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, 2017 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
27 #include <sys/sysmacros.h>
28 #include <sys/zfs_context.h>
29 #include <sys/fm/fs/zfs.h>
32 #include <sys/spa_impl.h>
33 #include <sys/vdev_impl.h>
34 #include <sys/zio_impl.h>
35 #include <sys/zio_compress.h>
36 #include <sys/zio_checksum.h>
37 #include <sys/dmu_objset.h>
40 #include <sys/blkptr.h>
41 #include <sys/zfeature.h>
42 #include <sys/dsl_scan.h>
43 #include <sys/metaslab_impl.h>
45 #include <sys/trace_zio.h>
47 #include <sys/dsl_crypt.h>
50 * ==========================================================================
51 * I/O type descriptions
52 * ==========================================================================
54 const char *zio_type_name[ZIO_TYPES] = {
56 * Note: Linux kernel thread name length is limited
57 * so these names will differ from upstream open zfs.
59 "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl"
62 int zio_dva_throttle_enabled = B_TRUE;
65 * ==========================================================================
67 * ==========================================================================
69 kmem_cache_t *zio_cache;
70 kmem_cache_t *zio_link_cache;
71 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
72 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
73 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
74 uint64_t zio_buf_cache_allocs[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
75 uint64_t zio_buf_cache_frees[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
78 int zio_delay_max = ZIO_DELAY_MAX;
80 #define ZIO_PIPELINE_CONTINUE 0x100
81 #define ZIO_PIPELINE_STOP 0x101
83 #define BP_SPANB(indblkshift, level) \
84 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
85 #define COMPARE_META_LEVEL 0x80000000ul
87 * The following actions directly effect the spa's sync-to-convergence logic.
88 * The values below define the sync pass when we start performing the action.
89 * Care should be taken when changing these values as they directly impact
90 * spa_sync() performance. Tuning these values may introduce subtle performance
91 * pathologies and should only be done in the context of performance analysis.
92 * These tunables will eventually be removed and replaced with #defines once
93 * enough analysis has been done to determine optimal values.
95 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
96 * regular blocks are not deferred.
98 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
99 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
100 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
103 * An allocating zio is one that either currently has the DVA allocate
104 * stage set or will have it later in its lifetime.
106 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
108 int zio_requeue_io_start_cut_in_line = 1;
111 int zio_buf_debug_limit = 16384;
113 int zio_buf_debug_limit = 0;
116 static inline void __zio_execute(zio_t *zio);
118 static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t);
124 vmem_t *data_alloc_arena = NULL;
126 zio_cache = kmem_cache_create("zio_cache",
127 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
128 zio_link_cache = kmem_cache_create("zio_link_cache",
129 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
132 * For small buffers, we want a cache for each multiple of
133 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
134 * for each quarter-power of 2.
136 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
137 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
140 size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
142 #if defined(_ILP32) && defined(_KERNEL)
144 * Cache size limited to 1M on 32-bit platforms until ARC
145 * buffers no longer require virtual address space.
147 if (size > zfs_max_recordsize)
156 * If we are using watchpoints, put each buffer on its own page,
157 * to eliminate the performance overhead of trapping to the
158 * kernel when modifying a non-watched buffer that shares the
159 * page with a watched buffer.
161 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
164 * Here's the problem - on 4K native devices in userland on
165 * Linux using O_DIRECT, buffers must be 4K aligned or I/O
166 * will fail with EINVAL, causing zdb (and others) to coredump.
167 * Since userland probably doesn't need optimized buffer caches,
168 * we just force 4K alignment on everything.
170 align = 8 * SPA_MINBLOCKSIZE;
172 if (size < PAGESIZE) {
173 align = SPA_MINBLOCKSIZE;
174 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
181 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
182 zio_buf_cache[c] = kmem_cache_create(name, size,
183 align, NULL, NULL, NULL, NULL, NULL, cflags);
185 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
186 zio_data_buf_cache[c] = kmem_cache_create(name, size,
187 align, NULL, NULL, NULL, NULL,
188 data_alloc_arena, cflags);
193 ASSERT(zio_buf_cache[c] != NULL);
194 if (zio_buf_cache[c - 1] == NULL)
195 zio_buf_cache[c - 1] = zio_buf_cache[c];
197 ASSERT(zio_data_buf_cache[c] != NULL);
198 if (zio_data_buf_cache[c - 1] == NULL)
199 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
211 kmem_cache_t *last_cache = NULL;
212 kmem_cache_t *last_data_cache = NULL;
214 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
217 * Cache size limited to 1M on 32-bit platforms until ARC
218 * buffers no longer require virtual address space.
220 if (((c + 1) << SPA_MINBLOCKSHIFT) > zfs_max_recordsize)
223 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
224 if (zio_buf_cache_allocs[c] != zio_buf_cache_frees[c])
225 (void) printf("zio_fini: [%d] %llu != %llu\n",
226 (int)((c + 1) << SPA_MINBLOCKSHIFT),
227 (long long unsigned)zio_buf_cache_allocs[c],
228 (long long unsigned)zio_buf_cache_frees[c]);
230 if (zio_buf_cache[c] != last_cache) {
231 last_cache = zio_buf_cache[c];
232 kmem_cache_destroy(zio_buf_cache[c]);
234 zio_buf_cache[c] = NULL;
236 if (zio_data_buf_cache[c] != last_data_cache) {
237 last_data_cache = zio_data_buf_cache[c];
238 kmem_cache_destroy(zio_data_buf_cache[c]);
240 zio_data_buf_cache[c] = NULL;
243 kmem_cache_destroy(zio_link_cache);
244 kmem_cache_destroy(zio_cache);
252 * ==========================================================================
253 * Allocate and free I/O buffers
254 * ==========================================================================
258 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
259 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
260 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
261 * excess / transient data in-core during a crashdump.
264 zio_buf_alloc(size_t size)
266 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
268 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
269 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
270 atomic_add_64(&zio_buf_cache_allocs[c], 1);
273 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
277 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
278 * crashdump if the kernel panics. This exists so that we will limit the amount
279 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
280 * of kernel heap dumped to disk when the kernel panics)
283 zio_data_buf_alloc(size_t size)
285 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
287 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
289 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
293 zio_buf_free(void *buf, size_t size)
295 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
297 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
298 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
299 atomic_add_64(&zio_buf_cache_frees[c], 1);
302 kmem_cache_free(zio_buf_cache[c], buf);
306 zio_data_buf_free(void *buf, size_t size)
308 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
310 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
312 kmem_cache_free(zio_data_buf_cache[c], buf);
316 zio_abd_free(void *abd, size_t size)
318 abd_free((abd_t *)abd);
322 * ==========================================================================
323 * Push and pop I/O transform buffers
324 * ==========================================================================
327 zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize,
328 zio_transform_func_t *transform)
330 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
333 * Ensure that anyone expecting this zio to contain a linear ABD isn't
334 * going to get a nasty surprise when they try to access the data.
336 IMPLY(abd_is_linear(zio->io_abd), abd_is_linear(data));
338 zt->zt_orig_abd = zio->io_abd;
339 zt->zt_orig_size = zio->io_size;
340 zt->zt_bufsize = bufsize;
341 zt->zt_transform = transform;
343 zt->zt_next = zio->io_transform_stack;
344 zio->io_transform_stack = zt;
351 zio_pop_transforms(zio_t *zio)
355 while ((zt = zio->io_transform_stack) != NULL) {
356 if (zt->zt_transform != NULL)
357 zt->zt_transform(zio,
358 zt->zt_orig_abd, zt->zt_orig_size);
360 if (zt->zt_bufsize != 0)
361 abd_free(zio->io_abd);
363 zio->io_abd = zt->zt_orig_abd;
364 zio->io_size = zt->zt_orig_size;
365 zio->io_transform_stack = zt->zt_next;
367 kmem_free(zt, sizeof (zio_transform_t));
372 * ==========================================================================
373 * I/O transform callbacks for subblocks, decompression, and decryption
374 * ==========================================================================
377 zio_subblock(zio_t *zio, abd_t *data, uint64_t size)
379 ASSERT(zio->io_size > size);
381 if (zio->io_type == ZIO_TYPE_READ)
382 abd_copy(data, zio->io_abd, size);
386 zio_decompress(zio_t *zio, abd_t *data, uint64_t size)
388 if (zio->io_error == 0) {
389 void *tmp = abd_borrow_buf(data, size);
390 int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
391 zio->io_abd, tmp, zio->io_size, size);
392 abd_return_buf_copy(data, tmp, size);
395 zio->io_error = SET_ERROR(EIO);
400 zio_decrypt(zio_t *zio, abd_t *data, uint64_t size)
404 blkptr_t *bp = zio->io_bp;
405 spa_t *spa = zio->io_spa;
406 uint64_t dsobj = zio->io_bookmark.zb_objset;
407 uint64_t lsize = BP_GET_LSIZE(bp);
408 dmu_object_type_t ot = BP_GET_TYPE(bp);
409 uint8_t salt[ZIO_DATA_SALT_LEN];
410 uint8_t iv[ZIO_DATA_IV_LEN];
411 uint8_t mac[ZIO_DATA_MAC_LEN];
412 boolean_t no_crypt = B_FALSE;
414 ASSERT(BP_USES_CRYPT(bp));
415 ASSERT3U(size, !=, 0);
417 if (zio->io_error != 0)
421 * Verify the cksum of MACs stored in an indirect bp. It will always
422 * be possible to verify this since it does not require an encryption
425 if (BP_HAS_INDIRECT_MAC_CKSUM(bp)) {
426 zio_crypt_decode_mac_bp(bp, mac);
428 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF) {
430 * We haven't decompressed the data yet, but
431 * zio_crypt_do_indirect_mac_checksum() requires
432 * decompressed data to be able to parse out the MACs
433 * from the indirect block. We decompress it now and
434 * throw away the result after we are finished.
436 tmp = zio_buf_alloc(lsize);
437 ret = zio_decompress_data(BP_GET_COMPRESS(bp),
438 zio->io_abd, tmp, zio->io_size, lsize);
440 ret = SET_ERROR(EIO);
443 ret = zio_crypt_do_indirect_mac_checksum(B_FALSE,
444 tmp, lsize, BP_SHOULD_BYTESWAP(bp), mac);
445 zio_buf_free(tmp, lsize);
447 ret = zio_crypt_do_indirect_mac_checksum_abd(B_FALSE,
448 zio->io_abd, size, BP_SHOULD_BYTESWAP(bp), mac);
450 abd_copy(data, zio->io_abd, size);
452 if (zio_injection_enabled && ot != DMU_OT_DNODE && ret == 0) {
453 ret = zio_handle_decrypt_injection(spa,
454 &zio->io_bookmark, ot, ECKSUM);
463 * If this is an authenticated block, just check the MAC. It would be
464 * nice to separate this out into its own flag, but for the moment
465 * enum zio_flag is out of bits.
467 if (BP_IS_AUTHENTICATED(bp)) {
468 if (ot == DMU_OT_OBJSET) {
469 ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa,
470 dsobj, zio->io_abd, size, BP_SHOULD_BYTESWAP(bp));
472 zio_crypt_decode_mac_bp(bp, mac);
473 ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj,
474 zio->io_abd, size, mac);
475 if (zio_injection_enabled && ret == 0) {
476 ret = zio_handle_decrypt_injection(spa,
477 &zio->io_bookmark, ot, ECKSUM);
480 abd_copy(data, zio->io_abd, size);
488 zio_crypt_decode_params_bp(bp, salt, iv);
490 if (ot == DMU_OT_INTENT_LOG) {
491 tmp = abd_borrow_buf_copy(zio->io_abd, sizeof (zil_chain_t));
492 zio_crypt_decode_mac_zil(tmp, mac);
493 abd_return_buf(zio->io_abd, tmp, sizeof (zil_chain_t));
495 zio_crypt_decode_mac_bp(bp, mac);
498 ret = spa_do_crypt_abd(B_FALSE, spa, &zio->io_bookmark, BP_GET_TYPE(bp),
499 BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp), salt, iv, mac, size, data,
500 zio->io_abd, &no_crypt);
502 abd_copy(data, zio->io_abd, size);
510 /* assert that the key was found unless this was speculative */
511 ASSERT(ret != EACCES || (zio->io_flags & ZIO_FLAG_SPECULATIVE));
514 * If there was a decryption / authentication error return EIO as
515 * the io_error. If this was not a speculative zio, create an ereport.
518 zio->io_error = SET_ERROR(EIO);
519 if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) {
520 spa_log_error(spa, &zio->io_bookmark);
521 zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION,
522 spa, NULL, &zio->io_bookmark, zio, 0, 0);
530 * ==========================================================================
531 * I/O parent/child relationships and pipeline interlocks
532 * ==========================================================================
535 zio_walk_parents(zio_t *cio, zio_link_t **zl)
537 list_t *pl = &cio->io_parent_list;
539 *zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl);
543 ASSERT((*zl)->zl_child == cio);
544 return ((*zl)->zl_parent);
548 zio_walk_children(zio_t *pio, zio_link_t **zl)
550 list_t *cl = &pio->io_child_list;
552 ASSERT(MUTEX_HELD(&pio->io_lock));
554 *zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl);
558 ASSERT((*zl)->zl_parent == pio);
559 return ((*zl)->zl_child);
563 zio_unique_parent(zio_t *cio)
565 zio_link_t *zl = NULL;
566 zio_t *pio = zio_walk_parents(cio, &zl);
568 VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL);
573 zio_add_child(zio_t *pio, zio_t *cio)
575 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
578 * Logical I/Os can have logical, gang, or vdev children.
579 * Gang I/Os can have gang or vdev children.
580 * Vdev I/Os can only have vdev children.
581 * The following ASSERT captures all of these constraints.
583 ASSERT3S(cio->io_child_type, <=, pio->io_child_type);
588 mutex_enter(&pio->io_lock);
589 mutex_enter(&cio->io_lock);
591 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
593 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
594 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
596 list_insert_head(&pio->io_child_list, zl);
597 list_insert_head(&cio->io_parent_list, zl);
599 pio->io_child_count++;
600 cio->io_parent_count++;
602 mutex_exit(&cio->io_lock);
603 mutex_exit(&pio->io_lock);
607 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
609 ASSERT(zl->zl_parent == pio);
610 ASSERT(zl->zl_child == cio);
612 mutex_enter(&pio->io_lock);
613 mutex_enter(&cio->io_lock);
615 list_remove(&pio->io_child_list, zl);
616 list_remove(&cio->io_parent_list, zl);
618 pio->io_child_count--;
619 cio->io_parent_count--;
621 mutex_exit(&cio->io_lock);
622 mutex_exit(&pio->io_lock);
623 kmem_cache_free(zio_link_cache, zl);
627 zio_wait_for_children(zio_t *zio, uint8_t childbits, enum zio_wait_type wait)
629 boolean_t waiting = B_FALSE;
631 mutex_enter(&zio->io_lock);
632 ASSERT(zio->io_stall == NULL);
633 for (int c = 0; c < ZIO_CHILD_TYPES; c++) {
634 if (!(ZIO_CHILD_BIT_IS_SET(childbits, c)))
637 uint64_t *countp = &zio->io_children[c][wait];
640 ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN);
641 zio->io_stall = countp;
646 mutex_exit(&zio->io_lock);
650 __attribute__((always_inline))
652 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
654 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
655 int *errorp = &pio->io_child_error[zio->io_child_type];
657 mutex_enter(&pio->io_lock);
658 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
659 *errorp = zio_worst_error(*errorp, zio->io_error);
660 pio->io_reexecute |= zio->io_reexecute;
661 ASSERT3U(*countp, >, 0);
665 if (*countp == 0 && pio->io_stall == countp) {
666 zio_taskq_type_t type =
667 pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE :
669 pio->io_stall = NULL;
670 mutex_exit(&pio->io_lock);
672 * Dispatch the parent zio in its own taskq so that
673 * the child can continue to make progress. This also
674 * prevents overflowing the stack when we have deeply nested
675 * parent-child relationships.
677 zio_taskq_dispatch(pio, type, B_FALSE);
679 mutex_exit(&pio->io_lock);
684 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
686 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
687 zio->io_error = zio->io_child_error[c];
691 zio_bookmark_compare(const void *x1, const void *x2)
693 const zio_t *z1 = x1;
694 const zio_t *z2 = x2;
696 if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset)
698 if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset)
701 if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object)
703 if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object)
706 if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level)
708 if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level)
711 if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid)
713 if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid)
725 * ==========================================================================
726 * Create the various types of I/O (read, write, free, etc)
727 * ==========================================================================
730 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
731 abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done,
732 void *private, zio_type_t type, zio_priority_t priority,
733 enum zio_flag flags, vdev_t *vd, uint64_t offset,
734 const zbookmark_phys_t *zb, enum zio_stage stage,
735 enum zio_stage pipeline)
739 ASSERT3U(psize, <=, SPA_MAXBLOCKSIZE);
740 ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0);
741 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
743 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
744 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
745 ASSERT(vd || stage == ZIO_STAGE_OPEN);
747 IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW_COMPRESS) != 0);
749 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
750 bzero(zio, sizeof (zio_t));
752 mutex_init(&zio->io_lock, NULL, MUTEX_NOLOCKDEP, NULL);
753 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
755 list_create(&zio->io_parent_list, sizeof (zio_link_t),
756 offsetof(zio_link_t, zl_parent_node));
757 list_create(&zio->io_child_list, sizeof (zio_link_t),
758 offsetof(zio_link_t, zl_child_node));
759 metaslab_trace_init(&zio->io_alloc_list);
762 zio->io_child_type = ZIO_CHILD_VDEV;
763 else if (flags & ZIO_FLAG_GANG_CHILD)
764 zio->io_child_type = ZIO_CHILD_GANG;
765 else if (flags & ZIO_FLAG_DDT_CHILD)
766 zio->io_child_type = ZIO_CHILD_DDT;
768 zio->io_child_type = ZIO_CHILD_LOGICAL;
771 zio->io_bp = (blkptr_t *)bp;
772 zio->io_bp_copy = *bp;
773 zio->io_bp_orig = *bp;
774 if (type != ZIO_TYPE_WRITE ||
775 zio->io_child_type == ZIO_CHILD_DDT)
776 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
777 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
778 zio->io_logical = zio;
779 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
780 pipeline |= ZIO_GANG_STAGES;
786 zio->io_private = private;
788 zio->io_priority = priority;
790 zio->io_offset = offset;
791 zio->io_orig_abd = zio->io_abd = data;
792 zio->io_orig_size = zio->io_size = psize;
793 zio->io_lsize = lsize;
794 zio->io_orig_flags = zio->io_flags = flags;
795 zio->io_orig_stage = zio->io_stage = stage;
796 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
797 zio->io_pipeline_trace = ZIO_STAGE_OPEN;
799 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
800 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
803 zio->io_bookmark = *zb;
806 if (zio->io_logical == NULL)
807 zio->io_logical = pio->io_logical;
808 if (zio->io_child_type == ZIO_CHILD_GANG)
809 zio->io_gang_leader = pio->io_gang_leader;
810 zio_add_child(pio, zio);
813 taskq_init_ent(&zio->io_tqent);
819 zio_destroy(zio_t *zio)
821 metaslab_trace_fini(&zio->io_alloc_list);
822 list_destroy(&zio->io_parent_list);
823 list_destroy(&zio->io_child_list);
824 mutex_destroy(&zio->io_lock);
825 cv_destroy(&zio->io_cv);
826 kmem_cache_free(zio_cache, zio);
830 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
831 void *private, enum zio_flag flags)
835 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
836 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
837 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
843 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
845 return (zio_null(NULL, spa, NULL, done, private, flags));
849 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
851 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
852 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
853 bp, (longlong_t)BP_GET_TYPE(bp));
855 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
856 BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
857 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
858 bp, (longlong_t)BP_GET_CHECKSUM(bp));
860 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
861 BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
862 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
863 bp, (longlong_t)BP_GET_COMPRESS(bp));
865 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
866 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
867 bp, (longlong_t)BP_GET_LSIZE(bp));
869 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
870 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
871 bp, (longlong_t)BP_GET_PSIZE(bp));
874 if (BP_IS_EMBEDDED(bp)) {
875 if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
876 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
877 bp, (longlong_t)BPE_GET_ETYPE(bp));
882 * Do not verify individual DVAs if the config is not trusted. This
883 * will be done once the zio is executed in vdev_mirror_map_alloc.
885 if (!spa->spa_trust_config)
889 * Pool-specific checks.
891 * Note: it would be nice to verify that the blk_birth and
892 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
893 * allows the birth time of log blocks (and dmu_sync()-ed blocks
894 * that are in the log) to be arbitrarily large.
896 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
897 uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
899 if (vdevid >= spa->spa_root_vdev->vdev_children) {
900 zfs_panic_recover("blkptr at %p DVA %u has invalid "
902 bp, i, (longlong_t)vdevid);
905 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
907 zfs_panic_recover("blkptr at %p DVA %u has invalid "
909 bp, i, (longlong_t)vdevid);
912 if (vd->vdev_ops == &vdev_hole_ops) {
913 zfs_panic_recover("blkptr at %p DVA %u has hole "
915 bp, i, (longlong_t)vdevid);
918 if (vd->vdev_ops == &vdev_missing_ops) {
920 * "missing" vdevs are valid during import, but we
921 * don't have their detailed info (e.g. asize), so
922 * we can't perform any more checks on them.
926 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
927 uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
929 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
930 if (offset + asize > vd->vdev_asize) {
931 zfs_panic_recover("blkptr at %p DVA %u has invalid "
933 bp, i, (longlong_t)offset);
939 zfs_dva_valid(spa_t *spa, const dva_t *dva, const blkptr_t *bp)
941 uint64_t vdevid = DVA_GET_VDEV(dva);
943 if (vdevid >= spa->spa_root_vdev->vdev_children)
946 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
950 if (vd->vdev_ops == &vdev_hole_ops)
953 if (vd->vdev_ops == &vdev_missing_ops) {
957 uint64_t offset = DVA_GET_OFFSET(dva);
958 uint64_t asize = DVA_GET_ASIZE(dva);
961 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
962 if (offset + asize > vd->vdev_asize)
969 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
970 abd_t *data, uint64_t size, zio_done_func_t *done, void *private,
971 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
975 zfs_blkptr_verify(spa, bp);
977 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
978 data, size, size, done, private,
979 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
980 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
981 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
987 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
988 abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp,
989 zio_done_func_t *ready, zio_done_func_t *children_ready,
990 zio_done_func_t *physdone, zio_done_func_t *done,
991 void *private, zio_priority_t priority, enum zio_flag flags,
992 const zbookmark_phys_t *zb)
996 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
997 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
998 zp->zp_compress >= ZIO_COMPRESS_OFF &&
999 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
1000 DMU_OT_IS_VALID(zp->zp_type) &&
1001 zp->zp_level < 32 &&
1002 zp->zp_copies > 0 &&
1003 zp->zp_copies <= spa_max_replication(spa));
1005 zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private,
1006 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
1007 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
1008 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
1010 zio->io_ready = ready;
1011 zio->io_children_ready = children_ready;
1012 zio->io_physdone = physdone;
1016 * Data can be NULL if we are going to call zio_write_override() to
1017 * provide the already-allocated BP. But we may need the data to
1018 * verify a dedup hit (if requested). In this case, don't try to
1019 * dedup (just take the already-allocated BP verbatim). Encrypted
1020 * dedup blocks need data as well so we also disable dedup in this
1024 (zio->io_prop.zp_dedup_verify || zio->io_prop.zp_encrypt)) {
1025 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
1032 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data,
1033 uint64_t size, zio_done_func_t *done, void *private,
1034 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
1038 zio = zio_create(pio, spa, txg, bp, data, size, size, done, private,
1039 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb,
1040 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
1046 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
1048 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
1049 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1050 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1051 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
1054 * We must reset the io_prop to match the values that existed
1055 * when the bp was first written by dmu_sync() keeping in mind
1056 * that nopwrite and dedup are mutually exclusive.
1058 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
1059 zio->io_prop.zp_nopwrite = nopwrite;
1060 zio->io_prop.zp_copies = copies;
1061 zio->io_bp_override = bp;
1065 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
1068 zfs_blkptr_verify(spa, bp);
1071 * The check for EMBEDDED is a performance optimization. We
1072 * process the free here (by ignoring it) rather than
1073 * putting it on the list and then processing it in zio_free_sync().
1075 if (BP_IS_EMBEDDED(bp))
1077 metaslab_check_free(spa, bp);
1080 * Frees that are for the currently-syncing txg, are not going to be
1081 * deferred, and which will not need to do a read (i.e. not GANG or
1082 * DEDUP), can be processed immediately. Otherwise, put them on the
1083 * in-memory list for later processing.
1085 if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
1086 txg != spa->spa_syncing_txg ||
1087 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
1088 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
1090 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 0)));
1095 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1096 enum zio_flag flags)
1099 enum zio_stage stage = ZIO_FREE_PIPELINE;
1101 ASSERT(!BP_IS_HOLE(bp));
1102 ASSERT(spa_syncing_txg(spa) == txg);
1103 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
1105 if (BP_IS_EMBEDDED(bp))
1106 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
1108 metaslab_check_free(spa, bp);
1110 dsl_scan_freed(spa, bp);
1113 * GANG and DEDUP blocks can induce a read (for the gang block header,
1114 * or the DDT), so issue them asynchronously so that this thread is
1117 if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
1118 stage |= ZIO_STAGE_ISSUE_ASYNC;
1120 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1121 BP_GET_PSIZE(bp), NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW,
1122 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
1128 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1129 zio_done_func_t *done, void *private, enum zio_flag flags)
1133 zfs_blkptr_verify(spa, bp);
1135 if (BP_IS_EMBEDDED(bp))
1136 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
1139 * A claim is an allocation of a specific block. Claims are needed
1140 * to support immediate writes in the intent log. The issue is that
1141 * immediate writes contain committed data, but in a txg that was
1142 * *not* committed. Upon opening the pool after an unclean shutdown,
1143 * the intent log claims all blocks that contain immediate write data
1144 * so that the SPA knows they're in use.
1146 * All claims *must* be resolved in the first txg -- before the SPA
1147 * starts allocating blocks -- so that nothing is allocated twice.
1148 * If txg == 0 we just verify that the block is claimable.
1150 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <,
1151 spa_min_claim_txg(spa));
1152 ASSERT(txg == spa_min_claim_txg(spa) || txg == 0);
1153 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
1155 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1156 BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW,
1157 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
1158 ASSERT0(zio->io_queued_timestamp);
1164 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
1165 zio_done_func_t *done, void *private, enum zio_flag flags)
1170 if (vd->vdev_children == 0) {
1171 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
1172 ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
1173 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
1177 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
1179 for (c = 0; c < vd->vdev_children; c++)
1180 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
1181 done, private, flags));
1188 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1189 abd_t *data, int checksum, zio_done_func_t *done, void *private,
1190 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1194 ASSERT(vd->vdev_children == 0);
1195 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1196 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1197 ASSERT3U(offset + size, <=, vd->vdev_psize);
1199 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1200 private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1201 offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
1203 zio->io_prop.zp_checksum = checksum;
1209 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1210 abd_t *data, int checksum, zio_done_func_t *done, void *private,
1211 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1215 ASSERT(vd->vdev_children == 0);
1216 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1217 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1218 ASSERT3U(offset + size, <=, vd->vdev_psize);
1220 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1221 private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1222 offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
1224 zio->io_prop.zp_checksum = checksum;
1226 if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
1228 * zec checksums are necessarily destructive -- they modify
1229 * the end of the write buffer to hold the verifier/checksum.
1230 * Therefore, we must make a local copy in case the data is
1231 * being written to multiple places in parallel.
1233 abd_t *wbuf = abd_alloc_sametype(data, size);
1234 abd_copy(wbuf, data, size);
1236 zio_push_transform(zio, wbuf, size, size, NULL);
1243 * Create a child I/O to do some work for us.
1246 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1247 abd_t *data, uint64_t size, int type, zio_priority_t priority,
1248 enum zio_flag flags, zio_done_func_t *done, void *private)
1250 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1254 * vdev child I/Os do not propagate their error to the parent.
1255 * Therefore, for correct operation the caller *must* check for
1256 * and handle the error in the child i/o's done callback.
1257 * The only exceptions are i/os that we don't care about
1258 * (OPTIONAL or REPAIR).
1260 ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) ||
1263 if (type == ZIO_TYPE_READ && bp != NULL) {
1265 * If we have the bp, then the child should perform the
1266 * checksum and the parent need not. This pushes error
1267 * detection as close to the leaves as possible and
1268 * eliminates redundant checksums in the interior nodes.
1270 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1271 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1274 if (vd->vdev_ops->vdev_op_leaf) {
1275 ASSERT0(vd->vdev_children);
1276 offset += VDEV_LABEL_START_SIZE;
1279 flags |= ZIO_VDEV_CHILD_FLAGS(pio);
1282 * If we've decided to do a repair, the write is not speculative --
1283 * even if the original read was.
1285 if (flags & ZIO_FLAG_IO_REPAIR)
1286 flags &= ~ZIO_FLAG_SPECULATIVE;
1289 * If we're creating a child I/O that is not associated with a
1290 * top-level vdev, then the child zio is not an allocating I/O.
1291 * If this is a retried I/O then we ignore it since we will
1292 * have already processed the original allocating I/O.
1294 if (flags & ZIO_FLAG_IO_ALLOCATING &&
1295 (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) {
1296 ASSERTV(metaslab_class_t *mc = spa_normal_class(pio->io_spa));
1298 ASSERT(mc->mc_alloc_throttle_enabled);
1299 ASSERT(type == ZIO_TYPE_WRITE);
1300 ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE);
1301 ASSERT(!(flags & ZIO_FLAG_IO_REPAIR));
1302 ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) ||
1303 pio->io_child_type == ZIO_CHILD_GANG);
1305 flags &= ~ZIO_FLAG_IO_ALLOCATING;
1309 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size,
1310 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1311 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1312 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
1314 zio->io_physdone = pio->io_physdone;
1315 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1316 zio->io_logical->io_phys_children++;
1322 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size,
1323 zio_type_t type, zio_priority_t priority, enum zio_flag flags,
1324 zio_done_func_t *done, void *private)
1328 ASSERT(vd->vdev_ops->vdev_op_leaf);
1330 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1331 data, size, size, done, private, type, priority,
1332 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1334 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1340 zio_flush(zio_t *zio, vdev_t *vd)
1342 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
1344 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1348 zio_shrink(zio_t *zio, uint64_t size)
1350 ASSERT3P(zio->io_executor, ==, NULL);
1351 ASSERT3U(zio->io_orig_size, ==, zio->io_size);
1352 ASSERT3U(size, <=, zio->io_size);
1355 * We don't shrink for raidz because of problems with the
1356 * reconstruction when reading back less than the block size.
1357 * Note, BP_IS_RAIDZ() assumes no compression.
1359 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1360 if (!BP_IS_RAIDZ(zio->io_bp)) {
1361 /* we are not doing a raw write */
1362 ASSERT3U(zio->io_size, ==, zio->io_lsize);
1363 zio->io_orig_size = zio->io_size = zio->io_lsize = size;
1368 * ==========================================================================
1369 * Prepare to read and write logical blocks
1370 * ==========================================================================
1374 zio_read_bp_init(zio_t *zio)
1376 blkptr_t *bp = zio->io_bp;
1378 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1380 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1382 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1383 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1384 !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) {
1385 zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1386 psize, psize, zio_decompress);
1389 if (((BP_IS_PROTECTED(bp) && !(zio->io_flags & ZIO_FLAG_RAW_ENCRYPT)) ||
1390 BP_HAS_INDIRECT_MAC_CKSUM(bp)) &&
1391 zio->io_child_type == ZIO_CHILD_LOGICAL) {
1392 zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1393 psize, psize, zio_decrypt);
1396 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1397 int psize = BPE_GET_PSIZE(bp);
1398 void *data = abd_borrow_buf(zio->io_abd, psize);
1400 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1401 decode_embedded_bp_compressed(bp, data);
1402 abd_return_buf_copy(zio->io_abd, data, psize);
1404 ASSERT(!BP_IS_EMBEDDED(bp));
1405 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1408 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1409 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1411 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1412 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1414 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1415 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1417 return (ZIO_PIPELINE_CONTINUE);
1421 zio_write_bp_init(zio_t *zio)
1423 if (!IO_IS_ALLOCATING(zio))
1424 return (ZIO_PIPELINE_CONTINUE);
1426 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1428 if (zio->io_bp_override) {
1429 blkptr_t *bp = zio->io_bp;
1430 zio_prop_t *zp = &zio->io_prop;
1432 ASSERT(bp->blk_birth != zio->io_txg);
1433 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1435 *bp = *zio->io_bp_override;
1436 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1438 if (BP_IS_EMBEDDED(bp))
1439 return (ZIO_PIPELINE_CONTINUE);
1442 * If we've been overridden and nopwrite is set then
1443 * set the flag accordingly to indicate that a nopwrite
1444 * has already occurred.
1446 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1447 ASSERT(!zp->zp_dedup);
1448 ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum);
1449 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1450 return (ZIO_PIPELINE_CONTINUE);
1453 ASSERT(!zp->zp_nopwrite);
1455 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1456 return (ZIO_PIPELINE_CONTINUE);
1458 ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags &
1459 ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify);
1461 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum &&
1463 BP_SET_DEDUP(bp, 1);
1464 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1465 return (ZIO_PIPELINE_CONTINUE);
1469 * We were unable to handle this as an override bp, treat
1470 * it as a regular write I/O.
1472 zio->io_bp_override = NULL;
1473 *bp = zio->io_bp_orig;
1474 zio->io_pipeline = zio->io_orig_pipeline;
1477 return (ZIO_PIPELINE_CONTINUE);
1481 zio_write_compress(zio_t *zio)
1483 spa_t *spa = zio->io_spa;
1484 zio_prop_t *zp = &zio->io_prop;
1485 enum zio_compress compress = zp->zp_compress;
1486 blkptr_t *bp = zio->io_bp;
1487 uint64_t lsize = zio->io_lsize;
1488 uint64_t psize = zio->io_size;
1492 * If our children haven't all reached the ready stage,
1493 * wait for them and then repeat this pipeline stage.
1495 if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT |
1496 ZIO_CHILD_GANG_BIT, ZIO_WAIT_READY)) {
1497 return (ZIO_PIPELINE_STOP);
1500 if (!IO_IS_ALLOCATING(zio))
1501 return (ZIO_PIPELINE_CONTINUE);
1503 if (zio->io_children_ready != NULL) {
1505 * Now that all our children are ready, run the callback
1506 * associated with this zio in case it wants to modify the
1507 * data to be written.
1509 ASSERT3U(zp->zp_level, >, 0);
1510 zio->io_children_ready(zio);
1513 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1514 ASSERT(zio->io_bp_override == NULL);
1516 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1518 * We're rewriting an existing block, which means we're
1519 * working on behalf of spa_sync(). For spa_sync() to
1520 * converge, it must eventually be the case that we don't
1521 * have to allocate new blocks. But compression changes
1522 * the blocksize, which forces a reallocate, and makes
1523 * convergence take longer. Therefore, after the first
1524 * few passes, stop compressing to ensure convergence.
1526 pass = spa_sync_pass(spa);
1528 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1529 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1530 ASSERT(!BP_GET_DEDUP(bp));
1532 if (pass >= zfs_sync_pass_dont_compress)
1533 compress = ZIO_COMPRESS_OFF;
1535 /* Make sure someone doesn't change their mind on overwrites */
1536 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1537 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1540 /* If it's a compressed write that is not raw, compress the buffer. */
1541 if (compress != ZIO_COMPRESS_OFF &&
1542 !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) {
1543 void *cbuf = zio_buf_alloc(lsize);
1544 psize = zio_compress_data(compress, zio->io_abd, cbuf, lsize);
1545 if (psize == 0 || psize == lsize) {
1546 compress = ZIO_COMPRESS_OFF;
1547 zio_buf_free(cbuf, lsize);
1548 } else if (!zp->zp_dedup && !zp->zp_encrypt &&
1549 psize <= BPE_PAYLOAD_SIZE &&
1550 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1551 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1552 encode_embedded_bp_compressed(bp,
1553 cbuf, compress, lsize, psize);
1554 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1555 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1556 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1557 zio_buf_free(cbuf, lsize);
1558 bp->blk_birth = zio->io_txg;
1559 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1560 ASSERT(spa_feature_is_active(spa,
1561 SPA_FEATURE_EMBEDDED_DATA));
1562 return (ZIO_PIPELINE_CONTINUE);
1565 * Round up compressed size up to the ashift
1566 * of the smallest-ashift device, and zero the tail.
1567 * This ensures that the compressed size of the BP
1568 * (and thus compressratio property) are correct,
1569 * in that we charge for the padding used to fill out
1572 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
1573 size_t rounded = (size_t)P2ROUNDUP(psize,
1574 1ULL << spa->spa_min_ashift);
1575 if (rounded >= lsize) {
1576 compress = ZIO_COMPRESS_OFF;
1577 zio_buf_free(cbuf, lsize);
1580 abd_t *cdata = abd_get_from_buf(cbuf, lsize);
1581 abd_take_ownership_of_buf(cdata, B_TRUE);
1582 abd_zero_off(cdata, psize, rounded - psize);
1584 zio_push_transform(zio, cdata,
1585 psize, lsize, NULL);
1590 * We were unable to handle this as an override bp, treat
1591 * it as a regular write I/O.
1593 zio->io_bp_override = NULL;
1594 *bp = zio->io_bp_orig;
1595 zio->io_pipeline = zio->io_orig_pipeline;
1597 } else if ((zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) != 0 &&
1598 zp->zp_type == DMU_OT_DNODE) {
1600 * The DMU actually relies on the zio layer's compression
1601 * to free metadnode blocks that have had all contained
1602 * dnodes freed. As a result, even when doing a raw
1603 * receive, we must check whether the block can be compressed
1606 psize = zio_compress_data(ZIO_COMPRESS_EMPTY,
1607 zio->io_abd, NULL, lsize);
1609 compress = ZIO_COMPRESS_OFF;
1611 ASSERT3U(psize, !=, 0);
1615 * The final pass of spa_sync() must be all rewrites, but the first
1616 * few passes offer a trade-off: allocating blocks defers convergence,
1617 * but newly allocated blocks are sequential, so they can be written
1618 * to disk faster. Therefore, we allow the first few passes of
1619 * spa_sync() to allocate new blocks, but force rewrites after that.
1620 * There should only be a handful of blocks after pass 1 in any case.
1622 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1623 BP_GET_PSIZE(bp) == psize &&
1624 pass >= zfs_sync_pass_rewrite) {
1626 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1627 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1628 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1631 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1635 if (zio->io_bp_orig.blk_birth != 0 &&
1636 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1637 BP_SET_LSIZE(bp, lsize);
1638 BP_SET_TYPE(bp, zp->zp_type);
1639 BP_SET_LEVEL(bp, zp->zp_level);
1640 BP_SET_BIRTH(bp, zio->io_txg, 0);
1642 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1644 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1645 BP_SET_LSIZE(bp, lsize);
1646 BP_SET_TYPE(bp, zp->zp_type);
1647 BP_SET_LEVEL(bp, zp->zp_level);
1648 BP_SET_PSIZE(bp, psize);
1649 BP_SET_COMPRESS(bp, compress);
1650 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1651 BP_SET_DEDUP(bp, zp->zp_dedup);
1652 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1654 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1655 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1656 ASSERT(!zp->zp_encrypt ||
1657 DMU_OT_IS_ENCRYPTED(zp->zp_type));
1658 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1660 if (zp->zp_nopwrite) {
1661 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1662 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1663 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1666 return (ZIO_PIPELINE_CONTINUE);
1670 zio_free_bp_init(zio_t *zio)
1672 blkptr_t *bp = zio->io_bp;
1674 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1675 if (BP_GET_DEDUP(bp))
1676 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1679 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1681 return (ZIO_PIPELINE_CONTINUE);
1685 * ==========================================================================
1686 * Execute the I/O pipeline
1687 * ==========================================================================
1691 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1693 spa_t *spa = zio->io_spa;
1694 zio_type_t t = zio->io_type;
1695 int flags = (cutinline ? TQ_FRONT : 0);
1698 * If we're a config writer or a probe, the normal issue and
1699 * interrupt threads may all be blocked waiting for the config lock.
1700 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1702 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1706 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1708 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1712 * If this is a high priority I/O, then use the high priority taskq if
1715 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1716 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1719 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1722 * NB: We are assuming that the zio can only be dispatched
1723 * to a single taskq at a time. It would be a grievous error
1724 * to dispatch the zio to another taskq at the same time.
1726 ASSERT(taskq_empty_ent(&zio->io_tqent));
1727 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1728 flags, &zio->io_tqent);
1732 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1734 kthread_t *executor = zio->io_executor;
1735 spa_t *spa = zio->io_spa;
1737 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1738 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1740 for (i = 0; i < tqs->stqs_count; i++) {
1741 if (taskq_member(tqs->stqs_taskq[i], executor))
1750 zio_issue_async(zio_t *zio)
1752 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1754 return (ZIO_PIPELINE_STOP);
1758 zio_interrupt(zio_t *zio)
1760 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1764 zio_delay_interrupt(zio_t *zio)
1767 * The timeout_generic() function isn't defined in userspace, so
1768 * rather than trying to implement the function, the zio delay
1769 * functionality has been disabled for userspace builds.
1774 * If io_target_timestamp is zero, then no delay has been registered
1775 * for this IO, thus jump to the end of this function and "skip" the
1776 * delay; issuing it directly to the zio layer.
1778 if (zio->io_target_timestamp != 0) {
1779 hrtime_t now = gethrtime();
1781 if (now >= zio->io_target_timestamp) {
1783 * This IO has already taken longer than the target
1784 * delay to complete, so we don't want to delay it
1785 * any longer; we "miss" the delay and issue it
1786 * directly to the zio layer. This is likely due to
1787 * the target latency being set to a value less than
1788 * the underlying hardware can satisfy (e.g. delay
1789 * set to 1ms, but the disks take 10ms to complete an
1793 DTRACE_PROBE2(zio__delay__miss, zio_t *, zio,
1799 hrtime_t diff = zio->io_target_timestamp - now;
1800 clock_t expire_at_tick = ddi_get_lbolt() +
1803 DTRACE_PROBE3(zio__delay__hit, zio_t *, zio,
1804 hrtime_t, now, hrtime_t, diff);
1806 if (NSEC_TO_TICK(diff) == 0) {
1807 /* Our delay is less than a jiffy - just spin */
1808 zfs_sleep_until(zio->io_target_timestamp);
1811 * Use taskq_dispatch_delay() in the place of
1812 * OpenZFS's timeout_generic().
1814 tid = taskq_dispatch_delay(system_taskq,
1815 (task_func_t *)zio_interrupt,
1816 zio, TQ_NOSLEEP, expire_at_tick);
1817 if (tid == TASKQID_INVALID) {
1819 * Couldn't allocate a task. Just
1820 * finish the zio without a delay.
1829 DTRACE_PROBE1(zio__delay__skip, zio_t *, zio);
1834 zio_deadman_impl(zio_t *pio)
1836 zio_t *cio, *cio_next;
1837 zio_link_t *zl = NULL;
1838 vdev_t *vd = pio->io_vd;
1840 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
1841 vdev_queue_t *vq = &vd->vdev_queue;
1842 zbookmark_phys_t *zb = &pio->io_bookmark;
1843 uint64_t delta = gethrtime() - pio->io_timestamp;
1844 uint64_t failmode = spa_get_deadman_failmode(pio->io_spa);
1846 zfs_dbgmsg("slow zio: zio=%p timestamp=%llu "
1847 "delta=%llu queued=%llu io=%llu "
1848 "path=%s last=%llu "
1849 "type=%d priority=%d flags=0x%x "
1850 "stage=0x%x pipeline=0x%x pipeline-trace=0x%x "
1851 "objset=%llu object=%llu level=%llu blkid=%llu "
1852 "offset=%llu size=%llu error=%d",
1853 pio, pio->io_timestamp,
1854 delta, pio->io_delta, pio->io_delay,
1855 vd->vdev_path, vq->vq_io_complete_ts,
1856 pio->io_type, pio->io_priority, pio->io_flags,
1857 pio->io_state, pio->io_pipeline, pio->io_pipeline_trace,
1858 zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
1859 pio->io_offset, pio->io_size, pio->io_error);
1860 zfs_ereport_post(FM_EREPORT_ZFS_DEADMAN,
1861 pio->io_spa, vd, zb, pio, 0, 0);
1863 if (failmode == ZIO_FAILURE_MODE_CONTINUE &&
1864 taskq_empty_ent(&pio->io_tqent)) {
1869 mutex_enter(&pio->io_lock);
1870 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
1871 cio_next = zio_walk_children(pio, &zl);
1872 zio_deadman_impl(cio);
1874 mutex_exit(&pio->io_lock);
1878 * Log the critical information describing this zio and all of its children
1879 * using the zfs_dbgmsg() interface then post deadman event for the ZED.
1882 zio_deadman(zio_t *pio, char *tag)
1884 spa_t *spa = pio->io_spa;
1885 char *name = spa_name(spa);
1887 if (!zfs_deadman_enabled || spa_suspended(spa))
1890 zio_deadman_impl(pio);
1892 switch (spa_get_deadman_failmode(spa)) {
1893 case ZIO_FAILURE_MODE_WAIT:
1894 zfs_dbgmsg("%s waiting for hung I/O to pool '%s'", tag, name);
1897 case ZIO_FAILURE_MODE_CONTINUE:
1898 zfs_dbgmsg("%s restarting hung I/O for pool '%s'", tag, name);
1901 case ZIO_FAILURE_MODE_PANIC:
1902 fm_panic("%s determined I/O to pool '%s' is hung.", tag, name);
1908 * Execute the I/O pipeline until one of the following occurs:
1909 * (1) the I/O completes; (2) the pipeline stalls waiting for
1910 * dependent child I/Os; (3) the I/O issues, so we're waiting
1911 * for an I/O completion interrupt; (4) the I/O is delegated by
1912 * vdev-level caching or aggregation; (5) the I/O is deferred
1913 * due to vdev-level queueing; (6) the I/O is handed off to
1914 * another thread. In all cases, the pipeline stops whenever
1915 * there's no CPU work; it never burns a thread in cv_wait_io().
1917 * There's no locking on io_stage because there's no legitimate way
1918 * for multiple threads to be attempting to process the same I/O.
1920 static zio_pipe_stage_t *zio_pipeline[];
1923 * zio_execute() is a wrapper around the static function
1924 * __zio_execute() so that we can force __zio_execute() to be
1925 * inlined. This reduces stack overhead which is important
1926 * because __zio_execute() is called recursively in several zio
1927 * code paths. zio_execute() itself cannot be inlined because
1928 * it is externally visible.
1931 zio_execute(zio_t *zio)
1933 fstrans_cookie_t cookie;
1935 cookie = spl_fstrans_mark();
1937 spl_fstrans_unmark(cookie);
1941 * Used to determine if in the current context the stack is sized large
1942 * enough to allow zio_execute() to be called recursively. A minimum
1943 * stack size of 16K is required to avoid needing to re-dispatch the zio.
1946 zio_execute_stack_check(zio_t *zio)
1948 #if !defined(HAVE_LARGE_STACKS)
1949 dsl_pool_t *dp = spa_get_dsl(zio->io_spa);
1951 /* Executing in txg_sync_thread() context. */
1952 if (dp && curthread == dp->dp_tx.tx_sync_thread)
1955 /* Pool initialization outside of zio_taskq context. */
1956 if (dp && spa_is_initializing(dp->dp_spa) &&
1957 !zio_taskq_member(zio, ZIO_TASKQ_ISSUE) &&
1958 !zio_taskq_member(zio, ZIO_TASKQ_ISSUE_HIGH))
1960 #endif /* HAVE_LARGE_STACKS */
1965 __attribute__((always_inline))
1967 __zio_execute(zio_t *zio)
1969 zio->io_executor = curthread;
1971 ASSERT3U(zio->io_queued_timestamp, >, 0);
1973 while (zio->io_stage < ZIO_STAGE_DONE) {
1974 enum zio_stage pipeline = zio->io_pipeline;
1975 enum zio_stage stage = zio->io_stage;
1978 ASSERT(!MUTEX_HELD(&zio->io_lock));
1979 ASSERT(ISP2(stage));
1980 ASSERT(zio->io_stall == NULL);
1984 } while ((stage & pipeline) == 0);
1986 ASSERT(stage <= ZIO_STAGE_DONE);
1989 * If we are in interrupt context and this pipeline stage
1990 * will grab a config lock that is held across I/O,
1991 * or may wait for an I/O that needs an interrupt thread
1992 * to complete, issue async to avoid deadlock.
1994 * For VDEV_IO_START, we cut in line so that the io will
1995 * be sent to disk promptly.
1997 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
1998 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
1999 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
2000 zio_requeue_io_start_cut_in_line : B_FALSE;
2001 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
2006 * If the current context doesn't have large enough stacks
2007 * the zio must be issued asynchronously to prevent overflow.
2009 if (zio_execute_stack_check(zio)) {
2010 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
2011 zio_requeue_io_start_cut_in_line : B_FALSE;
2012 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
2016 zio->io_stage = stage;
2017 zio->io_pipeline_trace |= zio->io_stage;
2018 rv = zio_pipeline[highbit64(stage) - 1](zio);
2020 if (rv == ZIO_PIPELINE_STOP)
2023 ASSERT(rv == ZIO_PIPELINE_CONTINUE);
2029 * ==========================================================================
2030 * Initiate I/O, either sync or async
2031 * ==========================================================================
2034 zio_wait(zio_t *zio)
2036 long timeout = MSEC_TO_TICK(zfs_deadman_ziotime_ms);
2039 ASSERT3S(zio->io_stage, ==, ZIO_STAGE_OPEN);
2040 ASSERT3P(zio->io_executor, ==, NULL);
2042 zio->io_waiter = curthread;
2043 ASSERT0(zio->io_queued_timestamp);
2044 zio->io_queued_timestamp = gethrtime();
2048 mutex_enter(&zio->io_lock);
2049 while (zio->io_executor != NULL) {
2050 error = cv_timedwait_io(&zio->io_cv, &zio->io_lock,
2051 ddi_get_lbolt() + timeout);
2053 if (zfs_deadman_enabled && error == -1 &&
2054 gethrtime() - zio->io_queued_timestamp >
2055 spa_deadman_ziotime(zio->io_spa)) {
2056 mutex_exit(&zio->io_lock);
2057 timeout = MSEC_TO_TICK(zfs_deadman_checktime_ms);
2058 zio_deadman(zio, FTAG);
2059 mutex_enter(&zio->io_lock);
2062 mutex_exit(&zio->io_lock);
2064 error = zio->io_error;
2071 zio_nowait(zio_t *zio)
2073 ASSERT3P(zio->io_executor, ==, NULL);
2075 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
2076 zio_unique_parent(zio) == NULL) {
2080 * This is a logical async I/O with no parent to wait for it.
2081 * We add it to the spa_async_root_zio "Godfather" I/O which
2082 * will ensure they complete prior to unloading the pool.
2084 spa_t *spa = zio->io_spa;
2086 pio = spa->spa_async_zio_root[CPU_SEQID];
2089 zio_add_child(pio, zio);
2092 ASSERT0(zio->io_queued_timestamp);
2093 zio->io_queued_timestamp = gethrtime();
2098 * ==========================================================================
2099 * Reexecute, cancel, or suspend/resume failed I/O
2100 * ==========================================================================
2104 zio_reexecute(zio_t *pio)
2106 zio_t *cio, *cio_next;
2108 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
2109 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
2110 ASSERT(pio->io_gang_leader == NULL);
2111 ASSERT(pio->io_gang_tree == NULL);
2113 pio->io_flags = pio->io_orig_flags;
2114 pio->io_stage = pio->io_orig_stage;
2115 pio->io_pipeline = pio->io_orig_pipeline;
2116 pio->io_reexecute = 0;
2117 pio->io_flags |= ZIO_FLAG_REEXECUTED;
2118 pio->io_pipeline_trace = 0;
2120 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2121 pio->io_state[w] = 0;
2122 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2123 pio->io_child_error[c] = 0;
2125 if (IO_IS_ALLOCATING(pio))
2126 BP_ZERO(pio->io_bp);
2129 * As we reexecute pio's children, new children could be created.
2130 * New children go to the head of pio's io_child_list, however,
2131 * so we will (correctly) not reexecute them. The key is that
2132 * the remainder of pio's io_child_list, from 'cio_next' onward,
2133 * cannot be affected by any side effects of reexecuting 'cio'.
2135 zio_link_t *zl = NULL;
2136 mutex_enter(&pio->io_lock);
2137 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
2138 cio_next = zio_walk_children(pio, &zl);
2139 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2140 pio->io_children[cio->io_child_type][w]++;
2141 mutex_exit(&pio->io_lock);
2143 mutex_enter(&pio->io_lock);
2145 mutex_exit(&pio->io_lock);
2148 * Now that all children have been reexecuted, execute the parent.
2149 * We don't reexecute "The Godfather" I/O here as it's the
2150 * responsibility of the caller to wait on it.
2152 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) {
2153 pio->io_queued_timestamp = gethrtime();
2159 zio_suspend(spa_t *spa, zio_t *zio, zio_suspend_reason_t reason)
2161 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
2162 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
2163 "failure and the failure mode property for this pool "
2164 "is set to panic.", spa_name(spa));
2166 cmn_err(CE_WARN, "Pool '%s' has encountered an uncorrectable I/O "
2167 "failure and has been suspended.\n", spa_name(spa));
2169 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL,
2172 mutex_enter(&spa->spa_suspend_lock);
2174 if (spa->spa_suspend_zio_root == NULL)
2175 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
2176 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
2177 ZIO_FLAG_GODFATHER);
2179 spa->spa_suspended = reason;
2182 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2183 ASSERT(zio != spa->spa_suspend_zio_root);
2184 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2185 ASSERT(zio_unique_parent(zio) == NULL);
2186 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
2187 zio_add_child(spa->spa_suspend_zio_root, zio);
2190 mutex_exit(&spa->spa_suspend_lock);
2194 zio_resume(spa_t *spa)
2199 * Reexecute all previously suspended i/o.
2201 mutex_enter(&spa->spa_suspend_lock);
2202 spa->spa_suspended = ZIO_SUSPEND_NONE;
2203 cv_broadcast(&spa->spa_suspend_cv);
2204 pio = spa->spa_suspend_zio_root;
2205 spa->spa_suspend_zio_root = NULL;
2206 mutex_exit(&spa->spa_suspend_lock);
2212 return (zio_wait(pio));
2216 zio_resume_wait(spa_t *spa)
2218 mutex_enter(&spa->spa_suspend_lock);
2219 while (spa_suspended(spa))
2220 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
2221 mutex_exit(&spa->spa_suspend_lock);
2225 * ==========================================================================
2228 * A gang block is a collection of small blocks that looks to the DMU
2229 * like one large block. When zio_dva_allocate() cannot find a block
2230 * of the requested size, due to either severe fragmentation or the pool
2231 * being nearly full, it calls zio_write_gang_block() to construct the
2232 * block from smaller fragments.
2234 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
2235 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
2236 * an indirect block: it's an array of block pointers. It consumes
2237 * only one sector and hence is allocatable regardless of fragmentation.
2238 * The gang header's bps point to its gang members, which hold the data.
2240 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
2241 * as the verifier to ensure uniqueness of the SHA256 checksum.
2242 * Critically, the gang block bp's blk_cksum is the checksum of the data,
2243 * not the gang header. This ensures that data block signatures (needed for
2244 * deduplication) are independent of how the block is physically stored.
2246 * Gang blocks can be nested: a gang member may itself be a gang block.
2247 * Thus every gang block is a tree in which root and all interior nodes are
2248 * gang headers, and the leaves are normal blocks that contain user data.
2249 * The root of the gang tree is called the gang leader.
2251 * To perform any operation (read, rewrite, free, claim) on a gang block,
2252 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
2253 * in the io_gang_tree field of the original logical i/o by recursively
2254 * reading the gang leader and all gang headers below it. This yields
2255 * an in-core tree containing the contents of every gang header and the
2256 * bps for every constituent of the gang block.
2258 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
2259 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
2260 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
2261 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
2262 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
2263 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
2264 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
2265 * of the gang header plus zio_checksum_compute() of the data to update the
2266 * gang header's blk_cksum as described above.
2268 * The two-phase assemble/issue model solves the problem of partial failure --
2269 * what if you'd freed part of a gang block but then couldn't read the
2270 * gang header for another part? Assembling the entire gang tree first
2271 * ensures that all the necessary gang header I/O has succeeded before
2272 * starting the actual work of free, claim, or write. Once the gang tree
2273 * is assembled, free and claim are in-memory operations that cannot fail.
2275 * In the event that a gang write fails, zio_dva_unallocate() walks the
2276 * gang tree to immediately free (i.e. insert back into the space map)
2277 * everything we've allocated. This ensures that we don't get ENOSPC
2278 * errors during repeated suspend/resume cycles due to a flaky device.
2280 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
2281 * the gang tree, we won't modify the block, so we can safely defer the free
2282 * (knowing that the block is still intact). If we *can* assemble the gang
2283 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
2284 * each constituent bp and we can allocate a new block on the next sync pass.
2286 * In all cases, the gang tree allows complete recovery from partial failure.
2287 * ==========================================================================
2291 zio_gang_issue_func_done(zio_t *zio)
2293 abd_put(zio->io_abd);
2297 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2303 return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset),
2304 BP_GET_PSIZE(bp), zio_gang_issue_func_done,
2305 NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2306 &pio->io_bookmark));
2310 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2317 abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2318 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2319 gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL,
2320 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2323 * As we rewrite each gang header, the pipeline will compute
2324 * a new gang block header checksum for it; but no one will
2325 * compute a new data checksum, so we do that here. The one
2326 * exception is the gang leader: the pipeline already computed
2327 * its data checksum because that stage precedes gang assembly.
2328 * (Presently, nothing actually uses interior data checksums;
2329 * this is just good hygiene.)
2331 if (gn != pio->io_gang_leader->io_gang_tree) {
2332 abd_t *buf = abd_get_offset(data, offset);
2334 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
2335 buf, BP_GET_PSIZE(bp));
2340 * If we are here to damage data for testing purposes,
2341 * leave the GBH alone so that we can detect the damage.
2343 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
2344 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2346 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2347 abd_get_offset(data, offset), BP_GET_PSIZE(bp),
2348 zio_gang_issue_func_done, NULL, pio->io_priority,
2349 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2357 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2360 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
2361 ZIO_GANG_CHILD_FLAGS(pio)));
2366 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2369 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
2370 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
2373 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
2382 static void zio_gang_tree_assemble_done(zio_t *zio);
2384 static zio_gang_node_t *
2385 zio_gang_node_alloc(zio_gang_node_t **gnpp)
2387 zio_gang_node_t *gn;
2389 ASSERT(*gnpp == NULL);
2391 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
2392 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
2399 zio_gang_node_free(zio_gang_node_t **gnpp)
2401 zio_gang_node_t *gn = *gnpp;
2403 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2404 ASSERT(gn->gn_child[g] == NULL);
2406 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2407 kmem_free(gn, sizeof (*gn));
2412 zio_gang_tree_free(zio_gang_node_t **gnpp)
2414 zio_gang_node_t *gn = *gnpp;
2419 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2420 zio_gang_tree_free(&gn->gn_child[g]);
2422 zio_gang_node_free(gnpp);
2426 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
2428 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
2429 abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2431 ASSERT(gio->io_gang_leader == gio);
2432 ASSERT(BP_IS_GANG(bp));
2434 zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2435 zio_gang_tree_assemble_done, gn, gio->io_priority,
2436 ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
2440 zio_gang_tree_assemble_done(zio_t *zio)
2442 zio_t *gio = zio->io_gang_leader;
2443 zio_gang_node_t *gn = zio->io_private;
2444 blkptr_t *bp = zio->io_bp;
2446 ASSERT(gio == zio_unique_parent(zio));
2447 ASSERT(zio->io_child_count == 0);
2452 /* this ABD was created from a linear buf in zio_gang_tree_assemble */
2453 if (BP_SHOULD_BYTESWAP(bp))
2454 byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size);
2456 ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh);
2457 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
2458 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2460 abd_put(zio->io_abd);
2462 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2463 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2464 if (!BP_IS_GANG(gbp))
2466 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
2471 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data,
2474 zio_t *gio = pio->io_gang_leader;
2477 ASSERT(BP_IS_GANG(bp) == !!gn);
2478 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
2479 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
2482 * If you're a gang header, your data is in gn->gn_gbh.
2483 * If you're a gang member, your data is in 'data' and gn == NULL.
2485 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset);
2488 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2490 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2491 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2492 if (BP_IS_HOLE(gbp))
2494 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data,
2496 offset += BP_GET_PSIZE(gbp);
2500 if (gn == gio->io_gang_tree)
2501 ASSERT3U(gio->io_size, ==, offset);
2508 zio_gang_assemble(zio_t *zio)
2510 blkptr_t *bp = zio->io_bp;
2512 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
2513 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2515 zio->io_gang_leader = zio;
2517 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
2519 return (ZIO_PIPELINE_CONTINUE);
2523 zio_gang_issue(zio_t *zio)
2525 blkptr_t *bp = zio->io_bp;
2527 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT, ZIO_WAIT_DONE)) {
2528 return (ZIO_PIPELINE_STOP);
2531 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
2532 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2534 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
2535 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd,
2538 zio_gang_tree_free(&zio->io_gang_tree);
2540 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2542 return (ZIO_PIPELINE_CONTINUE);
2546 zio_write_gang_member_ready(zio_t *zio)
2548 zio_t *pio = zio_unique_parent(zio);
2549 dva_t *cdva = zio->io_bp->blk_dva;
2550 dva_t *pdva = pio->io_bp->blk_dva;
2552 ASSERTV(zio_t *gio = zio->io_gang_leader);
2554 if (BP_IS_HOLE(zio->io_bp))
2557 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
2559 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
2560 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
2561 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2562 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
2563 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
2565 mutex_enter(&pio->io_lock);
2566 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
2567 ASSERT(DVA_GET_GANG(&pdva[d]));
2568 asize = DVA_GET_ASIZE(&pdva[d]);
2569 asize += DVA_GET_ASIZE(&cdva[d]);
2570 DVA_SET_ASIZE(&pdva[d], asize);
2572 mutex_exit(&pio->io_lock);
2576 zio_write_gang_done(zio_t *zio)
2578 abd_put(zio->io_abd);
2582 zio_write_gang_block(zio_t *pio)
2584 spa_t *spa = pio->io_spa;
2585 metaslab_class_t *mc = spa_normal_class(spa);
2586 blkptr_t *bp = pio->io_bp;
2587 zio_t *gio = pio->io_gang_leader;
2589 zio_gang_node_t *gn, **gnpp;
2590 zio_gbh_phys_t *gbh;
2592 uint64_t txg = pio->io_txg;
2593 uint64_t resid = pio->io_size;
2595 int copies = gio->io_prop.zp_copies;
2601 * encrypted blocks need DVA[2] free so encrypted gang headers can't
2602 * have a third copy.
2604 gbh_copies = MIN(copies + 1, spa_max_replication(spa));
2605 if (gio->io_prop.zp_encrypt && gbh_copies >= SPA_DVAS_PER_BP)
2606 gbh_copies = SPA_DVAS_PER_BP - 1;
2608 int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER;
2609 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2610 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2611 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2613 flags |= METASLAB_ASYNC_ALLOC;
2614 VERIFY(refcount_held(&mc->mc_alloc_slots, pio));
2617 * The logical zio has already placed a reservation for
2618 * 'copies' allocation slots but gang blocks may require
2619 * additional copies. These additional copies
2620 * (i.e. gbh_copies - copies) are guaranteed to succeed
2621 * since metaslab_class_throttle_reserve() always allows
2622 * additional reservations for gang blocks.
2624 VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies,
2628 error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE,
2629 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags,
2630 &pio->io_alloc_list, pio);
2632 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2633 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2634 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2637 * If we failed to allocate the gang block header then
2638 * we remove any additional allocation reservations that
2639 * we placed here. The original reservation will
2640 * be removed when the logical I/O goes to the ready
2643 metaslab_class_throttle_unreserve(mc,
2644 gbh_copies - copies, pio);
2647 pio->io_error = error;
2648 return (ZIO_PIPELINE_CONTINUE);
2652 gnpp = &gio->io_gang_tree;
2654 gnpp = pio->io_private;
2655 ASSERT(pio->io_ready == zio_write_gang_member_ready);
2658 gn = zio_gang_node_alloc(gnpp);
2660 bzero(gbh, SPA_GANGBLOCKSIZE);
2661 gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE);
2664 * Create the gang header.
2666 zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2667 zio_write_gang_done, NULL, pio->io_priority,
2668 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2671 * Create and nowait the gang children.
2673 for (int g = 0; resid != 0; resid -= lsize, g++) {
2674 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2676 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2678 zp.zp_checksum = gio->io_prop.zp_checksum;
2679 zp.zp_compress = ZIO_COMPRESS_OFF;
2680 zp.zp_type = DMU_OT_NONE;
2682 zp.zp_copies = gio->io_prop.zp_copies;
2683 zp.zp_dedup = B_FALSE;
2684 zp.zp_dedup_verify = B_FALSE;
2685 zp.zp_nopwrite = B_FALSE;
2686 zp.zp_encrypt = gio->io_prop.zp_encrypt;
2687 zp.zp_byteorder = gio->io_prop.zp_byteorder;
2688 bzero(zp.zp_salt, ZIO_DATA_SALT_LEN);
2689 bzero(zp.zp_iv, ZIO_DATA_IV_LEN);
2690 bzero(zp.zp_mac, ZIO_DATA_MAC_LEN);
2692 zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2693 abd_get_offset(pio->io_abd, pio->io_size - resid), lsize,
2694 lsize, &zp, zio_write_gang_member_ready, NULL, NULL,
2695 zio_write_gang_done, &gn->gn_child[g], pio->io_priority,
2696 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2698 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2699 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2700 ASSERT(!(pio->io_flags & ZIO_FLAG_NODATA));
2703 * Gang children won't throttle but we should
2704 * account for their work, so reserve an allocation
2705 * slot for them here.
2707 VERIFY(metaslab_class_throttle_reserve(mc,
2708 zp.zp_copies, cio, flags));
2714 * Set pio's pipeline to just wait for zio to finish.
2716 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2719 * We didn't allocate this bp, so make sure it doesn't get unmarked.
2721 pio->io_flags &= ~ZIO_FLAG_FASTWRITE;
2725 return (ZIO_PIPELINE_CONTINUE);
2729 * The zio_nop_write stage in the pipeline determines if allocating a
2730 * new bp is necessary. The nopwrite feature can handle writes in
2731 * either syncing or open context (i.e. zil writes) and as a result is
2732 * mutually exclusive with dedup.
2734 * By leveraging a cryptographically secure checksum, such as SHA256, we
2735 * can compare the checksums of the new data and the old to determine if
2736 * allocating a new block is required. Note that our requirements for
2737 * cryptographic strength are fairly weak: there can't be any accidental
2738 * hash collisions, but we don't need to be secure against intentional
2739 * (malicious) collisions. To trigger a nopwrite, you have to be able
2740 * to write the file to begin with, and triggering an incorrect (hash
2741 * collision) nopwrite is no worse than simply writing to the file.
2742 * That said, there are no known attacks against the checksum algorithms
2743 * used for nopwrite, assuming that the salt and the checksums
2744 * themselves remain secret.
2747 zio_nop_write(zio_t *zio)
2749 blkptr_t *bp = zio->io_bp;
2750 blkptr_t *bp_orig = &zio->io_bp_orig;
2751 zio_prop_t *zp = &zio->io_prop;
2753 ASSERT(BP_GET_LEVEL(bp) == 0);
2754 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2755 ASSERT(zp->zp_nopwrite);
2756 ASSERT(!zp->zp_dedup);
2757 ASSERT(zio->io_bp_override == NULL);
2758 ASSERT(IO_IS_ALLOCATING(zio));
2761 * Check to see if the original bp and the new bp have matching
2762 * characteristics (i.e. same checksum, compression algorithms, etc).
2763 * If they don't then just continue with the pipeline which will
2764 * allocate a new bp.
2766 if (BP_IS_HOLE(bp_orig) ||
2767 !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
2768 ZCHECKSUM_FLAG_NOPWRITE) ||
2769 BP_IS_ENCRYPTED(bp) || BP_IS_ENCRYPTED(bp_orig) ||
2770 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2771 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2772 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2773 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2774 return (ZIO_PIPELINE_CONTINUE);
2777 * If the checksums match then reset the pipeline so that we
2778 * avoid allocating a new bp and issuing any I/O.
2780 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2781 ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
2782 ZCHECKSUM_FLAG_NOPWRITE);
2783 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2784 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2785 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2786 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2787 sizeof (uint64_t)) == 0);
2790 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2791 zio->io_flags |= ZIO_FLAG_NOPWRITE;
2794 return (ZIO_PIPELINE_CONTINUE);
2798 * ==========================================================================
2800 * ==========================================================================
2803 zio_ddt_child_read_done(zio_t *zio)
2805 blkptr_t *bp = zio->io_bp;
2806 ddt_entry_t *dde = zio->io_private;
2808 zio_t *pio = zio_unique_parent(zio);
2810 mutex_enter(&pio->io_lock);
2811 ddp = ddt_phys_select(dde, bp);
2812 if (zio->io_error == 0)
2813 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2815 if (zio->io_error == 0 && dde->dde_repair_abd == NULL)
2816 dde->dde_repair_abd = zio->io_abd;
2818 abd_free(zio->io_abd);
2819 mutex_exit(&pio->io_lock);
2823 zio_ddt_read_start(zio_t *zio)
2825 blkptr_t *bp = zio->io_bp;
2827 ASSERT(BP_GET_DEDUP(bp));
2828 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2829 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2831 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2832 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2833 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2834 ddt_phys_t *ddp = dde->dde_phys;
2835 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2838 ASSERT(zio->io_vsd == NULL);
2841 if (ddp_self == NULL)
2842 return (ZIO_PIPELINE_CONTINUE);
2844 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2845 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2847 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2849 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2850 abd_alloc_for_io(zio->io_size, B_TRUE),
2851 zio->io_size, zio_ddt_child_read_done, dde,
2852 zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) |
2853 ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark));
2855 return (ZIO_PIPELINE_CONTINUE);
2858 zio_nowait(zio_read(zio, zio->io_spa, bp,
2859 zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority,
2860 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2862 return (ZIO_PIPELINE_CONTINUE);
2866 zio_ddt_read_done(zio_t *zio)
2868 blkptr_t *bp = zio->io_bp;
2870 if (zio_wait_for_children(zio, ZIO_CHILD_DDT_BIT, ZIO_WAIT_DONE)) {
2871 return (ZIO_PIPELINE_STOP);
2874 ASSERT(BP_GET_DEDUP(bp));
2875 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2876 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2878 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2879 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2880 ddt_entry_t *dde = zio->io_vsd;
2882 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2883 return (ZIO_PIPELINE_CONTINUE);
2886 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2887 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2888 return (ZIO_PIPELINE_STOP);
2890 if (dde->dde_repair_abd != NULL) {
2891 abd_copy(zio->io_abd, dde->dde_repair_abd,
2893 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2895 ddt_repair_done(ddt, dde);
2899 ASSERT(zio->io_vsd == NULL);
2901 return (ZIO_PIPELINE_CONTINUE);
2905 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2907 spa_t *spa = zio->io_spa;
2908 boolean_t do_raw = !!(zio->io_flags & ZIO_FLAG_RAW);
2910 ASSERT(!(zio->io_bp_override && do_raw));
2913 * Note: we compare the original data, not the transformed data,
2914 * because when zio->io_bp is an override bp, we will not have
2915 * pushed the I/O transforms. That's an important optimization
2916 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2917 * However, we should never get a raw, override zio so in these
2918 * cases we can compare the io_abd directly. This is useful because
2919 * it allows us to do dedup verification even if we don't have access
2920 * to the original data (for instance, if the encryption keys aren't
2924 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2925 zio_t *lio = dde->dde_lead_zio[p];
2927 if (lio != NULL && do_raw) {
2928 return (lio->io_size != zio->io_size ||
2929 abd_cmp(zio->io_abd, lio->io_abd) != 0);
2930 } else if (lio != NULL) {
2931 return (lio->io_orig_size != zio->io_orig_size ||
2932 abd_cmp(zio->io_orig_abd, lio->io_orig_abd) != 0);
2936 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2937 ddt_phys_t *ddp = &dde->dde_phys[p];
2939 if (ddp->ddp_phys_birth != 0 && do_raw) {
2940 blkptr_t blk = *zio->io_bp;
2945 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2946 psize = BP_GET_PSIZE(&blk);
2948 if (psize != zio->io_size)
2953 tmpabd = abd_alloc_for_io(psize, B_TRUE);
2955 error = zio_wait(zio_read(NULL, spa, &blk, tmpabd,
2956 psize, NULL, NULL, ZIO_PRIORITY_SYNC_READ,
2957 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
2958 ZIO_FLAG_RAW, &zio->io_bookmark));
2961 if (abd_cmp(tmpabd, zio->io_abd) != 0)
2962 error = SET_ERROR(ENOENT);
2967 return (error != 0);
2968 } else if (ddp->ddp_phys_birth != 0) {
2969 arc_buf_t *abuf = NULL;
2970 arc_flags_t aflags = ARC_FLAG_WAIT;
2971 blkptr_t blk = *zio->io_bp;
2974 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2976 if (BP_GET_LSIZE(&blk) != zio->io_orig_size)
2981 error = arc_read(NULL, spa, &blk,
2982 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
2983 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2984 &aflags, &zio->io_bookmark);
2987 if (abd_cmp_buf(zio->io_orig_abd, abuf->b_data,
2988 zio->io_orig_size) != 0)
2989 error = SET_ERROR(ENOENT);
2990 arc_buf_destroy(abuf, &abuf);
2994 return (error != 0);
3002 zio_ddt_child_write_ready(zio_t *zio)
3004 int p = zio->io_prop.zp_copies;
3005 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
3006 ddt_entry_t *dde = zio->io_private;
3007 ddt_phys_t *ddp = &dde->dde_phys[p];
3015 ASSERT(dde->dde_lead_zio[p] == zio);
3017 ddt_phys_fill(ddp, zio->io_bp);
3019 zio_link_t *zl = NULL;
3020 while ((pio = zio_walk_parents(zio, &zl)) != NULL)
3021 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
3027 zio_ddt_child_write_done(zio_t *zio)
3029 int p = zio->io_prop.zp_copies;
3030 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
3031 ddt_entry_t *dde = zio->io_private;
3032 ddt_phys_t *ddp = &dde->dde_phys[p];
3036 ASSERT(ddp->ddp_refcnt == 0);
3037 ASSERT(dde->dde_lead_zio[p] == zio);
3038 dde->dde_lead_zio[p] = NULL;
3040 if (zio->io_error == 0) {
3041 zio_link_t *zl = NULL;
3042 while (zio_walk_parents(zio, &zl) != NULL)
3043 ddt_phys_addref(ddp);
3045 ddt_phys_clear(ddp);
3052 zio_ddt_ditto_write_done(zio_t *zio)
3054 int p = DDT_PHYS_DITTO;
3055 ASSERTV(zio_prop_t *zp = &zio->io_prop);
3056 blkptr_t *bp = zio->io_bp;
3057 ddt_t *ddt = ddt_select(zio->io_spa, bp);
3058 ddt_entry_t *dde = zio->io_private;
3059 ddt_phys_t *ddp = &dde->dde_phys[p];
3060 ddt_key_t *ddk = &dde->dde_key;
3064 ASSERT(ddp->ddp_refcnt == 0);
3065 ASSERT(dde->dde_lead_zio[p] == zio);
3066 dde->dde_lead_zio[p] = NULL;
3068 if (zio->io_error == 0) {
3069 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
3070 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
3071 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
3072 if (ddp->ddp_phys_birth != 0)
3073 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
3074 ddt_phys_fill(ddp, bp);
3081 zio_ddt_write(zio_t *zio)
3083 spa_t *spa = zio->io_spa;
3084 blkptr_t *bp = zio->io_bp;
3085 uint64_t txg = zio->io_txg;
3086 zio_prop_t *zp = &zio->io_prop;
3087 int p = zp->zp_copies;
3091 ddt_t *ddt = ddt_select(spa, bp);
3095 ASSERT(BP_GET_DEDUP(bp));
3096 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
3097 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
3098 ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW)));
3101 dde = ddt_lookup(ddt, bp, B_TRUE);
3102 ddp = &dde->dde_phys[p];
3104 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
3106 * If we're using a weak checksum, upgrade to a strong checksum
3107 * and try again. If we're already using a strong checksum,
3108 * we can't resolve it, so just convert to an ordinary write.
3109 * (And automatically e-mail a paper to Nature?)
3111 if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
3112 ZCHECKSUM_FLAG_DEDUP)) {
3113 zp->zp_checksum = spa_dedup_checksum(spa);
3114 zio_pop_transforms(zio);
3115 zio->io_stage = ZIO_STAGE_OPEN;
3118 zp->zp_dedup = B_FALSE;
3120 zio->io_pipeline = ZIO_WRITE_PIPELINE;
3122 return (ZIO_PIPELINE_CONTINUE);
3125 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
3126 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
3128 if (ditto_copies > ddt_ditto_copies_present(dde) &&
3129 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
3130 zio_prop_t czp = *zp;
3132 czp.zp_copies = ditto_copies;
3135 * If we arrived here with an override bp, we won't have run
3136 * the transform stack, so we won't have the data we need to
3137 * generate a child i/o. So, toss the override bp and restart.
3138 * This is safe, because using the override bp is just an
3139 * optimization; and it's rare, so the cost doesn't matter.
3141 if (zio->io_bp_override) {
3142 zio_pop_transforms(zio);
3143 zio->io_stage = ZIO_STAGE_OPEN;
3144 zio->io_pipeline = ZIO_WRITE_PIPELINE;
3145 zio->io_bp_override = NULL;
3148 return (ZIO_PIPELINE_CONTINUE);
3151 dio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
3152 zio->io_orig_size, zio->io_orig_size, &czp, NULL, NULL,
3153 NULL, zio_ddt_ditto_write_done, dde, zio->io_priority,
3154 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
3156 zio_push_transform(dio, zio->io_abd, zio->io_size, 0, NULL);
3157 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
3160 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
3161 if (ddp->ddp_phys_birth != 0)
3162 ddt_bp_fill(ddp, bp, txg);
3163 if (dde->dde_lead_zio[p] != NULL)
3164 zio_add_child(zio, dde->dde_lead_zio[p]);
3166 ddt_phys_addref(ddp);
3167 } else if (zio->io_bp_override) {
3168 ASSERT(bp->blk_birth == txg);
3169 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
3170 ddt_phys_fill(ddp, bp);
3171 ddt_phys_addref(ddp);
3173 cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
3174 zio->io_orig_size, zio->io_orig_size, zp,
3175 zio_ddt_child_write_ready, NULL, NULL,
3176 zio_ddt_child_write_done, dde, zio->io_priority,
3177 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
3179 zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL);
3180 dde->dde_lead_zio[p] = cio;
3190 return (ZIO_PIPELINE_CONTINUE);
3193 ddt_entry_t *freedde; /* for debugging */
3196 zio_ddt_free(zio_t *zio)
3198 spa_t *spa = zio->io_spa;
3199 blkptr_t *bp = zio->io_bp;
3200 ddt_t *ddt = ddt_select(spa, bp);
3204 ASSERT(BP_GET_DEDUP(bp));
3205 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3208 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
3210 ddp = ddt_phys_select(dde, bp);
3212 ddt_phys_decref(ddp);
3216 return (ZIO_PIPELINE_CONTINUE);
3220 * ==========================================================================
3221 * Allocate and free blocks
3222 * ==========================================================================
3226 zio_io_to_allocate(spa_t *spa)
3230 ASSERT(MUTEX_HELD(&spa->spa_alloc_lock));
3232 zio = avl_first(&spa->spa_alloc_tree);
3236 ASSERT(IO_IS_ALLOCATING(zio));
3239 * Try to place a reservation for this zio. If we're unable to
3240 * reserve then we throttle.
3242 if (!metaslab_class_throttle_reserve(spa_normal_class(spa),
3243 zio->io_prop.zp_copies, zio, 0)) {
3247 avl_remove(&spa->spa_alloc_tree, zio);
3248 ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE);
3254 zio_dva_throttle(zio_t *zio)
3256 spa_t *spa = zio->io_spa;
3259 if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE ||
3260 !spa_normal_class(zio->io_spa)->mc_alloc_throttle_enabled ||
3261 zio->io_child_type == ZIO_CHILD_GANG ||
3262 zio->io_flags & ZIO_FLAG_NODATA) {
3263 return (ZIO_PIPELINE_CONTINUE);
3266 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3268 ASSERT3U(zio->io_queued_timestamp, >, 0);
3269 ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE);
3271 mutex_enter(&spa->spa_alloc_lock);
3273 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3274 avl_add(&spa->spa_alloc_tree, zio);
3276 nio = zio_io_to_allocate(zio->io_spa);
3277 mutex_exit(&spa->spa_alloc_lock);
3280 return (ZIO_PIPELINE_CONTINUE);
3283 ASSERT(nio->io_stage == ZIO_STAGE_DVA_THROTTLE);
3285 * We are passing control to a new zio so make sure that
3286 * it is processed by a different thread. We do this to
3287 * avoid stack overflows that can occur when parents are
3288 * throttled and children are making progress. We allow
3289 * it to go to the head of the taskq since it's already
3292 zio_taskq_dispatch(nio, ZIO_TASKQ_ISSUE, B_TRUE);
3294 return (ZIO_PIPELINE_STOP);
3298 zio_allocate_dispatch(spa_t *spa)
3302 mutex_enter(&spa->spa_alloc_lock);
3303 zio = zio_io_to_allocate(spa);
3304 mutex_exit(&spa->spa_alloc_lock);
3308 ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE);
3309 ASSERT0(zio->io_error);
3310 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE);
3314 zio_dva_allocate(zio_t *zio)
3316 spa_t *spa = zio->io_spa;
3317 metaslab_class_t *mc = spa_normal_class(spa);
3318 blkptr_t *bp = zio->io_bp;
3322 if (zio->io_gang_leader == NULL) {
3323 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3324 zio->io_gang_leader = zio;
3327 ASSERT(BP_IS_HOLE(bp));
3328 ASSERT0(BP_GET_NDVAS(bp));
3329 ASSERT3U(zio->io_prop.zp_copies, >, 0);
3330 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
3331 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
3333 flags |= (zio->io_flags & ZIO_FLAG_FASTWRITE) ? METASLAB_FASTWRITE : 0;
3334 if (zio->io_flags & ZIO_FLAG_NODATA)
3335 flags |= METASLAB_DONT_THROTTLE;
3336 if (zio->io_flags & ZIO_FLAG_GANG_CHILD)
3337 flags |= METASLAB_GANG_CHILD;
3338 if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE)
3339 flags |= METASLAB_ASYNC_ALLOC;
3341 error = metaslab_alloc(spa, mc, zio->io_size, bp,
3342 zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
3343 &zio->io_alloc_list, zio);
3346 zfs_dbgmsg("%s: metaslab allocation failure: zio %p, "
3347 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
3349 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
3350 return (zio_write_gang_block(zio));
3351 zio->io_error = error;
3354 return (ZIO_PIPELINE_CONTINUE);
3358 zio_dva_free(zio_t *zio)
3360 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
3362 return (ZIO_PIPELINE_CONTINUE);
3366 zio_dva_claim(zio_t *zio)
3370 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
3372 zio->io_error = error;
3374 return (ZIO_PIPELINE_CONTINUE);
3378 * Undo an allocation. This is used by zio_done() when an I/O fails
3379 * and we want to give back the block we just allocated.
3380 * This handles both normal blocks and gang blocks.
3383 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
3385 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
3386 ASSERT(zio->io_bp_override == NULL);
3388 if (!BP_IS_HOLE(bp))
3389 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
3392 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
3393 zio_dva_unallocate(zio, gn->gn_child[g],
3394 &gn->gn_gbh->zg_blkptr[g]);
3400 * Try to allocate an intent log block. Return 0 on success, errno on failure.
3403 zio_alloc_zil(spa_t *spa, objset_t *os, uint64_t txg, blkptr_t *new_bp,
3404 uint64_t size, boolean_t *slog)
3407 zio_alloc_list_t io_alloc_list;
3409 ASSERT(txg > spa_syncing_txg(spa));
3411 metaslab_trace_init(&io_alloc_list);
3412 error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1,
3413 txg, NULL, METASLAB_FASTWRITE, &io_alloc_list, NULL);
3417 error = metaslab_alloc(spa, spa_normal_class(spa), size,
3418 new_bp, 1, txg, NULL, METASLAB_FASTWRITE,
3419 &io_alloc_list, NULL);
3423 metaslab_trace_fini(&io_alloc_list);
3426 BP_SET_LSIZE(new_bp, size);
3427 BP_SET_PSIZE(new_bp, size);
3428 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
3429 BP_SET_CHECKSUM(new_bp,
3430 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
3431 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
3432 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3433 BP_SET_LEVEL(new_bp, 0);
3434 BP_SET_DEDUP(new_bp, 0);
3435 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
3438 * encrypted blocks will require an IV and salt. We generate
3439 * these now since we will not be rewriting the bp at
3442 if (os->os_encrypted) {
3443 uint8_t iv[ZIO_DATA_IV_LEN];
3444 uint8_t salt[ZIO_DATA_SALT_LEN];
3446 BP_SET_CRYPT(new_bp, B_TRUE);
3447 VERIFY0(spa_crypt_get_salt(spa,
3448 dmu_objset_id(os), salt));
3449 VERIFY0(zio_crypt_generate_iv(iv));
3451 zio_crypt_encode_params_bp(new_bp, salt, iv);
3454 zfs_dbgmsg("%s: zil block allocation failure: "
3455 "size %llu, error %d", spa_name(spa), size, error);
3462 * ==========================================================================
3463 * Read and write to physical devices
3464 * ==========================================================================
3469 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3470 * stops after this stage and will resume upon I/O completion.
3471 * However, there are instances where the vdev layer may need to
3472 * continue the pipeline when an I/O was not issued. Since the I/O
3473 * that was sent to the vdev layer might be different than the one
3474 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3475 * force the underlying vdev layers to call either zio_execute() or
3476 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3479 zio_vdev_io_start(zio_t *zio)
3481 vdev_t *vd = zio->io_vd;
3483 spa_t *spa = zio->io_spa;
3487 ASSERT(zio->io_error == 0);
3488 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
3491 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3492 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
3495 * The mirror_ops handle multiple DVAs in a single BP.
3497 vdev_mirror_ops.vdev_op_io_start(zio);
3498 return (ZIO_PIPELINE_STOP);
3501 ASSERT3P(zio->io_logical, !=, zio);
3502 if (zio->io_type == ZIO_TYPE_WRITE) {
3503 ASSERT(spa->spa_trust_config);
3506 * Note: the code can handle other kinds of writes,
3507 * but we don't expect them.
3509 if (zio->io_vd->vdev_removing) {
3510 ASSERT(zio->io_flags &
3511 (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL |
3512 ZIO_FLAG_RESILVER | ZIO_FLAG_INDUCE_DAMAGE));
3516 align = 1ULL << vd->vdev_top->vdev_ashift;
3518 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
3519 P2PHASE(zio->io_size, align) != 0) {
3520 /* Transform logical writes to be a full physical block size. */
3521 uint64_t asize = P2ROUNDUP(zio->io_size, align);
3522 abd_t *abuf = abd_alloc_sametype(zio->io_abd, asize);
3523 ASSERT(vd == vd->vdev_top);
3524 if (zio->io_type == ZIO_TYPE_WRITE) {
3525 abd_copy(abuf, zio->io_abd, zio->io_size);
3526 abd_zero_off(abuf, zio->io_size, asize - zio->io_size);
3528 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
3532 * If this is not a physical io, make sure that it is properly aligned
3533 * before proceeding.
3535 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
3536 ASSERT0(P2PHASE(zio->io_offset, align));
3537 ASSERT0(P2PHASE(zio->io_size, align));
3540 * For physical writes, we allow 512b aligned writes and assume
3541 * the device will perform a read-modify-write as necessary.
3543 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
3544 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
3547 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
3550 * If this is a repair I/O, and there's no self-healing involved --
3551 * that is, we're just resilvering what we expect to resilver --
3552 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3553 * This prevents spurious resilvering.
3555 * There are a few ways that we can end up creating these spurious
3558 * 1. A resilver i/o will be issued if any DVA in the BP has a
3559 * dirty DTL. The mirror code will issue resilver writes to
3560 * each DVA, including the one(s) that are not on vdevs with dirty
3563 * 2. With nested replication, which happens when we have a
3564 * "replacing" or "spare" vdev that's a child of a mirror or raidz.
3565 * For example, given mirror(replacing(A+B), C), it's likely that
3566 * only A is out of date (it's the new device). In this case, we'll
3567 * read from C, then use the data to resilver A+B -- but we don't
3568 * actually want to resilver B, just A. The top-level mirror has no
3569 * way to know this, so instead we just discard unnecessary repairs
3570 * as we work our way down the vdev tree.
3572 * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc.
3573 * The same logic applies to any form of nested replication: ditto
3574 * + mirror, RAID-Z + replacing, etc.
3576 * However, indirect vdevs point off to other vdevs which may have
3577 * DTL's, so we never bypass them. The child i/os on concrete vdevs
3578 * will be properly bypassed instead.
3580 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3581 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
3582 zio->io_txg != 0 && /* not a delegated i/o */
3583 vd->vdev_ops != &vdev_indirect_ops &&
3584 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
3585 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3586 zio_vdev_io_bypass(zio);
3587 return (ZIO_PIPELINE_CONTINUE);
3590 if (vd->vdev_ops->vdev_op_leaf &&
3591 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
3593 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio))
3594 return (ZIO_PIPELINE_CONTINUE);
3596 if ((zio = vdev_queue_io(zio)) == NULL)
3597 return (ZIO_PIPELINE_STOP);
3599 if (!vdev_accessible(vd, zio)) {
3600 zio->io_error = SET_ERROR(ENXIO);
3602 return (ZIO_PIPELINE_STOP);
3604 zio->io_delay = gethrtime();
3607 vd->vdev_ops->vdev_op_io_start(zio);
3608 return (ZIO_PIPELINE_STOP);
3612 zio_vdev_io_done(zio_t *zio)
3614 vdev_t *vd = zio->io_vd;
3615 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
3616 boolean_t unexpected_error = B_FALSE;
3618 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3619 return (ZIO_PIPELINE_STOP);
3622 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
3625 zio->io_delay = gethrtime() - zio->io_delay;
3627 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
3629 vdev_queue_io_done(zio);
3631 if (zio->io_type == ZIO_TYPE_WRITE)
3632 vdev_cache_write(zio);
3634 if (zio_injection_enabled && zio->io_error == 0)
3635 zio->io_error = zio_handle_device_injections(vd, zio,
3638 if (zio_injection_enabled && zio->io_error == 0)
3639 zio->io_error = zio_handle_label_injection(zio, EIO);
3641 if (zio->io_error) {
3642 if (!vdev_accessible(vd, zio)) {
3643 zio->io_error = SET_ERROR(ENXIO);
3645 unexpected_error = B_TRUE;
3650 ops->vdev_op_io_done(zio);
3652 if (unexpected_error)
3653 VERIFY(vdev_probe(vd, zio) == NULL);
3655 return (ZIO_PIPELINE_CONTINUE);
3659 * This function is used to change the priority of an existing zio that is
3660 * currently in-flight. This is used by the arc to upgrade priority in the
3661 * event that a demand read is made for a block that is currently queued
3662 * as a scrub or async read IO. Otherwise, the high priority read request
3663 * would end up having to wait for the lower priority IO.
3666 zio_change_priority(zio_t *pio, zio_priority_t priority)
3668 zio_t *cio, *cio_next;
3669 zio_link_t *zl = NULL;
3671 ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
3673 if (pio->io_vd != NULL && pio->io_vd->vdev_ops->vdev_op_leaf) {
3674 vdev_queue_change_io_priority(pio, priority);
3676 pio->io_priority = priority;
3679 mutex_enter(&pio->io_lock);
3680 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
3681 cio_next = zio_walk_children(pio, &zl);
3682 zio_change_priority(cio, priority);
3684 mutex_exit(&pio->io_lock);
3688 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3689 * disk, and use that to finish the checksum ereport later.
3692 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
3693 const abd_t *good_buf)
3695 /* no processing needed */
3696 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
3701 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
3703 void *abd = abd_alloc_sametype(zio->io_abd, zio->io_size);
3705 abd_copy(abd, zio->io_abd, zio->io_size);
3707 zcr->zcr_cbinfo = zio->io_size;
3708 zcr->zcr_cbdata = abd;
3709 zcr->zcr_finish = zio_vsd_default_cksum_finish;
3710 zcr->zcr_free = zio_abd_free;
3714 zio_vdev_io_assess(zio_t *zio)
3716 vdev_t *vd = zio->io_vd;
3718 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3719 return (ZIO_PIPELINE_STOP);
3722 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3723 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
3725 if (zio->io_vsd != NULL) {
3726 zio->io_vsd_ops->vsd_free(zio);
3730 if (zio_injection_enabled && zio->io_error == 0)
3731 zio->io_error = zio_handle_fault_injection(zio, EIO);
3734 * If the I/O failed, determine whether we should attempt to retry it.
3736 * On retry, we cut in line in the issue queue, since we don't want
3737 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3739 if (zio->io_error && vd == NULL &&
3740 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
3741 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
3742 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
3744 zio->io_flags |= ZIO_FLAG_IO_RETRY |
3745 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
3746 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
3747 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
3748 zio_requeue_io_start_cut_in_line);
3749 return (ZIO_PIPELINE_STOP);
3753 * If we got an error on a leaf device, convert it to ENXIO
3754 * if the device is not accessible at all.
3756 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3757 !vdev_accessible(vd, zio))
3758 zio->io_error = SET_ERROR(ENXIO);
3761 * If we can't write to an interior vdev (mirror or RAID-Z),
3762 * set vdev_cant_write so that we stop trying to allocate from it.
3764 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
3765 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
3766 vd->vdev_cant_write = B_TRUE;
3770 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
3771 * attempts will ever succeed. In this case we set a persistent bit so
3772 * that we don't bother with it in the future.
3774 if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) &&
3775 zio->io_type == ZIO_TYPE_IOCTL &&
3776 zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL)
3777 vd->vdev_nowritecache = B_TRUE;
3780 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3782 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3783 zio->io_physdone != NULL) {
3784 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
3785 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
3786 zio->io_physdone(zio->io_logical);
3789 return (ZIO_PIPELINE_CONTINUE);
3793 zio_vdev_io_reissue(zio_t *zio)
3795 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3796 ASSERT(zio->io_error == 0);
3798 zio->io_stage >>= 1;
3802 zio_vdev_io_redone(zio_t *zio)
3804 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
3806 zio->io_stage >>= 1;
3810 zio_vdev_io_bypass(zio_t *zio)
3812 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3813 ASSERT(zio->io_error == 0);
3815 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
3816 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
3820 * ==========================================================================
3821 * Encrypt and store encryption parameters
3822 * ==========================================================================
3827 * This function is used for ZIO_STAGE_ENCRYPT. It is responsible for
3828 * managing the storage of encryption parameters and passing them to the
3829 * lower-level encryption functions.
3832 zio_encrypt(zio_t *zio)
3834 zio_prop_t *zp = &zio->io_prop;
3835 spa_t *spa = zio->io_spa;
3836 blkptr_t *bp = zio->io_bp;
3837 uint64_t psize = BP_GET_PSIZE(bp);
3838 uint64_t dsobj = zio->io_bookmark.zb_objset;
3839 dmu_object_type_t ot = BP_GET_TYPE(bp);
3840 void *enc_buf = NULL;
3842 uint8_t salt[ZIO_DATA_SALT_LEN];
3843 uint8_t iv[ZIO_DATA_IV_LEN];
3844 uint8_t mac[ZIO_DATA_MAC_LEN];
3845 boolean_t no_crypt = B_FALSE;
3847 /* the root zio already encrypted the data */
3848 if (zio->io_child_type == ZIO_CHILD_GANG)
3849 return (ZIO_PIPELINE_CONTINUE);
3851 /* only ZIL blocks are re-encrypted on rewrite */
3852 if (!IO_IS_ALLOCATING(zio) && ot != DMU_OT_INTENT_LOG)
3853 return (ZIO_PIPELINE_CONTINUE);
3855 if (!(zp->zp_encrypt || BP_IS_ENCRYPTED(bp))) {
3856 BP_SET_CRYPT(bp, B_FALSE);
3857 return (ZIO_PIPELINE_CONTINUE);
3860 /* if we are doing raw encryption set the provided encryption params */
3861 if (zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) {
3862 ASSERT0(BP_GET_LEVEL(bp));
3863 BP_SET_CRYPT(bp, B_TRUE);
3864 BP_SET_BYTEORDER(bp, zp->zp_byteorder);
3865 if (ot != DMU_OT_OBJSET)
3866 zio_crypt_encode_mac_bp(bp, zp->zp_mac);
3868 /* dnode blocks must be written out in the provided byteorder */
3869 if (zp->zp_byteorder != ZFS_HOST_BYTEORDER &&
3870 ot == DMU_OT_DNODE) {
3871 void *bswap_buf = zio_buf_alloc(psize);
3872 abd_t *babd = abd_get_from_buf(bswap_buf, psize);
3874 ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
3875 abd_copy_to_buf(bswap_buf, zio->io_abd, psize);
3876 dmu_ot_byteswap[DMU_OT_BYTESWAP(ot)].ob_func(bswap_buf,
3879 abd_take_ownership_of_buf(babd, B_TRUE);
3880 zio_push_transform(zio, babd, psize, psize, NULL);
3883 if (DMU_OT_IS_ENCRYPTED(ot))
3884 zio_crypt_encode_params_bp(bp, zp->zp_salt, zp->zp_iv);
3885 return (ZIO_PIPELINE_CONTINUE);
3888 /* indirect blocks only maintain a cksum of the lower level MACs */
3889 if (BP_GET_LEVEL(bp) > 0) {
3890 BP_SET_CRYPT(bp, B_TRUE);
3891 VERIFY0(zio_crypt_do_indirect_mac_checksum_abd(B_TRUE,
3892 zio->io_orig_abd, BP_GET_LSIZE(bp), BP_SHOULD_BYTESWAP(bp),
3894 zio_crypt_encode_mac_bp(bp, mac);
3895 return (ZIO_PIPELINE_CONTINUE);
3899 * Objset blocks are a special case since they have 2 256-bit MACs
3900 * embedded within them.
3902 if (ot == DMU_OT_OBJSET) {
3903 ASSERT0(DMU_OT_IS_ENCRYPTED(ot));
3904 ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
3905 BP_SET_CRYPT(bp, B_TRUE);
3906 VERIFY0(spa_do_crypt_objset_mac_abd(B_TRUE, spa, dsobj,
3907 zio->io_abd, psize, BP_SHOULD_BYTESWAP(bp)));
3908 return (ZIO_PIPELINE_CONTINUE);
3911 /* unencrypted object types are only authenticated with a MAC */
3912 if (!DMU_OT_IS_ENCRYPTED(ot)) {
3913 BP_SET_CRYPT(bp, B_TRUE);
3914 VERIFY0(spa_do_crypt_mac_abd(B_TRUE, spa, dsobj,
3915 zio->io_abd, psize, mac));
3916 zio_crypt_encode_mac_bp(bp, mac);
3917 return (ZIO_PIPELINE_CONTINUE);
3921 * Later passes of sync-to-convergence may decide to rewrite data
3922 * in place to avoid more disk reallocations. This presents a problem
3923 * for encryption because this consitutes rewriting the new data with
3924 * the same encryption key and IV. However, this only applies to blocks
3925 * in the MOS (particularly the spacemaps) and we do not encrypt the
3926 * MOS. We assert that the zio is allocating or an intent log write
3929 ASSERT(IO_IS_ALLOCATING(zio) || ot == DMU_OT_INTENT_LOG);
3930 ASSERT(BP_GET_LEVEL(bp) == 0 || ot == DMU_OT_INTENT_LOG);
3931 ASSERT(spa_feature_is_active(spa, SPA_FEATURE_ENCRYPTION));
3932 ASSERT3U(psize, !=, 0);
3934 enc_buf = zio_buf_alloc(psize);
3935 eabd = abd_get_from_buf(enc_buf, psize);
3936 abd_take_ownership_of_buf(eabd, B_TRUE);
3939 * For an explanation of what encryption parameters are stored
3940 * where, see the block comment in zio_crypt.c.
3942 if (ot == DMU_OT_INTENT_LOG) {
3943 zio_crypt_decode_params_bp(bp, salt, iv);
3945 BP_SET_CRYPT(bp, B_TRUE);
3948 /* Perform the encryption. This should not fail */
3949 VERIFY0(spa_do_crypt_abd(B_TRUE, spa, &zio->io_bookmark,
3950 BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp),
3951 salt, iv, mac, psize, zio->io_abd, eabd, &no_crypt));
3953 /* encode encryption metadata into the bp */
3954 if (ot == DMU_OT_INTENT_LOG) {
3956 * ZIL blocks store the MAC in the embedded checksum, so the
3957 * transform must always be applied.
3959 zio_crypt_encode_mac_zil(enc_buf, mac);
3960 zio_push_transform(zio, eabd, psize, psize, NULL);
3962 BP_SET_CRYPT(bp, B_TRUE);
3963 zio_crypt_encode_params_bp(bp, salt, iv);
3964 zio_crypt_encode_mac_bp(bp, mac);
3967 ASSERT3U(ot, ==, DMU_OT_DNODE);
3970 zio_push_transform(zio, eabd, psize, psize, NULL);
3974 return (ZIO_PIPELINE_CONTINUE);
3978 * ==========================================================================
3979 * Generate and verify checksums
3980 * ==========================================================================
3983 zio_checksum_generate(zio_t *zio)
3985 blkptr_t *bp = zio->io_bp;
3986 enum zio_checksum checksum;
3990 * This is zio_write_phys().
3991 * We're either generating a label checksum, or none at all.
3993 checksum = zio->io_prop.zp_checksum;
3995 if (checksum == ZIO_CHECKSUM_OFF)
3996 return (ZIO_PIPELINE_CONTINUE);
3998 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
4000 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
4001 ASSERT(!IO_IS_ALLOCATING(zio));
4002 checksum = ZIO_CHECKSUM_GANG_HEADER;
4004 checksum = BP_GET_CHECKSUM(bp);
4008 zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size);
4010 return (ZIO_PIPELINE_CONTINUE);
4014 zio_checksum_verify(zio_t *zio)
4016 zio_bad_cksum_t info;
4017 blkptr_t *bp = zio->io_bp;
4020 ASSERT(zio->io_vd != NULL);
4024 * This is zio_read_phys().
4025 * We're either verifying a label checksum, or nothing at all.
4027 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
4028 return (ZIO_PIPELINE_CONTINUE);
4030 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
4033 if ((error = zio_checksum_error(zio, &info)) != 0) {
4034 zio->io_error = error;
4035 if (error == ECKSUM &&
4036 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
4037 zfs_ereport_start_checksum(zio->io_spa,
4038 zio->io_vd, &zio->io_bookmark, zio,
4039 zio->io_offset, zio->io_size, NULL, &info);
4043 return (ZIO_PIPELINE_CONTINUE);
4047 * Called by RAID-Z to ensure we don't compute the checksum twice.
4050 zio_checksum_verified(zio_t *zio)
4052 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
4056 * ==========================================================================
4057 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
4058 * An error of 0 indicates success. ENXIO indicates whole-device failure,
4059 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
4060 * indicate errors that are specific to one I/O, and most likely permanent.
4061 * Any other error is presumed to be worse because we weren't expecting it.
4062 * ==========================================================================
4065 zio_worst_error(int e1, int e2)
4067 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
4070 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
4071 if (e1 == zio_error_rank[r1])
4074 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
4075 if (e2 == zio_error_rank[r2])
4078 return (r1 > r2 ? e1 : e2);
4082 * ==========================================================================
4084 * ==========================================================================
4087 zio_ready(zio_t *zio)
4089 blkptr_t *bp = zio->io_bp;
4090 zio_t *pio, *pio_next;
4091 zio_link_t *zl = NULL;
4093 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT | ZIO_CHILD_DDT_BIT,
4095 return (ZIO_PIPELINE_STOP);
4098 if (zio->io_ready) {
4099 ASSERT(IO_IS_ALLOCATING(zio));
4100 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
4101 (zio->io_flags & ZIO_FLAG_NOPWRITE));
4102 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
4107 if (bp != NULL && bp != &zio->io_bp_copy)
4108 zio->io_bp_copy = *bp;
4110 if (zio->io_error != 0) {
4111 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
4113 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
4114 ASSERT(IO_IS_ALLOCATING(zio));
4115 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
4117 * We were unable to allocate anything, unreserve and
4118 * issue the next I/O to allocate.
4120 metaslab_class_throttle_unreserve(
4121 spa_normal_class(zio->io_spa),
4122 zio->io_prop.zp_copies, zio);
4123 zio_allocate_dispatch(zio->io_spa);
4127 mutex_enter(&zio->io_lock);
4128 zio->io_state[ZIO_WAIT_READY] = 1;
4129 pio = zio_walk_parents(zio, &zl);
4130 mutex_exit(&zio->io_lock);
4133 * As we notify zio's parents, new parents could be added.
4134 * New parents go to the head of zio's io_parent_list, however,
4135 * so we will (correctly) not notify them. The remainder of zio's
4136 * io_parent_list, from 'pio_next' onward, cannot change because
4137 * all parents must wait for us to be done before they can be done.
4139 for (; pio != NULL; pio = pio_next) {
4140 pio_next = zio_walk_parents(zio, &zl);
4141 zio_notify_parent(pio, zio, ZIO_WAIT_READY);
4144 if (zio->io_flags & ZIO_FLAG_NODATA) {
4145 if (BP_IS_GANG(bp)) {
4146 zio->io_flags &= ~ZIO_FLAG_NODATA;
4148 ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE);
4149 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
4153 if (zio_injection_enabled &&
4154 zio->io_spa->spa_syncing_txg == zio->io_txg)
4155 zio_handle_ignored_writes(zio);
4157 return (ZIO_PIPELINE_CONTINUE);
4161 * Update the allocation throttle accounting.
4164 zio_dva_throttle_done(zio_t *zio)
4166 ASSERTV(zio_t *lio = zio->io_logical);
4167 zio_t *pio = zio_unique_parent(zio);
4168 vdev_t *vd = zio->io_vd;
4169 int flags = METASLAB_ASYNC_ALLOC;
4171 ASSERT3P(zio->io_bp, !=, NULL);
4172 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
4173 ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE);
4174 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
4176 ASSERT3P(vd, ==, vd->vdev_top);
4177 ASSERT(zio_injection_enabled || !(zio->io_flags & ZIO_FLAG_IO_RETRY));
4178 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
4179 ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING);
4180 ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE));
4181 ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA));
4184 * Parents of gang children can have two flavors -- ones that
4185 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
4186 * and ones that allocated the constituent blocks. The allocation
4187 * throttle needs to know the allocating parent zio so we must find
4190 if (pio->io_child_type == ZIO_CHILD_GANG) {
4192 * If our parent is a rewrite gang child then our grandparent
4193 * would have been the one that performed the allocation.
4195 if (pio->io_flags & ZIO_FLAG_IO_REWRITE)
4196 pio = zio_unique_parent(pio);
4197 flags |= METASLAB_GANG_CHILD;
4200 ASSERT(IO_IS_ALLOCATING(pio));
4201 ASSERT3P(zio, !=, zio->io_logical);
4202 ASSERT(zio->io_logical != NULL);
4203 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
4204 ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE);
4206 mutex_enter(&pio->io_lock);
4207 metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags);
4208 mutex_exit(&pio->io_lock);
4210 metaslab_class_throttle_unreserve(spa_normal_class(zio->io_spa),
4214 * Call into the pipeline to see if there is more work that
4215 * needs to be done. If there is work to be done it will be
4216 * dispatched to another taskq thread.
4218 zio_allocate_dispatch(zio->io_spa);
4222 zio_done(zio_t *zio)
4225 * Always attempt to keep stack usage minimal here since
4226 * we can be called recurisvely up to 19 levels deep.
4228 const uint64_t psize = zio->io_size;
4229 zio_t *pio, *pio_next;
4230 zio_link_t *zl = NULL;
4233 * If our children haven't all completed,
4234 * wait for them and then repeat this pipeline stage.
4236 if (zio_wait_for_children(zio, ZIO_CHILD_ALL_BITS, ZIO_WAIT_DONE)) {
4237 return (ZIO_PIPELINE_STOP);
4241 * If the allocation throttle is enabled, then update the accounting.
4242 * We only track child I/Os that are part of an allocating async
4243 * write. We must do this since the allocation is performed
4244 * by the logical I/O but the actual write is done by child I/Os.
4246 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING &&
4247 zio->io_child_type == ZIO_CHILD_VDEV) {
4248 ASSERT(spa_normal_class(
4249 zio->io_spa)->mc_alloc_throttle_enabled);
4250 zio_dva_throttle_done(zio);
4254 * If the allocation throttle is enabled, verify that
4255 * we have decremented the refcounts for every I/O that was throttled.
4257 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
4258 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
4259 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
4260 ASSERT(zio->io_bp != NULL);
4261 metaslab_group_alloc_verify(zio->io_spa, zio->io_bp, zio);
4262 VERIFY(refcount_not_held(
4263 &(spa_normal_class(zio->io_spa)->mc_alloc_slots), zio));
4267 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
4268 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
4269 ASSERT(zio->io_children[c][w] == 0);
4271 if (zio->io_bp != NULL && !BP_IS_EMBEDDED(zio->io_bp)) {
4272 ASSERT(zio->io_bp->blk_pad[0] == 0);
4273 ASSERT(zio->io_bp->blk_pad[1] == 0);
4274 ASSERT(bcmp(zio->io_bp, &zio->io_bp_copy,
4275 sizeof (blkptr_t)) == 0 ||
4276 (zio->io_bp == zio_unique_parent(zio)->io_bp));
4277 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) &&
4278 zio->io_bp_override == NULL &&
4279 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
4280 ASSERT3U(zio->io_prop.zp_copies, <=,
4281 BP_GET_NDVAS(zio->io_bp));
4282 ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 ||
4283 (BP_COUNT_GANG(zio->io_bp) ==
4284 BP_GET_NDVAS(zio->io_bp)));
4286 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
4287 VERIFY(BP_EQUAL(zio->io_bp, &zio->io_bp_orig));
4291 * If there were child vdev/gang/ddt errors, they apply to us now.
4293 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
4294 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
4295 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
4298 * If the I/O on the transformed data was successful, generate any
4299 * checksum reports now while we still have the transformed data.
4301 if (zio->io_error == 0) {
4302 while (zio->io_cksum_report != NULL) {
4303 zio_cksum_report_t *zcr = zio->io_cksum_report;
4304 uint64_t align = zcr->zcr_align;
4305 uint64_t asize = P2ROUNDUP(psize, align);
4306 abd_t *adata = zio->io_abd;
4308 if (asize != psize) {
4309 adata = abd_alloc(asize, B_TRUE);
4310 abd_copy(adata, zio->io_abd, psize);
4311 abd_zero_off(adata, psize, asize - psize);
4314 zio->io_cksum_report = zcr->zcr_next;
4315 zcr->zcr_next = NULL;
4316 zcr->zcr_finish(zcr, adata);
4317 zfs_ereport_free_checksum(zcr);
4324 zio_pop_transforms(zio); /* note: may set zio->io_error */
4326 vdev_stat_update(zio, psize);
4329 * If this I/O is attached to a particular vdev is slow, exceeding
4330 * 30 seconds to complete, post an error described the I/O delay.
4331 * We ignore these errors if the device is currently unavailable.
4333 if (zio->io_delay >= MSEC2NSEC(zio_delay_max)) {
4334 if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd))
4335 zfs_ereport_post(FM_EREPORT_ZFS_DELAY, zio->io_spa,
4336 zio->io_vd, &zio->io_bookmark, zio, 0, 0);
4339 if (zio->io_error) {
4341 * If this I/O is attached to a particular vdev,
4342 * generate an error message describing the I/O failure
4343 * at the block level. We ignore these errors if the
4344 * device is currently unavailable.
4346 if (zio->io_error != ECKSUM && zio->io_vd != NULL &&
4347 !vdev_is_dead(zio->io_vd))
4348 zfs_ereport_post(FM_EREPORT_ZFS_IO, zio->io_spa,
4349 zio->io_vd, &zio->io_bookmark, zio, 0, 0);
4351 if ((zio->io_error == EIO || !(zio->io_flags &
4352 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
4353 zio == zio->io_logical) {
4355 * For logical I/O requests, tell the SPA to log the
4356 * error and generate a logical data ereport.
4358 spa_log_error(zio->io_spa, &zio->io_bookmark);
4359 zfs_ereport_post(FM_EREPORT_ZFS_DATA, zio->io_spa,
4360 NULL, &zio->io_bookmark, zio, 0, 0);
4364 if (zio->io_error && zio == zio->io_logical) {
4366 * Determine whether zio should be reexecuted. This will
4367 * propagate all the way to the root via zio_notify_parent().
4369 ASSERT(zio->io_vd == NULL && zio->io_bp != NULL);
4370 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
4372 if (IO_IS_ALLOCATING(zio) &&
4373 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
4374 if (zio->io_error != ENOSPC)
4375 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
4377 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4380 if ((zio->io_type == ZIO_TYPE_READ ||
4381 zio->io_type == ZIO_TYPE_FREE) &&
4382 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
4383 zio->io_error == ENXIO &&
4384 spa_load_state(zio->io_spa) == SPA_LOAD_NONE &&
4385 spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE)
4386 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4388 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
4389 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4392 * Here is a possibly good place to attempt to do
4393 * either combinatorial reconstruction or error correction
4394 * based on checksums. It also might be a good place
4395 * to send out preliminary ereports before we suspend
4401 * If there were logical child errors, they apply to us now.
4402 * We defer this until now to avoid conflating logical child
4403 * errors with errors that happened to the zio itself when
4404 * updating vdev stats and reporting FMA events above.
4406 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
4408 if ((zio->io_error || zio->io_reexecute) &&
4409 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
4410 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
4411 zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp);
4413 zio_gang_tree_free(&zio->io_gang_tree);
4416 * Godfather I/Os should never suspend.
4418 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
4419 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
4420 zio->io_reexecute &= ~ZIO_REEXECUTE_SUSPEND;
4422 if (zio->io_reexecute) {
4424 * This is a logical I/O that wants to reexecute.
4426 * Reexecute is top-down. When an i/o fails, if it's not
4427 * the root, it simply notifies its parent and sticks around.
4428 * The parent, seeing that it still has children in zio_done(),
4429 * does the same. This percolates all the way up to the root.
4430 * The root i/o will reexecute or suspend the entire tree.
4432 * This approach ensures that zio_reexecute() honors
4433 * all the original i/o dependency relationships, e.g.
4434 * parents not executing until children are ready.
4436 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
4438 zio->io_gang_leader = NULL;
4440 mutex_enter(&zio->io_lock);
4441 zio->io_state[ZIO_WAIT_DONE] = 1;
4442 mutex_exit(&zio->io_lock);
4445 * "The Godfather" I/O monitors its children but is
4446 * not a true parent to them. It will track them through
4447 * the pipeline but severs its ties whenever they get into
4448 * trouble (e.g. suspended). This allows "The Godfather"
4449 * I/O to return status without blocking.
4452 for (pio = zio_walk_parents(zio, &zl); pio != NULL;
4454 zio_link_t *remove_zl = zl;
4455 pio_next = zio_walk_parents(zio, &zl);
4457 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
4458 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
4459 zio_remove_child(pio, zio, remove_zl);
4460 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
4464 if ((pio = zio_unique_parent(zio)) != NULL) {
4466 * We're not a root i/o, so there's nothing to do
4467 * but notify our parent. Don't propagate errors
4468 * upward since we haven't permanently failed yet.
4470 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
4471 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
4472 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
4473 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
4475 * We'd fail again if we reexecuted now, so suspend
4476 * until conditions improve (e.g. device comes online).
4478 zio_suspend(zio->io_spa, zio, ZIO_SUSPEND_IOERR);
4481 * Reexecution is potentially a huge amount of work.
4482 * Hand it off to the otherwise-unused claim taskq.
4484 ASSERT(taskq_empty_ent(&zio->io_tqent));
4485 spa_taskq_dispatch_ent(zio->io_spa,
4486 ZIO_TYPE_CLAIM, ZIO_TASKQ_ISSUE,
4487 (task_func_t *)zio_reexecute, zio, 0,
4490 return (ZIO_PIPELINE_STOP);
4493 ASSERT(zio->io_child_count == 0);
4494 ASSERT(zio->io_reexecute == 0);
4495 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
4498 * Report any checksum errors, since the I/O is complete.
4500 while (zio->io_cksum_report != NULL) {
4501 zio_cksum_report_t *zcr = zio->io_cksum_report;
4502 zio->io_cksum_report = zcr->zcr_next;
4503 zcr->zcr_next = NULL;
4504 zcr->zcr_finish(zcr, NULL);
4505 zfs_ereport_free_checksum(zcr);
4508 if (zio->io_flags & ZIO_FLAG_FASTWRITE && zio->io_bp &&
4509 !BP_IS_HOLE(zio->io_bp) && !BP_IS_EMBEDDED(zio->io_bp) &&
4510 !(zio->io_flags & ZIO_FLAG_NOPWRITE)) {
4511 metaslab_fastwrite_unmark(zio->io_spa, zio->io_bp);
4515 * It is the responsibility of the done callback to ensure that this
4516 * particular zio is no longer discoverable for adoption, and as
4517 * such, cannot acquire any new parents.
4522 mutex_enter(&zio->io_lock);
4523 zio->io_state[ZIO_WAIT_DONE] = 1;
4524 mutex_exit(&zio->io_lock);
4527 for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) {
4528 zio_link_t *remove_zl = zl;
4529 pio_next = zio_walk_parents(zio, &zl);
4530 zio_remove_child(pio, zio, remove_zl);
4531 zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
4534 if (zio->io_waiter != NULL) {
4535 mutex_enter(&zio->io_lock);
4536 zio->io_executor = NULL;
4537 cv_broadcast(&zio->io_cv);
4538 mutex_exit(&zio->io_lock);
4543 return (ZIO_PIPELINE_STOP);
4547 * ==========================================================================
4548 * I/O pipeline definition
4549 * ==========================================================================
4551 static zio_pipe_stage_t *zio_pipeline[] = {
4559 zio_checksum_generate,
4575 zio_checksum_verify,
4583 * Compare two zbookmark_phys_t's to see which we would reach first in a
4584 * pre-order traversal of the object tree.
4586 * This is simple in every case aside from the meta-dnode object. For all other
4587 * objects, we traverse them in order (object 1 before object 2, and so on).
4588 * However, all of these objects are traversed while traversing object 0, since
4589 * the data it points to is the list of objects. Thus, we need to convert to a
4590 * canonical representation so we can compare meta-dnode bookmarks to
4591 * non-meta-dnode bookmarks.
4593 * We do this by calculating "equivalents" for each field of the zbookmark.
4594 * zbookmarks outside of the meta-dnode use their own object and level, and
4595 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
4596 * blocks this bookmark refers to) by multiplying their blkid by their span
4597 * (the number of L0 blocks contained within one block at their level).
4598 * zbookmarks inside the meta-dnode calculate their object equivalent
4599 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
4600 * level + 1<<31 (any value larger than a level could ever be) for their level.
4601 * This causes them to always compare before a bookmark in their object
4602 * equivalent, compare appropriately to bookmarks in other objects, and to
4603 * compare appropriately to other bookmarks in the meta-dnode.
4606 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
4607 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
4610 * These variables represent the "equivalent" values for the zbookmark,
4611 * after converting zbookmarks inside the meta dnode to their
4612 * normal-object equivalents.
4614 uint64_t zb1obj, zb2obj;
4615 uint64_t zb1L0, zb2L0;
4616 uint64_t zb1level, zb2level;
4618 if (zb1->zb_object == zb2->zb_object &&
4619 zb1->zb_level == zb2->zb_level &&
4620 zb1->zb_blkid == zb2->zb_blkid)
4624 * BP_SPANB calculates the span in blocks.
4626 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
4627 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
4629 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
4630 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4632 zb1level = zb1->zb_level + COMPARE_META_LEVEL;
4634 zb1obj = zb1->zb_object;
4635 zb1level = zb1->zb_level;
4638 if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
4639 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4641 zb2level = zb2->zb_level + COMPARE_META_LEVEL;
4643 zb2obj = zb2->zb_object;
4644 zb2level = zb2->zb_level;
4647 /* Now that we have a canonical representation, do the comparison. */
4648 if (zb1obj != zb2obj)
4649 return (zb1obj < zb2obj ? -1 : 1);
4650 else if (zb1L0 != zb2L0)
4651 return (zb1L0 < zb2L0 ? -1 : 1);
4652 else if (zb1level != zb2level)
4653 return (zb1level > zb2level ? -1 : 1);
4655 * This can (theoretically) happen if the bookmarks have the same object
4656 * and level, but different blkids, if the block sizes are not the same.
4657 * There is presently no way to change the indirect block sizes
4663 * This function checks the following: given that last_block is the place that
4664 * our traversal stopped last time, does that guarantee that we've visited
4665 * every node under subtree_root? Therefore, we can't just use the raw output
4666 * of zbookmark_compare. We have to pass in a modified version of
4667 * subtree_root; by incrementing the block id, and then checking whether
4668 * last_block is before or equal to that, we can tell whether or not having
4669 * visited last_block implies that all of subtree_root's children have been
4673 zbookmark_subtree_completed(const dnode_phys_t *dnp,
4674 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
4676 zbookmark_phys_t mod_zb = *subtree_root;
4678 ASSERT(last_block->zb_level == 0);
4680 /* The objset_phys_t isn't before anything. */
4685 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
4686 * data block size in sectors, because that variable is only used if
4687 * the bookmark refers to a block in the meta-dnode. Since we don't
4688 * know without examining it what object it refers to, and there's no
4689 * harm in passing in this value in other cases, we always pass it in.
4691 * We pass in 0 for the indirect block size shift because zb2 must be
4692 * level 0. The indirect block size is only used to calculate the span
4693 * of the bookmark, but since the bookmark must be level 0, the span is
4694 * always 1, so the math works out.
4696 * If you make changes to how the zbookmark_compare code works, be sure
4697 * to make sure that this code still works afterwards.
4699 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
4700 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,
4704 #if defined(_KERNEL)
4705 EXPORT_SYMBOL(zio_type_name);
4706 EXPORT_SYMBOL(zio_buf_alloc);
4707 EXPORT_SYMBOL(zio_data_buf_alloc);
4708 EXPORT_SYMBOL(zio_buf_free);
4709 EXPORT_SYMBOL(zio_data_buf_free);
4711 module_param(zio_delay_max, int, 0644);
4712 MODULE_PARM_DESC(zio_delay_max, "Max zio millisec delay before posting event");
4714 module_param(zio_requeue_io_start_cut_in_line, int, 0644);
4715 MODULE_PARM_DESC(zio_requeue_io_start_cut_in_line, "Prioritize requeued I/O");
4717 module_param(zfs_sync_pass_deferred_free, int, 0644);
4718 MODULE_PARM_DESC(zfs_sync_pass_deferred_free,
4719 "Defer frees starting in this pass");
4721 module_param(zfs_sync_pass_dont_compress, int, 0644);
4722 MODULE_PARM_DESC(zfs_sync_pass_dont_compress,
4723 "Don't compress starting in this pass");
4725 module_param(zfs_sync_pass_rewrite, int, 0644);
4726 MODULE_PARM_DESC(zfs_sync_pass_rewrite,
4727 "Rewrite new bps starting in this pass");
4729 module_param(zio_dva_throttle_enabled, int, 0644);
4730 MODULE_PARM_DESC(zio_dva_throttle_enabled,
4731 "Throttle block allocations in the ZIO pipeline");