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>
46 #include <sys/blkptr.h>
47 #include <sys/zfeature.h>
48 #include <sys/dsl_scan.h>
49 #include <sys/metaslab_impl.h>
51 #include <sys/trace_zfs.h>
53 #include <sys/dsl_crypt.h>
57 * ==========================================================================
58 * I/O type descriptions
59 * ==========================================================================
61 const char *const zio_type_name[ZIO_TYPES] = {
63 * Note: Linux kernel thread name length is limited
64 * so these names will differ from upstream open zfs.
66 "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl", "z_trim"
69 int zio_dva_throttle_enabled = B_TRUE;
70 static int zio_deadman_log_all = B_FALSE;
73 * ==========================================================================
75 * ==========================================================================
77 static kmem_cache_t *zio_cache;
78 static kmem_cache_t *zio_link_cache;
79 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
80 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
81 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
82 static uint64_t zio_buf_cache_allocs[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
83 static uint64_t zio_buf_cache_frees[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
86 /* Mark IOs as "slow" if they take longer than 30 seconds */
87 static uint_t zio_slow_io_ms = (30 * MILLISEC);
89 #define BP_SPANB(indblkshift, level) \
90 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
91 #define COMPARE_META_LEVEL 0x80000000ul
93 * The following actions directly effect the spa's sync-to-convergence logic.
94 * The values below define the sync pass when we start performing the action.
95 * Care should be taken when changing these values as they directly impact
96 * spa_sync() performance. Tuning these values may introduce subtle performance
97 * pathologies and should only be done in the context of performance analysis.
98 * These tunables will eventually be removed and replaced with #defines once
99 * enough analysis has been done to determine optimal values.
101 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
102 * regular blocks are not deferred.
104 * Starting in sync pass 8 (zfs_sync_pass_dont_compress), we disable
105 * compression (including of metadata). In practice, we don't have this
106 * many sync passes, so this has no effect.
108 * The original intent was that disabling compression would help the sync
109 * passes to converge. However, in practice disabling compression increases
110 * the average number of sync passes, because when we turn compression off, a
111 * lot of block's size will change and thus we have to re-allocate (not
112 * overwrite) them. It also increases the number of 128KB allocations (e.g.
113 * for indirect blocks and spacemaps) because these will not be compressed.
114 * The 128K allocations are especially detrimental to performance on highly
115 * fragmented systems, which may have very few free segments of this size,
116 * and may need to load new metaslabs to satisfy 128K allocations.
119 /* defer frees starting in this pass */
120 uint_t zfs_sync_pass_deferred_free = 2;
122 /* don't compress starting in this pass */
123 static uint_t zfs_sync_pass_dont_compress = 8;
125 /* rewrite new bps starting in this pass */
126 static uint_t zfs_sync_pass_rewrite = 2;
129 * An allocating zio is one that either currently has the DVA allocate
130 * stage set or will have it later in its lifetime.
132 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
135 * Enable smaller cores by excluding metadata
136 * allocations as well.
138 int zio_exclude_metadata = 0;
139 static int zio_requeue_io_start_cut_in_line = 1;
142 static const int zio_buf_debug_limit = 16384;
144 static const int zio_buf_debug_limit = 0;
147 static inline void __zio_execute(zio_t *zio);
149 static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t);
156 zio_cache = kmem_cache_create("zio_cache",
157 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
158 zio_link_cache = kmem_cache_create("zio_link_cache",
159 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
162 * For small buffers, we want a cache for each multiple of
163 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
164 * for each quarter-power of 2.
166 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
167 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
170 size_t data_cflags, cflags;
172 data_cflags = KMC_NODEBUG;
173 cflags = (zio_exclude_metadata || size > zio_buf_debug_limit) ?
181 * If we are using watchpoints, put each buffer on its own page,
182 * to eliminate the performance overhead of trapping to the
183 * kernel when modifying a non-watched buffer that shares the
184 * page with a watched buffer.
186 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
189 * Here's the problem - on 4K native devices in userland on
190 * Linux using O_DIRECT, buffers must be 4K aligned or I/O
191 * will fail with EINVAL, causing zdb (and others) to coredump.
192 * Since userland probably doesn't need optimized buffer caches,
193 * we just force 4K alignment on everything.
195 align = 8 * SPA_MINBLOCKSIZE;
197 if (size < PAGESIZE) {
198 align = SPA_MINBLOCKSIZE;
199 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
206 if (cflags == data_cflags) {
208 * Resulting kmem caches would be identical.
209 * Save memory by creating only one.
211 (void) snprintf(name, sizeof (name),
212 "zio_buf_comb_%lu", (ulong_t)size);
213 zio_buf_cache[c] = kmem_cache_create(name,
214 size, align, NULL, NULL, NULL, NULL, NULL,
216 zio_data_buf_cache[c] = zio_buf_cache[c];
219 (void) snprintf(name, sizeof (name), "zio_buf_%lu",
221 zio_buf_cache[c] = kmem_cache_create(name, size,
222 align, NULL, NULL, NULL, NULL, NULL, cflags);
224 (void) snprintf(name, sizeof (name), "zio_data_buf_%lu",
226 zio_data_buf_cache[c] = kmem_cache_create(name, size,
227 align, NULL, NULL, NULL, NULL, NULL, data_cflags);
232 ASSERT(zio_buf_cache[c] != NULL);
233 if (zio_buf_cache[c - 1] == NULL)
234 zio_buf_cache[c - 1] = zio_buf_cache[c];
236 ASSERT(zio_data_buf_cache[c] != NULL);
237 if (zio_data_buf_cache[c - 1] == NULL)
238 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
249 size_t n = SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT;
251 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
252 for (size_t i = 0; i < n; i++) {
253 if (zio_buf_cache_allocs[i] != zio_buf_cache_frees[i])
254 (void) printf("zio_fini: [%d] %llu != %llu\n",
255 (int)((i + 1) << SPA_MINBLOCKSHIFT),
256 (long long unsigned)zio_buf_cache_allocs[i],
257 (long long unsigned)zio_buf_cache_frees[i]);
262 * The same kmem cache can show up multiple times in both zio_buf_cache
263 * and zio_data_buf_cache. Do a wasteful but trivially correct scan to
266 for (size_t i = 0; i < n; i++) {
267 kmem_cache_t *cache = zio_buf_cache[i];
270 for (size_t j = i; j < n; j++) {
271 if (cache == zio_buf_cache[j])
272 zio_buf_cache[j] = NULL;
273 if (cache == zio_data_buf_cache[j])
274 zio_data_buf_cache[j] = NULL;
276 kmem_cache_destroy(cache);
279 for (size_t i = 0; i < n; i++) {
280 kmem_cache_t *cache = zio_data_buf_cache[i];
283 for (size_t j = i; j < n; j++) {
284 if (cache == zio_data_buf_cache[j])
285 zio_data_buf_cache[j] = NULL;
287 kmem_cache_destroy(cache);
290 for (size_t i = 0; i < n; i++) {
291 VERIFY3P(zio_buf_cache[i], ==, NULL);
292 VERIFY3P(zio_data_buf_cache[i], ==, NULL);
295 kmem_cache_destroy(zio_link_cache);
296 kmem_cache_destroy(zio_cache);
304 * ==========================================================================
305 * Allocate and free I/O buffers
306 * ==========================================================================
310 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
311 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
312 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
313 * excess / transient data in-core during a crashdump.
316 zio_buf_alloc(size_t size)
318 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
320 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
321 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
322 atomic_add_64(&zio_buf_cache_allocs[c], 1);
325 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
329 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
330 * crashdump if the kernel panics. This exists so that we will limit the amount
331 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
332 * of kernel heap dumped to disk when the kernel panics)
335 zio_data_buf_alloc(size_t size)
337 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
339 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
341 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
345 zio_buf_free(void *buf, size_t size)
347 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
349 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
350 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
351 atomic_add_64(&zio_buf_cache_frees[c], 1);
354 kmem_cache_free(zio_buf_cache[c], buf);
358 zio_data_buf_free(void *buf, size_t size)
360 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
362 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
364 kmem_cache_free(zio_data_buf_cache[c], buf);
368 zio_abd_free(void *abd, size_t size)
371 abd_free((abd_t *)abd);
375 * ==========================================================================
376 * Push and pop I/O transform buffers
377 * ==========================================================================
380 zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize,
381 zio_transform_func_t *transform)
383 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
385 zt->zt_orig_abd = zio->io_abd;
386 zt->zt_orig_size = zio->io_size;
387 zt->zt_bufsize = bufsize;
388 zt->zt_transform = transform;
390 zt->zt_next = zio->io_transform_stack;
391 zio->io_transform_stack = zt;
398 zio_pop_transforms(zio_t *zio)
402 while ((zt = zio->io_transform_stack) != NULL) {
403 if (zt->zt_transform != NULL)
404 zt->zt_transform(zio,
405 zt->zt_orig_abd, zt->zt_orig_size);
407 if (zt->zt_bufsize != 0)
408 abd_free(zio->io_abd);
410 zio->io_abd = zt->zt_orig_abd;
411 zio->io_size = zt->zt_orig_size;
412 zio->io_transform_stack = zt->zt_next;
414 kmem_free(zt, sizeof (zio_transform_t));
419 * ==========================================================================
420 * I/O transform callbacks for subblocks, decompression, and decryption
421 * ==========================================================================
424 zio_subblock(zio_t *zio, abd_t *data, uint64_t size)
426 ASSERT(zio->io_size > size);
428 if (zio->io_type == ZIO_TYPE_READ)
429 abd_copy(data, zio->io_abd, size);
433 zio_decompress(zio_t *zio, abd_t *data, uint64_t size)
435 if (zio->io_error == 0) {
436 void *tmp = abd_borrow_buf(data, size);
437 int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
438 zio->io_abd, tmp, zio->io_size, size,
439 &zio->io_prop.zp_complevel);
440 abd_return_buf_copy(data, tmp, size);
442 if (zio_injection_enabled && ret == 0)
443 ret = zio_handle_fault_injection(zio, EINVAL);
446 zio->io_error = SET_ERROR(EIO);
451 zio_decrypt(zio_t *zio, abd_t *data, uint64_t size)
455 blkptr_t *bp = zio->io_bp;
456 spa_t *spa = zio->io_spa;
457 uint64_t dsobj = zio->io_bookmark.zb_objset;
458 uint64_t lsize = BP_GET_LSIZE(bp);
459 dmu_object_type_t ot = BP_GET_TYPE(bp);
460 uint8_t salt[ZIO_DATA_SALT_LEN];
461 uint8_t iv[ZIO_DATA_IV_LEN];
462 uint8_t mac[ZIO_DATA_MAC_LEN];
463 boolean_t no_crypt = B_FALSE;
465 ASSERT(BP_USES_CRYPT(bp));
466 ASSERT3U(size, !=, 0);
468 if (zio->io_error != 0)
472 * Verify the cksum of MACs stored in an indirect bp. It will always
473 * be possible to verify this since it does not require an encryption
476 if (BP_HAS_INDIRECT_MAC_CKSUM(bp)) {
477 zio_crypt_decode_mac_bp(bp, mac);
479 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF) {
481 * We haven't decompressed the data yet, but
482 * zio_crypt_do_indirect_mac_checksum() requires
483 * decompressed data to be able to parse out the MACs
484 * from the indirect block. We decompress it now and
485 * throw away the result after we are finished.
487 tmp = zio_buf_alloc(lsize);
488 ret = zio_decompress_data(BP_GET_COMPRESS(bp),
489 zio->io_abd, tmp, zio->io_size, lsize,
490 &zio->io_prop.zp_complevel);
492 ret = SET_ERROR(EIO);
495 ret = zio_crypt_do_indirect_mac_checksum(B_FALSE,
496 tmp, lsize, BP_SHOULD_BYTESWAP(bp), mac);
497 zio_buf_free(tmp, lsize);
499 ret = zio_crypt_do_indirect_mac_checksum_abd(B_FALSE,
500 zio->io_abd, size, BP_SHOULD_BYTESWAP(bp), mac);
502 abd_copy(data, zio->io_abd, size);
504 if (zio_injection_enabled && ot != DMU_OT_DNODE && ret == 0) {
505 ret = zio_handle_decrypt_injection(spa,
506 &zio->io_bookmark, ot, ECKSUM);
515 * If this is an authenticated block, just check the MAC. It would be
516 * nice to separate this out into its own flag, but when this was done,
517 * we had run out of bits in what is now zio_flag_t. Future cleanup
518 * could make this a flag bit.
520 if (BP_IS_AUTHENTICATED(bp)) {
521 if (ot == DMU_OT_OBJSET) {
522 ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa,
523 dsobj, zio->io_abd, size, BP_SHOULD_BYTESWAP(bp));
525 zio_crypt_decode_mac_bp(bp, mac);
526 ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj,
527 zio->io_abd, size, mac);
528 if (zio_injection_enabled && ret == 0) {
529 ret = zio_handle_decrypt_injection(spa,
530 &zio->io_bookmark, ot, ECKSUM);
533 abd_copy(data, zio->io_abd, size);
541 zio_crypt_decode_params_bp(bp, salt, iv);
543 if (ot == DMU_OT_INTENT_LOG) {
544 tmp = abd_borrow_buf_copy(zio->io_abd, sizeof (zil_chain_t));
545 zio_crypt_decode_mac_zil(tmp, mac);
546 abd_return_buf(zio->io_abd, tmp, sizeof (zil_chain_t));
548 zio_crypt_decode_mac_bp(bp, mac);
551 ret = spa_do_crypt_abd(B_FALSE, spa, &zio->io_bookmark, BP_GET_TYPE(bp),
552 BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp), salt, iv, mac, size, data,
553 zio->io_abd, &no_crypt);
555 abd_copy(data, zio->io_abd, size);
563 /* assert that the key was found unless this was speculative */
564 ASSERT(ret != EACCES || (zio->io_flags & ZIO_FLAG_SPECULATIVE));
567 * If there was a decryption / authentication error return EIO as
568 * the io_error. If this was not a speculative zio, create an ereport.
571 zio->io_error = SET_ERROR(EIO);
572 if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) {
573 spa_log_error(spa, &zio->io_bookmark,
574 &zio->io_bp->blk_birth);
575 (void) zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION,
576 spa, NULL, &zio->io_bookmark, zio, 0);
584 * ==========================================================================
585 * I/O parent/child relationships and pipeline interlocks
586 * ==========================================================================
589 zio_walk_parents(zio_t *cio, zio_link_t **zl)
591 list_t *pl = &cio->io_parent_list;
593 *zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl);
597 ASSERT((*zl)->zl_child == cio);
598 return ((*zl)->zl_parent);
602 zio_walk_children(zio_t *pio, zio_link_t **zl)
604 list_t *cl = &pio->io_child_list;
606 ASSERT(MUTEX_HELD(&pio->io_lock));
608 *zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl);
612 ASSERT((*zl)->zl_parent == pio);
613 return ((*zl)->zl_child);
617 zio_unique_parent(zio_t *cio)
619 zio_link_t *zl = NULL;
620 zio_t *pio = zio_walk_parents(cio, &zl);
622 VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL);
627 zio_add_child(zio_t *pio, zio_t *cio)
629 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
632 * Logical I/Os can have logical, gang, or vdev children.
633 * Gang I/Os can have gang or vdev children.
634 * Vdev I/Os can only have vdev children.
635 * The following ASSERT captures all of these constraints.
637 ASSERT3S(cio->io_child_type, <=, pio->io_child_type);
642 mutex_enter(&pio->io_lock);
643 mutex_enter(&cio->io_lock);
645 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
647 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
648 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
650 list_insert_head(&pio->io_child_list, zl);
651 list_insert_head(&cio->io_parent_list, zl);
653 pio->io_child_count++;
654 cio->io_parent_count++;
656 mutex_exit(&cio->io_lock);
657 mutex_exit(&pio->io_lock);
661 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
663 ASSERT(zl->zl_parent == pio);
664 ASSERT(zl->zl_child == cio);
666 mutex_enter(&pio->io_lock);
667 mutex_enter(&cio->io_lock);
669 list_remove(&pio->io_child_list, zl);
670 list_remove(&cio->io_parent_list, zl);
672 pio->io_child_count--;
673 cio->io_parent_count--;
675 mutex_exit(&cio->io_lock);
676 mutex_exit(&pio->io_lock);
677 kmem_cache_free(zio_link_cache, zl);
681 zio_wait_for_children(zio_t *zio, uint8_t childbits, enum zio_wait_type wait)
683 boolean_t waiting = B_FALSE;
685 mutex_enter(&zio->io_lock);
686 ASSERT(zio->io_stall == NULL);
687 for (int c = 0; c < ZIO_CHILD_TYPES; c++) {
688 if (!(ZIO_CHILD_BIT_IS_SET(childbits, c)))
691 uint64_t *countp = &zio->io_children[c][wait];
694 ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN);
695 zio->io_stall = countp;
700 mutex_exit(&zio->io_lock);
704 __attribute__((always_inline))
706 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait,
707 zio_t **next_to_executep)
709 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
710 int *errorp = &pio->io_child_error[zio->io_child_type];
712 mutex_enter(&pio->io_lock);
713 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
714 *errorp = zio_worst_error(*errorp, zio->io_error);
715 pio->io_reexecute |= zio->io_reexecute;
716 ASSERT3U(*countp, >, 0);
720 if (*countp == 0 && pio->io_stall == countp) {
721 zio_taskq_type_t type =
722 pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE :
724 pio->io_stall = NULL;
725 mutex_exit(&pio->io_lock);
728 * If we can tell the caller to execute this parent next, do
729 * so. We only do this if the parent's zio type matches the
730 * child's type. Otherwise dispatch the parent zio in its
733 * Having the caller execute the parent when possible reduces
734 * locking on the zio taskq's, reduces context switch
735 * overhead, and has no recursion penalty. Note that one
736 * read from disk typically causes at least 3 zio's: a
737 * zio_null(), the logical zio_read(), and then a physical
738 * zio. When the physical ZIO completes, we are able to call
739 * zio_done() on all 3 of these zio's from one invocation of
740 * zio_execute() by returning the parent back to
741 * zio_execute(). Since the parent isn't executed until this
742 * thread returns back to zio_execute(), the caller should do
745 * In other cases, dispatching the parent prevents
746 * overflowing the stack when we have deeply nested
747 * parent-child relationships, as we do with the "mega zio"
748 * of writes for spa_sync(), and the chain of ZIL blocks.
750 if (next_to_executep != NULL && *next_to_executep == NULL &&
751 pio->io_type == zio->io_type) {
752 *next_to_executep = pio;
754 zio_taskq_dispatch(pio, type, B_FALSE);
757 mutex_exit(&pio->io_lock);
762 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
764 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
765 zio->io_error = zio->io_child_error[c];
769 zio_bookmark_compare(const void *x1, const void *x2)
771 const zio_t *z1 = x1;
772 const zio_t *z2 = x2;
774 if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset)
776 if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset)
779 if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object)
781 if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object)
784 if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level)
786 if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level)
789 if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid)
791 if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid)
803 * ==========================================================================
804 * Create the various types of I/O (read, write, free, etc)
805 * ==========================================================================
808 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
809 abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done,
810 void *private, zio_type_t type, zio_priority_t priority,
811 zio_flag_t flags, vdev_t *vd, uint64_t offset,
812 const zbookmark_phys_t *zb, enum zio_stage stage,
813 enum zio_stage pipeline)
817 IMPLY(type != ZIO_TYPE_TRIM, psize <= SPA_MAXBLOCKSIZE);
818 ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0);
819 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
821 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
822 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
823 ASSERT(vd || stage == ZIO_STAGE_OPEN);
825 IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW_COMPRESS) != 0);
827 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
828 memset(zio, 0, sizeof (zio_t));
830 mutex_init(&zio->io_lock, NULL, MUTEX_NOLOCKDEP, NULL);
831 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
833 list_create(&zio->io_parent_list, sizeof (zio_link_t),
834 offsetof(zio_link_t, zl_parent_node));
835 list_create(&zio->io_child_list, sizeof (zio_link_t),
836 offsetof(zio_link_t, zl_child_node));
837 metaslab_trace_init(&zio->io_alloc_list);
840 zio->io_child_type = ZIO_CHILD_VDEV;
841 else if (flags & ZIO_FLAG_GANG_CHILD)
842 zio->io_child_type = ZIO_CHILD_GANG;
843 else if (flags & ZIO_FLAG_DDT_CHILD)
844 zio->io_child_type = ZIO_CHILD_DDT;
846 zio->io_child_type = ZIO_CHILD_LOGICAL;
849 zio->io_bp = (blkptr_t *)bp;
850 zio->io_bp_copy = *bp;
851 zio->io_bp_orig = *bp;
852 if (type != ZIO_TYPE_WRITE ||
853 zio->io_child_type == ZIO_CHILD_DDT)
854 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
855 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
856 zio->io_logical = zio;
857 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
858 pipeline |= ZIO_GANG_STAGES;
864 zio->io_private = private;
866 zio->io_priority = priority;
868 zio->io_offset = offset;
869 zio->io_orig_abd = zio->io_abd = data;
870 zio->io_orig_size = zio->io_size = psize;
871 zio->io_lsize = lsize;
872 zio->io_orig_flags = zio->io_flags = flags;
873 zio->io_orig_stage = zio->io_stage = stage;
874 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
875 zio->io_pipeline_trace = ZIO_STAGE_OPEN;
877 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
878 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
881 zio->io_bookmark = *zb;
884 zio->io_metaslab_class = pio->io_metaslab_class;
885 if (zio->io_logical == NULL)
886 zio->io_logical = pio->io_logical;
887 if (zio->io_child_type == ZIO_CHILD_GANG)
888 zio->io_gang_leader = pio->io_gang_leader;
889 zio_add_child(pio, zio);
892 taskq_init_ent(&zio->io_tqent);
898 zio_destroy(zio_t *zio)
900 metaslab_trace_fini(&zio->io_alloc_list);
901 list_destroy(&zio->io_parent_list);
902 list_destroy(&zio->io_child_list);
903 mutex_destroy(&zio->io_lock);
904 cv_destroy(&zio->io_cv);
905 kmem_cache_free(zio_cache, zio);
909 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
910 void *private, zio_flag_t flags)
914 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
915 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
916 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
922 zio_root(spa_t *spa, zio_done_func_t *done, void *private, zio_flag_t flags)
924 return (zio_null(NULL, spa, NULL, done, private, flags));
928 zfs_blkptr_verify_log(spa_t *spa, const blkptr_t *bp,
929 enum blk_verify_flag blk_verify, const char *fmt, ...)
935 (void) vsnprintf(buf, sizeof (buf), fmt, adx);
938 zfs_dbgmsg("bad blkptr at %px: "
939 "DVA[0]=%#llx/%#llx "
940 "DVA[1]=%#llx/%#llx "
941 "DVA[2]=%#llx/%#llx "
947 "cksum=%#llx/%#llx/%#llx/%#llx",
949 (long long)bp->blk_dva[0].dva_word[0],
950 (long long)bp->blk_dva[0].dva_word[1],
951 (long long)bp->blk_dva[1].dva_word[0],
952 (long long)bp->blk_dva[1].dva_word[1],
953 (long long)bp->blk_dva[2].dva_word[0],
954 (long long)bp->blk_dva[2].dva_word[1],
955 (long long)bp->blk_prop,
956 (long long)bp->blk_pad[0],
957 (long long)bp->blk_pad[1],
958 (long long)bp->blk_phys_birth,
959 (long long)bp->blk_birth,
960 (long long)bp->blk_fill,
961 (long long)bp->blk_cksum.zc_word[0],
962 (long long)bp->blk_cksum.zc_word[1],
963 (long long)bp->blk_cksum.zc_word[2],
964 (long long)bp->blk_cksum.zc_word[3]);
965 switch (blk_verify) {
966 case BLK_VERIFY_HALT:
967 zfs_panic_recover("%s: %s", spa_name(spa), buf);
970 zfs_dbgmsg("%s: %s", spa_name(spa), buf);
972 case BLK_VERIFY_ONLY:
980 * Verify the block pointer fields contain reasonable values. This means
981 * it only contains known object types, checksum/compression identifiers,
982 * block sizes within the maximum allowed limits, valid DVAs, etc.
984 * If everything checks out B_TRUE is returned. The zfs_blkptr_verify
985 * argument controls the behavior when an invalid field is detected.
987 * Values for blk_verify_flag:
988 * BLK_VERIFY_ONLY: evaluate the block
989 * BLK_VERIFY_LOG: evaluate the block and log problems
990 * BLK_VERIFY_HALT: call zfs_panic_recover on error
992 * Values for blk_config_flag:
993 * BLK_CONFIG_HELD: caller holds SCL_VDEV for writer
994 * BLK_CONFIG_NEEDED: caller holds no config lock, SCL_VDEV will be
995 * obtained for reader
996 * BLK_CONFIG_SKIP: skip checks which require SCL_VDEV, for better
1000 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp,
1001 enum blk_config_flag blk_config, enum blk_verify_flag blk_verify)
1005 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
1006 errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1007 "blkptr at %px has invalid TYPE %llu",
1008 bp, (longlong_t)BP_GET_TYPE(bp));
1010 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS) {
1011 errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1012 "blkptr at %px has invalid CHECKSUM %llu",
1013 bp, (longlong_t)BP_GET_CHECKSUM(bp));
1015 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS) {
1016 errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1017 "blkptr at %px has invalid COMPRESS %llu",
1018 bp, (longlong_t)BP_GET_COMPRESS(bp));
1020 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
1021 errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1022 "blkptr at %px has invalid LSIZE %llu",
1023 bp, (longlong_t)BP_GET_LSIZE(bp));
1025 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
1026 errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1027 "blkptr at %px has invalid PSIZE %llu",
1028 bp, (longlong_t)BP_GET_PSIZE(bp));
1031 if (BP_IS_EMBEDDED(bp)) {
1032 if (BPE_GET_ETYPE(bp) >= NUM_BP_EMBEDDED_TYPES) {
1033 errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1034 "blkptr at %px has invalid ETYPE %llu",
1035 bp, (longlong_t)BPE_GET_ETYPE(bp));
1040 * Do not verify individual DVAs if the config is not trusted. This
1041 * will be done once the zio is executed in vdev_mirror_map_alloc.
1043 if (!spa->spa_trust_config)
1044 return (errors == 0);
1046 switch (blk_config) {
1047 case BLK_CONFIG_HELD:
1048 ASSERT(spa_config_held(spa, SCL_VDEV, RW_WRITER));
1050 case BLK_CONFIG_NEEDED:
1051 spa_config_enter(spa, SCL_VDEV, bp, RW_READER);
1053 case BLK_CONFIG_SKIP:
1054 return (errors == 0);
1056 panic("invalid blk_config %u", blk_config);
1060 * Pool-specific checks.
1062 * Note: it would be nice to verify that the blk_birth and
1063 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
1064 * allows the birth time of log blocks (and dmu_sync()-ed blocks
1065 * that are in the log) to be arbitrarily large.
1067 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
1068 const dva_t *dva = &bp->blk_dva[i];
1069 uint64_t vdevid = DVA_GET_VDEV(dva);
1071 if (vdevid >= spa->spa_root_vdev->vdev_children) {
1072 errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1073 "blkptr at %px DVA %u has invalid VDEV %llu",
1074 bp, i, (longlong_t)vdevid);
1077 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
1079 errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1080 "blkptr at %px DVA %u has invalid VDEV %llu",
1081 bp, i, (longlong_t)vdevid);
1084 if (vd->vdev_ops == &vdev_hole_ops) {
1085 errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1086 "blkptr at %px DVA %u has hole VDEV %llu",
1087 bp, i, (longlong_t)vdevid);
1090 if (vd->vdev_ops == &vdev_missing_ops) {
1092 * "missing" vdevs are valid during import, but we
1093 * don't have their detailed info (e.g. asize), so
1094 * we can't perform any more checks on them.
1098 uint64_t offset = DVA_GET_OFFSET(dva);
1099 uint64_t asize = DVA_GET_ASIZE(dva);
1100 if (DVA_GET_GANG(dva))
1101 asize = vdev_gang_header_asize(vd);
1102 if (offset + asize > vd->vdev_asize) {
1103 errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1104 "blkptr at %px DVA %u has invalid OFFSET %llu",
1105 bp, i, (longlong_t)offset);
1108 if (blk_config == BLK_CONFIG_NEEDED)
1109 spa_config_exit(spa, SCL_VDEV, bp);
1111 return (errors == 0);
1115 zfs_dva_valid(spa_t *spa, const dva_t *dva, const blkptr_t *bp)
1118 uint64_t vdevid = DVA_GET_VDEV(dva);
1120 if (vdevid >= spa->spa_root_vdev->vdev_children)
1123 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
1127 if (vd->vdev_ops == &vdev_hole_ops)
1130 if (vd->vdev_ops == &vdev_missing_ops) {
1134 uint64_t offset = DVA_GET_OFFSET(dva);
1135 uint64_t asize = DVA_GET_ASIZE(dva);
1137 if (DVA_GET_GANG(dva))
1138 asize = vdev_gang_header_asize(vd);
1139 if (offset + asize > vd->vdev_asize)
1146 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
1147 abd_t *data, uint64_t size, zio_done_func_t *done, void *private,
1148 zio_priority_t priority, zio_flag_t flags, const zbookmark_phys_t *zb)
1152 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
1153 data, size, size, done, private,
1154 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
1155 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
1156 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
1162 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
1163 abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp,
1164 zio_done_func_t *ready, zio_done_func_t *children_ready,
1165 zio_done_func_t *physdone, zio_done_func_t *done,
1166 void *private, zio_priority_t priority, zio_flag_t flags,
1167 const zbookmark_phys_t *zb)
1171 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
1172 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
1173 zp->zp_compress >= ZIO_COMPRESS_OFF &&
1174 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
1175 DMU_OT_IS_VALID(zp->zp_type) &&
1176 zp->zp_level < 32 &&
1177 zp->zp_copies > 0 &&
1178 zp->zp_copies <= spa_max_replication(spa));
1180 zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private,
1181 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
1182 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
1183 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
1185 zio->io_ready = ready;
1186 zio->io_children_ready = children_ready;
1187 zio->io_physdone = physdone;
1191 * Data can be NULL if we are going to call zio_write_override() to
1192 * provide the already-allocated BP. But we may need the data to
1193 * verify a dedup hit (if requested). In this case, don't try to
1194 * dedup (just take the already-allocated BP verbatim). Encrypted
1195 * dedup blocks need data as well so we also disable dedup in this
1199 (zio->io_prop.zp_dedup_verify || zio->io_prop.zp_encrypt)) {
1200 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
1207 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data,
1208 uint64_t size, zio_done_func_t *done, void *private,
1209 zio_priority_t priority, zio_flag_t flags, zbookmark_phys_t *zb)
1213 zio = zio_create(pio, spa, txg, bp, data, size, size, done, private,
1214 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb,
1215 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
1221 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite,
1224 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
1225 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1226 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1227 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
1228 ASSERT(!brtwrite || !nopwrite);
1231 * We must reset the io_prop to match the values that existed
1232 * when the bp was first written by dmu_sync() keeping in mind
1233 * that nopwrite and dedup are mutually exclusive.
1235 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
1236 zio->io_prop.zp_nopwrite = nopwrite;
1237 zio->io_prop.zp_brtwrite = brtwrite;
1238 zio->io_prop.zp_copies = copies;
1239 zio->io_bp_override = bp;
1243 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
1246 (void) zfs_blkptr_verify(spa, bp, BLK_CONFIG_NEEDED, BLK_VERIFY_HALT);
1249 * The check for EMBEDDED is a performance optimization. We
1250 * process the free here (by ignoring it) rather than
1251 * putting it on the list and then processing it in zio_free_sync().
1253 if (BP_IS_EMBEDDED(bp))
1257 * Frees that are for the currently-syncing txg, are not going to be
1258 * deferred, and which will not need to do a read (i.e. not GANG or
1259 * DEDUP), can be processed immediately. Otherwise, put them on the
1260 * in-memory list for later processing.
1262 * Note that we only defer frees after zfs_sync_pass_deferred_free
1263 * when the log space map feature is disabled. [see relevant comment
1264 * in spa_sync_iterate_to_convergence()]
1266 if (BP_IS_GANG(bp) ||
1268 txg != spa->spa_syncing_txg ||
1269 (spa_sync_pass(spa) >= zfs_sync_pass_deferred_free &&
1270 !spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP)) ||
1271 brt_maybe_exists(spa, bp)) {
1272 metaslab_check_free(spa, bp);
1273 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
1275 VERIFY3P(zio_free_sync(NULL, spa, txg, bp, 0), ==, NULL);
1280 * To improve performance, this function may return NULL if we were able
1281 * to do the free immediately. This avoids the cost of creating a zio
1282 * (and linking it to the parent, etc).
1285 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1288 ASSERT(!BP_IS_HOLE(bp));
1289 ASSERT(spa_syncing_txg(spa) == txg);
1291 if (BP_IS_EMBEDDED(bp))
1294 metaslab_check_free(spa, bp);
1296 dsl_scan_freed(spa, bp);
1298 if (BP_IS_GANG(bp) ||
1300 brt_maybe_exists(spa, bp)) {
1302 * GANG, DEDUP and BRT blocks can induce a read (for the gang
1303 * block header, the DDT or the BRT), so issue them
1304 * asynchronously so that this thread is not tied up.
1306 enum zio_stage stage =
1307 ZIO_FREE_PIPELINE | ZIO_STAGE_ISSUE_ASYNC;
1309 return (zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1310 BP_GET_PSIZE(bp), NULL, NULL,
1311 ZIO_TYPE_FREE, ZIO_PRIORITY_NOW,
1312 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage));
1314 metaslab_free(spa, bp, txg, B_FALSE);
1320 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1321 zio_done_func_t *done, void *private, zio_flag_t flags)
1325 (void) zfs_blkptr_verify(spa, bp, (flags & ZIO_FLAG_CONFIG_WRITER) ?
1326 BLK_CONFIG_HELD : BLK_CONFIG_NEEDED, BLK_VERIFY_HALT);
1328 if (BP_IS_EMBEDDED(bp))
1329 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
1332 * A claim is an allocation of a specific block. Claims are needed
1333 * to support immediate writes in the intent log. The issue is that
1334 * immediate writes contain committed data, but in a txg that was
1335 * *not* committed. Upon opening the pool after an unclean shutdown,
1336 * the intent log claims all blocks that contain immediate write data
1337 * so that the SPA knows they're in use.
1339 * All claims *must* be resolved in the first txg -- before the SPA
1340 * starts allocating blocks -- so that nothing is allocated twice.
1341 * If txg == 0 we just verify that the block is claimable.
1343 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <,
1344 spa_min_claim_txg(spa));
1345 ASSERT(txg == spa_min_claim_txg(spa) || txg == 0);
1346 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(8) */
1348 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1349 BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW,
1350 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
1351 ASSERT0(zio->io_queued_timestamp);
1357 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
1358 zio_done_func_t *done, void *private, zio_flag_t flags)
1363 if (vd->vdev_children == 0) {
1364 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
1365 ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
1366 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
1370 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
1372 for (c = 0; c < vd->vdev_children; c++)
1373 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
1374 done, private, flags));
1381 zio_trim(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1382 zio_done_func_t *done, void *private, zio_priority_t priority,
1383 zio_flag_t flags, enum trim_flag trim_flags)
1387 ASSERT0(vd->vdev_children);
1388 ASSERT0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
1389 ASSERT0(P2PHASE(size, 1ULL << vd->vdev_ashift));
1390 ASSERT3U(size, !=, 0);
1392 zio = zio_create(pio, vd->vdev_spa, 0, NULL, NULL, size, size, done,
1393 private, ZIO_TYPE_TRIM, priority, flags | ZIO_FLAG_PHYSICAL,
1394 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_TRIM_PIPELINE);
1395 zio->io_trim_flags = trim_flags;
1401 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1402 abd_t *data, int checksum, zio_done_func_t *done, void *private,
1403 zio_priority_t priority, zio_flag_t flags, boolean_t labels)
1407 ASSERT(vd->vdev_children == 0);
1408 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1409 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1410 ASSERT3U(offset + size, <=, vd->vdev_psize);
1412 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1413 private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1414 offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
1416 zio->io_prop.zp_checksum = checksum;
1422 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1423 abd_t *data, int checksum, zio_done_func_t *done, void *private,
1424 zio_priority_t priority, zio_flag_t flags, boolean_t labels)
1428 ASSERT(vd->vdev_children == 0);
1429 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1430 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1431 ASSERT3U(offset + size, <=, vd->vdev_psize);
1433 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1434 private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1435 offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
1437 zio->io_prop.zp_checksum = checksum;
1439 if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
1441 * zec checksums are necessarily destructive -- they modify
1442 * the end of the write buffer to hold the verifier/checksum.
1443 * Therefore, we must make a local copy in case the data is
1444 * being written to multiple places in parallel.
1446 abd_t *wbuf = abd_alloc_sametype(data, size);
1447 abd_copy(wbuf, data, size);
1449 zio_push_transform(zio, wbuf, size, size, NULL);
1456 * Create a child I/O to do some work for us.
1459 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1460 abd_t *data, uint64_t size, int type, zio_priority_t priority,
1461 zio_flag_t flags, zio_done_func_t *done, void *private)
1463 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1467 * vdev child I/Os do not propagate their error to the parent.
1468 * Therefore, for correct operation the caller *must* check for
1469 * and handle the error in the child i/o's done callback.
1470 * The only exceptions are i/os that we don't care about
1471 * (OPTIONAL or REPAIR).
1473 ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) ||
1476 if (type == ZIO_TYPE_READ && bp != NULL) {
1478 * If we have the bp, then the child should perform the
1479 * checksum and the parent need not. This pushes error
1480 * detection as close to the leaves as possible and
1481 * eliminates redundant checksums in the interior nodes.
1483 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1484 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1487 if (vd->vdev_ops->vdev_op_leaf) {
1488 ASSERT0(vd->vdev_children);
1489 offset += VDEV_LABEL_START_SIZE;
1492 flags |= ZIO_VDEV_CHILD_FLAGS(pio);
1495 * If we've decided to do a repair, the write is not speculative --
1496 * even if the original read was.
1498 if (flags & ZIO_FLAG_IO_REPAIR)
1499 flags &= ~ZIO_FLAG_SPECULATIVE;
1502 * If we're creating a child I/O that is not associated with a
1503 * top-level vdev, then the child zio is not an allocating I/O.
1504 * If this is a retried I/O then we ignore it since we will
1505 * have already processed the original allocating I/O.
1507 if (flags & ZIO_FLAG_IO_ALLOCATING &&
1508 (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) {
1509 ASSERT(pio->io_metaslab_class != NULL);
1510 ASSERT(pio->io_metaslab_class->mc_alloc_throttle_enabled);
1511 ASSERT(type == ZIO_TYPE_WRITE);
1512 ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE);
1513 ASSERT(!(flags & ZIO_FLAG_IO_REPAIR));
1514 ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) ||
1515 pio->io_child_type == ZIO_CHILD_GANG);
1517 flags &= ~ZIO_FLAG_IO_ALLOCATING;
1521 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size,
1522 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1523 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1524 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
1526 zio->io_physdone = pio->io_physdone;
1527 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1528 zio->io_logical->io_phys_children++;
1534 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size,
1535 zio_type_t type, zio_priority_t priority, zio_flag_t flags,
1536 zio_done_func_t *done, void *private)
1540 ASSERT(vd->vdev_ops->vdev_op_leaf);
1542 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1543 data, size, size, done, private, type, priority,
1544 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1546 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1552 zio_flush(zio_t *zio, vdev_t *vd)
1554 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
1556 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1560 zio_shrink(zio_t *zio, uint64_t size)
1562 ASSERT3P(zio->io_executor, ==, NULL);
1563 ASSERT3U(zio->io_orig_size, ==, zio->io_size);
1564 ASSERT3U(size, <=, zio->io_size);
1567 * We don't shrink for raidz because of problems with the
1568 * reconstruction when reading back less than the block size.
1569 * Note, BP_IS_RAIDZ() assumes no compression.
1571 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1572 if (!BP_IS_RAIDZ(zio->io_bp)) {
1573 /* we are not doing a raw write */
1574 ASSERT3U(zio->io_size, ==, zio->io_lsize);
1575 zio->io_orig_size = zio->io_size = zio->io_lsize = size;
1580 * ==========================================================================
1581 * Prepare to read and write logical blocks
1582 * ==========================================================================
1586 zio_read_bp_init(zio_t *zio)
1588 blkptr_t *bp = zio->io_bp;
1590 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1592 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1594 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1595 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1596 !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) {
1597 zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1598 psize, psize, zio_decompress);
1601 if (((BP_IS_PROTECTED(bp) && !(zio->io_flags & ZIO_FLAG_RAW_ENCRYPT)) ||
1602 BP_HAS_INDIRECT_MAC_CKSUM(bp)) &&
1603 zio->io_child_type == ZIO_CHILD_LOGICAL) {
1604 zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1605 psize, psize, zio_decrypt);
1608 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1609 int psize = BPE_GET_PSIZE(bp);
1610 void *data = abd_borrow_buf(zio->io_abd, psize);
1612 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1613 decode_embedded_bp_compressed(bp, data);
1614 abd_return_buf_copy(zio->io_abd, data, psize);
1616 ASSERT(!BP_IS_EMBEDDED(bp));
1617 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1620 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1621 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1627 zio_write_bp_init(zio_t *zio)
1629 if (!IO_IS_ALLOCATING(zio))
1632 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1634 if (zio->io_bp_override) {
1635 blkptr_t *bp = zio->io_bp;
1636 zio_prop_t *zp = &zio->io_prop;
1638 ASSERT(bp->blk_birth != zio->io_txg);
1640 *bp = *zio->io_bp_override;
1641 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1643 if (zp->zp_brtwrite)
1646 ASSERT(!BP_GET_DEDUP(zio->io_bp_override));
1648 if (BP_IS_EMBEDDED(bp))
1652 * If we've been overridden and nopwrite is set then
1653 * set the flag accordingly to indicate that a nopwrite
1654 * has already occurred.
1656 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1657 ASSERT(!zp->zp_dedup);
1658 ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum);
1659 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1663 ASSERT(!zp->zp_nopwrite);
1665 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1668 ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags &
1669 ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify);
1671 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum &&
1673 BP_SET_DEDUP(bp, 1);
1674 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1679 * We were unable to handle this as an override bp, treat
1680 * it as a regular write I/O.
1682 zio->io_bp_override = NULL;
1683 *bp = zio->io_bp_orig;
1684 zio->io_pipeline = zio->io_orig_pipeline;
1691 zio_write_compress(zio_t *zio)
1693 spa_t *spa = zio->io_spa;
1694 zio_prop_t *zp = &zio->io_prop;
1695 enum zio_compress compress = zp->zp_compress;
1696 blkptr_t *bp = zio->io_bp;
1697 uint64_t lsize = zio->io_lsize;
1698 uint64_t psize = zio->io_size;
1702 * If our children haven't all reached the ready stage,
1703 * wait for them and then repeat this pipeline stage.
1705 if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT |
1706 ZIO_CHILD_GANG_BIT, ZIO_WAIT_READY)) {
1710 if (!IO_IS_ALLOCATING(zio))
1713 if (zio->io_children_ready != NULL) {
1715 * Now that all our children are ready, run the callback
1716 * associated with this zio in case it wants to modify the
1717 * data to be written.
1719 ASSERT3U(zp->zp_level, >, 0);
1720 zio->io_children_ready(zio);
1723 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1724 ASSERT(zio->io_bp_override == NULL);
1726 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1728 * We're rewriting an existing block, which means we're
1729 * working on behalf of spa_sync(). For spa_sync() to
1730 * converge, it must eventually be the case that we don't
1731 * have to allocate new blocks. But compression changes
1732 * the blocksize, which forces a reallocate, and makes
1733 * convergence take longer. Therefore, after the first
1734 * few passes, stop compressing to ensure convergence.
1736 pass = spa_sync_pass(spa);
1738 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1739 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1740 ASSERT(!BP_GET_DEDUP(bp));
1742 if (pass >= zfs_sync_pass_dont_compress)
1743 compress = ZIO_COMPRESS_OFF;
1745 /* Make sure someone doesn't change their mind on overwrites */
1746 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1747 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1750 /* If it's a compressed write that is not raw, compress the buffer. */
1751 if (compress != ZIO_COMPRESS_OFF &&
1752 !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) {
1754 psize = zio_compress_data(compress, zio->io_abd, &cbuf, lsize,
1757 compress = ZIO_COMPRESS_OFF;
1758 } else if (psize >= lsize) {
1759 compress = ZIO_COMPRESS_OFF;
1761 zio_buf_free(cbuf, lsize);
1762 } else if (!zp->zp_dedup && !zp->zp_encrypt &&
1763 psize <= BPE_PAYLOAD_SIZE &&
1764 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1765 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1766 encode_embedded_bp_compressed(bp,
1767 cbuf, compress, lsize, psize);
1768 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1769 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1770 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1771 zio_buf_free(cbuf, lsize);
1772 bp->blk_birth = zio->io_txg;
1773 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1774 ASSERT(spa_feature_is_active(spa,
1775 SPA_FEATURE_EMBEDDED_DATA));
1779 * Round compressed size up to the minimum allocation
1780 * size of the smallest-ashift device, and zero the
1781 * tail. This ensures that the compressed size of the
1782 * BP (and thus compressratio property) are correct,
1783 * in that we charge for the padding used to fill out
1786 ASSERT3U(spa->spa_min_alloc, >=, SPA_MINBLOCKSHIFT);
1787 size_t rounded = (size_t)roundup(psize,
1788 spa->spa_min_alloc);
1789 if (rounded >= lsize) {
1790 compress = ZIO_COMPRESS_OFF;
1791 zio_buf_free(cbuf, lsize);
1794 abd_t *cdata = abd_get_from_buf(cbuf, lsize);
1795 abd_take_ownership_of_buf(cdata, B_TRUE);
1796 abd_zero_off(cdata, psize, rounded - psize);
1798 zio_push_transform(zio, cdata,
1799 psize, lsize, NULL);
1804 * We were unable to handle this as an override bp, treat
1805 * it as a regular write I/O.
1807 zio->io_bp_override = NULL;
1808 *bp = zio->io_bp_orig;
1809 zio->io_pipeline = zio->io_orig_pipeline;
1811 } else if ((zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) != 0 &&
1812 zp->zp_type == DMU_OT_DNODE) {
1814 * The DMU actually relies on the zio layer's compression
1815 * to free metadnode blocks that have had all contained
1816 * dnodes freed. As a result, even when doing a raw
1817 * receive, we must check whether the block can be compressed
1820 psize = zio_compress_data(ZIO_COMPRESS_EMPTY,
1821 zio->io_abd, NULL, lsize, zp->zp_complevel);
1822 if (psize == 0 || psize >= lsize)
1823 compress = ZIO_COMPRESS_OFF;
1824 } else if (zio->io_flags & ZIO_FLAG_RAW_COMPRESS &&
1825 !(zio->io_flags & ZIO_FLAG_RAW_ENCRYPT)) {
1827 * If we are raw receiving an encrypted dataset we should not
1828 * take this codepath because it will change the on-disk block
1829 * and decryption will fail.
1831 size_t rounded = MIN((size_t)roundup(psize,
1832 spa->spa_min_alloc), lsize);
1834 if (rounded != psize) {
1835 abd_t *cdata = abd_alloc_linear(rounded, B_TRUE);
1836 abd_zero_off(cdata, psize, rounded - psize);
1837 abd_copy_off(cdata, zio->io_abd, 0, 0, psize);
1839 zio_push_transform(zio, cdata,
1840 psize, rounded, NULL);
1843 ASSERT3U(psize, !=, 0);
1847 * The final pass of spa_sync() must be all rewrites, but the first
1848 * few passes offer a trade-off: allocating blocks defers convergence,
1849 * but newly allocated blocks are sequential, so they can be written
1850 * to disk faster. Therefore, we allow the first few passes of
1851 * spa_sync() to allocate new blocks, but force rewrites after that.
1852 * There should only be a handful of blocks after pass 1 in any case.
1854 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1855 BP_GET_PSIZE(bp) == psize &&
1856 pass >= zfs_sync_pass_rewrite) {
1857 VERIFY3U(psize, !=, 0);
1858 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1860 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1861 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1864 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1868 if (zio->io_bp_orig.blk_birth != 0 &&
1869 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1870 BP_SET_LSIZE(bp, lsize);
1871 BP_SET_TYPE(bp, zp->zp_type);
1872 BP_SET_LEVEL(bp, zp->zp_level);
1873 BP_SET_BIRTH(bp, zio->io_txg, 0);
1875 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1877 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1878 BP_SET_LSIZE(bp, lsize);
1879 BP_SET_TYPE(bp, zp->zp_type);
1880 BP_SET_LEVEL(bp, zp->zp_level);
1881 BP_SET_PSIZE(bp, psize);
1882 BP_SET_COMPRESS(bp, compress);
1883 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1884 BP_SET_DEDUP(bp, zp->zp_dedup);
1885 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1887 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1888 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1889 ASSERT(!zp->zp_encrypt ||
1890 DMU_OT_IS_ENCRYPTED(zp->zp_type));
1891 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1893 if (zp->zp_nopwrite) {
1894 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1895 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1896 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1903 zio_free_bp_init(zio_t *zio)
1905 blkptr_t *bp = zio->io_bp;
1907 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1908 if (BP_GET_DEDUP(bp))
1909 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1912 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1918 * ==========================================================================
1919 * Execute the I/O pipeline
1920 * ==========================================================================
1924 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1926 spa_t *spa = zio->io_spa;
1927 zio_type_t t = zio->io_type;
1928 int flags = (cutinline ? TQ_FRONT : 0);
1931 * If we're a config writer or a probe, the normal issue and
1932 * interrupt threads may all be blocked waiting for the config lock.
1933 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1935 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1939 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1941 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1945 * If this is a high priority I/O, then use the high priority taskq if
1948 if ((zio->io_priority == ZIO_PRIORITY_NOW ||
1949 zio->io_priority == ZIO_PRIORITY_SYNC_WRITE) &&
1950 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1953 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1956 * NB: We are assuming that the zio can only be dispatched
1957 * to a single taskq at a time. It would be a grievous error
1958 * to dispatch the zio to another taskq at the same time.
1960 ASSERT(taskq_empty_ent(&zio->io_tqent));
1961 spa_taskq_dispatch_ent(spa, t, q, zio_execute, zio, flags,
1966 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1968 spa_t *spa = zio->io_spa;
1970 taskq_t *tq = taskq_of_curthread();
1972 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1973 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1975 for (i = 0; i < tqs->stqs_count; i++) {
1976 if (tqs->stqs_taskq[i] == tq)
1985 zio_issue_async(zio_t *zio)
1987 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1993 zio_interrupt(void *zio)
1995 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1999 zio_delay_interrupt(zio_t *zio)
2002 * The timeout_generic() function isn't defined in userspace, so
2003 * rather than trying to implement the function, the zio delay
2004 * functionality has been disabled for userspace builds.
2009 * If io_target_timestamp is zero, then no delay has been registered
2010 * for this IO, thus jump to the end of this function and "skip" the
2011 * delay; issuing it directly to the zio layer.
2013 if (zio->io_target_timestamp != 0) {
2014 hrtime_t now = gethrtime();
2016 if (now >= zio->io_target_timestamp) {
2018 * This IO has already taken longer than the target
2019 * delay to complete, so we don't want to delay it
2020 * any longer; we "miss" the delay and issue it
2021 * directly to the zio layer. This is likely due to
2022 * the target latency being set to a value less than
2023 * the underlying hardware can satisfy (e.g. delay
2024 * set to 1ms, but the disks take 10ms to complete an
2028 DTRACE_PROBE2(zio__delay__miss, zio_t *, zio,
2034 hrtime_t diff = zio->io_target_timestamp - now;
2035 clock_t expire_at_tick = ddi_get_lbolt() +
2038 DTRACE_PROBE3(zio__delay__hit, zio_t *, zio,
2039 hrtime_t, now, hrtime_t, diff);
2041 if (NSEC_TO_TICK(diff) == 0) {
2042 /* Our delay is less than a jiffy - just spin */
2043 zfs_sleep_until(zio->io_target_timestamp);
2047 * Use taskq_dispatch_delay() in the place of
2048 * OpenZFS's timeout_generic().
2050 tid = taskq_dispatch_delay(system_taskq,
2051 zio_interrupt, zio, TQ_NOSLEEP,
2053 if (tid == TASKQID_INVALID) {
2055 * Couldn't allocate a task. Just
2056 * finish the zio without a delay.
2065 DTRACE_PROBE1(zio__delay__skip, zio_t *, zio);
2070 zio_deadman_impl(zio_t *pio, int ziodepth)
2072 zio_t *cio, *cio_next;
2073 zio_link_t *zl = NULL;
2074 vdev_t *vd = pio->io_vd;
2076 if (zio_deadman_log_all || (vd != NULL && vd->vdev_ops->vdev_op_leaf)) {
2077 vdev_queue_t *vq = vd ? &vd->vdev_queue : NULL;
2078 zbookmark_phys_t *zb = &pio->io_bookmark;
2079 uint64_t delta = gethrtime() - pio->io_timestamp;
2080 uint64_t failmode = spa_get_deadman_failmode(pio->io_spa);
2082 zfs_dbgmsg("slow zio[%d]: zio=%px timestamp=%llu "
2083 "delta=%llu queued=%llu io=%llu "
2085 "last=%llu type=%d "
2086 "priority=%d flags=0x%llx stage=0x%x "
2087 "pipeline=0x%x pipeline-trace=0x%x "
2088 "objset=%llu object=%llu "
2089 "level=%llu blkid=%llu "
2090 "offset=%llu size=%llu "
2092 ziodepth, pio, pio->io_timestamp,
2093 (u_longlong_t)delta, pio->io_delta, pio->io_delay,
2094 vd ? vd->vdev_path : "NULL",
2095 vq ? vq->vq_io_complete_ts : 0, pio->io_type,
2096 pio->io_priority, (u_longlong_t)pio->io_flags,
2097 pio->io_stage, pio->io_pipeline, pio->io_pipeline_trace,
2098 (u_longlong_t)zb->zb_objset, (u_longlong_t)zb->zb_object,
2099 (u_longlong_t)zb->zb_level, (u_longlong_t)zb->zb_blkid,
2100 (u_longlong_t)pio->io_offset, (u_longlong_t)pio->io_size,
2102 (void) zfs_ereport_post(FM_EREPORT_ZFS_DEADMAN,
2103 pio->io_spa, vd, zb, pio, 0);
2105 if (failmode == ZIO_FAILURE_MODE_CONTINUE &&
2106 taskq_empty_ent(&pio->io_tqent)) {
2111 mutex_enter(&pio->io_lock);
2112 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
2113 cio_next = zio_walk_children(pio, &zl);
2114 zio_deadman_impl(cio, ziodepth + 1);
2116 mutex_exit(&pio->io_lock);
2120 * Log the critical information describing this zio and all of its children
2121 * using the zfs_dbgmsg() interface then post deadman event for the ZED.
2124 zio_deadman(zio_t *pio, const char *tag)
2126 spa_t *spa = pio->io_spa;
2127 char *name = spa_name(spa);
2129 if (!zfs_deadman_enabled || spa_suspended(spa))
2132 zio_deadman_impl(pio, 0);
2134 switch (spa_get_deadman_failmode(spa)) {
2135 case ZIO_FAILURE_MODE_WAIT:
2136 zfs_dbgmsg("%s waiting for hung I/O to pool '%s'", tag, name);
2139 case ZIO_FAILURE_MODE_CONTINUE:
2140 zfs_dbgmsg("%s restarting hung I/O for pool '%s'", tag, name);
2143 case ZIO_FAILURE_MODE_PANIC:
2144 fm_panic("%s determined I/O to pool '%s' is hung.", tag, name);
2150 * Execute the I/O pipeline until one of the following occurs:
2151 * (1) the I/O completes; (2) the pipeline stalls waiting for
2152 * dependent child I/Os; (3) the I/O issues, so we're waiting
2153 * for an I/O completion interrupt; (4) the I/O is delegated by
2154 * vdev-level caching or aggregation; (5) the I/O is deferred
2155 * due to vdev-level queueing; (6) the I/O is handed off to
2156 * another thread. In all cases, the pipeline stops whenever
2157 * there's no CPU work; it never burns a thread in cv_wait_io().
2159 * There's no locking on io_stage because there's no legitimate way
2160 * for multiple threads to be attempting to process the same I/O.
2162 static zio_pipe_stage_t *zio_pipeline[];
2165 * zio_execute() is a wrapper around the static function
2166 * __zio_execute() so that we can force __zio_execute() to be
2167 * inlined. This reduces stack overhead which is important
2168 * because __zio_execute() is called recursively in several zio
2169 * code paths. zio_execute() itself cannot be inlined because
2170 * it is externally visible.
2173 zio_execute(void *zio)
2175 fstrans_cookie_t cookie;
2177 cookie = spl_fstrans_mark();
2179 spl_fstrans_unmark(cookie);
2183 * Used to determine if in the current context the stack is sized large
2184 * enough to allow zio_execute() to be called recursively. A minimum
2185 * stack size of 16K is required to avoid needing to re-dispatch the zio.
2188 zio_execute_stack_check(zio_t *zio)
2190 #if !defined(HAVE_LARGE_STACKS)
2191 dsl_pool_t *dp = spa_get_dsl(zio->io_spa);
2193 /* Executing in txg_sync_thread() context. */
2194 if (dp && curthread == dp->dp_tx.tx_sync_thread)
2197 /* Pool initialization outside of zio_taskq context. */
2198 if (dp && spa_is_initializing(dp->dp_spa) &&
2199 !zio_taskq_member(zio, ZIO_TASKQ_ISSUE) &&
2200 !zio_taskq_member(zio, ZIO_TASKQ_ISSUE_HIGH))
2204 #endif /* HAVE_LARGE_STACKS */
2209 __attribute__((always_inline))
2211 __zio_execute(zio_t *zio)
2213 ASSERT3U(zio->io_queued_timestamp, >, 0);
2215 while (zio->io_stage < ZIO_STAGE_DONE) {
2216 enum zio_stage pipeline = zio->io_pipeline;
2217 enum zio_stage stage = zio->io_stage;
2219 zio->io_executor = curthread;
2221 ASSERT(!MUTEX_HELD(&zio->io_lock));
2222 ASSERT(ISP2(stage));
2223 ASSERT(zio->io_stall == NULL);
2227 } while ((stage & pipeline) == 0);
2229 ASSERT(stage <= ZIO_STAGE_DONE);
2232 * If we are in interrupt context and this pipeline stage
2233 * will grab a config lock that is held across I/O,
2234 * or may wait for an I/O that needs an interrupt thread
2235 * to complete, issue async to avoid deadlock.
2237 * For VDEV_IO_START, we cut in line so that the io will
2238 * be sent to disk promptly.
2240 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
2241 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
2242 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
2243 zio_requeue_io_start_cut_in_line : B_FALSE;
2244 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
2249 * If the current context doesn't have large enough stacks
2250 * the zio must be issued asynchronously to prevent overflow.
2252 if (zio_execute_stack_check(zio)) {
2253 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
2254 zio_requeue_io_start_cut_in_line : B_FALSE;
2255 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
2259 zio->io_stage = stage;
2260 zio->io_pipeline_trace |= zio->io_stage;
2263 * The zio pipeline stage returns the next zio to execute
2264 * (typically the same as this one), or NULL if we should
2267 zio = zio_pipeline[highbit64(stage) - 1](zio);
2276 * ==========================================================================
2277 * Initiate I/O, either sync or async
2278 * ==========================================================================
2281 zio_wait(zio_t *zio)
2284 * Some routines, like zio_free_sync(), may return a NULL zio
2285 * to avoid the performance overhead of creating and then destroying
2286 * an unneeded zio. For the callers' simplicity, we accept a NULL
2287 * zio and ignore it.
2292 long timeout = MSEC_TO_TICK(zfs_deadman_ziotime_ms);
2295 ASSERT3S(zio->io_stage, ==, ZIO_STAGE_OPEN);
2296 ASSERT3P(zio->io_executor, ==, NULL);
2298 zio->io_waiter = curthread;
2299 ASSERT0(zio->io_queued_timestamp);
2300 zio->io_queued_timestamp = gethrtime();
2304 mutex_enter(&zio->io_lock);
2305 while (zio->io_executor != NULL) {
2306 error = cv_timedwait_io(&zio->io_cv, &zio->io_lock,
2307 ddi_get_lbolt() + timeout);
2309 if (zfs_deadman_enabled && error == -1 &&
2310 gethrtime() - zio->io_queued_timestamp >
2311 spa_deadman_ziotime(zio->io_spa)) {
2312 mutex_exit(&zio->io_lock);
2313 timeout = MSEC_TO_TICK(zfs_deadman_checktime_ms);
2314 zio_deadman(zio, FTAG);
2315 mutex_enter(&zio->io_lock);
2318 mutex_exit(&zio->io_lock);
2320 error = zio->io_error;
2327 zio_nowait(zio_t *zio)
2330 * See comment in zio_wait().
2335 ASSERT3P(zio->io_executor, ==, NULL);
2337 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
2338 list_is_empty(&zio->io_parent_list)) {
2342 * This is a logical async I/O with no parent to wait for it.
2343 * We add it to the spa_async_root_zio "Godfather" I/O which
2344 * will ensure they complete prior to unloading the pool.
2346 spa_t *spa = zio->io_spa;
2347 pio = spa->spa_async_zio_root[CPU_SEQID_UNSTABLE];
2349 zio_add_child(pio, zio);
2352 ASSERT0(zio->io_queued_timestamp);
2353 zio->io_queued_timestamp = gethrtime();
2358 * ==========================================================================
2359 * Reexecute, cancel, or suspend/resume failed I/O
2360 * ==========================================================================
2364 zio_reexecute(void *arg)
2367 zio_t *cio, *cio_next;
2369 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
2370 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
2371 ASSERT(pio->io_gang_leader == NULL);
2372 ASSERT(pio->io_gang_tree == NULL);
2374 pio->io_flags = pio->io_orig_flags;
2375 pio->io_stage = pio->io_orig_stage;
2376 pio->io_pipeline = pio->io_orig_pipeline;
2377 pio->io_reexecute = 0;
2378 pio->io_flags |= ZIO_FLAG_REEXECUTED;
2379 pio->io_pipeline_trace = 0;
2381 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2382 pio->io_state[w] = 0;
2383 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2384 pio->io_child_error[c] = 0;
2386 if (IO_IS_ALLOCATING(pio))
2387 BP_ZERO(pio->io_bp);
2390 * As we reexecute pio's children, new children could be created.
2391 * New children go to the head of pio's io_child_list, however,
2392 * so we will (correctly) not reexecute them. The key is that
2393 * the remainder of pio's io_child_list, from 'cio_next' onward,
2394 * cannot be affected by any side effects of reexecuting 'cio'.
2396 zio_link_t *zl = NULL;
2397 mutex_enter(&pio->io_lock);
2398 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
2399 cio_next = zio_walk_children(pio, &zl);
2400 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2401 pio->io_children[cio->io_child_type][w]++;
2402 mutex_exit(&pio->io_lock);
2404 mutex_enter(&pio->io_lock);
2406 mutex_exit(&pio->io_lock);
2409 * Now that all children have been reexecuted, execute the parent.
2410 * We don't reexecute "The Godfather" I/O here as it's the
2411 * responsibility of the caller to wait on it.
2413 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) {
2414 pio->io_queued_timestamp = gethrtime();
2420 zio_suspend(spa_t *spa, zio_t *zio, zio_suspend_reason_t reason)
2422 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
2423 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
2424 "failure and the failure mode property for this pool "
2425 "is set to panic.", spa_name(spa));
2427 cmn_err(CE_WARN, "Pool '%s' has encountered an uncorrectable I/O "
2428 "failure and has been suspended.\n", spa_name(spa));
2430 (void) zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL,
2433 mutex_enter(&spa->spa_suspend_lock);
2435 if (spa->spa_suspend_zio_root == NULL)
2436 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
2437 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
2438 ZIO_FLAG_GODFATHER);
2440 spa->spa_suspended = reason;
2443 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2444 ASSERT(zio != spa->spa_suspend_zio_root);
2445 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2446 ASSERT(zio_unique_parent(zio) == NULL);
2447 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
2448 zio_add_child(spa->spa_suspend_zio_root, zio);
2451 mutex_exit(&spa->spa_suspend_lock);
2455 zio_resume(spa_t *spa)
2460 * Reexecute all previously suspended i/o.
2462 mutex_enter(&spa->spa_suspend_lock);
2463 spa->spa_suspended = ZIO_SUSPEND_NONE;
2464 cv_broadcast(&spa->spa_suspend_cv);
2465 pio = spa->spa_suspend_zio_root;
2466 spa->spa_suspend_zio_root = NULL;
2467 mutex_exit(&spa->spa_suspend_lock);
2473 return (zio_wait(pio));
2477 zio_resume_wait(spa_t *spa)
2479 mutex_enter(&spa->spa_suspend_lock);
2480 while (spa_suspended(spa))
2481 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
2482 mutex_exit(&spa->spa_suspend_lock);
2486 * ==========================================================================
2489 * A gang block is a collection of small blocks that looks to the DMU
2490 * like one large block. When zio_dva_allocate() cannot find a block
2491 * of the requested size, due to either severe fragmentation or the pool
2492 * being nearly full, it calls zio_write_gang_block() to construct the
2493 * block from smaller fragments.
2495 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
2496 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
2497 * an indirect block: it's an array of block pointers. It consumes
2498 * only one sector and hence is allocatable regardless of fragmentation.
2499 * The gang header's bps point to its gang members, which hold the data.
2501 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
2502 * as the verifier to ensure uniqueness of the SHA256 checksum.
2503 * Critically, the gang block bp's blk_cksum is the checksum of the data,
2504 * not the gang header. This ensures that data block signatures (needed for
2505 * deduplication) are independent of how the block is physically stored.
2507 * Gang blocks can be nested: a gang member may itself be a gang block.
2508 * Thus every gang block is a tree in which root and all interior nodes are
2509 * gang headers, and the leaves are normal blocks that contain user data.
2510 * The root of the gang tree is called the gang leader.
2512 * To perform any operation (read, rewrite, free, claim) on a gang block,
2513 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
2514 * in the io_gang_tree field of the original logical i/o by recursively
2515 * reading the gang leader and all gang headers below it. This yields
2516 * an in-core tree containing the contents of every gang header and the
2517 * bps for every constituent of the gang block.
2519 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
2520 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
2521 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
2522 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
2523 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
2524 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
2525 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
2526 * of the gang header plus zio_checksum_compute() of the data to update the
2527 * gang header's blk_cksum as described above.
2529 * The two-phase assemble/issue model solves the problem of partial failure --
2530 * what if you'd freed part of a gang block but then couldn't read the
2531 * gang header for another part? Assembling the entire gang tree first
2532 * ensures that all the necessary gang header I/O has succeeded before
2533 * starting the actual work of free, claim, or write. Once the gang tree
2534 * is assembled, free and claim are in-memory operations that cannot fail.
2536 * In the event that a gang write fails, zio_dva_unallocate() walks the
2537 * gang tree to immediately free (i.e. insert back into the space map)
2538 * everything we've allocated. This ensures that we don't get ENOSPC
2539 * errors during repeated suspend/resume cycles due to a flaky device.
2541 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
2542 * the gang tree, we won't modify the block, so we can safely defer the free
2543 * (knowing that the block is still intact). If we *can* assemble the gang
2544 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
2545 * each constituent bp and we can allocate a new block on the next sync pass.
2547 * In all cases, the gang tree allows complete recovery from partial failure.
2548 * ==========================================================================
2552 zio_gang_issue_func_done(zio_t *zio)
2554 abd_free(zio->io_abd);
2558 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2564 return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset),
2565 BP_GET_PSIZE(bp), zio_gang_issue_func_done,
2566 NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2567 &pio->io_bookmark));
2571 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2578 abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2579 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2580 gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL,
2581 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2584 * As we rewrite each gang header, the pipeline will compute
2585 * a new gang block header checksum for it; but no one will
2586 * compute a new data checksum, so we do that here. The one
2587 * exception is the gang leader: the pipeline already computed
2588 * its data checksum because that stage precedes gang assembly.
2589 * (Presently, nothing actually uses interior data checksums;
2590 * this is just good hygiene.)
2592 if (gn != pio->io_gang_leader->io_gang_tree) {
2593 abd_t *buf = abd_get_offset(data, offset);
2595 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
2596 buf, BP_GET_PSIZE(bp));
2601 * If we are here to damage data for testing purposes,
2602 * leave the GBH alone so that we can detect the damage.
2604 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
2605 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2607 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2608 abd_get_offset(data, offset), BP_GET_PSIZE(bp),
2609 zio_gang_issue_func_done, NULL, pio->io_priority,
2610 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2617 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2620 (void) gn, (void) data, (void) offset;
2622 zio_t *zio = zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
2623 ZIO_GANG_CHILD_FLAGS(pio));
2625 zio = zio_null(pio, pio->io_spa,
2626 NULL, NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio));
2632 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2635 (void) gn, (void) data, (void) offset;
2636 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
2637 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
2640 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
2649 static void zio_gang_tree_assemble_done(zio_t *zio);
2651 static zio_gang_node_t *
2652 zio_gang_node_alloc(zio_gang_node_t **gnpp)
2654 zio_gang_node_t *gn;
2656 ASSERT(*gnpp == NULL);
2658 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
2659 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
2666 zio_gang_node_free(zio_gang_node_t **gnpp)
2668 zio_gang_node_t *gn = *gnpp;
2670 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2671 ASSERT(gn->gn_child[g] == NULL);
2673 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2674 kmem_free(gn, sizeof (*gn));
2679 zio_gang_tree_free(zio_gang_node_t **gnpp)
2681 zio_gang_node_t *gn = *gnpp;
2686 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2687 zio_gang_tree_free(&gn->gn_child[g]);
2689 zio_gang_node_free(gnpp);
2693 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
2695 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
2696 abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2698 ASSERT(gio->io_gang_leader == gio);
2699 ASSERT(BP_IS_GANG(bp));
2701 zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2702 zio_gang_tree_assemble_done, gn, gio->io_priority,
2703 ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
2707 zio_gang_tree_assemble_done(zio_t *zio)
2709 zio_t *gio = zio->io_gang_leader;
2710 zio_gang_node_t *gn = zio->io_private;
2711 blkptr_t *bp = zio->io_bp;
2713 ASSERT(gio == zio_unique_parent(zio));
2714 ASSERT(zio->io_child_count == 0);
2719 /* this ABD was created from a linear buf in zio_gang_tree_assemble */
2720 if (BP_SHOULD_BYTESWAP(bp))
2721 byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size);
2723 ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh);
2724 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
2725 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2727 abd_free(zio->io_abd);
2729 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2730 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2731 if (!BP_IS_GANG(gbp))
2733 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
2738 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data,
2741 zio_t *gio = pio->io_gang_leader;
2744 ASSERT(BP_IS_GANG(bp) == !!gn);
2745 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
2746 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
2749 * If you're a gang header, your data is in gn->gn_gbh.
2750 * If you're a gang member, your data is in 'data' and gn == NULL.
2752 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset);
2755 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2757 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2758 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2759 if (BP_IS_HOLE(gbp))
2761 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data,
2763 offset += BP_GET_PSIZE(gbp);
2767 if (gn == gio->io_gang_tree)
2768 ASSERT3U(gio->io_size, ==, offset);
2775 zio_gang_assemble(zio_t *zio)
2777 blkptr_t *bp = zio->io_bp;
2779 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
2780 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2782 zio->io_gang_leader = zio;
2784 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
2790 zio_gang_issue(zio_t *zio)
2792 blkptr_t *bp = zio->io_bp;
2794 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT, ZIO_WAIT_DONE)) {
2798 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
2799 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2801 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
2802 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd,
2805 zio_gang_tree_free(&zio->io_gang_tree);
2807 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2813 zio_write_gang_member_ready(zio_t *zio)
2815 zio_t *pio = zio_unique_parent(zio);
2816 dva_t *cdva = zio->io_bp->blk_dva;
2817 dva_t *pdva = pio->io_bp->blk_dva;
2819 zio_t *gio __maybe_unused = zio->io_gang_leader;
2821 if (BP_IS_HOLE(zio->io_bp))
2824 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
2826 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
2827 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
2828 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2829 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
2830 VERIFY3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
2832 mutex_enter(&pio->io_lock);
2833 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
2834 ASSERT(DVA_GET_GANG(&pdva[d]));
2835 asize = DVA_GET_ASIZE(&pdva[d]);
2836 asize += DVA_GET_ASIZE(&cdva[d]);
2837 DVA_SET_ASIZE(&pdva[d], asize);
2839 mutex_exit(&pio->io_lock);
2843 zio_write_gang_done(zio_t *zio)
2846 * The io_abd field will be NULL for a zio with no data. The io_flags
2847 * will initially have the ZIO_FLAG_NODATA bit flag set, but we can't
2848 * check for it here as it is cleared in zio_ready.
2850 if (zio->io_abd != NULL)
2851 abd_free(zio->io_abd);
2855 zio_write_gang_block(zio_t *pio, metaslab_class_t *mc)
2857 spa_t *spa = pio->io_spa;
2858 blkptr_t *bp = pio->io_bp;
2859 zio_t *gio = pio->io_gang_leader;
2861 zio_gang_node_t *gn, **gnpp;
2862 zio_gbh_phys_t *gbh;
2864 uint64_t txg = pio->io_txg;
2865 uint64_t resid = pio->io_size;
2867 int copies = gio->io_prop.zp_copies;
2870 boolean_t has_data = !(pio->io_flags & ZIO_FLAG_NODATA);
2873 * If one copy was requested, store 2 copies of the GBH, so that we
2874 * can still traverse all the data (e.g. to free or scrub) even if a
2875 * block is damaged. Note that we can't store 3 copies of the GBH in
2876 * all cases, e.g. with encryption, which uses DVA[2] for the IV+salt.
2878 int gbh_copies = copies;
2879 if (gbh_copies == 1) {
2880 gbh_copies = MIN(2, spa_max_replication(spa));
2883 int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER;
2884 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2885 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2888 flags |= METASLAB_ASYNC_ALLOC;
2889 VERIFY(zfs_refcount_held(&mc->mc_allocator[pio->io_allocator].
2890 mca_alloc_slots, pio));
2893 * The logical zio has already placed a reservation for
2894 * 'copies' allocation slots but gang blocks may require
2895 * additional copies. These additional copies
2896 * (i.e. gbh_copies - copies) are guaranteed to succeed
2897 * since metaslab_class_throttle_reserve() always allows
2898 * additional reservations for gang blocks.
2900 VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies,
2901 pio->io_allocator, pio, flags));
2904 error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE,
2905 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags,
2906 &pio->io_alloc_list, pio, pio->io_allocator);
2908 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2909 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2913 * If we failed to allocate the gang block header then
2914 * we remove any additional allocation reservations that
2915 * we placed here. The original reservation will
2916 * be removed when the logical I/O goes to the ready
2919 metaslab_class_throttle_unreserve(mc,
2920 gbh_copies - copies, pio->io_allocator, pio);
2923 pio->io_error = error;
2928 gnpp = &gio->io_gang_tree;
2930 gnpp = pio->io_private;
2931 ASSERT(pio->io_ready == zio_write_gang_member_ready);
2934 gn = zio_gang_node_alloc(gnpp);
2936 memset(gbh, 0, SPA_GANGBLOCKSIZE);
2937 gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE);
2940 * Create the gang header.
2942 zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2943 zio_write_gang_done, NULL, pio->io_priority,
2944 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2947 * Create and nowait the gang children.
2949 for (int g = 0; resid != 0; resid -= lsize, g++) {
2950 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2952 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2954 zp.zp_checksum = gio->io_prop.zp_checksum;
2955 zp.zp_compress = ZIO_COMPRESS_OFF;
2956 zp.zp_complevel = gio->io_prop.zp_complevel;
2957 zp.zp_type = DMU_OT_NONE;
2959 zp.zp_copies = gio->io_prop.zp_copies;
2960 zp.zp_dedup = B_FALSE;
2961 zp.zp_dedup_verify = B_FALSE;
2962 zp.zp_nopwrite = B_FALSE;
2963 zp.zp_encrypt = gio->io_prop.zp_encrypt;
2964 zp.zp_byteorder = gio->io_prop.zp_byteorder;
2965 memset(zp.zp_salt, 0, ZIO_DATA_SALT_LEN);
2966 memset(zp.zp_iv, 0, ZIO_DATA_IV_LEN);
2967 memset(zp.zp_mac, 0, ZIO_DATA_MAC_LEN);
2969 zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2970 has_data ? abd_get_offset(pio->io_abd, pio->io_size -
2971 resid) : NULL, lsize, lsize, &zp,
2972 zio_write_gang_member_ready, NULL, NULL,
2973 zio_write_gang_done, &gn->gn_child[g], pio->io_priority,
2974 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2976 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2977 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2981 * Gang children won't throttle but we should
2982 * account for their work, so reserve an allocation
2983 * slot for them here.
2985 VERIFY(metaslab_class_throttle_reserve(mc,
2986 zp.zp_copies, cio->io_allocator, cio, flags));
2992 * Set pio's pipeline to just wait for zio to finish.
2994 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2997 * We didn't allocate this bp, so make sure it doesn't get unmarked.
2999 pio->io_flags &= ~ZIO_FLAG_FASTWRITE;
3007 * The zio_nop_write stage in the pipeline determines if allocating a
3008 * new bp is necessary. The nopwrite feature can handle writes in
3009 * either syncing or open context (i.e. zil writes) and as a result is
3010 * mutually exclusive with dedup.
3012 * By leveraging a cryptographically secure checksum, such as SHA256, we
3013 * can compare the checksums of the new data and the old to determine if
3014 * allocating a new block is required. Note that our requirements for
3015 * cryptographic strength are fairly weak: there can't be any accidental
3016 * hash collisions, but we don't need to be secure against intentional
3017 * (malicious) collisions. To trigger a nopwrite, you have to be able
3018 * to write the file to begin with, and triggering an incorrect (hash
3019 * collision) nopwrite is no worse than simply writing to the file.
3020 * That said, there are no known attacks against the checksum algorithms
3021 * used for nopwrite, assuming that the salt and the checksums
3022 * themselves remain secret.
3025 zio_nop_write(zio_t *zio)
3027 blkptr_t *bp = zio->io_bp;
3028 blkptr_t *bp_orig = &zio->io_bp_orig;
3029 zio_prop_t *zp = &zio->io_prop;
3031 ASSERT(BP_IS_HOLE(bp));
3032 ASSERT(BP_GET_LEVEL(bp) == 0);
3033 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
3034 ASSERT(zp->zp_nopwrite);
3035 ASSERT(!zp->zp_dedup);
3036 ASSERT(zio->io_bp_override == NULL);
3037 ASSERT(IO_IS_ALLOCATING(zio));
3040 * Check to see if the original bp and the new bp have matching
3041 * characteristics (i.e. same checksum, compression algorithms, etc).
3042 * If they don't then just continue with the pipeline which will
3043 * allocate a new bp.
3045 if (BP_IS_HOLE(bp_orig) ||
3046 !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
3047 ZCHECKSUM_FLAG_NOPWRITE) ||
3048 BP_IS_ENCRYPTED(bp) || BP_IS_ENCRYPTED(bp_orig) ||
3049 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
3050 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
3051 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
3052 zp->zp_copies != BP_GET_NDVAS(bp_orig))
3056 * If the checksums match then reset the pipeline so that we
3057 * avoid allocating a new bp and issuing any I/O.
3059 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
3060 ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
3061 ZCHECKSUM_FLAG_NOPWRITE);
3062 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
3063 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
3064 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
3065 ASSERT3U(bp->blk_prop, ==, bp_orig->blk_prop);
3068 * If we're overwriting a block that is currently on an
3069 * indirect vdev, then ignore the nopwrite request and
3070 * allow a new block to be allocated on a concrete vdev.
3072 spa_config_enter(zio->io_spa, SCL_VDEV, FTAG, RW_READER);
3073 for (int d = 0; d < BP_GET_NDVAS(bp_orig); d++) {
3074 vdev_t *tvd = vdev_lookup_top(zio->io_spa,
3075 DVA_GET_VDEV(&bp_orig->blk_dva[d]));
3076 if (tvd->vdev_ops == &vdev_indirect_ops) {
3077 spa_config_exit(zio->io_spa, SCL_VDEV, FTAG);
3081 spa_config_exit(zio->io_spa, SCL_VDEV, FTAG);
3084 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3085 zio->io_flags |= ZIO_FLAG_NOPWRITE;
3092 * ==========================================================================
3093 * Block Reference Table
3094 * ==========================================================================
3097 zio_brt_free(zio_t *zio)
3103 if (BP_GET_LEVEL(bp) > 0 ||
3104 BP_IS_METADATA(bp) ||
3105 !brt_maybe_exists(zio->io_spa, bp)) {
3109 if (!brt_entry_decref(zio->io_spa, bp)) {
3111 * This isn't the last reference, so we cannot free
3114 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3121 * ==========================================================================
3123 * ==========================================================================
3126 zio_ddt_child_read_done(zio_t *zio)
3128 blkptr_t *bp = zio->io_bp;
3129 ddt_entry_t *dde = zio->io_private;
3131 zio_t *pio = zio_unique_parent(zio);
3133 mutex_enter(&pio->io_lock);
3134 ddp = ddt_phys_select(dde, bp);
3135 if (zio->io_error == 0)
3136 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
3138 if (zio->io_error == 0 && dde->dde_repair_abd == NULL)
3139 dde->dde_repair_abd = zio->io_abd;
3141 abd_free(zio->io_abd);
3142 mutex_exit(&pio->io_lock);
3146 zio_ddt_read_start(zio_t *zio)
3148 blkptr_t *bp = zio->io_bp;
3150 ASSERT(BP_GET_DEDUP(bp));
3151 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
3152 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3154 if (zio->io_child_error[ZIO_CHILD_DDT]) {
3155 ddt_t *ddt = ddt_select(zio->io_spa, bp);
3156 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
3157 ddt_phys_t *ddp = dde->dde_phys;
3158 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
3161 ASSERT(zio->io_vsd == NULL);
3164 if (ddp_self == NULL)
3167 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
3168 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
3170 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
3172 zio_nowait(zio_read(zio, zio->io_spa, &blk,
3173 abd_alloc_for_io(zio->io_size, B_TRUE),
3174 zio->io_size, zio_ddt_child_read_done, dde,
3175 zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) |
3176 ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark));
3181 zio_nowait(zio_read(zio, zio->io_spa, bp,
3182 zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority,
3183 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
3189 zio_ddt_read_done(zio_t *zio)
3191 blkptr_t *bp = zio->io_bp;
3193 if (zio_wait_for_children(zio, ZIO_CHILD_DDT_BIT, ZIO_WAIT_DONE)) {
3197 ASSERT(BP_GET_DEDUP(bp));
3198 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
3199 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3201 if (zio->io_child_error[ZIO_CHILD_DDT]) {
3202 ddt_t *ddt = ddt_select(zio->io_spa, bp);
3203 ddt_entry_t *dde = zio->io_vsd;
3205 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
3209 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
3210 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
3213 if (dde->dde_repair_abd != NULL) {
3214 abd_copy(zio->io_abd, dde->dde_repair_abd,
3216 zio->io_child_error[ZIO_CHILD_DDT] = 0;
3218 ddt_repair_done(ddt, dde);
3222 ASSERT(zio->io_vsd == NULL);
3228 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
3230 spa_t *spa = zio->io_spa;
3231 boolean_t do_raw = !!(zio->io_flags & ZIO_FLAG_RAW);
3233 ASSERT(!(zio->io_bp_override && do_raw));
3236 * Note: we compare the original data, not the transformed data,
3237 * because when zio->io_bp is an override bp, we will not have
3238 * pushed the I/O transforms. That's an important optimization
3239 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
3240 * However, we should never get a raw, override zio so in these
3241 * cases we can compare the io_abd directly. This is useful because
3242 * it allows us to do dedup verification even if we don't have access
3243 * to the original data (for instance, if the encryption keys aren't
3247 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
3248 zio_t *lio = dde->dde_lead_zio[p];
3250 if (lio != NULL && do_raw) {
3251 return (lio->io_size != zio->io_size ||
3252 abd_cmp(zio->io_abd, lio->io_abd) != 0);
3253 } else if (lio != NULL) {
3254 return (lio->io_orig_size != zio->io_orig_size ||
3255 abd_cmp(zio->io_orig_abd, lio->io_orig_abd) != 0);
3259 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
3260 ddt_phys_t *ddp = &dde->dde_phys[p];
3262 if (ddp->ddp_phys_birth != 0 && do_raw) {
3263 blkptr_t blk = *zio->io_bp;
3268 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
3269 psize = BP_GET_PSIZE(&blk);
3271 if (psize != zio->io_size)
3276 tmpabd = abd_alloc_for_io(psize, B_TRUE);
3278 error = zio_wait(zio_read(NULL, spa, &blk, tmpabd,
3279 psize, NULL, NULL, ZIO_PRIORITY_SYNC_READ,
3280 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
3281 ZIO_FLAG_RAW, &zio->io_bookmark));
3284 if (abd_cmp(tmpabd, zio->io_abd) != 0)
3285 error = SET_ERROR(ENOENT);
3290 return (error != 0);
3291 } else if (ddp->ddp_phys_birth != 0) {
3292 arc_buf_t *abuf = NULL;
3293 arc_flags_t aflags = ARC_FLAG_WAIT;
3294 blkptr_t blk = *zio->io_bp;
3297 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
3299 if (BP_GET_LSIZE(&blk) != zio->io_orig_size)
3304 error = arc_read(NULL, spa, &blk,
3305 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
3306 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
3307 &aflags, &zio->io_bookmark);
3310 if (abd_cmp_buf(zio->io_orig_abd, abuf->b_data,
3311 zio->io_orig_size) != 0)
3312 error = SET_ERROR(ENOENT);
3313 arc_buf_destroy(abuf, &abuf);
3317 return (error != 0);
3325 zio_ddt_child_write_ready(zio_t *zio)
3327 int p = zio->io_prop.zp_copies;
3328 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
3329 ddt_entry_t *dde = zio->io_private;
3330 ddt_phys_t *ddp = &dde->dde_phys[p];
3338 ASSERT(dde->dde_lead_zio[p] == zio);
3340 ddt_phys_fill(ddp, zio->io_bp);
3342 zio_link_t *zl = NULL;
3343 while ((pio = zio_walk_parents(zio, &zl)) != NULL)
3344 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
3350 zio_ddt_child_write_done(zio_t *zio)
3352 int p = zio->io_prop.zp_copies;
3353 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
3354 ddt_entry_t *dde = zio->io_private;
3355 ddt_phys_t *ddp = &dde->dde_phys[p];
3359 ASSERT(ddp->ddp_refcnt == 0);
3360 ASSERT(dde->dde_lead_zio[p] == zio);
3361 dde->dde_lead_zio[p] = NULL;
3363 if (zio->io_error == 0) {
3364 zio_link_t *zl = NULL;
3365 while (zio_walk_parents(zio, &zl) != NULL)
3366 ddt_phys_addref(ddp);
3368 ddt_phys_clear(ddp);
3375 zio_ddt_write(zio_t *zio)
3377 spa_t *spa = zio->io_spa;
3378 blkptr_t *bp = zio->io_bp;
3379 uint64_t txg = zio->io_txg;
3380 zio_prop_t *zp = &zio->io_prop;
3381 int p = zp->zp_copies;
3383 ddt_t *ddt = ddt_select(spa, bp);
3387 ASSERT(BP_GET_DEDUP(bp));
3388 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
3389 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
3390 ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW)));
3393 dde = ddt_lookup(ddt, bp, B_TRUE);
3394 ddp = &dde->dde_phys[p];
3396 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
3398 * If we're using a weak checksum, upgrade to a strong checksum
3399 * and try again. If we're already using a strong checksum,
3400 * we can't resolve it, so just convert to an ordinary write.
3401 * (And automatically e-mail a paper to Nature?)
3403 if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
3404 ZCHECKSUM_FLAG_DEDUP)) {
3405 zp->zp_checksum = spa_dedup_checksum(spa);
3406 zio_pop_transforms(zio);
3407 zio->io_stage = ZIO_STAGE_OPEN;
3410 zp->zp_dedup = B_FALSE;
3411 BP_SET_DEDUP(bp, B_FALSE);
3413 ASSERT(!BP_GET_DEDUP(bp));
3414 zio->io_pipeline = ZIO_WRITE_PIPELINE;
3419 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
3420 if (ddp->ddp_phys_birth != 0)
3421 ddt_bp_fill(ddp, bp, txg);
3422 if (dde->dde_lead_zio[p] != NULL)
3423 zio_add_child(zio, dde->dde_lead_zio[p]);
3425 ddt_phys_addref(ddp);
3426 } else if (zio->io_bp_override) {
3427 ASSERT(bp->blk_birth == txg);
3428 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
3429 ddt_phys_fill(ddp, bp);
3430 ddt_phys_addref(ddp);
3432 cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
3433 zio->io_orig_size, zio->io_orig_size, zp,
3434 zio_ddt_child_write_ready, NULL, NULL,
3435 zio_ddt_child_write_done, dde, zio->io_priority,
3436 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
3438 zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL);
3439 dde->dde_lead_zio[p] = cio;
3449 static ddt_entry_t *freedde; /* for debugging */
3452 zio_ddt_free(zio_t *zio)
3454 spa_t *spa = zio->io_spa;
3455 blkptr_t *bp = zio->io_bp;
3456 ddt_t *ddt = ddt_select(spa, bp);
3460 ASSERT(BP_GET_DEDUP(bp));
3461 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3464 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
3466 ddp = ddt_phys_select(dde, bp);
3468 ddt_phys_decref(ddp);
3476 * ==========================================================================
3477 * Allocate and free blocks
3478 * ==========================================================================
3482 zio_io_to_allocate(spa_t *spa, int allocator)
3486 ASSERT(MUTEX_HELD(&spa->spa_allocs[allocator].spaa_lock));
3488 zio = avl_first(&spa->spa_allocs[allocator].spaa_tree);
3492 ASSERT(IO_IS_ALLOCATING(zio));
3495 * Try to place a reservation for this zio. If we're unable to
3496 * reserve then we throttle.
3498 ASSERT3U(zio->io_allocator, ==, allocator);
3499 if (!metaslab_class_throttle_reserve(zio->io_metaslab_class,
3500 zio->io_prop.zp_copies, allocator, zio, 0)) {
3504 avl_remove(&spa->spa_allocs[allocator].spaa_tree, zio);
3505 ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE);
3511 zio_dva_throttle(zio_t *zio)
3513 spa_t *spa = zio->io_spa;
3515 metaslab_class_t *mc;
3517 /* locate an appropriate allocation class */
3518 mc = spa_preferred_class(spa, zio->io_size, zio->io_prop.zp_type,
3519 zio->io_prop.zp_level, zio->io_prop.zp_zpl_smallblk);
3521 if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE ||
3522 !mc->mc_alloc_throttle_enabled ||
3523 zio->io_child_type == ZIO_CHILD_GANG ||
3524 zio->io_flags & ZIO_FLAG_NODATA) {
3528 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3529 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3530 ASSERT3U(zio->io_queued_timestamp, >, 0);
3531 ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE);
3533 zbookmark_phys_t *bm = &zio->io_bookmark;
3535 * We want to try to use as many allocators as possible to help improve
3536 * performance, but we also want logically adjacent IOs to be physically
3537 * adjacent to improve sequential read performance. We chunk each object
3538 * into 2^20 block regions, and then hash based on the objset, object,
3539 * level, and region to accomplish both of these goals.
3541 int allocator = (uint_t)cityhash4(bm->zb_objset, bm->zb_object,
3542 bm->zb_level, bm->zb_blkid >> 20) % spa->spa_alloc_count;
3543 zio->io_allocator = allocator;
3544 zio->io_metaslab_class = mc;
3545 mutex_enter(&spa->spa_allocs[allocator].spaa_lock);
3546 avl_add(&spa->spa_allocs[allocator].spaa_tree, zio);
3547 nio = zio_io_to_allocate(spa, allocator);
3548 mutex_exit(&spa->spa_allocs[allocator].spaa_lock);
3553 zio_allocate_dispatch(spa_t *spa, int allocator)
3557 mutex_enter(&spa->spa_allocs[allocator].spaa_lock);
3558 zio = zio_io_to_allocate(spa, allocator);
3559 mutex_exit(&spa->spa_allocs[allocator].spaa_lock);
3563 ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE);
3564 ASSERT0(zio->io_error);
3565 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE);
3569 zio_dva_allocate(zio_t *zio)
3571 spa_t *spa = zio->io_spa;
3572 metaslab_class_t *mc;
3573 blkptr_t *bp = zio->io_bp;
3577 if (zio->io_gang_leader == NULL) {
3578 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3579 zio->io_gang_leader = zio;
3582 ASSERT(BP_IS_HOLE(bp));
3583 ASSERT0(BP_GET_NDVAS(bp));
3584 ASSERT3U(zio->io_prop.zp_copies, >, 0);
3585 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
3586 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
3588 flags |= (zio->io_flags & ZIO_FLAG_FASTWRITE) ? METASLAB_FASTWRITE : 0;
3589 if (zio->io_flags & ZIO_FLAG_NODATA)
3590 flags |= METASLAB_DONT_THROTTLE;
3591 if (zio->io_flags & ZIO_FLAG_GANG_CHILD)
3592 flags |= METASLAB_GANG_CHILD;
3593 if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE)
3594 flags |= METASLAB_ASYNC_ALLOC;
3597 * if not already chosen, locate an appropriate allocation class
3599 mc = zio->io_metaslab_class;
3601 mc = spa_preferred_class(spa, zio->io_size,
3602 zio->io_prop.zp_type, zio->io_prop.zp_level,
3603 zio->io_prop.zp_zpl_smallblk);
3604 zio->io_metaslab_class = mc;
3608 * Try allocating the block in the usual metaslab class.
3609 * If that's full, allocate it in the normal class.
3610 * If that's full, allocate as a gang block,
3611 * and if all are full, the allocation fails (which shouldn't happen).
3613 * Note that we do not fall back on embedded slog (ZIL) space, to
3614 * preserve unfragmented slog space, which is critical for decent
3615 * sync write performance. If a log allocation fails, we will fall
3616 * back to spa_sync() which is abysmal for performance.
3618 error = metaslab_alloc(spa, mc, zio->io_size, bp,
3619 zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
3620 &zio->io_alloc_list, zio, zio->io_allocator);
3623 * Fallback to normal class when an alloc class is full
3625 if (error == ENOSPC && mc != spa_normal_class(spa)) {
3627 * If throttling, transfer reservation over to normal class.
3628 * The io_allocator slot can remain the same even though we
3629 * are switching classes.
3631 if (mc->mc_alloc_throttle_enabled &&
3632 (zio->io_flags & ZIO_FLAG_IO_ALLOCATING)) {
3633 metaslab_class_throttle_unreserve(mc,
3634 zio->io_prop.zp_copies, zio->io_allocator, zio);
3635 zio->io_flags &= ~ZIO_FLAG_IO_ALLOCATING;
3637 VERIFY(metaslab_class_throttle_reserve(
3638 spa_normal_class(spa),
3639 zio->io_prop.zp_copies, zio->io_allocator, zio,
3640 flags | METASLAB_MUST_RESERVE));
3642 zio->io_metaslab_class = mc = spa_normal_class(spa);
3643 if (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC) {
3644 zfs_dbgmsg("%s: metaslab allocation failure, "
3645 "trying normal class: zio %px, size %llu, error %d",
3646 spa_name(spa), zio, (u_longlong_t)zio->io_size,
3650 error = metaslab_alloc(spa, mc, zio->io_size, bp,
3651 zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
3652 &zio->io_alloc_list, zio, zio->io_allocator);
3655 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE) {
3656 if (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC) {
3657 zfs_dbgmsg("%s: metaslab allocation failure, "
3658 "trying ganging: zio %px, size %llu, error %d",
3659 spa_name(spa), zio, (u_longlong_t)zio->io_size,
3662 return (zio_write_gang_block(zio, mc));
3665 if (error != ENOSPC ||
3666 (zfs_flags & ZFS_DEBUG_METASLAB_ALLOC)) {
3667 zfs_dbgmsg("%s: metaslab allocation failure: zio %px, "
3668 "size %llu, error %d",
3669 spa_name(spa), zio, (u_longlong_t)zio->io_size,
3672 zio->io_error = error;
3679 zio_dva_free(zio_t *zio)
3681 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
3687 zio_dva_claim(zio_t *zio)
3691 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
3693 zio->io_error = error;
3699 * Undo an allocation. This is used by zio_done() when an I/O fails
3700 * and we want to give back the block we just allocated.
3701 * This handles both normal blocks and gang blocks.
3704 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
3706 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
3707 ASSERT(zio->io_bp_override == NULL);
3709 if (!BP_IS_HOLE(bp))
3710 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
3713 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
3714 zio_dva_unallocate(zio, gn->gn_child[g],
3715 &gn->gn_gbh->zg_blkptr[g]);
3721 * Try to allocate an intent log block. Return 0 on success, errno on failure.
3724 zio_alloc_zil(spa_t *spa, objset_t *os, uint64_t txg, blkptr_t *new_bp,
3725 uint64_t size, boolean_t *slog)
3728 zio_alloc_list_t io_alloc_list;
3730 ASSERT(txg > spa_syncing_txg(spa));
3732 metaslab_trace_init(&io_alloc_list);
3735 * Block pointer fields are useful to metaslabs for stats and debugging.
3736 * Fill in the obvious ones before calling into metaslab_alloc().
3738 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3739 BP_SET_PSIZE(new_bp, size);
3740 BP_SET_LEVEL(new_bp, 0);
3743 * When allocating a zil block, we don't have information about
3744 * the final destination of the block except the objset it's part
3745 * of, so we just hash the objset ID to pick the allocator to get
3748 int flags = METASLAB_FASTWRITE | METASLAB_ZIL;
3749 int allocator = (uint_t)cityhash4(0, 0, 0,
3750 os->os_dsl_dataset->ds_object) % spa->spa_alloc_count;
3751 error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1,
3752 txg, NULL, flags, &io_alloc_list, NULL, allocator);
3753 *slog = (error == 0);
3755 error = metaslab_alloc(spa, spa_embedded_log_class(spa), size,
3756 new_bp, 1, txg, NULL, flags,
3757 &io_alloc_list, NULL, allocator);
3760 error = metaslab_alloc(spa, spa_normal_class(spa), size,
3761 new_bp, 1, txg, NULL, flags,
3762 &io_alloc_list, NULL, allocator);
3764 metaslab_trace_fini(&io_alloc_list);
3767 BP_SET_LSIZE(new_bp, size);
3768 BP_SET_PSIZE(new_bp, size);
3769 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
3770 BP_SET_CHECKSUM(new_bp,
3771 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
3772 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
3773 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3774 BP_SET_LEVEL(new_bp, 0);
3775 BP_SET_DEDUP(new_bp, 0);
3776 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
3779 * encrypted blocks will require an IV and salt. We generate
3780 * these now since we will not be rewriting the bp at
3783 if (os->os_encrypted) {
3784 uint8_t iv[ZIO_DATA_IV_LEN];
3785 uint8_t salt[ZIO_DATA_SALT_LEN];
3787 BP_SET_CRYPT(new_bp, B_TRUE);
3788 VERIFY0(spa_crypt_get_salt(spa,
3789 dmu_objset_id(os), salt));
3790 VERIFY0(zio_crypt_generate_iv(iv));
3792 zio_crypt_encode_params_bp(new_bp, salt, iv);
3795 zfs_dbgmsg("%s: zil block allocation failure: "
3796 "size %llu, error %d", spa_name(spa), (u_longlong_t)size,
3804 * ==========================================================================
3805 * Read and write to physical devices
3806 * ==========================================================================
3810 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3811 * stops after this stage and will resume upon I/O completion.
3812 * However, there are instances where the vdev layer may need to
3813 * continue the pipeline when an I/O was not issued. Since the I/O
3814 * that was sent to the vdev layer might be different than the one
3815 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3816 * force the underlying vdev layers to call either zio_execute() or
3817 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3820 zio_vdev_io_start(zio_t *zio)
3822 vdev_t *vd = zio->io_vd;
3824 spa_t *spa = zio->io_spa;
3828 ASSERT(zio->io_error == 0);
3829 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
3832 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3833 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
3836 * The mirror_ops handle multiple DVAs in a single BP.
3838 vdev_mirror_ops.vdev_op_io_start(zio);
3842 ASSERT3P(zio->io_logical, !=, zio);
3843 if (zio->io_type == ZIO_TYPE_WRITE) {
3844 ASSERT(spa->spa_trust_config);
3847 * Note: the code can handle other kinds of writes,
3848 * but we don't expect them.
3850 if (zio->io_vd->vdev_noalloc) {
3851 ASSERT(zio->io_flags &
3852 (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL |
3853 ZIO_FLAG_RESILVER | ZIO_FLAG_INDUCE_DAMAGE));
3857 align = 1ULL << vd->vdev_top->vdev_ashift;
3859 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
3860 P2PHASE(zio->io_size, align) != 0) {
3861 /* Transform logical writes to be a full physical block size. */
3862 uint64_t asize = P2ROUNDUP(zio->io_size, align);
3863 abd_t *abuf = abd_alloc_sametype(zio->io_abd, asize);
3864 ASSERT(vd == vd->vdev_top);
3865 if (zio->io_type == ZIO_TYPE_WRITE) {
3866 abd_copy(abuf, zio->io_abd, zio->io_size);
3867 abd_zero_off(abuf, zio->io_size, asize - zio->io_size);
3869 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
3873 * If this is not a physical io, make sure that it is properly aligned
3874 * before proceeding.
3876 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
3877 ASSERT0(P2PHASE(zio->io_offset, align));
3878 ASSERT0(P2PHASE(zio->io_size, align));
3881 * For physical writes, we allow 512b aligned writes and assume
3882 * the device will perform a read-modify-write as necessary.
3884 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
3885 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
3888 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
3891 * If this is a repair I/O, and there's no self-healing involved --
3892 * that is, we're just resilvering what we expect to resilver --
3893 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3894 * This prevents spurious resilvering.
3896 * There are a few ways that we can end up creating these spurious
3899 * 1. A resilver i/o will be issued if any DVA in the BP has a
3900 * dirty DTL. The mirror code will issue resilver writes to
3901 * each DVA, including the one(s) that are not on vdevs with dirty
3904 * 2. With nested replication, which happens when we have a
3905 * "replacing" or "spare" vdev that's a child of a mirror or raidz.
3906 * For example, given mirror(replacing(A+B), C), it's likely that
3907 * only A is out of date (it's the new device). In this case, we'll
3908 * read from C, then use the data to resilver A+B -- but we don't
3909 * actually want to resilver B, just A. The top-level mirror has no
3910 * way to know this, so instead we just discard unnecessary repairs
3911 * as we work our way down the vdev tree.
3913 * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc.
3914 * The same logic applies to any form of nested replication: ditto
3915 * + mirror, RAID-Z + replacing, etc.
3917 * However, indirect vdevs point off to other vdevs which may have
3918 * DTL's, so we never bypass them. The child i/os on concrete vdevs
3919 * will be properly bypassed instead.
3921 * Leaf DTL_PARTIAL can be empty when a legitimate write comes from
3922 * a dRAID spare vdev. For example, when a dRAID spare is first
3923 * used, its spare blocks need to be written to but the leaf vdev's
3924 * of such blocks can have empty DTL_PARTIAL.
3926 * There seemed no clean way to allow such writes while bypassing
3927 * spurious ones. At this point, just avoid all bypassing for dRAID
3930 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3931 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
3932 zio->io_txg != 0 && /* not a delegated i/o */
3933 vd->vdev_ops != &vdev_indirect_ops &&
3934 vd->vdev_top->vdev_ops != &vdev_draid_ops &&
3935 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
3936 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3937 zio_vdev_io_bypass(zio);
3942 * Select the next best leaf I/O to process. Distributed spares are
3943 * excluded since they dispatch the I/O directly to a leaf vdev after
3944 * applying the dRAID mapping.
3946 if (vd->vdev_ops->vdev_op_leaf &&
3947 vd->vdev_ops != &vdev_draid_spare_ops &&
3948 (zio->io_type == ZIO_TYPE_READ ||
3949 zio->io_type == ZIO_TYPE_WRITE ||
3950 zio->io_type == ZIO_TYPE_TRIM)) {
3952 if ((zio = vdev_queue_io(zio)) == NULL)
3955 if (!vdev_accessible(vd, zio)) {
3956 zio->io_error = SET_ERROR(ENXIO);
3960 zio->io_delay = gethrtime();
3963 vd->vdev_ops->vdev_op_io_start(zio);
3968 zio_vdev_io_done(zio_t *zio)
3970 vdev_t *vd = zio->io_vd;
3971 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
3972 boolean_t unexpected_error = B_FALSE;
3974 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3978 ASSERT(zio->io_type == ZIO_TYPE_READ ||
3979 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_TRIM);
3982 zio->io_delay = gethrtime() - zio->io_delay;
3984 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3985 vd->vdev_ops != &vdev_draid_spare_ops) {
3986 vdev_queue_io_done(zio);
3988 if (zio_injection_enabled && zio->io_error == 0)
3989 zio->io_error = zio_handle_device_injections(vd, zio,
3992 if (zio_injection_enabled && zio->io_error == 0)
3993 zio->io_error = zio_handle_label_injection(zio, EIO);
3995 if (zio->io_error && zio->io_type != ZIO_TYPE_TRIM) {
3996 if (!vdev_accessible(vd, zio)) {
3997 zio->io_error = SET_ERROR(ENXIO);
3999 unexpected_error = B_TRUE;
4004 ops->vdev_op_io_done(zio);
4006 if (unexpected_error && vd->vdev_remove_wanted == B_FALSE)
4007 VERIFY(vdev_probe(vd, zio) == NULL);
4013 * This function is used to change the priority of an existing zio that is
4014 * currently in-flight. This is used by the arc to upgrade priority in the
4015 * event that a demand read is made for a block that is currently queued
4016 * as a scrub or async read IO. Otherwise, the high priority read request
4017 * would end up having to wait for the lower priority IO.
4020 zio_change_priority(zio_t *pio, zio_priority_t priority)
4022 zio_t *cio, *cio_next;
4023 zio_link_t *zl = NULL;
4025 ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
4027 if (pio->io_vd != NULL && pio->io_vd->vdev_ops->vdev_op_leaf) {
4028 vdev_queue_change_io_priority(pio, priority);
4030 pio->io_priority = priority;
4033 mutex_enter(&pio->io_lock);
4034 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
4035 cio_next = zio_walk_children(pio, &zl);
4036 zio_change_priority(cio, priority);
4038 mutex_exit(&pio->io_lock);
4042 * For non-raidz ZIOs, we can just copy aside the bad data read from the
4043 * disk, and use that to finish the checksum ereport later.
4046 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
4047 const abd_t *good_buf)
4049 /* no processing needed */
4050 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
4054 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr)
4056 void *abd = abd_alloc_sametype(zio->io_abd, zio->io_size);
4058 abd_copy(abd, zio->io_abd, zio->io_size);
4060 zcr->zcr_cbinfo = zio->io_size;
4061 zcr->zcr_cbdata = abd;
4062 zcr->zcr_finish = zio_vsd_default_cksum_finish;
4063 zcr->zcr_free = zio_abd_free;
4067 zio_vdev_io_assess(zio_t *zio)
4069 vdev_t *vd = zio->io_vd;
4071 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
4075 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
4076 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
4078 if (zio->io_vsd != NULL) {
4079 zio->io_vsd_ops->vsd_free(zio);
4083 if (zio_injection_enabled && zio->io_error == 0)
4084 zio->io_error = zio_handle_fault_injection(zio, EIO);
4087 * If the I/O failed, determine whether we should attempt to retry it.
4089 * On retry, we cut in line in the issue queue, since we don't want
4090 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
4092 if (zio->io_error && vd == NULL &&
4093 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
4094 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
4095 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
4097 zio->io_flags |= ZIO_FLAG_IO_RETRY | ZIO_FLAG_DONT_AGGREGATE;
4098 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
4099 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
4100 zio_requeue_io_start_cut_in_line);
4105 * If we got an error on a leaf device, convert it to ENXIO
4106 * if the device is not accessible at all.
4108 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
4109 !vdev_accessible(vd, zio))
4110 zio->io_error = SET_ERROR(ENXIO);
4113 * If we can't write to an interior vdev (mirror or RAID-Z),
4114 * set vdev_cant_write so that we stop trying to allocate from it.
4116 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
4117 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
4118 vdev_dbgmsg(vd, "zio_vdev_io_assess(zio=%px) setting "
4119 "cant_write=TRUE due to write failure with ENXIO",
4121 vd->vdev_cant_write = B_TRUE;
4125 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
4126 * attempts will ever succeed. In this case we set a persistent
4127 * boolean flag so that we don't bother with it in the future.
4129 if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) &&
4130 zio->io_type == ZIO_TYPE_IOCTL &&
4131 zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL)
4132 vd->vdev_nowritecache = B_TRUE;
4135 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
4137 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
4138 zio->io_physdone != NULL) {
4139 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
4140 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
4141 zio->io_physdone(zio->io_logical);
4148 zio_vdev_io_reissue(zio_t *zio)
4150 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
4151 ASSERT(zio->io_error == 0);
4153 zio->io_stage >>= 1;
4157 zio_vdev_io_redone(zio_t *zio)
4159 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
4161 zio->io_stage >>= 1;
4165 zio_vdev_io_bypass(zio_t *zio)
4167 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
4168 ASSERT(zio->io_error == 0);
4170 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
4171 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
4175 * ==========================================================================
4176 * Encrypt and store encryption parameters
4177 * ==========================================================================
4182 * This function is used for ZIO_STAGE_ENCRYPT. It is responsible for
4183 * managing the storage of encryption parameters and passing them to the
4184 * lower-level encryption functions.
4187 zio_encrypt(zio_t *zio)
4189 zio_prop_t *zp = &zio->io_prop;
4190 spa_t *spa = zio->io_spa;
4191 blkptr_t *bp = zio->io_bp;
4192 uint64_t psize = BP_GET_PSIZE(bp);
4193 uint64_t dsobj = zio->io_bookmark.zb_objset;
4194 dmu_object_type_t ot = BP_GET_TYPE(bp);
4195 void *enc_buf = NULL;
4197 uint8_t salt[ZIO_DATA_SALT_LEN];
4198 uint8_t iv[ZIO_DATA_IV_LEN];
4199 uint8_t mac[ZIO_DATA_MAC_LEN];
4200 boolean_t no_crypt = B_FALSE;
4202 /* the root zio already encrypted the data */
4203 if (zio->io_child_type == ZIO_CHILD_GANG)
4206 /* only ZIL blocks are re-encrypted on rewrite */
4207 if (!IO_IS_ALLOCATING(zio) && ot != DMU_OT_INTENT_LOG)
4210 if (!(zp->zp_encrypt || BP_IS_ENCRYPTED(bp))) {
4211 BP_SET_CRYPT(bp, B_FALSE);
4215 /* if we are doing raw encryption set the provided encryption params */
4216 if (zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) {
4217 ASSERT0(BP_GET_LEVEL(bp));
4218 BP_SET_CRYPT(bp, B_TRUE);
4219 BP_SET_BYTEORDER(bp, zp->zp_byteorder);
4220 if (ot != DMU_OT_OBJSET)
4221 zio_crypt_encode_mac_bp(bp, zp->zp_mac);
4223 /* dnode blocks must be written out in the provided byteorder */
4224 if (zp->zp_byteorder != ZFS_HOST_BYTEORDER &&
4225 ot == DMU_OT_DNODE) {
4226 void *bswap_buf = zio_buf_alloc(psize);
4227 abd_t *babd = abd_get_from_buf(bswap_buf, psize);
4229 ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
4230 abd_copy_to_buf(bswap_buf, zio->io_abd, psize);
4231 dmu_ot_byteswap[DMU_OT_BYTESWAP(ot)].ob_func(bswap_buf,
4234 abd_take_ownership_of_buf(babd, B_TRUE);
4235 zio_push_transform(zio, babd, psize, psize, NULL);
4238 if (DMU_OT_IS_ENCRYPTED(ot))
4239 zio_crypt_encode_params_bp(bp, zp->zp_salt, zp->zp_iv);
4243 /* indirect blocks only maintain a cksum of the lower level MACs */
4244 if (BP_GET_LEVEL(bp) > 0) {
4245 BP_SET_CRYPT(bp, B_TRUE);
4246 VERIFY0(zio_crypt_do_indirect_mac_checksum_abd(B_TRUE,
4247 zio->io_orig_abd, BP_GET_LSIZE(bp), BP_SHOULD_BYTESWAP(bp),
4249 zio_crypt_encode_mac_bp(bp, mac);
4254 * Objset blocks are a special case since they have 2 256-bit MACs
4255 * embedded within them.
4257 if (ot == DMU_OT_OBJSET) {
4258 ASSERT0(DMU_OT_IS_ENCRYPTED(ot));
4259 ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
4260 BP_SET_CRYPT(bp, B_TRUE);
4261 VERIFY0(spa_do_crypt_objset_mac_abd(B_TRUE, spa, dsobj,
4262 zio->io_abd, psize, BP_SHOULD_BYTESWAP(bp)));
4266 /* unencrypted object types are only authenticated with a MAC */
4267 if (!DMU_OT_IS_ENCRYPTED(ot)) {
4268 BP_SET_CRYPT(bp, B_TRUE);
4269 VERIFY0(spa_do_crypt_mac_abd(B_TRUE, spa, dsobj,
4270 zio->io_abd, psize, mac));
4271 zio_crypt_encode_mac_bp(bp, mac);
4276 * Later passes of sync-to-convergence may decide to rewrite data
4277 * in place to avoid more disk reallocations. This presents a problem
4278 * for encryption because this constitutes rewriting the new data with
4279 * the same encryption key and IV. However, this only applies to blocks
4280 * in the MOS (particularly the spacemaps) and we do not encrypt the
4281 * MOS. We assert that the zio is allocating or an intent log write
4284 ASSERT(IO_IS_ALLOCATING(zio) || ot == DMU_OT_INTENT_LOG);
4285 ASSERT(BP_GET_LEVEL(bp) == 0 || ot == DMU_OT_INTENT_LOG);
4286 ASSERT(spa_feature_is_active(spa, SPA_FEATURE_ENCRYPTION));
4287 ASSERT3U(psize, !=, 0);
4289 enc_buf = zio_buf_alloc(psize);
4290 eabd = abd_get_from_buf(enc_buf, psize);
4291 abd_take_ownership_of_buf(eabd, B_TRUE);
4294 * For an explanation of what encryption parameters are stored
4295 * where, see the block comment in zio_crypt.c.
4297 if (ot == DMU_OT_INTENT_LOG) {
4298 zio_crypt_decode_params_bp(bp, salt, iv);
4300 BP_SET_CRYPT(bp, B_TRUE);
4303 /* Perform the encryption. This should not fail */
4304 VERIFY0(spa_do_crypt_abd(B_TRUE, spa, &zio->io_bookmark,
4305 BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp),
4306 salt, iv, mac, psize, zio->io_abd, eabd, &no_crypt));
4308 /* encode encryption metadata into the bp */
4309 if (ot == DMU_OT_INTENT_LOG) {
4311 * ZIL blocks store the MAC in the embedded checksum, so the
4312 * transform must always be applied.
4314 zio_crypt_encode_mac_zil(enc_buf, mac);
4315 zio_push_transform(zio, eabd, psize, psize, NULL);
4317 BP_SET_CRYPT(bp, B_TRUE);
4318 zio_crypt_encode_params_bp(bp, salt, iv);
4319 zio_crypt_encode_mac_bp(bp, mac);
4322 ASSERT3U(ot, ==, DMU_OT_DNODE);
4325 zio_push_transform(zio, eabd, psize, psize, NULL);
4333 * ==========================================================================
4334 * Generate and verify checksums
4335 * ==========================================================================
4338 zio_checksum_generate(zio_t *zio)
4340 blkptr_t *bp = zio->io_bp;
4341 enum zio_checksum checksum;
4345 * This is zio_write_phys().
4346 * We're either generating a label checksum, or none at all.
4348 checksum = zio->io_prop.zp_checksum;
4350 if (checksum == ZIO_CHECKSUM_OFF)
4353 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
4355 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
4356 ASSERT(!IO_IS_ALLOCATING(zio));
4357 checksum = ZIO_CHECKSUM_GANG_HEADER;
4359 checksum = BP_GET_CHECKSUM(bp);
4363 zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size);
4369 zio_checksum_verify(zio_t *zio)
4371 zio_bad_cksum_t info;
4372 blkptr_t *bp = zio->io_bp;
4375 ASSERT(zio->io_vd != NULL);
4379 * This is zio_read_phys().
4380 * We're either verifying a label checksum, or nothing at all.
4382 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
4385 ASSERT3U(zio->io_prop.zp_checksum, ==, ZIO_CHECKSUM_LABEL);
4388 if ((error = zio_checksum_error(zio, &info)) != 0) {
4389 zio->io_error = error;
4390 if (error == ECKSUM &&
4391 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
4392 mutex_enter(&zio->io_vd->vdev_stat_lock);
4393 zio->io_vd->vdev_stat.vs_checksum_errors++;
4394 mutex_exit(&zio->io_vd->vdev_stat_lock);
4395 (void) zfs_ereport_start_checksum(zio->io_spa,
4396 zio->io_vd, &zio->io_bookmark, zio,
4397 zio->io_offset, zio->io_size, &info);
4405 * Called by RAID-Z to ensure we don't compute the checksum twice.
4408 zio_checksum_verified(zio_t *zio)
4410 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
4414 * ==========================================================================
4415 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
4416 * An error of 0 indicates success. ENXIO indicates whole-device failure,
4417 * which may be transient (e.g. unplugged) or permanent. ECKSUM and EIO
4418 * indicate errors that are specific to one I/O, and most likely permanent.
4419 * Any other error is presumed to be worse because we weren't expecting it.
4420 * ==========================================================================
4423 zio_worst_error(int e1, int e2)
4425 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
4428 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
4429 if (e1 == zio_error_rank[r1])
4432 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
4433 if (e2 == zio_error_rank[r2])
4436 return (r1 > r2 ? e1 : e2);
4440 * ==========================================================================
4442 * ==========================================================================
4445 zio_ready(zio_t *zio)
4447 blkptr_t *bp = zio->io_bp;
4448 zio_t *pio, *pio_next;
4449 zio_link_t *zl = NULL;
4451 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT | ZIO_CHILD_DDT_BIT,
4456 if (zio->io_ready) {
4457 ASSERT(IO_IS_ALLOCATING(zio));
4458 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
4459 (zio->io_flags & ZIO_FLAG_NOPWRITE));
4460 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
4465 if (bp != NULL && bp != &zio->io_bp_copy)
4466 zio->io_bp_copy = *bp;
4468 if (zio->io_error != 0) {
4469 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
4471 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
4472 ASSERT(IO_IS_ALLOCATING(zio));
4473 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
4474 ASSERT(zio->io_metaslab_class != NULL);
4477 * We were unable to allocate anything, unreserve and
4478 * issue the next I/O to allocate.
4480 metaslab_class_throttle_unreserve(
4481 zio->io_metaslab_class, zio->io_prop.zp_copies,
4482 zio->io_allocator, zio);
4483 zio_allocate_dispatch(zio->io_spa, zio->io_allocator);
4487 mutex_enter(&zio->io_lock);
4488 zio->io_state[ZIO_WAIT_READY] = 1;
4489 pio = zio_walk_parents(zio, &zl);
4490 mutex_exit(&zio->io_lock);
4493 * As we notify zio's parents, new parents could be added.
4494 * New parents go to the head of zio's io_parent_list, however,
4495 * so we will (correctly) not notify them. The remainder of zio's
4496 * io_parent_list, from 'pio_next' onward, cannot change because
4497 * all parents must wait for us to be done before they can be done.
4499 for (; pio != NULL; pio = pio_next) {
4500 pio_next = zio_walk_parents(zio, &zl);
4501 zio_notify_parent(pio, zio, ZIO_WAIT_READY, NULL);
4504 if (zio->io_flags & ZIO_FLAG_NODATA) {
4505 if (bp != NULL && BP_IS_GANG(bp)) {
4506 zio->io_flags &= ~ZIO_FLAG_NODATA;
4508 ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE);
4509 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
4513 if (zio_injection_enabled &&
4514 zio->io_spa->spa_syncing_txg == zio->io_txg)
4515 zio_handle_ignored_writes(zio);
4521 * Update the allocation throttle accounting.
4524 zio_dva_throttle_done(zio_t *zio)
4526 zio_t *lio __maybe_unused = zio->io_logical;
4527 zio_t *pio = zio_unique_parent(zio);
4528 vdev_t *vd = zio->io_vd;
4529 int flags = METASLAB_ASYNC_ALLOC;
4531 ASSERT3P(zio->io_bp, !=, NULL);
4532 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
4533 ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE);
4534 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
4536 ASSERT3P(vd, ==, vd->vdev_top);
4537 ASSERT(zio_injection_enabled || !(zio->io_flags & ZIO_FLAG_IO_RETRY));
4538 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
4539 ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING);
4540 ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE));
4541 ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA));
4544 * Parents of gang children can have two flavors -- ones that
4545 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
4546 * and ones that allocated the constituent blocks. The allocation
4547 * throttle needs to know the allocating parent zio so we must find
4550 if (pio->io_child_type == ZIO_CHILD_GANG) {
4552 * If our parent is a rewrite gang child then our grandparent
4553 * would have been the one that performed the allocation.
4555 if (pio->io_flags & ZIO_FLAG_IO_REWRITE)
4556 pio = zio_unique_parent(pio);
4557 flags |= METASLAB_GANG_CHILD;
4560 ASSERT(IO_IS_ALLOCATING(pio));
4561 ASSERT3P(zio, !=, zio->io_logical);
4562 ASSERT(zio->io_logical != NULL);
4563 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
4564 ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE);
4565 ASSERT(zio->io_metaslab_class != NULL);
4567 mutex_enter(&pio->io_lock);
4568 metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags,
4569 pio->io_allocator, B_TRUE);
4570 mutex_exit(&pio->io_lock);
4572 metaslab_class_throttle_unreserve(zio->io_metaslab_class, 1,
4573 pio->io_allocator, pio);
4576 * Call into the pipeline to see if there is more work that
4577 * needs to be done. If there is work to be done it will be
4578 * dispatched to another taskq thread.
4580 zio_allocate_dispatch(zio->io_spa, pio->io_allocator);
4584 zio_done(zio_t *zio)
4587 * Always attempt to keep stack usage minimal here since
4588 * we can be called recursively up to 19 levels deep.
4590 const uint64_t psize = zio->io_size;
4591 zio_t *pio, *pio_next;
4592 zio_link_t *zl = NULL;
4595 * If our children haven't all completed,
4596 * wait for them and then repeat this pipeline stage.
4598 if (zio_wait_for_children(zio, ZIO_CHILD_ALL_BITS, ZIO_WAIT_DONE)) {
4603 * If the allocation throttle is enabled, then update the accounting.
4604 * We only track child I/Os that are part of an allocating async
4605 * write. We must do this since the allocation is performed
4606 * by the logical I/O but the actual write is done by child I/Os.
4608 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING &&
4609 zio->io_child_type == ZIO_CHILD_VDEV) {
4610 ASSERT(zio->io_metaslab_class != NULL);
4611 ASSERT(zio->io_metaslab_class->mc_alloc_throttle_enabled);
4612 zio_dva_throttle_done(zio);
4616 * If the allocation throttle is enabled, verify that
4617 * we have decremented the refcounts for every I/O that was throttled.
4619 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
4620 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
4621 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
4622 ASSERT(zio->io_bp != NULL);
4624 metaslab_group_alloc_verify(zio->io_spa, zio->io_bp, zio,
4626 VERIFY(zfs_refcount_not_held(&zio->io_metaslab_class->
4627 mc_allocator[zio->io_allocator].mca_alloc_slots, zio));
4631 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
4632 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
4633 ASSERT(zio->io_children[c][w] == 0);
4635 if (zio->io_bp != NULL && !BP_IS_EMBEDDED(zio->io_bp)) {
4636 ASSERT(zio->io_bp->blk_pad[0] == 0);
4637 ASSERT(zio->io_bp->blk_pad[1] == 0);
4638 ASSERT(memcmp(zio->io_bp, &zio->io_bp_copy,
4639 sizeof (blkptr_t)) == 0 ||
4640 (zio->io_bp == zio_unique_parent(zio)->io_bp));
4641 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) &&
4642 zio->io_bp_override == NULL &&
4643 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
4644 ASSERT3U(zio->io_prop.zp_copies, <=,
4645 BP_GET_NDVAS(zio->io_bp));
4646 ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 ||
4647 (BP_COUNT_GANG(zio->io_bp) ==
4648 BP_GET_NDVAS(zio->io_bp)));
4650 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
4651 VERIFY(BP_EQUAL(zio->io_bp, &zio->io_bp_orig));
4655 * If there were child vdev/gang/ddt errors, they apply to us now.
4657 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
4658 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
4659 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
4662 * If the I/O on the transformed data was successful, generate any
4663 * checksum reports now while we still have the transformed data.
4665 if (zio->io_error == 0) {
4666 while (zio->io_cksum_report != NULL) {
4667 zio_cksum_report_t *zcr = zio->io_cksum_report;
4668 uint64_t align = zcr->zcr_align;
4669 uint64_t asize = P2ROUNDUP(psize, align);
4670 abd_t *adata = zio->io_abd;
4672 if (adata != NULL && asize != psize) {
4673 adata = abd_alloc(asize, B_TRUE);
4674 abd_copy(adata, zio->io_abd, psize);
4675 abd_zero_off(adata, psize, asize - psize);
4678 zio->io_cksum_report = zcr->zcr_next;
4679 zcr->zcr_next = NULL;
4680 zcr->zcr_finish(zcr, adata);
4681 zfs_ereport_free_checksum(zcr);
4683 if (adata != NULL && asize != psize)
4688 zio_pop_transforms(zio); /* note: may set zio->io_error */
4690 vdev_stat_update(zio, psize);
4693 * If this I/O is attached to a particular vdev is slow, exceeding
4694 * 30 seconds to complete, post an error described the I/O delay.
4695 * We ignore these errors if the device is currently unavailable.
4697 if (zio->io_delay >= MSEC2NSEC(zio_slow_io_ms)) {
4698 if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd)) {
4700 * We want to only increment our slow IO counters if
4701 * the IO is valid (i.e. not if the drive is removed).
4703 * zfs_ereport_post() will also do these checks, but
4704 * it can also ratelimit and have other failures, so we
4705 * need to increment the slow_io counters independent
4708 if (zfs_ereport_is_valid(FM_EREPORT_ZFS_DELAY,
4709 zio->io_spa, zio->io_vd, zio)) {
4710 mutex_enter(&zio->io_vd->vdev_stat_lock);
4711 zio->io_vd->vdev_stat.vs_slow_ios++;
4712 mutex_exit(&zio->io_vd->vdev_stat_lock);
4714 (void) zfs_ereport_post(FM_EREPORT_ZFS_DELAY,
4715 zio->io_spa, zio->io_vd, &zio->io_bookmark,
4721 if (zio->io_error) {
4723 * If this I/O is attached to a particular vdev,
4724 * generate an error message describing the I/O failure
4725 * at the block level. We ignore these errors if the
4726 * device is currently unavailable.
4728 if (zio->io_error != ECKSUM && zio->io_vd != NULL &&
4729 !vdev_is_dead(zio->io_vd)) {
4730 int ret = zfs_ereport_post(FM_EREPORT_ZFS_IO,
4731 zio->io_spa, zio->io_vd, &zio->io_bookmark, zio, 0);
4732 if (ret != EALREADY) {
4733 mutex_enter(&zio->io_vd->vdev_stat_lock);
4734 if (zio->io_type == ZIO_TYPE_READ)
4735 zio->io_vd->vdev_stat.vs_read_errors++;
4736 else if (zio->io_type == ZIO_TYPE_WRITE)
4737 zio->io_vd->vdev_stat.vs_write_errors++;
4738 mutex_exit(&zio->io_vd->vdev_stat_lock);
4742 if ((zio->io_error == EIO || !(zio->io_flags &
4743 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
4744 zio == zio->io_logical) {
4746 * For logical I/O requests, tell the SPA to log the
4747 * error and generate a logical data ereport.
4749 spa_log_error(zio->io_spa, &zio->io_bookmark,
4750 &zio->io_bp->blk_birth);
4751 (void) zfs_ereport_post(FM_EREPORT_ZFS_DATA,
4752 zio->io_spa, NULL, &zio->io_bookmark, zio, 0);
4756 if (zio->io_error && zio == zio->io_logical) {
4758 * Determine whether zio should be reexecuted. This will
4759 * propagate all the way to the root via zio_notify_parent().
4761 ASSERT(zio->io_vd == NULL && zio->io_bp != NULL);
4762 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
4764 if (IO_IS_ALLOCATING(zio) &&
4765 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
4766 if (zio->io_error != ENOSPC)
4767 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
4769 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4772 if ((zio->io_type == ZIO_TYPE_READ ||
4773 zio->io_type == ZIO_TYPE_FREE) &&
4774 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
4775 zio->io_error == ENXIO &&
4776 spa_load_state(zio->io_spa) == SPA_LOAD_NONE &&
4777 spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE)
4778 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4780 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
4781 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4784 * Here is a possibly good place to attempt to do
4785 * either combinatorial reconstruction or error correction
4786 * based on checksums. It also might be a good place
4787 * to send out preliminary ereports before we suspend
4793 * If there were logical child errors, they apply to us now.
4794 * We defer this until now to avoid conflating logical child
4795 * errors with errors that happened to the zio itself when
4796 * updating vdev stats and reporting FMA events above.
4798 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
4800 if ((zio->io_error || zio->io_reexecute) &&
4801 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
4802 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
4803 zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp);
4805 zio_gang_tree_free(&zio->io_gang_tree);
4808 * Godfather I/Os should never suspend.
4810 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
4811 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
4812 zio->io_reexecute &= ~ZIO_REEXECUTE_SUSPEND;
4814 if (zio->io_reexecute) {
4816 * This is a logical I/O that wants to reexecute.
4818 * Reexecute is top-down. When an i/o fails, if it's not
4819 * the root, it simply notifies its parent and sticks around.
4820 * The parent, seeing that it still has children in zio_done(),
4821 * does the same. This percolates all the way up to the root.
4822 * The root i/o will reexecute or suspend the entire tree.
4824 * This approach ensures that zio_reexecute() honors
4825 * all the original i/o dependency relationships, e.g.
4826 * parents not executing until children are ready.
4828 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
4830 zio->io_gang_leader = NULL;
4832 mutex_enter(&zio->io_lock);
4833 zio->io_state[ZIO_WAIT_DONE] = 1;
4834 mutex_exit(&zio->io_lock);
4837 * "The Godfather" I/O monitors its children but is
4838 * not a true parent to them. It will track them through
4839 * the pipeline but severs its ties whenever they get into
4840 * trouble (e.g. suspended). This allows "The Godfather"
4841 * I/O to return status without blocking.
4844 for (pio = zio_walk_parents(zio, &zl); pio != NULL;
4846 zio_link_t *remove_zl = zl;
4847 pio_next = zio_walk_parents(zio, &zl);
4849 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
4850 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
4851 zio_remove_child(pio, zio, remove_zl);
4853 * This is a rare code path, so we don't
4854 * bother with "next_to_execute".
4856 zio_notify_parent(pio, zio, ZIO_WAIT_DONE,
4861 if ((pio = zio_unique_parent(zio)) != NULL) {
4863 * We're not a root i/o, so there's nothing to do
4864 * but notify our parent. Don't propagate errors
4865 * upward since we haven't permanently failed yet.
4867 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
4868 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
4870 * This is a rare code path, so we don't bother with
4871 * "next_to_execute".
4873 zio_notify_parent(pio, zio, ZIO_WAIT_DONE, NULL);
4874 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
4876 * We'd fail again if we reexecuted now, so suspend
4877 * until conditions improve (e.g. device comes online).
4879 zio_suspend(zio->io_spa, zio, ZIO_SUSPEND_IOERR);
4882 * Reexecution is potentially a huge amount of work.
4883 * Hand it off to the otherwise-unused claim taskq.
4885 ASSERT(taskq_empty_ent(&zio->io_tqent));
4886 spa_taskq_dispatch_ent(zio->io_spa,
4887 ZIO_TYPE_CLAIM, ZIO_TASKQ_ISSUE,
4888 zio_reexecute, zio, 0, &zio->io_tqent);
4893 ASSERT(zio->io_child_count == 0);
4894 ASSERT(zio->io_reexecute == 0);
4895 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
4898 * Report any checksum errors, since the I/O is complete.
4900 while (zio->io_cksum_report != NULL) {
4901 zio_cksum_report_t *zcr = zio->io_cksum_report;
4902 zio->io_cksum_report = zcr->zcr_next;
4903 zcr->zcr_next = NULL;
4904 zcr->zcr_finish(zcr, NULL);
4905 zfs_ereport_free_checksum(zcr);
4908 if (zio->io_flags & ZIO_FLAG_FASTWRITE && zio->io_bp &&
4909 !BP_IS_HOLE(zio->io_bp) && !BP_IS_EMBEDDED(zio->io_bp) &&
4910 !(zio->io_flags & ZIO_FLAG_NOPWRITE)) {
4911 metaslab_fastwrite_unmark(zio->io_spa, zio->io_bp);
4915 * It is the responsibility of the done callback to ensure that this
4916 * particular zio is no longer discoverable for adoption, and as
4917 * such, cannot acquire any new parents.
4922 mutex_enter(&zio->io_lock);
4923 zio->io_state[ZIO_WAIT_DONE] = 1;
4924 mutex_exit(&zio->io_lock);
4927 * We are done executing this zio. We may want to execute a parent
4928 * next. See the comment in zio_notify_parent().
4930 zio_t *next_to_execute = NULL;
4932 for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) {
4933 zio_link_t *remove_zl = zl;
4934 pio_next = zio_walk_parents(zio, &zl);
4935 zio_remove_child(pio, zio, remove_zl);
4936 zio_notify_parent(pio, zio, ZIO_WAIT_DONE, &next_to_execute);
4939 if (zio->io_waiter != NULL) {
4940 mutex_enter(&zio->io_lock);
4941 zio->io_executor = NULL;
4942 cv_broadcast(&zio->io_cv);
4943 mutex_exit(&zio->io_lock);
4948 return (next_to_execute);
4952 * ==========================================================================
4953 * I/O pipeline definition
4954 * ==========================================================================
4956 static zio_pipe_stage_t *zio_pipeline[] = {
4964 zio_checksum_generate,
4981 zio_checksum_verify,
4989 * Compare two zbookmark_phys_t's to see which we would reach first in a
4990 * pre-order traversal of the object tree.
4992 * This is simple in every case aside from the meta-dnode object. For all other
4993 * objects, we traverse them in order (object 1 before object 2, and so on).
4994 * However, all of these objects are traversed while traversing object 0, since
4995 * the data it points to is the list of objects. Thus, we need to convert to a
4996 * canonical representation so we can compare meta-dnode bookmarks to
4997 * non-meta-dnode bookmarks.
4999 * We do this by calculating "equivalents" for each field of the zbookmark.
5000 * zbookmarks outside of the meta-dnode use their own object and level, and
5001 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
5002 * blocks this bookmark refers to) by multiplying their blkid by their span
5003 * (the number of L0 blocks contained within one block at their level).
5004 * zbookmarks inside the meta-dnode calculate their object equivalent
5005 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
5006 * level + 1<<31 (any value larger than a level could ever be) for their level.
5007 * This causes them to always compare before a bookmark in their object
5008 * equivalent, compare appropriately to bookmarks in other objects, and to
5009 * compare appropriately to other bookmarks in the meta-dnode.
5012 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
5013 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
5016 * These variables represent the "equivalent" values for the zbookmark,
5017 * after converting zbookmarks inside the meta dnode to their
5018 * normal-object equivalents.
5020 uint64_t zb1obj, zb2obj;
5021 uint64_t zb1L0, zb2L0;
5022 uint64_t zb1level, zb2level;
5024 if (zb1->zb_object == zb2->zb_object &&
5025 zb1->zb_level == zb2->zb_level &&
5026 zb1->zb_blkid == zb2->zb_blkid)
5029 IMPLY(zb1->zb_level > 0, ibs1 >= SPA_MINBLOCKSHIFT);
5030 IMPLY(zb2->zb_level > 0, ibs2 >= SPA_MINBLOCKSHIFT);
5033 * BP_SPANB calculates the span in blocks.
5035 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
5036 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
5038 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
5039 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
5041 zb1level = zb1->zb_level + COMPARE_META_LEVEL;
5043 zb1obj = zb1->zb_object;
5044 zb1level = zb1->zb_level;
5047 if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
5048 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
5050 zb2level = zb2->zb_level + COMPARE_META_LEVEL;
5052 zb2obj = zb2->zb_object;
5053 zb2level = zb2->zb_level;
5056 /* Now that we have a canonical representation, do the comparison. */
5057 if (zb1obj != zb2obj)
5058 return (zb1obj < zb2obj ? -1 : 1);
5059 else if (zb1L0 != zb2L0)
5060 return (zb1L0 < zb2L0 ? -1 : 1);
5061 else if (zb1level != zb2level)
5062 return (zb1level > zb2level ? -1 : 1);
5064 * This can (theoretically) happen if the bookmarks have the same object
5065 * and level, but different blkids, if the block sizes are not the same.
5066 * There is presently no way to change the indirect block sizes
5072 * This function checks the following: given that last_block is the place that
5073 * our traversal stopped last time, does that guarantee that we've visited
5074 * every node under subtree_root? Therefore, we can't just use the raw output
5075 * of zbookmark_compare. We have to pass in a modified version of
5076 * subtree_root; by incrementing the block id, and then checking whether
5077 * last_block is before or equal to that, we can tell whether or not having
5078 * visited last_block implies that all of subtree_root's children have been
5082 zbookmark_subtree_completed(const dnode_phys_t *dnp,
5083 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
5085 zbookmark_phys_t mod_zb = *subtree_root;
5087 ASSERT0(last_block->zb_level);
5089 /* The objset_phys_t isn't before anything. */
5094 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
5095 * data block size in sectors, because that variable is only used if
5096 * the bookmark refers to a block in the meta-dnode. Since we don't
5097 * know without examining it what object it refers to, and there's no
5098 * harm in passing in this value in other cases, we always pass it in.
5100 * We pass in 0 for the indirect block size shift because zb2 must be
5101 * level 0. The indirect block size is only used to calculate the span
5102 * of the bookmark, but since the bookmark must be level 0, the span is
5103 * always 1, so the math works out.
5105 * If you make changes to how the zbookmark_compare code works, be sure
5106 * to make sure that this code still works afterwards.
5108 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
5109 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,
5114 * This function is similar to zbookmark_subtree_completed(), but returns true
5115 * if subtree_root is equal or ahead of last_block, i.e. still to be done.
5118 zbookmark_subtree_tbd(const dnode_phys_t *dnp,
5119 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
5121 ASSERT0(last_block->zb_level);
5124 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
5125 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, subtree_root,
5129 EXPORT_SYMBOL(zio_type_name);
5130 EXPORT_SYMBOL(zio_buf_alloc);
5131 EXPORT_SYMBOL(zio_data_buf_alloc);
5132 EXPORT_SYMBOL(zio_buf_free);
5133 EXPORT_SYMBOL(zio_data_buf_free);
5135 ZFS_MODULE_PARAM(zfs_zio, zio_, slow_io_ms, INT, ZMOD_RW,
5136 "Max I/O completion time (milliseconds) before marking it as slow");
5138 ZFS_MODULE_PARAM(zfs_zio, zio_, requeue_io_start_cut_in_line, INT, ZMOD_RW,
5139 "Prioritize requeued I/O");
5141 ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_deferred_free, UINT, ZMOD_RW,
5142 "Defer frees starting in this pass");
5144 ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_dont_compress, UINT, ZMOD_RW,
5145 "Don't compress starting in this pass");
5147 ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_rewrite, UINT, ZMOD_RW,
5148 "Rewrite new bps starting in this pass");
5150 ZFS_MODULE_PARAM(zfs_zio, zio_, dva_throttle_enabled, INT, ZMOD_RW,
5151 "Throttle block allocations in the ZIO pipeline");
5153 ZFS_MODULE_PARAM(zfs_zio, zio_, deadman_log_all, INT, ZMOD_RW,
5154 "Log all slow ZIOs, not just those with vdevs");