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, 2015 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
27 #include <sys/sysmacros.h>
28 #include <sys/zfs_context.h>
29 #include <sys/fm/fs/zfs.h>
32 #include <sys/spa_impl.h>
33 #include <sys/vdev_impl.h>
34 #include <sys/zio_impl.h>
35 #include <sys/zio_compress.h>
36 #include <sys/zio_checksum.h>
37 #include <sys/dmu_objset.h>
40 #include <sys/trim_map.h>
41 #include <sys/blkptr.h>
42 #include <sys/zfeature.h>
44 SYSCTL_DECL(_vfs_zfs);
45 SYSCTL_NODE(_vfs_zfs, OID_AUTO, zio, CTLFLAG_RW, 0, "ZFS ZIO");
46 #if defined(__amd64__)
47 static int zio_use_uma = 1;
49 static int zio_use_uma = 0;
51 TUNABLE_INT("vfs.zfs.zio.use_uma", &zio_use_uma);
52 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, use_uma, CTLFLAG_RDTUN, &zio_use_uma, 0,
53 "Use uma(9) for ZIO allocations");
54 static int zio_exclude_metadata = 0;
55 TUNABLE_INT("vfs.zfs.zio.exclude_metadata", &zio_exclude_metadata);
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",
83 * ==========================================================================
85 * ==========================================================================
87 kmem_cache_t *zio_cache;
88 kmem_cache_t *zio_link_cache;
89 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
90 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
93 extern vmem_t *zio_alloc_arena;
96 #define BP_SPANB(indblkshift, level) \
97 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
98 #define COMPARE_META_LEVEL 0x80000000ul
100 * The following actions directly effect the spa's sync-to-convergence logic.
101 * The values below define the sync pass when we start performing the action.
102 * Care should be taken when changing these values as they directly impact
103 * spa_sync() performance. Tuning these values may introduce subtle performance
104 * pathologies and should only be done in the context of performance analysis.
105 * These tunables will eventually be removed and replaced with #defines once
106 * enough analysis has been done to determine optimal values.
108 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
109 * regular blocks are not deferred.
111 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
112 TUNABLE_INT("vfs.zfs.sync_pass_deferred_free", &zfs_sync_pass_deferred_free);
113 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_deferred_free, CTLFLAG_RDTUN,
114 &zfs_sync_pass_deferred_free, 0, "defer frees starting in this pass");
115 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
116 TUNABLE_INT("vfs.zfs.sync_pass_dont_compress", &zfs_sync_pass_dont_compress);
117 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_dont_compress, CTLFLAG_RDTUN,
118 &zfs_sync_pass_dont_compress, 0, "don't compress starting in this pass");
119 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
120 TUNABLE_INT("vfs.zfs.sync_pass_rewrite", &zfs_sync_pass_rewrite);
121 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_rewrite, CTLFLAG_RDTUN,
122 &zfs_sync_pass_rewrite, 0, "rewrite new bps starting in this pass");
125 * An allocating zio is one that either currently has the DVA allocate
126 * stage set or will have it later in its lifetime.
128 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
130 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
133 int zio_buf_debug_limit = 16384;
135 int zio_buf_debug_limit = 0;
142 zio_cache = kmem_cache_create("zio_cache",
143 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
144 zio_link_cache = kmem_cache_create("zio_link_cache",
145 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
150 * For small buffers, we want a cache for each multiple of
151 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
152 * for each quarter-power of 2.
154 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
155 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
158 size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
166 * If we are using watchpoints, put each buffer on its own page,
167 * to eliminate the performance overhead of trapping to the
168 * kernel when modifying a non-watched buffer that shares the
169 * page with a watched buffer.
171 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
175 if (size <= 4 * SPA_MINBLOCKSIZE) {
176 align = SPA_MINBLOCKSIZE;
177 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
178 align = MIN(p2 >> 2, PAGESIZE);
183 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
184 zio_buf_cache[c] = kmem_cache_create(name, size,
185 align, NULL, NULL, NULL, NULL, NULL, cflags);
188 * Since zio_data bufs do not appear in crash dumps, we
189 * pass KMC_NOTOUCH so that no allocator metadata is
190 * stored with the buffers.
192 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
193 zio_data_buf_cache[c] = kmem_cache_create(name, size,
194 align, NULL, NULL, NULL, NULL, NULL,
195 cflags | KMC_NOTOUCH | KMC_NODEBUG);
200 ASSERT(zio_buf_cache[c] != NULL);
201 if (zio_buf_cache[c - 1] == NULL)
202 zio_buf_cache[c - 1] = zio_buf_cache[c];
204 ASSERT(zio_data_buf_cache[c] != NULL);
205 if (zio_data_buf_cache[c - 1] == NULL)
206 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
212 zio_trim_ksp = kstat_create("zfs", 0, "zio_trim", "misc",
214 sizeof(zio_trim_stats) / sizeof(kstat_named_t),
217 if (zio_trim_ksp != NULL) {
218 zio_trim_ksp->ks_data = &zio_trim_stats;
219 kstat_install(zio_trim_ksp);
227 kmem_cache_t *last_cache = NULL;
228 kmem_cache_t *last_data_cache = NULL;
230 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
231 if (zio_buf_cache[c] != last_cache) {
232 last_cache = zio_buf_cache[c];
233 kmem_cache_destroy(zio_buf_cache[c]);
235 zio_buf_cache[c] = NULL;
237 if (zio_data_buf_cache[c] != last_data_cache) {
238 last_data_cache = zio_data_buf_cache[c];
239 kmem_cache_destroy(zio_data_buf_cache[c]);
241 zio_data_buf_cache[c] = NULL;
244 kmem_cache_destroy(zio_link_cache);
245 kmem_cache_destroy(zio_cache);
249 if (zio_trim_ksp != NULL) {
250 kstat_delete(zio_trim_ksp);
256 * ==========================================================================
257 * Allocate and free I/O buffers
258 * ==========================================================================
262 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
263 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
264 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
265 * excess / transient data in-core during a crashdump.
268 zio_buf_alloc(size_t size)
270 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
271 int flags = zio_exclude_metadata ? KM_NODEBUG : 0;
273 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
276 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
278 return (kmem_alloc(size, KM_SLEEP|flags));
282 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
283 * crashdump if the kernel panics. This exists so that we will limit the amount
284 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
285 * of kernel heap dumped to disk when the kernel panics)
288 zio_data_buf_alloc(size_t size)
290 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
292 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
295 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
297 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG));
301 zio_buf_free(void *buf, size_t size)
303 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
305 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
308 kmem_cache_free(zio_buf_cache[c], buf);
310 kmem_free(buf, size);
314 zio_data_buf_free(void *buf, size_t size)
316 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
318 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
321 kmem_cache_free(zio_data_buf_cache[c], buf);
323 kmem_free(buf, size);
327 * ==========================================================================
328 * Push and pop I/O transform buffers
329 * ==========================================================================
332 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
333 zio_transform_func_t *transform)
335 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
337 zt->zt_orig_data = zio->io_data;
338 zt->zt_orig_size = zio->io_size;
339 zt->zt_bufsize = bufsize;
340 zt->zt_transform = transform;
342 zt->zt_next = zio->io_transform_stack;
343 zio->io_transform_stack = zt;
350 zio_pop_transforms(zio_t *zio)
354 while ((zt = zio->io_transform_stack) != NULL) {
355 if (zt->zt_transform != NULL)
356 zt->zt_transform(zio,
357 zt->zt_orig_data, zt->zt_orig_size);
359 if (zt->zt_bufsize != 0)
360 zio_buf_free(zio->io_data, zt->zt_bufsize);
362 zio->io_data = zt->zt_orig_data;
363 zio->io_size = zt->zt_orig_size;
364 zio->io_transform_stack = zt->zt_next;
366 kmem_free(zt, sizeof (zio_transform_t));
371 * ==========================================================================
372 * I/O transform callbacks for subblocks and decompression
373 * ==========================================================================
376 zio_subblock(zio_t *zio, void *data, uint64_t size)
378 ASSERT(zio->io_size > size);
380 if (zio->io_type == ZIO_TYPE_READ)
381 bcopy(zio->io_data, data, size);
385 zio_decompress(zio_t *zio, void *data, uint64_t size)
387 if (zio->io_error == 0 &&
388 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
389 zio->io_data, data, zio->io_size, size) != 0)
390 zio->io_error = SET_ERROR(EIO);
394 * ==========================================================================
395 * I/O parent/child relationships and pipeline interlocks
396 * ==========================================================================
399 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
400 * continue calling these functions until they return NULL.
401 * Otherwise, the next caller will pick up the list walk in
402 * some indeterminate state. (Otherwise every caller would
403 * have to pass in a cookie to keep the state represented by
404 * io_walk_link, which gets annoying.)
407 zio_walk_parents(zio_t *cio)
409 zio_link_t *zl = cio->io_walk_link;
410 list_t *pl = &cio->io_parent_list;
412 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
413 cio->io_walk_link = zl;
418 ASSERT(zl->zl_child == cio);
419 return (zl->zl_parent);
423 zio_walk_children(zio_t *pio)
425 zio_link_t *zl = pio->io_walk_link;
426 list_t *cl = &pio->io_child_list;
428 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
429 pio->io_walk_link = zl;
434 ASSERT(zl->zl_parent == pio);
435 return (zl->zl_child);
439 zio_unique_parent(zio_t *cio)
441 zio_t *pio = zio_walk_parents(cio);
443 VERIFY(zio_walk_parents(cio) == NULL);
448 zio_add_child(zio_t *pio, zio_t *cio)
450 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
453 * Logical I/Os can have logical, gang, or vdev children.
454 * Gang I/Os can have gang or vdev children.
455 * Vdev I/Os can only have vdev children.
456 * The following ASSERT captures all of these constraints.
458 ASSERT(cio->io_child_type <= pio->io_child_type);
463 mutex_enter(&cio->io_lock);
464 mutex_enter(&pio->io_lock);
466 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
468 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
469 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
471 list_insert_head(&pio->io_child_list, zl);
472 list_insert_head(&cio->io_parent_list, zl);
474 pio->io_child_count++;
475 cio->io_parent_count++;
477 mutex_exit(&pio->io_lock);
478 mutex_exit(&cio->io_lock);
482 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
484 ASSERT(zl->zl_parent == pio);
485 ASSERT(zl->zl_child == cio);
487 mutex_enter(&cio->io_lock);
488 mutex_enter(&pio->io_lock);
490 list_remove(&pio->io_child_list, zl);
491 list_remove(&cio->io_parent_list, zl);
493 pio->io_child_count--;
494 cio->io_parent_count--;
496 mutex_exit(&pio->io_lock);
497 mutex_exit(&cio->io_lock);
499 kmem_cache_free(zio_link_cache, zl);
503 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
505 uint64_t *countp = &zio->io_children[child][wait];
506 boolean_t waiting = B_FALSE;
508 mutex_enter(&zio->io_lock);
509 ASSERT(zio->io_stall == NULL);
512 zio->io_stall = countp;
515 mutex_exit(&zio->io_lock);
521 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
523 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
524 int *errorp = &pio->io_child_error[zio->io_child_type];
526 mutex_enter(&pio->io_lock);
527 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
528 *errorp = zio_worst_error(*errorp, zio->io_error);
529 pio->io_reexecute |= zio->io_reexecute;
530 ASSERT3U(*countp, >, 0);
534 if (*countp == 0 && pio->io_stall == countp) {
535 pio->io_stall = NULL;
536 mutex_exit(&pio->io_lock);
539 mutex_exit(&pio->io_lock);
544 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
546 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
547 zio->io_error = zio->io_child_error[c];
551 * ==========================================================================
552 * Create the various types of I/O (read, write, free, etc)
553 * ==========================================================================
556 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
557 void *data, uint64_t size, zio_done_func_t *done, void *private,
558 zio_type_t type, zio_priority_t priority, enum zio_flag flags,
559 vdev_t *vd, uint64_t offset, const zbookmark_phys_t *zb,
560 enum zio_stage stage, enum zio_stage pipeline)
564 ASSERT3U(type == ZIO_TYPE_FREE || size, <=, SPA_MAXBLOCKSIZE);
565 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
566 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
568 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
569 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
570 ASSERT(vd || stage == ZIO_STAGE_OPEN);
572 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
573 bzero(zio, sizeof (zio_t));
575 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
576 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
578 list_create(&zio->io_parent_list, sizeof (zio_link_t),
579 offsetof(zio_link_t, zl_parent_node));
580 list_create(&zio->io_child_list, sizeof (zio_link_t),
581 offsetof(zio_link_t, zl_child_node));
584 zio->io_child_type = ZIO_CHILD_VDEV;
585 else if (flags & ZIO_FLAG_GANG_CHILD)
586 zio->io_child_type = ZIO_CHILD_GANG;
587 else if (flags & ZIO_FLAG_DDT_CHILD)
588 zio->io_child_type = ZIO_CHILD_DDT;
590 zio->io_child_type = ZIO_CHILD_LOGICAL;
593 zio->io_bp = (blkptr_t *)bp;
594 zio->io_bp_copy = *bp;
595 zio->io_bp_orig = *bp;
596 if (type != ZIO_TYPE_WRITE ||
597 zio->io_child_type == ZIO_CHILD_DDT)
598 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
599 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
600 zio->io_logical = zio;
601 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
602 pipeline |= ZIO_GANG_STAGES;
608 zio->io_private = private;
610 zio->io_priority = priority;
612 zio->io_offset = offset;
613 zio->io_orig_data = zio->io_data = data;
614 zio->io_orig_size = zio->io_size = size;
615 zio->io_orig_flags = zio->io_flags = flags;
616 zio->io_orig_stage = zio->io_stage = stage;
617 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
619 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
620 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
623 zio->io_bookmark = *zb;
626 if (zio->io_logical == NULL)
627 zio->io_logical = pio->io_logical;
628 if (zio->io_child_type == ZIO_CHILD_GANG)
629 zio->io_gang_leader = pio->io_gang_leader;
630 zio_add_child(pio, zio);
637 zio_destroy(zio_t *zio)
639 list_destroy(&zio->io_parent_list);
640 list_destroy(&zio->io_child_list);
641 mutex_destroy(&zio->io_lock);
642 cv_destroy(&zio->io_cv);
643 kmem_cache_free(zio_cache, zio);
647 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
648 void *private, enum zio_flag flags)
652 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
653 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
654 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
660 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
662 return (zio_null(NULL, spa, NULL, done, private, flags));
666 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
668 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
669 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
670 bp, (longlong_t)BP_GET_TYPE(bp));
672 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
673 BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
674 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
675 bp, (longlong_t)BP_GET_CHECKSUM(bp));
677 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
678 BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
679 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
680 bp, (longlong_t)BP_GET_COMPRESS(bp));
682 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
683 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
684 bp, (longlong_t)BP_GET_LSIZE(bp));
686 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
687 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
688 bp, (longlong_t)BP_GET_PSIZE(bp));
691 if (BP_IS_EMBEDDED(bp)) {
692 if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
693 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
694 bp, (longlong_t)BPE_GET_ETYPE(bp));
699 * Pool-specific checks.
701 * Note: it would be nice to verify that the blk_birth and
702 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
703 * allows the birth time of log blocks (and dmu_sync()-ed blocks
704 * that are in the log) to be arbitrarily large.
706 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
707 uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
708 if (vdevid >= spa->spa_root_vdev->vdev_children) {
709 zfs_panic_recover("blkptr at %p DVA %u has invalid "
711 bp, i, (longlong_t)vdevid);
714 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
716 zfs_panic_recover("blkptr at %p DVA %u has invalid "
718 bp, i, (longlong_t)vdevid);
721 if (vd->vdev_ops == &vdev_hole_ops) {
722 zfs_panic_recover("blkptr at %p DVA %u has hole "
724 bp, i, (longlong_t)vdevid);
727 if (vd->vdev_ops == &vdev_missing_ops) {
729 * "missing" vdevs are valid during import, but we
730 * don't have their detailed info (e.g. asize), so
731 * we can't perform any more checks on them.
735 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
736 uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
738 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
739 if (offset + asize > vd->vdev_asize) {
740 zfs_panic_recover("blkptr at %p DVA %u has invalid "
742 bp, i, (longlong_t)offset);
748 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
749 void *data, uint64_t size, zio_done_func_t *done, void *private,
750 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
754 zfs_blkptr_verify(spa, bp);
756 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
757 data, size, done, private,
758 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
759 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
760 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
766 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
767 void *data, uint64_t size, const zio_prop_t *zp,
768 zio_done_func_t *ready, zio_done_func_t *physdone, zio_done_func_t *done,
770 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
774 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
775 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
776 zp->zp_compress >= ZIO_COMPRESS_OFF &&
777 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
778 DMU_OT_IS_VALID(zp->zp_type) &&
781 zp->zp_copies <= spa_max_replication(spa));
783 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
784 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
785 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
786 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
788 zio->io_ready = ready;
789 zio->io_physdone = physdone;
793 * Data can be NULL if we are going to call zio_write_override() to
794 * provide the already-allocated BP. But we may need the data to
795 * verify a dedup hit (if requested). In this case, don't try to
796 * dedup (just take the already-allocated BP verbatim).
798 if (data == NULL && zio->io_prop.zp_dedup_verify) {
799 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
806 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
807 uint64_t size, zio_done_func_t *done, void *private,
808 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
812 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
813 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
814 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
820 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
822 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
823 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
824 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
825 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
828 * We must reset the io_prop to match the values that existed
829 * when the bp was first written by dmu_sync() keeping in mind
830 * that nopwrite and dedup are mutually exclusive.
832 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
833 zio->io_prop.zp_nopwrite = nopwrite;
834 zio->io_prop.zp_copies = copies;
835 zio->io_bp_override = bp;
839 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
843 * The check for EMBEDDED is a performance optimization. We
844 * process the free here (by ignoring it) rather than
845 * putting it on the list and then processing it in zio_free_sync().
847 if (BP_IS_EMBEDDED(bp))
849 metaslab_check_free(spa, bp);
852 * Frees that are for the currently-syncing txg, are not going to be
853 * deferred, and which will not need to do a read (i.e. not GANG or
854 * DEDUP), can be processed immediately. Otherwise, put them on the
855 * in-memory list for later processing.
857 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
858 txg != spa->spa_syncing_txg ||
859 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
860 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
862 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp,
863 BP_GET_PSIZE(bp), 0)));
868 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
869 uint64_t size, enum zio_flag flags)
872 enum zio_stage stage = ZIO_FREE_PIPELINE;
874 ASSERT(!BP_IS_HOLE(bp));
875 ASSERT(spa_syncing_txg(spa) == txg);
876 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
878 if (BP_IS_EMBEDDED(bp))
879 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
881 metaslab_check_free(spa, bp);
884 if (zfs_trim_enabled)
885 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START |
886 ZIO_STAGE_VDEV_IO_ASSESS;
888 * GANG and DEDUP blocks can induce a read (for the gang block header,
889 * or the DDT), so issue them asynchronously so that this thread is
892 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
893 stage |= ZIO_STAGE_ISSUE_ASYNC;
895 flags |= ZIO_FLAG_DONT_QUEUE;
897 zio = zio_create(pio, spa, txg, bp, NULL, size,
898 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags,
899 NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
905 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
906 zio_done_func_t *done, void *private, enum zio_flag flags)
910 dprintf_bp(bp, "claiming in txg %llu", txg);
912 if (BP_IS_EMBEDDED(bp))
913 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
916 * A claim is an allocation of a specific block. Claims are needed
917 * to support immediate writes in the intent log. The issue is that
918 * immediate writes contain committed data, but in a txg that was
919 * *not* committed. Upon opening the pool after an unclean shutdown,
920 * the intent log claims all blocks that contain immediate write data
921 * so that the SPA knows they're in use.
923 * All claims *must* be resolved in the first txg -- before the SPA
924 * starts allocating blocks -- so that nothing is allocated twice.
925 * If txg == 0 we just verify that the block is claimable.
927 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
928 ASSERT(txg == spa_first_txg(spa) || txg == 0);
929 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
931 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
932 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
933 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
939 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset,
940 uint64_t size, zio_done_func_t *done, void *private,
941 zio_priority_t priority, enum zio_flag flags)
946 if (vd->vdev_children == 0) {
947 zio = zio_create(pio, spa, 0, NULL, NULL, size, done, private,
948 ZIO_TYPE_IOCTL, priority, flags, vd, offset, NULL,
949 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
953 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
955 for (c = 0; c < vd->vdev_children; c++)
956 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
957 offset, size, done, private, priority, flags));
964 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
965 void *data, int checksum, zio_done_func_t *done, void *private,
966 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
970 ASSERT(vd->vdev_children == 0);
971 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
972 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
973 ASSERT3U(offset + size, <=, vd->vdev_psize);
975 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
976 ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
977 NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
979 zio->io_prop.zp_checksum = checksum;
985 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
986 void *data, int checksum, zio_done_func_t *done, void *private,
987 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
991 ASSERT(vd->vdev_children == 0);
992 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
993 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
994 ASSERT3U(offset + size, <=, vd->vdev_psize);
996 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
997 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
998 NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
1000 zio->io_prop.zp_checksum = checksum;
1002 if (zio_checksum_table[checksum].ci_eck) {
1004 * zec checksums are necessarily destructive -- they modify
1005 * the end of the write buffer to hold the verifier/checksum.
1006 * Therefore, we must make a local copy in case the data is
1007 * being written to multiple places in parallel.
1009 void *wbuf = zio_buf_alloc(size);
1010 bcopy(data, wbuf, size);
1011 zio_push_transform(zio, wbuf, size, size, NULL);
1018 * Create a child I/O to do some work for us.
1021 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1022 void *data, uint64_t size, int type, zio_priority_t priority,
1023 enum zio_flag flags, zio_done_func_t *done, void *private)
1025 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1028 ASSERT(vd->vdev_parent ==
1029 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
1031 if (type == ZIO_TYPE_READ && bp != NULL) {
1033 * If we have the bp, then the child should perform the
1034 * checksum and the parent need not. This pushes error
1035 * detection as close to the leaves as possible and
1036 * eliminates redundant checksums in the interior nodes.
1038 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1039 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1042 /* Not all IO types require vdev io done stage e.g. free */
1043 if (!(pio->io_pipeline & ZIO_STAGE_VDEV_IO_DONE))
1044 pipeline &= ~ZIO_STAGE_VDEV_IO_DONE;
1046 if (vd->vdev_children == 0)
1047 offset += VDEV_LABEL_START_SIZE;
1049 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
1052 * If we've decided to do a repair, the write is not speculative --
1053 * even if the original read was.
1055 if (flags & ZIO_FLAG_IO_REPAIR)
1056 flags &= ~ZIO_FLAG_SPECULATIVE;
1058 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
1059 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1060 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1062 zio->io_physdone = pio->io_physdone;
1063 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1064 zio->io_logical->io_phys_children++;
1070 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
1071 int type, zio_priority_t priority, enum zio_flag flags,
1072 zio_done_func_t *done, void *private)
1076 ASSERT(vd->vdev_ops->vdev_op_leaf);
1078 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1079 data, size, done, private, type, priority,
1080 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1082 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1088 zio_flush(zio_t *zio, vdev_t *vd)
1090 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0,
1091 NULL, NULL, ZIO_PRIORITY_NOW,
1092 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1096 zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size)
1099 ASSERT(vd->vdev_ops->vdev_op_leaf);
1101 return (zio_create(zio, spa, 0, NULL, NULL, size, NULL, NULL,
1102 ZIO_TYPE_FREE, ZIO_PRIORITY_TRIM, ZIO_FLAG_DONT_AGGREGATE |
1103 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY,
1104 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PHYS_PIPELINE));
1108 zio_shrink(zio_t *zio, uint64_t size)
1110 ASSERT(zio->io_executor == NULL);
1111 ASSERT(zio->io_orig_size == zio->io_size);
1112 ASSERT(size <= zio->io_size);
1115 * We don't shrink for raidz because of problems with the
1116 * reconstruction when reading back less than the block size.
1117 * Note, BP_IS_RAIDZ() assumes no compression.
1119 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1120 if (!BP_IS_RAIDZ(zio->io_bp))
1121 zio->io_orig_size = zio->io_size = size;
1125 * ==========================================================================
1126 * Prepare to read and write logical blocks
1127 * ==========================================================================
1131 zio_read_bp_init(zio_t *zio)
1133 blkptr_t *bp = zio->io_bp;
1135 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1136 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1137 !(zio->io_flags & ZIO_FLAG_RAW)) {
1139 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1140 void *cbuf = zio_buf_alloc(psize);
1142 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
1145 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1146 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1147 decode_embedded_bp_compressed(bp, zio->io_data);
1149 ASSERT(!BP_IS_EMBEDDED(bp));
1152 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1153 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1155 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1156 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1158 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1159 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1161 return (ZIO_PIPELINE_CONTINUE);
1165 zio_write_bp_init(zio_t *zio)
1167 spa_t *spa = zio->io_spa;
1168 zio_prop_t *zp = &zio->io_prop;
1169 enum zio_compress compress = zp->zp_compress;
1170 blkptr_t *bp = zio->io_bp;
1171 uint64_t lsize = zio->io_size;
1172 uint64_t psize = lsize;
1176 * If our children haven't all reached the ready stage,
1177 * wait for them and then repeat this pipeline stage.
1179 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1180 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1181 return (ZIO_PIPELINE_STOP);
1183 if (!IO_IS_ALLOCATING(zio))
1184 return (ZIO_PIPELINE_CONTINUE);
1186 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1188 if (zio->io_bp_override) {
1189 ASSERT(bp->blk_birth != zio->io_txg);
1190 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1192 *bp = *zio->io_bp_override;
1193 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1195 if (BP_IS_EMBEDDED(bp))
1196 return (ZIO_PIPELINE_CONTINUE);
1199 * If we've been overridden and nopwrite is set then
1200 * set the flag accordingly to indicate that a nopwrite
1201 * has already occurred.
1203 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1204 ASSERT(!zp->zp_dedup);
1205 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1206 return (ZIO_PIPELINE_CONTINUE);
1209 ASSERT(!zp->zp_nopwrite);
1211 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1212 return (ZIO_PIPELINE_CONTINUE);
1214 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1215 zp->zp_dedup_verify);
1217 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1218 BP_SET_DEDUP(bp, 1);
1219 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1220 return (ZIO_PIPELINE_CONTINUE);
1224 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1226 * We're rewriting an existing block, which means we're
1227 * working on behalf of spa_sync(). For spa_sync() to
1228 * converge, it must eventually be the case that we don't
1229 * have to allocate new blocks. But compression changes
1230 * the blocksize, which forces a reallocate, and makes
1231 * convergence take longer. Therefore, after the first
1232 * few passes, stop compressing to ensure convergence.
1234 pass = spa_sync_pass(spa);
1236 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1237 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1238 ASSERT(!BP_GET_DEDUP(bp));
1240 if (pass >= zfs_sync_pass_dont_compress)
1241 compress = ZIO_COMPRESS_OFF;
1243 /* Make sure someone doesn't change their mind on overwrites */
1244 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1245 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1248 if (compress != ZIO_COMPRESS_OFF) {
1249 void *cbuf = zio_buf_alloc(lsize);
1250 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1251 if (psize == 0 || psize == lsize) {
1252 compress = ZIO_COMPRESS_OFF;
1253 zio_buf_free(cbuf, lsize);
1254 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1255 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1256 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1257 encode_embedded_bp_compressed(bp,
1258 cbuf, compress, lsize, psize);
1259 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1260 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1261 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1262 zio_buf_free(cbuf, lsize);
1263 bp->blk_birth = zio->io_txg;
1264 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1265 ASSERT(spa_feature_is_active(spa,
1266 SPA_FEATURE_EMBEDDED_DATA));
1267 return (ZIO_PIPELINE_CONTINUE);
1270 * Round up compressed size up to the ashift
1271 * of the smallest-ashift device, and zero the tail.
1272 * This ensures that the compressed size of the BP
1273 * (and thus compressratio property) are correct,
1274 * in that we charge for the padding used to fill out
1277 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
1278 size_t rounded = (size_t)P2ROUNDUP(psize,
1279 1ULL << spa->spa_min_ashift);
1280 if (rounded >= lsize) {
1281 compress = ZIO_COMPRESS_OFF;
1282 zio_buf_free(cbuf, lsize);
1285 bzero((char *)cbuf + psize, rounded - psize);
1287 zio_push_transform(zio, cbuf,
1288 psize, lsize, NULL);
1294 * The final pass of spa_sync() must be all rewrites, but the first
1295 * few passes offer a trade-off: allocating blocks defers convergence,
1296 * but newly allocated blocks are sequential, so they can be written
1297 * to disk faster. Therefore, we allow the first few passes of
1298 * spa_sync() to allocate new blocks, but force rewrites after that.
1299 * There should only be a handful of blocks after pass 1 in any case.
1301 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1302 BP_GET_PSIZE(bp) == psize &&
1303 pass >= zfs_sync_pass_rewrite) {
1305 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1306 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1307 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1310 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1314 if (zio->io_bp_orig.blk_birth != 0 &&
1315 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1316 BP_SET_LSIZE(bp, lsize);
1317 BP_SET_TYPE(bp, zp->zp_type);
1318 BP_SET_LEVEL(bp, zp->zp_level);
1319 BP_SET_BIRTH(bp, zio->io_txg, 0);
1321 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1323 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1324 BP_SET_LSIZE(bp, lsize);
1325 BP_SET_TYPE(bp, zp->zp_type);
1326 BP_SET_LEVEL(bp, zp->zp_level);
1327 BP_SET_PSIZE(bp, psize);
1328 BP_SET_COMPRESS(bp, compress);
1329 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1330 BP_SET_DEDUP(bp, zp->zp_dedup);
1331 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1333 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1334 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1335 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1337 if (zp->zp_nopwrite) {
1338 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1339 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1340 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1344 return (ZIO_PIPELINE_CONTINUE);
1348 zio_free_bp_init(zio_t *zio)
1350 blkptr_t *bp = zio->io_bp;
1352 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1353 if (BP_GET_DEDUP(bp))
1354 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1357 return (ZIO_PIPELINE_CONTINUE);
1361 * ==========================================================================
1362 * Execute the I/O pipeline
1363 * ==========================================================================
1367 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1369 spa_t *spa = zio->io_spa;
1370 zio_type_t t = zio->io_type;
1371 int flags = (cutinline ? TQ_FRONT : 0);
1373 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1376 * If we're a config writer or a probe, the normal issue and
1377 * interrupt threads may all be blocked waiting for the config lock.
1378 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1380 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1384 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1386 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1390 * If this is a high priority I/O, then use the high priority taskq if
1393 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1394 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1397 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1400 * NB: We are assuming that the zio can only be dispatched
1401 * to a single taskq at a time. It would be a grievous error
1402 * to dispatch the zio to another taskq at the same time.
1404 #if defined(illumos) || !defined(_KERNEL)
1405 ASSERT(zio->io_tqent.tqent_next == NULL);
1407 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
1409 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1410 flags, &zio->io_tqent);
1414 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1416 kthread_t *executor = zio->io_executor;
1417 spa_t *spa = zio->io_spa;
1419 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1420 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1422 for (i = 0; i < tqs->stqs_count; i++) {
1423 if (taskq_member(tqs->stqs_taskq[i], executor))
1432 zio_issue_async(zio_t *zio)
1434 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1436 return (ZIO_PIPELINE_STOP);
1440 zio_interrupt(zio_t *zio)
1442 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1446 * Execute the I/O pipeline until one of the following occurs:
1448 * (1) the I/O completes
1449 * (2) the pipeline stalls waiting for dependent child I/Os
1450 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1451 * (4) the I/O is delegated by vdev-level caching or aggregation
1452 * (5) the I/O is deferred due to vdev-level queueing
1453 * (6) the I/O is handed off to another thread.
1455 * In all cases, the pipeline stops whenever there's no CPU work; it never
1456 * burns a thread in cv_wait().
1458 * There's no locking on io_stage because there's no legitimate way
1459 * for multiple threads to be attempting to process the same I/O.
1461 static zio_pipe_stage_t *zio_pipeline[];
1464 zio_execute(zio_t *zio)
1466 zio->io_executor = curthread;
1468 while (zio->io_stage < ZIO_STAGE_DONE) {
1469 enum zio_stage pipeline = zio->io_pipeline;
1470 enum zio_stage stage = zio->io_stage;
1473 ASSERT(!MUTEX_HELD(&zio->io_lock));
1474 ASSERT(ISP2(stage));
1475 ASSERT(zio->io_stall == NULL);
1479 } while ((stage & pipeline) == 0);
1481 ASSERT(stage <= ZIO_STAGE_DONE);
1484 * If we are in interrupt context and this pipeline stage
1485 * will grab a config lock that is held across I/O,
1486 * or may wait for an I/O that needs an interrupt thread
1487 * to complete, issue async to avoid deadlock.
1489 * For VDEV_IO_START, we cut in line so that the io will
1490 * be sent to disk promptly.
1492 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1493 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1494 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1495 zio_requeue_io_start_cut_in_line : B_FALSE;
1496 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1500 zio->io_stage = stage;
1501 rv = zio_pipeline[highbit64(stage) - 1](zio);
1503 if (rv == ZIO_PIPELINE_STOP)
1506 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1511 * ==========================================================================
1512 * Initiate I/O, either sync or async
1513 * ==========================================================================
1516 zio_wait(zio_t *zio)
1520 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1521 ASSERT(zio->io_executor == NULL);
1523 zio->io_waiter = curthread;
1527 mutex_enter(&zio->io_lock);
1528 while (zio->io_executor != NULL)
1529 cv_wait(&zio->io_cv, &zio->io_lock);
1530 mutex_exit(&zio->io_lock);
1532 error = zio->io_error;
1539 zio_nowait(zio_t *zio)
1541 ASSERT(zio->io_executor == NULL);
1543 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1544 zio_unique_parent(zio) == NULL) {
1546 * This is a logical async I/O with no parent to wait for it.
1547 * We add it to the spa_async_root_zio "Godfather" I/O which
1548 * will ensure they complete prior to unloading the pool.
1550 spa_t *spa = zio->io_spa;
1552 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1559 * ==========================================================================
1560 * Reexecute or suspend/resume failed I/O
1561 * ==========================================================================
1565 zio_reexecute(zio_t *pio)
1567 zio_t *cio, *cio_next;
1569 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1570 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1571 ASSERT(pio->io_gang_leader == NULL);
1572 ASSERT(pio->io_gang_tree == NULL);
1574 pio->io_flags = pio->io_orig_flags;
1575 pio->io_stage = pio->io_orig_stage;
1576 pio->io_pipeline = pio->io_orig_pipeline;
1577 pio->io_reexecute = 0;
1578 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1580 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1581 pio->io_state[w] = 0;
1582 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1583 pio->io_child_error[c] = 0;
1585 if (IO_IS_ALLOCATING(pio))
1586 BP_ZERO(pio->io_bp);
1589 * As we reexecute pio's children, new children could be created.
1590 * New children go to the head of pio's io_child_list, however,
1591 * so we will (correctly) not reexecute them. The key is that
1592 * the remainder of pio's io_child_list, from 'cio_next' onward,
1593 * cannot be affected by any side effects of reexecuting 'cio'.
1595 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1596 cio_next = zio_walk_children(pio);
1597 mutex_enter(&pio->io_lock);
1598 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1599 pio->io_children[cio->io_child_type][w]++;
1600 mutex_exit(&pio->io_lock);
1605 * Now that all children have been reexecuted, execute the parent.
1606 * We don't reexecute "The Godfather" I/O here as it's the
1607 * responsibility of the caller to wait on him.
1609 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1614 zio_suspend(spa_t *spa, zio_t *zio)
1616 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1617 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1618 "failure and the failure mode property for this pool "
1619 "is set to panic.", spa_name(spa));
1621 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1623 mutex_enter(&spa->spa_suspend_lock);
1625 if (spa->spa_suspend_zio_root == NULL)
1626 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1627 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1628 ZIO_FLAG_GODFATHER);
1630 spa->spa_suspended = B_TRUE;
1633 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1634 ASSERT(zio != spa->spa_suspend_zio_root);
1635 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1636 ASSERT(zio_unique_parent(zio) == NULL);
1637 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1638 zio_add_child(spa->spa_suspend_zio_root, zio);
1641 mutex_exit(&spa->spa_suspend_lock);
1645 zio_resume(spa_t *spa)
1650 * Reexecute all previously suspended i/o.
1652 mutex_enter(&spa->spa_suspend_lock);
1653 spa->spa_suspended = B_FALSE;
1654 cv_broadcast(&spa->spa_suspend_cv);
1655 pio = spa->spa_suspend_zio_root;
1656 spa->spa_suspend_zio_root = NULL;
1657 mutex_exit(&spa->spa_suspend_lock);
1663 return (zio_wait(pio));
1667 zio_resume_wait(spa_t *spa)
1669 mutex_enter(&spa->spa_suspend_lock);
1670 while (spa_suspended(spa))
1671 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1672 mutex_exit(&spa->spa_suspend_lock);
1676 * ==========================================================================
1679 * A gang block is a collection of small blocks that looks to the DMU
1680 * like one large block. When zio_dva_allocate() cannot find a block
1681 * of the requested size, due to either severe fragmentation or the pool
1682 * being nearly full, it calls zio_write_gang_block() to construct the
1683 * block from smaller fragments.
1685 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1686 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1687 * an indirect block: it's an array of block pointers. It consumes
1688 * only one sector and hence is allocatable regardless of fragmentation.
1689 * The gang header's bps point to its gang members, which hold the data.
1691 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1692 * as the verifier to ensure uniqueness of the SHA256 checksum.
1693 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1694 * not the gang header. This ensures that data block signatures (needed for
1695 * deduplication) are independent of how the block is physically stored.
1697 * Gang blocks can be nested: a gang member may itself be a gang block.
1698 * Thus every gang block is a tree in which root and all interior nodes are
1699 * gang headers, and the leaves are normal blocks that contain user data.
1700 * The root of the gang tree is called the gang leader.
1702 * To perform any operation (read, rewrite, free, claim) on a gang block,
1703 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1704 * in the io_gang_tree field of the original logical i/o by recursively
1705 * reading the gang leader and all gang headers below it. This yields
1706 * an in-core tree containing the contents of every gang header and the
1707 * bps for every constituent of the gang block.
1709 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1710 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1711 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1712 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1713 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1714 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1715 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1716 * of the gang header plus zio_checksum_compute() of the data to update the
1717 * gang header's blk_cksum as described above.
1719 * The two-phase assemble/issue model solves the problem of partial failure --
1720 * what if you'd freed part of a gang block but then couldn't read the
1721 * gang header for another part? Assembling the entire gang tree first
1722 * ensures that all the necessary gang header I/O has succeeded before
1723 * starting the actual work of free, claim, or write. Once the gang tree
1724 * is assembled, free and claim are in-memory operations that cannot fail.
1726 * In the event that a gang write fails, zio_dva_unallocate() walks the
1727 * gang tree to immediately free (i.e. insert back into the space map)
1728 * everything we've allocated. This ensures that we don't get ENOSPC
1729 * errors during repeated suspend/resume cycles due to a flaky device.
1731 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1732 * the gang tree, we won't modify the block, so we can safely defer the free
1733 * (knowing that the block is still intact). If we *can* assemble the gang
1734 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1735 * each constituent bp and we can allocate a new block on the next sync pass.
1737 * In all cases, the gang tree allows complete recovery from partial failure.
1738 * ==========================================================================
1742 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1747 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1748 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1749 &pio->io_bookmark));
1753 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1758 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1759 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1760 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1762 * As we rewrite each gang header, the pipeline will compute
1763 * a new gang block header checksum for it; but no one will
1764 * compute a new data checksum, so we do that here. The one
1765 * exception is the gang leader: the pipeline already computed
1766 * its data checksum because that stage precedes gang assembly.
1767 * (Presently, nothing actually uses interior data checksums;
1768 * this is just good hygiene.)
1770 if (gn != pio->io_gang_leader->io_gang_tree) {
1771 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1772 data, BP_GET_PSIZE(bp));
1775 * If we are here to damage data for testing purposes,
1776 * leave the GBH alone so that we can detect the damage.
1778 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1779 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1781 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1782 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1783 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1791 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1793 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1794 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp),
1795 ZIO_GANG_CHILD_FLAGS(pio)));
1800 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1802 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1803 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1806 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1815 static void zio_gang_tree_assemble_done(zio_t *zio);
1817 static zio_gang_node_t *
1818 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1820 zio_gang_node_t *gn;
1822 ASSERT(*gnpp == NULL);
1824 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1825 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1832 zio_gang_node_free(zio_gang_node_t **gnpp)
1834 zio_gang_node_t *gn = *gnpp;
1836 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1837 ASSERT(gn->gn_child[g] == NULL);
1839 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1840 kmem_free(gn, sizeof (*gn));
1845 zio_gang_tree_free(zio_gang_node_t **gnpp)
1847 zio_gang_node_t *gn = *gnpp;
1852 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1853 zio_gang_tree_free(&gn->gn_child[g]);
1855 zio_gang_node_free(gnpp);
1859 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1861 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1863 ASSERT(gio->io_gang_leader == gio);
1864 ASSERT(BP_IS_GANG(bp));
1866 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1867 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1868 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1872 zio_gang_tree_assemble_done(zio_t *zio)
1874 zio_t *gio = zio->io_gang_leader;
1875 zio_gang_node_t *gn = zio->io_private;
1876 blkptr_t *bp = zio->io_bp;
1878 ASSERT(gio == zio_unique_parent(zio));
1879 ASSERT(zio->io_child_count == 0);
1884 if (BP_SHOULD_BYTESWAP(bp))
1885 byteswap_uint64_array(zio->io_data, zio->io_size);
1887 ASSERT(zio->io_data == gn->gn_gbh);
1888 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1889 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1891 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1892 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1893 if (!BP_IS_GANG(gbp))
1895 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1900 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1902 zio_t *gio = pio->io_gang_leader;
1905 ASSERT(BP_IS_GANG(bp) == !!gn);
1906 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1907 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1910 * If you're a gang header, your data is in gn->gn_gbh.
1911 * If you're a gang member, your data is in 'data' and gn == NULL.
1913 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1916 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1918 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1919 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1920 if (BP_IS_HOLE(gbp))
1922 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1923 data = (char *)data + BP_GET_PSIZE(gbp);
1927 if (gn == gio->io_gang_tree && gio->io_data != NULL)
1928 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1935 zio_gang_assemble(zio_t *zio)
1937 blkptr_t *bp = zio->io_bp;
1939 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1940 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1942 zio->io_gang_leader = zio;
1944 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1946 return (ZIO_PIPELINE_CONTINUE);
1950 zio_gang_issue(zio_t *zio)
1952 blkptr_t *bp = zio->io_bp;
1954 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1955 return (ZIO_PIPELINE_STOP);
1957 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1958 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1960 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1961 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1963 zio_gang_tree_free(&zio->io_gang_tree);
1965 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1967 return (ZIO_PIPELINE_CONTINUE);
1971 zio_write_gang_member_ready(zio_t *zio)
1973 zio_t *pio = zio_unique_parent(zio);
1974 zio_t *gio = zio->io_gang_leader;
1975 dva_t *cdva = zio->io_bp->blk_dva;
1976 dva_t *pdva = pio->io_bp->blk_dva;
1979 if (BP_IS_HOLE(zio->io_bp))
1982 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1984 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1985 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1986 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1987 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1988 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1990 mutex_enter(&pio->io_lock);
1991 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1992 ASSERT(DVA_GET_GANG(&pdva[d]));
1993 asize = DVA_GET_ASIZE(&pdva[d]);
1994 asize += DVA_GET_ASIZE(&cdva[d]);
1995 DVA_SET_ASIZE(&pdva[d], asize);
1997 mutex_exit(&pio->io_lock);
2001 zio_write_gang_block(zio_t *pio)
2003 spa_t *spa = pio->io_spa;
2004 blkptr_t *bp = pio->io_bp;
2005 zio_t *gio = pio->io_gang_leader;
2007 zio_gang_node_t *gn, **gnpp;
2008 zio_gbh_phys_t *gbh;
2009 uint64_t txg = pio->io_txg;
2010 uint64_t resid = pio->io_size;
2012 int copies = gio->io_prop.zp_copies;
2013 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
2017 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
2018 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
2019 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
2021 pio->io_error = error;
2022 return (ZIO_PIPELINE_CONTINUE);
2026 gnpp = &gio->io_gang_tree;
2028 gnpp = pio->io_private;
2029 ASSERT(pio->io_ready == zio_write_gang_member_ready);
2032 gn = zio_gang_node_alloc(gnpp);
2034 bzero(gbh, SPA_GANGBLOCKSIZE);
2037 * Create the gang header.
2039 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
2040 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2043 * Create and nowait the gang children.
2045 for (int g = 0; resid != 0; resid -= lsize, g++) {
2046 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2048 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2050 zp.zp_checksum = gio->io_prop.zp_checksum;
2051 zp.zp_compress = ZIO_COMPRESS_OFF;
2052 zp.zp_type = DMU_OT_NONE;
2054 zp.zp_copies = gio->io_prop.zp_copies;
2055 zp.zp_dedup = B_FALSE;
2056 zp.zp_dedup_verify = B_FALSE;
2057 zp.zp_nopwrite = B_FALSE;
2059 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2060 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
2061 zio_write_gang_member_ready, NULL, NULL, &gn->gn_child[g],
2062 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2063 &pio->io_bookmark));
2067 * Set pio's pipeline to just wait for zio to finish.
2069 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2073 return (ZIO_PIPELINE_CONTINUE);
2077 * The zio_nop_write stage in the pipeline determines if allocating
2078 * a new bp is necessary. By leveraging a cryptographically secure checksum,
2079 * such as SHA256, we can compare the checksums of the new data and the old
2080 * to determine if allocating a new block is required. The nopwrite
2081 * feature can handle writes in either syncing or open context (i.e. zil
2082 * writes) and as a result is mutually exclusive with dedup.
2085 zio_nop_write(zio_t *zio)
2087 blkptr_t *bp = zio->io_bp;
2088 blkptr_t *bp_orig = &zio->io_bp_orig;
2089 zio_prop_t *zp = &zio->io_prop;
2091 ASSERT(BP_GET_LEVEL(bp) == 0);
2092 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2093 ASSERT(zp->zp_nopwrite);
2094 ASSERT(!zp->zp_dedup);
2095 ASSERT(zio->io_bp_override == NULL);
2096 ASSERT(IO_IS_ALLOCATING(zio));
2099 * Check to see if the original bp and the new bp have matching
2100 * characteristics (i.e. same checksum, compression algorithms, etc).
2101 * If they don't then just continue with the pipeline which will
2102 * allocate a new bp.
2104 if (BP_IS_HOLE(bp_orig) ||
2105 !zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup ||
2106 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2107 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2108 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2109 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2110 return (ZIO_PIPELINE_CONTINUE);
2113 * If the checksums match then reset the pipeline so that we
2114 * avoid allocating a new bp and issuing any I/O.
2116 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2117 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup);
2118 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2119 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2120 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2121 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2122 sizeof (uint64_t)) == 0);
2125 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2126 zio->io_flags |= ZIO_FLAG_NOPWRITE;
2129 return (ZIO_PIPELINE_CONTINUE);
2133 * ==========================================================================
2135 * ==========================================================================
2138 zio_ddt_child_read_done(zio_t *zio)
2140 blkptr_t *bp = zio->io_bp;
2141 ddt_entry_t *dde = zio->io_private;
2143 zio_t *pio = zio_unique_parent(zio);
2145 mutex_enter(&pio->io_lock);
2146 ddp = ddt_phys_select(dde, bp);
2147 if (zio->io_error == 0)
2148 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2149 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
2150 dde->dde_repair_data = zio->io_data;
2152 zio_buf_free(zio->io_data, zio->io_size);
2153 mutex_exit(&pio->io_lock);
2157 zio_ddt_read_start(zio_t *zio)
2159 blkptr_t *bp = zio->io_bp;
2161 ASSERT(BP_GET_DEDUP(bp));
2162 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2163 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2165 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2166 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2167 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2168 ddt_phys_t *ddp = dde->dde_phys;
2169 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2172 ASSERT(zio->io_vsd == NULL);
2175 if (ddp_self == NULL)
2176 return (ZIO_PIPELINE_CONTINUE);
2178 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2179 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2181 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2183 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2184 zio_buf_alloc(zio->io_size), zio->io_size,
2185 zio_ddt_child_read_done, dde, zio->io_priority,
2186 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
2187 &zio->io_bookmark));
2189 return (ZIO_PIPELINE_CONTINUE);
2192 zio_nowait(zio_read(zio, zio->io_spa, bp,
2193 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
2194 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2196 return (ZIO_PIPELINE_CONTINUE);
2200 zio_ddt_read_done(zio_t *zio)
2202 blkptr_t *bp = zio->io_bp;
2204 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2205 return (ZIO_PIPELINE_STOP);
2207 ASSERT(BP_GET_DEDUP(bp));
2208 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2209 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2211 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2212 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2213 ddt_entry_t *dde = zio->io_vsd;
2215 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2216 return (ZIO_PIPELINE_CONTINUE);
2219 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2220 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2221 return (ZIO_PIPELINE_STOP);
2223 if (dde->dde_repair_data != NULL) {
2224 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2225 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2227 ddt_repair_done(ddt, dde);
2231 ASSERT(zio->io_vsd == NULL);
2233 return (ZIO_PIPELINE_CONTINUE);
2237 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2239 spa_t *spa = zio->io_spa;
2242 * Note: we compare the original data, not the transformed data,
2243 * because when zio->io_bp is an override bp, we will not have
2244 * pushed the I/O transforms. That's an important optimization
2245 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2247 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2248 zio_t *lio = dde->dde_lead_zio[p];
2251 return (lio->io_orig_size != zio->io_orig_size ||
2252 bcmp(zio->io_orig_data, lio->io_orig_data,
2253 zio->io_orig_size) != 0);
2257 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2258 ddt_phys_t *ddp = &dde->dde_phys[p];
2260 if (ddp->ddp_phys_birth != 0) {
2261 arc_buf_t *abuf = NULL;
2262 arc_flags_t aflags = ARC_FLAG_WAIT;
2263 blkptr_t blk = *zio->io_bp;
2266 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2270 error = arc_read(NULL, spa, &blk,
2271 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2272 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2273 &aflags, &zio->io_bookmark);
2276 if (arc_buf_size(abuf) != zio->io_orig_size ||
2277 bcmp(abuf->b_data, zio->io_orig_data,
2278 zio->io_orig_size) != 0)
2279 error = SET_ERROR(EEXIST);
2280 VERIFY(arc_buf_remove_ref(abuf, &abuf));
2284 return (error != 0);
2292 zio_ddt_child_write_ready(zio_t *zio)
2294 int p = zio->io_prop.zp_copies;
2295 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2296 ddt_entry_t *dde = zio->io_private;
2297 ddt_phys_t *ddp = &dde->dde_phys[p];
2305 ASSERT(dde->dde_lead_zio[p] == zio);
2307 ddt_phys_fill(ddp, zio->io_bp);
2309 while ((pio = zio_walk_parents(zio)) != NULL)
2310 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2316 zio_ddt_child_write_done(zio_t *zio)
2318 int p = zio->io_prop.zp_copies;
2319 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2320 ddt_entry_t *dde = zio->io_private;
2321 ddt_phys_t *ddp = &dde->dde_phys[p];
2325 ASSERT(ddp->ddp_refcnt == 0);
2326 ASSERT(dde->dde_lead_zio[p] == zio);
2327 dde->dde_lead_zio[p] = NULL;
2329 if (zio->io_error == 0) {
2330 while (zio_walk_parents(zio) != NULL)
2331 ddt_phys_addref(ddp);
2333 ddt_phys_clear(ddp);
2340 zio_ddt_ditto_write_done(zio_t *zio)
2342 int p = DDT_PHYS_DITTO;
2343 zio_prop_t *zp = &zio->io_prop;
2344 blkptr_t *bp = zio->io_bp;
2345 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2346 ddt_entry_t *dde = zio->io_private;
2347 ddt_phys_t *ddp = &dde->dde_phys[p];
2348 ddt_key_t *ddk = &dde->dde_key;
2352 ASSERT(ddp->ddp_refcnt == 0);
2353 ASSERT(dde->dde_lead_zio[p] == zio);
2354 dde->dde_lead_zio[p] = NULL;
2356 if (zio->io_error == 0) {
2357 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2358 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2359 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2360 if (ddp->ddp_phys_birth != 0)
2361 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2362 ddt_phys_fill(ddp, bp);
2369 zio_ddt_write(zio_t *zio)
2371 spa_t *spa = zio->io_spa;
2372 blkptr_t *bp = zio->io_bp;
2373 uint64_t txg = zio->io_txg;
2374 zio_prop_t *zp = &zio->io_prop;
2375 int p = zp->zp_copies;
2379 ddt_t *ddt = ddt_select(spa, bp);
2383 ASSERT(BP_GET_DEDUP(bp));
2384 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2385 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2388 dde = ddt_lookup(ddt, bp, B_TRUE);
2389 ddp = &dde->dde_phys[p];
2391 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2393 * If we're using a weak checksum, upgrade to a strong checksum
2394 * and try again. If we're already using a strong checksum,
2395 * we can't resolve it, so just convert to an ordinary write.
2396 * (And automatically e-mail a paper to Nature?)
2398 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2399 zp->zp_checksum = spa_dedup_checksum(spa);
2400 zio_pop_transforms(zio);
2401 zio->io_stage = ZIO_STAGE_OPEN;
2404 zp->zp_dedup = B_FALSE;
2406 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2408 return (ZIO_PIPELINE_CONTINUE);
2411 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2412 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2414 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2415 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2416 zio_prop_t czp = *zp;
2418 czp.zp_copies = ditto_copies;
2421 * If we arrived here with an override bp, we won't have run
2422 * the transform stack, so we won't have the data we need to
2423 * generate a child i/o. So, toss the override bp and restart.
2424 * This is safe, because using the override bp is just an
2425 * optimization; and it's rare, so the cost doesn't matter.
2427 if (zio->io_bp_override) {
2428 zio_pop_transforms(zio);
2429 zio->io_stage = ZIO_STAGE_OPEN;
2430 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2431 zio->io_bp_override = NULL;
2434 return (ZIO_PIPELINE_CONTINUE);
2437 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2438 zio->io_orig_size, &czp, NULL, NULL,
2439 zio_ddt_ditto_write_done, dde, zio->io_priority,
2440 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2442 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2443 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2446 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2447 if (ddp->ddp_phys_birth != 0)
2448 ddt_bp_fill(ddp, bp, txg);
2449 if (dde->dde_lead_zio[p] != NULL)
2450 zio_add_child(zio, dde->dde_lead_zio[p]);
2452 ddt_phys_addref(ddp);
2453 } else if (zio->io_bp_override) {
2454 ASSERT(bp->blk_birth == txg);
2455 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2456 ddt_phys_fill(ddp, bp);
2457 ddt_phys_addref(ddp);
2459 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2460 zio->io_orig_size, zp, zio_ddt_child_write_ready, NULL,
2461 zio_ddt_child_write_done, dde, zio->io_priority,
2462 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2464 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2465 dde->dde_lead_zio[p] = cio;
2475 return (ZIO_PIPELINE_CONTINUE);
2478 ddt_entry_t *freedde; /* for debugging */
2481 zio_ddt_free(zio_t *zio)
2483 spa_t *spa = zio->io_spa;
2484 blkptr_t *bp = zio->io_bp;
2485 ddt_t *ddt = ddt_select(spa, bp);
2489 ASSERT(BP_GET_DEDUP(bp));
2490 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2493 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2494 ddp = ddt_phys_select(dde, bp);
2495 ddt_phys_decref(ddp);
2498 return (ZIO_PIPELINE_CONTINUE);
2502 * ==========================================================================
2503 * Allocate and free blocks
2504 * ==========================================================================
2507 zio_dva_allocate(zio_t *zio)
2509 spa_t *spa = zio->io_spa;
2510 metaslab_class_t *mc = spa_normal_class(spa);
2511 blkptr_t *bp = zio->io_bp;
2515 if (zio->io_gang_leader == NULL) {
2516 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2517 zio->io_gang_leader = zio;
2520 ASSERT(BP_IS_HOLE(bp));
2521 ASSERT0(BP_GET_NDVAS(bp));
2522 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2523 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2524 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2527 * The dump device does not support gang blocks so allocation on
2528 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2529 * the "fast" gang feature.
2531 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2532 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2533 METASLAB_GANG_CHILD : 0;
2534 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2535 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2538 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2539 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2541 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2542 return (zio_write_gang_block(zio));
2543 zio->io_error = error;
2546 return (ZIO_PIPELINE_CONTINUE);
2550 zio_dva_free(zio_t *zio)
2552 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2554 return (ZIO_PIPELINE_CONTINUE);
2558 zio_dva_claim(zio_t *zio)
2562 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2564 zio->io_error = error;
2566 return (ZIO_PIPELINE_CONTINUE);
2570 * Undo an allocation. This is used by zio_done() when an I/O fails
2571 * and we want to give back the block we just allocated.
2572 * This handles both normal blocks and gang blocks.
2575 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2577 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2578 ASSERT(zio->io_bp_override == NULL);
2580 if (!BP_IS_HOLE(bp))
2581 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2584 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2585 zio_dva_unallocate(zio, gn->gn_child[g],
2586 &gn->gn_gbh->zg_blkptr[g]);
2592 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2595 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2596 uint64_t size, boolean_t use_slog)
2600 ASSERT(txg > spa_syncing_txg(spa));
2603 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2604 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2605 * when allocating them.
2608 error = metaslab_alloc(spa, spa_log_class(spa), size,
2609 new_bp, 1, txg, old_bp,
2610 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2614 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2615 new_bp, 1, txg, old_bp,
2616 METASLAB_HINTBP_AVOID);
2620 BP_SET_LSIZE(new_bp, size);
2621 BP_SET_PSIZE(new_bp, size);
2622 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2623 BP_SET_CHECKSUM(new_bp,
2624 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2625 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2626 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2627 BP_SET_LEVEL(new_bp, 0);
2628 BP_SET_DEDUP(new_bp, 0);
2629 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2636 * Free an intent log block.
2639 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2641 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2642 ASSERT(!BP_IS_GANG(bp));
2644 zio_free(spa, txg, bp);
2648 * ==========================================================================
2649 * Read, write and delete to physical devices
2650 * ==========================================================================
2653 zio_vdev_io_start(zio_t *zio)
2655 vdev_t *vd = zio->io_vd;
2657 spa_t *spa = zio->io_spa;
2660 ASSERT(zio->io_error == 0);
2661 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2664 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2665 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2668 * The mirror_ops handle multiple DVAs in a single BP.
2670 return (vdev_mirror_ops.vdev_op_io_start(zio));
2673 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE &&
2674 zio->io_priority == ZIO_PRIORITY_NOW) {
2675 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg);
2676 return (ZIO_PIPELINE_CONTINUE);
2680 * We keep track of time-sensitive I/Os so that the scan thread
2681 * can quickly react to certain workloads. In particular, we care
2682 * about non-scrubbing, top-level reads and writes with the following
2684 * - synchronous writes of user data to non-slog devices
2685 * - any reads of user data
2686 * When these conditions are met, adjust the timestamp of spa_last_io
2687 * which allows the scan thread to adjust its workload accordingly.
2689 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2690 vd == vd->vdev_top && !vd->vdev_islog &&
2691 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2692 zio->io_txg != spa_syncing_txg(spa)) {
2693 uint64_t old = spa->spa_last_io;
2694 uint64_t new = ddi_get_lbolt64();
2696 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2699 align = 1ULL << vd->vdev_top->vdev_ashift;
2701 if ((!(zio->io_flags & ZIO_FLAG_PHYSICAL) ||
2702 (vd->vdev_top->vdev_physical_ashift > SPA_MINBLOCKSHIFT)) &&
2703 P2PHASE(zio->io_size, align) != 0) {
2704 /* Transform logical writes to be a full physical block size. */
2705 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2707 if (zio->io_type == ZIO_TYPE_READ ||
2708 zio->io_type == ZIO_TYPE_WRITE)
2709 abuf = zio_buf_alloc(asize);
2710 ASSERT(vd == vd->vdev_top);
2711 if (zio->io_type == ZIO_TYPE_WRITE) {
2712 bcopy(zio->io_data, abuf, zio->io_size);
2713 bzero(abuf + zio->io_size, asize - zio->io_size);
2715 zio_push_transform(zio, abuf, asize, abuf ? asize : 0,
2720 * If this is not a physical io, make sure that it is properly aligned
2721 * before proceeding.
2723 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
2724 ASSERT0(P2PHASE(zio->io_offset, align));
2725 ASSERT0(P2PHASE(zio->io_size, align));
2728 * For physical writes, we allow 512b aligned writes and assume
2729 * the device will perform a read-modify-write as necessary.
2731 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
2732 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
2735 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa));
2738 * If this is a repair I/O, and there's no self-healing involved --
2739 * that is, we're just resilvering what we expect to resilver --
2740 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2741 * This prevents spurious resilvering with nested replication.
2742 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2743 * A is out of date, we'll read from C+D, then use the data to
2744 * resilver A+B -- but we don't actually want to resilver B, just A.
2745 * The top-level mirror has no way to know this, so instead we just
2746 * discard unnecessary repairs as we work our way down the vdev tree.
2747 * The same logic applies to any form of nested replication:
2748 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2750 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2751 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2752 zio->io_txg != 0 && /* not a delegated i/o */
2753 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2754 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2755 zio_vdev_io_bypass(zio);
2756 return (ZIO_PIPELINE_CONTINUE);
2759 if (vd->vdev_ops->vdev_op_leaf) {
2760 switch (zio->io_type) {
2762 if (vdev_cache_read(zio))
2763 return (ZIO_PIPELINE_CONTINUE);
2765 case ZIO_TYPE_WRITE:
2767 if ((zio = vdev_queue_io(zio)) == NULL)
2768 return (ZIO_PIPELINE_STOP);
2770 if (!vdev_accessible(vd, zio)) {
2771 zio->io_error = SET_ERROR(ENXIO);
2773 return (ZIO_PIPELINE_STOP);
2778 * Note that we ignore repair writes for TRIM because they can
2779 * conflict with normal writes. This isn't an issue because, by
2780 * definition, we only repair blocks that aren't freed.
2782 if (zio->io_type == ZIO_TYPE_WRITE &&
2783 !(zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2784 !trim_map_write_start(zio))
2785 return (ZIO_PIPELINE_STOP);
2788 ret = vd->vdev_ops->vdev_op_io_start(zio);
2789 ASSERT(ret == ZIO_PIPELINE_STOP);
2795 zio_vdev_io_done(zio_t *zio)
2797 vdev_t *vd = zio->io_vd;
2798 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2799 boolean_t unexpected_error = B_FALSE;
2801 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2802 return (ZIO_PIPELINE_STOP);
2804 ASSERT(zio->io_type == ZIO_TYPE_READ ||
2805 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE);
2807 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2808 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE ||
2809 zio->io_type == ZIO_TYPE_FREE)) {
2811 if (zio->io_type == ZIO_TYPE_WRITE &&
2812 !(zio->io_flags & ZIO_FLAG_IO_REPAIR))
2813 trim_map_write_done(zio);
2815 vdev_queue_io_done(zio);
2817 if (zio->io_type == ZIO_TYPE_WRITE)
2818 vdev_cache_write(zio);
2820 if (zio_injection_enabled && zio->io_error == 0)
2821 zio->io_error = zio_handle_device_injection(vd,
2824 if (zio_injection_enabled && zio->io_error == 0)
2825 zio->io_error = zio_handle_label_injection(zio, EIO);
2827 if (zio->io_error) {
2828 if (zio->io_error == ENOTSUP &&
2829 zio->io_type == ZIO_TYPE_FREE) {
2830 /* Not all devices support TRIM. */
2831 } else if (!vdev_accessible(vd, zio)) {
2832 zio->io_error = SET_ERROR(ENXIO);
2834 unexpected_error = B_TRUE;
2839 ops->vdev_op_io_done(zio);
2841 if (unexpected_error)
2842 VERIFY(vdev_probe(vd, zio) == NULL);
2844 return (ZIO_PIPELINE_CONTINUE);
2848 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2849 * disk, and use that to finish the checksum ereport later.
2852 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2853 const void *good_buf)
2855 /* no processing needed */
2856 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2861 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2863 void *buf = zio_buf_alloc(zio->io_size);
2865 bcopy(zio->io_data, buf, zio->io_size);
2867 zcr->zcr_cbinfo = zio->io_size;
2868 zcr->zcr_cbdata = buf;
2869 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2870 zcr->zcr_free = zio_buf_free;
2874 zio_vdev_io_assess(zio_t *zio)
2876 vdev_t *vd = zio->io_vd;
2878 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2879 return (ZIO_PIPELINE_STOP);
2881 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2882 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2884 if (zio->io_vsd != NULL) {
2885 zio->io_vsd_ops->vsd_free(zio);
2889 if (zio_injection_enabled && zio->io_error == 0)
2890 zio->io_error = zio_handle_fault_injection(zio, EIO);
2892 if (zio->io_type == ZIO_TYPE_FREE &&
2893 zio->io_priority != ZIO_PRIORITY_NOW) {
2894 switch (zio->io_error) {
2896 ZIO_TRIM_STAT_INCR(bytes, zio->io_size);
2897 ZIO_TRIM_STAT_BUMP(success);
2900 ZIO_TRIM_STAT_BUMP(unsupported);
2903 ZIO_TRIM_STAT_BUMP(failed);
2909 * If the I/O failed, determine whether we should attempt to retry it.
2911 * On retry, we cut in line in the issue queue, since we don't want
2912 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2914 if (zio->io_error && vd == NULL &&
2915 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2916 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2917 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2919 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2920 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2921 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2922 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2923 zio_requeue_io_start_cut_in_line);
2924 return (ZIO_PIPELINE_STOP);
2928 * If we got an error on a leaf device, convert it to ENXIO
2929 * if the device is not accessible at all.
2931 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2932 !vdev_accessible(vd, zio))
2933 zio->io_error = SET_ERROR(ENXIO);
2936 * If we can't write to an interior vdev (mirror or RAID-Z),
2937 * set vdev_cant_write so that we stop trying to allocate from it.
2939 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2940 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
2941 vd->vdev_cant_write = B_TRUE;
2945 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2947 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2948 zio->io_physdone != NULL) {
2949 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
2950 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
2951 zio->io_physdone(zio->io_logical);
2954 return (ZIO_PIPELINE_CONTINUE);
2958 zio_vdev_io_reissue(zio_t *zio)
2960 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2961 ASSERT(zio->io_error == 0);
2963 zio->io_stage >>= 1;
2967 zio_vdev_io_redone(zio_t *zio)
2969 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2971 zio->io_stage >>= 1;
2975 zio_vdev_io_bypass(zio_t *zio)
2977 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2978 ASSERT(zio->io_error == 0);
2980 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2981 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2985 * ==========================================================================
2986 * Generate and verify checksums
2987 * ==========================================================================
2990 zio_checksum_generate(zio_t *zio)
2992 blkptr_t *bp = zio->io_bp;
2993 enum zio_checksum checksum;
2997 * This is zio_write_phys().
2998 * We're either generating a label checksum, or none at all.
3000 checksum = zio->io_prop.zp_checksum;
3002 if (checksum == ZIO_CHECKSUM_OFF)
3003 return (ZIO_PIPELINE_CONTINUE);
3005 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
3007 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
3008 ASSERT(!IO_IS_ALLOCATING(zio));
3009 checksum = ZIO_CHECKSUM_GANG_HEADER;
3011 checksum = BP_GET_CHECKSUM(bp);
3015 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
3017 return (ZIO_PIPELINE_CONTINUE);
3021 zio_checksum_verify(zio_t *zio)
3023 zio_bad_cksum_t info;
3024 blkptr_t *bp = zio->io_bp;
3027 ASSERT(zio->io_vd != NULL);
3031 * This is zio_read_phys().
3032 * We're either verifying a label checksum, or nothing at all.
3034 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
3035 return (ZIO_PIPELINE_CONTINUE);
3037 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
3040 if ((error = zio_checksum_error(zio, &info)) != 0) {
3041 zio->io_error = error;
3042 if (error == ECKSUM &&
3043 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
3044 zfs_ereport_start_checksum(zio->io_spa,
3045 zio->io_vd, zio, zio->io_offset,
3046 zio->io_size, NULL, &info);
3050 return (ZIO_PIPELINE_CONTINUE);
3054 * Called by RAID-Z to ensure we don't compute the checksum twice.
3057 zio_checksum_verified(zio_t *zio)
3059 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
3063 * ==========================================================================
3064 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
3065 * An error of 0 indicates success. ENXIO indicates whole-device failure,
3066 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
3067 * indicate errors that are specific to one I/O, and most likely permanent.
3068 * Any other error is presumed to be worse because we weren't expecting it.
3069 * ==========================================================================
3072 zio_worst_error(int e1, int e2)
3074 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
3077 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
3078 if (e1 == zio_error_rank[r1])
3081 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
3082 if (e2 == zio_error_rank[r2])
3085 return (r1 > r2 ? e1 : e2);
3089 * ==========================================================================
3091 * ==========================================================================
3094 zio_ready(zio_t *zio)
3096 blkptr_t *bp = zio->io_bp;
3097 zio_t *pio, *pio_next;
3099 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
3100 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
3101 return (ZIO_PIPELINE_STOP);
3103 if (zio->io_ready) {
3104 ASSERT(IO_IS_ALLOCATING(zio));
3105 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
3106 (zio->io_flags & ZIO_FLAG_NOPWRITE));
3107 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
3112 if (bp != NULL && bp != &zio->io_bp_copy)
3113 zio->io_bp_copy = *bp;
3116 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3118 mutex_enter(&zio->io_lock);
3119 zio->io_state[ZIO_WAIT_READY] = 1;
3120 pio = zio_walk_parents(zio);
3121 mutex_exit(&zio->io_lock);
3124 * As we notify zio's parents, new parents could be added.
3125 * New parents go to the head of zio's io_parent_list, however,
3126 * so we will (correctly) not notify them. The remainder of zio's
3127 * io_parent_list, from 'pio_next' onward, cannot change because
3128 * all parents must wait for us to be done before they can be done.
3130 for (; pio != NULL; pio = pio_next) {
3131 pio_next = zio_walk_parents(zio);
3132 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
3135 if (zio->io_flags & ZIO_FLAG_NODATA) {
3136 if (BP_IS_GANG(bp)) {
3137 zio->io_flags &= ~ZIO_FLAG_NODATA;
3139 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
3140 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
3144 if (zio_injection_enabled &&
3145 zio->io_spa->spa_syncing_txg == zio->io_txg)
3146 zio_handle_ignored_writes(zio);
3148 return (ZIO_PIPELINE_CONTINUE);
3152 zio_done(zio_t *zio)
3154 spa_t *spa = zio->io_spa;
3155 zio_t *lio = zio->io_logical;
3156 blkptr_t *bp = zio->io_bp;
3157 vdev_t *vd = zio->io_vd;
3158 uint64_t psize = zio->io_size;
3159 zio_t *pio, *pio_next;
3162 * If our children haven't all completed,
3163 * wait for them and then repeat this pipeline stage.
3165 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
3166 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
3167 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
3168 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
3169 return (ZIO_PIPELINE_STOP);
3171 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3172 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3173 ASSERT(zio->io_children[c][w] == 0);
3175 if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
3176 ASSERT(bp->blk_pad[0] == 0);
3177 ASSERT(bp->blk_pad[1] == 0);
3178 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3179 (bp == zio_unique_parent(zio)->io_bp));
3180 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3181 zio->io_bp_override == NULL &&
3182 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3183 ASSERT(!BP_SHOULD_BYTESWAP(bp));
3184 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3185 ASSERT(BP_COUNT_GANG(bp) == 0 ||
3186 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3188 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3189 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3193 * If there were child vdev/gang/ddt errors, they apply to us now.
3195 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3196 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3197 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3200 * If the I/O on the transformed data was successful, generate any
3201 * checksum reports now while we still have the transformed data.
3203 if (zio->io_error == 0) {
3204 while (zio->io_cksum_report != NULL) {
3205 zio_cksum_report_t *zcr = zio->io_cksum_report;
3206 uint64_t align = zcr->zcr_align;
3207 uint64_t asize = P2ROUNDUP(psize, align);
3208 char *abuf = zio->io_data;
3210 if (asize != psize) {
3211 abuf = zio_buf_alloc(asize);
3212 bcopy(zio->io_data, abuf, psize);
3213 bzero(abuf + psize, asize - psize);
3216 zio->io_cksum_report = zcr->zcr_next;
3217 zcr->zcr_next = NULL;
3218 zcr->zcr_finish(zcr, abuf);
3219 zfs_ereport_free_checksum(zcr);
3222 zio_buf_free(abuf, asize);
3226 zio_pop_transforms(zio); /* note: may set zio->io_error */
3228 vdev_stat_update(zio, psize);
3230 if (zio->io_error) {
3232 * If this I/O is attached to a particular vdev,
3233 * generate an error message describing the I/O failure
3234 * at the block level. We ignore these errors if the
3235 * device is currently unavailable.
3237 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3238 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3240 if ((zio->io_error == EIO || !(zio->io_flags &
3241 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3244 * For logical I/O requests, tell the SPA to log the
3245 * error and generate a logical data ereport.
3247 spa_log_error(spa, zio);
3248 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3253 if (zio->io_error && zio == lio) {
3255 * Determine whether zio should be reexecuted. This will
3256 * propagate all the way to the root via zio_notify_parent().
3258 ASSERT(vd == NULL && bp != NULL);
3259 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3261 if (IO_IS_ALLOCATING(zio) &&
3262 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3263 if (zio->io_error != ENOSPC)
3264 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3266 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3269 if ((zio->io_type == ZIO_TYPE_READ ||
3270 zio->io_type == ZIO_TYPE_FREE) &&
3271 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3272 zio->io_error == ENXIO &&
3273 spa_load_state(spa) == SPA_LOAD_NONE &&
3274 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3275 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3277 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3278 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3281 * Here is a possibly good place to attempt to do
3282 * either combinatorial reconstruction or error correction
3283 * based on checksums. It also might be a good place
3284 * to send out preliminary ereports before we suspend
3290 * If there were logical child errors, they apply to us now.
3291 * We defer this until now to avoid conflating logical child
3292 * errors with errors that happened to the zio itself when
3293 * updating vdev stats and reporting FMA events above.
3295 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3297 if ((zio->io_error || zio->io_reexecute) &&
3298 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3299 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3300 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3302 zio_gang_tree_free(&zio->io_gang_tree);
3305 * Godfather I/Os should never suspend.
3307 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3308 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3309 zio->io_reexecute = 0;
3311 if (zio->io_reexecute) {
3313 * This is a logical I/O that wants to reexecute.
3315 * Reexecute is top-down. When an i/o fails, if it's not
3316 * the root, it simply notifies its parent and sticks around.
3317 * The parent, seeing that it still has children in zio_done(),
3318 * does the same. This percolates all the way up to the root.
3319 * The root i/o will reexecute or suspend the entire tree.
3321 * This approach ensures that zio_reexecute() honors
3322 * all the original i/o dependency relationships, e.g.
3323 * parents not executing until children are ready.
3325 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3327 zio->io_gang_leader = NULL;
3329 mutex_enter(&zio->io_lock);
3330 zio->io_state[ZIO_WAIT_DONE] = 1;
3331 mutex_exit(&zio->io_lock);
3334 * "The Godfather" I/O monitors its children but is
3335 * not a true parent to them. It will track them through
3336 * the pipeline but severs its ties whenever they get into
3337 * trouble (e.g. suspended). This allows "The Godfather"
3338 * I/O to return status without blocking.
3340 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3341 zio_link_t *zl = zio->io_walk_link;
3342 pio_next = zio_walk_parents(zio);
3344 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3345 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3346 zio_remove_child(pio, zio, zl);
3347 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3351 if ((pio = zio_unique_parent(zio)) != NULL) {
3353 * We're not a root i/o, so there's nothing to do
3354 * but notify our parent. Don't propagate errors
3355 * upward since we haven't permanently failed yet.
3357 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3358 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3359 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3360 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3362 * We'd fail again if we reexecuted now, so suspend
3363 * until conditions improve (e.g. device comes online).
3365 zio_suspend(spa, zio);
3368 * Reexecution is potentially a huge amount of work.
3369 * Hand it off to the otherwise-unused claim taskq.
3371 #if defined(illumos) || !defined(_KERNEL)
3372 ASSERT(zio->io_tqent.tqent_next == NULL);
3374 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
3376 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3377 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3380 return (ZIO_PIPELINE_STOP);
3383 ASSERT(zio->io_child_count == 0);
3384 ASSERT(zio->io_reexecute == 0);
3385 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3388 * Report any checksum errors, since the I/O is complete.
3390 while (zio->io_cksum_report != NULL) {
3391 zio_cksum_report_t *zcr = zio->io_cksum_report;
3392 zio->io_cksum_report = zcr->zcr_next;
3393 zcr->zcr_next = NULL;
3394 zcr->zcr_finish(zcr, NULL);
3395 zfs_ereport_free_checksum(zcr);
3399 * It is the responsibility of the done callback to ensure that this
3400 * particular zio is no longer discoverable for adoption, and as
3401 * such, cannot acquire any new parents.
3406 mutex_enter(&zio->io_lock);
3407 zio->io_state[ZIO_WAIT_DONE] = 1;
3408 mutex_exit(&zio->io_lock);
3410 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3411 zio_link_t *zl = zio->io_walk_link;
3412 pio_next = zio_walk_parents(zio);
3413 zio_remove_child(pio, zio, zl);
3414 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3417 if (zio->io_waiter != NULL) {
3418 mutex_enter(&zio->io_lock);
3419 zio->io_executor = NULL;
3420 cv_broadcast(&zio->io_cv);
3421 mutex_exit(&zio->io_lock);
3426 return (ZIO_PIPELINE_STOP);
3430 * ==========================================================================
3431 * I/O pipeline definition
3432 * ==========================================================================
3434 static zio_pipe_stage_t *zio_pipeline[] = {
3440 zio_checksum_generate,
3455 zio_checksum_verify,
3463 * Compare two zbookmark_phys_t's to see which we would reach first in a
3464 * pre-order traversal of the object tree.
3466 * This is simple in every case aside from the meta-dnode object. For all other
3467 * objects, we traverse them in order (object 1 before object 2, and so on).
3468 * However, all of these objects are traversed while traversing object 0, since
3469 * the data it points to is the list of objects. Thus, we need to convert to a
3470 * canonical representation so we can compare meta-dnode bookmarks to
3471 * non-meta-dnode bookmarks.
3473 * We do this by calculating "equivalents" for each field of the zbookmark.
3474 * zbookmarks outside of the meta-dnode use their own object and level, and
3475 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
3476 * blocks this bookmark refers to) by multiplying their blkid by their span
3477 * (the number of L0 blocks contained within one block at their level).
3478 * zbookmarks inside the meta-dnode calculate their object equivalent
3479 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
3480 * level + 1<<31 (any value larger than a level could ever be) for their level.
3481 * This causes them to always compare before a bookmark in their object
3482 * equivalent, compare appropriately to bookmarks in other objects, and to
3483 * compare appropriately to other bookmarks in the meta-dnode.
3486 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
3487 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
3490 * These variables represent the "equivalent" values for the zbookmark,
3491 * after converting zbookmarks inside the meta dnode to their
3492 * normal-object equivalents.
3494 uint64_t zb1obj, zb2obj;
3495 uint64_t zb1L0, zb2L0;
3496 uint64_t zb1level, zb2level;
3498 if (zb1->zb_object == zb2->zb_object &&
3499 zb1->zb_level == zb2->zb_level &&
3500 zb1->zb_blkid == zb2->zb_blkid)
3504 * BP_SPANB calculates the span in blocks.
3506 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
3507 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
3509 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3510 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
3512 zb1level = zb1->zb_level + COMPARE_META_LEVEL;
3514 zb1obj = zb1->zb_object;
3515 zb1level = zb1->zb_level;
3518 if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
3519 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
3521 zb2level = zb2->zb_level + COMPARE_META_LEVEL;
3523 zb2obj = zb2->zb_object;
3524 zb2level = zb2->zb_level;
3527 /* Now that we have a canonical representation, do the comparison. */
3528 if (zb1obj != zb2obj)
3529 return (zb1obj < zb2obj ? -1 : 1);
3530 else if (zb1L0 != zb2L0)
3531 return (zb1L0 < zb2L0 ? -1 : 1);
3532 else if (zb1level != zb2level)
3533 return (zb1level > zb2level ? -1 : 1);
3535 * This can (theoretically) happen if the bookmarks have the same object
3536 * and level, but different blkids, if the block sizes are not the same.
3537 * There is presently no way to change the indirect block sizes
3543 * This function checks the following: given that last_block is the place that
3544 * our traversal stopped last time, does that guarantee that we've visited
3545 * every node under subtree_root? Therefore, we can't just use the raw output
3546 * of zbookmark_compare. We have to pass in a modified version of
3547 * subtree_root; by incrementing the block id, and then checking whether
3548 * last_block is before or equal to that, we can tell whether or not having
3549 * visited last_block implies that all of subtree_root's children have been
3553 zbookmark_subtree_completed(const dnode_phys_t *dnp,
3554 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
3556 zbookmark_phys_t mod_zb = *subtree_root;
3558 ASSERT(last_block->zb_level == 0);
3560 /* The objset_phys_t isn't before anything. */
3565 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
3566 * data block size in sectors, because that variable is only used if
3567 * the bookmark refers to a block in the meta-dnode. Since we don't
3568 * know without examining it what object it refers to, and there's no
3569 * harm in passing in this value in other cases, we always pass it in.
3571 * We pass in 0 for the indirect block size shift because zb2 must be
3572 * level 0. The indirect block size is only used to calculate the span
3573 * of the bookmark, but since the bookmark must be level 0, the span is
3574 * always 1, so the math works out.
3576 * If you make changes to how the zbookmark_compare code works, be sure
3577 * to make sure that this code still works afterwards.
3579 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
3580 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,