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, 2016 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>
46 SYSCTL_DECL(_vfs_zfs);
47 SYSCTL_NODE(_vfs_zfs, OID_AUTO, zio, CTLFLAG_RW, 0, "ZFS ZIO");
48 #if defined(__amd64__)
49 static int zio_use_uma = 1;
51 static int zio_use_uma = 0;
53 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, use_uma, CTLFLAG_RDTUN, &zio_use_uma, 0,
54 "Use uma(9) for ZIO allocations");
55 static int zio_exclude_metadata = 0;
56 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, exclude_metadata, CTLFLAG_RDTUN, &zio_exclude_metadata, 0,
57 "Exclude metadata buffers from dumps as well");
59 zio_trim_stats_t zio_trim_stats = {
60 { "bytes", KSTAT_DATA_UINT64,
61 "Number of bytes successfully TRIMmed" },
62 { "success", KSTAT_DATA_UINT64,
63 "Number of successful TRIM requests" },
64 { "unsupported", KSTAT_DATA_UINT64,
65 "Number of TRIM requests that failed because TRIM is not supported" },
66 { "failed", KSTAT_DATA_UINT64,
67 "Number of TRIM requests that failed for reasons other than not supported" },
70 static kstat_t *zio_trim_ksp;
73 * ==========================================================================
74 * I/O type descriptions
75 * ==========================================================================
77 const char *zio_type_name[ZIO_TYPES] = {
78 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
82 boolean_t zio_dva_throttle_enabled = B_TRUE;
83 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, dva_throttle_enabled, CTLFLAG_RDTUN,
84 &zio_dva_throttle_enabled, 0, "");
87 * ==========================================================================
89 * ==========================================================================
91 kmem_cache_t *zio_cache;
92 kmem_cache_t *zio_link_cache;
93 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
94 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
97 extern vmem_t *zio_alloc_arena;
100 #define ZIO_PIPELINE_CONTINUE 0x100
101 #define ZIO_PIPELINE_STOP 0x101
103 #define BP_SPANB(indblkshift, level) \
104 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
105 #define COMPARE_META_LEVEL 0x80000000ul
107 * The following actions directly effect the spa's sync-to-convergence logic.
108 * The values below define the sync pass when we start performing the action.
109 * Care should be taken when changing these values as they directly impact
110 * spa_sync() performance. Tuning these values may introduce subtle performance
111 * pathologies and should only be done in the context of performance analysis.
112 * These tunables will eventually be removed and replaced with #defines once
113 * enough analysis has been done to determine optimal values.
115 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
116 * regular blocks are not deferred.
118 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
119 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_deferred_free, CTLFLAG_RDTUN,
120 &zfs_sync_pass_deferred_free, 0, "defer frees starting in this pass");
121 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
122 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_dont_compress, CTLFLAG_RDTUN,
123 &zfs_sync_pass_dont_compress, 0, "don't compress starting in this pass");
124 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
125 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_rewrite, CTLFLAG_RDTUN,
126 &zfs_sync_pass_rewrite, 0, "rewrite new bps starting in this pass");
129 * An allocating zio is one that either currently has the DVA allocate
130 * stage set or will have it later in its lifetime.
132 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
134 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
138 int zio_buf_debug_limit = 16384;
140 int zio_buf_debug_limit = 0;
144 static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t);
150 zio_cache = kmem_cache_create("zio_cache",
151 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
152 zio_link_cache = kmem_cache_create("zio_link_cache",
153 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
158 * For small buffers, we want a cache for each multiple of
159 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
160 * for each quarter-power of 2.
162 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
163 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
166 int cflags = zio_exclude_metadata ? KMC_NODEBUG : 0;
174 * If we are using watchpoints, put each buffer on its own page,
175 * to eliminate the performance overhead of trapping to the
176 * kernel when modifying a non-watched buffer that shares the
177 * page with a watched buffer.
179 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
183 if (size <= 4 * SPA_MINBLOCKSIZE) {
184 align = SPA_MINBLOCKSIZE;
185 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
186 align = MIN(p2 >> 2, PAGESIZE);
191 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
192 zio_buf_cache[c] = kmem_cache_create(name, size,
193 align, NULL, NULL, NULL, NULL, NULL, cflags);
196 * Since zio_data bufs do not appear in crash dumps, we
197 * pass KMC_NOTOUCH so that no allocator metadata is
198 * stored with the buffers.
200 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
201 zio_data_buf_cache[c] = kmem_cache_create(name, size,
202 align, NULL, NULL, NULL, NULL, NULL,
203 cflags | KMC_NOTOUCH | KMC_NODEBUG);
208 ASSERT(zio_buf_cache[c] != NULL);
209 if (zio_buf_cache[c - 1] == NULL)
210 zio_buf_cache[c - 1] = zio_buf_cache[c];
212 ASSERT(zio_data_buf_cache[c] != NULL);
213 if (zio_data_buf_cache[c - 1] == NULL)
214 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
220 zio_trim_ksp = kstat_create("zfs", 0, "zio_trim", "misc",
222 sizeof(zio_trim_stats) / sizeof(kstat_named_t),
225 if (zio_trim_ksp != NULL) {
226 zio_trim_ksp->ks_data = &zio_trim_stats;
227 kstat_install(zio_trim_ksp);
235 kmem_cache_t *last_cache = NULL;
236 kmem_cache_t *last_data_cache = NULL;
238 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
239 if (zio_buf_cache[c] != last_cache) {
240 last_cache = zio_buf_cache[c];
241 kmem_cache_destroy(zio_buf_cache[c]);
243 zio_buf_cache[c] = NULL;
245 if (zio_data_buf_cache[c] != last_data_cache) {
246 last_data_cache = zio_data_buf_cache[c];
247 kmem_cache_destroy(zio_data_buf_cache[c]);
249 zio_data_buf_cache[c] = NULL;
252 kmem_cache_destroy(zio_link_cache);
253 kmem_cache_destroy(zio_cache);
257 if (zio_trim_ksp != NULL) {
258 kstat_delete(zio_trim_ksp);
264 * ==========================================================================
265 * Allocate and free I/O buffers
266 * ==========================================================================
270 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
271 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
272 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
273 * excess / transient data in-core during a crashdump.
276 zio_buf_alloc_impl(size_t size, boolean_t canwait)
278 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
279 int flags = zio_exclude_metadata ? KM_NODEBUG : 0;
281 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
284 return (kmem_cache_alloc(zio_buf_cache[c],
285 canwait ? KM_PUSHPAGE : KM_NOSLEEP));
287 return (kmem_alloc(size,
288 (canwait ? KM_SLEEP : KM_NOSLEEP) | flags));
293 zio_buf_alloc(size_t size)
295 return (zio_buf_alloc_impl(size, B_TRUE));
299 zio_buf_alloc_nowait(size_t size)
301 return (zio_buf_alloc_impl(size, B_FALSE));
305 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
306 * crashdump if the kernel panics. This exists so that we will limit the amount
307 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
308 * of kernel heap dumped to disk when the kernel panics)
311 zio_data_buf_alloc(size_t size)
313 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
315 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
318 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
320 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG));
324 zio_buf_free(void *buf, size_t size)
326 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
328 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
331 kmem_cache_free(zio_buf_cache[c], buf);
333 kmem_free(buf, size);
337 zio_data_buf_free(void *buf, size_t size)
339 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
341 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
344 kmem_cache_free(zio_data_buf_cache[c], buf);
346 kmem_free(buf, size);
350 * ==========================================================================
351 * Push and pop I/O transform buffers
352 * ==========================================================================
355 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
356 zio_transform_func_t *transform)
358 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
360 zt->zt_orig_data = zio->io_data;
361 zt->zt_orig_size = zio->io_size;
362 zt->zt_bufsize = bufsize;
363 zt->zt_transform = transform;
365 zt->zt_next = zio->io_transform_stack;
366 zio->io_transform_stack = zt;
373 zio_pop_transforms(zio_t *zio)
377 while ((zt = zio->io_transform_stack) != NULL) {
378 if (zt->zt_transform != NULL)
379 zt->zt_transform(zio,
380 zt->zt_orig_data, zt->zt_orig_size);
382 if (zt->zt_bufsize != 0)
383 zio_buf_free(zio->io_data, zt->zt_bufsize);
385 zio->io_data = zt->zt_orig_data;
386 zio->io_size = zt->zt_orig_size;
387 zio->io_transform_stack = zt->zt_next;
389 kmem_free(zt, sizeof (zio_transform_t));
394 * ==========================================================================
395 * I/O transform callbacks for subblocks and decompression
396 * ==========================================================================
399 zio_subblock(zio_t *zio, void *data, uint64_t size)
401 ASSERT(zio->io_size > size);
403 if (zio->io_type == ZIO_TYPE_READ)
404 bcopy(zio->io_data, data, size);
408 zio_decompress(zio_t *zio, void *data, uint64_t size)
410 if (zio->io_error == 0 &&
411 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
412 zio->io_data, data, zio->io_size, size) != 0)
413 zio->io_error = SET_ERROR(EIO);
417 * ==========================================================================
418 * I/O parent/child relationships and pipeline interlocks
419 * ==========================================================================
422 zio_walk_parents(zio_t *cio, zio_link_t **zl)
424 list_t *pl = &cio->io_parent_list;
426 *zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl);
430 ASSERT((*zl)->zl_child == cio);
431 return ((*zl)->zl_parent);
435 zio_walk_children(zio_t *pio, zio_link_t **zl)
437 list_t *cl = &pio->io_child_list;
439 *zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl);
443 ASSERT((*zl)->zl_parent == pio);
444 return ((*zl)->zl_child);
448 zio_unique_parent(zio_t *cio)
450 zio_link_t *zl = NULL;
451 zio_t *pio = zio_walk_parents(cio, &zl);
453 VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL);
458 zio_add_child(zio_t *pio, zio_t *cio)
460 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
463 * Logical I/Os can have logical, gang, or vdev children.
464 * Gang I/Os can have gang or vdev children.
465 * Vdev I/Os can only have vdev children.
466 * The following ASSERT captures all of these constraints.
468 ASSERT(cio->io_child_type <= pio->io_child_type);
473 mutex_enter(&cio->io_lock);
474 mutex_enter(&pio->io_lock);
476 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
478 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
479 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
481 list_insert_head(&pio->io_child_list, zl);
482 list_insert_head(&cio->io_parent_list, zl);
484 pio->io_child_count++;
485 cio->io_parent_count++;
487 mutex_exit(&pio->io_lock);
488 mutex_exit(&cio->io_lock);
492 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
494 ASSERT(zl->zl_parent == pio);
495 ASSERT(zl->zl_child == cio);
497 mutex_enter(&cio->io_lock);
498 mutex_enter(&pio->io_lock);
500 list_remove(&pio->io_child_list, zl);
501 list_remove(&cio->io_parent_list, zl);
503 pio->io_child_count--;
504 cio->io_parent_count--;
506 mutex_exit(&pio->io_lock);
507 mutex_exit(&cio->io_lock);
509 kmem_cache_free(zio_link_cache, zl);
513 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
515 uint64_t *countp = &zio->io_children[child][wait];
516 boolean_t waiting = B_FALSE;
518 mutex_enter(&zio->io_lock);
519 ASSERT(zio->io_stall == NULL);
522 ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN);
523 zio->io_stall = countp;
526 mutex_exit(&zio->io_lock);
532 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
534 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
535 int *errorp = &pio->io_child_error[zio->io_child_type];
537 mutex_enter(&pio->io_lock);
538 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
539 *errorp = zio_worst_error(*errorp, zio->io_error);
540 pio->io_reexecute |= zio->io_reexecute;
541 ASSERT3U(*countp, >, 0);
545 if (*countp == 0 && pio->io_stall == countp) {
546 zio_taskq_type_t type =
547 pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE :
549 pio->io_stall = NULL;
550 mutex_exit(&pio->io_lock);
552 * Dispatch the parent zio in its own taskq so that
553 * the child can continue to make progress. This also
554 * prevents overflowing the stack when we have deeply nested
555 * parent-child relationships.
557 zio_taskq_dispatch(pio, type, B_FALSE);
559 mutex_exit(&pio->io_lock);
564 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
566 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
567 zio->io_error = zio->io_child_error[c];
571 zio_timestamp_compare(const void *x1, const void *x2)
573 const zio_t *z1 = x1;
574 const zio_t *z2 = x2;
576 if (z1->io_queued_timestamp < z2->io_queued_timestamp)
578 if (z1->io_queued_timestamp > z2->io_queued_timestamp)
581 if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset)
583 if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset)
586 if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object)
588 if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object)
591 if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level)
593 if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level)
596 if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid)
598 if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid)
610 * ==========================================================================
611 * Create the various types of I/O (read, write, free, etc)
612 * ==========================================================================
615 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
616 void *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done,
617 void *private, zio_type_t type, zio_priority_t priority,
618 enum zio_flag flags, vdev_t *vd, uint64_t offset,
619 const zbookmark_phys_t *zb, enum zio_stage stage, enum zio_stage pipeline)
623 ASSERT3U(type == ZIO_TYPE_FREE || psize, <=, SPA_MAXBLOCKSIZE);
624 ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0);
625 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
627 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
628 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
629 ASSERT(vd || stage == ZIO_STAGE_OPEN);
631 IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW) != 0);
633 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
634 bzero(zio, sizeof (zio_t));
636 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
637 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
639 list_create(&zio->io_parent_list, sizeof (zio_link_t),
640 offsetof(zio_link_t, zl_parent_node));
641 list_create(&zio->io_child_list, sizeof (zio_link_t),
642 offsetof(zio_link_t, zl_child_node));
643 metaslab_trace_init(&zio->io_alloc_list);
646 zio->io_child_type = ZIO_CHILD_VDEV;
647 else if (flags & ZIO_FLAG_GANG_CHILD)
648 zio->io_child_type = ZIO_CHILD_GANG;
649 else if (flags & ZIO_FLAG_DDT_CHILD)
650 zio->io_child_type = ZIO_CHILD_DDT;
652 zio->io_child_type = ZIO_CHILD_LOGICAL;
655 zio->io_bp = (blkptr_t *)bp;
656 zio->io_bp_copy = *bp;
657 zio->io_bp_orig = *bp;
658 if (type != ZIO_TYPE_WRITE ||
659 zio->io_child_type == ZIO_CHILD_DDT)
660 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
661 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
662 zio->io_logical = zio;
663 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
664 pipeline |= ZIO_GANG_STAGES;
670 zio->io_private = private;
672 zio->io_priority = priority;
674 zio->io_offset = offset;
675 zio->io_orig_data = zio->io_data = data;
676 zio->io_orig_size = zio->io_size = psize;
677 zio->io_lsize = lsize;
678 zio->io_orig_flags = zio->io_flags = flags;
679 zio->io_orig_stage = zio->io_stage = stage;
680 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
681 zio->io_pipeline_trace = ZIO_STAGE_OPEN;
683 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
684 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
687 zio->io_bookmark = *zb;
690 if (zio->io_logical == NULL)
691 zio->io_logical = pio->io_logical;
692 if (zio->io_child_type == ZIO_CHILD_GANG)
693 zio->io_gang_leader = pio->io_gang_leader;
694 zio_add_child(pio, zio);
701 zio_destroy(zio_t *zio)
703 metaslab_trace_fini(&zio->io_alloc_list);
704 list_destroy(&zio->io_parent_list);
705 list_destroy(&zio->io_child_list);
706 mutex_destroy(&zio->io_lock);
707 cv_destroy(&zio->io_cv);
708 kmem_cache_free(zio_cache, zio);
712 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
713 void *private, enum zio_flag flags)
717 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
718 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
719 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
725 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
727 return (zio_null(NULL, spa, NULL, done, private, flags));
731 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
733 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
734 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
735 bp, (longlong_t)BP_GET_TYPE(bp));
737 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
738 BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
739 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
740 bp, (longlong_t)BP_GET_CHECKSUM(bp));
742 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
743 BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
744 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
745 bp, (longlong_t)BP_GET_COMPRESS(bp));
747 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
748 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
749 bp, (longlong_t)BP_GET_LSIZE(bp));
751 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
752 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
753 bp, (longlong_t)BP_GET_PSIZE(bp));
756 if (BP_IS_EMBEDDED(bp)) {
757 if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
758 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
759 bp, (longlong_t)BPE_GET_ETYPE(bp));
764 * Pool-specific checks.
766 * Note: it would be nice to verify that the blk_birth and
767 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
768 * allows the birth time of log blocks (and dmu_sync()-ed blocks
769 * that are in the log) to be arbitrarily large.
771 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
772 uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
773 if (vdevid >= spa->spa_root_vdev->vdev_children) {
774 zfs_panic_recover("blkptr at %p DVA %u has invalid "
776 bp, i, (longlong_t)vdevid);
779 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
781 zfs_panic_recover("blkptr at %p DVA %u has invalid "
783 bp, i, (longlong_t)vdevid);
786 if (vd->vdev_ops == &vdev_hole_ops) {
787 zfs_panic_recover("blkptr at %p DVA %u has hole "
789 bp, i, (longlong_t)vdevid);
792 if (vd->vdev_ops == &vdev_missing_ops) {
794 * "missing" vdevs are valid during import, but we
795 * don't have their detailed info (e.g. asize), so
796 * we can't perform any more checks on them.
800 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
801 uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
803 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
804 if (offset + asize > vd->vdev_asize) {
805 zfs_panic_recover("blkptr at %p DVA %u has invalid "
807 bp, i, (longlong_t)offset);
813 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
814 void *data, uint64_t size, zio_done_func_t *done, void *private,
815 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
819 zfs_blkptr_verify(spa, bp);
821 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
822 data, size, size, done, private,
823 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
824 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
825 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
831 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
832 void *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp,
833 zio_done_func_t *ready, zio_done_func_t *children_ready,
834 zio_done_func_t *physdone, zio_done_func_t *done,
835 void *private, zio_priority_t priority, enum zio_flag flags,
836 const zbookmark_phys_t *zb)
840 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
841 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
842 zp->zp_compress >= ZIO_COMPRESS_OFF &&
843 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
844 DMU_OT_IS_VALID(zp->zp_type) &&
847 zp->zp_copies <= spa_max_replication(spa));
849 zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private,
850 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
851 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
852 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
854 zio->io_ready = ready;
855 zio->io_children_ready = children_ready;
856 zio->io_physdone = physdone;
860 * Data can be NULL if we are going to call zio_write_override() to
861 * provide the already-allocated BP. But we may need the data to
862 * verify a dedup hit (if requested). In this case, don't try to
863 * dedup (just take the already-allocated BP verbatim).
865 if (data == NULL && zio->io_prop.zp_dedup_verify) {
866 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
873 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
874 uint64_t size, zio_done_func_t *done, void *private,
875 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
879 zio = zio_create(pio, spa, txg, bp, data, size, size, done, private,
880 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb,
881 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
887 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
889 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
890 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
891 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
892 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
895 * We must reset the io_prop to match the values that existed
896 * when the bp was first written by dmu_sync() keeping in mind
897 * that nopwrite and dedup are mutually exclusive.
899 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
900 zio->io_prop.zp_nopwrite = nopwrite;
901 zio->io_prop.zp_copies = copies;
902 zio->io_bp_override = bp;
906 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
910 * The check for EMBEDDED is a performance optimization. We
911 * process the free here (by ignoring it) rather than
912 * putting it on the list and then processing it in zio_free_sync().
914 if (BP_IS_EMBEDDED(bp))
916 metaslab_check_free(spa, bp);
919 * Frees that are for the currently-syncing txg, are not going to be
920 * deferred, and which will not need to do a read (i.e. not GANG or
921 * DEDUP), can be processed immediately. Otherwise, put them on the
922 * in-memory list for later processing.
924 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
925 txg != spa->spa_syncing_txg ||
926 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
927 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
929 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp,
930 BP_GET_PSIZE(bp), 0)));
935 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
936 uint64_t size, enum zio_flag flags)
939 enum zio_stage stage = ZIO_FREE_PIPELINE;
941 ASSERT(!BP_IS_HOLE(bp));
942 ASSERT(spa_syncing_txg(spa) == txg);
943 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
945 if (BP_IS_EMBEDDED(bp))
946 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
948 metaslab_check_free(spa, bp);
951 if (zfs_trim_enabled)
952 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START |
953 ZIO_STAGE_VDEV_IO_ASSESS;
955 * GANG and DEDUP blocks can induce a read (for the gang block header,
956 * or the DDT), so issue them asynchronously so that this thread is
959 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
960 stage |= ZIO_STAGE_ISSUE_ASYNC;
962 flags |= ZIO_FLAG_DONT_QUEUE;
964 zio = zio_create(pio, spa, txg, bp, NULL, size,
965 size, NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW,
966 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
972 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
973 zio_done_func_t *done, void *private, enum zio_flag flags)
977 dprintf_bp(bp, "claiming in txg %llu", txg);
979 if (BP_IS_EMBEDDED(bp))
980 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
983 * A claim is an allocation of a specific block. Claims are needed
984 * to support immediate writes in the intent log. The issue is that
985 * immediate writes contain committed data, but in a txg that was
986 * *not* committed. Upon opening the pool after an unclean shutdown,
987 * the intent log claims all blocks that contain immediate write data
988 * so that the SPA knows they're in use.
990 * All claims *must* be resolved in the first txg -- before the SPA
991 * starts allocating blocks -- so that nothing is allocated twice.
992 * If txg == 0 we just verify that the block is claimable.
994 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
995 ASSERT(txg == spa_first_txg(spa) || txg == 0);
996 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
998 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
999 BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW,
1000 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
1001 ASSERT0(zio->io_queued_timestamp);
1007 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset,
1008 uint64_t size, zio_done_func_t *done, void *private,
1009 zio_priority_t priority, enum zio_flag flags)
1014 if (vd->vdev_children == 0) {
1015 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
1016 ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
1017 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
1021 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
1023 for (c = 0; c < vd->vdev_children; c++)
1024 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
1025 offset, size, done, private, priority, flags));
1032 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1033 void *data, int checksum, zio_done_func_t *done, void *private,
1034 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1038 ASSERT(vd->vdev_children == 0);
1039 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1040 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1041 ASSERT3U(offset + size, <=, vd->vdev_psize);
1043 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1044 private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1045 offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
1047 zio->io_prop.zp_checksum = checksum;
1053 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1054 void *data, int checksum, zio_done_func_t *done, void *private,
1055 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1059 ASSERT(vd->vdev_children == 0);
1060 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1061 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1062 ASSERT3U(offset + size, <=, vd->vdev_psize);
1064 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1065 private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1066 offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
1068 zio->io_prop.zp_checksum = checksum;
1070 if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
1072 * zec checksums are necessarily destructive -- they modify
1073 * the end of the write buffer to hold the verifier/checksum.
1074 * Therefore, we must make a local copy in case the data is
1075 * being written to multiple places in parallel.
1077 void *wbuf = zio_buf_alloc(size);
1078 bcopy(data, wbuf, size);
1079 zio_push_transform(zio, wbuf, size, size, NULL);
1086 * Create a child I/O to do some work for us.
1089 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1090 void *data, uint64_t size, int type, zio_priority_t priority,
1091 enum zio_flag flags, zio_done_func_t *done, void *private)
1093 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1096 ASSERT(vd->vdev_parent ==
1097 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
1099 if (type == ZIO_TYPE_READ && bp != NULL) {
1101 * If we have the bp, then the child should perform the
1102 * checksum and the parent need not. This pushes error
1103 * detection as close to the leaves as possible and
1104 * eliminates redundant checksums in the interior nodes.
1106 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1107 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1110 /* Not all IO types require vdev io done stage e.g. free */
1111 if (!(pio->io_pipeline & ZIO_STAGE_VDEV_IO_DONE))
1112 pipeline &= ~ZIO_STAGE_VDEV_IO_DONE;
1114 if (vd->vdev_children == 0)
1115 offset += VDEV_LABEL_START_SIZE;
1117 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
1120 * If we've decided to do a repair, the write is not speculative --
1121 * even if the original read was.
1123 if (flags & ZIO_FLAG_IO_REPAIR)
1124 flags &= ~ZIO_FLAG_SPECULATIVE;
1127 * If we're creating a child I/O that is not associated with a
1128 * top-level vdev, then the child zio is not an allocating I/O.
1129 * If this is a retried I/O then we ignore it since we will
1130 * have already processed the original allocating I/O.
1132 if (flags & ZIO_FLAG_IO_ALLOCATING &&
1133 (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) {
1134 metaslab_class_t *mc = spa_normal_class(pio->io_spa);
1136 ASSERT(mc->mc_alloc_throttle_enabled);
1137 ASSERT(type == ZIO_TYPE_WRITE);
1138 ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE);
1139 ASSERT(!(flags & ZIO_FLAG_IO_REPAIR));
1140 ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) ||
1141 pio->io_child_type == ZIO_CHILD_GANG);
1143 flags &= ~ZIO_FLAG_IO_ALLOCATING;
1146 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size,
1147 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1148 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1149 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
1151 zio->io_physdone = pio->io_physdone;
1152 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1153 zio->io_logical->io_phys_children++;
1159 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
1160 int type, zio_priority_t priority, enum zio_flag flags,
1161 zio_done_func_t *done, void *private)
1165 ASSERT(vd->vdev_ops->vdev_op_leaf);
1167 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1168 data, size, size, done, private, type, priority,
1169 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1171 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1177 zio_flush(zio_t *zio, vdev_t *vd)
1179 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0,
1180 NULL, NULL, ZIO_PRIORITY_NOW,
1181 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1185 zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size)
1188 ASSERT(vd->vdev_ops->vdev_op_leaf);
1190 return (zio_create(zio, spa, 0, NULL, NULL, size, size, NULL, NULL,
1191 ZIO_TYPE_FREE, ZIO_PRIORITY_TRIM, ZIO_FLAG_DONT_AGGREGATE |
1192 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY,
1193 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PHYS_PIPELINE));
1197 zio_shrink(zio_t *zio, uint64_t size)
1199 ASSERT(zio->io_executor == NULL);
1200 ASSERT(zio->io_orig_size == zio->io_size);
1201 ASSERT(size <= zio->io_size);
1204 * We don't shrink for raidz because of problems with the
1205 * reconstruction when reading back less than the block size.
1206 * Note, BP_IS_RAIDZ() assumes no compression.
1208 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1209 if (!BP_IS_RAIDZ(zio->io_bp)) {
1210 /* we are not doing a raw write */
1211 ASSERT3U(zio->io_size, ==, zio->io_lsize);
1212 zio->io_orig_size = zio->io_size = zio->io_lsize = size;
1217 * ==========================================================================
1218 * Prepare to read and write logical blocks
1219 * ==========================================================================
1223 zio_read_bp_init(zio_t *zio)
1225 blkptr_t *bp = zio->io_bp;
1227 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1228 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1229 !(zio->io_flags & ZIO_FLAG_RAW)) {
1231 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1232 void *cbuf = zio_buf_alloc(psize);
1234 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
1237 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1238 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1239 decode_embedded_bp_compressed(bp, zio->io_data);
1241 ASSERT(!BP_IS_EMBEDDED(bp));
1244 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1245 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1247 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1248 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1250 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1251 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1253 return (ZIO_PIPELINE_CONTINUE);
1257 zio_write_bp_init(zio_t *zio)
1259 if (!IO_IS_ALLOCATING(zio))
1260 return (ZIO_PIPELINE_CONTINUE);
1262 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1264 if (zio->io_bp_override) {
1265 blkptr_t *bp = zio->io_bp;
1266 zio_prop_t *zp = &zio->io_prop;
1268 ASSERT(bp->blk_birth != zio->io_txg);
1269 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1271 *bp = *zio->io_bp_override;
1272 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1274 if (BP_IS_EMBEDDED(bp))
1275 return (ZIO_PIPELINE_CONTINUE);
1278 * If we've been overridden and nopwrite is set then
1279 * set the flag accordingly to indicate that a nopwrite
1280 * has already occurred.
1282 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1283 ASSERT(!zp->zp_dedup);
1284 ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum);
1285 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1286 return (ZIO_PIPELINE_CONTINUE);
1289 ASSERT(!zp->zp_nopwrite);
1291 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1292 return (ZIO_PIPELINE_CONTINUE);
1294 ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags &
1295 ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify);
1297 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1298 BP_SET_DEDUP(bp, 1);
1299 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1300 return (ZIO_PIPELINE_CONTINUE);
1304 * We were unable to handle this as an override bp, treat
1305 * it as a regular write I/O.
1307 zio->io_bp_override = NULL;
1308 *bp = zio->io_bp_orig;
1309 zio->io_pipeline = zio->io_orig_pipeline;
1312 return (ZIO_PIPELINE_CONTINUE);
1316 zio_write_compress(zio_t *zio)
1318 spa_t *spa = zio->io_spa;
1319 zio_prop_t *zp = &zio->io_prop;
1320 enum zio_compress compress = zp->zp_compress;
1321 blkptr_t *bp = zio->io_bp;
1322 uint64_t lsize = zio->io_lsize;
1323 uint64_t psize = zio->io_size;
1326 EQUIV(lsize != psize, (zio->io_flags & ZIO_FLAG_RAW) != 0);
1329 * If our children haven't all reached the ready stage,
1330 * wait for them and then repeat this pipeline stage.
1332 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1333 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1334 return (ZIO_PIPELINE_STOP);
1336 if (!IO_IS_ALLOCATING(zio))
1337 return (ZIO_PIPELINE_CONTINUE);
1339 if (zio->io_children_ready != NULL) {
1341 * Now that all our children are ready, run the callback
1342 * associated with this zio in case it wants to modify the
1343 * data to be written.
1345 ASSERT3U(zp->zp_level, >, 0);
1346 zio->io_children_ready(zio);
1349 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1350 ASSERT(zio->io_bp_override == NULL);
1352 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1354 * We're rewriting an existing block, which means we're
1355 * working on behalf of spa_sync(). For spa_sync() to
1356 * converge, it must eventually be the case that we don't
1357 * have to allocate new blocks. But compression changes
1358 * the blocksize, which forces a reallocate, and makes
1359 * convergence take longer. Therefore, after the first
1360 * few passes, stop compressing to ensure convergence.
1362 pass = spa_sync_pass(spa);
1364 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1365 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1366 ASSERT(!BP_GET_DEDUP(bp));
1368 if (pass >= zfs_sync_pass_dont_compress)
1369 compress = ZIO_COMPRESS_OFF;
1371 /* Make sure someone doesn't change their mind on overwrites */
1372 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1373 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1376 /* If it's a compressed write that is not raw, compress the buffer. */
1377 if (compress != ZIO_COMPRESS_OFF && psize == lsize) {
1378 void *cbuf = zio_buf_alloc(lsize);
1379 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1380 if (psize == 0 || psize == lsize) {
1381 compress = ZIO_COMPRESS_OFF;
1382 zio_buf_free(cbuf, lsize);
1383 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1384 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1385 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1386 encode_embedded_bp_compressed(bp,
1387 cbuf, compress, lsize, psize);
1388 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1389 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1390 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1391 zio_buf_free(cbuf, lsize);
1392 bp->blk_birth = zio->io_txg;
1393 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1394 ASSERT(spa_feature_is_active(spa,
1395 SPA_FEATURE_EMBEDDED_DATA));
1396 return (ZIO_PIPELINE_CONTINUE);
1399 * Round up compressed size up to the ashift
1400 * of the smallest-ashift device, and zero the tail.
1401 * This ensures that the compressed size of the BP
1402 * (and thus compressratio property) are correct,
1403 * in that we charge for the padding used to fill out
1406 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
1407 size_t rounded = (size_t)P2ROUNDUP(psize,
1408 1ULL << spa->spa_min_ashift);
1409 if (rounded >= lsize) {
1410 compress = ZIO_COMPRESS_OFF;
1411 zio_buf_free(cbuf, lsize);
1414 bzero((char *)cbuf + psize, rounded - psize);
1416 zio_push_transform(zio, cbuf,
1417 psize, lsize, NULL);
1422 * We were unable to handle this as an override bp, treat
1423 * it as a regular write I/O.
1425 zio->io_bp_override = NULL;
1426 *bp = zio->io_bp_orig;
1427 zio->io_pipeline = zio->io_orig_pipeline;
1429 ASSERT3U(psize, !=, 0);
1433 * The final pass of spa_sync() must be all rewrites, but the first
1434 * few passes offer a trade-off: allocating blocks defers convergence,
1435 * but newly allocated blocks are sequential, so they can be written
1436 * to disk faster. Therefore, we allow the first few passes of
1437 * spa_sync() to allocate new blocks, but force rewrites after that.
1438 * There should only be a handful of blocks after pass 1 in any case.
1440 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1441 BP_GET_PSIZE(bp) == psize &&
1442 pass >= zfs_sync_pass_rewrite) {
1444 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1445 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1446 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1449 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1453 if (zio->io_bp_orig.blk_birth != 0 &&
1454 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1455 BP_SET_LSIZE(bp, lsize);
1456 BP_SET_TYPE(bp, zp->zp_type);
1457 BP_SET_LEVEL(bp, zp->zp_level);
1458 BP_SET_BIRTH(bp, zio->io_txg, 0);
1460 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1462 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1463 BP_SET_LSIZE(bp, lsize);
1464 BP_SET_TYPE(bp, zp->zp_type);
1465 BP_SET_LEVEL(bp, zp->zp_level);
1466 BP_SET_PSIZE(bp, psize);
1467 BP_SET_COMPRESS(bp, compress);
1468 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1469 BP_SET_DEDUP(bp, zp->zp_dedup);
1470 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1472 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1473 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1474 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1476 if (zp->zp_nopwrite) {
1477 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1478 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1479 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1482 return (ZIO_PIPELINE_CONTINUE);
1486 zio_free_bp_init(zio_t *zio)
1488 blkptr_t *bp = zio->io_bp;
1490 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1491 if (BP_GET_DEDUP(bp))
1492 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1495 return (ZIO_PIPELINE_CONTINUE);
1499 * ==========================================================================
1500 * Execute the I/O pipeline
1501 * ==========================================================================
1505 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1507 spa_t *spa = zio->io_spa;
1508 zio_type_t t = zio->io_type;
1509 int flags = (cutinline ? TQ_FRONT : 0);
1511 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1514 * If we're a config writer or a probe, the normal issue and
1515 * interrupt threads may all be blocked waiting for the config lock.
1516 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1518 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1522 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1524 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1528 * If this is a high priority I/O, then use the high priority taskq if
1531 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1532 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1535 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1538 * NB: We are assuming that the zio can only be dispatched
1539 * to a single taskq at a time. It would be a grievous error
1540 * to dispatch the zio to another taskq at the same time.
1542 #if defined(illumos) || !defined(_KERNEL)
1543 ASSERT(zio->io_tqent.tqent_next == NULL);
1545 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
1547 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1548 flags, &zio->io_tqent);
1552 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1554 kthread_t *executor = zio->io_executor;
1555 spa_t *spa = zio->io_spa;
1557 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1558 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1560 for (i = 0; i < tqs->stqs_count; i++) {
1561 if (taskq_member(tqs->stqs_taskq[i], executor))
1570 zio_issue_async(zio_t *zio)
1572 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1574 return (ZIO_PIPELINE_STOP);
1578 zio_interrupt(zio_t *zio)
1580 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1584 zio_delay_interrupt(zio_t *zio)
1587 * The timeout_generic() function isn't defined in userspace, so
1588 * rather than trying to implement the function, the zio delay
1589 * functionality has been disabled for userspace builds.
1594 * If io_target_timestamp is zero, then no delay has been registered
1595 * for this IO, thus jump to the end of this function and "skip" the
1596 * delay; issuing it directly to the zio layer.
1598 if (zio->io_target_timestamp != 0) {
1599 hrtime_t now = gethrtime();
1601 if (now >= zio->io_target_timestamp) {
1603 * This IO has already taken longer than the target
1604 * delay to complete, so we don't want to delay it
1605 * any longer; we "miss" the delay and issue it
1606 * directly to the zio layer. This is likely due to
1607 * the target latency being set to a value less than
1608 * the underlying hardware can satisfy (e.g. delay
1609 * set to 1ms, but the disks take 10ms to complete an
1613 DTRACE_PROBE2(zio__delay__miss, zio_t *, zio,
1618 hrtime_t diff = zio->io_target_timestamp - now;
1620 DTRACE_PROBE3(zio__delay__hit, zio_t *, zio,
1621 hrtime_t, now, hrtime_t, diff);
1623 (void) timeout_generic(CALLOUT_NORMAL,
1624 (void (*)(void *))zio_interrupt, zio, diff, 1, 0);
1631 DTRACE_PROBE1(zio__delay__skip, zio_t *, zio);
1636 * Execute the I/O pipeline until one of the following occurs:
1638 * (1) the I/O completes
1639 * (2) the pipeline stalls waiting for dependent child I/Os
1640 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1641 * (4) the I/O is delegated by vdev-level caching or aggregation
1642 * (5) the I/O is deferred due to vdev-level queueing
1643 * (6) the I/O is handed off to another thread.
1645 * In all cases, the pipeline stops whenever there's no CPU work; it never
1646 * burns a thread in cv_wait().
1648 * There's no locking on io_stage because there's no legitimate way
1649 * for multiple threads to be attempting to process the same I/O.
1651 static zio_pipe_stage_t *zio_pipeline[];
1654 zio_execute(zio_t *zio)
1656 zio->io_executor = curthread;
1658 ASSERT3U(zio->io_queued_timestamp, >, 0);
1660 while (zio->io_stage < ZIO_STAGE_DONE) {
1661 enum zio_stage pipeline = zio->io_pipeline;
1662 enum zio_stage stage = zio->io_stage;
1665 ASSERT(!MUTEX_HELD(&zio->io_lock));
1666 ASSERT(ISP2(stage));
1667 ASSERT(zio->io_stall == NULL);
1671 } while ((stage & pipeline) == 0);
1673 ASSERT(stage <= ZIO_STAGE_DONE);
1676 * If we are in interrupt context and this pipeline stage
1677 * will grab a config lock that is held across I/O,
1678 * or may wait for an I/O that needs an interrupt thread
1679 * to complete, issue async to avoid deadlock.
1681 * For VDEV_IO_START, we cut in line so that the io will
1682 * be sent to disk promptly.
1684 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1685 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1686 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1687 zio_requeue_io_start_cut_in_line : B_FALSE;
1688 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1692 zio->io_stage = stage;
1693 zio->io_pipeline_trace |= zio->io_stage;
1694 rv = zio_pipeline[highbit64(stage) - 1](zio);
1696 if (rv == ZIO_PIPELINE_STOP)
1699 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1704 * ==========================================================================
1705 * Initiate I/O, either sync or async
1706 * ==========================================================================
1709 zio_wait(zio_t *zio)
1713 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1714 ASSERT(zio->io_executor == NULL);
1716 zio->io_waiter = curthread;
1717 ASSERT0(zio->io_queued_timestamp);
1718 zio->io_queued_timestamp = gethrtime();
1722 mutex_enter(&zio->io_lock);
1723 while (zio->io_executor != NULL)
1724 cv_wait(&zio->io_cv, &zio->io_lock);
1725 mutex_exit(&zio->io_lock);
1727 error = zio->io_error;
1734 zio_nowait(zio_t *zio)
1736 ASSERT(zio->io_executor == NULL);
1738 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1739 zio_unique_parent(zio) == NULL) {
1741 * This is a logical async I/O with no parent to wait for it.
1742 * We add it to the spa_async_root_zio "Godfather" I/O which
1743 * will ensure they complete prior to unloading the pool.
1745 spa_t *spa = zio->io_spa;
1747 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1750 ASSERT0(zio->io_queued_timestamp);
1751 zio->io_queued_timestamp = gethrtime();
1756 * ==========================================================================
1757 * Reexecute or suspend/resume failed I/O
1758 * ==========================================================================
1762 zio_reexecute(zio_t *pio)
1764 zio_t *cio, *cio_next;
1766 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1767 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1768 ASSERT(pio->io_gang_leader == NULL);
1769 ASSERT(pio->io_gang_tree == NULL);
1771 pio->io_flags = pio->io_orig_flags;
1772 pio->io_stage = pio->io_orig_stage;
1773 pio->io_pipeline = pio->io_orig_pipeline;
1774 pio->io_reexecute = 0;
1775 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1776 pio->io_pipeline_trace = 0;
1778 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1779 pio->io_state[w] = 0;
1780 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1781 pio->io_child_error[c] = 0;
1783 if (IO_IS_ALLOCATING(pio))
1784 BP_ZERO(pio->io_bp);
1787 * As we reexecute pio's children, new children could be created.
1788 * New children go to the head of pio's io_child_list, however,
1789 * so we will (correctly) not reexecute them. The key is that
1790 * the remainder of pio's io_child_list, from 'cio_next' onward,
1791 * cannot be affected by any side effects of reexecuting 'cio'.
1793 zio_link_t *zl = NULL;
1794 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
1795 cio_next = zio_walk_children(pio, &zl);
1796 mutex_enter(&pio->io_lock);
1797 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1798 pio->io_children[cio->io_child_type][w]++;
1799 mutex_exit(&pio->io_lock);
1804 * Now that all children have been reexecuted, execute the parent.
1805 * We don't reexecute "The Godfather" I/O here as it's the
1806 * responsibility of the caller to wait on him.
1808 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) {
1809 pio->io_queued_timestamp = gethrtime();
1815 zio_suspend(spa_t *spa, zio_t *zio)
1817 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1818 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1819 "failure and the failure mode property for this pool "
1820 "is set to panic.", spa_name(spa));
1822 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1824 mutex_enter(&spa->spa_suspend_lock);
1826 if (spa->spa_suspend_zio_root == NULL)
1827 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1828 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1829 ZIO_FLAG_GODFATHER);
1831 spa->spa_suspended = B_TRUE;
1834 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1835 ASSERT(zio != spa->spa_suspend_zio_root);
1836 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1837 ASSERT(zio_unique_parent(zio) == NULL);
1838 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1839 zio_add_child(spa->spa_suspend_zio_root, zio);
1842 mutex_exit(&spa->spa_suspend_lock);
1846 zio_resume(spa_t *spa)
1851 * Reexecute all previously suspended i/o.
1853 mutex_enter(&spa->spa_suspend_lock);
1854 spa->spa_suspended = B_FALSE;
1855 cv_broadcast(&spa->spa_suspend_cv);
1856 pio = spa->spa_suspend_zio_root;
1857 spa->spa_suspend_zio_root = NULL;
1858 mutex_exit(&spa->spa_suspend_lock);
1864 return (zio_wait(pio));
1868 zio_resume_wait(spa_t *spa)
1870 mutex_enter(&spa->spa_suspend_lock);
1871 while (spa_suspended(spa))
1872 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1873 mutex_exit(&spa->spa_suspend_lock);
1877 * ==========================================================================
1880 * A gang block is a collection of small blocks that looks to the DMU
1881 * like one large block. When zio_dva_allocate() cannot find a block
1882 * of the requested size, due to either severe fragmentation or the pool
1883 * being nearly full, it calls zio_write_gang_block() to construct the
1884 * block from smaller fragments.
1886 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1887 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1888 * an indirect block: it's an array of block pointers. It consumes
1889 * only one sector and hence is allocatable regardless of fragmentation.
1890 * The gang header's bps point to its gang members, which hold the data.
1892 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1893 * as the verifier to ensure uniqueness of the SHA256 checksum.
1894 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1895 * not the gang header. This ensures that data block signatures (needed for
1896 * deduplication) are independent of how the block is physically stored.
1898 * Gang blocks can be nested: a gang member may itself be a gang block.
1899 * Thus every gang block is a tree in which root and all interior nodes are
1900 * gang headers, and the leaves are normal blocks that contain user data.
1901 * The root of the gang tree is called the gang leader.
1903 * To perform any operation (read, rewrite, free, claim) on a gang block,
1904 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1905 * in the io_gang_tree field of the original logical i/o by recursively
1906 * reading the gang leader and all gang headers below it. This yields
1907 * an in-core tree containing the contents of every gang header and the
1908 * bps for every constituent of the gang block.
1910 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1911 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1912 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1913 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1914 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1915 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1916 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1917 * of the gang header plus zio_checksum_compute() of the data to update the
1918 * gang header's blk_cksum as described above.
1920 * The two-phase assemble/issue model solves the problem of partial failure --
1921 * what if you'd freed part of a gang block but then couldn't read the
1922 * gang header for another part? Assembling the entire gang tree first
1923 * ensures that all the necessary gang header I/O has succeeded before
1924 * starting the actual work of free, claim, or write. Once the gang tree
1925 * is assembled, free and claim are in-memory operations that cannot fail.
1927 * In the event that a gang write fails, zio_dva_unallocate() walks the
1928 * gang tree to immediately free (i.e. insert back into the space map)
1929 * everything we've allocated. This ensures that we don't get ENOSPC
1930 * errors during repeated suspend/resume cycles due to a flaky device.
1932 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1933 * the gang tree, we won't modify the block, so we can safely defer the free
1934 * (knowing that the block is still intact). If we *can* assemble the gang
1935 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1936 * each constituent bp and we can allocate a new block on the next sync pass.
1938 * In all cases, the gang tree allows complete recovery from partial failure.
1939 * ==========================================================================
1943 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1948 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1949 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1950 &pio->io_bookmark));
1954 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1959 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1960 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1961 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1963 * As we rewrite each gang header, the pipeline will compute
1964 * a new gang block header checksum for it; but no one will
1965 * compute a new data checksum, so we do that here. The one
1966 * exception is the gang leader: the pipeline already computed
1967 * its data checksum because that stage precedes gang assembly.
1968 * (Presently, nothing actually uses interior data checksums;
1969 * this is just good hygiene.)
1971 if (gn != pio->io_gang_leader->io_gang_tree) {
1972 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1973 data, BP_GET_PSIZE(bp));
1976 * If we are here to damage data for testing purposes,
1977 * leave the GBH alone so that we can detect the damage.
1979 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1980 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1982 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1983 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1984 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1992 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1994 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1995 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp),
1996 ZIO_GANG_CHILD_FLAGS(pio)));
2001 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
2003 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
2004 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
2007 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
2016 static void zio_gang_tree_assemble_done(zio_t *zio);
2018 static zio_gang_node_t *
2019 zio_gang_node_alloc(zio_gang_node_t **gnpp)
2021 zio_gang_node_t *gn;
2023 ASSERT(*gnpp == NULL);
2025 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
2026 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
2033 zio_gang_node_free(zio_gang_node_t **gnpp)
2035 zio_gang_node_t *gn = *gnpp;
2037 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2038 ASSERT(gn->gn_child[g] == NULL);
2040 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2041 kmem_free(gn, sizeof (*gn));
2046 zio_gang_tree_free(zio_gang_node_t **gnpp)
2048 zio_gang_node_t *gn = *gnpp;
2053 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2054 zio_gang_tree_free(&gn->gn_child[g]);
2056 zio_gang_node_free(gnpp);
2060 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
2062 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
2064 ASSERT(gio->io_gang_leader == gio);
2065 ASSERT(BP_IS_GANG(bp));
2067 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
2068 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
2069 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
2073 zio_gang_tree_assemble_done(zio_t *zio)
2075 zio_t *gio = zio->io_gang_leader;
2076 zio_gang_node_t *gn = zio->io_private;
2077 blkptr_t *bp = zio->io_bp;
2079 ASSERT(gio == zio_unique_parent(zio));
2080 ASSERT(zio->io_child_count == 0);
2085 if (BP_SHOULD_BYTESWAP(bp))
2086 byteswap_uint64_array(zio->io_data, zio->io_size);
2088 ASSERT(zio->io_data == gn->gn_gbh);
2089 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
2090 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2092 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2093 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2094 if (!BP_IS_GANG(gbp))
2096 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
2101 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
2103 zio_t *gio = pio->io_gang_leader;
2106 ASSERT(BP_IS_GANG(bp) == !!gn);
2107 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
2108 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
2111 * If you're a gang header, your data is in gn->gn_gbh.
2112 * If you're a gang member, your data is in 'data' and gn == NULL.
2114 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
2117 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2119 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2120 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2121 if (BP_IS_HOLE(gbp))
2123 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
2124 data = (char *)data + BP_GET_PSIZE(gbp);
2128 if (gn == gio->io_gang_tree && gio->io_data != NULL)
2129 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
2136 zio_gang_assemble(zio_t *zio)
2138 blkptr_t *bp = zio->io_bp;
2140 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
2141 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2143 zio->io_gang_leader = zio;
2145 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
2147 return (ZIO_PIPELINE_CONTINUE);
2151 zio_gang_issue(zio_t *zio)
2153 blkptr_t *bp = zio->io_bp;
2155 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
2156 return (ZIO_PIPELINE_STOP);
2158 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
2159 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2161 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
2162 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
2164 zio_gang_tree_free(&zio->io_gang_tree);
2166 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2168 return (ZIO_PIPELINE_CONTINUE);
2172 zio_write_gang_member_ready(zio_t *zio)
2174 zio_t *pio = zio_unique_parent(zio);
2175 zio_t *gio = zio->io_gang_leader;
2176 dva_t *cdva = zio->io_bp->blk_dva;
2177 dva_t *pdva = pio->io_bp->blk_dva;
2180 if (BP_IS_HOLE(zio->io_bp))
2183 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
2185 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
2186 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
2187 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2188 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
2189 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
2191 mutex_enter(&pio->io_lock);
2192 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
2193 ASSERT(DVA_GET_GANG(&pdva[d]));
2194 asize = DVA_GET_ASIZE(&pdva[d]);
2195 asize += DVA_GET_ASIZE(&cdva[d]);
2196 DVA_SET_ASIZE(&pdva[d], asize);
2198 mutex_exit(&pio->io_lock);
2202 zio_write_gang_block(zio_t *pio)
2204 spa_t *spa = pio->io_spa;
2205 metaslab_class_t *mc = spa_normal_class(spa);
2206 blkptr_t *bp = pio->io_bp;
2207 zio_t *gio = pio->io_gang_leader;
2209 zio_gang_node_t *gn, **gnpp;
2210 zio_gbh_phys_t *gbh;
2211 uint64_t txg = pio->io_txg;
2212 uint64_t resid = pio->io_size;
2214 int copies = gio->io_prop.zp_copies;
2215 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
2219 int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER;
2220 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2221 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2222 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2224 flags |= METASLAB_ASYNC_ALLOC;
2225 VERIFY(refcount_held(&mc->mc_alloc_slots, pio));
2228 * The logical zio has already placed a reservation for
2229 * 'copies' allocation slots but gang blocks may require
2230 * additional copies. These additional copies
2231 * (i.e. gbh_copies - copies) are guaranteed to succeed
2232 * since metaslab_class_throttle_reserve() always allows
2233 * additional reservations for gang blocks.
2235 VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies,
2239 error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE,
2240 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags,
2241 &pio->io_alloc_list, pio);
2243 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2244 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2245 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2248 * If we failed to allocate the gang block header then
2249 * we remove any additional allocation reservations that
2250 * we placed here. The original reservation will
2251 * be removed when the logical I/O goes to the ready
2254 metaslab_class_throttle_unreserve(mc,
2255 gbh_copies - copies, pio);
2257 pio->io_error = error;
2258 return (ZIO_PIPELINE_CONTINUE);
2262 gnpp = &gio->io_gang_tree;
2264 gnpp = pio->io_private;
2265 ASSERT(pio->io_ready == zio_write_gang_member_ready);
2268 gn = zio_gang_node_alloc(gnpp);
2270 bzero(gbh, SPA_GANGBLOCKSIZE);
2273 * Create the gang header.
2275 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
2276 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2279 * Create and nowait the gang children.
2281 for (int g = 0; resid != 0; resid -= lsize, g++) {
2282 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2284 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2286 zp.zp_checksum = gio->io_prop.zp_checksum;
2287 zp.zp_compress = ZIO_COMPRESS_OFF;
2288 zp.zp_type = DMU_OT_NONE;
2290 zp.zp_copies = gio->io_prop.zp_copies;
2291 zp.zp_dedup = B_FALSE;
2292 zp.zp_dedup_verify = B_FALSE;
2293 zp.zp_nopwrite = B_FALSE;
2295 zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2296 (char *)pio->io_data + (pio->io_size - resid), lsize, lsize,
2297 &zp, zio_write_gang_member_ready, NULL, NULL, NULL,
2298 &gn->gn_child[g], pio->io_priority,
2299 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2301 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2302 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2303 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2306 * Gang children won't throttle but we should
2307 * account for their work, so reserve an allocation
2308 * slot for them here.
2310 VERIFY(metaslab_class_throttle_reserve(mc,
2311 zp.zp_copies, cio, flags));
2317 * Set pio's pipeline to just wait for zio to finish.
2319 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2323 return (ZIO_PIPELINE_CONTINUE);
2327 * The zio_nop_write stage in the pipeline determines if allocating a
2328 * new bp is necessary. The nopwrite feature can handle writes in
2329 * either syncing or open context (i.e. zil writes) and as a result is
2330 * mutually exclusive with dedup.
2332 * By leveraging a cryptographically secure checksum, such as SHA256, we
2333 * can compare the checksums of the new data and the old to determine if
2334 * allocating a new block is required. Note that our requirements for
2335 * cryptographic strength are fairly weak: there can't be any accidental
2336 * hash collisions, but we don't need to be secure against intentional
2337 * (malicious) collisions. To trigger a nopwrite, you have to be able
2338 * to write the file to begin with, and triggering an incorrect (hash
2339 * collision) nopwrite is no worse than simply writing to the file.
2340 * That said, there are no known attacks against the checksum algorithms
2341 * used for nopwrite, assuming that the salt and the checksums
2342 * themselves remain secret.
2345 zio_nop_write(zio_t *zio)
2347 blkptr_t *bp = zio->io_bp;
2348 blkptr_t *bp_orig = &zio->io_bp_orig;
2349 zio_prop_t *zp = &zio->io_prop;
2351 ASSERT(BP_GET_LEVEL(bp) == 0);
2352 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2353 ASSERT(zp->zp_nopwrite);
2354 ASSERT(!zp->zp_dedup);
2355 ASSERT(zio->io_bp_override == NULL);
2356 ASSERT(IO_IS_ALLOCATING(zio));
2359 * Check to see if the original bp and the new bp have matching
2360 * characteristics (i.e. same checksum, compression algorithms, etc).
2361 * If they don't then just continue with the pipeline which will
2362 * allocate a new bp.
2364 if (BP_IS_HOLE(bp_orig) ||
2365 !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
2366 ZCHECKSUM_FLAG_NOPWRITE) ||
2367 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2368 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2369 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2370 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2371 return (ZIO_PIPELINE_CONTINUE);
2374 * If the checksums match then reset the pipeline so that we
2375 * avoid allocating a new bp and issuing any I/O.
2377 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2378 ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
2379 ZCHECKSUM_FLAG_NOPWRITE);
2380 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2381 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2382 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2383 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2384 sizeof (uint64_t)) == 0);
2387 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2388 zio->io_flags |= ZIO_FLAG_NOPWRITE;
2391 return (ZIO_PIPELINE_CONTINUE);
2395 * ==========================================================================
2397 * ==========================================================================
2400 zio_ddt_child_read_done(zio_t *zio)
2402 blkptr_t *bp = zio->io_bp;
2403 ddt_entry_t *dde = zio->io_private;
2405 zio_t *pio = zio_unique_parent(zio);
2407 mutex_enter(&pio->io_lock);
2408 ddp = ddt_phys_select(dde, bp);
2409 if (zio->io_error == 0)
2410 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2411 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
2412 dde->dde_repair_data = zio->io_data;
2414 zio_buf_free(zio->io_data, zio->io_size);
2415 mutex_exit(&pio->io_lock);
2419 zio_ddt_read_start(zio_t *zio)
2421 blkptr_t *bp = zio->io_bp;
2423 ASSERT(BP_GET_DEDUP(bp));
2424 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2425 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2427 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2428 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2429 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2430 ddt_phys_t *ddp = dde->dde_phys;
2431 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2434 ASSERT(zio->io_vsd == NULL);
2437 if (ddp_self == NULL)
2438 return (ZIO_PIPELINE_CONTINUE);
2440 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2441 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2443 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2445 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2446 zio_buf_alloc(zio->io_size), zio->io_size,
2447 zio_ddt_child_read_done, dde, zio->io_priority,
2448 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
2449 &zio->io_bookmark));
2451 return (ZIO_PIPELINE_CONTINUE);
2454 zio_nowait(zio_read(zio, zio->io_spa, bp,
2455 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
2456 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2458 return (ZIO_PIPELINE_CONTINUE);
2462 zio_ddt_read_done(zio_t *zio)
2464 blkptr_t *bp = zio->io_bp;
2466 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2467 return (ZIO_PIPELINE_STOP);
2469 ASSERT(BP_GET_DEDUP(bp));
2470 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2471 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2473 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2474 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2475 ddt_entry_t *dde = zio->io_vsd;
2477 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2478 return (ZIO_PIPELINE_CONTINUE);
2481 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2482 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2483 return (ZIO_PIPELINE_STOP);
2485 if (dde->dde_repair_data != NULL) {
2486 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2487 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2489 ddt_repair_done(ddt, dde);
2493 ASSERT(zio->io_vsd == NULL);
2495 return (ZIO_PIPELINE_CONTINUE);
2499 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2501 spa_t *spa = zio->io_spa;
2502 boolean_t do_raw = (zio->io_flags & ZIO_FLAG_RAW);
2504 /* We should never get a raw, override zio */
2505 ASSERT(!(zio->io_bp_override && do_raw));
2508 * Note: we compare the original data, not the transformed data,
2509 * because when zio->io_bp is an override bp, we will not have
2510 * pushed the I/O transforms. That's an important optimization
2511 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2513 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2514 zio_t *lio = dde->dde_lead_zio[p];
2517 return (lio->io_orig_size != zio->io_orig_size ||
2518 bcmp(zio->io_orig_data, lio->io_orig_data,
2519 zio->io_orig_size) != 0);
2523 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2524 ddt_phys_t *ddp = &dde->dde_phys[p];
2526 if (ddp->ddp_phys_birth != 0) {
2527 arc_buf_t *abuf = NULL;
2528 arc_flags_t aflags = ARC_FLAG_WAIT;
2529 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE;
2530 blkptr_t blk = *zio->io_bp;
2533 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2538 * Intuitively, it would make more sense to compare
2539 * io_data than io_orig_data in the raw case since you
2540 * don't want to look at any transformations that have
2541 * happened to the data. However, for raw I/Os the
2542 * data will actually be the same in io_data and
2543 * io_orig_data, so all we have to do is issue this as
2547 zio_flags |= ZIO_FLAG_RAW;
2548 ASSERT3U(zio->io_size, ==, zio->io_orig_size);
2549 ASSERT0(bcmp(zio->io_data, zio->io_orig_data,
2551 ASSERT3P(zio->io_transform_stack, ==, NULL);
2554 error = arc_read(NULL, spa, &blk,
2555 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2556 zio_flags, &aflags, &zio->io_bookmark);
2559 if (arc_buf_size(abuf) != zio->io_orig_size ||
2560 bcmp(abuf->b_data, zio->io_orig_data,
2561 zio->io_orig_size) != 0)
2562 error = SET_ERROR(EEXIST);
2563 arc_buf_destroy(abuf, &abuf);
2567 return (error != 0);
2575 zio_ddt_child_write_ready(zio_t *zio)
2577 int p = zio->io_prop.zp_copies;
2578 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2579 ddt_entry_t *dde = zio->io_private;
2580 ddt_phys_t *ddp = &dde->dde_phys[p];
2588 ASSERT(dde->dde_lead_zio[p] == zio);
2590 ddt_phys_fill(ddp, zio->io_bp);
2592 zio_link_t *zl = NULL;
2593 while ((pio = zio_walk_parents(zio, &zl)) != NULL)
2594 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2600 zio_ddt_child_write_done(zio_t *zio)
2602 int p = zio->io_prop.zp_copies;
2603 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2604 ddt_entry_t *dde = zio->io_private;
2605 ddt_phys_t *ddp = &dde->dde_phys[p];
2609 ASSERT(ddp->ddp_refcnt == 0);
2610 ASSERT(dde->dde_lead_zio[p] == zio);
2611 dde->dde_lead_zio[p] = NULL;
2613 if (zio->io_error == 0) {
2614 zio_link_t *zl = NULL;
2615 while (zio_walk_parents(zio, &zl) != NULL)
2616 ddt_phys_addref(ddp);
2618 ddt_phys_clear(ddp);
2625 zio_ddt_ditto_write_done(zio_t *zio)
2627 int p = DDT_PHYS_DITTO;
2628 zio_prop_t *zp = &zio->io_prop;
2629 blkptr_t *bp = zio->io_bp;
2630 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2631 ddt_entry_t *dde = zio->io_private;
2632 ddt_phys_t *ddp = &dde->dde_phys[p];
2633 ddt_key_t *ddk = &dde->dde_key;
2637 ASSERT(ddp->ddp_refcnt == 0);
2638 ASSERT(dde->dde_lead_zio[p] == zio);
2639 dde->dde_lead_zio[p] = NULL;
2641 if (zio->io_error == 0) {
2642 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2643 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2644 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2645 if (ddp->ddp_phys_birth != 0)
2646 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2647 ddt_phys_fill(ddp, bp);
2654 zio_ddt_write(zio_t *zio)
2656 spa_t *spa = zio->io_spa;
2657 blkptr_t *bp = zio->io_bp;
2658 uint64_t txg = zio->io_txg;
2659 zio_prop_t *zp = &zio->io_prop;
2660 int p = zp->zp_copies;
2664 ddt_t *ddt = ddt_select(spa, bp);
2668 ASSERT(BP_GET_DEDUP(bp));
2669 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2670 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2671 ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW)));
2674 dde = ddt_lookup(ddt, bp, B_TRUE);
2675 ddp = &dde->dde_phys[p];
2677 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2679 * If we're using a weak checksum, upgrade to a strong checksum
2680 * and try again. If we're already using a strong checksum,
2681 * we can't resolve it, so just convert to an ordinary write.
2682 * (And automatically e-mail a paper to Nature?)
2684 if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
2685 ZCHECKSUM_FLAG_DEDUP)) {
2686 zp->zp_checksum = spa_dedup_checksum(spa);
2687 zio_pop_transforms(zio);
2688 zio->io_stage = ZIO_STAGE_OPEN;
2691 zp->zp_dedup = B_FALSE;
2692 BP_SET_DEDUP(bp, B_FALSE);
2694 ASSERT(!BP_GET_DEDUP(bp));
2695 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2697 return (ZIO_PIPELINE_CONTINUE);
2700 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2701 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2703 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2704 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2705 zio_prop_t czp = *zp;
2707 czp.zp_copies = ditto_copies;
2710 * If we arrived here with an override bp, we won't have run
2711 * the transform stack, so we won't have the data we need to
2712 * generate a child i/o. So, toss the override bp and restart.
2713 * This is safe, because using the override bp is just an
2714 * optimization; and it's rare, so the cost doesn't matter.
2716 if (zio->io_bp_override) {
2717 zio_pop_transforms(zio);
2718 zio->io_stage = ZIO_STAGE_OPEN;
2719 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2720 zio->io_bp_override = NULL;
2723 return (ZIO_PIPELINE_CONTINUE);
2726 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2727 zio->io_orig_size, zio->io_orig_size, &czp, NULL, NULL,
2728 NULL, zio_ddt_ditto_write_done, dde, zio->io_priority,
2729 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2731 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2732 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2735 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2736 if (ddp->ddp_phys_birth != 0)
2737 ddt_bp_fill(ddp, bp, txg);
2738 if (dde->dde_lead_zio[p] != NULL)
2739 zio_add_child(zio, dde->dde_lead_zio[p]);
2741 ddt_phys_addref(ddp);
2742 } else if (zio->io_bp_override) {
2743 ASSERT(bp->blk_birth == txg);
2744 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2745 ddt_phys_fill(ddp, bp);
2746 ddt_phys_addref(ddp);
2748 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2749 zio->io_orig_size, zio->io_orig_size, zp,
2750 zio_ddt_child_write_ready, NULL, NULL,
2751 zio_ddt_child_write_done, dde, zio->io_priority,
2752 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2754 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2755 dde->dde_lead_zio[p] = cio;
2765 return (ZIO_PIPELINE_CONTINUE);
2768 ddt_entry_t *freedde; /* for debugging */
2771 zio_ddt_free(zio_t *zio)
2773 spa_t *spa = zio->io_spa;
2774 blkptr_t *bp = zio->io_bp;
2775 ddt_t *ddt = ddt_select(spa, bp);
2779 ASSERT(BP_GET_DEDUP(bp));
2780 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2783 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2784 ddp = ddt_phys_select(dde, bp);
2785 ddt_phys_decref(ddp);
2788 return (ZIO_PIPELINE_CONTINUE);
2792 * ==========================================================================
2793 * Allocate and free blocks
2794 * ==========================================================================
2798 zio_io_to_allocate(spa_t *spa)
2802 ASSERT(MUTEX_HELD(&spa->spa_alloc_lock));
2804 zio = avl_first(&spa->spa_alloc_tree);
2808 ASSERT(IO_IS_ALLOCATING(zio));
2811 * Try to place a reservation for this zio. If we're unable to
2812 * reserve then we throttle.
2814 if (!metaslab_class_throttle_reserve(spa_normal_class(spa),
2815 zio->io_prop.zp_copies, zio, 0)) {
2819 avl_remove(&spa->spa_alloc_tree, zio);
2820 ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE);
2826 zio_dva_throttle(zio_t *zio)
2828 spa_t *spa = zio->io_spa;
2831 if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE ||
2832 !spa_normal_class(zio->io_spa)->mc_alloc_throttle_enabled ||
2833 zio->io_child_type == ZIO_CHILD_GANG ||
2834 zio->io_flags & ZIO_FLAG_NODATA) {
2835 return (ZIO_PIPELINE_CONTINUE);
2838 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2840 ASSERT3U(zio->io_queued_timestamp, >, 0);
2841 ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE);
2843 mutex_enter(&spa->spa_alloc_lock);
2845 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2846 avl_add(&spa->spa_alloc_tree, zio);
2848 nio = zio_io_to_allocate(zio->io_spa);
2849 mutex_exit(&spa->spa_alloc_lock);
2852 return (ZIO_PIPELINE_CONTINUE);
2855 ASSERT3U(nio->io_queued_timestamp, <=,
2856 zio->io_queued_timestamp);
2857 ASSERT(nio->io_stage == ZIO_STAGE_DVA_THROTTLE);
2859 * We are passing control to a new zio so make sure that
2860 * it is processed by a different thread. We do this to
2861 * avoid stack overflows that can occur when parents are
2862 * throttled and children are making progress. We allow
2863 * it to go to the head of the taskq since it's already
2866 zio_taskq_dispatch(nio, ZIO_TASKQ_ISSUE, B_TRUE);
2868 return (ZIO_PIPELINE_STOP);
2872 zio_allocate_dispatch(spa_t *spa)
2876 mutex_enter(&spa->spa_alloc_lock);
2877 zio = zio_io_to_allocate(spa);
2878 mutex_exit(&spa->spa_alloc_lock);
2882 ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE);
2883 ASSERT0(zio->io_error);
2884 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE);
2888 zio_dva_allocate(zio_t *zio)
2890 spa_t *spa = zio->io_spa;
2891 metaslab_class_t *mc = spa_normal_class(spa);
2892 blkptr_t *bp = zio->io_bp;
2896 if (zio->io_gang_leader == NULL) {
2897 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2898 zio->io_gang_leader = zio;
2901 ASSERT(BP_IS_HOLE(bp));
2902 ASSERT0(BP_GET_NDVAS(bp));
2903 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2904 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2905 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2907 if (zio->io_flags & ZIO_FLAG_NODATA) {
2908 flags |= METASLAB_DONT_THROTTLE;
2910 if (zio->io_flags & ZIO_FLAG_GANG_CHILD) {
2911 flags |= METASLAB_GANG_CHILD;
2913 if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE) {
2914 flags |= METASLAB_ASYNC_ALLOC;
2917 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2918 zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
2919 &zio->io_alloc_list, zio);
2922 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2923 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2925 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2926 return (zio_write_gang_block(zio));
2927 zio->io_error = error;
2930 return (ZIO_PIPELINE_CONTINUE);
2934 zio_dva_free(zio_t *zio)
2936 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2938 return (ZIO_PIPELINE_CONTINUE);
2942 zio_dva_claim(zio_t *zio)
2946 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2948 zio->io_error = error;
2950 return (ZIO_PIPELINE_CONTINUE);
2954 * Undo an allocation. This is used by zio_done() when an I/O fails
2955 * and we want to give back the block we just allocated.
2956 * This handles both normal blocks and gang blocks.
2959 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2961 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2962 ASSERT(zio->io_bp_override == NULL);
2964 if (!BP_IS_HOLE(bp))
2965 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2968 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2969 zio_dva_unallocate(zio, gn->gn_child[g],
2970 &gn->gn_gbh->zg_blkptr[g]);
2976 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2979 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2980 uint64_t size, boolean_t *slog)
2983 zio_alloc_list_t io_alloc_list;
2985 ASSERT(txg > spa_syncing_txg(spa));
2987 metaslab_trace_init(&io_alloc_list);
2988 error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1,
2989 txg, old_bp, METASLAB_HINTBP_AVOID, &io_alloc_list, NULL);
2993 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2994 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID,
2995 &io_alloc_list, NULL);
2999 metaslab_trace_fini(&io_alloc_list);
3002 BP_SET_LSIZE(new_bp, size);
3003 BP_SET_PSIZE(new_bp, size);
3004 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
3005 BP_SET_CHECKSUM(new_bp,
3006 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
3007 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
3008 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3009 BP_SET_LEVEL(new_bp, 0);
3010 BP_SET_DEDUP(new_bp, 0);
3011 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
3018 * Free an intent log block.
3021 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
3023 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
3024 ASSERT(!BP_IS_GANG(bp));
3026 zio_free(spa, txg, bp);
3030 * ==========================================================================
3031 * Read, write and delete to physical devices
3032 * ==========================================================================
3037 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3038 * stops after this stage and will resume upon I/O completion.
3039 * However, there are instances where the vdev layer may need to
3040 * continue the pipeline when an I/O was not issued. Since the I/O
3041 * that was sent to the vdev layer might be different than the one
3042 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3043 * force the underlying vdev layers to call either zio_execute() or
3044 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3047 zio_vdev_io_start(zio_t *zio)
3049 vdev_t *vd = zio->io_vd;
3051 spa_t *spa = zio->io_spa;
3054 ASSERT(zio->io_error == 0);
3055 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
3058 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3059 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
3062 * The mirror_ops handle multiple DVAs in a single BP.
3064 vdev_mirror_ops.vdev_op_io_start(zio);
3065 return (ZIO_PIPELINE_STOP);
3068 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE &&
3069 zio->io_priority == ZIO_PRIORITY_NOW) {
3070 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg);
3071 return (ZIO_PIPELINE_CONTINUE);
3074 ASSERT3P(zio->io_logical, !=, zio);
3077 * We keep track of time-sensitive I/Os so that the scan thread
3078 * can quickly react to certain workloads. In particular, we care
3079 * about non-scrubbing, top-level reads and writes with the following
3081 * - synchronous writes of user data to non-slog devices
3082 * - any reads of user data
3083 * When these conditions are met, adjust the timestamp of spa_last_io
3084 * which allows the scan thread to adjust its workload accordingly.
3086 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
3087 vd == vd->vdev_top && !vd->vdev_islog &&
3088 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
3089 zio->io_txg != spa_syncing_txg(spa)) {
3090 uint64_t old = spa->spa_last_io;
3091 uint64_t new = ddi_get_lbolt64();
3093 (void) atomic_cas_64(&spa->spa_last_io, old, new);
3096 align = 1ULL << vd->vdev_top->vdev_ashift;
3098 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
3099 P2PHASE(zio->io_size, align) != 0) {
3100 /* Transform logical writes to be a full physical block size. */
3101 uint64_t asize = P2ROUNDUP(zio->io_size, align);
3103 if (zio->io_type == ZIO_TYPE_READ ||
3104 zio->io_type == ZIO_TYPE_WRITE)
3105 abuf = zio_buf_alloc(asize);
3106 ASSERT(vd == vd->vdev_top);
3107 if (zio->io_type == ZIO_TYPE_WRITE) {
3108 bcopy(zio->io_data, abuf, zio->io_size);
3109 bzero(abuf + zio->io_size, asize - zio->io_size);
3111 zio_push_transform(zio, abuf, asize, abuf ? asize : 0,
3116 * If this is not a physical io, make sure that it is properly aligned
3117 * before proceeding.
3119 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
3120 ASSERT0(P2PHASE(zio->io_offset, align));
3121 ASSERT0(P2PHASE(zio->io_size, align));
3124 * For the physical io we allow alignment
3125 * to a logical block size.
3127 uint64_t log_align =
3128 1ULL << vd->vdev_top->vdev_logical_ashift;
3129 ASSERT0(P2PHASE(zio->io_offset, log_align));
3130 ASSERT0(P2PHASE(zio->io_size, log_align));
3133 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa));
3136 * If this is a repair I/O, and there's no self-healing involved --
3137 * that is, we're just resilvering what we expect to resilver --
3138 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3139 * This prevents spurious resilvering with nested replication.
3140 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
3141 * A is out of date, we'll read from C+D, then use the data to
3142 * resilver A+B -- but we don't actually want to resilver B, just A.
3143 * The top-level mirror has no way to know this, so instead we just
3144 * discard unnecessary repairs as we work our way down the vdev tree.
3145 * The same logic applies to any form of nested replication:
3146 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
3148 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3149 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
3150 zio->io_txg != 0 && /* not a delegated i/o */
3151 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
3152 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3153 zio_vdev_io_bypass(zio);
3154 return (ZIO_PIPELINE_CONTINUE);
3157 if (vd->vdev_ops->vdev_op_leaf) {
3158 switch (zio->io_type) {
3160 if (vdev_cache_read(zio))
3161 return (ZIO_PIPELINE_CONTINUE);
3163 case ZIO_TYPE_WRITE:
3165 if ((zio = vdev_queue_io(zio)) == NULL)
3166 return (ZIO_PIPELINE_STOP);
3168 if (!vdev_accessible(vd, zio)) {
3169 zio->io_error = SET_ERROR(ENXIO);
3171 return (ZIO_PIPELINE_STOP);
3176 * Note that we ignore repair writes for TRIM because they can
3177 * conflict with normal writes. This isn't an issue because, by
3178 * definition, we only repair blocks that aren't freed.
3180 if (zio->io_type == ZIO_TYPE_WRITE &&
3181 !(zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3182 !trim_map_write_start(zio))
3183 return (ZIO_PIPELINE_STOP);
3186 vd->vdev_ops->vdev_op_io_start(zio);
3187 return (ZIO_PIPELINE_STOP);
3191 zio_vdev_io_done(zio_t *zio)
3193 vdev_t *vd = zio->io_vd;
3194 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
3195 boolean_t unexpected_error = B_FALSE;
3197 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
3198 return (ZIO_PIPELINE_STOP);
3200 ASSERT(zio->io_type == ZIO_TYPE_READ ||
3201 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE);
3203 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3204 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE ||
3205 zio->io_type == ZIO_TYPE_FREE)) {
3207 if (zio->io_type == ZIO_TYPE_WRITE &&
3208 !(zio->io_flags & ZIO_FLAG_IO_REPAIR))
3209 trim_map_write_done(zio);
3211 vdev_queue_io_done(zio);
3213 if (zio->io_type == ZIO_TYPE_WRITE)
3214 vdev_cache_write(zio);
3216 if (zio_injection_enabled && zio->io_error == 0)
3217 zio->io_error = zio_handle_device_injection(vd,
3220 if (zio_injection_enabled && zio->io_error == 0)
3221 zio->io_error = zio_handle_label_injection(zio, EIO);
3223 if (zio->io_error) {
3224 if (zio->io_error == ENOTSUP &&
3225 zio->io_type == ZIO_TYPE_FREE) {
3226 /* Not all devices support TRIM. */
3227 } else if (!vdev_accessible(vd, zio)) {
3228 zio->io_error = SET_ERROR(ENXIO);
3230 unexpected_error = B_TRUE;
3235 ops->vdev_op_io_done(zio);
3237 if (unexpected_error)
3238 VERIFY(vdev_probe(vd, zio) == NULL);
3240 return (ZIO_PIPELINE_CONTINUE);
3244 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3245 * disk, and use that to finish the checksum ereport later.
3248 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
3249 const void *good_buf)
3251 /* no processing needed */
3252 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
3257 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
3259 void *buf = zio_buf_alloc(zio->io_size);
3261 bcopy(zio->io_data, buf, zio->io_size);
3263 zcr->zcr_cbinfo = zio->io_size;
3264 zcr->zcr_cbdata = buf;
3265 zcr->zcr_finish = zio_vsd_default_cksum_finish;
3266 zcr->zcr_free = zio_buf_free;
3270 zio_vdev_io_assess(zio_t *zio)
3272 vdev_t *vd = zio->io_vd;
3274 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
3275 return (ZIO_PIPELINE_STOP);
3277 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3278 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
3280 if (zio->io_vsd != NULL) {
3281 zio->io_vsd_ops->vsd_free(zio);
3285 if (zio_injection_enabled && zio->io_error == 0)
3286 zio->io_error = zio_handle_fault_injection(zio, EIO);
3288 if (zio->io_type == ZIO_TYPE_FREE &&
3289 zio->io_priority != ZIO_PRIORITY_NOW) {
3290 switch (zio->io_error) {
3292 ZIO_TRIM_STAT_INCR(bytes, zio->io_size);
3293 ZIO_TRIM_STAT_BUMP(success);
3296 ZIO_TRIM_STAT_BUMP(unsupported);
3299 ZIO_TRIM_STAT_BUMP(failed);
3305 * If the I/O failed, determine whether we should attempt to retry it.
3307 * On retry, we cut in line in the issue queue, since we don't want
3308 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3310 if (zio->io_error && vd == NULL &&
3311 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
3312 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
3313 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
3315 zio->io_flags |= ZIO_FLAG_IO_RETRY |
3316 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
3317 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
3318 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
3319 zio_requeue_io_start_cut_in_line);
3320 return (ZIO_PIPELINE_STOP);
3324 * If we got an error on a leaf device, convert it to ENXIO
3325 * if the device is not accessible at all.
3327 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3328 !vdev_accessible(vd, zio))
3329 zio->io_error = SET_ERROR(ENXIO);
3332 * If we can't write to an interior vdev (mirror or RAID-Z),
3333 * set vdev_cant_write so that we stop trying to allocate from it.
3335 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
3336 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
3337 vd->vdev_cant_write = B_TRUE;
3341 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
3342 * attempts will ever succeed. In this case we set a persistent bit so
3343 * that we don't bother with it in the future.
3345 if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) &&
3346 zio->io_type == ZIO_TYPE_IOCTL &&
3347 zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL)
3348 vd->vdev_nowritecache = B_TRUE;
3351 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3353 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3354 zio->io_physdone != NULL) {
3355 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
3356 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
3357 zio->io_physdone(zio->io_logical);
3360 return (ZIO_PIPELINE_CONTINUE);
3364 zio_vdev_io_reissue(zio_t *zio)
3366 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3367 ASSERT(zio->io_error == 0);
3369 zio->io_stage >>= 1;
3373 zio_vdev_io_redone(zio_t *zio)
3375 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
3377 zio->io_stage >>= 1;
3381 zio_vdev_io_bypass(zio_t *zio)
3383 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3384 ASSERT(zio->io_error == 0);
3386 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
3387 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
3391 * ==========================================================================
3392 * Generate and verify checksums
3393 * ==========================================================================
3396 zio_checksum_generate(zio_t *zio)
3398 blkptr_t *bp = zio->io_bp;
3399 enum zio_checksum checksum;
3403 * This is zio_write_phys().
3404 * We're either generating a label checksum, or none at all.
3406 checksum = zio->io_prop.zp_checksum;
3408 if (checksum == ZIO_CHECKSUM_OFF)
3409 return (ZIO_PIPELINE_CONTINUE);
3411 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
3413 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
3414 ASSERT(!IO_IS_ALLOCATING(zio));
3415 checksum = ZIO_CHECKSUM_GANG_HEADER;
3417 checksum = BP_GET_CHECKSUM(bp);
3421 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
3423 return (ZIO_PIPELINE_CONTINUE);
3427 zio_checksum_verify(zio_t *zio)
3429 zio_bad_cksum_t info;
3430 blkptr_t *bp = zio->io_bp;
3433 ASSERT(zio->io_vd != NULL);
3437 * This is zio_read_phys().
3438 * We're either verifying a label checksum, or nothing at all.
3440 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
3441 return (ZIO_PIPELINE_CONTINUE);
3443 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
3446 if ((error = zio_checksum_error(zio, &info)) != 0) {
3447 zio->io_error = error;
3448 if (error == ECKSUM &&
3449 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
3450 zfs_ereport_start_checksum(zio->io_spa,
3451 zio->io_vd, zio, zio->io_offset,
3452 zio->io_size, NULL, &info);
3456 return (ZIO_PIPELINE_CONTINUE);
3460 * Called by RAID-Z to ensure we don't compute the checksum twice.
3463 zio_checksum_verified(zio_t *zio)
3465 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
3469 * ==========================================================================
3470 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
3471 * An error of 0 indicates success. ENXIO indicates whole-device failure,
3472 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
3473 * indicate errors that are specific to one I/O, and most likely permanent.
3474 * Any other error is presumed to be worse because we weren't expecting it.
3475 * ==========================================================================
3478 zio_worst_error(int e1, int e2)
3480 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
3483 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
3484 if (e1 == zio_error_rank[r1])
3487 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
3488 if (e2 == zio_error_rank[r2])
3491 return (r1 > r2 ? e1 : e2);
3495 * ==========================================================================
3497 * ==========================================================================
3500 zio_ready(zio_t *zio)
3502 blkptr_t *bp = zio->io_bp;
3503 zio_t *pio, *pio_next;
3504 zio_link_t *zl = NULL;
3506 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
3507 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
3508 return (ZIO_PIPELINE_STOP);
3510 if (zio->io_ready) {
3511 ASSERT(IO_IS_ALLOCATING(zio));
3512 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
3513 (zio->io_flags & ZIO_FLAG_NOPWRITE));
3514 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
3519 if (bp != NULL && bp != &zio->io_bp_copy)
3520 zio->io_bp_copy = *bp;
3522 if (zio->io_error != 0) {
3523 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3525 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3526 ASSERT(IO_IS_ALLOCATING(zio));
3527 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3529 * We were unable to allocate anything, unreserve and
3530 * issue the next I/O to allocate.
3532 metaslab_class_throttle_unreserve(
3533 spa_normal_class(zio->io_spa),
3534 zio->io_prop.zp_copies, zio);
3535 zio_allocate_dispatch(zio->io_spa);
3539 mutex_enter(&zio->io_lock);
3540 zio->io_state[ZIO_WAIT_READY] = 1;
3541 pio = zio_walk_parents(zio, &zl);
3542 mutex_exit(&zio->io_lock);
3545 * As we notify zio's parents, new parents could be added.
3546 * New parents go to the head of zio's io_parent_list, however,
3547 * so we will (correctly) not notify them. The remainder of zio's
3548 * io_parent_list, from 'pio_next' onward, cannot change because
3549 * all parents must wait for us to be done before they can be done.
3551 for (; pio != NULL; pio = pio_next) {
3552 pio_next = zio_walk_parents(zio, &zl);
3553 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
3556 if (zio->io_flags & ZIO_FLAG_NODATA) {
3557 if (BP_IS_GANG(bp)) {
3558 zio->io_flags &= ~ZIO_FLAG_NODATA;
3560 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
3561 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
3565 if (zio_injection_enabled &&
3566 zio->io_spa->spa_syncing_txg == zio->io_txg)
3567 zio_handle_ignored_writes(zio);
3569 return (ZIO_PIPELINE_CONTINUE);
3573 * Update the allocation throttle accounting.
3576 zio_dva_throttle_done(zio_t *zio)
3578 zio_t *lio = zio->io_logical;
3579 zio_t *pio = zio_unique_parent(zio);
3580 vdev_t *vd = zio->io_vd;
3581 int flags = METASLAB_ASYNC_ALLOC;
3583 ASSERT3P(zio->io_bp, !=, NULL);
3584 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
3585 ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE);
3586 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
3588 ASSERT3P(vd, ==, vd->vdev_top);
3589 ASSERT(!(zio->io_flags & (ZIO_FLAG_IO_REPAIR | ZIO_FLAG_IO_RETRY)));
3590 ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING);
3591 ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE));
3592 ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA));
3595 * Parents of gang children can have two flavors -- ones that
3596 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
3597 * and ones that allocated the constituent blocks. The allocation
3598 * throttle needs to know the allocating parent zio so we must find
3601 if (pio->io_child_type == ZIO_CHILD_GANG) {
3603 * If our parent is a rewrite gang child then our grandparent
3604 * would have been the one that performed the allocation.
3606 if (pio->io_flags & ZIO_FLAG_IO_REWRITE)
3607 pio = zio_unique_parent(pio);
3608 flags |= METASLAB_GANG_CHILD;
3611 ASSERT(IO_IS_ALLOCATING(pio));
3612 ASSERT3P(zio, !=, zio->io_logical);
3613 ASSERT(zio->io_logical != NULL);
3614 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
3615 ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE);
3617 mutex_enter(&pio->io_lock);
3618 metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags);
3619 mutex_exit(&pio->io_lock);
3621 metaslab_class_throttle_unreserve(spa_normal_class(zio->io_spa),
3625 * Call into the pipeline to see if there is more work that
3626 * needs to be done. If there is work to be done it will be
3627 * dispatched to another taskq thread.
3629 zio_allocate_dispatch(zio->io_spa);
3633 zio_done(zio_t *zio)
3635 spa_t *spa = zio->io_spa;
3636 zio_t *lio = zio->io_logical;
3637 blkptr_t *bp = zio->io_bp;
3638 vdev_t *vd = zio->io_vd;
3639 uint64_t psize = zio->io_size;
3640 zio_t *pio, *pio_next;
3641 metaslab_class_t *mc = spa_normal_class(spa);
3642 zio_link_t *zl = NULL;
3645 * If our children haven't all completed,
3646 * wait for them and then repeat this pipeline stage.
3648 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
3649 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
3650 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
3651 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
3652 return (ZIO_PIPELINE_STOP);
3655 * If the allocation throttle is enabled, then update the accounting.
3656 * We only track child I/Os that are part of an allocating async
3657 * write. We must do this since the allocation is performed
3658 * by the logical I/O but the actual write is done by child I/Os.
3660 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING &&
3661 zio->io_child_type == ZIO_CHILD_VDEV) {
3662 ASSERT(mc->mc_alloc_throttle_enabled);
3663 zio_dva_throttle_done(zio);
3667 * If the allocation throttle is enabled, verify that
3668 * we have decremented the refcounts for every I/O that was throttled.
3670 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3671 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3672 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3674 metaslab_group_alloc_verify(spa, zio->io_bp, zio);
3675 VERIFY(refcount_not_held(&mc->mc_alloc_slots, zio));
3678 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3679 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3680 ASSERT(zio->io_children[c][w] == 0);
3682 if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
3683 ASSERT(bp->blk_pad[0] == 0);
3684 ASSERT(bp->blk_pad[1] == 0);
3685 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3686 (bp == zio_unique_parent(zio)->io_bp));
3687 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3688 zio->io_bp_override == NULL &&
3689 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3690 ASSERT(!BP_SHOULD_BYTESWAP(bp));
3691 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3692 ASSERT(BP_COUNT_GANG(bp) == 0 ||
3693 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3695 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3696 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3700 * If there were child vdev/gang/ddt errors, they apply to us now.
3702 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3703 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3704 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3707 * If the I/O on the transformed data was successful, generate any
3708 * checksum reports now while we still have the transformed data.
3710 if (zio->io_error == 0) {
3711 while (zio->io_cksum_report != NULL) {
3712 zio_cksum_report_t *zcr = zio->io_cksum_report;
3713 uint64_t align = zcr->zcr_align;
3714 uint64_t asize = P2ROUNDUP(psize, align);
3715 char *abuf = zio->io_data;
3717 if (asize != psize) {
3718 abuf = zio_buf_alloc(asize);
3719 bcopy(zio->io_data, abuf, psize);
3720 bzero(abuf + psize, asize - psize);
3723 zio->io_cksum_report = zcr->zcr_next;
3724 zcr->zcr_next = NULL;
3725 zcr->zcr_finish(zcr, abuf);
3726 zfs_ereport_free_checksum(zcr);
3729 zio_buf_free(abuf, asize);
3733 zio_pop_transforms(zio); /* note: may set zio->io_error */
3735 vdev_stat_update(zio, psize);
3737 if (zio->io_error) {
3739 * If this I/O is attached to a particular vdev,
3740 * generate an error message describing the I/O failure
3741 * at the block level. We ignore these errors if the
3742 * device is currently unavailable.
3744 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3745 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3747 if ((zio->io_error == EIO || !(zio->io_flags &
3748 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3751 * For logical I/O requests, tell the SPA to log the
3752 * error and generate a logical data ereport.
3754 spa_log_error(spa, zio);
3755 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3760 if (zio->io_error && zio == lio) {
3762 * Determine whether zio should be reexecuted. This will
3763 * propagate all the way to the root via zio_notify_parent().
3765 ASSERT(vd == NULL && bp != NULL);
3766 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3768 if (IO_IS_ALLOCATING(zio) &&
3769 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3770 if (zio->io_error != ENOSPC)
3771 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3773 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3776 if ((zio->io_type == ZIO_TYPE_READ ||
3777 zio->io_type == ZIO_TYPE_FREE) &&
3778 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3779 zio->io_error == ENXIO &&
3780 spa_load_state(spa) == SPA_LOAD_NONE &&
3781 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3782 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3784 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3785 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3788 * Here is a possibly good place to attempt to do
3789 * either combinatorial reconstruction or error correction
3790 * based on checksums. It also might be a good place
3791 * to send out preliminary ereports before we suspend
3797 * If there were logical child errors, they apply to us now.
3798 * We defer this until now to avoid conflating logical child
3799 * errors with errors that happened to the zio itself when
3800 * updating vdev stats and reporting FMA events above.
3802 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3804 if ((zio->io_error || zio->io_reexecute) &&
3805 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3806 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3807 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3809 zio_gang_tree_free(&zio->io_gang_tree);
3812 * Godfather I/Os should never suspend.
3814 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3815 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3816 zio->io_reexecute = 0;
3818 if (zio->io_reexecute) {
3820 * This is a logical I/O that wants to reexecute.
3822 * Reexecute is top-down. When an i/o fails, if it's not
3823 * the root, it simply notifies its parent and sticks around.
3824 * The parent, seeing that it still has children in zio_done(),
3825 * does the same. This percolates all the way up to the root.
3826 * The root i/o will reexecute or suspend the entire tree.
3828 * This approach ensures that zio_reexecute() honors
3829 * all the original i/o dependency relationships, e.g.
3830 * parents not executing until children are ready.
3832 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3834 zio->io_gang_leader = NULL;
3836 mutex_enter(&zio->io_lock);
3837 zio->io_state[ZIO_WAIT_DONE] = 1;
3838 mutex_exit(&zio->io_lock);
3841 * "The Godfather" I/O monitors its children but is
3842 * not a true parent to them. It will track them through
3843 * the pipeline but severs its ties whenever they get into
3844 * trouble (e.g. suspended). This allows "The Godfather"
3845 * I/O to return status without blocking.
3848 for (pio = zio_walk_parents(zio, &zl); pio != NULL;
3850 zio_link_t *remove_zl = zl;
3851 pio_next = zio_walk_parents(zio, &zl);
3853 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3854 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3855 zio_remove_child(pio, zio, remove_zl);
3856 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3860 if ((pio = zio_unique_parent(zio)) != NULL) {
3862 * We're not a root i/o, so there's nothing to do
3863 * but notify our parent. Don't propagate errors
3864 * upward since we haven't permanently failed yet.
3866 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3867 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3868 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3869 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3871 * We'd fail again if we reexecuted now, so suspend
3872 * until conditions improve (e.g. device comes online).
3874 zio_suspend(spa, zio);
3877 * Reexecution is potentially a huge amount of work.
3878 * Hand it off to the otherwise-unused claim taskq.
3880 #if defined(illumos) || !defined(_KERNEL)
3881 ASSERT(zio->io_tqent.tqent_next == NULL);
3883 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
3885 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3886 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3889 return (ZIO_PIPELINE_STOP);
3892 ASSERT(zio->io_child_count == 0);
3893 ASSERT(zio->io_reexecute == 0);
3894 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3897 * Report any checksum errors, since the I/O is complete.
3899 while (zio->io_cksum_report != NULL) {
3900 zio_cksum_report_t *zcr = zio->io_cksum_report;
3901 zio->io_cksum_report = zcr->zcr_next;
3902 zcr->zcr_next = NULL;
3903 zcr->zcr_finish(zcr, NULL);
3904 zfs_ereport_free_checksum(zcr);
3908 * It is the responsibility of the done callback to ensure that this
3909 * particular zio is no longer discoverable for adoption, and as
3910 * such, cannot acquire any new parents.
3915 mutex_enter(&zio->io_lock);
3916 zio->io_state[ZIO_WAIT_DONE] = 1;
3917 mutex_exit(&zio->io_lock);
3920 for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) {
3921 zio_link_t *remove_zl = zl;
3922 pio_next = zio_walk_parents(zio, &zl);
3923 zio_remove_child(pio, zio, remove_zl);
3924 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3927 if (zio->io_waiter != NULL) {
3928 mutex_enter(&zio->io_lock);
3929 zio->io_executor = NULL;
3930 cv_broadcast(&zio->io_cv);
3931 mutex_exit(&zio->io_lock);
3936 return (ZIO_PIPELINE_STOP);
3940 * ==========================================================================
3941 * I/O pipeline definition
3942 * ==========================================================================
3944 static zio_pipe_stage_t *zio_pipeline[] = {
3951 zio_checksum_generate,
3967 zio_checksum_verify,
3975 * Compare two zbookmark_phys_t's to see which we would reach first in a
3976 * pre-order traversal of the object tree.
3978 * This is simple in every case aside from the meta-dnode object. For all other
3979 * objects, we traverse them in order (object 1 before object 2, and so on).
3980 * However, all of these objects are traversed while traversing object 0, since
3981 * the data it points to is the list of objects. Thus, we need to convert to a
3982 * canonical representation so we can compare meta-dnode bookmarks to
3983 * non-meta-dnode bookmarks.
3985 * We do this by calculating "equivalents" for each field of the zbookmark.
3986 * zbookmarks outside of the meta-dnode use their own object and level, and
3987 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
3988 * blocks this bookmark refers to) by multiplying their blkid by their span
3989 * (the number of L0 blocks contained within one block at their level).
3990 * zbookmarks inside the meta-dnode calculate their object equivalent
3991 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
3992 * level + 1<<31 (any value larger than a level could ever be) for their level.
3993 * This causes them to always compare before a bookmark in their object
3994 * equivalent, compare appropriately to bookmarks in other objects, and to
3995 * compare appropriately to other bookmarks in the meta-dnode.
3998 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
3999 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
4002 * These variables represent the "equivalent" values for the zbookmark,
4003 * after converting zbookmarks inside the meta dnode to their
4004 * normal-object equivalents.
4006 uint64_t zb1obj, zb2obj;
4007 uint64_t zb1L0, zb2L0;
4008 uint64_t zb1level, zb2level;
4010 if (zb1->zb_object == zb2->zb_object &&
4011 zb1->zb_level == zb2->zb_level &&
4012 zb1->zb_blkid == zb2->zb_blkid)
4016 * BP_SPANB calculates the span in blocks.
4018 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
4019 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
4021 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
4022 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4024 zb1level = zb1->zb_level + COMPARE_META_LEVEL;
4026 zb1obj = zb1->zb_object;
4027 zb1level = zb1->zb_level;
4030 if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
4031 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4033 zb2level = zb2->zb_level + COMPARE_META_LEVEL;
4035 zb2obj = zb2->zb_object;
4036 zb2level = zb2->zb_level;
4039 /* Now that we have a canonical representation, do the comparison. */
4040 if (zb1obj != zb2obj)
4041 return (zb1obj < zb2obj ? -1 : 1);
4042 else if (zb1L0 != zb2L0)
4043 return (zb1L0 < zb2L0 ? -1 : 1);
4044 else if (zb1level != zb2level)
4045 return (zb1level > zb2level ? -1 : 1);
4047 * This can (theoretically) happen if the bookmarks have the same object
4048 * and level, but different blkids, if the block sizes are not the same.
4049 * There is presently no way to change the indirect block sizes
4055 * This function checks the following: given that last_block is the place that
4056 * our traversal stopped last time, does that guarantee that we've visited
4057 * every node under subtree_root? Therefore, we can't just use the raw output
4058 * of zbookmark_compare. We have to pass in a modified version of
4059 * subtree_root; by incrementing the block id, and then checking whether
4060 * last_block is before or equal to that, we can tell whether or not having
4061 * visited last_block implies that all of subtree_root's children have been
4065 zbookmark_subtree_completed(const dnode_phys_t *dnp,
4066 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
4068 zbookmark_phys_t mod_zb = *subtree_root;
4070 ASSERT(last_block->zb_level == 0);
4072 /* The objset_phys_t isn't before anything. */
4077 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
4078 * data block size in sectors, because that variable is only used if
4079 * the bookmark refers to a block in the meta-dnode. Since we don't
4080 * know without examining it what object it refers to, and there's no
4081 * harm in passing in this value in other cases, we always pass it in.
4083 * We pass in 0 for the indirect block size shift because zb2 must be
4084 * level 0. The indirect block size is only used to calculate the span
4085 * of the bookmark, but since the bookmark must be level 0, the span is
4086 * always 1, so the math works out.
4088 * If you make changes to how the zbookmark_compare code works, be sure
4089 * to make sure that this code still works afterwards.
4091 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
4092 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,