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, 2018 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
25 * Copyright (c) 2014 Integros [integros.com]
28 #include <sys/sysmacros.h>
29 #include <sys/zfs_context.h>
30 #include <sys/fm/fs/zfs.h>
33 #include <sys/spa_impl.h>
34 #include <sys/vdev_impl.h>
35 #include <sys/zio_impl.h>
36 #include <sys/zio_compress.h>
37 #include <sys/zio_checksum.h>
38 #include <sys/dmu_objset.h>
41 #include <sys/trim_map.h>
42 #include <sys/blkptr.h>
43 #include <sys/zfeature.h>
44 #include <sys/dsl_scan.h>
45 #include <sys/metaslab_impl.h>
47 #include <sys/cityhash.h>
49 SYSCTL_DECL(_vfs_zfs);
50 SYSCTL_NODE(_vfs_zfs, OID_AUTO, zio, CTLFLAG_RW, 0, "ZFS ZIO");
51 #if defined(__amd64__)
52 static int zio_use_uma = 1;
54 static int zio_use_uma = 0;
56 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, use_uma, CTLFLAG_RDTUN, &zio_use_uma, 0,
57 "Use uma(9) for ZIO allocations");
58 static int zio_exclude_metadata = 0;
59 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, exclude_metadata, CTLFLAG_RDTUN, &zio_exclude_metadata, 0,
60 "Exclude metadata buffers from dumps as well");
62 zio_trim_stats_t zio_trim_stats = {
63 { "bytes", KSTAT_DATA_UINT64,
64 "Number of bytes successfully TRIMmed" },
65 { "success", KSTAT_DATA_UINT64,
66 "Number of successful TRIM requests" },
67 { "unsupported", KSTAT_DATA_UINT64,
68 "Number of TRIM requests that failed because TRIM is not supported" },
69 { "failed", KSTAT_DATA_UINT64,
70 "Number of TRIM requests that failed for reasons other than not supported" },
73 static kstat_t *zio_trim_ksp;
76 * ==========================================================================
77 * I/O type descriptions
78 * ==========================================================================
80 const char *zio_type_name[ZIO_TYPES] = {
81 "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim",
85 boolean_t zio_dva_throttle_enabled = B_TRUE;
86 SYSCTL_INT(_vfs_zfs_zio, OID_AUTO, dva_throttle_enabled, CTLFLAG_RWTUN,
87 &zio_dva_throttle_enabled, 0, "Enable allocation throttling");
90 * ==========================================================================
92 * ==========================================================================
94 kmem_cache_t *zio_cache;
95 kmem_cache_t *zio_link_cache;
96 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
97 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
100 extern vmem_t *zio_alloc_arena;
103 #define BP_SPANB(indblkshift, level) \
104 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
105 #define COMPARE_META_LEVEL 0x80000000ul
107 * The following actions directly effect the spa's sync-to-convergence logic.
108 * The values below define the sync pass when we start performing the action.
109 * Care should be taken when changing these values as they directly impact
110 * spa_sync() performance. Tuning these values may introduce subtle performance
111 * pathologies and should only be done in the context of performance analysis.
112 * These tunables will eventually be removed and replaced with #defines once
113 * enough analysis has been done to determine optimal values.
115 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
116 * regular blocks are not deferred.
118 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
119 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_deferred_free, CTLFLAG_RDTUN,
120 &zfs_sync_pass_deferred_free, 0, "defer frees starting in this pass");
121 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
122 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_dont_compress, CTLFLAG_RDTUN,
123 &zfs_sync_pass_dont_compress, 0, "don't compress starting in this pass");
124 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
125 SYSCTL_INT(_vfs_zfs, OID_AUTO, sync_pass_rewrite, CTLFLAG_RDTUN,
126 &zfs_sync_pass_rewrite, 0, "rewrite new bps starting in this pass");
129 * An allocating zio is one that either currently has the DVA allocate
130 * stage set or will have it later in its lifetime.
132 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
134 boolean_t zio_requeue_io_start_cut_in_line = B_TRUE;
138 int zio_buf_debug_limit = 16384;
140 int zio_buf_debug_limit = 0;
144 static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t);
150 zio_cache = kmem_cache_create("zio_cache",
151 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
152 zio_link_cache = kmem_cache_create("zio_link_cache",
153 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
158 * For small buffers, we want a cache for each multiple of
159 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
160 * for each quarter-power of 2.
162 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
163 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
166 int cflags = zio_exclude_metadata ? KMC_NODEBUG : 0;
174 * If we are using watchpoints, put each buffer on its own page,
175 * to eliminate the performance overhead of trapping to the
176 * kernel when modifying a non-watched buffer that shares the
177 * page with a watched buffer.
179 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
183 if (size <= 4 * SPA_MINBLOCKSIZE) {
184 align = SPA_MINBLOCKSIZE;
185 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
186 align = MIN(p2 >> 2, PAGESIZE);
191 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
192 zio_buf_cache[c] = kmem_cache_create(name, size,
193 align, NULL, NULL, NULL, NULL, NULL, cflags);
196 * Since zio_data bufs do not appear in crash dumps, we
197 * pass KMC_NOTOUCH so that no allocator metadata is
198 * stored with the buffers.
200 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
201 zio_data_buf_cache[c] = kmem_cache_create(name, size,
202 align, NULL, NULL, NULL, NULL, NULL,
203 cflags | KMC_NOTOUCH | KMC_NODEBUG);
208 ASSERT(zio_buf_cache[c] != NULL);
209 if (zio_buf_cache[c - 1] == NULL)
210 zio_buf_cache[c - 1] = zio_buf_cache[c];
212 ASSERT(zio_data_buf_cache[c] != NULL);
213 if (zio_data_buf_cache[c - 1] == NULL)
214 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
220 zio_trim_ksp = kstat_create("zfs", 0, "zio_trim", "misc",
222 sizeof(zio_trim_stats) / sizeof(kstat_named_t),
225 if (zio_trim_ksp != NULL) {
226 zio_trim_ksp->ks_data = &zio_trim_stats;
227 kstat_install(zio_trim_ksp);
235 kmem_cache_t *last_cache = NULL;
236 kmem_cache_t *last_data_cache = NULL;
238 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
239 if (zio_buf_cache[c] != last_cache) {
240 last_cache = zio_buf_cache[c];
241 kmem_cache_destroy(zio_buf_cache[c]);
243 zio_buf_cache[c] = NULL;
245 if (zio_data_buf_cache[c] != last_data_cache) {
246 last_data_cache = zio_data_buf_cache[c];
247 kmem_cache_destroy(zio_data_buf_cache[c]);
249 zio_data_buf_cache[c] = NULL;
252 kmem_cache_destroy(zio_link_cache);
253 kmem_cache_destroy(zio_cache);
257 if (zio_trim_ksp != NULL) {
258 kstat_delete(zio_trim_ksp);
264 * ==========================================================================
265 * Allocate and free I/O buffers
266 * ==========================================================================
270 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
271 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
272 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
273 * excess / transient data in-core during a crashdump.
276 zio_buf_alloc(size_t size)
278 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
279 int flags = zio_exclude_metadata ? KM_NODEBUG : 0;
281 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
284 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
286 return (kmem_alloc(size, KM_SLEEP|flags));
290 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
291 * crashdump if the kernel panics. This exists so that we will limit the amount
292 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
293 * of kernel heap dumped to disk when the kernel panics)
296 zio_data_buf_alloc(size_t size)
298 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
300 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
303 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
305 return (kmem_alloc(size, KM_SLEEP | KM_NODEBUG));
309 zio_buf_free(void *buf, size_t size)
311 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
313 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
316 kmem_cache_free(zio_buf_cache[c], buf);
318 kmem_free(buf, size);
322 zio_data_buf_free(void *buf, size_t size)
324 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
326 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
329 kmem_cache_free(zio_data_buf_cache[c], buf);
331 kmem_free(buf, size);
335 * ==========================================================================
336 * Push and pop I/O transform buffers
337 * ==========================================================================
340 zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize,
341 zio_transform_func_t *transform)
343 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
346 * Ensure that anyone expecting this zio to contain a linear ABD isn't
347 * going to get a nasty surprise when they try to access the data.
350 IMPLY(abd_is_linear(zio->io_abd), abd_is_linear(data));
352 IMPLY(zio->io_abd != NULL && abd_is_linear(zio->io_abd),
353 abd_is_linear(data));
356 zt->zt_orig_abd = zio->io_abd;
357 zt->zt_orig_size = zio->io_size;
358 zt->zt_bufsize = bufsize;
359 zt->zt_transform = transform;
361 zt->zt_next = zio->io_transform_stack;
362 zio->io_transform_stack = zt;
369 zio_pop_transforms(zio_t *zio)
373 while ((zt = zio->io_transform_stack) != NULL) {
374 if (zt->zt_transform != NULL)
375 zt->zt_transform(zio,
376 zt->zt_orig_abd, zt->zt_orig_size);
378 if (zt->zt_bufsize != 0)
379 abd_free(zio->io_abd);
381 zio->io_abd = zt->zt_orig_abd;
382 zio->io_size = zt->zt_orig_size;
383 zio->io_transform_stack = zt->zt_next;
385 kmem_free(zt, sizeof (zio_transform_t));
390 * ==========================================================================
391 * I/O transform callbacks for subblocks and decompression
392 * ==========================================================================
395 zio_subblock(zio_t *zio, abd_t *data, uint64_t size)
397 ASSERT(zio->io_size > size);
399 if (zio->io_type == ZIO_TYPE_READ)
400 abd_copy(data, zio->io_abd, size);
404 zio_decompress(zio_t *zio, abd_t *data, uint64_t size)
406 if (zio->io_error == 0) {
407 void *tmp = abd_borrow_buf(data, size);
408 int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
409 zio->io_abd, tmp, zio->io_size, size);
410 abd_return_buf_copy(data, tmp, size);
413 zio->io_error = SET_ERROR(EIO);
418 * ==========================================================================
419 * I/O parent/child relationships and pipeline interlocks
420 * ==========================================================================
423 zio_walk_parents(zio_t *cio, zio_link_t **zl)
425 list_t *pl = &cio->io_parent_list;
427 *zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl);
431 ASSERT((*zl)->zl_child == cio);
432 return ((*zl)->zl_parent);
436 zio_walk_children(zio_t *pio, zio_link_t **zl)
438 list_t *cl = &pio->io_child_list;
440 ASSERT(MUTEX_HELD(&pio->io_lock));
442 *zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl);
446 ASSERT((*zl)->zl_parent == pio);
447 return ((*zl)->zl_child);
451 zio_unique_parent(zio_t *cio)
453 zio_link_t *zl = NULL;
454 zio_t *pio = zio_walk_parents(cio, &zl);
456 VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL);
461 zio_add_child(zio_t *pio, zio_t *cio)
463 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
466 * Logical I/Os can have logical, gang, or vdev children.
467 * Gang I/Os can have gang or vdev children.
468 * Vdev I/Os can only have vdev children.
469 * The following ASSERT captures all of these constraints.
471 ASSERT3S(cio->io_child_type, <=, pio->io_child_type);
476 mutex_enter(&pio->io_lock);
477 mutex_enter(&cio->io_lock);
479 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
481 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
482 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
484 list_insert_head(&pio->io_child_list, zl);
485 list_insert_head(&cio->io_parent_list, zl);
487 pio->io_child_count++;
488 cio->io_parent_count++;
490 mutex_exit(&cio->io_lock);
491 mutex_exit(&pio->io_lock);
495 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
497 ASSERT(zl->zl_parent == pio);
498 ASSERT(zl->zl_child == cio);
500 mutex_enter(&pio->io_lock);
501 mutex_enter(&cio->io_lock);
503 list_remove(&pio->io_child_list, zl);
504 list_remove(&cio->io_parent_list, zl);
506 pio->io_child_count--;
507 cio->io_parent_count--;
509 mutex_exit(&cio->io_lock);
510 mutex_exit(&pio->io_lock);
511 kmem_cache_free(zio_link_cache, zl);
515 zio_wait_for_children(zio_t *zio, uint8_t childbits, enum zio_wait_type wait)
517 boolean_t waiting = B_FALSE;
519 mutex_enter(&zio->io_lock);
520 ASSERT(zio->io_stall == NULL);
521 for (int c = 0; c < ZIO_CHILD_TYPES; c++) {
522 if (!(ZIO_CHILD_BIT_IS_SET(childbits, c)))
525 uint64_t *countp = &zio->io_children[c][wait];
528 ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN);
529 zio->io_stall = countp;
534 mutex_exit(&zio->io_lock);
539 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait,
540 zio_t **next_to_executep)
542 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
543 int *errorp = &pio->io_child_error[zio->io_child_type];
545 mutex_enter(&pio->io_lock);
546 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
547 *errorp = zio_worst_error(*errorp, zio->io_error);
548 pio->io_reexecute |= zio->io_reexecute;
549 ASSERT3U(*countp, >, 0);
553 if (*countp == 0 && pio->io_stall == countp) {
554 zio_taskq_type_t type =
555 pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE :
557 pio->io_stall = NULL;
558 mutex_exit(&pio->io_lock);
561 * If we can tell the caller to execute this parent next, do
562 * so. Otherwise dispatch the parent zio as its own task.
564 * Having the caller execute the parent when possible reduces
565 * locking on the zio taskq's, reduces context switch
566 * overhead, and has no recursion penalty. Note that one
567 * read from disk typically causes at least 3 zio's: a
568 * zio_null(), the logical zio_read(), and then a physical
569 * zio. When the physical ZIO completes, we are able to call
570 * zio_done() on all 3 of these zio's from one invocation of
571 * zio_execute() by returning the parent back to
572 * zio_execute(). Since the parent isn't executed until this
573 * thread returns back to zio_execute(), the caller should do
576 * In other cases, dispatching the parent prevents
577 * overflowing the stack when we have deeply nested
578 * parent-child relationships, as we do with the "mega zio"
579 * of writes for spa_sync(), and the chain of ZIL blocks.
581 if (next_to_executep != NULL && *next_to_executep == NULL) {
582 *next_to_executep = pio;
584 zio_taskq_dispatch(pio, type, B_FALSE);
587 mutex_exit(&pio->io_lock);
592 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
594 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
595 zio->io_error = zio->io_child_error[c];
599 zio_bookmark_compare(const void *x1, const void *x2)
601 const zio_t *z1 = x1;
602 const zio_t *z2 = x2;
604 if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset)
606 if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset)
609 if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object)
611 if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object)
614 if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level)
616 if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level)
619 if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid)
621 if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid)
633 * ==========================================================================
634 * Create the various types of I/O (read, write, free, etc)
635 * ==========================================================================
638 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
639 abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done,
640 void *private, zio_type_t type, zio_priority_t priority,
641 enum zio_flag flags, vdev_t *vd, uint64_t offset,
642 const zbookmark_phys_t *zb, enum zio_stage stage, enum zio_stage pipeline)
646 IMPLY(type != ZIO_TYPE_FREE, psize <= SPA_MAXBLOCKSIZE);
647 ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0);
648 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
650 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
651 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
652 ASSERT(vd || stage == ZIO_STAGE_OPEN);
654 IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW) != 0);
656 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
657 bzero(zio, sizeof (zio_t));
659 mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL);
660 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
662 list_create(&zio->io_parent_list, sizeof (zio_link_t),
663 offsetof(zio_link_t, zl_parent_node));
664 list_create(&zio->io_child_list, sizeof (zio_link_t),
665 offsetof(zio_link_t, zl_child_node));
666 metaslab_trace_init(&zio->io_alloc_list);
669 zio->io_child_type = ZIO_CHILD_VDEV;
670 else if (flags & ZIO_FLAG_GANG_CHILD)
671 zio->io_child_type = ZIO_CHILD_GANG;
672 else if (flags & ZIO_FLAG_DDT_CHILD)
673 zio->io_child_type = ZIO_CHILD_DDT;
675 zio->io_child_type = ZIO_CHILD_LOGICAL;
678 zio->io_bp = (blkptr_t *)bp;
679 zio->io_bp_copy = *bp;
680 zio->io_bp_orig = *bp;
681 if (type != ZIO_TYPE_WRITE ||
682 zio->io_child_type == ZIO_CHILD_DDT)
683 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
684 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
685 zio->io_logical = zio;
686 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
687 pipeline |= ZIO_GANG_STAGES;
693 zio->io_private = private;
695 zio->io_priority = priority;
697 zio->io_offset = offset;
698 zio->io_orig_abd = zio->io_abd = data;
699 zio->io_orig_size = zio->io_size = psize;
700 zio->io_lsize = lsize;
701 zio->io_orig_flags = zio->io_flags = flags;
702 zio->io_orig_stage = zio->io_stage = stage;
703 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
704 zio->io_pipeline_trace = ZIO_STAGE_OPEN;
706 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
707 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
710 zio->io_bookmark = *zb;
713 if (zio->io_logical == NULL)
714 zio->io_logical = pio->io_logical;
715 if (zio->io_child_type == ZIO_CHILD_GANG)
716 zio->io_gang_leader = pio->io_gang_leader;
717 zio_add_child(pio, zio);
724 zio_destroy(zio_t *zio)
726 metaslab_trace_fini(&zio->io_alloc_list);
727 list_destroy(&zio->io_parent_list);
728 list_destroy(&zio->io_child_list);
729 mutex_destroy(&zio->io_lock);
730 cv_destroy(&zio->io_cv);
731 kmem_cache_free(zio_cache, zio);
735 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
736 void *private, enum zio_flag flags)
740 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
741 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
742 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
748 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
750 return (zio_null(NULL, spa, NULL, done, private, flags));
754 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
756 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
757 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
758 bp, (longlong_t)BP_GET_TYPE(bp));
760 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
761 BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
762 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
763 bp, (longlong_t)BP_GET_CHECKSUM(bp));
765 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
766 BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
767 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
768 bp, (longlong_t)BP_GET_COMPRESS(bp));
770 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
771 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
772 bp, (longlong_t)BP_GET_LSIZE(bp));
774 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
775 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
776 bp, (longlong_t)BP_GET_PSIZE(bp));
779 if (BP_IS_EMBEDDED(bp)) {
780 if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
781 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
782 bp, (longlong_t)BPE_GET_ETYPE(bp));
787 * Do not verify individual DVAs if the config is not trusted. This
788 * will be done once the zio is executed in vdev_mirror_map_alloc.
790 if (!spa->spa_trust_config)
794 * Pool-specific checks.
796 * Note: it would be nice to verify that the blk_birth and
797 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
798 * allows the birth time of log blocks (and dmu_sync()-ed blocks
799 * that are in the log) to be arbitrarily large.
801 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
802 uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
803 if (vdevid >= spa->spa_root_vdev->vdev_children) {
804 zfs_panic_recover("blkptr at %p DVA %u has invalid "
806 bp, i, (longlong_t)vdevid);
809 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
811 zfs_panic_recover("blkptr at %p DVA %u has invalid "
813 bp, i, (longlong_t)vdevid);
816 if (vd->vdev_ops == &vdev_hole_ops) {
817 zfs_panic_recover("blkptr at %p DVA %u has hole "
819 bp, i, (longlong_t)vdevid);
822 if (vd->vdev_ops == &vdev_missing_ops) {
824 * "missing" vdevs are valid during import, but we
825 * don't have their detailed info (e.g. asize), so
826 * we can't perform any more checks on them.
830 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
831 uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
833 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
834 if (offset + asize > vd->vdev_asize) {
835 zfs_panic_recover("blkptr at %p DVA %u has invalid "
837 bp, i, (longlong_t)offset);
843 zfs_dva_valid(spa_t *spa, const dva_t *dva, const blkptr_t *bp)
845 uint64_t vdevid = DVA_GET_VDEV(dva);
847 if (vdevid >= spa->spa_root_vdev->vdev_children)
850 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
854 if (vd->vdev_ops == &vdev_hole_ops)
857 if (vd->vdev_ops == &vdev_missing_ops) {
861 uint64_t offset = DVA_GET_OFFSET(dva);
862 uint64_t asize = DVA_GET_ASIZE(dva);
865 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
866 if (offset + asize > vd->vdev_asize)
873 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
874 abd_t *data, uint64_t size, zio_done_func_t *done, void *private,
875 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
879 zfs_blkptr_verify(spa, bp);
881 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
882 data, size, size, done, private,
883 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
884 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
885 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
891 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
892 abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp,
893 zio_done_func_t *ready, zio_done_func_t *children_ready,
894 zio_done_func_t *physdone, zio_done_func_t *done,
895 void *private, zio_priority_t priority, enum zio_flag flags,
896 const zbookmark_phys_t *zb)
900 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
901 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
902 zp->zp_compress >= ZIO_COMPRESS_OFF &&
903 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
904 DMU_OT_IS_VALID(zp->zp_type) &&
907 zp->zp_copies <= spa_max_replication(spa));
909 zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private,
910 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
911 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
912 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
914 zio->io_ready = ready;
915 zio->io_children_ready = children_ready;
916 zio->io_physdone = physdone;
920 * Data can be NULL if we are going to call zio_write_override() to
921 * provide the already-allocated BP. But we may need the data to
922 * verify a dedup hit (if requested). In this case, don't try to
923 * dedup (just take the already-allocated BP verbatim).
925 if (data == NULL && zio->io_prop.zp_dedup_verify) {
926 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
933 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data,
934 uint64_t size, zio_done_func_t *done, void *private,
935 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
939 zio = zio_create(pio, spa, txg, bp, data, size, size, done, private,
940 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb,
941 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
947 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
949 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
950 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
951 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
952 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
955 * We must reset the io_prop to match the values that existed
956 * when the bp was first written by dmu_sync() keeping in mind
957 * that nopwrite and dedup are mutually exclusive.
959 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
960 zio->io_prop.zp_nopwrite = nopwrite;
961 zio->io_prop.zp_copies = copies;
962 zio->io_bp_override = bp;
966 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
969 zfs_blkptr_verify(spa, bp);
972 * The check for EMBEDDED is a performance optimization. We
973 * process the free here (by ignoring it) rather than
974 * putting it on the list and then processing it in zio_free_sync().
976 if (BP_IS_EMBEDDED(bp))
978 metaslab_check_free(spa, bp);
981 * Frees that are for the currently-syncing txg, are not going to be
982 * deferred, and which will not need to do a read (i.e. not GANG or
983 * DEDUP), can be processed immediately. Otherwise, put them on the
984 * in-memory list for later processing.
986 if (zfs_trim_enabled || BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
987 txg != spa->spa_syncing_txg ||
988 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
989 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
991 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp,
992 BP_GET_PSIZE(bp), 0)));
997 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
998 uint64_t size, enum zio_flag flags)
1001 enum zio_stage stage = ZIO_FREE_PIPELINE;
1003 ASSERT(!BP_IS_HOLE(bp));
1004 ASSERT(spa_syncing_txg(spa) == txg);
1005 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
1007 if (BP_IS_EMBEDDED(bp))
1008 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
1010 metaslab_check_free(spa, bp);
1012 dsl_scan_freed(spa, bp);
1014 if (zfs_trim_enabled)
1015 stage |= ZIO_STAGE_ISSUE_ASYNC | ZIO_STAGE_VDEV_IO_START |
1016 ZIO_STAGE_VDEV_IO_ASSESS;
1018 * GANG and DEDUP blocks can induce a read (for the gang block header,
1019 * or the DDT), so issue them asynchronously so that this thread is
1022 else if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
1023 stage |= ZIO_STAGE_ISSUE_ASYNC;
1025 flags |= ZIO_FLAG_DONT_QUEUE;
1027 zio = zio_create(pio, spa, txg, bp, NULL, size,
1028 size, NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW,
1029 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
1035 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1036 zio_done_func_t *done, void *private, enum zio_flag flags)
1040 zfs_blkptr_verify(spa, bp);
1042 if (BP_IS_EMBEDDED(bp))
1043 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
1046 * A claim is an allocation of a specific block. Claims are needed
1047 * to support immediate writes in the intent log. The issue is that
1048 * immediate writes contain committed data, but in a txg that was
1049 * *not* committed. Upon opening the pool after an unclean shutdown,
1050 * the intent log claims all blocks that contain immediate write data
1051 * so that the SPA knows they're in use.
1053 * All claims *must* be resolved in the first txg -- before the SPA
1054 * starts allocating blocks -- so that nothing is allocated twice.
1055 * If txg == 0 we just verify that the block is claimable.
1057 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <,
1058 spa_min_claim_txg(spa));
1059 ASSERT(txg == spa_min_claim_txg(spa) || txg == 0);
1060 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
1062 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1063 BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW,
1064 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
1065 ASSERT0(zio->io_queued_timestamp);
1071 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, uint64_t offset,
1072 uint64_t size, zio_done_func_t *done, void *private,
1073 zio_priority_t priority, enum zio_flag flags)
1078 if (vd->vdev_children == 0) {
1079 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
1080 ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
1081 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
1085 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
1087 for (c = 0; c < vd->vdev_children; c++)
1088 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
1089 offset, size, done, private, priority, flags));
1096 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1097 abd_t *data, int checksum, zio_done_func_t *done, void *private,
1098 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1102 ASSERT(vd->vdev_children == 0);
1103 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1104 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1105 ASSERT3U(offset + size, <=, vd->vdev_psize);
1107 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1108 private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1109 offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
1111 zio->io_prop.zp_checksum = checksum;
1117 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1118 abd_t *data, int checksum, zio_done_func_t *done, void *private,
1119 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1123 ASSERT(vd->vdev_children == 0);
1124 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1125 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1126 ASSERT3U(offset + size, <=, vd->vdev_psize);
1128 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1129 private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1130 offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
1132 zio->io_prop.zp_checksum = checksum;
1134 if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
1136 * zec checksums are necessarily destructive -- they modify
1137 * the end of the write buffer to hold the verifier/checksum.
1138 * Therefore, we must make a local copy in case the data is
1139 * being written to multiple places in parallel.
1141 abd_t *wbuf = abd_alloc_sametype(data, size);
1142 abd_copy(wbuf, data, size);
1144 zio_push_transform(zio, wbuf, size, size, NULL);
1151 * Create a child I/O to do some work for us.
1154 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1155 abd_t *data, uint64_t size, int type, zio_priority_t priority,
1156 enum zio_flag flags, zio_done_func_t *done, void *private)
1158 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1162 * vdev child I/Os do not propagate their error to the parent.
1163 * Therefore, for correct operation the caller *must* check for
1164 * and handle the error in the child i/o's done callback.
1165 * The only exceptions are i/os that we don't care about
1166 * (OPTIONAL or REPAIR).
1168 ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) ||
1171 if (type == ZIO_TYPE_READ && bp != NULL) {
1173 * If we have the bp, then the child should perform the
1174 * checksum and the parent need not. This pushes error
1175 * detection as close to the leaves as possible and
1176 * eliminates redundant checksums in the interior nodes.
1178 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1179 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1182 /* Not all IO types require vdev io done stage e.g. free */
1183 if (type == ZIO_TYPE_FREE &&
1184 !(pio->io_pipeline & ZIO_STAGE_VDEV_IO_DONE))
1185 pipeline &= ~ZIO_STAGE_VDEV_IO_DONE;
1187 if (vd->vdev_ops->vdev_op_leaf) {
1188 ASSERT0(vd->vdev_children);
1189 offset += VDEV_LABEL_START_SIZE;
1192 flags |= ZIO_VDEV_CHILD_FLAGS(pio);
1195 * If we've decided to do a repair, the write is not speculative --
1196 * even if the original read was.
1198 if (flags & ZIO_FLAG_IO_REPAIR)
1199 flags &= ~ZIO_FLAG_SPECULATIVE;
1202 * If we're creating a child I/O that is not associated with a
1203 * top-level vdev, then the child zio is not an allocating I/O.
1204 * If this is a retried I/O then we ignore it since we will
1205 * have already processed the original allocating I/O.
1207 if (flags & ZIO_FLAG_IO_ALLOCATING &&
1208 (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) {
1209 metaslab_class_t *mc = spa_normal_class(pio->io_spa);
1211 ASSERT(mc->mc_alloc_throttle_enabled);
1212 ASSERT(type == ZIO_TYPE_WRITE);
1213 ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE);
1214 ASSERT(!(flags & ZIO_FLAG_IO_REPAIR));
1215 ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) ||
1216 pio->io_child_type == ZIO_CHILD_GANG);
1218 flags &= ~ZIO_FLAG_IO_ALLOCATING;
1221 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size,
1222 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1223 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1224 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
1226 zio->io_physdone = pio->io_physdone;
1227 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1228 zio->io_logical->io_phys_children++;
1234 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size,
1235 zio_type_t type, zio_priority_t priority, enum zio_flag flags,
1236 zio_done_func_t *done, void *private)
1240 ASSERT(vd->vdev_ops->vdev_op_leaf);
1242 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1243 data, size, size, done, private, type, priority,
1244 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1246 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1252 zio_flush(zio_t *zio, vdev_t *vd)
1254 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 0, 0,
1255 NULL, NULL, ZIO_PRIORITY_NOW,
1256 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1260 zio_trim(zio_t *zio, spa_t *spa, vdev_t *vd, uint64_t offset, uint64_t size)
1263 ASSERT(vd->vdev_ops->vdev_op_leaf);
1265 return (zio_create(zio, spa, 0, NULL, NULL, size, size, NULL, NULL,
1266 ZIO_TYPE_FREE, ZIO_PRIORITY_TRIM, ZIO_FLAG_DONT_AGGREGATE |
1267 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY,
1268 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PHYS_PIPELINE));
1272 zio_shrink(zio_t *zio, uint64_t size)
1274 ASSERT3P(zio->io_executor, ==, NULL);
1275 ASSERT3P(zio->io_orig_size, ==, zio->io_size);
1276 ASSERT3U(size, <=, zio->io_size);
1279 * We don't shrink for raidz because of problems with the
1280 * reconstruction when reading back less than the block size.
1281 * Note, BP_IS_RAIDZ() assumes no compression.
1283 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1284 if (!BP_IS_RAIDZ(zio->io_bp)) {
1285 /* we are not doing a raw write */
1286 ASSERT3U(zio->io_size, ==, zio->io_lsize);
1287 zio->io_orig_size = zio->io_size = zio->io_lsize = size;
1292 * ==========================================================================
1293 * Prepare to read and write logical blocks
1294 * ==========================================================================
1298 zio_read_bp_init(zio_t *zio)
1300 blkptr_t *bp = zio->io_bp;
1302 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1304 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1305 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1306 !(zio->io_flags & ZIO_FLAG_RAW)) {
1308 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1309 zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1310 psize, psize, zio_decompress);
1313 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1314 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1316 int psize = BPE_GET_PSIZE(bp);
1317 void *data = abd_borrow_buf(zio->io_abd, psize);
1318 decode_embedded_bp_compressed(bp, data);
1319 abd_return_buf_copy(zio->io_abd, data, psize);
1321 ASSERT(!BP_IS_EMBEDDED(bp));
1322 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1325 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1326 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1328 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1329 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1331 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1332 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1338 zio_write_bp_init(zio_t *zio)
1340 if (!IO_IS_ALLOCATING(zio))
1343 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1345 if (zio->io_bp_override) {
1346 blkptr_t *bp = zio->io_bp;
1347 zio_prop_t *zp = &zio->io_prop;
1349 ASSERT(bp->blk_birth != zio->io_txg);
1350 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1352 *bp = *zio->io_bp_override;
1353 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1355 if (BP_IS_EMBEDDED(bp))
1359 * If we've been overridden and nopwrite is set then
1360 * set the flag accordingly to indicate that a nopwrite
1361 * has already occurred.
1363 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1364 ASSERT(!zp->zp_dedup);
1365 ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum);
1366 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1370 ASSERT(!zp->zp_nopwrite);
1372 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1375 ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags &
1376 ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify);
1378 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
1379 BP_SET_DEDUP(bp, 1);
1380 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1385 * We were unable to handle this as an override bp, treat
1386 * it as a regular write I/O.
1388 zio->io_bp_override = NULL;
1389 *bp = zio->io_bp_orig;
1390 zio->io_pipeline = zio->io_orig_pipeline;
1397 zio_write_compress(zio_t *zio)
1399 spa_t *spa = zio->io_spa;
1400 zio_prop_t *zp = &zio->io_prop;
1401 enum zio_compress compress = zp->zp_compress;
1402 blkptr_t *bp = zio->io_bp;
1403 uint64_t lsize = zio->io_lsize;
1404 uint64_t psize = zio->io_size;
1407 EQUIV(lsize != psize, (zio->io_flags & ZIO_FLAG_RAW) != 0);
1410 * If our children haven't all reached the ready stage,
1411 * wait for them and then repeat this pipeline stage.
1413 if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT |
1414 ZIO_CHILD_GANG_BIT, ZIO_WAIT_READY)) {
1418 if (!IO_IS_ALLOCATING(zio))
1421 if (zio->io_children_ready != NULL) {
1423 * Now that all our children are ready, run the callback
1424 * associated with this zio in case it wants to modify the
1425 * data to be written.
1427 ASSERT3U(zp->zp_level, >, 0);
1428 zio->io_children_ready(zio);
1431 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1432 ASSERT(zio->io_bp_override == NULL);
1434 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1436 * We're rewriting an existing block, which means we're
1437 * working on behalf of spa_sync(). For spa_sync() to
1438 * converge, it must eventually be the case that we don't
1439 * have to allocate new blocks. But compression changes
1440 * the blocksize, which forces a reallocate, and makes
1441 * convergence take longer. Therefore, after the first
1442 * few passes, stop compressing to ensure convergence.
1444 pass = spa_sync_pass(spa);
1446 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1447 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1448 ASSERT(!BP_GET_DEDUP(bp));
1450 if (pass >= zfs_sync_pass_dont_compress)
1451 compress = ZIO_COMPRESS_OFF;
1453 /* Make sure someone doesn't change their mind on overwrites */
1454 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1455 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1458 /* If it's a compressed write that is not raw, compress the buffer. */
1459 if (compress != ZIO_COMPRESS_OFF && psize == lsize) {
1460 void *cbuf = zio_buf_alloc(lsize);
1461 psize = zio_compress_data(compress, zio->io_abd, cbuf, lsize);
1462 if (psize == 0 || psize == lsize) {
1463 compress = ZIO_COMPRESS_OFF;
1464 zio_buf_free(cbuf, lsize);
1465 } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
1466 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1467 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1468 encode_embedded_bp_compressed(bp,
1469 cbuf, compress, lsize, psize);
1470 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1471 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1472 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1473 zio_buf_free(cbuf, lsize);
1474 bp->blk_birth = zio->io_txg;
1475 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1476 ASSERT(spa_feature_is_active(spa,
1477 SPA_FEATURE_EMBEDDED_DATA));
1481 * Round up compressed size up to the ashift
1482 * of the smallest-ashift device, and zero the tail.
1483 * This ensures that the compressed size of the BP
1484 * (and thus compressratio property) are correct,
1485 * in that we charge for the padding used to fill out
1488 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
1489 size_t rounded = (size_t)P2ROUNDUP(psize,
1490 1ULL << spa->spa_min_ashift);
1491 if (rounded >= lsize) {
1492 compress = ZIO_COMPRESS_OFF;
1493 zio_buf_free(cbuf, lsize);
1496 abd_t *cdata = abd_get_from_buf(cbuf, lsize);
1497 abd_take_ownership_of_buf(cdata, B_TRUE);
1498 abd_zero_off(cdata, psize, rounded - psize);
1500 zio_push_transform(zio, cdata,
1501 psize, lsize, NULL);
1506 * We were unable to handle this as an override bp, treat
1507 * it as a regular write I/O.
1509 zio->io_bp_override = NULL;
1510 *bp = zio->io_bp_orig;
1511 zio->io_pipeline = zio->io_orig_pipeline;
1513 ASSERT3U(psize, !=, 0);
1517 * The final pass of spa_sync() must be all rewrites, but the first
1518 * few passes offer a trade-off: allocating blocks defers convergence,
1519 * but newly allocated blocks are sequential, so they can be written
1520 * to disk faster. Therefore, we allow the first few passes of
1521 * spa_sync() to allocate new blocks, but force rewrites after that.
1522 * There should only be a handful of blocks after pass 1 in any case.
1524 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1525 BP_GET_PSIZE(bp) == psize &&
1526 pass >= zfs_sync_pass_rewrite) {
1528 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1529 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1530 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1533 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1537 if (zio->io_bp_orig.blk_birth != 0 &&
1538 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1539 BP_SET_LSIZE(bp, lsize);
1540 BP_SET_TYPE(bp, zp->zp_type);
1541 BP_SET_LEVEL(bp, zp->zp_level);
1542 BP_SET_BIRTH(bp, zio->io_txg, 0);
1544 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1546 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1547 BP_SET_LSIZE(bp, lsize);
1548 BP_SET_TYPE(bp, zp->zp_type);
1549 BP_SET_LEVEL(bp, zp->zp_level);
1550 BP_SET_PSIZE(bp, psize);
1551 BP_SET_COMPRESS(bp, compress);
1552 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1553 BP_SET_DEDUP(bp, zp->zp_dedup);
1554 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1556 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1557 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1558 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1560 if (zp->zp_nopwrite) {
1561 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1562 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1563 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1570 zio_free_bp_init(zio_t *zio)
1572 blkptr_t *bp = zio->io_bp;
1574 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1575 if (BP_GET_DEDUP(bp))
1576 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1579 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1585 * ==========================================================================
1586 * Execute the I/O pipeline
1587 * ==========================================================================
1591 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1593 spa_t *spa = zio->io_spa;
1594 zio_type_t t = zio->io_type;
1595 int flags = (cutinline ? TQ_FRONT : 0);
1597 ASSERT(q == ZIO_TASKQ_ISSUE || q == ZIO_TASKQ_INTERRUPT);
1600 * If we're a config writer or a probe, the normal issue and
1601 * interrupt threads may all be blocked waiting for the config lock.
1602 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1604 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1608 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1610 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1614 * If this is a high priority I/O, then use the high priority taskq if
1617 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1618 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1621 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1624 * NB: We are assuming that the zio can only be dispatched
1625 * to a single taskq at a time. It would be a grievous error
1626 * to dispatch the zio to another taskq at the same time.
1628 #if defined(illumos) || !defined(_KERNEL)
1629 ASSERT(zio->io_tqent.tqent_next == NULL);
1631 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
1633 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1634 flags, &zio->io_tqent);
1638 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1640 kthread_t *executor = zio->io_executor;
1641 spa_t *spa = zio->io_spa;
1643 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1644 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1646 for (i = 0; i < tqs->stqs_count; i++) {
1647 if (taskq_member(tqs->stqs_taskq[i], executor))
1656 zio_issue_async(zio_t *zio)
1658 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1664 zio_interrupt(zio_t *zio)
1666 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1670 zio_delay_interrupt(zio_t *zio)
1673 * The timeout_generic() function isn't defined in userspace, so
1674 * rather than trying to implement the function, the zio delay
1675 * functionality has been disabled for userspace builds.
1680 * If io_target_timestamp is zero, then no delay has been registered
1681 * for this IO, thus jump to the end of this function and "skip" the
1682 * delay; issuing it directly to the zio layer.
1684 if (zio->io_target_timestamp != 0) {
1685 hrtime_t now = gethrtime();
1687 if (now >= zio->io_target_timestamp) {
1689 * This IO has already taken longer than the target
1690 * delay to complete, so we don't want to delay it
1691 * any longer; we "miss" the delay and issue it
1692 * directly to the zio layer. This is likely due to
1693 * the target latency being set to a value less than
1694 * the underlying hardware can satisfy (e.g. delay
1695 * set to 1ms, but the disks take 10ms to complete an
1699 DTRACE_PROBE2(zio__delay__miss, zio_t *, zio,
1704 hrtime_t diff = zio->io_target_timestamp - now;
1706 DTRACE_PROBE3(zio__delay__hit, zio_t *, zio,
1707 hrtime_t, now, hrtime_t, diff);
1709 (void) timeout_generic(CALLOUT_NORMAL,
1710 (void (*)(void *))zio_interrupt, zio, diff, 1, 0);
1717 DTRACE_PROBE1(zio__delay__skip, zio_t *, zio);
1722 * Execute the I/O pipeline until one of the following occurs:
1724 * (1) the I/O completes
1725 * (2) the pipeline stalls waiting for dependent child I/Os
1726 * (3) the I/O issues, so we're waiting for an I/O completion interrupt
1727 * (4) the I/O is delegated by vdev-level caching or aggregation
1728 * (5) the I/O is deferred due to vdev-level queueing
1729 * (6) the I/O is handed off to another thread.
1731 * In all cases, the pipeline stops whenever there's no CPU work; it never
1732 * burns a thread in cv_wait().
1734 * There's no locking on io_stage because there's no legitimate way
1735 * for multiple threads to be attempting to process the same I/O.
1737 static zio_pipe_stage_t *zio_pipeline[];
1740 zio_execute(zio_t *zio)
1742 ASSERT3U(zio->io_queued_timestamp, >, 0);
1744 while (zio->io_stage < ZIO_STAGE_DONE) {
1745 enum zio_stage pipeline = zio->io_pipeline;
1746 enum zio_stage stage = zio->io_stage;
1748 zio->io_executor = curthread;
1750 ASSERT(!MUTEX_HELD(&zio->io_lock));
1751 ASSERT(ISP2(stage));
1752 ASSERT(zio->io_stall == NULL);
1756 } while ((stage & pipeline) == 0);
1758 ASSERT(stage <= ZIO_STAGE_DONE);
1761 * If we are in interrupt context and this pipeline stage
1762 * will grab a config lock that is held across I/O,
1763 * or may wait for an I/O that needs an interrupt thread
1764 * to complete, issue async to avoid deadlock.
1766 * For VDEV_IO_START, we cut in line so that the io will
1767 * be sent to disk promptly.
1769 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1770 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1771 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
1772 zio_requeue_io_start_cut_in_line : B_FALSE;
1773 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
1777 zio->io_stage = stage;
1778 zio->io_pipeline_trace |= zio->io_stage;
1781 * The zio pipeline stage returns the next zio to execute
1782 * (typically the same as this one), or NULL if we should
1785 zio = zio_pipeline[highbit64(stage) - 1](zio);
1793 * ==========================================================================
1794 * Initiate I/O, either sync or async
1795 * ==========================================================================
1798 zio_wait(zio_t *zio)
1802 ASSERT3P(zio->io_stage, ==, ZIO_STAGE_OPEN);
1803 ASSERT3P(zio->io_executor, ==, NULL);
1805 zio->io_waiter = curthread;
1806 ASSERT0(zio->io_queued_timestamp);
1807 zio->io_queued_timestamp = gethrtime();
1811 mutex_enter(&zio->io_lock);
1812 while (zio->io_executor != NULL)
1813 cv_wait(&zio->io_cv, &zio->io_lock);
1814 mutex_exit(&zio->io_lock);
1816 error = zio->io_error;
1823 zio_nowait(zio_t *zio)
1825 ASSERT3P(zio->io_executor, ==, NULL);
1827 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
1828 zio_unique_parent(zio) == NULL) {
1830 * This is a logical async I/O with no parent to wait for it.
1831 * We add it to the spa_async_root_zio "Godfather" I/O which
1832 * will ensure they complete prior to unloading the pool.
1834 spa_t *spa = zio->io_spa;
1836 zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio);
1839 ASSERT0(zio->io_queued_timestamp);
1840 zio->io_queued_timestamp = gethrtime();
1845 * ==========================================================================
1846 * Reexecute, cancel, or suspend/resume failed I/O
1847 * ==========================================================================
1851 zio_reexecute(zio_t *pio)
1853 zio_t *cio, *cio_next;
1855 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
1856 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
1857 ASSERT(pio->io_gang_leader == NULL);
1858 ASSERT(pio->io_gang_tree == NULL);
1860 pio->io_flags = pio->io_orig_flags;
1861 pio->io_stage = pio->io_orig_stage;
1862 pio->io_pipeline = pio->io_orig_pipeline;
1863 pio->io_reexecute = 0;
1864 pio->io_flags |= ZIO_FLAG_REEXECUTED;
1865 pio->io_pipeline_trace = 0;
1867 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1868 pio->io_state[w] = 0;
1869 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
1870 pio->io_child_error[c] = 0;
1872 if (IO_IS_ALLOCATING(pio))
1873 BP_ZERO(pio->io_bp);
1876 * As we reexecute pio's children, new children could be created.
1877 * New children go to the head of pio's io_child_list, however,
1878 * so we will (correctly) not reexecute them. The key is that
1879 * the remainder of pio's io_child_list, from 'cio_next' onward,
1880 * cannot be affected by any side effects of reexecuting 'cio'.
1882 zio_link_t *zl = NULL;
1883 mutex_enter(&pio->io_lock);
1884 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
1885 cio_next = zio_walk_children(pio, &zl);
1886 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
1887 pio->io_children[cio->io_child_type][w]++;
1888 mutex_exit(&pio->io_lock);
1890 mutex_enter(&pio->io_lock);
1892 mutex_exit(&pio->io_lock);
1895 * Now that all children have been reexecuted, execute the parent.
1896 * We don't reexecute "The Godfather" I/O here as it's the
1897 * responsibility of the caller to wait on it.
1899 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) {
1900 pio->io_queued_timestamp = gethrtime();
1906 zio_suspend(spa_t *spa, zio_t *zio)
1908 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
1909 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
1910 "failure and the failure mode property for this pool "
1911 "is set to panic.", spa_name(spa));
1913 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
1915 mutex_enter(&spa->spa_suspend_lock);
1917 if (spa->spa_suspend_zio_root == NULL)
1918 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
1919 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
1920 ZIO_FLAG_GODFATHER);
1922 spa->spa_suspended = B_TRUE;
1925 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
1926 ASSERT(zio != spa->spa_suspend_zio_root);
1927 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1928 ASSERT(zio_unique_parent(zio) == NULL);
1929 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
1930 zio_add_child(spa->spa_suspend_zio_root, zio);
1933 mutex_exit(&spa->spa_suspend_lock);
1937 zio_resume(spa_t *spa)
1942 * Reexecute all previously suspended i/o.
1944 mutex_enter(&spa->spa_suspend_lock);
1945 spa->spa_suspended = B_FALSE;
1946 cv_broadcast(&spa->spa_suspend_cv);
1947 pio = spa->spa_suspend_zio_root;
1948 spa->spa_suspend_zio_root = NULL;
1949 mutex_exit(&spa->spa_suspend_lock);
1955 return (zio_wait(pio));
1959 zio_resume_wait(spa_t *spa)
1961 mutex_enter(&spa->spa_suspend_lock);
1962 while (spa_suspended(spa))
1963 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
1964 mutex_exit(&spa->spa_suspend_lock);
1968 * ==========================================================================
1971 * A gang block is a collection of small blocks that looks to the DMU
1972 * like one large block. When zio_dva_allocate() cannot find a block
1973 * of the requested size, due to either severe fragmentation or the pool
1974 * being nearly full, it calls zio_write_gang_block() to construct the
1975 * block from smaller fragments.
1977 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
1978 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
1979 * an indirect block: it's an array of block pointers. It consumes
1980 * only one sector and hence is allocatable regardless of fragmentation.
1981 * The gang header's bps point to its gang members, which hold the data.
1983 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
1984 * as the verifier to ensure uniqueness of the SHA256 checksum.
1985 * Critically, the gang block bp's blk_cksum is the checksum of the data,
1986 * not the gang header. This ensures that data block signatures (needed for
1987 * deduplication) are independent of how the block is physically stored.
1989 * Gang blocks can be nested: a gang member may itself be a gang block.
1990 * Thus every gang block is a tree in which root and all interior nodes are
1991 * gang headers, and the leaves are normal blocks that contain user data.
1992 * The root of the gang tree is called the gang leader.
1994 * To perform any operation (read, rewrite, free, claim) on a gang block,
1995 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
1996 * in the io_gang_tree field of the original logical i/o by recursively
1997 * reading the gang leader and all gang headers below it. This yields
1998 * an in-core tree containing the contents of every gang header and the
1999 * bps for every constituent of the gang block.
2001 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
2002 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
2003 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
2004 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
2005 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
2006 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
2007 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
2008 * of the gang header plus zio_checksum_compute() of the data to update the
2009 * gang header's blk_cksum as described above.
2011 * The two-phase assemble/issue model solves the problem of partial failure --
2012 * what if you'd freed part of a gang block but then couldn't read the
2013 * gang header for another part? Assembling the entire gang tree first
2014 * ensures that all the necessary gang header I/O has succeeded before
2015 * starting the actual work of free, claim, or write. Once the gang tree
2016 * is assembled, free and claim are in-memory operations that cannot fail.
2018 * In the event that a gang write fails, zio_dva_unallocate() walks the
2019 * gang tree to immediately free (i.e. insert back into the space map)
2020 * everything we've allocated. This ensures that we don't get ENOSPC
2021 * errors during repeated suspend/resume cycles due to a flaky device.
2023 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
2024 * the gang tree, we won't modify the block, so we can safely defer the free
2025 * (knowing that the block is still intact). If we *can* assemble the gang
2026 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
2027 * each constituent bp and we can allocate a new block on the next sync pass.
2029 * In all cases, the gang tree allows complete recovery from partial failure.
2030 * ==========================================================================
2034 zio_gang_issue_func_done(zio_t *zio)
2036 abd_put(zio->io_abd);
2040 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2046 return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset),
2047 BP_GET_PSIZE(bp), zio_gang_issue_func_done,
2048 NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2049 &pio->io_bookmark));
2053 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2060 abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2061 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2062 gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL,
2063 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2066 * As we rewrite each gang header, the pipeline will compute
2067 * a new gang block header checksum for it; but no one will
2068 * compute a new data checksum, so we do that here. The one
2069 * exception is the gang leader: the pipeline already computed
2070 * its data checksum because that stage precedes gang assembly.
2071 * (Presently, nothing actually uses interior data checksums;
2072 * this is just good hygiene.)
2074 if (gn != pio->io_gang_leader->io_gang_tree) {
2075 abd_t *buf = abd_get_offset(data, offset);
2077 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
2078 buf, BP_GET_PSIZE(bp));
2083 * If we are here to damage data for testing purposes,
2084 * leave the GBH alone so that we can detect the damage.
2086 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
2087 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2089 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2090 abd_get_offset(data, offset), BP_GET_PSIZE(bp),
2091 zio_gang_issue_func_done, NULL, pio->io_priority,
2092 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2100 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2103 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
2104 BP_IS_GANG(bp) ? SPA_GANGBLOCKSIZE : BP_GET_PSIZE(bp),
2105 ZIO_GANG_CHILD_FLAGS(pio)));
2110 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2113 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
2114 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
2117 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
2126 static void zio_gang_tree_assemble_done(zio_t *zio);
2128 static zio_gang_node_t *
2129 zio_gang_node_alloc(zio_gang_node_t **gnpp)
2131 zio_gang_node_t *gn;
2133 ASSERT(*gnpp == NULL);
2135 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
2136 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
2143 zio_gang_node_free(zio_gang_node_t **gnpp)
2145 zio_gang_node_t *gn = *gnpp;
2147 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2148 ASSERT(gn->gn_child[g] == NULL);
2150 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2151 kmem_free(gn, sizeof (*gn));
2156 zio_gang_tree_free(zio_gang_node_t **gnpp)
2158 zio_gang_node_t *gn = *gnpp;
2163 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2164 zio_gang_tree_free(&gn->gn_child[g]);
2166 zio_gang_node_free(gnpp);
2170 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
2172 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
2173 abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2175 ASSERT(gio->io_gang_leader == gio);
2176 ASSERT(BP_IS_GANG(bp));
2178 zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2179 zio_gang_tree_assemble_done, gn, gio->io_priority,
2180 ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
2184 zio_gang_tree_assemble_done(zio_t *zio)
2186 zio_t *gio = zio->io_gang_leader;
2187 zio_gang_node_t *gn = zio->io_private;
2188 blkptr_t *bp = zio->io_bp;
2190 ASSERT(gio == zio_unique_parent(zio));
2191 ASSERT(zio->io_child_count == 0);
2196 /* this ABD was created from a linear buf in zio_gang_tree_assemble */
2197 if (BP_SHOULD_BYTESWAP(bp))
2198 byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size);
2200 ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh);
2201 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
2202 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2204 abd_put(zio->io_abd);
2206 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2207 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2208 if (!BP_IS_GANG(gbp))
2210 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
2215 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data,
2218 zio_t *gio = pio->io_gang_leader;
2221 ASSERT(BP_IS_GANG(bp) == !!gn);
2222 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
2223 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
2226 * If you're a gang header, your data is in gn->gn_gbh.
2227 * If you're a gang member, your data is in 'data' and gn == NULL.
2229 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset);
2232 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2234 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2235 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2236 if (BP_IS_HOLE(gbp))
2238 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data,
2240 offset += BP_GET_PSIZE(gbp);
2244 if (gn == gio->io_gang_tree && gio->io_abd != NULL)
2245 ASSERT3U(gio->io_size, ==, offset);
2252 zio_gang_assemble(zio_t *zio)
2254 blkptr_t *bp = zio->io_bp;
2256 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
2257 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2259 zio->io_gang_leader = zio;
2261 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
2267 zio_gang_issue(zio_t *zio)
2269 blkptr_t *bp = zio->io_bp;
2271 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT, ZIO_WAIT_DONE)) {
2275 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
2276 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2278 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
2279 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd,
2282 zio_gang_tree_free(&zio->io_gang_tree);
2284 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2290 zio_write_gang_member_ready(zio_t *zio)
2292 zio_t *pio = zio_unique_parent(zio);
2293 zio_t *gio = zio->io_gang_leader;
2294 dva_t *cdva = zio->io_bp->blk_dva;
2295 dva_t *pdva = pio->io_bp->blk_dva;
2298 if (BP_IS_HOLE(zio->io_bp))
2301 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
2303 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
2304 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
2305 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2306 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
2307 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
2309 mutex_enter(&pio->io_lock);
2310 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
2311 ASSERT(DVA_GET_GANG(&pdva[d]));
2312 asize = DVA_GET_ASIZE(&pdva[d]);
2313 asize += DVA_GET_ASIZE(&cdva[d]);
2314 DVA_SET_ASIZE(&pdva[d], asize);
2316 mutex_exit(&pio->io_lock);
2320 zio_write_gang_done(zio_t *zio)
2323 * The io_abd field will be NULL for a zio with no data. The io_flags
2324 * will initially have the ZIO_FLAG_NODATA bit flag set, but we can't
2325 * check for it here as it is cleared in zio_ready.
2327 if (zio->io_abd != NULL)
2328 abd_put(zio->io_abd);
2332 zio_write_gang_block(zio_t *pio)
2334 spa_t *spa = pio->io_spa;
2335 metaslab_class_t *mc = spa_normal_class(spa);
2336 blkptr_t *bp = pio->io_bp;
2337 zio_t *gio = pio->io_gang_leader;
2339 zio_gang_node_t *gn, **gnpp;
2340 zio_gbh_phys_t *gbh;
2342 uint64_t txg = pio->io_txg;
2343 uint64_t resid = pio->io_size;
2345 int copies = gio->io_prop.zp_copies;
2346 int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
2349 boolean_t has_data = !(pio->io_flags & ZIO_FLAG_NODATA);
2351 int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER;
2352 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2353 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2356 flags |= METASLAB_ASYNC_ALLOC;
2357 VERIFY(refcount_held(&mc->mc_alloc_slots[pio->io_allocator],
2361 * The logical zio has already placed a reservation for
2362 * 'copies' allocation slots but gang blocks may require
2363 * additional copies. These additional copies
2364 * (i.e. gbh_copies - copies) are guaranteed to succeed
2365 * since metaslab_class_throttle_reserve() always allows
2366 * additional reservations for gang blocks.
2368 VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies,
2369 pio->io_allocator, pio, flags));
2372 error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE,
2373 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags,
2374 &pio->io_alloc_list, pio, pio->io_allocator);
2376 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2377 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2381 * If we failed to allocate the gang block header then
2382 * we remove any additional allocation reservations that
2383 * we placed here. The original reservation will
2384 * be removed when the logical I/O goes to the ready
2387 metaslab_class_throttle_unreserve(mc,
2388 gbh_copies - copies, pio->io_allocator, pio);
2390 pio->io_error = error;
2395 gnpp = &gio->io_gang_tree;
2397 gnpp = pio->io_private;
2398 ASSERT(pio->io_ready == zio_write_gang_member_ready);
2401 gn = zio_gang_node_alloc(gnpp);
2403 bzero(gbh, SPA_GANGBLOCKSIZE);
2404 gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE);
2407 * Create the gang header.
2409 zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2410 zio_write_gang_done, NULL, pio->io_priority,
2411 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2414 * Create and nowait the gang children.
2416 for (int g = 0; resid != 0; resid -= lsize, g++) {
2417 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2419 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2421 zp.zp_checksum = gio->io_prop.zp_checksum;
2422 zp.zp_compress = ZIO_COMPRESS_OFF;
2423 zp.zp_type = DMU_OT_NONE;
2425 zp.zp_copies = gio->io_prop.zp_copies;
2426 zp.zp_dedup = B_FALSE;
2427 zp.zp_dedup_verify = B_FALSE;
2428 zp.zp_nopwrite = B_FALSE;
2430 zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2431 has_data ? abd_get_offset(pio->io_abd, pio->io_size -
2432 resid) : NULL, lsize, lsize, &zp,
2433 zio_write_gang_member_ready, NULL, NULL,
2434 zio_write_gang_done, &gn->gn_child[g], pio->io_priority,
2435 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2437 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2438 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2442 * Gang children won't throttle but we should
2443 * account for their work, so reserve an allocation
2444 * slot for them here.
2446 VERIFY(metaslab_class_throttle_reserve(mc,
2447 zp.zp_copies, cio->io_allocator, cio, flags));
2453 * Set pio's pipeline to just wait for zio to finish.
2455 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2463 * The zio_nop_write stage in the pipeline determines if allocating a
2464 * new bp is necessary. The nopwrite feature can handle writes in
2465 * either syncing or open context (i.e. zil writes) and as a result is
2466 * mutually exclusive with dedup.
2468 * By leveraging a cryptographically secure checksum, such as SHA256, we
2469 * can compare the checksums of the new data and the old to determine if
2470 * allocating a new block is required. Note that our requirements for
2471 * cryptographic strength are fairly weak: there can't be any accidental
2472 * hash collisions, but we don't need to be secure against intentional
2473 * (malicious) collisions. To trigger a nopwrite, you have to be able
2474 * to write the file to begin with, and triggering an incorrect (hash
2475 * collision) nopwrite is no worse than simply writing to the file.
2476 * That said, there are no known attacks against the checksum algorithms
2477 * used for nopwrite, assuming that the salt and the checksums
2478 * themselves remain secret.
2481 zio_nop_write(zio_t *zio)
2483 blkptr_t *bp = zio->io_bp;
2484 blkptr_t *bp_orig = &zio->io_bp_orig;
2485 zio_prop_t *zp = &zio->io_prop;
2487 ASSERT(BP_GET_LEVEL(bp) == 0);
2488 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2489 ASSERT(zp->zp_nopwrite);
2490 ASSERT(!zp->zp_dedup);
2491 ASSERT(zio->io_bp_override == NULL);
2492 ASSERT(IO_IS_ALLOCATING(zio));
2495 * Check to see if the original bp and the new bp have matching
2496 * characteristics (i.e. same checksum, compression algorithms, etc).
2497 * If they don't then just continue with the pipeline which will
2498 * allocate a new bp.
2500 if (BP_IS_HOLE(bp_orig) ||
2501 !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
2502 ZCHECKSUM_FLAG_NOPWRITE) ||
2503 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2504 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2505 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2506 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2510 * If the checksums match then reset the pipeline so that we
2511 * avoid allocating a new bp and issuing any I/O.
2513 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2514 ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
2515 ZCHECKSUM_FLAG_NOPWRITE);
2516 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2517 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2518 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2519 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2520 sizeof (uint64_t)) == 0);
2523 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2524 zio->io_flags |= ZIO_FLAG_NOPWRITE;
2531 * ==========================================================================
2533 * ==========================================================================
2536 zio_ddt_child_read_done(zio_t *zio)
2538 blkptr_t *bp = zio->io_bp;
2539 ddt_entry_t *dde = zio->io_private;
2541 zio_t *pio = zio_unique_parent(zio);
2543 mutex_enter(&pio->io_lock);
2544 ddp = ddt_phys_select(dde, bp);
2545 if (zio->io_error == 0)
2546 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2548 if (zio->io_error == 0 && dde->dde_repair_abd == NULL)
2549 dde->dde_repair_abd = zio->io_abd;
2551 abd_free(zio->io_abd);
2552 mutex_exit(&pio->io_lock);
2556 zio_ddt_read_start(zio_t *zio)
2558 blkptr_t *bp = zio->io_bp;
2560 ASSERT(BP_GET_DEDUP(bp));
2561 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2562 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2564 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2565 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2566 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2567 ddt_phys_t *ddp = dde->dde_phys;
2568 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2571 ASSERT(zio->io_vsd == NULL);
2574 if (ddp_self == NULL)
2577 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2578 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2580 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2582 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2583 abd_alloc_for_io(zio->io_size, B_TRUE),
2584 zio->io_size, zio_ddt_child_read_done, dde,
2585 zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) |
2586 ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark));
2591 zio_nowait(zio_read(zio, zio->io_spa, bp,
2592 zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority,
2593 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2599 zio_ddt_read_done(zio_t *zio)
2601 blkptr_t *bp = zio->io_bp;
2603 if (zio_wait_for_children(zio, ZIO_CHILD_DDT_BIT, ZIO_WAIT_DONE)) {
2607 ASSERT(BP_GET_DEDUP(bp));
2608 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2609 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2611 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2612 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2613 ddt_entry_t *dde = zio->io_vsd;
2615 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2619 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2620 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2623 if (dde->dde_repair_abd != NULL) {
2624 abd_copy(zio->io_abd, dde->dde_repair_abd,
2626 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2628 ddt_repair_done(ddt, dde);
2632 ASSERT(zio->io_vsd == NULL);
2638 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2640 spa_t *spa = zio->io_spa;
2641 boolean_t do_raw = (zio->io_flags & ZIO_FLAG_RAW);
2643 /* We should never get a raw, override zio */
2644 ASSERT(!(zio->io_bp_override && do_raw));
2647 * Note: we compare the original data, not the transformed data,
2648 * because when zio->io_bp is an override bp, we will not have
2649 * pushed the I/O transforms. That's an important optimization
2650 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2652 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2653 zio_t *lio = dde->dde_lead_zio[p];
2656 return (lio->io_orig_size != zio->io_orig_size ||
2657 abd_cmp(zio->io_orig_abd, lio->io_orig_abd,
2658 zio->io_orig_size) != 0);
2662 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2663 ddt_phys_t *ddp = &dde->dde_phys[p];
2665 if (ddp->ddp_phys_birth != 0) {
2666 arc_buf_t *abuf = NULL;
2667 arc_flags_t aflags = ARC_FLAG_WAIT;
2668 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE;
2669 blkptr_t blk = *zio->io_bp;
2672 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2677 * Intuitively, it would make more sense to compare
2678 * io_abd than io_orig_abd in the raw case since you
2679 * don't want to look at any transformations that have
2680 * happened to the data. However, for raw I/Os the
2681 * data will actually be the same in io_abd and
2682 * io_orig_abd, so all we have to do is issue this as
2686 zio_flags |= ZIO_FLAG_RAW;
2687 ASSERT3U(zio->io_size, ==, zio->io_orig_size);
2688 ASSERT0(abd_cmp(zio->io_abd, zio->io_orig_abd,
2690 ASSERT3P(zio->io_transform_stack, ==, NULL);
2693 error = arc_read(NULL, spa, &blk,
2694 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2695 zio_flags, &aflags, &zio->io_bookmark);
2698 if (arc_buf_size(abuf) != zio->io_orig_size ||
2699 abd_cmp_buf(zio->io_orig_abd, abuf->b_data,
2700 zio->io_orig_size) != 0)
2701 error = SET_ERROR(EEXIST);
2702 arc_buf_destroy(abuf, &abuf);
2706 return (error != 0);
2714 zio_ddt_child_write_ready(zio_t *zio)
2716 int p = zio->io_prop.zp_copies;
2717 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2718 ddt_entry_t *dde = zio->io_private;
2719 ddt_phys_t *ddp = &dde->dde_phys[p];
2727 ASSERT(dde->dde_lead_zio[p] == zio);
2729 ddt_phys_fill(ddp, zio->io_bp);
2731 zio_link_t *zl = NULL;
2732 while ((pio = zio_walk_parents(zio, &zl)) != NULL)
2733 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
2739 zio_ddt_child_write_done(zio_t *zio)
2741 int p = zio->io_prop.zp_copies;
2742 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
2743 ddt_entry_t *dde = zio->io_private;
2744 ddt_phys_t *ddp = &dde->dde_phys[p];
2748 ASSERT(ddp->ddp_refcnt == 0);
2749 ASSERT(dde->dde_lead_zio[p] == zio);
2750 dde->dde_lead_zio[p] = NULL;
2752 if (zio->io_error == 0) {
2753 zio_link_t *zl = NULL;
2754 while (zio_walk_parents(zio, &zl) != NULL)
2755 ddt_phys_addref(ddp);
2757 ddt_phys_clear(ddp);
2764 zio_ddt_ditto_write_done(zio_t *zio)
2766 int p = DDT_PHYS_DITTO;
2767 zio_prop_t *zp = &zio->io_prop;
2768 blkptr_t *bp = zio->io_bp;
2769 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2770 ddt_entry_t *dde = zio->io_private;
2771 ddt_phys_t *ddp = &dde->dde_phys[p];
2772 ddt_key_t *ddk = &dde->dde_key;
2776 ASSERT(ddp->ddp_refcnt == 0);
2777 ASSERT(dde->dde_lead_zio[p] == zio);
2778 dde->dde_lead_zio[p] = NULL;
2780 if (zio->io_error == 0) {
2781 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
2782 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
2783 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
2784 if (ddp->ddp_phys_birth != 0)
2785 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
2786 ddt_phys_fill(ddp, bp);
2793 zio_ddt_write(zio_t *zio)
2795 spa_t *spa = zio->io_spa;
2796 blkptr_t *bp = zio->io_bp;
2797 uint64_t txg = zio->io_txg;
2798 zio_prop_t *zp = &zio->io_prop;
2799 int p = zp->zp_copies;
2803 ddt_t *ddt = ddt_select(spa, bp);
2807 ASSERT(BP_GET_DEDUP(bp));
2808 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
2809 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
2810 ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW)));
2813 dde = ddt_lookup(ddt, bp, B_TRUE);
2814 ddp = &dde->dde_phys[p];
2816 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
2818 * If we're using a weak checksum, upgrade to a strong checksum
2819 * and try again. If we're already using a strong checksum,
2820 * we can't resolve it, so just convert to an ordinary write.
2821 * (And automatically e-mail a paper to Nature?)
2823 if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
2824 ZCHECKSUM_FLAG_DEDUP)) {
2825 zp->zp_checksum = spa_dedup_checksum(spa);
2826 zio_pop_transforms(zio);
2827 zio->io_stage = ZIO_STAGE_OPEN;
2830 zp->zp_dedup = B_FALSE;
2831 BP_SET_DEDUP(bp, B_FALSE);
2833 ASSERT(!BP_GET_DEDUP(bp));
2834 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2839 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
2840 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
2842 if (ditto_copies > ddt_ditto_copies_present(dde) &&
2843 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
2844 zio_prop_t czp = *zp;
2846 czp.zp_copies = ditto_copies;
2849 * If we arrived here with an override bp, we won't have run
2850 * the transform stack, so we won't have the data we need to
2851 * generate a child i/o. So, toss the override bp and restart.
2852 * This is safe, because using the override bp is just an
2853 * optimization; and it's rare, so the cost doesn't matter.
2855 if (zio->io_bp_override) {
2856 zio_pop_transforms(zio);
2857 zio->io_stage = ZIO_STAGE_OPEN;
2858 zio->io_pipeline = ZIO_WRITE_PIPELINE;
2859 zio->io_bp_override = NULL;
2865 dio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
2866 zio->io_orig_size, zio->io_orig_size, &czp, NULL, NULL,
2867 NULL, zio_ddt_ditto_write_done, dde, zio->io_priority,
2868 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2870 zio_push_transform(dio, zio->io_abd, zio->io_size, 0, NULL);
2871 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
2874 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
2875 if (ddp->ddp_phys_birth != 0)
2876 ddt_bp_fill(ddp, bp, txg);
2877 if (dde->dde_lead_zio[p] != NULL)
2878 zio_add_child(zio, dde->dde_lead_zio[p]);
2880 ddt_phys_addref(ddp);
2881 } else if (zio->io_bp_override) {
2882 ASSERT(bp->blk_birth == txg);
2883 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
2884 ddt_phys_fill(ddp, bp);
2885 ddt_phys_addref(ddp);
2887 cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
2888 zio->io_orig_size, zio->io_orig_size, zp,
2889 zio_ddt_child_write_ready, NULL, NULL,
2890 zio_ddt_child_write_done, dde, zio->io_priority,
2891 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
2893 zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL);
2894 dde->dde_lead_zio[p] = cio;
2907 ddt_entry_t *freedde; /* for debugging */
2910 zio_ddt_free(zio_t *zio)
2912 spa_t *spa = zio->io_spa;
2913 blkptr_t *bp = zio->io_bp;
2914 ddt_t *ddt = ddt_select(spa, bp);
2918 ASSERT(BP_GET_DEDUP(bp));
2919 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2922 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
2923 ddp = ddt_phys_select(dde, bp);
2924 ddt_phys_decref(ddp);
2931 * ==========================================================================
2932 * Allocate and free blocks
2933 * ==========================================================================
2937 zio_io_to_allocate(spa_t *spa, int allocator)
2941 ASSERT(MUTEX_HELD(&spa->spa_alloc_locks[allocator]));
2943 zio = avl_first(&spa->spa_alloc_trees[allocator]);
2947 ASSERT(IO_IS_ALLOCATING(zio));
2950 * Try to place a reservation for this zio. If we're unable to
2951 * reserve then we throttle.
2953 ASSERT3U(zio->io_allocator, ==, allocator);
2954 if (!metaslab_class_throttle_reserve(spa_normal_class(spa),
2955 zio->io_prop.zp_copies, zio->io_allocator, zio, 0)) {
2959 avl_remove(&spa->spa_alloc_trees[allocator], zio);
2960 ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE);
2966 zio_dva_throttle(zio_t *zio)
2968 spa_t *spa = zio->io_spa;
2971 if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE ||
2972 !spa_normal_class(zio->io_spa)->mc_alloc_throttle_enabled ||
2973 zio->io_child_type == ZIO_CHILD_GANG ||
2974 zio->io_flags & ZIO_FLAG_NODATA) {
2978 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2980 ASSERT3U(zio->io_queued_timestamp, >, 0);
2981 ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE);
2983 zbookmark_phys_t *bm = &zio->io_bookmark;
2985 * We want to try to use as many allocators as possible to help improve
2986 * performance, but we also want logically adjacent IOs to be physically
2987 * adjacent to improve sequential read performance. We chunk each object
2988 * into 2^20 block regions, and then hash based on the objset, object,
2989 * level, and region to accomplish both of these goals.
2991 zio->io_allocator = cityhash4(bm->zb_objset, bm->zb_object,
2992 bm->zb_level, bm->zb_blkid >> 20) % spa->spa_alloc_count;
2993 mutex_enter(&spa->spa_alloc_locks[zio->io_allocator]);
2995 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
2996 avl_add(&spa->spa_alloc_trees[zio->io_allocator], zio);
2998 nio = zio_io_to_allocate(zio->io_spa, zio->io_allocator);
2999 mutex_exit(&spa->spa_alloc_locks[zio->io_allocator]);
3005 zio_allocate_dispatch(spa_t *spa, int allocator)
3009 mutex_enter(&spa->spa_alloc_locks[allocator]);
3010 zio = zio_io_to_allocate(spa, allocator);
3011 mutex_exit(&spa->spa_alloc_locks[allocator]);
3015 ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE);
3016 ASSERT0(zio->io_error);
3017 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE);
3021 zio_dva_allocate(zio_t *zio)
3023 spa_t *spa = zio->io_spa;
3024 metaslab_class_t *mc = spa_normal_class(spa);
3025 blkptr_t *bp = zio->io_bp;
3029 if (zio->io_gang_leader == NULL) {
3030 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3031 zio->io_gang_leader = zio;
3034 ASSERT(BP_IS_HOLE(bp));
3035 ASSERT0(BP_GET_NDVAS(bp));
3036 ASSERT3U(zio->io_prop.zp_copies, >, 0);
3037 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
3038 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
3040 if (zio->io_flags & ZIO_FLAG_NODATA) {
3041 flags |= METASLAB_DONT_THROTTLE;
3043 if (zio->io_flags & ZIO_FLAG_GANG_CHILD) {
3044 flags |= METASLAB_GANG_CHILD;
3046 if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE) {
3047 flags |= METASLAB_ASYNC_ALLOC;
3050 error = metaslab_alloc(spa, mc, zio->io_size, bp,
3051 zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
3052 &zio->io_alloc_list, zio, zio->io_allocator);
3055 zfs_dbgmsg("%s: metaslab allocation failure: zio %p, "
3056 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
3058 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
3059 return (zio_write_gang_block(zio));
3060 zio->io_error = error;
3067 zio_dva_free(zio_t *zio)
3069 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
3075 zio_dva_claim(zio_t *zio)
3079 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
3081 zio->io_error = error;
3087 * Undo an allocation. This is used by zio_done() when an I/O fails
3088 * and we want to give back the block we just allocated.
3089 * This handles both normal blocks and gang blocks.
3092 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
3094 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
3095 ASSERT(zio->io_bp_override == NULL);
3097 if (!BP_IS_HOLE(bp))
3098 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
3101 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
3102 zio_dva_unallocate(zio, gn->gn_child[g],
3103 &gn->gn_gbh->zg_blkptr[g]);
3109 * Try to allocate an intent log block. Return 0 on success, errno on failure.
3112 zio_alloc_zil(spa_t *spa, uint64_t objset, uint64_t txg, blkptr_t *new_bp,
3113 blkptr_t *old_bp, uint64_t size, boolean_t *slog)
3116 zio_alloc_list_t io_alloc_list;
3118 ASSERT(txg > spa_syncing_txg(spa));
3120 metaslab_trace_init(&io_alloc_list);
3122 * When allocating a zil block, we don't have information about
3123 * the final destination of the block except the objset it's part
3124 * of, so we just hash the objset ID to pick the allocator to get
3127 error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1,
3128 txg, old_bp, METASLAB_HINTBP_AVOID, &io_alloc_list, NULL,
3129 cityhash4(0, 0, 0, objset) % spa->spa_alloc_count);
3133 error = metaslab_alloc(spa, spa_normal_class(spa), size,
3134 new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID,
3135 &io_alloc_list, NULL, cityhash4(0, 0, 0, objset) %
3136 spa->spa_alloc_count);
3140 metaslab_trace_fini(&io_alloc_list);
3143 BP_SET_LSIZE(new_bp, size);
3144 BP_SET_PSIZE(new_bp, size);
3145 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
3146 BP_SET_CHECKSUM(new_bp,
3147 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
3148 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
3149 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3150 BP_SET_LEVEL(new_bp, 0);
3151 BP_SET_DEDUP(new_bp, 0);
3152 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
3154 zfs_dbgmsg("%s: zil block allocation failure: "
3155 "size %llu, error %d", spa_name(spa), size, error);
3162 * ==========================================================================
3163 * Read, write and delete to physical devices
3164 * ==========================================================================
3169 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3170 * stops after this stage and will resume upon I/O completion.
3171 * However, there are instances where the vdev layer may need to
3172 * continue the pipeline when an I/O was not issued. Since the I/O
3173 * that was sent to the vdev layer might be different than the one
3174 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3175 * force the underlying vdev layers to call either zio_execute() or
3176 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3179 zio_vdev_io_start(zio_t *zio)
3181 vdev_t *vd = zio->io_vd;
3183 spa_t *spa = zio->io_spa;
3186 ASSERT(zio->io_error == 0);
3187 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
3190 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3191 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
3194 * The mirror_ops handle multiple DVAs in a single BP.
3196 vdev_mirror_ops.vdev_op_io_start(zio);
3200 if (vd->vdev_ops->vdev_op_leaf && zio->io_type == ZIO_TYPE_FREE &&
3201 zio->io_priority == ZIO_PRIORITY_NOW) {
3202 trim_map_free(vd, zio->io_offset, zio->io_size, zio->io_txg);
3206 ASSERT3P(zio->io_logical, !=, zio);
3207 if (zio->io_type == ZIO_TYPE_WRITE) {
3208 ASSERT(spa->spa_trust_config);
3210 if (zio->io_vd->vdev_removing) {
3212 * Note: the code can handle other kinds of writes,
3213 * but we don't expect them.
3215 ASSERT(zio->io_flags &
3216 (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL |
3217 ZIO_FLAG_RESILVER | ZIO_FLAG_INDUCE_DAMAGE));
3222 * We keep track of time-sensitive I/Os so that the scan thread
3223 * can quickly react to certain workloads. In particular, we care
3224 * about non-scrubbing, top-level reads and writes with the following
3226 * - synchronous writes of user data to non-slog devices
3227 * - any reads of user data
3228 * When these conditions are met, adjust the timestamp of spa_last_io
3229 * which allows the scan thread to adjust its workload accordingly.
3231 if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
3232 vd == vd->vdev_top && !vd->vdev_islog &&
3233 zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
3234 zio->io_txg != spa_syncing_txg(spa)) {
3235 uint64_t old = spa->spa_last_io;
3236 uint64_t new = ddi_get_lbolt64();
3238 (void) atomic_cas_64(&spa->spa_last_io, old, new);
3240 align = 1ULL << vd->vdev_top->vdev_ashift;
3242 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
3243 P2PHASE(zio->io_size, align) != 0) {
3244 /* Transform logical writes to be a full physical block size. */
3245 uint64_t asize = P2ROUNDUP(zio->io_size, align);
3247 if (zio->io_type == ZIO_TYPE_READ ||
3248 zio->io_type == ZIO_TYPE_WRITE)
3249 abuf = abd_alloc_sametype(zio->io_abd, asize);
3250 ASSERT(vd == vd->vdev_top);
3251 if (zio->io_type == ZIO_TYPE_WRITE) {
3252 abd_copy(abuf, zio->io_abd, zio->io_size);
3253 abd_zero_off(abuf, zio->io_size, asize - zio->io_size);
3255 zio_push_transform(zio, abuf, asize, abuf ? asize : 0,
3260 * If this is not a physical io, make sure that it is properly aligned
3261 * before proceeding.
3263 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
3264 ASSERT0(P2PHASE(zio->io_offset, align));
3265 ASSERT0(P2PHASE(zio->io_size, align));
3268 * For the physical io we allow alignment
3269 * to a logical block size.
3271 uint64_t log_align =
3272 1ULL << vd->vdev_top->vdev_logical_ashift;
3273 ASSERT0(P2PHASE(zio->io_offset, log_align));
3274 ASSERT0(P2PHASE(zio->io_size, log_align));
3277 VERIFY(zio->io_type == ZIO_TYPE_READ || spa_writeable(spa));
3280 * If this is a repair I/O, and there's no self-healing involved --
3281 * that is, we're just resilvering what we expect to resilver --
3282 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3283 * This prevents spurious resilvering.
3285 * There are a few ways that we can end up creating these spurious
3288 * 1. A resilver i/o will be issued if any DVA in the BP has a
3289 * dirty DTL. The mirror code will issue resilver writes to
3290 * each DVA, including the one(s) that are not on vdevs with dirty
3293 * 2. With nested replication, which happens when we have a
3294 * "replacing" or "spare" vdev that's a child of a mirror or raidz.
3295 * For example, given mirror(replacing(A+B), C), it's likely that
3296 * only A is out of date (it's the new device). In this case, we'll
3297 * read from C, then use the data to resilver A+B -- but we don't
3298 * actually want to resilver B, just A. The top-level mirror has no
3299 * way to know this, so instead we just discard unnecessary repairs
3300 * as we work our way down the vdev tree.
3302 * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc.
3303 * The same logic applies to any form of nested replication: ditto
3304 * + mirror, RAID-Z + replacing, etc.
3306 * However, indirect vdevs point off to other vdevs which may have
3307 * DTL's, so we never bypass them. The child i/os on concrete vdevs
3308 * will be properly bypassed instead.
3310 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3311 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
3312 zio->io_txg != 0 && /* not a delegated i/o */
3313 vd->vdev_ops != &vdev_indirect_ops &&
3314 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
3315 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3316 zio_vdev_io_bypass(zio);
3320 if (vd->vdev_ops->vdev_op_leaf) {
3321 switch (zio->io_type) {
3323 if (vdev_cache_read(zio))
3326 case ZIO_TYPE_WRITE:
3328 if ((zio = vdev_queue_io(zio)) == NULL)
3331 if (!vdev_accessible(vd, zio)) {
3332 zio->io_error = SET_ERROR(ENXIO);
3339 * Note that we ignore repair writes for TRIM because they can
3340 * conflict with normal writes. This isn't an issue because, by
3341 * definition, we only repair blocks that aren't freed.
3343 if (zio->io_type == ZIO_TYPE_WRITE &&
3344 !(zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3345 !trim_map_write_start(zio))
3349 vd->vdev_ops->vdev_op_io_start(zio);
3354 zio_vdev_io_done(zio_t *zio)
3356 vdev_t *vd = zio->io_vd;
3357 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
3358 boolean_t unexpected_error = B_FALSE;
3360 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3364 ASSERT(zio->io_type == ZIO_TYPE_READ ||
3365 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_FREE);
3367 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3368 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE ||
3369 zio->io_type == ZIO_TYPE_FREE)) {
3371 if (zio->io_type == ZIO_TYPE_WRITE &&
3372 !(zio->io_flags & ZIO_FLAG_IO_REPAIR))
3373 trim_map_write_done(zio);
3375 vdev_queue_io_done(zio);
3377 if (zio->io_type == ZIO_TYPE_WRITE)
3378 vdev_cache_write(zio);
3380 if (zio_injection_enabled && zio->io_error == 0)
3381 zio->io_error = zio_handle_device_injection(vd,
3384 if (zio_injection_enabled && zio->io_error == 0)
3385 zio->io_error = zio_handle_label_injection(zio, EIO);
3387 if (zio->io_error) {
3388 if (zio->io_error == ENOTSUP &&
3389 zio->io_type == ZIO_TYPE_FREE) {
3390 /* Not all devices support TRIM. */
3391 } else if (!vdev_accessible(vd, zio)) {
3392 zio->io_error = SET_ERROR(ENXIO);
3394 unexpected_error = B_TRUE;
3399 ops->vdev_op_io_done(zio);
3401 if (unexpected_error)
3402 VERIFY(vdev_probe(vd, zio) == NULL);
3408 * This function is used to change the priority of an existing zio that is
3409 * currently in-flight. This is used by the arc to upgrade priority in the
3410 * event that a demand read is made for a block that is currently queued
3411 * as a scrub or async read IO. Otherwise, the high priority read request
3412 * would end up having to wait for the lower priority IO.
3415 zio_change_priority(zio_t *pio, zio_priority_t priority)
3417 zio_t *cio, *cio_next;
3418 zio_link_t *zl = NULL;
3420 ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
3422 if (pio->io_vd != NULL && pio->io_vd->vdev_ops->vdev_op_leaf) {
3423 vdev_queue_change_io_priority(pio, priority);
3425 pio->io_priority = priority;
3428 mutex_enter(&pio->io_lock);
3429 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
3430 cio_next = zio_walk_children(pio, &zl);
3431 zio_change_priority(cio, priority);
3433 mutex_exit(&pio->io_lock);
3437 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3438 * disk, and use that to finish the checksum ereport later.
3441 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
3442 const void *good_buf)
3444 /* no processing needed */
3445 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
3450 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
3452 void *buf = zio_buf_alloc(zio->io_size);
3454 abd_copy_to_buf(buf, zio->io_abd, zio->io_size);
3456 zcr->zcr_cbinfo = zio->io_size;
3457 zcr->zcr_cbdata = buf;
3458 zcr->zcr_finish = zio_vsd_default_cksum_finish;
3459 zcr->zcr_free = zio_buf_free;
3463 zio_vdev_io_assess(zio_t *zio)
3465 vdev_t *vd = zio->io_vd;
3467 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3471 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3472 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
3474 if (zio->io_vsd != NULL) {
3475 zio->io_vsd_ops->vsd_free(zio);
3479 if (zio_injection_enabled && zio->io_error == 0)
3480 zio->io_error = zio_handle_fault_injection(zio, EIO);
3482 if (zio->io_type == ZIO_TYPE_FREE &&
3483 zio->io_priority != ZIO_PRIORITY_NOW) {
3484 switch (zio->io_error) {
3486 ZIO_TRIM_STAT_INCR(bytes, zio->io_size);
3487 ZIO_TRIM_STAT_BUMP(success);
3490 ZIO_TRIM_STAT_BUMP(unsupported);
3493 ZIO_TRIM_STAT_BUMP(failed);
3499 * If the I/O failed, determine whether we should attempt to retry it.
3501 * On retry, we cut in line in the issue queue, since we don't want
3502 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3504 if (zio->io_error && vd == NULL &&
3505 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
3506 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
3507 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
3509 zio->io_flags |= ZIO_FLAG_IO_RETRY |
3510 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
3511 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
3512 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
3513 zio_requeue_io_start_cut_in_line);
3518 * If we got an error on a leaf device, convert it to ENXIO
3519 * if the device is not accessible at all.
3521 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3522 !vdev_accessible(vd, zio))
3523 zio->io_error = SET_ERROR(ENXIO);
3526 * If we can't write to an interior vdev (mirror or RAID-Z),
3527 * set vdev_cant_write so that we stop trying to allocate from it.
3529 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
3530 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
3531 vd->vdev_cant_write = B_TRUE;
3535 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
3536 * attempts will ever succeed. In this case we set a persistent bit so
3537 * that we don't bother with it in the future.
3539 if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) &&
3540 zio->io_type == ZIO_TYPE_IOCTL &&
3541 zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL)
3542 vd->vdev_nowritecache = B_TRUE;
3545 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3547 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3548 zio->io_physdone != NULL) {
3549 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
3550 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
3551 zio->io_physdone(zio->io_logical);
3558 zio_vdev_io_reissue(zio_t *zio)
3560 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3561 ASSERT(zio->io_error == 0);
3563 zio->io_stage >>= 1;
3567 zio_vdev_io_redone(zio_t *zio)
3569 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
3571 zio->io_stage >>= 1;
3575 zio_vdev_io_bypass(zio_t *zio)
3577 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3578 ASSERT(zio->io_error == 0);
3580 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
3581 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
3585 * ==========================================================================
3586 * Generate and verify checksums
3587 * ==========================================================================
3590 zio_checksum_generate(zio_t *zio)
3592 blkptr_t *bp = zio->io_bp;
3593 enum zio_checksum checksum;
3597 * This is zio_write_phys().
3598 * We're either generating a label checksum, or none at all.
3600 checksum = zio->io_prop.zp_checksum;
3602 if (checksum == ZIO_CHECKSUM_OFF)
3605 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
3607 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
3608 ASSERT(!IO_IS_ALLOCATING(zio));
3609 checksum = ZIO_CHECKSUM_GANG_HEADER;
3611 checksum = BP_GET_CHECKSUM(bp);
3615 zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size);
3621 zio_checksum_verify(zio_t *zio)
3623 zio_bad_cksum_t info;
3624 blkptr_t *bp = zio->io_bp;
3627 ASSERT(zio->io_vd != NULL);
3631 * This is zio_read_phys().
3632 * We're either verifying a label checksum, or nothing at all.
3634 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
3637 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
3640 if ((error = zio_checksum_error(zio, &info)) != 0) {
3641 zio->io_error = error;
3642 if (error == ECKSUM &&
3643 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
3644 zfs_ereport_start_checksum(zio->io_spa,
3645 zio->io_vd, zio, zio->io_offset,
3646 zio->io_size, NULL, &info);
3654 * Called by RAID-Z to ensure we don't compute the checksum twice.
3657 zio_checksum_verified(zio_t *zio)
3659 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
3663 * ==========================================================================
3664 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
3665 * An error of 0 indicates success. ENXIO indicates whole-device failure,
3666 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
3667 * indicate errors that are specific to one I/O, and most likely permanent.
3668 * Any other error is presumed to be worse because we weren't expecting it.
3669 * ==========================================================================
3672 zio_worst_error(int e1, int e2)
3674 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
3677 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
3678 if (e1 == zio_error_rank[r1])
3681 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
3682 if (e2 == zio_error_rank[r2])
3685 return (r1 > r2 ? e1 : e2);
3689 * ==========================================================================
3691 * ==========================================================================
3694 zio_ready(zio_t *zio)
3696 blkptr_t *bp = zio->io_bp;
3697 zio_t *pio, *pio_next;
3698 zio_link_t *zl = NULL;
3700 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT | ZIO_CHILD_DDT_BIT,
3705 if (zio->io_ready) {
3706 ASSERT(IO_IS_ALLOCATING(zio));
3707 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
3708 (zio->io_flags & ZIO_FLAG_NOPWRITE));
3709 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
3714 if (bp != NULL && bp != &zio->io_bp_copy)
3715 zio->io_bp_copy = *bp;
3717 if (zio->io_error != 0) {
3718 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3720 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3721 ASSERT(IO_IS_ALLOCATING(zio));
3722 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3724 * We were unable to allocate anything, unreserve and
3725 * issue the next I/O to allocate.
3727 metaslab_class_throttle_unreserve(
3728 spa_normal_class(zio->io_spa),
3729 zio->io_prop.zp_copies, zio->io_allocator, zio);
3730 zio_allocate_dispatch(zio->io_spa, zio->io_allocator);
3734 mutex_enter(&zio->io_lock);
3735 zio->io_state[ZIO_WAIT_READY] = 1;
3736 pio = zio_walk_parents(zio, &zl);
3737 mutex_exit(&zio->io_lock);
3740 * As we notify zio's parents, new parents could be added.
3741 * New parents go to the head of zio's io_parent_list, however,
3742 * so we will (correctly) not notify them. The remainder of zio's
3743 * io_parent_list, from 'pio_next' onward, cannot change because
3744 * all parents must wait for us to be done before they can be done.
3746 for (; pio != NULL; pio = pio_next) {
3747 pio_next = zio_walk_parents(zio, &zl);
3748 zio_notify_parent(pio, zio, ZIO_WAIT_READY, NULL);
3751 if (zio->io_flags & ZIO_FLAG_NODATA) {
3752 if (BP_IS_GANG(bp)) {
3753 zio->io_flags &= ~ZIO_FLAG_NODATA;
3755 ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE);
3756 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
3760 if (zio_injection_enabled &&
3761 zio->io_spa->spa_syncing_txg == zio->io_txg)
3762 zio_handle_ignored_writes(zio);
3768 * Update the allocation throttle accounting.
3771 zio_dva_throttle_done(zio_t *zio)
3773 zio_t *lio = zio->io_logical;
3774 zio_t *pio = zio_unique_parent(zio);
3775 vdev_t *vd = zio->io_vd;
3776 int flags = METASLAB_ASYNC_ALLOC;
3778 ASSERT3P(zio->io_bp, !=, NULL);
3779 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
3780 ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE);
3781 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
3783 ASSERT3P(vd, ==, vd->vdev_top);
3784 ASSERT(!(zio->io_flags & (ZIO_FLAG_IO_REPAIR | ZIO_FLAG_IO_RETRY)));
3785 ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING);
3786 ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE));
3787 ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA));
3790 * Parents of gang children can have two flavors -- ones that
3791 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
3792 * and ones that allocated the constituent blocks. The allocation
3793 * throttle needs to know the allocating parent zio so we must find
3796 if (pio->io_child_type == ZIO_CHILD_GANG) {
3798 * If our parent is a rewrite gang child then our grandparent
3799 * would have been the one that performed the allocation.
3801 if (pio->io_flags & ZIO_FLAG_IO_REWRITE)
3802 pio = zio_unique_parent(pio);
3803 flags |= METASLAB_GANG_CHILD;
3806 ASSERT(IO_IS_ALLOCATING(pio));
3807 ASSERT3P(zio, !=, zio->io_logical);
3808 ASSERT(zio->io_logical != NULL);
3809 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
3810 ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE);
3812 mutex_enter(&pio->io_lock);
3813 metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags,
3814 pio->io_allocator, B_TRUE);
3815 mutex_exit(&pio->io_lock);
3817 metaslab_class_throttle_unreserve(spa_normal_class(zio->io_spa),
3818 1, pio->io_allocator, pio);
3821 * Call into the pipeline to see if there is more work that
3822 * needs to be done. If there is work to be done it will be
3823 * dispatched to another taskq thread.
3825 zio_allocate_dispatch(zio->io_spa, pio->io_allocator);
3829 zio_done(zio_t *zio)
3831 spa_t *spa = zio->io_spa;
3832 zio_t *lio = zio->io_logical;
3833 blkptr_t *bp = zio->io_bp;
3834 vdev_t *vd = zio->io_vd;
3835 uint64_t psize = zio->io_size;
3836 zio_t *pio, *pio_next;
3837 metaslab_class_t *mc = spa_normal_class(spa);
3838 zio_link_t *zl = NULL;
3841 * If our children haven't all completed,
3842 * wait for them and then repeat this pipeline stage.
3844 if (zio_wait_for_children(zio, ZIO_CHILD_ALL_BITS, ZIO_WAIT_DONE)) {
3849 * If the allocation throttle is enabled, then update the accounting.
3850 * We only track child I/Os that are part of an allocating async
3851 * write. We must do this since the allocation is performed
3852 * by the logical I/O but the actual write is done by child I/Os.
3854 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING &&
3855 zio->io_child_type == ZIO_CHILD_VDEV) {
3856 ASSERT(mc->mc_alloc_throttle_enabled);
3857 zio_dva_throttle_done(zio);
3861 * If the allocation throttle is enabled, verify that
3862 * we have decremented the refcounts for every I/O that was throttled.
3864 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
3865 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3866 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
3868 metaslab_group_alloc_verify(spa, zio->io_bp, zio,
3870 VERIFY(refcount_not_held(&mc->mc_alloc_slots[zio->io_allocator],
3874 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
3875 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
3876 ASSERT(zio->io_children[c][w] == 0);
3878 if (bp != NULL && !BP_IS_EMBEDDED(bp)) {
3879 ASSERT(bp->blk_pad[0] == 0);
3880 ASSERT(bp->blk_pad[1] == 0);
3881 ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 ||
3882 (bp == zio_unique_parent(zio)->io_bp));
3883 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) &&
3884 zio->io_bp_override == NULL &&
3885 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
3886 ASSERT(!BP_SHOULD_BYTESWAP(bp));
3887 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp));
3888 ASSERT(BP_COUNT_GANG(bp) == 0 ||
3889 (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp)));
3891 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
3892 VERIFY(BP_EQUAL(bp, &zio->io_bp_orig));
3896 * If there were child vdev/gang/ddt errors, they apply to us now.
3898 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
3899 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
3900 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
3903 * If the I/O on the transformed data was successful, generate any
3904 * checksum reports now while we still have the transformed data.
3906 if (zio->io_error == 0) {
3907 while (zio->io_cksum_report != NULL) {
3908 zio_cksum_report_t *zcr = zio->io_cksum_report;
3909 uint64_t align = zcr->zcr_align;
3910 uint64_t asize = P2ROUNDUP(psize, align);
3912 abd_t *adata = zio->io_abd;
3914 if (asize != psize) {
3915 adata = abd_alloc_linear(asize, B_TRUE);
3916 abd_copy(adata, zio->io_abd, psize);
3917 abd_zero_off(adata, psize, asize - psize);
3921 abuf = abd_borrow_buf_copy(adata, asize);
3923 zio->io_cksum_report = zcr->zcr_next;
3924 zcr->zcr_next = NULL;
3925 zcr->zcr_finish(zcr, abuf);
3926 zfs_ereport_free_checksum(zcr);
3929 abd_return_buf(adata, abuf, asize);
3936 zio_pop_transforms(zio); /* note: may set zio->io_error */
3938 vdev_stat_update(zio, psize);
3940 if (zio->io_error) {
3942 * If this I/O is attached to a particular vdev,
3943 * generate an error message describing the I/O failure
3944 * at the block level. We ignore these errors if the
3945 * device is currently unavailable.
3947 if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd))
3948 zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0);
3950 if ((zio->io_error == EIO || !(zio->io_flags &
3951 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
3954 * For logical I/O requests, tell the SPA to log the
3955 * error and generate a logical data ereport.
3957 spa_log_error(spa, zio);
3958 zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio,
3963 if (zio->io_error && zio == lio) {
3965 * Determine whether zio should be reexecuted. This will
3966 * propagate all the way to the root via zio_notify_parent().
3968 ASSERT(vd == NULL && bp != NULL);
3969 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3971 if (IO_IS_ALLOCATING(zio) &&
3972 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
3973 if (zio->io_error != ENOSPC)
3974 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
3976 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3979 if ((zio->io_type == ZIO_TYPE_READ ||
3980 zio->io_type == ZIO_TYPE_FREE) &&
3981 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
3982 zio->io_error == ENXIO &&
3983 spa_load_state(spa) == SPA_LOAD_NONE &&
3984 spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE)
3985 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3987 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
3988 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
3991 * Here is a possibly good place to attempt to do
3992 * either combinatorial reconstruction or error correction
3993 * based on checksums. It also might be a good place
3994 * to send out preliminary ereports before we suspend
4000 * If there were logical child errors, they apply to us now.
4001 * We defer this until now to avoid conflating logical child
4002 * errors with errors that happened to the zio itself when
4003 * updating vdev stats and reporting FMA events above.
4005 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
4007 if ((zio->io_error || zio->io_reexecute) &&
4008 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
4009 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
4010 zio_dva_unallocate(zio, zio->io_gang_tree, bp);
4012 zio_gang_tree_free(&zio->io_gang_tree);
4015 * Godfather I/Os should never suspend.
4017 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
4018 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
4019 zio->io_reexecute = 0;
4021 if (zio->io_reexecute) {
4023 * This is a logical I/O that wants to reexecute.
4025 * Reexecute is top-down. When an i/o fails, if it's not
4026 * the root, it simply notifies its parent and sticks around.
4027 * The parent, seeing that it still has children in zio_done(),
4028 * does the same. This percolates all the way up to the root.
4029 * The root i/o will reexecute or suspend the entire tree.
4031 * This approach ensures that zio_reexecute() honors
4032 * all the original i/o dependency relationships, e.g.
4033 * parents not executing until children are ready.
4035 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
4037 zio->io_gang_leader = NULL;
4039 mutex_enter(&zio->io_lock);
4040 zio->io_state[ZIO_WAIT_DONE] = 1;
4041 mutex_exit(&zio->io_lock);
4044 * "The Godfather" I/O monitors its children but is
4045 * not a true parent to them. It will track them through
4046 * the pipeline but severs its ties whenever they get into
4047 * trouble (e.g. suspended). This allows "The Godfather"
4048 * I/O to return status without blocking.
4051 for (pio = zio_walk_parents(zio, &zl); pio != NULL;
4053 zio_link_t *remove_zl = zl;
4054 pio_next = zio_walk_parents(zio, &zl);
4056 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
4057 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
4058 zio_remove_child(pio, zio, remove_zl);
4060 * This is a rare code path, so we don't
4061 * bother with "next_to_execute".
4063 zio_notify_parent(pio, zio, ZIO_WAIT_DONE,
4068 if ((pio = zio_unique_parent(zio)) != NULL) {
4070 * We're not a root i/o, so there's nothing to do
4071 * but notify our parent. Don't propagate errors
4072 * upward since we haven't permanently failed yet.
4074 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
4075 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
4077 * This is a rare code path, so we don't bother with
4078 * "next_to_execute".
4080 zio_notify_parent(pio, zio, ZIO_WAIT_DONE, NULL);
4081 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
4083 * We'd fail again if we reexecuted now, so suspend
4084 * until conditions improve (e.g. device comes online).
4086 zio_suspend(spa, zio);
4089 * Reexecution is potentially a huge amount of work.
4090 * Hand it off to the otherwise-unused claim taskq.
4092 #if defined(illumos) || !defined(_KERNEL)
4093 ASSERT(zio->io_tqent.tqent_next == NULL);
4095 ASSERT(zio->io_tqent.tqent_task.ta_pending == 0);
4097 spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM,
4098 ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio,
4104 ASSERT(zio->io_child_count == 0);
4105 ASSERT(zio->io_reexecute == 0);
4106 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
4109 * Report any checksum errors, since the I/O is complete.
4111 while (zio->io_cksum_report != NULL) {
4112 zio_cksum_report_t *zcr = zio->io_cksum_report;
4113 zio->io_cksum_report = zcr->zcr_next;
4114 zcr->zcr_next = NULL;
4115 zcr->zcr_finish(zcr, NULL);
4116 zfs_ereport_free_checksum(zcr);
4120 * It is the responsibility of the done callback to ensure that this
4121 * particular zio is no longer discoverable for adoption, and as
4122 * such, cannot acquire any new parents.
4127 mutex_enter(&zio->io_lock);
4128 zio->io_state[ZIO_WAIT_DONE] = 1;
4129 mutex_exit(&zio->io_lock);
4132 * We are done executing this zio. We may want to execute a parent
4133 * next. See the comment in zio_notify_parent().
4135 zio_t *next_to_execute = NULL;
4137 for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) {
4138 zio_link_t *remove_zl = zl;
4139 pio_next = zio_walk_parents(zio, &zl);
4140 zio_remove_child(pio, zio, remove_zl);
4141 zio_notify_parent(pio, zio, ZIO_WAIT_DONE, &next_to_execute);
4144 if (zio->io_waiter != NULL) {
4145 mutex_enter(&zio->io_lock);
4146 zio->io_executor = NULL;
4147 cv_broadcast(&zio->io_cv);
4148 mutex_exit(&zio->io_lock);
4153 return (next_to_execute);
4157 * ==========================================================================
4158 * I/O pipeline definition
4159 * ==========================================================================
4161 static zio_pipe_stage_t *zio_pipeline[] = {
4168 zio_checksum_generate,
4184 zio_checksum_verify,
4192 * Compare two zbookmark_phys_t's to see which we would reach first in a
4193 * pre-order traversal of the object tree.
4195 * This is simple in every case aside from the meta-dnode object. For all other
4196 * objects, we traverse them in order (object 1 before object 2, and so on).
4197 * However, all of these objects are traversed while traversing object 0, since
4198 * the data it points to is the list of objects. Thus, we need to convert to a
4199 * canonical representation so we can compare meta-dnode bookmarks to
4200 * non-meta-dnode bookmarks.
4202 * We do this by calculating "equivalents" for each field of the zbookmark.
4203 * zbookmarks outside of the meta-dnode use their own object and level, and
4204 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
4205 * blocks this bookmark refers to) by multiplying their blkid by their span
4206 * (the number of L0 blocks contained within one block at their level).
4207 * zbookmarks inside the meta-dnode calculate their object equivalent
4208 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
4209 * level + 1<<31 (any value larger than a level could ever be) for their level.
4210 * This causes them to always compare before a bookmark in their object
4211 * equivalent, compare appropriately to bookmarks in other objects, and to
4212 * compare appropriately to other bookmarks in the meta-dnode.
4215 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
4216 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
4219 * These variables represent the "equivalent" values for the zbookmark,
4220 * after converting zbookmarks inside the meta dnode to their
4221 * normal-object equivalents.
4223 uint64_t zb1obj, zb2obj;
4224 uint64_t zb1L0, zb2L0;
4225 uint64_t zb1level, zb2level;
4227 if (zb1->zb_object == zb2->zb_object &&
4228 zb1->zb_level == zb2->zb_level &&
4229 zb1->zb_blkid == zb2->zb_blkid)
4233 * BP_SPANB calculates the span in blocks.
4235 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
4236 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
4238 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
4239 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4241 zb1level = zb1->zb_level + COMPARE_META_LEVEL;
4243 zb1obj = zb1->zb_object;
4244 zb1level = zb1->zb_level;
4247 if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
4248 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4250 zb2level = zb2->zb_level + COMPARE_META_LEVEL;
4252 zb2obj = zb2->zb_object;
4253 zb2level = zb2->zb_level;
4256 /* Now that we have a canonical representation, do the comparison. */
4257 if (zb1obj != zb2obj)
4258 return (zb1obj < zb2obj ? -1 : 1);
4259 else if (zb1L0 != zb2L0)
4260 return (zb1L0 < zb2L0 ? -1 : 1);
4261 else if (zb1level != zb2level)
4262 return (zb1level > zb2level ? -1 : 1);
4264 * This can (theoretically) happen if the bookmarks have the same object
4265 * and level, but different blkids, if the block sizes are not the same.
4266 * There is presently no way to change the indirect block sizes
4272 * This function checks the following: given that last_block is the place that
4273 * our traversal stopped last time, does that guarantee that we've visited
4274 * every node under subtree_root? Therefore, we can't just use the raw output
4275 * of zbookmark_compare. We have to pass in a modified version of
4276 * subtree_root; by incrementing the block id, and then checking whether
4277 * last_block is before or equal to that, we can tell whether or not having
4278 * visited last_block implies that all of subtree_root's children have been
4282 zbookmark_subtree_completed(const dnode_phys_t *dnp,
4283 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
4285 zbookmark_phys_t mod_zb = *subtree_root;
4287 ASSERT(last_block->zb_level == 0);
4289 /* The objset_phys_t isn't before anything. */
4294 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
4295 * data block size in sectors, because that variable is only used if
4296 * the bookmark refers to a block in the meta-dnode. Since we don't
4297 * know without examining it what object it refers to, and there's no
4298 * harm in passing in this value in other cases, we always pass it in.
4300 * We pass in 0 for the indirect block size shift because zb2 must be
4301 * level 0. The indirect block size is only used to calculate the span
4302 * of the bookmark, but since the bookmark must be level 0, the span is
4303 * always 1, so the math works out.
4305 * If you make changes to how the zbookmark_compare code works, be sure
4306 * to make sure that this code still works afterwards.
4308 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
4309 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,