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) 2013 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
27 #include <sys/zfs_context.h>
28 #include <sys/fm/fs/zfs.h>
31 #include <sys/spa_impl.h>
32 #include <sys/vdev_impl.h>
33 #include <sys/zio_impl.h>
34 #include <sys/zio_compress.h>
35 #include <sys/zio_checksum.h>
36 #include <sys/dmu_objset.h>
39 #include <sys/trim_map.h>
40 #include <sys/zfeature.h>
42 SYSCTL_DECL(_vfs_zfs);
43 SYSCTL_NODE(_vfs_zfs, OID_AUTO, zio, CTLFLAG_RW, 0, "ZFS ZIO");
44 static int zio_use_uma = 0;
45 TUNABLE_INT("vfs.zfs.zio.use_uma", &zio_use_uma);
46 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, use_uma, CTLFLAG_RDTUN, &zio_use_uma, 0,
47 "Use uma(9) for ZIO allocations");
49 zio_trim_stats_t zio_trim_stats = {
50 { "bytes", KSTAT_DATA_UINT64,
51 "Number of bytes successfully TRIMmed" },
52 { "success", KSTAT_DATA_UINT64,
53 "Number of successful TRIM requests" },
54 { "unsupported", KSTAT_DATA_UINT64,
55 "Number of TRIM requests that failed because TRIM is not supported" },
56 { "failed", KSTAT_DATA_UINT64,
57 "Number of TRIM requests that failed for reasons other than not supported" },
60 static kstat_t *zio_trim_ksp;
63 * ==========================================================================
64 * I/O type descriptions
65 * ==========================================================================
67 const char *zio_type_name[ZIO_TYPES] = {
68 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
73 * ==========================================================================
75 * ==========================================================================
77 kmem_cache_t *zio_cache;
78 kmem_cache_t *zio_link_cache;
79 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
80 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
83 extern vmem_t *zio_alloc_arena;
85 extern int zfs_mg_alloc_failures;
88 * The following actions directly effect the spa's sync-to-convergence logic.
89 * The values below define the sync pass when we start performing the action.
90 * Care should be taken when changing these values as they directly impact
91 * spa_sync() performance. Tuning these values may introduce subtle performance
92 * pathologies and should only be done in the context of performance analysis.
93 * These tunables will eventually be removed and replaced with #defines once
94 * enough analysis has been done to determine optimal values.
96 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
97 * regular blocks are not deferred.
99 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
100 TUNABLE_INT("vfs.zfs.sync_pass_deferred_free", &zfs_sync_pass_deferred_free);
101 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_deferred_free, CTLFLAG_RDTUN,
102 &zfs_sync_pass_deferred_free, 0, "defer frees starting in this pass");
103 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
104 TUNABLE_INT("vfs.zfs.sync_pass_dont_compress", &zfs_sync_pass_dont_compress);
105 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_dont_compress, CTLFLAG_RDTUN,
106 &zfs_sync_pass_dont_compress, 0, "don't compress starting in this pass");
107 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
108 TUNABLE_INT("vfs.zfs.sync_pass_rewrite", &zfs_sync_pass_rewrite);
109 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_rewrite, CTLFLAG_RDTUN,
110 &zfs_sync_pass_rewrite, 0, "rewrite new bps starting in this pass");
113 * An allocating zio is one that either currently has the DVA allocate
114 * stage set or will have it later in its lifetime.
116 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
118 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
121 int zio_buf_debug_limit = 16384;
123 int zio_buf_debug_limit = 0;
130 zio_cache = kmem_cache_create("zio_cache",
131 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
132 zio_link_cache = kmem_cache_create("zio_link_cache",
133 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
138 * For small buffers, we want a cache for each multiple of
139 * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache
140 * for each quarter-power of 2. For large buffers, we want
141 * a cache for each multiple of PAGESIZE.
143 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
144 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
147 size_t cflags = (size > zio_buf_debug_limit) ? (KMC_NODEBUG|KMC_NOTOUCH) : 0;
149 while (p2 & (p2 - 1))
155 * If we are using watchpoints, put each buffer on its own page,
156 * to eliminate the performance overhead of trapping to the
157 * kernel when modifying a non-watched buffer that shares the
158 * page with a watched buffer.
160 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
164 if (size <= 4 * SPA_MINBLOCKSIZE) {
165 align = SPA_MINBLOCKSIZE;
166 } else if (IS_P2ALIGNED(size, PAGESIZE)) {
168 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
174 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
175 zio_buf_cache[c] = kmem_cache_create(name, size,
176 align, NULL, NULL, NULL, NULL, NULL, cflags);
179 * Since zio_data bufs do not appear in crash dumps, we
180 * pass KMC_NOTOUCH so that no allocator metadata is
181 * stored with the buffers.
183 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
184 zio_data_buf_cache[c] = kmem_cache_create(name, size,
185 align, NULL, NULL, NULL, NULL, NULL,
186 cflags | KMC_NOTOUCH);
191 ASSERT(zio_buf_cache[c] != NULL);
192 if (zio_buf_cache[c - 1] == NULL)
193 zio_buf_cache[c - 1] = zio_buf_cache[c];
195 ASSERT(zio_data_buf_cache[c] != NULL);
196 if (zio_data_buf_cache[c - 1] == NULL)
197 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
202 * The zio write taskqs have 1 thread per cpu, allow 1/2 of the taskqs
203 * to fail 3 times per txg or 8 failures, whichever is greater.
205 if (zfs_mg_alloc_failures == 0)
206 zfs_mg_alloc_failures = MAX((3 * max_ncpus / 2), 8);
207 else if (zfs_mg_alloc_failures < 8)
208 zfs_mg_alloc_failures = 8;
212 zio_trim_ksp = kstat_create("zfs", 0, "zio_trim", "misc",
214 sizeof(zio_trim_stats) / sizeof(kstat_named_t),
217 if (zio_trim_ksp != NULL) {
218 zio_trim_ksp->ks_data = &zio_trim_stats;
219 kstat_install(zio_trim_ksp);
227 kmem_cache_t *last_cache = NULL;
228 kmem_cache_t *last_data_cache = NULL;
230 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
231 if (zio_buf_cache[c] != last_cache) {
232 last_cache = zio_buf_cache[c];
233 kmem_cache_destroy(zio_buf_cache[c]);
235 zio_buf_cache[c] = NULL;
237 if (zio_data_buf_cache[c] != last_data_cache) {
238 last_data_cache = zio_data_buf_cache[c];
239 kmem_cache_destroy(zio_data_buf_cache[c]);
241 zio_data_buf_cache[c] = NULL;
244 kmem_cache_destroy(zio_link_cache);
245 kmem_cache_destroy(zio_cache);
249 if (zio_trim_ksp != NULL) {
250 kstat_delete(zio_trim_ksp);
256 * ==========================================================================
257 * Allocate and free I/O buffers
258 * ==========================================================================
262 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
263 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
264 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
265 * excess / transient data in-core during a crashdump.
268 zio_buf_alloc(size_t size)
270 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
272 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
275 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
277 return (kmem_alloc(size, KM_SLEEP));
281 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
282 * crashdump if the kernel panics. This exists so that we will limit the amount
283 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
284 * of kernel heap dumped to disk when the kernel panics)
287 zio_data_buf_alloc(size_t size)
289 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
291 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
294 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
296 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG));
300 zio_buf_free(void *buf, size_t size)
302 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
304 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
307 kmem_cache_free(zio_buf_cache[c], buf);
309 kmem_free(buf, size);
313 zio_data_buf_free(void *buf, size_t size)
315 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
317 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
320 kmem_cache_free(zio_data_buf_cache[c], buf);
322 kmem_free(buf, size);
326 * ==========================================================================
327 * Push and pop I/O transform buffers
328 * ==========================================================================
331 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
332 zio_transform_func_t *transform)
334 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
336 zt->zt_orig_data = zio->io_data;
337 zt->zt_orig_size = zio->io_size;
338 zt->zt_bufsize = bufsize;
339 zt->zt_transform = transform;
341 zt->zt_next = zio->io_transform_stack;
342 zio->io_transform_stack = zt;
349 zio_pop_transforms(zio_t *zio)
353 while ((zt = zio->io_transform_stack) != NULL) {
354 if (zt->zt_transform != NULL)
355 zt->zt_transform(zio,
356 zt->zt_orig_data, zt->zt_orig_size);
358 if (zt->zt_bufsize != 0)
359 zio_buf_free(zio->io_data, zt->zt_bufsize);
361 zio->io_data = zt->zt_orig_data;
362 zio->io_size = zt->zt_orig_size;
363 zio->io_transform_stack = zt->zt_next;
365 kmem_free(zt, sizeof (zio_transform_t));
370 * ==========================================================================
371 * I/O transform callbacks for subblocks and decompression
372 * ==========================================================================
375 zio_subblock(zio_t *zio, void *data, uint64_t size)
377 ASSERT(zio->io_size > size);
379 if (zio->io_type == ZIO_TYPE_READ)
380 bcopy(zio->io_data, data, size);
384 zio_decompress(zio_t *zio, void *data, uint64_t size)
386 if (zio->io_error == 0 &&
387 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
388 zio->io_data, data, zio->io_size, size) != 0)
389 zio->io_error = SET_ERROR(EIO);
393 * ==========================================================================
394 * I/O parent/child relationships and pipeline interlocks
395 * ==========================================================================
398 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
399 * continue calling these functions until they return NULL.
400 * Otherwise, the next caller will pick up the list walk in
401 * some indeterminate state. (Otherwise every caller would
402 * have to pass in a cookie to keep the state represented by
403 * io_walk_link, which gets annoying.)
406 zio_walk_parents(zio_t *cio)
408 zio_link_t *zl = cio->io_walk_link;
409 list_t *pl = &cio->io_parent_list;
411 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
412 cio->io_walk_link = zl;
417 ASSERT(zl->zl_child == cio);
418 return (zl->zl_parent);
422 zio_walk_children(zio_t *pio)
424 zio_link_t *zl = pio->io_walk_link;
425 list_t *cl = &pio->io_child_list;
427 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
428 pio->io_walk_link = zl;
433 ASSERT(zl->zl_parent == pio);
434 return (zl->zl_child);
438 zio_unique_parent(zio_t *cio)
440 zio_t *pio = zio_walk_parents(cio);
442 VERIFY(zio_walk_parents(cio) == NULL);
447 zio_add_child(zio_t *pio, zio_t *cio)
449 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
452 * Logical I/Os can have logical, gang, or vdev children.
453 * Gang I/Os can have gang or vdev children.
454 * Vdev I/Os can only have vdev children.
455 * The following ASSERT captures all of these constraints.
457 ASSERT(cio->io_child_type <= pio->io_child_type);
462 mutex_enter(&cio->io_lock);
463 mutex_enter(&pio->io_lock);
465 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
467 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
468 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
470 list_insert_head(&pio->io_child_list, zl);
471 list_insert_head(&cio->io_parent_list, zl);
473 pio->io_child_count++;
474 cio->io_parent_count++;
476 mutex_exit(&pio->io_lock);
477 mutex_exit(&cio->io_lock);
481 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
483 ASSERT(zl->zl_parent == pio);
484 ASSERT(zl->zl_child == cio);
486 mutex_enter(&cio->io_lock);
487 mutex_enter(&pio->io_lock);
489 list_remove(&pio->io_child_list, zl);
490 list_remove(&cio->io_parent_list, zl);
492 pio->io_child_count--;
493 cio->io_parent_count--;
495 mutex_exit(&pio->io_lock);
496 mutex_exit(&cio->io_lock);
498 kmem_cache_free(zio_link_cache, zl);
502 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
504 uint64_t *countp = &zio->io_children[child][wait];
505 boolean_t waiting = B_FALSE;
507 mutex_enter(&zio->io_lock);
508 ASSERT(zio->io_stall == NULL);
511 zio->io_stall = countp;
514 mutex_exit(&zio->io_lock);
520 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
522 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
523 int *errorp = &pio->io_child_error[zio->io_child_type];
525 mutex_enter(&pio->io_lock);
526 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
527 *errorp = zio_worst_error(*errorp, zio->io_error);
528 pio->io_reexecute |= zio->io_reexecute;
529 ASSERT3U(*countp, >, 0);
533 if (*countp == 0 && pio->io_stall == countp) {
534 pio->io_stall = NULL;
535 mutex_exit(&pio->io_lock);
538 mutex_exit(&pio->io_lock);
543 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
545 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
546 zio->io_error = zio->io_child_error[c];
550 * ==========================================================================
551 * Create the various types of I/O (read, write, free, etc)
552 * ==========================================================================
555 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
556 void *data, uint64_t size, zio_done_func_t *done, void *private,
557 zio_type_t type, zio_priority_t priority, enum zio_flag flags,
558 vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
559 enum zio_stage stage, enum zio_stage pipeline)
563 ASSERT3U(type == ZIO_TYPE_FREE || size, <=, SPA_MAXBLOCKSIZE);
564 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
565 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
567 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
568 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
569 ASSERT(vd || stage == ZIO_STAGE_OPEN);
571 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
572 bzero(zio, sizeof (zio_t));
574 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
575 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
577 list_create(&zio->io_parent_list, sizeof (zio_link_t),
578 offsetof(zio_link_t, zl_parent_node));
579 list_create(&zio->io_child_list, sizeof (zio_link_t),
580 offsetof(zio_link_t, zl_child_node));
583 zio->io_child_type = ZIO_CHILD_VDEV;
584 else if (flags & ZIO_FLAG_GANG_CHILD)
585 zio->io_child_type = ZIO_CHILD_GANG;
586 else if (flags & ZIO_FLAG_DDT_CHILD)
587 zio->io_child_type = ZIO_CHILD_DDT;
589 zio->io_child_type = ZIO_CHILD_LOGICAL;
592 zio->io_bp = (blkptr_t *)bp;
593 zio->io_bp_copy = *bp;
594 zio->io_bp_orig = *bp;
595 if (type != ZIO_TYPE_WRITE ||
596 zio->io_child_type == ZIO_CHILD_DDT)
597 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
598 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
599 zio->io_logical = zio;
600 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
601 pipeline |= ZIO_GANG_STAGES;
607 zio->io_private = private;
609 zio->io_priority = priority;
611 zio->io_offset = offset;
612 zio->io_orig_data = zio->io_data = data;
613 zio->io_orig_size = zio->io_size = size;
614 zio->io_orig_flags = zio->io_flags = flags;
615 zio->io_orig_stage = zio->io_stage = stage;
616 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
618 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
619 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
622 zio->io_bookmark = *zb;
625 if (zio->io_logical == NULL)
626 zio->io_logical = pio->io_logical;
627 if (zio->io_child_type == ZIO_CHILD_GANG)
628 zio->io_gang_leader = pio->io_gang_leader;
629 zio_add_child(pio, zio);
636 zio_destroy(zio_t *zio)
638 list_destroy(&zio->io_parent_list);
639 list_destroy(&zio->io_child_list);
640 mutex_destroy(&zio->io_lock);
641 cv_destroy(&zio->io_cv);
642 kmem_cache_free(zio_cache, zio);
646 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
647 void *private, enum zio_flag flags)
651 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
652 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
653 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
659 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
661 return (zio_null(NULL, spa, NULL, done, private, flags));
665 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
666 void *data, uint64_t size, zio_done_func_t *done, void *private,
667 zio_priority_t priority, enum zio_flag flags, const zbookmark_t *zb)
671 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
672 data, size, done, private,
673 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
674 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
675 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
681 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
682 void *data, uint64_t size, const zio_prop_t *zp,
683 zio_done_func_t *ready, zio_done_func_t *physdone, zio_done_func_t *done,
685 zio_priority_t priority, enum zio_flag flags, const zbookmark_t *zb)
689 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
690 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
691 zp->zp_compress >= ZIO_COMPRESS_OFF &&
692 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
693 DMU_OT_IS_VALID(zp->zp_type) &&
696 zp->zp_copies <= spa_max_replication(spa));
698 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
699 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
700 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
701 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
703 zio->io_ready = ready;
704 zio->io_physdone = physdone;
711 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
712 uint64_t size, zio_done_func_t *done, void *private,
713 zio_priority_t priority, enum zio_flag flags, zbookmark_t *zb)
717 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
718 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
719 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
725 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
727 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
728 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
729 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
730 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
733 * We must reset the io_prop to match the values that existed
734 * when the bp was first written by dmu_sync() keeping in mind
735 * that nopwrite and dedup are mutually exclusive.
737 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
738 zio->io_prop.zp_nopwrite = nopwrite;
739 zio->io_prop.zp_copies = copies;
740 zio->io_bp_override = bp;
744 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
746 metaslab_check_free(spa, bp);
749 * Frees that are for the currently-syncing txg, are not going to be
750 * deferred, and which will not need to do a read (i.e. not GANG or
751 * DEDUP), can be processed immediately. Otherwise, put them on the
752 * in-memory list for later processing.
754 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
755 txg != spa->spa_syncing_txg ||
756 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
757 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
759 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp,
760 BP_GET_PSIZE(bp), 0)));
765 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
766 uint64_t size, enum zio_flag flags)
769 enum zio_stage stage = ZIO_FREE_PIPELINE;
771 dprintf_bp(bp, "freeing in txg %llu, pass %u",
772 (longlong_t)txg, spa->spa_sync_pass);
774 ASSERT(!BP_IS_HOLE(bp));
775 ASSERT(spa_syncing_txg(spa) == txg);
776 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
778 metaslab_check_free(spa, bp);
781 if (zfs_trim_enabled)
782 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START |
783 ZIO_STAGE_VDEV_IO_ASSESS;
785 * GANG and DEDUP blocks can induce a read (for the gang block header,
786 * or the DDT), so issue them asynchronously so that this thread is
789 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
790 stage |= ZIO_STAGE_ISSUE_ASYNC;
792 zio = zio_create(pio, spa, txg, bp, NULL, size,
793 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags,
794 NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
800 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
801 zio_done_func_t *done, void *private, enum zio_flag flags)
806 * A claim is an allocation of a specific block. Claims are needed
807 * to support immediate writes in the intent log. The issue is that
808 * immediate writes contain committed data, but in a txg that was
809 * *not* committed. Upon opening the pool after an unclean shutdown,
810 * the intent log claims all blocks that contain immediate write data
811 * so that the SPA knows they're in use.
813 * All claims *must* be resolved in the first txg -- before the SPA
814 * starts allocating blocks -- so that nothing is allocated twice.
815 * If txg == 0 we just verify that the block is claimable.
817 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
818 ASSERT(txg == spa_first_txg(spa) || txg == 0);
819 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
821 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
822 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
823 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
829 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset,
830 uint64_t size, zio_done_func_t *done, void *private,
836 if (vd->vdev_children == 0) {
837 zio = zio_create(pio, spa, 0, NULL, NULL, size, done, private,
838 ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, offset, NULL,
839 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
843 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
845 for (c = 0; c < vd->vdev_children; c++)
846 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
847 offset, size, done, private, flags));
854 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
855 void *data, int checksum, zio_done_func_t *done, void *private,
856 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
860 ASSERT(vd->vdev_children == 0);
861 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
862 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
863 ASSERT3U(offset + size, <=, vd->vdev_psize);
865 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
866 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
867 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
869 zio->io_prop.zp_checksum = checksum;
875 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
876 void *data, int checksum, zio_done_func_t *done, void *private,
877 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
881 ASSERT(vd->vdev_children == 0);
882 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
883 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
884 ASSERT3U(offset + size, <=, vd->vdev_psize);
886 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
887 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
888 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
890 zio->io_prop.zp_checksum = checksum;
892 if (zio_checksum_table[checksum].ci_eck) {
894 * zec checksums are necessarily destructive -- they modify
895 * the end of the write buffer to hold the verifier/checksum.
896 * Therefore, we must make a local copy in case the data is
897 * being written to multiple places in parallel.
899 void *wbuf = zio_buf_alloc(size);
900 bcopy(data, wbuf, size);
901 zio_push_transform(zio, wbuf, size, size, NULL);
908 * Create a child I/O to do some work for us.
911 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
912 void *data, uint64_t size, int type, zio_priority_t priority,
913 enum zio_flag flags, zio_done_func_t *done, void *private)
915 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
918 ASSERT(vd->vdev_parent ==
919 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
921 if (type == ZIO_TYPE_READ && bp != NULL) {
923 * If we have the bp, then the child should perform the
924 * checksum and the parent need not. This pushes error
925 * detection as close to the leaves as possible and
926 * eliminates redundant checksums in the interior nodes.
928 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
929 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
932 if (vd->vdev_children == 0)
933 offset += VDEV_LABEL_START_SIZE;
935 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
938 * If we've decided to do a repair, the write is not speculative --
939 * even if the original read was.
941 if (flags & ZIO_FLAG_IO_REPAIR)
942 flags &= ~ZIO_FLAG_SPECULATIVE;
944 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
945 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
946 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
948 zio->io_physdone = pio->io_physdone;
949 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
950 zio->io_logical->io_phys_children++;
956 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
957 int type, zio_priority_t priority, enum zio_flag flags,
958 zio_done_func_t *done, void *private)
962 ASSERT(vd->vdev_ops->vdev_op_leaf);
964 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
965 data, size, done, private, type, priority,
966 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
968 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
974 zio_flush(zio_t *zio, vdev_t *vd)
976 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0,
978 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
982 zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size)
985 ASSERT(vd->vdev_ops->vdev_op_leaf);
987 return zio_ioctl(zio, spa, vd, DKIOCTRIM, offset, size,
989 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY);
993 zio_shrink(zio_t *zio, uint64_t size)
995 ASSERT(zio->io_executor == NULL);
996 ASSERT(zio->io_orig_size == zio->io_size);
997 ASSERT(size <= zio->io_size);
1000 * We don't shrink for raidz because of problems with the
1001 * reconstruction when reading back less than the block size.
1002 * Note, BP_IS_RAIDZ() assumes no compression.
1004 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1005 if (!BP_IS_RAIDZ(zio->io_bp))
1006 zio->io_orig_size = zio->io_size = size;
1010 * ==========================================================================
1011 * Prepare to read and write logical blocks
1012 * ==========================================================================
1016 zio_read_bp_init(zio_t *zio)
1018 blkptr_t *bp = zio->io_bp;
1020 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1021 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1022 !(zio->io_flags & ZIO_FLAG_RAW)) {
1023 uint64_t psize = BP_GET_PSIZE(bp);
1024 void *cbuf = zio_buf_alloc(psize);
1026 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
1029 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1030 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1032 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1033 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1035 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1036 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1038 return (ZIO_PIPELINE_CONTINUE);
1042 zio_write_bp_init(zio_t *zio)
1044 spa_t *spa = zio->io_spa;
1045 zio_prop_t *zp = &zio->io_prop;
1046 enum zio_compress compress = zp->zp_compress;
1047 blkptr_t *bp = zio->io_bp;
1048 uint64_t lsize = zio->io_size;
1049 uint64_t psize = lsize;
1053 * If our children haven't all reached the ready stage,
1054 * wait for them and then repeat this pipeline stage.
1056 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1057 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1058 return (ZIO_PIPELINE_STOP);
1060 if (!IO_IS_ALLOCATING(zio))
1061 return (ZIO_PIPELINE_CONTINUE);
1063 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1065 if (zio->io_bp_override) {
1066 ASSERT(bp->blk_birth != zio->io_txg);
1067 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1069 *bp = *zio->io_bp_override;
1070 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1073 * If we've been overridden and nopwrite is set then
1074 * set the flag accordingly to indicate that a nopwrite
1075 * has already occurred.
1077 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1078 ASSERT(!zp->zp_dedup);
1079 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1080 return (ZIO_PIPELINE_CONTINUE);
1083 ASSERT(!zp->zp_nopwrite);
1085 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1086 return (ZIO_PIPELINE_CONTINUE);
1088 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1089 zp->zp_dedup_verify);
1091 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1092 BP_SET_DEDUP(bp, 1);
1093 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1094 return (ZIO_PIPELINE_CONTINUE);
1096 zio->io_bp_override = NULL;
1100 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1102 * We're rewriting an existing block, which means we're
1103 * working on behalf of spa_sync(). For spa_sync() to
1104 * converge, it must eventually be the case that we don't
1105 * have to allocate new blocks. But compression changes
1106 * the blocksize, which forces a reallocate, and makes
1107 * convergence take longer. Therefore, after the first
1108 * few passes, stop compressing to ensure convergence.
1110 pass = spa_sync_pass(spa);
1112 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1113 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1114 ASSERT(!BP_GET_DEDUP(bp));
1116 if (pass >= zfs_sync_pass_dont_compress)
1117 compress = ZIO_COMPRESS_OFF;
1119 /* Make sure someone doesn't change their mind on overwrites */
1120 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
1121 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1124 if (compress != ZIO_COMPRESS_OFF) {
1125 metaslab_class_t *mc = spa_normal_class(spa);
1126 void *cbuf = zio_buf_alloc(lsize);
1127 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize,
1128 (size_t)metaslab_class_get_minblocksize(mc));
1129 if (psize == 0 || psize == lsize) {
1130 compress = ZIO_COMPRESS_OFF;
1131 zio_buf_free(cbuf, lsize);
1133 ASSERT(psize < lsize);
1134 zio_push_transform(zio, cbuf, psize, lsize, NULL);
1139 * The final pass of spa_sync() must be all rewrites, but the first
1140 * few passes offer a trade-off: allocating blocks defers convergence,
1141 * but newly allocated blocks are sequential, so they can be written
1142 * to disk faster. Therefore, we allow the first few passes of
1143 * spa_sync() to allocate new blocks, but force rewrites after that.
1144 * There should only be a handful of blocks after pass 1 in any case.
1146 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1147 BP_GET_PSIZE(bp) == psize &&
1148 pass >= zfs_sync_pass_rewrite) {
1150 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1151 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1152 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1155 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1159 if (zio->io_bp_orig.blk_birth != 0 &&
1160 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1161 BP_SET_LSIZE(bp, lsize);
1162 BP_SET_TYPE(bp, zp->zp_type);
1163 BP_SET_LEVEL(bp, zp->zp_level);
1164 BP_SET_BIRTH(bp, zio->io_txg, 0);
1166 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1168 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1169 BP_SET_LSIZE(bp, lsize);
1170 BP_SET_TYPE(bp, zp->zp_type);
1171 BP_SET_LEVEL(bp, zp->zp_level);
1172 BP_SET_PSIZE(bp, psize);
1173 BP_SET_COMPRESS(bp, compress);
1174 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1175 BP_SET_DEDUP(bp, zp->zp_dedup);
1176 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1178 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1179 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1180 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1182 if (zp->zp_nopwrite) {
1183 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1184 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1185 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1189 return (ZIO_PIPELINE_CONTINUE);
1193 zio_free_bp_init(zio_t *zio)
1195 blkptr_t *bp = zio->io_bp;
1197 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1198 if (BP_GET_DEDUP(bp))
1199 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1202 return (ZIO_PIPELINE_CONTINUE);
1206 * ==========================================================================
1207 * Execute the I/O pipeline
1208 * ==========================================================================
1212 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1214 spa_t *spa = zio->io_spa;
1215 zio_type_t t = zio->io_type;
1216 int flags = (cutinline ? TQ_FRONT : 0);
1218 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1221 * If we're a config writer or a probe, the normal issue and
1222 * interrupt threads may all be blocked waiting for the config lock.
1223 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1225 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1229 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1231 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1235 * If this is a high priority I/O, then use the high priority taskq if
1238 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1239 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1242 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1245 * NB: We are assuming that the zio can only be dispatched
1246 * to a single taskq at a time. It would be a grievous error
1247 * to dispatch the zio to another taskq at the same time.
1249 #if defined(illumos) || !defined(_KERNEL)
1250 ASSERT(zio->io_tqent.tqent_next == NULL);
1252 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
1254 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1255 flags, &zio->io_tqent);
1259 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1261 kthread_t *executor = zio->io_executor;
1262 spa_t *spa = zio->io_spa;
1264 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1265 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1267 for (i = 0; i < tqs->stqs_count; i++) {
1268 if (taskq_member(tqs->stqs_taskq[i], executor))
1277 zio_issue_async(zio_t *zio)
1279 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1281 return (ZIO_PIPELINE_STOP);
1285 zio_interrupt(zio_t *zio)
1287 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1291 * Execute the I/O pipeline until one of the following occurs:
1293 * (1) the I/O completes
1294 * (2) the pipeline stalls waiting for dependent child I/Os
1295 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1296 * (4) the I/O is delegated by vdev-level caching or aggregation
1297 * (5) the I/O is deferred due to vdev-level queueing
1298 * (6) the I/O is handed off to another thread.
1300 * In all cases, the pipeline stops whenever there's no CPU work; it never
1301 * burns a thread in cv_wait().
1303 * There's no locking on io_stage because there's no legitimate way
1304 * for multiple threads to be attempting to process the same I/O.
1306 static zio_pipe_stage_t *zio_pipeline[];
1309 zio_execute(zio_t *zio)
1311 zio->io_executor = curthread;
1313 while (zio->io_stage < ZIO_STAGE_DONE) {
1314 enum zio_stage pipeline = zio->io_pipeline;
1315 enum zio_stage stage = zio->io_stage;
1318 ASSERT(!MUTEX_HELD(&zio->io_lock));
1319 ASSERT(ISP2(stage));
1320 ASSERT(zio->io_stall == NULL);
1324 } while ((stage & pipeline) == 0);
1326 ASSERT(stage <= ZIO_STAGE_DONE);
1329 * If we are in interrupt context and this pipeline stage
1330 * will grab a config lock that is held across I/O,
1331 * or may wait for an I/O that needs an interrupt thread
1332 * to complete, issue async to avoid deadlock.
1334 * For VDEV_IO_START, we cut in line so that the io will
1335 * be sent to disk promptly.
1337 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1338 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1339 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1340 zio_requeue_io_start_cut_in_line : B_FALSE;
1341 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1345 zio->io_stage = stage;
1346 rv = zio_pipeline[highbit(stage) - 1](zio);
1348 if (rv == ZIO_PIPELINE_STOP)
1351 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1356 * ==========================================================================
1357 * Initiate I/O, either sync or async
1358 * ==========================================================================
1361 zio_wait(zio_t *zio)
1365 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1366 ASSERT(zio->io_executor == NULL);
1368 zio->io_waiter = curthread;
1372 mutex_enter(&zio->io_lock);
1373 while (zio->io_executor != NULL)
1374 cv_wait(&zio->io_cv, &zio->io_lock);
1375 mutex_exit(&zio->io_lock);
1377 error = zio->io_error;
1384 zio_nowait(zio_t *zio)
1386 ASSERT(zio->io_executor == NULL);
1388 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1389 zio_unique_parent(zio) == NULL) {
1391 * This is a logical async I/O with no parent to wait for it.
1392 * We add it to the spa_async_root_zio "Godfather" I/O which
1393 * will ensure they complete prior to unloading the pool.
1395 spa_t *spa = zio->io_spa;
1397 zio_add_child(spa->spa_async_zio_root, zio);
1404 * ==========================================================================
1405 * Reexecute or suspend/resume failed I/O
1406 * ==========================================================================
1410 zio_reexecute(zio_t *pio)
1412 zio_t *cio, *cio_next;
1414 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1415 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1416 ASSERT(pio->io_gang_leader == NULL);
1417 ASSERT(pio->io_gang_tree == NULL);
1419 pio->io_flags = pio->io_orig_flags;
1420 pio->io_stage = pio->io_orig_stage;
1421 pio->io_pipeline = pio->io_orig_pipeline;
1422 pio->io_reexecute = 0;
1423 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1425 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1426 pio->io_state[w] = 0;
1427 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1428 pio->io_child_error[c] = 0;
1430 if (IO_IS_ALLOCATING(pio))
1431 BP_ZERO(pio->io_bp);
1434 * As we reexecute pio's children, new children could be created.
1435 * New children go to the head of pio's io_child_list, however,
1436 * so we will (correctly) not reexecute them. The key is that
1437 * the remainder of pio's io_child_list, from 'cio_next' onward,
1438 * cannot be affected by any side effects of reexecuting 'cio'.
1440 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1441 cio_next = zio_walk_children(pio);
1442 mutex_enter(&pio->io_lock);
1443 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1444 pio->io_children[cio->io_child_type][w]++;
1445 mutex_exit(&pio->io_lock);
1450 * Now that all children have been reexecuted, execute the parent.
1451 * We don't reexecute "The Godfather" I/O here as it's the
1452 * responsibility of the caller to wait on him.
1454 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1459 zio_suspend(spa_t *spa, zio_t *zio)
1461 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1462 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1463 "failure and the failure mode property for this pool "
1464 "is set to panic.", spa_name(spa));
1466 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1468 mutex_enter(&spa->spa_suspend_lock);
1470 if (spa->spa_suspend_zio_root == NULL)
1471 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1472 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1473 ZIO_FLAG_GODFATHER);
1475 spa->spa_suspended = B_TRUE;
1478 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1479 ASSERT(zio != spa->spa_suspend_zio_root);
1480 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1481 ASSERT(zio_unique_parent(zio) == NULL);
1482 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1483 zio_add_child(spa->spa_suspend_zio_root, zio);
1486 mutex_exit(&spa->spa_suspend_lock);
1490 zio_resume(spa_t *spa)
1495 * Reexecute all previously suspended i/o.
1497 mutex_enter(&spa->spa_suspend_lock);
1498 spa->spa_suspended = B_FALSE;
1499 cv_broadcast(&spa->spa_suspend_cv);
1500 pio = spa->spa_suspend_zio_root;
1501 spa->spa_suspend_zio_root = NULL;
1502 mutex_exit(&spa->spa_suspend_lock);
1508 return (zio_wait(pio));
1512 zio_resume_wait(spa_t *spa)
1514 mutex_enter(&spa->spa_suspend_lock);
1515 while (spa_suspended(spa))
1516 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1517 mutex_exit(&spa->spa_suspend_lock);
1521 * ==========================================================================
1524 * A gang block is a collection of small blocks that looks to the DMU
1525 * like one large block. When zio_dva_allocate() cannot find a block
1526 * of the requested size, due to either severe fragmentation or the pool
1527 * being nearly full, it calls zio_write_gang_block() to construct the
1528 * block from smaller fragments.
1530 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1531 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1532 * an indirect block: it's an array of block pointers. It consumes
1533 * only one sector and hence is allocatable regardless of fragmentation.
1534 * The gang header's bps point to its gang members, which hold the data.
1536 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1537 * as the verifier to ensure uniqueness of the SHA256 checksum.
1538 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1539 * not the gang header. This ensures that data block signatures (needed for
1540 * deduplication) are independent of how the block is physically stored.
1542 * Gang blocks can be nested: a gang member may itself be a gang block.
1543 * Thus every gang block is a tree in which root and all interior nodes are
1544 * gang headers, and the leaves are normal blocks that contain user data.
1545 * The root of the gang tree is called the gang leader.
1547 * To perform any operation (read, rewrite, free, claim) on a gang block,
1548 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1549 * in the io_gang_tree field of the original logical i/o by recursively
1550 * reading the gang leader and all gang headers below it. This yields
1551 * an in-core tree containing the contents of every gang header and the
1552 * bps for every constituent of the gang block.
1554 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1555 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1556 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1557 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1558 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1559 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1560 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1561 * of the gang header plus zio_checksum_compute() of the data to update the
1562 * gang header's blk_cksum as described above.
1564 * The two-phase assemble/issue model solves the problem of partial failure --
1565 * what if you'd freed part of a gang block but then couldn't read the
1566 * gang header for another part? Assembling the entire gang tree first
1567 * ensures that all the necessary gang header I/O has succeeded before
1568 * starting the actual work of free, claim, or write. Once the gang tree
1569 * is assembled, free and claim are in-memory operations that cannot fail.
1571 * In the event that a gang write fails, zio_dva_unallocate() walks the
1572 * gang tree to immediately free (i.e. insert back into the space map)
1573 * everything we've allocated. This ensures that we don't get ENOSPC
1574 * errors during repeated suspend/resume cycles due to a flaky device.
1576 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1577 * the gang tree, we won't modify the block, so we can safely defer the free
1578 * (knowing that the block is still intact). If we *can* assemble the gang
1579 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1580 * each constituent bp and we can allocate a new block on the next sync pass.
1582 * In all cases, the gang tree allows complete recovery from partial failure.
1583 * ==========================================================================
1587 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1592 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1593 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1594 &pio->io_bookmark));
1598 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1603 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1604 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1605 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1607 * As we rewrite each gang header, the pipeline will compute
1608 * a new gang block header checksum for it; but no one will
1609 * compute a new data checksum, so we do that here. The one
1610 * exception is the gang leader: the pipeline already computed
1611 * its data checksum because that stage precedes gang assembly.
1612 * (Presently, nothing actually uses interior data checksums;
1613 * this is just good hygiene.)
1615 if (gn != pio->io_gang_leader->io_gang_tree) {
1616 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1617 data, BP_GET_PSIZE(bp));
1620 * If we are here to damage data for testing purposes,
1621 * leave the GBH alone so that we can detect the damage.
1623 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1624 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1626 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1627 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1628 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1636 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1638 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1639 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp),
1640 ZIO_GANG_CHILD_FLAGS(pio)));
1645 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1647 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1648 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1651 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1660 static void zio_gang_tree_assemble_done(zio_t *zio);
1662 static zio_gang_node_t *
1663 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1665 zio_gang_node_t *gn;
1667 ASSERT(*gnpp == NULL);
1669 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1670 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1677 zio_gang_node_free(zio_gang_node_t **gnpp)
1679 zio_gang_node_t *gn = *gnpp;
1681 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1682 ASSERT(gn->gn_child[g] == NULL);
1684 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1685 kmem_free(gn, sizeof (*gn));
1690 zio_gang_tree_free(zio_gang_node_t **gnpp)
1692 zio_gang_node_t *gn = *gnpp;
1697 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1698 zio_gang_tree_free(&gn->gn_child[g]);
1700 zio_gang_node_free(gnpp);
1704 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1706 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1708 ASSERT(gio->io_gang_leader == gio);
1709 ASSERT(BP_IS_GANG(bp));
1711 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1712 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1713 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1717 zio_gang_tree_assemble_done(zio_t *zio)
1719 zio_t *gio = zio->io_gang_leader;
1720 zio_gang_node_t *gn = zio->io_private;
1721 blkptr_t *bp = zio->io_bp;
1723 ASSERT(gio == zio_unique_parent(zio));
1724 ASSERT(zio->io_child_count == 0);
1729 if (BP_SHOULD_BYTESWAP(bp))
1730 byteswap_uint64_array(zio->io_data, zio->io_size);
1732 ASSERT(zio->io_data == gn->gn_gbh);
1733 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1734 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1736 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1737 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1738 if (!BP_IS_GANG(gbp))
1740 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1745 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1747 zio_t *gio = pio->io_gang_leader;
1750 ASSERT(BP_IS_GANG(bp) == !!gn);
1751 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1752 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1755 * If you're a gang header, your data is in gn->gn_gbh.
1756 * If you're a gang member, your data is in 'data' and gn == NULL.
1758 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1761 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1763 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1764 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1765 if (BP_IS_HOLE(gbp))
1767 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1768 data = (char *)data + BP_GET_PSIZE(gbp);
1772 if (gn == gio->io_gang_tree && gio->io_data != NULL)
1773 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1780 zio_gang_assemble(zio_t *zio)
1782 blkptr_t *bp = zio->io_bp;
1784 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1785 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1787 zio->io_gang_leader = zio;
1789 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1791 return (ZIO_PIPELINE_CONTINUE);
1795 zio_gang_issue(zio_t *zio)
1797 blkptr_t *bp = zio->io_bp;
1799 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1800 return (ZIO_PIPELINE_STOP);
1802 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1803 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1805 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1806 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1808 zio_gang_tree_free(&zio->io_gang_tree);
1810 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1812 return (ZIO_PIPELINE_CONTINUE);
1816 zio_write_gang_member_ready(zio_t *zio)
1818 zio_t *pio = zio_unique_parent(zio);
1819 zio_t *gio = zio->io_gang_leader;
1820 dva_t *cdva = zio->io_bp->blk_dva;
1821 dva_t *pdva = pio->io_bp->blk_dva;
1824 if (BP_IS_HOLE(zio->io_bp))
1827 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1829 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1830 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1831 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1832 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1833 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1835 mutex_enter(&pio->io_lock);
1836 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1837 ASSERT(DVA_GET_GANG(&pdva[d]));
1838 asize = DVA_GET_ASIZE(&pdva[d]);
1839 asize += DVA_GET_ASIZE(&cdva[d]);
1840 DVA_SET_ASIZE(&pdva[d], asize);
1842 mutex_exit(&pio->io_lock);
1846 zio_write_gang_block(zio_t *pio)
1848 spa_t *spa = pio->io_spa;
1849 blkptr_t *bp = pio->io_bp;
1850 zio_t *gio = pio->io_gang_leader;
1852 zio_gang_node_t *gn, **gnpp;
1853 zio_gbh_phys_t *gbh;
1854 uint64_t txg = pio->io_txg;
1855 uint64_t resid = pio->io_size;
1857 int copies = gio->io_prop.zp_copies;
1858 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1862 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1863 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1864 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1866 pio->io_error = error;
1867 return (ZIO_PIPELINE_CONTINUE);
1871 gnpp = &gio->io_gang_tree;
1873 gnpp = pio->io_private;
1874 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1877 gn = zio_gang_node_alloc(gnpp);
1879 bzero(gbh, SPA_GANGBLOCKSIZE);
1882 * Create the gang header.
1884 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1885 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1888 * Create and nowait the gang children.
1890 for (int g = 0; resid != 0; resid -= lsize, g++) {
1891 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1893 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1895 zp.zp_checksum = gio->io_prop.zp_checksum;
1896 zp.zp_compress = ZIO_COMPRESS_OFF;
1897 zp.zp_type = DMU_OT_NONE;
1899 zp.zp_copies = gio->io_prop.zp_copies;
1900 zp.zp_dedup = B_FALSE;
1901 zp.zp_dedup_verify = B_FALSE;
1902 zp.zp_nopwrite = B_FALSE;
1904 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1905 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1906 zio_write_gang_member_ready, NULL, NULL, &gn->gn_child[g],
1907 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1908 &pio->io_bookmark));
1912 * Set pio's pipeline to just wait for zio to finish.
1914 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1918 return (ZIO_PIPELINE_CONTINUE);
1922 * The zio_nop_write stage in the pipeline determines if allocating
1923 * a new bp is necessary. By leveraging a cryptographically secure checksum,
1924 * such as SHA256, we can compare the checksums of the new data and the old
1925 * to determine if allocating a new block is required. The nopwrite
1926 * feature can handle writes in either syncing or open context (i.e. zil
1927 * writes) and as a result is mutually exclusive with dedup.
1930 zio_nop_write(zio_t *zio)
1932 blkptr_t *bp = zio->io_bp;
1933 blkptr_t *bp_orig = &zio->io_bp_orig;
1934 zio_prop_t *zp = &zio->io_prop;
1936 ASSERT(BP_GET_LEVEL(bp) == 0);
1937 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1938 ASSERT(zp->zp_nopwrite);
1939 ASSERT(!zp->zp_dedup);
1940 ASSERT(zio->io_bp_override == NULL);
1941 ASSERT(IO_IS_ALLOCATING(zio));
1944 * Check to see if the original bp and the new bp have matching
1945 * characteristics (i.e. same checksum, compression algorithms, etc).
1946 * If they don't then just continue with the pipeline which will
1947 * allocate a new bp.
1949 if (BP_IS_HOLE(bp_orig) ||
1950 !zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup ||
1951 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
1952 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
1953 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
1954 zp->zp_copies != BP_GET_NDVAS(bp_orig))
1955 return (ZIO_PIPELINE_CONTINUE);
1958 * If the checksums match then reset the pipeline so that we
1959 * avoid allocating a new bp and issuing any I/O.
1961 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
1962 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup);
1963 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
1964 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
1965 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
1966 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
1967 sizeof (uint64_t)) == 0);
1970 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1971 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1974 return (ZIO_PIPELINE_CONTINUE);
1978 * ==========================================================================
1980 * ==========================================================================
1983 zio_ddt_child_read_done(zio_t *zio)
1985 blkptr_t *bp = zio->io_bp;
1986 ddt_entry_t *dde = zio->io_private;
1988 zio_t *pio = zio_unique_parent(zio);
1990 mutex_enter(&pio->io_lock);
1991 ddp = ddt_phys_select(dde, bp);
1992 if (zio->io_error == 0)
1993 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1994 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1995 dde->dde_repair_data = zio->io_data;
1997 zio_buf_free(zio->io_data, zio->io_size);
1998 mutex_exit(&pio->io_lock);
2002 zio_ddt_read_start(zio_t *zio)
2004 blkptr_t *bp = zio->io_bp;
2006 ASSERT(BP_GET_DEDUP(bp));
2007 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2008 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2010 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2011 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2012 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2013 ddt_phys_t *ddp = dde->dde_phys;
2014 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2017 ASSERT(zio->io_vsd == NULL);
2020 if (ddp_self == NULL)
2021 return (ZIO_PIPELINE_CONTINUE);
2023 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2024 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2026 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2028 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2029 zio_buf_alloc(zio->io_size), zio->io_size,
2030 zio_ddt_child_read_done, dde, zio->io_priority,
2031 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
2032 &zio->io_bookmark));
2034 return (ZIO_PIPELINE_CONTINUE);
2037 zio_nowait(zio_read(zio, zio->io_spa, bp,
2038 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
2039 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2041 return (ZIO_PIPELINE_CONTINUE);
2045 zio_ddt_read_done(zio_t *zio)
2047 blkptr_t *bp = zio->io_bp;
2049 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2050 return (ZIO_PIPELINE_STOP);
2052 ASSERT(BP_GET_DEDUP(bp));
2053 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2054 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2056 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2057 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2058 ddt_entry_t *dde = zio->io_vsd;
2060 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2061 return (ZIO_PIPELINE_CONTINUE);
2064 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2065 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2066 return (ZIO_PIPELINE_STOP);
2068 if (dde->dde_repair_data != NULL) {
2069 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2070 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2072 ddt_repair_done(ddt, dde);
2076 ASSERT(zio->io_vsd == NULL);
2078 return (ZIO_PIPELINE_CONTINUE);
2082 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2084 spa_t *spa = zio->io_spa;
2087 * Note: we compare the original data, not the transformed data,
2088 * because when zio->io_bp is an override bp, we will not have
2089 * pushed the I/O transforms. That's an important optimization
2090 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2092 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2093 zio_t *lio = dde->dde_lead_zio[p];
2096 return (lio->io_orig_size != zio->io_orig_size ||
2097 bcmp(zio->io_orig_data, lio->io_orig_data,
2098 zio->io_orig_size) != 0);
2102 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2103 ddt_phys_t *ddp = &dde->dde_phys[p];
2105 if (ddp->ddp_phys_birth != 0) {
2106 arc_buf_t *abuf = NULL;
2107 uint32_t aflags = ARC_WAIT;
2108 blkptr_t blk = *zio->io_bp;
2111 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2115 error = arc_read(NULL, spa, &blk,
2116 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2117 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2118 &aflags, &zio->io_bookmark);
2121 if (arc_buf_size(abuf) != zio->io_orig_size ||
2122 bcmp(abuf->b_data, zio->io_orig_data,
2123 zio->io_orig_size) != 0)
2124 error = SET_ERROR(EEXIST);
2125 VERIFY(arc_buf_remove_ref(abuf, &abuf));
2129 return (error != 0);
2137 zio_ddt_child_write_ready(zio_t *zio)
2139 int p = zio->io_prop.zp_copies;
2140 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2141 ddt_entry_t *dde = zio->io_private;
2142 ddt_phys_t *ddp = &dde->dde_phys[p];
2150 ASSERT(dde->dde_lead_zio[p] == zio);
2152 ddt_phys_fill(ddp, zio->io_bp);
2154 while ((pio = zio_walk_parents(zio)) != NULL)
2155 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2161 zio_ddt_child_write_done(zio_t *zio)
2163 int p = zio->io_prop.zp_copies;
2164 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2165 ddt_entry_t *dde = zio->io_private;
2166 ddt_phys_t *ddp = &dde->dde_phys[p];
2170 ASSERT(ddp->ddp_refcnt == 0);
2171 ASSERT(dde->dde_lead_zio[p] == zio);
2172 dde->dde_lead_zio[p] = NULL;
2174 if (zio->io_error == 0) {
2175 while (zio_walk_parents(zio) != NULL)
2176 ddt_phys_addref(ddp);
2178 ddt_phys_clear(ddp);
2185 zio_ddt_ditto_write_done(zio_t *zio)
2187 int p = DDT_PHYS_DITTO;
2188 zio_prop_t *zp = &zio->io_prop;
2189 blkptr_t *bp = zio->io_bp;
2190 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2191 ddt_entry_t *dde = zio->io_private;
2192 ddt_phys_t *ddp = &dde->dde_phys[p];
2193 ddt_key_t *ddk = &dde->dde_key;
2197 ASSERT(ddp->ddp_refcnt == 0);
2198 ASSERT(dde->dde_lead_zio[p] == zio);
2199 dde->dde_lead_zio[p] = NULL;
2201 if (zio->io_error == 0) {
2202 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2203 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2204 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2205 if (ddp->ddp_phys_birth != 0)
2206 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2207 ddt_phys_fill(ddp, bp);
2214 zio_ddt_write(zio_t *zio)
2216 spa_t *spa = zio->io_spa;
2217 blkptr_t *bp = zio->io_bp;
2218 uint64_t txg = zio->io_txg;
2219 zio_prop_t *zp = &zio->io_prop;
2220 int p = zp->zp_copies;
2224 ddt_t *ddt = ddt_select(spa, bp);
2228 ASSERT(BP_GET_DEDUP(bp));
2229 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2230 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2233 dde = ddt_lookup(ddt, bp, B_TRUE);
2234 ddp = &dde->dde_phys[p];
2236 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2238 * If we're using a weak checksum, upgrade to a strong checksum
2239 * and try again. If we're already using a strong checksum,
2240 * we can't resolve it, so just convert to an ordinary write.
2241 * (And automatically e-mail a paper to Nature?)
2243 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2244 zp->zp_checksum = spa_dedup_checksum(spa);
2245 zio_pop_transforms(zio);
2246 zio->io_stage = ZIO_STAGE_OPEN;
2249 zp->zp_dedup = B_FALSE;
2251 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2253 return (ZIO_PIPELINE_CONTINUE);
2256 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2257 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2259 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2260 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2261 zio_prop_t czp = *zp;
2263 czp.zp_copies = ditto_copies;
2266 * If we arrived here with an override bp, we won't have run
2267 * the transform stack, so we won't have the data we need to
2268 * generate a child i/o. So, toss the override bp and restart.
2269 * This is safe, because using the override bp is just an
2270 * optimization; and it's rare, so the cost doesn't matter.
2272 if (zio->io_bp_override) {
2273 zio_pop_transforms(zio);
2274 zio->io_stage = ZIO_STAGE_OPEN;
2275 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2276 zio->io_bp_override = NULL;
2279 return (ZIO_PIPELINE_CONTINUE);
2282 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2283 zio->io_orig_size, &czp, NULL, NULL,
2284 zio_ddt_ditto_write_done, dde, zio->io_priority,
2285 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2287 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2288 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2291 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2292 if (ddp->ddp_phys_birth != 0)
2293 ddt_bp_fill(ddp, bp, txg);
2294 if (dde->dde_lead_zio[p] != NULL)
2295 zio_add_child(zio, dde->dde_lead_zio[p]);
2297 ddt_phys_addref(ddp);
2298 } else if (zio->io_bp_override) {
2299 ASSERT(bp->blk_birth == txg);
2300 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2301 ddt_phys_fill(ddp, bp);
2302 ddt_phys_addref(ddp);
2304 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2305 zio->io_orig_size, zp, zio_ddt_child_write_ready, NULL,
2306 zio_ddt_child_write_done, dde, zio->io_priority,
2307 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2309 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2310 dde->dde_lead_zio[p] = cio;
2320 return (ZIO_PIPELINE_CONTINUE);
2323 ddt_entry_t *freedde; /* for debugging */
2326 zio_ddt_free(zio_t *zio)
2328 spa_t *spa = zio->io_spa;
2329 blkptr_t *bp = zio->io_bp;
2330 ddt_t *ddt = ddt_select(spa, bp);
2334 ASSERT(BP_GET_DEDUP(bp));
2335 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2338 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2339 ddp = ddt_phys_select(dde, bp);
2340 ddt_phys_decref(ddp);
2343 return (ZIO_PIPELINE_CONTINUE);
2347 * ==========================================================================
2348 * Allocate and free blocks
2349 * ==========================================================================
2352 zio_dva_allocate(zio_t *zio)
2354 spa_t *spa = zio->io_spa;
2355 metaslab_class_t *mc = spa_normal_class(spa);
2356 blkptr_t *bp = zio->io_bp;
2360 if (zio->io_gang_leader == NULL) {
2361 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2362 zio->io_gang_leader = zio;
2365 ASSERT(BP_IS_HOLE(bp));
2366 ASSERT0(BP_GET_NDVAS(bp));
2367 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2368 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2369 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2372 * The dump device does not support gang blocks so allocation on
2373 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2374 * the "fast" gang feature.
2376 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2377 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2378 METASLAB_GANG_CHILD : 0;
2379 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2380 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2383 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2384 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2386 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2387 return (zio_write_gang_block(zio));
2388 zio->io_error = error;
2391 return (ZIO_PIPELINE_CONTINUE);
2395 zio_dva_free(zio_t *zio)
2397 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2399 return (ZIO_PIPELINE_CONTINUE);
2403 zio_dva_claim(zio_t *zio)
2407 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2409 zio->io_error = error;
2411 return (ZIO_PIPELINE_CONTINUE);
2415 * Undo an allocation. This is used by zio_done() when an I/O fails
2416 * and we want to give back the block we just allocated.
2417 * This handles both normal blocks and gang blocks.
2420 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2422 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2423 ASSERT(zio->io_bp_override == NULL);
2425 if (!BP_IS_HOLE(bp))
2426 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2429 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2430 zio_dva_unallocate(zio, gn->gn_child[g],
2431 &gn->gn_gbh->zg_blkptr[g]);
2437 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2440 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2441 uint64_t size, boolean_t use_slog)
2445 ASSERT(txg > spa_syncing_txg(spa));
2448 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2449 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2450 * when allocating them.
2453 error = metaslab_alloc(spa, spa_log_class(spa), size,
2454 new_bp, 1, txg, old_bp,
2455 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2459 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2460 new_bp, 1, txg, old_bp,
2461 METASLAB_HINTBP_AVOID);
2465 BP_SET_LSIZE(new_bp, size);
2466 BP_SET_PSIZE(new_bp, size);
2467 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2468 BP_SET_CHECKSUM(new_bp,
2469 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2470 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2471 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2472 BP_SET_LEVEL(new_bp, 0);
2473 BP_SET_DEDUP(new_bp, 0);
2474 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2481 * Free an intent log block.
2484 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2486 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2487 ASSERT(!BP_IS_GANG(bp));
2489 zio_free(spa, txg, bp);
2493 * ==========================================================================
2494 * Read, write and delete to physical devices
2495 * ==========================================================================
2498 zio_vdev_io_start(zio_t *zio)
2500 vdev_t *vd = zio->io_vd;
2502 spa_t *spa = zio->io_spa;
2504 ASSERT(zio->io_error == 0);
2505 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2508 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2509 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2512 * The mirror_ops handle multiple DVAs in a single BP.
2514 return (vdev_mirror_ops.vdev_op_io_start(zio));
2517 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE) {
2518 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg);
2519 return (ZIO_PIPELINE_CONTINUE);
2523 * We keep track of time-sensitive I/Os so that the scan thread
2524 * can quickly react to certain workloads. In particular, we care
2525 * about non-scrubbing, top-level reads and writes with the following
2527 * - synchronous writes of user data to non-slog devices
2528 * - any reads of user data
2529 * When these conditions are met, adjust the timestamp of spa_last_io
2530 * which allows the scan thread to adjust its workload accordingly.
2532 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2533 vd == vd->vdev_top && !vd->vdev_islog &&
2534 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2535 zio->io_txg != spa_syncing_txg(spa)) {
2536 uint64_t old = spa->spa_last_io;
2537 uint64_t new = ddi_get_lbolt64();
2539 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2542 align = 1ULL << vd->vdev_top->vdev_ashift;
2544 if (P2PHASE(zio->io_size, align) != 0) {
2545 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2547 if (zio->io_type == ZIO_TYPE_READ ||
2548 zio->io_type == ZIO_TYPE_WRITE)
2549 abuf = zio_buf_alloc(asize);
2550 ASSERT(vd == vd->vdev_top);
2551 if (zio->io_type == ZIO_TYPE_WRITE) {
2552 bcopy(zio->io_data, abuf, zio->io_size);
2553 bzero(abuf + zio->io_size, asize - zio->io_size);
2555 zio_push_transform(zio, abuf, asize, abuf ? asize : 0,
2559 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2560 ASSERT(P2PHASE(zio->io_size, align) == 0);
2561 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa));
2564 * If this is a repair I/O, and there's no self-healing involved --
2565 * that is, we're just resilvering what we expect to resilver --
2566 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2567 * This prevents spurious resilvering with nested replication.
2568 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2569 * A is out of date, we'll read from C+D, then use the data to
2570 * resilver A+B -- but we don't actually want to resilver B, just A.
2571 * The top-level mirror has no way to know this, so instead we just
2572 * discard unnecessary repairs as we work our way down the vdev tree.
2573 * The same logic applies to any form of nested replication:
2574 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2576 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2577 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2578 zio->io_txg != 0 && /* not a delegated i/o */
2579 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2580 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2581 zio_vdev_io_bypass(zio);
2582 return (ZIO_PIPELINE_CONTINUE);
2585 if (vd->vdev_ops->vdev_op_leaf &&
2586 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2588 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio))
2589 return (ZIO_PIPELINE_CONTINUE);
2591 if ((zio = vdev_queue_io(zio)) == NULL)
2592 return (ZIO_PIPELINE_STOP);
2594 if (!vdev_accessible(vd, zio)) {
2595 zio->io_error = SET_ERROR(ENXIO);
2597 return (ZIO_PIPELINE_STOP);
2602 * Note that we ignore repair writes for TRIM because they can conflict
2603 * with normal writes. This isn't an issue because, by definition, we
2604 * only repair blocks that aren't freed.
2606 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_WRITE &&
2607 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2608 if (!trim_map_write_start(zio))
2609 return (ZIO_PIPELINE_STOP);
2612 return (vd->vdev_ops->vdev_op_io_start(zio));
2616 zio_vdev_io_done(zio_t *zio)
2618 vdev_t *vd = zio->io_vd;
2619 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2620 boolean_t unexpected_error = B_FALSE;
2622 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2623 return (ZIO_PIPELINE_STOP);
2625 ASSERT(zio->io_type == ZIO_TYPE_READ ||
2626 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE);
2628 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2629 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2631 if (zio->io_type == ZIO_TYPE_WRITE &&
2632 !(zio->io_flags & ZIO_FLAG_IO_REPAIR))
2633 trim_map_write_done(zio);
2635 vdev_queue_io_done(zio);
2637 if (zio->io_type == ZIO_TYPE_WRITE)
2638 vdev_cache_write(zio);
2640 if (zio_injection_enabled && zio->io_error == 0)
2641 zio->io_error = zio_handle_device_injection(vd,
2644 if (zio_injection_enabled && zio->io_error == 0)
2645 zio->io_error = zio_handle_label_injection(zio, EIO);
2647 if (zio->io_error) {
2648 if (!vdev_accessible(vd, zio)) {
2649 zio->io_error = SET_ERROR(ENXIO);
2651 unexpected_error = B_TRUE;
2656 ops->vdev_op_io_done(zio);
2658 if (unexpected_error)
2659 VERIFY(vdev_probe(vd, zio) == NULL);
2661 return (ZIO_PIPELINE_CONTINUE);
2665 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2666 * disk, and use that to finish the checksum ereport later.
2669 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2670 const void *good_buf)
2672 /* no processing needed */
2673 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2678 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2680 void *buf = zio_buf_alloc(zio->io_size);
2682 bcopy(zio->io_data, buf, zio->io_size);
2684 zcr->zcr_cbinfo = zio->io_size;
2685 zcr->zcr_cbdata = buf;
2686 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2687 zcr->zcr_free = zio_buf_free;
2691 zio_vdev_io_assess(zio_t *zio)
2693 vdev_t *vd = zio->io_vd;
2695 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2696 return (ZIO_PIPELINE_STOP);
2698 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2699 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2701 if (zio->io_vsd != NULL) {
2702 zio->io_vsd_ops->vsd_free(zio);
2706 if (zio_injection_enabled && zio->io_error == 0)
2707 zio->io_error = zio_handle_fault_injection(zio, EIO);
2709 if (zio->io_type == ZIO_TYPE_IOCTL && zio->io_cmd == DKIOCTRIM)
2710 switch (zio->io_error) {
2712 ZIO_TRIM_STAT_INCR(bytes, zio->io_size);
2713 ZIO_TRIM_STAT_BUMP(success);
2716 ZIO_TRIM_STAT_BUMP(unsupported);
2719 ZIO_TRIM_STAT_BUMP(failed);
2724 * If the I/O failed, determine whether we should attempt to retry it.
2726 * On retry, we cut in line in the issue queue, since we don't want
2727 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2729 if (zio->io_error && vd == NULL &&
2730 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2731 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2732 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2734 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2735 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2736 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2737 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2738 zio_requeue_io_start_cut_in_line);
2739 return (ZIO_PIPELINE_STOP);
2743 * If we got an error on a leaf device, convert it to ENXIO
2744 * if the device is not accessible at all.
2746 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2747 !vdev_accessible(vd, zio))
2748 zio->io_error = SET_ERROR(ENXIO);
2751 * If we can't write to an interior vdev (mirror or RAID-Z),
2752 * set vdev_cant_write so that we stop trying to allocate from it.
2754 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2755 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
2756 vd->vdev_cant_write = B_TRUE;
2760 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2762 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2763 zio->io_physdone != NULL) {
2764 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
2765 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
2766 zio->io_physdone(zio->io_logical);
2769 return (ZIO_PIPELINE_CONTINUE);
2773 zio_vdev_io_reissue(zio_t *zio)
2775 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2776 ASSERT(zio->io_error == 0);
2778 zio->io_stage >>= 1;
2782 zio_vdev_io_redone(zio_t *zio)
2784 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2786 zio->io_stage >>= 1;
2790 zio_vdev_io_bypass(zio_t *zio)
2792 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2793 ASSERT(zio->io_error == 0);
2795 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2796 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2800 * ==========================================================================
2801 * Generate and verify checksums
2802 * ==========================================================================
2805 zio_checksum_generate(zio_t *zio)
2807 blkptr_t *bp = zio->io_bp;
2808 enum zio_checksum checksum;
2812 * This is zio_write_phys().
2813 * We're either generating a label checksum, or none at all.
2815 checksum = zio->io_prop.zp_checksum;
2817 if (checksum == ZIO_CHECKSUM_OFF)
2818 return (ZIO_PIPELINE_CONTINUE);
2820 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2822 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2823 ASSERT(!IO_IS_ALLOCATING(zio));
2824 checksum = ZIO_CHECKSUM_GANG_HEADER;
2826 checksum = BP_GET_CHECKSUM(bp);
2830 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2832 return (ZIO_PIPELINE_CONTINUE);
2836 zio_checksum_verify(zio_t *zio)
2838 zio_bad_cksum_t info;
2839 blkptr_t *bp = zio->io_bp;
2842 ASSERT(zio->io_vd != NULL);
2846 * This is zio_read_phys().
2847 * We're either verifying a label checksum, or nothing at all.
2849 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2850 return (ZIO_PIPELINE_CONTINUE);
2852 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2855 if ((error = zio_checksum_error(zio, &info)) != 0) {
2856 zio->io_error = error;
2857 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2858 zfs_ereport_start_checksum(zio->io_spa,
2859 zio->io_vd, zio, zio->io_offset,
2860 zio->io_size, NULL, &info);
2864 return (ZIO_PIPELINE_CONTINUE);
2868 * Called by RAID-Z to ensure we don't compute the checksum twice.
2871 zio_checksum_verified(zio_t *zio)
2873 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2877 * ==========================================================================
2878 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2879 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2880 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2881 * indicate errors that are specific to one I/O, and most likely permanent.
2882 * Any other error is presumed to be worse because we weren't expecting it.
2883 * ==========================================================================
2886 zio_worst_error(int e1, int e2)
2888 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2891 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2892 if (e1 == zio_error_rank[r1])
2895 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2896 if (e2 == zio_error_rank[r2])
2899 return (r1 > r2 ? e1 : e2);
2903 * ==========================================================================
2905 * ==========================================================================
2908 zio_ready(zio_t *zio)
2910 blkptr_t *bp = zio->io_bp;
2911 zio_t *pio, *pio_next;
2913 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2914 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2915 return (ZIO_PIPELINE_STOP);
2917 if (zio->io_ready) {
2918 ASSERT(IO_IS_ALLOCATING(zio));
2919 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
2920 (zio->io_flags & ZIO_FLAG_NOPWRITE));
2921 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2926 if (bp != NULL && bp != &zio->io_bp_copy)
2927 zio->io_bp_copy = *bp;
2930 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2932 mutex_enter(&zio->io_lock);
2933 zio->io_state[ZIO_WAIT_READY] = 1;
2934 pio = zio_walk_parents(zio);
2935 mutex_exit(&zio->io_lock);
2938 * As we notify zio's parents, new parents could be added.
2939 * New parents go to the head of zio's io_parent_list, however,
2940 * so we will (correctly) not notify them. The remainder of zio's
2941 * io_parent_list, from 'pio_next' onward, cannot change because
2942 * all parents must wait for us to be done before they can be done.
2944 for (; pio != NULL; pio = pio_next) {
2945 pio_next = zio_walk_parents(zio);
2946 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2949 if (zio->io_flags & ZIO_FLAG_NODATA) {
2950 if (BP_IS_GANG(bp)) {
2951 zio->io_flags &= ~ZIO_FLAG_NODATA;
2953 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2954 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2958 if (zio_injection_enabled &&
2959 zio->io_spa->spa_syncing_txg == zio->io_txg)
2960 zio_handle_ignored_writes(zio);
2962 return (ZIO_PIPELINE_CONTINUE);
2966 zio_done(zio_t *zio)
2968 spa_t *spa = zio->io_spa;
2969 zio_t *lio = zio->io_logical;
2970 blkptr_t *bp = zio->io_bp;
2971 vdev_t *vd = zio->io_vd;
2972 uint64_t psize = zio->io_size;
2973 zio_t *pio, *pio_next;
2976 * If our children haven't all completed,
2977 * wait for them and then repeat this pipeline stage.
2979 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2980 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2981 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2982 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2983 return (ZIO_PIPELINE_STOP);
2985 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2986 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2987 ASSERT(zio->io_children[c][w] == 0);
2990 ASSERT(bp->blk_pad[0] == 0);
2991 ASSERT(bp->blk_pad[1] == 0);
2992 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2993 (bp == zio_unique_parent(zio)->io_bp));
2994 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
2995 zio->io_bp_override == NULL &&
2996 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2997 ASSERT(!BP_SHOULD_BYTESWAP(bp));
2998 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
2999 ASSERT(BP_COUNT_GANG(bp) == 0 ||
3000 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3002 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3003 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3007 * If there were child vdev/gang/ddt errors, they apply to us now.
3009 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3010 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3011 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3014 * If the I/O on the transformed data was successful, generate any
3015 * checksum reports now while we still have the transformed data.
3017 if (zio->io_error == 0) {
3018 while (zio->io_cksum_report != NULL) {
3019 zio_cksum_report_t *zcr = zio->io_cksum_report;
3020 uint64_t align = zcr->zcr_align;
3021 uint64_t asize = P2ROUNDUP(psize, align);
3022 char *abuf = zio->io_data;
3024 if (asize != psize) {
3025 abuf = zio_buf_alloc(asize);
3026 bcopy(zio->io_data, abuf, psize);
3027 bzero(abuf + psize, asize - psize);
3030 zio->io_cksum_report = zcr->zcr_next;
3031 zcr->zcr_next = NULL;
3032 zcr->zcr_finish(zcr, abuf);
3033 zfs_ereport_free_checksum(zcr);
3036 zio_buf_free(abuf, asize);
3040 zio_pop_transforms(zio); /* note: may set zio->io_error */
3042 vdev_stat_update(zio, psize);
3044 if (zio->io_error) {
3046 * If this I/O is attached to a particular vdev,
3047 * generate an error message describing the I/O failure
3048 * at the block level. We ignore these errors if the
3049 * device is currently unavailable.
3051 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3052 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3054 if ((zio->io_error == EIO || !(zio->io_flags &
3055 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3058 * For logical I/O requests, tell the SPA to log the
3059 * error and generate a logical data ereport.
3061 spa_log_error(spa, zio);
3062 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3067 if (zio->io_error && zio == lio) {
3069 * Determine whether zio should be reexecuted. This will
3070 * propagate all the way to the root via zio_notify_parent().
3072 ASSERT(vd == NULL && bp != NULL);
3073 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3075 if (IO_IS_ALLOCATING(zio) &&
3076 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3077 if (zio->io_error != ENOSPC)
3078 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3080 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3083 if ((zio->io_type == ZIO_TYPE_READ ||
3084 zio->io_type == ZIO_TYPE_FREE) &&
3085 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3086 zio->io_error == ENXIO &&
3087 spa_load_state(spa) == SPA_LOAD_NONE &&
3088 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3089 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3091 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3092 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3095 * Here is a possibly good place to attempt to do
3096 * either combinatorial reconstruction or error correction
3097 * based on checksums. It also might be a good place
3098 * to send out preliminary ereports before we suspend
3104 * If there were logical child errors, they apply to us now.
3105 * We defer this until now to avoid conflating logical child
3106 * errors with errors that happened to the zio itself when
3107 * updating vdev stats and reporting FMA events above.
3109 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3111 if ((zio->io_error || zio->io_reexecute) &&
3112 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3113 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3114 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3116 zio_gang_tree_free(&zio->io_gang_tree);
3119 * Godfather I/Os should never suspend.
3121 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3122 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3123 zio->io_reexecute = 0;
3125 if (zio->io_reexecute) {
3127 * This is a logical I/O that wants to reexecute.
3129 * Reexecute is top-down. When an i/o fails, if it's not
3130 * the root, it simply notifies its parent and sticks around.
3131 * The parent, seeing that it still has children in zio_done(),
3132 * does the same. This percolates all the way up to the root.
3133 * The root i/o will reexecute or suspend the entire tree.
3135 * This approach ensures that zio_reexecute() honors
3136 * all the original i/o dependency relationships, e.g.
3137 * parents not executing until children are ready.
3139 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3141 zio->io_gang_leader = NULL;
3143 mutex_enter(&zio->io_lock);
3144 zio->io_state[ZIO_WAIT_DONE] = 1;
3145 mutex_exit(&zio->io_lock);
3148 * "The Godfather" I/O monitors its children but is
3149 * not a true parent to them. It will track them through
3150 * the pipeline but severs its ties whenever they get into
3151 * trouble (e.g. suspended). This allows "The Godfather"
3152 * I/O to return status without blocking.
3154 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3155 zio_link_t *zl = zio->io_walk_link;
3156 pio_next = zio_walk_parents(zio);
3158 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3159 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3160 zio_remove_child(pio, zio, zl);
3161 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3165 if ((pio = zio_unique_parent(zio)) != NULL) {
3167 * We're not a root i/o, so there's nothing to do
3168 * but notify our parent. Don't propagate errors
3169 * upward since we haven't permanently failed yet.
3171 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3172 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3173 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3174 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3176 * We'd fail again if we reexecuted now, so suspend
3177 * until conditions improve (e.g. device comes online).
3179 zio_suspend(spa, zio);
3182 * Reexecution is potentially a huge amount of work.
3183 * Hand it off to the otherwise-unused claim taskq.
3185 #if defined(illumos) || !defined(_KERNEL)
3186 ASSERT(zio->io_tqent.tqent_next == NULL);
3188 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
3190 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3191 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3194 return (ZIO_PIPELINE_STOP);
3197 ASSERT(zio->io_child_count == 0);
3198 ASSERT(zio->io_reexecute == 0);
3199 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3202 * Report any checksum errors, since the I/O is complete.
3204 while (zio->io_cksum_report != NULL) {
3205 zio_cksum_report_t *zcr = zio->io_cksum_report;
3206 zio->io_cksum_report = zcr->zcr_next;
3207 zcr->zcr_next = NULL;
3208 zcr->zcr_finish(zcr, NULL);
3209 zfs_ereport_free_checksum(zcr);
3213 * It is the responsibility of the done callback to ensure that this
3214 * particular zio is no longer discoverable for adoption, and as
3215 * such, cannot acquire any new parents.
3220 mutex_enter(&zio->io_lock);
3221 zio->io_state[ZIO_WAIT_DONE] = 1;
3222 mutex_exit(&zio->io_lock);
3224 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3225 zio_link_t *zl = zio->io_walk_link;
3226 pio_next = zio_walk_parents(zio);
3227 zio_remove_child(pio, zio, zl);
3228 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3231 if (zio->io_waiter != NULL) {
3232 mutex_enter(&zio->io_lock);
3233 zio->io_executor = NULL;
3234 cv_broadcast(&zio->io_cv);
3235 mutex_exit(&zio->io_lock);
3240 return (ZIO_PIPELINE_STOP);
3244 * ==========================================================================
3245 * I/O pipeline definition
3246 * ==========================================================================
3248 static zio_pipe_stage_t *zio_pipeline[] = {
3254 zio_checksum_generate,
3269 zio_checksum_verify,
3273 /* dnp is the dnode for zb1->zb_object */
3275 zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_t *zb1,
3276 const zbookmark_t *zb2)
3278 uint64_t zb1nextL0, zb2thisobj;
3280 ASSERT(zb1->zb_objset == zb2->zb_objset);
3281 ASSERT(zb2->zb_level == 0);
3284 * A bookmark in the deadlist is considered to be after
3287 if (zb2->zb_object == DMU_DEADLIST_OBJECT)
3290 /* The objset_phys_t isn't before anything. */
3294 zb1nextL0 = (zb1->zb_blkid + 1) <<
3295 ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
3297 zb2thisobj = zb2->zb_object ? zb2->zb_object :
3298 zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
3300 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3301 uint64_t nextobj = zb1nextL0 *
3302 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
3303 return (nextobj <= zb2thisobj);
3306 if (zb1->zb_object < zb2thisobj)
3308 if (zb1->zb_object > zb2thisobj)
3310 if (zb2->zb_object == DMU_META_DNODE_OBJECT)
3312 return (zb1nextL0 <= zb2->zb_blkid);