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) 2011, 2014 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
27 #include <sys/zfs_context.h>
28 #include <sys/fm/fs/zfs.h>
31 #include <sys/spa_impl.h>
32 #include <sys/vdev_impl.h>
33 #include <sys/zio_impl.h>
34 #include <sys/zio_compress.h>
35 #include <sys/zio_checksum.h>
36 #include <sys/dmu_objset.h>
39 #include <sys/trim_map.h>
40 #include <sys/zfeature.h>
42 SYSCTL_DECL(_vfs_zfs);
43 SYSCTL_NODE(_vfs_zfs, OID_AUTO, zio, CTLFLAG_RW, 0, "ZFS ZIO");
44 static int zio_use_uma = 0;
45 TUNABLE_INT("vfs.zfs.zio.use_uma", &zio_use_uma);
46 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, use_uma, CTLFLAG_RDTUN, &zio_use_uma, 0,
47 "Use uma(9) for ZIO allocations");
49 zio_trim_stats_t zio_trim_stats = {
50 { "bytes", KSTAT_DATA_UINT64,
51 "Number of bytes successfully TRIMmed" },
52 { "success", KSTAT_DATA_UINT64,
53 "Number of successful TRIM requests" },
54 { "unsupported", KSTAT_DATA_UINT64,
55 "Number of TRIM requests that failed because TRIM is not supported" },
56 { "failed", KSTAT_DATA_UINT64,
57 "Number of TRIM requests that failed for reasons other than not supported" },
60 static kstat_t *zio_trim_ksp;
63 * ==========================================================================
64 * I/O type descriptions
65 * ==========================================================================
67 const char *zio_type_name[ZIO_TYPES] = {
68 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
73 * ==========================================================================
75 * ==========================================================================
77 kmem_cache_t *zio_cache;
78 kmem_cache_t *zio_link_cache;
79 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
80 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
83 extern vmem_t *zio_alloc_arena;
87 * The following actions directly effect the spa's sync-to-convergence logic.
88 * The values below define the sync pass when we start performing the action.
89 * Care should be taken when changing these values as they directly impact
90 * spa_sync() performance. Tuning these values may introduce subtle performance
91 * pathologies and should only be done in the context of performance analysis.
92 * These tunables will eventually be removed and replaced with #defines once
93 * enough analysis has been done to determine optimal values.
95 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
96 * regular blocks are not deferred.
98 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
99 TUNABLE_INT("vfs.zfs.sync_pass_deferred_free", &zfs_sync_pass_deferred_free);
100 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_deferred_free, CTLFLAG_RDTUN,
101 &zfs_sync_pass_deferred_free, 0, "defer frees starting in this pass");
102 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
103 TUNABLE_INT("vfs.zfs.sync_pass_dont_compress", &zfs_sync_pass_dont_compress);
104 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_dont_compress, CTLFLAG_RDTUN,
105 &zfs_sync_pass_dont_compress, 0, "don't compress starting in this pass");
106 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
107 TUNABLE_INT("vfs.zfs.sync_pass_rewrite", &zfs_sync_pass_rewrite);
108 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_rewrite, CTLFLAG_RDTUN,
109 &zfs_sync_pass_rewrite, 0, "rewrite new bps starting in this pass");
112 * An allocating zio is one that either currently has the DVA allocate
113 * stage set or will have it later in its lifetime.
115 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
117 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
120 int zio_buf_debug_limit = 16384;
122 int zio_buf_debug_limit = 0;
129 zio_cache = kmem_cache_create("zio_cache",
130 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
131 zio_link_cache = kmem_cache_create("zio_link_cache",
132 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
137 * For small buffers, we want a cache for each multiple of
138 * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache
139 * for each quarter-power of 2. For large buffers, we want
140 * a cache for each multiple of PAGESIZE.
142 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
143 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
146 size_t cflags = (size > zio_buf_debug_limit) ? (KMC_NODEBUG|KMC_NOTOUCH) : 0;
148 while (p2 & (p2 - 1))
154 * If we are using watchpoints, put each buffer on its own page,
155 * to eliminate the performance overhead of trapping to the
156 * kernel when modifying a non-watched buffer that shares the
157 * page with a watched buffer.
159 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
163 if (size <= 4 * SPA_MINBLOCKSIZE) {
164 align = SPA_MINBLOCKSIZE;
165 } else if (IS_P2ALIGNED(size, PAGESIZE)) {
167 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
173 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
174 zio_buf_cache[c] = kmem_cache_create(name, size,
175 align, NULL, NULL, NULL, NULL, NULL, cflags);
178 * Since zio_data bufs do not appear in crash dumps, we
179 * pass KMC_NOTOUCH so that no allocator metadata is
180 * stored with the buffers.
182 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
183 zio_data_buf_cache[c] = kmem_cache_create(name, size,
184 align, NULL, NULL, NULL, NULL, NULL,
185 cflags | KMC_NOTOUCH);
190 ASSERT(zio_buf_cache[c] != NULL);
191 if (zio_buf_cache[c - 1] == NULL)
192 zio_buf_cache[c - 1] = zio_buf_cache[c];
194 ASSERT(zio_data_buf_cache[c] != NULL);
195 if (zio_data_buf_cache[c - 1] == NULL)
196 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
202 zio_trim_ksp = kstat_create("zfs", 0, "zio_trim", "misc",
204 sizeof(zio_trim_stats) / sizeof(kstat_named_t),
207 if (zio_trim_ksp != NULL) {
208 zio_trim_ksp->ks_data = &zio_trim_stats;
209 kstat_install(zio_trim_ksp);
217 kmem_cache_t *last_cache = NULL;
218 kmem_cache_t *last_data_cache = NULL;
220 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
221 if (zio_buf_cache[c] != last_cache) {
222 last_cache = zio_buf_cache[c];
223 kmem_cache_destroy(zio_buf_cache[c]);
225 zio_buf_cache[c] = NULL;
227 if (zio_data_buf_cache[c] != last_data_cache) {
228 last_data_cache = zio_data_buf_cache[c];
229 kmem_cache_destroy(zio_data_buf_cache[c]);
231 zio_data_buf_cache[c] = NULL;
234 kmem_cache_destroy(zio_link_cache);
235 kmem_cache_destroy(zio_cache);
239 if (zio_trim_ksp != NULL) {
240 kstat_delete(zio_trim_ksp);
246 * ==========================================================================
247 * Allocate and free I/O buffers
248 * ==========================================================================
252 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
253 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
254 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
255 * excess / transient data in-core during a crashdump.
258 zio_buf_alloc(size_t size)
260 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
262 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
265 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
267 return (kmem_alloc(size, KM_SLEEP));
271 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
272 * crashdump if the kernel panics. This exists so that we will limit the amount
273 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
274 * of kernel heap dumped to disk when the kernel panics)
277 zio_data_buf_alloc(size_t size)
279 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
281 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
284 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
286 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG));
290 zio_buf_free(void *buf, size_t size)
292 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
294 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
297 kmem_cache_free(zio_buf_cache[c], buf);
299 kmem_free(buf, size);
303 zio_data_buf_free(void *buf, size_t size)
305 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
307 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
310 kmem_cache_free(zio_data_buf_cache[c], buf);
312 kmem_free(buf, size);
316 * ==========================================================================
317 * Push and pop I/O transform buffers
318 * ==========================================================================
321 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
322 zio_transform_func_t *transform)
324 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
326 zt->zt_orig_data = zio->io_data;
327 zt->zt_orig_size = zio->io_size;
328 zt->zt_bufsize = bufsize;
329 zt->zt_transform = transform;
331 zt->zt_next = zio->io_transform_stack;
332 zio->io_transform_stack = zt;
339 zio_pop_transforms(zio_t *zio)
343 while ((zt = zio->io_transform_stack) != NULL) {
344 if (zt->zt_transform != NULL)
345 zt->zt_transform(zio,
346 zt->zt_orig_data, zt->zt_orig_size);
348 if (zt->zt_bufsize != 0)
349 zio_buf_free(zio->io_data, zt->zt_bufsize);
351 zio->io_data = zt->zt_orig_data;
352 zio->io_size = zt->zt_orig_size;
353 zio->io_transform_stack = zt->zt_next;
355 kmem_free(zt, sizeof (zio_transform_t));
360 * ==========================================================================
361 * I/O transform callbacks for subblocks and decompression
362 * ==========================================================================
365 zio_subblock(zio_t *zio, void *data, uint64_t size)
367 ASSERT(zio->io_size > size);
369 if (zio->io_type == ZIO_TYPE_READ)
370 bcopy(zio->io_data, data, size);
374 zio_decompress(zio_t *zio, void *data, uint64_t size)
376 if (zio->io_error == 0 &&
377 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
378 zio->io_data, data, zio->io_size, size) != 0)
379 zio->io_error = SET_ERROR(EIO);
383 * ==========================================================================
384 * I/O parent/child relationships and pipeline interlocks
385 * ==========================================================================
388 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
389 * continue calling these functions until they return NULL.
390 * Otherwise, the next caller will pick up the list walk in
391 * some indeterminate state. (Otherwise every caller would
392 * have to pass in a cookie to keep the state represented by
393 * io_walk_link, which gets annoying.)
396 zio_walk_parents(zio_t *cio)
398 zio_link_t *zl = cio->io_walk_link;
399 list_t *pl = &cio->io_parent_list;
401 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
402 cio->io_walk_link = zl;
407 ASSERT(zl->zl_child == cio);
408 return (zl->zl_parent);
412 zio_walk_children(zio_t *pio)
414 zio_link_t *zl = pio->io_walk_link;
415 list_t *cl = &pio->io_child_list;
417 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
418 pio->io_walk_link = zl;
423 ASSERT(zl->zl_parent == pio);
424 return (zl->zl_child);
428 zio_unique_parent(zio_t *cio)
430 zio_t *pio = zio_walk_parents(cio);
432 VERIFY(zio_walk_parents(cio) == NULL);
437 zio_add_child(zio_t *pio, zio_t *cio)
439 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
442 * Logical I/Os can have logical, gang, or vdev children.
443 * Gang I/Os can have gang or vdev children.
444 * Vdev I/Os can only have vdev children.
445 * The following ASSERT captures all of these constraints.
447 ASSERT(cio->io_child_type <= pio->io_child_type);
452 mutex_enter(&cio->io_lock);
453 mutex_enter(&pio->io_lock);
455 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
457 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
458 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
460 list_insert_head(&pio->io_child_list, zl);
461 list_insert_head(&cio->io_parent_list, zl);
463 pio->io_child_count++;
464 cio->io_parent_count++;
466 mutex_exit(&pio->io_lock);
467 mutex_exit(&cio->io_lock);
471 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
473 ASSERT(zl->zl_parent == pio);
474 ASSERT(zl->zl_child == cio);
476 mutex_enter(&cio->io_lock);
477 mutex_enter(&pio->io_lock);
479 list_remove(&pio->io_child_list, zl);
480 list_remove(&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);
488 kmem_cache_free(zio_link_cache, zl);
492 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
494 uint64_t *countp = &zio->io_children[child][wait];
495 boolean_t waiting = B_FALSE;
497 mutex_enter(&zio->io_lock);
498 ASSERT(zio->io_stall == NULL);
501 zio->io_stall = countp;
504 mutex_exit(&zio->io_lock);
510 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
512 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
513 int *errorp = &pio->io_child_error[zio->io_child_type];
515 mutex_enter(&pio->io_lock);
516 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
517 *errorp = zio_worst_error(*errorp, zio->io_error);
518 pio->io_reexecute |= zio->io_reexecute;
519 ASSERT3U(*countp, >, 0);
523 if (*countp == 0 && pio->io_stall == countp) {
524 pio->io_stall = NULL;
525 mutex_exit(&pio->io_lock);
528 mutex_exit(&pio->io_lock);
533 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
535 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
536 zio->io_error = zio->io_child_error[c];
540 * ==========================================================================
541 * Create the various types of I/O (read, write, free, etc)
542 * ==========================================================================
545 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
546 void *data, uint64_t size, zio_done_func_t *done, void *private,
547 zio_type_t type, zio_priority_t priority, enum zio_flag flags,
548 vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
549 enum zio_stage stage, enum zio_stage pipeline)
553 ASSERT3U(type == ZIO_TYPE_FREE || size, <=, SPA_MAXBLOCKSIZE);
554 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
555 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
557 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
558 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
559 ASSERT(vd || stage == ZIO_STAGE_OPEN);
561 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
562 bzero(zio, sizeof (zio_t));
564 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
565 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
567 list_create(&zio->io_parent_list, sizeof (zio_link_t),
568 offsetof(zio_link_t, zl_parent_node));
569 list_create(&zio->io_child_list, sizeof (zio_link_t),
570 offsetof(zio_link_t, zl_child_node));
573 zio->io_child_type = ZIO_CHILD_VDEV;
574 else if (flags & ZIO_FLAG_GANG_CHILD)
575 zio->io_child_type = ZIO_CHILD_GANG;
576 else if (flags & ZIO_FLAG_DDT_CHILD)
577 zio->io_child_type = ZIO_CHILD_DDT;
579 zio->io_child_type = ZIO_CHILD_LOGICAL;
582 zio->io_bp = (blkptr_t *)bp;
583 zio->io_bp_copy = *bp;
584 zio->io_bp_orig = *bp;
585 if (type != ZIO_TYPE_WRITE ||
586 zio->io_child_type == ZIO_CHILD_DDT)
587 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
588 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
589 zio->io_logical = zio;
590 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
591 pipeline |= ZIO_GANG_STAGES;
597 zio->io_private = private;
599 zio->io_priority = priority;
601 zio->io_offset = offset;
602 zio->io_orig_data = zio->io_data = data;
603 zio->io_orig_size = zio->io_size = size;
604 zio->io_orig_flags = zio->io_flags = flags;
605 zio->io_orig_stage = zio->io_stage = stage;
606 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
608 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
609 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
612 zio->io_bookmark = *zb;
615 if (zio->io_logical == NULL)
616 zio->io_logical = pio->io_logical;
617 if (zio->io_child_type == ZIO_CHILD_GANG)
618 zio->io_gang_leader = pio->io_gang_leader;
619 zio_add_child(pio, zio);
626 zio_destroy(zio_t *zio)
628 list_destroy(&zio->io_parent_list);
629 list_destroy(&zio->io_child_list);
630 mutex_destroy(&zio->io_lock);
631 cv_destroy(&zio->io_cv);
632 kmem_cache_free(zio_cache, zio);
636 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
637 void *private, enum zio_flag flags)
641 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
642 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
643 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
649 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
651 return (zio_null(NULL, spa, NULL, done, private, flags));
655 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
656 void *data, uint64_t size, zio_done_func_t *done, void *private,
657 zio_priority_t priority, enum zio_flag flags, const zbookmark_t *zb)
661 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
662 data, size, done, private,
663 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
664 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
665 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
671 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
672 void *data, uint64_t size, const zio_prop_t *zp,
673 zio_done_func_t *ready, zio_done_func_t *physdone, zio_done_func_t *done,
675 zio_priority_t priority, enum zio_flag flags, const zbookmark_t *zb)
679 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
680 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
681 zp->zp_compress >= ZIO_COMPRESS_OFF &&
682 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
683 DMU_OT_IS_VALID(zp->zp_type) &&
686 zp->zp_copies <= spa_max_replication(spa));
688 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
689 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
690 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
691 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
693 zio->io_ready = ready;
694 zio->io_physdone = physdone;
701 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
702 uint64_t size, zio_done_func_t *done, void *private,
703 zio_priority_t priority, enum zio_flag flags, zbookmark_t *zb)
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, ZIO_REWRITE_PIPELINE);
715 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
717 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
718 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
719 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
720 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
723 * We must reset the io_prop to match the values that existed
724 * when the bp was first written by dmu_sync() keeping in mind
725 * that nopwrite and dedup are mutually exclusive.
727 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
728 zio->io_prop.zp_nopwrite = nopwrite;
729 zio->io_prop.zp_copies = copies;
730 zio->io_bp_override = bp;
734 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
736 metaslab_check_free(spa, bp);
739 * Frees that are for the currently-syncing txg, are not going to be
740 * deferred, and which will not need to do a read (i.e. not GANG or
741 * DEDUP), can be processed immediately. Otherwise, put them on the
742 * in-memory list for later processing.
744 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
745 txg != spa->spa_syncing_txg ||
746 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
747 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
749 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp,
750 BP_GET_PSIZE(bp), 0)));
755 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
756 uint64_t size, enum zio_flag flags)
759 enum zio_stage stage = ZIO_FREE_PIPELINE;
761 dprintf_bp(bp, "freeing in txg %llu, pass %u",
762 (longlong_t)txg, spa->spa_sync_pass);
764 ASSERT(!BP_IS_HOLE(bp));
765 ASSERT(spa_syncing_txg(spa) == txg);
766 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
768 metaslab_check_free(spa, bp);
771 if (zfs_trim_enabled)
772 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START |
773 ZIO_STAGE_VDEV_IO_ASSESS;
775 * GANG and DEDUP blocks can induce a read (for the gang block header,
776 * or the DDT), so issue them asynchronously so that this thread is
779 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
780 stage |= ZIO_STAGE_ISSUE_ASYNC;
782 zio = zio_create(pio, spa, txg, bp, NULL, size,
783 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags,
784 NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
790 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
791 zio_done_func_t *done, void *private, enum zio_flag flags)
796 * A claim is an allocation of a specific block. Claims are needed
797 * to support immediate writes in the intent log. The issue is that
798 * immediate writes contain committed data, but in a txg that was
799 * *not* committed. Upon opening the pool after an unclean shutdown,
800 * the intent log claims all blocks that contain immediate write data
801 * so that the SPA knows they're in use.
803 * All claims *must* be resolved in the first txg -- before the SPA
804 * starts allocating blocks -- so that nothing is allocated twice.
805 * If txg == 0 we just verify that the block is claimable.
807 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
808 ASSERT(txg == spa_first_txg(spa) || txg == 0);
809 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
811 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
812 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
813 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
819 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset,
820 uint64_t size, zio_done_func_t *done, void *private,
826 if (vd->vdev_children == 0) {
827 zio = zio_create(pio, spa, 0, NULL, NULL, size, done, private,
828 ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, offset, NULL,
829 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
833 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
835 for (c = 0; c < vd->vdev_children; c++)
836 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
837 offset, size, done, private, flags));
844 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
845 void *data, int checksum, zio_done_func_t *done, void *private,
846 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
850 ASSERT(vd->vdev_children == 0);
851 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
852 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
853 ASSERT3U(offset + size, <=, vd->vdev_psize);
855 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
856 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
857 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
859 zio->io_prop.zp_checksum = checksum;
865 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
866 void *data, int checksum, zio_done_func_t *done, void *private,
867 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
871 ASSERT(vd->vdev_children == 0);
872 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
873 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
874 ASSERT3U(offset + size, <=, vd->vdev_psize);
876 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
877 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
878 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
880 zio->io_prop.zp_checksum = checksum;
882 if (zio_checksum_table[checksum].ci_eck) {
884 * zec checksums are necessarily destructive -- they modify
885 * the end of the write buffer to hold the verifier/checksum.
886 * Therefore, we must make a local copy in case the data is
887 * being written to multiple places in parallel.
889 void *wbuf = zio_buf_alloc(size);
890 bcopy(data, wbuf, size);
891 zio_push_transform(zio, wbuf, size, size, NULL);
898 * Create a child I/O to do some work for us.
901 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
902 void *data, uint64_t size, int type, zio_priority_t priority,
903 enum zio_flag flags, zio_done_func_t *done, void *private)
905 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
908 ASSERT(vd->vdev_parent ==
909 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
911 if (type == ZIO_TYPE_READ && bp != NULL) {
913 * If we have the bp, then the child should perform the
914 * checksum and the parent need not. This pushes error
915 * detection as close to the leaves as possible and
916 * eliminates redundant checksums in the interior nodes.
918 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
919 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
922 if (vd->vdev_children == 0)
923 offset += VDEV_LABEL_START_SIZE;
925 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
928 * If we've decided to do a repair, the write is not speculative --
929 * even if the original read was.
931 if (flags & ZIO_FLAG_IO_REPAIR)
932 flags &= ~ZIO_FLAG_SPECULATIVE;
934 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
935 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
936 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
938 zio->io_physdone = pio->io_physdone;
939 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
940 zio->io_logical->io_phys_children++;
946 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
947 int type, zio_priority_t priority, enum zio_flag flags,
948 zio_done_func_t *done, void *private)
952 ASSERT(vd->vdev_ops->vdev_op_leaf);
954 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
955 data, size, done, private, type, priority,
956 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
958 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
964 zio_flush(zio_t *zio, vdev_t *vd)
966 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0,
968 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
972 zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size)
975 ASSERT(vd->vdev_ops->vdev_op_leaf);
977 return zio_ioctl(zio, spa, vd, DKIOCTRIM, offset, size,
979 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY);
983 zio_shrink(zio_t *zio, uint64_t size)
985 ASSERT(zio->io_executor == NULL);
986 ASSERT(zio->io_orig_size == zio->io_size);
987 ASSERT(size <= zio->io_size);
990 * We don't shrink for raidz because of problems with the
991 * reconstruction when reading back less than the block size.
992 * Note, BP_IS_RAIDZ() assumes no compression.
994 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
995 if (!BP_IS_RAIDZ(zio->io_bp))
996 zio->io_orig_size = zio->io_size = size;
1000 * ==========================================================================
1001 * Prepare to read and write logical blocks
1002 * ==========================================================================
1006 zio_read_bp_init(zio_t **ziop)
1009 blkptr_t *bp = zio->io_bp;
1011 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1012 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1013 !(zio->io_flags & ZIO_FLAG_RAW)) {
1014 uint64_t psize = BP_GET_PSIZE(bp);
1015 void *cbuf = zio_buf_alloc(psize);
1017 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
1020 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1021 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1023 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1024 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1026 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1027 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1029 return (ZIO_PIPELINE_CONTINUE);
1033 zio_write_bp_init(zio_t **ziop)
1036 spa_t *spa = zio->io_spa;
1037 zio_prop_t *zp = &zio->io_prop;
1038 enum zio_compress compress = zp->zp_compress;
1039 blkptr_t *bp = zio->io_bp;
1040 uint64_t lsize = zio->io_size;
1041 uint64_t psize = lsize;
1045 * If our children haven't all reached the ready stage,
1046 * wait for them and then repeat this pipeline stage.
1048 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1049 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1050 return (ZIO_PIPELINE_STOP);
1052 if (!IO_IS_ALLOCATING(zio))
1053 return (ZIO_PIPELINE_CONTINUE);
1055 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1057 if (zio->io_bp_override) {
1058 ASSERT(bp->blk_birth != zio->io_txg);
1059 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1061 *bp = *zio->io_bp_override;
1062 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1065 * If we've been overridden and nopwrite is set then
1066 * set the flag accordingly to indicate that a nopwrite
1067 * has already occurred.
1069 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1070 ASSERT(!zp->zp_dedup);
1071 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1072 return (ZIO_PIPELINE_CONTINUE);
1075 ASSERT(!zp->zp_nopwrite);
1077 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1078 return (ZIO_PIPELINE_CONTINUE);
1080 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1081 zp->zp_dedup_verify);
1083 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1084 BP_SET_DEDUP(bp, 1);
1085 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1086 return (ZIO_PIPELINE_CONTINUE);
1088 zio->io_bp_override = NULL;
1092 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1094 * We're rewriting an existing block, which means we're
1095 * working on behalf of spa_sync(). For spa_sync() to
1096 * converge, it must eventually be the case that we don't
1097 * have to allocate new blocks. But compression changes
1098 * the blocksize, which forces a reallocate, and makes
1099 * convergence take longer. Therefore, after the first
1100 * few passes, stop compressing to ensure convergence.
1102 pass = spa_sync_pass(spa);
1104 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1105 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1106 ASSERT(!BP_GET_DEDUP(bp));
1108 if (pass >= zfs_sync_pass_dont_compress)
1109 compress = ZIO_COMPRESS_OFF;
1111 /* Make sure someone doesn't change their mind on overwrites */
1112 ASSERT(MIN(zp->zp_copies + BP_IS_GANG(bp),
1113 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1116 if (compress != ZIO_COMPRESS_OFF) {
1117 metaslab_class_t *mc = spa_normal_class(spa);
1118 void *cbuf = zio_buf_alloc(lsize);
1119 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize,
1120 (size_t)metaslab_class_get_minblocksize(mc));
1121 if (psize == 0 || psize == lsize) {
1122 compress = ZIO_COMPRESS_OFF;
1123 zio_buf_free(cbuf, lsize);
1125 ASSERT(psize < lsize);
1126 zio_push_transform(zio, cbuf, psize, lsize, NULL);
1131 * The final pass of spa_sync() must be all rewrites, but the first
1132 * few passes offer a trade-off: allocating blocks defers convergence,
1133 * but newly allocated blocks are sequential, so they can be written
1134 * to disk faster. Therefore, we allow the first few passes of
1135 * spa_sync() to allocate new blocks, but force rewrites after that.
1136 * There should only be a handful of blocks after pass 1 in any case.
1138 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1139 BP_GET_PSIZE(bp) == psize &&
1140 pass >= zfs_sync_pass_rewrite) {
1142 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1143 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1144 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1147 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1151 if (zio->io_bp_orig.blk_birth != 0 &&
1152 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1153 BP_SET_LSIZE(bp, lsize);
1154 BP_SET_TYPE(bp, zp->zp_type);
1155 BP_SET_LEVEL(bp, zp->zp_level);
1156 BP_SET_BIRTH(bp, zio->io_txg, 0);
1158 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1160 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1161 BP_SET_LSIZE(bp, lsize);
1162 BP_SET_TYPE(bp, zp->zp_type);
1163 BP_SET_LEVEL(bp, zp->zp_level);
1164 BP_SET_PSIZE(bp, psize);
1165 BP_SET_COMPRESS(bp, compress);
1166 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1167 BP_SET_DEDUP(bp, zp->zp_dedup);
1168 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1170 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1171 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1172 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1174 if (zp->zp_nopwrite) {
1175 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1176 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1177 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1181 return (ZIO_PIPELINE_CONTINUE);
1185 zio_free_bp_init(zio_t **ziop)
1188 blkptr_t *bp = zio->io_bp;
1190 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1191 if (BP_GET_DEDUP(bp))
1192 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1195 return (ZIO_PIPELINE_CONTINUE);
1199 * ==========================================================================
1200 * Execute the I/O pipeline
1201 * ==========================================================================
1205 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1207 spa_t *spa = zio->io_spa;
1208 zio_type_t t = zio->io_type;
1209 int flags = (cutinline ? TQ_FRONT : 0);
1211 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1214 * If we're a config writer or a probe, the normal issue and
1215 * interrupt threads may all be blocked waiting for the config lock.
1216 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1218 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1222 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1224 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1228 * If this is a high priority I/O, then use the high priority taskq if
1231 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1232 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1235 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1238 * NB: We are assuming that the zio can only be dispatched
1239 * to a single taskq at a time. It would be a grievous error
1240 * to dispatch the zio to another taskq at the same time.
1242 #if defined(illumos) || !defined(_KERNEL)
1243 ASSERT(zio->io_tqent.tqent_next == NULL);
1245 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
1247 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1248 flags, &zio->io_tqent);
1252 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1254 kthread_t *executor = zio->io_executor;
1255 spa_t *spa = zio->io_spa;
1257 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1258 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1260 for (i = 0; i < tqs->stqs_count; i++) {
1261 if (taskq_member(tqs->stqs_taskq[i], executor))
1270 zio_issue_async(zio_t **ziop)
1274 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1276 return (ZIO_PIPELINE_STOP);
1280 zio_interrupt(zio_t *zio)
1282 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1286 * Execute the I/O pipeline until one of the following occurs:
1288 * (1) the I/O completes
1289 * (2) the pipeline stalls waiting for dependent child I/Os
1290 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1291 * (4) the I/O is delegated by vdev-level caching or aggregation
1292 * (5) the I/O is deferred due to vdev-level queueing
1293 * (6) the I/O is handed off to another thread.
1295 * In all cases, the pipeline stops whenever there's no CPU work; it never
1296 * burns a thread in cv_wait().
1298 * There's no locking on io_stage because there's no legitimate way
1299 * for multiple threads to be attempting to process the same I/O.
1301 static zio_pipe_stage_t *zio_pipeline[];
1304 zio_execute(zio_t *zio)
1306 zio->io_executor = curthread;
1308 while (zio->io_stage < ZIO_STAGE_DONE) {
1309 enum zio_stage pipeline = zio->io_pipeline;
1310 enum zio_stage stage = zio->io_stage;
1313 ASSERT(!MUTEX_HELD(&zio->io_lock));
1314 ASSERT(ISP2(stage));
1315 ASSERT(zio->io_stall == NULL);
1319 } while ((stage & pipeline) == 0);
1321 ASSERT(stage <= ZIO_STAGE_DONE);
1324 * If we are in interrupt context and this pipeline stage
1325 * will grab a config lock that is held across I/O,
1326 * or may wait for an I/O that needs an interrupt thread
1327 * to complete, issue async to avoid deadlock.
1329 * For VDEV_IO_START, we cut in line so that the io will
1330 * be sent to disk promptly.
1332 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1333 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1334 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1335 zio_requeue_io_start_cut_in_line : B_FALSE;
1336 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1340 zio->io_stage = stage;
1341 rv = zio_pipeline[highbit64(stage) - 1](&zio);
1343 if (rv == ZIO_PIPELINE_STOP)
1346 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1351 * ==========================================================================
1352 * Initiate I/O, either sync or async
1353 * ==========================================================================
1356 zio_wait(zio_t *zio)
1360 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1361 ASSERT(zio->io_executor == NULL);
1363 zio->io_waiter = curthread;
1367 mutex_enter(&zio->io_lock);
1368 while (zio->io_executor != NULL)
1369 cv_wait(&zio->io_cv, &zio->io_lock);
1370 mutex_exit(&zio->io_lock);
1372 error = zio->io_error;
1379 zio_nowait(zio_t *zio)
1381 ASSERT(zio->io_executor == NULL);
1383 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1384 zio_unique_parent(zio) == NULL) {
1386 * This is a logical async I/O with no parent to wait for it.
1387 * We add it to the spa_async_root_zio "Godfather" I/O which
1388 * will ensure they complete prior to unloading the pool.
1390 spa_t *spa = zio->io_spa;
1392 zio_add_child(spa->spa_async_zio_root, zio);
1399 * ==========================================================================
1400 * Reexecute or suspend/resume failed I/O
1401 * ==========================================================================
1405 zio_reexecute(zio_t *pio)
1407 zio_t *cio, *cio_next;
1409 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1410 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1411 ASSERT(pio->io_gang_leader == NULL);
1412 ASSERT(pio->io_gang_tree == NULL);
1414 pio->io_flags = pio->io_orig_flags;
1415 pio->io_stage = pio->io_orig_stage;
1416 pio->io_pipeline = pio->io_orig_pipeline;
1417 pio->io_reexecute = 0;
1418 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1420 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1421 pio->io_state[w] = 0;
1422 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1423 pio->io_child_error[c] = 0;
1425 if (IO_IS_ALLOCATING(pio))
1426 BP_ZERO(pio->io_bp);
1429 * As we reexecute pio's children, new children could be created.
1430 * New children go to the head of pio's io_child_list, however,
1431 * so we will (correctly) not reexecute them. The key is that
1432 * the remainder of pio's io_child_list, from 'cio_next' onward,
1433 * cannot be affected by any side effects of reexecuting 'cio'.
1435 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1436 cio_next = zio_walk_children(pio);
1437 mutex_enter(&pio->io_lock);
1438 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1439 pio->io_children[cio->io_child_type][w]++;
1440 mutex_exit(&pio->io_lock);
1445 * Now that all children have been reexecuted, execute the parent.
1446 * We don't reexecute "The Godfather" I/O here as it's the
1447 * responsibility of the caller to wait on him.
1449 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1454 zio_suspend(spa_t *spa, zio_t *zio)
1456 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1457 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1458 "failure and the failure mode property for this pool "
1459 "is set to panic.", spa_name(spa));
1461 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1463 mutex_enter(&spa->spa_suspend_lock);
1465 if (spa->spa_suspend_zio_root == NULL)
1466 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1467 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1468 ZIO_FLAG_GODFATHER);
1470 spa->spa_suspended = B_TRUE;
1473 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1474 ASSERT(zio != spa->spa_suspend_zio_root);
1475 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1476 ASSERT(zio_unique_parent(zio) == NULL);
1477 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1478 zio_add_child(spa->spa_suspend_zio_root, zio);
1481 mutex_exit(&spa->spa_suspend_lock);
1485 zio_resume(spa_t *spa)
1490 * Reexecute all previously suspended i/o.
1492 mutex_enter(&spa->spa_suspend_lock);
1493 spa->spa_suspended = B_FALSE;
1494 cv_broadcast(&spa->spa_suspend_cv);
1495 pio = spa->spa_suspend_zio_root;
1496 spa->spa_suspend_zio_root = NULL;
1497 mutex_exit(&spa->spa_suspend_lock);
1503 return (zio_wait(pio));
1507 zio_resume_wait(spa_t *spa)
1509 mutex_enter(&spa->spa_suspend_lock);
1510 while (spa_suspended(spa))
1511 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1512 mutex_exit(&spa->spa_suspend_lock);
1516 * ==========================================================================
1519 * A gang block is a collection of small blocks that looks to the DMU
1520 * like one large block. When zio_dva_allocate() cannot find a block
1521 * of the requested size, due to either severe fragmentation or the pool
1522 * being nearly full, it calls zio_write_gang_block() to construct the
1523 * block from smaller fragments.
1525 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1526 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1527 * an indirect block: it's an array of block pointers. It consumes
1528 * only one sector and hence is allocatable regardless of fragmentation.
1529 * The gang header's bps point to its gang members, which hold the data.
1531 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1532 * as the verifier to ensure uniqueness of the SHA256 checksum.
1533 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1534 * not the gang header. This ensures that data block signatures (needed for
1535 * deduplication) are independent of how the block is physically stored.
1537 * Gang blocks can be nested: a gang member may itself be a gang block.
1538 * Thus every gang block is a tree in which root and all interior nodes are
1539 * gang headers, and the leaves are normal blocks that contain user data.
1540 * The root of the gang tree is called the gang leader.
1542 * To perform any operation (read, rewrite, free, claim) on a gang block,
1543 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1544 * in the io_gang_tree field of the original logical i/o by recursively
1545 * reading the gang leader and all gang headers below it. This yields
1546 * an in-core tree containing the contents of every gang header and the
1547 * bps for every constituent of the gang block.
1549 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1550 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1551 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1552 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1553 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1554 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1555 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1556 * of the gang header plus zio_checksum_compute() of the data to update the
1557 * gang header's blk_cksum as described above.
1559 * The two-phase assemble/issue model solves the problem of partial failure --
1560 * what if you'd freed part of a gang block but then couldn't read the
1561 * gang header for another part? Assembling the entire gang tree first
1562 * ensures that all the necessary gang header I/O has succeeded before
1563 * starting the actual work of free, claim, or write. Once the gang tree
1564 * is assembled, free and claim are in-memory operations that cannot fail.
1566 * In the event that a gang write fails, zio_dva_unallocate() walks the
1567 * gang tree to immediately free (i.e. insert back into the space map)
1568 * everything we've allocated. This ensures that we don't get ENOSPC
1569 * errors during repeated suspend/resume cycles due to a flaky device.
1571 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1572 * the gang tree, we won't modify the block, so we can safely defer the free
1573 * (knowing that the block is still intact). If we *can* assemble the gang
1574 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1575 * each constituent bp and we can allocate a new block on the next sync pass.
1577 * In all cases, the gang tree allows complete recovery from partial failure.
1578 * ==========================================================================
1582 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1587 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1588 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1589 &pio->io_bookmark));
1593 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1598 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1599 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1600 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1602 * As we rewrite each gang header, the pipeline will compute
1603 * a new gang block header checksum for it; but no one will
1604 * compute a new data checksum, so we do that here. The one
1605 * exception is the gang leader: the pipeline already computed
1606 * its data checksum because that stage precedes gang assembly.
1607 * (Presently, nothing actually uses interior data checksums;
1608 * this is just good hygiene.)
1610 if (gn != pio->io_gang_leader->io_gang_tree) {
1611 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1612 data, BP_GET_PSIZE(bp));
1615 * If we are here to damage data for testing purposes,
1616 * leave the GBH alone so that we can detect the damage.
1618 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1619 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1621 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1622 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1623 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1631 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1633 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1634 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp),
1635 ZIO_GANG_CHILD_FLAGS(pio)));
1640 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1642 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1643 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1646 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1655 static void zio_gang_tree_assemble_done(zio_t *zio);
1657 static zio_gang_node_t *
1658 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1660 zio_gang_node_t *gn;
1662 ASSERT(*gnpp == NULL);
1664 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1665 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1672 zio_gang_node_free(zio_gang_node_t **gnpp)
1674 zio_gang_node_t *gn = *gnpp;
1676 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1677 ASSERT(gn->gn_child[g] == NULL);
1679 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1680 kmem_free(gn, sizeof (*gn));
1685 zio_gang_tree_free(zio_gang_node_t **gnpp)
1687 zio_gang_node_t *gn = *gnpp;
1692 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1693 zio_gang_tree_free(&gn->gn_child[g]);
1695 zio_gang_node_free(gnpp);
1699 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1701 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1703 ASSERT(gio->io_gang_leader == gio);
1704 ASSERT(BP_IS_GANG(bp));
1706 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1707 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1708 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1712 zio_gang_tree_assemble_done(zio_t *zio)
1714 zio_t *gio = zio->io_gang_leader;
1715 zio_gang_node_t *gn = zio->io_private;
1716 blkptr_t *bp = zio->io_bp;
1718 ASSERT(gio == zio_unique_parent(zio));
1719 ASSERT(zio->io_child_count == 0);
1724 if (BP_SHOULD_BYTESWAP(bp))
1725 byteswap_uint64_array(zio->io_data, zio->io_size);
1727 ASSERT(zio->io_data == gn->gn_gbh);
1728 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1729 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1731 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1732 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1733 if (!BP_IS_GANG(gbp))
1735 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1740 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1742 zio_t *gio = pio->io_gang_leader;
1745 ASSERT(BP_IS_GANG(bp) == !!gn);
1746 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1747 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1750 * If you're a gang header, your data is in gn->gn_gbh.
1751 * If you're a gang member, your data is in 'data' and gn == NULL.
1753 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1756 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1758 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1759 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1760 if (BP_IS_HOLE(gbp))
1762 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1763 data = (char *)data + BP_GET_PSIZE(gbp);
1767 if (gn == gio->io_gang_tree && gio->io_data != NULL)
1768 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1775 zio_gang_assemble(zio_t **ziop)
1778 blkptr_t *bp = zio->io_bp;
1780 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1781 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1783 zio->io_gang_leader = zio;
1785 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1787 return (ZIO_PIPELINE_CONTINUE);
1791 zio_gang_issue(zio_t **ziop)
1794 blkptr_t *bp = zio->io_bp;
1796 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1797 return (ZIO_PIPELINE_STOP);
1799 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1800 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1802 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1803 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1805 zio_gang_tree_free(&zio->io_gang_tree);
1807 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1809 return (ZIO_PIPELINE_CONTINUE);
1813 zio_write_gang_member_ready(zio_t *zio)
1815 zio_t *pio = zio_unique_parent(zio);
1816 zio_t *gio = zio->io_gang_leader;
1817 dva_t *cdva = zio->io_bp->blk_dva;
1818 dva_t *pdva = pio->io_bp->blk_dva;
1821 if (BP_IS_HOLE(zio->io_bp))
1824 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1826 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1827 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1828 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1829 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1830 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1832 mutex_enter(&pio->io_lock);
1833 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1834 ASSERT(DVA_GET_GANG(&pdva[d]));
1835 asize = DVA_GET_ASIZE(&pdva[d]);
1836 asize += DVA_GET_ASIZE(&cdva[d]);
1837 DVA_SET_ASIZE(&pdva[d], asize);
1839 mutex_exit(&pio->io_lock);
1843 zio_write_gang_block(zio_t *pio)
1845 spa_t *spa = pio->io_spa;
1846 blkptr_t *bp = pio->io_bp;
1847 zio_t *gio = pio->io_gang_leader;
1849 zio_gang_node_t *gn, **gnpp;
1850 zio_gbh_phys_t *gbh;
1851 uint64_t txg = pio->io_txg;
1852 uint64_t resid = pio->io_size;
1854 int copies = gio->io_prop.zp_copies;
1855 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1859 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1860 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1861 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1863 pio->io_error = error;
1864 return (ZIO_PIPELINE_CONTINUE);
1868 gnpp = &gio->io_gang_tree;
1870 gnpp = pio->io_private;
1871 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1874 gn = zio_gang_node_alloc(gnpp);
1876 bzero(gbh, SPA_GANGBLOCKSIZE);
1879 * Create the gang header.
1881 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1882 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1885 * Create and nowait the gang children.
1887 for (int g = 0; resid != 0; resid -= lsize, g++) {
1888 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1890 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1892 zp.zp_checksum = gio->io_prop.zp_checksum;
1893 zp.zp_compress = ZIO_COMPRESS_OFF;
1894 zp.zp_type = DMU_OT_NONE;
1896 zp.zp_copies = gio->io_prop.zp_copies;
1897 zp.zp_dedup = B_FALSE;
1898 zp.zp_dedup_verify = B_FALSE;
1899 zp.zp_nopwrite = B_FALSE;
1901 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1902 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1903 zio_write_gang_member_ready, NULL, NULL, &gn->gn_child[g],
1904 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1905 &pio->io_bookmark));
1909 * Set pio's pipeline to just wait for zio to finish.
1911 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1915 return (ZIO_PIPELINE_CONTINUE);
1919 * The zio_nop_write stage in the pipeline determines if allocating
1920 * a new bp is necessary. By leveraging a cryptographically secure checksum,
1921 * such as SHA256, we can compare the checksums of the new data and the old
1922 * to determine if allocating a new block is required. The nopwrite
1923 * feature can handle writes in either syncing or open context (i.e. zil
1924 * writes) and as a result is mutually exclusive with dedup.
1927 zio_nop_write(zio_t **ziop)
1930 blkptr_t *bp = zio->io_bp;
1931 blkptr_t *bp_orig = &zio->io_bp_orig;
1932 zio_prop_t *zp = &zio->io_prop;
1934 ASSERT(BP_GET_LEVEL(bp) == 0);
1935 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1936 ASSERT(zp->zp_nopwrite);
1937 ASSERT(!zp->zp_dedup);
1938 ASSERT(zio->io_bp_override == NULL);
1939 ASSERT(IO_IS_ALLOCATING(zio));
1942 * Check to see if the original bp and the new bp have matching
1943 * characteristics (i.e. same checksum, compression algorithms, etc).
1944 * If they don't then just continue with the pipeline which will
1945 * allocate a new bp.
1947 if (BP_IS_HOLE(bp_orig) ||
1948 !zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup ||
1949 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
1950 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
1951 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
1952 zp->zp_copies != BP_GET_NDVAS(bp_orig))
1953 return (ZIO_PIPELINE_CONTINUE);
1956 * If the checksums match then reset the pipeline so that we
1957 * avoid allocating a new bp and issuing any I/O.
1959 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
1960 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup);
1961 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
1962 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
1963 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
1964 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
1965 sizeof (uint64_t)) == 0);
1968 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1969 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1972 return (ZIO_PIPELINE_CONTINUE);
1976 * ==========================================================================
1978 * ==========================================================================
1981 zio_ddt_child_read_done(zio_t *zio)
1983 blkptr_t *bp = zio->io_bp;
1984 ddt_entry_t *dde = zio->io_private;
1986 zio_t *pio = zio_unique_parent(zio);
1988 mutex_enter(&pio->io_lock);
1989 ddp = ddt_phys_select(dde, bp);
1990 if (zio->io_error == 0)
1991 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
1992 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
1993 dde->dde_repair_data = zio->io_data;
1995 zio_buf_free(zio->io_data, zio->io_size);
1996 mutex_exit(&pio->io_lock);
2000 zio_ddt_read_start(zio_t **ziop)
2003 blkptr_t *bp = zio->io_bp;
2005 ASSERT(BP_GET_DEDUP(bp));
2006 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2007 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2009 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2010 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2011 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2012 ddt_phys_t *ddp = dde->dde_phys;
2013 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2016 ASSERT(zio->io_vsd == NULL);
2019 if (ddp_self == NULL)
2020 return (ZIO_PIPELINE_CONTINUE);
2022 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2023 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2025 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2027 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2028 zio_buf_alloc(zio->io_size), zio->io_size,
2029 zio_ddt_child_read_done, dde, zio->io_priority,
2030 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
2031 &zio->io_bookmark));
2033 return (ZIO_PIPELINE_CONTINUE);
2036 zio_nowait(zio_read(zio, zio->io_spa, bp,
2037 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
2038 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2040 return (ZIO_PIPELINE_CONTINUE);
2044 zio_ddt_read_done(zio_t **ziop)
2047 blkptr_t *bp = zio->io_bp;
2049 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2050 return (ZIO_PIPELINE_STOP);
2052 ASSERT(BP_GET_DEDUP(bp));
2053 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2054 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2056 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2057 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2058 ddt_entry_t *dde = zio->io_vsd;
2060 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2061 return (ZIO_PIPELINE_CONTINUE);
2064 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2065 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2066 return (ZIO_PIPELINE_STOP);
2068 if (dde->dde_repair_data != NULL) {
2069 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2070 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2072 ddt_repair_done(ddt, dde);
2076 ASSERT(zio->io_vsd == NULL);
2078 return (ZIO_PIPELINE_CONTINUE);
2082 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2084 spa_t *spa = zio->io_spa;
2087 * Note: we compare the original data, not the transformed data,
2088 * because when zio->io_bp is an override bp, we will not have
2089 * pushed the I/O transforms. That's an important optimization
2090 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2092 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2093 zio_t *lio = dde->dde_lead_zio[p];
2096 return (lio->io_orig_size != zio->io_orig_size ||
2097 bcmp(zio->io_orig_data, lio->io_orig_data,
2098 zio->io_orig_size) != 0);
2102 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2103 ddt_phys_t *ddp = &dde->dde_phys[p];
2105 if (ddp->ddp_phys_birth != 0) {
2106 arc_buf_t *abuf = NULL;
2107 uint32_t aflags = ARC_WAIT;
2108 blkptr_t blk = *zio->io_bp;
2111 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2115 error = arc_read(NULL, spa, &blk,
2116 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2117 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2118 &aflags, &zio->io_bookmark);
2121 if (arc_buf_size(abuf) != zio->io_orig_size ||
2122 bcmp(abuf->b_data, zio->io_orig_data,
2123 zio->io_orig_size) != 0)
2124 error = SET_ERROR(EEXIST);
2125 VERIFY(arc_buf_remove_ref(abuf, &abuf));
2129 return (error != 0);
2137 zio_ddt_child_write_ready(zio_t *zio)
2139 int p = zio->io_prop.zp_copies;
2140 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2141 ddt_entry_t *dde = zio->io_private;
2142 ddt_phys_t *ddp = &dde->dde_phys[p];
2150 ASSERT(dde->dde_lead_zio[p] == zio);
2152 ddt_phys_fill(ddp, zio->io_bp);
2154 while ((pio = zio_walk_parents(zio)) != NULL)
2155 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2161 zio_ddt_child_write_done(zio_t *zio)
2163 int p = zio->io_prop.zp_copies;
2164 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2165 ddt_entry_t *dde = zio->io_private;
2166 ddt_phys_t *ddp = &dde->dde_phys[p];
2170 ASSERT(ddp->ddp_refcnt == 0);
2171 ASSERT(dde->dde_lead_zio[p] == zio);
2172 dde->dde_lead_zio[p] = NULL;
2174 if (zio->io_error == 0) {
2175 while (zio_walk_parents(zio) != NULL)
2176 ddt_phys_addref(ddp);
2178 ddt_phys_clear(ddp);
2185 zio_ddt_ditto_write_done(zio_t *zio)
2187 int p = DDT_PHYS_DITTO;
2188 zio_prop_t *zp = &zio->io_prop;
2189 blkptr_t *bp = zio->io_bp;
2190 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2191 ddt_entry_t *dde = zio->io_private;
2192 ddt_phys_t *ddp = &dde->dde_phys[p];
2193 ddt_key_t *ddk = &dde->dde_key;
2197 ASSERT(ddp->ddp_refcnt == 0);
2198 ASSERT(dde->dde_lead_zio[p] == zio);
2199 dde->dde_lead_zio[p] = NULL;
2201 if (zio->io_error == 0) {
2202 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2203 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2204 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2205 if (ddp->ddp_phys_birth != 0)
2206 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2207 ddt_phys_fill(ddp, bp);
2214 zio_ddt_write(zio_t **ziop)
2217 spa_t *spa = zio->io_spa;
2218 blkptr_t *bp = zio->io_bp;
2219 uint64_t txg = zio->io_txg;
2220 zio_prop_t *zp = &zio->io_prop;
2221 int p = zp->zp_copies;
2225 ddt_t *ddt = ddt_select(spa, bp);
2229 ASSERT(BP_GET_DEDUP(bp));
2230 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2231 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2234 dde = ddt_lookup(ddt, bp, B_TRUE);
2235 ddp = &dde->dde_phys[p];
2237 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2239 * If we're using a weak checksum, upgrade to a strong checksum
2240 * and try again. If we're already using a strong checksum,
2241 * we can't resolve it, so just convert to an ordinary write.
2242 * (And automatically e-mail a paper to Nature?)
2244 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2245 zp->zp_checksum = spa_dedup_checksum(spa);
2246 zio_pop_transforms(zio);
2247 zio->io_stage = ZIO_STAGE_OPEN;
2250 zp->zp_dedup = B_FALSE;
2252 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2254 return (ZIO_PIPELINE_CONTINUE);
2257 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2258 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2260 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2261 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2262 zio_prop_t czp = *zp;
2264 czp.zp_copies = ditto_copies;
2267 * If we arrived here with an override bp, we won't have run
2268 * the transform stack, so we won't have the data we need to
2269 * generate a child i/o. So, toss the override bp and restart.
2270 * This is safe, because using the override bp is just an
2271 * optimization; and it's rare, so the cost doesn't matter.
2273 if (zio->io_bp_override) {
2274 zio_pop_transforms(zio);
2275 zio->io_stage = ZIO_STAGE_OPEN;
2276 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2277 zio->io_bp_override = NULL;
2280 return (ZIO_PIPELINE_CONTINUE);
2283 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2284 zio->io_orig_size, &czp, NULL, NULL,
2285 zio_ddt_ditto_write_done, dde, zio->io_priority,
2286 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2288 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2289 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2292 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2293 if (ddp->ddp_phys_birth != 0)
2294 ddt_bp_fill(ddp, bp, txg);
2295 if (dde->dde_lead_zio[p] != NULL)
2296 zio_add_child(zio, dde->dde_lead_zio[p]);
2298 ddt_phys_addref(ddp);
2299 } else if (zio->io_bp_override) {
2300 ASSERT(bp->blk_birth == txg);
2301 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2302 ddt_phys_fill(ddp, bp);
2303 ddt_phys_addref(ddp);
2305 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2306 zio->io_orig_size, zp, zio_ddt_child_write_ready, NULL,
2307 zio_ddt_child_write_done, dde, zio->io_priority,
2308 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2310 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2311 dde->dde_lead_zio[p] = cio;
2321 return (ZIO_PIPELINE_CONTINUE);
2324 ddt_entry_t *freedde; /* for debugging */
2327 zio_ddt_free(zio_t **ziop)
2330 spa_t *spa = zio->io_spa;
2331 blkptr_t *bp = zio->io_bp;
2332 ddt_t *ddt = ddt_select(spa, bp);
2336 ASSERT(BP_GET_DEDUP(bp));
2337 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2340 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2341 ddp = ddt_phys_select(dde, bp);
2342 ddt_phys_decref(ddp);
2345 return (ZIO_PIPELINE_CONTINUE);
2349 * ==========================================================================
2350 * Allocate and free blocks
2351 * ==========================================================================
2354 zio_dva_allocate(zio_t **ziop)
2357 spa_t *spa = zio->io_spa;
2358 metaslab_class_t *mc = spa_normal_class(spa);
2359 blkptr_t *bp = zio->io_bp;
2363 if (zio->io_gang_leader == NULL) {
2364 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2365 zio->io_gang_leader = zio;
2368 ASSERT(BP_IS_HOLE(bp));
2369 ASSERT0(BP_GET_NDVAS(bp));
2370 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2371 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2372 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2375 * The dump device does not support gang blocks so allocation on
2376 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2377 * the "fast" gang feature.
2379 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2380 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2381 METASLAB_GANG_CHILD : 0;
2382 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2383 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2386 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2387 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2389 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2390 return (zio_write_gang_block(zio));
2391 zio->io_error = error;
2394 return (ZIO_PIPELINE_CONTINUE);
2398 zio_dva_free(zio_t **ziop)
2402 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2404 return (ZIO_PIPELINE_CONTINUE);
2408 zio_dva_claim(zio_t **ziop)
2413 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2415 zio->io_error = error;
2417 return (ZIO_PIPELINE_CONTINUE);
2421 * Undo an allocation. This is used by zio_done() when an I/O fails
2422 * and we want to give back the block we just allocated.
2423 * This handles both normal blocks and gang blocks.
2426 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2428 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2429 ASSERT(zio->io_bp_override == NULL);
2431 if (!BP_IS_HOLE(bp))
2432 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2435 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2436 zio_dva_unallocate(zio, gn->gn_child[g],
2437 &gn->gn_gbh->zg_blkptr[g]);
2443 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2446 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2447 uint64_t size, boolean_t use_slog)
2451 ASSERT(txg > spa_syncing_txg(spa));
2454 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2455 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2456 * when allocating them.
2459 error = metaslab_alloc(spa, spa_log_class(spa), size,
2460 new_bp, 1, txg, old_bp,
2461 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2465 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2466 new_bp, 1, txg, old_bp,
2467 METASLAB_HINTBP_AVOID);
2471 BP_SET_LSIZE(new_bp, size);
2472 BP_SET_PSIZE(new_bp, size);
2473 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2474 BP_SET_CHECKSUM(new_bp,
2475 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2476 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2477 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2478 BP_SET_LEVEL(new_bp, 0);
2479 BP_SET_DEDUP(new_bp, 0);
2480 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2487 * Free an intent log block.
2490 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2492 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2493 ASSERT(!BP_IS_GANG(bp));
2495 zio_free(spa, txg, bp);
2499 * ==========================================================================
2500 * Read, write and delete to physical devices
2501 * ==========================================================================
2504 zio_vdev_io_start(zio_t **ziop)
2507 vdev_t *vd = zio->io_vd;
2509 spa_t *spa = zio->io_spa;
2511 ASSERT(zio->io_error == 0);
2512 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2515 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2516 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2519 * The mirror_ops handle multiple DVAs in a single BP.
2521 return (vdev_mirror_ops.vdev_op_io_start(zio));
2524 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE) {
2525 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg);
2526 return (ZIO_PIPELINE_CONTINUE);
2530 * We keep track of time-sensitive I/Os so that the scan thread
2531 * can quickly react to certain workloads. In particular, we care
2532 * about non-scrubbing, top-level reads and writes with the following
2534 * - synchronous writes of user data to non-slog devices
2535 * - any reads of user data
2536 * When these conditions are met, adjust the timestamp of spa_last_io
2537 * which allows the scan thread to adjust its workload accordingly.
2539 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2540 vd == vd->vdev_top && !vd->vdev_islog &&
2541 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2542 zio->io_txg != spa_syncing_txg(spa)) {
2543 uint64_t old = spa->spa_last_io;
2544 uint64_t new = ddi_get_lbolt64();
2546 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2549 align = 1ULL << vd->vdev_top->vdev_ashift;
2551 if (P2PHASE(zio->io_size, align) != 0) {
2552 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2554 if (zio->io_type == ZIO_TYPE_READ ||
2555 zio->io_type == ZIO_TYPE_WRITE)
2556 abuf = zio_buf_alloc(asize);
2557 ASSERT(vd == vd->vdev_top);
2558 if (zio->io_type == ZIO_TYPE_WRITE) {
2559 bcopy(zio->io_data, abuf, zio->io_size);
2560 bzero(abuf + zio->io_size, asize - zio->io_size);
2562 zio_push_transform(zio, abuf, asize, abuf ? asize : 0,
2566 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2567 ASSERT(P2PHASE(zio->io_size, align) == 0);
2568 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa));
2571 * If this is a repair I/O, and there's no self-healing involved --
2572 * that is, we're just resilvering what we expect to resilver --
2573 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2574 * This prevents spurious resilvering with nested replication.
2575 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2576 * A is out of date, we'll read from C+D, then use the data to
2577 * resilver A+B -- but we don't actually want to resilver B, just A.
2578 * The top-level mirror has no way to know this, so instead we just
2579 * discard unnecessary repairs as we work our way down the vdev tree.
2580 * The same logic applies to any form of nested replication:
2581 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2583 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2584 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2585 zio->io_txg != 0 && /* not a delegated i/o */
2586 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2587 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2588 zio_vdev_io_bypass(zio);
2589 return (ZIO_PIPELINE_CONTINUE);
2592 if (vd->vdev_ops->vdev_op_leaf &&
2593 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2595 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio))
2596 return (ZIO_PIPELINE_CONTINUE);
2598 if ((zio = vdev_queue_io(zio)) == NULL)
2599 return (ZIO_PIPELINE_STOP);
2602 if (!vdev_accessible(vd, zio)) {
2603 zio->io_error = SET_ERROR(ENXIO);
2605 return (ZIO_PIPELINE_STOP);
2610 * Note that we ignore repair writes for TRIM because they can conflict
2611 * with normal writes. This isn't an issue because, by definition, we
2612 * only repair blocks that aren't freed.
2614 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_WRITE &&
2615 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2616 if (!trim_map_write_start(zio))
2617 return (ZIO_PIPELINE_STOP);
2620 return (vd->vdev_ops->vdev_op_io_start(zio));
2624 zio_vdev_io_done(zio_t **ziop)
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 **ziop)
2703 vdev_t *vd = zio->io_vd;
2705 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2706 return (ZIO_PIPELINE_STOP);
2708 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2709 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2711 if (zio->io_vsd != NULL) {
2712 zio->io_vsd_ops->vsd_free(zio);
2716 if (zio_injection_enabled && zio->io_error == 0)
2717 zio->io_error = zio_handle_fault_injection(zio, EIO);
2719 if (zio->io_type == ZIO_TYPE_IOCTL && zio->io_cmd == DKIOCTRIM)
2720 switch (zio->io_error) {
2722 ZIO_TRIM_STAT_INCR(bytes, zio->io_size);
2723 ZIO_TRIM_STAT_BUMP(success);
2726 ZIO_TRIM_STAT_BUMP(unsupported);
2729 ZIO_TRIM_STAT_BUMP(failed);
2734 * If the I/O failed, determine whether we should attempt to retry it.
2736 * On retry, we cut in line in the issue queue, since we don't want
2737 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2739 if (zio->io_error && vd == NULL &&
2740 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2741 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2742 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2744 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2745 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2746 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2747 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2748 zio_requeue_io_start_cut_in_line);
2749 return (ZIO_PIPELINE_STOP);
2753 * If we got an error on a leaf device, convert it to ENXIO
2754 * if the device is not accessible at all.
2756 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2757 !vdev_accessible(vd, zio))
2758 zio->io_error = SET_ERROR(ENXIO);
2761 * If we can't write to an interior vdev (mirror or RAID-Z),
2762 * set vdev_cant_write so that we stop trying to allocate from it.
2764 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2765 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
2766 vd->vdev_cant_write = B_TRUE;
2770 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2772 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2773 zio->io_physdone != NULL) {
2774 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
2775 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
2776 zio->io_physdone(zio->io_logical);
2779 return (ZIO_PIPELINE_CONTINUE);
2783 zio_vdev_io_reissue(zio_t *zio)
2785 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2786 ASSERT(zio->io_error == 0);
2788 zio->io_stage >>= 1;
2792 zio_vdev_io_redone(zio_t *zio)
2794 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2796 zio->io_stage >>= 1;
2800 zio_vdev_io_bypass(zio_t *zio)
2802 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2803 ASSERT(zio->io_error == 0);
2805 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2806 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2810 * ==========================================================================
2811 * Generate and verify checksums
2812 * ==========================================================================
2815 zio_checksum_generate(zio_t **ziop)
2818 blkptr_t *bp = zio->io_bp;
2819 enum zio_checksum checksum;
2823 * This is zio_write_phys().
2824 * We're either generating a label checksum, or none at all.
2826 checksum = zio->io_prop.zp_checksum;
2828 if (checksum == ZIO_CHECKSUM_OFF)
2829 return (ZIO_PIPELINE_CONTINUE);
2831 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2833 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2834 ASSERT(!IO_IS_ALLOCATING(zio));
2835 checksum = ZIO_CHECKSUM_GANG_HEADER;
2837 checksum = BP_GET_CHECKSUM(bp);
2841 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2843 return (ZIO_PIPELINE_CONTINUE);
2847 zio_checksum_verify(zio_t **ziop)
2850 zio_bad_cksum_t info;
2851 blkptr_t *bp = zio->io_bp;
2854 ASSERT(zio->io_vd != NULL);
2858 * This is zio_read_phys().
2859 * We're either verifying a label checksum, or nothing at all.
2861 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2862 return (ZIO_PIPELINE_CONTINUE);
2864 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2867 if ((error = zio_checksum_error(zio, &info)) != 0) {
2868 zio->io_error = error;
2869 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2870 zfs_ereport_start_checksum(zio->io_spa,
2871 zio->io_vd, zio, zio->io_offset,
2872 zio->io_size, NULL, &info);
2876 return (ZIO_PIPELINE_CONTINUE);
2880 * Called by RAID-Z to ensure we don't compute the checksum twice.
2883 zio_checksum_verified(zio_t *zio)
2885 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2889 * ==========================================================================
2890 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2891 * An error of 0 indictes success. ENXIO indicates whole-device failure,
2892 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2893 * indicate errors that are specific to one I/O, and most likely permanent.
2894 * Any other error is presumed to be worse because we weren't expecting it.
2895 * ==========================================================================
2898 zio_worst_error(int e1, int e2)
2900 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2903 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2904 if (e1 == zio_error_rank[r1])
2907 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2908 if (e2 == zio_error_rank[r2])
2911 return (r1 > r2 ? e1 : e2);
2915 * ==========================================================================
2917 * ==========================================================================
2920 zio_ready(zio_t **ziop)
2923 blkptr_t *bp = zio->io_bp;
2924 zio_t *pio, *pio_next;
2926 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
2927 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2928 return (ZIO_PIPELINE_STOP);
2930 if (zio->io_ready) {
2931 ASSERT(IO_IS_ALLOCATING(zio));
2932 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
2933 (zio->io_flags & ZIO_FLAG_NOPWRITE));
2934 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2939 if (bp != NULL && bp != &zio->io_bp_copy)
2940 zio->io_bp_copy = *bp;
2943 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2945 mutex_enter(&zio->io_lock);
2946 zio->io_state[ZIO_WAIT_READY] = 1;
2947 pio = zio_walk_parents(zio);
2948 mutex_exit(&zio->io_lock);
2951 * As we notify zio's parents, new parents could be added.
2952 * New parents go to the head of zio's io_parent_list, however,
2953 * so we will (correctly) not notify them. The remainder of zio's
2954 * io_parent_list, from 'pio_next' onward, cannot change because
2955 * all parents must wait for us to be done before they can be done.
2957 for (; pio != NULL; pio = pio_next) {
2958 pio_next = zio_walk_parents(zio);
2959 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2962 if (zio->io_flags & ZIO_FLAG_NODATA) {
2963 if (BP_IS_GANG(bp)) {
2964 zio->io_flags &= ~ZIO_FLAG_NODATA;
2966 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
2967 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2971 if (zio_injection_enabled &&
2972 zio->io_spa->spa_syncing_txg == zio->io_txg)
2973 zio_handle_ignored_writes(zio);
2975 return (ZIO_PIPELINE_CONTINUE);
2979 zio_done(zio_t **ziop)
2982 spa_t *spa = zio->io_spa;
2983 zio_t *lio = zio->io_logical;
2984 blkptr_t *bp = zio->io_bp;
2985 vdev_t *vd = zio->io_vd;
2986 uint64_t psize = zio->io_size;
2987 zio_t *pio, *pio_next;
2990 * If our children haven't all completed,
2991 * wait for them and then repeat this pipeline stage.
2993 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
2994 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
2995 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
2996 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
2997 return (ZIO_PIPELINE_STOP);
2999 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3000 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3001 ASSERT(zio->io_children[c][w] == 0);
3004 ASSERT(bp->blk_pad[0] == 0);
3005 ASSERT(bp->blk_pad[1] == 0);
3006 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3007 (bp == zio_unique_parent(zio)->io_bp));
3008 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3009 zio->io_bp_override == NULL &&
3010 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3011 ASSERT(!BP_SHOULD_BYTESWAP(bp));
3012 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3013 ASSERT(BP_COUNT_GANG(bp) == 0 ||
3014 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3016 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3017 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3021 * If there were child vdev/gang/ddt errors, they apply to us now.
3023 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3024 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3025 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3028 * If the I/O on the transformed data was successful, generate any
3029 * checksum reports now while we still have the transformed data.
3031 if (zio->io_error == 0) {
3032 while (zio->io_cksum_report != NULL) {
3033 zio_cksum_report_t *zcr = zio->io_cksum_report;
3034 uint64_t align = zcr->zcr_align;
3035 uint64_t asize = P2ROUNDUP(psize, align);
3036 char *abuf = zio->io_data;
3038 if (asize != psize) {
3039 abuf = zio_buf_alloc(asize);
3040 bcopy(zio->io_data, abuf, psize);
3041 bzero(abuf + psize, asize - psize);
3044 zio->io_cksum_report = zcr->zcr_next;
3045 zcr->zcr_next = NULL;
3046 zcr->zcr_finish(zcr, abuf);
3047 zfs_ereport_free_checksum(zcr);
3050 zio_buf_free(abuf, asize);
3054 zio_pop_transforms(zio); /* note: may set zio->io_error */
3056 vdev_stat_update(zio, psize);
3058 if (zio->io_error) {
3060 * If this I/O is attached to a particular vdev,
3061 * generate an error message describing the I/O failure
3062 * at the block level. We ignore these errors if the
3063 * device is currently unavailable.
3065 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3066 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3068 if ((zio->io_error == EIO || !(zio->io_flags &
3069 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3072 * For logical I/O requests, tell the SPA to log the
3073 * error and generate a logical data ereport.
3075 spa_log_error(spa, zio);
3076 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3081 if (zio->io_error && zio == lio) {
3083 * Determine whether zio should be reexecuted. This will
3084 * propagate all the way to the root via zio_notify_parent().
3086 ASSERT(vd == NULL && bp != NULL);
3087 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3089 if (IO_IS_ALLOCATING(zio) &&
3090 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3091 if (zio->io_error != ENOSPC)
3092 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3094 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3097 if ((zio->io_type == ZIO_TYPE_READ ||
3098 zio->io_type == ZIO_TYPE_FREE) &&
3099 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3100 zio->io_error == ENXIO &&
3101 spa_load_state(spa) == SPA_LOAD_NONE &&
3102 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3103 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3105 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3106 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3109 * Here is a possibly good place to attempt to do
3110 * either combinatorial reconstruction or error correction
3111 * based on checksums. It also might be a good place
3112 * to send out preliminary ereports before we suspend
3118 * If there were logical child errors, they apply to us now.
3119 * We defer this until now to avoid conflating logical child
3120 * errors with errors that happened to the zio itself when
3121 * updating vdev stats and reporting FMA events above.
3123 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3125 if ((zio->io_error || zio->io_reexecute) &&
3126 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3127 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3128 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3130 zio_gang_tree_free(&zio->io_gang_tree);
3133 * Godfather I/Os should never suspend.
3135 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3136 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3137 zio->io_reexecute = 0;
3139 if (zio->io_reexecute) {
3141 * This is a logical I/O that wants to reexecute.
3143 * Reexecute is top-down. When an i/o fails, if it's not
3144 * the root, it simply notifies its parent and sticks around.
3145 * The parent, seeing that it still has children in zio_done(),
3146 * does the same. This percolates all the way up to the root.
3147 * The root i/o will reexecute or suspend the entire tree.
3149 * This approach ensures that zio_reexecute() honors
3150 * all the original i/o dependency relationships, e.g.
3151 * parents not executing until children are ready.
3153 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3155 zio->io_gang_leader = NULL;
3157 mutex_enter(&zio->io_lock);
3158 zio->io_state[ZIO_WAIT_DONE] = 1;
3159 mutex_exit(&zio->io_lock);
3162 * "The Godfather" I/O monitors its children but is
3163 * not a true parent to them. It will track them through
3164 * the pipeline but severs its ties whenever they get into
3165 * trouble (e.g. suspended). This allows "The Godfather"
3166 * I/O to return status without blocking.
3168 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3169 zio_link_t *zl = zio->io_walk_link;
3170 pio_next = zio_walk_parents(zio);
3172 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3173 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3174 zio_remove_child(pio, zio, zl);
3175 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3179 if ((pio = zio_unique_parent(zio)) != NULL) {
3181 * We're not a root i/o, so there's nothing to do
3182 * but notify our parent. Don't propagate errors
3183 * upward since we haven't permanently failed yet.
3185 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3186 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3187 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3188 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3190 * We'd fail again if we reexecuted now, so suspend
3191 * until conditions improve (e.g. device comes online).
3193 zio_suspend(spa, zio);
3196 * Reexecution is potentially a huge amount of work.
3197 * Hand it off to the otherwise-unused claim taskq.
3199 #if defined(illumos) || !defined(_KERNEL)
3200 ASSERT(zio->io_tqent.tqent_next == NULL);
3202 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
3204 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3205 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3208 return (ZIO_PIPELINE_STOP);
3211 ASSERT(zio->io_child_count == 0);
3212 ASSERT(zio->io_reexecute == 0);
3213 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3216 * Report any checksum errors, since the I/O is complete.
3218 while (zio->io_cksum_report != NULL) {
3219 zio_cksum_report_t *zcr = zio->io_cksum_report;
3220 zio->io_cksum_report = zcr->zcr_next;
3221 zcr->zcr_next = NULL;
3222 zcr->zcr_finish(zcr, NULL);
3223 zfs_ereport_free_checksum(zcr);
3227 * It is the responsibility of the done callback to ensure that this
3228 * particular zio is no longer discoverable for adoption, and as
3229 * such, cannot acquire any new parents.
3234 mutex_enter(&zio->io_lock);
3235 zio->io_state[ZIO_WAIT_DONE] = 1;
3236 mutex_exit(&zio->io_lock);
3238 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3239 zio_link_t *zl = zio->io_walk_link;
3240 pio_next = zio_walk_parents(zio);
3241 zio_remove_child(pio, zio, zl);
3242 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3245 if (zio->io_waiter != NULL) {
3246 mutex_enter(&zio->io_lock);
3247 zio->io_executor = NULL;
3248 cv_broadcast(&zio->io_cv);
3249 mutex_exit(&zio->io_lock);
3254 return (ZIO_PIPELINE_STOP);
3258 * ==========================================================================
3259 * I/O pipeline definition
3260 * ==========================================================================
3262 static zio_pipe_stage_t *zio_pipeline[] = {
3268 zio_checksum_generate,
3283 zio_checksum_verify,
3287 /* dnp is the dnode for zb1->zb_object */
3289 zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_t *zb1,
3290 const zbookmark_t *zb2)
3292 uint64_t zb1nextL0, zb2thisobj;
3294 ASSERT(zb1->zb_objset == zb2->zb_objset);
3295 ASSERT(zb2->zb_level == 0);
3298 * A bookmark in the deadlist is considered to be after
3301 if (zb2->zb_object == DMU_DEADLIST_OBJECT)
3304 /* The objset_phys_t isn't before anything. */
3308 zb1nextL0 = (zb1->zb_blkid + 1) <<
3309 ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
3311 zb2thisobj = zb2->zb_object ? zb2->zb_object :
3312 zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
3314 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3315 uint64_t nextobj = zb1nextL0 *
3316 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
3317 return (nextobj <= zb2thisobj);
3320 if (zb1->zb_object < zb2thisobj)
3322 if (zb1->zb_object > zb2thisobj)
3324 if (zb2->zb_object == DMU_META_DNODE_OBJECT)
3326 return (zb1nextL0 <= zb2->zb_blkid);