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 : 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));
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 (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 * GANG and DEDUP blocks can induce a read (for the gang block header,
782 * or the DDT), so issue them asynchronously so that this thread is
785 if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
786 stage |= ZIO_STAGE_ISSUE_ASYNC;
788 zio = zio_create(pio, spa, txg, bp, NULL, size,
789 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags,
790 NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
796 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
797 zio_done_func_t *done, void *private, enum zio_flag flags)
802 * A claim is an allocation of a specific block. Claims are needed
803 * to support immediate writes in the intent log. The issue is that
804 * immediate writes contain committed data, but in a txg that was
805 * *not* committed. Upon opening the pool after an unclean shutdown,
806 * the intent log claims all blocks that contain immediate write data
807 * so that the SPA knows they're in use.
809 * All claims *must* be resolved in the first txg -- before the SPA
810 * starts allocating blocks -- so that nothing is allocated twice.
811 * If txg == 0 we just verify that the block is claimable.
813 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
814 ASSERT(txg == spa_first_txg(spa) || txg == 0);
815 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
817 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
818 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
819 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
825 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset,
826 uint64_t size, zio_done_func_t *done, void *private,
832 if (vd->vdev_children == 0) {
833 zio = zio_create(pio, spa, 0, NULL, NULL, size, done, private,
834 ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, offset, NULL,
835 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
839 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
841 for (c = 0; c < vd->vdev_children; c++)
842 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
843 offset, size, done, private, flags));
850 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
851 void *data, int checksum, zio_done_func_t *done, void *private,
852 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
856 ASSERT(vd->vdev_children == 0);
857 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
858 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
859 ASSERT3U(offset + size, <=, vd->vdev_psize);
861 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
862 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
863 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
865 zio->io_prop.zp_checksum = checksum;
871 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
872 void *data, int checksum, zio_done_func_t *done, void *private,
873 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
877 ASSERT(vd->vdev_children == 0);
878 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
879 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
880 ASSERT3U(offset + size, <=, vd->vdev_psize);
882 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
883 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
884 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
886 zio->io_prop.zp_checksum = checksum;
888 if (zio_checksum_table[checksum].ci_eck) {
890 * zec checksums are necessarily destructive -- they modify
891 * the end of the write buffer to hold the verifier/checksum.
892 * Therefore, we must make a local copy in case the data is
893 * being written to multiple places in parallel.
895 void *wbuf = zio_buf_alloc(size);
896 bcopy(data, wbuf, size);
897 zio_push_transform(zio, wbuf, size, size, NULL);
904 * Create a child I/O to do some work for us.
907 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
908 void *data, uint64_t size, int type, zio_priority_t priority,
909 enum zio_flag flags, zio_done_func_t *done, void *private)
911 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
914 ASSERT(vd->vdev_parent ==
915 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
917 if (type == ZIO_TYPE_READ && bp != NULL) {
919 * If we have the bp, then the child should perform the
920 * checksum and the parent need not. This pushes error
921 * detection as close to the leaves as possible and
922 * eliminates redundant checksums in the interior nodes.
924 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
925 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
928 if (vd->vdev_children == 0)
929 offset += VDEV_LABEL_START_SIZE;
931 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
934 * If we've decided to do a repair, the write is not speculative --
935 * even if the original read was.
937 if (flags & ZIO_FLAG_IO_REPAIR)
938 flags &= ~ZIO_FLAG_SPECULATIVE;
940 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
941 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
942 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
944 zio->io_physdone = pio->io_physdone;
945 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
946 zio->io_logical->io_phys_children++;
952 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
953 int type, zio_priority_t priority, enum zio_flag flags,
954 zio_done_func_t *done, void *private)
958 ASSERT(vd->vdev_ops->vdev_op_leaf);
960 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
961 data, size, done, private, type, priority,
962 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
964 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
970 zio_flush(zio_t *zio, vdev_t *vd)
972 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0,
974 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
978 zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size)
981 ASSERT(vd->vdev_ops->vdev_op_leaf);
983 return zio_ioctl(zio, spa, vd, DKIOCTRIM, offset, size,
985 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY);
989 zio_shrink(zio_t *zio, uint64_t size)
991 ASSERT(zio->io_executor == NULL);
992 ASSERT(zio->io_orig_size == zio->io_size);
993 ASSERT(size <= zio->io_size);
996 * We don't shrink for raidz because of problems with the
997 * reconstruction when reading back less than the block size.
998 * Note, BP_IS_RAIDZ() assumes no compression.
1000 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1001 if (!BP_IS_RAIDZ(zio->io_bp))
1002 zio->io_orig_size = zio->io_size = size;
1006 * ==========================================================================
1007 * Prepare to read and write logical blocks
1008 * ==========================================================================
1012 zio_read_bp_init(zio_t **ziop)
1015 blkptr_t *bp = zio->io_bp;
1017 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1018 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1019 !(zio->io_flags & ZIO_FLAG_RAW)) {
1020 uint64_t psize = BP_GET_PSIZE(bp);
1021 void *cbuf = zio_buf_alloc(psize);
1023 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
1026 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1027 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1029 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1030 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1032 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1033 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1035 return (ZIO_PIPELINE_CONTINUE);
1039 zio_write_bp_init(zio_t **ziop)
1042 spa_t *spa = zio->io_spa;
1043 zio_prop_t *zp = &zio->io_prop;
1044 enum zio_compress compress = zp->zp_compress;
1045 blkptr_t *bp = zio->io_bp;
1046 uint64_t lsize = zio->io_size;
1047 uint64_t psize = lsize;
1051 * If our children haven't all reached the ready stage,
1052 * wait for them and then repeat this pipeline stage.
1054 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1055 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1056 return (ZIO_PIPELINE_STOP);
1058 if (!IO_IS_ALLOCATING(zio))
1059 return (ZIO_PIPELINE_CONTINUE);
1061 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1063 if (zio->io_bp_override) {
1064 ASSERT(bp->blk_birth != zio->io_txg);
1065 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1067 *bp = *zio->io_bp_override;
1068 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1071 * If we've been overridden and nopwrite is set then
1072 * set the flag accordingly to indicate that a nopwrite
1073 * has already occurred.
1075 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1076 ASSERT(!zp->zp_dedup);
1077 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1078 return (ZIO_PIPELINE_CONTINUE);
1081 ASSERT(!zp->zp_nopwrite);
1083 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1084 return (ZIO_PIPELINE_CONTINUE);
1086 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1087 zp->zp_dedup_verify);
1089 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1090 BP_SET_DEDUP(bp, 1);
1091 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1092 return (ZIO_PIPELINE_CONTINUE);
1094 zio->io_bp_override = NULL;
1098 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1100 * We're rewriting an existing block, which means we're
1101 * working on behalf of spa_sync(). For spa_sync() to
1102 * converge, it must eventually be the case that we don't
1103 * have to allocate new blocks. But compression changes
1104 * the blocksize, which forces a reallocate, and makes
1105 * convergence take longer. Therefore, after the first
1106 * few passes, stop compressing to ensure convergence.
1108 pass = spa_sync_pass(spa);
1110 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1111 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1112 ASSERT(!BP_GET_DEDUP(bp));
1114 if (pass >= zfs_sync_pass_dont_compress)
1115 compress = ZIO_COMPRESS_OFF;
1117 /* Make sure someone doesn't change their mind on overwrites */
1118 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
1119 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1122 if (compress != ZIO_COMPRESS_OFF) {
1123 metaslab_class_t *mc = spa_normal_class(spa);
1124 void *cbuf = zio_buf_alloc(lsize);
1125 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize,
1126 (size_t)metaslab_class_get_minblocksize(mc));
1127 if (psize == 0 || psize == lsize) {
1128 compress = ZIO_COMPRESS_OFF;
1129 zio_buf_free(cbuf, lsize);
1131 ASSERT(psize < lsize);
1132 zio_push_transform(zio, cbuf, psize, lsize, NULL);
1137 * The final pass of spa_sync() must be all rewrites, but the first
1138 * few passes offer a trade-off: allocating blocks defers convergence,
1139 * but newly allocated blocks are sequential, so they can be written
1140 * to disk faster. Therefore, we allow the first few passes of
1141 * spa_sync() to allocate new blocks, but force rewrites after that.
1142 * There should only be a handful of blocks after pass 1 in any case.
1144 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1145 BP_GET_PSIZE(bp) == psize &&
1146 pass >= zfs_sync_pass_rewrite) {
1148 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1149 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1150 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1153 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1157 if (zio->io_bp_orig.blk_birth != 0 &&
1158 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1159 BP_SET_LSIZE(bp, lsize);
1160 BP_SET_TYPE(bp, zp->zp_type);
1161 BP_SET_LEVEL(bp, zp->zp_level);
1162 BP_SET_BIRTH(bp, zio->io_txg, 0);
1164 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1166 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1167 BP_SET_LSIZE(bp, lsize);
1168 BP_SET_TYPE(bp, zp->zp_type);
1169 BP_SET_LEVEL(bp, zp->zp_level);
1170 BP_SET_PSIZE(bp, psize);
1171 BP_SET_COMPRESS(bp, compress);
1172 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1173 BP_SET_DEDUP(bp, zp->zp_dedup);
1174 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1176 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1177 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1178 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1180 if (zp->zp_nopwrite) {
1181 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1182 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1183 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1187 return (ZIO_PIPELINE_CONTINUE);
1191 zio_free_bp_init(zio_t **ziop)
1194 blkptr_t *bp = zio->io_bp;
1196 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1197 if (BP_GET_DEDUP(bp))
1198 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1201 return (ZIO_PIPELINE_CONTINUE);
1205 * ==========================================================================
1206 * Execute the I/O pipeline
1207 * ==========================================================================
1211 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1213 spa_t *spa = zio->io_spa;
1214 zio_type_t t = zio->io_type;
1215 int flags = (cutinline ? TQ_FRONT : 0);
1217 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1220 * If we're a config writer or a probe, the normal issue and
1221 * interrupt threads may all be blocked waiting for the config lock.
1222 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1224 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1228 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1230 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1234 * If this is a high priority I/O, then use the high priority taskq if
1237 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1238 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1241 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1244 * NB: We are assuming that the zio can only be dispatched
1245 * to a single taskq at a time. It would be a grievous error
1246 * to dispatch the zio to another taskq at the same time.
1248 #if defined(illumos) || !defined(_KERNEL)
1249 ASSERT(zio->io_tqent.tqent_next == NULL);
1251 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
1253 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1254 flags, &zio->io_tqent);
1258 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1260 kthread_t *executor = zio->io_executor;
1261 spa_t *spa = zio->io_spa;
1263 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1264 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1266 for (i = 0; i < tqs->stqs_count; i++) {
1267 if (taskq_member(tqs->stqs_taskq[i], executor))
1276 zio_issue_async(zio_t **ziop)
1280 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1282 return (ZIO_PIPELINE_STOP);
1286 zio_interrupt(zio_t *zio)
1288 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1292 * Execute the I/O pipeline until one of the following occurs:
1294 * (1) the I/O completes
1295 * (2) the pipeline stalls waiting for dependent child I/Os
1296 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1297 * (4) the I/O is delegated by vdev-level caching or aggregation
1298 * (5) the I/O is deferred due to vdev-level queueing
1299 * (6) the I/O is handed off to another thread.
1301 * In all cases, the pipeline stops whenever there's no CPU work; it never
1302 * burns a thread in cv_wait().
1304 * There's no locking on io_stage because there's no legitimate way
1305 * for multiple threads to be attempting to process the same I/O.
1307 static zio_pipe_stage_t *zio_pipeline[];
1310 zio_execute(zio_t *zio)
1312 zio->io_executor = curthread;
1314 while (zio->io_stage < ZIO_STAGE_DONE) {
1315 enum zio_stage pipeline = zio->io_pipeline;
1316 enum zio_stage stage = zio->io_stage;
1319 ASSERT(!MUTEX_HELD(&zio->io_lock));
1320 ASSERT(ISP2(stage));
1321 ASSERT(zio->io_stall == NULL);
1325 } while ((stage & pipeline) == 0);
1327 ASSERT(stage <= ZIO_STAGE_DONE);
1330 * If we are in interrupt context and this pipeline stage
1331 * will grab a config lock that is held across I/O,
1332 * or may wait for an I/O that needs an interrupt thread
1333 * to complete, issue async to avoid deadlock.
1335 * For VDEV_IO_START, we cut in line so that the io will
1336 * be sent to disk promptly.
1338 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1339 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1340 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1341 zio_requeue_io_start_cut_in_line : B_FALSE;
1342 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1346 zio->io_stage = stage;
1347 rv = zio_pipeline[highbit(stage) - 1](&zio);
1349 if (rv == ZIO_PIPELINE_STOP)
1352 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1357 * ==========================================================================
1358 * Initiate I/O, either sync or async
1359 * ==========================================================================
1362 zio_wait(zio_t *zio)
1366 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1367 ASSERT(zio->io_executor == NULL);
1369 zio->io_waiter = curthread;
1373 mutex_enter(&zio->io_lock);
1374 while (zio->io_executor != NULL)
1375 cv_wait(&zio->io_cv, &zio->io_lock);
1376 mutex_exit(&zio->io_lock);
1378 error = zio->io_error;
1385 zio_nowait(zio_t *zio)
1387 ASSERT(zio->io_executor == NULL);
1389 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1390 zio_unique_parent(zio) == NULL) {
1392 * This is a logical async I/O with no parent to wait for it.
1393 * We add it to the spa_async_root_zio "Godfather" I/O which
1394 * will ensure they complete prior to unloading the pool.
1396 spa_t *spa = zio->io_spa;
1398 zio_add_child(spa->spa_async_zio_root, zio);
1405 * ==========================================================================
1406 * Reexecute or suspend/resume failed I/O
1407 * ==========================================================================
1411 zio_reexecute(zio_t *pio)
1413 zio_t *cio, *cio_next;
1415 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1416 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1417 ASSERT(pio->io_gang_leader == NULL);
1418 ASSERT(pio->io_gang_tree == NULL);
1420 pio->io_flags = pio->io_orig_flags;
1421 pio->io_stage = pio->io_orig_stage;
1422 pio->io_pipeline = pio->io_orig_pipeline;
1423 pio->io_reexecute = 0;
1424 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1426 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1427 pio->io_state[w] = 0;
1428 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1429 pio->io_child_error[c] = 0;
1431 if (IO_IS_ALLOCATING(pio))
1432 BP_ZERO(pio->io_bp);
1435 * As we reexecute pio's children, new children could be created.
1436 * New children go to the head of pio's io_child_list, however,
1437 * so we will (correctly) not reexecute them. The key is that
1438 * the remainder of pio's io_child_list, from 'cio_next' onward,
1439 * cannot be affected by any side effects of reexecuting 'cio'.
1441 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1442 cio_next = zio_walk_children(pio);
1443 mutex_enter(&pio->io_lock);
1444 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1445 pio->io_children[cio->io_child_type][w]++;
1446 mutex_exit(&pio->io_lock);
1451 * Now that all children have been reexecuted, execute the parent.
1452 * We don't reexecute "The Godfather" I/O here as it's the
1453 * responsibility of the caller to wait on him.
1455 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1460 zio_suspend(spa_t *spa, zio_t *zio)
1462 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1463 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1464 "failure and the failure mode property for this pool "
1465 "is set to panic.", spa_name(spa));
1467 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1469 mutex_enter(&spa->spa_suspend_lock);
1471 if (spa->spa_suspend_zio_root == NULL)
1472 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1473 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1474 ZIO_FLAG_GODFATHER);
1476 spa->spa_suspended = B_TRUE;
1479 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1480 ASSERT(zio != spa->spa_suspend_zio_root);
1481 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1482 ASSERT(zio_unique_parent(zio) == NULL);
1483 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1484 zio_add_child(spa->spa_suspend_zio_root, zio);
1487 mutex_exit(&spa->spa_suspend_lock);
1491 zio_resume(spa_t *spa)
1496 * Reexecute all previously suspended i/o.
1498 mutex_enter(&spa->spa_suspend_lock);
1499 spa->spa_suspended = B_FALSE;
1500 cv_broadcast(&spa->spa_suspend_cv);
1501 pio = spa->spa_suspend_zio_root;
1502 spa->spa_suspend_zio_root = NULL;
1503 mutex_exit(&spa->spa_suspend_lock);
1509 return (zio_wait(pio));
1513 zio_resume_wait(spa_t *spa)
1515 mutex_enter(&spa->spa_suspend_lock);
1516 while (spa_suspended(spa))
1517 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1518 mutex_exit(&spa->spa_suspend_lock);
1522 * ==========================================================================
1525 * A gang block is a collection of small blocks that looks to the DMU
1526 * like one large block. When zio_dva_allocate() cannot find a block
1527 * of the requested size, due to either severe fragmentation or the pool
1528 * being nearly full, it calls zio_write_gang_block() to construct the
1529 * block from smaller fragments.
1531 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1532 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1533 * an indirect block: it's an array of block pointers. It consumes
1534 * only one sector and hence is allocatable regardless of fragmentation.
1535 * The gang header's bps point to its gang members, which hold the data.
1537 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1538 * as the verifier to ensure uniqueness of the SHA256 checksum.
1539 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1540 * not the gang header. This ensures that data block signatures (needed for
1541 * deduplication) are independent of how the block is physically stored.
1543 * Gang blocks can be nested: a gang member may itself be a gang block.
1544 * Thus every gang block is a tree in which root and all interior nodes are
1545 * gang headers, and the leaves are normal blocks that contain user data.
1546 * The root of the gang tree is called the gang leader.
1548 * To perform any operation (read, rewrite, free, claim) on a gang block,
1549 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1550 * in the io_gang_tree field of the original logical i/o by recursively
1551 * reading the gang leader and all gang headers below it. This yields
1552 * an in-core tree containing the contents of every gang header and the
1553 * bps for every constituent of the gang block.
1555 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1556 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1557 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1558 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1559 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1560 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1561 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1562 * of the gang header plus zio_checksum_compute() of the data to update the
1563 * gang header's blk_cksum as described above.
1565 * The two-phase assemble/issue model solves the problem of partial failure --
1566 * what if you'd freed part of a gang block but then couldn't read the
1567 * gang header for another part? Assembling the entire gang tree first
1568 * ensures that all the necessary gang header I/O has succeeded before
1569 * starting the actual work of free, claim, or write. Once the gang tree
1570 * is assembled, free and claim are in-memory operations that cannot fail.
1572 * In the event that a gang write fails, zio_dva_unallocate() walks the
1573 * gang tree to immediately free (i.e. insert back into the space map)
1574 * everything we've allocated. This ensures that we don't get ENOSPC
1575 * errors during repeated suspend/resume cycles due to a flaky device.
1577 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1578 * the gang tree, we won't modify the block, so we can safely defer the free
1579 * (knowing that the block is still intact). If we *can* assemble the gang
1580 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1581 * each constituent bp and we can allocate a new block on the next sync pass.
1583 * In all cases, the gang tree allows complete recovery from partial failure.
1584 * ==========================================================================
1588 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1593 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1594 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1595 &pio->io_bookmark));
1599 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1604 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1605 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1606 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1608 * As we rewrite each gang header, the pipeline will compute
1609 * a new gang block header checksum for it; but no one will
1610 * compute a new data checksum, so we do that here. The one
1611 * exception is the gang leader: the pipeline already computed
1612 * its data checksum because that stage precedes gang assembly.
1613 * (Presently, nothing actually uses interior data checksums;
1614 * this is just good hygiene.)
1616 if (gn != pio->io_gang_leader->io_gang_tree) {
1617 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1618 data, BP_GET_PSIZE(bp));
1621 * If we are here to damage data for testing purposes,
1622 * leave the GBH alone so that we can detect the damage.
1624 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1625 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1627 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1628 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1629 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1637 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1639 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1640 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp),
1641 ZIO_GANG_CHILD_FLAGS(pio)));
1646 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1648 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1649 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1652 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1661 static void zio_gang_tree_assemble_done(zio_t *zio);
1663 static zio_gang_node_t *
1664 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1666 zio_gang_node_t *gn;
1668 ASSERT(*gnpp == NULL);
1670 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1671 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1678 zio_gang_node_free(zio_gang_node_t **gnpp)
1680 zio_gang_node_t *gn = *gnpp;
1682 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1683 ASSERT(gn->gn_child[g] == NULL);
1685 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1686 kmem_free(gn, sizeof (*gn));
1691 zio_gang_tree_free(zio_gang_node_t **gnpp)
1693 zio_gang_node_t *gn = *gnpp;
1698 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1699 zio_gang_tree_free(&gn->gn_child[g]);
1701 zio_gang_node_free(gnpp);
1705 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1707 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1709 ASSERT(gio->io_gang_leader == gio);
1710 ASSERT(BP_IS_GANG(bp));
1712 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1713 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1714 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1718 zio_gang_tree_assemble_done(zio_t *zio)
1720 zio_t *gio = zio->io_gang_leader;
1721 zio_gang_node_t *gn = zio->io_private;
1722 blkptr_t *bp = zio->io_bp;
1724 ASSERT(gio == zio_unique_parent(zio));
1725 ASSERT(zio->io_child_count == 0);
1730 if (BP_SHOULD_BYTESWAP(bp))
1731 byteswap_uint64_array(zio->io_data, zio->io_size);
1733 ASSERT(zio->io_data == gn->gn_gbh);
1734 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1735 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1737 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1738 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1739 if (!BP_IS_GANG(gbp))
1741 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1746 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1748 zio_t *gio = pio->io_gang_leader;
1751 ASSERT(BP_IS_GANG(bp) == !!gn);
1752 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1753 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1756 * If you're a gang header, your data is in gn->gn_gbh.
1757 * If you're a gang member, your data is in 'data' and gn == NULL.
1759 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1762 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1764 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1765 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1766 if (BP_IS_HOLE(gbp))
1768 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1769 data = (char *)data + BP_GET_PSIZE(gbp);
1773 if (gn == gio->io_gang_tree && gio->io_data != NULL)
1774 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1781 zio_gang_assemble(zio_t **ziop)
1784 blkptr_t *bp = zio->io_bp;
1786 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1787 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1789 zio->io_gang_leader = zio;
1791 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1793 return (ZIO_PIPELINE_CONTINUE);
1797 zio_gang_issue(zio_t **ziop)
1800 blkptr_t *bp = zio->io_bp;
1802 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1803 return (ZIO_PIPELINE_STOP);
1805 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1806 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1808 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1809 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1811 zio_gang_tree_free(&zio->io_gang_tree);
1813 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1815 return (ZIO_PIPELINE_CONTINUE);
1819 zio_write_gang_member_ready(zio_t *zio)
1821 zio_t *pio = zio_unique_parent(zio);
1822 zio_t *gio = zio->io_gang_leader;
1823 dva_t *cdva = zio->io_bp->blk_dva;
1824 dva_t *pdva = pio->io_bp->blk_dva;
1827 if (BP_IS_HOLE(zio->io_bp))
1830 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1832 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1833 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1834 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1835 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1836 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1838 mutex_enter(&pio->io_lock);
1839 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1840 ASSERT(DVA_GET_GANG(&pdva[d]));
1841 asize = DVA_GET_ASIZE(&pdva[d]);
1842 asize += DVA_GET_ASIZE(&cdva[d]);
1843 DVA_SET_ASIZE(&pdva[d], asize);
1845 mutex_exit(&pio->io_lock);
1849 zio_write_gang_block(zio_t *pio)
1851 spa_t *spa = pio->io_spa;
1852 blkptr_t *bp = pio->io_bp;
1853 zio_t *gio = pio->io_gang_leader;
1855 zio_gang_node_t *gn, **gnpp;
1856 zio_gbh_phys_t *gbh;
1857 uint64_t txg = pio->io_txg;
1858 uint64_t resid = pio->io_size;
1860 int copies = gio->io_prop.zp_copies;
1861 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1865 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1866 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1867 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1869 pio->io_error = error;
1870 return (ZIO_PIPELINE_CONTINUE);
1874 gnpp = &gio->io_gang_tree;
1876 gnpp = pio->io_private;
1877 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1880 gn = zio_gang_node_alloc(gnpp);
1882 bzero(gbh, SPA_GANGBLOCKSIZE);
1885 * Create the gang header.
1887 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1888 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1891 * Create and nowait the gang children.
1893 for (int g = 0; resid != 0; resid -= lsize, g++) {
1894 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1896 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1898 zp.zp_checksum = gio->io_prop.zp_checksum;
1899 zp.zp_compress = ZIO_COMPRESS_OFF;
1900 zp.zp_type = DMU_OT_NONE;
1902 zp.zp_copies = gio->io_prop.zp_copies;
1903 zp.zp_dedup = B_FALSE;
1904 zp.zp_dedup_verify = B_FALSE;
1905 zp.zp_nopwrite = B_FALSE;
1907 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1908 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1909 zio_write_gang_member_ready, NULL, NULL, &gn->gn_child[g],
1910 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1911 &pio->io_bookmark));
1915 * Set pio's pipeline to just wait for zio to finish.
1917 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1921 return (ZIO_PIPELINE_CONTINUE);
1925 * The zio_nop_write stage in the pipeline determines if allocating
1926 * a new bp is necessary. By leveraging a cryptographically secure checksum,
1927 * such as SHA256, we can compare the checksums of the new data and the old
1928 * to determine if allocating a new block is required. The nopwrite
1929 * feature can handle writes in either syncing or open context (i.e. zil
1930 * writes) and as a result is mutually exclusive with dedup.
1933 zio_nop_write(zio_t **ziop)
1936 blkptr_t *bp = zio->io_bp;
1937 blkptr_t *bp_orig = &zio->io_bp_orig;
1938 zio_prop_t *zp = &zio->io_prop;
1940 ASSERT(BP_GET_LEVEL(bp) == 0);
1941 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1942 ASSERT(zp->zp_nopwrite);
1943 ASSERT(!zp->zp_dedup);
1944 ASSERT(zio->io_bp_override == NULL);
1945 ASSERT(IO_IS_ALLOCATING(zio));
1948 * Check to see if the original bp and the new bp have matching
1949 * characteristics (i.e. same checksum, compression algorithms, etc).
1950 * If they don't then just continue with the pipeline which will
1951 * allocate a new bp.
1953 if (BP_IS_HOLE(bp_orig) ||
1954 !zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup ||
1955 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
1956 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
1957 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
1958 zp->zp_copies != BP_GET_NDVAS(bp_orig))
1959 return (ZIO_PIPELINE_CONTINUE);
1962 * If the checksums match then reset the pipeline so that we
1963 * avoid allocating a new bp and issuing any I/O.
1965 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
1966 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup);
1967 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
1968 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
1969 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
1970 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
1971 sizeof (uint64_t)) == 0);
1974 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1975 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1978 return (ZIO_PIPELINE_CONTINUE);
1982 * ==========================================================================
1984 * ==========================================================================
1987 zio_ddt_child_read_done(zio_t *zio)
1989 blkptr_t *bp = zio->io_bp;
1990 ddt_entry_t *dde = zio->io_private;
1992 zio_t *pio = zio_unique_parent(zio);
1994 mutex_enter(&pio->io_lock);
1995 ddp = ddt_phys_select(dde, bp);
1996 if (zio->io_error == 0)
1997 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1998 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1999 dde->dde_repair_data = zio->io_data;
2001 zio_buf_free(zio->io_data, zio->io_size);
2002 mutex_exit(&pio->io_lock);
2006 zio_ddt_read_start(zio_t **ziop)
2009 blkptr_t *bp = zio->io_bp;
2011 ASSERT(BP_GET_DEDUP(bp));
2012 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2013 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2015 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2016 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2017 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2018 ddt_phys_t *ddp = dde->dde_phys;
2019 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2022 ASSERT(zio->io_vsd == NULL);
2025 if (ddp_self == NULL)
2026 return (ZIO_PIPELINE_CONTINUE);
2028 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2029 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2031 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2033 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2034 zio_buf_alloc(zio->io_size), zio->io_size,
2035 zio_ddt_child_read_done, dde, zio->io_priority,
2036 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
2037 &zio->io_bookmark));
2039 return (ZIO_PIPELINE_CONTINUE);
2042 zio_nowait(zio_read(zio, zio->io_spa, bp,
2043 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
2044 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2046 return (ZIO_PIPELINE_CONTINUE);
2050 zio_ddt_read_done(zio_t **ziop)
2053 blkptr_t *bp = zio->io_bp;
2055 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2056 return (ZIO_PIPELINE_STOP);
2058 ASSERT(BP_GET_DEDUP(bp));
2059 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2060 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2062 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2063 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2064 ddt_entry_t *dde = zio->io_vsd;
2066 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2067 return (ZIO_PIPELINE_CONTINUE);
2070 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2071 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2072 return (ZIO_PIPELINE_STOP);
2074 if (dde->dde_repair_data != NULL) {
2075 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2076 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2078 ddt_repair_done(ddt, dde);
2082 ASSERT(zio->io_vsd == NULL);
2084 return (ZIO_PIPELINE_CONTINUE);
2088 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2090 spa_t *spa = zio->io_spa;
2093 * Note: we compare the original data, not the transformed data,
2094 * because when zio->io_bp is an override bp, we will not have
2095 * pushed the I/O transforms. That's an important optimization
2096 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2098 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2099 zio_t *lio = dde->dde_lead_zio[p];
2102 return (lio->io_orig_size != zio->io_orig_size ||
2103 bcmp(zio->io_orig_data, lio->io_orig_data,
2104 zio->io_orig_size) != 0);
2108 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2109 ddt_phys_t *ddp = &dde->dde_phys[p];
2111 if (ddp->ddp_phys_birth != 0) {
2112 arc_buf_t *abuf = NULL;
2113 uint32_t aflags = ARC_WAIT;
2114 blkptr_t blk = *zio->io_bp;
2117 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2121 error = arc_read(NULL, spa, &blk,
2122 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2123 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2124 &aflags, &zio->io_bookmark);
2127 if (arc_buf_size(abuf) != zio->io_orig_size ||
2128 bcmp(abuf->b_data, zio->io_orig_data,
2129 zio->io_orig_size) != 0)
2130 error = SET_ERROR(EEXIST);
2131 VERIFY(arc_buf_remove_ref(abuf, &abuf));
2135 return (error != 0);
2143 zio_ddt_child_write_ready(zio_t *zio)
2145 int p = zio->io_prop.zp_copies;
2146 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2147 ddt_entry_t *dde = zio->io_private;
2148 ddt_phys_t *ddp = &dde->dde_phys[p];
2156 ASSERT(dde->dde_lead_zio[p] == zio);
2158 ddt_phys_fill(ddp, zio->io_bp);
2160 while ((pio = zio_walk_parents(zio)) != NULL)
2161 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2167 zio_ddt_child_write_done(zio_t *zio)
2169 int p = zio->io_prop.zp_copies;
2170 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2171 ddt_entry_t *dde = zio->io_private;
2172 ddt_phys_t *ddp = &dde->dde_phys[p];
2176 ASSERT(ddp->ddp_refcnt == 0);
2177 ASSERT(dde->dde_lead_zio[p] == zio);
2178 dde->dde_lead_zio[p] = NULL;
2180 if (zio->io_error == 0) {
2181 while (zio_walk_parents(zio) != NULL)
2182 ddt_phys_addref(ddp);
2184 ddt_phys_clear(ddp);
2191 zio_ddt_ditto_write_done(zio_t *zio)
2193 int p = DDT_PHYS_DITTO;
2194 zio_prop_t *zp = &zio->io_prop;
2195 blkptr_t *bp = zio->io_bp;
2196 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2197 ddt_entry_t *dde = zio->io_private;
2198 ddt_phys_t *ddp = &dde->dde_phys[p];
2199 ddt_key_t *ddk = &dde->dde_key;
2203 ASSERT(ddp->ddp_refcnt == 0);
2204 ASSERT(dde->dde_lead_zio[p] == zio);
2205 dde->dde_lead_zio[p] = NULL;
2207 if (zio->io_error == 0) {
2208 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2209 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2210 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2211 if (ddp->ddp_phys_birth != 0)
2212 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2213 ddt_phys_fill(ddp, bp);
2220 zio_ddt_write(zio_t **ziop)
2223 spa_t *spa = zio->io_spa;
2224 blkptr_t *bp = zio->io_bp;
2225 uint64_t txg = zio->io_txg;
2226 zio_prop_t *zp = &zio->io_prop;
2227 int p = zp->zp_copies;
2231 ddt_t *ddt = ddt_select(spa, bp);
2235 ASSERT(BP_GET_DEDUP(bp));
2236 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2237 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2240 dde = ddt_lookup(ddt, bp, B_TRUE);
2241 ddp = &dde->dde_phys[p];
2243 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2245 * If we're using a weak checksum, upgrade to a strong checksum
2246 * and try again. If we're already using a strong checksum,
2247 * we can't resolve it, so just convert to an ordinary write.
2248 * (And automatically e-mail a paper to Nature?)
2250 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2251 zp->zp_checksum = spa_dedup_checksum(spa);
2252 zio_pop_transforms(zio);
2253 zio->io_stage = ZIO_STAGE_OPEN;
2256 zp->zp_dedup = B_FALSE;
2258 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2260 return (ZIO_PIPELINE_CONTINUE);
2263 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2264 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2266 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2267 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2268 zio_prop_t czp = *zp;
2270 czp.zp_copies = ditto_copies;
2273 * If we arrived here with an override bp, we won't have run
2274 * the transform stack, so we won't have the data we need to
2275 * generate a child i/o. So, toss the override bp and restart.
2276 * This is safe, because using the override bp is just an
2277 * optimization; and it's rare, so the cost doesn't matter.
2279 if (zio->io_bp_override) {
2280 zio_pop_transforms(zio);
2281 zio->io_stage = ZIO_STAGE_OPEN;
2282 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2283 zio->io_bp_override = NULL;
2286 return (ZIO_PIPELINE_CONTINUE);
2289 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2290 zio->io_orig_size, &czp, NULL, NULL,
2291 zio_ddt_ditto_write_done, dde, zio->io_priority,
2292 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2294 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2295 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2298 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2299 if (ddp->ddp_phys_birth != 0)
2300 ddt_bp_fill(ddp, bp, txg);
2301 if (dde->dde_lead_zio[p] != NULL)
2302 zio_add_child(zio, dde->dde_lead_zio[p]);
2304 ddt_phys_addref(ddp);
2305 } else if (zio->io_bp_override) {
2306 ASSERT(bp->blk_birth == txg);
2307 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2308 ddt_phys_fill(ddp, bp);
2309 ddt_phys_addref(ddp);
2311 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2312 zio->io_orig_size, zp, zio_ddt_child_write_ready, NULL,
2313 zio_ddt_child_write_done, dde, zio->io_priority,
2314 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2316 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2317 dde->dde_lead_zio[p] = cio;
2327 return (ZIO_PIPELINE_CONTINUE);
2330 ddt_entry_t *freedde; /* for debugging */
2333 zio_ddt_free(zio_t **ziop)
2336 spa_t *spa = zio->io_spa;
2337 blkptr_t *bp = zio->io_bp;
2338 ddt_t *ddt = ddt_select(spa, bp);
2342 ASSERT(BP_GET_DEDUP(bp));
2343 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2346 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2347 ddp = ddt_phys_select(dde, bp);
2348 ddt_phys_decref(ddp);
2351 return (ZIO_PIPELINE_CONTINUE);
2355 * ==========================================================================
2356 * Allocate and free blocks
2357 * ==========================================================================
2360 zio_dva_allocate(zio_t **ziop)
2363 spa_t *spa = zio->io_spa;
2364 metaslab_class_t *mc = spa_normal_class(spa);
2365 blkptr_t *bp = zio->io_bp;
2369 if (zio->io_gang_leader == NULL) {
2370 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2371 zio->io_gang_leader = zio;
2374 ASSERT(BP_IS_HOLE(bp));
2375 ASSERT0(BP_GET_NDVAS(bp));
2376 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2377 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2378 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2381 * The dump device does not support gang blocks so allocation on
2382 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2383 * the "fast" gang feature.
2385 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2386 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2387 METASLAB_GANG_CHILD : 0;
2388 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2389 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2392 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2393 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2395 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2396 return (zio_write_gang_block(zio));
2397 zio->io_error = error;
2400 return (ZIO_PIPELINE_CONTINUE);
2404 zio_dva_free(zio_t **ziop)
2408 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2410 return (ZIO_PIPELINE_CONTINUE);
2414 zio_dva_claim(zio_t **ziop)
2419 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2421 zio->io_error = error;
2423 return (ZIO_PIPELINE_CONTINUE);
2427 * Undo an allocation. This is used by zio_done() when an I/O fails
2428 * and we want to give back the block we just allocated.
2429 * This handles both normal blocks and gang blocks.
2432 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2434 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2435 ASSERT(zio->io_bp_override == NULL);
2437 if (!BP_IS_HOLE(bp))
2438 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2441 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2442 zio_dva_unallocate(zio, gn->gn_child[g],
2443 &gn->gn_gbh->zg_blkptr[g]);
2449 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2452 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2453 uint64_t size, boolean_t use_slog)
2457 ASSERT(txg > spa_syncing_txg(spa));
2460 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2461 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2462 * when allocating them.
2465 error = metaslab_alloc(spa, spa_log_class(spa), size,
2466 new_bp, 1, txg, old_bp,
2467 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2471 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2472 new_bp, 1, txg, old_bp,
2473 METASLAB_HINTBP_AVOID);
2477 BP_SET_LSIZE(new_bp, size);
2478 BP_SET_PSIZE(new_bp, size);
2479 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2480 BP_SET_CHECKSUM(new_bp,
2481 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2482 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2483 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2484 BP_SET_LEVEL(new_bp, 0);
2485 BP_SET_DEDUP(new_bp, 0);
2486 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2493 * Free an intent log block.
2496 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2498 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2499 ASSERT(!BP_IS_GANG(bp));
2501 zio_free(spa, txg, bp);
2505 * ==========================================================================
2506 * Read, write and delete to physical devices
2507 * ==========================================================================
2510 zio_vdev_io_start(zio_t **ziop)
2513 vdev_t *vd = zio->io_vd;
2515 spa_t *spa = zio->io_spa;
2517 ASSERT(zio->io_error == 0);
2518 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2521 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2522 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2525 * The mirror_ops handle multiple DVAs in a single BP.
2527 return (vdev_mirror_ops.vdev_op_io_start(zio));
2530 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE) {
2531 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg);
2532 return (ZIO_PIPELINE_CONTINUE);
2536 * We keep track of time-sensitive I/Os so that the scan thread
2537 * can quickly react to certain workloads. In particular, we care
2538 * about non-scrubbing, top-level reads and writes with the following
2540 * - synchronous writes of user data to non-slog devices
2541 * - any reads of user data
2542 * When these conditions are met, adjust the timestamp of spa_last_io
2543 * which allows the scan thread to adjust its workload accordingly.
2545 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2546 vd == vd->vdev_top && !vd->vdev_islog &&
2547 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2548 zio->io_txg != spa_syncing_txg(spa)) {
2549 uint64_t old = spa->spa_last_io;
2550 uint64_t new = ddi_get_lbolt64();
2552 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2555 align = 1ULL << vd->vdev_top->vdev_ashift;
2557 if (P2PHASE(zio->io_size, align) != 0) {
2558 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2560 if (zio->io_type == ZIO_TYPE_READ ||
2561 zio->io_type == ZIO_TYPE_WRITE)
2562 abuf = zio_buf_alloc(asize);
2563 ASSERT(vd == vd->vdev_top);
2564 if (zio->io_type == ZIO_TYPE_WRITE) {
2565 bcopy(zio->io_data, abuf, zio->io_size);
2566 bzero(abuf + zio->io_size, asize - zio->io_size);
2568 zio_push_transform(zio, abuf, asize, abuf ? asize : 0,
2572 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2573 ASSERT(P2PHASE(zio->io_size, align) == 0);
2574 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa));
2577 * If this is a repair I/O, and there's no self-healing involved --
2578 * that is, we're just resilvering what we expect to resilver --
2579 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2580 * This prevents spurious resilvering with nested replication.
2581 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2582 * A is out of date, we'll read from C+D, then use the data to
2583 * resilver A+B -- but we don't actually want to resilver B, just A.
2584 * The top-level mirror has no way to know this, so instead we just
2585 * discard unnecessary repairs as we work our way down the vdev tree.
2586 * The same logic applies to any form of nested replication:
2587 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2589 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2590 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2591 zio->io_txg != 0 && /* not a delegated i/o */
2592 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2593 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2594 zio_vdev_io_bypass(zio);
2595 return (ZIO_PIPELINE_CONTINUE);
2598 if (vd->vdev_ops->vdev_op_leaf &&
2599 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2601 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio))
2602 return (ZIO_PIPELINE_CONTINUE);
2604 if ((zio = vdev_queue_io(zio)) == NULL)
2605 return (ZIO_PIPELINE_STOP);
2608 if (!vdev_accessible(vd, zio)) {
2609 zio->io_error = SET_ERROR(ENXIO);
2611 return (ZIO_PIPELINE_STOP);
2616 * Note that we ignore repair writes for TRIM because they can conflict
2617 * with normal writes. This isn't an issue because, by definition, we
2618 * only repair blocks that aren't freed.
2620 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_WRITE &&
2621 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2622 if (!trim_map_write_start(zio))
2623 return (ZIO_PIPELINE_STOP);
2626 return (vd->vdev_ops->vdev_op_io_start(zio));
2630 zio_vdev_io_done(zio_t **ziop)
2633 vdev_t *vd = zio->io_vd;
2634 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2635 boolean_t unexpected_error = B_FALSE;
2637 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2638 return (ZIO_PIPELINE_STOP);
2640 ASSERT(zio->io_type == ZIO_TYPE_READ ||
2641 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE);
2643 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2644 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2646 if (zio->io_type == ZIO_TYPE_WRITE &&
2647 !(zio->io_flags & ZIO_FLAG_IO_REPAIR))
2648 trim_map_write_done(zio);
2650 vdev_queue_io_done(zio);
2652 if (zio->io_type == ZIO_TYPE_WRITE)
2653 vdev_cache_write(zio);
2655 if (zio_injection_enabled && zio->io_error == 0)
2656 zio->io_error = zio_handle_device_injection(vd,
2659 if (zio_injection_enabled && zio->io_error == 0)
2660 zio->io_error = zio_handle_label_injection(zio, EIO);
2662 if (zio->io_error) {
2663 if (!vdev_accessible(vd, zio)) {
2664 zio->io_error = SET_ERROR(ENXIO);
2666 unexpected_error = B_TRUE;
2671 ops->vdev_op_io_done(zio);
2673 if (unexpected_error)
2674 VERIFY(vdev_probe(vd, zio) == NULL);
2676 return (ZIO_PIPELINE_CONTINUE);
2680 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2681 * disk, and use that to finish the checksum ereport later.
2684 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2685 const void *good_buf)
2687 /* no processing needed */
2688 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2693 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2695 void *buf = zio_buf_alloc(zio->io_size);
2697 bcopy(zio->io_data, buf, zio->io_size);
2699 zcr->zcr_cbinfo = zio->io_size;
2700 zcr->zcr_cbdata = buf;
2701 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2702 zcr->zcr_free = zio_buf_free;
2706 zio_vdev_io_assess(zio_t **ziop)
2709 vdev_t *vd = zio->io_vd;
2711 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2712 return (ZIO_PIPELINE_STOP);
2714 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2715 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2717 if (zio->io_vsd != NULL) {
2718 zio->io_vsd_ops->vsd_free(zio);
2722 if (zio_injection_enabled && zio->io_error == 0)
2723 zio->io_error = zio_handle_fault_injection(zio, EIO);
2725 if (zio->io_type == ZIO_TYPE_IOCTL && zio->io_cmd == DKIOCTRIM)
2726 switch (zio->io_error) {
2728 ZIO_TRIM_STAT_INCR(bytes, zio->io_size);
2729 ZIO_TRIM_STAT_BUMP(success);
2732 ZIO_TRIM_STAT_BUMP(unsupported);
2735 ZIO_TRIM_STAT_BUMP(failed);
2740 * If the I/O failed, determine whether we should attempt to retry it.
2742 * On retry, we cut in line in the issue queue, since we don't want
2743 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2745 if (zio->io_error && vd == NULL &&
2746 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2747 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2748 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2750 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2751 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2752 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2753 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2754 zio_requeue_io_start_cut_in_line);
2755 return (ZIO_PIPELINE_STOP);
2759 * If we got an error on a leaf device, convert it to ENXIO
2760 * if the device is not accessible at all.
2762 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2763 !vdev_accessible(vd, zio))
2764 zio->io_error = SET_ERROR(ENXIO);
2767 * If we can't write to an interior vdev (mirror or RAID-Z),
2768 * set vdev_cant_write so that we stop trying to allocate from it.
2770 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2771 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
2772 vd->vdev_cant_write = B_TRUE;
2776 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2778 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2779 zio->io_physdone != NULL) {
2780 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
2781 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
2782 zio->io_physdone(zio->io_logical);
2785 return (ZIO_PIPELINE_CONTINUE);
2789 zio_vdev_io_reissue(zio_t *zio)
2791 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2792 ASSERT(zio->io_error == 0);
2794 zio->io_stage >>= 1;
2798 zio_vdev_io_redone(zio_t *zio)
2800 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2802 zio->io_stage >>= 1;
2806 zio_vdev_io_bypass(zio_t *zio)
2808 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2809 ASSERT(zio->io_error == 0);
2811 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2812 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2816 * ==========================================================================
2817 * Generate and verify checksums
2818 * ==========================================================================
2821 zio_checksum_generate(zio_t **ziop)
2824 blkptr_t *bp = zio->io_bp;
2825 enum zio_checksum checksum;
2829 * This is zio_write_phys().
2830 * We're either generating a label checksum, or none at all.
2832 checksum = zio->io_prop.zp_checksum;
2834 if (checksum == ZIO_CHECKSUM_OFF)
2835 return (ZIO_PIPELINE_CONTINUE);
2837 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2839 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2840 ASSERT(!IO_IS_ALLOCATING(zio));
2841 checksum = ZIO_CHECKSUM_GANG_HEADER;
2843 checksum = BP_GET_CHECKSUM(bp);
2847 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2849 return (ZIO_PIPELINE_CONTINUE);
2853 zio_checksum_verify(zio_t **ziop)
2856 zio_bad_cksum_t info;
2857 blkptr_t *bp = zio->io_bp;
2860 ASSERT(zio->io_vd != NULL);
2864 * This is zio_read_phys().
2865 * We're either verifying a label checksum, or nothing at all.
2867 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2868 return (ZIO_PIPELINE_CONTINUE);
2870 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2873 if ((error = zio_checksum_error(zio, &info)) != 0) {
2874 zio->io_error = error;
2875 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2876 zfs_ereport_start_checksum(zio->io_spa,
2877 zio->io_vd, zio, zio->io_offset,
2878 zio->io_size, NULL, &info);
2882 return (ZIO_PIPELINE_CONTINUE);
2886 * Called by RAID-Z to ensure we don't compute the checksum twice.
2889 zio_checksum_verified(zio_t *zio)
2891 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2895 * ==========================================================================
2896 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2897 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2898 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2899 * indicate errors that are specific to one I/O, and most likely permanent.
2900 * Any other error is presumed to be worse because we weren't expecting it.
2901 * ==========================================================================
2904 zio_worst_error(int e1, int e2)
2906 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2909 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2910 if (e1 == zio_error_rank[r1])
2913 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2914 if (e2 == zio_error_rank[r2])
2917 return (r1 > r2 ? e1 : e2);
2921 * ==========================================================================
2923 * ==========================================================================
2926 zio_ready(zio_t **ziop)
2929 blkptr_t *bp = zio->io_bp;
2930 zio_t *pio, *pio_next;
2932 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2933 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2934 return (ZIO_PIPELINE_STOP);
2936 if (zio->io_ready) {
2937 ASSERT(IO_IS_ALLOCATING(zio));
2938 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
2939 (zio->io_flags & ZIO_FLAG_NOPWRITE));
2940 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2945 if (bp != NULL && bp != &zio->io_bp_copy)
2946 zio->io_bp_copy = *bp;
2949 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2951 mutex_enter(&zio->io_lock);
2952 zio->io_state[ZIO_WAIT_READY] = 1;
2953 pio = zio_walk_parents(zio);
2954 mutex_exit(&zio->io_lock);
2957 * As we notify zio's parents, new parents could be added.
2958 * New parents go to the head of zio's io_parent_list, however,
2959 * so we will (correctly) not notify them. The remainder of zio's
2960 * io_parent_list, from 'pio_next' onward, cannot change because
2961 * all parents must wait for us to be done before they can be done.
2963 for (; pio != NULL; pio = pio_next) {
2964 pio_next = zio_walk_parents(zio);
2965 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2968 if (zio->io_flags & ZIO_FLAG_NODATA) {
2969 if (BP_IS_GANG(bp)) {
2970 zio->io_flags &= ~ZIO_FLAG_NODATA;
2972 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2973 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2977 if (zio_injection_enabled &&
2978 zio->io_spa->spa_syncing_txg == zio->io_txg)
2979 zio_handle_ignored_writes(zio);
2981 return (ZIO_PIPELINE_CONTINUE);
2985 zio_done(zio_t **ziop)
2988 spa_t *spa = zio->io_spa;
2989 zio_t *lio = zio->io_logical;
2990 blkptr_t *bp = zio->io_bp;
2991 vdev_t *vd = zio->io_vd;
2992 uint64_t psize = zio->io_size;
2993 zio_t *pio, *pio_next;
2996 * If our children haven't all completed,
2997 * wait for them and then repeat this pipeline stage.
2999 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
3000 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
3001 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
3002 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
3003 return (ZIO_PIPELINE_STOP);
3005 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3006 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3007 ASSERT(zio->io_children[c][w] == 0);
3010 ASSERT(bp->blk_pad[0] == 0);
3011 ASSERT(bp->blk_pad[1] == 0);
3012 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3013 (bp == zio_unique_parent(zio)->io_bp));
3014 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3015 zio->io_bp_override == NULL &&
3016 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3017 ASSERT(!BP_SHOULD_BYTESWAP(bp));
3018 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3019 ASSERT(BP_COUNT_GANG(bp) == 0 ||
3020 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3022 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3023 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3027 * If there were child vdev/gang/ddt errors, they apply to us now.
3029 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3030 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3031 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3034 * If the I/O on the transformed data was successful, generate any
3035 * checksum reports now while we still have the transformed data.
3037 if (zio->io_error == 0) {
3038 while (zio->io_cksum_report != NULL) {
3039 zio_cksum_report_t *zcr = zio->io_cksum_report;
3040 uint64_t align = zcr->zcr_align;
3041 uint64_t asize = P2ROUNDUP(psize, align);
3042 char *abuf = zio->io_data;
3044 if (asize != psize) {
3045 abuf = zio_buf_alloc(asize);
3046 bcopy(zio->io_data, abuf, psize);
3047 bzero(abuf + psize, asize - psize);
3050 zio->io_cksum_report = zcr->zcr_next;
3051 zcr->zcr_next = NULL;
3052 zcr->zcr_finish(zcr, abuf);
3053 zfs_ereport_free_checksum(zcr);
3056 zio_buf_free(abuf, asize);
3060 zio_pop_transforms(zio); /* note: may set zio->io_error */
3062 vdev_stat_update(zio, psize);
3064 if (zio->io_error) {
3066 * If this I/O is attached to a particular vdev,
3067 * generate an error message describing the I/O failure
3068 * at the block level. We ignore these errors if the
3069 * device is currently unavailable.
3071 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3072 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3074 if ((zio->io_error == EIO || !(zio->io_flags &
3075 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3078 * For logical I/O requests, tell the SPA to log the
3079 * error and generate a logical data ereport.
3081 spa_log_error(spa, zio);
3082 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3087 if (zio->io_error && zio == lio) {
3089 * Determine whether zio should be reexecuted. This will
3090 * propagate all the way to the root via zio_notify_parent().
3092 ASSERT(vd == NULL && bp != NULL);
3093 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3095 if (IO_IS_ALLOCATING(zio) &&
3096 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3097 if (zio->io_error != ENOSPC)
3098 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3100 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3103 if ((zio->io_type == ZIO_TYPE_READ ||
3104 zio->io_type == ZIO_TYPE_FREE) &&
3105 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3106 zio->io_error == ENXIO &&
3107 spa_load_state(spa) == SPA_LOAD_NONE &&
3108 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3109 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3111 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3112 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3115 * Here is a possibly good place to attempt to do
3116 * either combinatorial reconstruction or error correction
3117 * based on checksums. It also might be a good place
3118 * to send out preliminary ereports before we suspend
3124 * If there were logical child errors, they apply to us now.
3125 * We defer this until now to avoid conflating logical child
3126 * errors with errors that happened to the zio itself when
3127 * updating vdev stats and reporting FMA events above.
3129 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3131 if ((zio->io_error || zio->io_reexecute) &&
3132 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3133 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3134 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3136 zio_gang_tree_free(&zio->io_gang_tree);
3139 * Godfather I/Os should never suspend.
3141 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3142 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3143 zio->io_reexecute = 0;
3145 if (zio->io_reexecute) {
3147 * This is a logical I/O that wants to reexecute.
3149 * Reexecute is top-down. When an i/o fails, if it's not
3150 * the root, it simply notifies its parent and sticks around.
3151 * The parent, seeing that it still has children in zio_done(),
3152 * does the same. This percolates all the way up to the root.
3153 * The root i/o will reexecute or suspend the entire tree.
3155 * This approach ensures that zio_reexecute() honors
3156 * all the original i/o dependency relationships, e.g.
3157 * parents not executing until children are ready.
3159 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3161 zio->io_gang_leader = NULL;
3163 mutex_enter(&zio->io_lock);
3164 zio->io_state[ZIO_WAIT_DONE] = 1;
3165 mutex_exit(&zio->io_lock);
3168 * "The Godfather" I/O monitors its children but is
3169 * not a true parent to them. It will track them through
3170 * the pipeline but severs its ties whenever they get into
3171 * trouble (e.g. suspended). This allows "The Godfather"
3172 * I/O to return status without blocking.
3174 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3175 zio_link_t *zl = zio->io_walk_link;
3176 pio_next = zio_walk_parents(zio);
3178 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3179 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3180 zio_remove_child(pio, zio, zl);
3181 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3185 if ((pio = zio_unique_parent(zio)) != NULL) {
3187 * We're not a root i/o, so there's nothing to do
3188 * but notify our parent. Don't propagate errors
3189 * upward since we haven't permanently failed yet.
3191 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3192 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3193 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3194 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3196 * We'd fail again if we reexecuted now, so suspend
3197 * until conditions improve (e.g. device comes online).
3199 zio_suspend(spa, zio);
3202 * Reexecution is potentially a huge amount of work.
3203 * Hand it off to the otherwise-unused claim taskq.
3205 #if defined(illumos) || !defined(_KERNEL)
3206 ASSERT(zio->io_tqent.tqent_next == NULL);
3208 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
3210 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3211 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3214 return (ZIO_PIPELINE_STOP);
3217 ASSERT(zio->io_child_count == 0);
3218 ASSERT(zio->io_reexecute == 0);
3219 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3222 * Report any checksum errors, since the I/O is complete.
3224 while (zio->io_cksum_report != NULL) {
3225 zio_cksum_report_t *zcr = zio->io_cksum_report;
3226 zio->io_cksum_report = zcr->zcr_next;
3227 zcr->zcr_next = NULL;
3228 zcr->zcr_finish(zcr, NULL);
3229 zfs_ereport_free_checksum(zcr);
3233 * It is the responsibility of the done callback to ensure that this
3234 * particular zio is no longer discoverable for adoption, and as
3235 * such, cannot acquire any new parents.
3240 mutex_enter(&zio->io_lock);
3241 zio->io_state[ZIO_WAIT_DONE] = 1;
3242 mutex_exit(&zio->io_lock);
3244 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3245 zio_link_t *zl = zio->io_walk_link;
3246 pio_next = zio_walk_parents(zio);
3247 zio_remove_child(pio, zio, zl);
3248 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3251 if (zio->io_waiter != NULL) {
3252 mutex_enter(&zio->io_lock);
3253 zio->io_executor = NULL;
3254 cv_broadcast(&zio->io_cv);
3255 mutex_exit(&zio->io_lock);
3260 return (ZIO_PIPELINE_STOP);
3264 * ==========================================================================
3265 * I/O pipeline definition
3266 * ==========================================================================
3268 static zio_pipe_stage_t *zio_pipeline[] = {
3274 zio_checksum_generate,
3289 zio_checksum_verify,
3293 /* dnp is the dnode for zb1->zb_object */
3295 zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_t *zb1,
3296 const zbookmark_t *zb2)
3298 uint64_t zb1nextL0, zb2thisobj;
3300 ASSERT(zb1->zb_objset == zb2->zb_objset);
3301 ASSERT(zb2->zb_level == 0);
3304 * A bookmark in the deadlist is considered to be after
3307 if (zb2->zb_object == DMU_DEADLIST_OBJECT)
3310 /* The objset_phys_t isn't before anything. */
3314 zb1nextL0 = (zb1->zb_blkid + 1) <<
3315 ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
3317 zb2thisobj = zb2->zb_object ? zb2->zb_object :
3318 zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
3320 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3321 uint64_t nextobj = zb1nextL0 *
3322 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
3323 return (nextobj <= zb2thisobj);
3326 if (zb1->zb_object < zb2thisobj)
3328 if (zb1->zb_object > zb2thisobj)
3330 if (zb2->zb_object == DMU_META_DNODE_OBJECT)
3332 return (zb1nextL0 <= zb2->zb_blkid);