4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2020 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
30 #include <sys/sysmacros.h>
31 #include <sys/zfs_context.h>
32 #include <sys/fm/fs/zfs.h>
35 #include <sys/spa_impl.h>
36 #include <sys/vdev_impl.h>
37 #include <sys/vdev_trim.h>
38 #include <sys/zio_impl.h>
39 #include <sys/zio_compress.h>
40 #include <sys/zio_checksum.h>
41 #include <sys/dmu_objset.h>
44 #include <sys/blkptr.h>
45 #include <sys/zfeature.h>
46 #include <sys/dsl_scan.h>
47 #include <sys/metaslab_impl.h>
49 #include <sys/trace_zfs.h>
51 #include <sys/dsl_crypt.h>
55 * ==========================================================================
56 * I/O type descriptions
57 * ==========================================================================
59 const char *zio_type_name[ZIO_TYPES] = {
61 * Note: Linux kernel thread name length is limited
62 * so these names will differ from upstream open zfs.
64 "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl", "z_trim"
67 int zio_dva_throttle_enabled = B_TRUE;
68 int zio_deadman_log_all = B_FALSE;
71 * ==========================================================================
73 * ==========================================================================
75 kmem_cache_t *zio_cache;
76 kmem_cache_t *zio_link_cache;
77 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
78 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
79 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
80 uint64_t zio_buf_cache_allocs[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
81 uint64_t zio_buf_cache_frees[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
84 /* Mark IOs as "slow" if they take longer than 30 seconds */
85 int zio_slow_io_ms = (30 * MILLISEC);
87 #define BP_SPANB(indblkshift, level) \
88 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
89 #define COMPARE_META_LEVEL 0x80000000ul
91 * The following actions directly effect the spa's sync-to-convergence logic.
92 * The values below define the sync pass when we start performing the action.
93 * Care should be taken when changing these values as they directly impact
94 * spa_sync() performance. Tuning these values may introduce subtle performance
95 * pathologies and should only be done in the context of performance analysis.
96 * These tunables will eventually be removed and replaced with #defines once
97 * enough analysis has been done to determine optimal values.
99 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
100 * regular blocks are not deferred.
102 * Starting in sync pass 8 (zfs_sync_pass_dont_compress), we disable
103 * compression (including of metadata). In practice, we don't have this
104 * many sync passes, so this has no effect.
106 * The original intent was that disabling compression would help the sync
107 * passes to converge. However, in practice disabling compression increases
108 * the average number of sync passes, because when we turn compression off, a
109 * lot of block's size will change and thus we have to re-allocate (not
110 * overwrite) them. It also increases the number of 128KB allocations (e.g.
111 * for indirect blocks and spacemaps) because these will not be compressed.
112 * The 128K allocations are especially detrimental to performance on highly
113 * fragmented systems, which may have very few free segments of this size,
114 * and may need to load new metaslabs to satisfy 128K allocations.
116 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
117 int zfs_sync_pass_dont_compress = 8; /* don't compress starting in this pass */
118 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
121 * An allocating zio is one that either currently has the DVA allocate
122 * stage set or will have it later in its lifetime.
124 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
127 * Enable smaller cores by excluding metadata
128 * allocations as well.
130 int zio_exclude_metadata = 0;
131 int zio_requeue_io_start_cut_in_line = 1;
134 int zio_buf_debug_limit = 16384;
136 int zio_buf_debug_limit = 0;
139 static inline void __zio_execute(zio_t *zio);
141 static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t);
148 zio_cache = kmem_cache_create("zio_cache",
149 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
150 zio_link_cache = kmem_cache_create("zio_link_cache",
151 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
154 * For small buffers, we want a cache for each multiple of
155 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
156 * for each quarter-power of 2.
158 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
159 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
162 size_t data_cflags, cflags;
164 data_cflags = KMC_NODEBUG;
165 cflags = (zio_exclude_metadata || size > zio_buf_debug_limit) ?
168 #if defined(_ILP32) && defined(_KERNEL)
170 * Cache size limited to 1M on 32-bit platforms until ARC
171 * buffers no longer require virtual address space.
173 if (size > zfs_max_recordsize)
182 * If we are using watchpoints, put each buffer on its own page,
183 * to eliminate the performance overhead of trapping to the
184 * kernel when modifying a non-watched buffer that shares the
185 * page with a watched buffer.
187 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
190 * Here's the problem - on 4K native devices in userland on
191 * Linux using O_DIRECT, buffers must be 4K aligned or I/O
192 * will fail with EINVAL, causing zdb (and others) to coredump.
193 * Since userland probably doesn't need optimized buffer caches,
194 * we just force 4K alignment on everything.
196 align = 8 * SPA_MINBLOCKSIZE;
198 if (size < PAGESIZE) {
199 align = SPA_MINBLOCKSIZE;
200 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
207 if (cflags == data_cflags) {
209 * Resulting kmem caches would be identical.
210 * Save memory by creating only one.
212 (void) snprintf(name, sizeof (name),
213 "zio_buf_comb_%lu", (ulong_t)size);
214 zio_buf_cache[c] = kmem_cache_create(name,
215 size, align, NULL, NULL, NULL, NULL, NULL,
217 zio_data_buf_cache[c] = zio_buf_cache[c];
220 (void) snprintf(name, sizeof (name), "zio_buf_%lu",
222 zio_buf_cache[c] = kmem_cache_create(name, size,
223 align, NULL, NULL, NULL, NULL, NULL, cflags);
225 (void) snprintf(name, sizeof (name), "zio_data_buf_%lu",
227 zio_data_buf_cache[c] = kmem_cache_create(name, size,
228 align, NULL, NULL, NULL, NULL, NULL, data_cflags);
233 ASSERT(zio_buf_cache[c] != NULL);
234 if (zio_buf_cache[c - 1] == NULL)
235 zio_buf_cache[c - 1] = zio_buf_cache[c];
237 ASSERT(zio_data_buf_cache[c] != NULL);
238 if (zio_data_buf_cache[c - 1] == NULL)
239 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
253 n = SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT;
255 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
256 for (i = 0; i < n; i++) {
257 if (zio_buf_cache_allocs[i] != zio_buf_cache_frees[i])
258 (void) printf("zio_fini: [%d] %llu != %llu\n",
259 (int)((i + 1) << SPA_MINBLOCKSHIFT),
260 (long long unsigned)zio_buf_cache_allocs[i],
261 (long long unsigned)zio_buf_cache_frees[i]);
266 * The same kmem cache can show up multiple times in both zio_buf_cache
267 * and zio_data_buf_cache. Do a wasteful but trivially correct scan to
270 for (i = 0; i < n; i++) {
271 cache = zio_buf_cache[i];
274 for (j = i; j < n; j++) {
275 if (cache == zio_buf_cache[j])
276 zio_buf_cache[j] = NULL;
277 if (cache == zio_data_buf_cache[j])
278 zio_data_buf_cache[j] = NULL;
280 kmem_cache_destroy(cache);
283 for (i = 0; i < n; i++) {
284 cache = zio_data_buf_cache[i];
287 for (j = i; j < n; j++) {
288 if (cache == zio_data_buf_cache[j])
289 zio_data_buf_cache[j] = NULL;
291 kmem_cache_destroy(cache);
294 for (i = 0; i < n; i++) {
295 if (zio_buf_cache[i] != NULL)
296 panic("zio_fini: zio_buf_cache[%d] != NULL", (int)i);
297 if (zio_data_buf_cache[i] != NULL)
298 panic("zio_fini: zio_data_buf_cache[%d] != NULL", (int)i);
301 kmem_cache_destroy(zio_link_cache);
302 kmem_cache_destroy(zio_cache);
310 * ==========================================================================
311 * Allocate and free I/O buffers
312 * ==========================================================================
316 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
317 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
318 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
319 * excess / transient data in-core during a crashdump.
322 zio_buf_alloc(size_t size)
324 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
326 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
327 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
328 atomic_add_64(&zio_buf_cache_allocs[c], 1);
331 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
335 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
336 * crashdump if the kernel panics. This exists so that we will limit the amount
337 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
338 * of kernel heap dumped to disk when the kernel panics)
341 zio_data_buf_alloc(size_t size)
343 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
345 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
347 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
351 zio_buf_free(void *buf, size_t size)
353 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
355 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
356 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
357 atomic_add_64(&zio_buf_cache_frees[c], 1);
360 kmem_cache_free(zio_buf_cache[c], buf);
364 zio_data_buf_free(void *buf, size_t size)
366 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
368 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
370 kmem_cache_free(zio_data_buf_cache[c], buf);
374 zio_abd_free(void *abd, size_t size)
376 abd_free((abd_t *)abd);
380 * ==========================================================================
381 * Push and pop I/O transform buffers
382 * ==========================================================================
385 zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize,
386 zio_transform_func_t *transform)
388 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
390 zt->zt_orig_abd = zio->io_abd;
391 zt->zt_orig_size = zio->io_size;
392 zt->zt_bufsize = bufsize;
393 zt->zt_transform = transform;
395 zt->zt_next = zio->io_transform_stack;
396 zio->io_transform_stack = zt;
403 zio_pop_transforms(zio_t *zio)
407 while ((zt = zio->io_transform_stack) != NULL) {
408 if (zt->zt_transform != NULL)
409 zt->zt_transform(zio,
410 zt->zt_orig_abd, zt->zt_orig_size);
412 if (zt->zt_bufsize != 0)
413 abd_free(zio->io_abd);
415 zio->io_abd = zt->zt_orig_abd;
416 zio->io_size = zt->zt_orig_size;
417 zio->io_transform_stack = zt->zt_next;
419 kmem_free(zt, sizeof (zio_transform_t));
424 * ==========================================================================
425 * I/O transform callbacks for subblocks, decompression, and decryption
426 * ==========================================================================
429 zio_subblock(zio_t *zio, abd_t *data, uint64_t size)
431 ASSERT(zio->io_size > size);
433 if (zio->io_type == ZIO_TYPE_READ)
434 abd_copy(data, zio->io_abd, size);
438 zio_decompress(zio_t *zio, abd_t *data, uint64_t size)
440 if (zio->io_error == 0) {
441 void *tmp = abd_borrow_buf(data, size);
442 int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
443 zio->io_abd, tmp, zio->io_size, size,
444 &zio->io_prop.zp_complevel);
445 abd_return_buf_copy(data, tmp, size);
447 if (zio_injection_enabled && ret == 0)
448 ret = zio_handle_fault_injection(zio, EINVAL);
451 zio->io_error = SET_ERROR(EIO);
456 zio_decrypt(zio_t *zio, abd_t *data, uint64_t size)
460 blkptr_t *bp = zio->io_bp;
461 spa_t *spa = zio->io_spa;
462 uint64_t dsobj = zio->io_bookmark.zb_objset;
463 uint64_t lsize = BP_GET_LSIZE(bp);
464 dmu_object_type_t ot = BP_GET_TYPE(bp);
465 uint8_t salt[ZIO_DATA_SALT_LEN];
466 uint8_t iv[ZIO_DATA_IV_LEN];
467 uint8_t mac[ZIO_DATA_MAC_LEN];
468 boolean_t no_crypt = B_FALSE;
470 ASSERT(BP_USES_CRYPT(bp));
471 ASSERT3U(size, !=, 0);
473 if (zio->io_error != 0)
477 * Verify the cksum of MACs stored in an indirect bp. It will always
478 * be possible to verify this since it does not require an encryption
481 if (BP_HAS_INDIRECT_MAC_CKSUM(bp)) {
482 zio_crypt_decode_mac_bp(bp, mac);
484 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF) {
486 * We haven't decompressed the data yet, but
487 * zio_crypt_do_indirect_mac_checksum() requires
488 * decompressed data to be able to parse out the MACs
489 * from the indirect block. We decompress it now and
490 * throw away the result after we are finished.
492 tmp = zio_buf_alloc(lsize);
493 ret = zio_decompress_data(BP_GET_COMPRESS(bp),
494 zio->io_abd, tmp, zio->io_size, lsize,
495 &zio->io_prop.zp_complevel);
497 ret = SET_ERROR(EIO);
500 ret = zio_crypt_do_indirect_mac_checksum(B_FALSE,
501 tmp, lsize, BP_SHOULD_BYTESWAP(bp), mac);
502 zio_buf_free(tmp, lsize);
504 ret = zio_crypt_do_indirect_mac_checksum_abd(B_FALSE,
505 zio->io_abd, size, BP_SHOULD_BYTESWAP(bp), mac);
507 abd_copy(data, zio->io_abd, size);
509 if (zio_injection_enabled && ot != DMU_OT_DNODE && ret == 0) {
510 ret = zio_handle_decrypt_injection(spa,
511 &zio->io_bookmark, ot, ECKSUM);
520 * If this is an authenticated block, just check the MAC. It would be
521 * nice to separate this out into its own flag, but for the moment
522 * enum zio_flag is out of bits.
524 if (BP_IS_AUTHENTICATED(bp)) {
525 if (ot == DMU_OT_OBJSET) {
526 ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa,
527 dsobj, zio->io_abd, size, BP_SHOULD_BYTESWAP(bp));
529 zio_crypt_decode_mac_bp(bp, mac);
530 ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj,
531 zio->io_abd, size, mac);
532 if (zio_injection_enabled && ret == 0) {
533 ret = zio_handle_decrypt_injection(spa,
534 &zio->io_bookmark, ot, ECKSUM);
537 abd_copy(data, zio->io_abd, size);
545 zio_crypt_decode_params_bp(bp, salt, iv);
547 if (ot == DMU_OT_INTENT_LOG) {
548 tmp = abd_borrow_buf_copy(zio->io_abd, sizeof (zil_chain_t));
549 zio_crypt_decode_mac_zil(tmp, mac);
550 abd_return_buf(zio->io_abd, tmp, sizeof (zil_chain_t));
552 zio_crypt_decode_mac_bp(bp, mac);
555 ret = spa_do_crypt_abd(B_FALSE, spa, &zio->io_bookmark, BP_GET_TYPE(bp),
556 BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp), salt, iv, mac, size, data,
557 zio->io_abd, &no_crypt);
559 abd_copy(data, zio->io_abd, size);
567 /* assert that the key was found unless this was speculative */
568 ASSERT(ret != EACCES || (zio->io_flags & ZIO_FLAG_SPECULATIVE));
571 * If there was a decryption / authentication error return EIO as
572 * the io_error. If this was not a speculative zio, create an ereport.
575 zio->io_error = SET_ERROR(EIO);
576 if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) {
577 spa_log_error(spa, &zio->io_bookmark);
578 (void) zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION,
579 spa, NULL, &zio->io_bookmark, zio, 0);
587 * ==========================================================================
588 * I/O parent/child relationships and pipeline interlocks
589 * ==========================================================================
592 zio_walk_parents(zio_t *cio, zio_link_t **zl)
594 list_t *pl = &cio->io_parent_list;
596 *zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl);
600 ASSERT((*zl)->zl_child == cio);
601 return ((*zl)->zl_parent);
605 zio_walk_children(zio_t *pio, zio_link_t **zl)
607 list_t *cl = &pio->io_child_list;
609 ASSERT(MUTEX_HELD(&pio->io_lock));
611 *zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl);
615 ASSERT((*zl)->zl_parent == pio);
616 return ((*zl)->zl_child);
620 zio_unique_parent(zio_t *cio)
622 zio_link_t *zl = NULL;
623 zio_t *pio = zio_walk_parents(cio, &zl);
625 VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL);
630 zio_add_child(zio_t *pio, zio_t *cio)
632 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
635 * Logical I/Os can have logical, gang, or vdev children.
636 * Gang I/Os can have gang or vdev children.
637 * Vdev I/Os can only have vdev children.
638 * The following ASSERT captures all of these constraints.
640 ASSERT3S(cio->io_child_type, <=, pio->io_child_type);
645 mutex_enter(&pio->io_lock);
646 mutex_enter(&cio->io_lock);
648 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
650 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
651 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
653 list_insert_head(&pio->io_child_list, zl);
654 list_insert_head(&cio->io_parent_list, zl);
656 pio->io_child_count++;
657 cio->io_parent_count++;
659 mutex_exit(&cio->io_lock);
660 mutex_exit(&pio->io_lock);
664 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
666 ASSERT(zl->zl_parent == pio);
667 ASSERT(zl->zl_child == cio);
669 mutex_enter(&pio->io_lock);
670 mutex_enter(&cio->io_lock);
672 list_remove(&pio->io_child_list, zl);
673 list_remove(&cio->io_parent_list, zl);
675 pio->io_child_count--;
676 cio->io_parent_count--;
678 mutex_exit(&cio->io_lock);
679 mutex_exit(&pio->io_lock);
680 kmem_cache_free(zio_link_cache, zl);
684 zio_wait_for_children(zio_t *zio, uint8_t childbits, enum zio_wait_type wait)
686 boolean_t waiting = B_FALSE;
688 mutex_enter(&zio->io_lock);
689 ASSERT(zio->io_stall == NULL);
690 for (int c = 0; c < ZIO_CHILD_TYPES; c++) {
691 if (!(ZIO_CHILD_BIT_IS_SET(childbits, c)))
694 uint64_t *countp = &zio->io_children[c][wait];
697 ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN);
698 zio->io_stall = countp;
703 mutex_exit(&zio->io_lock);
707 __attribute__((always_inline))
709 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait,
710 zio_t **next_to_executep)
712 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
713 int *errorp = &pio->io_child_error[zio->io_child_type];
715 mutex_enter(&pio->io_lock);
716 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
717 *errorp = zio_worst_error(*errorp, zio->io_error);
718 pio->io_reexecute |= zio->io_reexecute;
719 ASSERT3U(*countp, >, 0);
723 if (*countp == 0 && pio->io_stall == countp) {
724 zio_taskq_type_t type =
725 pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE :
727 pio->io_stall = NULL;
728 mutex_exit(&pio->io_lock);
731 * If we can tell the caller to execute this parent next, do
732 * so. Otherwise dispatch the parent zio as its own task.
734 * Having the caller execute the parent when possible reduces
735 * locking on the zio taskq's, reduces context switch
736 * overhead, and has no recursion penalty. Note that one
737 * read from disk typically causes at least 3 zio's: a
738 * zio_null(), the logical zio_read(), and then a physical
739 * zio. When the physical ZIO completes, we are able to call
740 * zio_done() on all 3 of these zio's from one invocation of
741 * zio_execute() by returning the parent back to
742 * zio_execute(). Since the parent isn't executed until this
743 * thread returns back to zio_execute(), the caller should do
746 * In other cases, dispatching the parent prevents
747 * overflowing the stack when we have deeply nested
748 * parent-child relationships, as we do with the "mega zio"
749 * of writes for spa_sync(), and the chain of ZIL blocks.
751 if (next_to_executep != NULL && *next_to_executep == NULL) {
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 enum zio_flag 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 bzero(zio, 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 if (zio->io_metaslab_class == NULL)
885 zio->io_metaslab_class = pio->io_metaslab_class;
886 if (zio->io_logical == NULL)
887 zio->io_logical = pio->io_logical;
888 if (zio->io_child_type == ZIO_CHILD_GANG)
889 zio->io_gang_leader = pio->io_gang_leader;
890 zio_add_child(pio, zio);
893 taskq_init_ent(&zio->io_tqent);
899 zio_destroy(zio_t *zio)
901 metaslab_trace_fini(&zio->io_alloc_list);
902 list_destroy(&zio->io_parent_list);
903 list_destroy(&zio->io_child_list);
904 mutex_destroy(&zio->io_lock);
905 cv_destroy(&zio->io_cv);
906 kmem_cache_free(zio_cache, zio);
910 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
911 void *private, enum zio_flag flags)
915 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
916 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
917 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
923 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
925 return (zio_null(NULL, spa, NULL, done, private, flags));
929 zfs_blkptr_verify_log(spa_t *spa, const blkptr_t *bp,
930 enum blk_verify_flag blk_verify, const char *fmt, ...)
936 (void) vsnprintf(buf, sizeof (buf), fmt, adx);
939 switch (blk_verify) {
940 case BLK_VERIFY_HALT:
941 dprintf_bp(bp, "blkptr at %p dprintf_bp():", bp);
942 zfs_panic_recover("%s: %s", spa_name(spa), buf);
945 zfs_dbgmsg("%s: %s", spa_name(spa), buf);
947 case BLK_VERIFY_ONLY:
955 * Verify the block pointer fields contain reasonable values. This means
956 * it only contains known object types, checksum/compression identifiers,
957 * block sizes within the maximum allowed limits, valid DVAs, etc.
959 * If everything checks out B_TRUE is returned. The zfs_blkptr_verify
960 * argument controls the behavior when an invalid field is detected.
962 * Modes for zfs_blkptr_verify:
963 * 1) BLK_VERIFY_ONLY (evaluate the block)
964 * 2) BLK_VERIFY_LOG (evaluate the block and log problems)
965 * 3) BLK_VERIFY_HALT (call zfs_panic_recover on error)
968 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp, boolean_t config_held,
969 enum blk_verify_flag blk_verify)
973 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
974 errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
975 "blkptr at %p has invalid TYPE %llu",
976 bp, (longlong_t)BP_GET_TYPE(bp));
978 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
979 BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
980 errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
981 "blkptr at %p has invalid CHECKSUM %llu",
982 bp, (longlong_t)BP_GET_CHECKSUM(bp));
984 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
985 BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
986 errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
987 "blkptr at %p has invalid COMPRESS %llu",
988 bp, (longlong_t)BP_GET_COMPRESS(bp));
990 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
991 errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
992 "blkptr at %p has invalid LSIZE %llu",
993 bp, (longlong_t)BP_GET_LSIZE(bp));
995 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
996 errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
997 "blkptr at %p has invalid PSIZE %llu",
998 bp, (longlong_t)BP_GET_PSIZE(bp));
1001 if (BP_IS_EMBEDDED(bp)) {
1002 if (BPE_GET_ETYPE(bp) >= NUM_BP_EMBEDDED_TYPES) {
1003 errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1004 "blkptr at %p has invalid ETYPE %llu",
1005 bp, (longlong_t)BPE_GET_ETYPE(bp));
1010 * Do not verify individual DVAs if the config is not trusted. This
1011 * will be done once the zio is executed in vdev_mirror_map_alloc.
1013 if (!spa->spa_trust_config)
1017 spa_config_enter(spa, SCL_VDEV, bp, RW_READER);
1019 ASSERT(spa_config_held(spa, SCL_VDEV, RW_WRITER));
1021 * Pool-specific checks.
1023 * Note: it would be nice to verify that the blk_birth and
1024 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
1025 * allows the birth time of log blocks (and dmu_sync()-ed blocks
1026 * that are in the log) to be arbitrarily large.
1028 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
1029 uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
1031 if (vdevid >= spa->spa_root_vdev->vdev_children) {
1032 errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1033 "blkptr at %p DVA %u has invalid VDEV %llu",
1034 bp, i, (longlong_t)vdevid);
1037 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
1039 errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1040 "blkptr at %p DVA %u has invalid VDEV %llu",
1041 bp, i, (longlong_t)vdevid);
1044 if (vd->vdev_ops == &vdev_hole_ops) {
1045 errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1046 "blkptr at %p DVA %u has hole VDEV %llu",
1047 bp, i, (longlong_t)vdevid);
1050 if (vd->vdev_ops == &vdev_missing_ops) {
1052 * "missing" vdevs are valid during import, but we
1053 * don't have their detailed info (e.g. asize), so
1054 * we can't perform any more checks on them.
1058 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
1059 uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
1061 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
1062 if (offset + asize > vd->vdev_asize) {
1063 errors += zfs_blkptr_verify_log(spa, bp, blk_verify,
1064 "blkptr at %p DVA %u has invalid OFFSET %llu",
1065 bp, i, (longlong_t)offset);
1069 dprintf_bp(bp, "blkptr at %p dprintf_bp():", bp);
1071 spa_config_exit(spa, SCL_VDEV, bp);
1073 return (errors == 0);
1077 zfs_dva_valid(spa_t *spa, const dva_t *dva, const blkptr_t *bp)
1079 uint64_t vdevid = DVA_GET_VDEV(dva);
1081 if (vdevid >= spa->spa_root_vdev->vdev_children)
1084 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
1088 if (vd->vdev_ops == &vdev_hole_ops)
1091 if (vd->vdev_ops == &vdev_missing_ops) {
1095 uint64_t offset = DVA_GET_OFFSET(dva);
1096 uint64_t asize = DVA_GET_ASIZE(dva);
1099 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
1100 if (offset + asize > vd->vdev_asize)
1107 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
1108 abd_t *data, uint64_t size, zio_done_func_t *done, void *private,
1109 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
1113 (void) zfs_blkptr_verify(spa, bp, flags & ZIO_FLAG_CONFIG_WRITER,
1116 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
1117 data, size, size, done, private,
1118 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
1119 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
1120 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
1126 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
1127 abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp,
1128 zio_done_func_t *ready, zio_done_func_t *children_ready,
1129 zio_done_func_t *physdone, zio_done_func_t *done,
1130 void *private, zio_priority_t priority, enum zio_flag flags,
1131 const zbookmark_phys_t *zb)
1135 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
1136 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
1137 zp->zp_compress >= ZIO_COMPRESS_OFF &&
1138 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
1139 DMU_OT_IS_VALID(zp->zp_type) &&
1140 zp->zp_level < 32 &&
1141 zp->zp_copies > 0 &&
1142 zp->zp_copies <= spa_max_replication(spa));
1144 zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private,
1145 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
1146 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
1147 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
1149 zio->io_ready = ready;
1150 zio->io_children_ready = children_ready;
1151 zio->io_physdone = physdone;
1155 * Data can be NULL if we are going to call zio_write_override() to
1156 * provide the already-allocated BP. But we may need the data to
1157 * verify a dedup hit (if requested). In this case, don't try to
1158 * dedup (just take the already-allocated BP verbatim). Encrypted
1159 * dedup blocks need data as well so we also disable dedup in this
1163 (zio->io_prop.zp_dedup_verify || zio->io_prop.zp_encrypt)) {
1164 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
1171 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data,
1172 uint64_t size, zio_done_func_t *done, void *private,
1173 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
1177 zio = zio_create(pio, spa, txg, bp, data, size, size, done, private,
1178 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb,
1179 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
1185 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
1187 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
1188 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1189 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1190 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
1193 * We must reset the io_prop to match the values that existed
1194 * when the bp was first written by dmu_sync() keeping in mind
1195 * that nopwrite and dedup are mutually exclusive.
1197 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
1198 zio->io_prop.zp_nopwrite = nopwrite;
1199 zio->io_prop.zp_copies = copies;
1200 zio->io_bp_override = bp;
1204 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
1207 (void) zfs_blkptr_verify(spa, bp, B_FALSE, BLK_VERIFY_HALT);
1210 * The check for EMBEDDED is a performance optimization. We
1211 * process the free here (by ignoring it) rather than
1212 * putting it on the list and then processing it in zio_free_sync().
1214 if (BP_IS_EMBEDDED(bp))
1216 metaslab_check_free(spa, bp);
1219 * Frees that are for the currently-syncing txg, are not going to be
1220 * deferred, and which will not need to do a read (i.e. not GANG or
1221 * DEDUP), can be processed immediately. Otherwise, put them on the
1222 * in-memory list for later processing.
1224 * Note that we only defer frees after zfs_sync_pass_deferred_free
1225 * when the log space map feature is disabled. [see relevant comment
1226 * in spa_sync_iterate_to_convergence()]
1228 if (BP_IS_GANG(bp) ||
1230 txg != spa->spa_syncing_txg ||
1231 (spa_sync_pass(spa) >= zfs_sync_pass_deferred_free &&
1232 !spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP))) {
1233 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
1235 VERIFY3P(zio_free_sync(NULL, spa, txg, bp, 0), ==, NULL);
1240 * To improve performance, this function may return NULL if we were able
1241 * to do the free immediately. This avoids the cost of creating a zio
1242 * (and linking it to the parent, etc).
1245 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1246 enum zio_flag flags)
1248 ASSERT(!BP_IS_HOLE(bp));
1249 ASSERT(spa_syncing_txg(spa) == txg);
1251 if (BP_IS_EMBEDDED(bp))
1254 metaslab_check_free(spa, bp);
1256 dsl_scan_freed(spa, bp);
1258 if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp)) {
1260 * GANG and DEDUP blocks can induce a read (for the gang block
1261 * header, or the DDT), so issue them asynchronously so that
1262 * this thread is not tied up.
1264 enum zio_stage stage =
1265 ZIO_FREE_PIPELINE | ZIO_STAGE_ISSUE_ASYNC;
1267 return (zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1268 BP_GET_PSIZE(bp), NULL, NULL,
1269 ZIO_TYPE_FREE, ZIO_PRIORITY_NOW,
1270 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage));
1272 metaslab_free(spa, bp, txg, B_FALSE);
1278 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1279 zio_done_func_t *done, void *private, enum zio_flag flags)
1283 (void) zfs_blkptr_verify(spa, bp, flags & ZIO_FLAG_CONFIG_WRITER,
1286 if (BP_IS_EMBEDDED(bp))
1287 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
1290 * A claim is an allocation of a specific block. Claims are needed
1291 * to support immediate writes in the intent log. The issue is that
1292 * immediate writes contain committed data, but in a txg that was
1293 * *not* committed. Upon opening the pool after an unclean shutdown,
1294 * the intent log claims all blocks that contain immediate write data
1295 * so that the SPA knows they're in use.
1297 * All claims *must* be resolved in the first txg -- before the SPA
1298 * starts allocating blocks -- so that nothing is allocated twice.
1299 * If txg == 0 we just verify that the block is claimable.
1301 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <,
1302 spa_min_claim_txg(spa));
1303 ASSERT(txg == spa_min_claim_txg(spa) || txg == 0);
1304 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
1306 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1307 BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW,
1308 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
1309 ASSERT0(zio->io_queued_timestamp);
1315 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
1316 zio_done_func_t *done, void *private, enum zio_flag flags)
1321 if (vd->vdev_children == 0) {
1322 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
1323 ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
1324 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
1328 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
1330 for (c = 0; c < vd->vdev_children; c++)
1331 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
1332 done, private, flags));
1339 zio_trim(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1340 zio_done_func_t *done, void *private, zio_priority_t priority,
1341 enum zio_flag flags, enum trim_flag trim_flags)
1345 ASSERT0(vd->vdev_children);
1346 ASSERT0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
1347 ASSERT0(P2PHASE(size, 1ULL << vd->vdev_ashift));
1348 ASSERT3U(size, !=, 0);
1350 zio = zio_create(pio, vd->vdev_spa, 0, NULL, NULL, size, size, done,
1351 private, ZIO_TYPE_TRIM, priority, flags | ZIO_FLAG_PHYSICAL,
1352 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_TRIM_PIPELINE);
1353 zio->io_trim_flags = trim_flags;
1359 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1360 abd_t *data, int checksum, zio_done_func_t *done, void *private,
1361 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1365 ASSERT(vd->vdev_children == 0);
1366 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1367 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1368 ASSERT3U(offset + size, <=, vd->vdev_psize);
1370 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1371 private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1372 offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
1374 zio->io_prop.zp_checksum = checksum;
1380 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1381 abd_t *data, int checksum, zio_done_func_t *done, void *private,
1382 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1386 ASSERT(vd->vdev_children == 0);
1387 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1388 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1389 ASSERT3U(offset + size, <=, vd->vdev_psize);
1391 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1392 private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1393 offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
1395 zio->io_prop.zp_checksum = checksum;
1397 if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
1399 * zec checksums are necessarily destructive -- they modify
1400 * the end of the write buffer to hold the verifier/checksum.
1401 * Therefore, we must make a local copy in case the data is
1402 * being written to multiple places in parallel.
1404 abd_t *wbuf = abd_alloc_sametype(data, size);
1405 abd_copy(wbuf, data, size);
1407 zio_push_transform(zio, wbuf, size, size, NULL);
1414 * Create a child I/O to do some work for us.
1417 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1418 abd_t *data, uint64_t size, int type, zio_priority_t priority,
1419 enum zio_flag flags, zio_done_func_t *done, void *private)
1421 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1425 * vdev child I/Os do not propagate their error to the parent.
1426 * Therefore, for correct operation the caller *must* check for
1427 * and handle the error in the child i/o's done callback.
1428 * The only exceptions are i/os that we don't care about
1429 * (OPTIONAL or REPAIR).
1431 ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) ||
1434 if (type == ZIO_TYPE_READ && bp != NULL) {
1436 * If we have the bp, then the child should perform the
1437 * checksum and the parent need not. This pushes error
1438 * detection as close to the leaves as possible and
1439 * eliminates redundant checksums in the interior nodes.
1441 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1442 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1445 if (vd->vdev_ops->vdev_op_leaf) {
1446 ASSERT0(vd->vdev_children);
1447 offset += VDEV_LABEL_START_SIZE;
1450 flags |= ZIO_VDEV_CHILD_FLAGS(pio);
1453 * If we've decided to do a repair, the write is not speculative --
1454 * even if the original read was.
1456 if (flags & ZIO_FLAG_IO_REPAIR)
1457 flags &= ~ZIO_FLAG_SPECULATIVE;
1460 * If we're creating a child I/O that is not associated with a
1461 * top-level vdev, then the child zio is not an allocating I/O.
1462 * If this is a retried I/O then we ignore it since we will
1463 * have already processed the original allocating I/O.
1465 if (flags & ZIO_FLAG_IO_ALLOCATING &&
1466 (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) {
1467 ASSERT(pio->io_metaslab_class != NULL);
1468 ASSERT(pio->io_metaslab_class->mc_alloc_throttle_enabled);
1469 ASSERT(type == ZIO_TYPE_WRITE);
1470 ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE);
1471 ASSERT(!(flags & ZIO_FLAG_IO_REPAIR));
1472 ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) ||
1473 pio->io_child_type == ZIO_CHILD_GANG);
1475 flags &= ~ZIO_FLAG_IO_ALLOCATING;
1479 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size,
1480 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1481 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1482 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
1484 zio->io_physdone = pio->io_physdone;
1485 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1486 zio->io_logical->io_phys_children++;
1492 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size,
1493 zio_type_t type, zio_priority_t priority, enum zio_flag flags,
1494 zio_done_func_t *done, void *private)
1498 ASSERT(vd->vdev_ops->vdev_op_leaf);
1500 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1501 data, size, size, done, private, type, priority,
1502 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1504 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1510 zio_flush(zio_t *zio, vdev_t *vd)
1512 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
1514 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1518 zio_shrink(zio_t *zio, uint64_t size)
1520 ASSERT3P(zio->io_executor, ==, NULL);
1521 ASSERT3U(zio->io_orig_size, ==, zio->io_size);
1522 ASSERT3U(size, <=, zio->io_size);
1525 * We don't shrink for raidz because of problems with the
1526 * reconstruction when reading back less than the block size.
1527 * Note, BP_IS_RAIDZ() assumes no compression.
1529 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1530 if (!BP_IS_RAIDZ(zio->io_bp)) {
1531 /* we are not doing a raw write */
1532 ASSERT3U(zio->io_size, ==, zio->io_lsize);
1533 zio->io_orig_size = zio->io_size = zio->io_lsize = size;
1538 * ==========================================================================
1539 * Prepare to read and write logical blocks
1540 * ==========================================================================
1544 zio_read_bp_init(zio_t *zio)
1546 blkptr_t *bp = zio->io_bp;
1548 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1550 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1552 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1553 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1554 !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) {
1555 zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1556 psize, psize, zio_decompress);
1559 if (((BP_IS_PROTECTED(bp) && !(zio->io_flags & ZIO_FLAG_RAW_ENCRYPT)) ||
1560 BP_HAS_INDIRECT_MAC_CKSUM(bp)) &&
1561 zio->io_child_type == ZIO_CHILD_LOGICAL) {
1562 zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1563 psize, psize, zio_decrypt);
1566 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1567 int psize = BPE_GET_PSIZE(bp);
1568 void *data = abd_borrow_buf(zio->io_abd, psize);
1570 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1571 decode_embedded_bp_compressed(bp, data);
1572 abd_return_buf_copy(zio->io_abd, data, psize);
1574 ASSERT(!BP_IS_EMBEDDED(bp));
1575 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1578 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1579 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1581 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1582 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1584 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1585 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1591 zio_write_bp_init(zio_t *zio)
1593 if (!IO_IS_ALLOCATING(zio))
1596 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1598 if (zio->io_bp_override) {
1599 blkptr_t *bp = zio->io_bp;
1600 zio_prop_t *zp = &zio->io_prop;
1602 ASSERT(bp->blk_birth != zio->io_txg);
1603 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1605 *bp = *zio->io_bp_override;
1606 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1608 if (BP_IS_EMBEDDED(bp))
1612 * If we've been overridden and nopwrite is set then
1613 * set the flag accordingly to indicate that a nopwrite
1614 * has already occurred.
1616 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1617 ASSERT(!zp->zp_dedup);
1618 ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum);
1619 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1623 ASSERT(!zp->zp_nopwrite);
1625 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1628 ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags &
1629 ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify);
1631 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum &&
1633 BP_SET_DEDUP(bp, 1);
1634 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1639 * We were unable to handle this as an override bp, treat
1640 * it as a regular write I/O.
1642 zio->io_bp_override = NULL;
1643 *bp = zio->io_bp_orig;
1644 zio->io_pipeline = zio->io_orig_pipeline;
1651 zio_write_compress(zio_t *zio)
1653 spa_t *spa = zio->io_spa;
1654 zio_prop_t *zp = &zio->io_prop;
1655 enum zio_compress compress = zp->zp_compress;
1656 blkptr_t *bp = zio->io_bp;
1657 uint64_t lsize = zio->io_lsize;
1658 uint64_t psize = zio->io_size;
1662 * If our children haven't all reached the ready stage,
1663 * wait for them and then repeat this pipeline stage.
1665 if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT |
1666 ZIO_CHILD_GANG_BIT, ZIO_WAIT_READY)) {
1670 if (!IO_IS_ALLOCATING(zio))
1673 if (zio->io_children_ready != NULL) {
1675 * Now that all our children are ready, run the callback
1676 * associated with this zio in case it wants to modify the
1677 * data to be written.
1679 ASSERT3U(zp->zp_level, >, 0);
1680 zio->io_children_ready(zio);
1683 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1684 ASSERT(zio->io_bp_override == NULL);
1686 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1688 * We're rewriting an existing block, which means we're
1689 * working on behalf of spa_sync(). For spa_sync() to
1690 * converge, it must eventually be the case that we don't
1691 * have to allocate new blocks. But compression changes
1692 * the blocksize, which forces a reallocate, and makes
1693 * convergence take longer. Therefore, after the first
1694 * few passes, stop compressing to ensure convergence.
1696 pass = spa_sync_pass(spa);
1698 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1699 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1700 ASSERT(!BP_GET_DEDUP(bp));
1702 if (pass >= zfs_sync_pass_dont_compress)
1703 compress = ZIO_COMPRESS_OFF;
1705 /* Make sure someone doesn't change their mind on overwrites */
1706 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1707 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1710 /* If it's a compressed write that is not raw, compress the buffer. */
1711 if (compress != ZIO_COMPRESS_OFF &&
1712 !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) {
1713 void *cbuf = zio_buf_alloc(lsize);
1714 psize = zio_compress_data(compress, zio->io_abd, cbuf, lsize,
1716 if (psize == 0 || psize >= lsize) {
1717 compress = ZIO_COMPRESS_OFF;
1718 zio_buf_free(cbuf, lsize);
1719 } else if (!zp->zp_dedup && !zp->zp_encrypt &&
1720 psize <= BPE_PAYLOAD_SIZE &&
1721 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1722 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1723 encode_embedded_bp_compressed(bp,
1724 cbuf, compress, lsize, psize);
1725 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1726 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1727 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1728 zio_buf_free(cbuf, lsize);
1729 bp->blk_birth = zio->io_txg;
1730 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1731 ASSERT(spa_feature_is_active(spa,
1732 SPA_FEATURE_EMBEDDED_DATA));
1736 * Round up compressed size up to the ashift
1737 * of the smallest-ashift device, and zero the tail.
1738 * This ensures that the compressed size of the BP
1739 * (and thus compressratio property) are correct,
1740 * in that we charge for the padding used to fill out
1743 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
1744 size_t rounded = (size_t)P2ROUNDUP(psize,
1745 1ULL << spa->spa_min_ashift);
1746 if (rounded >= lsize) {
1747 compress = ZIO_COMPRESS_OFF;
1748 zio_buf_free(cbuf, lsize);
1751 abd_t *cdata = abd_get_from_buf(cbuf, lsize);
1752 abd_take_ownership_of_buf(cdata, B_TRUE);
1753 abd_zero_off(cdata, psize, rounded - psize);
1755 zio_push_transform(zio, cdata,
1756 psize, lsize, NULL);
1761 * We were unable to handle this as an override bp, treat
1762 * it as a regular write I/O.
1764 zio->io_bp_override = NULL;
1765 *bp = zio->io_bp_orig;
1766 zio->io_pipeline = zio->io_orig_pipeline;
1768 } else if ((zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) != 0 &&
1769 zp->zp_type == DMU_OT_DNODE) {
1771 * The DMU actually relies on the zio layer's compression
1772 * to free metadnode blocks that have had all contained
1773 * dnodes freed. As a result, even when doing a raw
1774 * receive, we must check whether the block can be compressed
1777 psize = zio_compress_data(ZIO_COMPRESS_EMPTY,
1778 zio->io_abd, NULL, lsize, zp->zp_complevel);
1779 if (psize == 0 || psize >= lsize)
1780 compress = ZIO_COMPRESS_OFF;
1782 ASSERT3U(psize, !=, 0);
1786 * The final pass of spa_sync() must be all rewrites, but the first
1787 * few passes offer a trade-off: allocating blocks defers convergence,
1788 * but newly allocated blocks are sequential, so they can be written
1789 * to disk faster. Therefore, we allow the first few passes of
1790 * spa_sync() to allocate new blocks, but force rewrites after that.
1791 * There should only be a handful of blocks after pass 1 in any case.
1793 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1794 BP_GET_PSIZE(bp) == psize &&
1795 pass >= zfs_sync_pass_rewrite) {
1796 VERIFY3U(psize, !=, 0);
1797 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1799 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1800 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1803 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1807 if (zio->io_bp_orig.blk_birth != 0 &&
1808 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1809 BP_SET_LSIZE(bp, lsize);
1810 BP_SET_TYPE(bp, zp->zp_type);
1811 BP_SET_LEVEL(bp, zp->zp_level);
1812 BP_SET_BIRTH(bp, zio->io_txg, 0);
1814 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1816 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1817 BP_SET_LSIZE(bp, lsize);
1818 BP_SET_TYPE(bp, zp->zp_type);
1819 BP_SET_LEVEL(bp, zp->zp_level);
1820 BP_SET_PSIZE(bp, psize);
1821 BP_SET_COMPRESS(bp, compress);
1822 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1823 BP_SET_DEDUP(bp, zp->zp_dedup);
1824 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1826 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1827 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1828 ASSERT(!zp->zp_encrypt ||
1829 DMU_OT_IS_ENCRYPTED(zp->zp_type));
1830 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1832 if (zp->zp_nopwrite) {
1833 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1834 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1835 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1842 zio_free_bp_init(zio_t *zio)
1844 blkptr_t *bp = zio->io_bp;
1846 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1847 if (BP_GET_DEDUP(bp))
1848 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1851 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1857 * ==========================================================================
1858 * Execute the I/O pipeline
1859 * ==========================================================================
1863 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1865 spa_t *spa = zio->io_spa;
1866 zio_type_t t = zio->io_type;
1867 int flags = (cutinline ? TQ_FRONT : 0);
1870 * If we're a config writer or a probe, the normal issue and
1871 * interrupt threads may all be blocked waiting for the config lock.
1872 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1874 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1878 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1880 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1884 * If this is a high priority I/O, then use the high priority taskq if
1887 if ((zio->io_priority == ZIO_PRIORITY_NOW ||
1888 zio->io_priority == ZIO_PRIORITY_SYNC_WRITE) &&
1889 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1892 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1895 * NB: We are assuming that the zio can only be dispatched
1896 * to a single taskq at a time. It would be a grievous error
1897 * to dispatch the zio to another taskq at the same time.
1899 ASSERT(taskq_empty_ent(&zio->io_tqent));
1900 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1901 flags, &zio->io_tqent);
1905 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1907 spa_t *spa = zio->io_spa;
1909 taskq_t *tq = taskq_of_curthread();
1911 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1912 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1914 for (i = 0; i < tqs->stqs_count; i++) {
1915 if (tqs->stqs_taskq[i] == tq)
1924 zio_issue_async(zio_t *zio)
1926 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1932 zio_interrupt(zio_t *zio)
1934 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1938 zio_delay_interrupt(zio_t *zio)
1941 * The timeout_generic() function isn't defined in userspace, so
1942 * rather than trying to implement the function, the zio delay
1943 * functionality has been disabled for userspace builds.
1948 * If io_target_timestamp is zero, then no delay has been registered
1949 * for this IO, thus jump to the end of this function and "skip" the
1950 * delay; issuing it directly to the zio layer.
1952 if (zio->io_target_timestamp != 0) {
1953 hrtime_t now = gethrtime();
1955 if (now >= zio->io_target_timestamp) {
1957 * This IO has already taken longer than the target
1958 * delay to complete, so we don't want to delay it
1959 * any longer; we "miss" the delay and issue it
1960 * directly to the zio layer. This is likely due to
1961 * the target latency being set to a value less than
1962 * the underlying hardware can satisfy (e.g. delay
1963 * set to 1ms, but the disks take 10ms to complete an
1967 DTRACE_PROBE2(zio__delay__miss, zio_t *, zio,
1973 hrtime_t diff = zio->io_target_timestamp - now;
1974 clock_t expire_at_tick = ddi_get_lbolt() +
1977 DTRACE_PROBE3(zio__delay__hit, zio_t *, zio,
1978 hrtime_t, now, hrtime_t, diff);
1980 if (NSEC_TO_TICK(diff) == 0) {
1981 /* Our delay is less than a jiffy - just spin */
1982 zfs_sleep_until(zio->io_target_timestamp);
1986 * Use taskq_dispatch_delay() in the place of
1987 * OpenZFS's timeout_generic().
1989 tid = taskq_dispatch_delay(system_taskq,
1990 (task_func_t *)zio_interrupt,
1991 zio, TQ_NOSLEEP, expire_at_tick);
1992 if (tid == TASKQID_INVALID) {
1994 * Couldn't allocate a task. Just
1995 * finish the zio without a delay.
2004 DTRACE_PROBE1(zio__delay__skip, zio_t *, zio);
2009 zio_deadman_impl(zio_t *pio, int ziodepth)
2011 zio_t *cio, *cio_next;
2012 zio_link_t *zl = NULL;
2013 vdev_t *vd = pio->io_vd;
2015 if (zio_deadman_log_all || (vd != NULL && vd->vdev_ops->vdev_op_leaf)) {
2016 vdev_queue_t *vq = vd ? &vd->vdev_queue : NULL;
2017 zbookmark_phys_t *zb = &pio->io_bookmark;
2018 uint64_t delta = gethrtime() - pio->io_timestamp;
2019 uint64_t failmode = spa_get_deadman_failmode(pio->io_spa);
2021 zfs_dbgmsg("slow zio[%d]: zio=%px timestamp=%llu "
2022 "delta=%llu queued=%llu io=%llu "
2023 "path=%s last=%llu "
2024 "type=%d priority=%d flags=0x%x "
2025 "stage=0x%x pipeline=0x%x pipeline-trace=0x%x "
2026 "objset=%llu object=%llu level=%llu blkid=%llu "
2027 "offset=%llu size=%llu error=%d",
2028 ziodepth, pio, pio->io_timestamp,
2029 delta, pio->io_delta, pio->io_delay,
2030 vd ? vd->vdev_path : "NULL", vq ? vq->vq_io_complete_ts : 0,
2031 pio->io_type, pio->io_priority, pio->io_flags,
2032 pio->io_stage, pio->io_pipeline, pio->io_pipeline_trace,
2033 zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
2034 pio->io_offset, pio->io_size, pio->io_error);
2035 (void) zfs_ereport_post(FM_EREPORT_ZFS_DEADMAN,
2036 pio->io_spa, vd, zb, pio, 0);
2038 if (failmode == ZIO_FAILURE_MODE_CONTINUE &&
2039 taskq_empty_ent(&pio->io_tqent)) {
2044 mutex_enter(&pio->io_lock);
2045 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
2046 cio_next = zio_walk_children(pio, &zl);
2047 zio_deadman_impl(cio, ziodepth + 1);
2049 mutex_exit(&pio->io_lock);
2053 * Log the critical information describing this zio and all of its children
2054 * using the zfs_dbgmsg() interface then post deadman event for the ZED.
2057 zio_deadman(zio_t *pio, char *tag)
2059 spa_t *spa = pio->io_spa;
2060 char *name = spa_name(spa);
2062 if (!zfs_deadman_enabled || spa_suspended(spa))
2065 zio_deadman_impl(pio, 0);
2067 switch (spa_get_deadman_failmode(spa)) {
2068 case ZIO_FAILURE_MODE_WAIT:
2069 zfs_dbgmsg("%s waiting for hung I/O to pool '%s'", tag, name);
2072 case ZIO_FAILURE_MODE_CONTINUE:
2073 zfs_dbgmsg("%s restarting hung I/O for pool '%s'", tag, name);
2076 case ZIO_FAILURE_MODE_PANIC:
2077 fm_panic("%s determined I/O to pool '%s' is hung.", tag, name);
2083 * Execute the I/O pipeline until one of the following occurs:
2084 * (1) the I/O completes; (2) the pipeline stalls waiting for
2085 * dependent child I/Os; (3) the I/O issues, so we're waiting
2086 * for an I/O completion interrupt; (4) the I/O is delegated by
2087 * vdev-level caching or aggregation; (5) the I/O is deferred
2088 * due to vdev-level queueing; (6) the I/O is handed off to
2089 * another thread. In all cases, the pipeline stops whenever
2090 * there's no CPU work; it never burns a thread in cv_wait_io().
2092 * There's no locking on io_stage because there's no legitimate way
2093 * for multiple threads to be attempting to process the same I/O.
2095 static zio_pipe_stage_t *zio_pipeline[];
2098 * zio_execute() is a wrapper around the static function
2099 * __zio_execute() so that we can force __zio_execute() to be
2100 * inlined. This reduces stack overhead which is important
2101 * because __zio_execute() is called recursively in several zio
2102 * code paths. zio_execute() itself cannot be inlined because
2103 * it is externally visible.
2106 zio_execute(zio_t *zio)
2108 fstrans_cookie_t cookie;
2110 cookie = spl_fstrans_mark();
2112 spl_fstrans_unmark(cookie);
2116 * Used to determine if in the current context the stack is sized large
2117 * enough to allow zio_execute() to be called recursively. A minimum
2118 * stack size of 16K is required to avoid needing to re-dispatch the zio.
2121 zio_execute_stack_check(zio_t *zio)
2123 #if !defined(HAVE_LARGE_STACKS)
2124 dsl_pool_t *dp = spa_get_dsl(zio->io_spa);
2126 /* Executing in txg_sync_thread() context. */
2127 if (dp && curthread == dp->dp_tx.tx_sync_thread)
2130 /* Pool initialization outside of zio_taskq context. */
2131 if (dp && spa_is_initializing(dp->dp_spa) &&
2132 !zio_taskq_member(zio, ZIO_TASKQ_ISSUE) &&
2133 !zio_taskq_member(zio, ZIO_TASKQ_ISSUE_HIGH))
2135 #endif /* HAVE_LARGE_STACKS */
2140 __attribute__((always_inline))
2142 __zio_execute(zio_t *zio)
2144 ASSERT3U(zio->io_queued_timestamp, >, 0);
2146 while (zio->io_stage < ZIO_STAGE_DONE) {
2147 enum zio_stage pipeline = zio->io_pipeline;
2148 enum zio_stage stage = zio->io_stage;
2150 zio->io_executor = curthread;
2152 ASSERT(!MUTEX_HELD(&zio->io_lock));
2153 ASSERT(ISP2(stage));
2154 ASSERT(zio->io_stall == NULL);
2158 } while ((stage & pipeline) == 0);
2160 ASSERT(stage <= ZIO_STAGE_DONE);
2163 * If we are in interrupt context and this pipeline stage
2164 * will grab a config lock that is held across I/O,
2165 * or may wait for an I/O that needs an interrupt thread
2166 * to complete, issue async to avoid deadlock.
2168 * For VDEV_IO_START, we cut in line so that the io will
2169 * be sent to disk promptly.
2171 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
2172 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
2173 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
2174 zio_requeue_io_start_cut_in_line : B_FALSE;
2175 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
2180 * If the current context doesn't have large enough stacks
2181 * the zio must be issued asynchronously to prevent overflow.
2183 if (zio_execute_stack_check(zio)) {
2184 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
2185 zio_requeue_io_start_cut_in_line : B_FALSE;
2186 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
2190 zio->io_stage = stage;
2191 zio->io_pipeline_trace |= zio->io_stage;
2194 * The zio pipeline stage returns the next zio to execute
2195 * (typically the same as this one), or NULL if we should
2198 zio = zio_pipeline[highbit64(stage) - 1](zio);
2207 * ==========================================================================
2208 * Initiate I/O, either sync or async
2209 * ==========================================================================
2212 zio_wait(zio_t *zio)
2215 * Some routines, like zio_free_sync(), may return a NULL zio
2216 * to avoid the performance overhead of creating and then destroying
2217 * an unneeded zio. For the callers' simplicity, we accept a NULL
2218 * zio and ignore it.
2223 long timeout = MSEC_TO_TICK(zfs_deadman_ziotime_ms);
2226 ASSERT3S(zio->io_stage, ==, ZIO_STAGE_OPEN);
2227 ASSERT3P(zio->io_executor, ==, NULL);
2229 zio->io_waiter = curthread;
2230 ASSERT0(zio->io_queued_timestamp);
2231 zio->io_queued_timestamp = gethrtime();
2235 mutex_enter(&zio->io_lock);
2236 while (zio->io_executor != NULL) {
2237 error = cv_timedwait_io(&zio->io_cv, &zio->io_lock,
2238 ddi_get_lbolt() + timeout);
2240 if (zfs_deadman_enabled && error == -1 &&
2241 gethrtime() - zio->io_queued_timestamp >
2242 spa_deadman_ziotime(zio->io_spa)) {
2243 mutex_exit(&zio->io_lock);
2244 timeout = MSEC_TO_TICK(zfs_deadman_checktime_ms);
2245 zio_deadman(zio, FTAG);
2246 mutex_enter(&zio->io_lock);
2249 mutex_exit(&zio->io_lock);
2251 error = zio->io_error;
2258 zio_nowait(zio_t *zio)
2261 * See comment in zio_wait().
2266 ASSERT3P(zio->io_executor, ==, NULL);
2268 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
2269 zio_unique_parent(zio) == NULL) {
2273 * This is a logical async I/O with no parent to wait for it.
2274 * We add it to the spa_async_root_zio "Godfather" I/O which
2275 * will ensure they complete prior to unloading the pool.
2277 spa_t *spa = zio->io_spa;
2279 pio = spa->spa_async_zio_root[CPU_SEQID];
2282 zio_add_child(pio, zio);
2285 ASSERT0(zio->io_queued_timestamp);
2286 zio->io_queued_timestamp = gethrtime();
2291 * ==========================================================================
2292 * Reexecute, cancel, or suspend/resume failed I/O
2293 * ==========================================================================
2297 zio_reexecute(zio_t *pio)
2299 zio_t *cio, *cio_next;
2301 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
2302 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
2303 ASSERT(pio->io_gang_leader == NULL);
2304 ASSERT(pio->io_gang_tree == NULL);
2306 pio->io_flags = pio->io_orig_flags;
2307 pio->io_stage = pio->io_orig_stage;
2308 pio->io_pipeline = pio->io_orig_pipeline;
2309 pio->io_reexecute = 0;
2310 pio->io_flags |= ZIO_FLAG_REEXECUTED;
2311 pio->io_pipeline_trace = 0;
2313 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2314 pio->io_state[w] = 0;
2315 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2316 pio->io_child_error[c] = 0;
2318 if (IO_IS_ALLOCATING(pio))
2319 BP_ZERO(pio->io_bp);
2322 * As we reexecute pio's children, new children could be created.
2323 * New children go to the head of pio's io_child_list, however,
2324 * so we will (correctly) not reexecute them. The key is that
2325 * the remainder of pio's io_child_list, from 'cio_next' onward,
2326 * cannot be affected by any side effects of reexecuting 'cio'.
2328 zio_link_t *zl = NULL;
2329 mutex_enter(&pio->io_lock);
2330 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
2331 cio_next = zio_walk_children(pio, &zl);
2332 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2333 pio->io_children[cio->io_child_type][w]++;
2334 mutex_exit(&pio->io_lock);
2336 mutex_enter(&pio->io_lock);
2338 mutex_exit(&pio->io_lock);
2341 * Now that all children have been reexecuted, execute the parent.
2342 * We don't reexecute "The Godfather" I/O here as it's the
2343 * responsibility of the caller to wait on it.
2345 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) {
2346 pio->io_queued_timestamp = gethrtime();
2352 zio_suspend(spa_t *spa, zio_t *zio, zio_suspend_reason_t reason)
2354 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
2355 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
2356 "failure and the failure mode property for this pool "
2357 "is set to panic.", spa_name(spa));
2359 cmn_err(CE_WARN, "Pool '%s' has encountered an uncorrectable I/O "
2360 "failure and has been suspended.\n", spa_name(spa));
2362 (void) zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL,
2365 mutex_enter(&spa->spa_suspend_lock);
2367 if (spa->spa_suspend_zio_root == NULL)
2368 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
2369 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
2370 ZIO_FLAG_GODFATHER);
2372 spa->spa_suspended = reason;
2375 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2376 ASSERT(zio != spa->spa_suspend_zio_root);
2377 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2378 ASSERT(zio_unique_parent(zio) == NULL);
2379 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
2380 zio_add_child(spa->spa_suspend_zio_root, zio);
2383 mutex_exit(&spa->spa_suspend_lock);
2387 zio_resume(spa_t *spa)
2392 * Reexecute all previously suspended i/o.
2394 mutex_enter(&spa->spa_suspend_lock);
2395 spa->spa_suspended = ZIO_SUSPEND_NONE;
2396 cv_broadcast(&spa->spa_suspend_cv);
2397 pio = spa->spa_suspend_zio_root;
2398 spa->spa_suspend_zio_root = NULL;
2399 mutex_exit(&spa->spa_suspend_lock);
2405 return (zio_wait(pio));
2409 zio_resume_wait(spa_t *spa)
2411 mutex_enter(&spa->spa_suspend_lock);
2412 while (spa_suspended(spa))
2413 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
2414 mutex_exit(&spa->spa_suspend_lock);
2418 * ==========================================================================
2421 * A gang block is a collection of small blocks that looks to the DMU
2422 * like one large block. When zio_dva_allocate() cannot find a block
2423 * of the requested size, due to either severe fragmentation or the pool
2424 * being nearly full, it calls zio_write_gang_block() to construct the
2425 * block from smaller fragments.
2427 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
2428 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
2429 * an indirect block: it's an array of block pointers. It consumes
2430 * only one sector and hence is allocatable regardless of fragmentation.
2431 * The gang header's bps point to its gang members, which hold the data.
2433 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
2434 * as the verifier to ensure uniqueness of the SHA256 checksum.
2435 * Critically, the gang block bp's blk_cksum is the checksum of the data,
2436 * not the gang header. This ensures that data block signatures (needed for
2437 * deduplication) are independent of how the block is physically stored.
2439 * Gang blocks can be nested: a gang member may itself be a gang block.
2440 * Thus every gang block is a tree in which root and all interior nodes are
2441 * gang headers, and the leaves are normal blocks that contain user data.
2442 * The root of the gang tree is called the gang leader.
2444 * To perform any operation (read, rewrite, free, claim) on a gang block,
2445 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
2446 * in the io_gang_tree field of the original logical i/o by recursively
2447 * reading the gang leader and all gang headers below it. This yields
2448 * an in-core tree containing the contents of every gang header and the
2449 * bps for every constituent of the gang block.
2451 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
2452 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
2453 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
2454 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
2455 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
2456 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
2457 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
2458 * of the gang header plus zio_checksum_compute() of the data to update the
2459 * gang header's blk_cksum as described above.
2461 * The two-phase assemble/issue model solves the problem of partial failure --
2462 * what if you'd freed part of a gang block but then couldn't read the
2463 * gang header for another part? Assembling the entire gang tree first
2464 * ensures that all the necessary gang header I/O has succeeded before
2465 * starting the actual work of free, claim, or write. Once the gang tree
2466 * is assembled, free and claim are in-memory operations that cannot fail.
2468 * In the event that a gang write fails, zio_dva_unallocate() walks the
2469 * gang tree to immediately free (i.e. insert back into the space map)
2470 * everything we've allocated. This ensures that we don't get ENOSPC
2471 * errors during repeated suspend/resume cycles due to a flaky device.
2473 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
2474 * the gang tree, we won't modify the block, so we can safely defer the free
2475 * (knowing that the block is still intact). If we *can* assemble the gang
2476 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
2477 * each constituent bp and we can allocate a new block on the next sync pass.
2479 * In all cases, the gang tree allows complete recovery from partial failure.
2480 * ==========================================================================
2484 zio_gang_issue_func_done(zio_t *zio)
2486 abd_put(zio->io_abd);
2490 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2496 return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset),
2497 BP_GET_PSIZE(bp), zio_gang_issue_func_done,
2498 NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2499 &pio->io_bookmark));
2503 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2510 abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2511 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2512 gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL,
2513 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2516 * As we rewrite each gang header, the pipeline will compute
2517 * a new gang block header checksum for it; but no one will
2518 * compute a new data checksum, so we do that here. The one
2519 * exception is the gang leader: the pipeline already computed
2520 * its data checksum because that stage precedes gang assembly.
2521 * (Presently, nothing actually uses interior data checksums;
2522 * this is just good hygiene.)
2524 if (gn != pio->io_gang_leader->io_gang_tree) {
2525 abd_t *buf = abd_get_offset(data, offset);
2527 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
2528 buf, BP_GET_PSIZE(bp));
2533 * If we are here to damage data for testing purposes,
2534 * leave the GBH alone so that we can detect the damage.
2536 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
2537 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2539 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2540 abd_get_offset(data, offset), BP_GET_PSIZE(bp),
2541 zio_gang_issue_func_done, NULL, pio->io_priority,
2542 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2550 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2553 zio_t *zio = zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
2554 ZIO_GANG_CHILD_FLAGS(pio));
2556 zio = zio_null(pio, pio->io_spa,
2557 NULL, NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio));
2564 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2567 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
2568 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
2571 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
2580 static void zio_gang_tree_assemble_done(zio_t *zio);
2582 static zio_gang_node_t *
2583 zio_gang_node_alloc(zio_gang_node_t **gnpp)
2585 zio_gang_node_t *gn;
2587 ASSERT(*gnpp == NULL);
2589 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
2590 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
2597 zio_gang_node_free(zio_gang_node_t **gnpp)
2599 zio_gang_node_t *gn = *gnpp;
2601 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2602 ASSERT(gn->gn_child[g] == NULL);
2604 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2605 kmem_free(gn, sizeof (*gn));
2610 zio_gang_tree_free(zio_gang_node_t **gnpp)
2612 zio_gang_node_t *gn = *gnpp;
2617 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2618 zio_gang_tree_free(&gn->gn_child[g]);
2620 zio_gang_node_free(gnpp);
2624 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
2626 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
2627 abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2629 ASSERT(gio->io_gang_leader == gio);
2630 ASSERT(BP_IS_GANG(bp));
2632 zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2633 zio_gang_tree_assemble_done, gn, gio->io_priority,
2634 ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
2638 zio_gang_tree_assemble_done(zio_t *zio)
2640 zio_t *gio = zio->io_gang_leader;
2641 zio_gang_node_t *gn = zio->io_private;
2642 blkptr_t *bp = zio->io_bp;
2644 ASSERT(gio == zio_unique_parent(zio));
2645 ASSERT(zio->io_child_count == 0);
2650 /* this ABD was created from a linear buf in zio_gang_tree_assemble */
2651 if (BP_SHOULD_BYTESWAP(bp))
2652 byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size);
2654 ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh);
2655 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
2656 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2658 abd_put(zio->io_abd);
2660 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2661 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2662 if (!BP_IS_GANG(gbp))
2664 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
2669 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data,
2672 zio_t *gio = pio->io_gang_leader;
2675 ASSERT(BP_IS_GANG(bp) == !!gn);
2676 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
2677 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
2680 * If you're a gang header, your data is in gn->gn_gbh.
2681 * If you're a gang member, your data is in 'data' and gn == NULL.
2683 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset);
2686 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2688 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2689 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2690 if (BP_IS_HOLE(gbp))
2692 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data,
2694 offset += BP_GET_PSIZE(gbp);
2698 if (gn == gio->io_gang_tree)
2699 ASSERT3U(gio->io_size, ==, offset);
2706 zio_gang_assemble(zio_t *zio)
2708 blkptr_t *bp = zio->io_bp;
2710 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
2711 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2713 zio->io_gang_leader = zio;
2715 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
2721 zio_gang_issue(zio_t *zio)
2723 blkptr_t *bp = zio->io_bp;
2725 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT, ZIO_WAIT_DONE)) {
2729 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
2730 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2732 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
2733 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd,
2736 zio_gang_tree_free(&zio->io_gang_tree);
2738 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2744 zio_write_gang_member_ready(zio_t *zio)
2746 zio_t *pio = zio_unique_parent(zio);
2747 dva_t *cdva = zio->io_bp->blk_dva;
2748 dva_t *pdva = pio->io_bp->blk_dva;
2750 zio_t *gio __maybe_unused = zio->io_gang_leader;
2752 if (BP_IS_HOLE(zio->io_bp))
2755 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
2757 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
2758 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
2759 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2760 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
2761 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
2763 mutex_enter(&pio->io_lock);
2764 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
2765 ASSERT(DVA_GET_GANG(&pdva[d]));
2766 asize = DVA_GET_ASIZE(&pdva[d]);
2767 asize += DVA_GET_ASIZE(&cdva[d]);
2768 DVA_SET_ASIZE(&pdva[d], asize);
2770 mutex_exit(&pio->io_lock);
2774 zio_write_gang_done(zio_t *zio)
2777 * The io_abd field will be NULL for a zio with no data. The io_flags
2778 * will initially have the ZIO_FLAG_NODATA bit flag set, but we can't
2779 * check for it here as it is cleared in zio_ready.
2781 if (zio->io_abd != NULL)
2782 abd_put(zio->io_abd);
2786 zio_write_gang_block(zio_t *pio)
2788 spa_t *spa = pio->io_spa;
2789 metaslab_class_t *mc = spa_normal_class(spa);
2790 blkptr_t *bp = pio->io_bp;
2791 zio_t *gio = pio->io_gang_leader;
2793 zio_gang_node_t *gn, **gnpp;
2794 zio_gbh_phys_t *gbh;
2796 uint64_t txg = pio->io_txg;
2797 uint64_t resid = pio->io_size;
2799 int copies = gio->io_prop.zp_copies;
2803 boolean_t has_data = !(pio->io_flags & ZIO_FLAG_NODATA);
2806 * encrypted blocks need DVA[2] free so encrypted gang headers can't
2807 * have a third copy.
2809 gbh_copies = MIN(copies + 1, spa_max_replication(spa));
2810 if (gio->io_prop.zp_encrypt && gbh_copies >= SPA_DVAS_PER_BP)
2811 gbh_copies = SPA_DVAS_PER_BP - 1;
2813 int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER;
2814 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2815 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2818 flags |= METASLAB_ASYNC_ALLOC;
2819 VERIFY(zfs_refcount_held(&mc->mc_alloc_slots[pio->io_allocator],
2823 * The logical zio has already placed a reservation for
2824 * 'copies' allocation slots but gang blocks may require
2825 * additional copies. These additional copies
2826 * (i.e. gbh_copies - copies) are guaranteed to succeed
2827 * since metaslab_class_throttle_reserve() always allows
2828 * additional reservations for gang blocks.
2830 VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies,
2831 pio->io_allocator, pio, flags));
2834 error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE,
2835 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags,
2836 &pio->io_alloc_list, pio, pio->io_allocator);
2838 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2839 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2843 * If we failed to allocate the gang block header then
2844 * we remove any additional allocation reservations that
2845 * we placed here. The original reservation will
2846 * be removed when the logical I/O goes to the ready
2849 metaslab_class_throttle_unreserve(mc,
2850 gbh_copies - copies, pio->io_allocator, pio);
2853 pio->io_error = error;
2858 gnpp = &gio->io_gang_tree;
2860 gnpp = pio->io_private;
2861 ASSERT(pio->io_ready == zio_write_gang_member_ready);
2864 gn = zio_gang_node_alloc(gnpp);
2866 bzero(gbh, SPA_GANGBLOCKSIZE);
2867 gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE);
2870 * Create the gang header.
2872 zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2873 zio_write_gang_done, NULL, pio->io_priority,
2874 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2877 * Create and nowait the gang children.
2879 for (int g = 0; resid != 0; resid -= lsize, g++) {
2880 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2882 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2884 zp.zp_checksum = gio->io_prop.zp_checksum;
2885 zp.zp_compress = ZIO_COMPRESS_OFF;
2886 zp.zp_complevel = gio->io_prop.zp_complevel;
2887 zp.zp_type = DMU_OT_NONE;
2889 zp.zp_copies = gio->io_prop.zp_copies;
2890 zp.zp_dedup = B_FALSE;
2891 zp.zp_dedup_verify = B_FALSE;
2892 zp.zp_nopwrite = B_FALSE;
2893 zp.zp_encrypt = gio->io_prop.zp_encrypt;
2894 zp.zp_byteorder = gio->io_prop.zp_byteorder;
2895 bzero(zp.zp_salt, ZIO_DATA_SALT_LEN);
2896 bzero(zp.zp_iv, ZIO_DATA_IV_LEN);
2897 bzero(zp.zp_mac, ZIO_DATA_MAC_LEN);
2899 zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2900 has_data ? abd_get_offset(pio->io_abd, pio->io_size -
2901 resid) : NULL, lsize, lsize, &zp,
2902 zio_write_gang_member_ready, NULL, NULL,
2903 zio_write_gang_done, &gn->gn_child[g], pio->io_priority,
2904 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2906 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2907 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2911 * Gang children won't throttle but we should
2912 * account for their work, so reserve an allocation
2913 * slot for them here.
2915 VERIFY(metaslab_class_throttle_reserve(mc,
2916 zp.zp_copies, cio->io_allocator, cio, flags));
2922 * Set pio's pipeline to just wait for zio to finish.
2924 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2927 * We didn't allocate this bp, so make sure it doesn't get unmarked.
2929 pio->io_flags &= ~ZIO_FLAG_FASTWRITE;
2937 * The zio_nop_write stage in the pipeline determines if allocating a
2938 * new bp is necessary. The nopwrite feature can handle writes in
2939 * either syncing or open context (i.e. zil writes) and as a result is
2940 * mutually exclusive with dedup.
2942 * By leveraging a cryptographically secure checksum, such as SHA256, we
2943 * can compare the checksums of the new data and the old to determine if
2944 * allocating a new block is required. Note that our requirements for
2945 * cryptographic strength are fairly weak: there can't be any accidental
2946 * hash collisions, but we don't need to be secure against intentional
2947 * (malicious) collisions. To trigger a nopwrite, you have to be able
2948 * to write the file to begin with, and triggering an incorrect (hash
2949 * collision) nopwrite is no worse than simply writing to the file.
2950 * That said, there are no known attacks against the checksum algorithms
2951 * used for nopwrite, assuming that the salt and the checksums
2952 * themselves remain secret.
2955 zio_nop_write(zio_t *zio)
2957 blkptr_t *bp = zio->io_bp;
2958 blkptr_t *bp_orig = &zio->io_bp_orig;
2959 zio_prop_t *zp = &zio->io_prop;
2961 ASSERT(BP_GET_LEVEL(bp) == 0);
2962 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2963 ASSERT(zp->zp_nopwrite);
2964 ASSERT(!zp->zp_dedup);
2965 ASSERT(zio->io_bp_override == NULL);
2966 ASSERT(IO_IS_ALLOCATING(zio));
2969 * Check to see if the original bp and the new bp have matching
2970 * characteristics (i.e. same checksum, compression algorithms, etc).
2971 * If they don't then just continue with the pipeline which will
2972 * allocate a new bp.
2974 if (BP_IS_HOLE(bp_orig) ||
2975 !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
2976 ZCHECKSUM_FLAG_NOPWRITE) ||
2977 BP_IS_ENCRYPTED(bp) || BP_IS_ENCRYPTED(bp_orig) ||
2978 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2979 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2980 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2981 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2985 * If the checksums match then reset the pipeline so that we
2986 * avoid allocating a new bp and issuing any I/O.
2988 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2989 ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
2990 ZCHECKSUM_FLAG_NOPWRITE);
2991 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2992 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2993 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2994 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2995 sizeof (uint64_t)) == 0);
2998 * If we're overwriting a block that is currently on an
2999 * indirect vdev, then ignore the nopwrite request and
3000 * allow a new block to be allocated on a concrete vdev.
3002 spa_config_enter(zio->io_spa, SCL_VDEV, FTAG, RW_READER);
3003 vdev_t *tvd = vdev_lookup_top(zio->io_spa,
3004 DVA_GET_VDEV(&bp->blk_dva[0]));
3005 if (tvd->vdev_ops == &vdev_indirect_ops) {
3006 spa_config_exit(zio->io_spa, SCL_VDEV, FTAG);
3009 spa_config_exit(zio->io_spa, SCL_VDEV, FTAG);
3012 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3013 zio->io_flags |= ZIO_FLAG_NOPWRITE;
3020 * ==========================================================================
3022 * ==========================================================================
3025 zio_ddt_child_read_done(zio_t *zio)
3027 blkptr_t *bp = zio->io_bp;
3028 ddt_entry_t *dde = zio->io_private;
3030 zio_t *pio = zio_unique_parent(zio);
3032 mutex_enter(&pio->io_lock);
3033 ddp = ddt_phys_select(dde, bp);
3034 if (zio->io_error == 0)
3035 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
3037 if (zio->io_error == 0 && dde->dde_repair_abd == NULL)
3038 dde->dde_repair_abd = zio->io_abd;
3040 abd_free(zio->io_abd);
3041 mutex_exit(&pio->io_lock);
3045 zio_ddt_read_start(zio_t *zio)
3047 blkptr_t *bp = zio->io_bp;
3049 ASSERT(BP_GET_DEDUP(bp));
3050 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
3051 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3053 if (zio->io_child_error[ZIO_CHILD_DDT]) {
3054 ddt_t *ddt = ddt_select(zio->io_spa, bp);
3055 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
3056 ddt_phys_t *ddp = dde->dde_phys;
3057 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
3060 ASSERT(zio->io_vsd == NULL);
3063 if (ddp_self == NULL)
3066 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
3067 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
3069 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
3071 zio_nowait(zio_read(zio, zio->io_spa, &blk,
3072 abd_alloc_for_io(zio->io_size, B_TRUE),
3073 zio->io_size, zio_ddt_child_read_done, dde,
3074 zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) |
3075 ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark));
3080 zio_nowait(zio_read(zio, zio->io_spa, bp,
3081 zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority,
3082 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
3088 zio_ddt_read_done(zio_t *zio)
3090 blkptr_t *bp = zio->io_bp;
3092 if (zio_wait_for_children(zio, ZIO_CHILD_DDT_BIT, ZIO_WAIT_DONE)) {
3096 ASSERT(BP_GET_DEDUP(bp));
3097 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
3098 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3100 if (zio->io_child_error[ZIO_CHILD_DDT]) {
3101 ddt_t *ddt = ddt_select(zio->io_spa, bp);
3102 ddt_entry_t *dde = zio->io_vsd;
3104 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
3108 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
3109 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
3112 if (dde->dde_repair_abd != NULL) {
3113 abd_copy(zio->io_abd, dde->dde_repair_abd,
3115 zio->io_child_error[ZIO_CHILD_DDT] = 0;
3117 ddt_repair_done(ddt, dde);
3121 ASSERT(zio->io_vsd == NULL);
3127 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
3129 spa_t *spa = zio->io_spa;
3130 boolean_t do_raw = !!(zio->io_flags & ZIO_FLAG_RAW);
3132 ASSERT(!(zio->io_bp_override && do_raw));
3135 * Note: we compare the original data, not the transformed data,
3136 * because when zio->io_bp is an override bp, we will not have
3137 * pushed the I/O transforms. That's an important optimization
3138 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
3139 * However, we should never get a raw, override zio so in these
3140 * cases we can compare the io_abd directly. This is useful because
3141 * it allows us to do dedup verification even if we don't have access
3142 * to the original data (for instance, if the encryption keys aren't
3146 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
3147 zio_t *lio = dde->dde_lead_zio[p];
3149 if (lio != NULL && do_raw) {
3150 return (lio->io_size != zio->io_size ||
3151 abd_cmp(zio->io_abd, lio->io_abd) != 0);
3152 } else if (lio != NULL) {
3153 return (lio->io_orig_size != zio->io_orig_size ||
3154 abd_cmp(zio->io_orig_abd, lio->io_orig_abd) != 0);
3158 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
3159 ddt_phys_t *ddp = &dde->dde_phys[p];
3161 if (ddp->ddp_phys_birth != 0 && do_raw) {
3162 blkptr_t blk = *zio->io_bp;
3167 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
3168 psize = BP_GET_PSIZE(&blk);
3170 if (psize != zio->io_size)
3175 tmpabd = abd_alloc_for_io(psize, B_TRUE);
3177 error = zio_wait(zio_read(NULL, spa, &blk, tmpabd,
3178 psize, NULL, NULL, ZIO_PRIORITY_SYNC_READ,
3179 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
3180 ZIO_FLAG_RAW, &zio->io_bookmark));
3183 if (abd_cmp(tmpabd, zio->io_abd) != 0)
3184 error = SET_ERROR(ENOENT);
3189 return (error != 0);
3190 } else if (ddp->ddp_phys_birth != 0) {
3191 arc_buf_t *abuf = NULL;
3192 arc_flags_t aflags = ARC_FLAG_WAIT;
3193 blkptr_t blk = *zio->io_bp;
3196 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
3198 if (BP_GET_LSIZE(&blk) != zio->io_orig_size)
3203 error = arc_read(NULL, spa, &blk,
3204 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
3205 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
3206 &aflags, &zio->io_bookmark);
3209 if (abd_cmp_buf(zio->io_orig_abd, abuf->b_data,
3210 zio->io_orig_size) != 0)
3211 error = SET_ERROR(ENOENT);
3212 arc_buf_destroy(abuf, &abuf);
3216 return (error != 0);
3224 zio_ddt_child_write_ready(zio_t *zio)
3226 int p = zio->io_prop.zp_copies;
3227 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
3228 ddt_entry_t *dde = zio->io_private;
3229 ddt_phys_t *ddp = &dde->dde_phys[p];
3237 ASSERT(dde->dde_lead_zio[p] == zio);
3239 ddt_phys_fill(ddp, zio->io_bp);
3241 zio_link_t *zl = NULL;
3242 while ((pio = zio_walk_parents(zio, &zl)) != NULL)
3243 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
3249 zio_ddt_child_write_done(zio_t *zio)
3251 int p = zio->io_prop.zp_copies;
3252 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
3253 ddt_entry_t *dde = zio->io_private;
3254 ddt_phys_t *ddp = &dde->dde_phys[p];
3258 ASSERT(ddp->ddp_refcnt == 0);
3259 ASSERT(dde->dde_lead_zio[p] == zio);
3260 dde->dde_lead_zio[p] = NULL;
3262 if (zio->io_error == 0) {
3263 zio_link_t *zl = NULL;
3264 while (zio_walk_parents(zio, &zl) != NULL)
3265 ddt_phys_addref(ddp);
3267 ddt_phys_clear(ddp);
3274 zio_ddt_write(zio_t *zio)
3276 spa_t *spa = zio->io_spa;
3277 blkptr_t *bp = zio->io_bp;
3278 uint64_t txg = zio->io_txg;
3279 zio_prop_t *zp = &zio->io_prop;
3280 int p = zp->zp_copies;
3282 ddt_t *ddt = ddt_select(spa, bp);
3286 ASSERT(BP_GET_DEDUP(bp));
3287 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
3288 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
3289 ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW)));
3292 dde = ddt_lookup(ddt, bp, B_TRUE);
3293 ddp = &dde->dde_phys[p];
3295 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
3297 * If we're using a weak checksum, upgrade to a strong checksum
3298 * and try again. If we're already using a strong checksum,
3299 * we can't resolve it, so just convert to an ordinary write.
3300 * (And automatically e-mail a paper to Nature?)
3302 if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
3303 ZCHECKSUM_FLAG_DEDUP)) {
3304 zp->zp_checksum = spa_dedup_checksum(spa);
3305 zio_pop_transforms(zio);
3306 zio->io_stage = ZIO_STAGE_OPEN;
3309 zp->zp_dedup = B_FALSE;
3310 BP_SET_DEDUP(bp, B_FALSE);
3312 ASSERT(!BP_GET_DEDUP(bp));
3313 zio->io_pipeline = ZIO_WRITE_PIPELINE;
3318 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
3319 if (ddp->ddp_phys_birth != 0)
3320 ddt_bp_fill(ddp, bp, txg);
3321 if (dde->dde_lead_zio[p] != NULL)
3322 zio_add_child(zio, dde->dde_lead_zio[p]);
3324 ddt_phys_addref(ddp);
3325 } else if (zio->io_bp_override) {
3326 ASSERT(bp->blk_birth == txg);
3327 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
3328 ddt_phys_fill(ddp, bp);
3329 ddt_phys_addref(ddp);
3331 cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
3332 zio->io_orig_size, zio->io_orig_size, zp,
3333 zio_ddt_child_write_ready, NULL, NULL,
3334 zio_ddt_child_write_done, dde, zio->io_priority,
3335 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
3337 zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL);
3338 dde->dde_lead_zio[p] = cio;
3348 ddt_entry_t *freedde; /* for debugging */
3351 zio_ddt_free(zio_t *zio)
3353 spa_t *spa = zio->io_spa;
3354 blkptr_t *bp = zio->io_bp;
3355 ddt_t *ddt = ddt_select(spa, bp);
3359 ASSERT(BP_GET_DEDUP(bp));
3360 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3363 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
3365 ddp = ddt_phys_select(dde, bp);
3367 ddt_phys_decref(ddp);
3375 * ==========================================================================
3376 * Allocate and free blocks
3377 * ==========================================================================
3381 zio_io_to_allocate(spa_t *spa, int allocator)
3385 ASSERT(MUTEX_HELD(&spa->spa_alloc_locks[allocator]));
3387 zio = avl_first(&spa->spa_alloc_trees[allocator]);
3391 ASSERT(IO_IS_ALLOCATING(zio));
3394 * Try to place a reservation for this zio. If we're unable to
3395 * reserve then we throttle.
3397 ASSERT3U(zio->io_allocator, ==, allocator);
3398 if (!metaslab_class_throttle_reserve(zio->io_metaslab_class,
3399 zio->io_prop.zp_copies, zio->io_allocator, zio, 0)) {
3403 avl_remove(&spa->spa_alloc_trees[allocator], zio);
3404 ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE);
3410 zio_dva_throttle(zio_t *zio)
3412 spa_t *spa = zio->io_spa;
3414 metaslab_class_t *mc;
3416 /* locate an appropriate allocation class */
3417 mc = spa_preferred_class(spa, zio->io_size, zio->io_prop.zp_type,
3418 zio->io_prop.zp_level, zio->io_prop.zp_zpl_smallblk);
3420 if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE ||
3421 !mc->mc_alloc_throttle_enabled ||
3422 zio->io_child_type == ZIO_CHILD_GANG ||
3423 zio->io_flags & ZIO_FLAG_NODATA) {
3427 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3429 ASSERT3U(zio->io_queued_timestamp, >, 0);
3430 ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE);
3432 zbookmark_phys_t *bm = &zio->io_bookmark;
3434 * We want to try to use as many allocators as possible to help improve
3435 * performance, but we also want logically adjacent IOs to be physically
3436 * adjacent to improve sequential read performance. We chunk each object
3437 * into 2^20 block regions, and then hash based on the objset, object,
3438 * level, and region to accomplish both of these goals.
3440 zio->io_allocator = cityhash4(bm->zb_objset, bm->zb_object,
3441 bm->zb_level, bm->zb_blkid >> 20) % spa->spa_alloc_count;
3442 mutex_enter(&spa->spa_alloc_locks[zio->io_allocator]);
3443 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3444 zio->io_metaslab_class = mc;
3445 avl_add(&spa->spa_alloc_trees[zio->io_allocator], zio);
3446 nio = zio_io_to_allocate(spa, zio->io_allocator);
3447 mutex_exit(&spa->spa_alloc_locks[zio->io_allocator]);
3452 zio_allocate_dispatch(spa_t *spa, int allocator)
3456 mutex_enter(&spa->spa_alloc_locks[allocator]);
3457 zio = zio_io_to_allocate(spa, allocator);
3458 mutex_exit(&spa->spa_alloc_locks[allocator]);
3462 ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE);
3463 ASSERT0(zio->io_error);
3464 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE);
3468 zio_dva_allocate(zio_t *zio)
3470 spa_t *spa = zio->io_spa;
3471 metaslab_class_t *mc;
3472 blkptr_t *bp = zio->io_bp;
3476 if (zio->io_gang_leader == NULL) {
3477 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3478 zio->io_gang_leader = zio;
3481 ASSERT(BP_IS_HOLE(bp));
3482 ASSERT0(BP_GET_NDVAS(bp));
3483 ASSERT3U(zio->io_prop.zp_copies, >, 0);
3484 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
3485 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
3487 flags |= (zio->io_flags & ZIO_FLAG_FASTWRITE) ? METASLAB_FASTWRITE : 0;
3488 if (zio->io_flags & ZIO_FLAG_NODATA)
3489 flags |= METASLAB_DONT_THROTTLE;
3490 if (zio->io_flags & ZIO_FLAG_GANG_CHILD)
3491 flags |= METASLAB_GANG_CHILD;
3492 if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE)
3493 flags |= METASLAB_ASYNC_ALLOC;
3496 * if not already chosen, locate an appropriate allocation class
3498 mc = zio->io_metaslab_class;
3500 mc = spa_preferred_class(spa, zio->io_size,
3501 zio->io_prop.zp_type, zio->io_prop.zp_level,
3502 zio->io_prop.zp_zpl_smallblk);
3503 zio->io_metaslab_class = mc;
3506 error = metaslab_alloc(spa, mc, zio->io_size, bp,
3507 zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
3508 &zio->io_alloc_list, zio, zio->io_allocator);
3511 * Fallback to normal class when an alloc class is full
3513 if (error == ENOSPC && mc != spa_normal_class(spa)) {
3515 * If throttling, transfer reservation over to normal class.
3516 * The io_allocator slot can remain the same even though we
3517 * are switching classes.
3519 if (mc->mc_alloc_throttle_enabled &&
3520 (zio->io_flags & ZIO_FLAG_IO_ALLOCATING)) {
3521 metaslab_class_throttle_unreserve(mc,
3522 zio->io_prop.zp_copies, zio->io_allocator, zio);
3523 zio->io_flags &= ~ZIO_FLAG_IO_ALLOCATING;
3525 mc = spa_normal_class(spa);
3526 VERIFY(metaslab_class_throttle_reserve(mc,
3527 zio->io_prop.zp_copies, zio->io_allocator, zio,
3528 flags | METASLAB_MUST_RESERVE));
3530 mc = spa_normal_class(spa);
3532 zio->io_metaslab_class = mc;
3534 error = metaslab_alloc(spa, mc, zio->io_size, bp,
3535 zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
3536 &zio->io_alloc_list, zio, zio->io_allocator);
3540 zfs_dbgmsg("%s: metaslab allocation failure: zio %px, "
3541 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
3543 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
3544 return (zio_write_gang_block(zio));
3545 zio->io_error = error;
3552 zio_dva_free(zio_t *zio)
3554 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
3560 zio_dva_claim(zio_t *zio)
3564 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
3566 zio->io_error = error;
3572 * Undo an allocation. This is used by zio_done() when an I/O fails
3573 * and we want to give back the block we just allocated.
3574 * This handles both normal blocks and gang blocks.
3577 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
3579 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
3580 ASSERT(zio->io_bp_override == NULL);
3582 if (!BP_IS_HOLE(bp))
3583 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
3586 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
3587 zio_dva_unallocate(zio, gn->gn_child[g],
3588 &gn->gn_gbh->zg_blkptr[g]);
3594 * Try to allocate an intent log block. Return 0 on success, errno on failure.
3597 zio_alloc_zil(spa_t *spa, objset_t *os, uint64_t txg, blkptr_t *new_bp,
3598 uint64_t size, boolean_t *slog)
3601 zio_alloc_list_t io_alloc_list;
3603 ASSERT(txg > spa_syncing_txg(spa));
3605 metaslab_trace_init(&io_alloc_list);
3608 * Block pointer fields are useful to metaslabs for stats and debugging.
3609 * Fill in the obvious ones before calling into metaslab_alloc().
3611 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3612 BP_SET_PSIZE(new_bp, size);
3613 BP_SET_LEVEL(new_bp, 0);
3616 * When allocating a zil block, we don't have information about
3617 * the final destination of the block except the objset it's part
3618 * of, so we just hash the objset ID to pick the allocator to get
3621 error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1,
3622 txg, NULL, METASLAB_FASTWRITE, &io_alloc_list, NULL,
3623 cityhash4(0, 0, 0, os->os_dsl_dataset->ds_object) %
3624 spa->spa_alloc_count);
3628 error = metaslab_alloc(spa, spa_normal_class(spa), size,
3629 new_bp, 1, txg, NULL, METASLAB_FASTWRITE,
3630 &io_alloc_list, NULL, cityhash4(0, 0, 0,
3631 os->os_dsl_dataset->ds_object) % spa->spa_alloc_count);
3635 metaslab_trace_fini(&io_alloc_list);
3638 BP_SET_LSIZE(new_bp, size);
3639 BP_SET_PSIZE(new_bp, size);
3640 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
3641 BP_SET_CHECKSUM(new_bp,
3642 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
3643 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
3644 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3645 BP_SET_LEVEL(new_bp, 0);
3646 BP_SET_DEDUP(new_bp, 0);
3647 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
3650 * encrypted blocks will require an IV and salt. We generate
3651 * these now since we will not be rewriting the bp at
3654 if (os->os_encrypted) {
3655 uint8_t iv[ZIO_DATA_IV_LEN];
3656 uint8_t salt[ZIO_DATA_SALT_LEN];
3658 BP_SET_CRYPT(new_bp, B_TRUE);
3659 VERIFY0(spa_crypt_get_salt(spa,
3660 dmu_objset_id(os), salt));
3661 VERIFY0(zio_crypt_generate_iv(iv));
3663 zio_crypt_encode_params_bp(new_bp, salt, iv);
3666 zfs_dbgmsg("%s: zil block allocation failure: "
3667 "size %llu, error %d", spa_name(spa), size, error);
3674 * ==========================================================================
3675 * Read and write to physical devices
3676 * ==========================================================================
3680 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3681 * stops after this stage and will resume upon I/O completion.
3682 * However, there are instances where the vdev layer may need to
3683 * continue the pipeline when an I/O was not issued. Since the I/O
3684 * that was sent to the vdev layer might be different than the one
3685 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3686 * force the underlying vdev layers to call either zio_execute() or
3687 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3690 zio_vdev_io_start(zio_t *zio)
3692 vdev_t *vd = zio->io_vd;
3694 spa_t *spa = zio->io_spa;
3698 ASSERT(zio->io_error == 0);
3699 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
3702 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3703 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
3706 * The mirror_ops handle multiple DVAs in a single BP.
3708 vdev_mirror_ops.vdev_op_io_start(zio);
3712 ASSERT3P(zio->io_logical, !=, zio);
3713 if (zio->io_type == ZIO_TYPE_WRITE) {
3714 ASSERT(spa->spa_trust_config);
3717 * Note: the code can handle other kinds of writes,
3718 * but we don't expect them.
3720 if (zio->io_vd->vdev_removing) {
3721 ASSERT(zio->io_flags &
3722 (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL |
3723 ZIO_FLAG_RESILVER | ZIO_FLAG_INDUCE_DAMAGE));
3727 align = 1ULL << vd->vdev_top->vdev_ashift;
3729 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
3730 P2PHASE(zio->io_size, align) != 0) {
3731 /* Transform logical writes to be a full physical block size. */
3732 uint64_t asize = P2ROUNDUP(zio->io_size, align);
3733 abd_t *abuf = abd_alloc_sametype(zio->io_abd, asize);
3734 ASSERT(vd == vd->vdev_top);
3735 if (zio->io_type == ZIO_TYPE_WRITE) {
3736 abd_copy(abuf, zio->io_abd, zio->io_size);
3737 abd_zero_off(abuf, zio->io_size, asize - zio->io_size);
3739 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
3743 * If this is not a physical io, make sure that it is properly aligned
3744 * before proceeding.
3746 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
3747 ASSERT0(P2PHASE(zio->io_offset, align));
3748 ASSERT0(P2PHASE(zio->io_size, align));
3751 * For physical writes, we allow 512b aligned writes and assume
3752 * the device will perform a read-modify-write as necessary.
3754 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
3755 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
3758 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
3761 * If this is a repair I/O, and there's no self-healing involved --
3762 * that is, we're just resilvering what we expect to resilver --
3763 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3764 * This prevents spurious resilvering.
3766 * There are a few ways that we can end up creating these spurious
3769 * 1. A resilver i/o will be issued if any DVA in the BP has a
3770 * dirty DTL. The mirror code will issue resilver writes to
3771 * each DVA, including the one(s) that are not on vdevs with dirty
3774 * 2. With nested replication, which happens when we have a
3775 * "replacing" or "spare" vdev that's a child of a mirror or raidz.
3776 * For example, given mirror(replacing(A+B), C), it's likely that
3777 * only A is out of date (it's the new device). In this case, we'll
3778 * read from C, then use the data to resilver A+B -- but we don't
3779 * actually want to resilver B, just A. The top-level mirror has no
3780 * way to know this, so instead we just discard unnecessary repairs
3781 * as we work our way down the vdev tree.
3783 * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc.
3784 * The same logic applies to any form of nested replication: ditto
3785 * + mirror, RAID-Z + replacing, etc.
3787 * However, indirect vdevs point off to other vdevs which may have
3788 * DTL's, so we never bypass them. The child i/os on concrete vdevs
3789 * will be properly bypassed instead.
3791 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3792 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
3793 zio->io_txg != 0 && /* not a delegated i/o */
3794 vd->vdev_ops != &vdev_indirect_ops &&
3795 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
3796 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3797 zio_vdev_io_bypass(zio);
3801 if (vd->vdev_ops->vdev_op_leaf && (zio->io_type == ZIO_TYPE_READ ||
3802 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_TRIM)) {
3804 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio))
3807 if ((zio = vdev_queue_io(zio)) == NULL)
3810 if (!vdev_accessible(vd, zio)) {
3811 zio->io_error = SET_ERROR(ENXIO);
3815 zio->io_delay = gethrtime();
3818 vd->vdev_ops->vdev_op_io_start(zio);
3823 zio_vdev_io_done(zio_t *zio)
3825 vdev_t *vd = zio->io_vd;
3826 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
3827 boolean_t unexpected_error = B_FALSE;
3829 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3833 ASSERT(zio->io_type == ZIO_TYPE_READ ||
3834 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_TRIM);
3837 zio->io_delay = gethrtime() - zio->io_delay;
3839 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
3841 vdev_queue_io_done(zio);
3843 if (zio->io_type == ZIO_TYPE_WRITE)
3844 vdev_cache_write(zio);
3846 if (zio_injection_enabled && zio->io_error == 0)
3847 zio->io_error = zio_handle_device_injections(vd, zio,
3850 if (zio_injection_enabled && zio->io_error == 0)
3851 zio->io_error = zio_handle_label_injection(zio, EIO);
3853 if (zio->io_error && zio->io_type != ZIO_TYPE_TRIM) {
3854 if (!vdev_accessible(vd, zio)) {
3855 zio->io_error = SET_ERROR(ENXIO);
3857 unexpected_error = B_TRUE;
3862 ops->vdev_op_io_done(zio);
3864 if (unexpected_error)
3865 VERIFY(vdev_probe(vd, zio) == NULL);
3871 * This function is used to change the priority of an existing zio that is
3872 * currently in-flight. This is used by the arc to upgrade priority in the
3873 * event that a demand read is made for a block that is currently queued
3874 * as a scrub or async read IO. Otherwise, the high priority read request
3875 * would end up having to wait for the lower priority IO.
3878 zio_change_priority(zio_t *pio, zio_priority_t priority)
3880 zio_t *cio, *cio_next;
3881 zio_link_t *zl = NULL;
3883 ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
3885 if (pio->io_vd != NULL && pio->io_vd->vdev_ops->vdev_op_leaf) {
3886 vdev_queue_change_io_priority(pio, priority);
3888 pio->io_priority = priority;
3891 mutex_enter(&pio->io_lock);
3892 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
3893 cio_next = zio_walk_children(pio, &zl);
3894 zio_change_priority(cio, priority);
3896 mutex_exit(&pio->io_lock);
3900 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3901 * disk, and use that to finish the checksum ereport later.
3904 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
3905 const abd_t *good_buf)
3907 /* no processing needed */
3908 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
3913 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
3915 void *abd = abd_alloc_sametype(zio->io_abd, zio->io_size);
3917 abd_copy(abd, zio->io_abd, zio->io_size);
3919 zcr->zcr_cbinfo = zio->io_size;
3920 zcr->zcr_cbdata = abd;
3921 zcr->zcr_finish = zio_vsd_default_cksum_finish;
3922 zcr->zcr_free = zio_abd_free;
3926 zio_vdev_io_assess(zio_t *zio)
3928 vdev_t *vd = zio->io_vd;
3930 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3934 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3935 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
3937 if (zio->io_vsd != NULL) {
3938 zio->io_vsd_ops->vsd_free(zio);
3942 if (zio_injection_enabled && zio->io_error == 0)
3943 zio->io_error = zio_handle_fault_injection(zio, EIO);
3946 * If the I/O failed, determine whether we should attempt to retry it.
3948 * On retry, we cut in line in the issue queue, since we don't want
3949 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3951 if (zio->io_error && vd == NULL &&
3952 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
3953 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
3954 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
3956 zio->io_flags |= ZIO_FLAG_IO_RETRY |
3957 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
3958 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
3959 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
3960 zio_requeue_io_start_cut_in_line);
3965 * If we got an error on a leaf device, convert it to ENXIO
3966 * if the device is not accessible at all.
3968 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3969 !vdev_accessible(vd, zio))
3970 zio->io_error = SET_ERROR(ENXIO);
3973 * If we can't write to an interior vdev (mirror or RAID-Z),
3974 * set vdev_cant_write so that we stop trying to allocate from it.
3976 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
3977 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
3978 vd->vdev_cant_write = B_TRUE;
3982 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
3983 * attempts will ever succeed. In this case we set a persistent
3984 * boolean flag so that we don't bother with it in the future.
3986 if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) &&
3987 zio->io_type == ZIO_TYPE_IOCTL &&
3988 zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL)
3989 vd->vdev_nowritecache = B_TRUE;
3992 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3994 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3995 zio->io_physdone != NULL) {
3996 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
3997 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
3998 zio->io_physdone(zio->io_logical);
4005 zio_vdev_io_reissue(zio_t *zio)
4007 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
4008 ASSERT(zio->io_error == 0);
4010 zio->io_stage >>= 1;
4014 zio_vdev_io_redone(zio_t *zio)
4016 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
4018 zio->io_stage >>= 1;
4022 zio_vdev_io_bypass(zio_t *zio)
4024 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
4025 ASSERT(zio->io_error == 0);
4027 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
4028 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
4032 * ==========================================================================
4033 * Encrypt and store encryption parameters
4034 * ==========================================================================
4039 * This function is used for ZIO_STAGE_ENCRYPT. It is responsible for
4040 * managing the storage of encryption parameters and passing them to the
4041 * lower-level encryption functions.
4044 zio_encrypt(zio_t *zio)
4046 zio_prop_t *zp = &zio->io_prop;
4047 spa_t *spa = zio->io_spa;
4048 blkptr_t *bp = zio->io_bp;
4049 uint64_t psize = BP_GET_PSIZE(bp);
4050 uint64_t dsobj = zio->io_bookmark.zb_objset;
4051 dmu_object_type_t ot = BP_GET_TYPE(bp);
4052 void *enc_buf = NULL;
4054 uint8_t salt[ZIO_DATA_SALT_LEN];
4055 uint8_t iv[ZIO_DATA_IV_LEN];
4056 uint8_t mac[ZIO_DATA_MAC_LEN];
4057 boolean_t no_crypt = B_FALSE;
4059 /* the root zio already encrypted the data */
4060 if (zio->io_child_type == ZIO_CHILD_GANG)
4063 /* only ZIL blocks are re-encrypted on rewrite */
4064 if (!IO_IS_ALLOCATING(zio) && ot != DMU_OT_INTENT_LOG)
4067 if (!(zp->zp_encrypt || BP_IS_ENCRYPTED(bp))) {
4068 BP_SET_CRYPT(bp, B_FALSE);
4072 /* if we are doing raw encryption set the provided encryption params */
4073 if (zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) {
4074 ASSERT0(BP_GET_LEVEL(bp));
4075 BP_SET_CRYPT(bp, B_TRUE);
4076 BP_SET_BYTEORDER(bp, zp->zp_byteorder);
4077 if (ot != DMU_OT_OBJSET)
4078 zio_crypt_encode_mac_bp(bp, zp->zp_mac);
4080 /* dnode blocks must be written out in the provided byteorder */
4081 if (zp->zp_byteorder != ZFS_HOST_BYTEORDER &&
4082 ot == DMU_OT_DNODE) {
4083 void *bswap_buf = zio_buf_alloc(psize);
4084 abd_t *babd = abd_get_from_buf(bswap_buf, psize);
4086 ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
4087 abd_copy_to_buf(bswap_buf, zio->io_abd, psize);
4088 dmu_ot_byteswap[DMU_OT_BYTESWAP(ot)].ob_func(bswap_buf,
4091 abd_take_ownership_of_buf(babd, B_TRUE);
4092 zio_push_transform(zio, babd, psize, psize, NULL);
4095 if (DMU_OT_IS_ENCRYPTED(ot))
4096 zio_crypt_encode_params_bp(bp, zp->zp_salt, zp->zp_iv);
4100 /* indirect blocks only maintain a cksum of the lower level MACs */
4101 if (BP_GET_LEVEL(bp) > 0) {
4102 BP_SET_CRYPT(bp, B_TRUE);
4103 VERIFY0(zio_crypt_do_indirect_mac_checksum_abd(B_TRUE,
4104 zio->io_orig_abd, BP_GET_LSIZE(bp), BP_SHOULD_BYTESWAP(bp),
4106 zio_crypt_encode_mac_bp(bp, mac);
4111 * Objset blocks are a special case since they have 2 256-bit MACs
4112 * embedded within them.
4114 if (ot == DMU_OT_OBJSET) {
4115 ASSERT0(DMU_OT_IS_ENCRYPTED(ot));
4116 ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
4117 BP_SET_CRYPT(bp, B_TRUE);
4118 VERIFY0(spa_do_crypt_objset_mac_abd(B_TRUE, spa, dsobj,
4119 zio->io_abd, psize, BP_SHOULD_BYTESWAP(bp)));
4123 /* unencrypted object types are only authenticated with a MAC */
4124 if (!DMU_OT_IS_ENCRYPTED(ot)) {
4125 BP_SET_CRYPT(bp, B_TRUE);
4126 VERIFY0(spa_do_crypt_mac_abd(B_TRUE, spa, dsobj,
4127 zio->io_abd, psize, mac));
4128 zio_crypt_encode_mac_bp(bp, mac);
4133 * Later passes of sync-to-convergence may decide to rewrite data
4134 * in place to avoid more disk reallocations. This presents a problem
4135 * for encryption because this constitutes rewriting the new data with
4136 * the same encryption key and IV. However, this only applies to blocks
4137 * in the MOS (particularly the spacemaps) and we do not encrypt the
4138 * MOS. We assert that the zio is allocating or an intent log write
4141 ASSERT(IO_IS_ALLOCATING(zio) || ot == DMU_OT_INTENT_LOG);
4142 ASSERT(BP_GET_LEVEL(bp) == 0 || ot == DMU_OT_INTENT_LOG);
4143 ASSERT(spa_feature_is_active(spa, SPA_FEATURE_ENCRYPTION));
4144 ASSERT3U(psize, !=, 0);
4146 enc_buf = zio_buf_alloc(psize);
4147 eabd = abd_get_from_buf(enc_buf, psize);
4148 abd_take_ownership_of_buf(eabd, B_TRUE);
4151 * For an explanation of what encryption parameters are stored
4152 * where, see the block comment in zio_crypt.c.
4154 if (ot == DMU_OT_INTENT_LOG) {
4155 zio_crypt_decode_params_bp(bp, salt, iv);
4157 BP_SET_CRYPT(bp, B_TRUE);
4160 /* Perform the encryption. This should not fail */
4161 VERIFY0(spa_do_crypt_abd(B_TRUE, spa, &zio->io_bookmark,
4162 BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp),
4163 salt, iv, mac, psize, zio->io_abd, eabd, &no_crypt));
4165 /* encode encryption metadata into the bp */
4166 if (ot == DMU_OT_INTENT_LOG) {
4168 * ZIL blocks store the MAC in the embedded checksum, so the
4169 * transform must always be applied.
4171 zio_crypt_encode_mac_zil(enc_buf, mac);
4172 zio_push_transform(zio, eabd, psize, psize, NULL);
4174 BP_SET_CRYPT(bp, B_TRUE);
4175 zio_crypt_encode_params_bp(bp, salt, iv);
4176 zio_crypt_encode_mac_bp(bp, mac);
4179 ASSERT3U(ot, ==, DMU_OT_DNODE);
4182 zio_push_transform(zio, eabd, psize, psize, NULL);
4190 * ==========================================================================
4191 * Generate and verify checksums
4192 * ==========================================================================
4195 zio_checksum_generate(zio_t *zio)
4197 blkptr_t *bp = zio->io_bp;
4198 enum zio_checksum checksum;
4202 * This is zio_write_phys().
4203 * We're either generating a label checksum, or none at all.
4205 checksum = zio->io_prop.zp_checksum;
4207 if (checksum == ZIO_CHECKSUM_OFF)
4210 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
4212 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
4213 ASSERT(!IO_IS_ALLOCATING(zio));
4214 checksum = ZIO_CHECKSUM_GANG_HEADER;
4216 checksum = BP_GET_CHECKSUM(bp);
4220 zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size);
4226 zio_checksum_verify(zio_t *zio)
4228 zio_bad_cksum_t info;
4229 blkptr_t *bp = zio->io_bp;
4232 ASSERT(zio->io_vd != NULL);
4236 * This is zio_read_phys().
4237 * We're either verifying a label checksum, or nothing at all.
4239 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
4242 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
4245 if ((error = zio_checksum_error(zio, &info)) != 0) {
4246 zio->io_error = error;
4247 if (error == ECKSUM &&
4248 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
4249 int ret = zfs_ereport_start_checksum(zio->io_spa,
4250 zio->io_vd, &zio->io_bookmark, zio,
4251 zio->io_offset, zio->io_size, NULL, &info);
4253 if (ret != EALREADY) {
4254 mutex_enter(&zio->io_vd->vdev_stat_lock);
4255 zio->io_vd->vdev_stat.vs_checksum_errors++;
4256 mutex_exit(&zio->io_vd->vdev_stat_lock);
4265 * Called by RAID-Z to ensure we don't compute the checksum twice.
4268 zio_checksum_verified(zio_t *zio)
4270 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
4274 * ==========================================================================
4275 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
4276 * An error of 0 indicates success. ENXIO indicates whole-device failure,
4277 * which may be transient (e.g. unplugged) or permanent. ECKSUM and EIO
4278 * indicate errors that are specific to one I/O, and most likely permanent.
4279 * Any other error is presumed to be worse because we weren't expecting it.
4280 * ==========================================================================
4283 zio_worst_error(int e1, int e2)
4285 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
4288 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
4289 if (e1 == zio_error_rank[r1])
4292 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
4293 if (e2 == zio_error_rank[r2])
4296 return (r1 > r2 ? e1 : e2);
4300 * ==========================================================================
4302 * ==========================================================================
4305 zio_ready(zio_t *zio)
4307 blkptr_t *bp = zio->io_bp;
4308 zio_t *pio, *pio_next;
4309 zio_link_t *zl = NULL;
4311 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT | ZIO_CHILD_DDT_BIT,
4316 if (zio->io_ready) {
4317 ASSERT(IO_IS_ALLOCATING(zio));
4318 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
4319 (zio->io_flags & ZIO_FLAG_NOPWRITE));
4320 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
4325 if (bp != NULL && bp != &zio->io_bp_copy)
4326 zio->io_bp_copy = *bp;
4328 if (zio->io_error != 0) {
4329 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
4331 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
4332 ASSERT(IO_IS_ALLOCATING(zio));
4333 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
4334 ASSERT(zio->io_metaslab_class != NULL);
4337 * We were unable to allocate anything, unreserve and
4338 * issue the next I/O to allocate.
4340 metaslab_class_throttle_unreserve(
4341 zio->io_metaslab_class, zio->io_prop.zp_copies,
4342 zio->io_allocator, zio);
4343 zio_allocate_dispatch(zio->io_spa, zio->io_allocator);
4347 mutex_enter(&zio->io_lock);
4348 zio->io_state[ZIO_WAIT_READY] = 1;
4349 pio = zio_walk_parents(zio, &zl);
4350 mutex_exit(&zio->io_lock);
4353 * As we notify zio's parents, new parents could be added.
4354 * New parents go to the head of zio's io_parent_list, however,
4355 * so we will (correctly) not notify them. The remainder of zio's
4356 * io_parent_list, from 'pio_next' onward, cannot change because
4357 * all parents must wait for us to be done before they can be done.
4359 for (; pio != NULL; pio = pio_next) {
4360 pio_next = zio_walk_parents(zio, &zl);
4361 zio_notify_parent(pio, zio, ZIO_WAIT_READY, NULL);
4364 if (zio->io_flags & ZIO_FLAG_NODATA) {
4365 if (BP_IS_GANG(bp)) {
4366 zio->io_flags &= ~ZIO_FLAG_NODATA;
4368 ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE);
4369 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
4373 if (zio_injection_enabled &&
4374 zio->io_spa->spa_syncing_txg == zio->io_txg)
4375 zio_handle_ignored_writes(zio);
4381 * Update the allocation throttle accounting.
4384 zio_dva_throttle_done(zio_t *zio)
4386 zio_t *lio __maybe_unused = zio->io_logical;
4387 zio_t *pio = zio_unique_parent(zio);
4388 vdev_t *vd = zio->io_vd;
4389 int flags = METASLAB_ASYNC_ALLOC;
4391 ASSERT3P(zio->io_bp, !=, NULL);
4392 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
4393 ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE);
4394 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
4396 ASSERT3P(vd, ==, vd->vdev_top);
4397 ASSERT(zio_injection_enabled || !(zio->io_flags & ZIO_FLAG_IO_RETRY));
4398 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
4399 ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING);
4400 ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE));
4401 ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA));
4404 * Parents of gang children can have two flavors -- ones that
4405 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
4406 * and ones that allocated the constituent blocks. The allocation
4407 * throttle needs to know the allocating parent zio so we must find
4410 if (pio->io_child_type == ZIO_CHILD_GANG) {
4412 * If our parent is a rewrite gang child then our grandparent
4413 * would have been the one that performed the allocation.
4415 if (pio->io_flags & ZIO_FLAG_IO_REWRITE)
4416 pio = zio_unique_parent(pio);
4417 flags |= METASLAB_GANG_CHILD;
4420 ASSERT(IO_IS_ALLOCATING(pio));
4421 ASSERT3P(zio, !=, zio->io_logical);
4422 ASSERT(zio->io_logical != NULL);
4423 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
4424 ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE);
4425 ASSERT(zio->io_metaslab_class != NULL);
4427 mutex_enter(&pio->io_lock);
4428 metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags,
4429 pio->io_allocator, B_TRUE);
4430 mutex_exit(&pio->io_lock);
4432 metaslab_class_throttle_unreserve(zio->io_metaslab_class, 1,
4433 pio->io_allocator, pio);
4436 * Call into the pipeline to see if there is more work that
4437 * needs to be done. If there is work to be done it will be
4438 * dispatched to another taskq thread.
4440 zio_allocate_dispatch(zio->io_spa, pio->io_allocator);
4444 zio_done(zio_t *zio)
4447 * Always attempt to keep stack usage minimal here since
4448 * we can be called recursively up to 19 levels deep.
4450 const uint64_t psize = zio->io_size;
4451 zio_t *pio, *pio_next;
4452 zio_link_t *zl = NULL;
4455 * If our children haven't all completed,
4456 * wait for them and then repeat this pipeline stage.
4458 if (zio_wait_for_children(zio, ZIO_CHILD_ALL_BITS, ZIO_WAIT_DONE)) {
4463 * If the allocation throttle is enabled, then update the accounting.
4464 * We only track child I/Os that are part of an allocating async
4465 * write. We must do this since the allocation is performed
4466 * by the logical I/O but the actual write is done by child I/Os.
4468 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING &&
4469 zio->io_child_type == ZIO_CHILD_VDEV) {
4470 ASSERT(zio->io_metaslab_class != NULL);
4471 ASSERT(zio->io_metaslab_class->mc_alloc_throttle_enabled);
4472 zio_dva_throttle_done(zio);
4476 * If the allocation throttle is enabled, verify that
4477 * we have decremented the refcounts for every I/O that was throttled.
4479 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
4480 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
4481 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
4482 ASSERT(zio->io_bp != NULL);
4484 metaslab_group_alloc_verify(zio->io_spa, zio->io_bp, zio,
4486 VERIFY(zfs_refcount_not_held(
4487 &zio->io_metaslab_class->mc_alloc_slots[zio->io_allocator],
4492 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
4493 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
4494 ASSERT(zio->io_children[c][w] == 0);
4496 if (zio->io_bp != NULL && !BP_IS_EMBEDDED(zio->io_bp)) {
4497 ASSERT(zio->io_bp->blk_pad[0] == 0);
4498 ASSERT(zio->io_bp->blk_pad[1] == 0);
4499 ASSERT(bcmp(zio->io_bp, &zio->io_bp_copy,
4500 sizeof (blkptr_t)) == 0 ||
4501 (zio->io_bp == zio_unique_parent(zio)->io_bp));
4502 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) &&
4503 zio->io_bp_override == NULL &&
4504 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
4505 ASSERT3U(zio->io_prop.zp_copies, <=,
4506 BP_GET_NDVAS(zio->io_bp));
4507 ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 ||
4508 (BP_COUNT_GANG(zio->io_bp) ==
4509 BP_GET_NDVAS(zio->io_bp)));
4511 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
4512 VERIFY(BP_EQUAL(zio->io_bp, &zio->io_bp_orig));
4516 * If there were child vdev/gang/ddt errors, they apply to us now.
4518 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
4519 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
4520 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
4523 * If the I/O on the transformed data was successful, generate any
4524 * checksum reports now while we still have the transformed data.
4526 if (zio->io_error == 0) {
4527 while (zio->io_cksum_report != NULL) {
4528 zio_cksum_report_t *zcr = zio->io_cksum_report;
4529 uint64_t align = zcr->zcr_align;
4530 uint64_t asize = P2ROUNDUP(psize, align);
4531 abd_t *adata = zio->io_abd;
4533 if (asize != psize) {
4534 adata = abd_alloc(asize, B_TRUE);
4535 abd_copy(adata, zio->io_abd, psize);
4536 abd_zero_off(adata, psize, asize - psize);
4539 zio->io_cksum_report = zcr->zcr_next;
4540 zcr->zcr_next = NULL;
4541 zcr->zcr_finish(zcr, adata);
4542 zfs_ereport_free_checksum(zcr);
4549 zio_pop_transforms(zio); /* note: may set zio->io_error */
4551 vdev_stat_update(zio, psize);
4554 * If this I/O is attached to a particular vdev is slow, exceeding
4555 * 30 seconds to complete, post an error described the I/O delay.
4556 * We ignore these errors if the device is currently unavailable.
4558 if (zio->io_delay >= MSEC2NSEC(zio_slow_io_ms)) {
4559 if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd)) {
4561 * We want to only increment our slow IO counters if
4562 * the IO is valid (i.e. not if the drive is removed).
4564 * zfs_ereport_post() will also do these checks, but
4565 * it can also ratelimit and have other failures, so we
4566 * need to increment the slow_io counters independent
4569 if (zfs_ereport_is_valid(FM_EREPORT_ZFS_DELAY,
4570 zio->io_spa, zio->io_vd, zio)) {
4571 mutex_enter(&zio->io_vd->vdev_stat_lock);
4572 zio->io_vd->vdev_stat.vs_slow_ios++;
4573 mutex_exit(&zio->io_vd->vdev_stat_lock);
4575 (void) zfs_ereport_post(FM_EREPORT_ZFS_DELAY,
4576 zio->io_spa, zio->io_vd, &zio->io_bookmark,
4582 if (zio->io_error) {
4584 * If this I/O is attached to a particular vdev,
4585 * generate an error message describing the I/O failure
4586 * at the block level. We ignore these errors if the
4587 * device is currently unavailable.
4589 if (zio->io_error != ECKSUM && zio->io_vd != NULL &&
4590 !vdev_is_dead(zio->io_vd)) {
4591 int ret = zfs_ereport_post(FM_EREPORT_ZFS_IO,
4592 zio->io_spa, zio->io_vd, &zio->io_bookmark, zio, 0);
4593 if (ret != EALREADY) {
4594 mutex_enter(&zio->io_vd->vdev_stat_lock);
4595 if (zio->io_type == ZIO_TYPE_READ)
4596 zio->io_vd->vdev_stat.vs_read_errors++;
4597 else if (zio->io_type == ZIO_TYPE_WRITE)
4598 zio->io_vd->vdev_stat.vs_write_errors++;
4599 mutex_exit(&zio->io_vd->vdev_stat_lock);
4603 if ((zio->io_error == EIO || !(zio->io_flags &
4604 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
4605 zio == zio->io_logical) {
4607 * For logical I/O requests, tell the SPA to log the
4608 * error and generate a logical data ereport.
4610 spa_log_error(zio->io_spa, &zio->io_bookmark);
4611 (void) zfs_ereport_post(FM_EREPORT_ZFS_DATA,
4612 zio->io_spa, NULL, &zio->io_bookmark, zio, 0);
4616 if (zio->io_error && zio == zio->io_logical) {
4618 * Determine whether zio should be reexecuted. This will
4619 * propagate all the way to the root via zio_notify_parent().
4621 ASSERT(zio->io_vd == NULL && zio->io_bp != NULL);
4622 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
4624 if (IO_IS_ALLOCATING(zio) &&
4625 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
4626 if (zio->io_error != ENOSPC)
4627 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
4629 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4632 if ((zio->io_type == ZIO_TYPE_READ ||
4633 zio->io_type == ZIO_TYPE_FREE) &&
4634 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
4635 zio->io_error == ENXIO &&
4636 spa_load_state(zio->io_spa) == SPA_LOAD_NONE &&
4637 spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE)
4638 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4640 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
4641 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4644 * Here is a possibly good place to attempt to do
4645 * either combinatorial reconstruction or error correction
4646 * based on checksums. It also might be a good place
4647 * to send out preliminary ereports before we suspend
4653 * If there were logical child errors, they apply to us now.
4654 * We defer this until now to avoid conflating logical child
4655 * errors with errors that happened to the zio itself when
4656 * updating vdev stats and reporting FMA events above.
4658 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
4660 if ((zio->io_error || zio->io_reexecute) &&
4661 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
4662 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
4663 zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp);
4665 zio_gang_tree_free(&zio->io_gang_tree);
4668 * Godfather I/Os should never suspend.
4670 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
4671 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
4672 zio->io_reexecute &= ~ZIO_REEXECUTE_SUSPEND;
4674 if (zio->io_reexecute) {
4676 * This is a logical I/O that wants to reexecute.
4678 * Reexecute is top-down. When an i/o fails, if it's not
4679 * the root, it simply notifies its parent and sticks around.
4680 * The parent, seeing that it still has children in zio_done(),
4681 * does the same. This percolates all the way up to the root.
4682 * The root i/o will reexecute or suspend the entire tree.
4684 * This approach ensures that zio_reexecute() honors
4685 * all the original i/o dependency relationships, e.g.
4686 * parents not executing until children are ready.
4688 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
4690 zio->io_gang_leader = NULL;
4692 mutex_enter(&zio->io_lock);
4693 zio->io_state[ZIO_WAIT_DONE] = 1;
4694 mutex_exit(&zio->io_lock);
4697 * "The Godfather" I/O monitors its children but is
4698 * not a true parent to them. It will track them through
4699 * the pipeline but severs its ties whenever they get into
4700 * trouble (e.g. suspended). This allows "The Godfather"
4701 * I/O to return status without blocking.
4704 for (pio = zio_walk_parents(zio, &zl); pio != NULL;
4706 zio_link_t *remove_zl = zl;
4707 pio_next = zio_walk_parents(zio, &zl);
4709 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
4710 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
4711 zio_remove_child(pio, zio, remove_zl);
4713 * This is a rare code path, so we don't
4714 * bother with "next_to_execute".
4716 zio_notify_parent(pio, zio, ZIO_WAIT_DONE,
4721 if ((pio = zio_unique_parent(zio)) != NULL) {
4723 * We're not a root i/o, so there's nothing to do
4724 * but notify our parent. Don't propagate errors
4725 * upward since we haven't permanently failed yet.
4727 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
4728 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
4730 * This is a rare code path, so we don't bother with
4731 * "next_to_execute".
4733 zio_notify_parent(pio, zio, ZIO_WAIT_DONE, NULL);
4734 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
4736 * We'd fail again if we reexecuted now, so suspend
4737 * until conditions improve (e.g. device comes online).
4739 zio_suspend(zio->io_spa, zio, ZIO_SUSPEND_IOERR);
4742 * Reexecution is potentially a huge amount of work.
4743 * Hand it off to the otherwise-unused claim taskq.
4745 ASSERT(taskq_empty_ent(&zio->io_tqent));
4746 spa_taskq_dispatch_ent(zio->io_spa,
4747 ZIO_TYPE_CLAIM, ZIO_TASKQ_ISSUE,
4748 (task_func_t *)zio_reexecute, zio, 0,
4754 ASSERT(zio->io_child_count == 0);
4755 ASSERT(zio->io_reexecute == 0);
4756 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
4759 * Report any checksum errors, since the I/O is complete.
4761 while (zio->io_cksum_report != NULL) {
4762 zio_cksum_report_t *zcr = zio->io_cksum_report;
4763 zio->io_cksum_report = zcr->zcr_next;
4764 zcr->zcr_next = NULL;
4765 zcr->zcr_finish(zcr, NULL);
4766 zfs_ereport_free_checksum(zcr);
4769 if (zio->io_flags & ZIO_FLAG_FASTWRITE && zio->io_bp &&
4770 !BP_IS_HOLE(zio->io_bp) && !BP_IS_EMBEDDED(zio->io_bp) &&
4771 !(zio->io_flags & ZIO_FLAG_NOPWRITE)) {
4772 metaslab_fastwrite_unmark(zio->io_spa, zio->io_bp);
4776 * It is the responsibility of the done callback to ensure that this
4777 * particular zio is no longer discoverable for adoption, and as
4778 * such, cannot acquire any new parents.
4783 mutex_enter(&zio->io_lock);
4784 zio->io_state[ZIO_WAIT_DONE] = 1;
4785 mutex_exit(&zio->io_lock);
4788 * We are done executing this zio. We may want to execute a parent
4789 * next. See the comment in zio_notify_parent().
4791 zio_t *next_to_execute = NULL;
4793 for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) {
4794 zio_link_t *remove_zl = zl;
4795 pio_next = zio_walk_parents(zio, &zl);
4796 zio_remove_child(pio, zio, remove_zl);
4797 zio_notify_parent(pio, zio, ZIO_WAIT_DONE, &next_to_execute);
4800 if (zio->io_waiter != NULL) {
4801 mutex_enter(&zio->io_lock);
4802 zio->io_executor = NULL;
4803 cv_broadcast(&zio->io_cv);
4804 mutex_exit(&zio->io_lock);
4809 return (next_to_execute);
4813 * ==========================================================================
4814 * I/O pipeline definition
4815 * ==========================================================================
4817 static zio_pipe_stage_t *zio_pipeline[] = {
4825 zio_checksum_generate,
4841 zio_checksum_verify,
4849 * Compare two zbookmark_phys_t's to see which we would reach first in a
4850 * pre-order traversal of the object tree.
4852 * This is simple in every case aside from the meta-dnode object. For all other
4853 * objects, we traverse them in order (object 1 before object 2, and so on).
4854 * However, all of these objects are traversed while traversing object 0, since
4855 * the data it points to is the list of objects. Thus, we need to convert to a
4856 * canonical representation so we can compare meta-dnode bookmarks to
4857 * non-meta-dnode bookmarks.
4859 * We do this by calculating "equivalents" for each field of the zbookmark.
4860 * zbookmarks outside of the meta-dnode use their own object and level, and
4861 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
4862 * blocks this bookmark refers to) by multiplying their blkid by their span
4863 * (the number of L0 blocks contained within one block at their level).
4864 * zbookmarks inside the meta-dnode calculate their object equivalent
4865 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
4866 * level + 1<<31 (any value larger than a level could ever be) for their level.
4867 * This causes them to always compare before a bookmark in their object
4868 * equivalent, compare appropriately to bookmarks in other objects, and to
4869 * compare appropriately to other bookmarks in the meta-dnode.
4872 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
4873 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
4876 * These variables represent the "equivalent" values for the zbookmark,
4877 * after converting zbookmarks inside the meta dnode to their
4878 * normal-object equivalents.
4880 uint64_t zb1obj, zb2obj;
4881 uint64_t zb1L0, zb2L0;
4882 uint64_t zb1level, zb2level;
4884 if (zb1->zb_object == zb2->zb_object &&
4885 zb1->zb_level == zb2->zb_level &&
4886 zb1->zb_blkid == zb2->zb_blkid)
4889 IMPLY(zb1->zb_level > 0, ibs1 >= SPA_MINBLOCKSHIFT);
4890 IMPLY(zb2->zb_level > 0, ibs2 >= SPA_MINBLOCKSHIFT);
4893 * BP_SPANB calculates the span in blocks.
4895 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
4896 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
4898 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
4899 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4901 zb1level = zb1->zb_level + COMPARE_META_LEVEL;
4903 zb1obj = zb1->zb_object;
4904 zb1level = zb1->zb_level;
4907 if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
4908 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4910 zb2level = zb2->zb_level + COMPARE_META_LEVEL;
4912 zb2obj = zb2->zb_object;
4913 zb2level = zb2->zb_level;
4916 /* Now that we have a canonical representation, do the comparison. */
4917 if (zb1obj != zb2obj)
4918 return (zb1obj < zb2obj ? -1 : 1);
4919 else if (zb1L0 != zb2L0)
4920 return (zb1L0 < zb2L0 ? -1 : 1);
4921 else if (zb1level != zb2level)
4922 return (zb1level > zb2level ? -1 : 1);
4924 * This can (theoretically) happen if the bookmarks have the same object
4925 * and level, but different blkids, if the block sizes are not the same.
4926 * There is presently no way to change the indirect block sizes
4932 * This function checks the following: given that last_block is the place that
4933 * our traversal stopped last time, does that guarantee that we've visited
4934 * every node under subtree_root? Therefore, we can't just use the raw output
4935 * of zbookmark_compare. We have to pass in a modified version of
4936 * subtree_root; by incrementing the block id, and then checking whether
4937 * last_block is before or equal to that, we can tell whether or not having
4938 * visited last_block implies that all of subtree_root's children have been
4942 zbookmark_subtree_completed(const dnode_phys_t *dnp,
4943 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
4945 zbookmark_phys_t mod_zb = *subtree_root;
4947 ASSERT(last_block->zb_level == 0);
4949 /* The objset_phys_t isn't before anything. */
4954 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
4955 * data block size in sectors, because that variable is only used if
4956 * the bookmark refers to a block in the meta-dnode. Since we don't
4957 * know without examining it what object it refers to, and there's no
4958 * harm in passing in this value in other cases, we always pass it in.
4960 * We pass in 0 for the indirect block size shift because zb2 must be
4961 * level 0. The indirect block size is only used to calculate the span
4962 * of the bookmark, but since the bookmark must be level 0, the span is
4963 * always 1, so the math works out.
4965 * If you make changes to how the zbookmark_compare code works, be sure
4966 * to make sure that this code still works afterwards.
4968 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
4969 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,
4973 EXPORT_SYMBOL(zio_type_name);
4974 EXPORT_SYMBOL(zio_buf_alloc);
4975 EXPORT_SYMBOL(zio_data_buf_alloc);
4976 EXPORT_SYMBOL(zio_buf_free);
4977 EXPORT_SYMBOL(zio_data_buf_free);
4980 ZFS_MODULE_PARAM(zfs_zio, zio_, slow_io_ms, INT, ZMOD_RW,
4981 "Max I/O completion time (milliseconds) before marking it as slow");
4983 ZFS_MODULE_PARAM(zfs_zio, zio_, requeue_io_start_cut_in_line, INT, ZMOD_RW,
4984 "Prioritize requeued I/O");
4986 ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_deferred_free, INT, ZMOD_RW,
4987 "Defer frees starting in this pass");
4989 ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_dont_compress, INT, ZMOD_RW,
4990 "Don't compress starting in this pass");
4992 ZFS_MODULE_PARAM(zfs, zfs_, sync_pass_rewrite, INT, ZMOD_RW,
4993 "Rewrite new bps starting in this pass");
4995 ZFS_MODULE_PARAM(zfs_zio, zio_, dva_throttle_enabled, INT, ZMOD_RW,
4996 "Throttle block allocations in the ZIO pipeline");
4998 ZFS_MODULE_PARAM(zfs_zio, zio_, deadman_log_all, INT, ZMOD_RW,
4999 "Log all slow ZIOs, not just those with vdevs");