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 https://opensource.org/licenses/CDDL-1.0.
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, 2022 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
25 * Copyright (c) 2017, Intel Corporation.
26 * Copyright (c) 2019, Klara Inc.
27 * Copyright (c) 2019, Allan Jude
28 * Copyright (c) 2021, Datto, Inc.
31 #include <sys/sysmacros.h>
32 #include <sys/zfs_context.h>
33 #include <sys/fm/fs/zfs.h>
36 #include <sys/spa_impl.h>
37 #include <sys/vdev_impl.h>
38 #include <sys/vdev_trim.h>
39 #include <sys/zio_impl.h>
40 #include <sys/zio_compress.h>
41 #include <sys/zio_checksum.h>
42 #include <sys/dmu_objset.h>
45 #include <sys/blkptr.h>
46 #include <sys/zfeature.h>
47 #include <sys/dsl_scan.h>
48 #include <sys/metaslab_impl.h>
50 #include <sys/trace_zfs.h>
52 #include <sys/dsl_crypt.h>
56 * ==========================================================================
57 * I/O type descriptions
58 * ==========================================================================
60 const char *const zio_type_name[ZIO_TYPES] = {
62 * Note: Linux kernel thread name length is limited
63 * so these names will differ from upstream open zfs.
65 "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl", "z_trim"
68 int zio_dva_throttle_enabled = B_TRUE;
69 static int zio_deadman_log_all = B_FALSE;
72 * ==========================================================================
74 * ==========================================================================
76 static kmem_cache_t *zio_cache;
77 static kmem_cache_t *zio_link_cache;
78 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
79 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
80 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
81 static uint64_t zio_buf_cache_allocs[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
82 static uint64_t zio_buf_cache_frees[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
85 /* Mark IOs as "slow" if they take longer than 30 seconds */
86 static uint_t zio_slow_io_ms = (30 * MILLISEC);
88 #define BP_SPANB(indblkshift, level) \
89 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
90 #define COMPARE_META_LEVEL 0x80000000ul
92 * The following actions directly effect the spa's sync-to-convergence logic.
93 * The values below define the sync pass when we start performing the action.
94 * Care should be taken when changing these values as they directly impact
95 * spa_sync() performance. Tuning these values may introduce subtle performance
96 * pathologies and should only be done in the context of performance analysis.
97 * These tunables will eventually be removed and replaced with #defines once
98 * enough analysis has been done to determine optimal values.
100 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
101 * regular blocks are not deferred.
103 * Starting in sync pass 8 (zfs_sync_pass_dont_compress), we disable
104 * compression (including of metadata). In practice, we don't have this
105 * many sync passes, so this has no effect.
107 * The original intent was that disabling compression would help the sync
108 * passes to converge. However, in practice disabling compression increases
109 * the average number of sync passes, because when we turn compression off, a
110 * lot of block's size will change and thus we have to re-allocate (not
111 * overwrite) them. It also increases the number of 128KB allocations (e.g.
112 * for indirect blocks and spacemaps) because these will not be compressed.
113 * The 128K allocations are especially detrimental to performance on highly
114 * fragmented systems, which may have very few free segments of this size,
115 * and may need to load new metaslabs to satisfy 128K allocations.
118 /* defer frees starting in this pass */
119 uint_t zfs_sync_pass_deferred_free = 2;
121 /* don't compress starting in this pass */
122 static uint_t zfs_sync_pass_dont_compress = 8;
124 /* rewrite new bps starting in this pass */
125 static uint_t zfs_sync_pass_rewrite = 2;
128 * An allocating zio is one that either currently has the DVA allocate
129 * stage set or will have it later in its lifetime.
131 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
134 * Enable smaller cores by excluding metadata
135 * allocations as well.
137 int zio_exclude_metadata = 0;
138 static int zio_requeue_io_start_cut_in_line = 1;
141 static const int zio_buf_debug_limit = 16384;
143 static const int zio_buf_debug_limit = 0;
146 static inline void __zio_execute(zio_t *zio);
148 static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t);
155 zio_cache = kmem_cache_create("zio_cache",
156 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
157 zio_link_cache = kmem_cache_create("zio_link_cache",
158 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
161 * For small buffers, we want a cache for each multiple of
162 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
163 * for each quarter-power of 2.
165 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
166 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
169 size_t data_cflags, cflags;
171 data_cflags = KMC_NODEBUG;
172 cflags = (zio_exclude_metadata || size > zio_buf_debug_limit) ?
180 * If we are using watchpoints, put each buffer on its own page,
181 * to eliminate the performance overhead of trapping to the
182 * kernel when modifying a non-watched buffer that shares the
183 * page with a watched buffer.
185 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
188 * Here's the problem - on 4K native devices in userland on
189 * Linux using O_DIRECT, buffers must be 4K aligned or I/O
190 * will fail with EINVAL, causing zdb (and others) to coredump.
191 * Since userland probably doesn't need optimized buffer caches,
192 * we just force 4K alignment on everything.
194 align = 8 * SPA_MINBLOCKSIZE;
196 if (size < PAGESIZE) {
197 align = SPA_MINBLOCKSIZE;
198 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
205 if (cflags == data_cflags) {
207 * Resulting kmem caches would be identical.
208 * Save memory by creating only one.
210 (void) snprintf(name, sizeof (name),
211 "zio_buf_comb_%lu", (ulong_t)size);
212 zio_buf_cache[c] = kmem_cache_create(name,
213 size, align, NULL, NULL, NULL, NULL, NULL,
215 zio_data_buf_cache[c] = zio_buf_cache[c];
218 (void) snprintf(name, sizeof (name), "zio_buf_%lu",
220 zio_buf_cache[c] = kmem_cache_create(name, size,
221 align, NULL, NULL, NULL, NULL, NULL, cflags);
223 (void) snprintf(name, sizeof (name), "zio_data_buf_%lu",
225 zio_data_buf_cache[c] = kmem_cache_create(name, size,
226 align, NULL, NULL, NULL, NULL, NULL, data_cflags);
231 ASSERT(zio_buf_cache[c] != NULL);
232 if (zio_buf_cache[c - 1] == NULL)
233 zio_buf_cache[c - 1] = zio_buf_cache[c];
235 ASSERT(zio_data_buf_cache[c] != NULL);
236 if (zio_data_buf_cache[c - 1] == NULL)
237 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
248 size_t n = SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT;
250 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
251 for (size_t i = 0; i < n; i++) {
252 if (zio_buf_cache_allocs[i] != zio_buf_cache_frees[i])
253 (void) printf("zio_fini: [%d] %llu != %llu\n",
254 (int)((i + 1) << SPA_MINBLOCKSHIFT),
255 (long long unsigned)zio_buf_cache_allocs[i],
256 (long long unsigned)zio_buf_cache_frees[i]);
261 * The same kmem cache can show up multiple times in both zio_buf_cache
262 * and zio_data_buf_cache. Do a wasteful but trivially correct scan to
265 for (size_t i = 0; i < n; i++) {
266 kmem_cache_t *cache = zio_buf_cache[i];
269 for (size_t j = i; j < n; j++) {
270 if (cache == zio_buf_cache[j])
271 zio_buf_cache[j] = NULL;
272 if (cache == zio_data_buf_cache[j])
273 zio_data_buf_cache[j] = NULL;
275 kmem_cache_destroy(cache);
278 for (size_t i = 0; i < n; i++) {
279 kmem_cache_t *cache = zio_data_buf_cache[i];
282 for (size_t j = i; j < n; j++) {
283 if (cache == zio_data_buf_cache[j])
284 zio_data_buf_cache[j] = NULL;
286 kmem_cache_destroy(cache);
289 for (size_t i = 0; i < n; i++) {
290 VERIFY3P(zio_buf_cache[i], ==, NULL);
291 VERIFY3P(zio_data_buf_cache[i], ==, NULL);
294 kmem_cache_destroy(zio_link_cache);
295 kmem_cache_destroy(zio_cache);
303 * ==========================================================================
304 * Allocate and free I/O buffers
305 * ==========================================================================
309 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
310 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
311 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
312 * excess / transient data in-core during a crashdump.
315 zio_buf_alloc(size_t size)
317 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
319 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
320 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
321 atomic_add_64(&zio_buf_cache_allocs[c], 1);
324 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
328 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
329 * crashdump if the kernel panics. This exists so that we will limit the amount
330 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
331 * of kernel heap dumped to disk when the kernel panics)
334 zio_data_buf_alloc(size_t size)
336 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
338 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
340 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
344 zio_buf_free(void *buf, size_t size)
346 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
348 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
349 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
350 atomic_add_64(&zio_buf_cache_frees[c], 1);
353 kmem_cache_free(zio_buf_cache[c], buf);
357 zio_data_buf_free(void *buf, size_t size)
359 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
361 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
363 kmem_cache_free(zio_data_buf_cache[c], buf);
367 zio_abd_free(void *abd, size_t size)
370 abd_free((abd_t *)abd);
374 * ==========================================================================
375 * Push and pop I/O transform buffers
376 * ==========================================================================
379 zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize,
380 zio_transform_func_t *transform)
382 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
384 zt->zt_orig_abd = zio->io_abd;
385 zt->zt_orig_size = zio->io_size;
386 zt->zt_bufsize = bufsize;
387 zt->zt_transform = transform;
389 zt->zt_next = zio->io_transform_stack;
390 zio->io_transform_stack = zt;
397 zio_pop_transforms(zio_t *zio)
401 while ((zt = zio->io_transform_stack) != NULL) {
402 if (zt->zt_transform != NULL)
403 zt->zt_transform(zio,
404 zt->zt_orig_abd, zt->zt_orig_size);
406 if (zt->zt_bufsize != 0)
407 abd_free(zio->io_abd);
409 zio->io_abd = zt->zt_orig_abd;
410 zio->io_size = zt->zt_orig_size;
411 zio->io_transform_stack = zt->zt_next;
413 kmem_free(zt, sizeof (zio_transform_t));
418 * ==========================================================================
419 * I/O transform callbacks for subblocks, decompression, and decryption
420 * ==========================================================================
423 zio_subblock(zio_t *zio, abd_t *data, uint64_t size)
425 ASSERT(zio->io_size > size);
427 if (zio->io_type == ZIO_TYPE_READ)
428 abd_copy(data, zio->io_abd, size);
432 zio_decompress(zio_t *zio, abd_t *data, uint64_t size)
434 if (zio->io_error == 0) {
435 void *tmp = abd_borrow_buf(data, size);
436 int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
437 zio->io_abd, tmp, zio->io_size, size,
438 &zio->io_prop.zp_complevel);
439 abd_return_buf_copy(data, tmp, size);
441 if (zio_injection_enabled && ret == 0)
442 ret = zio_handle_fault_injection(zio, EINVAL);
445 zio->io_error = SET_ERROR(EIO);
450 zio_decrypt(zio_t *zio, abd_t *data, uint64_t size)
454 blkptr_t *bp = zio->io_bp;
455 spa_t *spa = zio->io_spa;
456 uint64_t dsobj = zio->io_bookmark.zb_objset;
457 uint64_t lsize = BP_GET_LSIZE(bp);
458 dmu_object_type_t ot = BP_GET_TYPE(bp);
459 uint8_t salt[ZIO_DATA_SALT_LEN];
460 uint8_t iv[ZIO_DATA_IV_LEN];
461 uint8_t mac[ZIO_DATA_MAC_LEN];
462 boolean_t no_crypt = B_FALSE;
464 ASSERT(BP_USES_CRYPT(bp));
465 ASSERT3U(size, !=, 0);
467 if (zio->io_error != 0)
471 * Verify the cksum of MACs stored in an indirect bp. It will always
472 * be possible to verify this since it does not require an encryption
475 if (BP_HAS_INDIRECT_MAC_CKSUM(bp)) {
476 zio_crypt_decode_mac_bp(bp, mac);
478 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF) {
480 * We haven't decompressed the data yet, but
481 * zio_crypt_do_indirect_mac_checksum() requires
482 * decompressed data to be able to parse out the MACs
483 * from the indirect block. We decompress it now and
484 * throw away the result after we are finished.
486 tmp = zio_buf_alloc(lsize);
487 ret = zio_decompress_data(BP_GET_COMPRESS(bp),
488 zio->io_abd, tmp, zio->io_size, lsize,
489 &zio->io_prop.zp_complevel);
491 ret = SET_ERROR(EIO);
494 ret = zio_crypt_do_indirect_mac_checksum(B_FALSE,
495 tmp, lsize, BP_SHOULD_BYTESWAP(bp), mac);
496 zio_buf_free(tmp, lsize);
498 ret = zio_crypt_do_indirect_mac_checksum_abd(B_FALSE,
499 zio->io_abd, size, BP_SHOULD_BYTESWAP(bp), mac);
501 abd_copy(data, zio->io_abd, size);
503 if (zio_injection_enabled && ot != DMU_OT_DNODE && ret == 0) {
504 ret = zio_handle_decrypt_injection(spa,
505 &zio->io_bookmark, ot, ECKSUM);
514 * If this is an authenticated block, just check the MAC. It would be
515 * nice to separate this out into its own flag, but when this was done,
516 * we had run out of bits in what is now zio_flag_t. Future cleanup
517 * could make this a flag bit.
519 if (BP_IS_AUTHENTICATED(bp)) {
520 if (ot == DMU_OT_OBJSET) {
521 ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa,
522 dsobj, zio->io_abd, size, BP_SHOULD_BYTESWAP(bp));
524 zio_crypt_decode_mac_bp(bp, mac);
525 ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj,
526 zio->io_abd, size, mac);
527 if (zio_injection_enabled && ret == 0) {
528 ret = zio_handle_decrypt_injection(spa,
529 &zio->io_bookmark, ot, ECKSUM);
532 abd_copy(data, zio->io_abd, size);
540 zio_crypt_decode_params_bp(bp, salt, iv);
542 if (ot == DMU_OT_INTENT_LOG) {
543 tmp = abd_borrow_buf_copy(zio->io_abd, sizeof (zil_chain_t));
544 zio_crypt_decode_mac_zil(tmp, mac);
545 abd_return_buf(zio->io_abd, tmp, sizeof (zil_chain_t));
547 zio_crypt_decode_mac_bp(bp, mac);
550 ret = spa_do_crypt_abd(B_FALSE, spa, &zio->io_bookmark, BP_GET_TYPE(bp),
551 BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp), salt, iv, mac, size, data,
552 zio->io_abd, &no_crypt);
554 abd_copy(data, zio->io_abd, size);
562 /* assert that the key was found unless this was speculative */
563 ASSERT(ret != EACCES || (zio->io_flags & ZIO_FLAG_SPECULATIVE));
566 * If there was a decryption / authentication error return EIO as
567 * the io_error. If this was not a speculative zio, create an ereport.
570 zio->io_error = SET_ERROR(EIO);
571 if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) {
572 spa_log_error(spa, &zio->io_bookmark);
573 (void) zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION,
574 spa, NULL, &zio->io_bookmark, zio, 0);
582 * ==========================================================================
583 * I/O parent/child relationships and pipeline interlocks
584 * ==========================================================================
587 zio_walk_parents(zio_t *cio, zio_link_t **zl)
589 list_t *pl = &cio->io_parent_list;
591 *zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl);
595 ASSERT((*zl)->zl_child == cio);
596 return ((*zl)->zl_parent);
600 zio_walk_children(zio_t *pio, zio_link_t **zl)
602 list_t *cl = &pio->io_child_list;
604 ASSERT(MUTEX_HELD(&pio->io_lock));
606 *zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl);
610 ASSERT((*zl)->zl_parent == pio);
611 return ((*zl)->zl_child);
615 zio_unique_parent(zio_t *cio)
617 zio_link_t *zl = NULL;
618 zio_t *pio = zio_walk_parents(cio, &zl);
620 VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL);
625 zio_add_child(zio_t *pio, zio_t *cio)
627 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
630 * Logical I/Os can have logical, gang, or vdev children.
631 * Gang I/Os can have gang or vdev children.
632 * Vdev I/Os can only have vdev children.
633 * The following ASSERT captures all of these constraints.
635 ASSERT3S(cio->io_child_type, <=, pio->io_child_type);
640 mutex_enter(&pio->io_lock);
641 mutex_enter(&cio->io_lock);
643 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
645 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
646 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
648 list_insert_head(&pio->io_child_list, zl);
649 list_insert_head(&cio->io_parent_list, zl);
651 pio->io_child_count++;
652 cio->io_parent_count++;
654 mutex_exit(&cio->io_lock);
655 mutex_exit(&pio->io_lock);
659 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
661 ASSERT(zl->zl_parent == pio);
662 ASSERT(zl->zl_child == cio);
664 mutex_enter(&pio->io_lock);
665 mutex_enter(&cio->io_lock);
667 list_remove(&pio->io_child_list, zl);
668 list_remove(&cio->io_parent_list, zl);
670 pio->io_child_count--;
671 cio->io_parent_count--;
673 mutex_exit(&cio->io_lock);
674 mutex_exit(&pio->io_lock);
675 kmem_cache_free(zio_link_cache, zl);
679 zio_wait_for_children(zio_t *zio, uint8_t childbits, enum zio_wait_type wait)
681 boolean_t waiting = B_FALSE;
683 mutex_enter(&zio->io_lock);
684 ASSERT(zio->io_stall == NULL);
685 for (int c = 0; c < ZIO_CHILD_TYPES; c++) {
686 if (!(ZIO_CHILD_BIT_IS_SET(childbits, c)))
689 uint64_t *countp = &zio->io_children[c][wait];
692 ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN);
693 zio->io_stall = countp;
698 mutex_exit(&zio->io_lock);
702 __attribute__((always_inline))
704 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait,
705 zio_t **next_to_executep)
707 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
708 int *errorp = &pio->io_child_error[zio->io_child_type];
710 mutex_enter(&pio->io_lock);
711 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
712 *errorp = zio_worst_error(*errorp, zio->io_error);
713 pio->io_reexecute |= zio->io_reexecute;
714 ASSERT3U(*countp, >, 0);
718 if (*countp == 0 && pio->io_stall == countp) {
719 zio_taskq_type_t type =
720 pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE :
722 pio->io_stall = NULL;
723 mutex_exit(&pio->io_lock);
726 * If we can tell the caller to execute this parent next, do
727 * so. We only do this if the parent's zio type matches the
728 * child's type. Otherwise dispatch the parent zio in its
731 * Having the caller execute the parent when possible reduces
732 * locking on the zio taskq's, reduces context switch
733 * overhead, and has no recursion penalty. Note that one
734 * read from disk typically causes at least 3 zio's: a
735 * zio_null(), the logical zio_read(), and then a physical
736 * zio. When the physical ZIO completes, we are able to call
737 * zio_done() on all 3 of these zio's from one invocation of
738 * zio_execute() by returning the parent back to
739 * zio_execute(). Since the parent isn't executed until this
740 * thread returns back to zio_execute(), the caller should do
743 * In other cases, dispatching the parent prevents
744 * overflowing the stack when we have deeply nested
745 * parent-child relationships, as we do with the "mega zio"
746 * of writes for spa_sync(), and the chain of ZIL blocks.
748 if (next_to_executep != NULL && *next_to_executep == NULL &&
749 pio->io_type == zio->io_type) {
750 *next_to_executep = pio;
752 zio_taskq_dispatch(pio, type, B_FALSE);
755 mutex_exit(&pio->io_lock);
760 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
762 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
763 zio->io_error = zio->io_child_error[c];
767 zio_bookmark_compare(const void *x1, const void *x2)
769 const zio_t *z1 = x1;
770 const zio_t *z2 = x2;
772 if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset)
774 if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset)
777 if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object)
779 if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object)
782 if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level)
784 if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level)
787 if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid)
789 if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid)
801 * ==========================================================================
802 * Create the various types of I/O (read, write, free, etc)
803 * ==========================================================================
806 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
807 abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done,
808 void *private, zio_type_t type, zio_priority_t priority,
809 zio_flag_t flags, vdev_t *vd, uint64_t offset,
810 const zbookmark_phys_t *zb, enum zio_stage stage,
811 enum zio_stage pipeline)
815 IMPLY(type != ZIO_TYPE_TRIM, psize <= SPA_MAXBLOCKSIZE);
816 ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0);
817 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
819 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
820 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
821 ASSERT(vd || stage == ZIO_STAGE_OPEN);
823 IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW_COMPRESS) != 0);
825 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
826 memset(zio, 0, sizeof (zio_t));
828 mutex_init(&zio->io_lock, NULL, MUTEX_NOLOCKDEP, NULL);
829 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
831 list_create(&zio->io_parent_list, sizeof (zio_link_t),
832 offsetof(zio_link_t, zl_parent_node));
833 list_create(&zio->io_child_list, sizeof (zio_link_t),
834 offsetof(zio_link_t, zl_child_node));
835 metaslab_trace_init(&zio->io_alloc_list);
838 zio->io_child_type = ZIO_CHILD_VDEV;
839 else if (flags & ZIO_FLAG_GANG_CHILD)
840 zio->io_child_type = ZIO_CHILD_GANG;
841 else if (flags & ZIO_FLAG_DDT_CHILD)
842 zio->io_child_type = ZIO_CHILD_DDT;
844 zio->io_child_type = ZIO_CHILD_LOGICAL;
847 zio->io_bp = (blkptr_t *)bp;
848 zio->io_bp_copy = *bp;
849 zio->io_bp_orig = *bp;
850 if (type != ZIO_TYPE_WRITE ||
851 zio->io_child_type == ZIO_CHILD_DDT)
852 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
853 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
854 zio->io_logical = zio;
855 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
856 pipeline |= ZIO_GANG_STAGES;
862 zio->io_private = private;
864 zio->io_priority = priority;
866 zio->io_offset = offset;
867 zio->io_orig_abd = zio->io_abd = data;
868 zio->io_orig_size = zio->io_size = psize;
869 zio->io_lsize = lsize;
870 zio->io_orig_flags = zio->io_flags = flags;
871 zio->io_orig_stage = zio->io_stage = stage;
872 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
873 zio->io_pipeline_trace = ZIO_STAGE_OPEN;
875 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
876 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
879 zio->io_bookmark = *zb;
882 zio->io_metaslab_class = pio->io_metaslab_class;
883 if (zio->io_logical == NULL)
884 zio->io_logical = pio->io_logical;
885 if (zio->io_child_type == ZIO_CHILD_GANG)
886 zio->io_gang_leader = pio->io_gang_leader;
887 zio_add_child(pio, zio);
890 taskq_init_ent(&zio->io_tqent);
896 zio_destroy(zio_t *zio)
898 metaslab_trace_fini(&zio->io_alloc_list);
899 list_destroy(&zio->io_parent_list);
900 list_destroy(&zio->io_child_list);
901 mutex_destroy(&zio->io_lock);
902 cv_destroy(&zio->io_cv);
903 kmem_cache_free(zio_cache, zio);
907 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
908 void *private, zio_flag_t flags)
912 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
913 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
914 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
920 zio_root(spa_t *spa, zio_done_func_t *done, void *private, zio_flag_t flags)
922 return (zio_null(NULL, spa, NULL, done, private, flags));
926 zfs_blkptr_verify_log(spa_t *spa, const blkptr_t *bp,
927 enum blk_verify_flag blk_verify, const char *fmt, ...)
933 (void) vsnprintf(buf, sizeof (buf), fmt, adx);
936 switch (blk_verify) {
937 case BLK_VERIFY_HALT:
938 dprintf_bp(bp, "blkptr at %p dprintf_bp():", bp);
939 zfs_panic_recover("%s: %s", spa_name(spa), buf);
942 zfs_dbgmsg("%s: %s", spa_name(spa), buf);
944 case BLK_VERIFY_ONLY:
952 * Verify the block pointer fields contain reasonable values. This means
953 * it only contains known object types, checksum/compression identifiers,
954 * block sizes within the maximum allowed limits, valid DVAs, etc.
956 * If everything checks out B_TRUE is returned. The zfs_blkptr_verify
957 * argument controls the behavior when an invalid field is detected.
959 * Modes for zfs_blkptr_verify:
960 * 1) BLK_VERIFY_ONLY (evaluate the block)
961 * 2) BLK_VERIFY_LOG (evaluate the block and log problems)
962 * 3) BLK_VERIFY_HALT (call zfs_panic_recover on error)
965 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp, boolean_t config_held,
966 enum blk_verify_flag blk_verify)
970 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
971 errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
972 "blkptr at %p has invalid TYPE %llu",
973 bp, (longlong_t)BP_GET_TYPE(bp));
975 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS) {
976 errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
977 "blkptr at %p has invalid CHECKSUM %llu",
978 bp, (longlong_t)BP_GET_CHECKSUM(bp));
980 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS) {
981 errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
982 "blkptr at %p has invalid COMPRESS %llu",
983 bp, (longlong_t)BP_GET_COMPRESS(bp));
985 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
986 errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
987 "blkptr at %p has invalid LSIZE %llu",
988 bp, (longlong_t)BP_GET_LSIZE(bp));
990 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
991 errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
992 "blkptr at %p has invalid PSIZE %llu",
993 bp, (longlong_t)BP_GET_PSIZE(bp));
996 if (BP_IS_EMBEDDED(bp)) {
997 if (BPE_GET_ETYPE(bp) >= NUM_BP_EMBEDDED_TYPES) {
998 errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
999 "blkptr at %p has invalid ETYPE %llu",
1000 bp, (longlong_t)BPE_GET_ETYPE(bp));
1005 * Do not verify individual DVAs if the config is not trusted. This
1006 * will be done once the zio is executed in vdev_mirror_map_alloc.
1008 if (!spa->spa_trust_config)
1009 return (errors == 0);
1012 spa_config_enter(spa, SCL_VDEV, bp, RW_READER);
1014 ASSERT(spa_config_held(spa, SCL_VDEV, RW_WRITER));
1016 * Pool-specific checks.
1018 * Note: it would be nice to verify that the blk_birth and
1019 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
1020 * allows the birth time of log blocks (and dmu_sync()-ed blocks
1021 * that are in the log) to be arbitrarily large.
1023 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
1024 const dva_t *dva = &bp->blk_dva[i];
1025 uint64_t vdevid = DVA_GET_VDEV(dva);
1027 if (vdevid >= spa->spa_root_vdev->vdev_children) {
1028 errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1029 "blkptr at %p DVA %u has invalid VDEV %llu",
1030 bp, i, (longlong_t)vdevid);
1033 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
1035 errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1036 "blkptr at %p DVA %u has invalid VDEV %llu",
1037 bp, i, (longlong_t)vdevid);
1040 if (vd->vdev_ops == &vdev_hole_ops) {
1041 errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1042 "blkptr at %p DVA %u has hole VDEV %llu",
1043 bp, i, (longlong_t)vdevid);
1046 if (vd->vdev_ops == &vdev_missing_ops) {
1048 * "missing" vdevs are valid during import, but we
1049 * don't have their detailed info (e.g. asize), so
1050 * we can't perform any more checks on them.
1054 uint64_t offset = DVA_GET_OFFSET(dva);
1055 uint64_t asize = DVA_GET_ASIZE(dva);
1056 if (DVA_GET_GANG(dva))
1057 asize = vdev_gang_header_asize(vd);
1058 if (offset + asize > vd->vdev_asize) {
1059 errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1060 "blkptr at %p DVA %u has invalid OFFSET %llu",
1061 bp, i, (longlong_t)offset);
1065 dprintf_bp(bp, "blkptr at %p dprintf_bp():", bp);
1067 spa_config_exit(spa, SCL_VDEV, bp);
1069 return (errors == 0);
1073 zfs_dva_valid(spa_t *spa, const dva_t *dva, const blkptr_t *bp)
1076 uint64_t vdevid = DVA_GET_VDEV(dva);
1078 if (vdevid >= spa->spa_root_vdev->vdev_children)
1081 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
1085 if (vd->vdev_ops == &vdev_hole_ops)
1088 if (vd->vdev_ops == &vdev_missing_ops) {
1092 uint64_t offset = DVA_GET_OFFSET(dva);
1093 uint64_t asize = DVA_GET_ASIZE(dva);
1095 if (DVA_GET_GANG(dva))
1096 asize = vdev_gang_header_asize(vd);
1097 if (offset + asize > vd->vdev_asize)
1104 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
1105 abd_t *data, uint64_t size, zio_done_func_t *done, void *private,
1106 zio_priority_t priority, zio_flag_t flags, const zbookmark_phys_t *zb)
1110 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
1111 data, size, size, done, private,
1112 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
1113 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
1114 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
1120 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
1121 abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp,
1122 zio_done_func_t *ready, zio_done_func_t *children_ready,
1123 zio_done_func_t *physdone, zio_done_func_t *done,
1124 void *private, zio_priority_t priority, zio_flag_t flags,
1125 const zbookmark_phys_t *zb)
1129 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
1130 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
1131 zp->zp_compress >= ZIO_COMPRESS_OFF &&
1132 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
1133 DMU_OT_IS_VALID(zp->zp_type) &&
1134 zp->zp_level < 32 &&
1135 zp->zp_copies > 0 &&
1136 zp->zp_copies <= spa_max_replication(spa));
1138 zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private,
1139 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
1140 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
1141 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
1143 zio->io_ready = ready;
1144 zio->io_children_ready = children_ready;
1145 zio->io_physdone = physdone;
1149 * Data can be NULL if we are going to call zio_write_override() to
1150 * provide the already-allocated BP. But we may need the data to
1151 * verify a dedup hit (if requested). In this case, don't try to
1152 * dedup (just take the already-allocated BP verbatim). Encrypted
1153 * dedup blocks need data as well so we also disable dedup in this
1157 (zio->io_prop.zp_dedup_verify || zio->io_prop.zp_encrypt)) {
1158 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
1165 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data,
1166 uint64_t size, zio_done_func_t *done, void *private,
1167 zio_priority_t priority, zio_flag_t flags, zbookmark_phys_t *zb)
1171 zio = zio_create(pio, spa, txg, bp, data, size, size, done, private,
1172 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb,
1173 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
1179 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
1181 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
1182 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1183 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1184 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
1187 * We must reset the io_prop to match the values that existed
1188 * when the bp was first written by dmu_sync() keeping in mind
1189 * that nopwrite and dedup are mutually exclusive.
1191 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
1192 zio->io_prop.zp_nopwrite = nopwrite;
1193 zio->io_prop.zp_copies = copies;
1194 zio->io_bp_override = bp;
1198 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
1201 (void) zfs_blkptr_verify(spa, bp, B_FALSE, BLK_VERIFY_HALT);
1204 * The check for EMBEDDED is a performance optimization. We
1205 * process the free here (by ignoring it) rather than
1206 * putting it on the list and then processing it in zio_free_sync().
1208 if (BP_IS_EMBEDDED(bp))
1212 * Frees that are for the currently-syncing txg, are not going to be
1213 * deferred, and which will not need to do a read (i.e. not GANG or
1214 * DEDUP), can be processed immediately. Otherwise, put them on the
1215 * in-memory list for later processing.
1217 * Note that we only defer frees after zfs_sync_pass_deferred_free
1218 * when the log space map feature is disabled. [see relevant comment
1219 * in spa_sync_iterate_to_convergence()]
1221 if (BP_IS_GANG(bp) ||
1223 txg != spa->spa_syncing_txg ||
1224 (spa_sync_pass(spa) >= zfs_sync_pass_deferred_free &&
1225 !spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP))) {
1226 metaslab_check_free(spa, bp);
1227 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
1229 VERIFY3P(zio_free_sync(NULL, spa, txg, bp, 0), ==, NULL);
1234 * To improve performance, this function may return NULL if we were able
1235 * to do the free immediately. This avoids the cost of creating a zio
1236 * (and linking it to the parent, etc).
1239 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1242 ASSERT(!BP_IS_HOLE(bp));
1243 ASSERT(spa_syncing_txg(spa) == txg);
1245 if (BP_IS_EMBEDDED(bp))
1248 metaslab_check_free(spa, bp);
1250 dsl_scan_freed(spa, bp);
1252 if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp)) {
1254 * GANG and DEDUP blocks can induce a read (for the gang block
1255 * header, or the DDT), so issue them asynchronously so that
1256 * this thread is not tied up.
1258 enum zio_stage stage =
1259 ZIO_FREE_PIPELINE | ZIO_STAGE_ISSUE_ASYNC;
1261 return (zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1262 BP_GET_PSIZE(bp), NULL, NULL,
1263 ZIO_TYPE_FREE, ZIO_PRIORITY_NOW,
1264 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage));
1266 metaslab_free(spa, bp, txg, B_FALSE);
1272 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1273 zio_done_func_t *done, void *private, zio_flag_t flags)
1277 (void) zfs_blkptr_verify(spa, bp, flags & ZIO_FLAG_CONFIG_WRITER,
1280 if (BP_IS_EMBEDDED(bp))
1281 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
1284 * A claim is an allocation of a specific block. Claims are needed
1285 * to support immediate writes in the intent log. The issue is that
1286 * immediate writes contain committed data, but in a txg that was
1287 * *not* committed. Upon opening the pool after an unclean shutdown,
1288 * the intent log claims all blocks that contain immediate write data
1289 * so that the SPA knows they're in use.
1291 * All claims *must* be resolved in the first txg -- before the SPA
1292 * starts allocating blocks -- so that nothing is allocated twice.
1293 * If txg == 0 we just verify that the block is claimable.
1295 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <,
1296 spa_min_claim_txg(spa));
1297 ASSERT(txg == spa_min_claim_txg(spa) || txg == 0);
1298 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(8) */
1300 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1301 BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW,
1302 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
1303 ASSERT0(zio->io_queued_timestamp);
1309 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
1310 zio_done_func_t *done, void *private, zio_flag_t flags)
1315 if (vd->vdev_children == 0) {
1316 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
1317 ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
1318 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
1322 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
1324 for (c = 0; c < vd->vdev_children; c++)
1325 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
1326 done, private, flags));
1333 zio_trim(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1334 zio_done_func_t *done, void *private, zio_priority_t priority,
1335 zio_flag_t flags, enum trim_flag trim_flags)
1339 ASSERT0(vd->vdev_children);
1340 ASSERT0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
1341 ASSERT0(P2PHASE(size, 1ULL << vd->vdev_ashift));
1342 ASSERT3U(size, !=, 0);
1344 zio = zio_create(pio, vd->vdev_spa, 0, NULL, NULL, size, size, done,
1345 private, ZIO_TYPE_TRIM, priority, flags | ZIO_FLAG_PHYSICAL,
1346 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_TRIM_PIPELINE);
1347 zio->io_trim_flags = trim_flags;
1353 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1354 abd_t *data, int checksum, zio_done_func_t *done, void *private,
1355 zio_priority_t priority, zio_flag_t flags, boolean_t labels)
1359 ASSERT(vd->vdev_children == 0);
1360 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1361 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1362 ASSERT3U(offset + size, <=, vd->vdev_psize);
1364 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1365 private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1366 offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
1368 zio->io_prop.zp_checksum = checksum;
1374 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1375 abd_t *data, int checksum, zio_done_func_t *done, void *private,
1376 zio_priority_t priority, zio_flag_t flags, boolean_t labels)
1380 ASSERT(vd->vdev_children == 0);
1381 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1382 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1383 ASSERT3U(offset + size, <=, vd->vdev_psize);
1385 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1386 private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1387 offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
1389 zio->io_prop.zp_checksum = checksum;
1391 if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
1393 * zec checksums are necessarily destructive -- they modify
1394 * the end of the write buffer to hold the verifier/checksum.
1395 * Therefore, we must make a local copy in case the data is
1396 * being written to multiple places in parallel.
1398 abd_t *wbuf = abd_alloc_sametype(data, size);
1399 abd_copy(wbuf, data, size);
1401 zio_push_transform(zio, wbuf, size, size, NULL);
1408 * Create a child I/O to do some work for us.
1411 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1412 abd_t *data, uint64_t size, int type, zio_priority_t priority,
1413 zio_flag_t flags, zio_done_func_t *done, void *private)
1415 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1419 * vdev child I/Os do not propagate their error to the parent.
1420 * Therefore, for correct operation the caller *must* check for
1421 * and handle the error in the child i/o's done callback.
1422 * The only exceptions are i/os that we don't care about
1423 * (OPTIONAL or REPAIR).
1425 ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) ||
1428 if (type == ZIO_TYPE_READ && bp != NULL) {
1430 * If we have the bp, then the child should perform the
1431 * checksum and the parent need not. This pushes error
1432 * detection as close to the leaves as possible and
1433 * eliminates redundant checksums in the interior nodes.
1435 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1436 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1439 if (vd->vdev_ops->vdev_op_leaf) {
1440 ASSERT0(vd->vdev_children);
1441 offset += VDEV_LABEL_START_SIZE;
1444 flags |= ZIO_VDEV_CHILD_FLAGS(pio);
1447 * If we've decided to do a repair, the write is not speculative --
1448 * even if the original read was.
1450 if (flags & ZIO_FLAG_IO_REPAIR)
1451 flags &= ~ZIO_FLAG_SPECULATIVE;
1454 * If we're creating a child I/O that is not associated with a
1455 * top-level vdev, then the child zio is not an allocating I/O.
1456 * If this is a retried I/O then we ignore it since we will
1457 * have already processed the original allocating I/O.
1459 if (flags & ZIO_FLAG_IO_ALLOCATING &&
1460 (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) {
1461 ASSERT(pio->io_metaslab_class != NULL);
1462 ASSERT(pio->io_metaslab_class->mc_alloc_throttle_enabled);
1463 ASSERT(type == ZIO_TYPE_WRITE);
1464 ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE);
1465 ASSERT(!(flags & ZIO_FLAG_IO_REPAIR));
1466 ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) ||
1467 pio->io_child_type == ZIO_CHILD_GANG);
1469 flags &= ~ZIO_FLAG_IO_ALLOCATING;
1473 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size,
1474 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1475 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1476 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
1478 zio->io_physdone = pio->io_physdone;
1479 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1480 zio->io_logical->io_phys_children++;
1486 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size,
1487 zio_type_t type, zio_priority_t priority, zio_flag_t flags,
1488 zio_done_func_t *done, void *private)
1492 ASSERT(vd->vdev_ops->vdev_op_leaf);
1494 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1495 data, size, size, done, private, type, priority,
1496 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1498 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1504 zio_flush(zio_t *zio, vdev_t *vd)
1506 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
1508 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1512 zio_shrink(zio_t *zio, uint64_t size)
1514 ASSERT3P(zio->io_executor, ==, NULL);
1515 ASSERT3U(zio->io_orig_size, ==, zio->io_size);
1516 ASSERT3U(size, <=, zio->io_size);
1519 * We don't shrink for raidz because of problems with the
1520 * reconstruction when reading back less than the block size.
1521 * Note, BP_IS_RAIDZ() assumes no compression.
1523 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1524 if (!BP_IS_RAIDZ(zio->io_bp)) {
1525 /* we are not doing a raw write */
1526 ASSERT3U(zio->io_size, ==, zio->io_lsize);
1527 zio->io_orig_size = zio->io_size = zio->io_lsize = size;
1532 * ==========================================================================
1533 * Prepare to read and write logical blocks
1534 * ==========================================================================
1538 zio_read_bp_init(zio_t *zio)
1540 blkptr_t *bp = zio->io_bp;
1542 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1544 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1546 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1547 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1548 !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) {
1549 zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1550 psize, psize, zio_decompress);
1553 if (((BP_IS_PROTECTED(bp) && !(zio->io_flags & ZIO_FLAG_RAW_ENCRYPT)) ||
1554 BP_HAS_INDIRECT_MAC_CKSUM(bp)) &&
1555 zio->io_child_type == ZIO_CHILD_LOGICAL) {
1556 zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1557 psize, psize, zio_decrypt);
1560 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1561 int psize = BPE_GET_PSIZE(bp);
1562 void *data = abd_borrow_buf(zio->io_abd, psize);
1564 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1565 decode_embedded_bp_compressed(bp, data);
1566 abd_return_buf_copy(zio->io_abd, data, psize);
1568 ASSERT(!BP_IS_EMBEDDED(bp));
1569 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1572 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1573 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1575 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1576 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1578 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1579 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1585 zio_write_bp_init(zio_t *zio)
1587 if (!IO_IS_ALLOCATING(zio))
1590 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1592 if (zio->io_bp_override) {
1593 blkptr_t *bp = zio->io_bp;
1594 zio_prop_t *zp = &zio->io_prop;
1596 ASSERT(bp->blk_birth != zio->io_txg);
1597 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1599 *bp = *zio->io_bp_override;
1600 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1602 if (BP_IS_EMBEDDED(bp))
1606 * If we've been overridden and nopwrite is set then
1607 * set the flag accordingly to indicate that a nopwrite
1608 * has already occurred.
1610 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1611 ASSERT(!zp->zp_dedup);
1612 ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum);
1613 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1617 ASSERT(!zp->zp_nopwrite);
1619 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1622 ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags &
1623 ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify);
1625 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum &&
1627 BP_SET_DEDUP(bp, 1);
1628 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1633 * We were unable to handle this as an override bp, treat
1634 * it as a regular write I/O.
1636 zio->io_bp_override = NULL;
1637 *bp = zio->io_bp_orig;
1638 zio->io_pipeline = zio->io_orig_pipeline;
1645 zio_write_compress(zio_t *zio)
1647 spa_t *spa = zio->io_spa;
1648 zio_prop_t *zp = &zio->io_prop;
1649 enum zio_compress compress = zp->zp_compress;
1650 blkptr_t *bp = zio->io_bp;
1651 uint64_t lsize = zio->io_lsize;
1652 uint64_t psize = zio->io_size;
1656 * If our children haven't all reached the ready stage,
1657 * wait for them and then repeat this pipeline stage.
1659 if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT |
1660 ZIO_CHILD_GANG_BIT, ZIO_WAIT_READY)) {
1664 if (!IO_IS_ALLOCATING(zio))
1667 if (zio->io_children_ready != NULL) {
1669 * Now that all our children are ready, run the callback
1670 * associated with this zio in case it wants to modify the
1671 * data to be written.
1673 ASSERT3U(zp->zp_level, >, 0);
1674 zio->io_children_ready(zio);
1677 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1678 ASSERT(zio->io_bp_override == NULL);
1680 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1682 * We're rewriting an existing block, which means we're
1683 * working on behalf of spa_sync(). For spa_sync() to
1684 * converge, it must eventually be the case that we don't
1685 * have to allocate new blocks. But compression changes
1686 * the blocksize, which forces a reallocate, and makes
1687 * convergence take longer. Therefore, after the first
1688 * few passes, stop compressing to ensure convergence.
1690 pass = spa_sync_pass(spa);
1692 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1693 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1694 ASSERT(!BP_GET_DEDUP(bp));
1696 if (pass >= zfs_sync_pass_dont_compress)
1697 compress = ZIO_COMPRESS_OFF;
1699 /* Make sure someone doesn't change their mind on overwrites */
1700 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1701 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1704 /* If it's a compressed write that is not raw, compress the buffer. */
1705 if (compress != ZIO_COMPRESS_OFF &&
1706 !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) {
1707 void *cbuf = zio_buf_alloc(lsize);
1708 psize = zio_compress_data(compress, zio->io_abd, cbuf, lsize,
1710 if (psize == 0 || psize >= lsize) {
1711 compress = ZIO_COMPRESS_OFF;
1712 zio_buf_free(cbuf, lsize);
1713 } else if (!zp->zp_dedup && !zp->zp_encrypt &&
1714 psize <= BPE_PAYLOAD_SIZE &&
1715 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1716 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1717 encode_embedded_bp_compressed(bp,
1718 cbuf, compress, lsize, psize);
1719 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1720 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1721 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1722 zio_buf_free(cbuf, lsize);
1723 bp->blk_birth = zio->io_txg;
1724 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1725 ASSERT(spa_feature_is_active(spa,
1726 SPA_FEATURE_EMBEDDED_DATA));
1730 * Round compressed size up to the minimum allocation
1731 * size of the smallest-ashift device, and zero the
1732 * tail. This ensures that the compressed size of the
1733 * BP (and thus compressratio property) are correct,
1734 * in that we charge for the padding used to fill out
1737 ASSERT3U(spa->spa_min_alloc, >=, SPA_MINBLOCKSHIFT);
1738 size_t rounded = (size_t)roundup(psize,
1739 spa->spa_min_alloc);
1740 if (rounded >= lsize) {
1741 compress = ZIO_COMPRESS_OFF;
1742 zio_buf_free(cbuf, lsize);
1745 abd_t *cdata = abd_get_from_buf(cbuf, lsize);
1746 abd_take_ownership_of_buf(cdata, B_TRUE);
1747 abd_zero_off(cdata, psize, rounded - psize);
1749 zio_push_transform(zio, cdata,
1750 psize, lsize, NULL);
1755 * We were unable to handle this as an override bp, treat
1756 * it as a regular write I/O.
1758 zio->io_bp_override = NULL;
1759 *bp = zio->io_bp_orig;
1760 zio->io_pipeline = zio->io_orig_pipeline;
1762 } else if ((zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) != 0 &&
1763 zp->zp_type == DMU_OT_DNODE) {
1765 * The DMU actually relies on the zio layer's compression
1766 * to free metadnode blocks that have had all contained
1767 * dnodes freed. As a result, even when doing a raw
1768 * receive, we must check whether the block can be compressed
1771 psize = zio_compress_data(ZIO_COMPRESS_EMPTY,
1772 zio->io_abd, NULL, lsize, zp->zp_complevel);
1773 if (psize == 0 || psize >= lsize)
1774 compress = ZIO_COMPRESS_OFF;
1775 } else if (zio->io_flags & ZIO_FLAG_RAW_COMPRESS &&
1776 !(zio->io_flags & ZIO_FLAG_RAW_ENCRYPT)) {
1778 * If we are raw receiving an encrypted dataset we should not
1779 * take this codepath because it will change the on-disk block
1780 * and decryption will fail.
1782 size_t rounded = MIN((size_t)roundup(psize,
1783 spa->spa_min_alloc), lsize);
1785 if (rounded != psize) {
1786 abd_t *cdata = abd_alloc_linear(rounded, B_TRUE);
1787 abd_zero_off(cdata, psize, rounded - psize);
1788 abd_copy_off(cdata, zio->io_abd, 0, 0, psize);
1790 zio_push_transform(zio, cdata,
1791 psize, rounded, NULL);
1794 ASSERT3U(psize, !=, 0);
1798 * The final pass of spa_sync() must be all rewrites, but the first
1799 * few passes offer a trade-off: allocating blocks defers convergence,
1800 * but newly allocated blocks are sequential, so they can be written
1801 * to disk faster. Therefore, we allow the first few passes of
1802 * spa_sync() to allocate new blocks, but force rewrites after that.
1803 * There should only be a handful of blocks after pass 1 in any case.
1805 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1806 BP_GET_PSIZE(bp) == psize &&
1807 pass >= zfs_sync_pass_rewrite) {
1808 VERIFY3U(psize, !=, 0);
1809 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1811 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1812 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1815 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1819 if (zio->io_bp_orig.blk_birth != 0 &&
1820 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1821 BP_SET_LSIZE(bp, lsize);
1822 BP_SET_TYPE(bp, zp->zp_type);
1823 BP_SET_LEVEL(bp, zp->zp_level);
1824 BP_SET_BIRTH(bp, zio->io_txg, 0);
1826 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1828 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1829 BP_SET_LSIZE(bp, lsize);
1830 BP_SET_TYPE(bp, zp->zp_type);
1831 BP_SET_LEVEL(bp, zp->zp_level);
1832 BP_SET_PSIZE(bp, psize);
1833 BP_SET_COMPRESS(bp, compress);
1834 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1835 BP_SET_DEDUP(bp, zp->zp_dedup);
1836 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1838 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1839 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1840 ASSERT(!zp->zp_encrypt ||
1841 DMU_OT_IS_ENCRYPTED(zp->zp_type));
1842 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1844 if (zp->zp_nopwrite) {
1845 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1846 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1847 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1854 zio_free_bp_init(zio_t *zio)
1856 blkptr_t *bp = zio->io_bp;
1858 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1859 if (BP_GET_DEDUP(bp))
1860 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1863 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1869 * ==========================================================================
1870 * Execute the I/O pipeline
1871 * ==========================================================================
1875 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1877 spa_t *spa = zio->io_spa;
1878 zio_type_t t = zio->io_type;
1879 int flags = (cutinline ? TQ_FRONT : 0);
1882 * If we're a config writer or a probe, the normal issue and
1883 * interrupt threads may all be blocked waiting for the config lock.
1884 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1886 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1890 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1892 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1896 * If this is a high priority I/O, then use the high priority taskq if
1899 if ((zio->io_priority == ZIO_PRIORITY_NOW ||
1900 zio->io_priority == ZIO_PRIORITY_SYNC_WRITE) &&
1901 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1904 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1907 * NB: We are assuming that the zio can only be dispatched
1908 * to a single taskq at a time. It would be a grievous error
1909 * to dispatch the zio to another taskq at the same time.
1911 ASSERT(taskq_empty_ent(&zio->io_tqent));
1912 spa_taskq_dispatch_ent(spa, t, q, zio_execute, zio, flags,
1917 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1919 spa_t *spa = zio->io_spa;
1921 taskq_t *tq = taskq_of_curthread();
1923 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1924 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1926 for (i = 0; i < tqs->stqs_count; i++) {
1927 if (tqs->stqs_taskq[i] == tq)
1936 zio_issue_async(zio_t *zio)
1938 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1944 zio_interrupt(void *zio)
1946 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1950 zio_delay_interrupt(zio_t *zio)
1953 * The timeout_generic() function isn't defined in userspace, so
1954 * rather than trying to implement the function, the zio delay
1955 * functionality has been disabled for userspace builds.
1960 * If io_target_timestamp is zero, then no delay has been registered
1961 * for this IO, thus jump to the end of this function and "skip" the
1962 * delay; issuing it directly to the zio layer.
1964 if (zio->io_target_timestamp != 0) {
1965 hrtime_t now = gethrtime();
1967 if (now >= zio->io_target_timestamp) {
1969 * This IO has already taken longer than the target
1970 * delay to complete, so we don't want to delay it
1971 * any longer; we "miss" the delay and issue it
1972 * directly to the zio layer. This is likely due to
1973 * the target latency being set to a value less than
1974 * the underlying hardware can satisfy (e.g. delay
1975 * set to 1ms, but the disks take 10ms to complete an
1979 DTRACE_PROBE2(zio__delay__miss, zio_t *, zio,
1985 hrtime_t diff = zio->io_target_timestamp - now;
1986 clock_t expire_at_tick = ddi_get_lbolt() +
1989 DTRACE_PROBE3(zio__delay__hit, zio_t *, zio,
1990 hrtime_t, now, hrtime_t, diff);
1992 if (NSEC_TO_TICK(diff) == 0) {
1993 /* Our delay is less than a jiffy - just spin */
1994 zfs_sleep_until(zio->io_target_timestamp);
1998 * Use taskq_dispatch_delay() in the place of
1999 * OpenZFS's timeout_generic().
2001 tid = taskq_dispatch_delay(system_taskq,
2002 zio_interrupt, zio, TQ_NOSLEEP,
2004 if (tid == TASKQID_INVALID) {
2006 * Couldn't allocate a task. Just
2007 * finish the zio without a delay.
2016 DTRACE_PROBE1(zio__delay__skip, zio_t *, zio);
2021 zio_deadman_impl(zio_t *pio, int ziodepth)
2023 zio_t *cio, *cio_next;
2024 zio_link_t *zl = NULL;
2025 vdev_t *vd = pio->io_vd;
2027 if (zio_deadman_log_all || (vd != NULL && vd->vdev_ops->vdev_op_leaf)) {
2028 vdev_queue_t *vq = vd ? &vd->vdev_queue : NULL;
2029 zbookmark_phys_t *zb = &pio->io_bookmark;
2030 uint64_t delta = gethrtime() - pio->io_timestamp;
2031 uint64_t failmode = spa_get_deadman_failmode(pio->io_spa);
2033 zfs_dbgmsg("slow zio[%d]: zio=%px timestamp=%llu "
2034 "delta=%llu queued=%llu io=%llu "
2036 "last=%llu type=%d "
2037 "priority=%d flags=0x%llx stage=0x%x "
2038 "pipeline=0x%x pipeline-trace=0x%x "
2039 "objset=%llu object=%llu "
2040 "level=%llu blkid=%llu "
2041 "offset=%llu size=%llu "
2043 ziodepth, pio, pio->io_timestamp,
2044 (u_longlong_t)delta, pio->io_delta, pio->io_delay,
2045 vd ? vd->vdev_path : "NULL",
2046 vq ? vq->vq_io_complete_ts : 0, pio->io_type,
2047 pio->io_priority, (u_longlong_t)pio->io_flags,
2048 pio->io_stage, pio->io_pipeline, pio->io_pipeline_trace,
2049 (u_longlong_t)zb->zb_objset, (u_longlong_t)zb->zb_object,
2050 (u_longlong_t)zb->zb_level, (u_longlong_t)zb->zb_blkid,
2051 (u_longlong_t)pio->io_offset, (u_longlong_t)pio->io_size,
2053 (void) zfs_ereport_post(FM_EREPORT_ZFS_DEADMAN,
2054 pio->io_spa, vd, zb, pio, 0);
2056 if (failmode == ZIO_FAILURE_MODE_CONTINUE &&
2057 taskq_empty_ent(&pio->io_tqent)) {
2062 mutex_enter(&pio->io_lock);
2063 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
2064 cio_next = zio_walk_children(pio, &zl);
2065 zio_deadman_impl(cio, ziodepth + 1);
2067 mutex_exit(&pio->io_lock);
2071 * Log the critical information describing this zio and all of its children
2072 * using the zfs_dbgmsg() interface then post deadman event for the ZED.
2075 zio_deadman(zio_t *pio, const char *tag)
2077 spa_t *spa = pio->io_spa;
2078 char *name = spa_name(spa);
2080 if (!zfs_deadman_enabled || spa_suspended(spa))
2083 zio_deadman_impl(pio, 0);
2085 switch (spa_get_deadman_failmode(spa)) {
2086 case ZIO_FAILURE_MODE_WAIT:
2087 zfs_dbgmsg("%s waiting for hung I/O to pool '%s'", tag, name);
2090 case ZIO_FAILURE_MODE_CONTINUE:
2091 zfs_dbgmsg("%s restarting hung I/O for pool '%s'", tag, name);
2094 case ZIO_FAILURE_MODE_PANIC:
2095 fm_panic("%s determined I/O to pool '%s' is hung.", tag, name);
2101 * Execute the I/O pipeline until one of the following occurs:
2102 * (1) the I/O completes; (2) the pipeline stalls waiting for
2103 * dependent child I/Os; (3) the I/O issues, so we're waiting
2104 * for an I/O completion interrupt; (4) the I/O is delegated by
2105 * vdev-level caching or aggregation; (5) the I/O is deferred
2106 * due to vdev-level queueing; (6) the I/O is handed off to
2107 * another thread. In all cases, the pipeline stops whenever
2108 * there's no CPU work; it never burns a thread in cv_wait_io().
2110 * There's no locking on io_stage because there's no legitimate way
2111 * for multiple threads to be attempting to process the same I/O.
2113 static zio_pipe_stage_t *zio_pipeline[];
2116 * zio_execute() is a wrapper around the static function
2117 * __zio_execute() so that we can force __zio_execute() to be
2118 * inlined. This reduces stack overhead which is important
2119 * because __zio_execute() is called recursively in several zio
2120 * code paths. zio_execute() itself cannot be inlined because
2121 * it is externally visible.
2124 zio_execute(void *zio)
2126 fstrans_cookie_t cookie;
2128 cookie = spl_fstrans_mark();
2130 spl_fstrans_unmark(cookie);
2134 * Used to determine if in the current context the stack is sized large
2135 * enough to allow zio_execute() to be called recursively. A minimum
2136 * stack size of 16K is required to avoid needing to re-dispatch the zio.
2139 zio_execute_stack_check(zio_t *zio)
2141 #if !defined(HAVE_LARGE_STACKS)
2142 dsl_pool_t *dp = spa_get_dsl(zio->io_spa);
2144 /* Executing in txg_sync_thread() context. */
2145 if (dp && curthread == dp->dp_tx.tx_sync_thread)
2148 /* Pool initialization outside of zio_taskq context. */
2149 if (dp && spa_is_initializing(dp->dp_spa) &&
2150 !zio_taskq_member(zio, ZIO_TASKQ_ISSUE) &&
2151 !zio_taskq_member(zio, ZIO_TASKQ_ISSUE_HIGH))
2155 #endif /* HAVE_LARGE_STACKS */
2160 __attribute__((always_inline))
2162 __zio_execute(zio_t *zio)
2164 ASSERT3U(zio->io_queued_timestamp, >, 0);
2166 while (zio->io_stage < ZIO_STAGE_DONE) {
2167 enum zio_stage pipeline = zio->io_pipeline;
2168 enum zio_stage stage = zio->io_stage;
2170 zio->io_executor = curthread;
2172 ASSERT(!MUTEX_HELD(&zio->io_lock));
2173 ASSERT(ISP2(stage));
2174 ASSERT(zio->io_stall == NULL);
2178 } while ((stage & pipeline) == 0);
2180 ASSERT(stage <= ZIO_STAGE_DONE);
2183 * If we are in interrupt context and this pipeline stage
2184 * will grab a config lock that is held across I/O,
2185 * or may wait for an I/O that needs an interrupt thread
2186 * to complete, issue async to avoid deadlock.
2188 * For VDEV_IO_START, we cut in line so that the io will
2189 * be sent to disk promptly.
2191 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
2192 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
2193 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
2194 zio_requeue_io_start_cut_in_line : B_FALSE;
2195 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
2200 * If the current context doesn't have large enough stacks
2201 * the zio must be issued asynchronously to prevent overflow.
2203 if (zio_execute_stack_check(zio)) {
2204 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
2205 zio_requeue_io_start_cut_in_line : B_FALSE;
2206 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
2210 zio->io_stage = stage;
2211 zio->io_pipeline_trace |= zio->io_stage;
2214 * The zio pipeline stage returns the next zio to execute
2215 * (typically the same as this one), or NULL if we should
2218 zio = zio_pipeline[highbit64(stage) - 1](zio);
2227 * ==========================================================================
2228 * Initiate I/O, either sync or async
2229 * ==========================================================================
2232 zio_wait(zio_t *zio)
2235 * Some routines, like zio_free_sync(), may return a NULL zio
2236 * to avoid the performance overhead of creating and then destroying
2237 * an unneeded zio. For the callers' simplicity, we accept a NULL
2238 * zio and ignore it.
2243 long timeout = MSEC_TO_TICK(zfs_deadman_ziotime_ms);
2246 ASSERT3S(zio->io_stage, ==, ZIO_STAGE_OPEN);
2247 ASSERT3P(zio->io_executor, ==, NULL);
2249 zio->io_waiter = curthread;
2250 ASSERT0(zio->io_queued_timestamp);
2251 zio->io_queued_timestamp = gethrtime();
2255 mutex_enter(&zio->io_lock);
2256 while (zio->io_executor != NULL) {
2257 error = cv_timedwait_io(&zio->io_cv, &zio->io_lock,
2258 ddi_get_lbolt() + timeout);
2260 if (zfs_deadman_enabled && error == -1 &&
2261 gethrtime() - zio->io_queued_timestamp >
2262 spa_deadman_ziotime(zio->io_spa)) {
2263 mutex_exit(&zio->io_lock);
2264 timeout = MSEC_TO_TICK(zfs_deadman_checktime_ms);
2265 zio_deadman(zio, FTAG);
2266 mutex_enter(&zio->io_lock);
2269 mutex_exit(&zio->io_lock);
2271 error = zio->io_error;
2278 zio_nowait(zio_t *zio)
2281 * See comment in zio_wait().
2286 ASSERT3P(zio->io_executor, ==, NULL);
2288 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
2289 zio_unique_parent(zio) == NULL) {
2293 * This is a logical async I/O with no parent to wait for it.
2294 * We add it to the spa_async_root_zio "Godfather" I/O which
2295 * will ensure they complete prior to unloading the pool.
2297 spa_t *spa = zio->io_spa;
2298 pio = spa->spa_async_zio_root[CPU_SEQID_UNSTABLE];
2300 zio_add_child(pio, zio);
2303 ASSERT0(zio->io_queued_timestamp);
2304 zio->io_queued_timestamp = gethrtime();
2309 * ==========================================================================
2310 * Reexecute, cancel, or suspend/resume failed I/O
2311 * ==========================================================================
2315 zio_reexecute(void *arg)
2318 zio_t *cio, *cio_next;
2320 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
2321 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
2322 ASSERT(pio->io_gang_leader == NULL);
2323 ASSERT(pio->io_gang_tree == NULL);
2325 pio->io_flags = pio->io_orig_flags;
2326 pio->io_stage = pio->io_orig_stage;
2327 pio->io_pipeline = pio->io_orig_pipeline;
2328 pio->io_reexecute = 0;
2329 pio->io_flags |= ZIO_FLAG_REEXECUTED;
2330 pio->io_pipeline_trace = 0;
2332 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2333 pio->io_state[w] = 0;
2334 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2335 pio->io_child_error[c] = 0;
2337 if (IO_IS_ALLOCATING(pio))
2338 BP_ZERO(pio->io_bp);
2341 * As we reexecute pio's children, new children could be created.
2342 * New children go to the head of pio's io_child_list, however,
2343 * so we will (correctly) not reexecute them. The key is that
2344 * the remainder of pio's io_child_list, from 'cio_next' onward,
2345 * cannot be affected by any side effects of reexecuting 'cio'.
2347 zio_link_t *zl = NULL;
2348 mutex_enter(&pio->io_lock);
2349 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
2350 cio_next = zio_walk_children(pio, &zl);
2351 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2352 pio->io_children[cio->io_child_type][w]++;
2353 mutex_exit(&pio->io_lock);
2355 mutex_enter(&pio->io_lock);
2357 mutex_exit(&pio->io_lock);
2360 * Now that all children have been reexecuted, execute the parent.
2361 * We don't reexecute "The Godfather" I/O here as it's the
2362 * responsibility of the caller to wait on it.
2364 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) {
2365 pio->io_queued_timestamp = gethrtime();
2371 zio_suspend(spa_t *spa, zio_t *zio, zio_suspend_reason_t reason)
2373 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
2374 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
2375 "failure and the failure mode property for this pool "
2376 "is set to panic.", spa_name(spa));
2378 cmn_err(CE_WARN, "Pool '%s' has encountered an uncorrectable I/O "
2379 "failure and has been suspended.\n", spa_name(spa));
2381 (void) zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL,
2384 mutex_enter(&spa->spa_suspend_lock);
2386 if (spa->spa_suspend_zio_root == NULL)
2387 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
2388 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
2389 ZIO_FLAG_GODFATHER);
2391 spa->spa_suspended = reason;
2394 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2395 ASSERT(zio != spa->spa_suspend_zio_root);
2396 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2397 ASSERT(zio_unique_parent(zio) == NULL);
2398 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
2399 zio_add_child(spa->spa_suspend_zio_root, zio);
2402 mutex_exit(&spa->spa_suspend_lock);
2406 zio_resume(spa_t *spa)
2411 * Reexecute all previously suspended i/o.
2413 mutex_enter(&spa->spa_suspend_lock);
2414 spa->spa_suspended = ZIO_SUSPEND_NONE;
2415 cv_broadcast(&spa->spa_suspend_cv);
2416 pio = spa->spa_suspend_zio_root;
2417 spa->spa_suspend_zio_root = NULL;
2418 mutex_exit(&spa->spa_suspend_lock);
2424 return (zio_wait(pio));
2428 zio_resume_wait(spa_t *spa)
2430 mutex_enter(&spa->spa_suspend_lock);
2431 while (spa_suspended(spa))
2432 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
2433 mutex_exit(&spa->spa_suspend_lock);
2437 * ==========================================================================
2440 * A gang block is a collection of small blocks that looks to the DMU
2441 * like one large block. When zio_dva_allocate() cannot find a block
2442 * of the requested size, due to either severe fragmentation or the pool
2443 * being nearly full, it calls zio_write_gang_block() to construct the
2444 * block from smaller fragments.
2446 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
2447 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
2448 * an indirect block: it's an array of block pointers. It consumes
2449 * only one sector and hence is allocatable regardless of fragmentation.
2450 * The gang header's bps point to its gang members, which hold the data.
2452 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
2453 * as the verifier to ensure uniqueness of the SHA256 checksum.
2454 * Critically, the gang block bp's blk_cksum is the checksum of the data,
2455 * not the gang header. This ensures that data block signatures (needed for
2456 * deduplication) are independent of how the block is physically stored.
2458 * Gang blocks can be nested: a gang member may itself be a gang block.
2459 * Thus every gang block is a tree in which root and all interior nodes are
2460 * gang headers, and the leaves are normal blocks that contain user data.
2461 * The root of the gang tree is called the gang leader.
2463 * To perform any operation (read, rewrite, free, claim) on a gang block,
2464 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
2465 * in the io_gang_tree field of the original logical i/o by recursively
2466 * reading the gang leader and all gang headers below it. This yields
2467 * an in-core tree containing the contents of every gang header and the
2468 * bps for every constituent of the gang block.
2470 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
2471 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
2472 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
2473 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
2474 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
2475 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
2476 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
2477 * of the gang header plus zio_checksum_compute() of the data to update the
2478 * gang header's blk_cksum as described above.
2480 * The two-phase assemble/issue model solves the problem of partial failure --
2481 * what if you'd freed part of a gang block but then couldn't read the
2482 * gang header for another part? Assembling the entire gang tree first
2483 * ensures that all the necessary gang header I/O has succeeded before
2484 * starting the actual work of free, claim, or write. Once the gang tree
2485 * is assembled, free and claim are in-memory operations that cannot fail.
2487 * In the event that a gang write fails, zio_dva_unallocate() walks the
2488 * gang tree to immediately free (i.e. insert back into the space map)
2489 * everything we've allocated. This ensures that we don't get ENOSPC
2490 * errors during repeated suspend/resume cycles due to a flaky device.
2492 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
2493 * the gang tree, we won't modify the block, so we can safely defer the free
2494 * (knowing that the block is still intact). If we *can* assemble the gang
2495 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
2496 * each constituent bp and we can allocate a new block on the next sync pass.
2498 * In all cases, the gang tree allows complete recovery from partial failure.
2499 * ==========================================================================
2503 zio_gang_issue_func_done(zio_t *zio)
2505 abd_free(zio->io_abd);
2509 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2515 return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset),
2516 BP_GET_PSIZE(bp), zio_gang_issue_func_done,
2517 NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2518 &pio->io_bookmark));
2522 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2529 abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2530 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2531 gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL,
2532 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2535 * As we rewrite each gang header, the pipeline will compute
2536 * a new gang block header checksum for it; but no one will
2537 * compute a new data checksum, so we do that here. The one
2538 * exception is the gang leader: the pipeline already computed
2539 * its data checksum because that stage precedes gang assembly.
2540 * (Presently, nothing actually uses interior data checksums;
2541 * this is just good hygiene.)
2543 if (gn != pio->io_gang_leader->io_gang_tree) {
2544 abd_t *buf = abd_get_offset(data, offset);
2546 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
2547 buf, BP_GET_PSIZE(bp));
2552 * If we are here to damage data for testing purposes,
2553 * leave the GBH alone so that we can detect the damage.
2555 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
2556 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2558 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2559 abd_get_offset(data, offset), BP_GET_PSIZE(bp),
2560 zio_gang_issue_func_done, NULL, pio->io_priority,
2561 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2568 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2571 (void) gn, (void) data, (void) offset;
2573 zio_t *zio = zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
2574 ZIO_GANG_CHILD_FLAGS(pio));
2576 zio = zio_null(pio, pio->io_spa,
2577 NULL, NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio));
2583 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2586 (void) gn, (void) data, (void) offset;
2587 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
2588 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
2591 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
2600 static void zio_gang_tree_assemble_done(zio_t *zio);
2602 static zio_gang_node_t *
2603 zio_gang_node_alloc(zio_gang_node_t **gnpp)
2605 zio_gang_node_t *gn;
2607 ASSERT(*gnpp == NULL);
2609 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
2610 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
2617 zio_gang_node_free(zio_gang_node_t **gnpp)
2619 zio_gang_node_t *gn = *gnpp;
2621 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2622 ASSERT(gn->gn_child[g] == NULL);
2624 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2625 kmem_free(gn, sizeof (*gn));
2630 zio_gang_tree_free(zio_gang_node_t **gnpp)
2632 zio_gang_node_t *gn = *gnpp;
2637 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2638 zio_gang_tree_free(&gn->gn_child[g]);
2640 zio_gang_node_free(gnpp);
2644 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
2646 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
2647 abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2649 ASSERT(gio->io_gang_leader == gio);
2650 ASSERT(BP_IS_GANG(bp));
2652 zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2653 zio_gang_tree_assemble_done, gn, gio->io_priority,
2654 ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
2658 zio_gang_tree_assemble_done(zio_t *zio)
2660 zio_t *gio = zio->io_gang_leader;
2661 zio_gang_node_t *gn = zio->io_private;
2662 blkptr_t *bp = zio->io_bp;
2664 ASSERT(gio == zio_unique_parent(zio));
2665 ASSERT(zio->io_child_count == 0);
2670 /* this ABD was created from a linear buf in zio_gang_tree_assemble */
2671 if (BP_SHOULD_BYTESWAP(bp))
2672 byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size);
2674 ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh);
2675 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
2676 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2678 abd_free(zio->io_abd);
2680 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2681 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2682 if (!BP_IS_GANG(gbp))
2684 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
2689 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data,
2692 zio_t *gio = pio->io_gang_leader;
2695 ASSERT(BP_IS_GANG(bp) == !!gn);
2696 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
2697 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
2700 * If you're a gang header, your data is in gn->gn_gbh.
2701 * If you're a gang member, your data is in 'data' and gn == NULL.
2703 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset);
2706 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2708 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2709 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2710 if (BP_IS_HOLE(gbp))
2712 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data,
2714 offset += BP_GET_PSIZE(gbp);
2718 if (gn == gio->io_gang_tree)
2719 ASSERT3U(gio->io_size, ==, offset);
2726 zio_gang_assemble(zio_t *zio)
2728 blkptr_t *bp = zio->io_bp;
2730 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
2731 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2733 zio->io_gang_leader = zio;
2735 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
2741 zio_gang_issue(zio_t *zio)
2743 blkptr_t *bp = zio->io_bp;
2745 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT, ZIO_WAIT_DONE)) {
2749 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
2750 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2752 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
2753 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd,
2756 zio_gang_tree_free(&zio->io_gang_tree);
2758 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2764 zio_write_gang_member_ready(zio_t *zio)
2766 zio_t *pio = zio_unique_parent(zio);
2767 dva_t *cdva = zio->io_bp->blk_dva;
2768 dva_t *pdva = pio->io_bp->blk_dva;
2770 zio_t *gio __maybe_unused = zio->io_gang_leader;
2772 if (BP_IS_HOLE(zio->io_bp))
2775 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
2777 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
2778 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
2779 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2780 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
2781 VERIFY3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
2783 mutex_enter(&pio->io_lock);
2784 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
2785 ASSERT(DVA_GET_GANG(&pdva[d]));
2786 asize = DVA_GET_ASIZE(&pdva[d]);
2787 asize += DVA_GET_ASIZE(&cdva[d]);
2788 DVA_SET_ASIZE(&pdva[d], asize);
2790 mutex_exit(&pio->io_lock);
2794 zio_write_gang_done(zio_t *zio)
2797 * The io_abd field will be NULL for a zio with no data. The io_flags
2798 * will initially have the ZIO_FLAG_NODATA bit flag set, but we can't
2799 * check for it here as it is cleared in zio_ready.
2801 if (zio->io_abd != NULL)
2802 abd_free(zio->io_abd);
2806 zio_write_gang_block(zio_t *pio, metaslab_class_t *mc)
2808 spa_t *spa = pio->io_spa;
2809 blkptr_t *bp = pio->io_bp;
2810 zio_t *gio = pio->io_gang_leader;
2812 zio_gang_node_t *gn, **gnpp;
2813 zio_gbh_phys_t *gbh;
2815 uint64_t txg = pio->io_txg;
2816 uint64_t resid = pio->io_size;
2818 int copies = gio->io_prop.zp_copies;
2821 boolean_t has_data = !(pio->io_flags & ZIO_FLAG_NODATA);
2824 * If one copy was requested, store 2 copies of the GBH, so that we
2825 * can still traverse all the data (e.g. to free or scrub) even if a
2826 * block is damaged. Note that we can't store 3 copies of the GBH in
2827 * all cases, e.g. with encryption, which uses DVA[2] for the IV+salt.
2829 int gbh_copies = copies;
2830 if (gbh_copies == 1) {
2831 gbh_copies = MIN(2, spa_max_replication(spa));
2834 int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER;
2835 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2836 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2839 flags |= METASLAB_ASYNC_ALLOC;
2840 VERIFY(zfs_refcount_held(&mc->mc_allocator[pio->io_allocator].
2841 mca_alloc_slots, pio));
2844 * The logical zio has already placed a reservation for
2845 * 'copies' allocation slots but gang blocks may require
2846 * additional copies. These additional copies
2847 * (i.e. gbh_copies - copies) are guaranteed to succeed
2848 * since metaslab_class_throttle_reserve() always allows
2849 * additional reservations for gang blocks.
2851 VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies,
2852 pio->io_allocator, pio, flags));
2855 error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE,
2856 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags,
2857 &pio->io_alloc_list, pio, pio->io_allocator);
2859 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2860 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2864 * If we failed to allocate the gang block header then
2865 * we remove any additional allocation reservations that
2866 * we placed here. The original reservation will
2867 * be removed when the logical I/O goes to the ready
2870 metaslab_class_throttle_unreserve(mc,
2871 gbh_copies - copies, pio->io_allocator, pio);
2874 pio->io_error = error;
2879 gnpp = &gio->io_gang_tree;
2881 gnpp = pio->io_private;
2882 ASSERT(pio->io_ready == zio_write_gang_member_ready);
2885 gn = zio_gang_node_alloc(gnpp);
2887 memset(gbh, 0, SPA_GANGBLOCKSIZE);
2888 gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE);
2891 * Create the gang header.
2893 zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2894 zio_write_gang_done, NULL, pio->io_priority,
2895 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2898 * Create and nowait the gang children.
2900 for (int g = 0; resid != 0; resid -= lsize, g++) {
2901 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2903 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2905 zp.zp_checksum = gio->io_prop.zp_checksum;
2906 zp.zp_compress = ZIO_COMPRESS_OFF;
2907 zp.zp_complevel = gio->io_prop.zp_complevel;
2908 zp.zp_type = DMU_OT_NONE;
2910 zp.zp_copies = gio->io_prop.zp_copies;
2911 zp.zp_dedup = B_FALSE;
2912 zp.zp_dedup_verify = B_FALSE;
2913 zp.zp_nopwrite = B_FALSE;
2914 zp.zp_encrypt = gio->io_prop.zp_encrypt;
2915 zp.zp_byteorder = gio->io_prop.zp_byteorder;
2916 memset(zp.zp_salt, 0, ZIO_DATA_SALT_LEN);
2917 memset(zp.zp_iv, 0, ZIO_DATA_IV_LEN);
2918 memset(zp.zp_mac, 0, ZIO_DATA_MAC_LEN);
2920 zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2921 has_data ? abd_get_offset(pio->io_abd, pio->io_size -
2922 resid) : NULL, lsize, lsize, &zp,
2923 zio_write_gang_member_ready, NULL, NULL,
2924 zio_write_gang_done, &gn->gn_child[g], pio->io_priority,
2925 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2927 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2928 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2932 * Gang children won't throttle but we should
2933 * account for their work, so reserve an allocation
2934 * slot for them here.
2936 VERIFY(metaslab_class_throttle_reserve(mc,
2937 zp.zp_copies, cio->io_allocator, cio, flags));
2943 * Set pio's pipeline to just wait for zio to finish.
2945 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2948 * We didn't allocate this bp, so make sure it doesn't get unmarked.
2950 pio->io_flags &= ~ZIO_FLAG_FASTWRITE;
2958 * The zio_nop_write stage in the pipeline determines if allocating a
2959 * new bp is necessary. The nopwrite feature can handle writes in
2960 * either syncing or open context (i.e. zil writes) and as a result is
2961 * mutually exclusive with dedup.
2963 * By leveraging a cryptographically secure checksum, such as SHA256, we
2964 * can compare the checksums of the new data and the old to determine if
2965 * allocating a new block is required. Note that our requirements for
2966 * cryptographic strength are fairly weak: there can't be any accidental
2967 * hash collisions, but we don't need to be secure against intentional
2968 * (malicious) collisions. To trigger a nopwrite, you have to be able
2969 * to write the file to begin with, and triggering an incorrect (hash
2970 * collision) nopwrite is no worse than simply writing to the file.
2971 * That said, there are no known attacks against the checksum algorithms
2972 * used for nopwrite, assuming that the salt and the checksums
2973 * themselves remain secret.
2976 zio_nop_write(zio_t *zio)
2978 blkptr_t *bp = zio->io_bp;
2979 blkptr_t *bp_orig = &zio->io_bp_orig;
2980 zio_prop_t *zp = &zio->io_prop;
2982 ASSERT(BP_IS_HOLE(bp));
2983 ASSERT(BP_GET_LEVEL(bp) == 0);
2984 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2985 ASSERT(zp->zp_nopwrite);
2986 ASSERT(!zp->zp_dedup);
2987 ASSERT(zio->io_bp_override == NULL);
2988 ASSERT(IO_IS_ALLOCATING(zio));
2991 * Check to see if the original bp and the new bp have matching
2992 * characteristics (i.e. same checksum, compression algorithms, etc).
2993 * If they don't then just continue with the pipeline which will
2994 * allocate a new bp.
2996 if (BP_IS_HOLE(bp_orig) ||
2997 !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
2998 ZCHECKSUM_FLAG_NOPWRITE) ||
2999 BP_IS_ENCRYPTED(bp) || BP_IS_ENCRYPTED(bp_orig) ||
3000 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
3001 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
3002 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
3003 zp->zp_copies != BP_GET_NDVAS(bp_orig))
3007 * If the checksums match then reset the pipeline so that we
3008 * avoid allocating a new bp and issuing any I/O.
3010 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
3011 ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
3012 ZCHECKSUM_FLAG_NOPWRITE);
3013 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
3014 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
3015 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
3016 ASSERT3U(bp->blk_prop, ==, bp_orig->blk_prop);
3019 * If we're overwriting a block that is currently on an
3020 * indirect vdev, then ignore the nopwrite request and
3021 * allow a new block to be allocated on a concrete vdev.
3023 spa_config_enter(zio->io_spa, SCL_VDEV, FTAG, RW_READER);
3024 for (int d = 0; d < BP_GET_NDVAS(bp_orig); d++) {
3025 vdev_t *tvd = vdev_lookup_top(zio->io_spa,
3026 DVA_GET_VDEV(&bp_orig->blk_dva[d]));
3027 if (tvd->vdev_ops == &vdev_indirect_ops) {
3028 spa_config_exit(zio->io_spa, SCL_VDEV, FTAG);
3032 spa_config_exit(zio->io_spa, SCL_VDEV, FTAG);
3035 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3036 zio->io_flags |= ZIO_FLAG_NOPWRITE;
3043 * ==========================================================================
3045 * ==========================================================================
3048 zio_ddt_child_read_done(zio_t *zio)
3050 blkptr_t *bp = zio->io_bp;
3051 ddt_entry_t *dde = zio->io_private;
3053 zio_t *pio = zio_unique_parent(zio);
3055 mutex_enter(&pio->io_lock);
3056 ddp = ddt_phys_select(dde, bp);
3057 if (zio->io_error == 0)
3058 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
3060 if (zio->io_error == 0 && dde->dde_repair_abd == NULL)
3061 dde->dde_repair_abd = zio->io_abd;
3063 abd_free(zio->io_abd);
3064 mutex_exit(&pio->io_lock);
3068 zio_ddt_read_start(zio_t *zio)
3070 blkptr_t *bp = zio->io_bp;
3072 ASSERT(BP_GET_DEDUP(bp));
3073 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
3074 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3076 if (zio->io_child_error[ZIO_CHILD_DDT]) {
3077 ddt_t *ddt = ddt_select(zio->io_spa, bp);
3078 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
3079 ddt_phys_t *ddp = dde->dde_phys;
3080 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
3083 ASSERT(zio->io_vsd == NULL);
3086 if (ddp_self == NULL)
3089 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
3090 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
3092 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
3094 zio_nowait(zio_read(zio, zio->io_spa, &blk,
3095 abd_alloc_for_io(zio->io_size, B_TRUE),
3096 zio->io_size, zio_ddt_child_read_done, dde,
3097 zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) |
3098 ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark));
3103 zio_nowait(zio_read(zio, zio->io_spa, bp,
3104 zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority,
3105 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
3111 zio_ddt_read_done(zio_t *zio)
3113 blkptr_t *bp = zio->io_bp;
3115 if (zio_wait_for_children(zio, ZIO_CHILD_DDT_BIT, ZIO_WAIT_DONE)) {
3119 ASSERT(BP_GET_DEDUP(bp));
3120 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
3121 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3123 if (zio->io_child_error[ZIO_CHILD_DDT]) {
3124 ddt_t *ddt = ddt_select(zio->io_spa, bp);
3125 ddt_entry_t *dde = zio->io_vsd;
3127 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
3131 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
3132 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
3135 if (dde->dde_repair_abd != NULL) {
3136 abd_copy(zio->io_abd, dde->dde_repair_abd,
3138 zio->io_child_error[ZIO_CHILD_DDT] = 0;
3140 ddt_repair_done(ddt, dde);
3144 ASSERT(zio->io_vsd == NULL);
3150 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
3152 spa_t *spa = zio->io_spa;
3153 boolean_t do_raw = !!(zio->io_flags & ZIO_FLAG_RAW);
3155 ASSERT(!(zio->io_bp_override && do_raw));
3158 * Note: we compare the original data, not the transformed data,
3159 * because when zio->io_bp is an override bp, we will not have
3160 * pushed the I/O transforms. That's an important optimization
3161 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
3162 * However, we should never get a raw, override zio so in these
3163 * cases we can compare the io_abd directly. This is useful because
3164 * it allows us to do dedup verification even if we don't have access
3165 * to the original data (for instance, if the encryption keys aren't
3169 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
3170 zio_t *lio = dde->dde_lead_zio[p];
3172 if (lio != NULL && do_raw) {
3173 return (lio->io_size != zio->io_size ||
3174 abd_cmp(zio->io_abd, lio->io_abd) != 0);
3175 } else if (lio != NULL) {
3176 return (lio->io_orig_size != zio->io_orig_size ||
3177 abd_cmp(zio->io_orig_abd, lio->io_orig_abd) != 0);
3181 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
3182 ddt_phys_t *ddp = &dde->dde_phys[p];
3184 if (ddp->ddp_phys_birth != 0 && do_raw) {
3185 blkptr_t blk = *zio->io_bp;
3190 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
3191 psize = BP_GET_PSIZE(&blk);
3193 if (psize != zio->io_size)
3198 tmpabd = abd_alloc_for_io(psize, B_TRUE);
3200 error = zio_wait(zio_read(NULL, spa, &blk, tmpabd,
3201 psize, NULL, NULL, ZIO_PRIORITY_SYNC_READ,
3202 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
3203 ZIO_FLAG_RAW, &zio->io_bookmark));
3206 if (abd_cmp(tmpabd, zio->io_abd) != 0)
3207 error = SET_ERROR(ENOENT);
3212 return (error != 0);
3213 } else if (ddp->ddp_phys_birth != 0) {
3214 arc_buf_t *abuf = NULL;
3215 arc_flags_t aflags = ARC_FLAG_WAIT;
3216 blkptr_t blk = *zio->io_bp;
3219 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
3221 if (BP_GET_LSIZE(&blk) != zio->io_orig_size)
3226 error = arc_read(NULL, spa, &blk,
3227 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
3228 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
3229 &aflags, &zio->io_bookmark);
3232 if (abd_cmp_buf(zio->io_orig_abd, abuf->b_data,
3233 zio->io_orig_size) != 0)
3234 error = SET_ERROR(ENOENT);
3235 arc_buf_destroy(abuf, &abuf);
3239 return (error != 0);
3247 zio_ddt_child_write_ready(zio_t *zio)
3249 int p = zio->io_prop.zp_copies;
3250 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
3251 ddt_entry_t *dde = zio->io_private;
3252 ddt_phys_t *ddp = &dde->dde_phys[p];
3260 ASSERT(dde->dde_lead_zio[p] == zio);
3262 ddt_phys_fill(ddp, zio->io_bp);
3264 zio_link_t *zl = NULL;
3265 while ((pio = zio_walk_parents(zio, &zl)) != NULL)
3266 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
3272 zio_ddt_child_write_done(zio_t *zio)
3274 int p = zio->io_prop.zp_copies;
3275 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
3276 ddt_entry_t *dde = zio->io_private;
3277 ddt_phys_t *ddp = &dde->dde_phys[p];
3281 ASSERT(ddp->ddp_refcnt == 0);
3282 ASSERT(dde->dde_lead_zio[p] == zio);
3283 dde->dde_lead_zio[p] = NULL;
3285 if (zio->io_error == 0) {
3286 zio_link_t *zl = NULL;
3287 while (zio_walk_parents(zio, &zl) != NULL)
3288 ddt_phys_addref(ddp);
3290 ddt_phys_clear(ddp);
3297 zio_ddt_write(zio_t *zio)
3299 spa_t *spa = zio->io_spa;
3300 blkptr_t *bp = zio->io_bp;
3301 uint64_t txg = zio->io_txg;
3302 zio_prop_t *zp = &zio->io_prop;
3303 int p = zp->zp_copies;
3305 ddt_t *ddt = ddt_select(spa, bp);
3309 ASSERT(BP_GET_DEDUP(bp));
3310 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
3311 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
3312 ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW)));
3315 dde = ddt_lookup(ddt, bp, B_TRUE);
3316 ddp = &dde->dde_phys[p];
3318 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
3320 * If we're using a weak checksum, upgrade to a strong checksum
3321 * and try again. If we're already using a strong checksum,
3322 * we can't resolve it, so just convert to an ordinary write.
3323 * (And automatically e-mail a paper to Nature?)
3325 if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
3326 ZCHECKSUM_FLAG_DEDUP)) {
3327 zp->zp_checksum = spa_dedup_checksum(spa);
3328 zio_pop_transforms(zio);
3329 zio->io_stage = ZIO_STAGE_OPEN;
3332 zp->zp_dedup = B_FALSE;
3333 BP_SET_DEDUP(bp, B_FALSE);
3335 ASSERT(!BP_GET_DEDUP(bp));
3336 zio->io_pipeline = ZIO_WRITE_PIPELINE;
3341 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
3342 if (ddp->ddp_phys_birth != 0)
3343 ddt_bp_fill(ddp, bp, txg);
3344 if (dde->dde_lead_zio[p] != NULL)
3345 zio_add_child(zio, dde->dde_lead_zio[p]);
3347 ddt_phys_addref(ddp);
3348 } else if (zio->io_bp_override) {
3349 ASSERT(bp->blk_birth == txg);
3350 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
3351 ddt_phys_fill(ddp, bp);
3352 ddt_phys_addref(ddp);
3354 cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
3355 zio->io_orig_size, zio->io_orig_size, zp,
3356 zio_ddt_child_write_ready, NULL, NULL,
3357 zio_ddt_child_write_done, dde, zio->io_priority,
3358 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
3360 zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL);
3361 dde->dde_lead_zio[p] = cio;
3371 static ddt_entry_t *freedde; /* for debugging */
3374 zio_ddt_free(zio_t *zio)
3376 spa_t *spa = zio->io_spa;
3377 blkptr_t *bp = zio->io_bp;
3378 ddt_t *ddt = ddt_select(spa, bp);
3382 ASSERT(BP_GET_DEDUP(bp));
3383 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3386 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
3388 ddp = ddt_phys_select(dde, bp);
3390 ddt_phys_decref(ddp);
3398 * ==========================================================================
3399 * Allocate and free blocks
3400 * ==========================================================================
3404 zio_io_to_allocate(spa_t *spa, int allocator)
3408 ASSERT(MUTEX_HELD(&spa->spa_allocs[allocator].spaa_lock));
3410 zio = avl_first(&spa->spa_allocs[allocator].spaa_tree);
3414 ASSERT(IO_IS_ALLOCATING(zio));
3417 * Try to place a reservation for this zio. If we're unable to
3418 * reserve then we throttle.
3420 ASSERT3U(zio->io_allocator, ==, allocator);
3421 if (!metaslab_class_throttle_reserve(zio->io_metaslab_class,
3422 zio->io_prop.zp_copies, allocator, zio, 0)) {
3426 avl_remove(&spa->spa_allocs[allocator].spaa_tree, zio);
3427 ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE);
3433 zio_dva_throttle(zio_t *zio)
3435 spa_t *spa = zio->io_spa;
3437 metaslab_class_t *mc;
3439 /* locate an appropriate allocation class */
3440 mc = spa_preferred_class(spa, zio->io_size, zio->io_prop.zp_type,
3441 zio->io_prop.zp_level, zio->io_prop.zp_zpl_smallblk);
3443 if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE ||
3444 !mc->mc_alloc_throttle_enabled ||
3445 zio->io_child_type == ZIO_CHILD_GANG ||
3446 zio->io_flags & ZIO_FLAG_NODATA) {
3450 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3451 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3452 ASSERT3U(zio->io_queued_timestamp, >, 0);
3453 ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE);
3455 zbookmark_phys_t *bm = &zio->io_bookmark;
3457 * We want to try to use as many allocators as possible to help improve
3458 * performance, but we also want logically adjacent IOs to be physically
3459 * adjacent to improve sequential read performance. We chunk each object
3460 * into 2^20 block regions, and then hash based on the objset, object,
3461 * level, and region to accomplish both of these goals.
3463 int allocator = (uint_t)cityhash4(bm->zb_objset, bm->zb_object,
3464 bm->zb_level, bm->zb_blkid >> 20) % spa->spa_alloc_count;
3465 zio->io_allocator = allocator;
3466 zio->io_metaslab_class = mc;
3467 mutex_enter(&spa->spa_allocs[allocator].spaa_lock);
3468 avl_add(&spa->spa_allocs[allocator].spaa_tree, zio);
3469 nio = zio_io_to_allocate(spa, allocator);
3470 mutex_exit(&spa->spa_allocs[allocator].spaa_lock);
3475 zio_allocate_dispatch(spa_t *spa, int allocator)
3479 mutex_enter(&spa->spa_allocs[allocator].spaa_lock);
3480 zio = zio_io_to_allocate(spa, allocator);
3481 mutex_exit(&spa->spa_allocs[allocator].spaa_lock);
3485 ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE);
3486 ASSERT0(zio->io_error);
3487 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE);
3491 zio_dva_allocate(zio_t *zio)
3493 spa_t *spa = zio->io_spa;
3494 metaslab_class_t *mc;
3495 blkptr_t *bp = zio->io_bp;
3499 if (zio->io_gang_leader == NULL) {
3500 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3501 zio->io_gang_leader = zio;
3504 ASSERT(BP_IS_HOLE(bp));
3505 ASSERT0(BP_GET_NDVAS(bp));
3506 ASSERT3U(zio->io_prop.zp_copies, >, 0);
3507 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
3508 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
3510 flags |= (zio->io_flags & ZIO_FLAG_FASTWRITE) ? METASLAB_FASTWRITE : 0;
3511 if (zio->io_flags & ZIO_FLAG_NODATA)
3512 flags |= METASLAB_DONT_THROTTLE;
3513 if (zio->io_flags & ZIO_FLAG_GANG_CHILD)
3514 flags |= METASLAB_GANG_CHILD;
3515 if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE)
3516 flags |= METASLAB_ASYNC_ALLOC;
3519 * if not already chosen, locate an appropriate allocation class
3521 mc = zio->io_metaslab_class;
3523 mc = spa_preferred_class(spa, zio->io_size,
3524 zio->io_prop.zp_type, zio->io_prop.zp_level,
3525 zio->io_prop.zp_zpl_smallblk);
3526 zio->io_metaslab_class = mc;
3530 * Try allocating the block in the usual metaslab class.
3531 * If that's full, allocate it in the normal class.
3532 * If that's full, allocate as a gang block,
3533 * and if all are full, the allocation fails (which shouldn't happen).
3535 * Note that we do not fall back on embedded slog (ZIL) space, to
3536 * preserve unfragmented slog space, which is critical for decent
3537 * sync write performance. If a log allocation fails, we will fall
3538 * back to spa_sync() which is abysmal for performance.
3540 error = metaslab_alloc(spa, mc, zio->io_size, bp,
3541 zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
3542 &zio->io_alloc_list, zio, zio->io_allocator);
3545 * Fallback to normal class when an alloc class is full
3547 if (error == ENOSPC && mc != spa_normal_class(spa)) {
3549 * If throttling, transfer reservation over to normal class.
3550 * The io_allocator slot can remain the same even though we
3551 * are switching classes.
3553 if (mc->mc_alloc_throttle_enabled &&
3554 (zio->io_flags & ZIO_FLAG_IO_ALLOCATING)) {
3555 metaslab_class_throttle_unreserve(mc,
3556 zio->io_prop.zp_copies, zio->io_allocator, zio);
3557 zio->io_flags &= ~ZIO_FLAG_IO_ALLOCATING;
3559 VERIFY(metaslab_class_throttle_reserve(
3560 spa_normal_class(spa),
3561 zio->io_prop.zp_copies, zio->io_allocator, zio,
3562 flags | METASLAB_MUST_RESERVE));
3564 zio->io_metaslab_class = mc = spa_normal_class(spa);
3565 if (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC) {
3566 zfs_dbgmsg("%s: metaslab allocation failure, "
3567 "trying normal class: zio %px, size %llu, error %d",
3568 spa_name(spa), zio, (u_longlong_t)zio->io_size,
3572 error = metaslab_alloc(spa, mc, zio->io_size, bp,
3573 zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
3574 &zio->io_alloc_list, zio, zio->io_allocator);
3577 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE) {
3578 if (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC) {
3579 zfs_dbgmsg("%s: metaslab allocation failure, "
3580 "trying ganging: zio %px, size %llu, error %d",
3581 spa_name(spa), zio, (u_longlong_t)zio->io_size,
3584 return (zio_write_gang_block(zio, mc));
3587 if (error != ENOSPC ||
3588 (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC)) {
3589 zfs_dbgmsg("%s: metaslab allocation failure: zio %px, "
3590 "size %llu, error %d",
3591 spa_name(spa), zio, (u_longlong_t)zio->io_size,
3594 zio->io_error = error;
3601 zio_dva_free(zio_t *zio)
3603 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
3609 zio_dva_claim(zio_t *zio)
3613 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
3615 zio->io_error = error;
3621 * Undo an allocation. This is used by zio_done() when an I/O fails
3622 * and we want to give back the block we just allocated.
3623 * This handles both normal blocks and gang blocks.
3626 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
3628 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
3629 ASSERT(zio->io_bp_override == NULL);
3631 if (!BP_IS_HOLE(bp))
3632 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
3635 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
3636 zio_dva_unallocate(zio, gn->gn_child[g],
3637 &gn->gn_gbh->zg_blkptr[g]);
3643 * Try to allocate an intent log block. Return 0 on success, errno on failure.
3646 zio_alloc_zil(spa_t *spa, objset_t *os, uint64_t txg, blkptr_t *new_bp,
3647 uint64_t size, boolean_t *slog)
3650 zio_alloc_list_t io_alloc_list;
3652 ASSERT(txg > spa_syncing_txg(spa));
3654 metaslab_trace_init(&io_alloc_list);
3657 * Block pointer fields are useful to metaslabs for stats and debugging.
3658 * Fill in the obvious ones before calling into metaslab_alloc().
3660 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3661 BP_SET_PSIZE(new_bp, size);
3662 BP_SET_LEVEL(new_bp, 0);
3665 * When allocating a zil block, we don't have information about
3666 * the final destination of the block except the objset it's part
3667 * of, so we just hash the objset ID to pick the allocator to get
3670 int flags = METASLAB_FASTWRITE | METASLAB_ZIL;
3671 int allocator = (uint_t)cityhash4(0, 0, 0,
3672 os->os_dsl_dataset->ds_object) % spa->spa_alloc_count;
3673 error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1,
3674 txg, NULL, flags, &io_alloc_list, NULL, allocator);
3675 *slog = (error == 0);
3677 error = metaslab_alloc(spa, spa_embedded_log_class(spa), size,
3678 new_bp, 1, txg, NULL, flags,
3679 &io_alloc_list, NULL, allocator);
3682 error = metaslab_alloc(spa, spa_normal_class(spa), size,
3683 new_bp, 1, txg, NULL, flags,
3684 &io_alloc_list, NULL, allocator);
3686 metaslab_trace_fini(&io_alloc_list);
3689 BP_SET_LSIZE(new_bp, size);
3690 BP_SET_PSIZE(new_bp, size);
3691 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
3692 BP_SET_CHECKSUM(new_bp,
3693 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
3694 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
3695 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3696 BP_SET_LEVEL(new_bp, 0);
3697 BP_SET_DEDUP(new_bp, 0);
3698 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
3701 * encrypted blocks will require an IV and salt. We generate
3702 * these now since we will not be rewriting the bp at
3705 if (os->os_encrypted) {
3706 uint8_t iv[ZIO_DATA_IV_LEN];
3707 uint8_t salt[ZIO_DATA_SALT_LEN];
3709 BP_SET_CRYPT(new_bp, B_TRUE);
3710 VERIFY0(spa_crypt_get_salt(spa,
3711 dmu_objset_id(os), salt));
3712 VERIFY0(zio_crypt_generate_iv(iv));
3714 zio_crypt_encode_params_bp(new_bp, salt, iv);
3717 zfs_dbgmsg("%s: zil block allocation failure: "
3718 "size %llu, error %d", spa_name(spa), (u_longlong_t)size,
3726 * ==========================================================================
3727 * Read and write to physical devices
3728 * ==========================================================================
3732 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3733 * stops after this stage and will resume upon I/O completion.
3734 * However, there are instances where the vdev layer may need to
3735 * continue the pipeline when an I/O was not issued. Since the I/O
3736 * that was sent to the vdev layer might be different than the one
3737 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3738 * force the underlying vdev layers to call either zio_execute() or
3739 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3742 zio_vdev_io_start(zio_t *zio)
3744 vdev_t *vd = zio->io_vd;
3746 spa_t *spa = zio->io_spa;
3750 ASSERT(zio->io_error == 0);
3751 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
3754 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3755 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
3758 * The mirror_ops handle multiple DVAs in a single BP.
3760 vdev_mirror_ops.vdev_op_io_start(zio);
3764 ASSERT3P(zio->io_logical, !=, zio);
3765 if (zio->io_type == ZIO_TYPE_WRITE) {
3766 ASSERT(spa->spa_trust_config);
3769 * Note: the code can handle other kinds of writes,
3770 * but we don't expect them.
3772 if (zio->io_vd->vdev_noalloc) {
3773 ASSERT(zio->io_flags &
3774 (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL |
3775 ZIO_FLAG_RESILVER | ZIO_FLAG_INDUCE_DAMAGE));
3779 align = 1ULL << vd->vdev_top->vdev_ashift;
3781 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
3782 P2PHASE(zio->io_size, align) != 0) {
3783 /* Transform logical writes to be a full physical block size. */
3784 uint64_t asize = P2ROUNDUP(zio->io_size, align);
3785 abd_t *abuf = abd_alloc_sametype(zio->io_abd, asize);
3786 ASSERT(vd == vd->vdev_top);
3787 if (zio->io_type == ZIO_TYPE_WRITE) {
3788 abd_copy(abuf, zio->io_abd, zio->io_size);
3789 abd_zero_off(abuf, zio->io_size, asize - zio->io_size);
3791 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
3795 * If this is not a physical io, make sure that it is properly aligned
3796 * before proceeding.
3798 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
3799 ASSERT0(P2PHASE(zio->io_offset, align));
3800 ASSERT0(P2PHASE(zio->io_size, align));
3803 * For physical writes, we allow 512b aligned writes and assume
3804 * the device will perform a read-modify-write as necessary.
3806 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
3807 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
3810 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
3813 * If this is a repair I/O, and there's no self-healing involved --
3814 * that is, we're just resilvering what we expect to resilver --
3815 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3816 * This prevents spurious resilvering.
3818 * There are a few ways that we can end up creating these spurious
3821 * 1. A resilver i/o will be issued if any DVA in the BP has a
3822 * dirty DTL. The mirror code will issue resilver writes to
3823 * each DVA, including the one(s) that are not on vdevs with dirty
3826 * 2. With nested replication, which happens when we have a
3827 * "replacing" or "spare" vdev that's a child of a mirror or raidz.
3828 * For example, given mirror(replacing(A+B), C), it's likely that
3829 * only A is out of date (it's the new device). In this case, we'll
3830 * read from C, then use the data to resilver A+B -- but we don't
3831 * actually want to resilver B, just A. The top-level mirror has no
3832 * way to know this, so instead we just discard unnecessary repairs
3833 * as we work our way down the vdev tree.
3835 * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc.
3836 * The same logic applies to any form of nested replication: ditto
3837 * + mirror, RAID-Z + replacing, etc.
3839 * However, indirect vdevs point off to other vdevs which may have
3840 * DTL's, so we never bypass them. The child i/os on concrete vdevs
3841 * will be properly bypassed instead.
3843 * Leaf DTL_PARTIAL can be empty when a legitimate write comes from
3844 * a dRAID spare vdev. For example, when a dRAID spare is first
3845 * used, its spare blocks need to be written to but the leaf vdev's
3846 * of such blocks can have empty DTL_PARTIAL.
3848 * There seemed no clean way to allow such writes while bypassing
3849 * spurious ones. At this point, just avoid all bypassing for dRAID
3852 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3853 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
3854 zio->io_txg != 0 && /* not a delegated i/o */
3855 vd->vdev_ops != &vdev_indirect_ops &&
3856 vd->vdev_top->vdev_ops != &vdev_draid_ops &&
3857 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
3858 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3859 zio_vdev_io_bypass(zio);
3864 * Select the next best leaf I/O to process. Distributed spares are
3865 * excluded since they dispatch the I/O directly to a leaf vdev after
3866 * applying the dRAID mapping.
3868 if (vd->vdev_ops->vdev_op_leaf &&
3869 vd->vdev_ops != &vdev_draid_spare_ops &&
3870 (zio->io_type == ZIO_TYPE_READ ||
3871 zio->io_type == ZIO_TYPE_WRITE ||
3872 zio->io_type == ZIO_TYPE_TRIM)) {
3874 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio))
3877 if ((zio = vdev_queue_io(zio)) == NULL)
3880 if (!vdev_accessible(vd, zio)) {
3881 zio->io_error = SET_ERROR(ENXIO);
3885 zio->io_delay = gethrtime();
3888 vd->vdev_ops->vdev_op_io_start(zio);
3893 zio_vdev_io_done(zio_t *zio)
3895 vdev_t *vd = zio->io_vd;
3896 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
3897 boolean_t unexpected_error = B_FALSE;
3899 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3903 ASSERT(zio->io_type == ZIO_TYPE_READ ||
3904 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_TRIM);
3907 zio->io_delay = gethrtime() - zio->io_delay;
3909 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3910 vd->vdev_ops != &vdev_draid_spare_ops) {
3911 vdev_queue_io_done(zio);
3913 if (zio->io_type == ZIO_TYPE_WRITE)
3914 vdev_cache_write(zio);
3916 if (zio_injection_enabled && zio->io_error == 0)
3917 zio->io_error = zio_handle_device_injections(vd, zio,
3920 if (zio_injection_enabled && zio->io_error == 0)
3921 zio->io_error = zio_handle_label_injection(zio, EIO);
3923 if (zio->io_error && zio->io_type != ZIO_TYPE_TRIM) {
3924 if (!vdev_accessible(vd, zio)) {
3925 zio->io_error = SET_ERROR(ENXIO);
3927 unexpected_error = B_TRUE;
3932 ops->vdev_op_io_done(zio);
3934 if (unexpected_error && vd->vdev_remove_wanted == B_FALSE)
3935 VERIFY(vdev_probe(vd, zio) == NULL);
3941 * This function is used to change the priority of an existing zio that is
3942 * currently in-flight. This is used by the arc to upgrade priority in the
3943 * event that a demand read is made for a block that is currently queued
3944 * as a scrub or async read IO. Otherwise, the high priority read request
3945 * would end up having to wait for the lower priority IO.
3948 zio_change_priority(zio_t *pio, zio_priority_t priority)
3950 zio_t *cio, *cio_next;
3951 zio_link_t *zl = NULL;
3953 ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
3955 if (pio->io_vd != NULL && pio->io_vd->vdev_ops->vdev_op_leaf) {
3956 vdev_queue_change_io_priority(pio, priority);
3958 pio->io_priority = priority;
3961 mutex_enter(&pio->io_lock);
3962 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
3963 cio_next = zio_walk_children(pio, &zl);
3964 zio_change_priority(cio, priority);
3966 mutex_exit(&pio->io_lock);
3970 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3971 * disk, and use that to finish the checksum ereport later.
3974 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
3975 const abd_t *good_buf)
3977 /* no processing needed */
3978 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
3982 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr)
3984 void *abd = abd_alloc_sametype(zio->io_abd, zio->io_size);
3986 abd_copy(abd, zio->io_abd, zio->io_size);
3988 zcr->zcr_cbinfo = zio->io_size;
3989 zcr->zcr_cbdata = abd;
3990 zcr->zcr_finish = zio_vsd_default_cksum_finish;
3991 zcr->zcr_free = zio_abd_free;
3995 zio_vdev_io_assess(zio_t *zio)
3997 vdev_t *vd = zio->io_vd;
3999 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
4003 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
4004 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
4006 if (zio->io_vsd != NULL) {
4007 zio->io_vsd_ops->vsd_free(zio);
4011 if (zio_injection_enabled && zio->io_error == 0)
4012 zio->io_error = zio_handle_fault_injection(zio, EIO);
4015 * If the I/O failed, determine whether we should attempt to retry it.
4017 * On retry, we cut in line in the issue queue, since we don't want
4018 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
4020 if (zio->io_error && vd == NULL &&
4021 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
4022 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
4023 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
4025 zio->io_flags |= ZIO_FLAG_IO_RETRY |
4026 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
4027 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
4028 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
4029 zio_requeue_io_start_cut_in_line);
4034 * If we got an error on a leaf device, convert it to ENXIO
4035 * if the device is not accessible at all.
4037 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
4038 !vdev_accessible(vd, zio))
4039 zio->io_error = SET_ERROR(ENXIO);
4042 * If we can't write to an interior vdev (mirror or RAID-Z),
4043 * set vdev_cant_write so that we stop trying to allocate from it.
4045 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
4046 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
4047 vdev_dbgmsg(vd, "zio_vdev_io_assess(zio=%px) setting "
4048 "cant_write=TRUE due to write failure with ENXIO",
4050 vd->vdev_cant_write = B_TRUE;
4054 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
4055 * attempts will ever succeed. In this case we set a persistent
4056 * boolean flag so that we don't bother with it in the future.
4058 if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) &&
4059 zio->io_type == ZIO_TYPE_IOCTL &&
4060 zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL)
4061 vd->vdev_nowritecache = B_TRUE;
4064 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
4066 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
4067 zio->io_physdone != NULL) {
4068 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
4069 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
4070 zio->io_physdone(zio->io_logical);
4077 zio_vdev_io_reissue(zio_t *zio)
4079 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
4080 ASSERT(zio->io_error == 0);
4082 zio->io_stage >>= 1;
4086 zio_vdev_io_redone(zio_t *zio)
4088 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
4090 zio->io_stage >>= 1;
4094 zio_vdev_io_bypass(zio_t *zio)
4096 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
4097 ASSERT(zio->io_error == 0);
4099 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
4100 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
4104 * ==========================================================================
4105 * Encrypt and store encryption parameters
4106 * ==========================================================================
4111 * This function is used for ZIO_STAGE_ENCRYPT. It is responsible for
4112 * managing the storage of encryption parameters and passing them to the
4113 * lower-level encryption functions.
4116 zio_encrypt(zio_t *zio)
4118 zio_prop_t *zp = &zio->io_prop;
4119 spa_t *spa = zio->io_spa;
4120 blkptr_t *bp = zio->io_bp;
4121 uint64_t psize = BP_GET_PSIZE(bp);
4122 uint64_t dsobj = zio->io_bookmark.zb_objset;
4123 dmu_object_type_t ot = BP_GET_TYPE(bp);
4124 void *enc_buf = NULL;
4126 uint8_t salt[ZIO_DATA_SALT_LEN];
4127 uint8_t iv[ZIO_DATA_IV_LEN];
4128 uint8_t mac[ZIO_DATA_MAC_LEN];
4129 boolean_t no_crypt = B_FALSE;
4131 /* the root zio already encrypted the data */
4132 if (zio->io_child_type == ZIO_CHILD_GANG)
4135 /* only ZIL blocks are re-encrypted on rewrite */
4136 if (!IO_IS_ALLOCATING(zio) && ot != DMU_OT_INTENT_LOG)
4139 if (!(zp->zp_encrypt || BP_IS_ENCRYPTED(bp))) {
4140 BP_SET_CRYPT(bp, B_FALSE);
4144 /* if we are doing raw encryption set the provided encryption params */
4145 if (zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) {
4146 ASSERT0(BP_GET_LEVEL(bp));
4147 BP_SET_CRYPT(bp, B_TRUE);
4148 BP_SET_BYTEORDER(bp, zp->zp_byteorder);
4149 if (ot != DMU_OT_OBJSET)
4150 zio_crypt_encode_mac_bp(bp, zp->zp_mac);
4152 /* dnode blocks must be written out in the provided byteorder */
4153 if (zp->zp_byteorder != ZFS_HOST_BYTEORDER &&
4154 ot == DMU_OT_DNODE) {
4155 void *bswap_buf = zio_buf_alloc(psize);
4156 abd_t *babd = abd_get_from_buf(bswap_buf, psize);
4158 ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
4159 abd_copy_to_buf(bswap_buf, zio->io_abd, psize);
4160 dmu_ot_byteswap[DMU_OT_BYTESWAP(ot)].ob_func(bswap_buf,
4163 abd_take_ownership_of_buf(babd, B_TRUE);
4164 zio_push_transform(zio, babd, psize, psize, NULL);
4167 if (DMU_OT_IS_ENCRYPTED(ot))
4168 zio_crypt_encode_params_bp(bp, zp->zp_salt, zp->zp_iv);
4172 /* indirect blocks only maintain a cksum of the lower level MACs */
4173 if (BP_GET_LEVEL(bp) > 0) {
4174 BP_SET_CRYPT(bp, B_TRUE);
4175 VERIFY0(zio_crypt_do_indirect_mac_checksum_abd(B_TRUE,
4176 zio->io_orig_abd, BP_GET_LSIZE(bp), BP_SHOULD_BYTESWAP(bp),
4178 zio_crypt_encode_mac_bp(bp, mac);
4183 * Objset blocks are a special case since they have 2 256-bit MACs
4184 * embedded within them.
4186 if (ot == DMU_OT_OBJSET) {
4187 ASSERT0(DMU_OT_IS_ENCRYPTED(ot));
4188 ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
4189 BP_SET_CRYPT(bp, B_TRUE);
4190 VERIFY0(spa_do_crypt_objset_mac_abd(B_TRUE, spa, dsobj,
4191 zio->io_abd, psize, BP_SHOULD_BYTESWAP(bp)));
4195 /* unencrypted object types are only authenticated with a MAC */
4196 if (!DMU_OT_IS_ENCRYPTED(ot)) {
4197 BP_SET_CRYPT(bp, B_TRUE);
4198 VERIFY0(spa_do_crypt_mac_abd(B_TRUE, spa, dsobj,
4199 zio->io_abd, psize, mac));
4200 zio_crypt_encode_mac_bp(bp, mac);
4205 * Later passes of sync-to-convergence may decide to rewrite data
4206 * in place to avoid more disk reallocations. This presents a problem
4207 * for encryption because this constitutes rewriting the new data with
4208 * the same encryption key and IV. However, this only applies to blocks
4209 * in the MOS (particularly the spacemaps) and we do not encrypt the
4210 * MOS. We assert that the zio is allocating or an intent log write
4213 ASSERT(IO_IS_ALLOCATING(zio) || ot == DMU_OT_INTENT_LOG);
4214 ASSERT(BP_GET_LEVEL(bp) == 0 || ot == DMU_OT_INTENT_LOG);
4215 ASSERT(spa_feature_is_active(spa, SPA_FEATURE_ENCRYPTION));
4216 ASSERT3U(psize, !=, 0);
4218 enc_buf = zio_buf_alloc(psize);
4219 eabd = abd_get_from_buf(enc_buf, psize);
4220 abd_take_ownership_of_buf(eabd, B_TRUE);
4223 * For an explanation of what encryption parameters are stored
4224 * where, see the block comment in zio_crypt.c.
4226 if (ot == DMU_OT_INTENT_LOG) {
4227 zio_crypt_decode_params_bp(bp, salt, iv);
4229 BP_SET_CRYPT(bp, B_TRUE);
4232 /* Perform the encryption. This should not fail */
4233 VERIFY0(spa_do_crypt_abd(B_TRUE, spa, &zio->io_bookmark,
4234 BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp),
4235 salt, iv, mac, psize, zio->io_abd, eabd, &no_crypt));
4237 /* encode encryption metadata into the bp */
4238 if (ot == DMU_OT_INTENT_LOG) {
4240 * ZIL blocks store the MAC in the embedded checksum, so the
4241 * transform must always be applied.
4243 zio_crypt_encode_mac_zil(enc_buf, mac);
4244 zio_push_transform(zio, eabd, psize, psize, NULL);
4246 BP_SET_CRYPT(bp, B_TRUE);
4247 zio_crypt_encode_params_bp(bp, salt, iv);
4248 zio_crypt_encode_mac_bp(bp, mac);
4251 ASSERT3U(ot, ==, DMU_OT_DNODE);
4254 zio_push_transform(zio, eabd, psize, psize, NULL);
4262 * ==========================================================================
4263 * Generate and verify checksums
4264 * ==========================================================================
4267 zio_checksum_generate(zio_t *zio)
4269 blkptr_t *bp = zio->io_bp;
4270 enum zio_checksum checksum;
4274 * This is zio_write_phys().
4275 * We're either generating a label checksum, or none at all.
4277 checksum = zio->io_prop.zp_checksum;
4279 if (checksum == ZIO_CHECKSUM_OFF)
4282 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
4284 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
4285 ASSERT(!IO_IS_ALLOCATING(zio));
4286 checksum = ZIO_CHECKSUM_GANG_HEADER;
4288 checksum = BP_GET_CHECKSUM(bp);
4292 zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size);
4298 zio_checksum_verify(zio_t *zio)
4300 zio_bad_cksum_t info;
4301 blkptr_t *bp = zio->io_bp;
4304 ASSERT(zio->io_vd != NULL);
4308 * This is zio_read_phys().
4309 * We're either verifying a label checksum, or nothing at all.
4311 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
4314 ASSERT3U(zio->io_prop.zp_checksum, ==, ZIO_CHECKSUM_LABEL);
4317 if ((error = zio_checksum_error(zio, &info)) != 0) {
4318 zio->io_error = error;
4319 if (error == ECKSUM &&
4320 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
4321 mutex_enter(&zio->io_vd->vdev_stat_lock);
4322 zio->io_vd->vdev_stat.vs_checksum_errors++;
4323 mutex_exit(&zio->io_vd->vdev_stat_lock);
4324 (void) zfs_ereport_start_checksum(zio->io_spa,
4325 zio->io_vd, &zio->io_bookmark, zio,
4326 zio->io_offset, zio->io_size, &info);
4334 * Called by RAID-Z to ensure we don't compute the checksum twice.
4337 zio_checksum_verified(zio_t *zio)
4339 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
4343 * ==========================================================================
4344 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
4345 * An error of 0 indicates success. ENXIO indicates whole-device failure,
4346 * which may be transient (e.g. unplugged) or permanent. ECKSUM and EIO
4347 * indicate errors that are specific to one I/O, and most likely permanent.
4348 * Any other error is presumed to be worse because we weren't expecting it.
4349 * ==========================================================================
4352 zio_worst_error(int e1, int e2)
4354 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
4357 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
4358 if (e1 == zio_error_rank[r1])
4361 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
4362 if (e2 == zio_error_rank[r2])
4365 return (r1 > r2 ? e1 : e2);
4369 * ==========================================================================
4371 * ==========================================================================
4374 zio_ready(zio_t *zio)
4376 blkptr_t *bp = zio->io_bp;
4377 zio_t *pio, *pio_next;
4378 zio_link_t *zl = NULL;
4380 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT | ZIO_CHILD_DDT_BIT,
4385 if (zio->io_ready) {
4386 ASSERT(IO_IS_ALLOCATING(zio));
4387 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
4388 (zio->io_flags & ZIO_FLAG_NOPWRITE));
4389 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
4394 if (bp != NULL && bp != &zio->io_bp_copy)
4395 zio->io_bp_copy = *bp;
4397 if (zio->io_error != 0) {
4398 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
4400 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
4401 ASSERT(IO_IS_ALLOCATING(zio));
4402 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
4403 ASSERT(zio->io_metaslab_class != NULL);
4406 * We were unable to allocate anything, unreserve and
4407 * issue the next I/O to allocate.
4409 metaslab_class_throttle_unreserve(
4410 zio->io_metaslab_class, zio->io_prop.zp_copies,
4411 zio->io_allocator, zio);
4412 zio_allocate_dispatch(zio->io_spa, zio->io_allocator);
4416 mutex_enter(&zio->io_lock);
4417 zio->io_state[ZIO_WAIT_READY] = 1;
4418 pio = zio_walk_parents(zio, &zl);
4419 mutex_exit(&zio->io_lock);
4422 * As we notify zio's parents, new parents could be added.
4423 * New parents go to the head of zio's io_parent_list, however,
4424 * so we will (correctly) not notify them. The remainder of zio's
4425 * io_parent_list, from 'pio_next' onward, cannot change because
4426 * all parents must wait for us to be done before they can be done.
4428 for (; pio != NULL; pio = pio_next) {
4429 pio_next = zio_walk_parents(zio, &zl);
4430 zio_notify_parent(pio, zio, ZIO_WAIT_READY, NULL);
4433 if (zio->io_flags & ZIO_FLAG_NODATA) {
4434 if (BP_IS_GANG(bp)) {
4435 zio->io_flags &= ~ZIO_FLAG_NODATA;
4437 ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE);
4438 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
4442 if (zio_injection_enabled &&
4443 zio->io_spa->spa_syncing_txg == zio->io_txg)
4444 zio_handle_ignored_writes(zio);
4450 * Update the allocation throttle accounting.
4453 zio_dva_throttle_done(zio_t *zio)
4455 zio_t *lio __maybe_unused = zio->io_logical;
4456 zio_t *pio = zio_unique_parent(zio);
4457 vdev_t *vd = zio->io_vd;
4458 int flags = METASLAB_ASYNC_ALLOC;
4460 ASSERT3P(zio->io_bp, !=, NULL);
4461 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
4462 ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE);
4463 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
4465 ASSERT3P(vd, ==, vd->vdev_top);
4466 ASSERT(zio_injection_enabled || !(zio->io_flags & ZIO_FLAG_IO_RETRY));
4467 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
4468 ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING);
4469 ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE));
4470 ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA));
4473 * Parents of gang children can have two flavors -- ones that
4474 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
4475 * and ones that allocated the constituent blocks. The allocation
4476 * throttle needs to know the allocating parent zio so we must find
4479 if (pio->io_child_type == ZIO_CHILD_GANG) {
4481 * If our parent is a rewrite gang child then our grandparent
4482 * would have been the one that performed the allocation.
4484 if (pio->io_flags & ZIO_FLAG_IO_REWRITE)
4485 pio = zio_unique_parent(pio);
4486 flags |= METASLAB_GANG_CHILD;
4489 ASSERT(IO_IS_ALLOCATING(pio));
4490 ASSERT3P(zio, !=, zio->io_logical);
4491 ASSERT(zio->io_logical != NULL);
4492 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
4493 ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE);
4494 ASSERT(zio->io_metaslab_class != NULL);
4496 mutex_enter(&pio->io_lock);
4497 metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags,
4498 pio->io_allocator, B_TRUE);
4499 mutex_exit(&pio->io_lock);
4501 metaslab_class_throttle_unreserve(zio->io_metaslab_class, 1,
4502 pio->io_allocator, pio);
4505 * Call into the pipeline to see if there is more work that
4506 * needs to be done. If there is work to be done it will be
4507 * dispatched to another taskq thread.
4509 zio_allocate_dispatch(zio->io_spa, pio->io_allocator);
4513 zio_done(zio_t *zio)
4516 * Always attempt to keep stack usage minimal here since
4517 * we can be called recursively up to 19 levels deep.
4519 const uint64_t psize = zio->io_size;
4520 zio_t *pio, *pio_next;
4521 zio_link_t *zl = NULL;
4524 * If our children haven't all completed,
4525 * wait for them and then repeat this pipeline stage.
4527 if (zio_wait_for_children(zio, ZIO_CHILD_ALL_BITS, ZIO_WAIT_DONE)) {
4532 * If the allocation throttle is enabled, then update the accounting.
4533 * We only track child I/Os that are part of an allocating async
4534 * write. We must do this since the allocation is performed
4535 * by the logical I/O but the actual write is done by child I/Os.
4537 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING &&
4538 zio->io_child_type == ZIO_CHILD_VDEV) {
4539 ASSERT(zio->io_metaslab_class != NULL);
4540 ASSERT(zio->io_metaslab_class->mc_alloc_throttle_enabled);
4541 zio_dva_throttle_done(zio);
4545 * If the allocation throttle is enabled, verify that
4546 * we have decremented the refcounts for every I/O that was throttled.
4548 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
4549 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
4550 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
4551 ASSERT(zio->io_bp != NULL);
4553 metaslab_group_alloc_verify(zio->io_spa, zio->io_bp, zio,
4555 VERIFY(zfs_refcount_not_held(&zio->io_metaslab_class->
4556 mc_allocator[zio->io_allocator].mca_alloc_slots, zio));
4560 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
4561 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
4562 ASSERT(zio->io_children[c][w] == 0);
4564 if (zio->io_bp != NULL && !BP_IS_EMBEDDED(zio->io_bp)) {
4565 ASSERT(zio->io_bp->blk_pad[0] == 0);
4566 ASSERT(zio->io_bp->blk_pad[1] == 0);
4567 ASSERT(memcmp(zio->io_bp, &zio->io_bp_copy,
4568 sizeof (blkptr_t)) == 0 ||
4569 (zio->io_bp == zio_unique_parent(zio)->io_bp));
4570 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) &&
4571 zio->io_bp_override == NULL &&
4572 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
4573 ASSERT3U(zio->io_prop.zp_copies, <=,
4574 BP_GET_NDVAS(zio->io_bp));
4575 ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 ||
4576 (BP_COUNT_GANG(zio->io_bp) ==
4577 BP_GET_NDVAS(zio->io_bp)));
4579 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
4580 VERIFY(BP_EQUAL(zio->io_bp, &zio->io_bp_orig));
4584 * If there were child vdev/gang/ddt errors, they apply to us now.
4586 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
4587 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
4588 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
4591 * If the I/O on the transformed data was successful, generate any
4592 * checksum reports now while we still have the transformed data.
4594 if (zio->io_error == 0) {
4595 while (zio->io_cksum_report != NULL) {
4596 zio_cksum_report_t *zcr = zio->io_cksum_report;
4597 uint64_t align = zcr->zcr_align;
4598 uint64_t asize = P2ROUNDUP(psize, align);
4599 abd_t *adata = zio->io_abd;
4601 if (adata != NULL && asize != psize) {
4602 adata = abd_alloc(asize, B_TRUE);
4603 abd_copy(adata, zio->io_abd, psize);
4604 abd_zero_off(adata, psize, asize - psize);
4607 zio->io_cksum_report = zcr->zcr_next;
4608 zcr->zcr_next = NULL;
4609 zcr->zcr_finish(zcr, adata);
4610 zfs_ereport_free_checksum(zcr);
4612 if (adata != NULL && asize != psize)
4617 zio_pop_transforms(zio); /* note: may set zio->io_error */
4619 vdev_stat_update(zio, psize);
4622 * If this I/O is attached to a particular vdev is slow, exceeding
4623 * 30 seconds to complete, post an error described the I/O delay.
4624 * We ignore these errors if the device is currently unavailable.
4626 if (zio->io_delay >= MSEC2NSEC(zio_slow_io_ms)) {
4627 if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd)) {
4629 * We want to only increment our slow IO counters if
4630 * the IO is valid (i.e. not if the drive is removed).
4632 * zfs_ereport_post() will also do these checks, but
4633 * it can also ratelimit and have other failures, so we
4634 * need to increment the slow_io counters independent
4637 if (zfs_ereport_is_valid(FM_EREPORT_ZFS_DELAY,
4638 zio->io_spa, zio->io_vd, zio)) {
4639 mutex_enter(&zio->io_vd->vdev_stat_lock);
4640 zio->io_vd->vdev_stat.vs_slow_ios++;
4641 mutex_exit(&zio->io_vd->vdev_stat_lock);
4643 (void) zfs_ereport_post(FM_EREPORT_ZFS_DELAY,
4644 zio->io_spa, zio->io_vd, &zio->io_bookmark,
4650 if (zio->io_error) {
4652 * If this I/O is attached to a particular vdev,
4653 * generate an error message describing the I/O failure
4654 * at the block level. We ignore these errors if the
4655 * device is currently unavailable.
4657 if (zio->io_error != ECKSUM && zio->io_vd != NULL &&
4658 !vdev_is_dead(zio->io_vd)) {
4659 int ret = zfs_ereport_post(FM_EREPORT_ZFS_IO,
4660 zio->io_spa, zio->io_vd, &zio->io_bookmark, zio, 0);
4661 if (ret != EALREADY) {
4662 mutex_enter(&zio->io_vd->vdev_stat_lock);
4663 if (zio->io_type == ZIO_TYPE_READ)
4664 zio->io_vd->vdev_stat.vs_read_errors++;
4665 else if (zio->io_type == ZIO_TYPE_WRITE)
4666 zio->io_vd->vdev_stat.vs_write_errors++;
4667 mutex_exit(&zio->io_vd->vdev_stat_lock);
4671 if ((zio->io_error == EIO || !(zio->io_flags &
4672 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
4673 zio == zio->io_logical) {
4675 * For logical I/O requests, tell the SPA to log the
4676 * error and generate a logical data ereport.
4678 spa_log_error(zio->io_spa, &zio->io_bookmark);
4679 (void) zfs_ereport_post(FM_EREPORT_ZFS_DATA,
4680 zio->io_spa, NULL, &zio->io_bookmark, zio, 0);
4684 if (zio->io_error && zio == zio->io_logical) {
4686 * Determine whether zio should be reexecuted. This will
4687 * propagate all the way to the root via zio_notify_parent().
4689 ASSERT(zio->io_vd == NULL && zio->io_bp != NULL);
4690 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
4692 if (IO_IS_ALLOCATING(zio) &&
4693 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
4694 if (zio->io_error != ENOSPC)
4695 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
4697 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4700 if ((zio->io_type == ZIO_TYPE_READ ||
4701 zio->io_type == ZIO_TYPE_FREE) &&
4702 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
4703 zio->io_error == ENXIO &&
4704 spa_load_state(zio->io_spa) == SPA_LOAD_NONE &&
4705 spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE)
4706 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4708 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
4709 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4712 * Here is a possibly good place to attempt to do
4713 * either combinatorial reconstruction or error correction
4714 * based on checksums. It also might be a good place
4715 * to send out preliminary ereports before we suspend
4721 * If there were logical child errors, they apply to us now.
4722 * We defer this until now to avoid conflating logical child
4723 * errors with errors that happened to the zio itself when
4724 * updating vdev stats and reporting FMA events above.
4726 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
4728 if ((zio->io_error || zio->io_reexecute) &&
4729 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
4730 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
4731 zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp);
4733 zio_gang_tree_free(&zio->io_gang_tree);
4736 * Godfather I/Os should never suspend.
4738 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
4739 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
4740 zio->io_reexecute &= ~ZIO_REEXECUTE_SUSPEND;
4742 if (zio->io_reexecute) {
4744 * This is a logical I/O that wants to reexecute.
4746 * Reexecute is top-down. When an i/o fails, if it's not
4747 * the root, it simply notifies its parent and sticks around.
4748 * The parent, seeing that it still has children in zio_done(),
4749 * does the same. This percolates all the way up to the root.
4750 * The root i/o will reexecute or suspend the entire tree.
4752 * This approach ensures that zio_reexecute() honors
4753 * all the original i/o dependency relationships, e.g.
4754 * parents not executing until children are ready.
4756 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
4758 zio->io_gang_leader = NULL;
4760 mutex_enter(&zio->io_lock);
4761 zio->io_state[ZIO_WAIT_DONE] = 1;
4762 mutex_exit(&zio->io_lock);
4765 * "The Godfather" I/O monitors its children but is
4766 * not a true parent to them. It will track them through
4767 * the pipeline but severs its ties whenever they get into
4768 * trouble (e.g. suspended). This allows "The Godfather"
4769 * I/O to return status without blocking.
4772 for (pio = zio_walk_parents(zio, &zl); pio != NULL;
4774 zio_link_t *remove_zl = zl;
4775 pio_next = zio_walk_parents(zio, &zl);
4777 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
4778 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
4779 zio_remove_child(pio, zio, remove_zl);
4781 * This is a rare code path, so we don't
4782 * bother with "next_to_execute".
4784 zio_notify_parent(pio, zio, ZIO_WAIT_DONE,
4789 if ((pio = zio_unique_parent(zio)) != NULL) {
4791 * We're not a root i/o, so there's nothing to do
4792 * but notify our parent. Don't propagate errors
4793 * upward since we haven't permanently failed yet.
4795 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
4796 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
4798 * This is a rare code path, so we don't bother with
4799 * "next_to_execute".
4801 zio_notify_parent(pio, zio, ZIO_WAIT_DONE, NULL);
4802 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
4804 * We'd fail again if we reexecuted now, so suspend
4805 * until conditions improve (e.g. device comes online).
4807 zio_suspend(zio->io_spa, zio, ZIO_SUSPEND_IOERR);
4810 * Reexecution is potentially a huge amount of work.
4811 * Hand it off to the otherwise-unused claim taskq.
4813 ASSERT(taskq_empty_ent(&zio->io_tqent));
4814 spa_taskq_dispatch_ent(zio->io_spa,
4815 ZIO_TYPE_CLAIM, ZIO_TASKQ_ISSUE,
4816 zio_reexecute, zio, 0, &zio->io_tqent);
4821 ASSERT(zio->io_child_count == 0);
4822 ASSERT(zio->io_reexecute == 0);
4823 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
4826 * Report any checksum errors, since the I/O is complete.
4828 while (zio->io_cksum_report != NULL) {
4829 zio_cksum_report_t *zcr = zio->io_cksum_report;
4830 zio->io_cksum_report = zcr->zcr_next;
4831 zcr->zcr_next = NULL;
4832 zcr->zcr_finish(zcr, NULL);
4833 zfs_ereport_free_checksum(zcr);
4836 if (zio->io_flags & ZIO_FLAG_FASTWRITE && zio->io_bp &&
4837 !BP_IS_HOLE(zio->io_bp) && !BP_IS_EMBEDDED(zio->io_bp) &&
4838 !(zio->io_flags & ZIO_FLAG_NOPWRITE)) {
4839 metaslab_fastwrite_unmark(zio->io_spa, zio->io_bp);
4843 * It is the responsibility of the done callback to ensure that this
4844 * particular zio is no longer discoverable for adoption, and as
4845 * such, cannot acquire any new parents.
4850 mutex_enter(&zio->io_lock);
4851 zio->io_state[ZIO_WAIT_DONE] = 1;
4852 mutex_exit(&zio->io_lock);
4855 * We are done executing this zio. We may want to execute a parent
4856 * next. See the comment in zio_notify_parent().
4858 zio_t *next_to_execute = NULL;
4860 for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) {
4861 zio_link_t *remove_zl = zl;
4862 pio_next = zio_walk_parents(zio, &zl);
4863 zio_remove_child(pio, zio, remove_zl);
4864 zio_notify_parent(pio, zio, ZIO_WAIT_DONE, &next_to_execute);
4867 if (zio->io_waiter != NULL) {
4868 mutex_enter(&zio->io_lock);
4869 zio->io_executor = NULL;
4870 cv_broadcast(&zio->io_cv);
4871 mutex_exit(&zio->io_lock);
4876 return (next_to_execute);
4880 * ==========================================================================
4881 * I/O pipeline definition
4882 * ==========================================================================
4884 static zio_pipe_stage_t *zio_pipeline[] = {
4892 zio_checksum_generate,
4908 zio_checksum_verify,
4916 * Compare two zbookmark_phys_t's to see which we would reach first in a
4917 * pre-order traversal of the object tree.
4919 * This is simple in every case aside from the meta-dnode object. For all other
4920 * objects, we traverse them in order (object 1 before object 2, and so on).
4921 * However, all of these objects are traversed while traversing object 0, since
4922 * the data it points to is the list of objects. Thus, we need to convert to a
4923 * canonical representation so we can compare meta-dnode bookmarks to
4924 * non-meta-dnode bookmarks.
4926 * We do this by calculating "equivalents" for each field of the zbookmark.
4927 * zbookmarks outside of the meta-dnode use their own object and level, and
4928 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
4929 * blocks this bookmark refers to) by multiplying their blkid by their span
4930 * (the number of L0 blocks contained within one block at their level).
4931 * zbookmarks inside the meta-dnode calculate their object equivalent
4932 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
4933 * level + 1<<31 (any value larger than a level could ever be) for their level.
4934 * This causes them to always compare before a bookmark in their object
4935 * equivalent, compare appropriately to bookmarks in other objects, and to
4936 * compare appropriately to other bookmarks in the meta-dnode.
4939 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
4940 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
4943 * These variables represent the "equivalent" values for the zbookmark,
4944 * after converting zbookmarks inside the meta dnode to their
4945 * normal-object equivalents.
4947 uint64_t zb1obj, zb2obj;
4948 uint64_t zb1L0, zb2L0;
4949 uint64_t zb1level, zb2level;
4951 if (zb1->zb_object == zb2->zb_object &&
4952 zb1->zb_level == zb2->zb_level &&
4953 zb1->zb_blkid == zb2->zb_blkid)
4956 IMPLY(zb1->zb_level > 0, ibs1 >= SPA_MINBLOCKSHIFT);
4957 IMPLY(zb2->zb_level > 0, ibs2 >= SPA_MINBLOCKSHIFT);
4960 * BP_SPANB calculates the span in blocks.
4962 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
4963 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
4965 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
4966 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4968 zb1level = zb1->zb_level + COMPARE_META_LEVEL;
4970 zb1obj = zb1->zb_object;
4971 zb1level = zb1->zb_level;
4974 if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
4975 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4977 zb2level = zb2->zb_level + COMPARE_META_LEVEL;
4979 zb2obj = zb2->zb_object;
4980 zb2level = zb2->zb_level;
4983 /* Now that we have a canonical representation, do the comparison. */
4984 if (zb1obj != zb2obj)
4985 return (zb1obj < zb2obj ? -1 : 1);
4986 else if (zb1L0 != zb2L0)
4987 return (zb1L0 < zb2L0 ? -1 : 1);
4988 else if (zb1level != zb2level)
4989 return (zb1level > zb2level ? -1 : 1);
4991 * This can (theoretically) happen if the bookmarks have the same object
4992 * and level, but different blkids, if the block sizes are not the same.
4993 * There is presently no way to change the indirect block sizes
4999 * This function checks the following: given that last_block is the place that
5000 * our traversal stopped last time, does that guarantee that we've visited
5001 * every node under subtree_root? Therefore, we can't just use the raw output
5002 * of zbookmark_compare. We have to pass in a modified version of
5003 * subtree_root; by incrementing the block id, and then checking whether
5004 * last_block is before or equal to that, we can tell whether or not having
5005 * visited last_block implies that all of subtree_root's children have been
5009 zbookmark_subtree_completed(const dnode_phys_t *dnp,
5010 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
5012 zbookmark_phys_t mod_zb = *subtree_root;
5014 ASSERT0(last_block->zb_level);
5016 /* The objset_phys_t isn't before anything. */
5021 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
5022 * data block size in sectors, because that variable is only used if
5023 * the bookmark refers to a block in the meta-dnode. Since we don't
5024 * know without examining it what object it refers to, and there's no
5025 * harm in passing in this value in other cases, we always pass it in.
5027 * We pass in 0 for the indirect block size shift because zb2 must be
5028 * level 0. The indirect block size is only used to calculate the span
5029 * of the bookmark, but since the bookmark must be level 0, the span is
5030 * always 1, so the math works out.
5032 * If you make changes to how the zbookmark_compare code works, be sure
5033 * to make sure that this code still works afterwards.
5035 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
5036 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,
5041 * This function is similar to zbookmark_subtree_completed(), but returns true
5042 * if subtree_root is equal or ahead of last_block, i.e. still to be done.
5045 zbookmark_subtree_tbd(const dnode_phys_t *dnp,
5046 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
5048 ASSERT0(last_block->zb_level);
5051 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
5052 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, subtree_root,
5056 EXPORT_SYMBOL(zio_type_name);
5057 EXPORT_SYMBOL(zio_buf_alloc);
5058 EXPORT_SYMBOL(zio_data_buf_alloc);
5059 EXPORT_SYMBOL(zio_buf_free);
5060 EXPORT_SYMBOL(zio_data_buf_free);
5062 ZFS_MODULE_PARAM(zfs_zio, zio_, slow_io_ms, INT, ZMOD_RW,
5063 "Max I/O completion time (milliseconds) before marking it as slow");
5065 ZFS_MODULE_PARAM(zfs_zio, zio_, requeue_io_start_cut_in_line, INT, ZMOD_RW,
5066 "Prioritize requeued I/O");
5068 ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_deferred_free, UINT, ZMOD_RW,
5069 "Defer frees starting in this pass");
5071 ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_dont_compress, UINT, ZMOD_RW,
5072 "Don't compress starting in this pass");
5074 ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_rewrite, UINT, ZMOD_RW,
5075 "Rewrite new bps starting in this pass");
5077 ZFS_MODULE_PARAM(zfs_zio, zio_, dva_throttle_enabled, INT, ZMOD_RW,
5078 "Throttle block allocations in the ZIO pipeline");
5080 ZFS_MODULE_PARAM(zfs_zio, zio_, deadman_log_all, INT, ZMOD_RW,
5081 "Log all slow ZIOs, not just those with vdevs");