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/blkptr.h>
41 #include <sys/zfeature.h>
43 SYSCTL_DECL(_vfs_zfs);
44 SYSCTL_NODE(_vfs_zfs, OID_AUTO, zio, CTLFLAG_RW, 0, "ZFS ZIO");
45 #if defined(__amd64__)
46 static int zio_use_uma = 1;
48 static int zio_use_uma = 0;
50 TUNABLE_INT("vfs.zfs.zio.use_uma", &zio_use_uma);
51 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, use_uma, CTLFLAG_RDTUN, &zio_use_uma, 0,
52 "Use uma(9) for ZIO allocations");
53 static int zio_exclude_metadata = 0;
54 TUNABLE_INT("vfs.zfs.zio.exclude_metadata", &zio_exclude_metadata);
55 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, exclude_metadata, CTLFLAG_RDTUN, &zio_exclude_metadata, 0,
56 "Exclude metadata buffers from dumps as well");
58 zio_trim_stats_t zio_trim_stats = {
59 { "bytes", KSTAT_DATA_UINT64,
60 "Number of bytes successfully TRIMmed" },
61 { "success", KSTAT_DATA_UINT64,
62 "Number of successful TRIM requests" },
63 { "unsupported", KSTAT_DATA_UINT64,
64 "Number of TRIM requests that failed because TRIM is not supported" },
65 { "failed", KSTAT_DATA_UINT64,
66 "Number of TRIM requests that failed for reasons other than not supported" },
69 static kstat_t *zio_trim_ksp;
72 * ==========================================================================
73 * I/O type descriptions
74 * ==========================================================================
76 const char *zio_type_name[ZIO_TYPES] = {
77 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
82 * ==========================================================================
84 * ==========================================================================
86 kmem_cache_t *zio_cache;
87 kmem_cache_t *zio_link_cache;
88 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
89 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
92 extern vmem_t *zio_alloc_arena;
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];
211 zio_trim_ksp = kstat_create("zfs", 0, "zio_trim", "misc",
213 sizeof(zio_trim_stats) / sizeof(kstat_named_t),
216 if (zio_trim_ksp != NULL) {
217 zio_trim_ksp->ks_data = &zio_trim_stats;
218 kstat_install(zio_trim_ksp);
226 kmem_cache_t *last_cache = NULL;
227 kmem_cache_t *last_data_cache = NULL;
229 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
230 if (zio_buf_cache[c] != last_cache) {
231 last_cache = zio_buf_cache[c];
232 kmem_cache_destroy(zio_buf_cache[c]);
234 zio_buf_cache[c] = NULL;
236 if (zio_data_buf_cache[c] != last_data_cache) {
237 last_data_cache = zio_data_buf_cache[c];
238 kmem_cache_destroy(zio_data_buf_cache[c]);
240 zio_data_buf_cache[c] = NULL;
243 kmem_cache_destroy(zio_link_cache);
244 kmem_cache_destroy(zio_cache);
248 if (zio_trim_ksp != NULL) {
249 kstat_delete(zio_trim_ksp);
255 * ==========================================================================
256 * Allocate and free I/O buffers
257 * ==========================================================================
261 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
262 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
263 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
264 * excess / transient data in-core during a crashdump.
267 zio_buf_alloc(size_t size)
269 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
270 int flags = zio_exclude_metadata ? KM_NODEBUG : 0;
272 ASSERT3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
275 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
277 return (kmem_alloc(size, KM_SLEEP|flags));
281 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
282 * crashdump if the kernel panics. This exists so that we will limit the amount
283 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
284 * of kernel heap dumped to disk when the kernel panics)
287 zio_data_buf_alloc(size_t size)
289 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
291 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
294 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
296 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG));
300 zio_buf_free(void *buf, size_t size)
302 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
304 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
307 kmem_cache_free(zio_buf_cache[c], buf);
309 kmem_free(buf, size);
313 zio_data_buf_free(void *buf, size_t size)
315 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
317 ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
320 kmem_cache_free(zio_data_buf_cache[c], buf);
322 kmem_free(buf, size);
326 * ==========================================================================
327 * Push and pop I/O transform buffers
328 * ==========================================================================
331 zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
332 zio_transform_func_t *transform)
334 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
336 zt->zt_orig_data = zio->io_data;
337 zt->zt_orig_size = zio->io_size;
338 zt->zt_bufsize = bufsize;
339 zt->zt_transform = transform;
341 zt->zt_next = zio->io_transform_stack;
342 zio->io_transform_stack = zt;
349 zio_pop_transforms(zio_t *zio)
353 while ((zt = zio->io_transform_stack) != NULL) {
354 if (zt->zt_transform != NULL)
355 zt->zt_transform(zio,
356 zt->zt_orig_data, zt->zt_orig_size);
358 if (zt->zt_bufsize != 0)
359 zio_buf_free(zio->io_data, zt->zt_bufsize);
361 zio->io_data = zt->zt_orig_data;
362 zio->io_size = zt->zt_orig_size;
363 zio->io_transform_stack = zt->zt_next;
365 kmem_free(zt, sizeof (zio_transform_t));
370 * ==========================================================================
371 * I/O transform callbacks for subblocks and decompression
372 * ==========================================================================
375 zio_subblock(zio_t *zio, void *data, uint64_t size)
377 ASSERT(zio->io_size > size);
379 if (zio->io_type == ZIO_TYPE_READ)
380 bcopy(zio->io_data, data, size);
384 zio_decompress(zio_t *zio, void *data, uint64_t size)
386 if (zio->io_error == 0 &&
387 zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
388 zio->io_data, data, zio->io_size, size) != 0)
389 zio->io_error = SET_ERROR(EIO);
393 * ==========================================================================
394 * I/O parent/child relationships and pipeline interlocks
395 * ==========================================================================
398 * NOTE - Callers to zio_walk_parents() and zio_walk_children must
399 * continue calling these functions until they return NULL.
400 * Otherwise, the next caller will pick up the list walk in
401 * some indeterminate state. (Otherwise every caller would
402 * have to pass in a cookie to keep the state represented by
403 * io_walk_link, which gets annoying.)
406 zio_walk_parents(zio_t *cio)
408 zio_link_t *zl = cio->io_walk_link;
409 list_t *pl = &cio->io_parent_list;
411 zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
412 cio->io_walk_link = zl;
417 ASSERT(zl->zl_child == cio);
418 return (zl->zl_parent);
422 zio_walk_children(zio_t *pio)
424 zio_link_t *zl = pio->io_walk_link;
425 list_t *cl = &pio->io_child_list;
427 zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
428 pio->io_walk_link = zl;
433 ASSERT(zl->zl_parent == pio);
434 return (zl->zl_child);
438 zio_unique_parent(zio_t *cio)
440 zio_t *pio = zio_walk_parents(cio);
442 VERIFY(zio_walk_parents(cio) == NULL);
447 zio_add_child(zio_t *pio, zio_t *cio)
449 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
452 * Logical I/Os can have logical, gang, or vdev children.
453 * Gang I/Os can have gang or vdev children.
454 * Vdev I/Os can only have vdev children.
455 * The following ASSERT captures all of these constraints.
457 ASSERT(cio->io_child_type <= pio->io_child_type);
462 mutex_enter(&cio->io_lock);
463 mutex_enter(&pio->io_lock);
465 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
467 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
468 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
470 list_insert_head(&pio->io_child_list, zl);
471 list_insert_head(&cio->io_parent_list, zl);
473 pio->io_child_count++;
474 cio->io_parent_count++;
476 mutex_exit(&pio->io_lock);
477 mutex_exit(&cio->io_lock);
481 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
483 ASSERT(zl->zl_parent == pio);
484 ASSERT(zl->zl_child == cio);
486 mutex_enter(&cio->io_lock);
487 mutex_enter(&pio->io_lock);
489 list_remove(&pio->io_child_list, zl);
490 list_remove(&cio->io_parent_list, zl);
492 pio->io_child_count--;
493 cio->io_parent_count--;
495 mutex_exit(&pio->io_lock);
496 mutex_exit(&cio->io_lock);
498 kmem_cache_free(zio_link_cache, zl);
502 zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
504 uint64_t *countp = &zio->io_children[child][wait];
505 boolean_t waiting = B_FALSE;
507 mutex_enter(&zio->io_lock);
508 ASSERT(zio->io_stall == NULL);
511 zio->io_stall = countp;
514 mutex_exit(&zio->io_lock);
520 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
522 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
523 int *errorp = &pio->io_child_error[zio->io_child_type];
525 mutex_enter(&pio->io_lock);
526 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
527 *errorp = zio_worst_error(*errorp, zio->io_error);
528 pio->io_reexecute |= zio->io_reexecute;
529 ASSERT3U(*countp, >, 0);
533 if (*countp == 0 && pio->io_stall == countp) {
534 pio->io_stall = NULL;
535 mutex_exit(&pio->io_lock);
538 mutex_exit(&pio->io_lock);
543 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
545 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
546 zio->io_error = zio->io_child_error[c];
550 * ==========================================================================
551 * Create the various types of I/O (read, write, free, etc)
552 * ==========================================================================
555 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
556 void *data, uint64_t size, zio_done_func_t *done, void *private,
557 zio_type_t type, zio_priority_t priority, enum zio_flag flags,
558 vdev_t *vd, uint64_t offset, const zbookmark_t *zb,
559 enum zio_stage stage, enum zio_stage pipeline)
563 ASSERT3U(type == ZIO_TYPE_FREE || size, <=, SPA_MAXBLOCKSIZE);
564 ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0);
565 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
567 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
568 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
569 ASSERT(vd || stage == ZIO_STAGE_OPEN);
571 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
572 bzero(zio, sizeof (zio_t));
574 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
575 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
577 list_create(&zio->io_parent_list, sizeof (zio_link_t),
578 offsetof(zio_link_t, zl_parent_node));
579 list_create(&zio->io_child_list, sizeof (zio_link_t),
580 offsetof(zio_link_t, zl_child_node));
583 zio->io_child_type = ZIO_CHILD_VDEV;
584 else if (flags & ZIO_FLAG_GANG_CHILD)
585 zio->io_child_type = ZIO_CHILD_GANG;
586 else if (flags & ZIO_FLAG_DDT_CHILD)
587 zio->io_child_type = ZIO_CHILD_DDT;
589 zio->io_child_type = ZIO_CHILD_LOGICAL;
592 zio->io_bp = (blkptr_t *)bp;
593 zio->io_bp_copy = *bp;
594 zio->io_bp_orig = *bp;
595 if (type != ZIO_TYPE_WRITE ||
596 zio->io_child_type == ZIO_CHILD_DDT)
597 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
598 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
599 zio->io_logical = zio;
600 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
601 pipeline |= ZIO_GANG_STAGES;
607 zio->io_private = private;
609 zio->io_priority = priority;
611 zio->io_offset = offset;
612 zio->io_orig_data = zio->io_data = data;
613 zio->io_orig_size = zio->io_size = size;
614 zio->io_orig_flags = zio->io_flags = flags;
615 zio->io_orig_stage = zio->io_stage = stage;
616 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
618 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
619 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
622 zio->io_bookmark = *zb;
625 if (zio->io_logical == NULL)
626 zio->io_logical = pio->io_logical;
627 if (zio->io_child_type == ZIO_CHILD_GANG)
628 zio->io_gang_leader = pio->io_gang_leader;
629 zio_add_child(pio, zio);
636 zio_destroy(zio_t *zio)
638 list_destroy(&zio->io_parent_list);
639 list_destroy(&zio->io_child_list);
640 mutex_destroy(&zio->io_lock);
641 cv_destroy(&zio->io_cv);
642 kmem_cache_free(zio_cache, zio);
646 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
647 void *private, enum zio_flag flags)
651 zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private,
652 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
653 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
659 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
661 return (zio_null(NULL, spa, NULL, done, private, flags));
665 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
666 void *data, uint64_t size, zio_done_func_t *done, void *private,
667 zio_priority_t priority, enum zio_flag flags, const zbookmark_t *zb)
671 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
672 data, size, done, private,
673 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
674 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
675 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
681 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
682 void *data, uint64_t size, const zio_prop_t *zp,
683 zio_done_func_t *ready, zio_done_func_t *physdone, zio_done_func_t *done,
685 zio_priority_t priority, enum zio_flag flags, const zbookmark_t *zb)
689 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
690 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
691 zp->zp_compress >= ZIO_COMPRESS_OFF &&
692 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
693 DMU_OT_IS_VALID(zp->zp_type) &&
696 zp->zp_copies <= spa_max_replication(spa));
698 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
699 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
700 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
701 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
703 zio->io_ready = ready;
704 zio->io_physdone = physdone;
708 * Data can be NULL if we are going to call zio_write_override() to
709 * provide the already-allocated BP. But we may need the data to
710 * verify a dedup hit (if requested). In this case, don't try to
711 * dedup (just take the already-allocated BP verbatim).
713 if (data == NULL && zio->io_prop.zp_dedup_verify) {
714 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
721 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
722 uint64_t size, zio_done_func_t *done, void *private,
723 zio_priority_t priority, enum zio_flag flags, zbookmark_t *zb)
727 zio = zio_create(pio, spa, txg, bp, data, size, done, private,
728 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
729 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
735 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
737 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
738 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
739 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
740 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
743 * We must reset the io_prop to match the values that existed
744 * when the bp was first written by dmu_sync() keeping in mind
745 * that nopwrite and dedup are mutually exclusive.
747 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
748 zio->io_prop.zp_nopwrite = nopwrite;
749 zio->io_prop.zp_copies = copies;
750 zio->io_bp_override = bp;
754 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
758 * The check for EMBEDDED is a performance optimization. We
759 * process the free here (by ignoring it) rather than
760 * putting it on the list and then processing it in zio_free_sync().
762 if (BP_IS_EMBEDDED(bp))
764 metaslab_check_free(spa, bp);
767 * Frees that are for the currently-syncing txg, are not going to be
768 * deferred, and which will not need to do a read (i.e. not GANG or
769 * DEDUP), can be processed immediately. Otherwise, put them on the
770 * in-memory list for later processing.
772 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
773 txg != spa->spa_syncing_txg ||
774 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
775 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
777 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp,
778 BP_GET_PSIZE(bp), 0)));
783 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
784 uint64_t size, enum zio_flag flags)
787 enum zio_stage stage = ZIO_FREE_PIPELINE;
789 ASSERT(!BP_IS_HOLE(bp));
790 ASSERT(spa_syncing_txg(spa) == txg);
791 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
793 if (BP_IS_EMBEDDED(bp))
794 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
796 metaslab_check_free(spa, bp);
799 if (zfs_trim_enabled)
800 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START |
801 ZIO_STAGE_VDEV_IO_ASSESS;
803 * GANG and DEDUP blocks can induce a read (for the gang block header,
804 * or the DDT), so issue them asynchronously so that this thread is
807 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
808 stage |= ZIO_STAGE_ISSUE_ASYNC;
810 zio = zio_create(pio, spa, txg, bp, NULL, size,
811 NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags,
812 NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
818 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
819 zio_done_func_t *done, void *private, enum zio_flag flags)
823 dprintf_bp(bp, "claiming in txg %llu", txg);
825 if (BP_IS_EMBEDDED(bp))
826 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
829 * A claim is an allocation of a specific block. Claims are needed
830 * to support immediate writes in the intent log. The issue is that
831 * immediate writes contain committed data, but in a txg that was
832 * *not* committed. Upon opening the pool after an unclean shutdown,
833 * the intent log claims all blocks that contain immediate write data
834 * so that the SPA knows they're in use.
836 * All claims *must* be resolved in the first txg -- before the SPA
837 * starts allocating blocks -- so that nothing is allocated twice.
838 * If txg == 0 we just verify that the block is claimable.
840 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
841 ASSERT(txg == spa_first_txg(spa) || txg == 0);
842 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
844 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
845 done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags,
846 NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
852 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset,
853 uint64_t size, zio_done_func_t *done, void *private,
859 if (vd->vdev_children == 0) {
860 zio = zio_create(pio, spa, 0, NULL, NULL, size, done, private,
861 ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, offset, NULL,
862 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
866 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
868 for (c = 0; c < vd->vdev_children; c++)
869 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
870 offset, size, done, private, flags));
877 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
878 void *data, int checksum, zio_done_func_t *done, void *private,
879 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
883 ASSERT(vd->vdev_children == 0);
884 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
885 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
886 ASSERT3U(offset + size, <=, vd->vdev_psize);
888 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
889 ZIO_TYPE_READ, priority, flags, vd, offset, NULL,
890 ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
892 zio->io_prop.zp_checksum = checksum;
898 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
899 void *data, int checksum, zio_done_func_t *done, void *private,
900 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
904 ASSERT(vd->vdev_children == 0);
905 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
906 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
907 ASSERT3U(offset + size, <=, vd->vdev_psize);
909 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private,
910 ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL,
911 ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
913 zio->io_prop.zp_checksum = checksum;
915 if (zio_checksum_table[checksum].ci_eck) {
917 * zec checksums are necessarily destructive -- they modify
918 * the end of the write buffer to hold the verifier/checksum.
919 * Therefore, we must make a local copy in case the data is
920 * being written to multiple places in parallel.
922 void *wbuf = zio_buf_alloc(size);
923 bcopy(data, wbuf, size);
924 zio_push_transform(zio, wbuf, size, size, NULL);
931 * Create a child I/O to do some work for us.
934 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
935 void *data, uint64_t size, int type, zio_priority_t priority,
936 enum zio_flag flags, zio_done_func_t *done, void *private)
938 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
941 ASSERT(vd->vdev_parent ==
942 (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
944 if (type == ZIO_TYPE_READ && bp != NULL) {
946 * If we have the bp, then the child should perform the
947 * checksum and the parent need not. This pushes error
948 * detection as close to the leaves as possible and
949 * eliminates redundant checksums in the interior nodes.
951 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
952 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
955 if (vd->vdev_children == 0)
956 offset += VDEV_LABEL_START_SIZE;
958 flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
961 * If we've decided to do a repair, the write is not speculative --
962 * even if the original read was.
964 if (flags & ZIO_FLAG_IO_REPAIR)
965 flags &= ~ZIO_FLAG_SPECULATIVE;
967 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size,
968 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
969 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
971 zio->io_physdone = pio->io_physdone;
972 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
973 zio->io_logical->io_phys_children++;
979 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
980 int type, zio_priority_t priority, enum zio_flag flags,
981 zio_done_func_t *done, void *private)
985 ASSERT(vd->vdev_ops->vdev_op_leaf);
987 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
988 data, size, done, private, type, priority,
989 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
991 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
997 zio_flush(zio_t *zio, vdev_t *vd)
999 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0,
1001 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1005 zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size)
1008 ASSERT(vd->vdev_ops->vdev_op_leaf);
1010 return zio_ioctl(zio, spa, vd, DKIOCTRIM, offset, size,
1012 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY);
1016 zio_shrink(zio_t *zio, uint64_t size)
1018 ASSERT(zio->io_executor == NULL);
1019 ASSERT(zio->io_orig_size == zio->io_size);
1020 ASSERT(size <= zio->io_size);
1023 * We don't shrink for raidz because of problems with the
1024 * reconstruction when reading back less than the block size.
1025 * Note, BP_IS_RAIDZ() assumes no compression.
1027 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1028 if (!BP_IS_RAIDZ(zio->io_bp))
1029 zio->io_orig_size = zio->io_size = size;
1033 * ==========================================================================
1034 * Prepare to read and write logical blocks
1035 * ==========================================================================
1039 zio_read_bp_init(zio_t **ziop)
1042 blkptr_t *bp = zio->io_bp;
1044 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1045 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1046 !(zio->io_flags & ZIO_FLAG_RAW)) {
1048 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1049 void *cbuf = zio_buf_alloc(psize);
1051 zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
1054 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1055 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1056 decode_embedded_bp_compressed(bp, zio->io_data);
1058 ASSERT(!BP_IS_EMBEDDED(bp));
1061 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1062 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1064 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1065 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1067 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1068 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1070 return (ZIO_PIPELINE_CONTINUE);
1074 zio_write_bp_init(zio_t **ziop)
1077 spa_t *spa = zio->io_spa;
1078 zio_prop_t *zp = &zio->io_prop;
1079 enum zio_compress compress = zp->zp_compress;
1080 blkptr_t *bp = zio->io_bp;
1081 uint64_t lsize = zio->io_size;
1082 uint64_t psize = lsize;
1086 * If our children haven't all reached the ready stage,
1087 * wait for them and then repeat this pipeline stage.
1089 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
1090 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
1091 return (ZIO_PIPELINE_STOP);
1093 if (!IO_IS_ALLOCATING(zio))
1094 return (ZIO_PIPELINE_CONTINUE);
1096 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1098 if (zio->io_bp_override) {
1099 ASSERT(bp->blk_birth != zio->io_txg);
1100 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1102 *bp = *zio->io_bp_override;
1103 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1105 if (BP_IS_EMBEDDED(bp))
1106 return (ZIO_PIPELINE_CONTINUE);
1109 * If we've been overridden and nopwrite is set then
1110 * set the flag accordingly to indicate that a nopwrite
1111 * has already occurred.
1113 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1114 ASSERT(!zp->zp_dedup);
1115 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1116 return (ZIO_PIPELINE_CONTINUE);
1119 ASSERT(!zp->zp_nopwrite);
1121 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1122 return (ZIO_PIPELINE_CONTINUE);
1124 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
1125 zp->zp_dedup_verify);
1127 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1128 BP_SET_DEDUP(bp, 1);
1129 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1130 return (ZIO_PIPELINE_CONTINUE);
1132 zio->io_bp_override = NULL;
1136 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1138 * We're rewriting an existing block, which means we're
1139 * working on behalf of spa_sync(). For spa_sync() to
1140 * converge, it must eventually be the case that we don't
1141 * have to allocate new blocks. But compression changes
1142 * the blocksize, which forces a reallocate, and makes
1143 * convergence take longer. Therefore, after the first
1144 * few passes, stop compressing to ensure convergence.
1146 pass = spa_sync_pass(spa);
1148 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1149 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1150 ASSERT(!BP_GET_DEDUP(bp));
1152 if (pass >= zfs_sync_pass_dont_compress)
1153 compress = ZIO_COMPRESS_OFF;
1155 /* Make sure someone doesn't change their mind on overwrites */
1156 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1157 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1160 if (compress != ZIO_COMPRESS_OFF) {
1161 metaslab_class_t *mc = spa_normal_class(spa);
1162 void *cbuf = zio_buf_alloc(lsize);
1163 psize = zio_compress_data(compress, zio->io_data, cbuf, lsize,
1164 (size_t)metaslab_class_get_minblocksize(mc));
1165 if (psize == 0 || psize == lsize) {
1166 compress = ZIO_COMPRESS_OFF;
1167 zio_buf_free(cbuf, lsize);
1168 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1169 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1170 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1171 encode_embedded_bp_compressed(bp,
1172 cbuf, compress, lsize, psize);
1173 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1174 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1175 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1176 zio_buf_free(cbuf, lsize);
1177 bp->blk_birth = zio->io_txg;
1178 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1179 ASSERT(spa_feature_is_active(spa,
1180 SPA_FEATURE_EMBEDDED_DATA));
1181 return (ZIO_PIPELINE_CONTINUE);
1184 * Round up compressed size to MINBLOCKSIZE and
1188 P2ROUNDUP(psize, (size_t)SPA_MINBLOCKSIZE);
1189 if (rounded > psize) {
1190 bzero((char *)cbuf + psize, rounded - psize);
1193 if (psize == lsize) {
1194 compress = ZIO_COMPRESS_OFF;
1195 zio_buf_free(cbuf, lsize);
1197 zio_push_transform(zio, cbuf,
1198 psize, lsize, NULL);
1204 * The final pass of spa_sync() must be all rewrites, but the first
1205 * few passes offer a trade-off: allocating blocks defers convergence,
1206 * but newly allocated blocks are sequential, so they can be written
1207 * to disk faster. Therefore, we allow the first few passes of
1208 * spa_sync() to allocate new blocks, but force rewrites after that.
1209 * There should only be a handful of blocks after pass 1 in any case.
1211 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1212 BP_GET_PSIZE(bp) == psize &&
1213 pass >= zfs_sync_pass_rewrite) {
1215 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1216 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1217 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1220 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1224 if (zio->io_bp_orig.blk_birth != 0 &&
1225 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1226 BP_SET_LSIZE(bp, lsize);
1227 BP_SET_TYPE(bp, zp->zp_type);
1228 BP_SET_LEVEL(bp, zp->zp_level);
1229 BP_SET_BIRTH(bp, zio->io_txg, 0);
1231 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1233 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1234 BP_SET_LSIZE(bp, lsize);
1235 BP_SET_TYPE(bp, zp->zp_type);
1236 BP_SET_LEVEL(bp, zp->zp_level);
1237 BP_SET_PSIZE(bp, psize);
1238 BP_SET_COMPRESS(bp, compress);
1239 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1240 BP_SET_DEDUP(bp, zp->zp_dedup);
1241 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1243 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1244 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1245 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1247 if (zp->zp_nopwrite) {
1248 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1249 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1250 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1254 return (ZIO_PIPELINE_CONTINUE);
1258 zio_free_bp_init(zio_t **ziop)
1261 blkptr_t *bp = zio->io_bp;
1263 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1264 if (BP_GET_DEDUP(bp))
1265 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1268 return (ZIO_PIPELINE_CONTINUE);
1272 * ==========================================================================
1273 * Execute the I/O pipeline
1274 * ==========================================================================
1278 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1280 spa_t *spa = zio->io_spa;
1281 zio_type_t t = zio->io_type;
1282 int flags = (cutinline ? TQ_FRONT : 0);
1284 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1287 * If we're a config writer or a probe, the normal issue and
1288 * interrupt threads may all be blocked waiting for the config lock.
1289 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1291 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1295 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1297 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1301 * If this is a high priority I/O, then use the high priority taskq if
1304 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1305 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1308 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1311 * NB: We are assuming that the zio can only be dispatched
1312 * to a single taskq at a time. It would be a grievous error
1313 * to dispatch the zio to another taskq at the same time.
1315 #if defined(illumos) || !defined(_KERNEL)
1316 ASSERT(zio->io_tqent.tqent_next == NULL);
1318 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
1320 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1321 flags, &zio->io_tqent);
1325 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1327 kthread_t *executor = zio->io_executor;
1328 spa_t *spa = zio->io_spa;
1330 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1331 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1333 for (i = 0; i < tqs->stqs_count; i++) {
1334 if (taskq_member(tqs->stqs_taskq[i], executor))
1343 zio_issue_async(zio_t **ziop)
1347 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1349 return (ZIO_PIPELINE_STOP);
1353 zio_interrupt(zio_t *zio)
1355 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1359 * Execute the I/O pipeline until one of the following occurs:
1361 * (1) the I/O completes
1362 * (2) the pipeline stalls waiting for dependent child I/Os
1363 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1364 * (4) the I/O is delegated by vdev-level caching or aggregation
1365 * (5) the I/O is deferred due to vdev-level queueing
1366 * (6) the I/O is handed off to another thread.
1368 * In all cases, the pipeline stops whenever there's no CPU work; it never
1369 * burns a thread in cv_wait().
1371 * There's no locking on io_stage because there's no legitimate way
1372 * for multiple threads to be attempting to process the same I/O.
1374 static zio_pipe_stage_t *zio_pipeline[];
1377 zio_execute(zio_t *zio)
1379 zio->io_executor = curthread;
1381 while (zio->io_stage < ZIO_STAGE_DONE) {
1382 enum zio_stage pipeline = zio->io_pipeline;
1383 enum zio_stage stage = zio->io_stage;
1386 ASSERT(!MUTEX_HELD(&zio->io_lock));
1387 ASSERT(ISP2(stage));
1388 ASSERT(zio->io_stall == NULL);
1392 } while ((stage & pipeline) == 0);
1394 ASSERT(stage <= ZIO_STAGE_DONE);
1397 * If we are in interrupt context and this pipeline stage
1398 * will grab a config lock that is held across I/O,
1399 * or may wait for an I/O that needs an interrupt thread
1400 * to complete, issue async to avoid deadlock.
1402 * For VDEV_IO_START, we cut in line so that the io will
1403 * be sent to disk promptly.
1405 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1406 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1407 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1408 zio_requeue_io_start_cut_in_line : B_FALSE;
1409 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1413 zio->io_stage = stage;
1414 rv = zio_pipeline[highbit64(stage) - 1](&zio);
1416 if (rv == ZIO_PIPELINE_STOP)
1419 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
1424 * ==========================================================================
1425 * Initiate I/O, either sync or async
1426 * ==========================================================================
1429 zio_wait(zio_t *zio)
1433 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1434 ASSERT(zio->io_executor == NULL);
1436 zio->io_waiter = curthread;
1440 mutex_enter(&zio->io_lock);
1441 while (zio->io_executor != NULL)
1442 cv_wait(&zio->io_cv, &zio->io_lock);
1443 mutex_exit(&zio->io_lock);
1445 error = zio->io_error;
1452 zio_nowait(zio_t *zio)
1454 ASSERT(zio->io_executor == NULL);
1456 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1457 zio_unique_parent(zio) == NULL) {
1459 * This is a logical async I/O with no parent to wait for it.
1460 * We add it to the spa_async_root_zio "Godfather" I/O which
1461 * will ensure they complete prior to unloading the pool.
1463 spa_t *spa = zio->io_spa;
1465 zio_add_child(spa->spa_async_zio_root, zio);
1472 * ==========================================================================
1473 * Reexecute or suspend/resume failed I/O
1474 * ==========================================================================
1478 zio_reexecute(zio_t *pio)
1480 zio_t *cio, *cio_next;
1482 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1483 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1484 ASSERT(pio->io_gang_leader == NULL);
1485 ASSERT(pio->io_gang_tree == NULL);
1487 pio->io_flags = pio->io_orig_flags;
1488 pio->io_stage = pio->io_orig_stage;
1489 pio->io_pipeline = pio->io_orig_pipeline;
1490 pio->io_reexecute = 0;
1491 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1493 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1494 pio->io_state[w] = 0;
1495 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1496 pio->io_child_error[c] = 0;
1498 if (IO_IS_ALLOCATING(pio))
1499 BP_ZERO(pio->io_bp);
1502 * As we reexecute pio's children, new children could be created.
1503 * New children go to the head of pio's io_child_list, however,
1504 * so we will (correctly) not reexecute them. The key is that
1505 * the remainder of pio's io_child_list, from 'cio_next' onward,
1506 * cannot be affected by any side effects of reexecuting 'cio'.
1508 for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
1509 cio_next = zio_walk_children(pio);
1510 mutex_enter(&pio->io_lock);
1511 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1512 pio->io_children[cio->io_child_type][w]++;
1513 mutex_exit(&pio->io_lock);
1518 * Now that all children have been reexecuted, execute the parent.
1519 * We don't reexecute "The Godfather" I/O here as it's the
1520 * responsibility of the caller to wait on him.
1522 if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
1527 zio_suspend(spa_t *spa, zio_t *zio)
1529 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1530 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1531 "failure and the failure mode property for this pool "
1532 "is set to panic.", spa_name(spa));
1534 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1536 mutex_enter(&spa->spa_suspend_lock);
1538 if (spa->spa_suspend_zio_root == NULL)
1539 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1540 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1541 ZIO_FLAG_GODFATHER);
1543 spa->spa_suspended = B_TRUE;
1546 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1547 ASSERT(zio != spa->spa_suspend_zio_root);
1548 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1549 ASSERT(zio_unique_parent(zio) == NULL);
1550 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1551 zio_add_child(spa->spa_suspend_zio_root, zio);
1554 mutex_exit(&spa->spa_suspend_lock);
1558 zio_resume(spa_t *spa)
1563 * Reexecute all previously suspended i/o.
1565 mutex_enter(&spa->spa_suspend_lock);
1566 spa->spa_suspended = B_FALSE;
1567 cv_broadcast(&spa->spa_suspend_cv);
1568 pio = spa->spa_suspend_zio_root;
1569 spa->spa_suspend_zio_root = NULL;
1570 mutex_exit(&spa->spa_suspend_lock);
1576 return (zio_wait(pio));
1580 zio_resume_wait(spa_t *spa)
1582 mutex_enter(&spa->spa_suspend_lock);
1583 while (spa_suspended(spa))
1584 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1585 mutex_exit(&spa->spa_suspend_lock);
1589 * ==========================================================================
1592 * A gang block is a collection of small blocks that looks to the DMU
1593 * like one large block. When zio_dva_allocate() cannot find a block
1594 * of the requested size, due to either severe fragmentation or the pool
1595 * being nearly full, it calls zio_write_gang_block() to construct the
1596 * block from smaller fragments.
1598 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1599 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1600 * an indirect block: it's an array of block pointers. It consumes
1601 * only one sector and hence is allocatable regardless of fragmentation.
1602 * The gang header's bps point to its gang members, which hold the data.
1604 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1605 * as the verifier to ensure uniqueness of the SHA256 checksum.
1606 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1607 * not the gang header. This ensures that data block signatures (needed for
1608 * deduplication) are independent of how the block is physically stored.
1610 * Gang blocks can be nested: a gang member may itself be a gang block.
1611 * Thus every gang block is a tree in which root and all interior nodes are
1612 * gang headers, and the leaves are normal blocks that contain user data.
1613 * The root of the gang tree is called the gang leader.
1615 * To perform any operation (read, rewrite, free, claim) on a gang block,
1616 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1617 * in the io_gang_tree field of the original logical i/o by recursively
1618 * reading the gang leader and all gang headers below it. This yields
1619 * an in-core tree containing the contents of every gang header and the
1620 * bps for every constituent of the gang block.
1622 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
1623 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
1624 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
1625 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
1626 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
1627 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
1628 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
1629 * of the gang header plus zio_checksum_compute() of the data to update the
1630 * gang header's blk_cksum as described above.
1632 * The two-phase assemble/issue model solves the problem of partial failure --
1633 * what if you'd freed part of a gang block but then couldn't read the
1634 * gang header for another part? Assembling the entire gang tree first
1635 * ensures that all the necessary gang header I/O has succeeded before
1636 * starting the actual work of free, claim, or write. Once the gang tree
1637 * is assembled, free and claim are in-memory operations that cannot fail.
1639 * In the event that a gang write fails, zio_dva_unallocate() walks the
1640 * gang tree to immediately free (i.e. insert back into the space map)
1641 * everything we've allocated. This ensures that we don't get ENOSPC
1642 * errors during repeated suspend/resume cycles due to a flaky device.
1644 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
1645 * the gang tree, we won't modify the block, so we can safely defer the free
1646 * (knowing that the block is still intact). If we *can* assemble the gang
1647 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
1648 * each constituent bp and we can allocate a new block on the next sync pass.
1650 * In all cases, the gang tree allows complete recovery from partial failure.
1651 * ==========================================================================
1655 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1660 return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
1661 NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1662 &pio->io_bookmark));
1666 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1671 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1672 gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
1673 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1675 * As we rewrite each gang header, the pipeline will compute
1676 * a new gang block header checksum for it; but no one will
1677 * compute a new data checksum, so we do that here. The one
1678 * exception is the gang leader: the pipeline already computed
1679 * its data checksum because that stage precedes gang assembly.
1680 * (Presently, nothing actually uses interior data checksums;
1681 * this is just good hygiene.)
1683 if (gn != pio->io_gang_leader->io_gang_tree) {
1684 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
1685 data, BP_GET_PSIZE(bp));
1688 * If we are here to damage data for testing purposes,
1689 * leave the GBH alone so that we can detect the damage.
1691 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
1692 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
1694 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
1695 data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
1696 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1704 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1706 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
1707 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp),
1708 ZIO_GANG_CHILD_FLAGS(pio)));
1713 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
1715 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
1716 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
1719 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
1728 static void zio_gang_tree_assemble_done(zio_t *zio);
1730 static zio_gang_node_t *
1731 zio_gang_node_alloc(zio_gang_node_t **gnpp)
1733 zio_gang_node_t *gn;
1735 ASSERT(*gnpp == NULL);
1737 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
1738 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
1745 zio_gang_node_free(zio_gang_node_t **gnpp)
1747 zio_gang_node_t *gn = *gnpp;
1749 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1750 ASSERT(gn->gn_child[g] == NULL);
1752 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
1753 kmem_free(gn, sizeof (*gn));
1758 zio_gang_tree_free(zio_gang_node_t **gnpp)
1760 zio_gang_node_t *gn = *gnpp;
1765 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
1766 zio_gang_tree_free(&gn->gn_child[g]);
1768 zio_gang_node_free(gnpp);
1772 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
1774 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
1776 ASSERT(gio->io_gang_leader == gio);
1777 ASSERT(BP_IS_GANG(bp));
1779 zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
1780 SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
1781 gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
1785 zio_gang_tree_assemble_done(zio_t *zio)
1787 zio_t *gio = zio->io_gang_leader;
1788 zio_gang_node_t *gn = zio->io_private;
1789 blkptr_t *bp = zio->io_bp;
1791 ASSERT(gio == zio_unique_parent(zio));
1792 ASSERT(zio->io_child_count == 0);
1797 if (BP_SHOULD_BYTESWAP(bp))
1798 byteswap_uint64_array(zio->io_data, zio->io_size);
1800 ASSERT(zio->io_data == gn->gn_gbh);
1801 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
1802 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1804 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1805 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1806 if (!BP_IS_GANG(gbp))
1808 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
1813 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
1815 zio_t *gio = pio->io_gang_leader;
1818 ASSERT(BP_IS_GANG(bp) == !!gn);
1819 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
1820 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
1823 * If you're a gang header, your data is in gn->gn_gbh.
1824 * If you're a gang member, your data is in 'data' and gn == NULL.
1826 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
1829 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
1831 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
1832 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
1833 if (BP_IS_HOLE(gbp))
1835 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
1836 data = (char *)data + BP_GET_PSIZE(gbp);
1840 if (gn == gio->io_gang_tree && gio->io_data != NULL)
1841 ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
1848 zio_gang_assemble(zio_t **ziop)
1851 blkptr_t *bp = zio->io_bp;
1853 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
1854 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1856 zio->io_gang_leader = zio;
1858 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
1860 return (ZIO_PIPELINE_CONTINUE);
1864 zio_gang_issue(zio_t **ziop)
1867 blkptr_t *bp = zio->io_bp;
1869 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
1870 return (ZIO_PIPELINE_STOP);
1872 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
1873 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
1875 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
1876 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
1878 zio_gang_tree_free(&zio->io_gang_tree);
1880 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1882 return (ZIO_PIPELINE_CONTINUE);
1886 zio_write_gang_member_ready(zio_t *zio)
1888 zio_t *pio = zio_unique_parent(zio);
1889 zio_t *gio = zio->io_gang_leader;
1890 dva_t *cdva = zio->io_bp->blk_dva;
1891 dva_t *pdva = pio->io_bp->blk_dva;
1894 if (BP_IS_HOLE(zio->io_bp))
1897 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
1899 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
1900 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
1901 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
1902 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
1903 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
1905 mutex_enter(&pio->io_lock);
1906 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
1907 ASSERT(DVA_GET_GANG(&pdva[d]));
1908 asize = DVA_GET_ASIZE(&pdva[d]);
1909 asize += DVA_GET_ASIZE(&cdva[d]);
1910 DVA_SET_ASIZE(&pdva[d], asize);
1912 mutex_exit(&pio->io_lock);
1916 zio_write_gang_block(zio_t *pio)
1918 spa_t *spa = pio->io_spa;
1919 blkptr_t *bp = pio->io_bp;
1920 zio_t *gio = pio->io_gang_leader;
1922 zio_gang_node_t *gn, **gnpp;
1923 zio_gbh_phys_t *gbh;
1924 uint64_t txg = pio->io_txg;
1925 uint64_t resid = pio->io_size;
1927 int copies = gio->io_prop.zp_copies;
1928 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
1932 error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
1933 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
1934 METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
1936 pio->io_error = error;
1937 return (ZIO_PIPELINE_CONTINUE);
1941 gnpp = &gio->io_gang_tree;
1943 gnpp = pio->io_private;
1944 ASSERT(pio->io_ready == zio_write_gang_member_ready);
1947 gn = zio_gang_node_alloc(gnpp);
1949 bzero(gbh, SPA_GANGBLOCKSIZE);
1952 * Create the gang header.
1954 zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
1955 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
1958 * Create and nowait the gang children.
1960 for (int g = 0; resid != 0; resid -= lsize, g++) {
1961 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
1963 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
1965 zp.zp_checksum = gio->io_prop.zp_checksum;
1966 zp.zp_compress = ZIO_COMPRESS_OFF;
1967 zp.zp_type = DMU_OT_NONE;
1969 zp.zp_copies = gio->io_prop.zp_copies;
1970 zp.zp_dedup = B_FALSE;
1971 zp.zp_dedup_verify = B_FALSE;
1972 zp.zp_nopwrite = B_FALSE;
1974 zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
1975 (char *)pio->io_data + (pio->io_size - resid), lsize, &zp,
1976 zio_write_gang_member_ready, NULL, NULL, &gn->gn_child[g],
1977 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
1978 &pio->io_bookmark));
1982 * Set pio's pipeline to just wait for zio to finish.
1984 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1988 return (ZIO_PIPELINE_CONTINUE);
1992 * The zio_nop_write stage in the pipeline determines if allocating
1993 * a new bp is necessary. By leveraging a cryptographically secure checksum,
1994 * such as SHA256, we can compare the checksums of the new data and the old
1995 * to determine if allocating a new block is required. The nopwrite
1996 * feature can handle writes in either syncing or open context (i.e. zil
1997 * writes) and as a result is mutually exclusive with dedup.
2000 zio_nop_write(zio_t **ziop)
2003 blkptr_t *bp = zio->io_bp;
2004 blkptr_t *bp_orig = &zio->io_bp_orig;
2005 zio_prop_t *zp = &zio->io_prop;
2007 ASSERT(BP_GET_LEVEL(bp) == 0);
2008 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2009 ASSERT(zp->zp_nopwrite);
2010 ASSERT(!zp->zp_dedup);
2011 ASSERT(zio->io_bp_override == NULL);
2012 ASSERT(IO_IS_ALLOCATING(zio));
2015 * Check to see if the original bp and the new bp have matching
2016 * characteristics (i.e. same checksum, compression algorithms, etc).
2017 * If they don't then just continue with the pipeline which will
2018 * allocate a new bp.
2020 if (BP_IS_HOLE(bp_orig) ||
2021 !zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup ||
2022 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2023 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2024 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2025 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2026 return (ZIO_PIPELINE_CONTINUE);
2029 * If the checksums match then reset the pipeline so that we
2030 * avoid allocating a new bp and issuing any I/O.
2032 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2033 ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup);
2034 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2035 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2036 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2037 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2038 sizeof (uint64_t)) == 0);
2041 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2042 zio->io_flags |= ZIO_FLAG_NOPWRITE;
2045 return (ZIO_PIPELINE_CONTINUE);
2049 * ==========================================================================
2051 * ==========================================================================
2054 zio_ddt_child_read_done(zio_t *zio)
2056 blkptr_t *bp = zio->io_bp;
2057 ddt_entry_t *dde = zio->io_private;
2059 zio_t *pio = zio_unique_parent(zio);
2061 mutex_enter(&pio->io_lock);
2062 ddp = ddt_phys_select(dde, bp);
2063 if (zio->io_error == 0)
2064 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2065 if (zio->io_error == 0 && dde->dde_repair_data == NULL)
2066 dde->dde_repair_data = zio->io_data;
2068 zio_buf_free(zio->io_data, zio->io_size);
2069 mutex_exit(&pio->io_lock);
2073 zio_ddt_read_start(zio_t **ziop)
2076 blkptr_t *bp = zio->io_bp;
2078 ASSERT(BP_GET_DEDUP(bp));
2079 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2080 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2082 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2083 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2084 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2085 ddt_phys_t *ddp = dde->dde_phys;
2086 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2089 ASSERT(zio->io_vsd == NULL);
2092 if (ddp_self == NULL)
2093 return (ZIO_PIPELINE_CONTINUE);
2095 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2096 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2098 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2100 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2101 zio_buf_alloc(zio->io_size), zio->io_size,
2102 zio_ddt_child_read_done, dde, zio->io_priority,
2103 ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
2104 &zio->io_bookmark));
2106 return (ZIO_PIPELINE_CONTINUE);
2109 zio_nowait(zio_read(zio, zio->io_spa, bp,
2110 zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
2111 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2113 return (ZIO_PIPELINE_CONTINUE);
2117 zio_ddt_read_done(zio_t **ziop)
2120 blkptr_t *bp = zio->io_bp;
2122 if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
2123 return (ZIO_PIPELINE_STOP);
2125 ASSERT(BP_GET_DEDUP(bp));
2126 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2127 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2129 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2130 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2131 ddt_entry_t *dde = zio->io_vsd;
2133 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2134 return (ZIO_PIPELINE_CONTINUE);
2137 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2138 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2139 return (ZIO_PIPELINE_STOP);
2141 if (dde->dde_repair_data != NULL) {
2142 bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
2143 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2145 ddt_repair_done(ddt, dde);
2149 ASSERT(zio->io_vsd == NULL);
2151 return (ZIO_PIPELINE_CONTINUE);
2155 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2157 spa_t *spa = zio->io_spa;
2160 * Note: we compare the original data, not the transformed data,
2161 * because when zio->io_bp is an override bp, we will not have
2162 * pushed the I/O transforms. That's an important optimization
2163 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2165 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2166 zio_t *lio = dde->dde_lead_zio[p];
2169 return (lio->io_orig_size != zio->io_orig_size ||
2170 bcmp(zio->io_orig_data, lio->io_orig_data,
2171 zio->io_orig_size) != 0);
2175 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2176 ddt_phys_t *ddp = &dde->dde_phys[p];
2178 if (ddp->ddp_phys_birth != 0) {
2179 arc_buf_t *abuf = NULL;
2180 uint32_t aflags = ARC_WAIT;
2181 blkptr_t blk = *zio->io_bp;
2184 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2188 error = arc_read(NULL, spa, &blk,
2189 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2190 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2191 &aflags, &zio->io_bookmark);
2194 if (arc_buf_size(abuf) != zio->io_orig_size ||
2195 bcmp(abuf->b_data, zio->io_orig_data,
2196 zio->io_orig_size) != 0)
2197 error = SET_ERROR(EEXIST);
2198 VERIFY(arc_buf_remove_ref(abuf, &abuf));
2202 return (error != 0);
2210 zio_ddt_child_write_ready(zio_t *zio)
2212 int p = zio->io_prop.zp_copies;
2213 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2214 ddt_entry_t *dde = zio->io_private;
2215 ddt_phys_t *ddp = &dde->dde_phys[p];
2223 ASSERT(dde->dde_lead_zio[p] == zio);
2225 ddt_phys_fill(ddp, zio->io_bp);
2227 while ((pio = zio_walk_parents(zio)) != NULL)
2228 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2234 zio_ddt_child_write_done(zio_t *zio)
2236 int p = zio->io_prop.zp_copies;
2237 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2238 ddt_entry_t *dde = zio->io_private;
2239 ddt_phys_t *ddp = &dde->dde_phys[p];
2243 ASSERT(ddp->ddp_refcnt == 0);
2244 ASSERT(dde->dde_lead_zio[p] == zio);
2245 dde->dde_lead_zio[p] = NULL;
2247 if (zio->io_error == 0) {
2248 while (zio_walk_parents(zio) != NULL)
2249 ddt_phys_addref(ddp);
2251 ddt_phys_clear(ddp);
2258 zio_ddt_ditto_write_done(zio_t *zio)
2260 int p = DDT_PHYS_DITTO;
2261 zio_prop_t *zp = &zio->io_prop;
2262 blkptr_t *bp = zio->io_bp;
2263 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2264 ddt_entry_t *dde = zio->io_private;
2265 ddt_phys_t *ddp = &dde->dde_phys[p];
2266 ddt_key_t *ddk = &dde->dde_key;
2270 ASSERT(ddp->ddp_refcnt == 0);
2271 ASSERT(dde->dde_lead_zio[p] == zio);
2272 dde->dde_lead_zio[p] = NULL;
2274 if (zio->io_error == 0) {
2275 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2276 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2277 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2278 if (ddp->ddp_phys_birth != 0)
2279 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2280 ddt_phys_fill(ddp, bp);
2287 zio_ddt_write(zio_t **ziop)
2290 spa_t *spa = zio->io_spa;
2291 blkptr_t *bp = zio->io_bp;
2292 uint64_t txg = zio->io_txg;
2293 zio_prop_t *zp = &zio->io_prop;
2294 int p = zp->zp_copies;
2298 ddt_t *ddt = ddt_select(spa, bp);
2302 ASSERT(BP_GET_DEDUP(bp));
2303 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2304 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2307 dde = ddt_lookup(ddt, bp, B_TRUE);
2308 ddp = &dde->dde_phys[p];
2310 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2312 * If we're using a weak checksum, upgrade to a strong checksum
2313 * and try again. If we're already using a strong checksum,
2314 * we can't resolve it, so just convert to an ordinary write.
2315 * (And automatically e-mail a paper to Nature?)
2317 if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
2318 zp->zp_checksum = spa_dedup_checksum(spa);
2319 zio_pop_transforms(zio);
2320 zio->io_stage = ZIO_STAGE_OPEN;
2323 zp->zp_dedup = B_FALSE;
2325 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2327 return (ZIO_PIPELINE_CONTINUE);
2330 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2331 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2333 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2334 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2335 zio_prop_t czp = *zp;
2337 czp.zp_copies = ditto_copies;
2340 * If we arrived here with an override bp, we won't have run
2341 * the transform stack, so we won't have the data we need to
2342 * generate a child i/o. So, toss the override bp and restart.
2343 * This is safe, because using the override bp is just an
2344 * optimization; and it's rare, so the cost doesn't matter.
2346 if (zio->io_bp_override) {
2347 zio_pop_transforms(zio);
2348 zio->io_stage = ZIO_STAGE_OPEN;
2349 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2350 zio->io_bp_override = NULL;
2353 return (ZIO_PIPELINE_CONTINUE);
2356 dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2357 zio->io_orig_size, &czp, NULL, NULL,
2358 zio_ddt_ditto_write_done, dde, zio->io_priority,
2359 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2361 zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
2362 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2365 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2366 if (ddp->ddp_phys_birth != 0)
2367 ddt_bp_fill(ddp, bp, txg);
2368 if (dde->dde_lead_zio[p] != NULL)
2369 zio_add_child(zio, dde->dde_lead_zio[p]);
2371 ddt_phys_addref(ddp);
2372 } else if (zio->io_bp_override) {
2373 ASSERT(bp->blk_birth == txg);
2374 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2375 ddt_phys_fill(ddp, bp);
2376 ddt_phys_addref(ddp);
2378 cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
2379 zio->io_orig_size, zp, zio_ddt_child_write_ready, NULL,
2380 zio_ddt_child_write_done, dde, zio->io_priority,
2381 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2383 zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
2384 dde->dde_lead_zio[p] = cio;
2394 return (ZIO_PIPELINE_CONTINUE);
2397 ddt_entry_t *freedde; /* for debugging */
2400 zio_ddt_free(zio_t **ziop)
2403 spa_t *spa = zio->io_spa;
2404 blkptr_t *bp = zio->io_bp;
2405 ddt_t *ddt = ddt_select(spa, bp);
2409 ASSERT(BP_GET_DEDUP(bp));
2410 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2413 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2414 ddp = ddt_phys_select(dde, bp);
2415 ddt_phys_decref(ddp);
2418 return (ZIO_PIPELINE_CONTINUE);
2422 * ==========================================================================
2423 * Allocate and free blocks
2424 * ==========================================================================
2427 zio_dva_allocate(zio_t **ziop)
2430 spa_t *spa = zio->io_spa;
2431 metaslab_class_t *mc = spa_normal_class(spa);
2432 blkptr_t *bp = zio->io_bp;
2436 if (zio->io_gang_leader == NULL) {
2437 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2438 zio->io_gang_leader = zio;
2441 ASSERT(BP_IS_HOLE(bp));
2442 ASSERT0(BP_GET_NDVAS(bp));
2443 ASSERT3U(zio->io_prop.zp_copies, >, 0);
2444 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2445 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
2448 * The dump device does not support gang blocks so allocation on
2449 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
2450 * the "fast" gang feature.
2452 flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
2453 flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
2454 METASLAB_GANG_CHILD : 0;
2455 error = metaslab_alloc(spa, mc, zio->io_size, bp,
2456 zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2459 spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
2460 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
2462 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
2463 return (zio_write_gang_block(zio));
2464 zio->io_error = error;
2467 return (ZIO_PIPELINE_CONTINUE);
2471 zio_dva_free(zio_t **ziop)
2475 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2477 return (ZIO_PIPELINE_CONTINUE);
2481 zio_dva_claim(zio_t **ziop)
2486 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
2488 zio->io_error = error;
2490 return (ZIO_PIPELINE_CONTINUE);
2494 * Undo an allocation. This is used by zio_done() when an I/O fails
2495 * and we want to give back the block we just allocated.
2496 * This handles both normal blocks and gang blocks.
2499 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
2501 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
2502 ASSERT(zio->io_bp_override == NULL);
2504 if (!BP_IS_HOLE(bp))
2505 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
2508 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2509 zio_dva_unallocate(zio, gn->gn_child[g],
2510 &gn->gn_gbh->zg_blkptr[g]);
2516 * Try to allocate an intent log block. Return 0 on success, errno on failure.
2519 zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp,
2520 uint64_t size, boolean_t use_slog)
2524 ASSERT(txg > spa_syncing_txg(spa));
2527 * ZIL blocks are always contiguous (i.e. not gang blocks) so we
2528 * set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
2529 * when allocating them.
2532 error = metaslab_alloc(spa, spa_log_class(spa), size,
2533 new_bp, 1, txg, old_bp,
2534 METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID);
2538 error = metaslab_alloc(spa, spa_normal_class(spa), size,
2539 new_bp, 1, txg, old_bp,
2540 METASLAB_HINTBP_AVOID);
2544 BP_SET_LSIZE(new_bp, size);
2545 BP_SET_PSIZE(new_bp, size);
2546 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
2547 BP_SET_CHECKSUM(new_bp,
2548 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
2549 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
2550 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
2551 BP_SET_LEVEL(new_bp, 0);
2552 BP_SET_DEDUP(new_bp, 0);
2553 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
2560 * Free an intent log block.
2563 zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
2565 ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
2566 ASSERT(!BP_IS_GANG(bp));
2568 zio_free(spa, txg, bp);
2572 * ==========================================================================
2573 * Read, write and delete to physical devices
2574 * ==========================================================================
2577 zio_vdev_io_start(zio_t **ziop)
2580 vdev_t *vd = zio->io_vd;
2582 spa_t *spa = zio->io_spa;
2584 ASSERT(zio->io_error == 0);
2585 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2588 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2589 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2592 * The mirror_ops handle multiple DVAs in a single BP.
2594 return (vdev_mirror_ops.vdev_op_io_start(zio));
2597 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE) {
2598 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg);
2599 return (ZIO_PIPELINE_CONTINUE);
2603 * We keep track of time-sensitive I/Os so that the scan thread
2604 * can quickly react to certain workloads. In particular, we care
2605 * about non-scrubbing, top-level reads and writes with the following
2607 * - synchronous writes of user data to non-slog devices
2608 * - any reads of user data
2609 * When these conditions are met, adjust the timestamp of spa_last_io
2610 * which allows the scan thread to adjust its workload accordingly.
2612 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
2613 vd == vd->vdev_top && !vd->vdev_islog &&
2614 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
2615 zio->io_txg != spa_syncing_txg(spa)) {
2616 uint64_t old = spa->spa_last_io;
2617 uint64_t new = ddi_get_lbolt64();
2619 (void) atomic_cas_64(&spa->spa_last_io, old, new);
2622 align = 1ULL << vd->vdev_top->vdev_ashift;
2624 if (P2PHASE(zio->io_size, align) != 0) {
2625 uint64_t asize = P2ROUNDUP(zio->io_size, align);
2627 if (zio->io_type == ZIO_TYPE_READ ||
2628 zio->io_type == ZIO_TYPE_WRITE)
2629 abuf = zio_buf_alloc(asize);
2630 ASSERT(vd == vd->vdev_top);
2631 if (zio->io_type == ZIO_TYPE_WRITE) {
2632 bcopy(zio->io_data, abuf, zio->io_size);
2633 bzero(abuf + zio->io_size, asize - zio->io_size);
2635 zio_push_transform(zio, abuf, asize, abuf ? asize : 0,
2639 ASSERT(P2PHASE(zio->io_offset, align) == 0);
2640 ASSERT(P2PHASE(zio->io_size, align) == 0);
2641 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa));
2644 * If this is a repair I/O, and there's no self-healing involved --
2645 * that is, we're just resilvering what we expect to resilver --
2646 * then don't do the I/O unless zio's txg is actually in vd's DTL.
2647 * This prevents spurious resilvering with nested replication.
2648 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
2649 * A is out of date, we'll read from C+D, then use the data to
2650 * resilver A+B -- but we don't actually want to resilver B, just A.
2651 * The top-level mirror has no way to know this, so instead we just
2652 * discard unnecessary repairs as we work our way down the vdev tree.
2653 * The same logic applies to any form of nested replication:
2654 * ditto + mirror, RAID-Z + replacing, etc. This covers them all.
2656 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
2657 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
2658 zio->io_txg != 0 && /* not a delegated i/o */
2659 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
2660 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2661 zio_vdev_io_bypass(zio);
2662 return (ZIO_PIPELINE_CONTINUE);
2665 if (vd->vdev_ops->vdev_op_leaf &&
2666 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2668 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio))
2669 return (ZIO_PIPELINE_CONTINUE);
2671 if ((zio = vdev_queue_io(zio)) == NULL)
2672 return (ZIO_PIPELINE_STOP);
2675 if (!vdev_accessible(vd, zio)) {
2676 zio->io_error = SET_ERROR(ENXIO);
2678 return (ZIO_PIPELINE_STOP);
2683 * Note that we ignore repair writes for TRIM because they can conflict
2684 * with normal writes. This isn't an issue because, by definition, we
2685 * only repair blocks that aren't freed.
2687 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_WRITE &&
2688 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
2689 if (!trim_map_write_start(zio))
2690 return (ZIO_PIPELINE_STOP);
2693 return (vd->vdev_ops->vdev_op_io_start(zio));
2697 zio_vdev_io_done(zio_t **ziop)
2700 vdev_t *vd = zio->io_vd;
2701 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
2702 boolean_t unexpected_error = B_FALSE;
2704 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2705 return (ZIO_PIPELINE_STOP);
2707 ASSERT(zio->io_type == ZIO_TYPE_READ ||
2708 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE);
2710 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2711 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
2713 if (zio->io_type == ZIO_TYPE_WRITE &&
2714 !(zio->io_flags & ZIO_FLAG_IO_REPAIR))
2715 trim_map_write_done(zio);
2717 vdev_queue_io_done(zio);
2719 if (zio->io_type == ZIO_TYPE_WRITE)
2720 vdev_cache_write(zio);
2722 if (zio_injection_enabled && zio->io_error == 0)
2723 zio->io_error = zio_handle_device_injection(vd,
2726 if (zio_injection_enabled && zio->io_error == 0)
2727 zio->io_error = zio_handle_label_injection(zio, EIO);
2729 if (zio->io_error) {
2730 if (!vdev_accessible(vd, zio)) {
2731 zio->io_error = SET_ERROR(ENXIO);
2733 unexpected_error = B_TRUE;
2738 ops->vdev_op_io_done(zio);
2740 if (unexpected_error)
2741 VERIFY(vdev_probe(vd, zio) == NULL);
2743 return (ZIO_PIPELINE_CONTINUE);
2747 * For non-raidz ZIOs, we can just copy aside the bad data read from the
2748 * disk, and use that to finish the checksum ereport later.
2751 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
2752 const void *good_buf)
2754 /* no processing needed */
2755 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
2760 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
2762 void *buf = zio_buf_alloc(zio->io_size);
2764 bcopy(zio->io_data, buf, zio->io_size);
2766 zcr->zcr_cbinfo = zio->io_size;
2767 zcr->zcr_cbdata = buf;
2768 zcr->zcr_finish = zio_vsd_default_cksum_finish;
2769 zcr->zcr_free = zio_buf_free;
2773 zio_vdev_io_assess(zio_t **ziop)
2776 vdev_t *vd = zio->io_vd;
2778 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
2779 return (ZIO_PIPELINE_STOP);
2781 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
2782 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
2784 if (zio->io_vsd != NULL) {
2785 zio->io_vsd_ops->vsd_free(zio);
2789 if (zio_injection_enabled && zio->io_error == 0)
2790 zio->io_error = zio_handle_fault_injection(zio, EIO);
2792 if (zio->io_type == ZIO_TYPE_IOCTL && zio->io_cmd == DKIOCTRIM)
2793 switch (zio->io_error) {
2795 ZIO_TRIM_STAT_INCR(bytes, zio->io_size);
2796 ZIO_TRIM_STAT_BUMP(success);
2799 ZIO_TRIM_STAT_BUMP(unsupported);
2802 ZIO_TRIM_STAT_BUMP(failed);
2807 * If the I/O failed, determine whether we should attempt to retry it.
2809 * On retry, we cut in line in the issue queue, since we don't want
2810 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2812 if (zio->io_error && vd == NULL &&
2813 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
2814 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
2815 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2817 zio->io_flags |= ZIO_FLAG_IO_RETRY |
2818 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
2819 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
2820 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
2821 zio_requeue_io_start_cut_in_line);
2822 return (ZIO_PIPELINE_STOP);
2826 * If we got an error on a leaf device, convert it to ENXIO
2827 * if the device is not accessible at all.
2829 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2830 !vdev_accessible(vd, zio))
2831 zio->io_error = SET_ERROR(ENXIO);
2834 * If we can't write to an interior vdev (mirror or RAID-Z),
2835 * set vdev_cant_write so that we stop trying to allocate from it.
2837 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
2838 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
2839 vd->vdev_cant_write = B_TRUE;
2843 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2845 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
2846 zio->io_physdone != NULL) {
2847 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
2848 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
2849 zio->io_physdone(zio->io_logical);
2852 return (ZIO_PIPELINE_CONTINUE);
2856 zio_vdev_io_reissue(zio_t *zio)
2858 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2859 ASSERT(zio->io_error == 0);
2861 zio->io_stage >>= 1;
2865 zio_vdev_io_redone(zio_t *zio)
2867 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
2869 zio->io_stage >>= 1;
2873 zio_vdev_io_bypass(zio_t *zio)
2875 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
2876 ASSERT(zio->io_error == 0);
2878 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
2879 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2883 * ==========================================================================
2884 * Generate and verify checksums
2885 * ==========================================================================
2888 zio_checksum_generate(zio_t **ziop)
2891 blkptr_t *bp = zio->io_bp;
2892 enum zio_checksum checksum;
2896 * This is zio_write_phys().
2897 * We're either generating a label checksum, or none at all.
2899 checksum = zio->io_prop.zp_checksum;
2901 if (checksum == ZIO_CHECKSUM_OFF)
2902 return (ZIO_PIPELINE_CONTINUE);
2904 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
2906 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
2907 ASSERT(!IO_IS_ALLOCATING(zio));
2908 checksum = ZIO_CHECKSUM_GANG_HEADER;
2910 checksum = BP_GET_CHECKSUM(bp);
2914 zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2916 return (ZIO_PIPELINE_CONTINUE);
2920 zio_checksum_verify(zio_t **ziop)
2923 zio_bad_cksum_t info;
2924 blkptr_t *bp = zio->io_bp;
2927 ASSERT(zio->io_vd != NULL);
2931 * This is zio_read_phys().
2932 * We're either verifying a label checksum, or nothing at all.
2934 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
2935 return (ZIO_PIPELINE_CONTINUE);
2937 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
2940 if ((error = zio_checksum_error(zio, &info)) != 0) {
2941 zio->io_error = error;
2942 if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
2943 zfs_ereport_start_checksum(zio->io_spa,
2944 zio->io_vd, zio, zio->io_offset,
2945 zio->io_size, NULL, &info);
2949 return (ZIO_PIPELINE_CONTINUE);
2953 * Called by RAID-Z to ensure we don't compute the checksum twice.
2956 zio_checksum_verified(zio_t *zio)
2958 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2962 * ==========================================================================
2963 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
2964 * An error of 0 indicates success. ENXIO indicates whole-device failure,
2965 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
2966 * indicate errors that are specific to one I/O, and most likely permanent.
2967 * Any other error is presumed to be worse because we weren't expecting it.
2968 * ==========================================================================
2971 zio_worst_error(int e1, int e2)
2973 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
2976 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
2977 if (e1 == zio_error_rank[r1])
2980 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
2981 if (e2 == zio_error_rank[r2])
2984 return (r1 > r2 ? e1 : e2);
2988 * ==========================================================================
2990 * ==========================================================================
2993 zio_ready(zio_t **ziop)
2996 blkptr_t *bp = zio->io_bp;
2997 zio_t *pio, *pio_next;
2999 if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
3000 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
3001 return (ZIO_PIPELINE_STOP);
3003 if (zio->io_ready) {
3004 ASSERT(IO_IS_ALLOCATING(zio));
3005 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
3006 (zio->io_flags & ZIO_FLAG_NOPWRITE));
3007 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
3012 if (bp != NULL && bp != &zio->io_bp_copy)
3013 zio->io_bp_copy = *bp;
3016 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3018 mutex_enter(&zio->io_lock);
3019 zio->io_state[ZIO_WAIT_READY] = 1;
3020 pio = zio_walk_parents(zio);
3021 mutex_exit(&zio->io_lock);
3024 * As we notify zio's parents, new parents could be added.
3025 * New parents go to the head of zio's io_parent_list, however,
3026 * so we will (correctly) not notify them. The remainder of zio's
3027 * io_parent_list, from 'pio_next' onward, cannot change because
3028 * all parents must wait for us to be done before they can be done.
3030 for (; pio != NULL; pio = pio_next) {
3031 pio_next = zio_walk_parents(zio);
3032 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
3035 if (zio->io_flags & ZIO_FLAG_NODATA) {
3036 if (BP_IS_GANG(bp)) {
3037 zio->io_flags &= ~ZIO_FLAG_NODATA;
3039 ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
3040 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
3044 if (zio_injection_enabled &&
3045 zio->io_spa->spa_syncing_txg == zio->io_txg)
3046 zio_handle_ignored_writes(zio);
3048 return (ZIO_PIPELINE_CONTINUE);
3052 zio_done(zio_t **ziop)
3055 spa_t *spa = zio->io_spa;
3056 zio_t *lio = zio->io_logical;
3057 blkptr_t *bp = zio->io_bp;
3058 vdev_t *vd = zio->io_vd;
3059 uint64_t psize = zio->io_size;
3060 zio_t *pio, *pio_next;
3063 * If our children haven't all completed,
3064 * wait for them and then repeat this pipeline stage.
3066 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
3067 zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
3068 zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
3069 zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
3070 return (ZIO_PIPELINE_STOP);
3072 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3073 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3074 ASSERT(zio->io_children[c][w] == 0);
3076 if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
3077 ASSERT(bp->blk_pad[0] == 0);
3078 ASSERT(bp->blk_pad[1] == 0);
3079 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3080 (bp == zio_unique_parent(zio)->io_bp));
3081 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3082 zio->io_bp_override == NULL &&
3083 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3084 ASSERT(!BP_SHOULD_BYTESWAP(bp));
3085 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3086 ASSERT(BP_COUNT_GANG(bp) == 0 ||
3087 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3089 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3090 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3094 * If there were child vdev/gang/ddt errors, they apply to us now.
3096 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3097 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3098 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3101 * If the I/O on the transformed data was successful, generate any
3102 * checksum reports now while we still have the transformed data.
3104 if (zio->io_error == 0) {
3105 while (zio->io_cksum_report != NULL) {
3106 zio_cksum_report_t *zcr = zio->io_cksum_report;
3107 uint64_t align = zcr->zcr_align;
3108 uint64_t asize = P2ROUNDUP(psize, align);
3109 char *abuf = zio->io_data;
3111 if (asize != psize) {
3112 abuf = zio_buf_alloc(asize);
3113 bcopy(zio->io_data, abuf, psize);
3114 bzero(abuf + psize, asize - psize);
3117 zio->io_cksum_report = zcr->zcr_next;
3118 zcr->zcr_next = NULL;
3119 zcr->zcr_finish(zcr, abuf);
3120 zfs_ereport_free_checksum(zcr);
3123 zio_buf_free(abuf, asize);
3127 zio_pop_transforms(zio); /* note: may set zio->io_error */
3129 vdev_stat_update(zio, psize);
3131 if (zio->io_error) {
3133 * If this I/O is attached to a particular vdev,
3134 * generate an error message describing the I/O failure
3135 * at the block level. We ignore these errors if the
3136 * device is currently unavailable.
3138 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3139 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3141 if ((zio->io_error == EIO || !(zio->io_flags &
3142 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3145 * For logical I/O requests, tell the SPA to log the
3146 * error and generate a logical data ereport.
3148 spa_log_error(spa, zio);
3149 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3154 if (zio->io_error && zio == lio) {
3156 * Determine whether zio should be reexecuted. This will
3157 * propagate all the way to the root via zio_notify_parent().
3159 ASSERT(vd == NULL && bp != NULL);
3160 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3162 if (IO_IS_ALLOCATING(zio) &&
3163 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3164 if (zio->io_error != ENOSPC)
3165 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3167 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3170 if ((zio->io_type == ZIO_TYPE_READ ||
3171 zio->io_type == ZIO_TYPE_FREE) &&
3172 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3173 zio->io_error == ENXIO &&
3174 spa_load_state(spa) == SPA_LOAD_NONE &&
3175 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3176 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3178 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3179 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3182 * Here is a possibly good place to attempt to do
3183 * either combinatorial reconstruction or error correction
3184 * based on checksums. It also might be a good place
3185 * to send out preliminary ereports before we suspend
3191 * If there were logical child errors, they apply to us now.
3192 * We defer this until now to avoid conflating logical child
3193 * errors with errors that happened to the zio itself when
3194 * updating vdev stats and reporting FMA events above.
3196 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
3198 if ((zio->io_error || zio->io_reexecute) &&
3199 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
3200 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
3201 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
3203 zio_gang_tree_free(&zio->io_gang_tree);
3206 * Godfather I/Os should never suspend.
3208 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
3209 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
3210 zio->io_reexecute = 0;
3212 if (zio->io_reexecute) {
3214 * This is a logical I/O that wants to reexecute.
3216 * Reexecute is top-down. When an i/o fails, if it's not
3217 * the root, it simply notifies its parent and sticks around.
3218 * The parent, seeing that it still has children in zio_done(),
3219 * does the same. This percolates all the way up to the root.
3220 * The root i/o will reexecute or suspend the entire tree.
3222 * This approach ensures that zio_reexecute() honors
3223 * all the original i/o dependency relationships, e.g.
3224 * parents not executing until children are ready.
3226 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3228 zio->io_gang_leader = NULL;
3230 mutex_enter(&zio->io_lock);
3231 zio->io_state[ZIO_WAIT_DONE] = 1;
3232 mutex_exit(&zio->io_lock);
3235 * "The Godfather" I/O monitors its children but is
3236 * not a true parent to them. It will track them through
3237 * the pipeline but severs its ties whenever they get into
3238 * trouble (e.g. suspended). This allows "The Godfather"
3239 * I/O to return status without blocking.
3241 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3242 zio_link_t *zl = zio->io_walk_link;
3243 pio_next = zio_walk_parents(zio);
3245 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
3246 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
3247 zio_remove_child(pio, zio, zl);
3248 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3252 if ((pio = zio_unique_parent(zio)) != NULL) {
3254 * We're not a root i/o, so there's nothing to do
3255 * but notify our parent. Don't propagate errors
3256 * upward since we haven't permanently failed yet.
3258 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
3259 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
3260 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3261 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
3263 * We'd fail again if we reexecuted now, so suspend
3264 * until conditions improve (e.g. device comes online).
3266 zio_suspend(spa, zio);
3269 * Reexecution is potentially a huge amount of work.
3270 * Hand it off to the otherwise-unused claim taskq.
3272 #if defined(illumos) || !defined(_KERNEL)
3273 ASSERT(zio->io_tqent.tqent_next == NULL);
3275 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
3277 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
3278 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
3281 return (ZIO_PIPELINE_STOP);
3284 ASSERT(zio->io_child_count == 0);
3285 ASSERT(zio->io_reexecute == 0);
3286 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
3289 * Report any checksum errors, since the I/O is complete.
3291 while (zio->io_cksum_report != NULL) {
3292 zio_cksum_report_t *zcr = zio->io_cksum_report;
3293 zio->io_cksum_report = zcr->zcr_next;
3294 zcr->zcr_next = NULL;
3295 zcr->zcr_finish(zcr, NULL);
3296 zfs_ereport_free_checksum(zcr);
3300 * It is the responsibility of the done callback to ensure that this
3301 * particular zio is no longer discoverable for adoption, and as
3302 * such, cannot acquire any new parents.
3307 mutex_enter(&zio->io_lock);
3308 zio->io_state[ZIO_WAIT_DONE] = 1;
3309 mutex_exit(&zio->io_lock);
3311 for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
3312 zio_link_t *zl = zio->io_walk_link;
3313 pio_next = zio_walk_parents(zio);
3314 zio_remove_child(pio, zio, zl);
3315 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
3318 if (zio->io_waiter != NULL) {
3319 mutex_enter(&zio->io_lock);
3320 zio->io_executor = NULL;
3321 cv_broadcast(&zio->io_cv);
3322 mutex_exit(&zio->io_lock);
3327 return (ZIO_PIPELINE_STOP);
3331 * ==========================================================================
3332 * I/O pipeline definition
3333 * ==========================================================================
3335 static zio_pipe_stage_t *zio_pipeline[] = {
3341 zio_checksum_generate,
3356 zio_checksum_verify,
3360 /* dnp is the dnode for zb1->zb_object */
3362 zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_t *zb1,
3363 const zbookmark_t *zb2)
3365 uint64_t zb1nextL0, zb2thisobj;
3367 ASSERT(zb1->zb_objset == zb2->zb_objset);
3368 ASSERT(zb2->zb_level == 0);
3371 * A bookmark in the deadlist is considered to be after
3374 if (zb2->zb_object == DMU_DEADLIST_OBJECT)
3377 /* The objset_phys_t isn't before anything. */
3381 zb1nextL0 = (zb1->zb_blkid + 1) <<
3382 ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
3384 zb2thisobj = zb2->zb_object ? zb2->zb_object :
3385 zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
3387 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
3388 uint64_t nextobj = zb1nextL0 *
3389 (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
3390 return (nextobj <= zb2thisobj);
3393 if (zb1->zb_object < zb2thisobj)
3395 if (zb1->zb_object > zb2thisobj)
3397 if (zb2->zb_object == DMU_META_DNODE_OBJECT)
3399 return (zb1nextL0 <= zb2->zb_blkid);