io_txg;
1389 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1390 ASSERT(spa_feature_is_active(spa,
1391 SPA_FEATURE_EMBEDDED_DATA));
1392 return (ZIO_PIPELINE_CONTINUE);
1395 * Round up compressed size up to the ashift
1396 * of the smallest-ashift device, and zero the tail.
1397 * This ensures that the compressed size of the BP
1398 * (and thus compressratio property) are correct,
1399 * in that we charge for the padding used to fill out
1402 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
1403 size_t rounded = (size_t)P2ROUNDUP(psize,
1404 1ULL << spa->spa_min_ashift);
1405 if (rounded >= lsize) {
1406 compress = ZIO_COMPRESS_OFF;
1407 zio_buf_free(cbuf, lsize);
1410 abd_t *cdata = abd_get_from_buf(cbuf, lsize);
1411 abd_take_ownership_of_buf(cdata, B_TRUE);
1412 abd_zero_off(cdata, psize, rounded - psize);
1414 zio_push_transform(zio, cdata,
1415 psize, lsize, NULL);
1420 * We were unable to handle this as an override bp, treat
1421 * it as a regular write I/O.
1423 zio->io_bp_override = NULL;
1424 *bp = zio->io_bp_orig;
1425 zio->io_pipeline = zio->io_orig_pipeline;
1427 ASSERT3U(psize, !=, 0);
1431 * The final pass of spa_sync() must be all rewrites, but the first
1432 * few passes offer a trade-off: allocating blocks defers convergence,
1433 * but newly allocated blocks are sequential, so they can be written
1434 * to disk faster. Therefore, we allow the first few passes of
1435 * spa_sync() to allocate new blocks, but force rewrites after that.
1436 * There should only be a handful of blocks after pass 1 in any case.
1438 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1439 BP_GET_PSIZE(bp) == psize &&
1440 pass >= zfs_sync_pass_rewrite) {
1442 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1443 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1444 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1447 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1451 if (zio->io_bp_orig.blk_birth != 0 &&
1452 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1453 BP_SET_LSIZE(bp, lsize);
1454 BP_SET_TYPE(bp, zp->zp_type);
1455 BP_SET_LEVEL(bp, zp->zp_level);
1456 BP_SET_BIRTH(bp, zio->io_txg, 0);
1458 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1460 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1461 BP_SET_LSIZE(bp, lsize);
1462 BP_SET_TYPE(bp, zp->zp_type);
1463 BP_SET_LEVEL(bp, zp->zp_level);
1464 BP_SET_PSIZE(bp, psize);
1465 BP_SET_COMPRESS(bp, compress);
1466 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1467 BP_SET_DEDUP(bp, zp->zp_dedup);
1468 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1470 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1471 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1472 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1474 if (zp->zp_nopwrite) {
1475 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1476 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1477 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1480 return (ZIO_PIPELINE_CONTINUE);
1484 zio_free_bp_init(zio_t *zio)
1486 blkptr_t *bp = zio->io_bp;
1488 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1489 if (BP_GET_DEDUP(bp))
1490 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1493 return (ZIO_PIPELINE_CONTINUE);
1497 * ==========================================================================
1498 * Execute the I/O pipeline
1499 * ==========================================================================
1503 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1505 spa_t *spa = zio->io_spa;
1506 zio_type_t t = zio->io_type;
1507 int flags = (cutinline ? TQ_FRONT : 0);
1509 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1512 * If we're a config writer or a probe, the normal issue and
1513 * interrupt threads may all be blocked waiting for the config lock.
1514 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1516 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1520 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1522 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1526 * If this is a high priority I/O, then use the high priority taskq if
1529 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1530 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1533 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1536 * NB: We are assuming that the zio can only be dispatched
1537 * to a single taskq at a time. It would be a grievous error
1538 * to dispatch the zio to another taskq at the same time.
1540 #if defined(illumos) || !defined(_KERNEL)
1541 ASSERT(zio->io_tqent.tqent_next == NULL);
1543 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
1545 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1546 flags, &zio->io_tqent);
1550 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1552 kthread_t *executor = zio->io_executor;
1553 spa_t *spa = zio->io_spa;
1555 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1556 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1558 for (i = 0; i < tqs->stqs_count; i++) {
1559 if (taskq_member(tqs->stqs_taskq[i], executor))
1568 zio_issue_async(zio_t *zio)
1570 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1572 return (ZIO_PIPELINE_STOP);
1576 zio_interrupt(zio_t *zio)
1578 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1582 zio_delay_interrupt(zio_t *zio)
1585 * The timeout_generic() function isn't defined in userspace, so
1586 * rather than trying to implement the function, the zio delay
1587 * functionality has been disabled for userspace builds.
1592 * If io_target_timestamp is zero, then no delay has been registered
1593 * for this IO, thus jump to the end of this function and "skip" the
1594 * delay; issuing it directly to the zio layer.
1596 if (zio->io_target_timestamp != 0) {
1597 hrtime_t now = gethrtime();
1599 if (now >= zio->io_target_timestamp) {
1601 * This IO has already taken longer than the target
1602 * delay to complete, so we don't want to delay it
1603 * any longer; we "miss" the delay and issue it
1604 * directly to the zio layer. This is likely due to
1605 * the target latency being set to a value less than
1606 * the underlying hardware can satisfy (e.g. delay
1607 * set to 1ms, but the disks take 10ms to complete an
1611 DTRACE_PROBE2(zio__delay__miss, zio_t *, zio,
1616 hrtime_t diff = zio->io_target_timestamp - now;
1618 DTRACE_PROBE3(zio__delay__hit, zio_t *, zio,
1619 hrtime_t, now, hrtime_t, diff);
1621 (void) timeout_generic(CALLOUT_NORMAL,
1622 (void (*)(void *))zio_interrupt, zio, diff, 1, 0);
1629 DTRACE_PROBE1(zio__delay__skip, zio_t *, zio);
1634 * Execute the I/O pipeline until one of the following occurs:
1636 * (1) the I/O completes
1637 * (2) the pipeline stalls waiting for dependent child I/Os
1638 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1639 * (4) the I/O is delegated by vdev-level caching or aggregation
1640 * (5) the I/O is deferred due to vdev-level queueing
1641 * (6) the I/O is handed off to another thread.
1643 * In all cases, the pipeline stops whenever there's no CPU work; it never
1644 * burns a thread in cv_wait().
1646 * There's no locking on io_stage because there's no legitimate way
1647 * for multiple threads to be attempting to process the same I/O.
1649 static zio_pipe_stage_t *zio_pipeline[];
1652 zio_execute(zio_t *zio)
1654 zio->io_executor = curthread;
1656 ASSERT3U(zio->io_queued_timestamp, >, 0);
1658 while (zio->io_stage < ZIO_STAGE_DONE) {
1659 enum zio_stage pipeline = zio->io_pipeline;
1660 enum zio_stage stage = zio->io_stage;
1663 ASSERT(!MUTEX_HELD(&zio->io_lock));
1664 ASSERT(ISP2(stage));
1665 ASSERT(zio->io_stall == NULL);
1669 } while ((stage & pipeline) == 0);
1671 ASSERT(stage <= ZIO_STAGE_DONE);
1674 * If we are in interrupt context and this pipeline stage
1675 * will grab a config lock that is held across I/O,
1676 * or may wait for an I/O that needs an interrupt thread
1677 * to complete, issue async to avoid deadlock.
1679 * For VDEV_IO_START, we cut in line so that the io will
1680 * be sent to disk promptly.
1682 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1683 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1684 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1685 zio_requeue_io_start_cut_in_line : B_FALSE;
1686 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1690 zio->io_stage = stage;
1691 zio->io_pipeline_trace |= zio->io_stage;
1692 rv = zio_pipeline[highbit64(stage) - 1](zio);
1694 if (rv == ZIO_PIPELINE_STOP)
1697 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1702 * ==========================================================================
1703 * Initiate I/O, either sync or async
1704 * ==========================================================================
1707 zio_wait(zio_t *zio)
1711 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1712 ASSERT(zio->io_executor == NULL);
1714 zio->io_waiter = curthread;
1715 ASSERT0(zio->io_queued_timestamp);
1716 zio->io_queued_timestamp = gethrtime();
1720 mutex_enter(&zio->io_lock);
1721 while (zio->io_executor != NULL)
1722 cv_wait(&zio->io_cv, &zio->io_lock);
1723 mutex_exit(&zio->io_lock);
1725 error = zio->io_error;
1732 zio_nowait(zio_t *zio)
1734 ASSERT(zio->io_executor == NULL);
1736 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1737 zio_unique_parent(zio) == NULL) {
1739 * This is a logical async I/O with no parent to wait for it.
1740 * We add it to the spa_async_root_zio "Godfather" I/O which
1741 * will ensure they complete prior to unloading the pool.
1743 spa_t *spa = zio->io_spa;
1745 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1748 ASSERT0(zio->io_queued_timestamp);
1749 zio->io_queued_timestamp = gethrtime();
1754 * ==========================================================================
1755 * Reexecute or suspend/resume failed I/O
1756 * ==========================================================================
1760 zio_reexecute(zio_t *pio)
1762 zio_t *cio, *cio_next;
1764 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1765 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1766 ASSERT(pio->io_gang_leader == NULL);
1767 ASSERT(pio->io_gang_tree == NULL);
1769 pio->io_flags = pio->io_orig_flags;
1770 pio->io_stage = pio->io_orig_stage;
1771 pio->io_pipeline = pio->io_orig_pipeline;
1772 pio->io_reexecute = 0;
1773 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1774 pio->io_pipeline_trace = 0;
1776 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1777 pio->io_state[w] = 0;
1778 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1779 pio->io_child_error[c] = 0;
1781 if (IO_IS_ALLOCATING(pio))
1782 BP_ZERO(pio->io_bp);
1785 * As we reexecute pio's children, new children could be created.
1786 * New children go to the head of pio's io_child_list, however,
1787 * so we will (correctly) not reexecute them. The key is that
1788 * the remainder of pio's io_child_list, from 'cio_next' onward,
1789 * cannot be affected by any side effects of reexecuting 'cio'.
1791 zio_link_t *zl = NULL;
1792 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
1793 cio_next = zio_walk_children(pio, &zl);
1794 mutex_enter(&pio->io_lock);
1795 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1796 pio->io_children[cio->io_child_type][w]++;
1797 mutex_exit(&pio->io_lock);
1802 * Now that all children have been reexecuted, execute the parent.
1803 * We don't reexecute "The Godfather" I/O here as it's the
1804 * responsibility of the caller to wait on him.
1806 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) {
1807 pio->io_queued_timestamp = gethrtime();
1813 zio_suspend(spa_t *spa, zio_t *zio)
1815 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1816 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1817 "failure and the failure mode property for this pool "
1818 "is set to panic.", spa_name(spa));
1820 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1822 mutex_enter(&spa->spa_suspend_lock);
1824 if (spa->spa_suspend_zio_root == NULL)
1825 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1826 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1827 ZIO_FLAG_GODFATHER);
1829 spa->spa_suspended = B_TRUE;
1832 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1833 ASSERT(zio != spa->spa_suspend_zio_root);
1834 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1835 ASSERT(zio_unique_parent(zio) == NULL);
1836 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1837 zio_add_child(spa->spa_suspend_zio_root, zio);
1840 mutex_exit(&spa->spa_suspend_lock);
1844 zio_resume(spa_t *spa)
1849 * Reexecute all previously suspended i/o.
1851 mutex_enter(&spa->spa_suspend_lock);
1852 spa->spa_suspended = B_FALSE;
1853 cv_broadcast(&spa->spa_suspend_cv);
1854 pio = spa->spa_suspend_zio_root;
1855 spa->spa_suspend_zio_root = NULL;
1856 mutex_exit(&spa->spa_suspend_lock);
1862 return (zio_wait(pio));
1866 zio_resume_wait(spa_t *spa)
1868 mutex_enter(&spa->spa_suspend_lock);
1869 while (spa_suspended(spa))
1870 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1871 mutex_exit(&spa->spa_suspend_lock);
1875 * ==========================================================================
1878 * A gang block is a collection of small blocks that looks to the DMU
1879 * like one large block. When zio_dva_allocate() cannot find a block
1880 * of the requested size, due to either severe fragmentation or the pool
1881 * being nearly full, it calls zio_write_gang_block() to construct the
1882 * block from smaller fragments.
1884 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1885 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1886 * an indirect block: it's an array of block pointers. It consumes
1887 * only one sector and hence is allocatable regardless of fragmentation.
1888 * The gang header's bps point to its gang members, which hold the data.
1890 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1891 * as the verifier to ensure uniqueness of the SHA256 checksum.
1892 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1893 * not the gang header. This ensures that data block signatures (needed for
1894 * deduplication) are independent of how the block is physically stored.
1896 * Gang blocks can be nested: a gang member may itself be a gang block.
1897 * Thus every gang block is a tree in which root and all interior nodes are
1898 * gang headers, and the leaves are normal blocks that contain user data.
1899 * The root of the gang tree is called the gang leader.
1901 * To perform any operation (read, rewrite, free, claim) on a gang block,
1902 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1903 * in the io_gang_tree field of the original logical i/o by recursively
1904 * reading the gang leader and all gang headers below it. This yields
1905 * an in-core tree containing the contents of every gang header and the
1906 * bps for every constituent of the gang block.
1908 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1909 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1910 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1911 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1912 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1913 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1914 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1915 * of the gang header plus zio_checksum_compute() of the data to update the
1916 * gang header's blk_cksum as described above.
1918 * The two-phase assemble/issue model solves the problem of partial failure --
1919 * what if you'd freed part of a gang block but then couldn't read the
1920 * gang header for another part? Assembling the entire gang tree first
1921 * ensures that all the necessary gang header I/O has succeeded before
1922 * starting the actual work of free, claim, or write. Once the gang tree
1923 * is assembled, free and claim are in-memory operations that cannot fail.
1925 * In the event that a gang write fails, zio_dva_unallocate() walks the
1926 * gang tree to immediately free (i.e. insert back into the space map)
1927 * everything we've allocated. This ensures that we don't get ENOSPC
1928 * errors during repeated suspend/resume cycles due to a flaky device.
1930 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1931 * the gang tree, we won't modify the block, so we can safely defer the free
1932 * (knowing that the block is still intact). If we *can* assemble the gang
1933 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1934 * each constituent bp and we can allocate a new block on the next sync pass.
1936 * In all cases, the gang tree allows complete recovery from partial failure.
1937 * ==========================================================================
1941 zio_gang_issue_func_done(zio_t *zio)
1943 abd_put(zio->io_abd);
1947 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
1953 return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset),
1954 BP_GET_PSIZE(bp), zio_gang_issue_func_done,
1955 NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1956 &pio->io_bookmark));
1960 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
1967 abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1968 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1969 gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL,
1970 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1973 * As we rewrite each gang header, the pipeline will compute
1974 * a new gang block header checksum for it; but no one will
1975 * compute a new data checksum, so we do that here. The one
1976 * exception is the gang leader: the pipeline already computed
1977 * its data checksum because that stage precedes gang assembly.
1978 * (Presently, nothing actually uses interior data checksums;
1979 * this is just good hygiene.)
1981 if (gn != pio->io_gang_leader->io_gang_tree) {
1982 abd_t *buf = abd_get_offset(data, offset);
1984 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1985 buf, BP_GET_PSIZE(bp));
1990 * If we are here to damage data for testing purposes,
1991 * leave the GBH alone so that we can detect the damage.
1993 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1994 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1996 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1997 abd_get_offset(data, offset), BP_GET_PSIZE(bp),
1998 zio_gang_issue_func_done, NULL, pio->io_priority,
1999 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2007 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2010 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
2011 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp),
2012 ZIO_GANG_CHILD_FLAGS(pio)));
2017 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2020 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
2021 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
2024 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
2033 static void zio_gang_tree_assemble_done(zio_t *zio);
2035 static zio_gang_node_t *
2036 zio_gang_node_alloc(zio_gang_node_t **gnpp)
2038 zio_gang_node_t *gn;
2040 ASSERT(*gnpp == NULL);
2042 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
2043 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
2050 zio_gang_node_free(zio_gang_node_t **gnpp)
2052 zio_gang_node_t *gn = *gnpp;
2054 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2055 ASSERT(gn->gn_child[g] == NULL);
2057 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2058 kmem_free(gn, sizeof (*gn));
2063 zio_gang_tree_free(zio_gang_node_t **gnpp)
2065 zio_gang_node_t *gn = *gnpp;
2070 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2071 zio_gang_tree_free(&gn->gn_child[g]);
2073 zio_gang_node_free(gnpp);
2077 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
2079 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
2080 abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2082 ASSERT(gio->io_gang_leader == gio);
2083 ASSERT(BP_IS_GANG(bp));
2085 zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2086 zio_gang_tree_assemble_done, gn, gio->io_priority,
2087 ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
2091 zio_gang_tree_assemble_done(zio_t *zio)
2093 zio_t *gio = zio->io_gang_leader;
2094 zio_gang_node_t *gn = zio->io_private;
2095 blkptr_t *bp = zio->io_bp;
2097 ASSERT(gio == zio_unique_parent(zio));
2098 ASSERT(zio->io_child_count == 0);
2103 /* this ABD was created from a linear buf in zio_gang_tree_assemble */
2104 if (BP_SHOULD_BYTESWAP(bp))
2105 byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size);
2107 ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh);
2108 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
2109 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2111 abd_put(zio->io_abd);
2113 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2114 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2115 if (!BP_IS_GANG(gbp))
2117 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
2122 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data,
2125 zio_t *gio = pio->io_gang_leader;
2128 ASSERT(BP_IS_GANG(bp) == !!gn);
2129 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
2130 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
2133 * If you're a gang header, your data is in gn->gn_gbh.
2134 * If you're a gang member, your data is in 'data' and gn == NULL.
2136 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset);
2139 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2141 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2142 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2143 if (BP_IS_HOLE(gbp))
2145 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data,
2147 offset += BP_GET_PSIZE(gbp);
2151 if (gn == gio->io_gang_tree && gio->io_abd != NULL)
2152 ASSERT3U(gio->io_size, ==, offset);
2159 zio_gang_assemble(zio_t *zio)
2161 blkptr_t *bp = zio->io_bp;
2163 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
2164 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2166 zio->io_gang_leader = zio;
2168 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
2170 return (ZIO_PIPELINE_CONTINUE);
2174 zio_gang_issue(zio_t *zio)
2176 blkptr_t *bp = zio->io_bp;
2178 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
2179 return (ZIO_PIPELINE_STOP);
2181 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
2182 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2184 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
2185 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd,
2188 zio_gang_tree_free(&zio->io_gang_tree);
2190 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2192 return (ZIO_PIPELINE_CONTINUE);
2196 zio_write_gang_member_ready(zio_t *zio)
2198 zio_t *pio = zio_unique_parent(zio);
2199 zio_t *gio = zio->io_gang_leader;
2200 dva_t *cdva = zio->io_bp->blk_dva;
2201 dva_t *pdva = pio->io_bp->blk_dva;
2204 if (BP_IS_HOLE(zio->io_bp))
2207 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
2209 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
2210 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
2211 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2212 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
2213 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
2215 mutex_enter(&pio->io_lock);
2216 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
2217 ASSERT(DVA_GET_GANG(&pdva[d]));
2218 asize = DVA_GET_ASIZE(&pdva[d]);
2219 asize += DVA_GET_ASIZE(&cdva[d]);
2220 DVA_SET_ASIZE(&pdva[d], asize);
2222 mutex_exit(&pio->io_lock);
2226 zio_write_gang_done(zio_t *zio)
2228 abd_put(zio->io_abd);
2232 zio_write_gang_block(zio_t *pio)
2234 spa_t *spa = pio->io_spa;
2235 metaslab_class_t *mc = spa_normal_class(spa);
2236 blkptr_t *bp = pio->io_bp;
2237 zio_t *gio = pio->io_gang_leader;
2239 zio_gang_node_t *gn, **gnpp;
2240 zio_gbh_phys_t *gbh;
2242 uint64_t txg = pio->io_txg;
2243 uint64_t resid = pio->io_size;
2245 int copies = gio->io_prop.zp_copies;
2246 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
2250 int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER;
2251 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2252 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2253 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2255 flags |= METASLAB_ASYNC_ALLOC;
2256 VERIFY(refcount_held(&mc->mc_alloc_slots, pio));
2259 * The logical zio has already placed a reservation for
2260 * 'copies' allocation slots but gang blocks may require
2261 * additional copies. These additional copies
2262 * (i.e. gbh_copies - copies) are guaranteed to succeed
2263 * since metaslab_class_throttle_reserve() always allows
2264 * additional reservations for gang blocks.
2266 VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies,
2270 error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE,
2271 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags,
2272 &pio->io_alloc_list, pio);
2274 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2275 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2276 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2279 * If we failed to allocate the gang block header then
2280 * we remove any additional allocation reservations that
2281 * we placed here. The original reservation will
2282 * be removed when the logical I/O goes to the ready
2285 metaslab_class_throttle_unreserve(mc,
2286 gbh_copies - copies, pio);
2288 pio->io_error = error;
2289 return (ZIO_PIPELINE_CONTINUE);
2293 gnpp = &gio->io_gang_tree;
2295 gnpp = pio->io_private;
2296 ASSERT(pio->io_ready == zio_write_gang_member_ready);
2299 gn = zio_gang_node_alloc(gnpp);
2301 bzero(gbh, SPA_GANGBLOCKSIZE);
2302 gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE);
2305 * Create the gang header.
2307 zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2308 zio_write_gang_done, NULL, pio->io_priority,
2309 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2312 * Create and nowait the gang children.
2314 for (int g = 0; resid != 0; resid -= lsize, g++) {
2315 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2317 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2319 zp.zp_checksum = gio->io_prop.zp_checksum;
2320 zp.zp_compress = ZIO_COMPRESS_OFF;
2321 zp.zp_type = DMU_OT_NONE;
2323 zp.zp_copies = gio->io_prop.zp_copies;
2324 zp.zp_dedup = B_FALSE;
2325 zp.zp_dedup_verify = B_FALSE;
2326 zp.zp_nopwrite = B_FALSE;
2328 zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2329 abd_get_offset(pio->io_abd, pio->io_size - resid), lsize,
2330 lsize, &zp, zio_write_gang_member_ready, NULL, NULL,
2331 zio_write_gang_done, &gn->gn_child[g], pio->io_priority,
2332 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2334 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2335 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2336 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2339 * Gang children won't throttle but we should
2340 * account for their work, so reserve an allocation
2341 * slot for them here.
2343 VERIFY(metaslab_class_throttle_reserve(mc,
2344 zp.zp_copies, cio, flags));
2350 * Set pio's pipeline to just wait for zio to finish.
2352 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2356 return (ZIO_PIPELINE_CONTINUE);
2360 * The zio_nop_write stage in the pipeline determines if allocating a
2361 * new bp is necessary. The nopwrite feature can handle writes in
2362 * either syncing or open context (i.e. zil writes) and as a result is
2363 * mutually exclusive with dedup.
2365 * By leveraging a cryptographically secure checksum, such as SHA256, we
2366 * can compare the checksums of the new data and the old to determine if
2367 * allocating a new block is required. Note that our requirements for
2368 * cryptographic strength are fairly weak: there can't be any accidental
2369 * hash collisions, but we don't need to be secure against intentional
2370 * (malicious) collisions. To trigger a nopwrite, you have to be able
2371 * to write the file to begin with, and triggering an incorrect (hash
2372 * collision) nopwrite is no worse than simply writing to the file.
2373 * That said, there are no known attacks against the checksum algorithms
2374 * used for nopwrite, assuming that the salt and the checksums
2375 * themselves remain secret.
2378 zio_nop_write(zio_t *zio)
2380 blkptr_t *bp = zio->io_bp;
2381 blkptr_t *bp_orig = &zio->io_bp_orig;
2382 zio_prop_t *zp = &zio->io_prop;
2384 ASSERT(BP_GET_LEVEL(bp) == 0);
2385 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2386 ASSERT(zp->zp_nopwrite);
2387 ASSERT(!zp->zp_dedup);
2388 ASSERT(zio->io_bp_override == NULL);
2389 ASSERT(IO_IS_ALLOCATING(zio));
2392 * Check to see if the original bp and the new bp have matching
2393 * characteristics (i.e. same checksum, compression algorithms, etc).
2394 * If they don't then just continue with the pipeline which will
2395 * allocate a new bp.
2397 if (BP_IS_HOLE(bp_orig) ||
2398 !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
2399 ZCHECKSUM_FLAG_NOPWRITE) ||
2400 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2401 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2402 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2403 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2404 return (ZIO_PIPELINE_CONTINUE);
2407 * If the checksums match then reset the pipeline so that we
2408 * avoid allocating a new bp and issuing any I/O.
2410 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2411 ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
2412 ZCHECKSUM_FLAG_NOPWRITE);
2413 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2414 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2415 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2416 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2417 sizeof (uint64_t)) == 0);
2420 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2421 zio->io_flags |= ZIO_FLAG_NOPWRITE;
2424 return (ZIO_PIPELINE_CONTINUE);
2428 * ==========================================================================
2430 * ==========================================================================
2433 zio_ddt_child_read_done(zio_t *zio)
2435 blkptr_t *bp = zio->io_bp;
2436 ddt_entry_t *dde = zio->io_private;
2438 zio_t *pio = zio_unique_parent(zio);
2440 mutex_enter(&pio->io_lock);
2441 ddp = ddt_phys_select(dde, bp);
2442 if (zio->io_error == 0)
2443 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2445 if (zio->io_error == 0 && dde->dde_repair_abd == NULL)
2446 dde->dde_repair_abd = zio->io_abd;
2448 abd_free(zio->io_abd);
2449 mutex_exit(&pio->io_lock);
2453 zio_ddt_read_start(zio_t *zio)
2455 blkptr_t *bp = zio->io_bp;
2457 ASSERT(BP_GET_DEDUP(bp));
2458 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2459 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2461 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2462 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2463 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2464 ddt_phys_t *ddp = dde->dde_phys;
2465 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2468 ASSERT(zio->io_vsd == NULL);
2471 if (ddp_self == NULL)
2472 return (ZIO_PIPELINE_CONTINUE);
2474 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2475 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2477 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2479 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2480 abd_alloc_for_io(zio->io_size, B_TRUE),
2481 zio->io_size, zio_ddt_child_read_done, dde,
2482 zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) |
2483 ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark));
2485 return (ZIO_PIPELINE_CONTINUE);
2488 zio_nowait(zio_read(zio, zio->io_spa, bp,
2489 zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority,
2490 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2492 return (ZIO_PIPELINE_CONTINUE);
2496 zio_ddt_read_done(zio_t *zio)
2498 blkptr_t *bp = zio->io_bp;
2500 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2501 return (ZIO_PIPELINE_STOP);
2503 ASSERT(BP_GET_DEDUP(bp));
2504 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2505 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2507 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2508 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2509 ddt_entry_t *dde = zio->io_vsd;
2511 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2512 return (ZIO_PIPELINE_CONTINUE);
2515 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2516 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2517 return (ZIO_PIPELINE_STOP);
2519 if (dde->dde_repair_abd != NULL) {
2520 abd_copy(zio->io_abd, dde->dde_repair_abd,
2522 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2524 ddt_repair_done(ddt, dde);
2528 ASSERT(zio->io_vsd == NULL);
2530 return (ZIO_PIPELINE_CONTINUE);
2534 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2536 spa_t *spa = zio->io_spa;
2537 boolean_t do_raw = (zio->io_flags & ZIO_FLAG_RAW);
2539 /* We should never get a raw, override zio */
2540 ASSERT(!(zio->io_bp_override && do_raw));
2543 * Note: we compare the original data, not the transformed data,
2544 * because when zio->io_bp is an override bp, we will not have
2545 * pushed the I/O transforms. That's an important optimization
2546 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2548 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2549 zio_t *lio = dde->dde_lead_zio[p];
2552 return (lio->io_orig_size != zio->io_orig_size ||
2553 abd_cmp(zio->io_orig_abd, lio->io_orig_abd,
2554 zio->io_orig_size) != 0);
2558 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2559 ddt_phys_t *ddp = &dde->dde_phys[p];
2561 if (ddp->ddp_phys_birth != 0) {
2562 arc_buf_t *abuf = NULL;
2563 arc_flags_t aflags = ARC_FLAG_WAIT;
2564 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE;
2565 blkptr_t blk = *zio->io_bp;
2568 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2573 * Intuitively, it would make more sense to compare
2574 * io_abd than io_orig_abd in the raw case since you
2575 * don't want to look at any transformations that have
2576 * happened to the data. However, for raw I/Os the
2577 * data will actually be the same in io_abd and
2578 * io_orig_abd, so all we have to do is issue this as
2582 zio_flags |= ZIO_FLAG_RAW;
2583 ASSERT3U(zio->io_size, ==, zio->io_orig_size);
2584 ASSERT0(abd_cmp(zio->io_abd, zio->io_orig_abd,
2586 ASSERT3P(zio->io_transform_stack, ==, NULL);
2589 error = arc_read(NULL, spa, &blk,
2590 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2591 zio_flags, &aflags, &zio->io_bookmark);
2594 if (arc_buf_size(abuf) != zio->io_orig_size ||
2595 abd_cmp_buf(zio->io_orig_abd, abuf->b_data,
2596 zio->io_orig_size) != 0)
2597 error = SET_ERROR(EEXIST);
2598 arc_buf_destroy(abuf, &abuf);
2602 return (error != 0);
2610 zio_ddt_child_write_ready(zio_t *zio)
2612 int p = zio->io_prop.zp_copies;
2613 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2614 ddt_entry_t *dde = zio->io_private;
2615 ddt_phys_t *ddp = &dde->dde_phys[p];
2623 ASSERT(dde->dde_lead_zio[p] == zio);
2625 ddt_phys_fill(ddp, zio->io_bp);
2627 zio_link_t *zl = NULL;
2628 while ((pio = zio_walk_parents(zio, &zl)) != NULL)
2629 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2635 zio_ddt_child_write_done(zio_t *zio)
2637 int p = zio->io_prop.zp_copies;
2638 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2639 ddt_entry_t *dde = zio->io_private;
2640 ddt_phys_t *ddp = &dde->dde_phys[p];
2644 ASSERT(ddp->ddp_refcnt == 0);
2645 ASSERT(dde->dde_lead_zio[p] == zio);
2646 dde->dde_lead_zio[p] = NULL;
2648 if (zio->io_error == 0) {
2649 zio_link_t *zl = NULL;
2650 while (zio_walk_parents(zio, &zl) != NULL)
2651 ddt_phys_addref(ddp);
2653 ddt_phys_clear(ddp);
2660 zio_ddt_ditto_write_done(zio_t *zio)
2662 int p = DDT_PHYS_DITTO;
2663 zio_prop_t *zp = &zio->io_prop;
2664 blkptr_t *bp = zio->io_bp;
2665 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2666 ddt_entry_t *dde = zio->io_private;
2667 ddt_phys_t *ddp = &dde->dde_phys[p];
2668 ddt_key_t *ddk = &dde->dde_key;
2672 ASSERT(ddp->ddp_refcnt == 0);
2673 ASSERT(dde->dde_lead_zio[p] == zio);
2674 dde->dde_lead_zio[p] = NULL;
2676 if (zio->io_error == 0) {
2677 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2678 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2679 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2680 if (ddp->ddp_phys_birth != 0)
2681 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2682 ddt_phys_fill(ddp, bp);
2689 zio_ddt_write(zio_t *zio)
2691 spa_t *spa = zio->io_spa;
2692 blkptr_t *bp = zio->io_bp;
2693 uint64_t txg = zio->io_txg;
2694 zio_prop_t *zp = &zio->io_prop;
2695 int p = zp->zp_copies;
2699 ddt_t *ddt = ddt_select(spa, bp);
2703 ASSERT(BP_GET_DEDUP(bp));
2704 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2705 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2706 ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW)));
2709 dde = ddt_lookup(ddt, bp, B_TRUE);
2710 ddp = &dde->dde_phys[p];
2712 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2714 * If we're using a weak checksum, upgrade to a strong checksum
2715 * and try again. If we're already using a strong checksum,
2716 * we can't resolve it, so just convert to an ordinary write.
2717 * (And automatically e-mail a paper to Nature?)
2719 if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
2720 ZCHECKSUM_FLAG_DEDUP)) {
2721 zp->zp_checksum = spa_dedup_checksum(spa);
2722 zio_pop_transforms(zio);
2723 zio->io_stage = ZIO_STAGE_OPEN;
2726 zp->zp_dedup = B_FALSE;
2727 BP_SET_DEDUP(bp, B_FALSE);
2729 ASSERT(!BP_GET_DEDUP(bp));
2730 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2732 return (ZIO_PIPELINE_CONTINUE);
2735 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2736 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2738 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2739 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2740 zio_prop_t czp = *zp;
2742 czp.zp_copies = ditto_copies;
2745 * If we arrived here with an override bp, we won't have run
2746 * the transform stack, so we won't have the data we need to
2747 * generate a child i/o. So, toss the override bp and restart.
2748 * This is safe, because using the override bp is just an
2749 * optimization; and it's rare, so the cost doesn't matter.
2751 if (zio->io_bp_override) {
2752 zio_pop_transforms(zio);
2753 zio->io_stage = ZIO_STAGE_OPEN;
2754 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2755 zio->io_bp_override = NULL;
2758 return (ZIO_PIPELINE_CONTINUE);
2761 dio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
2762 zio->io_orig_size, zio->io_orig_size, &czp, NULL, NULL,
2763 NULL, zio_ddt_ditto_write_done, dde, zio->io_priority,
2764 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2766 zio_push_transform(dio, zio->io_abd, zio->io_size, 0, NULL);
2767 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2770 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2771 if (ddp->ddp_phys_birth != 0)
2772 ddt_bp_fill(ddp, bp, txg);
2773 if (dde->dde_lead_zio[p] != NULL)
2774 zio_add_child(zio, dde->dde_lead_zio[p]);
2776 ddt_phys_addref(ddp);
2777 } else if (zio->io_bp_override) {
2778 ASSERT(bp->blk_birth == txg);
2779 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2780 ddt_phys_fill(ddp, bp);
2781 ddt_phys_addref(ddp);
2783 cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
2784 zio->io_orig_size, zio->io_orig_size, zp,
2785 zio_ddt_child_write_ready, NULL, NULL,
2786 zio_ddt_child_write_done, dde, zio->io_priority,
2787 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2789 zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL);
2790 dde->dde_lead_zio[p] = cio;
2800 return (ZIO_PIPELINE_CONTINUE);
2803 ddt_entry_t *freedde; /* for debugging */
2806 zio_ddt_free(zio_t *zio)
2808 spa_t *spa = zio->io_spa;
2809 blkptr_t *bp = zio->io_bp;
2810 ddt_t *ddt = ddt_select(spa, bp);
2814 ASSERT(BP_GET_DEDUP(bp));
2815 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2818 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2819 ddp = ddt_phys_select(dde, bp);
2820 ddt_phys_decref(ddp);
2823 return (ZIO_PIPELINE_CONTINUE);
2827 * ==========================================================================
2828 * Allocate and free blocks
2829 * ==========================================================================
2833 zio_io_to_allocate(spa_t *spa)
2837 ASSERT(MUTEX_HELD(&spa->spa_alloc_lock));
2839 zio = avl_first(&spa->spa_alloc_tree);
2843 ASSERT(IO_IS_ALLOCATING(zio));
2846 * Try to place a reservation for this zio. If we're unable to
2847 * reserve then we throttle.
2849 if (!metaslab_class_throttle_reserve(spa_normal_class(spa),
2850 zio->io_prop.zp_copies, zio, 0)) {
2854 avl_remove(&spa->spa_alloc_tree, zio);
2855 ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE);
2861 zio_dva_throttle(zio_t *zio)
2863 spa_t *spa = zio->io_spa;
2866 if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE ||
2867 !spa_normal_class(zio->io_spa)->mc_alloc_throttle_enabled ||
2868 zio->io_child_type == ZIO_CHILD_GANG ||
2869 zio->io_flags & ZIO_FLAG_NODATA) {
2870 return (ZIO_PIPELINE_CONTINUE);
2873 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2875 ASSERT3U(zio->io_queued_timestamp, >, 0);
2876 ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE);
2878 mutex_enter(&spa->spa_alloc_lock);
2880 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2881 avl_add(&spa->spa_alloc_tree, zio);
2883 nio = zio_io_to_allocate(zio->io_spa);
2884 mutex_exit(&spa->spa_alloc_lock);
2887 return (ZIO_PIPELINE_CONTINUE);
2890 ASSERT(nio->io_stage == ZIO_STAGE_DVA_THROTTLE);
2892 * We are passing control to a new zio so make sure that
2893 * it is processed by a different thread. We do this to
2894 * avoid stack overflows that can occur when parents are
2895 * throttled and children are making progress. We allow
2896 * it to go to the head of the taskq since it's already
2899 zio_taskq_dispatch(nio, ZIO_TASKQ_ISSUE, B_TRUE);
2901 return (ZIO_PIPELINE_STOP);
2905 zio_allocate_dispatch(spa_t *spa)
2909 mutex_enter(&spa->spa_alloc_lock);
2910 zio = zio_io_to_allocate(spa);
2911 mutex_exit(&spa->spa_alloc_lock);
2915 ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE);
2916 ASSERT0(zio->io_error);
2917 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE);
2921 zio_dva_allocate(zio_t *zio)
2923 spa_t *spa = zio->io_spa;
2924 metaslab_class_t *mc = spa_normal_class(spa);
2925 blkptr_t *bp = zio->io_bp;
2929 if (zio->io_gang_leader == NULL) {
2930 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2931 zio->io_gang_leader = zio;
2934 ASSERT(BP_IS_HOLE(bp));
2935 ASSERT0(BP_GET_NDVAS(bp));
2936 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2937 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2938 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2940 if (zio->io_flags & ZIO_FLAG_NODATA) {
2941 flags |= METASLAB_DONT_THROTTLE;
2943 if (zio->io_flags & ZIO_FLAG_GANG_CHILD) {
2944 flags |= METASLAB_GANG_CHILD;
2946 if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE) {
2947 flags |= METASLAB_ASYNC_ALLOC;
2950 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2951 zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
2952 &zio->io_alloc_list, zio);
2955 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2956 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2958 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2959 return (zio_write_gang_block(zio));
2960 zio->io_error = error;
2963 return (ZIO_PIPELINE_CONTINUE);
2967 zio_dva_free(zio_t *zio)
2969 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2971 return (ZIO_PIPELINE_CONTINUE);
2975 zio_dva_claim(zio_t *zio)
2979 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2981 zio->io_error = error;
2983 return (ZIO_PIPELINE_CONTINUE);
2987 * Undo an allocation. This is used by zio_done() when an I/O fails
2988 * and we want to give back the block we just allocated.
2989 * This handles both normal blocks and gang blocks.
2992 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2994 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2995 ASSERT(zio->io_bp_override == NULL);
2997 if (!BP_IS_HOLE(bp))
2998 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
3001 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
3002 zio_dva_unallocate(zio, gn->gn_child[g],
3003 &gn->gn_gbh->zg_blkptr[g]);
3009 * Try to allocate an intent log block. Return 0 on success, errno on failure.
3012 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
3013 uint64_t size, boolean_t *slog)
3016 zio_alloc_list_t io_alloc_list;
3018 ASSERT(txg > spa_syncing_txg(spa));
3020 metaslab_trace_init(&io_alloc_list);
3021 error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1,
3022 txg, old_bp, METASLAB_HINTBP_AVOID, &io_alloc_list, NULL);
3026 error = metaslab_alloc(spa, spa_normal_class(spa), size,
3027 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID,
3028 &io_alloc_list, NULL);
3032 metaslab_trace_fini(&io_alloc_list);
3035 BP_SET_LSIZE(new_bp, size);
3036 BP_SET_PSIZE(new_bp, size);
3037 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
3038 BP_SET_CHECKSUM(new_bp,
3039 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
3040 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
3041 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3042 BP_SET_LEVEL(new_bp, 0);
3043 BP_SET_DEDUP(new_bp, 0);
3044 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
3051 * Free an intent log block.
3054 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
3056 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
3057 ASSERT(!BP_IS_GANG(bp));
3059 zio_free(spa, txg, bp);
3063 * ==========================================================================
3064 * Read, write and delete to physical devices
3065 * ==========================================================================
3070 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3071 * stops after this stage and will resume upon I/O completion.
3072 * However, there are instances where the vdev layer may need to
3073 * continue the pipeline when an I/O was not issued. Since the I/O
3074 * that was sent to the vdev layer might be different than the one
3075 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3076 * force the underlying vdev layers to call either zio_execute() or
3077 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3080 zio_vdev_io_start(zio_t *zio)
3082 vdev_t *vd = zio->io_vd;
3084 spa_t *spa = zio->io_spa;
3087 ASSERT(zio->io_error == 0);
3088 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
3091 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3092 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
3095 * The mirror_ops handle multiple DVAs in a single BP.
3097 vdev_mirror_ops.vdev_op_io_start(zio);
3098 return (ZIO_PIPELINE_STOP);
3101 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE &&
3102 zio->io_priority == ZIO_PRIORITY_NOW) {
3103 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg);
3104 return (ZIO_PIPELINE_CONTINUE);
3107 ASSERT3P(zio->io_logical, !=, zio);
3110 * We keep track of time-sensitive I/Os so that the scan thread
3111 * can quickly react to certain workloads. In particular, we care
3112 * about non-scrubbing, top-level reads and writes with the following
3114 * - synchronous writes of user data to non-slog devices
3115 * - any reads of user data
3116 * When these conditions are met, adjust the timestamp of spa_last_io
3117 * which allows the scan thread to adjust its workload accordingly.
3119 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
3120 vd == vd->vdev_top && !vd->vdev_islog &&
3121 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
3122 zio->io_txg != spa_syncing_txg(spa)) {
3123 uint64_t old = spa->spa_last_io;
3124 uint64_t new = ddi_get_lbolt64();
3126 (void) atomic_cas_64(&spa->spa_last_io, old, new);
3129 align = 1ULL << vd->vdev_top->vdev_ashift;
3131 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
3132 P2PHASE(zio->io_size, align) != 0) {
3133 /* Transform logical writes to be a full physical block size. */
3134 uint64_t asize = P2ROUNDUP(zio->io_size, align);
3136 if (zio->io_type == ZIO_TYPE_READ ||
3137 zio->io_type == ZIO_TYPE_WRITE)
3138 abuf = abd_alloc_sametype(zio->io_abd, asize);
3139 ASSERT(vd == vd->vdev_top);
3140 if (zio->io_type == ZIO_TYPE_WRITE) {
3141 abd_copy(abuf, zio->io_abd, zio->io_size);
3142 abd_zero_off(abuf, zio->io_size, asize - zio->io_size);
3144 zio_push_transform(zio, abuf, asize, abuf ? asize : 0,
3149 * If this is not a physical io, make sure that it is properly aligned
3150 * before proceeding.
3152 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
3153 ASSERT0(P2PHASE(zio->io_offset, align));
3154 ASSERT0(P2PHASE(zio->io_size, align));
3157 * For the physical io we allow alignment
3158 * to a logical block size.
3160 uint64_t log_align =
3161 1ULL << vd->vdev_top->vdev_logical_ashift;
3162 ASSERT0(P2PHASE(zio->io_offset, log_align));
3163 ASSERT0(P2PHASE(zio->io_size, log_align));
3166 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa));
3169 * If this is a repair I/O, and there's no self-healing involved --
3170 * that is, we're just resilvering what we expect to resilver --
3171 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3172 * This prevents spurious resilvering with nested replication.
3173 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
3174 * A is out of date, we'll read from C+D, then use the data to
3175 * resilver A+B -- but we don't actually want to resilver B, just A.
3176 * The top-level mirror has no way to know this, so instead we just
3177 * discard unnecessary repairs as we work our way down the vdev tree.
3178 * The same logic applies to any form of nested replication:
3179 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
3181 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3182 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
3183 zio->io_txg != 0 && /* not a delegated i/o */
3184 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
3185 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3186 zio_vdev_io_bypass(zio);
3187 return (ZIO_PIPELINE_CONTINUE);
3190 if (vd->vdev_ops->vdev_op_leaf) {
3191 switch (zio->io_type) {
3193 if (vdev_cache_read(zio))
3194 return (ZIO_PIPELINE_CONTINUE);
3196 case ZIO_TYPE_WRITE:
3198 if ((zio = vdev_queue_io(zio)) == NULL)
3199 return (ZIO_PIPELINE_STOP);
3201 if (!vdev_accessible(vd, zio)) {
3202 zio->io_error = SET_ERROR(ENXIO);
3204 return (ZIO_PIPELINE_STOP);
3209 * Note that we ignore repair writes for TRIM because they can
3210 * conflict with normal writes. This isn't an issue because, by
3211 * definition, we only repair blocks that aren't freed.
3213 if (zio->io_type == ZIO_TYPE_WRITE &&
3214 !(zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3215 !trim_map_write_start(zio))
3216 return (ZIO_PIPELINE_STOP);
3219 vd->vdev_ops->vdev_op_io_start(zio);
3220 return (ZIO_PIPELINE_STOP);
3224 zio_vdev_io_done(zio_t *zio)
3226 vdev_t *vd = zio->io_vd;
3227 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
3228 boolean_t unexpected_error = B_FALSE;
3230 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
3231 return (ZIO_PIPELINE_STOP);
3233 ASSERT(zio->io_type == ZIO_TYPE_READ ||
3234 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE);
3236 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3237 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE ||
3238 zio->io_type == ZIO_TYPE_FREE)) {
3240 if (zio->io_type == ZIO_TYPE_WRITE &&
3241 !(zio->io_flags & ZIO_FLAG_IO_REPAIR))
3242 trim_map_write_done(zio);
3244 vdev_queue_io_done(zio);
3246 if (zio->io_type == ZIO_TYPE_WRITE)
3247 vdev_cache_write(zio);
3249 if (zio_injection_enabled && zio->io_error == 0)
3250 zio->io_error = zio_handle_device_injection(vd,
3253 if (zio_injection_enabled && zio->io_error == 0)
3254 zio->io_error = zio_handle_label_injection(zio, EIO);
3256 if (zio->io_error) {
3257 if (zio->io_error == ENOTSUP &&
3258 zio->io_type == ZIO_TYPE_FREE) {
3259 /* Not all devices support TRIM. */
3260 } else if (!vdev_accessible(vd, zio)) {
3261 zio->io_error = SET_ERROR(ENXIO);
3263 unexpected_error = B_TRUE;
3268 ops->vdev_op_io_done(zio);
3270 if (unexpected_error)
3271 VERIFY(vdev_probe(vd, zio) == NULL);
3273 return (ZIO_PIPELINE_CONTINUE);
3277 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3278 * disk, and use that to finish the checksum ereport later.
3281 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
3282 const void *good_buf)
3284 /* no processing needed */
3285 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
3290 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
3292 void *buf = zio_buf_alloc(zio->io_size);
3294 abd_copy_to_buf(buf, zio->io_abd, zio->io_size);
3296 zcr->zcr_cbinfo = zio->io_size;
3297 zcr->zcr_cbdata = buf;
3298 zcr->zcr_finish = zio_vsd_default_cksum_finish;
3299 zcr->zcr_free = zio_buf_free;
3303 zio_vdev_io_assess(zio_t *zio)
3305 vdev_t *vd = zio->io_vd;
3307 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
3308 return (ZIO_PIPELINE_STOP);
3310 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3311 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
3313 if (zio->io_vsd != NULL) {
3314 zio->io_vsd_ops->vsd_free(zio);
3318 if (zio_injection_enabled && zio->io_error == 0)
3319 zio->io_error = zio_handle_fault_injection(zio, EIO);
3321 if (zio->io_type == ZIO_TYPE_FREE &&
3322 zio->io_priority != ZIO_PRIORITY_NOW) {
3323 switch (zio->io_error) {
3325 ZIO_TRIM_STAT_INCR(bytes, zio->io_size);
3326 ZIO_TRIM_STAT_BUMP(success);
3329 ZIO_TRIM_STAT_BUMP(unsupported);
3332 ZIO_TRIM_STAT_BUMP(failed);
3338 * If the I/O failed, determine whether we should attempt to retry it.
3340 * On retry, we cut in line in the issue queue, since we don't want
3341 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3343 if (zio->io_error && vd == NULL &&
3344 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
3345 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
3346 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
3348 zio->io_flags |= ZIO_FLAG_IO_RETRY |
3349 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
3350 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
3351 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
3352 zio_requeue_io_start_cut_in_line);
3353 return (ZIO_PIPELINE_STOP);
3357 * If we got an error on a leaf device, convert it to ENXIO
3358 * if the device is not accessible at all.
3360 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3361 !vdev_accessible(vd, zio))
3362 zio->io_error = SET_ERROR(ENXIO);
3365 * If we can't write to an interior vdev (mirror or RAID-Z),
3366 * set vdev_cant_write so that we stop trying to allocate from it.
3368 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
3369 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
3370 vd->vdev_cant_write = B_TRUE;
3374 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
3375 * attempts will ever succeed. In this case we set a persistent bit so
3376 * that we don't bother with it in the future.
3378 if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) &&
3379 zio->io_type == ZIO_TYPE_IOCTL &&
3380 zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL)
3381 vd->vdev_nowritecache = B_TRUE;
3384 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3386 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3387 zio->io_physdone != NULL) {
3388 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
3389 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
3390 zio->io_physdone(zio->io_logical);
3393 return (ZIO_PIPELINE_CONTINUE);
3397 zio_vdev_io_reissue(zio_t *zio)
3399 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3400 ASSERT(zio->io_error == 0);
3402 zio->io_stage >>= 1;
3406 zio_vdev_io_redone(zio_t *zio)
3408 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
3410 zio->io_stage >>= 1;
3414 zio_vdev_io_bypass(zio_t *zio)
3416 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3417 ASSERT(zio->io_error == 0);
3419 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
3420 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
3424 * ==========================================================================
3425 * Generate and verify checksums
3426 * ==========================================================================
3429 zio_checksum_generate(zio_t *zio)
3431 blkptr_t *bp = zio->io_bp;
3432 enum zio_checksum checksum;
3436 * This is zio_write_phys().
3437 * We're either generating a label checksum, or none at all.
3439 checksum = zio->io_prop.zp_checksum;
3441 if (checksum == ZIO_CHECKSUM_OFF)
3442 return (ZIO_PIPELINE_CONTINUE);
3444 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
3446 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
3447 ASSERT(!IO_IS_ALLOCATING(zio));
3448 checksum = ZIO_CHECKSUM_GANG_HEADER;
3450 checksum = BP_GET_CHECKSUM(bp);
3454 zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size);
3456 return (ZIO_PIPELINE_CONTINUE);
3460 zio_checksum_verify(zio_t *zio)
3462 zio_bad_cksum_t info;
3463 blkptr_t *bp = zio->io_bp;
3466 ASSERT(zio->io_vd != NULL);
3470 * This is zio_read_phys().
3471 * We're either verifying a label checksum, or nothing at all.
3473 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
3474 return (ZIO_PIPELINE_CONTINUE);
3476 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
3479 if ((error = zio_checksum_error(zio, &info)) != 0) {
3480 zio->io_error = error;
3481 if (error == ECKSUM &&
3482 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
3483 zfs_ereport_start_checksum(zio->io_spa,
3484 zio->io_vd, zio, zio->io_offset,
3485 zio->io_size, NULL, &info);
3489 return (ZIO_PIPELINE_CONTINUE);
3493 * Called by RAID-Z to ensure we don't compute the checksum twice.
3496 zio_checksum_verified(zio_t *zio)
3498 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
3502 * ==========================================================================
3503 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
3504 * An error of 0 indicates success. ENXIO indicates whole-device failure,
3505 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
3506 * indicate errors that are specific to one I/O, and most likely permanent.
3507 * Any other error is presumed to be worse because we weren't expecting it.
3508 * ==========================================================================
3511 zio_worst_error(int e1, int e2)
3513 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
3516 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
3517 if (e1 == zio_error_rank[r1])
3520 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
3521 if (e2 == zio_error_rank[r2])
3524 return (r1 > r2 ? e1 : e2);
3528 * ==========================================================================
3530 * ==========================================================================
3533 zio_ready(zio_t *zio)
3535 blkptr_t *bp = zio->io_bp;
3536 zio_t *pio, *pio_next;
3537 zio_link_t *zl = NULL;
3539 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
3540 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
3541 return (ZIO_PIPELINE_STOP);
3543 if (zio->io_ready) {
3544 ASSERT(IO_IS_ALLOCATING(zio));
3545 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
3546 (zio->io_flags & ZIO_FLAG_NOPWRITE));
3547 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
3552 if (bp != NULL && bp != &zio->io_bp_copy)
3553 zio->io_bp_copy = *bp;
3555 if (zio->io_error != 0) {
3556 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3558 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3559 ASSERT(IO_IS_ALLOCATING(zio));
3560 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3562 * We were unable to allocate anything, unreserve and
3563 * issue the next I/O to allocate.
3565 metaslab_class_throttle_unreserve(
3566 spa_normal_class(zio->io_spa),
3567 zio->io_prop.zp_copies, zio);
3568 zio_allocate_dispatch(zio->io_spa);
3572 mutex_enter(&zio->io_lock);
3573 zio->io_state[ZIO_WAIT_READY] = 1;
3574 pio = zio_walk_parents(zio, &zl);
3575 mutex_exit(&zio->io_lock);
3578 * As we notify zio's parents, new parents could be added.
3579 * New parents go to the head of zio's io_parent_list, however,
3580 * so we will (correctly) not notify them. The remainder of zio's
3581 * io_parent_list, from 'pio_next' onward, cannot change because
3582 * all parents must wait for us to be done before they can be done.
3584 for (; pio != NULL; pio = pio_next) {
3585 pio_next = zio_walk_parents(zio, &zl);
3586 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
3589 if (zio->io_flags & ZIO_FLAG_NODATA) {
3590 if (BP_IS_GANG(bp)) {
3591 zio->io_flags &= ~ZIO_FLAG_NODATA;
3593 ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE);
3594 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
3598 if (zio_injection_enabled &&
3599 zio->io_spa->spa_syncing_txg == zio->io_txg)
3600 zio_handle_ignored_writes(zio);
3602 return (ZIO_PIPELINE_CONTINUE);
3606 * Update the allocation throttle accounting.
3609 zio_dva_throttle_done(zio_t *zio)
3611 zio_t *lio = zio->io_logical;
3612 zio_t *pio = zio_unique_parent(zio);
3613 vdev_t *vd = zio->io_vd;
3614 int flags = METASLAB_ASYNC_ALLOC;
3616 ASSERT3P(zio->io_bp, !=, NULL);
3617 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
3618 ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE);
3619 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
3621 ASSERT3P(vd, ==, vd->vdev_top);
3622 ASSERT(!(zio->io_flags & (ZIO_FLAG_IO_REPAIR | ZIO_FLAG_IO_RETRY)));
3623 ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING);
3624 ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE));
3625 ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA));
3628 * Parents of gang children can have two flavors -- ones that
3629 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
3630 * and ones that allocated the constituent blocks. The allocation
3631 * throttle needs to know the allocating parent zio so we must find
3634 if (pio->io_child_type == ZIO_CHILD_GANG) {
3636 * If our parent is a rewrite gang child then our grandparent
3637 * would have been the one that performed the allocation.
3639 if (pio->io_flags & ZIO_FLAG_IO_REWRITE)
3640 pio = zio_unique_parent(pio);
3641 flags |= METASLAB_GANG_CHILD;
3644 ASSERT(IO_IS_ALLOCATING(pio));
3645 ASSERT3P(zio, !=, zio->io_logical);
3646 ASSERT(zio->io_logical != NULL);
3647 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
3648 ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE);
3650 mutex_enter(&pio->io_lock);
3651 metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags);
3652 mutex_exit(&pio->io_lock);
3654 metaslab_class_throttle_unreserve(spa_normal_class(zio->io_spa),
3658 * Call into the pipeline to see if there is more work that
3659 * needs to be done. If there is work to be done it will be
3660 * dispatched to another taskq thread.
3662 zio_allocate_dispatch(zio->io_spa);
3666 zio_done(zio_t *zio)
3668 spa_t *spa = zio->io_spa;
3669 zio_t *lio = zio->io_logical;
3670 blkptr_t *bp = zio->io_bp;
3671 vdev_t *vd = zio->io_vd;
3672 uint64_t psize = zio->io_size;
3673 zio_t *pio, *pio_next;
3674 metaslab_class_t *mc = spa_normal_class(spa);
3675 zio_link_t *zl = NULL;
3678 * If our children haven't all completed,
3679 * wait for them and then repeat this pipeline stage.
3681 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
3682 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
3683 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
3684 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
3685 return (ZIO_PIPELINE_STOP);
3688 * If the allocation throttle is enabled, then update the accounting.
3689 * We only track child I/Os that are part of an allocating async
3690 * write. We must do this since the allocation is performed
3691 * by the logical I/O but the actual write is done by child I/Os.
3693 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING &&
3694 zio->io_child_type == ZIO_CHILD_VDEV) {
3695 ASSERT(mc->mc_alloc_throttle_enabled);
3696 zio_dva_throttle_done(zio);
3700 * If the allocation throttle is enabled, verify that
3701 * we have decremented the refcounts for every I/O that was throttled.
3703 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3704 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3705 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3707 metaslab_group_alloc_verify(spa, zio->io_bp, zio);
3708 VERIFY(refcount_not_held(&mc->mc_alloc_slots, zio));
3711 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3712 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3713 ASSERT(zio->io_children[c][w] == 0);
3715 if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
3716 ASSERT(bp->blk_pad[0] == 0);
3717 ASSERT(bp->blk_pad[1] == 0);
3718 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3719 (bp == zio_unique_parent(zio)->io_bp));
3720 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3721 zio->io_bp_override == NULL &&
3722 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3723 ASSERT(!BP_SHOULD_BYTESWAP(bp));
3724 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3725 ASSERT(BP_COUNT_GANG(bp) == 0 ||
3726 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3728 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3729 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3733 * If there were child vdev/gang/ddt errors, they apply to us now.
3735 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3736 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3737 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3740 * If the I/O on the transformed data was successful, generate any
3741 * checksum reports now while we still have the transformed data.
3743 if (zio->io_error == 0) {
3744 while (zio->io_cksum_report != NULL) {
3745 zio_cksum_report_t *zcr = zio->io_cksum_report;
3746 uint64_t align = zcr->zcr_align;
3747 uint64_t asize = P2ROUNDUP(psize, align);
3749 abd_t *adata = zio->io_abd;
3751 if (asize != psize) {
3752 adata = abd_alloc_linear(asize, B_TRUE);
3753 abd_copy(adata, zio->io_abd, psize);
3754 abd_zero_off(adata, psize, asize - psize);
3758 abuf = abd_borrow_buf_copy(adata, asize);
3760 zio->io_cksum_report = zcr->zcr_next;
3761 zcr->zcr_next = NULL;
3762 zcr->zcr_finish(zcr, abuf);
3763 zfs_ereport_free_checksum(zcr);
3766 abd_return_buf(adata, abuf, asize);
3773 zio_pop_transforms(zio); /* note: may set zio->io_error */
3775 vdev_stat_update(zio, psize);
3777 if (zio->io_error) {
3779 * If this I/O is attached to a particular vdev,
3780 * generate an error message describing the I/O failure
3781 * at the block level. We ignore these errors if the
3782 * device is currently unavailable.
3784 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3785 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3787 if ((zio->io_error == EIO || !(zio->io_flags &
3788 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3791 * For logical I/O requests, tell the SPA to log the
3792 * error and generate a logical data ereport.
3794 spa_log_error(spa, zio);
3795 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3800 if (zio->io_error && zio == lio) {
3802 * Determine whether zio should be reexecuted. This will
3803 * propagate all the way to the root via zio_notify_parent().
3805 ASSERT(vd == NULL && bp != NULL);
3806 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3808 if (IO_IS_ALLOCATING(zio) &&
3809 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3810 if (zio->io_error != ENOSPC)
3811 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3813 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3816 if ((zio->io_type == ZIO_TYPE_READ ||
3817 zio->io_type == ZIO_TYPE_FREE) &&
3818 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3819 zio->io_error == ENXIO &&
3820 spa_load_state(spa) == SPA_LOAD_NONE &&
3821 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3822 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3824 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3825 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3828 * Here is a possibly good place to attempt to do
3829 * either combinatorial reconstruction or error correction
3830 * based on checksums. It also might be a good place
3831 * to send out preliminary ereports before we suspend
3837 * If there were logical child errors, they apply to us now.
3838 * We defer this until now to avoid conflating logical child
3839 * errors with errors that happened to the zio itself when
3840 * updating vdev stats and reporting FMA events above.
3842 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3844 if ((zio->io_error || zio->io_reexecute) &&
3845 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3846 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3847 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3849 zio_gang_tree_free(&zio->io_gang_tree);
3852 * Godfather I/Os should never suspend.
3854 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3855 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3856 zio->io_reexecute = 0;
3858 if (zio->io_reexecute) {
3860 * This is a logical I/O that wants to reexecute.
3862 * Reexecute is top-down. When an i/o fails, if it's not
3863 * the root, it simply notifies its parent and sticks around.
3864 * The parent, seeing that it still has children in zio_done(),
3865 * does the same. This percolates all the way up to the root.
3866 * The root i/o will reexecute or suspend the entire tree.
3868 * This approach ensures that zio_reexecute() honors
3869 * all the original i/o dependency relationships, e.g.
3870 * parents not executing until children are ready.
3872 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3874 zio->io_gang_leader = NULL;
3876 mutex_enter(&zio->io_lock);
3877 zio->io_state[ZIO_WAIT_DONE] = 1;
3878 mutex_exit(&zio->io_lock);
3881 * "The Godfather" I/O monitors its children but is
3882 * not a true parent to them. It will track them through
3883 * the pipeline but severs its ties whenever they get into
3884 * trouble (e.g. suspended). This allows "The Godfather"
3885 * I/O to return status without blocking.
3888 for (pio = zio_walk_parents(zio, &zl); pio != NULL;
3890 zio_link_t *remove_zl = zl;
3891 pio_next = zio_walk_parents(zio, &zl);
3893 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3894 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3895 zio_remove_child(pio, zio, remove_zl);
3896 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3900 if ((pio = zio_unique_parent(zio)) != NULL) {
3902 * We're not a root i/o, so there's nothing to do
3903 * but notify our parent. Don't propagate errors
3904 * upward since we haven't permanently failed yet.
3906 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3907 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3908 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3909 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3911 * We'd fail again if we reexecuted now, so suspend
3912 * until conditions improve (e.g. device comes online).
3914 zio_suspend(spa, zio);
3917 * Reexecution is potentially a huge amount of work.
3918 * Hand it off to the otherwise-unused claim taskq.
3920 #if defined(illumos) || !defined(_KERNEL)
3921 ASSERT(zio->io_tqent.tqent_next == NULL);
3923 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
3925 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3926 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3929 return (ZIO_PIPELINE_STOP);
3932 ASSERT(zio->io_child_count == 0);
3933 ASSERT(zio->io_reexecute == 0);
3934 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3937 * Report any checksum errors, since the I/O is complete.
3939 while (zio->io_cksum_report != NULL) {
3940 zio_cksum_report_t *zcr = zio->io_cksum_report;
3941 zio->io_cksum_report = zcr->zcr_next;
3942 zcr->zcr_next = NULL;
3943 zcr->zcr_finish(zcr, NULL);
3944 zfs_ereport_free_checksum(zcr);
3948 * It is the responsibility of the done callback to ensure that this
3949 * particular zio is no longer discoverable for adoption, and as
3950 * such, cannot acquire any new parents.
3955 mutex_enter(&zio->io_lock);
3956 zio->io_state[ZIO_WAIT_DONE] = 1;
3957 mutex_exit(&zio->io_lock);
3960 for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) {
3961 zio_link_t *remove_zl = zl;
3962 pio_next = zio_walk_parents(zio, &zl);
3963 zio_remove_child(pio, zio, remove_zl);
3964 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3967 if (zio->io_waiter != NULL) {
3968 mutex_enter(&zio->io_lock);
3969 zio->io_executor = NULL;
3970 cv_broadcast(&zio->io_cv);
3971 mutex_exit(&zio->io_lock);
3976 return (ZIO_PIPELINE_STOP);
3980 * ==========================================================================
3981 * I/O pipeline definition
3982 * ==========================================================================
3984 static zio_pipe_stage_t *zio_pipeline[] = {
3991 zio_checksum_generate,
4007 zio_checksum_verify,
4015 * Compare two zbookmark_phys_t's to see which we would reach first in a
4016 * pre-order traversal of the object tree.
4018 * This is simple in every case aside from the meta-dnode object. For all other
4019 * objects, we traverse them in order (object 1 before object 2, and so on).
4020 * However, all of these objects are traversed while traversing object 0, since
4021 * the data it points to is the list of objects. Thus, we need to convert to a
4022 * canonical representation so we can compare meta-dnode bookmarks to
4023 * non-meta-dnode bookmarks.
4025 * We do this by calculating "equivalents" for each field of the zbookmark.
4026 * zbookmarks outside of the meta-dnode use their own object and level, and
4027 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
4028 * blocks this bookmark refers to) by multiplying their blkid by their span
4029 * (the number of L0 blocks contained within one block at their level).
4030 * zbookmarks inside the meta-dnode calculate their object equivalent
4031 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
4032 * level + 1<<31 (any value larger than a level could ever be) for their level.
4033 * This causes them to always compare before a bookmark in their object
4034 * equivalent, compare appropriately to bookmarks in other objects, and to
4035 * compare appropriately to other bookmarks in the meta-dnode.
4038 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
4039 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
4042 * These variables represent the "equivalent" values for the zbookmark,
4043 * after converting zbookmarks inside the meta dnode to their
4044 * normal-object equivalents.
4046 uint64_t zb1obj, zb2obj;
4047 uint64_t zb1L0, zb2L0;
4048 uint64_t zb1level, zb2level;
4050 if (zb1->zb_object == zb2->zb_object &&
4051 zb1->zb_level == zb2->zb_level &&
4052 zb1->zb_blkid == zb2->zb_blkid)
4056 * BP_SPANB calculates the span in blocks.
4058 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
4059 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
4061 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
4062 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4064 zb1level = zb1->zb_level + COMPARE_META_LEVEL;
4066 zb1obj = zb1->zb_object;
4067 zb1level = zb1->zb_level;
4070 if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
4071 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4073 zb2level = zb2->zb_level + COMPARE_META_LEVEL;
4075 zb2obj = zb2->zb_object;
4076 zb2level = zb2->zb_level;
4079 /* Now that we have a canonical representation, do the comparison. */
4080 if (zb1obj != zb2obj)
4081 return (zb1obj < zb2obj ? -1 : 1);
4082 else if (zb1L0 != zb2L0)
4083 return (zb1L0 < zb2L0 ? -1 : 1);
4084 else if (zb1level != zb2level)
4085 return (zb1level > zb2level ? -1 : 1);
4087 * This can (theoretically) happen if the bookmarks have the same object
4088 * and level, but different blkids, if the block sizes are not the same.
4089 * There is presently no way to change the indirect block sizes
4095 * This function checks the following: given that last_block is the place that
4096 * our traversal stopped last time, does that guarantee that we've visited
4097 * every node under subtree_root? Therefore, we can't just use the raw output
4098 * of zbookmark_compare. We have to pass in a modified version of
4099 * subtree_root; by incrementing the block id, and then checking whether
4100 * last_block is before or equal to that, we can tell whether or not having
4101 * visited last_block implies that all of subtree_root's children have been
4105 zbookmark_subtree_completed(const dnode_phys_t *dnp,
4106 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
4108 zbookmark_phys_t mod_zb = *subtree_root;
4110 ASSERT(last_block->zb_level == 0);
4112 /* The objset_phys_t isn't before anything. */
4117 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
4118 * data block size in sectors, because that variable is only used if
4119 * the bookmark refers to a block in the meta-dnode. Since we don't
4120 * know without examining it what object it refers to, and there's no
4121 * harm in passing in this value in other cases, we always pass it in.
4123 * We pass in 0 for the indirect block size shift because zb2 must be
4124 * level 0. The indirect block size is only used to calculate the span
4125 * of the bookmark, but since the bookmark must be level 0, the span is
4126 * always 1, so the math works out.
4128 * If you make changes to how the zbookmark_compare code works, be sure
4129 * to make sure that this code still works afterwards.
4131 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
4132 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,