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.
26 #include <sys/zfs_context.h>
27 #include <sys/fm/fs/zfs.h>
30 #include <sys/spa_impl.h>
31 #include <sys/vdev_impl.h>
32 #include <sys/zio_impl.h>
33 #include <sys/zio_compress.h>
34 #include <sys/zio_checksum.h>
35 #include <sys/dmu_objset.h>
38 #include <sys/trim_map.h>
40 SYSCTL_DECL(_vfs_zfs);
41 SYSCTL_NODE(_vfs_zfs, OID_AUTO, zio, CTLFLAG_RW, 0, "ZFS ZIO");
42 static int zio_use_uma = 0;
43 TUNABLE_INT("vfs.zfs.zio.use_uma", &zio_use_uma);
44 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, use_uma, CTLFLAG_RDTUN, &zio_use_uma, 0,
45 "Use uma(9) for ZIO allocations");
47 zio_trim_stats_t zio_trim_stats = {
48 { "bytes", KSTAT_DATA_UINT64,
49 "Number of bytes successfully TRIMmed" },
50 { "success", KSTAT_DATA_UINT64,
51 "Number of successful TRIM requests" },
52 { "unsupported", KSTAT_DATA_UINT64,
53 "Number of TRIM requests that failed because TRIM is not supported" },
54 { "failed", KSTAT_DATA_UINT64,
55 "Number of TRIM requests that failed for reasons other than not supported" },
58 static kstat_t *zio_trim_ksp;
61 * ==========================================================================
63 * ==========================================================================
65 uint8_t zio_priority_table[ZIO_PRIORITY_TABLE_SIZE] = {
66 0, /* ZIO_PRIORITY_NOW */
67 0, /* ZIO_PRIORITY_SYNC_READ */
68 0, /* ZIO_PRIORITY_SYNC_WRITE */
69 0, /* ZIO_PRIORITY_LOG_WRITE */
70 1, /* ZIO_PRIORITY_CACHE_FILL */
71 1, /* ZIO_PRIORITY_AGG */
72 4, /* ZIO_PRIORITY_FREE */
73 4, /* ZIO_PRIORITY_ASYNC_WRITE */
74 6, /* ZIO_PRIORITY_ASYNC_READ */
75 10, /* ZIO_PRIORITY_RESILVER */
76 20, /* ZIO_PRIORITY_SCRUB */
77 2, /* ZIO_PRIORITY_DDT_PREFETCH */
78 30, /* ZIO_PRIORITY_TRIM */
82 * ==========================================================================
83 * I/O type descriptions
84 * ==========================================================================
86 char *zio_type_name[ZIO_TYPES] = {
87 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
92 * ==========================================================================
94 * ==========================================================================
96 kmem_cache_t *zio_cache;
97 kmem_cache_t *zio_link_cache;
98 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
99 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
102 extern vmem_t *zio_alloc_arena;
104 extern int zfs_mg_alloc_failures;
107 * The following actions directly effect the spa's sync-to-convergence logic.
108 * The values below define the sync pass when we start performing the action.
109 * Care should be taken when changing these values as they directly impact
110 * spa_sync() performance. Tuning these values may introduce subtle performance
111 * pathologies and should only be done in the context of performance analysis.
112 * These tunables will eventually be removed and replaced with #defines once
113 * enough analysis has been done to determine optimal values.
115 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
116 * regular blocks are not deferred.
118 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
119 TUNABLE_INT("vfs.zfs.sync_pass_deferred_free", &zfs_sync_pass_deferred_free);
120 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_deferred_free, CTLFLAG_RDTUN,
121 &zfs_sync_pass_deferred_free, 0, "defer frees starting in this pass");
122 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
123 TUNABLE_INT("vfs.zfs.sync_pass_dont_compress", &zfs_sync_pass_dont_compress);
124 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_dont_compress, CTLFLAG_RDTUN,
125 &zfs_sync_pass_dont_compress, 0, "don't compress starting in this pass");
126 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
127 TUNABLE_INT("vfs.zfs.sync_pass_rewrite", &zfs_sync_pass_rewrite);
128 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_rewrite, CTLFLAG_RDTUN,
129 &zfs_sync_pass_rewrite, 0, "rewrite new bps starting in this pass");
132 * An allocating zio is one that either currently has the DVA allocate
133 * stage set or will have it later in its lifetime.
135 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
137 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
140 int zio_buf_debug_limit = 16384;
142 int zio_buf_debug_limit = 0;
149 zio_cache = kmem_cache_create("zio_cache",
150 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
151 zio_link_cache = kmem_cache_create("zio_link_cache",
152 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
155 * For small buffers, we want a cache for each multiple of
156 * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache
157 * for each quarter-power of 2. For large buffers, we want
158 * a cache for each multiple of PAGESIZE.
160 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
161 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
164 size_t cflags = (size > zio_buf_debug_limit) ? (KMC_NODEBUG|KMC_NOTOUCH) : 0;
166 while (p2 & (p2 - 1))
172 * If we are using watchpoints, put each buffer on its own page,
173 * to eliminate the performance overhead of trapping to the
174 * kernel when modifying a non-watched buffer that shares the
175 * page with a watched buffer.
177 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
181 if (size <= 4 * SPA_MINBLOCKSIZE) {
182 align = SPA_MINBLOCKSIZE;
183 } else if (IS_P2ALIGNED(size, PAGESIZE)) {
185 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
191 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
192 zio_buf_cache[c] = kmem_cache_create(name, size,
193 align, NULL, NULL, NULL, NULL, NULL, cflags);
196 * Since zio_data bufs do not appear in crash dumps, we
197 * pass KMC_NOTOUCH so that no allocator metadata is
198 * stored with the buffers.
200 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
201 zio_data_buf_cache[c] = kmem_cache_create(name, size,
202 align, NULL, NULL, NULL, NULL, NULL,
203 cflags | KMC_NOTOUCH);
208 ASSERT(zio_buf_cache[c] != NULL);
209 if (zio_buf_cache[c - 1] == NULL)
210 zio_buf_cache[c - 1] = zio_buf_cache[c];
212 ASSERT(zio_data_buf_cache[c] != NULL);
213 if (zio_data_buf_cache[c - 1] == NULL)
214 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
218 * The zio write taskqs have 1 thread per cpu, allow 1/2 of the taskqs
219 * to fail 3 times per txg or 8 failures, whichever is greater.
221 if (zfs_mg_alloc_failures == 0)
222 zfs_mg_alloc_failures = MAX((3 * max_ncpus / 2), 8);
223 else if (zfs_mg_alloc_failures < 8)
224 zfs_mg_alloc_failures = 8;
228 zio_trim_ksp = kstat_create("zfs", 0, "zio_trim", "misc",
230 sizeof(zio_trim_stats) / sizeof(kstat_named_t),
233 if (zio_trim_ksp != NULL) {
234 zio_trim_ksp->ks_data = &zio_trim_stats;
235 kstat_install(zio_trim_ksp);
243 kmem_cache_t *last_cache = NULL;
244 kmem_cache_t *last_data_cache = NULL;
246 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
247 if (zio_buf_cache[c] != last_cache) {
248 last_cache = zio_buf_cache[c];
249 kmem_cache_destroy(zio_buf_cache[c]);
251 zio_buf_cache[c] = NULL;
253 if (zio_data_buf_cache[c] != last_data_cache) {
254 last_data_cache = zio_data_buf_cache[c];
255 kmem_cache_destroy(zio_data_buf_cache[c]);
257 zio_data_buf_cache[c] = NULL;
260 kmem_cache_destroy(zio_link_cache);
261 kmem_cache_destroy(zio_cache);
265 if (zio_trim_ksp != NULL) {
266 kstat_delete(zio_trim_ksp);
272 * ==========================================================================
273 * Allocate and free I/O buffers
274 * ==========================================================================
278 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
279 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
280 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
281 * excess / transient data in-core during a crashdump.
284 zio_buf_alloc(size_t size)
286 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
288 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
291 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
293 return (kmem_alloc(size, KM_SLEEP));
297 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
298 * crashdump if the kernel panics. This exists so that we will limit the amount
299 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
300 * of kernel heap dumped to disk when the kernel panics)
303 zio_data_buf_alloc(size_t size)
305 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
307 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
310 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
312 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG));
316 zio_buf_free(void *buf, size_t size)
318 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
320 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
323 kmem_cache_free(zio_buf_cache[c], buf);
325 kmem_free(buf, size);
329 zio_data_buf_free(void *buf, size_t size)
331 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
333 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
336 kmem_cache_free(zio_data_buf_cache[c], buf);
338 kmem_free(buf, size);
342 * ==========================================================================
343 * Push and pop I/O transform buffers
344 * ==========================================================================
347 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
348 zio_transform_func_t *transform)
350 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
352 zt->zt_orig_data = zio->io_data;
353 zt->zt_orig_size = zio->io_size;
354 zt->zt_bufsize = bufsize;
355 zt->zt_transform = transform;
357 zt->zt_next = zio->io_transform_stack;
358 zio->io_transform_stack = zt;
365 zio_pop_transforms(zio_t *zio)
369 while ((zt = zio->io_transform_stack) != NULL) {
370 if (zt->zt_transform != NULL)
371 zt->zt_transform(zio,
372 zt->zt_orig_data, zt->zt_orig_size);
374 if (zt->zt_bufsize != 0)
375 zio_buf_free(zio->io_data, zt->zt_bufsize);
377 zio->io_data = zt->zt_orig_data;
378 zio->io_size = zt->zt_orig_size;
379 zio->io_transform_stack = zt->zt_next;
381 kmem_free(zt, sizeof (zio_transform_t));
386 * ==========================================================================
387 * I/O transform callbacks for subblocks and decompression
388 * ==========================================================================
391 zio_subblock(zio_t *zio, void *data, uint64_t size)
393 ASSERT(zio->io_size > size);
395 if (zio->io_type == ZIO_TYPE_READ)
396 bcopy(zio->io_data, data, size);
400 zio_decompress(zio_t *zio, void *data, uint64_t size)
402 if (zio->io_error == 0 &&
403 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
404 zio->io_data, data, zio->io_size, size) != 0)
405 zio->io_error = SET_ERROR(EIO);
409 * ==========================================================================
410 * I/O parent/child relationships and pipeline interlocks
411 * ==========================================================================
414 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
415 * continue calling these functions until they return NULL.
416 * Otherwise, the next caller will pick up the list walk in
417 * some indeterminate state. (Otherwise every caller would
418 * have to pass in a cookie to keep the state represented by
419 * io_walk_link, which gets annoying.)
422 zio_walk_parents(zio_t *cio)
424 zio_link_t *zl = cio->io_walk_link;
425 list_t *pl = &cio->io_parent_list;
427 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
428 cio->io_walk_link = zl;
433 ASSERT(zl->zl_child == cio);
434 return (zl->zl_parent);
438 zio_walk_children(zio_t *pio)
440 zio_link_t *zl = pio->io_walk_link;
441 list_t *cl = &pio->io_child_list;
443 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
444 pio->io_walk_link = zl;
449 ASSERT(zl->zl_parent == pio);
450 return (zl->zl_child);
454 zio_unique_parent(zio_t *cio)
456 zio_t *pio = zio_walk_parents(cio);
458 VERIFY(zio_walk_parents(cio) == NULL);
463 zio_add_child(zio_t *pio, zio_t *cio)
465 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
468 * Logical I/Os can have logical, gang, or vdev children.
469 * Gang I/Os can have gang or vdev children.
470 * Vdev I/Os can only have vdev children.
471 * The following ASSERT captures all of these constraints.
473 ASSERT(cio->io_child_type <= pio->io_child_type);
478 mutex_enter(&cio->io_lock);
479 mutex_enter(&pio->io_lock);
481 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
483 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
484 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
486 list_insert_head(&pio->io_child_list, zl);
487 list_insert_head(&cio->io_parent_list, zl);
489 pio->io_child_count++;
490 cio->io_parent_count++;
492 mutex_exit(&pio->io_lock);
493 mutex_exit(&cio->io_lock);
497 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
499 ASSERT(zl->zl_parent == pio);
500 ASSERT(zl->zl_child == cio);
502 mutex_enter(&cio->io_lock);
503 mutex_enter(&pio->io_lock);
505 list_remove(&pio->io_child_list, zl);
506 list_remove(&cio->io_parent_list, zl);
508 pio->io_child_count--;
509 cio->io_parent_count--;
511 mutex_exit(&pio->io_lock);
512 mutex_exit(&cio->io_lock);
514 kmem_cache_free(zio_link_cache, zl);
518 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
520 uint64_t *countp = &zio->io_children[child][wait];
521 boolean_t waiting = B_FALSE;
523 mutex_enter(&zio->io_lock);
524 ASSERT(zio->io_stall == NULL);
527 zio->io_stall = countp;
530 mutex_exit(&zio->io_lock);
536 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
538 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
539 int *errorp = &pio->io_child_error[zio->io_child_type];
541 mutex_enter(&pio->io_lock);
542 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
543 *errorp = zio_worst_error(*errorp, zio->io_error);
544 pio->io_reexecute |= zio->io_reexecute;
545 ASSERT3U(*countp, >, 0);
546 if (--*countp == 0 && pio->io_stall == countp) {
547 pio->io_stall = NULL;
548 mutex_exit(&pio->io_lock);
551 mutex_exit(&pio->io_lock);
556 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
558 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
559 zio->io_error = zio->io_child_error[c];
563 * ==========================================================================
564 * Create the various types of I/O (read, write, free, etc)
565 * ==========================================================================
568 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
569 void *data, uint64_t size, zio_done_func_t *done, void *private,
570 zio_type_t type, int priority, enum zio_flag flags,
571 vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
572 enum zio_stage stage, enum zio_stage pipeline)
576 ASSERT3U(type == ZIO_TYPE_FREE || size, <=, SPA_MAXBLOCKSIZE);
577 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
578 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
580 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
581 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
582 ASSERT(vd || stage == ZIO_STAGE_OPEN);
584 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
585 bzero(zio, sizeof (zio_t));
587 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
588 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
590 list_create(&zio->io_parent_list, sizeof (zio_link_t),
591 offsetof(zio_link_t, zl_parent_node));
592 list_create(&zio->io_child_list, sizeof (zio_link_t),
593 offsetof(zio_link_t, zl_child_node));
596 zio->io_child_type = ZIO_CHILD_VDEV;
597 else if (flags & ZIO_FLAG_GANG_CHILD)
598 zio->io_child_type = ZIO_CHILD_GANG;
599 else if (flags & ZIO_FLAG_DDT_CHILD)
600 zio->io_child_type = ZIO_CHILD_DDT;
602 zio->io_child_type = ZIO_CHILD_LOGICAL;
605 zio->io_bp = (blkptr_t *)bp;
606 zio->io_bp_copy = *bp;
607 zio->io_bp_orig = *bp;
608 if (type != ZIO_TYPE_WRITE ||
609 zio->io_child_type == ZIO_CHILD_DDT)
610 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
611 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
612 zio->io_logical = zio;
613 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
614 pipeline |= ZIO_GANG_STAGES;
620 zio->io_private = private;
622 zio->io_priority = priority;
624 zio->io_offset = offset;
625 zio->io_orig_data = zio->io_data = data;
626 zio->io_orig_size = zio->io_size = size;
627 zio->io_orig_flags = zio->io_flags = flags;
628 zio->io_orig_stage = zio->io_stage = stage;
629 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
631 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
632 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
635 zio->io_bookmark = *zb;
638 if (zio->io_logical == NULL)
639 zio->io_logical = pio->io_logical;
640 if (zio->io_child_type == ZIO_CHILD_GANG)
641 zio->io_gang_leader = pio->io_gang_leader;
642 zio_add_child(pio, zio);
649 zio_destroy(zio_t *zio)
651 list_destroy(&zio->io_parent_list);
652 list_destroy(&zio->io_child_list);
653 mutex_destroy(&zio->io_lock);
654 cv_destroy(&zio->io_cv);
655 kmem_cache_free(zio_cache, zio);
659 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
660 void *private, enum zio_flag flags)
664 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
665 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
666 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
672 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
674 return (zio_null(NULL, spa, NULL, done, private, flags));
678 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
679 void *data, uint64_t size, zio_done_func_t *done, void *private,
680 int priority, enum zio_flag flags, const zbookmark_t *zb)
684 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
685 data, size, done, private,
686 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
687 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
688 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
694 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
695 void *data, uint64_t size, const zio_prop_t *zp,
696 zio_done_func_t *ready, zio_done_func_t *done, void *private,
697 int priority, enum zio_flag flags, const zbookmark_t *zb)
701 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
702 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
703 zp->zp_compress >= ZIO_COMPRESS_OFF &&
704 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
705 DMU_OT_IS_VALID(zp->zp_type) &&
708 zp->zp_copies <= spa_max_replication(spa));
710 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
711 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
712 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
713 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
715 zio->io_ready = ready;
722 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
723 uint64_t size, zio_done_func_t *done, void *private, int priority,
724 enum zio_flag flags, zbookmark_t *zb)
728 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
729 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
730 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
736 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
738 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
739 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
740 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
741 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
744 * We must reset the io_prop to match the values that existed
745 * when the bp was first written by dmu_sync() keeping in mind
746 * that nopwrite and dedup are mutually exclusive.
748 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
749 zio->io_prop.zp_nopwrite = nopwrite;
750 zio->io_prop.zp_copies = copies;
751 zio->io_bp_override = bp;
755 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
757 metaslab_check_free(spa, bp);
758 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
762 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
763 uint64_t size, enum zio_flag flags)
767 dprintf_bp(bp, "freeing in txg %llu, pass %u",
768 (longlong_t)txg, spa->spa_sync_pass);
770 ASSERT(!BP_IS_HOLE(bp));
771 ASSERT(spa_syncing_txg(spa) == txg);
772 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
774 metaslab_check_free(spa, bp);
776 zio = zio_create(pio, spa, txg, bp, NULL, size,
777 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
778 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
784 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
785 zio_done_func_t *done, void *private, enum zio_flag flags)
790 * A claim is an allocation of a specific block. Claims are needed
791 * to support immediate writes in the intent log. The issue is that
792 * immediate writes contain committed data, but in a txg that was
793 * *not* committed. Upon opening the pool after an unclean shutdown,
794 * the intent log claims all blocks that contain immediate write data
795 * so that the SPA knows they're in use.
797 * All claims *must* be resolved in the first txg -- before the SPA
798 * starts allocating blocks -- so that nothing is allocated twice.
799 * If txg == 0 we just verify that the block is claimable.
801 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
802 ASSERT(txg == spa_first_txg(spa) || txg == 0);
803 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
805 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
806 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
807 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
813 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset,
814 uint64_t size, zio_done_func_t *done, void *private, int priority,
820 if (vd->vdev_children == 0) {
821 zio = zio_create(pio, spa, 0, NULL, NULL, size, done, private,
822 ZIO_TYPE_IOCTL, priority, flags, vd, offset, NULL,
823 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
827 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
829 for (c = 0; c < vd->vdev_children; c++)
830 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
831 offset, size, done, private, priority, flags));
838 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
839 void *data, int checksum, zio_done_func_t *done, void *private,
840 int priority, enum zio_flag flags, boolean_t labels)
844 ASSERT(vd->vdev_children == 0);
845 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
846 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
847 ASSERT3U(offset + size, <=, vd->vdev_psize);
849 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
850 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
851 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
853 zio->io_prop.zp_checksum = checksum;
859 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
860 void *data, int checksum, zio_done_func_t *done, void *private,
861 int priority, enum zio_flag flags, boolean_t labels)
865 ASSERT(vd->vdev_children == 0);
866 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
867 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
868 ASSERT3U(offset + size, <=, vd->vdev_psize);
870 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
871 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
872 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
874 zio->io_prop.zp_checksum = checksum;
876 if (zio_checksum_table[checksum].ci_eck) {
878 * zec checksums are necessarily destructive -- they modify
879 * the end of the write buffer to hold the verifier/checksum.
880 * Therefore, we must make a local copy in case the data is
881 * being written to multiple places in parallel.
883 void *wbuf = zio_buf_alloc(size);
884 bcopy(data, wbuf, size);
885 zio_push_transform(zio, wbuf, size, size, NULL);
892 * Create a child I/O to do some work for us.
895 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
896 void *data, uint64_t size, int type, int priority, enum zio_flag flags,
897 zio_done_func_t *done, void *private)
899 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
902 ASSERT(vd->vdev_parent ==
903 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
905 if (type == ZIO_TYPE_READ && bp != NULL) {
907 * If we have the bp, then the child should perform the
908 * checksum and the parent need not. This pushes error
909 * detection as close to the leaves as possible and
910 * eliminates redundant checksums in the interior nodes.
912 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
913 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
916 if (vd->vdev_children == 0)
917 offset += VDEV_LABEL_START_SIZE;
919 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
922 * If we've decided to do a repair, the write is not speculative --
923 * even if the original read was.
925 if (flags & ZIO_FLAG_IO_REPAIR)
926 flags &= ~ZIO_FLAG_SPECULATIVE;
928 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
929 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
930 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
936 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
937 int type, int priority, enum zio_flag flags,
938 zio_done_func_t *done, void *private)
942 ASSERT(vd->vdev_ops->vdev_op_leaf);
944 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
945 data, size, done, private, type, priority,
946 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
948 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
954 zio_flush(zio_t *zio, vdev_t *vd)
956 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0,
957 NULL, NULL, ZIO_PRIORITY_NOW,
958 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
962 zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size)
965 ASSERT(vd->vdev_ops->vdev_op_leaf);
967 return zio_ioctl(zio, spa, vd, DKIOCTRIM, offset, size,
968 NULL, NULL, ZIO_PRIORITY_TRIM,
969 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY);
973 zio_shrink(zio_t *zio, uint64_t size)
975 ASSERT(zio->io_executor == NULL);
976 ASSERT(zio->io_orig_size == zio->io_size);
977 ASSERT(size <= zio->io_size);
980 * We don't shrink for raidz because of problems with the
981 * reconstruction when reading back less than the block size.
982 * Note, BP_IS_RAIDZ() assumes no compression.
984 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
985 if (!BP_IS_RAIDZ(zio->io_bp))
986 zio->io_orig_size = zio->io_size = size;
990 * ==========================================================================
991 * Prepare to read and write logical blocks
992 * ==========================================================================
996 zio_read_bp_init(zio_t *zio)
998 blkptr_t *bp = zio->io_bp;
1000 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1001 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1002 !(zio->io_flags & ZIO_FLAG_RAW)) {
1003 uint64_t psize = BP_GET_PSIZE(bp);
1004 void *cbuf = zio_buf_alloc(psize);
1006 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
1009 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1010 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1012 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1013 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1015 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1016 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1018 return (ZIO_PIPELINE_CONTINUE);
1022 zio_write_bp_init(zio_t *zio)
1024 spa_t *spa = zio->io_spa;
1025 zio_prop_t *zp = &zio->io_prop;
1026 enum zio_compress compress = zp->zp_compress;
1027 blkptr_t *bp = zio->io_bp;
1028 uint64_t lsize = zio->io_size;
1029 uint64_t psize = lsize;
1033 * If our children haven't all reached the ready stage,
1034 * wait for them and then repeat this pipeline stage.
1036 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1037 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1038 return (ZIO_PIPELINE_STOP);
1040 if (!IO_IS_ALLOCATING(zio))
1041 return (ZIO_PIPELINE_CONTINUE);
1043 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1045 if (zio->io_bp_override) {
1046 ASSERT(bp->blk_birth != zio->io_txg);
1047 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1049 *bp = *zio->io_bp_override;
1050 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1053 * If we've been overridden and nopwrite is set then
1054 * set the flag accordingly to indicate that a nopwrite
1055 * has already occurred.
1057 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1058 ASSERT(!zp->zp_dedup);
1059 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1060 return (ZIO_PIPELINE_CONTINUE);
1063 ASSERT(!zp->zp_nopwrite);
1065 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1066 return (ZIO_PIPELINE_CONTINUE);
1068 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1069 zp->zp_dedup_verify);
1071 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1072 BP_SET_DEDUP(bp, 1);
1073 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1074 return (ZIO_PIPELINE_CONTINUE);
1076 zio->io_bp_override = NULL;
1080 if (bp->blk_birth == zio->io_txg) {
1082 * We're rewriting an existing block, which means we're
1083 * working on behalf of spa_sync(). For spa_sync() to
1084 * converge, it must eventually be the case that we don't
1085 * have to allocate new blocks. But compression changes
1086 * the blocksize, which forces a reallocate, and makes
1087 * convergence take longer. Therefore, after the first
1088 * few passes, stop compressing to ensure convergence.
1090 pass = spa_sync_pass(spa);
1092 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1093 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1094 ASSERT(!BP_GET_DEDUP(bp));
1096 if (pass >= zfs_sync_pass_dont_compress)
1097 compress = ZIO_COMPRESS_OFF;
1099 /* Make sure someone doesn't change their mind on overwrites */
1100 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
1101 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1104 if (compress != ZIO_COMPRESS_OFF) {
1105 void *cbuf = zio_buf_alloc(lsize);
1106 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1107 if (psize == 0 || psize == lsize) {
1108 compress = ZIO_COMPRESS_OFF;
1109 zio_buf_free(cbuf, lsize);
1111 ASSERT(psize < lsize);
1112 zio_push_transform(zio, cbuf, psize, lsize, NULL);
1117 * The final pass of spa_sync() must be all rewrites, but the first
1118 * few passes offer a trade-off: allocating blocks defers convergence,
1119 * but newly allocated blocks are sequential, so they can be written
1120 * to disk faster. Therefore, we allow the first few passes of
1121 * spa_sync() to allocate new blocks, but force rewrites after that.
1122 * There should only be a handful of blocks after pass 1 in any case.
1124 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
1125 pass >= zfs_sync_pass_rewrite) {
1127 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1128 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1129 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1132 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1136 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1138 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1139 BP_SET_LSIZE(bp, lsize);
1140 BP_SET_PSIZE(bp, psize);
1141 BP_SET_COMPRESS(bp, compress);
1142 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1143 BP_SET_TYPE(bp, zp->zp_type);
1144 BP_SET_LEVEL(bp, zp->zp_level);
1145 BP_SET_DEDUP(bp, zp->zp_dedup);
1146 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1148 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1149 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1150 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1152 if (zp->zp_nopwrite) {
1153 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1154 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1155 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1159 return (ZIO_PIPELINE_CONTINUE);
1163 zio_free_bp_init(zio_t *zio)
1165 blkptr_t *bp = zio->io_bp;
1167 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1168 if (BP_GET_DEDUP(bp))
1169 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1172 return (ZIO_PIPELINE_CONTINUE);
1176 * ==========================================================================
1177 * Execute the I/O pipeline
1178 * ==========================================================================
1182 zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q, boolean_t cutinline)
1184 spa_t *spa = zio->io_spa;
1185 zio_type_t t = zio->io_type;
1186 int flags = TQ_SLEEP | (cutinline ? TQ_FRONT : 0);
1188 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1191 * If we're a config writer or a probe, the normal issue and
1192 * interrupt threads may all be blocked waiting for the config lock.
1193 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1195 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1199 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1201 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1205 * If this is a high priority I/O, then use the high priority taskq.
1207 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1208 spa->spa_zio_taskq[t][q + 1] != NULL)
1211 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1213 (void) taskq_dispatch_safe(spa->spa_zio_taskq[t][q],
1214 (task_func_t *)zio_execute, zio, flags, &zio->io_task);
1216 (void) taskq_dispatch(spa->spa_zio_taskq[t][q],
1217 (task_func_t *)zio_execute, zio, flags);
1222 zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
1224 kthread_t *executor = zio->io_executor;
1225 spa_t *spa = zio->io_spa;
1227 for (zio_type_t t = 0; t < ZIO_TYPES; t++)
1228 if (taskq_member(spa->spa_zio_taskq[t][q], executor))
1235 zio_issue_async(zio_t *zio)
1237 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1239 return (ZIO_PIPELINE_STOP);
1243 zio_interrupt(zio_t *zio)
1245 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1249 * Execute the I/O pipeline until one of the following occurs:
1250 * (1) the I/O completes; (2) the pipeline stalls waiting for
1251 * dependent child I/Os; (3) the I/O issues, so we're waiting
1252 * for an I/O completion interrupt; (4) the I/O is delegated by
1253 * vdev-level caching or aggregation; (5) the I/O is deferred
1254 * due to vdev-level queueing; (6) the I/O is handed off to
1255 * another thread. In all cases, the pipeline stops whenever
1256 * there's no CPU work; it never burns a thread in cv_wait().
1258 * There's no locking on io_stage because there's no legitimate way
1259 * for multiple threads to be attempting to process the same I/O.
1261 static zio_pipe_stage_t *zio_pipeline[];
1264 zio_execute(zio_t *zio)
1266 zio->io_executor = curthread;
1268 while (zio->io_stage < ZIO_STAGE_DONE) {
1269 enum zio_stage pipeline = zio->io_pipeline;
1270 enum zio_stage stage = zio->io_stage;
1273 ASSERT(!MUTEX_HELD(&zio->io_lock));
1274 ASSERT(ISP2(stage));
1275 ASSERT(zio->io_stall == NULL);
1279 } while ((stage & pipeline) == 0);
1281 ASSERT(stage <= ZIO_STAGE_DONE);
1284 * If we are in interrupt context and this pipeline stage
1285 * will grab a config lock that is held across I/O,
1286 * or may wait for an I/O that needs an interrupt thread
1287 * to complete, issue async to avoid deadlock.
1289 * For VDEV_IO_START, we cut in line so that the io will
1290 * be sent to disk promptly.
1292 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1293 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1294 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1295 zio_requeue_io_start_cut_in_line : B_FALSE;
1296 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1300 zio->io_stage = stage;
1301 rv = zio_pipeline[highbit(stage) - 1](zio);
1303 if (rv == ZIO_PIPELINE_STOP)
1306 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1311 * ==========================================================================
1312 * Initiate I/O, either sync or async
1313 * ==========================================================================
1316 zio_wait(zio_t *zio)
1320 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1321 ASSERT(zio->io_executor == NULL);
1323 zio->io_waiter = curthread;
1327 mutex_enter(&zio->io_lock);
1328 while (zio->io_executor != NULL)
1329 cv_wait(&zio->io_cv, &zio->io_lock);
1330 mutex_exit(&zio->io_lock);
1332 error = zio->io_error;
1339 zio_nowait(zio_t *zio)
1341 ASSERT(zio->io_executor == NULL);
1343 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1344 zio_unique_parent(zio) == NULL) {
1346 * This is a logical async I/O with no parent to wait for it.
1347 * We add it to the spa_async_root_zio "Godfather" I/O which
1348 * will ensure they complete prior to unloading the pool.
1350 spa_t *spa = zio->io_spa;
1352 zio_add_child(spa->spa_async_zio_root, zio);
1359 * ==========================================================================
1360 * Reexecute or suspend/resume failed I/O
1361 * ==========================================================================
1365 zio_reexecute(zio_t *pio)
1367 zio_t *cio, *cio_next;
1369 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1370 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1371 ASSERT(pio->io_gang_leader == NULL);
1372 ASSERT(pio->io_gang_tree == NULL);
1374 pio->io_flags = pio->io_orig_flags;
1375 pio->io_stage = pio->io_orig_stage;
1376 pio->io_pipeline = pio->io_orig_pipeline;
1377 pio->io_reexecute = 0;
1378 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1380 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1381 pio->io_state[w] = 0;
1382 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1383 pio->io_child_error[c] = 0;
1385 if (IO_IS_ALLOCATING(pio))
1386 BP_ZERO(pio->io_bp);
1389 * As we reexecute pio's children, new children could be created.
1390 * New children go to the head of pio's io_child_list, however,
1391 * so we will (correctly) not reexecute them. The key is that
1392 * the remainder of pio's io_child_list, from 'cio_next' onward,
1393 * cannot be affected by any side effects of reexecuting 'cio'.
1395 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1396 cio_next = zio_walk_children(pio);
1397 mutex_enter(&pio->io_lock);
1398 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1399 pio->io_children[cio->io_child_type][w]++;
1400 mutex_exit(&pio->io_lock);
1405 * Now that all children have been reexecuted, execute the parent.
1406 * We don't reexecute "The Godfather" I/O here as it's the
1407 * responsibility of the caller to wait on him.
1409 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1414 zio_suspend(spa_t *spa, zio_t *zio)
1416 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1417 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1418 "failure and the failure mode property for this pool "
1419 "is set to panic.", spa_name(spa));
1421 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1423 mutex_enter(&spa->spa_suspend_lock);
1425 if (spa->spa_suspend_zio_root == NULL)
1426 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1427 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1428 ZIO_FLAG_GODFATHER);
1430 spa->spa_suspended = B_TRUE;
1433 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1434 ASSERT(zio != spa->spa_suspend_zio_root);
1435 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1436 ASSERT(zio_unique_parent(zio) == NULL);
1437 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1438 zio_add_child(spa->spa_suspend_zio_root, zio);
1441 mutex_exit(&spa->spa_suspend_lock);
1445 zio_resume(spa_t *spa)
1450 * Reexecute all previously suspended i/o.
1452 mutex_enter(&spa->spa_suspend_lock);
1453 spa->spa_suspended = B_FALSE;
1454 cv_broadcast(&spa->spa_suspend_cv);
1455 pio = spa->spa_suspend_zio_root;
1456 spa->spa_suspend_zio_root = NULL;
1457 mutex_exit(&spa->spa_suspend_lock);
1463 return (zio_wait(pio));
1467 zio_resume_wait(spa_t *spa)
1469 mutex_enter(&spa->spa_suspend_lock);
1470 while (spa_suspended(spa))
1471 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1472 mutex_exit(&spa->spa_suspend_lock);
1476 * ==========================================================================
1479 * A gang block is a collection of small blocks that looks to the DMU
1480 * like one large block. When zio_dva_allocate() cannot find a block
1481 * of the requested size, due to either severe fragmentation or the pool
1482 * being nearly full, it calls zio_write_gang_block() to construct the
1483 * block from smaller fragments.
1485 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1486 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1487 * an indirect block: it's an array of block pointers. It consumes
1488 * only one sector and hence is allocatable regardless of fragmentation.
1489 * The gang header's bps point to its gang members, which hold the data.
1491 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1492 * as the verifier to ensure uniqueness of the SHA256 checksum.
1493 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1494 * not the gang header. This ensures that data block signatures (needed for
1495 * deduplication) are independent of how the block is physically stored.
1497 * Gang blocks can be nested: a gang member may itself be a gang block.
1498 * Thus every gang block is a tree in which root and all interior nodes are
1499 * gang headers, and the leaves are normal blocks that contain user data.
1500 * The root of the gang tree is called the gang leader.
1502 * To perform any operation (read, rewrite, free, claim) on a gang block,
1503 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1504 * in the io_gang_tree field of the original logical i/o by recursively
1505 * reading the gang leader and all gang headers below it. This yields
1506 * an in-core tree containing the contents of every gang header and the
1507 * bps for every constituent of the gang block.
1509 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1510 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1511 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1512 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1513 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1514 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1515 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1516 * of the gang header plus zio_checksum_compute() of the data to update the
1517 * gang header's blk_cksum as described above.
1519 * The two-phase assemble/issue model solves the problem of partial failure --
1520 * what if you'd freed part of a gang block but then couldn't read the
1521 * gang header for another part? Assembling the entire gang tree first
1522 * ensures that all the necessary gang header I/O has succeeded before
1523 * starting the actual work of free, claim, or write. Once the gang tree
1524 * is assembled, free and claim are in-memory operations that cannot fail.
1526 * In the event that a gang write fails, zio_dva_unallocate() walks the
1527 * gang tree to immediately free (i.e. insert back into the space map)
1528 * everything we've allocated. This ensures that we don't get ENOSPC
1529 * errors during repeated suspend/resume cycles due to a flaky device.
1531 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1532 * the gang tree, we won't modify the block, so we can safely defer the free
1533 * (knowing that the block is still intact). If we *can* assemble the gang
1534 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1535 * each constituent bp and we can allocate a new block on the next sync pass.
1537 * In all cases, the gang tree allows complete recovery from partial failure.
1538 * ==========================================================================
1542 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1547 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1548 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1549 &pio->io_bookmark));
1553 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1558 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1559 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1560 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1562 * As we rewrite each gang header, the pipeline will compute
1563 * a new gang block header checksum for it; but no one will
1564 * compute a new data checksum, so we do that here. The one
1565 * exception is the gang leader: the pipeline already computed
1566 * its data checksum because that stage precedes gang assembly.
1567 * (Presently, nothing actually uses interior data checksums;
1568 * this is just good hygiene.)
1570 if (gn != pio->io_gang_leader->io_gang_tree) {
1571 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1572 data, BP_GET_PSIZE(bp));
1575 * If we are here to damage data for testing purposes,
1576 * leave the GBH alone so that we can detect the damage.
1578 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1579 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1581 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1582 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1583 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1591 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1593 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1594 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp),
1595 ZIO_GANG_CHILD_FLAGS(pio)));
1600 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1602 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1603 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1606 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1615 static void zio_gang_tree_assemble_done(zio_t *zio);
1617 static zio_gang_node_t *
1618 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1620 zio_gang_node_t *gn;
1622 ASSERT(*gnpp == NULL);
1624 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1625 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1632 zio_gang_node_free(zio_gang_node_t **gnpp)
1634 zio_gang_node_t *gn = *gnpp;
1636 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1637 ASSERT(gn->gn_child[g] == NULL);
1639 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1640 kmem_free(gn, sizeof (*gn));
1645 zio_gang_tree_free(zio_gang_node_t **gnpp)
1647 zio_gang_node_t *gn = *gnpp;
1652 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1653 zio_gang_tree_free(&gn->gn_child[g]);
1655 zio_gang_node_free(gnpp);
1659 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1661 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1663 ASSERT(gio->io_gang_leader == gio);
1664 ASSERT(BP_IS_GANG(bp));
1666 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1667 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1668 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1672 zio_gang_tree_assemble_done(zio_t *zio)
1674 zio_t *gio = zio->io_gang_leader;
1675 zio_gang_node_t *gn = zio->io_private;
1676 blkptr_t *bp = zio->io_bp;
1678 ASSERT(gio == zio_unique_parent(zio));
1679 ASSERT(zio->io_child_count == 0);
1684 if (BP_SHOULD_BYTESWAP(bp))
1685 byteswap_uint64_array(zio->io_data, zio->io_size);
1687 ASSERT(zio->io_data == gn->gn_gbh);
1688 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1689 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1691 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1692 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1693 if (!BP_IS_GANG(gbp))
1695 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1700 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1702 zio_t *gio = pio->io_gang_leader;
1705 ASSERT(BP_IS_GANG(bp) == !!gn);
1706 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1707 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1710 * If you're a gang header, your data is in gn->gn_gbh.
1711 * If you're a gang member, your data is in 'data' and gn == NULL.
1713 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1716 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1718 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1719 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1720 if (BP_IS_HOLE(gbp))
1722 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1723 data = (char *)data + BP_GET_PSIZE(gbp);
1727 if (gn == gio->io_gang_tree && gio->io_data != NULL)
1728 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1735 zio_gang_assemble(zio_t *zio)
1737 blkptr_t *bp = zio->io_bp;
1739 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1740 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1742 zio->io_gang_leader = zio;
1744 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1746 return (ZIO_PIPELINE_CONTINUE);
1750 zio_gang_issue(zio_t *zio)
1752 blkptr_t *bp = zio->io_bp;
1754 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1755 return (ZIO_PIPELINE_STOP);
1757 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1758 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1760 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1761 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1763 zio_gang_tree_free(&zio->io_gang_tree);
1765 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1767 return (ZIO_PIPELINE_CONTINUE);
1771 zio_write_gang_member_ready(zio_t *zio)
1773 zio_t *pio = zio_unique_parent(zio);
1774 zio_t *gio = zio->io_gang_leader;
1775 dva_t *cdva = zio->io_bp->blk_dva;
1776 dva_t *pdva = pio->io_bp->blk_dva;
1779 if (BP_IS_HOLE(zio->io_bp))
1782 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1784 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1785 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1786 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1787 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1788 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1790 mutex_enter(&pio->io_lock);
1791 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1792 ASSERT(DVA_GET_GANG(&pdva[d]));
1793 asize = DVA_GET_ASIZE(&pdva[d]);
1794 asize += DVA_GET_ASIZE(&cdva[d]);
1795 DVA_SET_ASIZE(&pdva[d], asize);
1797 mutex_exit(&pio->io_lock);
1801 zio_write_gang_block(zio_t *pio)
1803 spa_t *spa = pio->io_spa;
1804 blkptr_t *bp = pio->io_bp;
1805 zio_t *gio = pio->io_gang_leader;
1807 zio_gang_node_t *gn, **gnpp;
1808 zio_gbh_phys_t *gbh;
1809 uint64_t txg = pio->io_txg;
1810 uint64_t resid = pio->io_size;
1812 int copies = gio->io_prop.zp_copies;
1813 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1817 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1818 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1819 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1821 pio->io_error = error;
1822 return (ZIO_PIPELINE_CONTINUE);
1826 gnpp = &gio->io_gang_tree;
1828 gnpp = pio->io_private;
1829 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1832 gn = zio_gang_node_alloc(gnpp);
1834 bzero(gbh, SPA_GANGBLOCKSIZE);
1837 * Create the gang header.
1839 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1840 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1843 * Create and nowait the gang children.
1845 for (int g = 0; resid != 0; resid -= lsize, g++) {
1846 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1848 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1850 zp.zp_checksum = gio->io_prop.zp_checksum;
1851 zp.zp_compress = ZIO_COMPRESS_OFF;
1852 zp.zp_type = DMU_OT_NONE;
1854 zp.zp_copies = gio->io_prop.zp_copies;
1855 zp.zp_dedup = B_FALSE;
1856 zp.zp_dedup_verify = B_FALSE;
1857 zp.zp_nopwrite = B_FALSE;
1859 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1860 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1861 zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1862 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1863 &pio->io_bookmark));
1867 * Set pio's pipeline to just wait for zio to finish.
1869 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1873 return (ZIO_PIPELINE_CONTINUE);
1877 * The zio_nop_write stage in the pipeline determines if allocating
1878 * a new bp is necessary. By leveraging a cryptographically secure checksum,
1879 * such as SHA256, we can compare the checksums of the new data and the old
1880 * to determine if allocating a new block is required. The nopwrite
1881 * feature can handle writes in either syncing or open context (i.e. zil
1882 * writes) and as a result is mutually exclusive with dedup.
1885 zio_nop_write(zio_t *zio)
1887 blkptr_t *bp = zio->io_bp;
1888 blkptr_t *bp_orig = &zio->io_bp_orig;
1889 zio_prop_t *zp = &zio->io_prop;
1891 ASSERT(BP_GET_LEVEL(bp) == 0);
1892 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1893 ASSERT(zp->zp_nopwrite);
1894 ASSERT(!zp->zp_dedup);
1895 ASSERT(zio->io_bp_override == NULL);
1896 ASSERT(IO_IS_ALLOCATING(zio));
1899 * Check to see if the original bp and the new bp have matching
1900 * characteristics (i.e. same checksum, compression algorithms, etc).
1901 * If they don't then just continue with the pipeline which will
1902 * allocate a new bp.
1904 if (BP_IS_HOLE(bp_orig) ||
1905 !zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup ||
1906 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
1907 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
1908 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
1909 zp->zp_copies != BP_GET_NDVAS(bp_orig))
1910 return (ZIO_PIPELINE_CONTINUE);
1913 * If the checksums match then reset the pipeline so that we
1914 * avoid allocating a new bp and issuing any I/O.
1916 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
1917 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup);
1918 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
1919 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
1920 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
1921 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
1922 sizeof (uint64_t)) == 0);
1925 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1926 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1929 return (ZIO_PIPELINE_CONTINUE);
1933 * ==========================================================================
1935 * ==========================================================================
1938 zio_ddt_child_read_done(zio_t *zio)
1940 blkptr_t *bp = zio->io_bp;
1941 ddt_entry_t *dde = zio->io_private;
1943 zio_t *pio = zio_unique_parent(zio);
1945 mutex_enter(&pio->io_lock);
1946 ddp = ddt_phys_select(dde, bp);
1947 if (zio->io_error == 0)
1948 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1949 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1950 dde->dde_repair_data = zio->io_data;
1952 zio_buf_free(zio->io_data, zio->io_size);
1953 mutex_exit(&pio->io_lock);
1957 zio_ddt_read_start(zio_t *zio)
1959 blkptr_t *bp = zio->io_bp;
1961 ASSERT(BP_GET_DEDUP(bp));
1962 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1963 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1965 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1966 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1967 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1968 ddt_phys_t *ddp = dde->dde_phys;
1969 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1972 ASSERT(zio->io_vsd == NULL);
1975 if (ddp_self == NULL)
1976 return (ZIO_PIPELINE_CONTINUE);
1978 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1979 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1981 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1983 zio_nowait(zio_read(zio, zio->io_spa, &blk,
1984 zio_buf_alloc(zio->io_size), zio->io_size,
1985 zio_ddt_child_read_done, dde, zio->io_priority,
1986 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1987 &zio->io_bookmark));
1989 return (ZIO_PIPELINE_CONTINUE);
1992 zio_nowait(zio_read(zio, zio->io_spa, bp,
1993 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1994 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
1996 return (ZIO_PIPELINE_CONTINUE);
2000 zio_ddt_read_done(zio_t *zio)
2002 blkptr_t *bp = zio->io_bp;
2004 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2005 return (ZIO_PIPELINE_STOP);
2007 ASSERT(BP_GET_DEDUP(bp));
2008 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2009 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2011 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2012 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2013 ddt_entry_t *dde = zio->io_vsd;
2015 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2016 return (ZIO_PIPELINE_CONTINUE);
2019 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2020 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2021 return (ZIO_PIPELINE_STOP);
2023 if (dde->dde_repair_data != NULL) {
2024 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2025 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2027 ddt_repair_done(ddt, dde);
2031 ASSERT(zio->io_vsd == NULL);
2033 return (ZIO_PIPELINE_CONTINUE);
2037 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2039 spa_t *spa = zio->io_spa;
2042 * Note: we compare the original data, not the transformed data,
2043 * because when zio->io_bp is an override bp, we will not have
2044 * pushed the I/O transforms. That's an important optimization
2045 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2047 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2048 zio_t *lio = dde->dde_lead_zio[p];
2051 return (lio->io_orig_size != zio->io_orig_size ||
2052 bcmp(zio->io_orig_data, lio->io_orig_data,
2053 zio->io_orig_size) != 0);
2057 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2058 ddt_phys_t *ddp = &dde->dde_phys[p];
2060 if (ddp->ddp_phys_birth != 0) {
2061 arc_buf_t *abuf = NULL;
2062 uint32_t aflags = ARC_WAIT;
2063 blkptr_t blk = *zio->io_bp;
2066 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2070 error = arc_read(NULL, spa, &blk,
2071 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2072 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2073 &aflags, &zio->io_bookmark);
2076 if (arc_buf_size(abuf) != zio->io_orig_size ||
2077 bcmp(abuf->b_data, zio->io_orig_data,
2078 zio->io_orig_size) != 0)
2079 error = SET_ERROR(EEXIST);
2080 VERIFY(arc_buf_remove_ref(abuf, &abuf));
2084 return (error != 0);
2092 zio_ddt_child_write_ready(zio_t *zio)
2094 int p = zio->io_prop.zp_copies;
2095 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2096 ddt_entry_t *dde = zio->io_private;
2097 ddt_phys_t *ddp = &dde->dde_phys[p];
2105 ASSERT(dde->dde_lead_zio[p] == zio);
2107 ddt_phys_fill(ddp, zio->io_bp);
2109 while ((pio = zio_walk_parents(zio)) != NULL)
2110 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2116 zio_ddt_child_write_done(zio_t *zio)
2118 int p = zio->io_prop.zp_copies;
2119 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2120 ddt_entry_t *dde = zio->io_private;
2121 ddt_phys_t *ddp = &dde->dde_phys[p];
2125 ASSERT(ddp->ddp_refcnt == 0);
2126 ASSERT(dde->dde_lead_zio[p] == zio);
2127 dde->dde_lead_zio[p] = NULL;
2129 if (zio->io_error == 0) {
2130 while (zio_walk_parents(zio) != NULL)
2131 ddt_phys_addref(ddp);
2133 ddt_phys_clear(ddp);
2140 zio_ddt_ditto_write_done(zio_t *zio)
2142 int p = DDT_PHYS_DITTO;
2143 zio_prop_t *zp = &zio->io_prop;
2144 blkptr_t *bp = zio->io_bp;
2145 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2146 ddt_entry_t *dde = zio->io_private;
2147 ddt_phys_t *ddp = &dde->dde_phys[p];
2148 ddt_key_t *ddk = &dde->dde_key;
2152 ASSERT(ddp->ddp_refcnt == 0);
2153 ASSERT(dde->dde_lead_zio[p] == zio);
2154 dde->dde_lead_zio[p] = NULL;
2156 if (zio->io_error == 0) {
2157 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2158 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2159 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2160 if (ddp->ddp_phys_birth != 0)
2161 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2162 ddt_phys_fill(ddp, bp);
2169 zio_ddt_write(zio_t *zio)
2171 spa_t *spa = zio->io_spa;
2172 blkptr_t *bp = zio->io_bp;
2173 uint64_t txg = zio->io_txg;
2174 zio_prop_t *zp = &zio->io_prop;
2175 int p = zp->zp_copies;
2179 ddt_t *ddt = ddt_select(spa, bp);
2183 ASSERT(BP_GET_DEDUP(bp));
2184 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2185 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2188 dde = ddt_lookup(ddt, bp, B_TRUE);
2189 ddp = &dde->dde_phys[p];
2191 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2193 * If we're using a weak checksum, upgrade to a strong checksum
2194 * and try again. If we're already using a strong checksum,
2195 * we can't resolve it, so just convert to an ordinary write.
2196 * (And automatically e-mail a paper to Nature?)
2198 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2199 zp->zp_checksum = spa_dedup_checksum(spa);
2200 zio_pop_transforms(zio);
2201 zio->io_stage = ZIO_STAGE_OPEN;
2204 zp->zp_dedup = B_FALSE;
2206 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2208 return (ZIO_PIPELINE_CONTINUE);
2211 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2212 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2214 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2215 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2216 zio_prop_t czp = *zp;
2218 czp.zp_copies = ditto_copies;
2221 * If we arrived here with an override bp, we won't have run
2222 * the transform stack, so we won't have the data we need to
2223 * generate a child i/o. So, toss the override bp and restart.
2224 * This is safe, because using the override bp is just an
2225 * optimization; and it's rare, so the cost doesn't matter.
2227 if (zio->io_bp_override) {
2228 zio_pop_transforms(zio);
2229 zio->io_stage = ZIO_STAGE_OPEN;
2230 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2231 zio->io_bp_override = NULL;
2234 return (ZIO_PIPELINE_CONTINUE);
2237 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2238 zio->io_orig_size, &czp, NULL,
2239 zio_ddt_ditto_write_done, dde, zio->io_priority,
2240 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2242 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2243 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2246 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2247 if (ddp->ddp_phys_birth != 0)
2248 ddt_bp_fill(ddp, bp, txg);
2249 if (dde->dde_lead_zio[p] != NULL)
2250 zio_add_child(zio, dde->dde_lead_zio[p]);
2252 ddt_phys_addref(ddp);
2253 } else if (zio->io_bp_override) {
2254 ASSERT(bp->blk_birth == txg);
2255 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2256 ddt_phys_fill(ddp, bp);
2257 ddt_phys_addref(ddp);
2259 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2260 zio->io_orig_size, zp, zio_ddt_child_write_ready,
2261 zio_ddt_child_write_done, dde, zio->io_priority,
2262 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2264 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2265 dde->dde_lead_zio[p] = cio;
2275 return (ZIO_PIPELINE_CONTINUE);
2278 ddt_entry_t *freedde; /* for debugging */
2281 zio_ddt_free(zio_t *zio)
2283 spa_t *spa = zio->io_spa;
2284 blkptr_t *bp = zio->io_bp;
2285 ddt_t *ddt = ddt_select(spa, bp);
2289 ASSERT(BP_GET_DEDUP(bp));
2290 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2293 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2294 ddp = ddt_phys_select(dde, bp);
2295 ddt_phys_decref(ddp);
2298 return (ZIO_PIPELINE_CONTINUE);
2302 * ==========================================================================
2303 * Allocate and free blocks
2304 * ==========================================================================
2307 zio_dva_allocate(zio_t *zio)
2309 spa_t *spa = zio->io_spa;
2310 metaslab_class_t *mc = spa_normal_class(spa);
2311 blkptr_t *bp = zio->io_bp;
2315 if (zio->io_gang_leader == NULL) {
2316 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2317 zio->io_gang_leader = zio;
2320 ASSERT(BP_IS_HOLE(bp));
2321 ASSERT0(BP_GET_NDVAS(bp));
2322 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2323 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2324 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2327 * The dump device does not support gang blocks so allocation on
2328 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2329 * the "fast" gang feature.
2331 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2332 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2333 METASLAB_GANG_CHILD : 0;
2334 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2335 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2338 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2339 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2341 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2342 return (zio_write_gang_block(zio));
2343 zio->io_error = error;
2346 return (ZIO_PIPELINE_CONTINUE);
2350 zio_dva_free(zio_t *zio)
2352 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2354 return (ZIO_PIPELINE_CONTINUE);
2358 zio_dva_claim(zio_t *zio)
2362 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2364 zio->io_error = error;
2366 return (ZIO_PIPELINE_CONTINUE);
2370 * Undo an allocation. This is used by zio_done() when an I/O fails
2371 * and we want to give back the block we just allocated.
2372 * This handles both normal blocks and gang blocks.
2375 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2377 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2378 ASSERT(zio->io_bp_override == NULL);
2380 if (!BP_IS_HOLE(bp))
2381 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2384 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2385 zio_dva_unallocate(zio, gn->gn_child[g],
2386 &gn->gn_gbh->zg_blkptr[g]);
2392 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2395 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2396 uint64_t size, boolean_t use_slog)
2400 ASSERT(txg > spa_syncing_txg(spa));
2403 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2404 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2405 * when allocating them.
2408 error = metaslab_alloc(spa, spa_log_class(spa), size,
2409 new_bp, 1, txg, old_bp,
2410 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2414 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2415 new_bp, 1, txg, old_bp,
2416 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2420 BP_SET_LSIZE(new_bp, size);
2421 BP_SET_PSIZE(new_bp, size);
2422 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2423 BP_SET_CHECKSUM(new_bp,
2424 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2425 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2426 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2427 BP_SET_LEVEL(new_bp, 0);
2428 BP_SET_DEDUP(new_bp, 0);
2429 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2436 * Free an intent log block.
2439 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2441 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2442 ASSERT(!BP_IS_GANG(bp));
2444 zio_free(spa, txg, bp);
2448 * ==========================================================================
2449 * Read, write and delete to physical devices
2450 * ==========================================================================
2453 zio_vdev_io_start(zio_t *zio)
2455 vdev_t *vd = zio->io_vd;
2457 spa_t *spa = zio->io_spa;
2459 ASSERT(zio->io_error == 0);
2460 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2463 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2464 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2467 * The mirror_ops handle multiple DVAs in a single BP.
2469 return (vdev_mirror_ops.vdev_op_io_start(zio));
2472 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE) {
2473 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg);
2474 return (ZIO_PIPELINE_CONTINUE);
2478 * We keep track of time-sensitive I/Os so that the scan thread
2479 * can quickly react to certain workloads. In particular, we care
2480 * about non-scrubbing, top-level reads and writes with the following
2482 * - synchronous writes of user data to non-slog devices
2483 * - any reads of user data
2484 * When these conditions are met, adjust the timestamp of spa_last_io
2485 * which allows the scan thread to adjust its workload accordingly.
2487 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2488 vd == vd->vdev_top && !vd->vdev_islog &&
2489 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2490 zio->io_txg != spa_syncing_txg(spa)) {
2491 uint64_t old = spa->spa_last_io;
2492 uint64_t new = ddi_get_lbolt64();
2494 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2497 align = 1ULL << vd->vdev_top->vdev_ashift;
2499 if (P2PHASE(zio->io_size, align) != 0) {
2500 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2502 if (zio->io_type == ZIO_TYPE_READ ||
2503 zio->io_type == ZIO_TYPE_WRITE)
2504 abuf = zio_buf_alloc(asize);
2505 ASSERT(vd == vd->vdev_top);
2506 if (zio->io_type == ZIO_TYPE_WRITE) {
2507 bcopy(zio->io_data, abuf, zio->io_size);
2508 bzero(abuf + zio->io_size, asize - zio->io_size);
2510 zio_push_transform(zio, abuf, asize, abuf ? asize : 0,
2514 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2515 ASSERT(P2PHASE(zio->io_size, align) == 0);
2516 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa));
2519 * If this is a repair I/O, and there's no self-healing involved --
2520 * that is, we're just resilvering what we expect to resilver --
2521 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2522 * This prevents spurious resilvering with nested replication.
2523 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2524 * A is out of date, we'll read from C+D, then use the data to
2525 * resilver A+B -- but we don't actually want to resilver B, just A.
2526 * The top-level mirror has no way to know this, so instead we just
2527 * discard unnecessary repairs as we work our way down the vdev tree.
2528 * The same logic applies to any form of nested replication:
2529 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2531 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2532 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2533 zio->io_txg != 0 && /* not a delegated i/o */
2534 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2535 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2536 zio_vdev_io_bypass(zio);
2537 return (ZIO_PIPELINE_CONTINUE);
2540 if (vd->vdev_ops->vdev_op_leaf &&
2541 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2543 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2544 return (ZIO_PIPELINE_CONTINUE);
2546 if ((zio = vdev_queue_io(zio)) == NULL)
2547 return (ZIO_PIPELINE_STOP);
2549 if (!vdev_accessible(vd, zio)) {
2550 zio->io_error = SET_ERROR(ENXIO);
2552 return (ZIO_PIPELINE_STOP);
2556 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_WRITE) {
2557 if (!trim_map_write_start(zio))
2558 return (ZIO_PIPELINE_STOP);
2561 return (vd->vdev_ops->vdev_op_io_start(zio));
2565 zio_vdev_io_done(zio_t *zio)
2567 vdev_t *vd = zio->io_vd;
2568 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2569 boolean_t unexpected_error = B_FALSE;
2571 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2572 return (ZIO_PIPELINE_STOP);
2574 ASSERT(zio->io_type == ZIO_TYPE_READ ||
2575 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE);
2577 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2578 zio->io_type == ZIO_TYPE_WRITE) {
2579 trim_map_write_done(zio);
2582 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2583 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2585 vdev_queue_io_done(zio);
2587 if (zio->io_type == ZIO_TYPE_WRITE)
2588 vdev_cache_write(zio);
2590 if (zio_injection_enabled && zio->io_error == 0)
2591 zio->io_error = zio_handle_device_injection(vd,
2594 if (zio_injection_enabled && zio->io_error == 0)
2595 zio->io_error = zio_handle_label_injection(zio, EIO);
2597 if (zio->io_error) {
2598 if (!vdev_accessible(vd, zio)) {
2599 zio->io_error = SET_ERROR(ENXIO);
2601 unexpected_error = B_TRUE;
2606 ops->vdev_op_io_done(zio);
2608 if (unexpected_error)
2609 VERIFY(vdev_probe(vd, zio) == NULL);
2611 return (ZIO_PIPELINE_CONTINUE);
2615 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2616 * disk, and use that to finish the checksum ereport later.
2619 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2620 const void *good_buf)
2622 /* no processing needed */
2623 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2628 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2630 void *buf = zio_buf_alloc(zio->io_size);
2632 bcopy(zio->io_data, buf, zio->io_size);
2634 zcr->zcr_cbinfo = zio->io_size;
2635 zcr->zcr_cbdata = buf;
2636 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2637 zcr->zcr_free = zio_buf_free;
2641 zio_vdev_io_assess(zio_t *zio)
2643 vdev_t *vd = zio->io_vd;
2645 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2646 return (ZIO_PIPELINE_STOP);
2648 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2649 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2651 if (zio->io_vsd != NULL) {
2652 zio->io_vsd_ops->vsd_free(zio);
2656 if (zio_injection_enabled && zio->io_error == 0)
2657 zio->io_error = zio_handle_fault_injection(zio, EIO);
2659 if (zio->io_type == ZIO_TYPE_IOCTL && zio->io_cmd == DKIOCTRIM)
2660 switch (zio->io_error) {
2662 ZIO_TRIM_STAT_INCR(bytes, zio->io_size);
2663 ZIO_TRIM_STAT_BUMP(success);
2666 ZIO_TRIM_STAT_BUMP(unsupported);
2669 ZIO_TRIM_STAT_BUMP(failed);
2674 * If the I/O failed, determine whether we should attempt to retry it.
2676 * On retry, we cut in line in the issue queue, since we don't want
2677 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2679 if (zio->io_error && vd == NULL &&
2680 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2681 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2682 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2684 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2685 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2686 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2687 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2688 zio_requeue_io_start_cut_in_line);
2689 return (ZIO_PIPELINE_STOP);
2693 * If we got an error on a leaf device, convert it to ENXIO
2694 * if the device is not accessible at all.
2696 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2697 !vdev_accessible(vd, zio))
2698 zio->io_error = SET_ERROR(ENXIO);
2701 * If we can't write to an interior vdev (mirror or RAID-Z),
2702 * set vdev_cant_write so that we stop trying to allocate from it.
2704 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2705 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
2706 vd->vdev_cant_write = B_TRUE;
2710 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2712 return (ZIO_PIPELINE_CONTINUE);
2716 zio_vdev_io_reissue(zio_t *zio)
2718 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2719 ASSERT(zio->io_error == 0);
2721 zio->io_stage >>= 1;
2725 zio_vdev_io_redone(zio_t *zio)
2727 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2729 zio->io_stage >>= 1;
2733 zio_vdev_io_bypass(zio_t *zio)
2735 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2736 ASSERT(zio->io_error == 0);
2738 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2739 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2743 * ==========================================================================
2744 * Generate and verify checksums
2745 * ==========================================================================
2748 zio_checksum_generate(zio_t *zio)
2750 blkptr_t *bp = zio->io_bp;
2751 enum zio_checksum checksum;
2755 * This is zio_write_phys().
2756 * We're either generating a label checksum, or none at all.
2758 checksum = zio->io_prop.zp_checksum;
2760 if (checksum == ZIO_CHECKSUM_OFF)
2761 return (ZIO_PIPELINE_CONTINUE);
2763 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2765 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2766 ASSERT(!IO_IS_ALLOCATING(zio));
2767 checksum = ZIO_CHECKSUM_GANG_HEADER;
2769 checksum = BP_GET_CHECKSUM(bp);
2773 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2775 return (ZIO_PIPELINE_CONTINUE);
2779 zio_checksum_verify(zio_t *zio)
2781 zio_bad_cksum_t info;
2782 blkptr_t *bp = zio->io_bp;
2785 ASSERT(zio->io_vd != NULL);
2789 * This is zio_read_phys().
2790 * We're either verifying a label checksum, or nothing at all.
2792 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2793 return (ZIO_PIPELINE_CONTINUE);
2795 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2798 if ((error = zio_checksum_error(zio, &info)) != 0) {
2799 zio->io_error = error;
2800 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2801 zfs_ereport_start_checksum(zio->io_spa,
2802 zio->io_vd, zio, zio->io_offset,
2803 zio->io_size, NULL, &info);
2807 return (ZIO_PIPELINE_CONTINUE);
2811 * Called by RAID-Z to ensure we don't compute the checksum twice.
2814 zio_checksum_verified(zio_t *zio)
2816 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2820 * ==========================================================================
2821 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2822 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2823 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2824 * indicate errors that are specific to one I/O, and most likely permanent.
2825 * Any other error is presumed to be worse because we weren't expecting it.
2826 * ==========================================================================
2829 zio_worst_error(int e1, int e2)
2831 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2834 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2835 if (e1 == zio_error_rank[r1])
2838 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2839 if (e2 == zio_error_rank[r2])
2842 return (r1 > r2 ? e1 : e2);
2846 * ==========================================================================
2848 * ==========================================================================
2851 zio_ready(zio_t *zio)
2853 blkptr_t *bp = zio->io_bp;
2854 zio_t *pio, *pio_next;
2856 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2857 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2858 return (ZIO_PIPELINE_STOP);
2860 if (zio->io_ready) {
2861 ASSERT(IO_IS_ALLOCATING(zio));
2862 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
2863 (zio->io_flags & ZIO_FLAG_NOPWRITE));
2864 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2869 if (bp != NULL && bp != &zio->io_bp_copy)
2870 zio->io_bp_copy = *bp;
2873 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2875 mutex_enter(&zio->io_lock);
2876 zio->io_state[ZIO_WAIT_READY] = 1;
2877 pio = zio_walk_parents(zio);
2878 mutex_exit(&zio->io_lock);
2881 * As we notify zio's parents, new parents could be added.
2882 * New parents go to the head of zio's io_parent_list, however,
2883 * so we will (correctly) not notify them. The remainder of zio's
2884 * io_parent_list, from 'pio_next' onward, cannot change because
2885 * all parents must wait for us to be done before they can be done.
2887 for (; pio != NULL; pio = pio_next) {
2888 pio_next = zio_walk_parents(zio);
2889 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2892 if (zio->io_flags & ZIO_FLAG_NODATA) {
2893 if (BP_IS_GANG(bp)) {
2894 zio->io_flags &= ~ZIO_FLAG_NODATA;
2896 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2897 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2901 if (zio_injection_enabled &&
2902 zio->io_spa->spa_syncing_txg == zio->io_txg)
2903 zio_handle_ignored_writes(zio);
2905 return (ZIO_PIPELINE_CONTINUE);
2909 zio_done(zio_t *zio)
2911 spa_t *spa = zio->io_spa;
2912 zio_t *lio = zio->io_logical;
2913 blkptr_t *bp = zio->io_bp;
2914 vdev_t *vd = zio->io_vd;
2915 uint64_t psize = zio->io_size;
2916 zio_t *pio, *pio_next;
2919 * If our children haven't all completed,
2920 * wait for them and then repeat this pipeline stage.
2922 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2923 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2924 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2925 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2926 return (ZIO_PIPELINE_STOP);
2928 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2929 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2930 ASSERT(zio->io_children[c][w] == 0);
2933 ASSERT(bp->blk_pad[0] == 0);
2934 ASSERT(bp->blk_pad[1] == 0);
2935 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2936 (bp == zio_unique_parent(zio)->io_bp));
2937 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
2938 zio->io_bp_override == NULL &&
2939 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2940 ASSERT(!BP_SHOULD_BYTESWAP(bp));
2941 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
2942 ASSERT(BP_COUNT_GANG(bp) == 0 ||
2943 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
2945 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
2946 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
2950 * If there were child vdev/gang/ddt errors, they apply to us now.
2952 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2953 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2954 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2957 * If the I/O on the transformed data was successful, generate any
2958 * checksum reports now while we still have the transformed data.
2960 if (zio->io_error == 0) {
2961 while (zio->io_cksum_report != NULL) {
2962 zio_cksum_report_t *zcr = zio->io_cksum_report;
2963 uint64_t align = zcr->zcr_align;
2964 uint64_t asize = P2ROUNDUP(psize, align);
2965 char *abuf = zio->io_data;
2967 if (asize != psize) {
2968 abuf = zio_buf_alloc(asize);
2969 bcopy(zio->io_data, abuf, psize);
2970 bzero(abuf + psize, asize - psize);
2973 zio->io_cksum_report = zcr->zcr_next;
2974 zcr->zcr_next = NULL;
2975 zcr->zcr_finish(zcr, abuf);
2976 zfs_ereport_free_checksum(zcr);
2979 zio_buf_free(abuf, asize);
2983 zio_pop_transforms(zio); /* note: may set zio->io_error */
2985 vdev_stat_update(zio, psize);
2987 if (zio->io_error) {
2989 * If this I/O is attached to a particular vdev,
2990 * generate an error message describing the I/O failure
2991 * at the block level. We ignore these errors if the
2992 * device is currently unavailable.
2994 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
2995 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
2997 if ((zio->io_error == EIO || !(zio->io_flags &
2998 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3001 * For logical I/O requests, tell the SPA to log the
3002 * error and generate a logical data ereport.
3004 spa_log_error(spa, zio);
3005 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3010 if (zio->io_error && zio == lio) {
3012 * Determine whether zio should be reexecuted. This will
3013 * propagate all the way to the root via zio_notify_parent().
3015 ASSERT(vd == NULL && bp != NULL);
3016 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3018 if (IO_IS_ALLOCATING(zio) &&
3019 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3020 if (zio->io_error != ENOSPC)
3021 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3023 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3026 if ((zio->io_type == ZIO_TYPE_READ ||
3027 zio->io_type == ZIO_TYPE_FREE) &&
3028 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3029 zio->io_error == ENXIO &&
3030 spa_load_state(spa) == SPA_LOAD_NONE &&
3031 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3032 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3034 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3035 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3038 * Here is a possibly good place to attempt to do
3039 * either combinatorial reconstruction or error correction
3040 * based on checksums. It also might be a good place
3041 * to send out preliminary ereports before we suspend
3047 * If there were logical child errors, they apply to us now.
3048 * We defer this until now to avoid conflating logical child
3049 * errors with errors that happened to the zio itself when
3050 * updating vdev stats and reporting FMA events above.
3052 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3054 if ((zio->io_error || zio->io_reexecute) &&
3055 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3056 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3057 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3059 zio_gang_tree_free(&zio->io_gang_tree);
3062 * Godfather I/Os should never suspend.
3064 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3065 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3066 zio->io_reexecute = 0;
3068 if (zio->io_reexecute) {
3070 * This is a logical I/O that wants to reexecute.
3072 * Reexecute is top-down. When an i/o fails, if it's not
3073 * the root, it simply notifies its parent and sticks around.
3074 * The parent, seeing that it still has children in zio_done(),
3075 * does the same. This percolates all the way up to the root.
3076 * The root i/o will reexecute or suspend the entire tree.
3078 * This approach ensures that zio_reexecute() honors
3079 * all the original i/o dependency relationships, e.g.
3080 * parents not executing until children are ready.
3082 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3084 zio->io_gang_leader = NULL;
3086 mutex_enter(&zio->io_lock);
3087 zio->io_state[ZIO_WAIT_DONE] = 1;
3088 mutex_exit(&zio->io_lock);
3091 * "The Godfather" I/O monitors its children but is
3092 * not a true parent to them. It will track them through
3093 * the pipeline but severs its ties whenever they get into
3094 * trouble (e.g. suspended). This allows "The Godfather"
3095 * I/O to return status without blocking.
3097 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3098 zio_link_t *zl = zio->io_walk_link;
3099 pio_next = zio_walk_parents(zio);
3101 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3102 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3103 zio_remove_child(pio, zio, zl);
3104 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3108 if ((pio = zio_unique_parent(zio)) != NULL) {
3110 * We're not a root i/o, so there's nothing to do
3111 * but notify our parent. Don't propagate errors
3112 * upward since we haven't permanently failed yet.
3114 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3115 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3116 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3117 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3119 * We'd fail again if we reexecuted now, so suspend
3120 * until conditions improve (e.g. device comes online).
3122 zio_suspend(spa, zio);
3125 * Reexecution is potentially a huge amount of work.
3126 * Hand it off to the otherwise-unused claim taskq.
3129 (void) taskq_dispatch_safe(
3130 spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
3131 (task_func_t *)zio_reexecute, zio, TQ_SLEEP,
3134 (void) taskq_dispatch(
3135 spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
3136 (task_func_t *)zio_reexecute, zio, TQ_SLEEP);
3139 return (ZIO_PIPELINE_STOP);
3142 ASSERT(zio->io_child_count == 0);
3143 ASSERT(zio->io_reexecute == 0);
3144 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3147 * Report any checksum errors, since the I/O is complete.
3149 while (zio->io_cksum_report != NULL) {
3150 zio_cksum_report_t *zcr = zio->io_cksum_report;
3151 zio->io_cksum_report = zcr->zcr_next;
3152 zcr->zcr_next = NULL;
3153 zcr->zcr_finish(zcr, NULL);
3154 zfs_ereport_free_checksum(zcr);
3158 * It is the responsibility of the done callback to ensure that this
3159 * particular zio is no longer discoverable for adoption, and as
3160 * such, cannot acquire any new parents.
3165 mutex_enter(&zio->io_lock);
3166 zio->io_state[ZIO_WAIT_DONE] = 1;
3167 mutex_exit(&zio->io_lock);
3169 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3170 zio_link_t *zl = zio->io_walk_link;
3171 pio_next = zio_walk_parents(zio);
3172 zio_remove_child(pio, zio, zl);
3173 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3176 if (zio->io_waiter != NULL) {
3177 mutex_enter(&zio->io_lock);
3178 zio->io_executor = NULL;
3179 cv_broadcast(&zio->io_cv);
3180 mutex_exit(&zio->io_lock);
3185 return (ZIO_PIPELINE_STOP);
3189 * ==========================================================================
3190 * I/O pipeline definition
3191 * ==========================================================================
3193 static zio_pipe_stage_t *zio_pipeline[] = {
3199 zio_checksum_generate,
3214 zio_checksum_verify,
3218 /* dnp is the dnode for zb1->zb_object */
3220 zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_t *zb1,
3221 const zbookmark_t *zb2)
3223 uint64_t zb1nextL0, zb2thisobj;
3225 ASSERT(zb1->zb_objset == zb2->zb_objset);
3226 ASSERT(zb2->zb_level == 0);
3229 * A bookmark in the deadlist is considered to be after
3232 if (zb2->zb_object == DMU_DEADLIST_OBJECT)
3235 /* The objset_phys_t isn't before anything. */
3239 zb1nextL0 = (zb1->zb_blkid + 1) <<
3240 ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
3242 zb2thisobj = zb2->zb_object ? zb2->zb_object :
3243 zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
3245 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3246 uint64_t nextobj = zb1nextL0 *
3247 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
3248 return (nextobj <= zb2thisobj);
3251 if (zb1->zb_object < zb2thisobj)
3253 if (zb1->zb_object > zb2thisobj)
3255 if (zb2->zb_object == DMU_META_DNODE_OBJECT)
3257 return (zb1nextL0 <= zb2->zb_blkid);