4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2016 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
25 * Copyright (c) 2014 Integros [integros.com]
28 #include <sys/sysmacros.h>
29 #include <sys/zfs_context.h>
30 #include <sys/fm/fs/zfs.h>
33 #include <sys/spa_impl.h>
34 #include <sys/vdev_impl.h>
35 #include <sys/zio_impl.h>
36 #include <sys/zio_compress.h>
37 #include <sys/zio_checksum.h>
38 #include <sys/dmu_objset.h>
41 #include <sys/trim_map.h>
42 #include <sys/blkptr.h>
43 #include <sys/zfeature.h>
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 *children_ready,
775 zio_done_func_t *physdone, zio_done_func_t *done,
776 void *private, zio_priority_t priority, enum zio_flag flags,
777 const zbookmark_phys_t *zb)
781 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
782 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
783 zp->zp_compress >= ZIO_COMPRESS_OFF &&
784 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
785 DMU_OT_IS_VALID(zp->zp_type) &&
788 zp->zp_copies <= spa_max_replication(spa));
790 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
791 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
792 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
793 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
795 zio->io_ready = ready;
796 zio->io_children_ready = children_ready;
797 zio->io_physdone = physdone;
801 * Data can be NULL if we are going to call zio_write_override() to
802 * provide the already-allocated BP. But we may need the data to
803 * verify a dedup hit (if requested). In this case, don't try to
804 * dedup (just take the already-allocated BP verbatim).
806 if (data == NULL && zio->io_prop.zp_dedup_verify) {
807 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
814 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
815 uint64_t size, zio_done_func_t *done, void *private,
816 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
820 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
821 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
822 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
828 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
830 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
831 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
832 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
833 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
836 * We must reset the io_prop to match the values that existed
837 * when the bp was first written by dmu_sync() keeping in mind
838 * that nopwrite and dedup are mutually exclusive.
840 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
841 zio->io_prop.zp_nopwrite = nopwrite;
842 zio->io_prop.zp_copies = copies;
843 zio->io_bp_override = bp;
847 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
851 * The check for EMBEDDED is a performance optimization. We
852 * process the free here (by ignoring it) rather than
853 * putting it on the list and then processing it in zio_free_sync().
855 if (BP_IS_EMBEDDED(bp))
857 metaslab_check_free(spa, bp);
860 * Frees that are for the currently-syncing txg, are not going to be
861 * deferred, and which will not need to do a read (i.e. not GANG or
862 * DEDUP), can be processed immediately. Otherwise, put them on the
863 * in-memory list for later processing.
865 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
866 txg != spa->spa_syncing_txg ||
867 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
868 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
870 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp,
871 BP_GET_PSIZE(bp), 0)));
876 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
877 uint64_t size, enum zio_flag flags)
880 enum zio_stage stage = ZIO_FREE_PIPELINE;
882 ASSERT(!BP_IS_HOLE(bp));
883 ASSERT(spa_syncing_txg(spa) == txg);
884 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
886 if (BP_IS_EMBEDDED(bp))
887 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
889 metaslab_check_free(spa, bp);
892 if (zfs_trim_enabled)
893 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START |
894 ZIO_STAGE_VDEV_IO_ASSESS;
896 * GANG and DEDUP blocks can induce a read (for the gang block header,
897 * or the DDT), so issue them asynchronously so that this thread is
900 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
901 stage |= ZIO_STAGE_ISSUE_ASYNC;
903 flags |= ZIO_FLAG_DONT_QUEUE;
905 zio = zio_create(pio, spa, txg, bp, NULL, size,
906 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags,
907 NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
913 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
914 zio_done_func_t *done, void *private, enum zio_flag flags)
918 dprintf_bp(bp, "claiming in txg %llu", txg);
920 if (BP_IS_EMBEDDED(bp))
921 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
924 * A claim is an allocation of a specific block. Claims are needed
925 * to support immediate writes in the intent log. The issue is that
926 * immediate writes contain committed data, but in a txg that was
927 * *not* committed. Upon opening the pool after an unclean shutdown,
928 * the intent log claims all blocks that contain immediate write data
929 * so that the SPA knows they're in use.
931 * All claims *must* be resolved in the first txg -- before the SPA
932 * starts allocating blocks -- so that nothing is allocated twice.
933 * If txg == 0 we just verify that the block is claimable.
935 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
936 ASSERT(txg == spa_first_txg(spa) || txg == 0);
937 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
939 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
940 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
941 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
947 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset,
948 uint64_t size, zio_done_func_t *done, void *private,
949 zio_priority_t priority, enum zio_flag flags)
954 if (vd->vdev_children == 0) {
955 zio = zio_create(pio, spa, 0, NULL, NULL, size, done, private,
956 ZIO_TYPE_IOCTL, priority, flags, vd, offset, NULL,
957 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
961 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
963 for (c = 0; c < vd->vdev_children; c++)
964 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
965 offset, size, done, private, priority, flags));
972 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
973 void *data, int checksum, zio_done_func_t *done, void *private,
974 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
978 ASSERT(vd->vdev_children == 0);
979 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
980 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
981 ASSERT3U(offset + size, <=, vd->vdev_psize);
983 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
984 ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
985 NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
987 zio->io_prop.zp_checksum = checksum;
993 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
994 void *data, int checksum, zio_done_func_t *done, void *private,
995 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
999 ASSERT(vd->vdev_children == 0);
1000 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1001 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1002 ASSERT3U(offset + size, <=, vd->vdev_psize);
1004 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
1005 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
1006 NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
1008 zio->io_prop.zp_checksum = checksum;
1010 if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
1012 * zec checksums are necessarily destructive -- they modify
1013 * the end of the write buffer to hold the verifier/checksum.
1014 * Therefore, we must make a local copy in case the data is
1015 * being written to multiple places in parallel.
1017 void *wbuf = zio_buf_alloc(size);
1018 bcopy(data, wbuf, size);
1019 zio_push_transform(zio, wbuf, size, size, NULL);
1026 * Create a child I/O to do some work for us.
1029 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1030 void *data, uint64_t size, int type, zio_priority_t priority,
1031 enum zio_flag flags, zio_done_func_t *done, void *private)
1033 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1036 ASSERT(vd->vdev_parent ==
1037 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
1039 if (type == ZIO_TYPE_READ && bp != NULL) {
1041 * If we have the bp, then the child should perform the
1042 * checksum and the parent need not. This pushes error
1043 * detection as close to the leaves as possible and
1044 * eliminates redundant checksums in the interior nodes.
1046 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1047 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1050 /* Not all IO types require vdev io done stage e.g. free */
1051 if (!(pio->io_pipeline & ZIO_STAGE_VDEV_IO_DONE))
1052 pipeline &= ~ZIO_STAGE_VDEV_IO_DONE;
1054 if (vd->vdev_children == 0)
1055 offset += VDEV_LABEL_START_SIZE;
1057 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
1060 * If we've decided to do a repair, the write is not speculative --
1061 * even if the original read was.
1063 if (flags & ZIO_FLAG_IO_REPAIR)
1064 flags &= ~ZIO_FLAG_SPECULATIVE;
1066 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
1067 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1068 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1070 zio->io_physdone = pio->io_physdone;
1071 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1072 zio->io_logical->io_phys_children++;
1078 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
1079 int type, zio_priority_t priority, enum zio_flag flags,
1080 zio_done_func_t *done, void *private)
1084 ASSERT(vd->vdev_ops->vdev_op_leaf);
1086 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1087 data, size, done, private, type, priority,
1088 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1090 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1096 zio_flush(zio_t *zio, vdev_t *vd)
1098 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0,
1099 NULL, NULL, ZIO_PRIORITY_NOW,
1100 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1104 zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size)
1107 ASSERT(vd->vdev_ops->vdev_op_leaf);
1109 return (zio_create(zio, spa, 0, NULL, NULL, size, NULL, NULL,
1110 ZIO_TYPE_FREE, ZIO_PRIORITY_TRIM, ZIO_FLAG_DONT_AGGREGATE |
1111 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY,
1112 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PHYS_PIPELINE));
1116 zio_shrink(zio_t *zio, uint64_t size)
1118 ASSERT(zio->io_executor == NULL);
1119 ASSERT(zio->io_orig_size == zio->io_size);
1120 ASSERT(size <= zio->io_size);
1123 * We don't shrink for raidz because of problems with the
1124 * reconstruction when reading back less than the block size.
1125 * Note, BP_IS_RAIDZ() assumes no compression.
1127 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1128 if (!BP_IS_RAIDZ(zio->io_bp))
1129 zio->io_orig_size = zio->io_size = size;
1133 * ==========================================================================
1134 * Prepare to read and write logical blocks
1135 * ==========================================================================
1139 zio_read_bp_init(zio_t *zio)
1141 blkptr_t *bp = zio->io_bp;
1143 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1144 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1145 !(zio->io_flags & ZIO_FLAG_RAW)) {
1147 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1148 void *cbuf = zio_buf_alloc(psize);
1150 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
1153 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1154 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1155 decode_embedded_bp_compressed(bp, zio->io_data);
1157 ASSERT(!BP_IS_EMBEDDED(bp));
1160 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1161 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1163 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1164 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1166 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1167 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1169 return (ZIO_PIPELINE_CONTINUE);
1173 zio_write_bp_init(zio_t *zio)
1175 spa_t *spa = zio->io_spa;
1176 zio_prop_t *zp = &zio->io_prop;
1177 enum zio_compress compress = zp->zp_compress;
1178 blkptr_t *bp = zio->io_bp;
1179 uint64_t lsize = zio->io_size;
1180 uint64_t psize = lsize;
1184 * If our children haven't all reached the ready stage,
1185 * wait for them and then repeat this pipeline stage.
1187 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1188 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1189 return (ZIO_PIPELINE_STOP);
1191 if (!IO_IS_ALLOCATING(zio))
1192 return (ZIO_PIPELINE_CONTINUE);
1194 if (zio->io_children_ready != NULL) {
1196 * Now that all our children are ready, run the callback
1197 * associated with this zio in case it wants to modify the
1198 * data to be written.
1200 ASSERT3U(zp->zp_level, >, 0);
1201 zio->io_children_ready(zio);
1204 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1206 if (zio->io_bp_override) {
1207 ASSERT(bp->blk_birth != zio->io_txg);
1208 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1210 *bp = *zio->io_bp_override;
1211 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1213 if (BP_IS_EMBEDDED(bp))
1214 return (ZIO_PIPELINE_CONTINUE);
1217 * If we've been overridden and nopwrite is set then
1218 * set the flag accordingly to indicate that a nopwrite
1219 * has already occurred.
1221 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1222 ASSERT(!zp->zp_dedup);
1223 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1224 return (ZIO_PIPELINE_CONTINUE);
1227 ASSERT(!zp->zp_nopwrite);
1229 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1230 return (ZIO_PIPELINE_CONTINUE);
1232 ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags &
1233 ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify);
1235 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1236 BP_SET_DEDUP(bp, 1);
1237 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1238 return (ZIO_PIPELINE_CONTINUE);
1240 zio->io_bp_override = NULL;
1244 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1246 * We're rewriting an existing block, which means we're
1247 * working on behalf of spa_sync(). For spa_sync() to
1248 * converge, it must eventually be the case that we don't
1249 * have to allocate new blocks. But compression changes
1250 * the blocksize, which forces a reallocate, and makes
1251 * convergence take longer. Therefore, after the first
1252 * few passes, stop compressing to ensure convergence.
1254 pass = spa_sync_pass(spa);
1256 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1257 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1258 ASSERT(!BP_GET_DEDUP(bp));
1260 if (pass >= zfs_sync_pass_dont_compress)
1261 compress = ZIO_COMPRESS_OFF;
1263 /* Make sure someone doesn't change their mind on overwrites */
1264 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1265 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1268 if (compress != ZIO_COMPRESS_OFF) {
1269 void *cbuf = zio_buf_alloc(lsize);
1270 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1271 if (psize == 0 || psize == lsize) {
1272 compress = ZIO_COMPRESS_OFF;
1273 zio_buf_free(cbuf, lsize);
1274 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1275 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1276 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1277 encode_embedded_bp_compressed(bp,
1278 cbuf, compress, lsize, psize);
1279 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1280 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1281 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1282 zio_buf_free(cbuf, lsize);
1283 bp->blk_birth = zio->io_txg;
1284 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1285 ASSERT(spa_feature_is_active(spa,
1286 SPA_FEATURE_EMBEDDED_DATA));
1287 return (ZIO_PIPELINE_CONTINUE);
1290 * Round up compressed size up to the ashift
1291 * of the smallest-ashift device, and zero the tail.
1292 * This ensures that the compressed size of the BP
1293 * (and thus compressratio property) are correct,
1294 * in that we charge for the padding used to fill out
1297 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
1298 size_t rounded = (size_t)P2ROUNDUP(psize,
1299 1ULL << spa->spa_min_ashift);
1300 if (rounded >= lsize) {
1301 compress = ZIO_COMPRESS_OFF;
1302 zio_buf_free(cbuf, lsize);
1305 bzero((char *)cbuf + psize, rounded - psize);
1307 zio_push_transform(zio, cbuf,
1308 psize, lsize, NULL);
1314 * The final pass of spa_sync() must be all rewrites, but the first
1315 * few passes offer a trade-off: allocating blocks defers convergence,
1316 * but newly allocated blocks are sequential, so they can be written
1317 * to disk faster. Therefore, we allow the first few passes of
1318 * spa_sync() to allocate new blocks, but force rewrites after that.
1319 * There should only be a handful of blocks after pass 1 in any case.
1321 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1322 BP_GET_PSIZE(bp) == psize &&
1323 pass >= zfs_sync_pass_rewrite) {
1325 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1326 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1327 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1330 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1334 if (zio->io_bp_orig.blk_birth != 0 &&
1335 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1336 BP_SET_LSIZE(bp, lsize);
1337 BP_SET_TYPE(bp, zp->zp_type);
1338 BP_SET_LEVEL(bp, zp->zp_level);
1339 BP_SET_BIRTH(bp, zio->io_txg, 0);
1341 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1343 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1344 BP_SET_LSIZE(bp, lsize);
1345 BP_SET_TYPE(bp, zp->zp_type);
1346 BP_SET_LEVEL(bp, zp->zp_level);
1347 BP_SET_PSIZE(bp, psize);
1348 BP_SET_COMPRESS(bp, compress);
1349 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1350 BP_SET_DEDUP(bp, zp->zp_dedup);
1351 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1353 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1354 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1355 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1357 if (zp->zp_nopwrite) {
1358 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1359 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1360 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1364 return (ZIO_PIPELINE_CONTINUE);
1368 zio_free_bp_init(zio_t *zio)
1370 blkptr_t *bp = zio->io_bp;
1372 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1373 if (BP_GET_DEDUP(bp))
1374 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1377 return (ZIO_PIPELINE_CONTINUE);
1381 * ==========================================================================
1382 * Execute the I/O pipeline
1383 * ==========================================================================
1387 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1389 spa_t *spa = zio->io_spa;
1390 zio_type_t t = zio->io_type;
1391 int flags = (cutinline ? TQ_FRONT : 0);
1393 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1396 * If we're a config writer or a probe, the normal issue and
1397 * interrupt threads may all be blocked waiting for the config lock.
1398 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1400 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1404 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1406 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1410 * If this is a high priority I/O, then use the high priority taskq if
1413 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1414 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1417 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1420 * NB: We are assuming that the zio can only be dispatched
1421 * to a single taskq at a time. It would be a grievous error
1422 * to dispatch the zio to another taskq at the same time.
1424 #if defined(illumos) || !defined(_KERNEL)
1425 ASSERT(zio->io_tqent.tqent_next == NULL);
1427 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
1429 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1430 flags, &zio->io_tqent);
1434 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1436 kthread_t *executor = zio->io_executor;
1437 spa_t *spa = zio->io_spa;
1439 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1440 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1442 for (i = 0; i < tqs->stqs_count; i++) {
1443 if (taskq_member(tqs->stqs_taskq[i], executor))
1452 zio_issue_async(zio_t *zio)
1454 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1456 return (ZIO_PIPELINE_STOP);
1460 zio_interrupt(zio_t *zio)
1462 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1466 zio_delay_interrupt(zio_t *zio)
1469 * The timeout_generic() function isn't defined in userspace, so
1470 * rather than trying to implement the function, the zio delay
1471 * functionality has been disabled for userspace builds.
1476 * If io_target_timestamp is zero, then no delay has been registered
1477 * for this IO, thus jump to the end of this function and "skip" the
1478 * delay; issuing it directly to the zio layer.
1480 if (zio->io_target_timestamp != 0) {
1481 hrtime_t now = gethrtime();
1483 if (now >= zio->io_target_timestamp) {
1485 * This IO has already taken longer than the target
1486 * delay to complete, so we don't want to delay it
1487 * any longer; we "miss" the delay and issue it
1488 * directly to the zio layer. This is likely due to
1489 * the target latency being set to a value less than
1490 * the underlying hardware can satisfy (e.g. delay
1491 * set to 1ms, but the disks take 10ms to complete an
1495 DTRACE_PROBE2(zio__delay__miss, zio_t *, zio,
1500 hrtime_t diff = zio->io_target_timestamp - now;
1502 DTRACE_PROBE3(zio__delay__hit, zio_t *, zio,
1503 hrtime_t, now, hrtime_t, diff);
1505 (void) timeout_generic(CALLOUT_NORMAL,
1506 (void (*)(void *))zio_interrupt, zio, diff, 1, 0);
1513 DTRACE_PROBE1(zio__delay__skip, zio_t *, zio);
1518 * Execute the I/O pipeline until one of the following occurs:
1520 * (1) the I/O completes
1521 * (2) the pipeline stalls waiting for dependent child I/Os
1522 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1523 * (4) the I/O is delegated by vdev-level caching or aggregation
1524 * (5) the I/O is deferred due to vdev-level queueing
1525 * (6) the I/O is handed off to another thread.
1527 * In all cases, the pipeline stops whenever there's no CPU work; it never
1528 * burns a thread in cv_wait().
1530 * There's no locking on io_stage because there's no legitimate way
1531 * for multiple threads to be attempting to process the same I/O.
1533 static zio_pipe_stage_t *zio_pipeline[];
1536 zio_execute(zio_t *zio)
1538 zio->io_executor = curthread;
1540 while (zio->io_stage < ZIO_STAGE_DONE) {
1541 enum zio_stage pipeline = zio->io_pipeline;
1542 enum zio_stage stage = zio->io_stage;
1545 ASSERT(!MUTEX_HELD(&zio->io_lock));
1546 ASSERT(ISP2(stage));
1547 ASSERT(zio->io_stall == NULL);
1551 } while ((stage & pipeline) == 0);
1553 ASSERT(stage <= ZIO_STAGE_DONE);
1556 * If we are in interrupt context and this pipeline stage
1557 * will grab a config lock that is held across I/O,
1558 * or may wait for an I/O that needs an interrupt thread
1559 * to complete, issue async to avoid deadlock.
1561 * For VDEV_IO_START, we cut in line so that the io will
1562 * be sent to disk promptly.
1564 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1565 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1566 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1567 zio_requeue_io_start_cut_in_line : B_FALSE;
1568 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1572 zio->io_stage = stage;
1573 rv = zio_pipeline[highbit64(stage) - 1](zio);
1575 if (rv == ZIO_PIPELINE_STOP)
1578 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1583 * ==========================================================================
1584 * Initiate I/O, either sync or async
1585 * ==========================================================================
1588 zio_wait(zio_t *zio)
1592 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1593 ASSERT(zio->io_executor == NULL);
1595 zio->io_waiter = curthread;
1599 mutex_enter(&zio->io_lock);
1600 while (zio->io_executor != NULL)
1601 cv_wait(&zio->io_cv, &zio->io_lock);
1602 mutex_exit(&zio->io_lock);
1604 error = zio->io_error;
1611 zio_nowait(zio_t *zio)
1613 ASSERT(zio->io_executor == NULL);
1615 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1616 zio_unique_parent(zio) == NULL) {
1618 * This is a logical async I/O with no parent to wait for it.
1619 * We add it to the spa_async_root_zio "Godfather" I/O which
1620 * will ensure they complete prior to unloading the pool.
1622 spa_t *spa = zio->io_spa;
1624 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1631 * ==========================================================================
1632 * Reexecute or suspend/resume failed I/O
1633 * ==========================================================================
1637 zio_reexecute(zio_t *pio)
1639 zio_t *cio, *cio_next;
1641 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1642 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1643 ASSERT(pio->io_gang_leader == NULL);
1644 ASSERT(pio->io_gang_tree == NULL);
1646 pio->io_flags = pio->io_orig_flags;
1647 pio->io_stage = pio->io_orig_stage;
1648 pio->io_pipeline = pio->io_orig_pipeline;
1649 pio->io_reexecute = 0;
1650 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1652 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1653 pio->io_state[w] = 0;
1654 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1655 pio->io_child_error[c] = 0;
1657 if (IO_IS_ALLOCATING(pio))
1658 BP_ZERO(pio->io_bp);
1661 * As we reexecute pio's children, new children could be created.
1662 * New children go to the head of pio's io_child_list, however,
1663 * so we will (correctly) not reexecute them. The key is that
1664 * the remainder of pio's io_child_list, from 'cio_next' onward,
1665 * cannot be affected by any side effects of reexecuting 'cio'.
1667 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1668 cio_next = zio_walk_children(pio);
1669 mutex_enter(&pio->io_lock);
1670 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1671 pio->io_children[cio->io_child_type][w]++;
1672 mutex_exit(&pio->io_lock);
1677 * Now that all children have been reexecuted, execute the parent.
1678 * We don't reexecute "The Godfather" I/O here as it's the
1679 * responsibility of the caller to wait on him.
1681 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1686 zio_suspend(spa_t *spa, zio_t *zio)
1688 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1689 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1690 "failure and the failure mode property for this pool "
1691 "is set to panic.", spa_name(spa));
1693 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1695 mutex_enter(&spa->spa_suspend_lock);
1697 if (spa->spa_suspend_zio_root == NULL)
1698 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1699 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1700 ZIO_FLAG_GODFATHER);
1702 spa->spa_suspended = B_TRUE;
1705 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1706 ASSERT(zio != spa->spa_suspend_zio_root);
1707 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1708 ASSERT(zio_unique_parent(zio) == NULL);
1709 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1710 zio_add_child(spa->spa_suspend_zio_root, zio);
1713 mutex_exit(&spa->spa_suspend_lock);
1717 zio_resume(spa_t *spa)
1722 * Reexecute all previously suspended i/o.
1724 mutex_enter(&spa->spa_suspend_lock);
1725 spa->spa_suspended = B_FALSE;
1726 cv_broadcast(&spa->spa_suspend_cv);
1727 pio = spa->spa_suspend_zio_root;
1728 spa->spa_suspend_zio_root = NULL;
1729 mutex_exit(&spa->spa_suspend_lock);
1735 return (zio_wait(pio));
1739 zio_resume_wait(spa_t *spa)
1741 mutex_enter(&spa->spa_suspend_lock);
1742 while (spa_suspended(spa))
1743 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1744 mutex_exit(&spa->spa_suspend_lock);
1748 * ==========================================================================
1751 * A gang block is a collection of small blocks that looks to the DMU
1752 * like one large block. When zio_dva_allocate() cannot find a block
1753 * of the requested size, due to either severe fragmentation or the pool
1754 * being nearly full, it calls zio_write_gang_block() to construct the
1755 * block from smaller fragments.
1757 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1758 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1759 * an indirect block: it's an array of block pointers. It consumes
1760 * only one sector and hence is allocatable regardless of fragmentation.
1761 * The gang header's bps point to its gang members, which hold the data.
1763 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1764 * as the verifier to ensure uniqueness of the SHA256 checksum.
1765 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1766 * not the gang header. This ensures that data block signatures (needed for
1767 * deduplication) are independent of how the block is physically stored.
1769 * Gang blocks can be nested: a gang member may itself be a gang block.
1770 * Thus every gang block is a tree in which root and all interior nodes are
1771 * gang headers, and the leaves are normal blocks that contain user data.
1772 * The root of the gang tree is called the gang leader.
1774 * To perform any operation (read, rewrite, free, claim) on a gang block,
1775 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1776 * in the io_gang_tree field of the original logical i/o by recursively
1777 * reading the gang leader and all gang headers below it. This yields
1778 * an in-core tree containing the contents of every gang header and the
1779 * bps for every constituent of the gang block.
1781 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1782 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1783 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1784 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1785 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1786 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1787 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1788 * of the gang header plus zio_checksum_compute() of the data to update the
1789 * gang header's blk_cksum as described above.
1791 * The two-phase assemble/issue model solves the problem of partial failure --
1792 * what if you'd freed part of a gang block but then couldn't read the
1793 * gang header for another part? Assembling the entire gang tree first
1794 * ensures that all the necessary gang header I/O has succeeded before
1795 * starting the actual work of free, claim, or write. Once the gang tree
1796 * is assembled, free and claim are in-memory operations that cannot fail.
1798 * In the event that a gang write fails, zio_dva_unallocate() walks the
1799 * gang tree to immediately free (i.e. insert back into the space map)
1800 * everything we've allocated. This ensures that we don't get ENOSPC
1801 * errors during repeated suspend/resume cycles due to a flaky device.
1803 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1804 * the gang tree, we won't modify the block, so we can safely defer the free
1805 * (knowing that the block is still intact). If we *can* assemble the gang
1806 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1807 * each constituent bp and we can allocate a new block on the next sync pass.
1809 * In all cases, the gang tree allows complete recovery from partial failure.
1810 * ==========================================================================
1814 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1819 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1820 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1821 &pio->io_bookmark));
1825 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1830 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1831 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1832 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1834 * As we rewrite each gang header, the pipeline will compute
1835 * a new gang block header checksum for it; but no one will
1836 * compute a new data checksum, so we do that here. The one
1837 * exception is the gang leader: the pipeline already computed
1838 * its data checksum because that stage precedes gang assembly.
1839 * (Presently, nothing actually uses interior data checksums;
1840 * this is just good hygiene.)
1842 if (gn != pio->io_gang_leader->io_gang_tree) {
1843 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1844 data, BP_GET_PSIZE(bp));
1847 * If we are here to damage data for testing purposes,
1848 * leave the GBH alone so that we can detect the damage.
1850 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1851 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1853 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1854 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1855 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1863 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1865 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1866 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp),
1867 ZIO_GANG_CHILD_FLAGS(pio)));
1872 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1874 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1875 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1878 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1887 static void zio_gang_tree_assemble_done(zio_t *zio);
1889 static zio_gang_node_t *
1890 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1892 zio_gang_node_t *gn;
1894 ASSERT(*gnpp == NULL);
1896 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1897 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1904 zio_gang_node_free(zio_gang_node_t **gnpp)
1906 zio_gang_node_t *gn = *gnpp;
1908 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1909 ASSERT(gn->gn_child[g] == NULL);
1911 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1912 kmem_free(gn, sizeof (*gn));
1917 zio_gang_tree_free(zio_gang_node_t **gnpp)
1919 zio_gang_node_t *gn = *gnpp;
1924 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1925 zio_gang_tree_free(&gn->gn_child[g]);
1927 zio_gang_node_free(gnpp);
1931 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1933 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1935 ASSERT(gio->io_gang_leader == gio);
1936 ASSERT(BP_IS_GANG(bp));
1938 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1939 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1940 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1944 zio_gang_tree_assemble_done(zio_t *zio)
1946 zio_t *gio = zio->io_gang_leader;
1947 zio_gang_node_t *gn = zio->io_private;
1948 blkptr_t *bp = zio->io_bp;
1950 ASSERT(gio == zio_unique_parent(zio));
1951 ASSERT(zio->io_child_count == 0);
1956 if (BP_SHOULD_BYTESWAP(bp))
1957 byteswap_uint64_array(zio->io_data, zio->io_size);
1959 ASSERT(zio->io_data == gn->gn_gbh);
1960 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1961 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1963 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1964 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1965 if (!BP_IS_GANG(gbp))
1967 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1972 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1974 zio_t *gio = pio->io_gang_leader;
1977 ASSERT(BP_IS_GANG(bp) == !!gn);
1978 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1979 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1982 * If you're a gang header, your data is in gn->gn_gbh.
1983 * If you're a gang member, your data is in 'data' and gn == NULL.
1985 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1988 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1990 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1991 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1992 if (BP_IS_HOLE(gbp))
1994 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1995 data = (char *)data + BP_GET_PSIZE(gbp);
1999 if (gn == gio->io_gang_tree && gio->io_data != NULL)
2000 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
2007 zio_gang_assemble(zio_t *zio)
2009 blkptr_t *bp = zio->io_bp;
2011 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
2012 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2014 zio->io_gang_leader = zio;
2016 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
2018 return (ZIO_PIPELINE_CONTINUE);
2022 zio_gang_issue(zio_t *zio)
2024 blkptr_t *bp = zio->io_bp;
2026 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
2027 return (ZIO_PIPELINE_STOP);
2029 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
2030 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2032 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
2033 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
2035 zio_gang_tree_free(&zio->io_gang_tree);
2037 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2039 return (ZIO_PIPELINE_CONTINUE);
2043 zio_write_gang_member_ready(zio_t *zio)
2045 zio_t *pio = zio_unique_parent(zio);
2046 zio_t *gio = zio->io_gang_leader;
2047 dva_t *cdva = zio->io_bp->blk_dva;
2048 dva_t *pdva = pio->io_bp->blk_dva;
2051 if (BP_IS_HOLE(zio->io_bp))
2054 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
2056 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
2057 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
2058 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2059 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
2060 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
2062 mutex_enter(&pio->io_lock);
2063 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
2064 ASSERT(DVA_GET_GANG(&pdva[d]));
2065 asize = DVA_GET_ASIZE(&pdva[d]);
2066 asize += DVA_GET_ASIZE(&cdva[d]);
2067 DVA_SET_ASIZE(&pdva[d], asize);
2069 mutex_exit(&pio->io_lock);
2073 zio_write_gang_block(zio_t *pio)
2075 spa_t *spa = pio->io_spa;
2076 blkptr_t *bp = pio->io_bp;
2077 zio_t *gio = pio->io_gang_leader;
2079 zio_gang_node_t *gn, **gnpp;
2080 zio_gbh_phys_t *gbh;
2081 uint64_t txg = pio->io_txg;
2082 uint64_t resid = pio->io_size;
2084 int copies = gio->io_prop.zp_copies;
2085 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
2089 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
2090 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
2091 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
2093 pio->io_error = error;
2094 return (ZIO_PIPELINE_CONTINUE);
2098 gnpp = &gio->io_gang_tree;
2100 gnpp = pio->io_private;
2101 ASSERT(pio->io_ready == zio_write_gang_member_ready);
2104 gn = zio_gang_node_alloc(gnpp);
2106 bzero(gbh, SPA_GANGBLOCKSIZE);
2109 * Create the gang header.
2111 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
2112 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2115 * Create and nowait the gang children.
2117 for (int g = 0; resid != 0; resid -= lsize, g++) {
2118 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2120 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2122 zp.zp_checksum = gio->io_prop.zp_checksum;
2123 zp.zp_compress = ZIO_COMPRESS_OFF;
2124 zp.zp_type = DMU_OT_NONE;
2126 zp.zp_copies = gio->io_prop.zp_copies;
2127 zp.zp_dedup = B_FALSE;
2128 zp.zp_dedup_verify = B_FALSE;
2129 zp.zp_nopwrite = B_FALSE;
2131 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2132 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
2133 zio_write_gang_member_ready, NULL, NULL, NULL,
2134 &gn->gn_child[g], pio->io_priority,
2135 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark));
2139 * Set pio's pipeline to just wait for zio to finish.
2141 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2145 return (ZIO_PIPELINE_CONTINUE);
2149 * The zio_nop_write stage in the pipeline determines if allocating a
2150 * new bp is necessary. The nopwrite feature can handle writes in
2151 * either syncing or open context (i.e. zil writes) and as a result is
2152 * mutually exclusive with dedup.
2154 * By leveraging a cryptographically secure checksum, such as SHA256, we
2155 * can compare the checksums of the new data and the old to determine if
2156 * allocating a new block is required. Note that our requirements for
2157 * cryptographic strength are fairly weak: there can't be any accidental
2158 * hash collisions, but we don't need to be secure against intentional
2159 * (malicious) collisions. To trigger a nopwrite, you have to be able
2160 * to write the file to begin with, and triggering an incorrect (hash
2161 * collision) nopwrite is no worse than simply writing to the file.
2162 * That said, there are no known attacks against the checksum algorithms
2163 * used for nopwrite, assuming that the salt and the checksums
2164 * themselves remain secret.
2167 zio_nop_write(zio_t *zio)
2169 blkptr_t *bp = zio->io_bp;
2170 blkptr_t *bp_orig = &zio->io_bp_orig;
2171 zio_prop_t *zp = &zio->io_prop;
2173 ASSERT(BP_GET_LEVEL(bp) == 0);
2174 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2175 ASSERT(zp->zp_nopwrite);
2176 ASSERT(!zp->zp_dedup);
2177 ASSERT(zio->io_bp_override == NULL);
2178 ASSERT(IO_IS_ALLOCATING(zio));
2181 * Check to see if the original bp and the new bp have matching
2182 * characteristics (i.e. same checksum, compression algorithms, etc).
2183 * If they don't then just continue with the pipeline which will
2184 * allocate a new bp.
2186 if (BP_IS_HOLE(bp_orig) ||
2187 !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
2188 ZCHECKSUM_FLAG_NOPWRITE) ||
2189 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2190 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2191 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2192 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2193 return (ZIO_PIPELINE_CONTINUE);
2196 * If the checksums match then reset the pipeline so that we
2197 * avoid allocating a new bp and issuing any I/O.
2199 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2200 ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
2201 ZCHECKSUM_FLAG_NOPWRITE);
2202 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2203 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2204 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2205 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2206 sizeof (uint64_t)) == 0);
2209 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2210 zio->io_flags |= ZIO_FLAG_NOPWRITE;
2213 return (ZIO_PIPELINE_CONTINUE);
2217 * ==========================================================================
2219 * ==========================================================================
2222 zio_ddt_child_read_done(zio_t *zio)
2224 blkptr_t *bp = zio->io_bp;
2225 ddt_entry_t *dde = zio->io_private;
2227 zio_t *pio = zio_unique_parent(zio);
2229 mutex_enter(&pio->io_lock);
2230 ddp = ddt_phys_select(dde, bp);
2231 if (zio->io_error == 0)
2232 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2233 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
2234 dde->dde_repair_data = zio->io_data;
2236 zio_buf_free(zio->io_data, zio->io_size);
2237 mutex_exit(&pio->io_lock);
2241 zio_ddt_read_start(zio_t *zio)
2243 blkptr_t *bp = zio->io_bp;
2245 ASSERT(BP_GET_DEDUP(bp));
2246 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2247 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2249 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2250 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2251 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2252 ddt_phys_t *ddp = dde->dde_phys;
2253 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2256 ASSERT(zio->io_vsd == NULL);
2259 if (ddp_self == NULL)
2260 return (ZIO_PIPELINE_CONTINUE);
2262 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2263 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2265 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2267 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2268 zio_buf_alloc(zio->io_size), zio->io_size,
2269 zio_ddt_child_read_done, dde, zio->io_priority,
2270 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
2271 &zio->io_bookmark));
2273 return (ZIO_PIPELINE_CONTINUE);
2276 zio_nowait(zio_read(zio, zio->io_spa, bp,
2277 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
2278 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2280 return (ZIO_PIPELINE_CONTINUE);
2284 zio_ddt_read_done(zio_t *zio)
2286 blkptr_t *bp = zio->io_bp;
2288 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2289 return (ZIO_PIPELINE_STOP);
2291 ASSERT(BP_GET_DEDUP(bp));
2292 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2293 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2295 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2296 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2297 ddt_entry_t *dde = zio->io_vsd;
2299 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2300 return (ZIO_PIPELINE_CONTINUE);
2303 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2304 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2305 return (ZIO_PIPELINE_STOP);
2307 if (dde->dde_repair_data != NULL) {
2308 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2309 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2311 ddt_repair_done(ddt, dde);
2315 ASSERT(zio->io_vsd == NULL);
2317 return (ZIO_PIPELINE_CONTINUE);
2321 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2323 spa_t *spa = zio->io_spa;
2326 * Note: we compare the original data, not the transformed data,
2327 * because when zio->io_bp is an override bp, we will not have
2328 * pushed the I/O transforms. That's an important optimization
2329 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2331 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2332 zio_t *lio = dde->dde_lead_zio[p];
2335 return (lio->io_orig_size != zio->io_orig_size ||
2336 bcmp(zio->io_orig_data, lio->io_orig_data,
2337 zio->io_orig_size) != 0);
2341 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2342 ddt_phys_t *ddp = &dde->dde_phys[p];
2344 if (ddp->ddp_phys_birth != 0) {
2345 arc_buf_t *abuf = NULL;
2346 arc_flags_t aflags = ARC_FLAG_WAIT;
2347 blkptr_t blk = *zio->io_bp;
2350 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2354 error = arc_read(NULL, spa, &blk,
2355 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2356 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2357 &aflags, &zio->io_bookmark);
2360 if (arc_buf_size(abuf) != zio->io_orig_size ||
2361 bcmp(abuf->b_data, zio->io_orig_data,
2362 zio->io_orig_size) != 0)
2363 error = SET_ERROR(EEXIST);
2364 arc_buf_destroy(abuf, &abuf);
2368 return (error != 0);
2376 zio_ddt_child_write_ready(zio_t *zio)
2378 int p = zio->io_prop.zp_copies;
2379 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2380 ddt_entry_t *dde = zio->io_private;
2381 ddt_phys_t *ddp = &dde->dde_phys[p];
2389 ASSERT(dde->dde_lead_zio[p] == zio);
2391 ddt_phys_fill(ddp, zio->io_bp);
2393 while ((pio = zio_walk_parents(zio)) != NULL)
2394 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2400 zio_ddt_child_write_done(zio_t *zio)
2402 int p = zio->io_prop.zp_copies;
2403 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2404 ddt_entry_t *dde = zio->io_private;
2405 ddt_phys_t *ddp = &dde->dde_phys[p];
2409 ASSERT(ddp->ddp_refcnt == 0);
2410 ASSERT(dde->dde_lead_zio[p] == zio);
2411 dde->dde_lead_zio[p] = NULL;
2413 if (zio->io_error == 0) {
2414 while (zio_walk_parents(zio) != NULL)
2415 ddt_phys_addref(ddp);
2417 ddt_phys_clear(ddp);
2424 zio_ddt_ditto_write_done(zio_t *zio)
2426 int p = DDT_PHYS_DITTO;
2427 zio_prop_t *zp = &zio->io_prop;
2428 blkptr_t *bp = zio->io_bp;
2429 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2430 ddt_entry_t *dde = zio->io_private;
2431 ddt_phys_t *ddp = &dde->dde_phys[p];
2432 ddt_key_t *ddk = &dde->dde_key;
2436 ASSERT(ddp->ddp_refcnt == 0);
2437 ASSERT(dde->dde_lead_zio[p] == zio);
2438 dde->dde_lead_zio[p] = NULL;
2440 if (zio->io_error == 0) {
2441 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2442 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2443 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2444 if (ddp->ddp_phys_birth != 0)
2445 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2446 ddt_phys_fill(ddp, bp);
2453 zio_ddt_write(zio_t *zio)
2455 spa_t *spa = zio->io_spa;
2456 blkptr_t *bp = zio->io_bp;
2457 uint64_t txg = zio->io_txg;
2458 zio_prop_t *zp = &zio->io_prop;
2459 int p = zp->zp_copies;
2463 ddt_t *ddt = ddt_select(spa, bp);
2467 ASSERT(BP_GET_DEDUP(bp));
2468 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2469 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2472 dde = ddt_lookup(ddt, bp, B_TRUE);
2473 ddp = &dde->dde_phys[p];
2475 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2477 * If we're using a weak checksum, upgrade to a strong checksum
2478 * and try again. If we're already using a strong checksum,
2479 * we can't resolve it, so just convert to an ordinary write.
2480 * (And automatically e-mail a paper to Nature?)
2482 if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
2483 ZCHECKSUM_FLAG_DEDUP)) {
2484 zp->zp_checksum = spa_dedup_checksum(spa);
2485 zio_pop_transforms(zio);
2486 zio->io_stage = ZIO_STAGE_OPEN;
2489 zp->zp_dedup = B_FALSE;
2491 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2493 return (ZIO_PIPELINE_CONTINUE);
2496 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2497 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2499 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2500 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2501 zio_prop_t czp = *zp;
2503 czp.zp_copies = ditto_copies;
2506 * If we arrived here with an override bp, we won't have run
2507 * the transform stack, so we won't have the data we need to
2508 * generate a child i/o. So, toss the override bp and restart.
2509 * This is safe, because using the override bp is just an
2510 * optimization; and it's rare, so the cost doesn't matter.
2512 if (zio->io_bp_override) {
2513 zio_pop_transforms(zio);
2514 zio->io_stage = ZIO_STAGE_OPEN;
2515 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2516 zio->io_bp_override = NULL;
2519 return (ZIO_PIPELINE_CONTINUE);
2522 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2523 zio->io_orig_size, &czp, NULL, NULL,
2524 NULL, zio_ddt_ditto_write_done, dde, zio->io_priority,
2525 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2527 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2528 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2531 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2532 if (ddp->ddp_phys_birth != 0)
2533 ddt_bp_fill(ddp, bp, txg);
2534 if (dde->dde_lead_zio[p] != NULL)
2535 zio_add_child(zio, dde->dde_lead_zio[p]);
2537 ddt_phys_addref(ddp);
2538 } else if (zio->io_bp_override) {
2539 ASSERT(bp->blk_birth == txg);
2540 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2541 ddt_phys_fill(ddp, bp);
2542 ddt_phys_addref(ddp);
2544 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2545 zio->io_orig_size, zp,
2546 zio_ddt_child_write_ready, NULL, NULL,
2547 zio_ddt_child_write_done, dde, zio->io_priority,
2548 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2550 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2551 dde->dde_lead_zio[p] = cio;
2561 return (ZIO_PIPELINE_CONTINUE);
2564 ddt_entry_t *freedde; /* for debugging */
2567 zio_ddt_free(zio_t *zio)
2569 spa_t *spa = zio->io_spa;
2570 blkptr_t *bp = zio->io_bp;
2571 ddt_t *ddt = ddt_select(spa, bp);
2575 ASSERT(BP_GET_DEDUP(bp));
2576 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2579 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2580 ddp = ddt_phys_select(dde, bp);
2581 ddt_phys_decref(ddp);
2584 return (ZIO_PIPELINE_CONTINUE);
2588 * ==========================================================================
2589 * Allocate and free blocks
2590 * ==========================================================================
2593 zio_dva_allocate(zio_t *zio)
2595 spa_t *spa = zio->io_spa;
2596 metaslab_class_t *mc = spa_normal_class(spa);
2597 blkptr_t *bp = zio->io_bp;
2601 if (zio->io_gang_leader == NULL) {
2602 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2603 zio->io_gang_leader = zio;
2606 ASSERT(BP_IS_HOLE(bp));
2607 ASSERT0(BP_GET_NDVAS(bp));
2608 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2609 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2610 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2613 * The dump device does not support gang blocks so allocation on
2614 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2615 * the "fast" gang feature.
2617 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2618 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2619 METASLAB_GANG_CHILD : 0;
2620 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2621 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2624 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2625 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2627 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2628 return (zio_write_gang_block(zio));
2629 zio->io_error = error;
2632 return (ZIO_PIPELINE_CONTINUE);
2636 zio_dva_free(zio_t *zio)
2638 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2640 return (ZIO_PIPELINE_CONTINUE);
2644 zio_dva_claim(zio_t *zio)
2648 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2650 zio->io_error = error;
2652 return (ZIO_PIPELINE_CONTINUE);
2656 * Undo an allocation. This is used by zio_done() when an I/O fails
2657 * and we want to give back the block we just allocated.
2658 * This handles both normal blocks and gang blocks.
2661 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2663 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2664 ASSERT(zio->io_bp_override == NULL);
2666 if (!BP_IS_HOLE(bp))
2667 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2670 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2671 zio_dva_unallocate(zio, gn->gn_child[g],
2672 &gn->gn_gbh->zg_blkptr[g]);
2678 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2681 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2682 uint64_t size, boolean_t use_slog)
2686 ASSERT(txg > spa_syncing_txg(spa));
2689 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2690 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2691 * when allocating them.
2694 error = metaslab_alloc(spa, spa_log_class(spa), size,
2695 new_bp, 1, txg, old_bp,
2696 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2700 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2701 new_bp, 1, txg, old_bp,
2702 METASLAB_HINTBP_AVOID);
2706 BP_SET_LSIZE(new_bp, size);
2707 BP_SET_PSIZE(new_bp, size);
2708 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2709 BP_SET_CHECKSUM(new_bp,
2710 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2711 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2712 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2713 BP_SET_LEVEL(new_bp, 0);
2714 BP_SET_DEDUP(new_bp, 0);
2715 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2722 * Free an intent log block.
2725 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2727 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2728 ASSERT(!BP_IS_GANG(bp));
2730 zio_free(spa, txg, bp);
2734 * ==========================================================================
2735 * Read, write and delete to physical devices
2736 * ==========================================================================
2741 * Issue an I/O to the underlying vdev. Typically the issue pipeline
2742 * stops after this stage and will resume upon I/O completion.
2743 * However, there are instances where the vdev layer may need to
2744 * continue the pipeline when an I/O was not issued. Since the I/O
2745 * that was sent to the vdev layer might be different than the one
2746 * currently active in the pipeline (see vdev_queue_io()), we explicitly
2747 * force the underlying vdev layers to call either zio_execute() or
2748 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
2751 zio_vdev_io_start(zio_t *zio)
2753 vdev_t *vd = zio->io_vd;
2755 spa_t *spa = zio->io_spa;
2758 ASSERT(zio->io_error == 0);
2759 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2762 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2763 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2766 * The mirror_ops handle multiple DVAs in a single BP.
2768 vdev_mirror_ops.vdev_op_io_start(zio);
2769 return (ZIO_PIPELINE_STOP);
2772 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE &&
2773 zio->io_priority == ZIO_PRIORITY_NOW) {
2774 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg);
2775 return (ZIO_PIPELINE_CONTINUE);
2779 * We keep track of time-sensitive I/Os so that the scan thread
2780 * can quickly react to certain workloads. In particular, we care
2781 * about non-scrubbing, top-level reads and writes with the following
2783 * - synchronous writes of user data to non-slog devices
2784 * - any reads of user data
2785 * When these conditions are met, adjust the timestamp of spa_last_io
2786 * which allows the scan thread to adjust its workload accordingly.
2788 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2789 vd == vd->vdev_top && !vd->vdev_islog &&
2790 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2791 zio->io_txg != spa_syncing_txg(spa)) {
2792 uint64_t old = spa->spa_last_io;
2793 uint64_t new = ddi_get_lbolt64();
2795 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2798 align = 1ULL << vd->vdev_top->vdev_ashift;
2800 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
2801 P2PHASE(zio->io_size, align) != 0) {
2802 /* Transform logical writes to be a full physical block size. */
2803 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2805 if (zio->io_type == ZIO_TYPE_READ ||
2806 zio->io_type == ZIO_TYPE_WRITE)
2807 abuf = zio_buf_alloc(asize);
2808 ASSERT(vd == vd->vdev_top);
2809 if (zio->io_type == ZIO_TYPE_WRITE) {
2810 bcopy(zio->io_data, abuf, zio->io_size);
2811 bzero(abuf + zio->io_size, asize - zio->io_size);
2813 zio_push_transform(zio, abuf, asize, abuf ? asize : 0,
2818 * If this is not a physical io, make sure that it is properly aligned
2819 * before proceeding.
2821 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
2822 ASSERT0(P2PHASE(zio->io_offset, align));
2823 ASSERT0(P2PHASE(zio->io_size, align));
2826 * For the physical io we allow alignment
2827 * to a logical block size.
2829 uint64_t log_align =
2830 1ULL << vd->vdev_top->vdev_logical_ashift;
2831 ASSERT0(P2PHASE(zio->io_offset, log_align));
2832 ASSERT0(P2PHASE(zio->io_size, log_align));
2835 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa));
2838 * If this is a repair I/O, and there's no self-healing involved --
2839 * that is, we're just resilvering what we expect to resilver --
2840 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2841 * This prevents spurious resilvering with nested replication.
2842 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2843 * A is out of date, we'll read from C+D, then use the data to
2844 * resilver A+B -- but we don't actually want to resilver B, just A.
2845 * The top-level mirror has no way to know this, so instead we just
2846 * discard unnecessary repairs as we work our way down the vdev tree.
2847 * The same logic applies to any form of nested replication:
2848 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2850 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2851 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2852 zio->io_txg != 0 && /* not a delegated i/o */
2853 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2854 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2855 zio_vdev_io_bypass(zio);
2856 return (ZIO_PIPELINE_CONTINUE);
2859 if (vd->vdev_ops->vdev_op_leaf) {
2860 switch (zio->io_type) {
2862 if (vdev_cache_read(zio))
2863 return (ZIO_PIPELINE_CONTINUE);
2865 case ZIO_TYPE_WRITE:
2867 if ((zio = vdev_queue_io(zio)) == NULL)
2868 return (ZIO_PIPELINE_STOP);
2870 if (!vdev_accessible(vd, zio)) {
2871 zio->io_error = SET_ERROR(ENXIO);
2873 return (ZIO_PIPELINE_STOP);
2878 * Note that we ignore repair writes for TRIM because they can
2879 * conflict with normal writes. This isn't an issue because, by
2880 * definition, we only repair blocks that aren't freed.
2882 if (zio->io_type == ZIO_TYPE_WRITE &&
2883 !(zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2884 !trim_map_write_start(zio))
2885 return (ZIO_PIPELINE_STOP);
2888 vd->vdev_ops->vdev_op_io_start(zio);
2889 return (ZIO_PIPELINE_STOP);
2893 zio_vdev_io_done(zio_t *zio)
2895 vdev_t *vd = zio->io_vd;
2896 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2897 boolean_t unexpected_error = B_FALSE;
2899 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2900 return (ZIO_PIPELINE_STOP);
2902 ASSERT(zio->io_type == ZIO_TYPE_READ ||
2903 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE);
2905 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2906 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE ||
2907 zio->io_type == ZIO_TYPE_FREE)) {
2909 if (zio->io_type == ZIO_TYPE_WRITE &&
2910 !(zio->io_flags & ZIO_FLAG_IO_REPAIR))
2911 trim_map_write_done(zio);
2913 vdev_queue_io_done(zio);
2915 if (zio->io_type == ZIO_TYPE_WRITE)
2916 vdev_cache_write(zio);
2918 if (zio_injection_enabled && zio->io_error == 0)
2919 zio->io_error = zio_handle_device_injection(vd,
2922 if (zio_injection_enabled && zio->io_error == 0)
2923 zio->io_error = zio_handle_label_injection(zio, EIO);
2925 if (zio->io_error) {
2926 if (zio->io_error == ENOTSUP &&
2927 zio->io_type == ZIO_TYPE_FREE) {
2928 /* Not all devices support TRIM. */
2929 } else if (!vdev_accessible(vd, zio)) {
2930 zio->io_error = SET_ERROR(ENXIO);
2932 unexpected_error = B_TRUE;
2937 ops->vdev_op_io_done(zio);
2939 if (unexpected_error)
2940 VERIFY(vdev_probe(vd, zio) == NULL);
2942 return (ZIO_PIPELINE_CONTINUE);
2946 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2947 * disk, and use that to finish the checksum ereport later.
2950 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2951 const void *good_buf)
2953 /* no processing needed */
2954 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2959 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2961 void *buf = zio_buf_alloc(zio->io_size);
2963 bcopy(zio->io_data, buf, zio->io_size);
2965 zcr->zcr_cbinfo = zio->io_size;
2966 zcr->zcr_cbdata = buf;
2967 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2968 zcr->zcr_free = zio_buf_free;
2972 zio_vdev_io_assess(zio_t *zio)
2974 vdev_t *vd = zio->io_vd;
2976 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2977 return (ZIO_PIPELINE_STOP);
2979 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2980 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2982 if (zio->io_vsd != NULL) {
2983 zio->io_vsd_ops->vsd_free(zio);
2987 if (zio_injection_enabled && zio->io_error == 0)
2988 zio->io_error = zio_handle_fault_injection(zio, EIO);
2990 if (zio->io_type == ZIO_TYPE_FREE &&
2991 zio->io_priority != ZIO_PRIORITY_NOW) {
2992 switch (zio->io_error) {
2994 ZIO_TRIM_STAT_INCR(bytes, zio->io_size);
2995 ZIO_TRIM_STAT_BUMP(success);
2998 ZIO_TRIM_STAT_BUMP(unsupported);
3001 ZIO_TRIM_STAT_BUMP(failed);
3007 * If the I/O failed, determine whether we should attempt to retry it.
3009 * On retry, we cut in line in the issue queue, since we don't want
3010 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3012 if (zio->io_error && vd == NULL &&
3013 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
3014 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
3015 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
3017 zio->io_flags |= ZIO_FLAG_IO_RETRY |
3018 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
3019 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
3020 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
3021 zio_requeue_io_start_cut_in_line);
3022 return (ZIO_PIPELINE_STOP);
3026 * If we got an error on a leaf device, convert it to ENXIO
3027 * if the device is not accessible at all.
3029 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3030 !vdev_accessible(vd, zio))
3031 zio->io_error = SET_ERROR(ENXIO);
3034 * If we can't write to an interior vdev (mirror or RAID-Z),
3035 * set vdev_cant_write so that we stop trying to allocate from it.
3037 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
3038 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
3039 vd->vdev_cant_write = B_TRUE;
3043 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3045 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3046 zio->io_physdone != NULL) {
3047 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
3048 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
3049 zio->io_physdone(zio->io_logical);
3052 return (ZIO_PIPELINE_CONTINUE);
3056 zio_vdev_io_reissue(zio_t *zio)
3058 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3059 ASSERT(zio->io_error == 0);
3061 zio->io_stage >>= 1;
3065 zio_vdev_io_redone(zio_t *zio)
3067 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
3069 zio->io_stage >>= 1;
3073 zio_vdev_io_bypass(zio_t *zio)
3075 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3076 ASSERT(zio->io_error == 0);
3078 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
3079 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
3083 * ==========================================================================
3084 * Generate and verify checksums
3085 * ==========================================================================
3088 zio_checksum_generate(zio_t *zio)
3090 blkptr_t *bp = zio->io_bp;
3091 enum zio_checksum checksum;
3095 * This is zio_write_phys().
3096 * We're either generating a label checksum, or none at all.
3098 checksum = zio->io_prop.zp_checksum;
3100 if (checksum == ZIO_CHECKSUM_OFF)
3101 return (ZIO_PIPELINE_CONTINUE);
3103 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
3105 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
3106 ASSERT(!IO_IS_ALLOCATING(zio));
3107 checksum = ZIO_CHECKSUM_GANG_HEADER;
3109 checksum = BP_GET_CHECKSUM(bp);
3113 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
3115 return (ZIO_PIPELINE_CONTINUE);
3119 zio_checksum_verify(zio_t *zio)
3121 zio_bad_cksum_t info;
3122 blkptr_t *bp = zio->io_bp;
3125 ASSERT(zio->io_vd != NULL);
3129 * This is zio_read_phys().
3130 * We're either verifying a label checksum, or nothing at all.
3132 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
3133 return (ZIO_PIPELINE_CONTINUE);
3135 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
3138 if ((error = zio_checksum_error(zio, &info)) != 0) {
3139 zio->io_error = error;
3140 if (error == ECKSUM &&
3141 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
3142 zfs_ereport_start_checksum(zio->io_spa,
3143 zio->io_vd, zio, zio->io_offset,
3144 zio->io_size, NULL, &info);
3148 return (ZIO_PIPELINE_CONTINUE);
3152 * Called by RAID-Z to ensure we don't compute the checksum twice.
3155 zio_checksum_verified(zio_t *zio)
3157 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
3161 * ==========================================================================
3162 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
3163 * An error of 0 indicates success. ENXIO indicates whole-device failure,
3164 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
3165 * indicate errors that are specific to one I/O, and most likely permanent.
3166 * Any other error is presumed to be worse because we weren't expecting it.
3167 * ==========================================================================
3170 zio_worst_error(int e1, int e2)
3172 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
3175 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
3176 if (e1 == zio_error_rank[r1])
3179 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
3180 if (e2 == zio_error_rank[r2])
3183 return (r1 > r2 ? e1 : e2);
3187 * ==========================================================================
3189 * ==========================================================================
3192 zio_ready(zio_t *zio)
3194 blkptr_t *bp = zio->io_bp;
3195 zio_t *pio, *pio_next;
3197 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
3198 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
3199 return (ZIO_PIPELINE_STOP);
3201 if (zio->io_ready) {
3202 ASSERT(IO_IS_ALLOCATING(zio));
3203 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
3204 (zio->io_flags & ZIO_FLAG_NOPWRITE));
3205 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
3210 if (bp != NULL && bp != &zio->io_bp_copy)
3211 zio->io_bp_copy = *bp;
3214 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3216 mutex_enter(&zio->io_lock);
3217 zio->io_state[ZIO_WAIT_READY] = 1;
3218 pio = zio_walk_parents(zio);
3219 mutex_exit(&zio->io_lock);
3222 * As we notify zio's parents, new parents could be added.
3223 * New parents go to the head of zio's io_parent_list, however,
3224 * so we will (correctly) not notify them. The remainder of zio's
3225 * io_parent_list, from 'pio_next' onward, cannot change because
3226 * all parents must wait for us to be done before they can be done.
3228 for (; pio != NULL; pio = pio_next) {
3229 pio_next = zio_walk_parents(zio);
3230 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
3233 if (zio->io_flags & ZIO_FLAG_NODATA) {
3234 if (BP_IS_GANG(bp)) {
3235 zio->io_flags &= ~ZIO_FLAG_NODATA;
3237 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
3238 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
3242 if (zio_injection_enabled &&
3243 zio->io_spa->spa_syncing_txg == zio->io_txg)
3244 zio_handle_ignored_writes(zio);
3246 return (ZIO_PIPELINE_CONTINUE);
3250 zio_done(zio_t *zio)
3252 spa_t *spa = zio->io_spa;
3253 zio_t *lio = zio->io_logical;
3254 blkptr_t *bp = zio->io_bp;
3255 vdev_t *vd = zio->io_vd;
3256 uint64_t psize = zio->io_size;
3257 zio_t *pio, *pio_next;
3260 * If our children haven't all completed,
3261 * wait for them and then repeat this pipeline stage.
3263 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
3264 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
3265 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
3266 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
3267 return (ZIO_PIPELINE_STOP);
3269 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3270 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3271 ASSERT(zio->io_children[c][w] == 0);
3273 if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
3274 ASSERT(bp->blk_pad[0] == 0);
3275 ASSERT(bp->blk_pad[1] == 0);
3276 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3277 (bp == zio_unique_parent(zio)->io_bp));
3278 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3279 zio->io_bp_override == NULL &&
3280 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3281 ASSERT(!BP_SHOULD_BYTESWAP(bp));
3282 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3283 ASSERT(BP_COUNT_GANG(bp) == 0 ||
3284 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3286 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3287 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3291 * If there were child vdev/gang/ddt errors, they apply to us now.
3293 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3294 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3295 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3298 * If the I/O on the transformed data was successful, generate any
3299 * checksum reports now while we still have the transformed data.
3301 if (zio->io_error == 0) {
3302 while (zio->io_cksum_report != NULL) {
3303 zio_cksum_report_t *zcr = zio->io_cksum_report;
3304 uint64_t align = zcr->zcr_align;
3305 uint64_t asize = P2ROUNDUP(psize, align);
3306 char *abuf = zio->io_data;
3308 if (asize != psize) {
3309 abuf = zio_buf_alloc(asize);
3310 bcopy(zio->io_data, abuf, psize);
3311 bzero(abuf + psize, asize - psize);
3314 zio->io_cksum_report = zcr->zcr_next;
3315 zcr->zcr_next = NULL;
3316 zcr->zcr_finish(zcr, abuf);
3317 zfs_ereport_free_checksum(zcr);
3320 zio_buf_free(abuf, asize);
3324 zio_pop_transforms(zio); /* note: may set zio->io_error */
3326 vdev_stat_update(zio, psize);
3328 if (zio->io_error) {
3330 * If this I/O is attached to a particular vdev,
3331 * generate an error message describing the I/O failure
3332 * at the block level. We ignore these errors if the
3333 * device is currently unavailable.
3335 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3336 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3338 if ((zio->io_error == EIO || !(zio->io_flags &
3339 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3342 * For logical I/O requests, tell the SPA to log the
3343 * error and generate a logical data ereport.
3345 spa_log_error(spa, zio);
3346 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3351 if (zio->io_error && zio == lio) {
3353 * Determine whether zio should be reexecuted. This will
3354 * propagate all the way to the root via zio_notify_parent().
3356 ASSERT(vd == NULL && bp != NULL);
3357 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3359 if (IO_IS_ALLOCATING(zio) &&
3360 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3361 if (zio->io_error != ENOSPC)
3362 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3364 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3367 if ((zio->io_type == ZIO_TYPE_READ ||
3368 zio->io_type == ZIO_TYPE_FREE) &&
3369 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3370 zio->io_error == ENXIO &&
3371 spa_load_state(spa) == SPA_LOAD_NONE &&
3372 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3373 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3375 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3376 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3379 * Here is a possibly good place to attempt to do
3380 * either combinatorial reconstruction or error correction
3381 * based on checksums. It also might be a good place
3382 * to send out preliminary ereports before we suspend
3388 * If there were logical child errors, they apply to us now.
3389 * We defer this until now to avoid conflating logical child
3390 * errors with errors that happened to the zio itself when
3391 * updating vdev stats and reporting FMA events above.
3393 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3395 if ((zio->io_error || zio->io_reexecute) &&
3396 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3397 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3398 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3400 zio_gang_tree_free(&zio->io_gang_tree);
3403 * Godfather I/Os should never suspend.
3405 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3406 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3407 zio->io_reexecute = 0;
3409 if (zio->io_reexecute) {
3411 * This is a logical I/O that wants to reexecute.
3413 * Reexecute is top-down. When an i/o fails, if it's not
3414 * the root, it simply notifies its parent and sticks around.
3415 * The parent, seeing that it still has children in zio_done(),
3416 * does the same. This percolates all the way up to the root.
3417 * The root i/o will reexecute or suspend the entire tree.
3419 * This approach ensures that zio_reexecute() honors
3420 * all the original i/o dependency relationships, e.g.
3421 * parents not executing until children are ready.
3423 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3425 zio->io_gang_leader = NULL;
3427 mutex_enter(&zio->io_lock);
3428 zio->io_state[ZIO_WAIT_DONE] = 1;
3429 mutex_exit(&zio->io_lock);
3432 * "The Godfather" I/O monitors its children but is
3433 * not a true parent to them. It will track them through
3434 * the pipeline but severs its ties whenever they get into
3435 * trouble (e.g. suspended). This allows "The Godfather"
3436 * I/O to return status without blocking.
3438 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3439 zio_link_t *zl = zio->io_walk_link;
3440 pio_next = zio_walk_parents(zio);
3442 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3443 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3444 zio_remove_child(pio, zio, zl);
3445 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3449 if ((pio = zio_unique_parent(zio)) != NULL) {
3451 * We're not a root i/o, so there's nothing to do
3452 * but notify our parent. Don't propagate errors
3453 * upward since we haven't permanently failed yet.
3455 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3456 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3457 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3458 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3460 * We'd fail again if we reexecuted now, so suspend
3461 * until conditions improve (e.g. device comes online).
3463 zio_suspend(spa, zio);
3466 * Reexecution is potentially a huge amount of work.
3467 * Hand it off to the otherwise-unused claim taskq.
3469 #if defined(illumos) || !defined(_KERNEL)
3470 ASSERT(zio->io_tqent.tqent_next == NULL);
3472 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
3474 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3475 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3478 return (ZIO_PIPELINE_STOP);
3481 ASSERT(zio->io_child_count == 0);
3482 ASSERT(zio->io_reexecute == 0);
3483 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3486 * Report any checksum errors, since the I/O is complete.
3488 while (zio->io_cksum_report != NULL) {
3489 zio_cksum_report_t *zcr = zio->io_cksum_report;
3490 zio->io_cksum_report = zcr->zcr_next;
3491 zcr->zcr_next = NULL;
3492 zcr->zcr_finish(zcr, NULL);
3493 zfs_ereport_free_checksum(zcr);
3497 * It is the responsibility of the done callback to ensure that this
3498 * particular zio is no longer discoverable for adoption, and as
3499 * such, cannot acquire any new parents.
3504 mutex_enter(&zio->io_lock);
3505 zio->io_state[ZIO_WAIT_DONE] = 1;
3506 mutex_exit(&zio->io_lock);
3508 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3509 zio_link_t *zl = zio->io_walk_link;
3510 pio_next = zio_walk_parents(zio);
3511 zio_remove_child(pio, zio, zl);
3512 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3515 if (zio->io_waiter != NULL) {
3516 mutex_enter(&zio->io_lock);
3517 zio->io_executor = NULL;
3518 cv_broadcast(&zio->io_cv);
3519 mutex_exit(&zio->io_lock);
3524 return (ZIO_PIPELINE_STOP);
3528 * ==========================================================================
3529 * I/O pipeline definition
3530 * ==========================================================================
3532 static zio_pipe_stage_t *zio_pipeline[] = {
3538 zio_checksum_generate,
3553 zio_checksum_verify,
3561 * Compare two zbookmark_phys_t's to see which we would reach first in a
3562 * pre-order traversal of the object tree.
3564 * This is simple in every case aside from the meta-dnode object. For all other
3565 * objects, we traverse them in order (object 1 before object 2, and so on).
3566 * However, all of these objects are traversed while traversing object 0, since
3567 * the data it points to is the list of objects. Thus, we need to convert to a
3568 * canonical representation so we can compare meta-dnode bookmarks to
3569 * non-meta-dnode bookmarks.
3571 * We do this by calculating "equivalents" for each field of the zbookmark.
3572 * zbookmarks outside of the meta-dnode use their own object and level, and
3573 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
3574 * blocks this bookmark refers to) by multiplying their blkid by their span
3575 * (the number of L0 blocks contained within one block at their level).
3576 * zbookmarks inside the meta-dnode calculate their object equivalent
3577 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
3578 * level + 1<<31 (any value larger than a level could ever be) for their level.
3579 * This causes them to always compare before a bookmark in their object
3580 * equivalent, compare appropriately to bookmarks in other objects, and to
3581 * compare appropriately to other bookmarks in the meta-dnode.
3584 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
3585 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
3588 * These variables represent the "equivalent" values for the zbookmark,
3589 * after converting zbookmarks inside the meta dnode to their
3590 * normal-object equivalents.
3592 uint64_t zb1obj, zb2obj;
3593 uint64_t zb1L0, zb2L0;
3594 uint64_t zb1level, zb2level;
3596 if (zb1->zb_object == zb2->zb_object &&
3597 zb1->zb_level == zb2->zb_level &&
3598 zb1->zb_blkid == zb2->zb_blkid)
3602 * BP_SPANB calculates the span in blocks.
3604 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
3605 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
3607 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3608 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
3610 zb1level = zb1->zb_level + COMPARE_META_LEVEL;
3612 zb1obj = zb1->zb_object;
3613 zb1level = zb1->zb_level;
3616 if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
3617 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
3619 zb2level = zb2->zb_level + COMPARE_META_LEVEL;
3621 zb2obj = zb2->zb_object;
3622 zb2level = zb2->zb_level;
3625 /* Now that we have a canonical representation, do the comparison. */
3626 if (zb1obj != zb2obj)
3627 return (zb1obj < zb2obj ? -1 : 1);
3628 else if (zb1L0 != zb2L0)
3629 return (zb1L0 < zb2L0 ? -1 : 1);
3630 else if (zb1level != zb2level)
3631 return (zb1level > zb2level ? -1 : 1);
3633 * This can (theoretically) happen if the bookmarks have the same object
3634 * and level, but different blkids, if the block sizes are not the same.
3635 * There is presently no way to change the indirect block sizes
3641 * This function checks the following: given that last_block is the place that
3642 * our traversal stopped last time, does that guarantee that we've visited
3643 * every node under subtree_root? Therefore, we can't just use the raw output
3644 * of zbookmark_compare. We have to pass in a modified version of
3645 * subtree_root; by incrementing the block id, and then checking whether
3646 * last_block is before or equal to that, we can tell whether or not having
3647 * visited last_block implies that all of subtree_root's children have been
3651 zbookmark_subtree_completed(const dnode_phys_t *dnp,
3652 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
3654 zbookmark_phys_t mod_zb = *subtree_root;
3656 ASSERT(last_block->zb_level == 0);
3658 /* The objset_phys_t isn't before anything. */
3663 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
3664 * data block size in sectors, because that variable is only used if
3665 * the bookmark refers to a block in the meta-dnode. Since we don't
3666 * know without examining it what object it refers to, and there's no
3667 * harm in passing in this value in other cases, we always pass it in.
3669 * We pass in 0 for the indirect block size shift because zb2 must be
3670 * level 0. The indirect block size is only used to calculate the span
3671 * of the bookmark, but since the bookmark must be level 0, the span is
3672 * always 1, so the math works out.
3674 * If you make changes to how the zbookmark_compare code works, be sure
3675 * to make sure that this code still works afterwards.
3677 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
3678 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,