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 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, use_uma, CTLFLAG_RDTUN, &zio_use_uma, 0,
53 "Use uma(9) for ZIO allocations");
54 static int zio_exclude_metadata = 0;
55 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, exclude_metadata, CTLFLAG_RDTUN, &zio_exclude_metadata, 0,
56 "Exclude metadata buffers from dumps as well");
58 zio_trim_stats_t zio_trim_stats = {
59 { "bytes", KSTAT_DATA_UINT64,
60 "Number of bytes successfully TRIMmed" },
61 { "success", KSTAT_DATA_UINT64,
62 "Number of successful TRIM requests" },
63 { "unsupported", KSTAT_DATA_UINT64,
64 "Number of TRIM requests that failed because TRIM is not supported" },
65 { "failed", KSTAT_DATA_UINT64,
66 "Number of TRIM requests that failed for reasons other than not supported" },
69 static kstat_t *zio_trim_ksp;
72 * ==========================================================================
73 * I/O type descriptions
74 * ==========================================================================
76 const char *zio_type_name[ZIO_TYPES] = {
77 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
82 * ==========================================================================
84 * ==========================================================================
86 kmem_cache_t *zio_cache;
87 kmem_cache_t *zio_link_cache;
88 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
89 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
92 extern vmem_t *zio_alloc_arena;
95 #define ZIO_PIPELINE_CONTINUE 0x100
96 #define ZIO_PIPELINE_STOP 0x101
98 #define BP_SPANB(indblkshift, level) \
99 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
100 #define COMPARE_META_LEVEL 0x80000000ul
102 * The following actions directly effect the spa's sync-to-convergence logic.
103 * The values below define the sync pass when we start performing the action.
104 * Care should be taken when changing these values as they directly impact
105 * spa_sync() performance. Tuning these values may introduce subtle performance
106 * pathologies and should only be done in the context of performance analysis.
107 * These tunables will eventually be removed and replaced with #defines once
108 * enough analysis has been done to determine optimal values.
110 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
111 * regular blocks are not deferred.
113 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
114 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_deferred_free, CTLFLAG_RDTUN,
115 &zfs_sync_pass_deferred_free, 0, "defer frees starting in this pass");
116 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
117 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_dont_compress, CTLFLAG_RDTUN,
118 &zfs_sync_pass_dont_compress, 0, "don't compress starting in this pass");
119 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
120 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_rewrite, CTLFLAG_RDTUN,
121 &zfs_sync_pass_rewrite, 0, "rewrite new bps starting in this pass");
124 * An allocating zio is one that either currently has the DVA allocate
125 * stage set or will have it later in its lifetime.
127 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
129 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
133 int zio_buf_debug_limit = 16384;
135 int zio_buf_debug_limit = 0;
143 zio_cache = kmem_cache_create("zio_cache",
144 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
145 zio_link_cache = kmem_cache_create("zio_link_cache",
146 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
151 * For small buffers, we want a cache for each multiple of
152 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
153 * for each quarter-power of 2.
155 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
156 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
159 int cflags = zio_exclude_metadata ? KMC_NODEBUG : 0;
167 * If we are using watchpoints, put each buffer on its own page,
168 * to eliminate the performance overhead of trapping to the
169 * kernel when modifying a non-watched buffer that shares the
170 * page with a watched buffer.
172 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
176 if (size <= 4 * SPA_MINBLOCKSIZE) {
177 align = SPA_MINBLOCKSIZE;
178 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
179 align = MIN(p2 >> 2, PAGESIZE);
184 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
185 zio_buf_cache[c] = kmem_cache_create(name, size,
186 align, NULL, NULL, NULL, NULL, NULL, cflags);
189 * Since zio_data bufs do not appear in crash dumps, we
190 * pass KMC_NOTOUCH so that no allocator metadata is
191 * stored with the buffers.
193 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
194 zio_data_buf_cache[c] = kmem_cache_create(name, size,
195 align, NULL, NULL, NULL, NULL, NULL,
196 cflags | KMC_NOTOUCH | KMC_NODEBUG);
201 ASSERT(zio_buf_cache[c] != NULL);
202 if (zio_buf_cache[c - 1] == NULL)
203 zio_buf_cache[c - 1] = zio_buf_cache[c];
205 ASSERT(zio_data_buf_cache[c] != NULL);
206 if (zio_data_buf_cache[c - 1] == NULL)
207 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
213 zio_trim_ksp = kstat_create("zfs", 0, "zio_trim", "misc",
215 sizeof(zio_trim_stats) / sizeof(kstat_named_t),
218 if (zio_trim_ksp != NULL) {
219 zio_trim_ksp->ks_data = &zio_trim_stats;
220 kstat_install(zio_trim_ksp);
228 kmem_cache_t *last_cache = NULL;
229 kmem_cache_t *last_data_cache = NULL;
231 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
232 if (zio_buf_cache[c] != last_cache) {
233 last_cache = zio_buf_cache[c];
234 kmem_cache_destroy(zio_buf_cache[c]);
236 zio_buf_cache[c] = NULL;
238 if (zio_data_buf_cache[c] != last_data_cache) {
239 last_data_cache = zio_data_buf_cache[c];
240 kmem_cache_destroy(zio_data_buf_cache[c]);
242 zio_data_buf_cache[c] = NULL;
245 kmem_cache_destroy(zio_link_cache);
246 kmem_cache_destroy(zio_cache);
250 if (zio_trim_ksp != NULL) {
251 kstat_delete(zio_trim_ksp);
257 * ==========================================================================
258 * Allocate and free I/O buffers
259 * ==========================================================================
263 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
264 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
265 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
266 * excess / transient data in-core during a crashdump.
269 zio_buf_alloc(size_t size)
271 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
272 int flags = zio_exclude_metadata ? KM_NODEBUG : 0;
274 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
277 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
279 return (kmem_alloc(size, KM_SLEEP|flags));
283 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
284 * crashdump if the kernel panics. This exists so that we will limit the amount
285 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
286 * of kernel heap dumped to disk when the kernel panics)
289 zio_data_buf_alloc(size_t size)
291 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
293 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
296 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
298 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG));
302 zio_buf_free(void *buf, size_t size)
304 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
306 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
309 kmem_cache_free(zio_buf_cache[c], buf);
311 kmem_free(buf, size);
315 zio_data_buf_free(void *buf, size_t size)
317 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
319 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
322 kmem_cache_free(zio_data_buf_cache[c], buf);
324 kmem_free(buf, size);
328 * ==========================================================================
329 * Push and pop I/O transform buffers
330 * ==========================================================================
333 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
334 zio_transform_func_t *transform)
336 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
338 zt->zt_orig_data = zio->io_data;
339 zt->zt_orig_size = zio->io_size;
340 zt->zt_bufsize = bufsize;
341 zt->zt_transform = transform;
343 zt->zt_next = zio->io_transform_stack;
344 zio->io_transform_stack = zt;
351 zio_pop_transforms(zio_t *zio)
355 while ((zt = zio->io_transform_stack) != NULL) {
356 if (zt->zt_transform != NULL)
357 zt->zt_transform(zio,
358 zt->zt_orig_data, zt->zt_orig_size);
360 if (zt->zt_bufsize != 0)
361 zio_buf_free(zio->io_data, zt->zt_bufsize);
363 zio->io_data = zt->zt_orig_data;
364 zio->io_size = zt->zt_orig_size;
365 zio->io_transform_stack = zt->zt_next;
367 kmem_free(zt, sizeof (zio_transform_t));
372 * ==========================================================================
373 * I/O transform callbacks for subblocks and decompression
374 * ==========================================================================
377 zio_subblock(zio_t *zio, void *data, uint64_t size)
379 ASSERT(zio->io_size > size);
381 if (zio->io_type == ZIO_TYPE_READ)
382 bcopy(zio->io_data, data, size);
386 zio_decompress(zio_t *zio, void *data, uint64_t size)
388 if (zio->io_error == 0 &&
389 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
390 zio->io_data, data, zio->io_size, size) != 0)
391 zio->io_error = SET_ERROR(EIO);
395 * ==========================================================================
396 * I/O parent/child relationships and pipeline interlocks
397 * ==========================================================================
400 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
401 * continue calling these functions until they return NULL.
402 * Otherwise, the next caller will pick up the list walk in
403 * some indeterminate state. (Otherwise every caller would
404 * have to pass in a cookie to keep the state represented by
405 * io_walk_link, which gets annoying.)
408 zio_walk_parents(zio_t *cio)
410 zio_link_t *zl = cio->io_walk_link;
411 list_t *pl = &cio->io_parent_list;
413 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
414 cio->io_walk_link = zl;
419 ASSERT(zl->zl_child == cio);
420 return (zl->zl_parent);
424 zio_walk_children(zio_t *pio)
426 zio_link_t *zl = pio->io_walk_link;
427 list_t *cl = &pio->io_child_list;
429 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
430 pio->io_walk_link = zl;
435 ASSERT(zl->zl_parent == pio);
436 return (zl->zl_child);
440 zio_unique_parent(zio_t *cio)
442 zio_t *pio = zio_walk_parents(cio);
444 VERIFY(zio_walk_parents(cio) == NULL);
449 zio_add_child(zio_t *pio, zio_t *cio)
451 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
454 * Logical I/Os can have logical, gang, or vdev children.
455 * Gang I/Os can have gang or vdev children.
456 * Vdev I/Os can only have vdev children.
457 * The following ASSERT captures all of these constraints.
459 ASSERT(cio->io_child_type <= pio->io_child_type);
464 mutex_enter(&cio->io_lock);
465 mutex_enter(&pio->io_lock);
467 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
469 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
470 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
472 list_insert_head(&pio->io_child_list, zl);
473 list_insert_head(&cio->io_parent_list, zl);
475 pio->io_child_count++;
476 cio->io_parent_count++;
478 mutex_exit(&pio->io_lock);
479 mutex_exit(&cio->io_lock);
483 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
485 ASSERT(zl->zl_parent == pio);
486 ASSERT(zl->zl_child == cio);
488 mutex_enter(&cio->io_lock);
489 mutex_enter(&pio->io_lock);
491 list_remove(&pio->io_child_list, zl);
492 list_remove(&cio->io_parent_list, zl);
494 pio->io_child_count--;
495 cio->io_parent_count--;
497 mutex_exit(&pio->io_lock);
498 mutex_exit(&cio->io_lock);
500 kmem_cache_free(zio_link_cache, zl);
504 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
506 uint64_t *countp = &zio->io_children[child][wait];
507 boolean_t waiting = B_FALSE;
509 mutex_enter(&zio->io_lock);
510 ASSERT(zio->io_stall == NULL);
513 zio->io_stall = countp;
516 mutex_exit(&zio->io_lock);
522 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
524 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
525 int *errorp = &pio->io_child_error[zio->io_child_type];
527 mutex_enter(&pio->io_lock);
528 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
529 *errorp = zio_worst_error(*errorp, zio->io_error);
530 pio->io_reexecute |= zio->io_reexecute;
531 ASSERT3U(*countp, >, 0);
535 if (*countp == 0 && pio->io_stall == countp) {
536 pio->io_stall = NULL;
537 mutex_exit(&pio->io_lock);
540 mutex_exit(&pio->io_lock);
545 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
547 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
548 zio->io_error = zio->io_child_error[c];
552 * ==========================================================================
553 * Create the various types of I/O (read, write, free, etc)
554 * ==========================================================================
557 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
558 void *data, uint64_t size, zio_done_func_t *done, void *private,
559 zio_type_t type, zio_priority_t priority, enum zio_flag flags,
560 vdev_t *vd, uint64_t offset, const zbookmark_phys_t *zb,
561 enum zio_stage stage, enum zio_stage pipeline)
565 ASSERT3U(type == ZIO_TYPE_FREE || size, <=, SPA_MAXBLOCKSIZE);
566 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
567 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
569 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
570 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
571 ASSERT(vd || stage == ZIO_STAGE_OPEN);
573 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
574 bzero(zio, sizeof (zio_t));
576 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
577 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
579 list_create(&zio->io_parent_list, sizeof (zio_link_t),
580 offsetof(zio_link_t, zl_parent_node));
581 list_create(&zio->io_child_list, sizeof (zio_link_t),
582 offsetof(zio_link_t, zl_child_node));
585 zio->io_child_type = ZIO_CHILD_VDEV;
586 else if (flags & ZIO_FLAG_GANG_CHILD)
587 zio->io_child_type = ZIO_CHILD_GANG;
588 else if (flags & ZIO_FLAG_DDT_CHILD)
589 zio->io_child_type = ZIO_CHILD_DDT;
591 zio->io_child_type = ZIO_CHILD_LOGICAL;
594 zio->io_bp = (blkptr_t *)bp;
595 zio->io_bp_copy = *bp;
596 zio->io_bp_orig = *bp;
597 if (type != ZIO_TYPE_WRITE ||
598 zio->io_child_type == ZIO_CHILD_DDT)
599 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
600 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
601 zio->io_logical = zio;
602 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
603 pipeline |= ZIO_GANG_STAGES;
609 zio->io_private = private;
611 zio->io_priority = priority;
613 zio->io_offset = offset;
614 zio->io_orig_data = zio->io_data = data;
615 zio->io_orig_size = zio->io_size = size;
616 zio->io_orig_flags = zio->io_flags = flags;
617 zio->io_orig_stage = zio->io_stage = stage;
618 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
620 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
621 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
624 zio->io_bookmark = *zb;
627 if (zio->io_logical == NULL)
628 zio->io_logical = pio->io_logical;
629 if (zio->io_child_type == ZIO_CHILD_GANG)
630 zio->io_gang_leader = pio->io_gang_leader;
631 zio_add_child(pio, zio);
638 zio_destroy(zio_t *zio)
640 list_destroy(&zio->io_parent_list);
641 list_destroy(&zio->io_child_list);
642 mutex_destroy(&zio->io_lock);
643 cv_destroy(&zio->io_cv);
644 kmem_cache_free(zio_cache, zio);
648 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
649 void *private, enum zio_flag flags)
653 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
654 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
655 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
661 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
663 return (zio_null(NULL, spa, NULL, done, private, flags));
667 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
669 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
670 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
671 bp, (longlong_t)BP_GET_TYPE(bp));
673 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
674 BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
675 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
676 bp, (longlong_t)BP_GET_CHECKSUM(bp));
678 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
679 BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
680 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
681 bp, (longlong_t)BP_GET_COMPRESS(bp));
683 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
684 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
685 bp, (longlong_t)BP_GET_LSIZE(bp));
687 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
688 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
689 bp, (longlong_t)BP_GET_PSIZE(bp));
692 if (BP_IS_EMBEDDED(bp)) {
693 if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
694 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
695 bp, (longlong_t)BPE_GET_ETYPE(bp));
700 * Pool-specific checks.
702 * Note: it would be nice to verify that the blk_birth and
703 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
704 * allows the birth time of log blocks (and dmu_sync()-ed blocks
705 * that are in the log) to be arbitrarily large.
707 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
708 uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
709 if (vdevid >= spa->spa_root_vdev->vdev_children) {
710 zfs_panic_recover("blkptr at %p DVA %u has invalid "
712 bp, i, (longlong_t)vdevid);
715 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
717 zfs_panic_recover("blkptr at %p DVA %u has invalid "
719 bp, i, (longlong_t)vdevid);
722 if (vd->vdev_ops == &vdev_hole_ops) {
723 zfs_panic_recover("blkptr at %p DVA %u has hole "
725 bp, i, (longlong_t)vdevid);
728 if (vd->vdev_ops == &vdev_missing_ops) {
730 * "missing" vdevs are valid during import, but we
731 * don't have their detailed info (e.g. asize), so
732 * we can't perform any more checks on them.
736 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
737 uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
739 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
740 if (offset + asize > vd->vdev_asize) {
741 zfs_panic_recover("blkptr at %p DVA %u has invalid "
743 bp, i, (longlong_t)offset);
749 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
750 void *data, uint64_t size, zio_done_func_t *done, void *private,
751 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
755 zfs_blkptr_verify(spa, bp);
757 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
758 data, size, done, private,
759 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
760 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
761 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
767 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
768 void *data, uint64_t size, const zio_prop_t *zp,
769 zio_done_func_t *ready, zio_done_func_t *children_ready,
770 zio_done_func_t *physdone, zio_done_func_t *done,
771 void *private, zio_priority_t priority, enum zio_flag flags,
772 const zbookmark_phys_t *zb)
776 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
777 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
778 zp->zp_compress >= ZIO_COMPRESS_OFF &&
779 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
780 DMU_OT_IS_VALID(zp->zp_type) &&
783 zp->zp_copies <= spa_max_replication(spa));
785 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
786 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
787 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
788 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
790 zio->io_ready = ready;
791 zio->io_children_ready = children_ready;
792 zio->io_physdone = physdone;
796 * Data can be NULL if we are going to call zio_write_override() to
797 * provide the already-allocated BP. But we may need the data to
798 * verify a dedup hit (if requested). In this case, don't try to
799 * dedup (just take the already-allocated BP verbatim).
801 if (data == NULL && zio->io_prop.zp_dedup_verify) {
802 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
809 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
810 uint64_t size, zio_done_func_t *done, void *private,
811 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
815 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
816 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
817 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
823 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
825 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
826 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
827 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
828 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
831 * We must reset the io_prop to match the values that existed
832 * when the bp was first written by dmu_sync() keeping in mind
833 * that nopwrite and dedup are mutually exclusive.
835 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
836 zio->io_prop.zp_nopwrite = nopwrite;
837 zio->io_prop.zp_copies = copies;
838 zio->io_bp_override = bp;
842 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
846 * The check for EMBEDDED is a performance optimization. We
847 * process the free here (by ignoring it) rather than
848 * putting it on the list and then processing it in zio_free_sync().
850 if (BP_IS_EMBEDDED(bp))
852 metaslab_check_free(spa, bp);
855 * Frees that are for the currently-syncing txg, are not going to be
856 * deferred, and which will not need to do a read (i.e. not GANG or
857 * DEDUP), can be processed immediately. Otherwise, put them on the
858 * in-memory list for later processing.
860 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
861 txg != spa->spa_syncing_txg ||
862 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
863 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
865 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp,
866 BP_GET_PSIZE(bp), 0)));
871 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
872 uint64_t size, enum zio_flag flags)
875 enum zio_stage stage = ZIO_FREE_PIPELINE;
877 ASSERT(!BP_IS_HOLE(bp));
878 ASSERT(spa_syncing_txg(spa) == txg);
879 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
881 if (BP_IS_EMBEDDED(bp))
882 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
884 metaslab_check_free(spa, bp);
887 if (zfs_trim_enabled)
888 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START |
889 ZIO_STAGE_VDEV_IO_ASSESS;
891 * GANG and DEDUP blocks can induce a read (for the gang block header,
892 * or the DDT), so issue them asynchronously so that this thread is
895 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
896 stage |= ZIO_STAGE_ISSUE_ASYNC;
898 flags |= ZIO_FLAG_DONT_QUEUE;
900 zio = zio_create(pio, spa, txg, bp, NULL, size,
901 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags,
902 NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
908 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
909 zio_done_func_t *done, void *private, enum zio_flag flags)
913 dprintf_bp(bp, "claiming in txg %llu", txg);
915 if (BP_IS_EMBEDDED(bp))
916 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
919 * A claim is an allocation of a specific block. Claims are needed
920 * to support immediate writes in the intent log. The issue is that
921 * immediate writes contain committed data, but in a txg that was
922 * *not* committed. Upon opening the pool after an unclean shutdown,
923 * the intent log claims all blocks that contain immediate write data
924 * so that the SPA knows they're in use.
926 * All claims *must* be resolved in the first txg -- before the SPA
927 * starts allocating blocks -- so that nothing is allocated twice.
928 * If txg == 0 we just verify that the block is claimable.
930 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
931 ASSERT(txg == spa_first_txg(spa) || txg == 0);
932 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
934 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
935 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
936 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
942 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset,
943 uint64_t size, zio_done_func_t *done, void *private,
944 zio_priority_t priority, enum zio_flag flags)
949 if (vd->vdev_children == 0) {
950 zio = zio_create(pio, spa, 0, NULL, NULL, size, done, private,
951 ZIO_TYPE_IOCTL, priority, flags, vd, offset, NULL,
952 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
956 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
958 for (c = 0; c < vd->vdev_children; c++)
959 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
960 offset, size, done, private, priority, flags));
967 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
968 void *data, int checksum, zio_done_func_t *done, void *private,
969 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
973 ASSERT(vd->vdev_children == 0);
974 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
975 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
976 ASSERT3U(offset + size, <=, vd->vdev_psize);
978 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
979 ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
980 NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
982 zio->io_prop.zp_checksum = checksum;
988 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
989 void *data, int checksum, zio_done_func_t *done, void *private,
990 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
994 ASSERT(vd->vdev_children == 0);
995 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
996 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
997 ASSERT3U(offset + size, <=, vd->vdev_psize);
999 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
1000 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset,
1001 NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
1003 zio->io_prop.zp_checksum = checksum;
1005 if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
1007 * zec checksums are necessarily destructive -- they modify
1008 * the end of the write buffer to hold the verifier/checksum.
1009 * Therefore, we must make a local copy in case the data is
1010 * being written to multiple places in parallel.
1012 void *wbuf = zio_buf_alloc(size);
1013 bcopy(data, wbuf, size);
1014 zio_push_transform(zio, wbuf, size, size, NULL);
1021 * Create a child I/O to do some work for us.
1024 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1025 void *data, uint64_t size, int type, zio_priority_t priority,
1026 enum zio_flag flags, zio_done_func_t *done, void *private)
1028 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1031 ASSERT(vd->vdev_parent ==
1032 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
1034 if (type == ZIO_TYPE_READ && bp != NULL) {
1036 * If we have the bp, then the child should perform the
1037 * checksum and the parent need not. This pushes error
1038 * detection as close to the leaves as possible and
1039 * eliminates redundant checksums in the interior nodes.
1041 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1042 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1045 /* Not all IO types require vdev io done stage e.g. free */
1046 if (!(pio->io_pipeline & ZIO_STAGE_VDEV_IO_DONE))
1047 pipeline &= ~ZIO_STAGE_VDEV_IO_DONE;
1049 if (vd->vdev_children == 0)
1050 offset += VDEV_LABEL_START_SIZE;
1052 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
1055 * If we've decided to do a repair, the write is not speculative --
1056 * even if the original read was.
1058 if (flags & ZIO_FLAG_IO_REPAIR)
1059 flags &= ~ZIO_FLAG_SPECULATIVE;
1061 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
1062 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1063 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1065 zio->io_physdone = pio->io_physdone;
1066 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1067 zio->io_logical->io_phys_children++;
1073 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
1074 int type, zio_priority_t priority, enum zio_flag flags,
1075 zio_done_func_t *done, void *private)
1079 ASSERT(vd->vdev_ops->vdev_op_leaf);
1081 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1082 data, size, done, private, type, priority,
1083 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1085 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1091 zio_flush(zio_t *zio, vdev_t *vd)
1093 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0,
1094 NULL, NULL, ZIO_PRIORITY_NOW,
1095 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1099 zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size)
1102 ASSERT(vd->vdev_ops->vdev_op_leaf);
1104 return (zio_create(zio, spa, 0, NULL, NULL, size, NULL, NULL,
1105 ZIO_TYPE_FREE, ZIO_PRIORITY_TRIM, ZIO_FLAG_DONT_AGGREGATE |
1106 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY,
1107 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PHYS_PIPELINE));
1111 zio_shrink(zio_t *zio, uint64_t size)
1113 ASSERT(zio->io_executor == NULL);
1114 ASSERT(zio->io_orig_size == zio->io_size);
1115 ASSERT(size <= zio->io_size);
1118 * We don't shrink for raidz because of problems with the
1119 * reconstruction when reading back less than the block size.
1120 * Note, BP_IS_RAIDZ() assumes no compression.
1122 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1123 if (!BP_IS_RAIDZ(zio->io_bp))
1124 zio->io_orig_size = zio->io_size = size;
1128 * ==========================================================================
1129 * Prepare to read and write logical blocks
1130 * ==========================================================================
1134 zio_read_bp_init(zio_t *zio)
1136 blkptr_t *bp = zio->io_bp;
1138 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1139 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1140 !(zio->io_flags & ZIO_FLAG_RAW)) {
1142 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1143 void *cbuf = zio_buf_alloc(psize);
1145 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
1148 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1149 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1150 decode_embedded_bp_compressed(bp, zio->io_data);
1152 ASSERT(!BP_IS_EMBEDDED(bp));
1155 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1156 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1158 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1159 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1161 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1162 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1164 return (ZIO_PIPELINE_CONTINUE);
1168 zio_write_bp_init(zio_t *zio)
1170 spa_t *spa = zio->io_spa;
1171 zio_prop_t *zp = &zio->io_prop;
1172 enum zio_compress compress = zp->zp_compress;
1173 blkptr_t *bp = zio->io_bp;
1174 uint64_t lsize = zio->io_size;
1175 uint64_t psize = lsize;
1179 * If our children haven't all reached the ready stage,
1180 * wait for them and then repeat this pipeline stage.
1182 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1183 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1184 return (ZIO_PIPELINE_STOP);
1186 if (!IO_IS_ALLOCATING(zio))
1187 return (ZIO_PIPELINE_CONTINUE);
1189 if (zio->io_children_ready != NULL) {
1191 * Now that all our children are ready, run the callback
1192 * associated with this zio in case it wants to modify the
1193 * data to be written.
1195 ASSERT3U(zp->zp_level, >, 0);
1196 zio->io_children_ready(zio);
1199 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1201 if (zio->io_bp_override) {
1202 ASSERT(bp->blk_birth != zio->io_txg);
1203 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1205 *bp = *zio->io_bp_override;
1206 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1208 if (BP_IS_EMBEDDED(bp))
1209 return (ZIO_PIPELINE_CONTINUE);
1212 * If we've been overridden and nopwrite is set then
1213 * set the flag accordingly to indicate that a nopwrite
1214 * has already occurred.
1216 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1217 ASSERT(!zp->zp_dedup);
1218 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1219 return (ZIO_PIPELINE_CONTINUE);
1222 ASSERT(!zp->zp_nopwrite);
1224 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1225 return (ZIO_PIPELINE_CONTINUE);
1227 ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags &
1228 ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify);
1230 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1231 BP_SET_DEDUP(bp, 1);
1232 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1233 return (ZIO_PIPELINE_CONTINUE);
1235 zio->io_bp_override = NULL;
1239 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1241 * We're rewriting an existing block, which means we're
1242 * working on behalf of spa_sync(). For spa_sync() to
1243 * converge, it must eventually be the case that we don't
1244 * have to allocate new blocks. But compression changes
1245 * the blocksize, which forces a reallocate, and makes
1246 * convergence take longer. Therefore, after the first
1247 * few passes, stop compressing to ensure convergence.
1249 pass = spa_sync_pass(spa);
1251 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1252 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1253 ASSERT(!BP_GET_DEDUP(bp));
1255 if (pass >= zfs_sync_pass_dont_compress)
1256 compress = ZIO_COMPRESS_OFF;
1258 /* Make sure someone doesn't change their mind on overwrites */
1259 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1260 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1263 if (compress != ZIO_COMPRESS_OFF) {
1264 void *cbuf = zio_buf_alloc(lsize);
1265 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1266 if (psize == 0 || psize == lsize) {
1267 compress = ZIO_COMPRESS_OFF;
1268 zio_buf_free(cbuf, lsize);
1269 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1270 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1271 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1272 encode_embedded_bp_compressed(bp,
1273 cbuf, compress, lsize, psize);
1274 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1275 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1276 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1277 zio_buf_free(cbuf, lsize);
1278 bp->blk_birth = zio->io_txg;
1279 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1280 ASSERT(spa_feature_is_active(spa,
1281 SPA_FEATURE_EMBEDDED_DATA));
1282 return (ZIO_PIPELINE_CONTINUE);
1285 * Round up compressed size up to the ashift
1286 * of the smallest-ashift device, and zero the tail.
1287 * This ensures that the compressed size of the BP
1288 * (and thus compressratio property) are correct,
1289 * in that we charge for the padding used to fill out
1292 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
1293 size_t rounded = (size_t)P2ROUNDUP(psize,
1294 1ULL << spa->spa_min_ashift);
1295 if (rounded >= lsize) {
1296 compress = ZIO_COMPRESS_OFF;
1297 zio_buf_free(cbuf, lsize);
1300 bzero((char *)cbuf + psize, rounded - psize);
1302 zio_push_transform(zio, cbuf,
1303 psize, lsize, NULL);
1309 * The final pass of spa_sync() must be all rewrites, but the first
1310 * few passes offer a trade-off: allocating blocks defers convergence,
1311 * but newly allocated blocks are sequential, so they can be written
1312 * to disk faster. Therefore, we allow the first few passes of
1313 * spa_sync() to allocate new blocks, but force rewrites after that.
1314 * There should only be a handful of blocks after pass 1 in any case.
1316 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1317 BP_GET_PSIZE(bp) == psize &&
1318 pass >= zfs_sync_pass_rewrite) {
1320 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1321 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1322 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1325 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1329 if (zio->io_bp_orig.blk_birth != 0 &&
1330 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1331 BP_SET_LSIZE(bp, lsize);
1332 BP_SET_TYPE(bp, zp->zp_type);
1333 BP_SET_LEVEL(bp, zp->zp_level);
1334 BP_SET_BIRTH(bp, zio->io_txg, 0);
1336 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1338 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1339 BP_SET_LSIZE(bp, lsize);
1340 BP_SET_TYPE(bp, zp->zp_type);
1341 BP_SET_LEVEL(bp, zp->zp_level);
1342 BP_SET_PSIZE(bp, psize);
1343 BP_SET_COMPRESS(bp, compress);
1344 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1345 BP_SET_DEDUP(bp, zp->zp_dedup);
1346 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1348 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1349 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1350 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1352 if (zp->zp_nopwrite) {
1353 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1354 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1355 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1359 return (ZIO_PIPELINE_CONTINUE);
1363 zio_free_bp_init(zio_t *zio)
1365 blkptr_t *bp = zio->io_bp;
1367 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1368 if (BP_GET_DEDUP(bp))
1369 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1372 return (ZIO_PIPELINE_CONTINUE);
1376 * ==========================================================================
1377 * Execute the I/O pipeline
1378 * ==========================================================================
1382 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1384 spa_t *spa = zio->io_spa;
1385 zio_type_t t = zio->io_type;
1386 int flags = (cutinline ? TQ_FRONT : 0);
1388 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1391 * If we're a config writer or a probe, the normal issue and
1392 * interrupt threads may all be blocked waiting for the config lock.
1393 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1395 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1399 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1401 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1405 * If this is a high priority I/O, then use the high priority taskq if
1408 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1409 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1412 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1415 * NB: We are assuming that the zio can only be dispatched
1416 * to a single taskq at a time. It would be a grievous error
1417 * to dispatch the zio to another taskq at the same time.
1419 #if defined(illumos) || !defined(_KERNEL)
1420 ASSERT(zio->io_tqent.tqent_next == NULL);
1422 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
1424 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1425 flags, &zio->io_tqent);
1429 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1431 kthread_t *executor = zio->io_executor;
1432 spa_t *spa = zio->io_spa;
1434 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1435 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1437 for (i = 0; i < tqs->stqs_count; i++) {
1438 if (taskq_member(tqs->stqs_taskq[i], executor))
1447 zio_issue_async(zio_t *zio)
1449 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1451 return (ZIO_PIPELINE_STOP);
1455 zio_interrupt(zio_t *zio)
1457 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1461 zio_delay_interrupt(zio_t *zio)
1464 * The timeout_generic() function isn't defined in userspace, so
1465 * rather than trying to implement the function, the zio delay
1466 * functionality has been disabled for userspace builds.
1471 * If io_target_timestamp is zero, then no delay has been registered
1472 * for this IO, thus jump to the end of this function and "skip" the
1473 * delay; issuing it directly to the zio layer.
1475 if (zio->io_target_timestamp != 0) {
1476 hrtime_t now = gethrtime();
1478 if (now >= zio->io_target_timestamp) {
1480 * This IO has already taken longer than the target
1481 * delay to complete, so we don't want to delay it
1482 * any longer; we "miss" the delay and issue it
1483 * directly to the zio layer. This is likely due to
1484 * the target latency being set to a value less than
1485 * the underlying hardware can satisfy (e.g. delay
1486 * set to 1ms, but the disks take 10ms to complete an
1490 DTRACE_PROBE2(zio__delay__miss, zio_t *, zio,
1495 hrtime_t diff = zio->io_target_timestamp - now;
1497 DTRACE_PROBE3(zio__delay__hit, zio_t *, zio,
1498 hrtime_t, now, hrtime_t, diff);
1500 (void) timeout_generic(CALLOUT_NORMAL,
1501 (void (*)(void *))zio_interrupt, zio, diff, 1, 0);
1508 DTRACE_PROBE1(zio__delay__skip, zio_t *, zio);
1513 * Execute the I/O pipeline until one of the following occurs:
1515 * (1) the I/O completes
1516 * (2) the pipeline stalls waiting for dependent child I/Os
1517 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1518 * (4) the I/O is delegated by vdev-level caching or aggregation
1519 * (5) the I/O is deferred due to vdev-level queueing
1520 * (6) the I/O is handed off to another thread.
1522 * In all cases, the pipeline stops whenever there's no CPU work; it never
1523 * burns a thread in cv_wait().
1525 * There's no locking on io_stage because there's no legitimate way
1526 * for multiple threads to be attempting to process the same I/O.
1528 static zio_pipe_stage_t *zio_pipeline[];
1531 zio_execute(zio_t *zio)
1533 zio->io_executor = curthread;
1535 while (zio->io_stage < ZIO_STAGE_DONE) {
1536 enum zio_stage pipeline = zio->io_pipeline;
1537 enum zio_stage stage = zio->io_stage;
1540 ASSERT(!MUTEX_HELD(&zio->io_lock));
1541 ASSERT(ISP2(stage));
1542 ASSERT(zio->io_stall == NULL);
1546 } while ((stage & pipeline) == 0);
1548 ASSERT(stage <= ZIO_STAGE_DONE);
1551 * If we are in interrupt context and this pipeline stage
1552 * will grab a config lock that is held across I/O,
1553 * or may wait for an I/O that needs an interrupt thread
1554 * to complete, issue async to avoid deadlock.
1556 * For VDEV_IO_START, we cut in line so that the io will
1557 * be sent to disk promptly.
1559 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1560 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1561 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1562 zio_requeue_io_start_cut_in_line : B_FALSE;
1563 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1567 zio->io_stage = stage;
1568 rv = zio_pipeline[highbit64(stage) - 1](zio);
1570 if (rv == ZIO_PIPELINE_STOP)
1573 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1578 * ==========================================================================
1579 * Initiate I/O, either sync or async
1580 * ==========================================================================
1583 zio_wait(zio_t *zio)
1587 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1588 ASSERT(zio->io_executor == NULL);
1590 zio->io_waiter = curthread;
1594 mutex_enter(&zio->io_lock);
1595 while (zio->io_executor != NULL)
1596 cv_wait(&zio->io_cv, &zio->io_lock);
1597 mutex_exit(&zio->io_lock);
1599 error = zio->io_error;
1606 zio_nowait(zio_t *zio)
1608 ASSERT(zio->io_executor == NULL);
1610 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1611 zio_unique_parent(zio) == NULL) {
1613 * This is a logical async I/O with no parent to wait for it.
1614 * We add it to the spa_async_root_zio "Godfather" I/O which
1615 * will ensure they complete prior to unloading the pool.
1617 spa_t *spa = zio->io_spa;
1619 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1626 * ==========================================================================
1627 * Reexecute or suspend/resume failed I/O
1628 * ==========================================================================
1632 zio_reexecute(zio_t *pio)
1634 zio_t *cio, *cio_next;
1636 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1637 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1638 ASSERT(pio->io_gang_leader == NULL);
1639 ASSERT(pio->io_gang_tree == NULL);
1641 pio->io_flags = pio->io_orig_flags;
1642 pio->io_stage = pio->io_orig_stage;
1643 pio->io_pipeline = pio->io_orig_pipeline;
1644 pio->io_reexecute = 0;
1645 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1647 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1648 pio->io_state[w] = 0;
1649 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1650 pio->io_child_error[c] = 0;
1652 if (IO_IS_ALLOCATING(pio))
1653 BP_ZERO(pio->io_bp);
1656 * As we reexecute pio's children, new children could be created.
1657 * New children go to the head of pio's io_child_list, however,
1658 * so we will (correctly) not reexecute them. The key is that
1659 * the remainder of pio's io_child_list, from 'cio_next' onward,
1660 * cannot be affected by any side effects of reexecuting 'cio'.
1662 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1663 cio_next = zio_walk_children(pio);
1664 mutex_enter(&pio->io_lock);
1665 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1666 pio->io_children[cio->io_child_type][w]++;
1667 mutex_exit(&pio->io_lock);
1672 * Now that all children have been reexecuted, execute the parent.
1673 * We don't reexecute "The Godfather" I/O here as it's the
1674 * responsibility of the caller to wait on him.
1676 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1681 zio_suspend(spa_t *spa, zio_t *zio)
1683 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1684 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1685 "failure and the failure mode property for this pool "
1686 "is set to panic.", spa_name(spa));
1688 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1690 mutex_enter(&spa->spa_suspend_lock);
1692 if (spa->spa_suspend_zio_root == NULL)
1693 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1694 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1695 ZIO_FLAG_GODFATHER);
1697 spa->spa_suspended = B_TRUE;
1700 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1701 ASSERT(zio != spa->spa_suspend_zio_root);
1702 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1703 ASSERT(zio_unique_parent(zio) == NULL);
1704 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1705 zio_add_child(spa->spa_suspend_zio_root, zio);
1708 mutex_exit(&spa->spa_suspend_lock);
1712 zio_resume(spa_t *spa)
1717 * Reexecute all previously suspended i/o.
1719 mutex_enter(&spa->spa_suspend_lock);
1720 spa->spa_suspended = B_FALSE;
1721 cv_broadcast(&spa->spa_suspend_cv);
1722 pio = spa->spa_suspend_zio_root;
1723 spa->spa_suspend_zio_root = NULL;
1724 mutex_exit(&spa->spa_suspend_lock);
1730 return (zio_wait(pio));
1734 zio_resume_wait(spa_t *spa)
1736 mutex_enter(&spa->spa_suspend_lock);
1737 while (spa_suspended(spa))
1738 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1739 mutex_exit(&spa->spa_suspend_lock);
1743 * ==========================================================================
1746 * A gang block is a collection of small blocks that looks to the DMU
1747 * like one large block. When zio_dva_allocate() cannot find a block
1748 * of the requested size, due to either severe fragmentation or the pool
1749 * being nearly full, it calls zio_write_gang_block() to construct the
1750 * block from smaller fragments.
1752 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1753 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1754 * an indirect block: it's an array of block pointers. It consumes
1755 * only one sector and hence is allocatable regardless of fragmentation.
1756 * The gang header's bps point to its gang members, which hold the data.
1758 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1759 * as the verifier to ensure uniqueness of the SHA256 checksum.
1760 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1761 * not the gang header. This ensures that data block signatures (needed for
1762 * deduplication) are independent of how the block is physically stored.
1764 * Gang blocks can be nested: a gang member may itself be a gang block.
1765 * Thus every gang block is a tree in which root and all interior nodes are
1766 * gang headers, and the leaves are normal blocks that contain user data.
1767 * The root of the gang tree is called the gang leader.
1769 * To perform any operation (read, rewrite, free, claim) on a gang block,
1770 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1771 * in the io_gang_tree field of the original logical i/o by recursively
1772 * reading the gang leader and all gang headers below it. This yields
1773 * an in-core tree containing the contents of every gang header and the
1774 * bps for every constituent of the gang block.
1776 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1777 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1778 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1779 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1780 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1781 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1782 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1783 * of the gang header plus zio_checksum_compute() of the data to update the
1784 * gang header's blk_cksum as described above.
1786 * The two-phase assemble/issue model solves the problem of partial failure --
1787 * what if you'd freed part of a gang block but then couldn't read the
1788 * gang header for another part? Assembling the entire gang tree first
1789 * ensures that all the necessary gang header I/O has succeeded before
1790 * starting the actual work of free, claim, or write. Once the gang tree
1791 * is assembled, free and claim are in-memory operations that cannot fail.
1793 * In the event that a gang write fails, zio_dva_unallocate() walks the
1794 * gang tree to immediately free (i.e. insert back into the space map)
1795 * everything we've allocated. This ensures that we don't get ENOSPC
1796 * errors during repeated suspend/resume cycles due to a flaky device.
1798 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1799 * the gang tree, we won't modify the block, so we can safely defer the free
1800 * (knowing that the block is still intact). If we *can* assemble the gang
1801 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1802 * each constituent bp and we can allocate a new block on the next sync pass.
1804 * In all cases, the gang tree allows complete recovery from partial failure.
1805 * ==========================================================================
1809 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1814 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1815 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1816 &pio->io_bookmark));
1820 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1825 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1826 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1827 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1829 * As we rewrite each gang header, the pipeline will compute
1830 * a new gang block header checksum for it; but no one will
1831 * compute a new data checksum, so we do that here. The one
1832 * exception is the gang leader: the pipeline already computed
1833 * its data checksum because that stage precedes gang assembly.
1834 * (Presently, nothing actually uses interior data checksums;
1835 * this is just good hygiene.)
1837 if (gn != pio->io_gang_leader->io_gang_tree) {
1838 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1839 data, BP_GET_PSIZE(bp));
1842 * If we are here to damage data for testing purposes,
1843 * leave the GBH alone so that we can detect the damage.
1845 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1846 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1848 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1849 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1850 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1858 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1860 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1861 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp),
1862 ZIO_GANG_CHILD_FLAGS(pio)));
1867 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1869 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1870 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1873 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1882 static void zio_gang_tree_assemble_done(zio_t *zio);
1884 static zio_gang_node_t *
1885 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1887 zio_gang_node_t *gn;
1889 ASSERT(*gnpp == NULL);
1891 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1892 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1899 zio_gang_node_free(zio_gang_node_t **gnpp)
1901 zio_gang_node_t *gn = *gnpp;
1903 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1904 ASSERT(gn->gn_child[g] == NULL);
1906 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1907 kmem_free(gn, sizeof (*gn));
1912 zio_gang_tree_free(zio_gang_node_t **gnpp)
1914 zio_gang_node_t *gn = *gnpp;
1919 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1920 zio_gang_tree_free(&gn->gn_child[g]);
1922 zio_gang_node_free(gnpp);
1926 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1928 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1930 ASSERT(gio->io_gang_leader == gio);
1931 ASSERT(BP_IS_GANG(bp));
1933 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1934 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1935 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1939 zio_gang_tree_assemble_done(zio_t *zio)
1941 zio_t *gio = zio->io_gang_leader;
1942 zio_gang_node_t *gn = zio->io_private;
1943 blkptr_t *bp = zio->io_bp;
1945 ASSERT(gio == zio_unique_parent(zio));
1946 ASSERT(zio->io_child_count == 0);
1951 if (BP_SHOULD_BYTESWAP(bp))
1952 byteswap_uint64_array(zio->io_data, zio->io_size);
1954 ASSERT(zio->io_data == gn->gn_gbh);
1955 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1956 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1958 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1959 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1960 if (!BP_IS_GANG(gbp))
1962 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1967 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1969 zio_t *gio = pio->io_gang_leader;
1972 ASSERT(BP_IS_GANG(bp) == !!gn);
1973 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1974 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1977 * If you're a gang header, your data is in gn->gn_gbh.
1978 * If you're a gang member, your data is in 'data' and gn == NULL.
1980 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1983 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1985 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1986 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1987 if (BP_IS_HOLE(gbp))
1989 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1990 data = (char *)data + BP_GET_PSIZE(gbp);
1994 if (gn == gio->io_gang_tree && gio->io_data != NULL)
1995 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
2002 zio_gang_assemble(zio_t *zio)
2004 blkptr_t *bp = zio->io_bp;
2006 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
2007 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2009 zio->io_gang_leader = zio;
2011 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
2013 return (ZIO_PIPELINE_CONTINUE);
2017 zio_gang_issue(zio_t *zio)
2019 blkptr_t *bp = zio->io_bp;
2021 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
2022 return (ZIO_PIPELINE_STOP);
2024 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
2025 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2027 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
2028 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
2030 zio_gang_tree_free(&zio->io_gang_tree);
2032 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2034 return (ZIO_PIPELINE_CONTINUE);
2038 zio_write_gang_member_ready(zio_t *zio)
2040 zio_t *pio = zio_unique_parent(zio);
2041 zio_t *gio = zio->io_gang_leader;
2042 dva_t *cdva = zio->io_bp->blk_dva;
2043 dva_t *pdva = pio->io_bp->blk_dva;
2046 if (BP_IS_HOLE(zio->io_bp))
2049 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
2051 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
2052 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
2053 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2054 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
2055 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
2057 mutex_enter(&pio->io_lock);
2058 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
2059 ASSERT(DVA_GET_GANG(&pdva[d]));
2060 asize = DVA_GET_ASIZE(&pdva[d]);
2061 asize += DVA_GET_ASIZE(&cdva[d]);
2062 DVA_SET_ASIZE(&pdva[d], asize);
2064 mutex_exit(&pio->io_lock);
2068 zio_write_gang_block(zio_t *pio)
2070 spa_t *spa = pio->io_spa;
2071 blkptr_t *bp = pio->io_bp;
2072 zio_t *gio = pio->io_gang_leader;
2074 zio_gang_node_t *gn, **gnpp;
2075 zio_gbh_phys_t *gbh;
2076 uint64_t txg = pio->io_txg;
2077 uint64_t resid = pio->io_size;
2079 int copies = gio->io_prop.zp_copies;
2080 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
2084 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
2085 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
2086 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
2088 pio->io_error = error;
2089 return (ZIO_PIPELINE_CONTINUE);
2093 gnpp = &gio->io_gang_tree;
2095 gnpp = pio->io_private;
2096 ASSERT(pio->io_ready == zio_write_gang_member_ready);
2099 gn = zio_gang_node_alloc(gnpp);
2101 bzero(gbh, SPA_GANGBLOCKSIZE);
2104 * Create the gang header.
2106 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
2107 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2110 * Create and nowait the gang children.
2112 for (int g = 0; resid != 0; resid -= lsize, g++) {
2113 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2115 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2117 zp.zp_checksum = gio->io_prop.zp_checksum;
2118 zp.zp_compress = ZIO_COMPRESS_OFF;
2119 zp.zp_type = DMU_OT_NONE;
2121 zp.zp_copies = gio->io_prop.zp_copies;
2122 zp.zp_dedup = B_FALSE;
2123 zp.zp_dedup_verify = B_FALSE;
2124 zp.zp_nopwrite = B_FALSE;
2126 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2127 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
2128 zio_write_gang_member_ready, NULL, NULL, NULL,
2129 &gn->gn_child[g], pio->io_priority,
2130 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark));
2134 * Set pio's pipeline to just wait for zio to finish.
2136 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2140 return (ZIO_PIPELINE_CONTINUE);
2144 * The zio_nop_write stage in the pipeline determines if allocating a
2145 * new bp is necessary. The nopwrite feature can handle writes in
2146 * either syncing or open context (i.e. zil writes) and as a result is
2147 * mutually exclusive with dedup.
2149 * By leveraging a cryptographically secure checksum, such as SHA256, we
2150 * can compare the checksums of the new data and the old to determine if
2151 * allocating a new block is required. Note that our requirements for
2152 * cryptographic strength are fairly weak: there can't be any accidental
2153 * hash collisions, but we don't need to be secure against intentional
2154 * (malicious) collisions. To trigger a nopwrite, you have to be able
2155 * to write the file to begin with, and triggering an incorrect (hash
2156 * collision) nopwrite is no worse than simply writing to the file.
2157 * That said, there are no known attacks against the checksum algorithms
2158 * used for nopwrite, assuming that the salt and the checksums
2159 * themselves remain secret.
2162 zio_nop_write(zio_t *zio)
2164 blkptr_t *bp = zio->io_bp;
2165 blkptr_t *bp_orig = &zio->io_bp_orig;
2166 zio_prop_t *zp = &zio->io_prop;
2168 ASSERT(BP_GET_LEVEL(bp) == 0);
2169 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2170 ASSERT(zp->zp_nopwrite);
2171 ASSERT(!zp->zp_dedup);
2172 ASSERT(zio->io_bp_override == NULL);
2173 ASSERT(IO_IS_ALLOCATING(zio));
2176 * Check to see if the original bp and the new bp have matching
2177 * characteristics (i.e. same checksum, compression algorithms, etc).
2178 * If they don't then just continue with the pipeline which will
2179 * allocate a new bp.
2181 if (BP_IS_HOLE(bp_orig) ||
2182 !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
2183 ZCHECKSUM_FLAG_NOPWRITE) ||
2184 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2185 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2186 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2187 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2188 return (ZIO_PIPELINE_CONTINUE);
2191 * If the checksums match then reset the pipeline so that we
2192 * avoid allocating a new bp and issuing any I/O.
2194 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2195 ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
2196 ZCHECKSUM_FLAG_NOPWRITE);
2197 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2198 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2199 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2200 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2201 sizeof (uint64_t)) == 0);
2204 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2205 zio->io_flags |= ZIO_FLAG_NOPWRITE;
2208 return (ZIO_PIPELINE_CONTINUE);
2212 * ==========================================================================
2214 * ==========================================================================
2217 zio_ddt_child_read_done(zio_t *zio)
2219 blkptr_t *bp = zio->io_bp;
2220 ddt_entry_t *dde = zio->io_private;
2222 zio_t *pio = zio_unique_parent(zio);
2224 mutex_enter(&pio->io_lock);
2225 ddp = ddt_phys_select(dde, bp);
2226 if (zio->io_error == 0)
2227 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2228 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
2229 dde->dde_repair_data = zio->io_data;
2231 zio_buf_free(zio->io_data, zio->io_size);
2232 mutex_exit(&pio->io_lock);
2236 zio_ddt_read_start(zio_t *zio)
2238 blkptr_t *bp = zio->io_bp;
2240 ASSERT(BP_GET_DEDUP(bp));
2241 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2242 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2244 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2245 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2246 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2247 ddt_phys_t *ddp = dde->dde_phys;
2248 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2251 ASSERT(zio->io_vsd == NULL);
2254 if (ddp_self == NULL)
2255 return (ZIO_PIPELINE_CONTINUE);
2257 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2258 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2260 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2262 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2263 zio_buf_alloc(zio->io_size), zio->io_size,
2264 zio_ddt_child_read_done, dde, zio->io_priority,
2265 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
2266 &zio->io_bookmark));
2268 return (ZIO_PIPELINE_CONTINUE);
2271 zio_nowait(zio_read(zio, zio->io_spa, bp,
2272 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
2273 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2275 return (ZIO_PIPELINE_CONTINUE);
2279 zio_ddt_read_done(zio_t *zio)
2281 blkptr_t *bp = zio->io_bp;
2283 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2284 return (ZIO_PIPELINE_STOP);
2286 ASSERT(BP_GET_DEDUP(bp));
2287 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2288 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2290 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2291 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2292 ddt_entry_t *dde = zio->io_vsd;
2294 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2295 return (ZIO_PIPELINE_CONTINUE);
2298 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2299 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2300 return (ZIO_PIPELINE_STOP);
2302 if (dde->dde_repair_data != NULL) {
2303 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2304 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2306 ddt_repair_done(ddt, dde);
2310 ASSERT(zio->io_vsd == NULL);
2312 return (ZIO_PIPELINE_CONTINUE);
2316 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2318 spa_t *spa = zio->io_spa;
2321 * Note: we compare the original data, not the transformed data,
2322 * because when zio->io_bp is an override bp, we will not have
2323 * pushed the I/O transforms. That's an important optimization
2324 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2326 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2327 zio_t *lio = dde->dde_lead_zio[p];
2330 return (lio->io_orig_size != zio->io_orig_size ||
2331 bcmp(zio->io_orig_data, lio->io_orig_data,
2332 zio->io_orig_size) != 0);
2336 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2337 ddt_phys_t *ddp = &dde->dde_phys[p];
2339 if (ddp->ddp_phys_birth != 0) {
2340 arc_buf_t *abuf = NULL;
2341 arc_flags_t aflags = ARC_FLAG_WAIT;
2342 blkptr_t blk = *zio->io_bp;
2345 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2349 error = arc_read(NULL, spa, &blk,
2350 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2351 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2352 &aflags, &zio->io_bookmark);
2355 if (arc_buf_size(abuf) != zio->io_orig_size ||
2356 bcmp(abuf->b_data, zio->io_orig_data,
2357 zio->io_orig_size) != 0)
2358 error = SET_ERROR(EEXIST);
2359 VERIFY(arc_buf_remove_ref(abuf, &abuf));
2363 return (error != 0);
2371 zio_ddt_child_write_ready(zio_t *zio)
2373 int p = zio->io_prop.zp_copies;
2374 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2375 ddt_entry_t *dde = zio->io_private;
2376 ddt_phys_t *ddp = &dde->dde_phys[p];
2384 ASSERT(dde->dde_lead_zio[p] == zio);
2386 ddt_phys_fill(ddp, zio->io_bp);
2388 while ((pio = zio_walk_parents(zio)) != NULL)
2389 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2395 zio_ddt_child_write_done(zio_t *zio)
2397 int p = zio->io_prop.zp_copies;
2398 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2399 ddt_entry_t *dde = zio->io_private;
2400 ddt_phys_t *ddp = &dde->dde_phys[p];
2404 ASSERT(ddp->ddp_refcnt == 0);
2405 ASSERT(dde->dde_lead_zio[p] == zio);
2406 dde->dde_lead_zio[p] = NULL;
2408 if (zio->io_error == 0) {
2409 while (zio_walk_parents(zio) != NULL)
2410 ddt_phys_addref(ddp);
2412 ddt_phys_clear(ddp);
2419 zio_ddt_ditto_write_done(zio_t *zio)
2421 int p = DDT_PHYS_DITTO;
2422 zio_prop_t *zp = &zio->io_prop;
2423 blkptr_t *bp = zio->io_bp;
2424 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2425 ddt_entry_t *dde = zio->io_private;
2426 ddt_phys_t *ddp = &dde->dde_phys[p];
2427 ddt_key_t *ddk = &dde->dde_key;
2431 ASSERT(ddp->ddp_refcnt == 0);
2432 ASSERT(dde->dde_lead_zio[p] == zio);
2433 dde->dde_lead_zio[p] = NULL;
2435 if (zio->io_error == 0) {
2436 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2437 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2438 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2439 if (ddp->ddp_phys_birth != 0)
2440 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2441 ddt_phys_fill(ddp, bp);
2448 zio_ddt_write(zio_t *zio)
2450 spa_t *spa = zio->io_spa;
2451 blkptr_t *bp = zio->io_bp;
2452 uint64_t txg = zio->io_txg;
2453 zio_prop_t *zp = &zio->io_prop;
2454 int p = zp->zp_copies;
2458 ddt_t *ddt = ddt_select(spa, bp);
2462 ASSERT(BP_GET_DEDUP(bp));
2463 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2464 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2467 dde = ddt_lookup(ddt, bp, B_TRUE);
2468 ddp = &dde->dde_phys[p];
2470 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2472 * If we're using a weak checksum, upgrade to a strong checksum
2473 * and try again. If we're already using a strong checksum,
2474 * we can't resolve it, so just convert to an ordinary write.
2475 * (And automatically e-mail a paper to Nature?)
2477 if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
2478 ZCHECKSUM_FLAG_DEDUP)) {
2479 zp->zp_checksum = spa_dedup_checksum(spa);
2480 zio_pop_transforms(zio);
2481 zio->io_stage = ZIO_STAGE_OPEN;
2484 zp->zp_dedup = B_FALSE;
2486 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2488 return (ZIO_PIPELINE_CONTINUE);
2491 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2492 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2494 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2495 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2496 zio_prop_t czp = *zp;
2498 czp.zp_copies = ditto_copies;
2501 * If we arrived here with an override bp, we won't have run
2502 * the transform stack, so we won't have the data we need to
2503 * generate a child i/o. So, toss the override bp and restart.
2504 * This is safe, because using the override bp is just an
2505 * optimization; and it's rare, so the cost doesn't matter.
2507 if (zio->io_bp_override) {
2508 zio_pop_transforms(zio);
2509 zio->io_stage = ZIO_STAGE_OPEN;
2510 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2511 zio->io_bp_override = NULL;
2514 return (ZIO_PIPELINE_CONTINUE);
2517 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2518 zio->io_orig_size, &czp, NULL, NULL,
2519 NULL, zio_ddt_ditto_write_done, dde, zio->io_priority,
2520 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2522 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2523 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2526 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2527 if (ddp->ddp_phys_birth != 0)
2528 ddt_bp_fill(ddp, bp, txg);
2529 if (dde->dde_lead_zio[p] != NULL)
2530 zio_add_child(zio, dde->dde_lead_zio[p]);
2532 ddt_phys_addref(ddp);
2533 } else if (zio->io_bp_override) {
2534 ASSERT(bp->blk_birth == txg);
2535 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2536 ddt_phys_fill(ddp, bp);
2537 ddt_phys_addref(ddp);
2539 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2540 zio->io_orig_size, zp,
2541 zio_ddt_child_write_ready, NULL, NULL,
2542 zio_ddt_child_write_done, dde, zio->io_priority,
2543 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2545 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2546 dde->dde_lead_zio[p] = cio;
2556 return (ZIO_PIPELINE_CONTINUE);
2559 ddt_entry_t *freedde; /* for debugging */
2562 zio_ddt_free(zio_t *zio)
2564 spa_t *spa = zio->io_spa;
2565 blkptr_t *bp = zio->io_bp;
2566 ddt_t *ddt = ddt_select(spa, bp);
2570 ASSERT(BP_GET_DEDUP(bp));
2571 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2574 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2575 ddp = ddt_phys_select(dde, bp);
2576 ddt_phys_decref(ddp);
2579 return (ZIO_PIPELINE_CONTINUE);
2583 * ==========================================================================
2584 * Allocate and free blocks
2585 * ==========================================================================
2588 zio_dva_allocate(zio_t *zio)
2590 spa_t *spa = zio->io_spa;
2591 metaslab_class_t *mc = spa_normal_class(spa);
2592 blkptr_t *bp = zio->io_bp;
2596 if (zio->io_gang_leader == NULL) {
2597 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2598 zio->io_gang_leader = zio;
2601 ASSERT(BP_IS_HOLE(bp));
2602 ASSERT0(BP_GET_NDVAS(bp));
2603 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2604 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2605 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2608 * The dump device does not support gang blocks so allocation on
2609 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2610 * the "fast" gang feature.
2612 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2613 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2614 METASLAB_GANG_CHILD : 0;
2615 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2616 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2619 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2620 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2622 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2623 return (zio_write_gang_block(zio));
2624 zio->io_error = error;
2627 return (ZIO_PIPELINE_CONTINUE);
2631 zio_dva_free(zio_t *zio)
2633 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2635 return (ZIO_PIPELINE_CONTINUE);
2639 zio_dva_claim(zio_t *zio)
2643 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2645 zio->io_error = error;
2647 return (ZIO_PIPELINE_CONTINUE);
2651 * Undo an allocation. This is used by zio_done() when an I/O fails
2652 * and we want to give back the block we just allocated.
2653 * This handles both normal blocks and gang blocks.
2656 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2658 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2659 ASSERT(zio->io_bp_override == NULL);
2661 if (!BP_IS_HOLE(bp))
2662 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2665 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2666 zio_dva_unallocate(zio, gn->gn_child[g],
2667 &gn->gn_gbh->zg_blkptr[g]);
2673 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2676 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2677 uint64_t size, boolean_t use_slog)
2681 ASSERT(txg > spa_syncing_txg(spa));
2684 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2685 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2686 * when allocating them.
2689 error = metaslab_alloc(spa, spa_log_class(spa), size,
2690 new_bp, 1, txg, old_bp,
2691 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2695 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2696 new_bp, 1, txg, old_bp,
2697 METASLAB_HINTBP_AVOID);
2701 BP_SET_LSIZE(new_bp, size);
2702 BP_SET_PSIZE(new_bp, size);
2703 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2704 BP_SET_CHECKSUM(new_bp,
2705 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2706 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2707 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2708 BP_SET_LEVEL(new_bp, 0);
2709 BP_SET_DEDUP(new_bp, 0);
2710 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2717 * Free an intent log block.
2720 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2722 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2723 ASSERT(!BP_IS_GANG(bp));
2725 zio_free(spa, txg, bp);
2729 * ==========================================================================
2730 * Read, write and delete to physical devices
2731 * ==========================================================================
2736 * Issue an I/O to the underlying vdev. Typically the issue pipeline
2737 * stops after this stage and will resume upon I/O completion.
2738 * However, there are instances where the vdev layer may need to
2739 * continue the pipeline when an I/O was not issued. Since the I/O
2740 * that was sent to the vdev layer might be different than the one
2741 * currently active in the pipeline (see vdev_queue_io()), we explicitly
2742 * force the underlying vdev layers to call either zio_execute() or
2743 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
2746 zio_vdev_io_start(zio_t *zio)
2748 vdev_t *vd = zio->io_vd;
2750 spa_t *spa = zio->io_spa;
2753 ASSERT(zio->io_error == 0);
2754 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2757 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2758 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2761 * The mirror_ops handle multiple DVAs in a single BP.
2763 vdev_mirror_ops.vdev_op_io_start(zio);
2764 return (ZIO_PIPELINE_STOP);
2767 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE &&
2768 zio->io_priority == ZIO_PRIORITY_NOW) {
2769 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg);
2770 return (ZIO_PIPELINE_CONTINUE);
2774 * We keep track of time-sensitive I/Os so that the scan thread
2775 * can quickly react to certain workloads. In particular, we care
2776 * about non-scrubbing, top-level reads and writes with the following
2778 * - synchronous writes of user data to non-slog devices
2779 * - any reads of user data
2780 * When these conditions are met, adjust the timestamp of spa_last_io
2781 * which allows the scan thread to adjust its workload accordingly.
2783 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2784 vd == vd->vdev_top && !vd->vdev_islog &&
2785 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2786 zio->io_txg != spa_syncing_txg(spa)) {
2787 uint64_t old = spa->spa_last_io;
2788 uint64_t new = ddi_get_lbolt64();
2790 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2793 align = 1ULL << vd->vdev_top->vdev_ashift;
2795 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
2796 P2PHASE(zio->io_size, align) != 0) {
2797 /* Transform logical writes to be a full physical block size. */
2798 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2800 if (zio->io_type == ZIO_TYPE_READ ||
2801 zio->io_type == ZIO_TYPE_WRITE)
2802 abuf = zio_buf_alloc(asize);
2803 ASSERT(vd == vd->vdev_top);
2804 if (zio->io_type == ZIO_TYPE_WRITE) {
2805 bcopy(zio->io_data, abuf, zio->io_size);
2806 bzero(abuf + zio->io_size, asize - zio->io_size);
2808 zio_push_transform(zio, abuf, asize, abuf ? asize : 0,
2813 * If this is not a physical io, make sure that it is properly aligned
2814 * before proceeding.
2816 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
2817 ASSERT0(P2PHASE(zio->io_offset, align));
2818 ASSERT0(P2PHASE(zio->io_size, align));
2821 * For the physical io we allow alignment
2822 * to a logical block size.
2824 uint64_t log_align =
2825 1ULL << vd->vdev_top->vdev_logical_ashift;
2826 ASSERT0(P2PHASE(zio->io_offset, log_align));
2827 ASSERT0(P2PHASE(zio->io_size, log_align));
2830 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa));
2833 * If this is a repair I/O, and there's no self-healing involved --
2834 * that is, we're just resilvering what we expect to resilver --
2835 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2836 * This prevents spurious resilvering with nested replication.
2837 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2838 * A is out of date, we'll read from C+D, then use the data to
2839 * resilver A+B -- but we don't actually want to resilver B, just A.
2840 * The top-level mirror has no way to know this, so instead we just
2841 * discard unnecessary repairs as we work our way down the vdev tree.
2842 * The same logic applies to any form of nested replication:
2843 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2845 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2846 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2847 zio->io_txg != 0 && /* not a delegated i/o */
2848 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2849 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2850 zio_vdev_io_bypass(zio);
2851 return (ZIO_PIPELINE_CONTINUE);
2854 if (vd->vdev_ops->vdev_op_leaf) {
2855 switch (zio->io_type) {
2857 if (vdev_cache_read(zio))
2858 return (ZIO_PIPELINE_CONTINUE);
2860 case ZIO_TYPE_WRITE:
2862 if ((zio = vdev_queue_io(zio)) == NULL)
2863 return (ZIO_PIPELINE_STOP);
2865 if (!vdev_accessible(vd, zio)) {
2866 zio->io_error = SET_ERROR(ENXIO);
2868 return (ZIO_PIPELINE_STOP);
2873 * Note that we ignore repair writes for TRIM because they can
2874 * conflict with normal writes. This isn't an issue because, by
2875 * definition, we only repair blocks that aren't freed.
2877 if (zio->io_type == ZIO_TYPE_WRITE &&
2878 !(zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2879 !trim_map_write_start(zio))
2880 return (ZIO_PIPELINE_STOP);
2883 vd->vdev_ops->vdev_op_io_start(zio);
2884 return (ZIO_PIPELINE_STOP);
2888 zio_vdev_io_done(zio_t *zio)
2890 vdev_t *vd = zio->io_vd;
2891 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2892 boolean_t unexpected_error = B_FALSE;
2894 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2895 return (ZIO_PIPELINE_STOP);
2897 ASSERT(zio->io_type == ZIO_TYPE_READ ||
2898 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE);
2900 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2901 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE ||
2902 zio->io_type == ZIO_TYPE_FREE)) {
2904 if (zio->io_type == ZIO_TYPE_WRITE &&
2905 !(zio->io_flags & ZIO_FLAG_IO_REPAIR))
2906 trim_map_write_done(zio);
2908 vdev_queue_io_done(zio);
2910 if (zio->io_type == ZIO_TYPE_WRITE)
2911 vdev_cache_write(zio);
2913 if (zio_injection_enabled && zio->io_error == 0)
2914 zio->io_error = zio_handle_device_injection(vd,
2917 if (zio_injection_enabled && zio->io_error == 0)
2918 zio->io_error = zio_handle_label_injection(zio, EIO);
2920 if (zio->io_error) {
2921 if (zio->io_error == ENOTSUP &&
2922 zio->io_type == ZIO_TYPE_FREE) {
2923 /* Not all devices support TRIM. */
2924 } else if (!vdev_accessible(vd, zio)) {
2925 zio->io_error = SET_ERROR(ENXIO);
2927 unexpected_error = B_TRUE;
2932 ops->vdev_op_io_done(zio);
2934 if (unexpected_error)
2935 VERIFY(vdev_probe(vd, zio) == NULL);
2937 return (ZIO_PIPELINE_CONTINUE);
2941 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2942 * disk, and use that to finish the checksum ereport later.
2945 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2946 const void *good_buf)
2948 /* no processing needed */
2949 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2954 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2956 void *buf = zio_buf_alloc(zio->io_size);
2958 bcopy(zio->io_data, buf, zio->io_size);
2960 zcr->zcr_cbinfo = zio->io_size;
2961 zcr->zcr_cbdata = buf;
2962 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2963 zcr->zcr_free = zio_buf_free;
2967 zio_vdev_io_assess(zio_t *zio)
2969 vdev_t *vd = zio->io_vd;
2971 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2972 return (ZIO_PIPELINE_STOP);
2974 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2975 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2977 if (zio->io_vsd != NULL) {
2978 zio->io_vsd_ops->vsd_free(zio);
2982 if (zio_injection_enabled && zio->io_error == 0)
2983 zio->io_error = zio_handle_fault_injection(zio, EIO);
2985 if (zio->io_type == ZIO_TYPE_FREE &&
2986 zio->io_priority != ZIO_PRIORITY_NOW) {
2987 switch (zio->io_error) {
2989 ZIO_TRIM_STAT_INCR(bytes, zio->io_size);
2990 ZIO_TRIM_STAT_BUMP(success);
2993 ZIO_TRIM_STAT_BUMP(unsupported);
2996 ZIO_TRIM_STAT_BUMP(failed);
3002 * If the I/O failed, determine whether we should attempt to retry it.
3004 * On retry, we cut in line in the issue queue, since we don't want
3005 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3007 if (zio->io_error && vd == NULL &&
3008 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
3009 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
3010 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
3012 zio->io_flags |= ZIO_FLAG_IO_RETRY |
3013 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
3014 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
3015 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
3016 zio_requeue_io_start_cut_in_line);
3017 return (ZIO_PIPELINE_STOP);
3021 * If we got an error on a leaf device, convert it to ENXIO
3022 * if the device is not accessible at all.
3024 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3025 !vdev_accessible(vd, zio))
3026 zio->io_error = SET_ERROR(ENXIO);
3029 * If we can't write to an interior vdev (mirror or RAID-Z),
3030 * set vdev_cant_write so that we stop trying to allocate from it.
3032 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
3033 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
3034 vd->vdev_cant_write = B_TRUE;
3038 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3040 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3041 zio->io_physdone != NULL) {
3042 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
3043 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
3044 zio->io_physdone(zio->io_logical);
3047 return (ZIO_PIPELINE_CONTINUE);
3051 zio_vdev_io_reissue(zio_t *zio)
3053 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3054 ASSERT(zio->io_error == 0);
3056 zio->io_stage >>= 1;
3060 zio_vdev_io_redone(zio_t *zio)
3062 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
3064 zio->io_stage >>= 1;
3068 zio_vdev_io_bypass(zio_t *zio)
3070 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3071 ASSERT(zio->io_error == 0);
3073 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
3074 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
3078 * ==========================================================================
3079 * Generate and verify checksums
3080 * ==========================================================================
3083 zio_checksum_generate(zio_t *zio)
3085 blkptr_t *bp = zio->io_bp;
3086 enum zio_checksum checksum;
3090 * This is zio_write_phys().
3091 * We're either generating a label checksum, or none at all.
3093 checksum = zio->io_prop.zp_checksum;
3095 if (checksum == ZIO_CHECKSUM_OFF)
3096 return (ZIO_PIPELINE_CONTINUE);
3098 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
3100 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
3101 ASSERT(!IO_IS_ALLOCATING(zio));
3102 checksum = ZIO_CHECKSUM_GANG_HEADER;
3104 checksum = BP_GET_CHECKSUM(bp);
3108 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
3110 return (ZIO_PIPELINE_CONTINUE);
3114 zio_checksum_verify(zio_t *zio)
3116 zio_bad_cksum_t info;
3117 blkptr_t *bp = zio->io_bp;
3120 ASSERT(zio->io_vd != NULL);
3124 * This is zio_read_phys().
3125 * We're either verifying a label checksum, or nothing at all.
3127 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
3128 return (ZIO_PIPELINE_CONTINUE);
3130 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
3133 if ((error = zio_checksum_error(zio, &info)) != 0) {
3134 zio->io_error = error;
3135 if (error == ECKSUM &&
3136 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
3137 zfs_ereport_start_checksum(zio->io_spa,
3138 zio->io_vd, zio, zio->io_offset,
3139 zio->io_size, NULL, &info);
3143 return (ZIO_PIPELINE_CONTINUE);
3147 * Called by RAID-Z to ensure we don't compute the checksum twice.
3150 zio_checksum_verified(zio_t *zio)
3152 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
3156 * ==========================================================================
3157 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
3158 * An error of 0 indicates success. ENXIO indicates whole-device failure,
3159 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
3160 * indicate errors that are specific to one I/O, and most likely permanent.
3161 * Any other error is presumed to be worse because we weren't expecting it.
3162 * ==========================================================================
3165 zio_worst_error(int e1, int e2)
3167 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
3170 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
3171 if (e1 == zio_error_rank[r1])
3174 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
3175 if (e2 == zio_error_rank[r2])
3178 return (r1 > r2 ? e1 : e2);
3182 * ==========================================================================
3184 * ==========================================================================
3187 zio_ready(zio_t *zio)
3189 blkptr_t *bp = zio->io_bp;
3190 zio_t *pio, *pio_next;
3192 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
3193 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
3194 return (ZIO_PIPELINE_STOP);
3196 if (zio->io_ready) {
3197 ASSERT(IO_IS_ALLOCATING(zio));
3198 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
3199 (zio->io_flags & ZIO_FLAG_NOPWRITE));
3200 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
3205 if (bp != NULL && bp != &zio->io_bp_copy)
3206 zio->io_bp_copy = *bp;
3209 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3211 mutex_enter(&zio->io_lock);
3212 zio->io_state[ZIO_WAIT_READY] = 1;
3213 pio = zio_walk_parents(zio);
3214 mutex_exit(&zio->io_lock);
3217 * As we notify zio's parents, new parents could be added.
3218 * New parents go to the head of zio's io_parent_list, however,
3219 * so we will (correctly) not notify them. The remainder of zio's
3220 * io_parent_list, from 'pio_next' onward, cannot change because
3221 * all parents must wait for us to be done before they can be done.
3223 for (; pio != NULL; pio = pio_next) {
3224 pio_next = zio_walk_parents(zio);
3225 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
3228 if (zio->io_flags & ZIO_FLAG_NODATA) {
3229 if (BP_IS_GANG(bp)) {
3230 zio->io_flags &= ~ZIO_FLAG_NODATA;
3232 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
3233 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
3237 if (zio_injection_enabled &&
3238 zio->io_spa->spa_syncing_txg == zio->io_txg)
3239 zio_handle_ignored_writes(zio);
3241 return (ZIO_PIPELINE_CONTINUE);
3245 zio_done(zio_t *zio)
3247 spa_t *spa = zio->io_spa;
3248 zio_t *lio = zio->io_logical;
3249 blkptr_t *bp = zio->io_bp;
3250 vdev_t *vd = zio->io_vd;
3251 uint64_t psize = zio->io_size;
3252 zio_t *pio, *pio_next;
3255 * If our children haven't all completed,
3256 * wait for them and then repeat this pipeline stage.
3258 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
3259 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
3260 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
3261 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
3262 return (ZIO_PIPELINE_STOP);
3264 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3265 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3266 ASSERT(zio->io_children[c][w] == 0);
3268 if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
3269 ASSERT(bp->blk_pad[0] == 0);
3270 ASSERT(bp->blk_pad[1] == 0);
3271 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3272 (bp == zio_unique_parent(zio)->io_bp));
3273 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3274 zio->io_bp_override == NULL &&
3275 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3276 ASSERT(!BP_SHOULD_BYTESWAP(bp));
3277 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3278 ASSERT(BP_COUNT_GANG(bp) == 0 ||
3279 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3281 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3282 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3286 * If there were child vdev/gang/ddt errors, they apply to us now.
3288 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3289 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3290 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3293 * If the I/O on the transformed data was successful, generate any
3294 * checksum reports now while we still have the transformed data.
3296 if (zio->io_error == 0) {
3297 while (zio->io_cksum_report != NULL) {
3298 zio_cksum_report_t *zcr = zio->io_cksum_report;
3299 uint64_t align = zcr->zcr_align;
3300 uint64_t asize = P2ROUNDUP(psize, align);
3301 char *abuf = zio->io_data;
3303 if (asize != psize) {
3304 abuf = zio_buf_alloc(asize);
3305 bcopy(zio->io_data, abuf, psize);
3306 bzero(abuf + psize, asize - psize);
3309 zio->io_cksum_report = zcr->zcr_next;
3310 zcr->zcr_next = NULL;
3311 zcr->zcr_finish(zcr, abuf);
3312 zfs_ereport_free_checksum(zcr);
3315 zio_buf_free(abuf, asize);
3319 zio_pop_transforms(zio); /* note: may set zio->io_error */
3321 vdev_stat_update(zio, psize);
3323 if (zio->io_error) {
3325 * If this I/O is attached to a particular vdev,
3326 * generate an error message describing the I/O failure
3327 * at the block level. We ignore these errors if the
3328 * device is currently unavailable.
3330 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3331 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3333 if ((zio->io_error == EIO || !(zio->io_flags &
3334 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3337 * For logical I/O requests, tell the SPA to log the
3338 * error and generate a logical data ereport.
3340 spa_log_error(spa, zio);
3341 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3346 if (zio->io_error && zio == lio) {
3348 * Determine whether zio should be reexecuted. This will
3349 * propagate all the way to the root via zio_notify_parent().
3351 ASSERT(vd == NULL && bp != NULL);
3352 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3354 if (IO_IS_ALLOCATING(zio) &&
3355 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3356 if (zio->io_error != ENOSPC)
3357 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3359 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3362 if ((zio->io_type == ZIO_TYPE_READ ||
3363 zio->io_type == ZIO_TYPE_FREE) &&
3364 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3365 zio->io_error == ENXIO &&
3366 spa_load_state(spa) == SPA_LOAD_NONE &&
3367 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3368 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3370 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3371 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3374 * Here is a possibly good place to attempt to do
3375 * either combinatorial reconstruction or error correction
3376 * based on checksums. It also might be a good place
3377 * to send out preliminary ereports before we suspend
3383 * If there were logical child errors, they apply to us now.
3384 * We defer this until now to avoid conflating logical child
3385 * errors with errors that happened to the zio itself when
3386 * updating vdev stats and reporting FMA events above.
3388 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3390 if ((zio->io_error || zio->io_reexecute) &&
3391 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3392 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3393 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3395 zio_gang_tree_free(&zio->io_gang_tree);
3398 * Godfather I/Os should never suspend.
3400 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3401 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3402 zio->io_reexecute = 0;
3404 if (zio->io_reexecute) {
3406 * This is a logical I/O that wants to reexecute.
3408 * Reexecute is top-down. When an i/o fails, if it's not
3409 * the root, it simply notifies its parent and sticks around.
3410 * The parent, seeing that it still has children in zio_done(),
3411 * does the same. This percolates all the way up to the root.
3412 * The root i/o will reexecute or suspend the entire tree.
3414 * This approach ensures that zio_reexecute() honors
3415 * all the original i/o dependency relationships, e.g.
3416 * parents not executing until children are ready.
3418 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3420 zio->io_gang_leader = NULL;
3422 mutex_enter(&zio->io_lock);
3423 zio->io_state[ZIO_WAIT_DONE] = 1;
3424 mutex_exit(&zio->io_lock);
3427 * "The Godfather" I/O monitors its children but is
3428 * not a true parent to them. It will track them through
3429 * the pipeline but severs its ties whenever they get into
3430 * trouble (e.g. suspended). This allows "The Godfather"
3431 * I/O to return status without blocking.
3433 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3434 zio_link_t *zl = zio->io_walk_link;
3435 pio_next = zio_walk_parents(zio);
3437 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3438 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3439 zio_remove_child(pio, zio, zl);
3440 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3444 if ((pio = zio_unique_parent(zio)) != NULL) {
3446 * We're not a root i/o, so there's nothing to do
3447 * but notify our parent. Don't propagate errors
3448 * upward since we haven't permanently failed yet.
3450 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3451 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3452 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3453 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3455 * We'd fail again if we reexecuted now, so suspend
3456 * until conditions improve (e.g. device comes online).
3458 zio_suspend(spa, zio);
3461 * Reexecution is potentially a huge amount of work.
3462 * Hand it off to the otherwise-unused claim taskq.
3464 #if defined(illumos) || !defined(_KERNEL)
3465 ASSERT(zio->io_tqent.tqent_next == NULL);
3467 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
3469 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3470 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3473 return (ZIO_PIPELINE_STOP);
3476 ASSERT(zio->io_child_count == 0);
3477 ASSERT(zio->io_reexecute == 0);
3478 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3481 * Report any checksum errors, since the I/O is complete.
3483 while (zio->io_cksum_report != NULL) {
3484 zio_cksum_report_t *zcr = zio->io_cksum_report;
3485 zio->io_cksum_report = zcr->zcr_next;
3486 zcr->zcr_next = NULL;
3487 zcr->zcr_finish(zcr, NULL);
3488 zfs_ereport_free_checksum(zcr);
3492 * It is the responsibility of the done callback to ensure that this
3493 * particular zio is no longer discoverable for adoption, and as
3494 * such, cannot acquire any new parents.
3499 mutex_enter(&zio->io_lock);
3500 zio->io_state[ZIO_WAIT_DONE] = 1;
3501 mutex_exit(&zio->io_lock);
3503 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3504 zio_link_t *zl = zio->io_walk_link;
3505 pio_next = zio_walk_parents(zio);
3506 zio_remove_child(pio, zio, zl);
3507 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3510 if (zio->io_waiter != NULL) {
3511 mutex_enter(&zio->io_lock);
3512 zio->io_executor = NULL;
3513 cv_broadcast(&zio->io_cv);
3514 mutex_exit(&zio->io_lock);
3519 return (ZIO_PIPELINE_STOP);
3523 * ==========================================================================
3524 * I/O pipeline definition
3525 * ==========================================================================
3527 static zio_pipe_stage_t *zio_pipeline[] = {
3533 zio_checksum_generate,
3548 zio_checksum_verify,
3556 * Compare two zbookmark_phys_t's to see which we would reach first in a
3557 * pre-order traversal of the object tree.
3559 * This is simple in every case aside from the meta-dnode object. For all other
3560 * objects, we traverse them in order (object 1 before object 2, and so on).
3561 * However, all of these objects are traversed while traversing object 0, since
3562 * the data it points to is the list of objects. Thus, we need to convert to a
3563 * canonical representation so we can compare meta-dnode bookmarks to
3564 * non-meta-dnode bookmarks.
3566 * We do this by calculating "equivalents" for each field of the zbookmark.
3567 * zbookmarks outside of the meta-dnode use their own object and level, and
3568 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
3569 * blocks this bookmark refers to) by multiplying their blkid by their span
3570 * (the number of L0 blocks contained within one block at their level).
3571 * zbookmarks inside the meta-dnode calculate their object equivalent
3572 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
3573 * level + 1<<31 (any value larger than a level could ever be) for their level.
3574 * This causes them to always compare before a bookmark in their object
3575 * equivalent, compare appropriately to bookmarks in other objects, and to
3576 * compare appropriately to other bookmarks in the meta-dnode.
3579 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
3580 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
3583 * These variables represent the "equivalent" values for the zbookmark,
3584 * after converting zbookmarks inside the meta dnode to their
3585 * normal-object equivalents.
3587 uint64_t zb1obj, zb2obj;
3588 uint64_t zb1L0, zb2L0;
3589 uint64_t zb1level, zb2level;
3591 if (zb1->zb_object == zb2->zb_object &&
3592 zb1->zb_level == zb2->zb_level &&
3593 zb1->zb_blkid == zb2->zb_blkid)
3597 * BP_SPANB calculates the span in blocks.
3599 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
3600 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
3602 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3603 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
3605 zb1level = zb1->zb_level + COMPARE_META_LEVEL;
3607 zb1obj = zb1->zb_object;
3608 zb1level = zb1->zb_level;
3611 if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
3612 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
3614 zb2level = zb2->zb_level + COMPARE_META_LEVEL;
3616 zb2obj = zb2->zb_object;
3617 zb2level = zb2->zb_level;
3620 /* Now that we have a canonical representation, do the comparison. */
3621 if (zb1obj != zb2obj)
3622 return (zb1obj < zb2obj ? -1 : 1);
3623 else if (zb1L0 != zb2L0)
3624 return (zb1L0 < zb2L0 ? -1 : 1);
3625 else if (zb1level != zb2level)
3626 return (zb1level > zb2level ? -1 : 1);
3628 * This can (theoretically) happen if the bookmarks have the same object
3629 * and level, but different blkids, if the block sizes are not the same.
3630 * There is presently no way to change the indirect block sizes
3636 * This function checks the following: given that last_block is the place that
3637 * our traversal stopped last time, does that guarantee that we've visited
3638 * every node under subtree_root? Therefore, we can't just use the raw output
3639 * of zbookmark_compare. We have to pass in a modified version of
3640 * subtree_root; by incrementing the block id, and then checking whether
3641 * last_block is before or equal to that, we can tell whether or not having
3642 * visited last_block implies that all of subtree_root's children have been
3646 zbookmark_subtree_completed(const dnode_phys_t *dnp,
3647 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
3649 zbookmark_phys_t mod_zb = *subtree_root;
3651 ASSERT(last_block->zb_level == 0);
3653 /* The objset_phys_t isn't before anything. */
3658 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
3659 * data block size in sectors, because that variable is only used if
3660 * the bookmark refers to a block in the meta-dnode. Since we don't
3661 * know without examining it what object it refers to, and there's no
3662 * harm in passing in this value in other cases, we always pass it in.
3664 * We pass in 0 for the indirect block size shift because zb2 must be
3665 * level 0. The indirect block size is only used to calculate the span
3666 * of the bookmark, but since the bookmark must be level 0, the span is
3667 * always 1, so the math works out.
3669 * If you make changes to how the zbookmark_compare code works, be sure
3670 * to make sure that this code still works afterwards.
3672 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
3673 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,