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.
25 * Copyright (c) 2014 Integros [integros.com]
28 #include <sys/sysmacros.h>
29 #include <sys/zfs_context.h>
30 #include <sys/fm/fs/zfs.h>
33 #include <sys/spa_impl.h>
34 #include <sys/vdev_impl.h>
35 #include <sys/zio_impl.h>
36 #include <sys/zio_compress.h>
37 #include <sys/zio_checksum.h>
38 #include <sys/dmu_objset.h>
41 #include <sys/trim_map.h>
42 #include <sys/blkptr.h>
43 #include <sys/zfeature.h>
45 SYSCTL_DECL(_vfs_zfs);
46 SYSCTL_NODE(_vfs_zfs, OID_AUTO, zio, CTLFLAG_RW, 0, "ZFS ZIO");
47 #if defined(__amd64__)
48 static int zio_use_uma = 1;
50 static int zio_use_uma = 0;
52 TUNABLE_INT("vfs.zfs.zio.use_uma", &zio_use_uma);
53 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, use_uma, CTLFLAG_RDTUN, &zio_use_uma, 0,
54 "Use uma(9) for ZIO allocations");
55 static int zio_exclude_metadata = 0;
56 TUNABLE_INT("vfs.zfs.zio.exclude_metadata", &zio_exclude_metadata);
57 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, exclude_metadata, CTLFLAG_RDTUN, &zio_exclude_metadata, 0,
58 "Exclude metadata buffers from dumps as well");
60 zio_trim_stats_t zio_trim_stats = {
61 { "bytes", KSTAT_DATA_UINT64,
62 "Number of bytes successfully TRIMmed" },
63 { "success", KSTAT_DATA_UINT64,
64 "Number of successful TRIM requests" },
65 { "unsupported", KSTAT_DATA_UINT64,
66 "Number of TRIM requests that failed because TRIM is not supported" },
67 { "failed", KSTAT_DATA_UINT64,
68 "Number of TRIM requests that failed for reasons other than not supported" },
71 static kstat_t *zio_trim_ksp;
74 * ==========================================================================
75 * I/O type descriptions
76 * ==========================================================================
78 const char *zio_type_name[ZIO_TYPES] = {
79 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
84 * ==========================================================================
86 * ==========================================================================
88 kmem_cache_t *zio_cache;
89 kmem_cache_t *zio_link_cache;
90 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
91 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
94 extern vmem_t *zio_alloc_arena;
97 #define ZIO_PIPELINE_CONTINUE 0x100
98 #define ZIO_PIPELINE_STOP 0x101
100 #define BP_SPANB(indblkshift, level) \
101 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
102 #define COMPARE_META_LEVEL 0x80000000ul
104 * The following actions directly effect the spa's sync-to-convergence logic.
105 * The values below define the sync pass when we start performing the action.
106 * Care should be taken when changing these values as they directly impact
107 * spa_sync() performance. Tuning these values may introduce subtle performance
108 * pathologies and should only be done in the context of performance analysis.
109 * These tunables will eventually be removed and replaced with #defines once
110 * enough analysis has been done to determine optimal values.
112 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
113 * regular blocks are not deferred.
115 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
116 TUNABLE_INT("vfs.zfs.sync_pass_deferred_free", &zfs_sync_pass_deferred_free);
117 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_deferred_free, CTLFLAG_RDTUN,
118 &zfs_sync_pass_deferred_free, 0, "defer frees starting in this pass");
119 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
120 TUNABLE_INT("vfs.zfs.sync_pass_dont_compress", &zfs_sync_pass_dont_compress);
121 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_dont_compress, CTLFLAG_RDTUN,
122 &zfs_sync_pass_dont_compress, 0, "don't compress starting in this pass");
123 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
124 TUNABLE_INT("vfs.zfs.sync_pass_rewrite", &zfs_sync_pass_rewrite);
125 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_rewrite, CTLFLAG_RDTUN,
126 &zfs_sync_pass_rewrite, 0, "rewrite new bps starting in this pass");
129 * An allocating zio is one that either currently has the DVA allocate
130 * stage set or will have it later in its lifetime.
132 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
134 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
138 int zio_buf_debug_limit = 16384;
140 int zio_buf_debug_limit = 0;
148 zio_cache = kmem_cache_create("zio_cache",
149 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
150 zio_link_cache = kmem_cache_create("zio_link_cache",
151 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
156 * For small buffers, we want a cache for each multiple of
157 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
158 * for each quarter-power of 2.
160 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
161 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
164 int cflags = zio_exclude_metadata ? KMC_NODEBUG : 0;
172 * If we are using watchpoints, put each buffer on its own page,
173 * to eliminate the performance overhead of trapping to the
174 * kernel when modifying a non-watched buffer that shares the
175 * page with a watched buffer.
177 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
181 if (size <= 4 * SPA_MINBLOCKSIZE) {
182 align = SPA_MINBLOCKSIZE;
183 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
184 align = MIN(p2 >> 2, PAGESIZE);
189 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
190 zio_buf_cache[c] = kmem_cache_create(name, size,
191 align, NULL, NULL, NULL, NULL, NULL, cflags);
194 * Since zio_data bufs do not appear in crash dumps, we
195 * pass KMC_NOTOUCH so that no allocator metadata is
196 * stored with the buffers.
198 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
199 zio_data_buf_cache[c] = kmem_cache_create(name, size,
200 align, NULL, NULL, NULL, NULL, NULL,
201 cflags | KMC_NOTOUCH | KMC_NODEBUG);
206 ASSERT(zio_buf_cache[c] != NULL);
207 if (zio_buf_cache[c - 1] == NULL)
208 zio_buf_cache[c - 1] = zio_buf_cache[c];
210 ASSERT(zio_data_buf_cache[c] != NULL);
211 if (zio_data_buf_cache[c - 1] == NULL)
212 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
218 zio_trim_ksp = kstat_create("zfs", 0, "zio_trim", "misc",
220 sizeof(zio_trim_stats) / sizeof(kstat_named_t),
223 if (zio_trim_ksp != NULL) {
224 zio_trim_ksp->ks_data = &zio_trim_stats;
225 kstat_install(zio_trim_ksp);
233 kmem_cache_t *last_cache = NULL;
234 kmem_cache_t *last_data_cache = NULL;
236 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
237 if (zio_buf_cache[c] != last_cache) {
238 last_cache = zio_buf_cache[c];
239 kmem_cache_destroy(zio_buf_cache[c]);
241 zio_buf_cache[c] = NULL;
243 if (zio_data_buf_cache[c] != last_data_cache) {
244 last_data_cache = zio_data_buf_cache[c];
245 kmem_cache_destroy(zio_data_buf_cache[c]);
247 zio_data_buf_cache[c] = NULL;
250 kmem_cache_destroy(zio_link_cache);
251 kmem_cache_destroy(zio_cache);
255 if (zio_trim_ksp != NULL) {
256 kstat_delete(zio_trim_ksp);
262 * ==========================================================================
263 * Allocate and free I/O buffers
264 * ==========================================================================
268 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
269 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
270 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
271 * excess / transient data in-core during a crashdump.
274 zio_buf_alloc(size_t size)
276 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
277 int flags = zio_exclude_metadata ? KM_NODEBUG : 0;
279 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
282 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
284 return (kmem_alloc(size, KM_SLEEP|flags));
288 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
289 * crashdump if the kernel panics. This exists so that we will limit the amount
290 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
291 * of kernel heap dumped to disk when the kernel panics)
294 zio_data_buf_alloc(size_t size)
296 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
298 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
301 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
303 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG));
307 zio_buf_free(void *buf, size_t size)
309 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
311 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
314 kmem_cache_free(zio_buf_cache[c], buf);
316 kmem_free(buf, size);
320 zio_data_buf_free(void *buf, size_t size)
322 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
324 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
327 kmem_cache_free(zio_data_buf_cache[c], buf);
329 kmem_free(buf, size);
333 * ==========================================================================
334 * Push and pop I/O transform buffers
335 * ==========================================================================
338 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
339 zio_transform_func_t *transform)
341 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
343 zt->zt_orig_data = zio->io_data;
344 zt->zt_orig_size = zio->io_size;
345 zt->zt_bufsize = bufsize;
346 zt->zt_transform = transform;
348 zt->zt_next = zio->io_transform_stack;
349 zio->io_transform_stack = zt;
356 zio_pop_transforms(zio_t *zio)
360 while ((zt = zio->io_transform_stack) != NULL) {
361 if (zt->zt_transform != NULL)
362 zt->zt_transform(zio,
363 zt->zt_orig_data, zt->zt_orig_size);
365 if (zt->zt_bufsize != 0)
366 zio_buf_free(zio->io_data, zt->zt_bufsize);
368 zio->io_data = zt->zt_orig_data;
369 zio->io_size = zt->zt_orig_size;
370 zio->io_transform_stack = zt->zt_next;
372 kmem_free(zt, sizeof (zio_transform_t));
377 * ==========================================================================
378 * I/O transform callbacks for subblocks and decompression
379 * ==========================================================================
382 zio_subblock(zio_t *zio, void *data, uint64_t size)
384 ASSERT(zio->io_size > size);
386 if (zio->io_type == ZIO_TYPE_READ)
387 bcopy(zio->io_data, data, size);
391 zio_decompress(zio_t *zio, void *data, uint64_t size)
393 if (zio->io_error == 0 &&
394 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
395 zio->io_data, data, zio->io_size, size) != 0)
396 zio->io_error = SET_ERROR(EIO);
400 * ==========================================================================
401 * I/O parent/child relationships and pipeline interlocks
402 * ==========================================================================
405 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
406 * continue calling these functions until they return NULL.
407 * Otherwise, the next caller will pick up the list walk in
408 * some indeterminate state. (Otherwise every caller would
409 * have to pass in a cookie to keep the state represented by
410 * io_walk_link, which gets annoying.)
413 zio_walk_parents(zio_t *cio)
415 zio_link_t *zl = cio->io_walk_link;
416 list_t *pl = &cio->io_parent_list;
418 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
419 cio->io_walk_link = zl;
424 ASSERT(zl->zl_child == cio);
425 return (zl->zl_parent);
429 zio_walk_children(zio_t *pio)
431 zio_link_t *zl = pio->io_walk_link;
432 list_t *cl = &pio->io_child_list;
434 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
435 pio->io_walk_link = zl;
440 ASSERT(zl->zl_parent == pio);
441 return (zl->zl_child);
445 zio_unique_parent(zio_t *cio)
447 zio_t *pio = zio_walk_parents(cio);
449 VERIFY(zio_walk_parents(cio) == NULL);
454 zio_add_child(zio_t *pio, zio_t *cio)
456 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
459 * Logical I/Os can have logical, gang, or vdev children.
460 * Gang I/Os can have gang or vdev children.
461 * Vdev I/Os can only have vdev children.
462 * The following ASSERT captures all of these constraints.
464 ASSERT(cio->io_child_type <= pio->io_child_type);
469 mutex_enter(&cio->io_lock);
470 mutex_enter(&pio->io_lock);
472 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
474 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
475 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
477 list_insert_head(&pio->io_child_list, zl);
478 list_insert_head(&cio->io_parent_list, zl);
480 pio->io_child_count++;
481 cio->io_parent_count++;
483 mutex_exit(&pio->io_lock);
484 mutex_exit(&cio->io_lock);
488 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
490 ASSERT(zl->zl_parent == pio);
491 ASSERT(zl->zl_child == cio);
493 mutex_enter(&cio->io_lock);
494 mutex_enter(&pio->io_lock);
496 list_remove(&pio->io_child_list, zl);
497 list_remove(&cio->io_parent_list, zl);
499 pio->io_child_count--;
500 cio->io_parent_count--;
502 mutex_exit(&pio->io_lock);
503 mutex_exit(&cio->io_lock);
505 kmem_cache_free(zio_link_cache, zl);
509 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
511 uint64_t *countp = &zio->io_children[child][wait];
512 boolean_t waiting = B_FALSE;
514 mutex_enter(&zio->io_lock);
515 ASSERT(zio->io_stall == NULL);
518 zio->io_stall = countp;
521 mutex_exit(&zio->io_lock);
527 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
529 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
530 int *errorp = &pio->io_child_error[zio->io_child_type];
532 mutex_enter(&pio->io_lock);
533 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
534 *errorp = zio_worst_error(*errorp, zio->io_error);
535 pio->io_reexecute |= zio->io_reexecute;
536 ASSERT3U(*countp, >, 0);
540 if (*countp == 0 && pio->io_stall == countp) {
541 pio->io_stall = NULL;
542 mutex_exit(&pio->io_lock);
545 mutex_exit(&pio->io_lock);
550 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
552 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
553 zio->io_error = zio->io_child_error[c];
557 * ==========================================================================
558 * Create the various types of I/O (read, write, free, etc)
559 * ==========================================================================
562 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
563 void *data, uint64_t size, zio_done_func_t *done, void *private,
564 zio_type_t type, zio_priority_t priority, enum zio_flag flags,
565 vdev_t *vd, uint64_t offset, const zbookmark_phys_t *zb,
566 enum zio_stage stage, enum zio_stage pipeline)
570 ASSERT3U(type == ZIO_TYPE_FREE || size, <=, SPA_MAXBLOCKSIZE);
571 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
572 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
574 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
575 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
576 ASSERT(vd || stage == ZIO_STAGE_OPEN);
578 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
579 bzero(zio, sizeof (zio_t));
581 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
582 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
584 list_create(&zio->io_parent_list, sizeof (zio_link_t),
585 offsetof(zio_link_t, zl_parent_node));
586 list_create(&zio->io_child_list, sizeof (zio_link_t),
587 offsetof(zio_link_t, zl_child_node));
590 zio->io_child_type = ZIO_CHILD_VDEV;
591 else if (flags & ZIO_FLAG_GANG_CHILD)
592 zio->io_child_type = ZIO_CHILD_GANG;
593 else if (flags & ZIO_FLAG_DDT_CHILD)
594 zio->io_child_type = ZIO_CHILD_DDT;
596 zio->io_child_type = ZIO_CHILD_LOGICAL;
599 zio->io_bp = (blkptr_t *)bp;
600 zio->io_bp_copy = *bp;
601 zio->io_bp_orig = *bp;
602 if (type != ZIO_TYPE_WRITE ||
603 zio->io_child_type == ZIO_CHILD_DDT)
604 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
605 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
606 zio->io_logical = zio;
607 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
608 pipeline |= ZIO_GANG_STAGES;
614 zio->io_private = private;
616 zio->io_priority = priority;
618 zio->io_offset = offset;
619 zio->io_orig_data = zio->io_data = data;
620 zio->io_orig_size = zio->io_size = size;
621 zio->io_orig_flags = zio->io_flags = flags;
622 zio->io_orig_stage = zio->io_stage = stage;
623 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
625 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
626 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
629 zio->io_bookmark = *zb;
632 if (zio->io_logical == NULL)
633 zio->io_logical = pio->io_logical;
634 if (zio->io_child_type == ZIO_CHILD_GANG)
635 zio->io_gang_leader = pio->io_gang_leader;
636 zio_add_child(pio, zio);
643 zio_destroy(zio_t *zio)
645 list_destroy(&zio->io_parent_list);
646 list_destroy(&zio->io_child_list);
647 mutex_destroy(&zio->io_lock);
648 cv_destroy(&zio->io_cv);
649 kmem_cache_free(zio_cache, zio);
653 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
654 void *private, enum zio_flag flags)
658 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
659 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
660 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
666 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
668 return (zio_null(NULL, spa, NULL, done, private, flags));
672 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
674 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
675 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
676 bp, (longlong_t)BP_GET_TYPE(bp));
678 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
679 BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
680 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
681 bp, (longlong_t)BP_GET_CHECKSUM(bp));
683 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
684 BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
685 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
686 bp, (longlong_t)BP_GET_COMPRESS(bp));
688 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
689 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
690 bp, (longlong_t)BP_GET_LSIZE(bp));
692 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
693 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
694 bp, (longlong_t)BP_GET_PSIZE(bp));
697 if (BP_IS_EMBEDDED(bp)) {
698 if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
699 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
700 bp, (longlong_t)BPE_GET_ETYPE(bp));
705 * Pool-specific checks.
707 * Note: it would be nice to verify that the blk_birth and
708 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
709 * allows the birth time of log blocks (and dmu_sync()-ed blocks
710 * that are in the log) to be arbitrarily large.
712 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
713 uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
714 if (vdevid >= spa->spa_root_vdev->vdev_children) {
715 zfs_panic_recover("blkptr at %p DVA %u has invalid "
717 bp, i, (longlong_t)vdevid);
720 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
722 zfs_panic_recover("blkptr at %p DVA %u has invalid "
724 bp, i, (longlong_t)vdevid);
727 if (vd->vdev_ops == &vdev_hole_ops) {
728 zfs_panic_recover("blkptr at %p DVA %u has hole "
730 bp, i, (longlong_t)vdevid);
733 if (vd->vdev_ops == &vdev_missing_ops) {
735 * "missing" vdevs are valid during import, but we
736 * don't have their detailed info (e.g. asize), so
737 * we can't perform any more checks on them.
741 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
742 uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
744 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
745 if (offset + asize > vd->vdev_asize) {
746 zfs_panic_recover("blkptr at %p DVA %u has invalid "
748 bp, i, (longlong_t)offset);
754 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
755 void *data, uint64_t size, zio_done_func_t *done, void *private,
756 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
760 zfs_blkptr_verify(spa, bp);
762 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
763 data, size, done, private,
764 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
765 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
766 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
772 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
773 void *data, uint64_t size, const zio_prop_t *zp,
774 zio_done_func_t *ready, zio_done_func_t *physdone, zio_done_func_t *done,
776 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
780 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
781 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
782 zp->zp_compress >= ZIO_COMPRESS_OFF &&
783 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
784 DMU_OT_IS_VALID(zp->zp_type) &&
787 zp->zp_copies <= spa_max_replication(spa));
789 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
790 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
791 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
792 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
794 zio->io_ready = ready;
795 zio->io_physdone = physdone;
799 * Data can be NULL if we are going to call zio_write_override() to
800 * provide the already-allocated BP. But we may need the data to
801 * verify a dedup hit (if requested). In this case, don't try to
802 * dedup (just take the already-allocated BP verbatim).
804 if (data == NULL && zio->io_prop.zp_dedup_verify) {
805 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
812 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
813 uint64_t size, zio_done_func_t *done, void *private,
814 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
818 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
819 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
820 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
826 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
828 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
829 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
830 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
831 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
834 * We must reset the io_prop to match the values that existed
835 * when the bp was first written by dmu_sync() keeping in mind
836 * that nopwrite and dedup are mutually exclusive.
838 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
839 zio->io_prop.zp_nopwrite = nopwrite;
840 zio->io_prop.zp_copies = copies;
841 zio->io_bp_override = bp;
845 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
849 * The check for EMBEDDED is a performance optimization. We
850 * process the free here (by ignoring it) rather than
851 * putting it on the list and then processing it in zio_free_sync().
853 if (BP_IS_EMBEDDED(bp))
855 metaslab_check_free(spa, bp);
858 * Frees that are for the currently-syncing txg, are not going to be
859 * deferred, and which will not need to do a read (i.e. not GANG or
860 * DEDUP), can be processed immediately. Otherwise, put them on the
861 * in-memory list for later processing.
863 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
864 txg != spa->spa_syncing_txg ||
865 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
866 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
868 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp,
869 BP_GET_PSIZE(bp), 0)));
874 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
875 uint64_t size, enum zio_flag flags)
878 enum zio_stage stage = ZIO_FREE_PIPELINE;
880 ASSERT(!BP_IS_HOLE(bp));
881 ASSERT(spa_syncing_txg(spa) == txg);
882 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
884 if (BP_IS_EMBEDDED(bp))
885 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
887 metaslab_check_free(spa, bp);
890 if (zfs_trim_enabled)
891 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START |
892 ZIO_STAGE_VDEV_IO_ASSESS;
894 * GANG and DEDUP blocks can induce a read (for the gang block header,
895 * or the DDT), so issue them asynchronously so that this thread is
898 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
899 stage |= ZIO_STAGE_ISSUE_ASYNC;
901 flags |= ZIO_FLAG_DONT_QUEUE;
903 zio = zio_create(pio, spa, txg, bp, NULL, size,
904 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags,
905 NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
911 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
912 zio_done_func_t *done, void *private, enum zio_flag flags)
916 dprintf_bp(bp, "claiming in txg %llu", txg);
918 if (BP_IS_EMBEDDED(bp))
919 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
922 * A claim is an allocation of a specific block. Claims are needed
923 * to support immediate writes in the intent log. The issue is that
924 * immediate writes contain committed data, but in a txg that was
925 * *not* committed. Upon opening the pool after an unclean shutdown,
926 * the intent log claims all blocks that contain immediate write data
927 * so that the SPA knows they're in use.
929 * All claims *must* be resolved in the first txg -- before the SPA
930 * starts allocating blocks -- so that nothing is allocated twice.
931 * If txg == 0 we just verify that the block is claimable.
933 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
934 ASSERT(txg == spa_first_txg(spa) || txg == 0);
935 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
937 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
938 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
939 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
945 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset,
946 uint64_t size, zio_done_func_t *done, void *private,
947 zio_priority_t priority, enum zio_flag flags)
952 if (vd->vdev_children == 0) {
953 zio = zio_create(pio, spa, 0, NULL, NULL, size, done, private,
954 ZIO_TYPE_IOCTL, priority, flags, vd, offset, NULL,
955 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
959 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
961 for (c = 0; c < vd->vdev_children; c++)
962 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
963 offset, size, done, private, priority, flags));
970 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
971 void *data, int checksum, zio_done_func_t *done, void *private,
972 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
976 ASSERT(vd->vdev_children == 0);
977 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
978 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
979 ASSERT3U(offset + size, <=, vd->vdev_psize);
981 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
982 ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
983 NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
985 zio->io_prop.zp_checksum = checksum;
991 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
992 void *data, int checksum, zio_done_func_t *done, void *private,
993 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
997 ASSERT(vd->vdev_children == 0);
998 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
999 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1000 ASSERT3U(offset + size, <=, vd->vdev_psize);
1002 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
1003 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
1004 NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
1006 zio->io_prop.zp_checksum = checksum;
1008 if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
1010 * zec checksums are necessarily destructive -- they modify
1011 * the end of the write buffer to hold the verifier/checksum.
1012 * Therefore, we must make a local copy in case the data is
1013 * being written to multiple places in parallel.
1015 void *wbuf = zio_buf_alloc(size);
1016 bcopy(data, wbuf, size);
1017 zio_push_transform(zio, wbuf, size, size, NULL);
1024 * Create a child I/O to do some work for us.
1027 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1028 void *data, uint64_t size, int type, zio_priority_t priority,
1029 enum zio_flag flags, zio_done_func_t *done, void *private)
1031 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1034 ASSERT(vd->vdev_parent ==
1035 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
1037 if (type == ZIO_TYPE_READ && bp != NULL) {
1039 * If we have the bp, then the child should perform the
1040 * checksum and the parent need not. This pushes error
1041 * detection as close to the leaves as possible and
1042 * eliminates redundant checksums in the interior nodes.
1044 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1045 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1048 /* Not all IO types require vdev io done stage e.g. free */
1049 if (!(pio->io_pipeline & ZIO_STAGE_VDEV_IO_DONE))
1050 pipeline &= ~ZIO_STAGE_VDEV_IO_DONE;
1052 if (vd->vdev_children == 0)
1053 offset += VDEV_LABEL_START_SIZE;
1055 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
1058 * If we've decided to do a repair, the write is not speculative --
1059 * even if the original read was.
1061 if (flags & ZIO_FLAG_IO_REPAIR)
1062 flags &= ~ZIO_FLAG_SPECULATIVE;
1064 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
1065 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1066 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1068 zio->io_physdone = pio->io_physdone;
1069 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1070 zio->io_logical->io_phys_children++;
1076 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
1077 int type, zio_priority_t priority, enum zio_flag flags,
1078 zio_done_func_t *done, void *private)
1082 ASSERT(vd->vdev_ops->vdev_op_leaf);
1084 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1085 data, size, done, private, type, priority,
1086 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1088 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1094 zio_flush(zio_t *zio, vdev_t *vd)
1096 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0,
1097 NULL, NULL, ZIO_PRIORITY_NOW,
1098 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1102 zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size)
1105 ASSERT(vd->vdev_ops->vdev_op_leaf);
1107 return (zio_create(zio, spa, 0, NULL, NULL, size, NULL, NULL,
1108 ZIO_TYPE_FREE, ZIO_PRIORITY_TRIM, ZIO_FLAG_DONT_AGGREGATE |
1109 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY,
1110 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PHYS_PIPELINE));
1114 zio_shrink(zio_t *zio, uint64_t size)
1116 ASSERT(zio->io_executor == NULL);
1117 ASSERT(zio->io_orig_size == zio->io_size);
1118 ASSERT(size <= zio->io_size);
1121 * We don't shrink for raidz because of problems with the
1122 * reconstruction when reading back less than the block size.
1123 * Note, BP_IS_RAIDZ() assumes no compression.
1125 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1126 if (!BP_IS_RAIDZ(zio->io_bp))
1127 zio->io_orig_size = zio->io_size = size;
1131 * ==========================================================================
1132 * Prepare to read and write logical blocks
1133 * ==========================================================================
1137 zio_read_bp_init(zio_t *zio)
1139 blkptr_t *bp = zio->io_bp;
1141 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1142 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1143 !(zio->io_flags & ZIO_FLAG_RAW)) {
1145 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1146 void *cbuf = zio_buf_alloc(psize);
1148 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
1151 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1152 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1153 decode_embedded_bp_compressed(bp, zio->io_data);
1155 ASSERT(!BP_IS_EMBEDDED(bp));
1158 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1159 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1161 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1162 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1164 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1165 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1167 return (ZIO_PIPELINE_CONTINUE);
1171 zio_write_bp_init(zio_t *zio)
1173 spa_t *spa = zio->io_spa;
1174 zio_prop_t *zp = &zio->io_prop;
1175 enum zio_compress compress = zp->zp_compress;
1176 blkptr_t *bp = zio->io_bp;
1177 uint64_t lsize = zio->io_size;
1178 uint64_t psize = lsize;
1182 * If our children haven't all reached the ready stage,
1183 * wait for them and then repeat this pipeline stage.
1185 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1186 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1187 return (ZIO_PIPELINE_STOP);
1189 if (!IO_IS_ALLOCATING(zio))
1190 return (ZIO_PIPELINE_CONTINUE);
1192 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1194 if (zio->io_bp_override) {
1195 ASSERT(bp->blk_birth != zio->io_txg);
1196 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1198 *bp = *zio->io_bp_override;
1199 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1201 if (BP_IS_EMBEDDED(bp))
1202 return (ZIO_PIPELINE_CONTINUE);
1205 * If we've been overridden and nopwrite is set then
1206 * set the flag accordingly to indicate that a nopwrite
1207 * has already occurred.
1209 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1210 ASSERT(!zp->zp_dedup);
1211 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1212 return (ZIO_PIPELINE_CONTINUE);
1215 ASSERT(!zp->zp_nopwrite);
1217 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1218 return (ZIO_PIPELINE_CONTINUE);
1220 ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags &
1221 ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify);
1223 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1224 BP_SET_DEDUP(bp, 1);
1225 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1226 return (ZIO_PIPELINE_CONTINUE);
1228 zio->io_bp_override = NULL;
1232 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1234 * We're rewriting an existing block, which means we're
1235 * working on behalf of spa_sync(). For spa_sync() to
1236 * converge, it must eventually be the case that we don't
1237 * have to allocate new blocks. But compression changes
1238 * the blocksize, which forces a reallocate, and makes
1239 * convergence take longer. Therefore, after the first
1240 * few passes, stop compressing to ensure convergence.
1242 pass = spa_sync_pass(spa);
1244 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1245 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1246 ASSERT(!BP_GET_DEDUP(bp));
1248 if (pass >= zfs_sync_pass_dont_compress)
1249 compress = ZIO_COMPRESS_OFF;
1251 /* Make sure someone doesn't change their mind on overwrites */
1252 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1253 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1256 if (compress != ZIO_COMPRESS_OFF) {
1257 void *cbuf = zio_buf_alloc(lsize);
1258 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1259 if (psize == 0 || psize == lsize) {
1260 compress = ZIO_COMPRESS_OFF;
1261 zio_buf_free(cbuf, lsize);
1262 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1263 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1264 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1265 encode_embedded_bp_compressed(bp,
1266 cbuf, compress, lsize, psize);
1267 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1268 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1269 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1270 zio_buf_free(cbuf, lsize);
1271 bp->blk_birth = zio->io_txg;
1272 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1273 ASSERT(spa_feature_is_active(spa,
1274 SPA_FEATURE_EMBEDDED_DATA));
1275 return (ZIO_PIPELINE_CONTINUE);
1278 * Round up compressed size up to the ashift
1279 * of the smallest-ashift device, and zero the tail.
1280 * This ensures that the compressed size of the BP
1281 * (and thus compressratio property) are correct,
1282 * in that we charge for the padding used to fill out
1285 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
1286 size_t rounded = (size_t)P2ROUNDUP(psize,
1287 1ULL << spa->spa_min_ashift);
1288 if (rounded >= lsize) {
1289 compress = ZIO_COMPRESS_OFF;
1290 zio_buf_free(cbuf, lsize);
1293 bzero((char *)cbuf + psize, rounded - psize);
1295 zio_push_transform(zio, cbuf,
1296 psize, lsize, NULL);
1302 * The final pass of spa_sync() must be all rewrites, but the first
1303 * few passes offer a trade-off: allocating blocks defers convergence,
1304 * but newly allocated blocks are sequential, so they can be written
1305 * to disk faster. Therefore, we allow the first few passes of
1306 * spa_sync() to allocate new blocks, but force rewrites after that.
1307 * There should only be a handful of blocks after pass 1 in any case.
1309 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1310 BP_GET_PSIZE(bp) == psize &&
1311 pass >= zfs_sync_pass_rewrite) {
1313 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1314 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1315 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1318 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1322 if (zio->io_bp_orig.blk_birth != 0 &&
1323 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
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_BIRTH(bp, zio->io_txg, 0);
1329 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1331 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1332 BP_SET_LSIZE(bp, lsize);
1333 BP_SET_TYPE(bp, zp->zp_type);
1334 BP_SET_LEVEL(bp, zp->zp_level);
1335 BP_SET_PSIZE(bp, psize);
1336 BP_SET_COMPRESS(bp, compress);
1337 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1338 BP_SET_DEDUP(bp, zp->zp_dedup);
1339 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1341 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1342 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1343 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1345 if (zp->zp_nopwrite) {
1346 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1347 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1348 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1352 return (ZIO_PIPELINE_CONTINUE);
1356 zio_free_bp_init(zio_t *zio)
1358 blkptr_t *bp = zio->io_bp;
1360 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1361 if (BP_GET_DEDUP(bp))
1362 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1365 return (ZIO_PIPELINE_CONTINUE);
1369 * ==========================================================================
1370 * Execute the I/O pipeline
1371 * ==========================================================================
1375 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1377 spa_t *spa = zio->io_spa;
1378 zio_type_t t = zio->io_type;
1379 int flags = (cutinline ? TQ_FRONT : 0);
1381 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1384 * If we're a config writer or a probe, the normal issue and
1385 * interrupt threads may all be blocked waiting for the config lock.
1386 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1388 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1392 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1394 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1398 * If this is a high priority I/O, then use the high priority taskq if
1401 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1402 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1405 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1408 * NB: We are assuming that the zio can only be dispatched
1409 * to a single taskq at a time. It would be a grievous error
1410 * to dispatch the zio to another taskq at the same time.
1412 #if defined(illumos) || !defined(_KERNEL)
1413 ASSERT(zio->io_tqent.tqent_next == NULL);
1415 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
1417 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1418 flags, &zio->io_tqent);
1422 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1424 kthread_t *executor = zio->io_executor;
1425 spa_t *spa = zio->io_spa;
1427 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1428 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1430 for (i = 0; i < tqs->stqs_count; i++) {
1431 if (taskq_member(tqs->stqs_taskq[i], executor))
1440 zio_issue_async(zio_t *zio)
1442 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1444 return (ZIO_PIPELINE_STOP);
1448 zio_interrupt(zio_t *zio)
1450 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1454 zio_delay_interrupt(zio_t *zio)
1457 * The timeout_generic() function isn't defined in userspace, so
1458 * rather than trying to implement the function, the zio delay
1459 * functionality has been disabled for userspace builds.
1464 * If io_target_timestamp is zero, then no delay has been registered
1465 * for this IO, thus jump to the end of this function and "skip" the
1466 * delay; issuing it directly to the zio layer.
1468 if (zio->io_target_timestamp != 0) {
1469 hrtime_t now = gethrtime();
1471 if (now >= zio->io_target_timestamp) {
1473 * This IO has already taken longer than the target
1474 * delay to complete, so we don't want to delay it
1475 * any longer; we "miss" the delay and issue it
1476 * directly to the zio layer. This is likely due to
1477 * the target latency being set to a value less than
1478 * the underlying hardware can satisfy (e.g. delay
1479 * set to 1ms, but the disks take 10ms to complete an
1483 DTRACE_PROBE2(zio__delay__miss, zio_t *, zio,
1488 hrtime_t diff = zio->io_target_timestamp - now;
1490 DTRACE_PROBE3(zio__delay__hit, zio_t *, zio,
1491 hrtime_t, now, hrtime_t, diff);
1493 (void) timeout_generic(CALLOUT_NORMAL,
1494 (void (*)(void *))zio_interrupt, zio, diff, 1, 0);
1501 DTRACE_PROBE1(zio__delay__skip, zio_t *, zio);
1506 * Execute the I/O pipeline until one of the following occurs:
1508 * (1) the I/O completes
1509 * (2) the pipeline stalls waiting for dependent child I/Os
1510 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1511 * (4) the I/O is delegated by vdev-level caching or aggregation
1512 * (5) the I/O is deferred due to vdev-level queueing
1513 * (6) the I/O is handed off to another thread.
1515 * In all cases, the pipeline stops whenever there's no CPU work; it never
1516 * burns a thread in cv_wait().
1518 * There's no locking on io_stage because there's no legitimate way
1519 * for multiple threads to be attempting to process the same I/O.
1521 static zio_pipe_stage_t *zio_pipeline[];
1524 zio_execute(zio_t *zio)
1526 zio->io_executor = curthread;
1528 while (zio->io_stage < ZIO_STAGE_DONE) {
1529 enum zio_stage pipeline = zio->io_pipeline;
1530 enum zio_stage stage = zio->io_stage;
1533 ASSERT(!MUTEX_HELD(&zio->io_lock));
1534 ASSERT(ISP2(stage));
1535 ASSERT(zio->io_stall == NULL);
1539 } while ((stage & pipeline) == 0);
1541 ASSERT(stage <= ZIO_STAGE_DONE);
1544 * If we are in interrupt context and this pipeline stage
1545 * will grab a config lock that is held across I/O,
1546 * or may wait for an I/O that needs an interrupt thread
1547 * to complete, issue async to avoid deadlock.
1549 * For VDEV_IO_START, we cut in line so that the io will
1550 * be sent to disk promptly.
1552 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1553 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1554 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1555 zio_requeue_io_start_cut_in_line : B_FALSE;
1556 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1560 zio->io_stage = stage;
1561 rv = zio_pipeline[highbit64(stage) - 1](zio);
1563 if (rv == ZIO_PIPELINE_STOP)
1566 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1571 * ==========================================================================
1572 * Initiate I/O, either sync or async
1573 * ==========================================================================
1576 zio_wait(zio_t *zio)
1580 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1581 ASSERT(zio->io_executor == NULL);
1583 zio->io_waiter = curthread;
1587 mutex_enter(&zio->io_lock);
1588 while (zio->io_executor != NULL)
1589 cv_wait(&zio->io_cv, &zio->io_lock);
1590 mutex_exit(&zio->io_lock);
1592 error = zio->io_error;
1599 zio_nowait(zio_t *zio)
1601 ASSERT(zio->io_executor == NULL);
1603 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1604 zio_unique_parent(zio) == NULL) {
1606 * This is a logical async I/O with no parent to wait for it.
1607 * We add it to the spa_async_root_zio "Godfather" I/O which
1608 * will ensure they complete prior to unloading the pool.
1610 spa_t *spa = zio->io_spa;
1612 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1619 * ==========================================================================
1620 * Reexecute or suspend/resume failed I/O
1621 * ==========================================================================
1625 zio_reexecute(zio_t *pio)
1627 zio_t *cio, *cio_next;
1629 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1630 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1631 ASSERT(pio->io_gang_leader == NULL);
1632 ASSERT(pio->io_gang_tree == NULL);
1634 pio->io_flags = pio->io_orig_flags;
1635 pio->io_stage = pio->io_orig_stage;
1636 pio->io_pipeline = pio->io_orig_pipeline;
1637 pio->io_reexecute = 0;
1638 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1640 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1641 pio->io_state[w] = 0;
1642 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1643 pio->io_child_error[c] = 0;
1645 if (IO_IS_ALLOCATING(pio))
1646 BP_ZERO(pio->io_bp);
1649 * As we reexecute pio's children, new children could be created.
1650 * New children go to the head of pio's io_child_list, however,
1651 * so we will (correctly) not reexecute them. The key is that
1652 * the remainder of pio's io_child_list, from 'cio_next' onward,
1653 * cannot be affected by any side effects of reexecuting 'cio'.
1655 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1656 cio_next = zio_walk_children(pio);
1657 mutex_enter(&pio->io_lock);
1658 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1659 pio->io_children[cio->io_child_type][w]++;
1660 mutex_exit(&pio->io_lock);
1665 * Now that all children have been reexecuted, execute the parent.
1666 * We don't reexecute "The Godfather" I/O here as it's the
1667 * responsibility of the caller to wait on him.
1669 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1674 zio_suspend(spa_t *spa, zio_t *zio)
1676 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1677 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1678 "failure and the failure mode property for this pool "
1679 "is set to panic.", spa_name(spa));
1681 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1683 mutex_enter(&spa->spa_suspend_lock);
1685 if (spa->spa_suspend_zio_root == NULL)
1686 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1687 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1688 ZIO_FLAG_GODFATHER);
1690 spa->spa_suspended = B_TRUE;
1693 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1694 ASSERT(zio != spa->spa_suspend_zio_root);
1695 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1696 ASSERT(zio_unique_parent(zio) == NULL);
1697 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1698 zio_add_child(spa->spa_suspend_zio_root, zio);
1701 mutex_exit(&spa->spa_suspend_lock);
1705 zio_resume(spa_t *spa)
1710 * Reexecute all previously suspended i/o.
1712 mutex_enter(&spa->spa_suspend_lock);
1713 spa->spa_suspended = B_FALSE;
1714 cv_broadcast(&spa->spa_suspend_cv);
1715 pio = spa->spa_suspend_zio_root;
1716 spa->spa_suspend_zio_root = NULL;
1717 mutex_exit(&spa->spa_suspend_lock);
1723 return (zio_wait(pio));
1727 zio_resume_wait(spa_t *spa)
1729 mutex_enter(&spa->spa_suspend_lock);
1730 while (spa_suspended(spa))
1731 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1732 mutex_exit(&spa->spa_suspend_lock);
1736 * ==========================================================================
1739 * A gang block is a collection of small blocks that looks to the DMU
1740 * like one large block. When zio_dva_allocate() cannot find a block
1741 * of the requested size, due to either severe fragmentation or the pool
1742 * being nearly full, it calls zio_write_gang_block() to construct the
1743 * block from smaller fragments.
1745 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1746 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1747 * an indirect block: it's an array of block pointers. It consumes
1748 * only one sector and hence is allocatable regardless of fragmentation.
1749 * The gang header's bps point to its gang members, which hold the data.
1751 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1752 * as the verifier to ensure uniqueness of the SHA256 checksum.
1753 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1754 * not the gang header. This ensures that data block signatures (needed for
1755 * deduplication) are independent of how the block is physically stored.
1757 * Gang blocks can be nested: a gang member may itself be a gang block.
1758 * Thus every gang block is a tree in which root and all interior nodes are
1759 * gang headers, and the leaves are normal blocks that contain user data.
1760 * The root of the gang tree is called the gang leader.
1762 * To perform any operation (read, rewrite, free, claim) on a gang block,
1763 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1764 * in the io_gang_tree field of the original logical i/o by recursively
1765 * reading the gang leader and all gang headers below it. This yields
1766 * an in-core tree containing the contents of every gang header and the
1767 * bps for every constituent of the gang block.
1769 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1770 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1771 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1772 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1773 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1774 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1775 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1776 * of the gang header plus zio_checksum_compute() of the data to update the
1777 * gang header's blk_cksum as described above.
1779 * The two-phase assemble/issue model solves the problem of partial failure --
1780 * what if you'd freed part of a gang block but then couldn't read the
1781 * gang header for another part? Assembling the entire gang tree first
1782 * ensures that all the necessary gang header I/O has succeeded before
1783 * starting the actual work of free, claim, or write. Once the gang tree
1784 * is assembled, free and claim are in-memory operations that cannot fail.
1786 * In the event that a gang write fails, zio_dva_unallocate() walks the
1787 * gang tree to immediately free (i.e. insert back into the space map)
1788 * everything we've allocated. This ensures that we don't get ENOSPC
1789 * errors during repeated suspend/resume cycles due to a flaky device.
1791 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1792 * the gang tree, we won't modify the block, so we can safely defer the free
1793 * (knowing that the block is still intact). If we *can* assemble the gang
1794 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1795 * each constituent bp and we can allocate a new block on the next sync pass.
1797 * In all cases, the gang tree allows complete recovery from partial failure.
1798 * ==========================================================================
1802 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1807 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1808 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1809 &pio->io_bookmark));
1813 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1818 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1819 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1820 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1822 * As we rewrite each gang header, the pipeline will compute
1823 * a new gang block header checksum for it; but no one will
1824 * compute a new data checksum, so we do that here. The one
1825 * exception is the gang leader: the pipeline already computed
1826 * its data checksum because that stage precedes gang assembly.
1827 * (Presently, nothing actually uses interior data checksums;
1828 * this is just good hygiene.)
1830 if (gn != pio->io_gang_leader->io_gang_tree) {
1831 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1832 data, BP_GET_PSIZE(bp));
1835 * If we are here to damage data for testing purposes,
1836 * leave the GBH alone so that we can detect the damage.
1838 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1839 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1841 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1842 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1843 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1851 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1853 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1854 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp),
1855 ZIO_GANG_CHILD_FLAGS(pio)));
1860 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1862 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1863 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1866 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1875 static void zio_gang_tree_assemble_done(zio_t *zio);
1877 static zio_gang_node_t *
1878 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1880 zio_gang_node_t *gn;
1882 ASSERT(*gnpp == NULL);
1884 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1885 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1892 zio_gang_node_free(zio_gang_node_t **gnpp)
1894 zio_gang_node_t *gn = *gnpp;
1896 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1897 ASSERT(gn->gn_child[g] == NULL);
1899 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1900 kmem_free(gn, sizeof (*gn));
1905 zio_gang_tree_free(zio_gang_node_t **gnpp)
1907 zio_gang_node_t *gn = *gnpp;
1912 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1913 zio_gang_tree_free(&gn->gn_child[g]);
1915 zio_gang_node_free(gnpp);
1919 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1921 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1923 ASSERT(gio->io_gang_leader == gio);
1924 ASSERT(BP_IS_GANG(bp));
1926 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1927 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1928 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1932 zio_gang_tree_assemble_done(zio_t *zio)
1934 zio_t *gio = zio->io_gang_leader;
1935 zio_gang_node_t *gn = zio->io_private;
1936 blkptr_t *bp = zio->io_bp;
1938 ASSERT(gio == zio_unique_parent(zio));
1939 ASSERT(zio->io_child_count == 0);
1944 if (BP_SHOULD_BYTESWAP(bp))
1945 byteswap_uint64_array(zio->io_data, zio->io_size);
1947 ASSERT(zio->io_data == gn->gn_gbh);
1948 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1949 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1951 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1952 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1953 if (!BP_IS_GANG(gbp))
1955 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1960 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1962 zio_t *gio = pio->io_gang_leader;
1965 ASSERT(BP_IS_GANG(bp) == !!gn);
1966 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1967 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1970 * If you're a gang header, your data is in gn->gn_gbh.
1971 * If you're a gang member, your data is in 'data' and gn == NULL.
1973 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1976 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1978 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1979 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1980 if (BP_IS_HOLE(gbp))
1982 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1983 data = (char *)data + BP_GET_PSIZE(gbp);
1987 if (gn == gio->io_gang_tree && gio->io_data != NULL)
1988 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1995 zio_gang_assemble(zio_t *zio)
1997 blkptr_t *bp = zio->io_bp;
1999 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
2000 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2002 zio->io_gang_leader = zio;
2004 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
2006 return (ZIO_PIPELINE_CONTINUE);
2010 zio_gang_issue(zio_t *zio)
2012 blkptr_t *bp = zio->io_bp;
2014 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
2015 return (ZIO_PIPELINE_STOP);
2017 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
2018 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2020 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
2021 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
2023 zio_gang_tree_free(&zio->io_gang_tree);
2025 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2027 return (ZIO_PIPELINE_CONTINUE);
2031 zio_write_gang_member_ready(zio_t *zio)
2033 zio_t *pio = zio_unique_parent(zio);
2034 zio_t *gio = zio->io_gang_leader;
2035 dva_t *cdva = zio->io_bp->blk_dva;
2036 dva_t *pdva = pio->io_bp->blk_dva;
2039 if (BP_IS_HOLE(zio->io_bp))
2042 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
2044 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
2045 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
2046 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2047 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
2048 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
2050 mutex_enter(&pio->io_lock);
2051 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
2052 ASSERT(DVA_GET_GANG(&pdva[d]));
2053 asize = DVA_GET_ASIZE(&pdva[d]);
2054 asize += DVA_GET_ASIZE(&cdva[d]);
2055 DVA_SET_ASIZE(&pdva[d], asize);
2057 mutex_exit(&pio->io_lock);
2061 zio_write_gang_block(zio_t *pio)
2063 spa_t *spa = pio->io_spa;
2064 blkptr_t *bp = pio->io_bp;
2065 zio_t *gio = pio->io_gang_leader;
2067 zio_gang_node_t *gn, **gnpp;
2068 zio_gbh_phys_t *gbh;
2069 uint64_t txg = pio->io_txg;
2070 uint64_t resid = pio->io_size;
2072 int copies = gio->io_prop.zp_copies;
2073 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
2077 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
2078 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
2079 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
2081 pio->io_error = error;
2082 return (ZIO_PIPELINE_CONTINUE);
2086 gnpp = &gio->io_gang_tree;
2088 gnpp = pio->io_private;
2089 ASSERT(pio->io_ready == zio_write_gang_member_ready);
2092 gn = zio_gang_node_alloc(gnpp);
2094 bzero(gbh, SPA_GANGBLOCKSIZE);
2097 * Create the gang header.
2099 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
2100 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2103 * Create and nowait the gang children.
2105 for (int g = 0; resid != 0; resid -= lsize, g++) {
2106 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2108 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2110 zp.zp_checksum = gio->io_prop.zp_checksum;
2111 zp.zp_compress = ZIO_COMPRESS_OFF;
2112 zp.zp_type = DMU_OT_NONE;
2114 zp.zp_copies = gio->io_prop.zp_copies;
2115 zp.zp_dedup = B_FALSE;
2116 zp.zp_dedup_verify = B_FALSE;
2117 zp.zp_nopwrite = B_FALSE;
2119 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2120 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
2121 zio_write_gang_member_ready, NULL, NULL, &gn->gn_child[g],
2122 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2123 &pio->io_bookmark));
2127 * Set pio's pipeline to just wait for zio to finish.
2129 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2133 return (ZIO_PIPELINE_CONTINUE);
2137 * The zio_nop_write stage in the pipeline determines if allocating a
2138 * new bp is necessary. The nopwrite feature can handle writes in
2139 * either syncing or open context (i.e. zil writes) and as a result is
2140 * mutually exclusive with dedup.
2142 * By leveraging a cryptographically secure checksum, such as SHA256, we
2143 * can compare the checksums of the new data and the old to determine if
2144 * allocating a new block is required. Note that our requirements for
2145 * cryptographic strength are fairly weak: there can't be any accidental
2146 * hash collisions, but we don't need to be secure against intentional
2147 * (malicious) collisions. To trigger a nopwrite, you have to be able
2148 * to write the file to begin with, and triggering an incorrect (hash
2149 * collision) nopwrite is no worse than simply writing to the file.
2150 * That said, there are no known attacks against the checksum algorithms
2151 * used for nopwrite, assuming that the salt and the checksums
2152 * themselves remain secret.
2155 zio_nop_write(zio_t *zio)
2157 blkptr_t *bp = zio->io_bp;
2158 blkptr_t *bp_orig = &zio->io_bp_orig;
2159 zio_prop_t *zp = &zio->io_prop;
2161 ASSERT(BP_GET_LEVEL(bp) == 0);
2162 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2163 ASSERT(zp->zp_nopwrite);
2164 ASSERT(!zp->zp_dedup);
2165 ASSERT(zio->io_bp_override == NULL);
2166 ASSERT(IO_IS_ALLOCATING(zio));
2169 * Check to see if the original bp and the new bp have matching
2170 * characteristics (i.e. same checksum, compression algorithms, etc).
2171 * If they don't then just continue with the pipeline which will
2172 * allocate a new bp.
2174 if (BP_IS_HOLE(bp_orig) ||
2175 !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
2176 ZCHECKSUM_FLAG_NOPWRITE) ||
2177 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2178 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2179 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2180 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2181 return (ZIO_PIPELINE_CONTINUE);
2184 * If the checksums match then reset the pipeline so that we
2185 * avoid allocating a new bp and issuing any I/O.
2187 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2188 ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
2189 ZCHECKSUM_FLAG_NOPWRITE);
2190 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2191 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2192 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2193 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2194 sizeof (uint64_t)) == 0);
2197 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2198 zio->io_flags |= ZIO_FLAG_NOPWRITE;
2201 return (ZIO_PIPELINE_CONTINUE);
2205 * ==========================================================================
2207 * ==========================================================================
2210 zio_ddt_child_read_done(zio_t *zio)
2212 blkptr_t *bp = zio->io_bp;
2213 ddt_entry_t *dde = zio->io_private;
2215 zio_t *pio = zio_unique_parent(zio);
2217 mutex_enter(&pio->io_lock);
2218 ddp = ddt_phys_select(dde, bp);
2219 if (zio->io_error == 0)
2220 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2221 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
2222 dde->dde_repair_data = zio->io_data;
2224 zio_buf_free(zio->io_data, zio->io_size);
2225 mutex_exit(&pio->io_lock);
2229 zio_ddt_read_start(zio_t *zio)
2231 blkptr_t *bp = zio->io_bp;
2233 ASSERT(BP_GET_DEDUP(bp));
2234 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2235 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2237 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2238 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2239 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2240 ddt_phys_t *ddp = dde->dde_phys;
2241 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2244 ASSERT(zio->io_vsd == NULL);
2247 if (ddp_self == NULL)
2248 return (ZIO_PIPELINE_CONTINUE);
2250 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2251 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2253 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2255 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2256 zio_buf_alloc(zio->io_size), zio->io_size,
2257 zio_ddt_child_read_done, dde, zio->io_priority,
2258 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
2259 &zio->io_bookmark));
2261 return (ZIO_PIPELINE_CONTINUE);
2264 zio_nowait(zio_read(zio, zio->io_spa, bp,
2265 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
2266 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2268 return (ZIO_PIPELINE_CONTINUE);
2272 zio_ddt_read_done(zio_t *zio)
2274 blkptr_t *bp = zio->io_bp;
2276 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2277 return (ZIO_PIPELINE_STOP);
2279 ASSERT(BP_GET_DEDUP(bp));
2280 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2281 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2283 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2284 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2285 ddt_entry_t *dde = zio->io_vsd;
2287 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2288 return (ZIO_PIPELINE_CONTINUE);
2291 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2292 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2293 return (ZIO_PIPELINE_STOP);
2295 if (dde->dde_repair_data != NULL) {
2296 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2297 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2299 ddt_repair_done(ddt, dde);
2303 ASSERT(zio->io_vsd == NULL);
2305 return (ZIO_PIPELINE_CONTINUE);
2309 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2311 spa_t *spa = zio->io_spa;
2314 * Note: we compare the original data, not the transformed data,
2315 * because when zio->io_bp is an override bp, we will not have
2316 * pushed the I/O transforms. That's an important optimization
2317 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2319 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2320 zio_t *lio = dde->dde_lead_zio[p];
2323 return (lio->io_orig_size != zio->io_orig_size ||
2324 bcmp(zio->io_orig_data, lio->io_orig_data,
2325 zio->io_orig_size) != 0);
2329 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2330 ddt_phys_t *ddp = &dde->dde_phys[p];
2332 if (ddp->ddp_phys_birth != 0) {
2333 arc_buf_t *abuf = NULL;
2334 arc_flags_t aflags = ARC_FLAG_WAIT;
2335 blkptr_t blk = *zio->io_bp;
2338 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2342 error = arc_read(NULL, spa, &blk,
2343 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2344 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2345 &aflags, &zio->io_bookmark);
2348 if (arc_buf_size(abuf) != zio->io_orig_size ||
2349 bcmp(abuf->b_data, zio->io_orig_data,
2350 zio->io_orig_size) != 0)
2351 error = SET_ERROR(EEXIST);
2352 VERIFY(arc_buf_remove_ref(abuf, &abuf));
2356 return (error != 0);
2364 zio_ddt_child_write_ready(zio_t *zio)
2366 int p = zio->io_prop.zp_copies;
2367 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2368 ddt_entry_t *dde = zio->io_private;
2369 ddt_phys_t *ddp = &dde->dde_phys[p];
2377 ASSERT(dde->dde_lead_zio[p] == zio);
2379 ddt_phys_fill(ddp, zio->io_bp);
2381 while ((pio = zio_walk_parents(zio)) != NULL)
2382 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2388 zio_ddt_child_write_done(zio_t *zio)
2390 int p = zio->io_prop.zp_copies;
2391 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2392 ddt_entry_t *dde = zio->io_private;
2393 ddt_phys_t *ddp = &dde->dde_phys[p];
2397 ASSERT(ddp->ddp_refcnt == 0);
2398 ASSERT(dde->dde_lead_zio[p] == zio);
2399 dde->dde_lead_zio[p] = NULL;
2401 if (zio->io_error == 0) {
2402 while (zio_walk_parents(zio) != NULL)
2403 ddt_phys_addref(ddp);
2405 ddt_phys_clear(ddp);
2412 zio_ddt_ditto_write_done(zio_t *zio)
2414 int p = DDT_PHYS_DITTO;
2415 zio_prop_t *zp = &zio->io_prop;
2416 blkptr_t *bp = zio->io_bp;
2417 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2418 ddt_entry_t *dde = zio->io_private;
2419 ddt_phys_t *ddp = &dde->dde_phys[p];
2420 ddt_key_t *ddk = &dde->dde_key;
2424 ASSERT(ddp->ddp_refcnt == 0);
2425 ASSERT(dde->dde_lead_zio[p] == zio);
2426 dde->dde_lead_zio[p] = NULL;
2428 if (zio->io_error == 0) {
2429 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2430 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2431 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2432 if (ddp->ddp_phys_birth != 0)
2433 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2434 ddt_phys_fill(ddp, bp);
2441 zio_ddt_write(zio_t *zio)
2443 spa_t *spa = zio->io_spa;
2444 blkptr_t *bp = zio->io_bp;
2445 uint64_t txg = zio->io_txg;
2446 zio_prop_t *zp = &zio->io_prop;
2447 int p = zp->zp_copies;
2451 ddt_t *ddt = ddt_select(spa, bp);
2455 ASSERT(BP_GET_DEDUP(bp));
2456 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2457 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2460 dde = ddt_lookup(ddt, bp, B_TRUE);
2461 ddp = &dde->dde_phys[p];
2463 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2465 * If we're using a weak checksum, upgrade to a strong checksum
2466 * and try again. If we're already using a strong checksum,
2467 * we can't resolve it, so just convert to an ordinary write.
2468 * (And automatically e-mail a paper to Nature?)
2470 if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
2471 ZCHECKSUM_FLAG_DEDUP)) {
2472 zp->zp_checksum = spa_dedup_checksum(spa);
2473 zio_pop_transforms(zio);
2474 zio->io_stage = ZIO_STAGE_OPEN;
2477 zp->zp_dedup = B_FALSE;
2479 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2481 return (ZIO_PIPELINE_CONTINUE);
2484 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2485 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2487 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2488 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2489 zio_prop_t czp = *zp;
2491 czp.zp_copies = ditto_copies;
2494 * If we arrived here with an override bp, we won't have run
2495 * the transform stack, so we won't have the data we need to
2496 * generate a child i/o. So, toss the override bp and restart.
2497 * This is safe, because using the override bp is just an
2498 * optimization; and it's rare, so the cost doesn't matter.
2500 if (zio->io_bp_override) {
2501 zio_pop_transforms(zio);
2502 zio->io_stage = ZIO_STAGE_OPEN;
2503 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2504 zio->io_bp_override = NULL;
2507 return (ZIO_PIPELINE_CONTINUE);
2510 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2511 zio->io_orig_size, &czp, NULL, NULL,
2512 zio_ddt_ditto_write_done, dde, zio->io_priority,
2513 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2515 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2516 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2519 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2520 if (ddp->ddp_phys_birth != 0)
2521 ddt_bp_fill(ddp, bp, txg);
2522 if (dde->dde_lead_zio[p] != NULL)
2523 zio_add_child(zio, dde->dde_lead_zio[p]);
2525 ddt_phys_addref(ddp);
2526 } else if (zio->io_bp_override) {
2527 ASSERT(bp->blk_birth == txg);
2528 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2529 ddt_phys_fill(ddp, bp);
2530 ddt_phys_addref(ddp);
2532 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2533 zio->io_orig_size, zp, zio_ddt_child_write_ready, NULL,
2534 zio_ddt_child_write_done, dde, zio->io_priority,
2535 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2537 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2538 dde->dde_lead_zio[p] = cio;
2548 return (ZIO_PIPELINE_CONTINUE);
2551 ddt_entry_t *freedde; /* for debugging */
2554 zio_ddt_free(zio_t *zio)
2556 spa_t *spa = zio->io_spa;
2557 blkptr_t *bp = zio->io_bp;
2558 ddt_t *ddt = ddt_select(spa, bp);
2562 ASSERT(BP_GET_DEDUP(bp));
2563 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2566 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2567 ddp = ddt_phys_select(dde, bp);
2568 ddt_phys_decref(ddp);
2571 return (ZIO_PIPELINE_CONTINUE);
2575 * ==========================================================================
2576 * Allocate and free blocks
2577 * ==========================================================================
2580 zio_dva_allocate(zio_t *zio)
2582 spa_t *spa = zio->io_spa;
2583 metaslab_class_t *mc = spa_normal_class(spa);
2584 blkptr_t *bp = zio->io_bp;
2588 if (zio->io_gang_leader == NULL) {
2589 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2590 zio->io_gang_leader = zio;
2593 ASSERT(BP_IS_HOLE(bp));
2594 ASSERT0(BP_GET_NDVAS(bp));
2595 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2596 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2597 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2600 * The dump device does not support gang blocks so allocation on
2601 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2602 * the "fast" gang feature.
2604 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2605 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2606 METASLAB_GANG_CHILD : 0;
2607 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2608 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2611 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2612 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2614 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2615 return (zio_write_gang_block(zio));
2616 zio->io_error = error;
2619 return (ZIO_PIPELINE_CONTINUE);
2623 zio_dva_free(zio_t *zio)
2625 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2627 return (ZIO_PIPELINE_CONTINUE);
2631 zio_dva_claim(zio_t *zio)
2635 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2637 zio->io_error = error;
2639 return (ZIO_PIPELINE_CONTINUE);
2643 * Undo an allocation. This is used by zio_done() when an I/O fails
2644 * and we want to give back the block we just allocated.
2645 * This handles both normal blocks and gang blocks.
2648 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2650 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2651 ASSERT(zio->io_bp_override == NULL);
2653 if (!BP_IS_HOLE(bp))
2654 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2657 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2658 zio_dva_unallocate(zio, gn->gn_child[g],
2659 &gn->gn_gbh->zg_blkptr[g]);
2665 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2668 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2669 uint64_t size, boolean_t use_slog)
2673 ASSERT(txg > spa_syncing_txg(spa));
2676 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2677 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2678 * when allocating them.
2681 error = metaslab_alloc(spa, spa_log_class(spa), size,
2682 new_bp, 1, txg, old_bp,
2683 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2687 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2688 new_bp, 1, txg, old_bp,
2689 METASLAB_HINTBP_AVOID);
2693 BP_SET_LSIZE(new_bp, size);
2694 BP_SET_PSIZE(new_bp, size);
2695 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2696 BP_SET_CHECKSUM(new_bp,
2697 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2698 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2699 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2700 BP_SET_LEVEL(new_bp, 0);
2701 BP_SET_DEDUP(new_bp, 0);
2702 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2709 * Free an intent log block.
2712 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2714 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2715 ASSERT(!BP_IS_GANG(bp));
2717 zio_free(spa, txg, bp);
2721 * ==========================================================================
2722 * Read, write and delete to physical devices
2723 * ==========================================================================
2728 * Issue an I/O to the underlying vdev. Typically the issue pipeline
2729 * stops after this stage and will resume upon I/O completion.
2730 * However, there are instances where the vdev layer may need to
2731 * continue the pipeline when an I/O was not issued. Since the I/O
2732 * that was sent to the vdev layer might be different than the one
2733 * currently active in the pipeline (see vdev_queue_io()), we explicitly
2734 * force the underlying vdev layers to call either zio_execute() or
2735 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
2738 zio_vdev_io_start(zio_t *zio)
2740 vdev_t *vd = zio->io_vd;
2742 spa_t *spa = zio->io_spa;
2745 ASSERT(zio->io_error == 0);
2746 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2749 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2750 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2753 * The mirror_ops handle multiple DVAs in a single BP.
2755 vdev_mirror_ops.vdev_op_io_start(zio);
2756 return (ZIO_PIPELINE_STOP);
2759 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE &&
2760 zio->io_priority == ZIO_PRIORITY_NOW) {
2761 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg);
2762 return (ZIO_PIPELINE_CONTINUE);
2766 * We keep track of time-sensitive I/Os so that the scan thread
2767 * can quickly react to certain workloads. In particular, we care
2768 * about non-scrubbing, top-level reads and writes with the following
2770 * - synchronous writes of user data to non-slog devices
2771 * - any reads of user data
2772 * When these conditions are met, adjust the timestamp of spa_last_io
2773 * which allows the scan thread to adjust its workload accordingly.
2775 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2776 vd == vd->vdev_top && !vd->vdev_islog &&
2777 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2778 zio->io_txg != spa_syncing_txg(spa)) {
2779 uint64_t old = spa->spa_last_io;
2780 uint64_t new = ddi_get_lbolt64();
2782 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2785 align = 1ULL << vd->vdev_top->vdev_ashift;
2787 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
2788 P2PHASE(zio->io_size, align) != 0) {
2789 /* Transform logical writes to be a full physical block size. */
2790 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2792 if (zio->io_type == ZIO_TYPE_READ ||
2793 zio->io_type == ZIO_TYPE_WRITE)
2794 abuf = zio_buf_alloc(asize);
2795 ASSERT(vd == vd->vdev_top);
2796 if (zio->io_type == ZIO_TYPE_WRITE) {
2797 bcopy(zio->io_data, abuf, zio->io_size);
2798 bzero(abuf + zio->io_size, asize - zio->io_size);
2800 zio_push_transform(zio, abuf, asize, abuf ? asize : 0,
2805 * If this is not a physical io, make sure that it is properly aligned
2806 * before proceeding.
2808 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
2809 ASSERT0(P2PHASE(zio->io_offset, align));
2810 ASSERT0(P2PHASE(zio->io_size, align));
2813 * For physical writes, we allow 512b aligned writes and assume
2814 * the device will perform a read-modify-write as necessary.
2816 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
2817 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
2820 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa));
2823 * If this is a repair I/O, and there's no self-healing involved --
2824 * that is, we're just resilvering what we expect to resilver --
2825 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2826 * This prevents spurious resilvering with nested replication.
2827 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2828 * A is out of date, we'll read from C+D, then use the data to
2829 * resilver A+B -- but we don't actually want to resilver B, just A.
2830 * The top-level mirror has no way to know this, so instead we just
2831 * discard unnecessary repairs as we work our way down the vdev tree.
2832 * The same logic applies to any form of nested replication:
2833 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2835 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2836 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2837 zio->io_txg != 0 && /* not a delegated i/o */
2838 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2839 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2840 zio_vdev_io_bypass(zio);
2841 return (ZIO_PIPELINE_CONTINUE);
2844 if (vd->vdev_ops->vdev_op_leaf) {
2845 switch (zio->io_type) {
2847 if (vdev_cache_read(zio))
2848 return (ZIO_PIPELINE_CONTINUE);
2850 case ZIO_TYPE_WRITE:
2852 if ((zio = vdev_queue_io(zio)) == NULL)
2853 return (ZIO_PIPELINE_STOP);
2855 if (!vdev_accessible(vd, zio)) {
2856 zio->io_error = SET_ERROR(ENXIO);
2858 return (ZIO_PIPELINE_STOP);
2863 * Note that we ignore repair writes for TRIM because they can
2864 * conflict with normal writes. This isn't an issue because, by
2865 * definition, we only repair blocks that aren't freed.
2867 if (zio->io_type == ZIO_TYPE_WRITE &&
2868 !(zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2869 !trim_map_write_start(zio))
2870 return (ZIO_PIPELINE_STOP);
2873 vd->vdev_ops->vdev_op_io_start(zio);
2874 return (ZIO_PIPELINE_STOP);
2878 zio_vdev_io_done(zio_t *zio)
2880 vdev_t *vd = zio->io_vd;
2881 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2882 boolean_t unexpected_error = B_FALSE;
2884 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2885 return (ZIO_PIPELINE_STOP);
2887 ASSERT(zio->io_type == ZIO_TYPE_READ ||
2888 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE);
2890 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2891 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE ||
2892 zio->io_type == ZIO_TYPE_FREE)) {
2894 if (zio->io_type == ZIO_TYPE_WRITE &&
2895 !(zio->io_flags & ZIO_FLAG_IO_REPAIR))
2896 trim_map_write_done(zio);
2898 vdev_queue_io_done(zio);
2900 if (zio->io_type == ZIO_TYPE_WRITE)
2901 vdev_cache_write(zio);
2903 if (zio_injection_enabled && zio->io_error == 0)
2904 zio->io_error = zio_handle_device_injection(vd,
2907 if (zio_injection_enabled && zio->io_error == 0)
2908 zio->io_error = zio_handle_label_injection(zio, EIO);
2910 if (zio->io_error) {
2911 if (zio->io_error == ENOTSUP &&
2912 zio->io_type == ZIO_TYPE_FREE) {
2913 /* Not all devices support TRIM. */
2914 } else if (!vdev_accessible(vd, zio)) {
2915 zio->io_error = SET_ERROR(ENXIO);
2917 unexpected_error = B_TRUE;
2922 ops->vdev_op_io_done(zio);
2924 if (unexpected_error)
2925 VERIFY(vdev_probe(vd, zio) == NULL);
2927 return (ZIO_PIPELINE_CONTINUE);
2931 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2932 * disk, and use that to finish the checksum ereport later.
2935 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2936 const void *good_buf)
2938 /* no processing needed */
2939 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2944 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2946 void *buf = zio_buf_alloc(zio->io_size);
2948 bcopy(zio->io_data, buf, zio->io_size);
2950 zcr->zcr_cbinfo = zio->io_size;
2951 zcr->zcr_cbdata = buf;
2952 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2953 zcr->zcr_free = zio_buf_free;
2957 zio_vdev_io_assess(zio_t *zio)
2959 vdev_t *vd = zio->io_vd;
2961 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2962 return (ZIO_PIPELINE_STOP);
2964 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2965 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2967 if (zio->io_vsd != NULL) {
2968 zio->io_vsd_ops->vsd_free(zio);
2972 if (zio_injection_enabled && zio->io_error == 0)
2973 zio->io_error = zio_handle_fault_injection(zio, EIO);
2975 if (zio->io_type == ZIO_TYPE_FREE &&
2976 zio->io_priority != ZIO_PRIORITY_NOW) {
2977 switch (zio->io_error) {
2979 ZIO_TRIM_STAT_INCR(bytes, zio->io_size);
2980 ZIO_TRIM_STAT_BUMP(success);
2983 ZIO_TRIM_STAT_BUMP(unsupported);
2986 ZIO_TRIM_STAT_BUMP(failed);
2992 * If the I/O failed, determine whether we should attempt to retry it.
2994 * On retry, we cut in line in the issue queue, since we don't want
2995 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2997 if (zio->io_error && vd == NULL &&
2998 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2999 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
3000 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
3002 zio->io_flags |= ZIO_FLAG_IO_RETRY |
3003 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
3004 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
3005 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
3006 zio_requeue_io_start_cut_in_line);
3007 return (ZIO_PIPELINE_STOP);
3011 * If we got an error on a leaf device, convert it to ENXIO
3012 * if the device is not accessible at all.
3014 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3015 !vdev_accessible(vd, zio))
3016 zio->io_error = SET_ERROR(ENXIO);
3019 * If we can't write to an interior vdev (mirror or RAID-Z),
3020 * set vdev_cant_write so that we stop trying to allocate from it.
3022 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
3023 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
3024 vd->vdev_cant_write = B_TRUE;
3028 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3030 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3031 zio->io_physdone != NULL) {
3032 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
3033 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
3034 zio->io_physdone(zio->io_logical);
3037 return (ZIO_PIPELINE_CONTINUE);
3041 zio_vdev_io_reissue(zio_t *zio)
3043 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3044 ASSERT(zio->io_error == 0);
3046 zio->io_stage >>= 1;
3050 zio_vdev_io_redone(zio_t *zio)
3052 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
3054 zio->io_stage >>= 1;
3058 zio_vdev_io_bypass(zio_t *zio)
3060 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3061 ASSERT(zio->io_error == 0);
3063 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
3064 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
3068 * ==========================================================================
3069 * Generate and verify checksums
3070 * ==========================================================================
3073 zio_checksum_generate(zio_t *zio)
3075 blkptr_t *bp = zio->io_bp;
3076 enum zio_checksum checksum;
3080 * This is zio_write_phys().
3081 * We're either generating a label checksum, or none at all.
3083 checksum = zio->io_prop.zp_checksum;
3085 if (checksum == ZIO_CHECKSUM_OFF)
3086 return (ZIO_PIPELINE_CONTINUE);
3088 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
3090 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
3091 ASSERT(!IO_IS_ALLOCATING(zio));
3092 checksum = ZIO_CHECKSUM_GANG_HEADER;
3094 checksum = BP_GET_CHECKSUM(bp);
3098 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
3100 return (ZIO_PIPELINE_CONTINUE);
3104 zio_checksum_verify(zio_t *zio)
3106 zio_bad_cksum_t info;
3107 blkptr_t *bp = zio->io_bp;
3110 ASSERT(zio->io_vd != NULL);
3114 * This is zio_read_phys().
3115 * We're either verifying a label checksum, or nothing at all.
3117 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
3118 return (ZIO_PIPELINE_CONTINUE);
3120 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
3123 if ((error = zio_checksum_error(zio, &info)) != 0) {
3124 zio->io_error = error;
3125 if (error == ECKSUM &&
3126 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
3127 zfs_ereport_start_checksum(zio->io_spa,
3128 zio->io_vd, zio, zio->io_offset,
3129 zio->io_size, NULL, &info);
3133 return (ZIO_PIPELINE_CONTINUE);
3137 * Called by RAID-Z to ensure we don't compute the checksum twice.
3140 zio_checksum_verified(zio_t *zio)
3142 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
3146 * ==========================================================================
3147 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
3148 * An error of 0 indicates success. ENXIO indicates whole-device failure,
3149 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
3150 * indicate errors that are specific to one I/O, and most likely permanent.
3151 * Any other error is presumed to be worse because we weren't expecting it.
3152 * ==========================================================================
3155 zio_worst_error(int e1, int e2)
3157 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
3160 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
3161 if (e1 == zio_error_rank[r1])
3164 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
3165 if (e2 == zio_error_rank[r2])
3168 return (r1 > r2 ? e1 : e2);
3172 * ==========================================================================
3174 * ==========================================================================
3177 zio_ready(zio_t *zio)
3179 blkptr_t *bp = zio->io_bp;
3180 zio_t *pio, *pio_next;
3182 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
3183 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
3184 return (ZIO_PIPELINE_STOP);
3186 if (zio->io_ready) {
3187 ASSERT(IO_IS_ALLOCATING(zio));
3188 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
3189 (zio->io_flags & ZIO_FLAG_NOPWRITE));
3190 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
3195 if (bp != NULL && bp != &zio->io_bp_copy)
3196 zio->io_bp_copy = *bp;
3199 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3201 mutex_enter(&zio->io_lock);
3202 zio->io_state[ZIO_WAIT_READY] = 1;
3203 pio = zio_walk_parents(zio);
3204 mutex_exit(&zio->io_lock);
3207 * As we notify zio's parents, new parents could be added.
3208 * New parents go to the head of zio's io_parent_list, however,
3209 * so we will (correctly) not notify them. The remainder of zio's
3210 * io_parent_list, from 'pio_next' onward, cannot change because
3211 * all parents must wait for us to be done before they can be done.
3213 for (; pio != NULL; pio = pio_next) {
3214 pio_next = zio_walk_parents(zio);
3215 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
3218 if (zio->io_flags & ZIO_FLAG_NODATA) {
3219 if (BP_IS_GANG(bp)) {
3220 zio->io_flags &= ~ZIO_FLAG_NODATA;
3222 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
3223 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
3227 if (zio_injection_enabled &&
3228 zio->io_spa->spa_syncing_txg == zio->io_txg)
3229 zio_handle_ignored_writes(zio);
3231 return (ZIO_PIPELINE_CONTINUE);
3235 zio_done(zio_t *zio)
3237 spa_t *spa = zio->io_spa;
3238 zio_t *lio = zio->io_logical;
3239 blkptr_t *bp = zio->io_bp;
3240 vdev_t *vd = zio->io_vd;
3241 uint64_t psize = zio->io_size;
3242 zio_t *pio, *pio_next;
3245 * If our children haven't all completed,
3246 * wait for them and then repeat this pipeline stage.
3248 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
3249 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
3250 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
3251 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
3252 return (ZIO_PIPELINE_STOP);
3254 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3255 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3256 ASSERT(zio->io_children[c][w] == 0);
3258 if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
3259 ASSERT(bp->blk_pad[0] == 0);
3260 ASSERT(bp->blk_pad[1] == 0);
3261 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3262 (bp == zio_unique_parent(zio)->io_bp));
3263 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3264 zio->io_bp_override == NULL &&
3265 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3266 ASSERT(!BP_SHOULD_BYTESWAP(bp));
3267 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3268 ASSERT(BP_COUNT_GANG(bp) == 0 ||
3269 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3271 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3272 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3276 * If there were child vdev/gang/ddt errors, they apply to us now.
3278 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3279 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3280 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3283 * If the I/O on the transformed data was successful, generate any
3284 * checksum reports now while we still have the transformed data.
3286 if (zio->io_error == 0) {
3287 while (zio->io_cksum_report != NULL) {
3288 zio_cksum_report_t *zcr = zio->io_cksum_report;
3289 uint64_t align = zcr->zcr_align;
3290 uint64_t asize = P2ROUNDUP(psize, align);
3291 char *abuf = zio->io_data;
3293 if (asize != psize) {
3294 abuf = zio_buf_alloc(asize);
3295 bcopy(zio->io_data, abuf, psize);
3296 bzero(abuf + psize, asize - psize);
3299 zio->io_cksum_report = zcr->zcr_next;
3300 zcr->zcr_next = NULL;
3301 zcr->zcr_finish(zcr, abuf);
3302 zfs_ereport_free_checksum(zcr);
3305 zio_buf_free(abuf, asize);
3309 zio_pop_transforms(zio); /* note: may set zio->io_error */
3311 vdev_stat_update(zio, psize);
3313 if (zio->io_error) {
3315 * If this I/O is attached to a particular vdev,
3316 * generate an error message describing the I/O failure
3317 * at the block level. We ignore these errors if the
3318 * device is currently unavailable.
3320 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3321 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3323 if ((zio->io_error == EIO || !(zio->io_flags &
3324 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3327 * For logical I/O requests, tell the SPA to log the
3328 * error and generate a logical data ereport.
3330 spa_log_error(spa, zio);
3331 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3336 if (zio->io_error && zio == lio) {
3338 * Determine whether zio should be reexecuted. This will
3339 * propagate all the way to the root via zio_notify_parent().
3341 ASSERT(vd == NULL && bp != NULL);
3342 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3344 if (IO_IS_ALLOCATING(zio) &&
3345 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3346 if (zio->io_error != ENOSPC)
3347 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3349 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3352 if ((zio->io_type == ZIO_TYPE_READ ||
3353 zio->io_type == ZIO_TYPE_FREE) &&
3354 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3355 zio->io_error == ENXIO &&
3356 spa_load_state(spa) == SPA_LOAD_NONE &&
3357 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3358 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3360 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3361 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3364 * Here is a possibly good place to attempt to do
3365 * either combinatorial reconstruction or error correction
3366 * based on checksums. It also might be a good place
3367 * to send out preliminary ereports before we suspend
3373 * If there were logical child errors, they apply to us now.
3374 * We defer this until now to avoid conflating logical child
3375 * errors with errors that happened to the zio itself when
3376 * updating vdev stats and reporting FMA events above.
3378 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3380 if ((zio->io_error || zio->io_reexecute) &&
3381 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3382 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3383 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3385 zio_gang_tree_free(&zio->io_gang_tree);
3388 * Godfather I/Os should never suspend.
3390 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3391 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3392 zio->io_reexecute = 0;
3394 if (zio->io_reexecute) {
3396 * This is a logical I/O that wants to reexecute.
3398 * Reexecute is top-down. When an i/o fails, if it's not
3399 * the root, it simply notifies its parent and sticks around.
3400 * The parent, seeing that it still has children in zio_done(),
3401 * does the same. This percolates all the way up to the root.
3402 * The root i/o will reexecute or suspend the entire tree.
3404 * This approach ensures that zio_reexecute() honors
3405 * all the original i/o dependency relationships, e.g.
3406 * parents not executing until children are ready.
3408 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3410 zio->io_gang_leader = NULL;
3412 mutex_enter(&zio->io_lock);
3413 zio->io_state[ZIO_WAIT_DONE] = 1;
3414 mutex_exit(&zio->io_lock);
3417 * "The Godfather" I/O monitors its children but is
3418 * not a true parent to them. It will track them through
3419 * the pipeline but severs its ties whenever they get into
3420 * trouble (e.g. suspended). This allows "The Godfather"
3421 * I/O to return status without blocking.
3423 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3424 zio_link_t *zl = zio->io_walk_link;
3425 pio_next = zio_walk_parents(zio);
3427 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3428 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3429 zio_remove_child(pio, zio, zl);
3430 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3434 if ((pio = zio_unique_parent(zio)) != NULL) {
3436 * We're not a root i/o, so there's nothing to do
3437 * but notify our parent. Don't propagate errors
3438 * upward since we haven't permanently failed yet.
3440 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3441 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3442 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3443 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3445 * We'd fail again if we reexecuted now, so suspend
3446 * until conditions improve (e.g. device comes online).
3448 zio_suspend(spa, zio);
3451 * Reexecution is potentially a huge amount of work.
3452 * Hand it off to the otherwise-unused claim taskq.
3454 #if defined(illumos) || !defined(_KERNEL)
3455 ASSERT(zio->io_tqent.tqent_next == NULL);
3457 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
3459 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3460 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3463 return (ZIO_PIPELINE_STOP);
3466 ASSERT(zio->io_child_count == 0);
3467 ASSERT(zio->io_reexecute == 0);
3468 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3471 * Report any checksum errors, since the I/O is complete.
3473 while (zio->io_cksum_report != NULL) {
3474 zio_cksum_report_t *zcr = zio->io_cksum_report;
3475 zio->io_cksum_report = zcr->zcr_next;
3476 zcr->zcr_next = NULL;
3477 zcr->zcr_finish(zcr, NULL);
3478 zfs_ereport_free_checksum(zcr);
3482 * It is the responsibility of the done callback to ensure that this
3483 * particular zio is no longer discoverable for adoption, and as
3484 * such, cannot acquire any new parents.
3489 mutex_enter(&zio->io_lock);
3490 zio->io_state[ZIO_WAIT_DONE] = 1;
3491 mutex_exit(&zio->io_lock);
3493 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3494 zio_link_t *zl = zio->io_walk_link;
3495 pio_next = zio_walk_parents(zio);
3496 zio_remove_child(pio, zio, zl);
3497 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3500 if (zio->io_waiter != NULL) {
3501 mutex_enter(&zio->io_lock);
3502 zio->io_executor = NULL;
3503 cv_broadcast(&zio->io_cv);
3504 mutex_exit(&zio->io_lock);
3509 return (ZIO_PIPELINE_STOP);
3513 * ==========================================================================
3514 * I/O pipeline definition
3515 * ==========================================================================
3517 static zio_pipe_stage_t *zio_pipeline[] = {
3523 zio_checksum_generate,
3538 zio_checksum_verify,
3546 * Compare two zbookmark_phys_t's to see which we would reach first in a
3547 * pre-order traversal of the object tree.
3549 * This is simple in every case aside from the meta-dnode object. For all other
3550 * objects, we traverse them in order (object 1 before object 2, and so on).
3551 * However, all of these objects are traversed while traversing object 0, since
3552 * the data it points to is the list of objects. Thus, we need to convert to a
3553 * canonical representation so we can compare meta-dnode bookmarks to
3554 * non-meta-dnode bookmarks.
3556 * We do this by calculating "equivalents" for each field of the zbookmark.
3557 * zbookmarks outside of the meta-dnode use their own object and level, and
3558 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
3559 * blocks this bookmark refers to) by multiplying their blkid by their span
3560 * (the number of L0 blocks contained within one block at their level).
3561 * zbookmarks inside the meta-dnode calculate their object equivalent
3562 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
3563 * level + 1<<31 (any value larger than a level could ever be) for their level.
3564 * This causes them to always compare before a bookmark in their object
3565 * equivalent, compare appropriately to bookmarks in other objects, and to
3566 * compare appropriately to other bookmarks in the meta-dnode.
3569 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
3570 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
3573 * These variables represent the "equivalent" values for the zbookmark,
3574 * after converting zbookmarks inside the meta dnode to their
3575 * normal-object equivalents.
3577 uint64_t zb1obj, zb2obj;
3578 uint64_t zb1L0, zb2L0;
3579 uint64_t zb1level, zb2level;
3581 if (zb1->zb_object == zb2->zb_object &&
3582 zb1->zb_level == zb2->zb_level &&
3583 zb1->zb_blkid == zb2->zb_blkid)
3587 * BP_SPANB calculates the span in blocks.
3589 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
3590 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
3592 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3593 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
3595 zb1level = zb1->zb_level + COMPARE_META_LEVEL;
3597 zb1obj = zb1->zb_object;
3598 zb1level = zb1->zb_level;
3601 if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
3602 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
3604 zb2level = zb2->zb_level + COMPARE_META_LEVEL;
3606 zb2obj = zb2->zb_object;
3607 zb2level = zb2->zb_level;
3610 /* Now that we have a canonical representation, do the comparison. */
3611 if (zb1obj != zb2obj)
3612 return (zb1obj < zb2obj ? -1 : 1);
3613 else if (zb1L0 != zb2L0)
3614 return (zb1L0 < zb2L0 ? -1 : 1);
3615 else if (zb1level != zb2level)
3616 return (zb1level > zb2level ? -1 : 1);
3618 * This can (theoretically) happen if the bookmarks have the same object
3619 * and level, but different blkids, if the block sizes are not the same.
3620 * There is presently no way to change the indirect block sizes
3626 * This function checks the following: given that last_block is the place that
3627 * our traversal stopped last time, does that guarantee that we've visited
3628 * every node under subtree_root? Therefore, we can't just use the raw output
3629 * of zbookmark_compare. We have to pass in a modified version of
3630 * subtree_root; by incrementing the block id, and then checking whether
3631 * last_block is before or equal to that, we can tell whether or not having
3632 * visited last_block implies that all of subtree_root's children have been
3636 zbookmark_subtree_completed(const dnode_phys_t *dnp,
3637 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
3639 zbookmark_phys_t mod_zb = *subtree_root;
3641 ASSERT(last_block->zb_level == 0);
3643 /* The objset_phys_t isn't before anything. */
3648 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
3649 * data block size in sectors, because that variable is only used if
3650 * the bookmark refers to a block in the meta-dnode. Since we don't
3651 * know without examining it what object it refers to, and there's no
3652 * harm in passing in this value in other cases, we always pass it in.
3654 * We pass in 0 for the indirect block size shift because zb2 must be
3655 * level 0. The indirect block size is only used to calculate the span
3656 * of the bookmark, but since the bookmark must be level 0, the span is
3657 * always 1, so the math works out.
3659 * If you make changes to how the zbookmark_compare code works, be sure
3660 * to make sure that this code still works afterwards.
3662 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
3663 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,