4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2013 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
27 #include <sys/zfs_context.h>
28 #include <sys/fm/fs/zfs.h>
31 #include <sys/spa_impl.h>
32 #include <sys/vdev_impl.h>
33 #include <sys/zio_impl.h>
34 #include <sys/zio_compress.h>
35 #include <sys/zio_checksum.h>
36 #include <sys/dmu_objset.h>
39 #include <sys/trim_map.h>
40 #include <sys/zfeature.h>
42 SYSCTL_DECL(_vfs_zfs);
43 SYSCTL_NODE(_vfs_zfs, OID_AUTO, zio, CTLFLAG_RW, 0, "ZFS ZIO");
44 #if defined(__amd64__)
45 static int zio_use_uma = 1;
47 static int zio_use_uma = 0;
49 TUNABLE_INT("vfs.zfs.zio.use_uma", &zio_use_uma);
50 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, use_uma, CTLFLAG_RDTUN, &zio_use_uma, 0,
51 "Use uma(9) for ZIO allocations");
52 static int zio_exclude_metadata = 0;
53 TUNABLE_INT("vfs.zfs.zio.exclude_metadata", &zio_exclude_metadata);
54 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, exclude_metadata, CTLFLAG_RDTUN, &zio_exclude_metadata, 0,
55 "Exclude metadata buffers from dumps as well");
57 zio_trim_stats_t zio_trim_stats = {
58 { "bytes", KSTAT_DATA_UINT64,
59 "Number of bytes successfully TRIMmed" },
60 { "success", KSTAT_DATA_UINT64,
61 "Number of successful TRIM requests" },
62 { "unsupported", KSTAT_DATA_UINT64,
63 "Number of TRIM requests that failed because TRIM is not supported" },
64 { "failed", KSTAT_DATA_UINT64,
65 "Number of TRIM requests that failed for reasons other than not supported" },
68 static kstat_t *zio_trim_ksp;
71 * ==========================================================================
72 * I/O type descriptions
73 * ==========================================================================
75 const char *zio_type_name[ZIO_TYPES] = {
76 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
81 * ==========================================================================
83 * ==========================================================================
85 kmem_cache_t *zio_cache;
86 kmem_cache_t *zio_link_cache;
87 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
88 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
91 extern vmem_t *zio_alloc_arena;
93 extern int zfs_mg_alloc_failures;
96 * The following actions directly effect the spa's sync-to-convergence logic.
97 * The values below define the sync pass when we start performing the action.
98 * Care should be taken when changing these values as they directly impact
99 * spa_sync() performance. Tuning these values may introduce subtle performance
100 * pathologies and should only be done in the context of performance analysis.
101 * These tunables will eventually be removed and replaced with #defines once
102 * enough analysis has been done to determine optimal values.
104 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
105 * regular blocks are not deferred.
107 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
108 TUNABLE_INT("vfs.zfs.sync_pass_deferred_free", &zfs_sync_pass_deferred_free);
109 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_deferred_free, CTLFLAG_RDTUN,
110 &zfs_sync_pass_deferred_free, 0, "defer frees starting in this pass");
111 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
112 TUNABLE_INT("vfs.zfs.sync_pass_dont_compress", &zfs_sync_pass_dont_compress);
113 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_dont_compress, CTLFLAG_RDTUN,
114 &zfs_sync_pass_dont_compress, 0, "don't compress starting in this pass");
115 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
116 TUNABLE_INT("vfs.zfs.sync_pass_rewrite", &zfs_sync_pass_rewrite);
117 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_rewrite, CTLFLAG_RDTUN,
118 &zfs_sync_pass_rewrite, 0, "rewrite new bps starting in this pass");
121 * An allocating zio is one that either currently has the DVA allocate
122 * stage set or will have it later in its lifetime.
124 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
126 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
129 int zio_buf_debug_limit = 16384;
131 int zio_buf_debug_limit = 0;
138 zio_cache = kmem_cache_create("zio_cache",
139 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
140 zio_link_cache = kmem_cache_create("zio_link_cache",
141 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
146 * For small buffers, we want a cache for each multiple of
147 * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache
148 * for each quarter-power of 2. For large buffers, we want
149 * a cache for each multiple of PAGESIZE.
151 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
152 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
155 size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
157 while (p2 & (p2 - 1))
163 * If we are using watchpoints, put each buffer on its own page,
164 * to eliminate the performance overhead of trapping to the
165 * kernel when modifying a non-watched buffer that shares the
166 * page with a watched buffer.
168 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
172 if (size <= 4 * SPA_MINBLOCKSIZE) {
173 align = SPA_MINBLOCKSIZE;
174 } else if (IS_P2ALIGNED(size, PAGESIZE)) {
176 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
182 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
183 zio_buf_cache[c] = kmem_cache_create(name, size,
184 align, NULL, NULL, NULL, NULL, NULL, cflags);
187 * Since zio_data bufs do not appear in crash dumps, we
188 * pass KMC_NOTOUCH so that no allocator metadata is
189 * stored with the buffers.
191 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
192 zio_data_buf_cache[c] = kmem_cache_create(name, size,
193 align, NULL, NULL, NULL, NULL, NULL,
194 cflags | KMC_NOTOUCH | KMC_NODEBUG);
199 ASSERT(zio_buf_cache[c] != NULL);
200 if (zio_buf_cache[c - 1] == NULL)
201 zio_buf_cache[c - 1] = zio_buf_cache[c];
203 ASSERT(zio_data_buf_cache[c] != NULL);
204 if (zio_data_buf_cache[c - 1] == NULL)
205 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
210 * The zio write taskqs have 1 thread per cpu, allow 1/2 of the taskqs
211 * to fail 3 times per txg or 8 failures, whichever is greater.
213 if (zfs_mg_alloc_failures == 0)
214 zfs_mg_alloc_failures = MAX((3 * max_ncpus / 2), 8);
215 else if (zfs_mg_alloc_failures < 8)
216 zfs_mg_alloc_failures = 8;
220 zio_trim_ksp = kstat_create("zfs", 0, "zio_trim", "misc",
222 sizeof(zio_trim_stats) / sizeof(kstat_named_t),
225 if (zio_trim_ksp != NULL) {
226 zio_trim_ksp->ks_data = &zio_trim_stats;
227 kstat_install(zio_trim_ksp);
235 kmem_cache_t *last_cache = NULL;
236 kmem_cache_t *last_data_cache = NULL;
238 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
239 if (zio_buf_cache[c] != last_cache) {
240 last_cache = zio_buf_cache[c];
241 kmem_cache_destroy(zio_buf_cache[c]);
243 zio_buf_cache[c] = NULL;
245 if (zio_data_buf_cache[c] != last_data_cache) {
246 last_data_cache = zio_data_buf_cache[c];
247 kmem_cache_destroy(zio_data_buf_cache[c]);
249 zio_data_buf_cache[c] = NULL;
252 kmem_cache_destroy(zio_link_cache);
253 kmem_cache_destroy(zio_cache);
257 if (zio_trim_ksp != NULL) {
258 kstat_delete(zio_trim_ksp);
264 * ==========================================================================
265 * Allocate and free I/O buffers
266 * ==========================================================================
270 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
271 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
272 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
273 * excess / transient data in-core during a crashdump.
276 zio_buf_alloc(size_t size)
278 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
279 int flags = zio_exclude_metadata ? KM_NODEBUG : 0;
281 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
284 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
286 return (kmem_alloc(size, KM_SLEEP|flags));
290 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
291 * crashdump if the kernel panics. This exists so that we will limit the amount
292 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
293 * of kernel heap dumped to disk when the kernel panics)
296 zio_data_buf_alloc(size_t size)
298 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
300 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
303 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
305 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG));
309 zio_buf_free(void *buf, size_t size)
311 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
313 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
316 kmem_cache_free(zio_buf_cache[c], buf);
318 kmem_free(buf, size);
322 zio_data_buf_free(void *buf, size_t size)
324 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
326 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
329 kmem_cache_free(zio_data_buf_cache[c], buf);
331 kmem_free(buf, size);
335 * ==========================================================================
336 * Push and pop I/O transform buffers
337 * ==========================================================================
340 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
341 zio_transform_func_t *transform)
343 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
345 zt->zt_orig_data = zio->io_data;
346 zt->zt_orig_size = zio->io_size;
347 zt->zt_bufsize = bufsize;
348 zt->zt_transform = transform;
350 zt->zt_next = zio->io_transform_stack;
351 zio->io_transform_stack = zt;
358 zio_pop_transforms(zio_t *zio)
362 while ((zt = zio->io_transform_stack) != NULL) {
363 if (zt->zt_transform != NULL)
364 zt->zt_transform(zio,
365 zt->zt_orig_data, zt->zt_orig_size);
367 if (zt->zt_bufsize != 0)
368 zio_buf_free(zio->io_data, zt->zt_bufsize);
370 zio->io_data = zt->zt_orig_data;
371 zio->io_size = zt->zt_orig_size;
372 zio->io_transform_stack = zt->zt_next;
374 kmem_free(zt, sizeof (zio_transform_t));
379 * ==========================================================================
380 * I/O transform callbacks for subblocks and decompression
381 * ==========================================================================
384 zio_subblock(zio_t *zio, void *data, uint64_t size)
386 ASSERT(zio->io_size > size);
388 if (zio->io_type == ZIO_TYPE_READ)
389 bcopy(zio->io_data, data, size);
393 zio_decompress(zio_t *zio, void *data, uint64_t size)
395 if (zio->io_error == 0 &&
396 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
397 zio->io_data, data, zio->io_size, size) != 0)
398 zio->io_error = SET_ERROR(EIO);
402 * ==========================================================================
403 * I/O parent/child relationships and pipeline interlocks
404 * ==========================================================================
407 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
408 * continue calling these functions until they return NULL.
409 * Otherwise, the next caller will pick up the list walk in
410 * some indeterminate state. (Otherwise every caller would
411 * have to pass in a cookie to keep the state represented by
412 * io_walk_link, which gets annoying.)
415 zio_walk_parents(zio_t *cio)
417 zio_link_t *zl = cio->io_walk_link;
418 list_t *pl = &cio->io_parent_list;
420 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
421 cio->io_walk_link = zl;
426 ASSERT(zl->zl_child == cio);
427 return (zl->zl_parent);
431 zio_walk_children(zio_t *pio)
433 zio_link_t *zl = pio->io_walk_link;
434 list_t *cl = &pio->io_child_list;
436 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
437 pio->io_walk_link = zl;
442 ASSERT(zl->zl_parent == pio);
443 return (zl->zl_child);
447 zio_unique_parent(zio_t *cio)
449 zio_t *pio = zio_walk_parents(cio);
451 VERIFY(zio_walk_parents(cio) == NULL);
456 zio_add_child(zio_t *pio, zio_t *cio)
458 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
461 * Logical I/Os can have logical, gang, or vdev children.
462 * Gang I/Os can have gang or vdev children.
463 * Vdev I/Os can only have vdev children.
464 * The following ASSERT captures all of these constraints.
466 ASSERT(cio->io_child_type <= pio->io_child_type);
471 mutex_enter(&cio->io_lock);
472 mutex_enter(&pio->io_lock);
474 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
476 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
477 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
479 list_insert_head(&pio->io_child_list, zl);
480 list_insert_head(&cio->io_parent_list, zl);
482 pio->io_child_count++;
483 cio->io_parent_count++;
485 mutex_exit(&pio->io_lock);
486 mutex_exit(&cio->io_lock);
490 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
492 ASSERT(zl->zl_parent == pio);
493 ASSERT(zl->zl_child == cio);
495 mutex_enter(&cio->io_lock);
496 mutex_enter(&pio->io_lock);
498 list_remove(&pio->io_child_list, zl);
499 list_remove(&cio->io_parent_list, zl);
501 pio->io_child_count--;
502 cio->io_parent_count--;
504 mutex_exit(&pio->io_lock);
505 mutex_exit(&cio->io_lock);
507 kmem_cache_free(zio_link_cache, zl);
511 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
513 uint64_t *countp = &zio->io_children[child][wait];
514 boolean_t waiting = B_FALSE;
516 mutex_enter(&zio->io_lock);
517 ASSERT(zio->io_stall == NULL);
520 zio->io_stall = countp;
523 mutex_exit(&zio->io_lock);
529 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
531 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
532 int *errorp = &pio->io_child_error[zio->io_child_type];
534 mutex_enter(&pio->io_lock);
535 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
536 *errorp = zio_worst_error(*errorp, zio->io_error);
537 pio->io_reexecute |= zio->io_reexecute;
538 ASSERT3U(*countp, >, 0);
542 if (*countp == 0 && pio->io_stall == countp) {
543 pio->io_stall = NULL;
544 mutex_exit(&pio->io_lock);
547 mutex_exit(&pio->io_lock);
552 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
554 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
555 zio->io_error = zio->io_child_error[c];
559 * ==========================================================================
560 * Create the various types of I/O (read, write, free, etc)
561 * ==========================================================================
564 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
565 void *data, uint64_t size, zio_done_func_t *done, void *private,
566 zio_type_t type, zio_priority_t priority, enum zio_flag flags,
567 vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
568 enum zio_stage stage, enum zio_stage pipeline)
572 ASSERT3U(type == ZIO_TYPE_FREE || size, <=, SPA_MAXBLOCKSIZE);
573 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
574 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
576 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
577 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
578 ASSERT(vd || stage == ZIO_STAGE_OPEN);
580 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
581 bzero(zio, sizeof (zio_t));
583 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
584 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
586 list_create(&zio->io_parent_list, sizeof (zio_link_t),
587 offsetof(zio_link_t, zl_parent_node));
588 list_create(&zio->io_child_list, sizeof (zio_link_t),
589 offsetof(zio_link_t, zl_child_node));
592 zio->io_child_type = ZIO_CHILD_VDEV;
593 else if (flags & ZIO_FLAG_GANG_CHILD)
594 zio->io_child_type = ZIO_CHILD_GANG;
595 else if (flags & ZIO_FLAG_DDT_CHILD)
596 zio->io_child_type = ZIO_CHILD_DDT;
598 zio->io_child_type = ZIO_CHILD_LOGICAL;
601 zio->io_bp = (blkptr_t *)bp;
602 zio->io_bp_copy = *bp;
603 zio->io_bp_orig = *bp;
604 if (type != ZIO_TYPE_WRITE ||
605 zio->io_child_type == ZIO_CHILD_DDT)
606 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
607 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
608 zio->io_logical = zio;
609 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
610 pipeline |= ZIO_GANG_STAGES;
616 zio->io_private = private;
618 zio->io_priority = priority;
620 zio->io_offset = offset;
621 zio->io_orig_data = zio->io_data = data;
622 zio->io_orig_size = zio->io_size = size;
623 zio->io_orig_flags = zio->io_flags = flags;
624 zio->io_orig_stage = zio->io_stage = stage;
625 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
627 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
628 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
631 zio->io_bookmark = *zb;
634 if (zio->io_logical == NULL)
635 zio->io_logical = pio->io_logical;
636 if (zio->io_child_type == ZIO_CHILD_GANG)
637 zio->io_gang_leader = pio->io_gang_leader;
638 zio_add_child(pio, zio);
645 zio_destroy(zio_t *zio)
647 list_destroy(&zio->io_parent_list);
648 list_destroy(&zio->io_child_list);
649 mutex_destroy(&zio->io_lock);
650 cv_destroy(&zio->io_cv);
651 kmem_cache_free(zio_cache, zio);
655 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
656 void *private, enum zio_flag flags)
660 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
661 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
662 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
668 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
670 return (zio_null(NULL, spa, NULL, done, private, flags));
674 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
675 void *data, uint64_t size, zio_done_func_t *done, void *private,
676 zio_priority_t priority, enum zio_flag flags, const zbookmark_t *zb)
680 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
681 data, size, done, private,
682 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
683 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
684 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
690 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
691 void *data, uint64_t size, const zio_prop_t *zp,
692 zio_done_func_t *ready, zio_done_func_t *physdone, zio_done_func_t *done,
694 zio_priority_t priority, enum zio_flag flags, const zbookmark_t *zb)
698 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
699 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
700 zp->zp_compress >= ZIO_COMPRESS_OFF &&
701 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
702 DMU_OT_IS_VALID(zp->zp_type) &&
705 zp->zp_copies <= spa_max_replication(spa));
707 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
708 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
709 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
710 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
712 zio->io_ready = ready;
713 zio->io_physdone = physdone;
720 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
721 uint64_t size, zio_done_func_t *done, void *private,
722 zio_priority_t priority, enum zio_flag flags, zbookmark_t *zb)
726 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
727 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
728 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
734 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
736 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
737 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
738 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
739 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
742 * We must reset the io_prop to match the values that existed
743 * when the bp was first written by dmu_sync() keeping in mind
744 * that nopwrite and dedup are mutually exclusive.
746 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
747 zio->io_prop.zp_nopwrite = nopwrite;
748 zio->io_prop.zp_copies = copies;
749 zio->io_bp_override = bp;
753 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
755 metaslab_check_free(spa, bp);
758 * Frees that are for the currently-syncing txg, are not going to be
759 * deferred, and which will not need to do a read (i.e. not GANG or
760 * DEDUP), can be processed immediately. Otherwise, put them on the
761 * in-memory list for later processing.
763 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
764 txg != spa->spa_syncing_txg ||
765 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
766 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
768 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp,
769 BP_GET_PSIZE(bp), 0)));
774 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
775 uint64_t size, enum zio_flag flags)
778 enum zio_stage stage = ZIO_FREE_PIPELINE;
780 dprintf_bp(bp, "freeing in txg %llu, pass %u",
781 (longlong_t)txg, spa->spa_sync_pass);
783 ASSERT(!BP_IS_HOLE(bp));
784 ASSERT(spa_syncing_txg(spa) == txg);
785 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
787 metaslab_check_free(spa, bp);
790 if (zfs_trim_enabled)
791 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START |
792 ZIO_STAGE_VDEV_IO_ASSESS;
794 * GANG and DEDUP blocks can induce a read (for the gang block header,
795 * or the DDT), so issue them asynchronously so that this thread is
798 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
799 stage |= ZIO_STAGE_ISSUE_ASYNC;
801 zio = zio_create(pio, spa, txg, bp, NULL, size,
802 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags,
803 NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
809 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
810 zio_done_func_t *done, void *private, enum zio_flag flags)
815 * A claim is an allocation of a specific block. Claims are needed
816 * to support immediate writes in the intent log. The issue is that
817 * immediate writes contain committed data, but in a txg that was
818 * *not* committed. Upon opening the pool after an unclean shutdown,
819 * the intent log claims all blocks that contain immediate write data
820 * so that the SPA knows they're in use.
822 * All claims *must* be resolved in the first txg -- before the SPA
823 * starts allocating blocks -- so that nothing is allocated twice.
824 * If txg == 0 we just verify that the block is claimable.
826 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
827 ASSERT(txg == spa_first_txg(spa) || txg == 0);
828 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
830 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
831 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
832 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
838 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset,
839 uint64_t size, zio_done_func_t *done, void *private,
845 if (vd->vdev_children == 0) {
846 zio = zio_create(pio, spa, 0, NULL, NULL, size, done, private,
847 ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, offset, NULL,
848 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
852 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
854 for (c = 0; c < vd->vdev_children; c++)
855 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
856 offset, size, done, private, flags));
863 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
864 void *data, int checksum, zio_done_func_t *done, void *private,
865 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
869 ASSERT(vd->vdev_children == 0);
870 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
871 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
872 ASSERT3U(offset + size, <=, vd->vdev_psize);
874 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
875 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
876 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
878 zio->io_prop.zp_checksum = checksum;
884 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
885 void *data, int checksum, zio_done_func_t *done, void *private,
886 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
890 ASSERT(vd->vdev_children == 0);
891 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
892 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
893 ASSERT3U(offset + size, <=, vd->vdev_psize);
895 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
896 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
897 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
899 zio->io_prop.zp_checksum = checksum;
901 if (zio_checksum_table[checksum].ci_eck) {
903 * zec checksums are necessarily destructive -- they modify
904 * the end of the write buffer to hold the verifier/checksum.
905 * Therefore, we must make a local copy in case the data is
906 * being written to multiple places in parallel.
908 void *wbuf = zio_buf_alloc(size);
909 bcopy(data, wbuf, size);
910 zio_push_transform(zio, wbuf, size, size, NULL);
917 * Create a child I/O to do some work for us.
920 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
921 void *data, uint64_t size, int type, zio_priority_t priority,
922 enum zio_flag flags, zio_done_func_t *done, void *private)
924 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
927 ASSERT(vd->vdev_parent ==
928 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
930 if (type == ZIO_TYPE_READ && bp != NULL) {
932 * If we have the bp, then the child should perform the
933 * checksum and the parent need not. This pushes error
934 * detection as close to the leaves as possible and
935 * eliminates redundant checksums in the interior nodes.
937 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
938 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
941 if (vd->vdev_children == 0)
942 offset += VDEV_LABEL_START_SIZE;
944 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
947 * If we've decided to do a repair, the write is not speculative --
948 * even if the original read was.
950 if (flags & ZIO_FLAG_IO_REPAIR)
951 flags &= ~ZIO_FLAG_SPECULATIVE;
953 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
954 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
955 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
957 zio->io_physdone = pio->io_physdone;
958 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
959 zio->io_logical->io_phys_children++;
965 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
966 int type, zio_priority_t priority, enum zio_flag flags,
967 zio_done_func_t *done, void *private)
971 ASSERT(vd->vdev_ops->vdev_op_leaf);
973 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
974 data, size, done, private, type, priority,
975 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
977 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
983 zio_flush(zio_t *zio, vdev_t *vd)
985 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0,
987 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
991 zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size)
994 ASSERT(vd->vdev_ops->vdev_op_leaf);
996 return zio_ioctl(zio, spa, vd, DKIOCTRIM, offset, size,
998 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY);
1002 zio_shrink(zio_t *zio, uint64_t size)
1004 ASSERT(zio->io_executor == NULL);
1005 ASSERT(zio->io_orig_size == zio->io_size);
1006 ASSERT(size <= zio->io_size);
1009 * We don't shrink for raidz because of problems with the
1010 * reconstruction when reading back less than the block size.
1011 * Note, BP_IS_RAIDZ() assumes no compression.
1013 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1014 if (!BP_IS_RAIDZ(zio->io_bp))
1015 zio->io_orig_size = zio->io_size = size;
1019 * ==========================================================================
1020 * Prepare to read and write logical blocks
1021 * ==========================================================================
1025 zio_read_bp_init(zio_t *zio)
1027 blkptr_t *bp = zio->io_bp;
1029 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1030 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1031 !(zio->io_flags & ZIO_FLAG_RAW)) {
1032 uint64_t psize = BP_GET_PSIZE(bp);
1033 void *cbuf = zio_buf_alloc(psize);
1035 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
1038 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1039 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1041 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1042 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1044 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1045 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1047 return (ZIO_PIPELINE_CONTINUE);
1051 zio_write_bp_init(zio_t *zio)
1053 spa_t *spa = zio->io_spa;
1054 zio_prop_t *zp = &zio->io_prop;
1055 enum zio_compress compress = zp->zp_compress;
1056 blkptr_t *bp = zio->io_bp;
1057 uint64_t lsize = zio->io_size;
1058 uint64_t psize = lsize;
1062 * If our children haven't all reached the ready stage,
1063 * wait for them and then repeat this pipeline stage.
1065 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1066 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1067 return (ZIO_PIPELINE_STOP);
1069 if (!IO_IS_ALLOCATING(zio))
1070 return (ZIO_PIPELINE_CONTINUE);
1072 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1074 if (zio->io_bp_override) {
1075 ASSERT(bp->blk_birth != zio->io_txg);
1076 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1078 *bp = *zio->io_bp_override;
1079 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1082 * If we've been overridden and nopwrite is set then
1083 * set the flag accordingly to indicate that a nopwrite
1084 * has already occurred.
1086 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1087 ASSERT(!zp->zp_dedup);
1088 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1089 return (ZIO_PIPELINE_CONTINUE);
1092 ASSERT(!zp->zp_nopwrite);
1094 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1095 return (ZIO_PIPELINE_CONTINUE);
1097 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1098 zp->zp_dedup_verify);
1100 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1101 BP_SET_DEDUP(bp, 1);
1102 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1103 return (ZIO_PIPELINE_CONTINUE);
1105 zio->io_bp_override = NULL;
1109 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1111 * We're rewriting an existing block, which means we're
1112 * working on behalf of spa_sync(). For spa_sync() to
1113 * converge, it must eventually be the case that we don't
1114 * have to allocate new blocks. But compression changes
1115 * the blocksize, which forces a reallocate, and makes
1116 * convergence take longer. Therefore, after the first
1117 * few passes, stop compressing to ensure convergence.
1119 pass = spa_sync_pass(spa);
1121 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1122 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1123 ASSERT(!BP_GET_DEDUP(bp));
1125 if (pass >= zfs_sync_pass_dont_compress)
1126 compress = ZIO_COMPRESS_OFF;
1128 /* Make sure someone doesn't change their mind on overwrites */
1129 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
1130 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1133 if (compress != ZIO_COMPRESS_OFF) {
1134 metaslab_class_t *mc = spa_normal_class(spa);
1135 void *cbuf = zio_buf_alloc(lsize);
1136 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize,
1137 (size_t)metaslab_class_get_minblocksize(mc));
1138 if (psize == 0 || psize == lsize) {
1139 compress = ZIO_COMPRESS_OFF;
1140 zio_buf_free(cbuf, lsize);
1142 ASSERT(psize < lsize);
1143 zio_push_transform(zio, cbuf, psize, lsize, NULL);
1148 * The final pass of spa_sync() must be all rewrites, but the first
1149 * few passes offer a trade-off: allocating blocks defers convergence,
1150 * but newly allocated blocks are sequential, so they can be written
1151 * to disk faster. Therefore, we allow the first few passes of
1152 * spa_sync() to allocate new blocks, but force rewrites after that.
1153 * There should only be a handful of blocks after pass 1 in any case.
1155 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1156 BP_GET_PSIZE(bp) == psize &&
1157 pass >= zfs_sync_pass_rewrite) {
1159 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1160 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1161 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1164 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1168 if (zio->io_bp_orig.blk_birth != 0 &&
1169 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1170 BP_SET_LSIZE(bp, lsize);
1171 BP_SET_TYPE(bp, zp->zp_type);
1172 BP_SET_LEVEL(bp, zp->zp_level);
1173 BP_SET_BIRTH(bp, zio->io_txg, 0);
1175 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1177 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1178 BP_SET_LSIZE(bp, lsize);
1179 BP_SET_TYPE(bp, zp->zp_type);
1180 BP_SET_LEVEL(bp, zp->zp_level);
1181 BP_SET_PSIZE(bp, psize);
1182 BP_SET_COMPRESS(bp, compress);
1183 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1184 BP_SET_DEDUP(bp, zp->zp_dedup);
1185 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1187 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1188 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1189 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1191 if (zp->zp_nopwrite) {
1192 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1193 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1194 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1198 return (ZIO_PIPELINE_CONTINUE);
1202 zio_free_bp_init(zio_t *zio)
1204 blkptr_t *bp = zio->io_bp;
1206 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1207 if (BP_GET_DEDUP(bp))
1208 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1211 return (ZIO_PIPELINE_CONTINUE);
1215 * ==========================================================================
1216 * Execute the I/O pipeline
1217 * ==========================================================================
1221 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1223 spa_t *spa = zio->io_spa;
1224 zio_type_t t = zio->io_type;
1225 int flags = (cutinline ? TQ_FRONT : 0);
1227 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1230 * If we're a config writer or a probe, the normal issue and
1231 * interrupt threads may all be blocked waiting for the config lock.
1232 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1234 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1238 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1240 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1244 * If this is a high priority I/O, then use the high priority taskq if
1247 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1248 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1251 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1254 * NB: We are assuming that the zio can only be dispatched
1255 * to a single taskq at a time. It would be a grievous error
1256 * to dispatch the zio to another taskq at the same time.
1258 #if defined(illumos) || !defined(_KERNEL)
1259 ASSERT(zio->io_tqent.tqent_next == NULL);
1261 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
1263 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1264 flags, &zio->io_tqent);
1268 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1270 kthread_t *executor = zio->io_executor;
1271 spa_t *spa = zio->io_spa;
1273 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1274 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1276 for (i = 0; i < tqs->stqs_count; i++) {
1277 if (taskq_member(tqs->stqs_taskq[i], executor))
1286 zio_issue_async(zio_t *zio)
1288 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1290 return (ZIO_PIPELINE_STOP);
1294 zio_interrupt(zio_t *zio)
1296 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1300 * Execute the I/O pipeline until one of the following occurs:
1302 * (1) the I/O completes
1303 * (2) the pipeline stalls waiting for dependent child I/Os
1304 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1305 * (4) the I/O is delegated by vdev-level caching or aggregation
1306 * (5) the I/O is deferred due to vdev-level queueing
1307 * (6) the I/O is handed off to another thread.
1309 * In all cases, the pipeline stops whenever there's no CPU work; it never
1310 * burns a thread in cv_wait().
1312 * There's no locking on io_stage because there's no legitimate way
1313 * for multiple threads to be attempting to process the same I/O.
1315 static zio_pipe_stage_t *zio_pipeline[];
1318 zio_execute(zio_t *zio)
1320 zio->io_executor = curthread;
1322 while (zio->io_stage < ZIO_STAGE_DONE) {
1323 enum zio_stage pipeline = zio->io_pipeline;
1324 enum zio_stage stage = zio->io_stage;
1327 ASSERT(!MUTEX_HELD(&zio->io_lock));
1328 ASSERT(ISP2(stage));
1329 ASSERT(zio->io_stall == NULL);
1333 } while ((stage & pipeline) == 0);
1335 ASSERT(stage <= ZIO_STAGE_DONE);
1338 * If we are in interrupt context and this pipeline stage
1339 * will grab a config lock that is held across I/O,
1340 * or may wait for an I/O that needs an interrupt thread
1341 * to complete, issue async to avoid deadlock.
1343 * For VDEV_IO_START, we cut in line so that the io will
1344 * be sent to disk promptly.
1346 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1347 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1348 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1349 zio_requeue_io_start_cut_in_line : B_FALSE;
1350 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1354 zio->io_stage = stage;
1355 rv = zio_pipeline[highbit(stage) - 1](zio);
1357 if (rv == ZIO_PIPELINE_STOP)
1360 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1365 * ==========================================================================
1366 * Initiate I/O, either sync or async
1367 * ==========================================================================
1370 zio_wait(zio_t *zio)
1374 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1375 ASSERT(zio->io_executor == NULL);
1377 zio->io_waiter = curthread;
1381 mutex_enter(&zio->io_lock);
1382 while (zio->io_executor != NULL)
1383 cv_wait(&zio->io_cv, &zio->io_lock);
1384 mutex_exit(&zio->io_lock);
1386 error = zio->io_error;
1393 zio_nowait(zio_t *zio)
1395 ASSERT(zio->io_executor == NULL);
1397 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1398 zio_unique_parent(zio) == NULL) {
1400 * This is a logical async I/O with no parent to wait for it.
1401 * We add it to the spa_async_root_zio "Godfather" I/O which
1402 * will ensure they complete prior to unloading the pool.
1404 spa_t *spa = zio->io_spa;
1406 zio_add_child(spa->spa_async_zio_root, zio);
1413 * ==========================================================================
1414 * Reexecute or suspend/resume failed I/O
1415 * ==========================================================================
1419 zio_reexecute(zio_t *pio)
1421 zio_t *cio, *cio_next;
1423 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1424 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1425 ASSERT(pio->io_gang_leader == NULL);
1426 ASSERT(pio->io_gang_tree == NULL);
1428 pio->io_flags = pio->io_orig_flags;
1429 pio->io_stage = pio->io_orig_stage;
1430 pio->io_pipeline = pio->io_orig_pipeline;
1431 pio->io_reexecute = 0;
1432 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1434 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1435 pio->io_state[w] = 0;
1436 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1437 pio->io_child_error[c] = 0;
1439 if (IO_IS_ALLOCATING(pio))
1440 BP_ZERO(pio->io_bp);
1443 * As we reexecute pio's children, new children could be created.
1444 * New children go to the head of pio's io_child_list, however,
1445 * so we will (correctly) not reexecute them. The key is that
1446 * the remainder of pio's io_child_list, from 'cio_next' onward,
1447 * cannot be affected by any side effects of reexecuting 'cio'.
1449 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1450 cio_next = zio_walk_children(pio);
1451 mutex_enter(&pio->io_lock);
1452 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1453 pio->io_children[cio->io_child_type][w]++;
1454 mutex_exit(&pio->io_lock);
1459 * Now that all children have been reexecuted, execute the parent.
1460 * We don't reexecute "The Godfather" I/O here as it's the
1461 * responsibility of the caller to wait on him.
1463 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1468 zio_suspend(spa_t *spa, zio_t *zio)
1470 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1471 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1472 "failure and the failure mode property for this pool "
1473 "is set to panic.", spa_name(spa));
1475 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1477 mutex_enter(&spa->spa_suspend_lock);
1479 if (spa->spa_suspend_zio_root == NULL)
1480 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1481 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1482 ZIO_FLAG_GODFATHER);
1484 spa->spa_suspended = B_TRUE;
1487 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1488 ASSERT(zio != spa->spa_suspend_zio_root);
1489 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1490 ASSERT(zio_unique_parent(zio) == NULL);
1491 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1492 zio_add_child(spa->spa_suspend_zio_root, zio);
1495 mutex_exit(&spa->spa_suspend_lock);
1499 zio_resume(spa_t *spa)
1504 * Reexecute all previously suspended i/o.
1506 mutex_enter(&spa->spa_suspend_lock);
1507 spa->spa_suspended = B_FALSE;
1508 cv_broadcast(&spa->spa_suspend_cv);
1509 pio = spa->spa_suspend_zio_root;
1510 spa->spa_suspend_zio_root = NULL;
1511 mutex_exit(&spa->spa_suspend_lock);
1517 return (zio_wait(pio));
1521 zio_resume_wait(spa_t *spa)
1523 mutex_enter(&spa->spa_suspend_lock);
1524 while (spa_suspended(spa))
1525 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1526 mutex_exit(&spa->spa_suspend_lock);
1530 * ==========================================================================
1533 * A gang block is a collection of small blocks that looks to the DMU
1534 * like one large block. When zio_dva_allocate() cannot find a block
1535 * of the requested size, due to either severe fragmentation or the pool
1536 * being nearly full, it calls zio_write_gang_block() to construct the
1537 * block from smaller fragments.
1539 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1540 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1541 * an indirect block: it's an array of block pointers. It consumes
1542 * only one sector and hence is allocatable regardless of fragmentation.
1543 * The gang header's bps point to its gang members, which hold the data.
1545 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1546 * as the verifier to ensure uniqueness of the SHA256 checksum.
1547 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1548 * not the gang header. This ensures that data block signatures (needed for
1549 * deduplication) are independent of how the block is physically stored.
1551 * Gang blocks can be nested: a gang member may itself be a gang block.
1552 * Thus every gang block is a tree in which root and all interior nodes are
1553 * gang headers, and the leaves are normal blocks that contain user data.
1554 * The root of the gang tree is called the gang leader.
1556 * To perform any operation (read, rewrite, free, claim) on a gang block,
1557 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1558 * in the io_gang_tree field of the original logical i/o by recursively
1559 * reading the gang leader and all gang headers below it. This yields
1560 * an in-core tree containing the contents of every gang header and the
1561 * bps for every constituent of the gang block.
1563 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1564 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1565 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1566 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1567 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1568 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1569 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1570 * of the gang header plus zio_checksum_compute() of the data to update the
1571 * gang header's blk_cksum as described above.
1573 * The two-phase assemble/issue model solves the problem of partial failure --
1574 * what if you'd freed part of a gang block but then couldn't read the
1575 * gang header for another part? Assembling the entire gang tree first
1576 * ensures that all the necessary gang header I/O has succeeded before
1577 * starting the actual work of free, claim, or write. Once the gang tree
1578 * is assembled, free and claim are in-memory operations that cannot fail.
1580 * In the event that a gang write fails, zio_dva_unallocate() walks the
1581 * gang tree to immediately free (i.e. insert back into the space map)
1582 * everything we've allocated. This ensures that we don't get ENOSPC
1583 * errors during repeated suspend/resume cycles due to a flaky device.
1585 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1586 * the gang tree, we won't modify the block, so we can safely defer the free
1587 * (knowing that the block is still intact). If we *can* assemble the gang
1588 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1589 * each constituent bp and we can allocate a new block on the next sync pass.
1591 * In all cases, the gang tree allows complete recovery from partial failure.
1592 * ==========================================================================
1596 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1601 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1602 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1603 &pio->io_bookmark));
1607 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1612 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1613 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1614 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1616 * As we rewrite each gang header, the pipeline will compute
1617 * a new gang block header checksum for it; but no one will
1618 * compute a new data checksum, so we do that here. The one
1619 * exception is the gang leader: the pipeline already computed
1620 * its data checksum because that stage precedes gang assembly.
1621 * (Presently, nothing actually uses interior data checksums;
1622 * this is just good hygiene.)
1624 if (gn != pio->io_gang_leader->io_gang_tree) {
1625 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1626 data, BP_GET_PSIZE(bp));
1629 * If we are here to damage data for testing purposes,
1630 * leave the GBH alone so that we can detect the damage.
1632 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1633 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1635 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1636 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1637 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1645 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1647 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1648 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp),
1649 ZIO_GANG_CHILD_FLAGS(pio)));
1654 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1656 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1657 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1660 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1669 static void zio_gang_tree_assemble_done(zio_t *zio);
1671 static zio_gang_node_t *
1672 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1674 zio_gang_node_t *gn;
1676 ASSERT(*gnpp == NULL);
1678 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1679 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1686 zio_gang_node_free(zio_gang_node_t **gnpp)
1688 zio_gang_node_t *gn = *gnpp;
1690 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1691 ASSERT(gn->gn_child[g] == NULL);
1693 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1694 kmem_free(gn, sizeof (*gn));
1699 zio_gang_tree_free(zio_gang_node_t **gnpp)
1701 zio_gang_node_t *gn = *gnpp;
1706 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1707 zio_gang_tree_free(&gn->gn_child[g]);
1709 zio_gang_node_free(gnpp);
1713 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1715 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1717 ASSERT(gio->io_gang_leader == gio);
1718 ASSERT(BP_IS_GANG(bp));
1720 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1721 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1722 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1726 zio_gang_tree_assemble_done(zio_t *zio)
1728 zio_t *gio = zio->io_gang_leader;
1729 zio_gang_node_t *gn = zio->io_private;
1730 blkptr_t *bp = zio->io_bp;
1732 ASSERT(gio == zio_unique_parent(zio));
1733 ASSERT(zio->io_child_count == 0);
1738 if (BP_SHOULD_BYTESWAP(bp))
1739 byteswap_uint64_array(zio->io_data, zio->io_size);
1741 ASSERT(zio->io_data == gn->gn_gbh);
1742 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1743 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1745 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1746 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1747 if (!BP_IS_GANG(gbp))
1749 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1754 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1756 zio_t *gio = pio->io_gang_leader;
1759 ASSERT(BP_IS_GANG(bp) == !!gn);
1760 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1761 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1764 * If you're a gang header, your data is in gn->gn_gbh.
1765 * If you're a gang member, your data is in 'data' and gn == NULL.
1767 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1770 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1772 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1773 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1774 if (BP_IS_HOLE(gbp))
1776 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1777 data = (char *)data + BP_GET_PSIZE(gbp);
1781 if (gn == gio->io_gang_tree && gio->io_data != NULL)
1782 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1789 zio_gang_assemble(zio_t *zio)
1791 blkptr_t *bp = zio->io_bp;
1793 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1794 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1796 zio->io_gang_leader = zio;
1798 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1800 return (ZIO_PIPELINE_CONTINUE);
1804 zio_gang_issue(zio_t *zio)
1806 blkptr_t *bp = zio->io_bp;
1808 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1809 return (ZIO_PIPELINE_STOP);
1811 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1812 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1814 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1815 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1817 zio_gang_tree_free(&zio->io_gang_tree);
1819 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1821 return (ZIO_PIPELINE_CONTINUE);
1825 zio_write_gang_member_ready(zio_t *zio)
1827 zio_t *pio = zio_unique_parent(zio);
1828 zio_t *gio = zio->io_gang_leader;
1829 dva_t *cdva = zio->io_bp->blk_dva;
1830 dva_t *pdva = pio->io_bp->blk_dva;
1833 if (BP_IS_HOLE(zio->io_bp))
1836 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1838 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1839 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1840 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1841 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1842 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1844 mutex_enter(&pio->io_lock);
1845 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1846 ASSERT(DVA_GET_GANG(&pdva[d]));
1847 asize = DVA_GET_ASIZE(&pdva[d]);
1848 asize += DVA_GET_ASIZE(&cdva[d]);
1849 DVA_SET_ASIZE(&pdva[d], asize);
1851 mutex_exit(&pio->io_lock);
1855 zio_write_gang_block(zio_t *pio)
1857 spa_t *spa = pio->io_spa;
1858 blkptr_t *bp = pio->io_bp;
1859 zio_t *gio = pio->io_gang_leader;
1861 zio_gang_node_t *gn, **gnpp;
1862 zio_gbh_phys_t *gbh;
1863 uint64_t txg = pio->io_txg;
1864 uint64_t resid = pio->io_size;
1866 int copies = gio->io_prop.zp_copies;
1867 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1871 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1872 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1873 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1875 pio->io_error = error;
1876 return (ZIO_PIPELINE_CONTINUE);
1880 gnpp = &gio->io_gang_tree;
1882 gnpp = pio->io_private;
1883 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1886 gn = zio_gang_node_alloc(gnpp);
1888 bzero(gbh, SPA_GANGBLOCKSIZE);
1891 * Create the gang header.
1893 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1894 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1897 * Create and nowait the gang children.
1899 for (int g = 0; resid != 0; resid -= lsize, g++) {
1900 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1902 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1904 zp.zp_checksum = gio->io_prop.zp_checksum;
1905 zp.zp_compress = ZIO_COMPRESS_OFF;
1906 zp.zp_type = DMU_OT_NONE;
1908 zp.zp_copies = gio->io_prop.zp_copies;
1909 zp.zp_dedup = B_FALSE;
1910 zp.zp_dedup_verify = B_FALSE;
1911 zp.zp_nopwrite = B_FALSE;
1913 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1914 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1915 zio_write_gang_member_ready, NULL, NULL, &gn->gn_child[g],
1916 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1917 &pio->io_bookmark));
1921 * Set pio's pipeline to just wait for zio to finish.
1923 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1927 return (ZIO_PIPELINE_CONTINUE);
1931 * The zio_nop_write stage in the pipeline determines if allocating
1932 * a new bp is necessary. By leveraging a cryptographically secure checksum,
1933 * such as SHA256, we can compare the checksums of the new data and the old
1934 * to determine if allocating a new block is required. The nopwrite
1935 * feature can handle writes in either syncing or open context (i.e. zil
1936 * writes) and as a result is mutually exclusive with dedup.
1939 zio_nop_write(zio_t *zio)
1941 blkptr_t *bp = zio->io_bp;
1942 blkptr_t *bp_orig = &zio->io_bp_orig;
1943 zio_prop_t *zp = &zio->io_prop;
1945 ASSERT(BP_GET_LEVEL(bp) == 0);
1946 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1947 ASSERT(zp->zp_nopwrite);
1948 ASSERT(!zp->zp_dedup);
1949 ASSERT(zio->io_bp_override == NULL);
1950 ASSERT(IO_IS_ALLOCATING(zio));
1953 * Check to see if the original bp and the new bp have matching
1954 * characteristics (i.e. same checksum, compression algorithms, etc).
1955 * If they don't then just continue with the pipeline which will
1956 * allocate a new bp.
1958 if (BP_IS_HOLE(bp_orig) ||
1959 !zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup ||
1960 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
1961 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
1962 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
1963 zp->zp_copies != BP_GET_NDVAS(bp_orig))
1964 return (ZIO_PIPELINE_CONTINUE);
1967 * If the checksums match then reset the pipeline so that we
1968 * avoid allocating a new bp and issuing any I/O.
1970 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
1971 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup);
1972 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
1973 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
1974 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
1975 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
1976 sizeof (uint64_t)) == 0);
1979 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1980 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1983 return (ZIO_PIPELINE_CONTINUE);
1987 * ==========================================================================
1989 * ==========================================================================
1992 zio_ddt_child_read_done(zio_t *zio)
1994 blkptr_t *bp = zio->io_bp;
1995 ddt_entry_t *dde = zio->io_private;
1997 zio_t *pio = zio_unique_parent(zio);
1999 mutex_enter(&pio->io_lock);
2000 ddp = ddt_phys_select(dde, bp);
2001 if (zio->io_error == 0)
2002 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2003 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
2004 dde->dde_repair_data = zio->io_data;
2006 zio_buf_free(zio->io_data, zio->io_size);
2007 mutex_exit(&pio->io_lock);
2011 zio_ddt_read_start(zio_t *zio)
2013 blkptr_t *bp = zio->io_bp;
2015 ASSERT(BP_GET_DEDUP(bp));
2016 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2017 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2019 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2020 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2021 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2022 ddt_phys_t *ddp = dde->dde_phys;
2023 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2026 ASSERT(zio->io_vsd == NULL);
2029 if (ddp_self == NULL)
2030 return (ZIO_PIPELINE_CONTINUE);
2032 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2033 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2035 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2037 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2038 zio_buf_alloc(zio->io_size), zio->io_size,
2039 zio_ddt_child_read_done, dde, zio->io_priority,
2040 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
2041 &zio->io_bookmark));
2043 return (ZIO_PIPELINE_CONTINUE);
2046 zio_nowait(zio_read(zio, zio->io_spa, bp,
2047 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
2048 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2050 return (ZIO_PIPELINE_CONTINUE);
2054 zio_ddt_read_done(zio_t *zio)
2056 blkptr_t *bp = zio->io_bp;
2058 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2059 return (ZIO_PIPELINE_STOP);
2061 ASSERT(BP_GET_DEDUP(bp));
2062 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2063 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2065 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2066 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2067 ddt_entry_t *dde = zio->io_vsd;
2069 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2070 return (ZIO_PIPELINE_CONTINUE);
2073 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2074 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2075 return (ZIO_PIPELINE_STOP);
2077 if (dde->dde_repair_data != NULL) {
2078 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2079 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2081 ddt_repair_done(ddt, dde);
2085 ASSERT(zio->io_vsd == NULL);
2087 return (ZIO_PIPELINE_CONTINUE);
2091 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2093 spa_t *spa = zio->io_spa;
2096 * Note: we compare the original data, not the transformed data,
2097 * because when zio->io_bp is an override bp, we will not have
2098 * pushed the I/O transforms. That's an important optimization
2099 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2101 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2102 zio_t *lio = dde->dde_lead_zio[p];
2105 return (lio->io_orig_size != zio->io_orig_size ||
2106 bcmp(zio->io_orig_data, lio->io_orig_data,
2107 zio->io_orig_size) != 0);
2111 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2112 ddt_phys_t *ddp = &dde->dde_phys[p];
2114 if (ddp->ddp_phys_birth != 0) {
2115 arc_buf_t *abuf = NULL;
2116 uint32_t aflags = ARC_WAIT;
2117 blkptr_t blk = *zio->io_bp;
2120 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2124 error = arc_read(NULL, spa, &blk,
2125 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2126 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2127 &aflags, &zio->io_bookmark);
2130 if (arc_buf_size(abuf) != zio->io_orig_size ||
2131 bcmp(abuf->b_data, zio->io_orig_data,
2132 zio->io_orig_size) != 0)
2133 error = SET_ERROR(EEXIST);
2134 VERIFY(arc_buf_remove_ref(abuf, &abuf));
2138 return (error != 0);
2146 zio_ddt_child_write_ready(zio_t *zio)
2148 int p = zio->io_prop.zp_copies;
2149 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2150 ddt_entry_t *dde = zio->io_private;
2151 ddt_phys_t *ddp = &dde->dde_phys[p];
2159 ASSERT(dde->dde_lead_zio[p] == zio);
2161 ddt_phys_fill(ddp, zio->io_bp);
2163 while ((pio = zio_walk_parents(zio)) != NULL)
2164 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2170 zio_ddt_child_write_done(zio_t *zio)
2172 int p = zio->io_prop.zp_copies;
2173 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2174 ddt_entry_t *dde = zio->io_private;
2175 ddt_phys_t *ddp = &dde->dde_phys[p];
2179 ASSERT(ddp->ddp_refcnt == 0);
2180 ASSERT(dde->dde_lead_zio[p] == zio);
2181 dde->dde_lead_zio[p] = NULL;
2183 if (zio->io_error == 0) {
2184 while (zio_walk_parents(zio) != NULL)
2185 ddt_phys_addref(ddp);
2187 ddt_phys_clear(ddp);
2194 zio_ddt_ditto_write_done(zio_t *zio)
2196 int p = DDT_PHYS_DITTO;
2197 zio_prop_t *zp = &zio->io_prop;
2198 blkptr_t *bp = zio->io_bp;
2199 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2200 ddt_entry_t *dde = zio->io_private;
2201 ddt_phys_t *ddp = &dde->dde_phys[p];
2202 ddt_key_t *ddk = &dde->dde_key;
2206 ASSERT(ddp->ddp_refcnt == 0);
2207 ASSERT(dde->dde_lead_zio[p] == zio);
2208 dde->dde_lead_zio[p] = NULL;
2210 if (zio->io_error == 0) {
2211 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2212 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2213 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2214 if (ddp->ddp_phys_birth != 0)
2215 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2216 ddt_phys_fill(ddp, bp);
2223 zio_ddt_write(zio_t *zio)
2225 spa_t *spa = zio->io_spa;
2226 blkptr_t *bp = zio->io_bp;
2227 uint64_t txg = zio->io_txg;
2228 zio_prop_t *zp = &zio->io_prop;
2229 int p = zp->zp_copies;
2233 ddt_t *ddt = ddt_select(spa, bp);
2237 ASSERT(BP_GET_DEDUP(bp));
2238 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2239 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2242 dde = ddt_lookup(ddt, bp, B_TRUE);
2243 ddp = &dde->dde_phys[p];
2245 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2247 * If we're using a weak checksum, upgrade to a strong checksum
2248 * and try again. If we're already using a strong checksum,
2249 * we can't resolve it, so just convert to an ordinary write.
2250 * (And automatically e-mail a paper to Nature?)
2252 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2253 zp->zp_checksum = spa_dedup_checksum(spa);
2254 zio_pop_transforms(zio);
2255 zio->io_stage = ZIO_STAGE_OPEN;
2258 zp->zp_dedup = B_FALSE;
2260 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2262 return (ZIO_PIPELINE_CONTINUE);
2265 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2266 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2268 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2269 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2270 zio_prop_t czp = *zp;
2272 czp.zp_copies = ditto_copies;
2275 * If we arrived here with an override bp, we won't have run
2276 * the transform stack, so we won't have the data we need to
2277 * generate a child i/o. So, toss the override bp and restart.
2278 * This is safe, because using the override bp is just an
2279 * optimization; and it's rare, so the cost doesn't matter.
2281 if (zio->io_bp_override) {
2282 zio_pop_transforms(zio);
2283 zio->io_stage = ZIO_STAGE_OPEN;
2284 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2285 zio->io_bp_override = NULL;
2288 return (ZIO_PIPELINE_CONTINUE);
2291 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2292 zio->io_orig_size, &czp, NULL, NULL,
2293 zio_ddt_ditto_write_done, dde, zio->io_priority,
2294 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2296 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2297 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2300 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2301 if (ddp->ddp_phys_birth != 0)
2302 ddt_bp_fill(ddp, bp, txg);
2303 if (dde->dde_lead_zio[p] != NULL)
2304 zio_add_child(zio, dde->dde_lead_zio[p]);
2306 ddt_phys_addref(ddp);
2307 } else if (zio->io_bp_override) {
2308 ASSERT(bp->blk_birth == txg);
2309 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2310 ddt_phys_fill(ddp, bp);
2311 ddt_phys_addref(ddp);
2313 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2314 zio->io_orig_size, zp, zio_ddt_child_write_ready, NULL,
2315 zio_ddt_child_write_done, dde, zio->io_priority,
2316 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2318 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2319 dde->dde_lead_zio[p] = cio;
2329 return (ZIO_PIPELINE_CONTINUE);
2332 ddt_entry_t *freedde; /* for debugging */
2335 zio_ddt_free(zio_t *zio)
2337 spa_t *spa = zio->io_spa;
2338 blkptr_t *bp = zio->io_bp;
2339 ddt_t *ddt = ddt_select(spa, bp);
2343 ASSERT(BP_GET_DEDUP(bp));
2344 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2347 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2348 ddp = ddt_phys_select(dde, bp);
2349 ddt_phys_decref(ddp);
2352 return (ZIO_PIPELINE_CONTINUE);
2356 * ==========================================================================
2357 * Allocate and free blocks
2358 * ==========================================================================
2361 zio_dva_allocate(zio_t *zio)
2363 spa_t *spa = zio->io_spa;
2364 metaslab_class_t *mc = spa_normal_class(spa);
2365 blkptr_t *bp = zio->io_bp;
2369 if (zio->io_gang_leader == NULL) {
2370 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2371 zio->io_gang_leader = zio;
2374 ASSERT(BP_IS_HOLE(bp));
2375 ASSERT0(BP_GET_NDVAS(bp));
2376 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2377 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2378 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2381 * The dump device does not support gang blocks so allocation on
2382 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2383 * the "fast" gang feature.
2385 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2386 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2387 METASLAB_GANG_CHILD : 0;
2388 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2389 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2392 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2393 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2395 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2396 return (zio_write_gang_block(zio));
2397 zio->io_error = error;
2400 return (ZIO_PIPELINE_CONTINUE);
2404 zio_dva_free(zio_t *zio)
2406 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2408 return (ZIO_PIPELINE_CONTINUE);
2412 zio_dva_claim(zio_t *zio)
2416 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2418 zio->io_error = error;
2420 return (ZIO_PIPELINE_CONTINUE);
2424 * Undo an allocation. This is used by zio_done() when an I/O fails
2425 * and we want to give back the block we just allocated.
2426 * This handles both normal blocks and gang blocks.
2429 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2431 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2432 ASSERT(zio->io_bp_override == NULL);
2434 if (!BP_IS_HOLE(bp))
2435 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2438 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2439 zio_dva_unallocate(zio, gn->gn_child[g],
2440 &gn->gn_gbh->zg_blkptr[g]);
2446 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2449 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2450 uint64_t size, boolean_t use_slog)
2454 ASSERT(txg > spa_syncing_txg(spa));
2457 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2458 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2459 * when allocating them.
2462 error = metaslab_alloc(spa, spa_log_class(spa), size,
2463 new_bp, 1, txg, old_bp,
2464 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2468 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2469 new_bp, 1, txg, old_bp,
2470 METASLAB_HINTBP_AVOID);
2474 BP_SET_LSIZE(new_bp, size);
2475 BP_SET_PSIZE(new_bp, size);
2476 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2477 BP_SET_CHECKSUM(new_bp,
2478 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2479 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2480 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2481 BP_SET_LEVEL(new_bp, 0);
2482 BP_SET_DEDUP(new_bp, 0);
2483 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2490 * Free an intent log block.
2493 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2495 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2496 ASSERT(!BP_IS_GANG(bp));
2498 zio_free(spa, txg, bp);
2502 * ==========================================================================
2503 * Read, write and delete to physical devices
2504 * ==========================================================================
2507 zio_vdev_io_start(zio_t *zio)
2509 vdev_t *vd = zio->io_vd;
2511 spa_t *spa = zio->io_spa;
2513 ASSERT(zio->io_error == 0);
2514 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2517 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2518 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2521 * The mirror_ops handle multiple DVAs in a single BP.
2523 return (vdev_mirror_ops.vdev_op_io_start(zio));
2526 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE) {
2527 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg);
2528 return (ZIO_PIPELINE_CONTINUE);
2532 * We keep track of time-sensitive I/Os so that the scan thread
2533 * can quickly react to certain workloads. In particular, we care
2534 * about non-scrubbing, top-level reads and writes with the following
2536 * - synchronous writes of user data to non-slog devices
2537 * - any reads of user data
2538 * When these conditions are met, adjust the timestamp of spa_last_io
2539 * which allows the scan thread to adjust its workload accordingly.
2541 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2542 vd == vd->vdev_top && !vd->vdev_islog &&
2543 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2544 zio->io_txg != spa_syncing_txg(spa)) {
2545 uint64_t old = spa->spa_last_io;
2546 uint64_t new = ddi_get_lbolt64();
2548 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2551 align = 1ULL << vd->vdev_top->vdev_ashift;
2553 if (P2PHASE(zio->io_size, align) != 0) {
2554 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2556 if (zio->io_type == ZIO_TYPE_READ ||
2557 zio->io_type == ZIO_TYPE_WRITE)
2558 abuf = zio_buf_alloc(asize);
2559 ASSERT(vd == vd->vdev_top);
2560 if (zio->io_type == ZIO_TYPE_WRITE) {
2561 bcopy(zio->io_data, abuf, zio->io_size);
2562 bzero(abuf + zio->io_size, asize - zio->io_size);
2564 zio_push_transform(zio, abuf, asize, abuf ? asize : 0,
2568 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2569 ASSERT(P2PHASE(zio->io_size, align) == 0);
2570 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa));
2573 * If this is a repair I/O, and there's no self-healing involved --
2574 * that is, we're just resilvering what we expect to resilver --
2575 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2576 * This prevents spurious resilvering with nested replication.
2577 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2578 * A is out of date, we'll read from C+D, then use the data to
2579 * resilver A+B -- but we don't actually want to resilver B, just A.
2580 * The top-level mirror has no way to know this, so instead we just
2581 * discard unnecessary repairs as we work our way down the vdev tree.
2582 * The same logic applies to any form of nested replication:
2583 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2585 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2586 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2587 zio->io_txg != 0 && /* not a delegated i/o */
2588 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2589 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2590 zio_vdev_io_bypass(zio);
2591 return (ZIO_PIPELINE_CONTINUE);
2594 if (vd->vdev_ops->vdev_op_leaf &&
2595 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2597 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio))
2598 return (ZIO_PIPELINE_CONTINUE);
2600 if ((zio = vdev_queue_io(zio)) == NULL)
2601 return (ZIO_PIPELINE_STOP);
2603 if (!vdev_accessible(vd, zio)) {
2604 zio->io_error = SET_ERROR(ENXIO);
2606 return (ZIO_PIPELINE_STOP);
2611 * Note that we ignore repair writes for TRIM because they can conflict
2612 * with normal writes. This isn't an issue because, by definition, we
2613 * only repair blocks that aren't freed.
2615 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_WRITE &&
2616 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2617 if (!trim_map_write_start(zio))
2618 return (ZIO_PIPELINE_STOP);
2621 return (vd->vdev_ops->vdev_op_io_start(zio));
2625 zio_vdev_io_done(zio_t *zio)
2627 vdev_t *vd = zio->io_vd;
2628 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2629 boolean_t unexpected_error = B_FALSE;
2631 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2632 return (ZIO_PIPELINE_STOP);
2634 ASSERT(zio->io_type == ZIO_TYPE_READ ||
2635 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE);
2637 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2638 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2640 if (zio->io_type == ZIO_TYPE_WRITE &&
2641 !(zio->io_flags & ZIO_FLAG_IO_REPAIR))
2642 trim_map_write_done(zio);
2644 vdev_queue_io_done(zio);
2646 if (zio->io_type == ZIO_TYPE_WRITE)
2647 vdev_cache_write(zio);
2649 if (zio_injection_enabled && zio->io_error == 0)
2650 zio->io_error = zio_handle_device_injection(vd,
2653 if (zio_injection_enabled && zio->io_error == 0)
2654 zio->io_error = zio_handle_label_injection(zio, EIO);
2656 if (zio->io_error) {
2657 if (!vdev_accessible(vd, zio)) {
2658 zio->io_error = SET_ERROR(ENXIO);
2660 unexpected_error = B_TRUE;
2665 ops->vdev_op_io_done(zio);
2667 if (unexpected_error)
2668 VERIFY(vdev_probe(vd, zio) == NULL);
2670 return (ZIO_PIPELINE_CONTINUE);
2674 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2675 * disk, and use that to finish the checksum ereport later.
2678 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2679 const void *good_buf)
2681 /* no processing needed */
2682 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2687 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2689 void *buf = zio_buf_alloc(zio->io_size);
2691 bcopy(zio->io_data, buf, zio->io_size);
2693 zcr->zcr_cbinfo = zio->io_size;
2694 zcr->zcr_cbdata = buf;
2695 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2696 zcr->zcr_free = zio_buf_free;
2700 zio_vdev_io_assess(zio_t *zio)
2702 vdev_t *vd = zio->io_vd;
2704 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2705 return (ZIO_PIPELINE_STOP);
2707 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2708 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2710 if (zio->io_vsd != NULL) {
2711 zio->io_vsd_ops->vsd_free(zio);
2715 if (zio_injection_enabled && zio->io_error == 0)
2716 zio->io_error = zio_handle_fault_injection(zio, EIO);
2718 if (zio->io_type == ZIO_TYPE_IOCTL && zio->io_cmd == DKIOCTRIM)
2719 switch (zio->io_error) {
2721 ZIO_TRIM_STAT_INCR(bytes, zio->io_size);
2722 ZIO_TRIM_STAT_BUMP(success);
2725 ZIO_TRIM_STAT_BUMP(unsupported);
2728 ZIO_TRIM_STAT_BUMP(failed);
2733 * If the I/O failed, determine whether we should attempt to retry it.
2735 * On retry, we cut in line in the issue queue, since we don't want
2736 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2738 if (zio->io_error && vd == NULL &&
2739 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2740 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2741 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2743 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2744 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2745 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2746 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2747 zio_requeue_io_start_cut_in_line);
2748 return (ZIO_PIPELINE_STOP);
2752 * If we got an error on a leaf device, convert it to ENXIO
2753 * if the device is not accessible at all.
2755 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2756 !vdev_accessible(vd, zio))
2757 zio->io_error = SET_ERROR(ENXIO);
2760 * If we can't write to an interior vdev (mirror or RAID-Z),
2761 * set vdev_cant_write so that we stop trying to allocate from it.
2763 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2764 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
2765 vd->vdev_cant_write = B_TRUE;
2769 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2771 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2772 zio->io_physdone != NULL) {
2773 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
2774 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
2775 zio->io_physdone(zio->io_logical);
2778 return (ZIO_PIPELINE_CONTINUE);
2782 zio_vdev_io_reissue(zio_t *zio)
2784 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2785 ASSERT(zio->io_error == 0);
2787 zio->io_stage >>= 1;
2791 zio_vdev_io_redone(zio_t *zio)
2793 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2795 zio->io_stage >>= 1;
2799 zio_vdev_io_bypass(zio_t *zio)
2801 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2802 ASSERT(zio->io_error == 0);
2804 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2805 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2809 * ==========================================================================
2810 * Generate and verify checksums
2811 * ==========================================================================
2814 zio_checksum_generate(zio_t *zio)
2816 blkptr_t *bp = zio->io_bp;
2817 enum zio_checksum checksum;
2821 * This is zio_write_phys().
2822 * We're either generating a label checksum, or none at all.
2824 checksum = zio->io_prop.zp_checksum;
2826 if (checksum == ZIO_CHECKSUM_OFF)
2827 return (ZIO_PIPELINE_CONTINUE);
2829 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2831 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2832 ASSERT(!IO_IS_ALLOCATING(zio));
2833 checksum = ZIO_CHECKSUM_GANG_HEADER;
2835 checksum = BP_GET_CHECKSUM(bp);
2839 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2841 return (ZIO_PIPELINE_CONTINUE);
2845 zio_checksum_verify(zio_t *zio)
2847 zio_bad_cksum_t info;
2848 blkptr_t *bp = zio->io_bp;
2851 ASSERT(zio->io_vd != NULL);
2855 * This is zio_read_phys().
2856 * We're either verifying a label checksum, or nothing at all.
2858 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2859 return (ZIO_PIPELINE_CONTINUE);
2861 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2864 if ((error = zio_checksum_error(zio, &info)) != 0) {
2865 zio->io_error = error;
2866 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2867 zfs_ereport_start_checksum(zio->io_spa,
2868 zio->io_vd, zio, zio->io_offset,
2869 zio->io_size, NULL, &info);
2873 return (ZIO_PIPELINE_CONTINUE);
2877 * Called by RAID-Z to ensure we don't compute the checksum twice.
2880 zio_checksum_verified(zio_t *zio)
2882 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2886 * ==========================================================================
2887 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2888 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2889 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2890 * indicate errors that are specific to one I/O, and most likely permanent.
2891 * Any other error is presumed to be worse because we weren't expecting it.
2892 * ==========================================================================
2895 zio_worst_error(int e1, int e2)
2897 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2900 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2901 if (e1 == zio_error_rank[r1])
2904 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2905 if (e2 == zio_error_rank[r2])
2908 return (r1 > r2 ? e1 : e2);
2912 * ==========================================================================
2914 * ==========================================================================
2917 zio_ready(zio_t *zio)
2919 blkptr_t *bp = zio->io_bp;
2920 zio_t *pio, *pio_next;
2922 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2923 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2924 return (ZIO_PIPELINE_STOP);
2926 if (zio->io_ready) {
2927 ASSERT(IO_IS_ALLOCATING(zio));
2928 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
2929 (zio->io_flags & ZIO_FLAG_NOPWRITE));
2930 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2935 if (bp != NULL && bp != &zio->io_bp_copy)
2936 zio->io_bp_copy = *bp;
2939 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2941 mutex_enter(&zio->io_lock);
2942 zio->io_state[ZIO_WAIT_READY] = 1;
2943 pio = zio_walk_parents(zio);
2944 mutex_exit(&zio->io_lock);
2947 * As we notify zio's parents, new parents could be added.
2948 * New parents go to the head of zio's io_parent_list, however,
2949 * so we will (correctly) not notify them. The remainder of zio's
2950 * io_parent_list, from 'pio_next' onward, cannot change because
2951 * all parents must wait for us to be done before they can be done.
2953 for (; pio != NULL; pio = pio_next) {
2954 pio_next = zio_walk_parents(zio);
2955 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2958 if (zio->io_flags & ZIO_FLAG_NODATA) {
2959 if (BP_IS_GANG(bp)) {
2960 zio->io_flags &= ~ZIO_FLAG_NODATA;
2962 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2963 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2967 if (zio_injection_enabled &&
2968 zio->io_spa->spa_syncing_txg == zio->io_txg)
2969 zio_handle_ignored_writes(zio);
2971 return (ZIO_PIPELINE_CONTINUE);
2975 zio_done(zio_t *zio)
2977 spa_t *spa = zio->io_spa;
2978 zio_t *lio = zio->io_logical;
2979 blkptr_t *bp = zio->io_bp;
2980 vdev_t *vd = zio->io_vd;
2981 uint64_t psize = zio->io_size;
2982 zio_t *pio, *pio_next;
2985 * If our children haven't all completed,
2986 * wait for them and then repeat this pipeline stage.
2988 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2989 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2990 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2991 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2992 return (ZIO_PIPELINE_STOP);
2994 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2995 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2996 ASSERT(zio->io_children[c][w] == 0);
2999 ASSERT(bp->blk_pad[0] == 0);
3000 ASSERT(bp->blk_pad[1] == 0);
3001 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3002 (bp == zio_unique_parent(zio)->io_bp));
3003 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3004 zio->io_bp_override == NULL &&
3005 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3006 ASSERT(!BP_SHOULD_BYTESWAP(bp));
3007 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3008 ASSERT(BP_COUNT_GANG(bp) == 0 ||
3009 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3011 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3012 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3016 * If there were child vdev/gang/ddt errors, they apply to us now.
3018 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3019 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3020 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3023 * If the I/O on the transformed data was successful, generate any
3024 * checksum reports now while we still have the transformed data.
3026 if (zio->io_error == 0) {
3027 while (zio->io_cksum_report != NULL) {
3028 zio_cksum_report_t *zcr = zio->io_cksum_report;
3029 uint64_t align = zcr->zcr_align;
3030 uint64_t asize = P2ROUNDUP(psize, align);
3031 char *abuf = zio->io_data;
3033 if (asize != psize) {
3034 abuf = zio_buf_alloc(asize);
3035 bcopy(zio->io_data, abuf, psize);
3036 bzero(abuf + psize, asize - psize);
3039 zio->io_cksum_report = zcr->zcr_next;
3040 zcr->zcr_next = NULL;
3041 zcr->zcr_finish(zcr, abuf);
3042 zfs_ereport_free_checksum(zcr);
3045 zio_buf_free(abuf, asize);
3049 zio_pop_transforms(zio); /* note: may set zio->io_error */
3051 vdev_stat_update(zio, psize);
3053 if (zio->io_error) {
3055 * If this I/O is attached to a particular vdev,
3056 * generate an error message describing the I/O failure
3057 * at the block level. We ignore these errors if the
3058 * device is currently unavailable.
3060 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3061 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3063 if ((zio->io_error == EIO || !(zio->io_flags &
3064 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3067 * For logical I/O requests, tell the SPA to log the
3068 * error and generate a logical data ereport.
3070 spa_log_error(spa, zio);
3071 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3076 if (zio->io_error && zio == lio) {
3078 * Determine whether zio should be reexecuted. This will
3079 * propagate all the way to the root via zio_notify_parent().
3081 ASSERT(vd == NULL && bp != NULL);
3082 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3084 if (IO_IS_ALLOCATING(zio) &&
3085 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3086 if (zio->io_error != ENOSPC)
3087 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3089 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3092 if ((zio->io_type == ZIO_TYPE_READ ||
3093 zio->io_type == ZIO_TYPE_FREE) &&
3094 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3095 zio->io_error == ENXIO &&
3096 spa_load_state(spa) == SPA_LOAD_NONE &&
3097 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3098 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3100 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3101 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3104 * Here is a possibly good place to attempt to do
3105 * either combinatorial reconstruction or error correction
3106 * based on checksums. It also might be a good place
3107 * to send out preliminary ereports before we suspend
3113 * If there were logical child errors, they apply to us now.
3114 * We defer this until now to avoid conflating logical child
3115 * errors with errors that happened to the zio itself when
3116 * updating vdev stats and reporting FMA events above.
3118 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3120 if ((zio->io_error || zio->io_reexecute) &&
3121 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3122 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3123 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3125 zio_gang_tree_free(&zio->io_gang_tree);
3128 * Godfather I/Os should never suspend.
3130 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3131 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3132 zio->io_reexecute = 0;
3134 if (zio->io_reexecute) {
3136 * This is a logical I/O that wants to reexecute.
3138 * Reexecute is top-down. When an i/o fails, if it's not
3139 * the root, it simply notifies its parent and sticks around.
3140 * The parent, seeing that it still has children in zio_done(),
3141 * does the same. This percolates all the way up to the root.
3142 * The root i/o will reexecute or suspend the entire tree.
3144 * This approach ensures that zio_reexecute() honors
3145 * all the original i/o dependency relationships, e.g.
3146 * parents not executing until children are ready.
3148 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3150 zio->io_gang_leader = NULL;
3152 mutex_enter(&zio->io_lock);
3153 zio->io_state[ZIO_WAIT_DONE] = 1;
3154 mutex_exit(&zio->io_lock);
3157 * "The Godfather" I/O monitors its children but is
3158 * not a true parent to them. It will track them through
3159 * the pipeline but severs its ties whenever they get into
3160 * trouble (e.g. suspended). This allows "The Godfather"
3161 * I/O to return status without blocking.
3163 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3164 zio_link_t *zl = zio->io_walk_link;
3165 pio_next = zio_walk_parents(zio);
3167 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3168 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3169 zio_remove_child(pio, zio, zl);
3170 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3174 if ((pio = zio_unique_parent(zio)) != NULL) {
3176 * We're not a root i/o, so there's nothing to do
3177 * but notify our parent. Don't propagate errors
3178 * upward since we haven't permanently failed yet.
3180 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3181 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3182 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3183 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3185 * We'd fail again if we reexecuted now, so suspend
3186 * until conditions improve (e.g. device comes online).
3188 zio_suspend(spa, zio);
3191 * Reexecution is potentially a huge amount of work.
3192 * Hand it off to the otherwise-unused claim taskq.
3194 #if defined(illumos) || !defined(_KERNEL)
3195 ASSERT(zio->io_tqent.tqent_next == NULL);
3197 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
3199 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3200 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3203 return (ZIO_PIPELINE_STOP);
3206 ASSERT(zio->io_child_count == 0);
3207 ASSERT(zio->io_reexecute == 0);
3208 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3211 * Report any checksum errors, since the I/O is complete.
3213 while (zio->io_cksum_report != NULL) {
3214 zio_cksum_report_t *zcr = zio->io_cksum_report;
3215 zio->io_cksum_report = zcr->zcr_next;
3216 zcr->zcr_next = NULL;
3217 zcr->zcr_finish(zcr, NULL);
3218 zfs_ereport_free_checksum(zcr);
3222 * It is the responsibility of the done callback to ensure that this
3223 * particular zio is no longer discoverable for adoption, and as
3224 * such, cannot acquire any new parents.
3229 mutex_enter(&zio->io_lock);
3230 zio->io_state[ZIO_WAIT_DONE] = 1;
3231 mutex_exit(&zio->io_lock);
3233 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3234 zio_link_t *zl = zio->io_walk_link;
3235 pio_next = zio_walk_parents(zio);
3236 zio_remove_child(pio, zio, zl);
3237 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3240 if (zio->io_waiter != NULL) {
3241 mutex_enter(&zio->io_lock);
3242 zio->io_executor = NULL;
3243 cv_broadcast(&zio->io_cv);
3244 mutex_exit(&zio->io_lock);
3249 return (ZIO_PIPELINE_STOP);
3253 * ==========================================================================
3254 * I/O pipeline definition
3255 * ==========================================================================
3257 static zio_pipe_stage_t *zio_pipeline[] = {
3263 zio_checksum_generate,
3278 zio_checksum_verify,
3282 /* dnp is the dnode for zb1->zb_object */
3284 zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_t *zb1,
3285 const zbookmark_t *zb2)
3287 uint64_t zb1nextL0, zb2thisobj;
3289 ASSERT(zb1->zb_objset == zb2->zb_objset);
3290 ASSERT(zb2->zb_level == 0);
3293 * A bookmark in the deadlist is considered to be after
3296 if (zb2->zb_object == DMU_DEADLIST_OBJECT)
3299 /* The objset_phys_t isn't before anything. */
3303 zb1nextL0 = (zb1->zb_blkid + 1) <<
3304 ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
3306 zb2thisobj = zb2->zb_object ? zb2->zb_object :
3307 zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
3309 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3310 uint64_t nextobj = zb1nextL0 *
3311 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
3312 return (nextobj <= zb2thisobj);
3315 if (zb1->zb_object < zb2thisobj)
3317 if (zb1->zb_object > zb2thisobj)
3319 if (zb2->zb_object == DMU_META_DNODE_OBJECT)
3321 return (zb1nextL0 <= zb2->zb_blkid);