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);
777 zio = zio_create(pio, spa, txg, bp, NULL, size,
778 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
779 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
785 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
786 zio_done_func_t *done, void *private, enum zio_flag flags)
791 * A claim is an allocation of a specific block. Claims are needed
792 * to support immediate writes in the intent log. The issue is that
793 * immediate writes contain committed data, but in a txg that was
794 * *not* committed. Upon opening the pool after an unclean shutdown,
795 * the intent log claims all blocks that contain immediate write data
796 * so that the SPA knows they're in use.
798 * All claims *must* be resolved in the first txg -- before the SPA
799 * starts allocating blocks -- so that nothing is allocated twice.
800 * If txg == 0 we just verify that the block is claimable.
802 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
803 ASSERT(txg == spa_first_txg(spa) || txg == 0);
804 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
806 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
807 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
808 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
814 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset,
815 uint64_t size, zio_done_func_t *done, void *private, int priority,
821 if (vd->vdev_children == 0) {
822 zio = zio_create(pio, spa, 0, NULL, NULL, size, done, private,
823 ZIO_TYPE_IOCTL, priority, flags, vd, offset, NULL,
824 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
828 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
830 for (c = 0; c < vd->vdev_children; c++)
831 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
832 offset, size, done, private, priority, flags));
839 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
840 void *data, int checksum, zio_done_func_t *done, void *private,
841 int priority, enum zio_flag flags, boolean_t labels)
845 ASSERT(vd->vdev_children == 0);
846 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
847 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
848 ASSERT3U(offset + size, <=, vd->vdev_psize);
850 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
851 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
852 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
854 zio->io_prop.zp_checksum = checksum;
860 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
861 void *data, int checksum, zio_done_func_t *done, void *private,
862 int priority, enum zio_flag flags, boolean_t labels)
866 ASSERT(vd->vdev_children == 0);
867 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
868 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
869 ASSERT3U(offset + size, <=, vd->vdev_psize);
871 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
872 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
873 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
875 zio->io_prop.zp_checksum = checksum;
877 if (zio_checksum_table[checksum].ci_eck) {
879 * zec checksums are necessarily destructive -- they modify
880 * the end of the write buffer to hold the verifier/checksum.
881 * Therefore, we must make a local copy in case the data is
882 * being written to multiple places in parallel.
884 void *wbuf = zio_buf_alloc(size);
885 bcopy(data, wbuf, size);
886 zio_push_transform(zio, wbuf, size, size, NULL);
893 * Create a child I/O to do some work for us.
896 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
897 void *data, uint64_t size, int type, int priority, enum zio_flag flags,
898 zio_done_func_t *done, void *private)
900 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
903 ASSERT(vd->vdev_parent ==
904 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
906 if (type == ZIO_TYPE_READ && bp != NULL) {
908 * If we have the bp, then the child should perform the
909 * checksum and the parent need not. This pushes error
910 * detection as close to the leaves as possible and
911 * eliminates redundant checksums in the interior nodes.
913 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
914 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
917 if (vd->vdev_children == 0)
918 offset += VDEV_LABEL_START_SIZE;
920 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
923 * If we've decided to do a repair, the write is not speculative --
924 * even if the original read was.
926 if (flags & ZIO_FLAG_IO_REPAIR)
927 flags &= ~ZIO_FLAG_SPECULATIVE;
929 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
930 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
931 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
937 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
938 int type, int priority, enum zio_flag flags,
939 zio_done_func_t *done, void *private)
943 ASSERT(vd->vdev_ops->vdev_op_leaf);
945 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
946 data, size, done, private, type, priority,
947 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
949 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
955 zio_flush(zio_t *zio, vdev_t *vd)
957 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0,
958 NULL, NULL, ZIO_PRIORITY_NOW,
959 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
963 zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size)
966 ASSERT(vd->vdev_ops->vdev_op_leaf);
968 return zio_ioctl(zio, spa, vd, DKIOCTRIM, offset, size,
969 NULL, NULL, ZIO_PRIORITY_TRIM,
970 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY);
974 zio_shrink(zio_t *zio, uint64_t size)
976 ASSERT(zio->io_executor == NULL);
977 ASSERT(zio->io_orig_size == zio->io_size);
978 ASSERT(size <= zio->io_size);
981 * We don't shrink for raidz because of problems with the
982 * reconstruction when reading back less than the block size.
983 * Note, BP_IS_RAIDZ() assumes no compression.
985 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
986 if (!BP_IS_RAIDZ(zio->io_bp))
987 zio->io_orig_size = zio->io_size = size;
991 * ==========================================================================
992 * Prepare to read and write logical blocks
993 * ==========================================================================
997 zio_read_bp_init(zio_t *zio)
999 blkptr_t *bp = zio->io_bp;
1001 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1002 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1003 !(zio->io_flags & ZIO_FLAG_RAW)) {
1004 uint64_t psize = BP_GET_PSIZE(bp);
1005 void *cbuf = zio_buf_alloc(psize);
1007 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
1010 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1011 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1013 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1014 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1016 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1017 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1019 return (ZIO_PIPELINE_CONTINUE);
1023 zio_write_bp_init(zio_t *zio)
1025 spa_t *spa = zio->io_spa;
1026 zio_prop_t *zp = &zio->io_prop;
1027 enum zio_compress compress = zp->zp_compress;
1028 blkptr_t *bp = zio->io_bp;
1029 uint64_t lsize = zio->io_size;
1030 uint64_t psize = lsize;
1034 * If our children haven't all reached the ready stage,
1035 * wait for them and then repeat this pipeline stage.
1037 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1038 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1039 return (ZIO_PIPELINE_STOP);
1041 if (!IO_IS_ALLOCATING(zio))
1042 return (ZIO_PIPELINE_CONTINUE);
1044 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1046 if (zio->io_bp_override) {
1047 ASSERT(bp->blk_birth != zio->io_txg);
1048 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1050 *bp = *zio->io_bp_override;
1051 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1054 * If we've been overridden and nopwrite is set then
1055 * set the flag accordingly to indicate that a nopwrite
1056 * has already occurred.
1058 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1059 ASSERT(!zp->zp_dedup);
1060 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1061 return (ZIO_PIPELINE_CONTINUE);
1064 ASSERT(!zp->zp_nopwrite);
1066 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1067 return (ZIO_PIPELINE_CONTINUE);
1069 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1070 zp->zp_dedup_verify);
1072 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1073 BP_SET_DEDUP(bp, 1);
1074 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1075 return (ZIO_PIPELINE_CONTINUE);
1077 zio->io_bp_override = NULL;
1081 if (bp->blk_birth == zio->io_txg) {
1083 * We're rewriting an existing block, which means we're
1084 * working on behalf of spa_sync(). For spa_sync() to
1085 * converge, it must eventually be the case that we don't
1086 * have to allocate new blocks. But compression changes
1087 * the blocksize, which forces a reallocate, and makes
1088 * convergence take longer. Therefore, after the first
1089 * few passes, stop compressing to ensure convergence.
1091 pass = spa_sync_pass(spa);
1093 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1094 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1095 ASSERT(!BP_GET_DEDUP(bp));
1097 if (pass >= zfs_sync_pass_dont_compress)
1098 compress = ZIO_COMPRESS_OFF;
1100 /* Make sure someone doesn't change their mind on overwrites */
1101 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
1102 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1105 if (compress != ZIO_COMPRESS_OFF) {
1106 void *cbuf = zio_buf_alloc(lsize);
1107 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1108 if (psize == 0 || psize == lsize) {
1109 compress = ZIO_COMPRESS_OFF;
1110 zio_buf_free(cbuf, lsize);
1112 ASSERT(psize < lsize);
1113 zio_push_transform(zio, cbuf, psize, lsize, NULL);
1118 * The final pass of spa_sync() must be all rewrites, but the first
1119 * few passes offer a trade-off: allocating blocks defers convergence,
1120 * but newly allocated blocks are sequential, so they can be written
1121 * to disk faster. Therefore, we allow the first few passes of
1122 * spa_sync() to allocate new blocks, but force rewrites after that.
1123 * There should only be a handful of blocks after pass 1 in any case.
1125 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
1126 pass >= zfs_sync_pass_rewrite) {
1128 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1129 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1130 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1133 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1137 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1139 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1140 BP_SET_LSIZE(bp, lsize);
1141 BP_SET_PSIZE(bp, psize);
1142 BP_SET_COMPRESS(bp, compress);
1143 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1144 BP_SET_TYPE(bp, zp->zp_type);
1145 BP_SET_LEVEL(bp, zp->zp_level);
1146 BP_SET_DEDUP(bp, zp->zp_dedup);
1147 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1149 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1150 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1151 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1153 if (zp->zp_nopwrite) {
1154 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1155 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1156 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1160 return (ZIO_PIPELINE_CONTINUE);
1164 zio_free_bp_init(zio_t *zio)
1166 blkptr_t *bp = zio->io_bp;
1168 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1169 if (BP_GET_DEDUP(bp))
1170 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1173 return (ZIO_PIPELINE_CONTINUE);
1177 * ==========================================================================
1178 * Execute the I/O pipeline
1179 * ==========================================================================
1183 zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q, boolean_t cutinline)
1185 spa_t *spa = zio->io_spa;
1186 zio_type_t t = zio->io_type;
1187 int flags = TQ_SLEEP | (cutinline ? TQ_FRONT : 0);
1189 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1192 * If we're a config writer or a probe, the normal issue and
1193 * interrupt threads may all be blocked waiting for the config lock.
1194 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1196 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1200 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1202 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1206 * If this is a high priority I/O, then use the high priority taskq.
1208 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1209 spa->spa_zio_taskq[t][q + 1] != NULL)
1212 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1214 (void) taskq_dispatch_safe(spa->spa_zio_taskq[t][q],
1215 (task_func_t *)zio_execute, zio, flags, &zio->io_task);
1217 (void) taskq_dispatch(spa->spa_zio_taskq[t][q],
1218 (task_func_t *)zio_execute, zio, flags);
1223 zio_taskq_member(zio_t *zio, enum zio_taskq_type q)
1225 kthread_t *executor = zio->io_executor;
1226 spa_t *spa = zio->io_spa;
1228 for (zio_type_t t = 0; t < ZIO_TYPES; t++)
1229 if (taskq_member(spa->spa_zio_taskq[t][q], executor))
1236 zio_issue_async(zio_t *zio)
1238 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1240 return (ZIO_PIPELINE_STOP);
1244 zio_interrupt(zio_t *zio)
1246 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1250 * Execute the I/O pipeline until one of the following occurs:
1252 * (1) the I/O completes
1253 * (2) the pipeline stalls waiting for dependent child I/Os
1254 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1255 * (4) the I/O is delegated by vdev-level caching or aggregation
1256 * (5) the I/O is deferred due to vdev-level queueing
1257 * (6) the I/O is handed off to another thread.
1259 * In all cases, the pipeline stops whenever there's no CPU work; it never
1260 * burns a thread in cv_wait().
1262 * There's no locking on io_stage because there's no legitimate way
1263 * for multiple threads to be attempting to process the same I/O.
1265 static zio_pipe_stage_t *zio_pipeline[];
1268 zio_execute(zio_t *zio)
1270 zio->io_executor = curthread;
1272 while (zio->io_stage < ZIO_STAGE_DONE) {
1273 enum zio_stage pipeline = zio->io_pipeline;
1274 enum zio_stage stage = zio->io_stage;
1277 ASSERT(!MUTEX_HELD(&zio->io_lock));
1278 ASSERT(ISP2(stage));
1279 ASSERT(zio->io_stall == NULL);
1283 } while ((stage & pipeline) == 0);
1285 ASSERT(stage <= ZIO_STAGE_DONE);
1288 * If we are in interrupt context and this pipeline stage
1289 * will grab a config lock that is held across I/O,
1290 * or may wait for an I/O that needs an interrupt thread
1291 * to complete, issue async to avoid deadlock.
1293 * For VDEV_IO_START, we cut in line so that the io will
1294 * be sent to disk promptly.
1296 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1297 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1298 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1299 zio_requeue_io_start_cut_in_line : B_FALSE;
1300 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1304 zio->io_stage = stage;
1305 rv = zio_pipeline[highbit(stage) - 1](zio);
1307 if (rv == ZIO_PIPELINE_STOP)
1310 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1315 * ==========================================================================
1316 * Initiate I/O, either sync or async
1317 * ==========================================================================
1320 zio_wait(zio_t *zio)
1324 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1325 ASSERT(zio->io_executor == NULL);
1327 zio->io_waiter = curthread;
1331 mutex_enter(&zio->io_lock);
1332 while (zio->io_executor != NULL)
1333 cv_wait(&zio->io_cv, &zio->io_lock);
1334 mutex_exit(&zio->io_lock);
1336 error = zio->io_error;
1343 zio_nowait(zio_t *zio)
1345 ASSERT(zio->io_executor == NULL);
1347 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1348 zio_unique_parent(zio) == NULL) {
1350 * This is a logical async I/O with no parent to wait for it.
1351 * We add it to the spa_async_root_zio "Godfather" I/O which
1352 * will ensure they complete prior to unloading the pool.
1354 spa_t *spa = zio->io_spa;
1356 zio_add_child(spa->spa_async_zio_root, zio);
1363 * ==========================================================================
1364 * Reexecute or suspend/resume failed I/O
1365 * ==========================================================================
1369 zio_reexecute(zio_t *pio)
1371 zio_t *cio, *cio_next;
1373 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1374 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1375 ASSERT(pio->io_gang_leader == NULL);
1376 ASSERT(pio->io_gang_tree == NULL);
1378 pio->io_flags = pio->io_orig_flags;
1379 pio->io_stage = pio->io_orig_stage;
1380 pio->io_pipeline = pio->io_orig_pipeline;
1381 pio->io_reexecute = 0;
1382 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1384 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1385 pio->io_state[w] = 0;
1386 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1387 pio->io_child_error[c] = 0;
1389 if (IO_IS_ALLOCATING(pio))
1390 BP_ZERO(pio->io_bp);
1393 * As we reexecute pio's children, new children could be created.
1394 * New children go to the head of pio's io_child_list, however,
1395 * so we will (correctly) not reexecute them. The key is that
1396 * the remainder of pio's io_child_list, from 'cio_next' onward,
1397 * cannot be affected by any side effects of reexecuting 'cio'.
1399 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1400 cio_next = zio_walk_children(pio);
1401 mutex_enter(&pio->io_lock);
1402 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1403 pio->io_children[cio->io_child_type][w]++;
1404 mutex_exit(&pio->io_lock);
1409 * Now that all children have been reexecuted, execute the parent.
1410 * We don't reexecute "The Godfather" I/O here as it's the
1411 * responsibility of the caller to wait on him.
1413 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1418 zio_suspend(spa_t *spa, zio_t *zio)
1420 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1421 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1422 "failure and the failure mode property for this pool "
1423 "is set to panic.", spa_name(spa));
1425 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1427 mutex_enter(&spa->spa_suspend_lock);
1429 if (spa->spa_suspend_zio_root == NULL)
1430 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1431 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1432 ZIO_FLAG_GODFATHER);
1434 spa->spa_suspended = B_TRUE;
1437 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1438 ASSERT(zio != spa->spa_suspend_zio_root);
1439 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1440 ASSERT(zio_unique_parent(zio) == NULL);
1441 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1442 zio_add_child(spa->spa_suspend_zio_root, zio);
1445 mutex_exit(&spa->spa_suspend_lock);
1449 zio_resume(spa_t *spa)
1454 * Reexecute all previously suspended i/o.
1456 mutex_enter(&spa->spa_suspend_lock);
1457 spa->spa_suspended = B_FALSE;
1458 cv_broadcast(&spa->spa_suspend_cv);
1459 pio = spa->spa_suspend_zio_root;
1460 spa->spa_suspend_zio_root = NULL;
1461 mutex_exit(&spa->spa_suspend_lock);
1467 return (zio_wait(pio));
1471 zio_resume_wait(spa_t *spa)
1473 mutex_enter(&spa->spa_suspend_lock);
1474 while (spa_suspended(spa))
1475 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1476 mutex_exit(&spa->spa_suspend_lock);
1480 * ==========================================================================
1483 * A gang block is a collection of small blocks that looks to the DMU
1484 * like one large block. When zio_dva_allocate() cannot find a block
1485 * of the requested size, due to either severe fragmentation or the pool
1486 * being nearly full, it calls zio_write_gang_block() to construct the
1487 * block from smaller fragments.
1489 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1490 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1491 * an indirect block: it's an array of block pointers. It consumes
1492 * only one sector and hence is allocatable regardless of fragmentation.
1493 * The gang header's bps point to its gang members, which hold the data.
1495 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1496 * as the verifier to ensure uniqueness of the SHA256 checksum.
1497 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1498 * not the gang header. This ensures that data block signatures (needed for
1499 * deduplication) are independent of how the block is physically stored.
1501 * Gang blocks can be nested: a gang member may itself be a gang block.
1502 * Thus every gang block is a tree in which root and all interior nodes are
1503 * gang headers, and the leaves are normal blocks that contain user data.
1504 * The root of the gang tree is called the gang leader.
1506 * To perform any operation (read, rewrite, free, claim) on a gang block,
1507 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1508 * in the io_gang_tree field of the original logical i/o by recursively
1509 * reading the gang leader and all gang headers below it. This yields
1510 * an in-core tree containing the contents of every gang header and the
1511 * bps for every constituent of the gang block.
1513 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1514 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1515 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1516 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1517 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1518 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1519 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1520 * of the gang header plus zio_checksum_compute() of the data to update the
1521 * gang header's blk_cksum as described above.
1523 * The two-phase assemble/issue model solves the problem of partial failure --
1524 * what if you'd freed part of a gang block but then couldn't read the
1525 * gang header for another part? Assembling the entire gang tree first
1526 * ensures that all the necessary gang header I/O has succeeded before
1527 * starting the actual work of free, claim, or write. Once the gang tree
1528 * is assembled, free and claim are in-memory operations that cannot fail.
1530 * In the event that a gang write fails, zio_dva_unallocate() walks the
1531 * gang tree to immediately free (i.e. insert back into the space map)
1532 * everything we've allocated. This ensures that we don't get ENOSPC
1533 * errors during repeated suspend/resume cycles due to a flaky device.
1535 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1536 * the gang tree, we won't modify the block, so we can safely defer the free
1537 * (knowing that the block is still intact). If we *can* assemble the gang
1538 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1539 * each constituent bp and we can allocate a new block on the next sync pass.
1541 * In all cases, the gang tree allows complete recovery from partial failure.
1542 * ==========================================================================
1546 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1551 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1552 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1553 &pio->io_bookmark));
1557 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1562 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1563 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1564 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1566 * As we rewrite each gang header, the pipeline will compute
1567 * a new gang block header checksum for it; but no one will
1568 * compute a new data checksum, so we do that here. The one
1569 * exception is the gang leader: the pipeline already computed
1570 * its data checksum because that stage precedes gang assembly.
1571 * (Presently, nothing actually uses interior data checksums;
1572 * this is just good hygiene.)
1574 if (gn != pio->io_gang_leader->io_gang_tree) {
1575 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1576 data, BP_GET_PSIZE(bp));
1579 * If we are here to damage data for testing purposes,
1580 * leave the GBH alone so that we can detect the damage.
1582 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1583 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1585 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1586 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1587 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1595 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1597 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1598 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp),
1599 ZIO_GANG_CHILD_FLAGS(pio)));
1604 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1606 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1607 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1610 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1619 static void zio_gang_tree_assemble_done(zio_t *zio);
1621 static zio_gang_node_t *
1622 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1624 zio_gang_node_t *gn;
1626 ASSERT(*gnpp == NULL);
1628 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1629 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1636 zio_gang_node_free(zio_gang_node_t **gnpp)
1638 zio_gang_node_t *gn = *gnpp;
1640 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1641 ASSERT(gn->gn_child[g] == NULL);
1643 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1644 kmem_free(gn, sizeof (*gn));
1649 zio_gang_tree_free(zio_gang_node_t **gnpp)
1651 zio_gang_node_t *gn = *gnpp;
1656 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1657 zio_gang_tree_free(&gn->gn_child[g]);
1659 zio_gang_node_free(gnpp);
1663 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1665 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1667 ASSERT(gio->io_gang_leader == gio);
1668 ASSERT(BP_IS_GANG(bp));
1670 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1671 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1672 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1676 zio_gang_tree_assemble_done(zio_t *zio)
1678 zio_t *gio = zio->io_gang_leader;
1679 zio_gang_node_t *gn = zio->io_private;
1680 blkptr_t *bp = zio->io_bp;
1682 ASSERT(gio == zio_unique_parent(zio));
1683 ASSERT(zio->io_child_count == 0);
1688 if (BP_SHOULD_BYTESWAP(bp))
1689 byteswap_uint64_array(zio->io_data, zio->io_size);
1691 ASSERT(zio->io_data == gn->gn_gbh);
1692 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1693 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1695 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1696 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1697 if (!BP_IS_GANG(gbp))
1699 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1704 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1706 zio_t *gio = pio->io_gang_leader;
1709 ASSERT(BP_IS_GANG(bp) == !!gn);
1710 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1711 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1714 * If you're a gang header, your data is in gn->gn_gbh.
1715 * If you're a gang member, your data is in 'data' and gn == NULL.
1717 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1720 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1722 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1723 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1724 if (BP_IS_HOLE(gbp))
1726 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1727 data = (char *)data + BP_GET_PSIZE(gbp);
1731 if (gn == gio->io_gang_tree && gio->io_data != NULL)
1732 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1739 zio_gang_assemble(zio_t *zio)
1741 blkptr_t *bp = zio->io_bp;
1743 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1744 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1746 zio->io_gang_leader = zio;
1748 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1750 return (ZIO_PIPELINE_CONTINUE);
1754 zio_gang_issue(zio_t *zio)
1756 blkptr_t *bp = zio->io_bp;
1758 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1759 return (ZIO_PIPELINE_STOP);
1761 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1762 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1764 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1765 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1767 zio_gang_tree_free(&zio->io_gang_tree);
1769 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1771 return (ZIO_PIPELINE_CONTINUE);
1775 zio_write_gang_member_ready(zio_t *zio)
1777 zio_t *pio = zio_unique_parent(zio);
1778 zio_t *gio = zio->io_gang_leader;
1779 dva_t *cdva = zio->io_bp->blk_dva;
1780 dva_t *pdva = pio->io_bp->blk_dva;
1783 if (BP_IS_HOLE(zio->io_bp))
1786 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1788 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1789 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1790 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1791 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1792 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1794 mutex_enter(&pio->io_lock);
1795 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1796 ASSERT(DVA_GET_GANG(&pdva[d]));
1797 asize = DVA_GET_ASIZE(&pdva[d]);
1798 asize += DVA_GET_ASIZE(&cdva[d]);
1799 DVA_SET_ASIZE(&pdva[d], asize);
1801 mutex_exit(&pio->io_lock);
1805 zio_write_gang_block(zio_t *pio)
1807 spa_t *spa = pio->io_spa;
1808 blkptr_t *bp = pio->io_bp;
1809 zio_t *gio = pio->io_gang_leader;
1811 zio_gang_node_t *gn, **gnpp;
1812 zio_gbh_phys_t *gbh;
1813 uint64_t txg = pio->io_txg;
1814 uint64_t resid = pio->io_size;
1816 int copies = gio->io_prop.zp_copies;
1817 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1821 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1822 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1823 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1825 pio->io_error = error;
1826 return (ZIO_PIPELINE_CONTINUE);
1830 gnpp = &gio->io_gang_tree;
1832 gnpp = pio->io_private;
1833 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1836 gn = zio_gang_node_alloc(gnpp);
1838 bzero(gbh, SPA_GANGBLOCKSIZE);
1841 * Create the gang header.
1843 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1844 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1847 * Create and nowait the gang children.
1849 for (int g = 0; resid != 0; resid -= lsize, g++) {
1850 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1852 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1854 zp.zp_checksum = gio->io_prop.zp_checksum;
1855 zp.zp_compress = ZIO_COMPRESS_OFF;
1856 zp.zp_type = DMU_OT_NONE;
1858 zp.zp_copies = gio->io_prop.zp_copies;
1859 zp.zp_dedup = B_FALSE;
1860 zp.zp_dedup_verify = B_FALSE;
1861 zp.zp_nopwrite = B_FALSE;
1863 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1864 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1865 zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1866 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1867 &pio->io_bookmark));
1871 * Set pio's pipeline to just wait for zio to finish.
1873 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1877 return (ZIO_PIPELINE_CONTINUE);
1881 * The zio_nop_write stage in the pipeline determines if allocating
1882 * a new bp is necessary. By leveraging a cryptographically secure checksum,
1883 * such as SHA256, we can compare the checksums of the new data and the old
1884 * to determine if allocating a new block is required. The nopwrite
1885 * feature can handle writes in either syncing or open context (i.e. zil
1886 * writes) and as a result is mutually exclusive with dedup.
1889 zio_nop_write(zio_t *zio)
1891 blkptr_t *bp = zio->io_bp;
1892 blkptr_t *bp_orig = &zio->io_bp_orig;
1893 zio_prop_t *zp = &zio->io_prop;
1895 ASSERT(BP_GET_LEVEL(bp) == 0);
1896 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1897 ASSERT(zp->zp_nopwrite);
1898 ASSERT(!zp->zp_dedup);
1899 ASSERT(zio->io_bp_override == NULL);
1900 ASSERT(IO_IS_ALLOCATING(zio));
1903 * Check to see if the original bp and the new bp have matching
1904 * characteristics (i.e. same checksum, compression algorithms, etc).
1905 * If they don't then just continue with the pipeline which will
1906 * allocate a new bp.
1908 if (BP_IS_HOLE(bp_orig) ||
1909 !zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup ||
1910 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
1911 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
1912 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
1913 zp->zp_copies != BP_GET_NDVAS(bp_orig))
1914 return (ZIO_PIPELINE_CONTINUE);
1917 * If the checksums match then reset the pipeline so that we
1918 * avoid allocating a new bp and issuing any I/O.
1920 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
1921 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup);
1922 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
1923 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
1924 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
1925 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
1926 sizeof (uint64_t)) == 0);
1929 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1930 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1933 return (ZIO_PIPELINE_CONTINUE);
1937 * ==========================================================================
1939 * ==========================================================================
1942 zio_ddt_child_read_done(zio_t *zio)
1944 blkptr_t *bp = zio->io_bp;
1945 ddt_entry_t *dde = zio->io_private;
1947 zio_t *pio = zio_unique_parent(zio);
1949 mutex_enter(&pio->io_lock);
1950 ddp = ddt_phys_select(dde, bp);
1951 if (zio->io_error == 0)
1952 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1953 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1954 dde->dde_repair_data = zio->io_data;
1956 zio_buf_free(zio->io_data, zio->io_size);
1957 mutex_exit(&pio->io_lock);
1961 zio_ddt_read_start(zio_t *zio)
1963 blkptr_t *bp = zio->io_bp;
1965 ASSERT(BP_GET_DEDUP(bp));
1966 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1967 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1969 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1970 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1971 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1972 ddt_phys_t *ddp = dde->dde_phys;
1973 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1976 ASSERT(zio->io_vsd == NULL);
1979 if (ddp_self == NULL)
1980 return (ZIO_PIPELINE_CONTINUE);
1982 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1983 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1985 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
1987 zio_nowait(zio_read(zio, zio->io_spa, &blk,
1988 zio_buf_alloc(zio->io_size), zio->io_size,
1989 zio_ddt_child_read_done, dde, zio->io_priority,
1990 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
1991 &zio->io_bookmark));
1993 return (ZIO_PIPELINE_CONTINUE);
1996 zio_nowait(zio_read(zio, zio->io_spa, bp,
1997 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
1998 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2000 return (ZIO_PIPELINE_CONTINUE);
2004 zio_ddt_read_done(zio_t *zio)
2006 blkptr_t *bp = zio->io_bp;
2008 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2009 return (ZIO_PIPELINE_STOP);
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 = zio->io_vsd;
2019 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2020 return (ZIO_PIPELINE_CONTINUE);
2023 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2024 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2025 return (ZIO_PIPELINE_STOP);
2027 if (dde->dde_repair_data != NULL) {
2028 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2029 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2031 ddt_repair_done(ddt, dde);
2035 ASSERT(zio->io_vsd == NULL);
2037 return (ZIO_PIPELINE_CONTINUE);
2041 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2043 spa_t *spa = zio->io_spa;
2046 * Note: we compare the original data, not the transformed data,
2047 * because when zio->io_bp is an override bp, we will not have
2048 * pushed the I/O transforms. That's an important optimization
2049 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2051 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2052 zio_t *lio = dde->dde_lead_zio[p];
2055 return (lio->io_orig_size != zio->io_orig_size ||
2056 bcmp(zio->io_orig_data, lio->io_orig_data,
2057 zio->io_orig_size) != 0);
2061 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2062 ddt_phys_t *ddp = &dde->dde_phys[p];
2064 if (ddp->ddp_phys_birth != 0) {
2065 arc_buf_t *abuf = NULL;
2066 uint32_t aflags = ARC_WAIT;
2067 blkptr_t blk = *zio->io_bp;
2070 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2074 error = arc_read(NULL, spa, &blk,
2075 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2076 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2077 &aflags, &zio->io_bookmark);
2080 if (arc_buf_size(abuf) != zio->io_orig_size ||
2081 bcmp(abuf->b_data, zio->io_orig_data,
2082 zio->io_orig_size) != 0)
2083 error = SET_ERROR(EEXIST);
2084 VERIFY(arc_buf_remove_ref(abuf, &abuf));
2088 return (error != 0);
2096 zio_ddt_child_write_ready(zio_t *zio)
2098 int p = zio->io_prop.zp_copies;
2099 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2100 ddt_entry_t *dde = zio->io_private;
2101 ddt_phys_t *ddp = &dde->dde_phys[p];
2109 ASSERT(dde->dde_lead_zio[p] == zio);
2111 ddt_phys_fill(ddp, zio->io_bp);
2113 while ((pio = zio_walk_parents(zio)) != NULL)
2114 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2120 zio_ddt_child_write_done(zio_t *zio)
2122 int p = zio->io_prop.zp_copies;
2123 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2124 ddt_entry_t *dde = zio->io_private;
2125 ddt_phys_t *ddp = &dde->dde_phys[p];
2129 ASSERT(ddp->ddp_refcnt == 0);
2130 ASSERT(dde->dde_lead_zio[p] == zio);
2131 dde->dde_lead_zio[p] = NULL;
2133 if (zio->io_error == 0) {
2134 while (zio_walk_parents(zio) != NULL)
2135 ddt_phys_addref(ddp);
2137 ddt_phys_clear(ddp);
2144 zio_ddt_ditto_write_done(zio_t *zio)
2146 int p = DDT_PHYS_DITTO;
2147 zio_prop_t *zp = &zio->io_prop;
2148 blkptr_t *bp = zio->io_bp;
2149 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2150 ddt_entry_t *dde = zio->io_private;
2151 ddt_phys_t *ddp = &dde->dde_phys[p];
2152 ddt_key_t *ddk = &dde->dde_key;
2156 ASSERT(ddp->ddp_refcnt == 0);
2157 ASSERT(dde->dde_lead_zio[p] == zio);
2158 dde->dde_lead_zio[p] = NULL;
2160 if (zio->io_error == 0) {
2161 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2162 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2163 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2164 if (ddp->ddp_phys_birth != 0)
2165 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2166 ddt_phys_fill(ddp, bp);
2173 zio_ddt_write(zio_t *zio)
2175 spa_t *spa = zio->io_spa;
2176 blkptr_t *bp = zio->io_bp;
2177 uint64_t txg = zio->io_txg;
2178 zio_prop_t *zp = &zio->io_prop;
2179 int p = zp->zp_copies;
2183 ddt_t *ddt = ddt_select(spa, bp);
2187 ASSERT(BP_GET_DEDUP(bp));
2188 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2189 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2192 dde = ddt_lookup(ddt, bp, B_TRUE);
2193 ddp = &dde->dde_phys[p];
2195 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2197 * If we're using a weak checksum, upgrade to a strong checksum
2198 * and try again. If we're already using a strong checksum,
2199 * we can't resolve it, so just convert to an ordinary write.
2200 * (And automatically e-mail a paper to Nature?)
2202 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2203 zp->zp_checksum = spa_dedup_checksum(spa);
2204 zio_pop_transforms(zio);
2205 zio->io_stage = ZIO_STAGE_OPEN;
2208 zp->zp_dedup = B_FALSE;
2210 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2212 return (ZIO_PIPELINE_CONTINUE);
2215 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2216 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2218 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2219 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2220 zio_prop_t czp = *zp;
2222 czp.zp_copies = ditto_copies;
2225 * If we arrived here with an override bp, we won't have run
2226 * the transform stack, so we won't have the data we need to
2227 * generate a child i/o. So, toss the override bp and restart.
2228 * This is safe, because using the override bp is just an
2229 * optimization; and it's rare, so the cost doesn't matter.
2231 if (zio->io_bp_override) {
2232 zio_pop_transforms(zio);
2233 zio->io_stage = ZIO_STAGE_OPEN;
2234 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2235 zio->io_bp_override = NULL;
2238 return (ZIO_PIPELINE_CONTINUE);
2241 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2242 zio->io_orig_size, &czp, NULL,
2243 zio_ddt_ditto_write_done, dde, zio->io_priority,
2244 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2246 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2247 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2250 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2251 if (ddp->ddp_phys_birth != 0)
2252 ddt_bp_fill(ddp, bp, txg);
2253 if (dde->dde_lead_zio[p] != NULL)
2254 zio_add_child(zio, dde->dde_lead_zio[p]);
2256 ddt_phys_addref(ddp);
2257 } else if (zio->io_bp_override) {
2258 ASSERT(bp->blk_birth == txg);
2259 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2260 ddt_phys_fill(ddp, bp);
2261 ddt_phys_addref(ddp);
2263 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2264 zio->io_orig_size, zp, zio_ddt_child_write_ready,
2265 zio_ddt_child_write_done, dde, zio->io_priority,
2266 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2268 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2269 dde->dde_lead_zio[p] = cio;
2279 return (ZIO_PIPELINE_CONTINUE);
2282 ddt_entry_t *freedde; /* for debugging */
2285 zio_ddt_free(zio_t *zio)
2287 spa_t *spa = zio->io_spa;
2288 blkptr_t *bp = zio->io_bp;
2289 ddt_t *ddt = ddt_select(spa, bp);
2293 ASSERT(BP_GET_DEDUP(bp));
2294 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2297 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2298 ddp = ddt_phys_select(dde, bp);
2299 ddt_phys_decref(ddp);
2302 return (ZIO_PIPELINE_CONTINUE);
2306 * ==========================================================================
2307 * Allocate and free blocks
2308 * ==========================================================================
2311 zio_dva_allocate(zio_t *zio)
2313 spa_t *spa = zio->io_spa;
2314 metaslab_class_t *mc = spa_normal_class(spa);
2315 blkptr_t *bp = zio->io_bp;
2319 if (zio->io_gang_leader == NULL) {
2320 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2321 zio->io_gang_leader = zio;
2324 ASSERT(BP_IS_HOLE(bp));
2325 ASSERT0(BP_GET_NDVAS(bp));
2326 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2327 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2328 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2331 * The dump device does not support gang blocks so allocation on
2332 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2333 * the "fast" gang feature.
2335 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2336 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2337 METASLAB_GANG_CHILD : 0;
2338 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2339 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2342 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2343 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2345 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2346 return (zio_write_gang_block(zio));
2347 zio->io_error = error;
2350 return (ZIO_PIPELINE_CONTINUE);
2354 zio_dva_free(zio_t *zio)
2356 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2358 return (ZIO_PIPELINE_CONTINUE);
2362 zio_dva_claim(zio_t *zio)
2366 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2368 zio->io_error = error;
2370 return (ZIO_PIPELINE_CONTINUE);
2374 * Undo an allocation. This is used by zio_done() when an I/O fails
2375 * and we want to give back the block we just allocated.
2376 * This handles both normal blocks and gang blocks.
2379 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2381 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2382 ASSERT(zio->io_bp_override == NULL);
2384 if (!BP_IS_HOLE(bp))
2385 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2388 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2389 zio_dva_unallocate(zio, gn->gn_child[g],
2390 &gn->gn_gbh->zg_blkptr[g]);
2396 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2399 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2400 uint64_t size, boolean_t use_slog)
2404 ASSERT(txg > spa_syncing_txg(spa));
2407 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2408 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2409 * when allocating them.
2412 error = metaslab_alloc(spa, spa_log_class(spa), size,
2413 new_bp, 1, txg, old_bp,
2414 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2418 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2419 new_bp, 1, txg, old_bp,
2420 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2424 BP_SET_LSIZE(new_bp, size);
2425 BP_SET_PSIZE(new_bp, size);
2426 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2427 BP_SET_CHECKSUM(new_bp,
2428 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2429 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2430 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2431 BP_SET_LEVEL(new_bp, 0);
2432 BP_SET_DEDUP(new_bp, 0);
2433 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2440 * Free an intent log block.
2443 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2445 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2446 ASSERT(!BP_IS_GANG(bp));
2448 zio_free(spa, txg, bp);
2452 * ==========================================================================
2453 * Read, write and delete to physical devices
2454 * ==========================================================================
2457 zio_vdev_io_start(zio_t *zio)
2459 vdev_t *vd = zio->io_vd;
2461 spa_t *spa = zio->io_spa;
2463 ASSERT(zio->io_error == 0);
2464 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2467 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2468 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2471 * The mirror_ops handle multiple DVAs in a single BP.
2473 return (vdev_mirror_ops.vdev_op_io_start(zio));
2476 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE) {
2477 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg);
2478 return (ZIO_PIPELINE_CONTINUE);
2482 * We keep track of time-sensitive I/Os so that the scan thread
2483 * can quickly react to certain workloads. In particular, we care
2484 * about non-scrubbing, top-level reads and writes with the following
2486 * - synchronous writes of user data to non-slog devices
2487 * - any reads of user data
2488 * When these conditions are met, adjust the timestamp of spa_last_io
2489 * which allows the scan thread to adjust its workload accordingly.
2491 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2492 vd == vd->vdev_top && !vd->vdev_islog &&
2493 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2494 zio->io_txg != spa_syncing_txg(spa)) {
2495 uint64_t old = spa->spa_last_io;
2496 uint64_t new = ddi_get_lbolt64();
2498 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2501 align = 1ULL << vd->vdev_top->vdev_ashift;
2503 if (P2PHASE(zio->io_size, align) != 0) {
2504 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2506 if (zio->io_type == ZIO_TYPE_READ ||
2507 zio->io_type == ZIO_TYPE_WRITE)
2508 abuf = zio_buf_alloc(asize);
2509 ASSERT(vd == vd->vdev_top);
2510 if (zio->io_type == ZIO_TYPE_WRITE) {
2511 bcopy(zio->io_data, abuf, zio->io_size);
2512 bzero(abuf + zio->io_size, asize - zio->io_size);
2514 zio_push_transform(zio, abuf, asize, abuf ? asize : 0,
2518 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2519 ASSERT(P2PHASE(zio->io_size, align) == 0);
2520 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa));
2523 * If this is a repair I/O, and there's no self-healing involved --
2524 * that is, we're just resilvering what we expect to resilver --
2525 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2526 * This prevents spurious resilvering with nested replication.
2527 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2528 * A is out of date, we'll read from C+D, then use the data to
2529 * resilver A+B -- but we don't actually want to resilver B, just A.
2530 * The top-level mirror has no way to know this, so instead we just
2531 * discard unnecessary repairs as we work our way down the vdev tree.
2532 * The same logic applies to any form of nested replication:
2533 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2535 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2536 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2537 zio->io_txg != 0 && /* not a delegated i/o */
2538 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2539 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2540 zio_vdev_io_bypass(zio);
2541 return (ZIO_PIPELINE_CONTINUE);
2544 if (vd->vdev_ops->vdev_op_leaf &&
2545 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2547 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2548 return (ZIO_PIPELINE_CONTINUE);
2550 if ((zio = vdev_queue_io(zio)) == NULL)
2551 return (ZIO_PIPELINE_STOP);
2553 if (!vdev_accessible(vd, zio)) {
2554 zio->io_error = SET_ERROR(ENXIO);
2556 return (ZIO_PIPELINE_STOP);
2561 * Note that we ignore repair writes for TRIM because they can conflict
2562 * with normal writes. This isn't an issue because, by definition, we
2563 * only repair blocks that aren't freed.
2565 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_WRITE &&
2566 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2567 if (!trim_map_write_start(zio))
2568 return (ZIO_PIPELINE_STOP);
2571 return (vd->vdev_ops->vdev_op_io_start(zio));
2575 zio_vdev_io_done(zio_t *zio)
2577 vdev_t *vd = zio->io_vd;
2578 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2579 boolean_t unexpected_error = B_FALSE;
2581 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2582 return (ZIO_PIPELINE_STOP);
2584 ASSERT(zio->io_type == ZIO_TYPE_READ ||
2585 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE);
2587 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2588 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2590 if (zio->io_type == ZIO_TYPE_WRITE &&
2591 !(zio->io_flags & ZIO_FLAG_IO_REPAIR))
2592 trim_map_write_done(zio);
2594 vdev_queue_io_done(zio);
2596 if (zio->io_type == ZIO_TYPE_WRITE)
2597 vdev_cache_write(zio);
2599 if (zio_injection_enabled && zio->io_error == 0)
2600 zio->io_error = zio_handle_device_injection(vd,
2603 if (zio_injection_enabled && zio->io_error == 0)
2604 zio->io_error = zio_handle_label_injection(zio, EIO);
2606 if (zio->io_error) {
2607 if (!vdev_accessible(vd, zio)) {
2608 zio->io_error = SET_ERROR(ENXIO);
2610 unexpected_error = B_TRUE;
2615 ops->vdev_op_io_done(zio);
2617 if (unexpected_error)
2618 VERIFY(vdev_probe(vd, zio) == NULL);
2620 return (ZIO_PIPELINE_CONTINUE);
2624 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2625 * disk, and use that to finish the checksum ereport later.
2628 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2629 const void *good_buf)
2631 /* no processing needed */
2632 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2637 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2639 void *buf = zio_buf_alloc(zio->io_size);
2641 bcopy(zio->io_data, buf, zio->io_size);
2643 zcr->zcr_cbinfo = zio->io_size;
2644 zcr->zcr_cbdata = buf;
2645 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2646 zcr->zcr_free = zio_buf_free;
2650 zio_vdev_io_assess(zio_t *zio)
2652 vdev_t *vd = zio->io_vd;
2654 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2655 return (ZIO_PIPELINE_STOP);
2657 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2658 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2660 if (zio->io_vsd != NULL) {
2661 zio->io_vsd_ops->vsd_free(zio);
2665 if (zio_injection_enabled && zio->io_error == 0)
2666 zio->io_error = zio_handle_fault_injection(zio, EIO);
2668 if (zio->io_type == ZIO_TYPE_IOCTL && zio->io_cmd == DKIOCTRIM)
2669 switch (zio->io_error) {
2671 ZIO_TRIM_STAT_INCR(bytes, zio->io_size);
2672 ZIO_TRIM_STAT_BUMP(success);
2675 ZIO_TRIM_STAT_BUMP(unsupported);
2678 ZIO_TRIM_STAT_BUMP(failed);
2683 * If the I/O failed, determine whether we should attempt to retry it.
2685 * On retry, we cut in line in the issue queue, since we don't want
2686 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2688 if (zio->io_error && vd == NULL &&
2689 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2690 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2691 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2693 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2694 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2695 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2696 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2697 zio_requeue_io_start_cut_in_line);
2698 return (ZIO_PIPELINE_STOP);
2702 * If we got an error on a leaf device, convert it to ENXIO
2703 * if the device is not accessible at all.
2705 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2706 !vdev_accessible(vd, zio))
2707 zio->io_error = SET_ERROR(ENXIO);
2710 * If we can't write to an interior vdev (mirror or RAID-Z),
2711 * set vdev_cant_write so that we stop trying to allocate from it.
2713 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2714 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
2715 vd->vdev_cant_write = B_TRUE;
2719 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2721 return (ZIO_PIPELINE_CONTINUE);
2725 zio_vdev_io_reissue(zio_t *zio)
2727 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2728 ASSERT(zio->io_error == 0);
2730 zio->io_stage >>= 1;
2734 zio_vdev_io_redone(zio_t *zio)
2736 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2738 zio->io_stage >>= 1;
2742 zio_vdev_io_bypass(zio_t *zio)
2744 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2745 ASSERT(zio->io_error == 0);
2747 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2748 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2752 * ==========================================================================
2753 * Generate and verify checksums
2754 * ==========================================================================
2757 zio_checksum_generate(zio_t *zio)
2759 blkptr_t *bp = zio->io_bp;
2760 enum zio_checksum checksum;
2764 * This is zio_write_phys().
2765 * We're either generating a label checksum, or none at all.
2767 checksum = zio->io_prop.zp_checksum;
2769 if (checksum == ZIO_CHECKSUM_OFF)
2770 return (ZIO_PIPELINE_CONTINUE);
2772 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2774 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2775 ASSERT(!IO_IS_ALLOCATING(zio));
2776 checksum = ZIO_CHECKSUM_GANG_HEADER;
2778 checksum = BP_GET_CHECKSUM(bp);
2782 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2784 return (ZIO_PIPELINE_CONTINUE);
2788 zio_checksum_verify(zio_t *zio)
2790 zio_bad_cksum_t info;
2791 blkptr_t *bp = zio->io_bp;
2794 ASSERT(zio->io_vd != NULL);
2798 * This is zio_read_phys().
2799 * We're either verifying a label checksum, or nothing at all.
2801 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2802 return (ZIO_PIPELINE_CONTINUE);
2804 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2807 if ((error = zio_checksum_error(zio, &info)) != 0) {
2808 zio->io_error = error;
2809 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2810 zfs_ereport_start_checksum(zio->io_spa,
2811 zio->io_vd, zio, zio->io_offset,
2812 zio->io_size, NULL, &info);
2816 return (ZIO_PIPELINE_CONTINUE);
2820 * Called by RAID-Z to ensure we don't compute the checksum twice.
2823 zio_checksum_verified(zio_t *zio)
2825 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2829 * ==========================================================================
2830 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2831 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2832 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2833 * indicate errors that are specific to one I/O, and most likely permanent.
2834 * Any other error is presumed to be worse because we weren't expecting it.
2835 * ==========================================================================
2838 zio_worst_error(int e1, int e2)
2840 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2843 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2844 if (e1 == zio_error_rank[r1])
2847 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2848 if (e2 == zio_error_rank[r2])
2851 return (r1 > r2 ? e1 : e2);
2855 * ==========================================================================
2857 * ==========================================================================
2860 zio_ready(zio_t *zio)
2862 blkptr_t *bp = zio->io_bp;
2863 zio_t *pio, *pio_next;
2865 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2866 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2867 return (ZIO_PIPELINE_STOP);
2869 if (zio->io_ready) {
2870 ASSERT(IO_IS_ALLOCATING(zio));
2871 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
2872 (zio->io_flags & ZIO_FLAG_NOPWRITE));
2873 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2878 if (bp != NULL && bp != &zio->io_bp_copy)
2879 zio->io_bp_copy = *bp;
2882 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2884 mutex_enter(&zio->io_lock);
2885 zio->io_state[ZIO_WAIT_READY] = 1;
2886 pio = zio_walk_parents(zio);
2887 mutex_exit(&zio->io_lock);
2890 * As we notify zio's parents, new parents could be added.
2891 * New parents go to the head of zio's io_parent_list, however,
2892 * so we will (correctly) not notify them. The remainder of zio's
2893 * io_parent_list, from 'pio_next' onward, cannot change because
2894 * all parents must wait for us to be done before they can be done.
2896 for (; pio != NULL; pio = pio_next) {
2897 pio_next = zio_walk_parents(zio);
2898 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2901 if (zio->io_flags & ZIO_FLAG_NODATA) {
2902 if (BP_IS_GANG(bp)) {
2903 zio->io_flags &= ~ZIO_FLAG_NODATA;
2905 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2906 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2910 if (zio_injection_enabled &&
2911 zio->io_spa->spa_syncing_txg == zio->io_txg)
2912 zio_handle_ignored_writes(zio);
2914 return (ZIO_PIPELINE_CONTINUE);
2918 zio_done(zio_t *zio)
2920 spa_t *spa = zio->io_spa;
2921 zio_t *lio = zio->io_logical;
2922 blkptr_t *bp = zio->io_bp;
2923 vdev_t *vd = zio->io_vd;
2924 uint64_t psize = zio->io_size;
2925 zio_t *pio, *pio_next;
2928 * If our children haven't all completed,
2929 * wait for them and then repeat this pipeline stage.
2931 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2932 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2933 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2934 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2935 return (ZIO_PIPELINE_STOP);
2937 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2938 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2939 ASSERT(zio->io_children[c][w] == 0);
2942 ASSERT(bp->blk_pad[0] == 0);
2943 ASSERT(bp->blk_pad[1] == 0);
2944 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2945 (bp == zio_unique_parent(zio)->io_bp));
2946 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
2947 zio->io_bp_override == NULL &&
2948 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2949 ASSERT(!BP_SHOULD_BYTESWAP(bp));
2950 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
2951 ASSERT(BP_COUNT_GANG(bp) == 0 ||
2952 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
2954 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
2955 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
2959 * If there were child vdev/gang/ddt errors, they apply to us now.
2961 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2962 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2963 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2966 * If the I/O on the transformed data was successful, generate any
2967 * checksum reports now while we still have the transformed data.
2969 if (zio->io_error == 0) {
2970 while (zio->io_cksum_report != NULL) {
2971 zio_cksum_report_t *zcr = zio->io_cksum_report;
2972 uint64_t align = zcr->zcr_align;
2973 uint64_t asize = P2ROUNDUP(psize, align);
2974 char *abuf = zio->io_data;
2976 if (asize != psize) {
2977 abuf = zio_buf_alloc(asize);
2978 bcopy(zio->io_data, abuf, psize);
2979 bzero(abuf + psize, asize - psize);
2982 zio->io_cksum_report = zcr->zcr_next;
2983 zcr->zcr_next = NULL;
2984 zcr->zcr_finish(zcr, abuf);
2985 zfs_ereport_free_checksum(zcr);
2988 zio_buf_free(abuf, asize);
2992 zio_pop_transforms(zio); /* note: may set zio->io_error */
2994 vdev_stat_update(zio, psize);
2996 if (zio->io_error) {
2998 * If this I/O is attached to a particular vdev,
2999 * generate an error message describing the I/O failure
3000 * at the block level. We ignore these errors if the
3001 * device is currently unavailable.
3003 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3004 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3006 if ((zio->io_error == EIO || !(zio->io_flags &
3007 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3010 * For logical I/O requests, tell the SPA to log the
3011 * error and generate a logical data ereport.
3013 spa_log_error(spa, zio);
3014 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3019 if (zio->io_error && zio == lio) {
3021 * Determine whether zio should be reexecuted. This will
3022 * propagate all the way to the root via zio_notify_parent().
3024 ASSERT(vd == NULL && bp != NULL);
3025 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3027 if (IO_IS_ALLOCATING(zio) &&
3028 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3029 if (zio->io_error != ENOSPC)
3030 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3032 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3035 if ((zio->io_type == ZIO_TYPE_READ ||
3036 zio->io_type == ZIO_TYPE_FREE) &&
3037 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3038 zio->io_error == ENXIO &&
3039 spa_load_state(spa) == SPA_LOAD_NONE &&
3040 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3041 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3043 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3044 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3047 * Here is a possibly good place to attempt to do
3048 * either combinatorial reconstruction or error correction
3049 * based on checksums. It also might be a good place
3050 * to send out preliminary ereports before we suspend
3056 * If there were logical child errors, they apply to us now.
3057 * We defer this until now to avoid conflating logical child
3058 * errors with errors that happened to the zio itself when
3059 * updating vdev stats and reporting FMA events above.
3061 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3063 if ((zio->io_error || zio->io_reexecute) &&
3064 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3065 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3066 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3068 zio_gang_tree_free(&zio->io_gang_tree);
3071 * Godfather I/Os should never suspend.
3073 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3074 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3075 zio->io_reexecute = 0;
3077 if (zio->io_reexecute) {
3079 * This is a logical I/O that wants to reexecute.
3081 * Reexecute is top-down. When an i/o fails, if it's not
3082 * the root, it simply notifies its parent and sticks around.
3083 * The parent, seeing that it still has children in zio_done(),
3084 * does the same. This percolates all the way up to the root.
3085 * The root i/o will reexecute or suspend the entire tree.
3087 * This approach ensures that zio_reexecute() honors
3088 * all the original i/o dependency relationships, e.g.
3089 * parents not executing until children are ready.
3091 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3093 zio->io_gang_leader = NULL;
3095 mutex_enter(&zio->io_lock);
3096 zio->io_state[ZIO_WAIT_DONE] = 1;
3097 mutex_exit(&zio->io_lock);
3100 * "The Godfather" I/O monitors its children but is
3101 * not a true parent to them. It will track them through
3102 * the pipeline but severs its ties whenever they get into
3103 * trouble (e.g. suspended). This allows "The Godfather"
3104 * I/O to return status without blocking.
3106 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3107 zio_link_t *zl = zio->io_walk_link;
3108 pio_next = zio_walk_parents(zio);
3110 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3111 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3112 zio_remove_child(pio, zio, zl);
3113 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3117 if ((pio = zio_unique_parent(zio)) != NULL) {
3119 * We're not a root i/o, so there's nothing to do
3120 * but notify our parent. Don't propagate errors
3121 * upward since we haven't permanently failed yet.
3123 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3124 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3125 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3126 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3128 * We'd fail again if we reexecuted now, so suspend
3129 * until conditions improve (e.g. device comes online).
3131 zio_suspend(spa, zio);
3134 * Reexecution is potentially a huge amount of work.
3135 * Hand it off to the otherwise-unused claim taskq.
3138 (void) taskq_dispatch_safe(
3139 spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
3140 (task_func_t *)zio_reexecute, zio, TQ_SLEEP,
3143 (void) taskq_dispatch(
3144 spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE],
3145 (task_func_t *)zio_reexecute, zio, TQ_SLEEP);
3148 return (ZIO_PIPELINE_STOP);
3151 ASSERT(zio->io_child_count == 0);
3152 ASSERT(zio->io_reexecute == 0);
3153 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3156 * Report any checksum errors, since the I/O is complete.
3158 while (zio->io_cksum_report != NULL) {
3159 zio_cksum_report_t *zcr = zio->io_cksum_report;
3160 zio->io_cksum_report = zcr->zcr_next;
3161 zcr->zcr_next = NULL;
3162 zcr->zcr_finish(zcr, NULL);
3163 zfs_ereport_free_checksum(zcr);
3167 * It is the responsibility of the done callback to ensure that this
3168 * particular zio is no longer discoverable for adoption, and as
3169 * such, cannot acquire any new parents.
3174 mutex_enter(&zio->io_lock);
3175 zio->io_state[ZIO_WAIT_DONE] = 1;
3176 mutex_exit(&zio->io_lock);
3178 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3179 zio_link_t *zl = zio->io_walk_link;
3180 pio_next = zio_walk_parents(zio);
3181 zio_remove_child(pio, zio, zl);
3182 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3185 if (zio->io_waiter != NULL) {
3186 mutex_enter(&zio->io_lock);
3187 zio->io_executor = NULL;
3188 cv_broadcast(&zio->io_cv);
3189 mutex_exit(&zio->io_lock);
3194 return (ZIO_PIPELINE_STOP);
3198 * ==========================================================================
3199 * I/O pipeline definition
3200 * ==========================================================================
3202 static zio_pipe_stage_t *zio_pipeline[] = {
3208 zio_checksum_generate,
3223 zio_checksum_verify,
3227 /* dnp is the dnode for zb1->zb_object */
3229 zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_t *zb1,
3230 const zbookmark_t *zb2)
3232 uint64_t zb1nextL0, zb2thisobj;
3234 ASSERT(zb1->zb_objset == zb2->zb_objset);
3235 ASSERT(zb2->zb_level == 0);
3238 * A bookmark in the deadlist is considered to be after
3241 if (zb2->zb_object == DMU_DEADLIST_OBJECT)
3244 /* The objset_phys_t isn't before anything. */
3248 zb1nextL0 = (zb1->zb_blkid + 1) <<
3249 ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
3251 zb2thisobj = zb2->zb_object ? zb2->zb_object :
3252 zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
3254 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3255 uint64_t nextobj = zb1nextL0 *
3256 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
3257 return (nextobj <= zb2thisobj);
3260 if (zb1->zb_object < zb2thisobj)
3262 if (zb1->zb_object > zb2thisobj)
3264 if (zb2->zb_object == DMU_META_DNODE_OBJECT)
3266 return (zb1nextL0 <= zb2->zb_blkid);