4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2013 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
27 #include <sys/zfs_context.h>
28 #include <sys/fm/fs/zfs.h>
31 #include <sys/spa_impl.h>
32 #include <sys/vdev_impl.h>
33 #include <sys/zio_impl.h>
34 #include <sys/zio_compress.h>
35 #include <sys/zio_checksum.h>
36 #include <sys/dmu_objset.h>
39 #include <sys/trim_map.h>
41 SYSCTL_DECL(_vfs_zfs);
42 SYSCTL_NODE(_vfs_zfs, OID_AUTO, zio, CTLFLAG_RW, 0, "ZFS ZIO");
43 static int zio_use_uma = 0;
44 TUNABLE_INT("vfs.zfs.zio.use_uma", &zio_use_uma);
45 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, use_uma, CTLFLAG_RDTUN, &zio_use_uma, 0,
46 "Use uma(9) for ZIO allocations");
48 zio_trim_stats_t zio_trim_stats = {
49 { "bytes", KSTAT_DATA_UINT64,
50 "Number of bytes successfully TRIMmed" },
51 { "success", KSTAT_DATA_UINT64,
52 "Number of successful TRIM requests" },
53 { "unsupported", KSTAT_DATA_UINT64,
54 "Number of TRIM requests that failed because TRIM is not supported" },
55 { "failed", KSTAT_DATA_UINT64,
56 "Number of TRIM requests that failed for reasons other than not supported" },
59 static kstat_t *zio_trim_ksp;
62 * ==========================================================================
64 * ==========================================================================
66 uint8_t zio_priority_table[ZIO_PRIORITY_TABLE_SIZE] = {
67 0, /* ZIO_PRIORITY_NOW */
68 0, /* ZIO_PRIORITY_SYNC_READ */
69 0, /* ZIO_PRIORITY_SYNC_WRITE */
70 0, /* ZIO_PRIORITY_LOG_WRITE */
71 1, /* ZIO_PRIORITY_CACHE_FILL */
72 1, /* ZIO_PRIORITY_AGG */
73 4, /* ZIO_PRIORITY_FREE */
74 4, /* ZIO_PRIORITY_ASYNC_WRITE */
75 6, /* ZIO_PRIORITY_ASYNC_READ */
76 10, /* ZIO_PRIORITY_RESILVER */
77 20, /* ZIO_PRIORITY_SCRUB */
78 2, /* ZIO_PRIORITY_DDT_PREFETCH */
79 30, /* ZIO_PRIORITY_TRIM */
83 * ==========================================================================
84 * I/O type descriptions
85 * ==========================================================================
87 char *zio_type_name[ZIO_TYPES] = {
88 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
93 * ==========================================================================
95 * ==========================================================================
97 kmem_cache_t *zio_cache;
98 kmem_cache_t *zio_link_cache;
99 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
100 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
103 extern vmem_t *zio_alloc_arena;
105 extern int zfs_mg_alloc_failures;
108 * The following actions directly effect the spa's sync-to-convergence logic.
109 * The values below define the sync pass when we start performing the action.
110 * Care should be taken when changing these values as they directly impact
111 * spa_sync() performance. Tuning these values may introduce subtle performance
112 * pathologies and should only be done in the context of performance analysis.
113 * These tunables will eventually be removed and replaced with #defines once
114 * enough analysis has been done to determine optimal values.
116 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
117 * regular blocks are not deferred.
119 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
120 TUNABLE_INT("vfs.zfs.sync_pass_deferred_free", &zfs_sync_pass_deferred_free);
121 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_deferred_free, CTLFLAG_RDTUN,
122 &zfs_sync_pass_deferred_free, 0, "defer frees starting in this pass");
123 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
124 TUNABLE_INT("vfs.zfs.sync_pass_dont_compress", &zfs_sync_pass_dont_compress);
125 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_dont_compress, CTLFLAG_RDTUN,
126 &zfs_sync_pass_dont_compress, 0, "don't compress starting in this pass");
127 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
128 TUNABLE_INT("vfs.zfs.sync_pass_rewrite", &zfs_sync_pass_rewrite);
129 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_rewrite, CTLFLAG_RDTUN,
130 &zfs_sync_pass_rewrite, 0, "rewrite new bps starting in this pass");
133 * An allocating zio is one that either currently has the DVA allocate
134 * stage set or will have it later in its lifetime.
136 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
138 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
141 int zio_buf_debug_limit = 16384;
143 int zio_buf_debug_limit = 0;
150 zio_cache = kmem_cache_create("zio_cache",
151 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
152 zio_link_cache = kmem_cache_create("zio_link_cache",
153 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
156 * For small buffers, we want a cache for each multiple of
157 * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache
158 * for each quarter-power of 2. For large buffers, we want
159 * a cache for each multiple of PAGESIZE.
161 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
162 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
165 size_t cflags = (size > zio_buf_debug_limit) ? (KMC_NODEBUG|KMC_NOTOUCH) : 0;
167 while (p2 & (p2 - 1))
173 * If we are using watchpoints, put each buffer on its own page,
174 * to eliminate the performance overhead of trapping to the
175 * kernel when modifying a non-watched buffer that shares the
176 * page with a watched buffer.
178 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
182 if (size <= 4 * SPA_MINBLOCKSIZE) {
183 align = SPA_MINBLOCKSIZE;
184 } else if (IS_P2ALIGNED(size, PAGESIZE)) {
186 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
192 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
193 zio_buf_cache[c] = kmem_cache_create(name, size,
194 align, NULL, NULL, NULL, NULL, NULL, cflags);
197 * Since zio_data bufs do not appear in crash dumps, we
198 * pass KMC_NOTOUCH so that no allocator metadata is
199 * stored with the buffers.
201 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
202 zio_data_buf_cache[c] = kmem_cache_create(name, size,
203 align, NULL, NULL, NULL, NULL, NULL,
204 cflags | KMC_NOTOUCH);
209 ASSERT(zio_buf_cache[c] != NULL);
210 if (zio_buf_cache[c - 1] == NULL)
211 zio_buf_cache[c - 1] = zio_buf_cache[c];
213 ASSERT(zio_data_buf_cache[c] != NULL);
214 if (zio_data_buf_cache[c - 1] == NULL)
215 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
219 * The zio write taskqs have 1 thread per cpu, allow 1/2 of the taskqs
220 * to fail 3 times per txg or 8 failures, whichever is greater.
222 if (zfs_mg_alloc_failures == 0)
223 zfs_mg_alloc_failures = MAX((3 * max_ncpus / 2), 8);
224 else if (zfs_mg_alloc_failures < 8)
225 zfs_mg_alloc_failures = 8;
229 zio_trim_ksp = kstat_create("zfs", 0, "zio_trim", "misc",
231 sizeof(zio_trim_stats) / sizeof(kstat_named_t),
234 if (zio_trim_ksp != NULL) {
235 zio_trim_ksp->ks_data = &zio_trim_stats;
236 kstat_install(zio_trim_ksp);
244 kmem_cache_t *last_cache = NULL;
245 kmem_cache_t *last_data_cache = NULL;
247 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
248 if (zio_buf_cache[c] != last_cache) {
249 last_cache = zio_buf_cache[c];
250 kmem_cache_destroy(zio_buf_cache[c]);
252 zio_buf_cache[c] = NULL;
254 if (zio_data_buf_cache[c] != last_data_cache) {
255 last_data_cache = zio_data_buf_cache[c];
256 kmem_cache_destroy(zio_data_buf_cache[c]);
258 zio_data_buf_cache[c] = NULL;
261 kmem_cache_destroy(zio_link_cache);
262 kmem_cache_destroy(zio_cache);
266 if (zio_trim_ksp != NULL) {
267 kstat_delete(zio_trim_ksp);
273 * ==========================================================================
274 * Allocate and free I/O buffers
275 * ==========================================================================
279 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
280 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
281 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
282 * excess / transient data in-core during a crashdump.
285 zio_buf_alloc(size_t size)
287 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
289 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
292 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
294 return (kmem_alloc(size, KM_SLEEP));
298 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
299 * crashdump if the kernel panics. This exists so that we will limit the amount
300 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
301 * of kernel heap dumped to disk when the kernel panics)
304 zio_data_buf_alloc(size_t size)
306 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
308 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
311 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
313 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG));
317 zio_buf_free(void *buf, size_t size)
319 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
321 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
324 kmem_cache_free(zio_buf_cache[c], buf);
326 kmem_free(buf, size);
330 zio_data_buf_free(void *buf, size_t size)
332 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
334 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
337 kmem_cache_free(zio_data_buf_cache[c], buf);
339 kmem_free(buf, size);
343 * ==========================================================================
344 * Push and pop I/O transform buffers
345 * ==========================================================================
348 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
349 zio_transform_func_t *transform)
351 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
353 zt->zt_orig_data = zio->io_data;
354 zt->zt_orig_size = zio->io_size;
355 zt->zt_bufsize = bufsize;
356 zt->zt_transform = transform;
358 zt->zt_next = zio->io_transform_stack;
359 zio->io_transform_stack = zt;
366 zio_pop_transforms(zio_t *zio)
370 while ((zt = zio->io_transform_stack) != NULL) {
371 if (zt->zt_transform != NULL)
372 zt->zt_transform(zio,
373 zt->zt_orig_data, zt->zt_orig_size);
375 if (zt->zt_bufsize != 0)
376 zio_buf_free(zio->io_data, zt->zt_bufsize);
378 zio->io_data = zt->zt_orig_data;
379 zio->io_size = zt->zt_orig_size;
380 zio->io_transform_stack = zt->zt_next;
382 kmem_free(zt, sizeof (zio_transform_t));
387 * ==========================================================================
388 * I/O transform callbacks for subblocks and decompression
389 * ==========================================================================
392 zio_subblock(zio_t *zio, void *data, uint64_t size)
394 ASSERT(zio->io_size > size);
396 if (zio->io_type == ZIO_TYPE_READ)
397 bcopy(zio->io_data, data, size);
401 zio_decompress(zio_t *zio, void *data, uint64_t size)
403 if (zio->io_error == 0 &&
404 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
405 zio->io_data, data, zio->io_size, size) != 0)
406 zio->io_error = SET_ERROR(EIO);
410 * ==========================================================================
411 * I/O parent/child relationships and pipeline interlocks
412 * ==========================================================================
415 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
416 * continue calling these functions until they return NULL.
417 * Otherwise, the next caller will pick up the list walk in
418 * some indeterminate state. (Otherwise every caller would
419 * have to pass in a cookie to keep the state represented by
420 * io_walk_link, which gets annoying.)
423 zio_walk_parents(zio_t *cio)
425 zio_link_t *zl = cio->io_walk_link;
426 list_t *pl = &cio->io_parent_list;
428 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
429 cio->io_walk_link = zl;
434 ASSERT(zl->zl_child == cio);
435 return (zl->zl_parent);
439 zio_walk_children(zio_t *pio)
441 zio_link_t *zl = pio->io_walk_link;
442 list_t *cl = &pio->io_child_list;
444 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
445 pio->io_walk_link = zl;
450 ASSERT(zl->zl_parent == pio);
451 return (zl->zl_child);
455 zio_unique_parent(zio_t *cio)
457 zio_t *pio = zio_walk_parents(cio);
459 VERIFY(zio_walk_parents(cio) == NULL);
464 zio_add_child(zio_t *pio, zio_t *cio)
466 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
469 * Logical I/Os can have logical, gang, or vdev children.
470 * Gang I/Os can have gang or vdev children.
471 * Vdev I/Os can only have vdev children.
472 * The following ASSERT captures all of these constraints.
474 ASSERT(cio->io_child_type <= pio->io_child_type);
479 mutex_enter(&cio->io_lock);
480 mutex_enter(&pio->io_lock);
482 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
484 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
485 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
487 list_insert_head(&pio->io_child_list, zl);
488 list_insert_head(&cio->io_parent_list, zl);
490 pio->io_child_count++;
491 cio->io_parent_count++;
493 mutex_exit(&pio->io_lock);
494 mutex_exit(&cio->io_lock);
498 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
500 ASSERT(zl->zl_parent == pio);
501 ASSERT(zl->zl_child == cio);
503 mutex_enter(&cio->io_lock);
504 mutex_enter(&pio->io_lock);
506 list_remove(&pio->io_child_list, zl);
507 list_remove(&cio->io_parent_list, zl);
509 pio->io_child_count--;
510 cio->io_parent_count--;
512 mutex_exit(&pio->io_lock);
513 mutex_exit(&cio->io_lock);
515 kmem_cache_free(zio_link_cache, zl);
519 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
521 uint64_t *countp = &zio->io_children[child][wait];
522 boolean_t waiting = B_FALSE;
524 mutex_enter(&zio->io_lock);
525 ASSERT(zio->io_stall == NULL);
528 zio->io_stall = countp;
531 mutex_exit(&zio->io_lock);
537 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
539 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
540 int *errorp = &pio->io_child_error[zio->io_child_type];
542 mutex_enter(&pio->io_lock);
543 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
544 *errorp = zio_worst_error(*errorp, zio->io_error);
545 pio->io_reexecute |= zio->io_reexecute;
546 ASSERT3U(*countp, >, 0);
547 if (--*countp == 0 && pio->io_stall == countp) {
548 pio->io_stall = NULL;
549 mutex_exit(&pio->io_lock);
552 mutex_exit(&pio->io_lock);
557 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
559 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
560 zio->io_error = zio->io_child_error[c];
564 * ==========================================================================
565 * Create the various types of I/O (read, write, free, etc)
566 * ==========================================================================
569 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
570 void *data, uint64_t size, zio_done_func_t *done, void *private,
571 zio_type_t type, int priority, enum zio_flag flags,
572 vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
573 enum zio_stage stage, enum zio_stage pipeline)
577 ASSERT3U(type == ZIO_TYPE_FREE || size, <=, SPA_MAXBLOCKSIZE);
578 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
579 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
581 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
582 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
583 ASSERT(vd || stage == ZIO_STAGE_OPEN);
585 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
586 bzero(zio, sizeof (zio_t));
588 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
589 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
591 list_create(&zio->io_parent_list, sizeof (zio_link_t),
592 offsetof(zio_link_t, zl_parent_node));
593 list_create(&zio->io_child_list, sizeof (zio_link_t),
594 offsetof(zio_link_t, zl_child_node));
597 zio->io_child_type = ZIO_CHILD_VDEV;
598 else if (flags & ZIO_FLAG_GANG_CHILD)
599 zio->io_child_type = ZIO_CHILD_GANG;
600 else if (flags & ZIO_FLAG_DDT_CHILD)
601 zio->io_child_type = ZIO_CHILD_DDT;
603 zio->io_child_type = ZIO_CHILD_LOGICAL;
606 zio->io_bp = (blkptr_t *)bp;
607 zio->io_bp_copy = *bp;
608 zio->io_bp_orig = *bp;
609 if (type != ZIO_TYPE_WRITE ||
610 zio->io_child_type == ZIO_CHILD_DDT)
611 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
612 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
613 zio->io_logical = zio;
614 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
615 pipeline |= ZIO_GANG_STAGES;
621 zio->io_private = private;
623 zio->io_priority = priority;
625 zio->io_offset = offset;
626 zio->io_orig_data = zio->io_data = data;
627 zio->io_orig_size = zio->io_size = size;
628 zio->io_orig_flags = zio->io_flags = flags;
629 zio->io_orig_stage = zio->io_stage = stage;
630 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
632 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
633 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
636 zio->io_bookmark = *zb;
639 if (zio->io_logical == NULL)
640 zio->io_logical = pio->io_logical;
641 if (zio->io_child_type == ZIO_CHILD_GANG)
642 zio->io_gang_leader = pio->io_gang_leader;
643 zio_add_child(pio, zio);
650 zio_destroy(zio_t *zio)
652 list_destroy(&zio->io_parent_list);
653 list_destroy(&zio->io_child_list);
654 mutex_destroy(&zio->io_lock);
655 cv_destroy(&zio->io_cv);
656 kmem_cache_free(zio_cache, zio);
660 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
661 void *private, enum zio_flag flags)
665 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
666 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
667 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
673 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
675 return (zio_null(NULL, spa, NULL, done, private, flags));
679 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
680 void *data, uint64_t size, zio_done_func_t *done, void *private,
681 int priority, enum zio_flag flags, const zbookmark_t *zb)
685 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
686 data, size, done, private,
687 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
688 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
689 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
695 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
696 void *data, uint64_t size, const zio_prop_t *zp,
697 zio_done_func_t *ready, zio_done_func_t *done, void *private,
698 int priority, enum zio_flag flags, const zbookmark_t *zb)
702 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
703 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
704 zp->zp_compress >= ZIO_COMPRESS_OFF &&
705 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
706 DMU_OT_IS_VALID(zp->zp_type) &&
709 zp->zp_copies <= spa_max_replication(spa));
711 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
712 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
713 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
714 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
716 zio->io_ready = ready;
723 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
724 uint64_t size, zio_done_func_t *done, void *private, int priority,
725 enum zio_flag flags, zbookmark_t *zb)
729 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
730 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
731 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
737 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
739 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
740 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
741 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
742 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
745 * We must reset the io_prop to match the values that existed
746 * when the bp was first written by dmu_sync() keeping in mind
747 * that nopwrite and dedup are mutually exclusive.
749 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
750 zio->io_prop.zp_nopwrite = nopwrite;
751 zio->io_prop.zp_copies = copies;
752 zio->io_bp_override = bp;
756 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
758 metaslab_check_free(spa, bp);
759 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
763 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
764 uint64_t size, enum zio_flag flags)
768 dprintf_bp(bp, "freeing in txg %llu, pass %u",
769 (longlong_t)txg, spa->spa_sync_pass);
771 ASSERT(!BP_IS_HOLE(bp));
772 ASSERT(spa_syncing_txg(spa) == txg);
773 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
775 metaslab_check_free(spa, bp);
778 zio = zio_create(pio, spa, txg, bp, NULL, size,
779 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags,
780 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE);
786 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
787 zio_done_func_t *done, void *private, enum zio_flag flags)
792 * A claim is an allocation of a specific block. Claims are needed
793 * to support immediate writes in the intent log. The issue is that
794 * immediate writes contain committed data, but in a txg that was
795 * *not* committed. Upon opening the pool after an unclean shutdown,
796 * the intent log claims all blocks that contain immediate write data
797 * so that the SPA knows they're in use.
799 * All claims *must* be resolved in the first txg -- before the SPA
800 * starts allocating blocks -- so that nothing is allocated twice.
801 * If txg == 0 we just verify that the block is claimable.
803 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
804 ASSERT(txg == spa_first_txg(spa) || txg == 0);
805 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
807 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
808 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
809 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
815 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset,
816 uint64_t size, zio_done_func_t *done, void *private, int priority,
822 if (vd->vdev_children == 0) {
823 zio = zio_create(pio, spa, 0, NULL, NULL, size, done, private,
824 ZIO_TYPE_IOCTL, priority, flags, vd, offset, NULL,
825 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
829 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
831 for (c = 0; c < vd->vdev_children; c++)
832 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
833 offset, size, done, private, priority, flags));
840 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
841 void *data, int checksum, zio_done_func_t *done, void *private,
842 int priority, enum zio_flag flags, boolean_t labels)
846 ASSERT(vd->vdev_children == 0);
847 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
848 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
849 ASSERT3U(offset + size, <=, vd->vdev_psize);
851 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
852 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
853 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
855 zio->io_prop.zp_checksum = checksum;
861 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
862 void *data, int checksum, zio_done_func_t *done, void *private,
863 int priority, enum zio_flag flags, boolean_t labels)
867 ASSERT(vd->vdev_children == 0);
868 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
869 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
870 ASSERT3U(offset + size, <=, vd->vdev_psize);
872 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
873 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
874 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
876 zio->io_prop.zp_checksum = checksum;
878 if (zio_checksum_table[checksum].ci_eck) {
880 * zec checksums are necessarily destructive -- they modify
881 * the end of the write buffer to hold the verifier/checksum.
882 * Therefore, we must make a local copy in case the data is
883 * being written to multiple places in parallel.
885 void *wbuf = zio_buf_alloc(size);
886 bcopy(data, wbuf, size);
887 zio_push_transform(zio, wbuf, size, size, NULL);
894 * Create a child I/O to do some work for us.
897 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
898 void *data, uint64_t size, int type, int priority, enum zio_flag flags,
899 zio_done_func_t *done, void *private)
901 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
904 ASSERT(vd->vdev_parent ==
905 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
907 if (type == ZIO_TYPE_READ && bp != NULL) {
909 * If we have the bp, then the child should perform the
910 * checksum and the parent need not. This pushes error
911 * detection as close to the leaves as possible and
912 * eliminates redundant checksums in the interior nodes.
914 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
915 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
918 if (vd->vdev_children == 0)
919 offset += VDEV_LABEL_START_SIZE;
921 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
924 * If we've decided to do a repair, the write is not speculative --
925 * even if the original read was.
927 if (flags & ZIO_FLAG_IO_REPAIR)
928 flags &= ~ZIO_FLAG_SPECULATIVE;
930 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
931 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
932 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
938 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
939 int type, int priority, enum zio_flag flags,
940 zio_done_func_t *done, void *private)
944 ASSERT(vd->vdev_ops->vdev_op_leaf);
946 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
947 data, size, done, private, type, priority,
948 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY,
950 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
956 zio_flush(zio_t *zio, vdev_t *vd)
958 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0,
959 NULL, NULL, ZIO_PRIORITY_NOW,
960 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
964 zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size)
967 ASSERT(vd->vdev_ops->vdev_op_leaf);
969 return zio_ioctl(zio, spa, vd, DKIOCTRIM, offset, size,
970 NULL, NULL, ZIO_PRIORITY_TRIM,
971 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY);
975 zio_shrink(zio_t *zio, uint64_t size)
977 ASSERT(zio->io_executor == NULL);
978 ASSERT(zio->io_orig_size == zio->io_size);
979 ASSERT(size <= zio->io_size);
982 * We don't shrink for raidz because of problems with the
983 * reconstruction when reading back less than the block size.
984 * Note, BP_IS_RAIDZ() assumes no compression.
986 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
987 if (!BP_IS_RAIDZ(zio->io_bp))
988 zio->io_orig_size = zio->io_size = size;
992 * ==========================================================================
993 * Prepare to read and write logical blocks
994 * ==========================================================================
998 zio_read_bp_init(zio_t *zio)
1000 blkptr_t *bp = zio->io_bp;
1002 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1003 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1004 !(zio->io_flags & ZIO_FLAG_RAW)) {
1005 uint64_t psize = BP_GET_PSIZE(bp);
1006 void *cbuf = zio_buf_alloc(psize);
1008 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
1011 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1012 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1014 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1015 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1017 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1018 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1020 return (ZIO_PIPELINE_CONTINUE);
1024 zio_write_bp_init(zio_t *zio)
1026 spa_t *spa = zio->io_spa;
1027 zio_prop_t *zp = &zio->io_prop;
1028 enum zio_compress compress = zp->zp_compress;
1029 blkptr_t *bp = zio->io_bp;
1030 uint64_t lsize = zio->io_size;
1031 uint64_t psize = lsize;
1035 * If our children haven't all reached the ready stage,
1036 * wait for them and then repeat this pipeline stage.
1038 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1039 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1040 return (ZIO_PIPELINE_STOP);
1042 if (!IO_IS_ALLOCATING(zio))
1043 return (ZIO_PIPELINE_CONTINUE);
1045 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1047 if (zio->io_bp_override) {
1048 ASSERT(bp->blk_birth != zio->io_txg);
1049 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1051 *bp = *zio->io_bp_override;
1052 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1055 * If we've been overridden and nopwrite is set then
1056 * set the flag accordingly to indicate that a nopwrite
1057 * has already occurred.
1059 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1060 ASSERT(!zp->zp_dedup);
1061 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1062 return (ZIO_PIPELINE_CONTINUE);
1065 ASSERT(!zp->zp_nopwrite);
1067 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1068 return (ZIO_PIPELINE_CONTINUE);
1070 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1071 zp->zp_dedup_verify);
1073 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1074 BP_SET_DEDUP(bp, 1);
1075 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1076 return (ZIO_PIPELINE_CONTINUE);
1078 zio->io_bp_override = NULL;
1082 if (bp->blk_birth == zio->io_txg) {
1084 * We're rewriting an existing block, which means we're
1085 * working on behalf of spa_sync(). For spa_sync() to
1086 * converge, it must eventually be the case that we don't
1087 * have to allocate new blocks. But compression changes
1088 * the blocksize, which forces a reallocate, and makes
1089 * convergence take longer. Therefore, after the first
1090 * few passes, stop compressing to ensure convergence.
1092 pass = spa_sync_pass(spa);
1094 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1095 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1096 ASSERT(!BP_GET_DEDUP(bp));
1098 if (pass >= zfs_sync_pass_dont_compress)
1099 compress = ZIO_COMPRESS_OFF;
1101 /* Make sure someone doesn't change their mind on overwrites */
1102 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
1103 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1106 if (compress != ZIO_COMPRESS_OFF) {
1107 void *cbuf = zio_buf_alloc(lsize);
1108 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
1109 if (psize == 0 || psize == lsize) {
1110 compress = ZIO_COMPRESS_OFF;
1111 zio_buf_free(cbuf, lsize);
1113 ASSERT(psize < lsize);
1114 zio_push_transform(zio, cbuf, psize, lsize, NULL);
1119 * The final pass of spa_sync() must be all rewrites, but the first
1120 * few passes offer a trade-off: allocating blocks defers convergence,
1121 * but newly allocated blocks are sequential, so they can be written
1122 * to disk faster. Therefore, we allow the first few passes of
1123 * spa_sync() to allocate new blocks, but force rewrites after that.
1124 * There should only be a handful of blocks after pass 1 in any case.
1126 if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == psize &&
1127 pass >= zfs_sync_pass_rewrite) {
1129 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1130 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1131 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1134 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1138 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1140 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1141 BP_SET_LSIZE(bp, lsize);
1142 BP_SET_PSIZE(bp, psize);
1143 BP_SET_COMPRESS(bp, compress);
1144 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1145 BP_SET_TYPE(bp, zp->zp_type);
1146 BP_SET_LEVEL(bp, zp->zp_level);
1147 BP_SET_DEDUP(bp, zp->zp_dedup);
1148 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1150 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1151 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1152 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1154 if (zp->zp_nopwrite) {
1155 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1156 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1157 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1161 return (ZIO_PIPELINE_CONTINUE);
1165 zio_free_bp_init(zio_t *zio)
1167 blkptr_t *bp = zio->io_bp;
1169 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1170 if (BP_GET_DEDUP(bp))
1171 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1174 return (ZIO_PIPELINE_CONTINUE);
1178 * ==========================================================================
1179 * Execute the I/O pipeline
1180 * ==========================================================================
1184 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1186 spa_t *spa = zio->io_spa;
1187 zio_type_t t = zio->io_type;
1188 int flags = (cutinline ? TQ_FRONT : 0);
1190 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1193 * If we're a config writer or a probe, the normal issue and
1194 * interrupt threads may all be blocked waiting for the config lock.
1195 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1197 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1201 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1203 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1207 * If this is a high priority I/O, then use the high priority taskq if
1210 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1211 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1214 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1217 * NB: We are assuming that the zio can only be dispatched
1218 * to a single taskq at a time. It would be a grievous error
1219 * to dispatch the zio to another taskq at the same time.
1221 #if defined(illumos) || !defined(_KERNEL)
1222 ASSERT(zio->io_tqent.tqent_next == NULL);
1224 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
1226 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1227 flags, &zio->io_tqent);
1231 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1233 kthread_t *executor = zio->io_executor;
1234 spa_t *spa = zio->io_spa;
1236 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1237 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1239 for (i = 0; i < tqs->stqs_count; i++) {
1240 if (taskq_member(tqs->stqs_taskq[i], executor))
1249 zio_issue_async(zio_t *zio)
1251 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1253 return (ZIO_PIPELINE_STOP);
1257 zio_interrupt(zio_t *zio)
1259 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1263 * Execute the I/O pipeline until one of the following occurs:
1265 * (1) the I/O completes
1266 * (2) the pipeline stalls waiting for dependent child I/Os
1267 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1268 * (4) the I/O is delegated by vdev-level caching or aggregation
1269 * (5) the I/O is deferred due to vdev-level queueing
1270 * (6) the I/O is handed off to another thread.
1272 * In all cases, the pipeline stops whenever there's no CPU work; it never
1273 * burns a thread in cv_wait().
1275 * There's no locking on io_stage because there's no legitimate way
1276 * for multiple threads to be attempting to process the same I/O.
1278 static zio_pipe_stage_t *zio_pipeline[];
1281 zio_execute(zio_t *zio)
1283 zio->io_executor = curthread;
1285 while (zio->io_stage < ZIO_STAGE_DONE) {
1286 enum zio_stage pipeline = zio->io_pipeline;
1287 enum zio_stage stage = zio->io_stage;
1290 ASSERT(!MUTEX_HELD(&zio->io_lock));
1291 ASSERT(ISP2(stage));
1292 ASSERT(zio->io_stall == NULL);
1296 } while ((stage & pipeline) == 0);
1298 ASSERT(stage <= ZIO_STAGE_DONE);
1301 * If we are in interrupt context and this pipeline stage
1302 * will grab a config lock that is held across I/O,
1303 * or may wait for an I/O that needs an interrupt thread
1304 * to complete, issue async to avoid deadlock.
1306 * For VDEV_IO_START, we cut in line so that the io will
1307 * be sent to disk promptly.
1309 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1310 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1311 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1312 zio_requeue_io_start_cut_in_line : B_FALSE;
1313 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1317 zio->io_stage = stage;
1318 rv = zio_pipeline[highbit(stage) - 1](zio);
1320 if (rv == ZIO_PIPELINE_STOP)
1323 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1328 * ==========================================================================
1329 * Initiate I/O, either sync or async
1330 * ==========================================================================
1333 zio_wait(zio_t *zio)
1337 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1338 ASSERT(zio->io_executor == NULL);
1340 zio->io_waiter = curthread;
1344 mutex_enter(&zio->io_lock);
1345 while (zio->io_executor != NULL)
1346 cv_wait(&zio->io_cv, &zio->io_lock);
1347 mutex_exit(&zio->io_lock);
1349 error = zio->io_error;
1356 zio_nowait(zio_t *zio)
1358 ASSERT(zio->io_executor == NULL);
1360 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1361 zio_unique_parent(zio) == NULL) {
1363 * This is a logical async I/O with no parent to wait for it.
1364 * We add it to the spa_async_root_zio "Godfather" I/O which
1365 * will ensure they complete prior to unloading the pool.
1367 spa_t *spa = zio->io_spa;
1369 zio_add_child(spa->spa_async_zio_root, zio);
1376 * ==========================================================================
1377 * Reexecute or suspend/resume failed I/O
1378 * ==========================================================================
1382 zio_reexecute(zio_t *pio)
1384 zio_t *cio, *cio_next;
1386 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1387 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1388 ASSERT(pio->io_gang_leader == NULL);
1389 ASSERT(pio->io_gang_tree == NULL);
1391 pio->io_flags = pio->io_orig_flags;
1392 pio->io_stage = pio->io_orig_stage;
1393 pio->io_pipeline = pio->io_orig_pipeline;
1394 pio->io_reexecute = 0;
1395 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1397 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1398 pio->io_state[w] = 0;
1399 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1400 pio->io_child_error[c] = 0;
1402 if (IO_IS_ALLOCATING(pio))
1403 BP_ZERO(pio->io_bp);
1406 * As we reexecute pio's children, new children could be created.
1407 * New children go to the head of pio's io_child_list, however,
1408 * so we will (correctly) not reexecute them. The key is that
1409 * the remainder of pio's io_child_list, from 'cio_next' onward,
1410 * cannot be affected by any side effects of reexecuting 'cio'.
1412 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1413 cio_next = zio_walk_children(pio);
1414 mutex_enter(&pio->io_lock);
1415 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1416 pio->io_children[cio->io_child_type][w]++;
1417 mutex_exit(&pio->io_lock);
1422 * Now that all children have been reexecuted, execute the parent.
1423 * We don't reexecute "The Godfather" I/O here as it's the
1424 * responsibility of the caller to wait on him.
1426 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1431 zio_suspend(spa_t *spa, zio_t *zio)
1433 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1434 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1435 "failure and the failure mode property for this pool "
1436 "is set to panic.", spa_name(spa));
1438 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1440 mutex_enter(&spa->spa_suspend_lock);
1442 if (spa->spa_suspend_zio_root == NULL)
1443 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1444 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1445 ZIO_FLAG_GODFATHER);
1447 spa->spa_suspended = B_TRUE;
1450 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1451 ASSERT(zio != spa->spa_suspend_zio_root);
1452 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1453 ASSERT(zio_unique_parent(zio) == NULL);
1454 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1455 zio_add_child(spa->spa_suspend_zio_root, zio);
1458 mutex_exit(&spa->spa_suspend_lock);
1462 zio_resume(spa_t *spa)
1467 * Reexecute all previously suspended i/o.
1469 mutex_enter(&spa->spa_suspend_lock);
1470 spa->spa_suspended = B_FALSE;
1471 cv_broadcast(&spa->spa_suspend_cv);
1472 pio = spa->spa_suspend_zio_root;
1473 spa->spa_suspend_zio_root = NULL;
1474 mutex_exit(&spa->spa_suspend_lock);
1480 return (zio_wait(pio));
1484 zio_resume_wait(spa_t *spa)
1486 mutex_enter(&spa->spa_suspend_lock);
1487 while (spa_suspended(spa))
1488 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1489 mutex_exit(&spa->spa_suspend_lock);
1493 * ==========================================================================
1496 * A gang block is a collection of small blocks that looks to the DMU
1497 * like one large block. When zio_dva_allocate() cannot find a block
1498 * of the requested size, due to either severe fragmentation or the pool
1499 * being nearly full, it calls zio_write_gang_block() to construct the
1500 * block from smaller fragments.
1502 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1503 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1504 * an indirect block: it's an array of block pointers. It consumes
1505 * only one sector and hence is allocatable regardless of fragmentation.
1506 * The gang header's bps point to its gang members, which hold the data.
1508 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1509 * as the verifier to ensure uniqueness of the SHA256 checksum.
1510 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1511 * not the gang header. This ensures that data block signatures (needed for
1512 * deduplication) are independent of how the block is physically stored.
1514 * Gang blocks can be nested: a gang member may itself be a gang block.
1515 * Thus every gang block is a tree in which root and all interior nodes are
1516 * gang headers, and the leaves are normal blocks that contain user data.
1517 * The root of the gang tree is called the gang leader.
1519 * To perform any operation (read, rewrite, free, claim) on a gang block,
1520 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1521 * in the io_gang_tree field of the original logical i/o by recursively
1522 * reading the gang leader and all gang headers below it. This yields
1523 * an in-core tree containing the contents of every gang header and the
1524 * bps for every constituent of the gang block.
1526 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1527 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1528 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1529 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1530 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1531 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1532 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1533 * of the gang header plus zio_checksum_compute() of the data to update the
1534 * gang header's blk_cksum as described above.
1536 * The two-phase assemble/issue model solves the problem of partial failure --
1537 * what if you'd freed part of a gang block but then couldn't read the
1538 * gang header for another part? Assembling the entire gang tree first
1539 * ensures that all the necessary gang header I/O has succeeded before
1540 * starting the actual work of free, claim, or write. Once the gang tree
1541 * is assembled, free and claim are in-memory operations that cannot fail.
1543 * In the event that a gang write fails, zio_dva_unallocate() walks the
1544 * gang tree to immediately free (i.e. insert back into the space map)
1545 * everything we've allocated. This ensures that we don't get ENOSPC
1546 * errors during repeated suspend/resume cycles due to a flaky device.
1548 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1549 * the gang tree, we won't modify the block, so we can safely defer the free
1550 * (knowing that the block is still intact). If we *can* assemble the gang
1551 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1552 * each constituent bp and we can allocate a new block on the next sync pass.
1554 * In all cases, the gang tree allows complete recovery from partial failure.
1555 * ==========================================================================
1559 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1564 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1565 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1566 &pio->io_bookmark));
1570 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1575 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1576 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1577 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1579 * As we rewrite each gang header, the pipeline will compute
1580 * a new gang block header checksum for it; but no one will
1581 * compute a new data checksum, so we do that here. The one
1582 * exception is the gang leader: the pipeline already computed
1583 * its data checksum because that stage precedes gang assembly.
1584 * (Presently, nothing actually uses interior data checksums;
1585 * this is just good hygiene.)
1587 if (gn != pio->io_gang_leader->io_gang_tree) {
1588 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1589 data, BP_GET_PSIZE(bp));
1592 * If we are here to damage data for testing purposes,
1593 * leave the GBH alone so that we can detect the damage.
1595 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1596 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1598 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1599 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1600 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1608 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1610 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1611 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp),
1612 ZIO_GANG_CHILD_FLAGS(pio)));
1617 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1619 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1620 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1623 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1632 static void zio_gang_tree_assemble_done(zio_t *zio);
1634 static zio_gang_node_t *
1635 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1637 zio_gang_node_t *gn;
1639 ASSERT(*gnpp == NULL);
1641 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1642 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1649 zio_gang_node_free(zio_gang_node_t **gnpp)
1651 zio_gang_node_t *gn = *gnpp;
1653 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1654 ASSERT(gn->gn_child[g] == NULL);
1656 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1657 kmem_free(gn, sizeof (*gn));
1662 zio_gang_tree_free(zio_gang_node_t **gnpp)
1664 zio_gang_node_t *gn = *gnpp;
1669 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1670 zio_gang_tree_free(&gn->gn_child[g]);
1672 zio_gang_node_free(gnpp);
1676 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1678 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1680 ASSERT(gio->io_gang_leader == gio);
1681 ASSERT(BP_IS_GANG(bp));
1683 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1684 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1685 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1689 zio_gang_tree_assemble_done(zio_t *zio)
1691 zio_t *gio = zio->io_gang_leader;
1692 zio_gang_node_t *gn = zio->io_private;
1693 blkptr_t *bp = zio->io_bp;
1695 ASSERT(gio == zio_unique_parent(zio));
1696 ASSERT(zio->io_child_count == 0);
1701 if (BP_SHOULD_BYTESWAP(bp))
1702 byteswap_uint64_array(zio->io_data, zio->io_size);
1704 ASSERT(zio->io_data == gn->gn_gbh);
1705 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1706 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1708 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1709 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1710 if (!BP_IS_GANG(gbp))
1712 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1717 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1719 zio_t *gio = pio->io_gang_leader;
1722 ASSERT(BP_IS_GANG(bp) == !!gn);
1723 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1724 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1727 * If you're a gang header, your data is in gn->gn_gbh.
1728 * If you're a gang member, your data is in 'data' and gn == NULL.
1730 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1733 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1735 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1736 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1737 if (BP_IS_HOLE(gbp))
1739 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1740 data = (char *)data + BP_GET_PSIZE(gbp);
1744 if (gn == gio->io_gang_tree && gio->io_data != NULL)
1745 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1752 zio_gang_assemble(zio_t *zio)
1754 blkptr_t *bp = zio->io_bp;
1756 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1757 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1759 zio->io_gang_leader = zio;
1761 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1763 return (ZIO_PIPELINE_CONTINUE);
1767 zio_gang_issue(zio_t *zio)
1769 blkptr_t *bp = zio->io_bp;
1771 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1772 return (ZIO_PIPELINE_STOP);
1774 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1775 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1777 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1778 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1780 zio_gang_tree_free(&zio->io_gang_tree);
1782 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1784 return (ZIO_PIPELINE_CONTINUE);
1788 zio_write_gang_member_ready(zio_t *zio)
1790 zio_t *pio = zio_unique_parent(zio);
1791 zio_t *gio = zio->io_gang_leader;
1792 dva_t *cdva = zio->io_bp->blk_dva;
1793 dva_t *pdva = pio->io_bp->blk_dva;
1796 if (BP_IS_HOLE(zio->io_bp))
1799 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1801 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1802 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1803 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1804 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1805 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1807 mutex_enter(&pio->io_lock);
1808 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1809 ASSERT(DVA_GET_GANG(&pdva[d]));
1810 asize = DVA_GET_ASIZE(&pdva[d]);
1811 asize += DVA_GET_ASIZE(&cdva[d]);
1812 DVA_SET_ASIZE(&pdva[d], asize);
1814 mutex_exit(&pio->io_lock);
1818 zio_write_gang_block(zio_t *pio)
1820 spa_t *spa = pio->io_spa;
1821 blkptr_t *bp = pio->io_bp;
1822 zio_t *gio = pio->io_gang_leader;
1824 zio_gang_node_t *gn, **gnpp;
1825 zio_gbh_phys_t *gbh;
1826 uint64_t txg = pio->io_txg;
1827 uint64_t resid = pio->io_size;
1829 int copies = gio->io_prop.zp_copies;
1830 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1834 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1835 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1836 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1838 pio->io_error = error;
1839 return (ZIO_PIPELINE_CONTINUE);
1843 gnpp = &gio->io_gang_tree;
1845 gnpp = pio->io_private;
1846 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1849 gn = zio_gang_node_alloc(gnpp);
1851 bzero(gbh, SPA_GANGBLOCKSIZE);
1854 * Create the gang header.
1856 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1857 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1860 * Create and nowait the gang children.
1862 for (int g = 0; resid != 0; resid -= lsize, g++) {
1863 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1865 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1867 zp.zp_checksum = gio->io_prop.zp_checksum;
1868 zp.zp_compress = ZIO_COMPRESS_OFF;
1869 zp.zp_type = DMU_OT_NONE;
1871 zp.zp_copies = gio->io_prop.zp_copies;
1872 zp.zp_dedup = B_FALSE;
1873 zp.zp_dedup_verify = B_FALSE;
1874 zp.zp_nopwrite = B_FALSE;
1876 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1877 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1878 zio_write_gang_member_ready, NULL, &gn->gn_child[g],
1879 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1880 &pio->io_bookmark));
1884 * Set pio's pipeline to just wait for zio to finish.
1886 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1890 return (ZIO_PIPELINE_CONTINUE);
1894 * The zio_nop_write stage in the pipeline determines if allocating
1895 * a new bp is necessary. By leveraging a cryptographically secure checksum,
1896 * such as SHA256, we can compare the checksums of the new data and the old
1897 * to determine if allocating a new block is required. The nopwrite
1898 * feature can handle writes in either syncing or open context (i.e. zil
1899 * writes) and as a result is mutually exclusive with dedup.
1902 zio_nop_write(zio_t *zio)
1904 blkptr_t *bp = zio->io_bp;
1905 blkptr_t *bp_orig = &zio->io_bp_orig;
1906 zio_prop_t *zp = &zio->io_prop;
1908 ASSERT(BP_GET_LEVEL(bp) == 0);
1909 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1910 ASSERT(zp->zp_nopwrite);
1911 ASSERT(!zp->zp_dedup);
1912 ASSERT(zio->io_bp_override == NULL);
1913 ASSERT(IO_IS_ALLOCATING(zio));
1916 * Check to see if the original bp and the new bp have matching
1917 * characteristics (i.e. same checksum, compression algorithms, etc).
1918 * If they don't then just continue with the pipeline which will
1919 * allocate a new bp.
1921 if (BP_IS_HOLE(bp_orig) ||
1922 !zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup ||
1923 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
1924 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
1925 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
1926 zp->zp_copies != BP_GET_NDVAS(bp_orig))
1927 return (ZIO_PIPELINE_CONTINUE);
1930 * If the checksums match then reset the pipeline so that we
1931 * avoid allocating a new bp and issuing any I/O.
1933 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
1934 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup);
1935 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
1936 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
1937 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
1938 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
1939 sizeof (uint64_t)) == 0);
1942 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1943 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1946 return (ZIO_PIPELINE_CONTINUE);
1950 * ==========================================================================
1952 * ==========================================================================
1955 zio_ddt_child_read_done(zio_t *zio)
1957 blkptr_t *bp = zio->io_bp;
1958 ddt_entry_t *dde = zio->io_private;
1960 zio_t *pio = zio_unique_parent(zio);
1962 mutex_enter(&pio->io_lock);
1963 ddp = ddt_phys_select(dde, bp);
1964 if (zio->io_error == 0)
1965 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1966 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1967 dde->dde_repair_data = zio->io_data;
1969 zio_buf_free(zio->io_data, zio->io_size);
1970 mutex_exit(&pio->io_lock);
1974 zio_ddt_read_start(zio_t *zio)
1976 blkptr_t *bp = zio->io_bp;
1978 ASSERT(BP_GET_DEDUP(bp));
1979 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
1980 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1982 if (zio->io_child_error[ZIO_CHILD_DDT]) {
1983 ddt_t *ddt = ddt_select(zio->io_spa, bp);
1984 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
1985 ddt_phys_t *ddp = dde->dde_phys;
1986 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
1989 ASSERT(zio->io_vsd == NULL);
1992 if (ddp_self == NULL)
1993 return (ZIO_PIPELINE_CONTINUE);
1995 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1996 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
1998 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2000 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2001 zio_buf_alloc(zio->io_size), zio->io_size,
2002 zio_ddt_child_read_done, dde, zio->io_priority,
2003 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
2004 &zio->io_bookmark));
2006 return (ZIO_PIPELINE_CONTINUE);
2009 zio_nowait(zio_read(zio, zio->io_spa, bp,
2010 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
2011 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2013 return (ZIO_PIPELINE_CONTINUE);
2017 zio_ddt_read_done(zio_t *zio)
2019 blkptr_t *bp = zio->io_bp;
2021 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2022 return (ZIO_PIPELINE_STOP);
2024 ASSERT(BP_GET_DEDUP(bp));
2025 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2026 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2028 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2029 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2030 ddt_entry_t *dde = zio->io_vsd;
2032 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2033 return (ZIO_PIPELINE_CONTINUE);
2036 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2037 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2038 return (ZIO_PIPELINE_STOP);
2040 if (dde->dde_repair_data != NULL) {
2041 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2042 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2044 ddt_repair_done(ddt, dde);
2048 ASSERT(zio->io_vsd == NULL);
2050 return (ZIO_PIPELINE_CONTINUE);
2054 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2056 spa_t *spa = zio->io_spa;
2059 * Note: we compare the original data, not the transformed data,
2060 * because when zio->io_bp is an override bp, we will not have
2061 * pushed the I/O transforms. That's an important optimization
2062 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2064 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2065 zio_t *lio = dde->dde_lead_zio[p];
2068 return (lio->io_orig_size != zio->io_orig_size ||
2069 bcmp(zio->io_orig_data, lio->io_orig_data,
2070 zio->io_orig_size) != 0);
2074 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2075 ddt_phys_t *ddp = &dde->dde_phys[p];
2077 if (ddp->ddp_phys_birth != 0) {
2078 arc_buf_t *abuf = NULL;
2079 uint32_t aflags = ARC_WAIT;
2080 blkptr_t blk = *zio->io_bp;
2083 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2087 error = arc_read(NULL, spa, &blk,
2088 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2089 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2090 &aflags, &zio->io_bookmark);
2093 if (arc_buf_size(abuf) != zio->io_orig_size ||
2094 bcmp(abuf->b_data, zio->io_orig_data,
2095 zio->io_orig_size) != 0)
2096 error = SET_ERROR(EEXIST);
2097 VERIFY(arc_buf_remove_ref(abuf, &abuf));
2101 return (error != 0);
2109 zio_ddt_child_write_ready(zio_t *zio)
2111 int p = zio->io_prop.zp_copies;
2112 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2113 ddt_entry_t *dde = zio->io_private;
2114 ddt_phys_t *ddp = &dde->dde_phys[p];
2122 ASSERT(dde->dde_lead_zio[p] == zio);
2124 ddt_phys_fill(ddp, zio->io_bp);
2126 while ((pio = zio_walk_parents(zio)) != NULL)
2127 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2133 zio_ddt_child_write_done(zio_t *zio)
2135 int p = zio->io_prop.zp_copies;
2136 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2137 ddt_entry_t *dde = zio->io_private;
2138 ddt_phys_t *ddp = &dde->dde_phys[p];
2142 ASSERT(ddp->ddp_refcnt == 0);
2143 ASSERT(dde->dde_lead_zio[p] == zio);
2144 dde->dde_lead_zio[p] = NULL;
2146 if (zio->io_error == 0) {
2147 while (zio_walk_parents(zio) != NULL)
2148 ddt_phys_addref(ddp);
2150 ddt_phys_clear(ddp);
2157 zio_ddt_ditto_write_done(zio_t *zio)
2159 int p = DDT_PHYS_DITTO;
2160 zio_prop_t *zp = &zio->io_prop;
2161 blkptr_t *bp = zio->io_bp;
2162 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2163 ddt_entry_t *dde = zio->io_private;
2164 ddt_phys_t *ddp = &dde->dde_phys[p];
2165 ddt_key_t *ddk = &dde->dde_key;
2169 ASSERT(ddp->ddp_refcnt == 0);
2170 ASSERT(dde->dde_lead_zio[p] == zio);
2171 dde->dde_lead_zio[p] = NULL;
2173 if (zio->io_error == 0) {
2174 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2175 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2176 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2177 if (ddp->ddp_phys_birth != 0)
2178 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2179 ddt_phys_fill(ddp, bp);
2186 zio_ddt_write(zio_t *zio)
2188 spa_t *spa = zio->io_spa;
2189 blkptr_t *bp = zio->io_bp;
2190 uint64_t txg = zio->io_txg;
2191 zio_prop_t *zp = &zio->io_prop;
2192 int p = zp->zp_copies;
2196 ddt_t *ddt = ddt_select(spa, bp);
2200 ASSERT(BP_GET_DEDUP(bp));
2201 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2202 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2205 dde = ddt_lookup(ddt, bp, B_TRUE);
2206 ddp = &dde->dde_phys[p];
2208 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2210 * If we're using a weak checksum, upgrade to a strong checksum
2211 * and try again. If we're already using a strong checksum,
2212 * we can't resolve it, so just convert to an ordinary write.
2213 * (And automatically e-mail a paper to Nature?)
2215 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2216 zp->zp_checksum = spa_dedup_checksum(spa);
2217 zio_pop_transforms(zio);
2218 zio->io_stage = ZIO_STAGE_OPEN;
2221 zp->zp_dedup = B_FALSE;
2223 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2225 return (ZIO_PIPELINE_CONTINUE);
2228 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2229 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2231 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2232 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2233 zio_prop_t czp = *zp;
2235 czp.zp_copies = ditto_copies;
2238 * If we arrived here with an override bp, we won't have run
2239 * the transform stack, so we won't have the data we need to
2240 * generate a child i/o. So, toss the override bp and restart.
2241 * This is safe, because using the override bp is just an
2242 * optimization; and it's rare, so the cost doesn't matter.
2244 if (zio->io_bp_override) {
2245 zio_pop_transforms(zio);
2246 zio->io_stage = ZIO_STAGE_OPEN;
2247 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2248 zio->io_bp_override = NULL;
2251 return (ZIO_PIPELINE_CONTINUE);
2254 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2255 zio->io_orig_size, &czp, NULL,
2256 zio_ddt_ditto_write_done, dde, zio->io_priority,
2257 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2259 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2260 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2263 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2264 if (ddp->ddp_phys_birth != 0)
2265 ddt_bp_fill(ddp, bp, txg);
2266 if (dde->dde_lead_zio[p] != NULL)
2267 zio_add_child(zio, dde->dde_lead_zio[p]);
2269 ddt_phys_addref(ddp);
2270 } else if (zio->io_bp_override) {
2271 ASSERT(bp->blk_birth == txg);
2272 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2273 ddt_phys_fill(ddp, bp);
2274 ddt_phys_addref(ddp);
2276 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2277 zio->io_orig_size, zp, zio_ddt_child_write_ready,
2278 zio_ddt_child_write_done, dde, zio->io_priority,
2279 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2281 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2282 dde->dde_lead_zio[p] = cio;
2292 return (ZIO_PIPELINE_CONTINUE);
2295 ddt_entry_t *freedde; /* for debugging */
2298 zio_ddt_free(zio_t *zio)
2300 spa_t *spa = zio->io_spa;
2301 blkptr_t *bp = zio->io_bp;
2302 ddt_t *ddt = ddt_select(spa, bp);
2306 ASSERT(BP_GET_DEDUP(bp));
2307 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2310 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2311 ddp = ddt_phys_select(dde, bp);
2312 ddt_phys_decref(ddp);
2315 return (ZIO_PIPELINE_CONTINUE);
2319 * ==========================================================================
2320 * Allocate and free blocks
2321 * ==========================================================================
2324 zio_dva_allocate(zio_t *zio)
2326 spa_t *spa = zio->io_spa;
2327 metaslab_class_t *mc = spa_normal_class(spa);
2328 blkptr_t *bp = zio->io_bp;
2332 if (zio->io_gang_leader == NULL) {
2333 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2334 zio->io_gang_leader = zio;
2337 ASSERT(BP_IS_HOLE(bp));
2338 ASSERT0(BP_GET_NDVAS(bp));
2339 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2340 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2341 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2344 * The dump device does not support gang blocks so allocation on
2345 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2346 * the "fast" gang feature.
2348 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2349 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2350 METASLAB_GANG_CHILD : 0;
2351 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2352 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2355 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2356 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2358 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2359 return (zio_write_gang_block(zio));
2360 zio->io_error = error;
2363 return (ZIO_PIPELINE_CONTINUE);
2367 zio_dva_free(zio_t *zio)
2369 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2371 return (ZIO_PIPELINE_CONTINUE);
2375 zio_dva_claim(zio_t *zio)
2379 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2381 zio->io_error = error;
2383 return (ZIO_PIPELINE_CONTINUE);
2387 * Undo an allocation. This is used by zio_done() when an I/O fails
2388 * and we want to give back the block we just allocated.
2389 * This handles both normal blocks and gang blocks.
2392 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2394 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2395 ASSERT(zio->io_bp_override == NULL);
2397 if (!BP_IS_HOLE(bp))
2398 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2401 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2402 zio_dva_unallocate(zio, gn->gn_child[g],
2403 &gn->gn_gbh->zg_blkptr[g]);
2409 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2412 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2413 uint64_t size, boolean_t use_slog)
2417 ASSERT(txg > spa_syncing_txg(spa));
2420 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2421 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2422 * when allocating them.
2425 error = metaslab_alloc(spa, spa_log_class(spa), size,
2426 new_bp, 1, txg, old_bp,
2427 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2431 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2432 new_bp, 1, txg, old_bp,
2433 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2437 BP_SET_LSIZE(new_bp, size);
2438 BP_SET_PSIZE(new_bp, size);
2439 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2440 BP_SET_CHECKSUM(new_bp,
2441 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2442 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2443 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2444 BP_SET_LEVEL(new_bp, 0);
2445 BP_SET_DEDUP(new_bp, 0);
2446 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2453 * Free an intent log block.
2456 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2458 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2459 ASSERT(!BP_IS_GANG(bp));
2461 zio_free(spa, txg, bp);
2465 * ==========================================================================
2466 * Read, write and delete to physical devices
2467 * ==========================================================================
2470 zio_vdev_io_start(zio_t *zio)
2472 vdev_t *vd = zio->io_vd;
2474 spa_t *spa = zio->io_spa;
2476 ASSERT(zio->io_error == 0);
2477 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2480 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2481 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2484 * The mirror_ops handle multiple DVAs in a single BP.
2486 return (vdev_mirror_ops.vdev_op_io_start(zio));
2489 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE) {
2490 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg);
2491 return (ZIO_PIPELINE_CONTINUE);
2495 * We keep track of time-sensitive I/Os so that the scan thread
2496 * can quickly react to certain workloads. In particular, we care
2497 * about non-scrubbing, top-level reads and writes with the following
2499 * - synchronous writes of user data to non-slog devices
2500 * - any reads of user data
2501 * When these conditions are met, adjust the timestamp of spa_last_io
2502 * which allows the scan thread to adjust its workload accordingly.
2504 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2505 vd == vd->vdev_top && !vd->vdev_islog &&
2506 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2507 zio->io_txg != spa_syncing_txg(spa)) {
2508 uint64_t old = spa->spa_last_io;
2509 uint64_t new = ddi_get_lbolt64();
2511 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2514 align = 1ULL << vd->vdev_top->vdev_ashift;
2516 if (P2PHASE(zio->io_size, align) != 0) {
2517 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2519 if (zio->io_type == ZIO_TYPE_READ ||
2520 zio->io_type == ZIO_TYPE_WRITE)
2521 abuf = zio_buf_alloc(asize);
2522 ASSERT(vd == vd->vdev_top);
2523 if (zio->io_type == ZIO_TYPE_WRITE) {
2524 bcopy(zio->io_data, abuf, zio->io_size);
2525 bzero(abuf + zio->io_size, asize - zio->io_size);
2527 zio_push_transform(zio, abuf, asize, abuf ? asize : 0,
2531 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2532 ASSERT(P2PHASE(zio->io_size, align) == 0);
2533 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa));
2536 * If this is a repair I/O, and there's no self-healing involved --
2537 * that is, we're just resilvering what we expect to resilver --
2538 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2539 * This prevents spurious resilvering with nested replication.
2540 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2541 * A is out of date, we'll read from C+D, then use the data to
2542 * resilver A+B -- but we don't actually want to resilver B, just A.
2543 * The top-level mirror has no way to know this, so instead we just
2544 * discard unnecessary repairs as we work our way down the vdev tree.
2545 * The same logic applies to any form of nested replication:
2546 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2548 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2549 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2550 zio->io_txg != 0 && /* not a delegated i/o */
2551 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2552 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2553 zio_vdev_io_bypass(zio);
2554 return (ZIO_PIPELINE_CONTINUE);
2557 if (vd->vdev_ops->vdev_op_leaf &&
2558 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2560 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0)
2561 return (ZIO_PIPELINE_CONTINUE);
2563 if ((zio = vdev_queue_io(zio)) == NULL)
2564 return (ZIO_PIPELINE_STOP);
2566 if (!vdev_accessible(vd, zio)) {
2567 zio->io_error = SET_ERROR(ENXIO);
2569 return (ZIO_PIPELINE_STOP);
2574 * Note that we ignore repair writes for TRIM because they can conflict
2575 * with normal writes. This isn't an issue because, by definition, we
2576 * only repair blocks that aren't freed.
2578 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_WRITE &&
2579 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2580 if (!trim_map_write_start(zio))
2581 return (ZIO_PIPELINE_STOP);
2584 return (vd->vdev_ops->vdev_op_io_start(zio));
2588 zio_vdev_io_done(zio_t *zio)
2590 vdev_t *vd = zio->io_vd;
2591 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2592 boolean_t unexpected_error = B_FALSE;
2594 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2595 return (ZIO_PIPELINE_STOP);
2597 ASSERT(zio->io_type == ZIO_TYPE_READ ||
2598 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE);
2600 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2601 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2603 if (zio->io_type == ZIO_TYPE_WRITE &&
2604 !(zio->io_flags & ZIO_FLAG_IO_REPAIR))
2605 trim_map_write_done(zio);
2607 vdev_queue_io_done(zio);
2609 if (zio->io_type == ZIO_TYPE_WRITE)
2610 vdev_cache_write(zio);
2612 if (zio_injection_enabled && zio->io_error == 0)
2613 zio->io_error = zio_handle_device_injection(vd,
2616 if (zio_injection_enabled && zio->io_error == 0)
2617 zio->io_error = zio_handle_label_injection(zio, EIO);
2619 if (zio->io_error) {
2620 if (!vdev_accessible(vd, zio)) {
2621 zio->io_error = SET_ERROR(ENXIO);
2623 unexpected_error = B_TRUE;
2628 ops->vdev_op_io_done(zio);
2630 if (unexpected_error)
2631 VERIFY(vdev_probe(vd, zio) == NULL);
2633 return (ZIO_PIPELINE_CONTINUE);
2637 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2638 * disk, and use that to finish the checksum ereport later.
2641 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2642 const void *good_buf)
2644 /* no processing needed */
2645 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2650 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2652 void *buf = zio_buf_alloc(zio->io_size);
2654 bcopy(zio->io_data, buf, zio->io_size);
2656 zcr->zcr_cbinfo = zio->io_size;
2657 zcr->zcr_cbdata = buf;
2658 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2659 zcr->zcr_free = zio_buf_free;
2663 zio_vdev_io_assess(zio_t *zio)
2665 vdev_t *vd = zio->io_vd;
2667 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2668 return (ZIO_PIPELINE_STOP);
2670 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2671 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2673 if (zio->io_vsd != NULL) {
2674 zio->io_vsd_ops->vsd_free(zio);
2678 if (zio_injection_enabled && zio->io_error == 0)
2679 zio->io_error = zio_handle_fault_injection(zio, EIO);
2681 if (zio->io_type == ZIO_TYPE_IOCTL && zio->io_cmd == DKIOCTRIM)
2682 switch (zio->io_error) {
2684 ZIO_TRIM_STAT_INCR(bytes, zio->io_size);
2685 ZIO_TRIM_STAT_BUMP(success);
2688 ZIO_TRIM_STAT_BUMP(unsupported);
2691 ZIO_TRIM_STAT_BUMP(failed);
2696 * If the I/O failed, determine whether we should attempt to retry it.
2698 * On retry, we cut in line in the issue queue, since we don't want
2699 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2701 if (zio->io_error && vd == NULL &&
2702 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2703 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2704 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2706 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2707 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2708 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2709 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2710 zio_requeue_io_start_cut_in_line);
2711 return (ZIO_PIPELINE_STOP);
2715 * If we got an error on a leaf device, convert it to ENXIO
2716 * if the device is not accessible at all.
2718 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2719 !vdev_accessible(vd, zio))
2720 zio->io_error = SET_ERROR(ENXIO);
2723 * If we can't write to an interior vdev (mirror or RAID-Z),
2724 * set vdev_cant_write so that we stop trying to allocate from it.
2726 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2727 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
2728 vd->vdev_cant_write = B_TRUE;
2732 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2734 return (ZIO_PIPELINE_CONTINUE);
2738 zio_vdev_io_reissue(zio_t *zio)
2740 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2741 ASSERT(zio->io_error == 0);
2743 zio->io_stage >>= 1;
2747 zio_vdev_io_redone(zio_t *zio)
2749 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2751 zio->io_stage >>= 1;
2755 zio_vdev_io_bypass(zio_t *zio)
2757 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2758 ASSERT(zio->io_error == 0);
2760 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2761 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2765 * ==========================================================================
2766 * Generate and verify checksums
2767 * ==========================================================================
2770 zio_checksum_generate(zio_t *zio)
2772 blkptr_t *bp = zio->io_bp;
2773 enum zio_checksum checksum;
2777 * This is zio_write_phys().
2778 * We're either generating a label checksum, or none at all.
2780 checksum = zio->io_prop.zp_checksum;
2782 if (checksum == ZIO_CHECKSUM_OFF)
2783 return (ZIO_PIPELINE_CONTINUE);
2785 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2787 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2788 ASSERT(!IO_IS_ALLOCATING(zio));
2789 checksum = ZIO_CHECKSUM_GANG_HEADER;
2791 checksum = BP_GET_CHECKSUM(bp);
2795 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2797 return (ZIO_PIPELINE_CONTINUE);
2801 zio_checksum_verify(zio_t *zio)
2803 zio_bad_cksum_t info;
2804 blkptr_t *bp = zio->io_bp;
2807 ASSERT(zio->io_vd != NULL);
2811 * This is zio_read_phys().
2812 * We're either verifying a label checksum, or nothing at all.
2814 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2815 return (ZIO_PIPELINE_CONTINUE);
2817 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2820 if ((error = zio_checksum_error(zio, &info)) != 0) {
2821 zio->io_error = error;
2822 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2823 zfs_ereport_start_checksum(zio->io_spa,
2824 zio->io_vd, zio, zio->io_offset,
2825 zio->io_size, NULL, &info);
2829 return (ZIO_PIPELINE_CONTINUE);
2833 * Called by RAID-Z to ensure we don't compute the checksum twice.
2836 zio_checksum_verified(zio_t *zio)
2838 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2842 * ==========================================================================
2843 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2844 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2845 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2846 * indicate errors that are specific to one I/O, and most likely permanent.
2847 * Any other error is presumed to be worse because we weren't expecting it.
2848 * ==========================================================================
2851 zio_worst_error(int e1, int e2)
2853 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2856 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2857 if (e1 == zio_error_rank[r1])
2860 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2861 if (e2 == zio_error_rank[r2])
2864 return (r1 > r2 ? e1 : e2);
2868 * ==========================================================================
2870 * ==========================================================================
2873 zio_ready(zio_t *zio)
2875 blkptr_t *bp = zio->io_bp;
2876 zio_t *pio, *pio_next;
2878 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2879 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2880 return (ZIO_PIPELINE_STOP);
2882 if (zio->io_ready) {
2883 ASSERT(IO_IS_ALLOCATING(zio));
2884 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
2885 (zio->io_flags & ZIO_FLAG_NOPWRITE));
2886 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2891 if (bp != NULL && bp != &zio->io_bp_copy)
2892 zio->io_bp_copy = *bp;
2895 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2897 mutex_enter(&zio->io_lock);
2898 zio->io_state[ZIO_WAIT_READY] = 1;
2899 pio = zio_walk_parents(zio);
2900 mutex_exit(&zio->io_lock);
2903 * As we notify zio's parents, new parents could be added.
2904 * New parents go to the head of zio's io_parent_list, however,
2905 * so we will (correctly) not notify them. The remainder of zio's
2906 * io_parent_list, from 'pio_next' onward, cannot change because
2907 * all parents must wait for us to be done before they can be done.
2909 for (; pio != NULL; pio = pio_next) {
2910 pio_next = zio_walk_parents(zio);
2911 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2914 if (zio->io_flags & ZIO_FLAG_NODATA) {
2915 if (BP_IS_GANG(bp)) {
2916 zio->io_flags &= ~ZIO_FLAG_NODATA;
2918 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2919 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2923 if (zio_injection_enabled &&
2924 zio->io_spa->spa_syncing_txg == zio->io_txg)
2925 zio_handle_ignored_writes(zio);
2927 return (ZIO_PIPELINE_CONTINUE);
2931 zio_done(zio_t *zio)
2933 spa_t *spa = zio->io_spa;
2934 zio_t *lio = zio->io_logical;
2935 blkptr_t *bp = zio->io_bp;
2936 vdev_t *vd = zio->io_vd;
2937 uint64_t psize = zio->io_size;
2938 zio_t *pio, *pio_next;
2941 * If our children haven't all completed,
2942 * wait for them and then repeat this pipeline stage.
2944 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2945 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2946 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2947 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2948 return (ZIO_PIPELINE_STOP);
2950 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2951 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2952 ASSERT(zio->io_children[c][w] == 0);
2955 ASSERT(bp->blk_pad[0] == 0);
2956 ASSERT(bp->blk_pad[1] == 0);
2957 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
2958 (bp == zio_unique_parent(zio)->io_bp));
2959 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
2960 zio->io_bp_override == NULL &&
2961 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2962 ASSERT(!BP_SHOULD_BYTESWAP(bp));
2963 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
2964 ASSERT(BP_COUNT_GANG(bp) == 0 ||
2965 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
2967 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
2968 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
2972 * If there were child vdev/gang/ddt errors, they apply to us now.
2974 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
2975 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
2976 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
2979 * If the I/O on the transformed data was successful, generate any
2980 * checksum reports now while we still have the transformed data.
2982 if (zio->io_error == 0) {
2983 while (zio->io_cksum_report != NULL) {
2984 zio_cksum_report_t *zcr = zio->io_cksum_report;
2985 uint64_t align = zcr->zcr_align;
2986 uint64_t asize = P2ROUNDUP(psize, align);
2987 char *abuf = zio->io_data;
2989 if (asize != psize) {
2990 abuf = zio_buf_alloc(asize);
2991 bcopy(zio->io_data, abuf, psize);
2992 bzero(abuf + psize, asize - psize);
2995 zio->io_cksum_report = zcr->zcr_next;
2996 zcr->zcr_next = NULL;
2997 zcr->zcr_finish(zcr, abuf);
2998 zfs_ereport_free_checksum(zcr);
3001 zio_buf_free(abuf, asize);
3005 zio_pop_transforms(zio); /* note: may set zio->io_error */
3007 vdev_stat_update(zio, psize);
3009 if (zio->io_error) {
3011 * If this I/O is attached to a particular vdev,
3012 * generate an error message describing the I/O failure
3013 * at the block level. We ignore these errors if the
3014 * device is currently unavailable.
3016 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3017 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3019 if ((zio->io_error == EIO || !(zio->io_flags &
3020 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3023 * For logical I/O requests, tell the SPA to log the
3024 * error and generate a logical data ereport.
3026 spa_log_error(spa, zio);
3027 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3032 if (zio->io_error && zio == lio) {
3034 * Determine whether zio should be reexecuted. This will
3035 * propagate all the way to the root via zio_notify_parent().
3037 ASSERT(vd == NULL && bp != NULL);
3038 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3040 if (IO_IS_ALLOCATING(zio) &&
3041 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3042 if (zio->io_error != ENOSPC)
3043 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3045 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3048 if ((zio->io_type == ZIO_TYPE_READ ||
3049 zio->io_type == ZIO_TYPE_FREE) &&
3050 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3051 zio->io_error == ENXIO &&
3052 spa_load_state(spa) == SPA_LOAD_NONE &&
3053 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3054 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3056 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3057 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3060 * Here is a possibly good place to attempt to do
3061 * either combinatorial reconstruction or error correction
3062 * based on checksums. It also might be a good place
3063 * to send out preliminary ereports before we suspend
3069 * If there were logical child errors, they apply to us now.
3070 * We defer this until now to avoid conflating logical child
3071 * errors with errors that happened to the zio itself when
3072 * updating vdev stats and reporting FMA events above.
3074 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3076 if ((zio->io_error || zio->io_reexecute) &&
3077 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3078 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3079 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3081 zio_gang_tree_free(&zio->io_gang_tree);
3084 * Godfather I/Os should never suspend.
3086 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3087 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3088 zio->io_reexecute = 0;
3090 if (zio->io_reexecute) {
3092 * This is a logical I/O that wants to reexecute.
3094 * Reexecute is top-down. When an i/o fails, if it's not
3095 * the root, it simply notifies its parent and sticks around.
3096 * The parent, seeing that it still has children in zio_done(),
3097 * does the same. This percolates all the way up to the root.
3098 * The root i/o will reexecute or suspend the entire tree.
3100 * This approach ensures that zio_reexecute() honors
3101 * all the original i/o dependency relationships, e.g.
3102 * parents not executing until children are ready.
3104 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3106 zio->io_gang_leader = NULL;
3108 mutex_enter(&zio->io_lock);
3109 zio->io_state[ZIO_WAIT_DONE] = 1;
3110 mutex_exit(&zio->io_lock);
3113 * "The Godfather" I/O monitors its children but is
3114 * not a true parent to them. It will track them through
3115 * the pipeline but severs its ties whenever they get into
3116 * trouble (e.g. suspended). This allows "The Godfather"
3117 * I/O to return status without blocking.
3119 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3120 zio_link_t *zl = zio->io_walk_link;
3121 pio_next = zio_walk_parents(zio);
3123 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3124 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3125 zio_remove_child(pio, zio, zl);
3126 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3130 if ((pio = zio_unique_parent(zio)) != NULL) {
3132 * We're not a root i/o, so there's nothing to do
3133 * but notify our parent. Don't propagate errors
3134 * upward since we haven't permanently failed yet.
3136 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3137 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3138 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3139 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3141 * We'd fail again if we reexecuted now, so suspend
3142 * until conditions improve (e.g. device comes online).
3144 zio_suspend(spa, zio);
3147 * Reexecution is potentially a huge amount of work.
3148 * Hand it off to the otherwise-unused claim taskq.
3150 #if defined(illumos) || !defined(_KERNEL)
3151 ASSERT(zio->io_tqent.tqent_next == NULL);
3153 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
3155 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3156 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3159 return (ZIO_PIPELINE_STOP);
3162 ASSERT(zio->io_child_count == 0);
3163 ASSERT(zio->io_reexecute == 0);
3164 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3167 * Report any checksum errors, since the I/O is complete.
3169 while (zio->io_cksum_report != NULL) {
3170 zio_cksum_report_t *zcr = zio->io_cksum_report;
3171 zio->io_cksum_report = zcr->zcr_next;
3172 zcr->zcr_next = NULL;
3173 zcr->zcr_finish(zcr, NULL);
3174 zfs_ereport_free_checksum(zcr);
3178 * It is the responsibility of the done callback to ensure that this
3179 * particular zio is no longer discoverable for adoption, and as
3180 * such, cannot acquire any new parents.
3185 mutex_enter(&zio->io_lock);
3186 zio->io_state[ZIO_WAIT_DONE] = 1;
3187 mutex_exit(&zio->io_lock);
3189 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3190 zio_link_t *zl = zio->io_walk_link;
3191 pio_next = zio_walk_parents(zio);
3192 zio_remove_child(pio, zio, zl);
3193 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3196 if (zio->io_waiter != NULL) {
3197 mutex_enter(&zio->io_lock);
3198 zio->io_executor = NULL;
3199 cv_broadcast(&zio->io_cv);
3200 mutex_exit(&zio->io_lock);
3205 return (ZIO_PIPELINE_STOP);
3209 * ==========================================================================
3210 * I/O pipeline definition
3211 * ==========================================================================
3213 static zio_pipe_stage_t *zio_pipeline[] = {
3219 zio_checksum_generate,
3234 zio_checksum_verify,
3238 /* dnp is the dnode for zb1->zb_object */
3240 zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_t *zb1,
3241 const zbookmark_t *zb2)
3243 uint64_t zb1nextL0, zb2thisobj;
3245 ASSERT(zb1->zb_objset == zb2->zb_objset);
3246 ASSERT(zb2->zb_level == 0);
3249 * A bookmark in the deadlist is considered to be after
3252 if (zb2->zb_object == DMU_DEADLIST_OBJECT)
3255 /* The objset_phys_t isn't before anything. */
3259 zb1nextL0 = (zb1->zb_blkid + 1) <<
3260 ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
3262 zb2thisobj = zb2->zb_object ? zb2->zb_object :
3263 zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
3265 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3266 uint64_t nextobj = zb1nextL0 *
3267 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
3268 return (nextobj <= zb2thisobj);
3271 if (zb1->zb_object < zb2thisobj)
3273 if (zb1->zb_object > zb2thisobj)
3275 if (zb2->zb_object == DMU_META_DNODE_OBJECT)
3277 return (zb1nextL0 <= zb2->zb_blkid);