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, 2019 by Delphix. All rights reserved.
24 * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
25 * Copyright (c) 2017, Intel Corporation.
28 #include <sys/sysmacros.h>
29 #include <sys/zfs_context.h>
30 #include <sys/fm/fs/zfs.h>
33 #include <sys/spa_impl.h>
34 #include <sys/vdev_impl.h>
35 #include <sys/vdev_trim.h>
36 #include <sys/zio_impl.h>
37 #include <sys/zio_compress.h>
38 #include <sys/zio_checksum.h>
39 #include <sys/dmu_objset.h>
42 #include <sys/blkptr.h>
43 #include <sys/zfeature.h>
44 #include <sys/dsl_scan.h>
45 #include <sys/metaslab_impl.h>
47 #include <sys/trace_zio.h>
49 #include <sys/dsl_crypt.h>
50 #include <sys/cityhash.h>
53 * ==========================================================================
54 * I/O type descriptions
55 * ==========================================================================
57 const char *zio_type_name[ZIO_TYPES] = {
59 * Note: Linux kernel thread name length is limited
60 * so these names will differ from upstream open zfs.
62 "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl", "z_trim"
65 int zio_dva_throttle_enabled = B_TRUE;
66 int zio_deadman_log_all = B_FALSE;
69 * ==========================================================================
71 * ==========================================================================
73 kmem_cache_t *zio_cache;
74 kmem_cache_t *zio_link_cache;
75 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
76 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
77 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
78 uint64_t zio_buf_cache_allocs[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
79 uint64_t zio_buf_cache_frees[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
82 /* Mark IOs as "slow" if they take longer than 30 seconds */
83 int zio_slow_io_ms = (30 * MILLISEC);
85 #define BP_SPANB(indblkshift, level) \
86 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
87 #define COMPARE_META_LEVEL 0x80000000ul
89 * The following actions directly effect the spa's sync-to-convergence logic.
90 * The values below define the sync pass when we start performing the action.
91 * Care should be taken when changing these values as they directly impact
92 * spa_sync() performance. Tuning these values may introduce subtle performance
93 * pathologies and should only be done in the context of performance analysis.
94 * These tunables will eventually be removed and replaced with #defines once
95 * enough analysis has been done to determine optimal values.
97 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
98 * regular blocks are not deferred.
100 * Starting in sync pass 8 (zfs_sync_pass_dont_compress), we disable
101 * compression (including of metadata). In practice, we don't have this
102 * many sync passes, so this has no effect.
104 * The original intent was that disabling compression would help the sync
105 * passes to converge. However, in practice disabling compression increases
106 * the average number of sync passes, because when we turn compression off, a
107 * lot of block's size will change and thus we have to re-allocate (not
108 * overwrite) them. It also increases the number of 128KB allocations (e.g.
109 * for indirect blocks and spacemaps) because these will not be compressed.
110 * The 128K allocations are especially detrimental to performance on highly
111 * fragmented systems, which may have very few free segments of this size,
112 * and may need to load new metaslabs to satisfy 128K allocations.
114 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
115 int zfs_sync_pass_dont_compress = 8; /* don't compress starting in this pass */
116 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
119 * An allocating zio is one that either currently has the DVA allocate
120 * stage set or will have it later in its lifetime.
122 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
124 int zio_requeue_io_start_cut_in_line = 1;
127 int zio_buf_debug_limit = 16384;
129 int zio_buf_debug_limit = 0;
132 static inline void __zio_execute(zio_t *zio);
134 static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t);
140 vmem_t *data_alloc_arena = NULL;
142 zio_cache = kmem_cache_create("zio_cache",
143 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
144 zio_link_cache = kmem_cache_create("zio_link_cache",
145 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
148 * For small buffers, we want a cache for each multiple of
149 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
150 * for each quarter-power of 2.
152 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
153 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
156 size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
158 #if defined(_ILP32) && defined(_KERNEL)
160 * Cache size limited to 1M on 32-bit platforms until ARC
161 * buffers no longer require virtual address space.
163 if (size > zfs_max_recordsize)
172 * If we are using watchpoints, put each buffer on its own page,
173 * to eliminate the performance overhead of trapping to the
174 * kernel when modifying a non-watched buffer that shares the
175 * page with a watched buffer.
177 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
180 * Here's the problem - on 4K native devices in userland on
181 * Linux using O_DIRECT, buffers must be 4K aligned or I/O
182 * will fail with EINVAL, causing zdb (and others) to coredump.
183 * Since userland probably doesn't need optimized buffer caches,
184 * we just force 4K alignment on everything.
186 align = 8 * SPA_MINBLOCKSIZE;
188 if (size < PAGESIZE) {
189 align = SPA_MINBLOCKSIZE;
190 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
197 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
198 zio_buf_cache[c] = kmem_cache_create(name, size,
199 align, NULL, NULL, NULL, NULL, NULL, cflags);
201 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
202 zio_data_buf_cache[c] = kmem_cache_create(name, size,
203 align, NULL, NULL, NULL, NULL,
204 data_alloc_arena, cflags);
209 ASSERT(zio_buf_cache[c] != NULL);
210 if (zio_buf_cache[c - 1] == NULL)
211 zio_buf_cache[c - 1] = zio_buf_cache[c];
213 ASSERT(zio_data_buf_cache[c] != NULL);
214 if (zio_data_buf_cache[c - 1] == NULL)
215 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
227 kmem_cache_t *last_cache = NULL;
228 kmem_cache_t *last_data_cache = NULL;
230 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
233 * Cache size limited to 1M on 32-bit platforms until ARC
234 * buffers no longer require virtual address space.
236 if (((c + 1) << SPA_MINBLOCKSHIFT) > zfs_max_recordsize)
239 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
240 if (zio_buf_cache_allocs[c] != zio_buf_cache_frees[c])
241 (void) printf("zio_fini: [%d] %llu != %llu\n",
242 (int)((c + 1) << SPA_MINBLOCKSHIFT),
243 (long long unsigned)zio_buf_cache_allocs[c],
244 (long long unsigned)zio_buf_cache_frees[c]);
246 if (zio_buf_cache[c] != last_cache) {
247 last_cache = zio_buf_cache[c];
248 kmem_cache_destroy(zio_buf_cache[c]);
250 zio_buf_cache[c] = NULL;
252 if (zio_data_buf_cache[c] != last_data_cache) {
253 last_data_cache = zio_data_buf_cache[c];
254 kmem_cache_destroy(zio_data_buf_cache[c]);
256 zio_data_buf_cache[c] = NULL;
259 kmem_cache_destroy(zio_link_cache);
260 kmem_cache_destroy(zio_cache);
268 * ==========================================================================
269 * Allocate and free I/O buffers
270 * ==========================================================================
274 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
275 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
276 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
277 * excess / transient data in-core during a crashdump.
280 zio_buf_alloc(size_t size)
282 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
284 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
285 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
286 atomic_add_64(&zio_buf_cache_allocs[c], 1);
289 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
293 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
294 * crashdump if the kernel panics. This exists so that we will limit the amount
295 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
296 * of kernel heap dumped to disk when the kernel panics)
299 zio_data_buf_alloc(size_t size)
301 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
303 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
305 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
309 zio_buf_free(void *buf, size_t size)
311 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
313 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
314 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
315 atomic_add_64(&zio_buf_cache_frees[c], 1);
318 kmem_cache_free(zio_buf_cache[c], buf);
322 zio_data_buf_free(void *buf, size_t size)
324 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
326 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
328 kmem_cache_free(zio_data_buf_cache[c], buf);
332 zio_abd_free(void *abd, size_t size)
334 abd_free((abd_t *)abd);
338 * ==========================================================================
339 * Push and pop I/O transform buffers
340 * ==========================================================================
343 zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize,
344 zio_transform_func_t *transform)
346 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
348 zt->zt_orig_abd = zio->io_abd;
349 zt->zt_orig_size = zio->io_size;
350 zt->zt_bufsize = bufsize;
351 zt->zt_transform = transform;
353 zt->zt_next = zio->io_transform_stack;
354 zio->io_transform_stack = zt;
361 zio_pop_transforms(zio_t *zio)
365 while ((zt = zio->io_transform_stack) != NULL) {
366 if (zt->zt_transform != NULL)
367 zt->zt_transform(zio,
368 zt->zt_orig_abd, zt->zt_orig_size);
370 if (zt->zt_bufsize != 0)
371 abd_free(zio->io_abd);
373 zio->io_abd = zt->zt_orig_abd;
374 zio->io_size = zt->zt_orig_size;
375 zio->io_transform_stack = zt->zt_next;
377 kmem_free(zt, sizeof (zio_transform_t));
382 * ==========================================================================
383 * I/O transform callbacks for subblocks, decompression, and decryption
384 * ==========================================================================
387 zio_subblock(zio_t *zio, abd_t *data, uint64_t size)
389 ASSERT(zio->io_size > size);
391 if (zio->io_type == ZIO_TYPE_READ)
392 abd_copy(data, zio->io_abd, size);
396 zio_decompress(zio_t *zio, abd_t *data, uint64_t size)
398 if (zio->io_error == 0) {
399 void *tmp = abd_borrow_buf(data, size);
400 int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
401 zio->io_abd, tmp, zio->io_size, size);
402 abd_return_buf_copy(data, tmp, size);
404 if (zio_injection_enabled && ret == 0)
405 ret = zio_handle_fault_injection(zio, EINVAL);
408 zio->io_error = SET_ERROR(EIO);
413 zio_decrypt(zio_t *zio, abd_t *data, uint64_t size)
417 blkptr_t *bp = zio->io_bp;
418 spa_t *spa = zio->io_spa;
419 uint64_t dsobj = zio->io_bookmark.zb_objset;
420 uint64_t lsize = BP_GET_LSIZE(bp);
421 dmu_object_type_t ot = BP_GET_TYPE(bp);
422 uint8_t salt[ZIO_DATA_SALT_LEN];
423 uint8_t iv[ZIO_DATA_IV_LEN];
424 uint8_t mac[ZIO_DATA_MAC_LEN];
425 boolean_t no_crypt = B_FALSE;
427 ASSERT(BP_USES_CRYPT(bp));
428 ASSERT3U(size, !=, 0);
430 if (zio->io_error != 0)
434 * Verify the cksum of MACs stored in an indirect bp. It will always
435 * be possible to verify this since it does not require an encryption
438 if (BP_HAS_INDIRECT_MAC_CKSUM(bp)) {
439 zio_crypt_decode_mac_bp(bp, mac);
441 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF) {
443 * We haven't decompressed the data yet, but
444 * zio_crypt_do_indirect_mac_checksum() requires
445 * decompressed data to be able to parse out the MACs
446 * from the indirect block. We decompress it now and
447 * throw away the result after we are finished.
449 tmp = zio_buf_alloc(lsize);
450 ret = zio_decompress_data(BP_GET_COMPRESS(bp),
451 zio->io_abd, tmp, zio->io_size, lsize);
453 ret = SET_ERROR(EIO);
456 ret = zio_crypt_do_indirect_mac_checksum(B_FALSE,
457 tmp, lsize, BP_SHOULD_BYTESWAP(bp), mac);
458 zio_buf_free(tmp, lsize);
460 ret = zio_crypt_do_indirect_mac_checksum_abd(B_FALSE,
461 zio->io_abd, size, BP_SHOULD_BYTESWAP(bp), mac);
463 abd_copy(data, zio->io_abd, size);
465 if (zio_injection_enabled && ot != DMU_OT_DNODE && ret == 0) {
466 ret = zio_handle_decrypt_injection(spa,
467 &zio->io_bookmark, ot, ECKSUM);
476 * If this is an authenticated block, just check the MAC. It would be
477 * nice to separate this out into its own flag, but for the moment
478 * enum zio_flag is out of bits.
480 if (BP_IS_AUTHENTICATED(bp)) {
481 if (ot == DMU_OT_OBJSET) {
482 ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa,
483 dsobj, zio->io_abd, size, BP_SHOULD_BYTESWAP(bp));
485 zio_crypt_decode_mac_bp(bp, mac);
486 ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj,
487 zio->io_abd, size, mac);
488 if (zio_injection_enabled && ret == 0) {
489 ret = zio_handle_decrypt_injection(spa,
490 &zio->io_bookmark, ot, ECKSUM);
493 abd_copy(data, zio->io_abd, size);
501 zio_crypt_decode_params_bp(bp, salt, iv);
503 if (ot == DMU_OT_INTENT_LOG) {
504 tmp = abd_borrow_buf_copy(zio->io_abd, sizeof (zil_chain_t));
505 zio_crypt_decode_mac_zil(tmp, mac);
506 abd_return_buf(zio->io_abd, tmp, sizeof (zil_chain_t));
508 zio_crypt_decode_mac_bp(bp, mac);
511 ret = spa_do_crypt_abd(B_FALSE, spa, &zio->io_bookmark, BP_GET_TYPE(bp),
512 BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp), salt, iv, mac, size, data,
513 zio->io_abd, &no_crypt);
515 abd_copy(data, zio->io_abd, size);
523 /* assert that the key was found unless this was speculative */
524 ASSERT(ret != EACCES || (zio->io_flags & ZIO_FLAG_SPECULATIVE));
527 * If there was a decryption / authentication error return EIO as
528 * the io_error. If this was not a speculative zio, create an ereport.
531 zio->io_error = SET_ERROR(EIO);
532 if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) {
533 spa_log_error(spa, &zio->io_bookmark);
534 zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION,
535 spa, NULL, &zio->io_bookmark, zio, 0, 0);
543 * ==========================================================================
544 * I/O parent/child relationships and pipeline interlocks
545 * ==========================================================================
548 zio_walk_parents(zio_t *cio, zio_link_t **zl)
550 list_t *pl = &cio->io_parent_list;
552 *zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl);
556 ASSERT((*zl)->zl_child == cio);
557 return ((*zl)->zl_parent);
561 zio_walk_children(zio_t *pio, zio_link_t **zl)
563 list_t *cl = &pio->io_child_list;
565 ASSERT(MUTEX_HELD(&pio->io_lock));
567 *zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl);
571 ASSERT((*zl)->zl_parent == pio);
572 return ((*zl)->zl_child);
576 zio_unique_parent(zio_t *cio)
578 zio_link_t *zl = NULL;
579 zio_t *pio = zio_walk_parents(cio, &zl);
581 VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL);
586 zio_add_child(zio_t *pio, zio_t *cio)
588 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
591 * Logical I/Os can have logical, gang, or vdev children.
592 * Gang I/Os can have gang or vdev children.
593 * Vdev I/Os can only have vdev children.
594 * The following ASSERT captures all of these constraints.
596 ASSERT3S(cio->io_child_type, <=, pio->io_child_type);
601 mutex_enter(&pio->io_lock);
602 mutex_enter(&cio->io_lock);
604 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
606 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
607 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
609 list_insert_head(&pio->io_child_list, zl);
610 list_insert_head(&cio->io_parent_list, zl);
612 pio->io_child_count++;
613 cio->io_parent_count++;
615 mutex_exit(&cio->io_lock);
616 mutex_exit(&pio->io_lock);
620 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
622 ASSERT(zl->zl_parent == pio);
623 ASSERT(zl->zl_child == cio);
625 mutex_enter(&pio->io_lock);
626 mutex_enter(&cio->io_lock);
628 list_remove(&pio->io_child_list, zl);
629 list_remove(&cio->io_parent_list, zl);
631 pio->io_child_count--;
632 cio->io_parent_count--;
634 mutex_exit(&cio->io_lock);
635 mutex_exit(&pio->io_lock);
636 kmem_cache_free(zio_link_cache, zl);
640 zio_wait_for_children(zio_t *zio, uint8_t childbits, enum zio_wait_type wait)
642 boolean_t waiting = B_FALSE;
644 mutex_enter(&zio->io_lock);
645 ASSERT(zio->io_stall == NULL);
646 for (int c = 0; c < ZIO_CHILD_TYPES; c++) {
647 if (!(ZIO_CHILD_BIT_IS_SET(childbits, c)))
650 uint64_t *countp = &zio->io_children[c][wait];
653 ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN);
654 zio->io_stall = countp;
659 mutex_exit(&zio->io_lock);
663 __attribute__((always_inline))
665 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait,
666 zio_t **next_to_executep)
668 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
669 int *errorp = &pio->io_child_error[zio->io_child_type];
671 mutex_enter(&pio->io_lock);
672 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
673 *errorp = zio_worst_error(*errorp, zio->io_error);
674 pio->io_reexecute |= zio->io_reexecute;
675 ASSERT3U(*countp, >, 0);
679 if (*countp == 0 && pio->io_stall == countp) {
680 zio_taskq_type_t type =
681 pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE :
683 pio->io_stall = NULL;
684 mutex_exit(&pio->io_lock);
687 * If we can tell the caller to execute this parent next, do
688 * so. Otherwise dispatch the parent zio as its own task.
690 * Having the caller execute the parent when possible reduces
691 * locking on the zio taskq's, reduces context switch
692 * overhead, and has no recursion penalty. Note that one
693 * read from disk typically causes at least 3 zio's: a
694 * zio_null(), the logical zio_read(), and then a physical
695 * zio. When the physical ZIO completes, we are able to call
696 * zio_done() on all 3 of these zio's from one invocation of
697 * zio_execute() by returning the parent back to
698 * zio_execute(). Since the parent isn't executed until this
699 * thread returns back to zio_execute(), the caller should do
702 * In other cases, dispatching the parent prevents
703 * overflowing the stack when we have deeply nested
704 * parent-child relationships, as we do with the "mega zio"
705 * of writes for spa_sync(), and the chain of ZIL blocks.
707 if (next_to_executep != NULL && *next_to_executep == NULL) {
708 *next_to_executep = pio;
710 zio_taskq_dispatch(pio, type, B_FALSE);
713 mutex_exit(&pio->io_lock);
718 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
720 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
721 zio->io_error = zio->io_child_error[c];
725 zio_bookmark_compare(const void *x1, const void *x2)
727 const zio_t *z1 = x1;
728 const zio_t *z2 = x2;
730 if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset)
732 if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset)
735 if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object)
737 if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object)
740 if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level)
742 if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level)
745 if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid)
747 if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid)
759 * ==========================================================================
760 * Create the various types of I/O (read, write, free, etc)
761 * ==========================================================================
764 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
765 abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done,
766 void *private, zio_type_t type, zio_priority_t priority,
767 enum zio_flag flags, vdev_t *vd, uint64_t offset,
768 const zbookmark_phys_t *zb, enum zio_stage stage,
769 enum zio_stage pipeline)
773 IMPLY(type != ZIO_TYPE_TRIM, psize <= SPA_MAXBLOCKSIZE);
774 ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0);
775 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
777 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
778 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
779 ASSERT(vd || stage == ZIO_STAGE_OPEN);
781 IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW_COMPRESS) != 0);
783 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
784 bzero(zio, sizeof (zio_t));
786 mutex_init(&zio->io_lock, NULL, MUTEX_NOLOCKDEP, NULL);
787 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
789 list_create(&zio->io_parent_list, sizeof (zio_link_t),
790 offsetof(zio_link_t, zl_parent_node));
791 list_create(&zio->io_child_list, sizeof (zio_link_t),
792 offsetof(zio_link_t, zl_child_node));
793 metaslab_trace_init(&zio->io_alloc_list);
796 zio->io_child_type = ZIO_CHILD_VDEV;
797 else if (flags & ZIO_FLAG_GANG_CHILD)
798 zio->io_child_type = ZIO_CHILD_GANG;
799 else if (flags & ZIO_FLAG_DDT_CHILD)
800 zio->io_child_type = ZIO_CHILD_DDT;
802 zio->io_child_type = ZIO_CHILD_LOGICAL;
805 zio->io_bp = (blkptr_t *)bp;
806 zio->io_bp_copy = *bp;
807 zio->io_bp_orig = *bp;
808 if (type != ZIO_TYPE_WRITE ||
809 zio->io_child_type == ZIO_CHILD_DDT)
810 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
811 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
812 zio->io_logical = zio;
813 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
814 pipeline |= ZIO_GANG_STAGES;
820 zio->io_private = private;
822 zio->io_priority = priority;
824 zio->io_offset = offset;
825 zio->io_orig_abd = zio->io_abd = data;
826 zio->io_orig_size = zio->io_size = psize;
827 zio->io_lsize = lsize;
828 zio->io_orig_flags = zio->io_flags = flags;
829 zio->io_orig_stage = zio->io_stage = stage;
830 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
831 zio->io_pipeline_trace = ZIO_STAGE_OPEN;
833 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
834 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
837 zio->io_bookmark = *zb;
840 if (zio->io_metaslab_class == NULL)
841 zio->io_metaslab_class = pio->io_metaslab_class;
842 if (zio->io_logical == NULL)
843 zio->io_logical = pio->io_logical;
844 if (zio->io_child_type == ZIO_CHILD_GANG)
845 zio->io_gang_leader = pio->io_gang_leader;
846 zio_add_child(pio, zio);
849 taskq_init_ent(&zio->io_tqent);
855 zio_destroy(zio_t *zio)
857 metaslab_trace_fini(&zio->io_alloc_list);
858 list_destroy(&zio->io_parent_list);
859 list_destroy(&zio->io_child_list);
860 mutex_destroy(&zio->io_lock);
861 cv_destroy(&zio->io_cv);
862 kmem_cache_free(zio_cache, zio);
866 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
867 void *private, enum zio_flag flags)
871 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
872 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
873 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
879 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
881 return (zio_null(NULL, spa, NULL, done, private, flags));
885 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
887 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
888 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
889 bp, (longlong_t)BP_GET_TYPE(bp));
891 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
892 BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
893 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
894 bp, (longlong_t)BP_GET_CHECKSUM(bp));
896 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
897 BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
898 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
899 bp, (longlong_t)BP_GET_COMPRESS(bp));
901 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
902 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
903 bp, (longlong_t)BP_GET_LSIZE(bp));
905 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
906 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
907 bp, (longlong_t)BP_GET_PSIZE(bp));
910 if (BP_IS_EMBEDDED(bp)) {
911 if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
912 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
913 bp, (longlong_t)BPE_GET_ETYPE(bp));
918 * Do not verify individual DVAs if the config is not trusted. This
919 * will be done once the zio is executed in vdev_mirror_map_alloc.
921 if (!spa->spa_trust_config)
925 * Pool-specific checks.
927 * Note: it would be nice to verify that the blk_birth and
928 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
929 * allows the birth time of log blocks (and dmu_sync()-ed blocks
930 * that are in the log) to be arbitrarily large.
932 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
933 uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
935 if (vdevid >= spa->spa_root_vdev->vdev_children) {
936 zfs_panic_recover("blkptr at %p DVA %u has invalid "
938 bp, i, (longlong_t)vdevid);
941 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
943 zfs_panic_recover("blkptr at %p DVA %u has invalid "
945 bp, i, (longlong_t)vdevid);
948 if (vd->vdev_ops == &vdev_hole_ops) {
949 zfs_panic_recover("blkptr at %p DVA %u has hole "
951 bp, i, (longlong_t)vdevid);
954 if (vd->vdev_ops == &vdev_missing_ops) {
956 * "missing" vdevs are valid during import, but we
957 * don't have their detailed info (e.g. asize), so
958 * we can't perform any more checks on them.
962 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
963 uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
965 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
966 if (offset + asize > vd->vdev_asize) {
967 zfs_panic_recover("blkptr at %p DVA %u has invalid "
969 bp, i, (longlong_t)offset);
975 zfs_dva_valid(spa_t *spa, const dva_t *dva, const blkptr_t *bp)
977 uint64_t vdevid = DVA_GET_VDEV(dva);
979 if (vdevid >= spa->spa_root_vdev->vdev_children)
982 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
986 if (vd->vdev_ops == &vdev_hole_ops)
989 if (vd->vdev_ops == &vdev_missing_ops) {
993 uint64_t offset = DVA_GET_OFFSET(dva);
994 uint64_t asize = DVA_GET_ASIZE(dva);
997 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
998 if (offset + asize > vd->vdev_asize)
1005 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
1006 abd_t *data, uint64_t size, zio_done_func_t *done, void *private,
1007 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
1011 zfs_blkptr_verify(spa, bp);
1013 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
1014 data, size, size, done, private,
1015 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
1016 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
1017 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
1023 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
1024 abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp,
1025 zio_done_func_t *ready, zio_done_func_t *children_ready,
1026 zio_done_func_t *physdone, zio_done_func_t *done,
1027 void *private, zio_priority_t priority, enum zio_flag flags,
1028 const zbookmark_phys_t *zb)
1032 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
1033 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
1034 zp->zp_compress >= ZIO_COMPRESS_OFF &&
1035 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
1036 DMU_OT_IS_VALID(zp->zp_type) &&
1037 zp->zp_level < 32 &&
1038 zp->zp_copies > 0 &&
1039 zp->zp_copies <= spa_max_replication(spa));
1041 zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private,
1042 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
1043 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
1044 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
1046 zio->io_ready = ready;
1047 zio->io_children_ready = children_ready;
1048 zio->io_physdone = physdone;
1052 * Data can be NULL if we are going to call zio_write_override() to
1053 * provide the already-allocated BP. But we may need the data to
1054 * verify a dedup hit (if requested). In this case, don't try to
1055 * dedup (just take the already-allocated BP verbatim). Encrypted
1056 * dedup blocks need data as well so we also disable dedup in this
1060 (zio->io_prop.zp_dedup_verify || zio->io_prop.zp_encrypt)) {
1061 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
1068 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data,
1069 uint64_t size, zio_done_func_t *done, void *private,
1070 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
1074 zio = zio_create(pio, spa, txg, bp, data, size, size, done, private,
1075 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb,
1076 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
1082 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
1084 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
1085 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1086 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1087 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
1090 * We must reset the io_prop to match the values that existed
1091 * when the bp was first written by dmu_sync() keeping in mind
1092 * that nopwrite and dedup are mutually exclusive.
1094 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
1095 zio->io_prop.zp_nopwrite = nopwrite;
1096 zio->io_prop.zp_copies = copies;
1097 zio->io_bp_override = bp;
1101 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
1104 zfs_blkptr_verify(spa, bp);
1107 * The check for EMBEDDED is a performance optimization. We
1108 * process the free here (by ignoring it) rather than
1109 * putting it on the list and then processing it in zio_free_sync().
1111 if (BP_IS_EMBEDDED(bp))
1113 metaslab_check_free(spa, bp);
1116 * Frees that are for the currently-syncing txg, are not going to be
1117 * deferred, and which will not need to do a read (i.e. not GANG or
1118 * DEDUP), can be processed immediately. Otherwise, put them on the
1119 * in-memory list for later processing.
1121 if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
1122 txg != spa->spa_syncing_txg ||
1123 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
1124 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
1126 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 0)));
1131 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1132 enum zio_flag flags)
1135 enum zio_stage stage = ZIO_FREE_PIPELINE;
1137 ASSERT(!BP_IS_HOLE(bp));
1138 ASSERT(spa_syncing_txg(spa) == txg);
1139 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
1141 if (BP_IS_EMBEDDED(bp))
1142 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
1144 metaslab_check_free(spa, bp);
1146 dsl_scan_freed(spa, bp);
1149 * GANG and DEDUP blocks can induce a read (for the gang block header,
1150 * or the DDT), so issue them asynchronously so that this thread is
1153 if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
1154 stage |= ZIO_STAGE_ISSUE_ASYNC;
1156 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1157 BP_GET_PSIZE(bp), NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW,
1158 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
1164 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1165 zio_done_func_t *done, void *private, enum zio_flag flags)
1169 zfs_blkptr_verify(spa, bp);
1171 if (BP_IS_EMBEDDED(bp))
1172 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
1175 * A claim is an allocation of a specific block. Claims are needed
1176 * to support immediate writes in the intent log. The issue is that
1177 * immediate writes contain committed data, but in a txg that was
1178 * *not* committed. Upon opening the pool after an unclean shutdown,
1179 * the intent log claims all blocks that contain immediate write data
1180 * so that the SPA knows they're in use.
1182 * All claims *must* be resolved in the first txg -- before the SPA
1183 * starts allocating blocks -- so that nothing is allocated twice.
1184 * If txg == 0 we just verify that the block is claimable.
1186 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <,
1187 spa_min_claim_txg(spa));
1188 ASSERT(txg == spa_min_claim_txg(spa) || txg == 0);
1189 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
1191 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1192 BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW,
1193 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
1194 ASSERT0(zio->io_queued_timestamp);
1200 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
1201 zio_done_func_t *done, void *private, enum zio_flag flags)
1206 if (vd->vdev_children == 0) {
1207 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
1208 ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
1209 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
1213 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
1215 for (c = 0; c < vd->vdev_children; c++)
1216 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
1217 done, private, flags));
1224 zio_trim(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1225 zio_done_func_t *done, void *private, zio_priority_t priority,
1226 enum zio_flag flags, enum trim_flag trim_flags)
1230 ASSERT0(vd->vdev_children);
1231 ASSERT0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
1232 ASSERT0(P2PHASE(size, 1ULL << vd->vdev_ashift));
1233 ASSERT3U(size, !=, 0);
1235 zio = zio_create(pio, vd->vdev_spa, 0, NULL, NULL, size, size, done,
1236 private, ZIO_TYPE_TRIM, priority, flags | ZIO_FLAG_PHYSICAL,
1237 vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_TRIM_PIPELINE);
1238 zio->io_trim_flags = trim_flags;
1244 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1245 abd_t *data, int checksum, zio_done_func_t *done, void *private,
1246 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1250 ASSERT(vd->vdev_children == 0);
1251 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1252 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1253 ASSERT3U(offset + size, <=, vd->vdev_psize);
1255 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1256 private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1257 offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
1259 zio->io_prop.zp_checksum = checksum;
1265 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1266 abd_t *data, int checksum, zio_done_func_t *done, void *private,
1267 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1271 ASSERT(vd->vdev_children == 0);
1272 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1273 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1274 ASSERT3U(offset + size, <=, vd->vdev_psize);
1276 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1277 private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1278 offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
1280 zio->io_prop.zp_checksum = checksum;
1282 if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
1284 * zec checksums are necessarily destructive -- they modify
1285 * the end of the write buffer to hold the verifier/checksum.
1286 * Therefore, we must make a local copy in case the data is
1287 * being written to multiple places in parallel.
1289 abd_t *wbuf = abd_alloc_sametype(data, size);
1290 abd_copy(wbuf, data, size);
1292 zio_push_transform(zio, wbuf, size, size, NULL);
1299 * Create a child I/O to do some work for us.
1302 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1303 abd_t *data, uint64_t size, int type, zio_priority_t priority,
1304 enum zio_flag flags, zio_done_func_t *done, void *private)
1306 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1310 * vdev child I/Os do not propagate their error to the parent.
1311 * Therefore, for correct operation the caller *must* check for
1312 * and handle the error in the child i/o's done callback.
1313 * The only exceptions are i/os that we don't care about
1314 * (OPTIONAL or REPAIR).
1316 ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) ||
1319 if (type == ZIO_TYPE_READ && bp != NULL) {
1321 * If we have the bp, then the child should perform the
1322 * checksum and the parent need not. This pushes error
1323 * detection as close to the leaves as possible and
1324 * eliminates redundant checksums in the interior nodes.
1326 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1327 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1330 if (vd->vdev_ops->vdev_op_leaf) {
1331 ASSERT0(vd->vdev_children);
1332 offset += VDEV_LABEL_START_SIZE;
1335 flags |= ZIO_VDEV_CHILD_FLAGS(pio);
1338 * If we've decided to do a repair, the write is not speculative --
1339 * even if the original read was.
1341 if (flags & ZIO_FLAG_IO_REPAIR)
1342 flags &= ~ZIO_FLAG_SPECULATIVE;
1345 * If we're creating a child I/O that is not associated with a
1346 * top-level vdev, then the child zio is not an allocating I/O.
1347 * If this is a retried I/O then we ignore it since we will
1348 * have already processed the original allocating I/O.
1350 if (flags & ZIO_FLAG_IO_ALLOCATING &&
1351 (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) {
1352 ASSERT(pio->io_metaslab_class != NULL);
1353 ASSERT(pio->io_metaslab_class->mc_alloc_throttle_enabled);
1354 ASSERT(type == ZIO_TYPE_WRITE);
1355 ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE);
1356 ASSERT(!(flags & ZIO_FLAG_IO_REPAIR));
1357 ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) ||
1358 pio->io_child_type == ZIO_CHILD_GANG);
1360 flags &= ~ZIO_FLAG_IO_ALLOCATING;
1364 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size,
1365 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1366 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1367 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
1369 zio->io_physdone = pio->io_physdone;
1370 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1371 zio->io_logical->io_phys_children++;
1377 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size,
1378 zio_type_t type, zio_priority_t priority, enum zio_flag flags,
1379 zio_done_func_t *done, void *private)
1383 ASSERT(vd->vdev_ops->vdev_op_leaf);
1385 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1386 data, size, size, done, private, type, priority,
1387 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1389 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1395 zio_flush(zio_t *zio, vdev_t *vd)
1397 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
1399 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1403 zio_shrink(zio_t *zio, uint64_t size)
1405 ASSERT3P(zio->io_executor, ==, NULL);
1406 ASSERT3U(zio->io_orig_size, ==, zio->io_size);
1407 ASSERT3U(size, <=, zio->io_size);
1410 * We don't shrink for raidz because of problems with the
1411 * reconstruction when reading back less than the block size.
1412 * Note, BP_IS_RAIDZ() assumes no compression.
1414 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1415 if (!BP_IS_RAIDZ(zio->io_bp)) {
1416 /* we are not doing a raw write */
1417 ASSERT3U(zio->io_size, ==, zio->io_lsize);
1418 zio->io_orig_size = zio->io_size = zio->io_lsize = size;
1423 * ==========================================================================
1424 * Prepare to read and write logical blocks
1425 * ==========================================================================
1429 zio_read_bp_init(zio_t *zio)
1431 blkptr_t *bp = zio->io_bp;
1433 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1435 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1437 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1438 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1439 !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) {
1440 zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1441 psize, psize, zio_decompress);
1444 if (((BP_IS_PROTECTED(bp) && !(zio->io_flags & ZIO_FLAG_RAW_ENCRYPT)) ||
1445 BP_HAS_INDIRECT_MAC_CKSUM(bp)) &&
1446 zio->io_child_type == ZIO_CHILD_LOGICAL) {
1447 zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1448 psize, psize, zio_decrypt);
1451 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1452 int psize = BPE_GET_PSIZE(bp);
1453 void *data = abd_borrow_buf(zio->io_abd, psize);
1455 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1456 decode_embedded_bp_compressed(bp, data);
1457 abd_return_buf_copy(zio->io_abd, data, psize);
1459 ASSERT(!BP_IS_EMBEDDED(bp));
1460 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1463 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1464 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1466 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1467 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1469 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1470 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1476 zio_write_bp_init(zio_t *zio)
1478 if (!IO_IS_ALLOCATING(zio))
1481 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1483 if (zio->io_bp_override) {
1484 blkptr_t *bp = zio->io_bp;
1485 zio_prop_t *zp = &zio->io_prop;
1487 ASSERT(bp->blk_birth != zio->io_txg);
1488 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1490 *bp = *zio->io_bp_override;
1491 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1493 if (BP_IS_EMBEDDED(bp))
1497 * If we've been overridden and nopwrite is set then
1498 * set the flag accordingly to indicate that a nopwrite
1499 * has already occurred.
1501 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1502 ASSERT(!zp->zp_dedup);
1503 ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum);
1504 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1508 ASSERT(!zp->zp_nopwrite);
1510 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1513 ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags &
1514 ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify);
1516 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum &&
1518 BP_SET_DEDUP(bp, 1);
1519 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1524 * We were unable to handle this as an override bp, treat
1525 * it as a regular write I/O.
1527 zio->io_bp_override = NULL;
1528 *bp = zio->io_bp_orig;
1529 zio->io_pipeline = zio->io_orig_pipeline;
1536 zio_write_compress(zio_t *zio)
1538 spa_t *spa = zio->io_spa;
1539 zio_prop_t *zp = &zio->io_prop;
1540 enum zio_compress compress = zp->zp_compress;
1541 blkptr_t *bp = zio->io_bp;
1542 uint64_t lsize = zio->io_lsize;
1543 uint64_t psize = zio->io_size;
1547 * If our children haven't all reached the ready stage,
1548 * wait for them and then repeat this pipeline stage.
1550 if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT |
1551 ZIO_CHILD_GANG_BIT, ZIO_WAIT_READY)) {
1555 if (!IO_IS_ALLOCATING(zio))
1558 if (zio->io_children_ready != NULL) {
1560 * Now that all our children are ready, run the callback
1561 * associated with this zio in case it wants to modify the
1562 * data to be written.
1564 ASSERT3U(zp->zp_level, >, 0);
1565 zio->io_children_ready(zio);
1568 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1569 ASSERT(zio->io_bp_override == NULL);
1571 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1573 * We're rewriting an existing block, which means we're
1574 * working on behalf of spa_sync(). For spa_sync() to
1575 * converge, it must eventually be the case that we don't
1576 * have to allocate new blocks. But compression changes
1577 * the blocksize, which forces a reallocate, and makes
1578 * convergence take longer. Therefore, after the first
1579 * few passes, stop compressing to ensure convergence.
1581 pass = spa_sync_pass(spa);
1583 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1584 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1585 ASSERT(!BP_GET_DEDUP(bp));
1587 if (pass >= zfs_sync_pass_dont_compress)
1588 compress = ZIO_COMPRESS_OFF;
1590 /* Make sure someone doesn't change their mind on overwrites */
1591 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1592 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1595 /* If it's a compressed write that is not raw, compress the buffer. */
1596 if (compress != ZIO_COMPRESS_OFF &&
1597 !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) {
1598 void *cbuf = zio_buf_alloc(lsize);
1599 psize = zio_compress_data(compress, zio->io_abd, cbuf, lsize);
1600 if (psize == 0 || psize == lsize) {
1601 compress = ZIO_COMPRESS_OFF;
1602 zio_buf_free(cbuf, lsize);
1603 } else if (!zp->zp_dedup && !zp->zp_encrypt &&
1604 psize <= BPE_PAYLOAD_SIZE &&
1605 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1606 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1607 encode_embedded_bp_compressed(bp,
1608 cbuf, compress, lsize, psize);
1609 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1610 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1611 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1612 zio_buf_free(cbuf, lsize);
1613 bp->blk_birth = zio->io_txg;
1614 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1615 ASSERT(spa_feature_is_active(spa,
1616 SPA_FEATURE_EMBEDDED_DATA));
1620 * Round up compressed size up to the ashift
1621 * of the smallest-ashift device, and zero the tail.
1622 * This ensures that the compressed size of the BP
1623 * (and thus compressratio property) are correct,
1624 * in that we charge for the padding used to fill out
1627 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
1628 size_t rounded = (size_t)P2ROUNDUP(psize,
1629 1ULL << spa->spa_min_ashift);
1630 if (rounded >= lsize) {
1631 compress = ZIO_COMPRESS_OFF;
1632 zio_buf_free(cbuf, lsize);
1635 abd_t *cdata = abd_get_from_buf(cbuf, lsize);
1636 abd_take_ownership_of_buf(cdata, B_TRUE);
1637 abd_zero_off(cdata, psize, rounded - psize);
1639 zio_push_transform(zio, cdata,
1640 psize, lsize, NULL);
1645 * We were unable to handle this as an override bp, treat
1646 * it as a regular write I/O.
1648 zio->io_bp_override = NULL;
1649 *bp = zio->io_bp_orig;
1650 zio->io_pipeline = zio->io_orig_pipeline;
1652 } else if ((zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) != 0 &&
1653 zp->zp_type == DMU_OT_DNODE) {
1655 * The DMU actually relies on the zio layer's compression
1656 * to free metadnode blocks that have had all contained
1657 * dnodes freed. As a result, even when doing a raw
1658 * receive, we must check whether the block can be compressed
1661 psize = zio_compress_data(ZIO_COMPRESS_EMPTY,
1662 zio->io_abd, NULL, lsize);
1664 compress = ZIO_COMPRESS_OFF;
1666 ASSERT3U(psize, !=, 0);
1670 * The final pass of spa_sync() must be all rewrites, but the first
1671 * few passes offer a trade-off: allocating blocks defers convergence,
1672 * but newly allocated blocks are sequential, so they can be written
1673 * to disk faster. Therefore, we allow the first few passes of
1674 * spa_sync() to allocate new blocks, but force rewrites after that.
1675 * There should only be a handful of blocks after pass 1 in any case.
1677 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1678 BP_GET_PSIZE(bp) == psize &&
1679 pass >= zfs_sync_pass_rewrite) {
1680 VERIFY3U(psize, !=, 0);
1681 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1683 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1684 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1687 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1691 if (zio->io_bp_orig.blk_birth != 0 &&
1692 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1693 BP_SET_LSIZE(bp, lsize);
1694 BP_SET_TYPE(bp, zp->zp_type);
1695 BP_SET_LEVEL(bp, zp->zp_level);
1696 BP_SET_BIRTH(bp, zio->io_txg, 0);
1698 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1700 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1701 BP_SET_LSIZE(bp, lsize);
1702 BP_SET_TYPE(bp, zp->zp_type);
1703 BP_SET_LEVEL(bp, zp->zp_level);
1704 BP_SET_PSIZE(bp, psize);
1705 BP_SET_COMPRESS(bp, compress);
1706 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1707 BP_SET_DEDUP(bp, zp->zp_dedup);
1708 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1710 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1711 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1712 ASSERT(!zp->zp_encrypt ||
1713 DMU_OT_IS_ENCRYPTED(zp->zp_type));
1714 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1716 if (zp->zp_nopwrite) {
1717 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1718 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1719 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1726 zio_free_bp_init(zio_t *zio)
1728 blkptr_t *bp = zio->io_bp;
1730 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1731 if (BP_GET_DEDUP(bp))
1732 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1735 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1741 * ==========================================================================
1742 * Execute the I/O pipeline
1743 * ==========================================================================
1747 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1749 spa_t *spa = zio->io_spa;
1750 zio_type_t t = zio->io_type;
1751 int flags = (cutinline ? TQ_FRONT : 0);
1754 * If we're a config writer or a probe, the normal issue and
1755 * interrupt threads may all be blocked waiting for the config lock.
1756 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1758 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1762 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1764 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1768 * If this is a high priority I/O, then use the high priority taskq if
1771 if ((zio->io_priority == ZIO_PRIORITY_NOW ||
1772 zio->io_priority == ZIO_PRIORITY_SYNC_WRITE) &&
1773 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1776 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1779 * NB: We are assuming that the zio can only be dispatched
1780 * to a single taskq at a time. It would be a grievous error
1781 * to dispatch the zio to another taskq at the same time.
1783 ASSERT(taskq_empty_ent(&zio->io_tqent));
1784 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1785 flags, &zio->io_tqent);
1789 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1791 kthread_t *executor = zio->io_executor;
1792 spa_t *spa = zio->io_spa;
1794 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1795 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1797 for (i = 0; i < tqs->stqs_count; i++) {
1798 if (taskq_member(tqs->stqs_taskq[i], executor))
1807 zio_issue_async(zio_t *zio)
1809 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1815 zio_interrupt(zio_t *zio)
1817 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1821 zio_delay_interrupt(zio_t *zio)
1824 * The timeout_generic() function isn't defined in userspace, so
1825 * rather than trying to implement the function, the zio delay
1826 * functionality has been disabled for userspace builds.
1831 * If io_target_timestamp is zero, then no delay has been registered
1832 * for this IO, thus jump to the end of this function and "skip" the
1833 * delay; issuing it directly to the zio layer.
1835 if (zio->io_target_timestamp != 0) {
1836 hrtime_t now = gethrtime();
1838 if (now >= zio->io_target_timestamp) {
1840 * This IO has already taken longer than the target
1841 * delay to complete, so we don't want to delay it
1842 * any longer; we "miss" the delay and issue it
1843 * directly to the zio layer. This is likely due to
1844 * the target latency being set to a value less than
1845 * the underlying hardware can satisfy (e.g. delay
1846 * set to 1ms, but the disks take 10ms to complete an
1850 DTRACE_PROBE2(zio__delay__miss, zio_t *, zio,
1856 hrtime_t diff = zio->io_target_timestamp - now;
1857 clock_t expire_at_tick = ddi_get_lbolt() +
1860 DTRACE_PROBE3(zio__delay__hit, zio_t *, zio,
1861 hrtime_t, now, hrtime_t, diff);
1863 if (NSEC_TO_TICK(diff) == 0) {
1864 /* Our delay is less than a jiffy - just spin */
1865 zfs_sleep_until(zio->io_target_timestamp);
1869 * Use taskq_dispatch_delay() in the place of
1870 * OpenZFS's timeout_generic().
1872 tid = taskq_dispatch_delay(system_taskq,
1873 (task_func_t *)zio_interrupt,
1874 zio, TQ_NOSLEEP, expire_at_tick);
1875 if (tid == TASKQID_INVALID) {
1877 * Couldn't allocate a task. Just
1878 * finish the zio without a delay.
1887 DTRACE_PROBE1(zio__delay__skip, zio_t *, zio);
1892 zio_deadman_impl(zio_t *pio, int ziodepth)
1894 zio_t *cio, *cio_next;
1895 zio_link_t *zl = NULL;
1896 vdev_t *vd = pio->io_vd;
1898 if (zio_deadman_log_all || (vd != NULL && vd->vdev_ops->vdev_op_leaf)) {
1899 vdev_queue_t *vq = vd ? &vd->vdev_queue : NULL;
1900 zbookmark_phys_t *zb = &pio->io_bookmark;
1901 uint64_t delta = gethrtime() - pio->io_timestamp;
1902 uint64_t failmode = spa_get_deadman_failmode(pio->io_spa);
1904 zfs_dbgmsg("slow zio[%d]: zio=%px timestamp=%llu "
1905 "delta=%llu queued=%llu io=%llu "
1906 "path=%s last=%llu "
1907 "type=%d priority=%d flags=0x%x "
1908 "stage=0x%x pipeline=0x%x pipeline-trace=0x%x "
1909 "objset=%llu object=%llu level=%llu blkid=%llu "
1910 "offset=%llu size=%llu error=%d",
1911 ziodepth, pio, pio->io_timestamp,
1912 delta, pio->io_delta, pio->io_delay,
1913 vd ? vd->vdev_path : "NULL", vq ? vq->vq_io_complete_ts : 0,
1914 pio->io_type, pio->io_priority, pio->io_flags,
1915 pio->io_stage, pio->io_pipeline, pio->io_pipeline_trace,
1916 zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
1917 pio->io_offset, pio->io_size, pio->io_error);
1918 zfs_ereport_post(FM_EREPORT_ZFS_DEADMAN,
1919 pio->io_spa, vd, zb, pio, 0, 0);
1921 if (failmode == ZIO_FAILURE_MODE_CONTINUE &&
1922 taskq_empty_ent(&pio->io_tqent)) {
1927 mutex_enter(&pio->io_lock);
1928 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
1929 cio_next = zio_walk_children(pio, &zl);
1930 zio_deadman_impl(cio, ziodepth + 1);
1932 mutex_exit(&pio->io_lock);
1936 * Log the critical information describing this zio and all of its children
1937 * using the zfs_dbgmsg() interface then post deadman event for the ZED.
1940 zio_deadman(zio_t *pio, char *tag)
1942 spa_t *spa = pio->io_spa;
1943 char *name = spa_name(spa);
1945 if (!zfs_deadman_enabled || spa_suspended(spa))
1948 zio_deadman_impl(pio, 0);
1950 switch (spa_get_deadman_failmode(spa)) {
1951 case ZIO_FAILURE_MODE_WAIT:
1952 zfs_dbgmsg("%s waiting for hung I/O to pool '%s'", tag, name);
1955 case ZIO_FAILURE_MODE_CONTINUE:
1956 zfs_dbgmsg("%s restarting hung I/O for pool '%s'", tag, name);
1959 case ZIO_FAILURE_MODE_PANIC:
1960 fm_panic("%s determined I/O to pool '%s' is hung.", tag, name);
1966 * Execute the I/O pipeline until one of the following occurs:
1967 * (1) the I/O completes; (2) the pipeline stalls waiting for
1968 * dependent child I/Os; (3) the I/O issues, so we're waiting
1969 * for an I/O completion interrupt; (4) the I/O is delegated by
1970 * vdev-level caching or aggregation; (5) the I/O is deferred
1971 * due to vdev-level queueing; (6) the I/O is handed off to
1972 * another thread. In all cases, the pipeline stops whenever
1973 * there's no CPU work; it never burns a thread in cv_wait_io().
1975 * There's no locking on io_stage because there's no legitimate way
1976 * for multiple threads to be attempting to process the same I/O.
1978 static zio_pipe_stage_t *zio_pipeline[];
1981 * zio_execute() is a wrapper around the static function
1982 * __zio_execute() so that we can force __zio_execute() to be
1983 * inlined. This reduces stack overhead which is important
1984 * because __zio_execute() is called recursively in several zio
1985 * code paths. zio_execute() itself cannot be inlined because
1986 * it is externally visible.
1989 zio_execute(zio_t *zio)
1991 fstrans_cookie_t cookie;
1993 cookie = spl_fstrans_mark();
1995 spl_fstrans_unmark(cookie);
1999 * Used to determine if in the current context the stack is sized large
2000 * enough to allow zio_execute() to be called recursively. A minimum
2001 * stack size of 16K is required to avoid needing to re-dispatch the zio.
2004 zio_execute_stack_check(zio_t *zio)
2006 #if !defined(HAVE_LARGE_STACKS)
2007 dsl_pool_t *dp = spa_get_dsl(zio->io_spa);
2009 /* Executing in txg_sync_thread() context. */
2010 if (dp && curthread == dp->dp_tx.tx_sync_thread)
2013 /* Pool initialization outside of zio_taskq context. */
2014 if (dp && spa_is_initializing(dp->dp_spa) &&
2015 !zio_taskq_member(zio, ZIO_TASKQ_ISSUE) &&
2016 !zio_taskq_member(zio, ZIO_TASKQ_ISSUE_HIGH))
2018 #endif /* HAVE_LARGE_STACKS */
2023 __attribute__((always_inline))
2025 __zio_execute(zio_t *zio)
2027 ASSERT3U(zio->io_queued_timestamp, >, 0);
2029 while (zio->io_stage < ZIO_STAGE_DONE) {
2030 enum zio_stage pipeline = zio->io_pipeline;
2031 enum zio_stage stage = zio->io_stage;
2033 zio->io_executor = curthread;
2035 ASSERT(!MUTEX_HELD(&zio->io_lock));
2036 ASSERT(ISP2(stage));
2037 ASSERT(zio->io_stall == NULL);
2041 } while ((stage & pipeline) == 0);
2043 ASSERT(stage <= ZIO_STAGE_DONE);
2046 * If we are in interrupt context and this pipeline stage
2047 * will grab a config lock that is held across I/O,
2048 * or may wait for an I/O that needs an interrupt thread
2049 * to complete, issue async to avoid deadlock.
2051 * For VDEV_IO_START, we cut in line so that the io will
2052 * be sent to disk promptly.
2054 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
2055 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
2056 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
2057 zio_requeue_io_start_cut_in_line : B_FALSE;
2058 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
2063 * If the current context doesn't have large enough stacks
2064 * the zio must be issued asynchronously to prevent overflow.
2066 if (zio_execute_stack_check(zio)) {
2067 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
2068 zio_requeue_io_start_cut_in_line : B_FALSE;
2069 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
2073 zio->io_stage = stage;
2074 zio->io_pipeline_trace |= zio->io_stage;
2077 * The zio pipeline stage returns the next zio to execute
2078 * (typically the same as this one), or NULL if we should
2081 zio = zio_pipeline[highbit64(stage) - 1](zio);
2090 * ==========================================================================
2091 * Initiate I/O, either sync or async
2092 * ==========================================================================
2095 zio_wait(zio_t *zio)
2097 long timeout = MSEC_TO_TICK(zfs_deadman_ziotime_ms);
2100 ASSERT3S(zio->io_stage, ==, ZIO_STAGE_OPEN);
2101 ASSERT3P(zio->io_executor, ==, NULL);
2103 zio->io_waiter = curthread;
2104 ASSERT0(zio->io_queued_timestamp);
2105 zio->io_queued_timestamp = gethrtime();
2109 mutex_enter(&zio->io_lock);
2110 while (zio->io_executor != NULL) {
2111 error = cv_timedwait_io(&zio->io_cv, &zio->io_lock,
2112 ddi_get_lbolt() + timeout);
2114 if (zfs_deadman_enabled && error == -1 &&
2115 gethrtime() - zio->io_queued_timestamp >
2116 spa_deadman_ziotime(zio->io_spa)) {
2117 mutex_exit(&zio->io_lock);
2118 timeout = MSEC_TO_TICK(zfs_deadman_checktime_ms);
2119 zio_deadman(zio, FTAG);
2120 mutex_enter(&zio->io_lock);
2123 mutex_exit(&zio->io_lock);
2125 error = zio->io_error;
2132 zio_nowait(zio_t *zio)
2134 ASSERT3P(zio->io_executor, ==, NULL);
2136 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
2137 zio_unique_parent(zio) == NULL) {
2141 * This is a logical async I/O with no parent to wait for it.
2142 * We add it to the spa_async_root_zio "Godfather" I/O which
2143 * will ensure they complete prior to unloading the pool.
2145 spa_t *spa = zio->io_spa;
2147 pio = spa->spa_async_zio_root[CPU_SEQID];
2150 zio_add_child(pio, zio);
2153 ASSERT0(zio->io_queued_timestamp);
2154 zio->io_queued_timestamp = gethrtime();
2159 * ==========================================================================
2160 * Reexecute, cancel, or suspend/resume failed I/O
2161 * ==========================================================================
2165 zio_reexecute(zio_t *pio)
2167 zio_t *cio, *cio_next;
2169 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
2170 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
2171 ASSERT(pio->io_gang_leader == NULL);
2172 ASSERT(pio->io_gang_tree == NULL);
2174 pio->io_flags = pio->io_orig_flags;
2175 pio->io_stage = pio->io_orig_stage;
2176 pio->io_pipeline = pio->io_orig_pipeline;
2177 pio->io_reexecute = 0;
2178 pio->io_flags |= ZIO_FLAG_REEXECUTED;
2179 pio->io_pipeline_trace = 0;
2181 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2182 pio->io_state[w] = 0;
2183 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2184 pio->io_child_error[c] = 0;
2186 if (IO_IS_ALLOCATING(pio))
2187 BP_ZERO(pio->io_bp);
2190 * As we reexecute pio's children, new children could be created.
2191 * New children go to the head of pio's io_child_list, however,
2192 * so we will (correctly) not reexecute them. The key is that
2193 * the remainder of pio's io_child_list, from 'cio_next' onward,
2194 * cannot be affected by any side effects of reexecuting 'cio'.
2196 zio_link_t *zl = NULL;
2197 mutex_enter(&pio->io_lock);
2198 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
2199 cio_next = zio_walk_children(pio, &zl);
2200 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2201 pio->io_children[cio->io_child_type][w]++;
2202 mutex_exit(&pio->io_lock);
2204 mutex_enter(&pio->io_lock);
2206 mutex_exit(&pio->io_lock);
2209 * Now that all children have been reexecuted, execute the parent.
2210 * We don't reexecute "The Godfather" I/O here as it's the
2211 * responsibility of the caller to wait on it.
2213 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) {
2214 pio->io_queued_timestamp = gethrtime();
2220 zio_suspend(spa_t *spa, zio_t *zio, zio_suspend_reason_t reason)
2222 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
2223 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
2224 "failure and the failure mode property for this pool "
2225 "is set to panic.", spa_name(spa));
2227 cmn_err(CE_WARN, "Pool '%s' has encountered an uncorrectable I/O "
2228 "failure and has been suspended.\n", spa_name(spa));
2230 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL,
2233 mutex_enter(&spa->spa_suspend_lock);
2235 if (spa->spa_suspend_zio_root == NULL)
2236 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
2237 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
2238 ZIO_FLAG_GODFATHER);
2240 spa->spa_suspended = reason;
2243 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2244 ASSERT(zio != spa->spa_suspend_zio_root);
2245 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2246 ASSERT(zio_unique_parent(zio) == NULL);
2247 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
2248 zio_add_child(spa->spa_suspend_zio_root, zio);
2251 mutex_exit(&spa->spa_suspend_lock);
2255 zio_resume(spa_t *spa)
2260 * Reexecute all previously suspended i/o.
2262 mutex_enter(&spa->spa_suspend_lock);
2263 spa->spa_suspended = ZIO_SUSPEND_NONE;
2264 cv_broadcast(&spa->spa_suspend_cv);
2265 pio = spa->spa_suspend_zio_root;
2266 spa->spa_suspend_zio_root = NULL;
2267 mutex_exit(&spa->spa_suspend_lock);
2273 return (zio_wait(pio));
2277 zio_resume_wait(spa_t *spa)
2279 mutex_enter(&spa->spa_suspend_lock);
2280 while (spa_suspended(spa))
2281 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
2282 mutex_exit(&spa->spa_suspend_lock);
2286 * ==========================================================================
2289 * A gang block is a collection of small blocks that looks to the DMU
2290 * like one large block. When zio_dva_allocate() cannot find a block
2291 * of the requested size, due to either severe fragmentation or the pool
2292 * being nearly full, it calls zio_write_gang_block() to construct the
2293 * block from smaller fragments.
2295 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
2296 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
2297 * an indirect block: it's an array of block pointers. It consumes
2298 * only one sector and hence is allocatable regardless of fragmentation.
2299 * The gang header's bps point to its gang members, which hold the data.
2301 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
2302 * as the verifier to ensure uniqueness of the SHA256 checksum.
2303 * Critically, the gang block bp's blk_cksum is the checksum of the data,
2304 * not the gang header. This ensures that data block signatures (needed for
2305 * deduplication) are independent of how the block is physically stored.
2307 * Gang blocks can be nested: a gang member may itself be a gang block.
2308 * Thus every gang block is a tree in which root and all interior nodes are
2309 * gang headers, and the leaves are normal blocks that contain user data.
2310 * The root of the gang tree is called the gang leader.
2312 * To perform any operation (read, rewrite, free, claim) on a gang block,
2313 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
2314 * in the io_gang_tree field of the original logical i/o by recursively
2315 * reading the gang leader and all gang headers below it. This yields
2316 * an in-core tree containing the contents of every gang header and the
2317 * bps for every constituent of the gang block.
2319 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
2320 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
2321 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
2322 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
2323 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
2324 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
2325 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
2326 * of the gang header plus zio_checksum_compute() of the data to update the
2327 * gang header's blk_cksum as described above.
2329 * The two-phase assemble/issue model solves the problem of partial failure --
2330 * what if you'd freed part of a gang block but then couldn't read the
2331 * gang header for another part? Assembling the entire gang tree first
2332 * ensures that all the necessary gang header I/O has succeeded before
2333 * starting the actual work of free, claim, or write. Once the gang tree
2334 * is assembled, free and claim are in-memory operations that cannot fail.
2336 * In the event that a gang write fails, zio_dva_unallocate() walks the
2337 * gang tree to immediately free (i.e. insert back into the space map)
2338 * everything we've allocated. This ensures that we don't get ENOSPC
2339 * errors during repeated suspend/resume cycles due to a flaky device.
2341 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
2342 * the gang tree, we won't modify the block, so we can safely defer the free
2343 * (knowing that the block is still intact). If we *can* assemble the gang
2344 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
2345 * each constituent bp and we can allocate a new block on the next sync pass.
2347 * In all cases, the gang tree allows complete recovery from partial failure.
2348 * ==========================================================================
2352 zio_gang_issue_func_done(zio_t *zio)
2354 abd_put(zio->io_abd);
2358 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2364 return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset),
2365 BP_GET_PSIZE(bp), zio_gang_issue_func_done,
2366 NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2367 &pio->io_bookmark));
2371 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2378 abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2379 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2380 gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL,
2381 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2384 * As we rewrite each gang header, the pipeline will compute
2385 * a new gang block header checksum for it; but no one will
2386 * compute a new data checksum, so we do that here. The one
2387 * exception is the gang leader: the pipeline already computed
2388 * its data checksum because that stage precedes gang assembly.
2389 * (Presently, nothing actually uses interior data checksums;
2390 * this is just good hygiene.)
2392 if (gn != pio->io_gang_leader->io_gang_tree) {
2393 abd_t *buf = abd_get_offset(data, offset);
2395 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
2396 buf, BP_GET_PSIZE(bp));
2401 * If we are here to damage data for testing purposes,
2402 * leave the GBH alone so that we can detect the damage.
2404 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
2405 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2407 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2408 abd_get_offset(data, offset), BP_GET_PSIZE(bp),
2409 zio_gang_issue_func_done, NULL, pio->io_priority,
2410 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2418 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2421 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
2422 ZIO_GANG_CHILD_FLAGS(pio)));
2427 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2430 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
2431 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
2434 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
2443 static void zio_gang_tree_assemble_done(zio_t *zio);
2445 static zio_gang_node_t *
2446 zio_gang_node_alloc(zio_gang_node_t **gnpp)
2448 zio_gang_node_t *gn;
2450 ASSERT(*gnpp == NULL);
2452 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
2453 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
2460 zio_gang_node_free(zio_gang_node_t **gnpp)
2462 zio_gang_node_t *gn = *gnpp;
2464 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2465 ASSERT(gn->gn_child[g] == NULL);
2467 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2468 kmem_free(gn, sizeof (*gn));
2473 zio_gang_tree_free(zio_gang_node_t **gnpp)
2475 zio_gang_node_t *gn = *gnpp;
2480 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2481 zio_gang_tree_free(&gn->gn_child[g]);
2483 zio_gang_node_free(gnpp);
2487 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
2489 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
2490 abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2492 ASSERT(gio->io_gang_leader == gio);
2493 ASSERT(BP_IS_GANG(bp));
2495 zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2496 zio_gang_tree_assemble_done, gn, gio->io_priority,
2497 ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
2501 zio_gang_tree_assemble_done(zio_t *zio)
2503 zio_t *gio = zio->io_gang_leader;
2504 zio_gang_node_t *gn = zio->io_private;
2505 blkptr_t *bp = zio->io_bp;
2507 ASSERT(gio == zio_unique_parent(zio));
2508 ASSERT(zio->io_child_count == 0);
2513 /* this ABD was created from a linear buf in zio_gang_tree_assemble */
2514 if (BP_SHOULD_BYTESWAP(bp))
2515 byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size);
2517 ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh);
2518 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
2519 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2521 abd_put(zio->io_abd);
2523 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2524 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2525 if (!BP_IS_GANG(gbp))
2527 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
2532 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data,
2535 zio_t *gio = pio->io_gang_leader;
2538 ASSERT(BP_IS_GANG(bp) == !!gn);
2539 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
2540 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
2543 * If you're a gang header, your data is in gn->gn_gbh.
2544 * If you're a gang member, your data is in 'data' and gn == NULL.
2546 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset);
2549 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2551 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2552 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2553 if (BP_IS_HOLE(gbp))
2555 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data,
2557 offset += BP_GET_PSIZE(gbp);
2561 if (gn == gio->io_gang_tree)
2562 ASSERT3U(gio->io_size, ==, offset);
2569 zio_gang_assemble(zio_t *zio)
2571 blkptr_t *bp = zio->io_bp;
2573 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
2574 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2576 zio->io_gang_leader = zio;
2578 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
2584 zio_gang_issue(zio_t *zio)
2586 blkptr_t *bp = zio->io_bp;
2588 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT, ZIO_WAIT_DONE)) {
2592 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
2593 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2595 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
2596 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd,
2599 zio_gang_tree_free(&zio->io_gang_tree);
2601 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2607 zio_write_gang_member_ready(zio_t *zio)
2609 zio_t *pio = zio_unique_parent(zio);
2610 dva_t *cdva = zio->io_bp->blk_dva;
2611 dva_t *pdva = pio->io_bp->blk_dva;
2613 ASSERTV(zio_t *gio = zio->io_gang_leader);
2615 if (BP_IS_HOLE(zio->io_bp))
2618 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
2620 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
2621 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
2622 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2623 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
2624 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
2626 mutex_enter(&pio->io_lock);
2627 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
2628 ASSERT(DVA_GET_GANG(&pdva[d]));
2629 asize = DVA_GET_ASIZE(&pdva[d]);
2630 asize += DVA_GET_ASIZE(&cdva[d]);
2631 DVA_SET_ASIZE(&pdva[d], asize);
2633 mutex_exit(&pio->io_lock);
2637 zio_write_gang_done(zio_t *zio)
2640 * The io_abd field will be NULL for a zio with no data. The io_flags
2641 * will initially have the ZIO_FLAG_NODATA bit flag set, but we can't
2642 * check for it here as it is cleared in zio_ready.
2644 if (zio->io_abd != NULL)
2645 abd_put(zio->io_abd);
2649 zio_write_gang_block(zio_t *pio)
2651 spa_t *spa = pio->io_spa;
2652 metaslab_class_t *mc = spa_normal_class(spa);
2653 blkptr_t *bp = pio->io_bp;
2654 zio_t *gio = pio->io_gang_leader;
2656 zio_gang_node_t *gn, **gnpp;
2657 zio_gbh_phys_t *gbh;
2659 uint64_t txg = pio->io_txg;
2660 uint64_t resid = pio->io_size;
2662 int copies = gio->io_prop.zp_copies;
2666 boolean_t has_data = !(pio->io_flags & ZIO_FLAG_NODATA);
2669 * encrypted blocks need DVA[2] free so encrypted gang headers can't
2670 * have a third copy.
2672 gbh_copies = MIN(copies + 1, spa_max_replication(spa));
2673 if (gio->io_prop.zp_encrypt && gbh_copies >= SPA_DVAS_PER_BP)
2674 gbh_copies = SPA_DVAS_PER_BP - 1;
2676 int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER;
2677 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2678 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2681 flags |= METASLAB_ASYNC_ALLOC;
2682 VERIFY(zfs_refcount_held(&mc->mc_alloc_slots[pio->io_allocator],
2686 * The logical zio has already placed a reservation for
2687 * 'copies' allocation slots but gang blocks may require
2688 * additional copies. These additional copies
2689 * (i.e. gbh_copies - copies) are guaranteed to succeed
2690 * since metaslab_class_throttle_reserve() always allows
2691 * additional reservations for gang blocks.
2693 VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies,
2694 pio->io_allocator, pio, flags));
2697 error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE,
2698 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags,
2699 &pio->io_alloc_list, pio, pio->io_allocator);
2701 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2702 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2706 * If we failed to allocate the gang block header then
2707 * we remove any additional allocation reservations that
2708 * we placed here. The original reservation will
2709 * be removed when the logical I/O goes to the ready
2712 metaslab_class_throttle_unreserve(mc,
2713 gbh_copies - copies, pio->io_allocator, pio);
2716 pio->io_error = error;
2721 gnpp = &gio->io_gang_tree;
2723 gnpp = pio->io_private;
2724 ASSERT(pio->io_ready == zio_write_gang_member_ready);
2727 gn = zio_gang_node_alloc(gnpp);
2729 bzero(gbh, SPA_GANGBLOCKSIZE);
2730 gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE);
2733 * Create the gang header.
2735 zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2736 zio_write_gang_done, NULL, pio->io_priority,
2737 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2740 * Create and nowait the gang children.
2742 for (int g = 0; resid != 0; resid -= lsize, g++) {
2743 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2745 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2747 zp.zp_checksum = gio->io_prop.zp_checksum;
2748 zp.zp_compress = ZIO_COMPRESS_OFF;
2749 zp.zp_type = DMU_OT_NONE;
2751 zp.zp_copies = gio->io_prop.zp_copies;
2752 zp.zp_dedup = B_FALSE;
2753 zp.zp_dedup_verify = B_FALSE;
2754 zp.zp_nopwrite = B_FALSE;
2755 zp.zp_encrypt = gio->io_prop.zp_encrypt;
2756 zp.zp_byteorder = gio->io_prop.zp_byteorder;
2757 bzero(zp.zp_salt, ZIO_DATA_SALT_LEN);
2758 bzero(zp.zp_iv, ZIO_DATA_IV_LEN);
2759 bzero(zp.zp_mac, ZIO_DATA_MAC_LEN);
2761 zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2762 has_data ? abd_get_offset(pio->io_abd, pio->io_size -
2763 resid) : NULL, lsize, lsize, &zp,
2764 zio_write_gang_member_ready, NULL, NULL,
2765 zio_write_gang_done, &gn->gn_child[g], pio->io_priority,
2766 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2768 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2769 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2773 * Gang children won't throttle but we should
2774 * account for their work, so reserve an allocation
2775 * slot for them here.
2777 VERIFY(metaslab_class_throttle_reserve(mc,
2778 zp.zp_copies, cio->io_allocator, cio, flags));
2784 * Set pio's pipeline to just wait for zio to finish.
2786 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2789 * We didn't allocate this bp, so make sure it doesn't get unmarked.
2791 pio->io_flags &= ~ZIO_FLAG_FASTWRITE;
2799 * The zio_nop_write stage in the pipeline determines if allocating a
2800 * new bp is necessary. The nopwrite feature can handle writes in
2801 * either syncing or open context (i.e. zil writes) and as a result is
2802 * mutually exclusive with dedup.
2804 * By leveraging a cryptographically secure checksum, such as SHA256, we
2805 * can compare the checksums of the new data and the old to determine if
2806 * allocating a new block is required. Note that our requirements for
2807 * cryptographic strength are fairly weak: there can't be any accidental
2808 * hash collisions, but we don't need to be secure against intentional
2809 * (malicious) collisions. To trigger a nopwrite, you have to be able
2810 * to write the file to begin with, and triggering an incorrect (hash
2811 * collision) nopwrite is no worse than simply writing to the file.
2812 * That said, there are no known attacks against the checksum algorithms
2813 * used for nopwrite, assuming that the salt and the checksums
2814 * themselves remain secret.
2817 zio_nop_write(zio_t *zio)
2819 blkptr_t *bp = zio->io_bp;
2820 blkptr_t *bp_orig = &zio->io_bp_orig;
2821 zio_prop_t *zp = &zio->io_prop;
2823 ASSERT(BP_GET_LEVEL(bp) == 0);
2824 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2825 ASSERT(zp->zp_nopwrite);
2826 ASSERT(!zp->zp_dedup);
2827 ASSERT(zio->io_bp_override == NULL);
2828 ASSERT(IO_IS_ALLOCATING(zio));
2831 * Check to see if the original bp and the new bp have matching
2832 * characteristics (i.e. same checksum, compression algorithms, etc).
2833 * If they don't then just continue with the pipeline which will
2834 * allocate a new bp.
2836 if (BP_IS_HOLE(bp_orig) ||
2837 !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
2838 ZCHECKSUM_FLAG_NOPWRITE) ||
2839 BP_IS_ENCRYPTED(bp) || BP_IS_ENCRYPTED(bp_orig) ||
2840 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2841 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2842 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2843 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2847 * If the checksums match then reset the pipeline so that we
2848 * avoid allocating a new bp and issuing any I/O.
2850 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2851 ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
2852 ZCHECKSUM_FLAG_NOPWRITE);
2853 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2854 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2855 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2856 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2857 sizeof (uint64_t)) == 0);
2860 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2861 zio->io_flags |= ZIO_FLAG_NOPWRITE;
2868 * ==========================================================================
2870 * ==========================================================================
2873 zio_ddt_child_read_done(zio_t *zio)
2875 blkptr_t *bp = zio->io_bp;
2876 ddt_entry_t *dde = zio->io_private;
2878 zio_t *pio = zio_unique_parent(zio);
2880 mutex_enter(&pio->io_lock);
2881 ddp = ddt_phys_select(dde, bp);
2882 if (zio->io_error == 0)
2883 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2885 if (zio->io_error == 0 && dde->dde_repair_abd == NULL)
2886 dde->dde_repair_abd = zio->io_abd;
2888 abd_free(zio->io_abd);
2889 mutex_exit(&pio->io_lock);
2893 zio_ddt_read_start(zio_t *zio)
2895 blkptr_t *bp = zio->io_bp;
2897 ASSERT(BP_GET_DEDUP(bp));
2898 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2899 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2901 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2902 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2903 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2904 ddt_phys_t *ddp = dde->dde_phys;
2905 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2908 ASSERT(zio->io_vsd == NULL);
2911 if (ddp_self == NULL)
2914 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2915 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2917 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2919 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2920 abd_alloc_for_io(zio->io_size, B_TRUE),
2921 zio->io_size, zio_ddt_child_read_done, dde,
2922 zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) |
2923 ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark));
2928 zio_nowait(zio_read(zio, zio->io_spa, bp,
2929 zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority,
2930 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2936 zio_ddt_read_done(zio_t *zio)
2938 blkptr_t *bp = zio->io_bp;
2940 if (zio_wait_for_children(zio, ZIO_CHILD_DDT_BIT, ZIO_WAIT_DONE)) {
2944 ASSERT(BP_GET_DEDUP(bp));
2945 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2946 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2948 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2949 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2950 ddt_entry_t *dde = zio->io_vsd;
2952 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2956 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2957 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2960 if (dde->dde_repair_abd != NULL) {
2961 abd_copy(zio->io_abd, dde->dde_repair_abd,
2963 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2965 ddt_repair_done(ddt, dde);
2969 ASSERT(zio->io_vsd == NULL);
2975 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2977 spa_t *spa = zio->io_spa;
2978 boolean_t do_raw = !!(zio->io_flags & ZIO_FLAG_RAW);
2980 ASSERT(!(zio->io_bp_override && do_raw));
2983 * Note: we compare the original data, not the transformed data,
2984 * because when zio->io_bp is an override bp, we will not have
2985 * pushed the I/O transforms. That's an important optimization
2986 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2987 * However, we should never get a raw, override zio so in these
2988 * cases we can compare the io_abd directly. This is useful because
2989 * it allows us to do dedup verification even if we don't have access
2990 * to the original data (for instance, if the encryption keys aren't
2994 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2995 zio_t *lio = dde->dde_lead_zio[p];
2997 if (lio != NULL && do_raw) {
2998 return (lio->io_size != zio->io_size ||
2999 abd_cmp(zio->io_abd, lio->io_abd) != 0);
3000 } else if (lio != NULL) {
3001 return (lio->io_orig_size != zio->io_orig_size ||
3002 abd_cmp(zio->io_orig_abd, lio->io_orig_abd) != 0);
3006 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
3007 ddt_phys_t *ddp = &dde->dde_phys[p];
3009 if (ddp->ddp_phys_birth != 0 && do_raw) {
3010 blkptr_t blk = *zio->io_bp;
3015 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
3016 psize = BP_GET_PSIZE(&blk);
3018 if (psize != zio->io_size)
3023 tmpabd = abd_alloc_for_io(psize, B_TRUE);
3025 error = zio_wait(zio_read(NULL, spa, &blk, tmpabd,
3026 psize, NULL, NULL, ZIO_PRIORITY_SYNC_READ,
3027 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
3028 ZIO_FLAG_RAW, &zio->io_bookmark));
3031 if (abd_cmp(tmpabd, zio->io_abd) != 0)
3032 error = SET_ERROR(ENOENT);
3037 return (error != 0);
3038 } else if (ddp->ddp_phys_birth != 0) {
3039 arc_buf_t *abuf = NULL;
3040 arc_flags_t aflags = ARC_FLAG_WAIT;
3041 blkptr_t blk = *zio->io_bp;
3044 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
3046 if (BP_GET_LSIZE(&blk) != zio->io_orig_size)
3051 error = arc_read(NULL, spa, &blk,
3052 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
3053 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
3054 &aflags, &zio->io_bookmark);
3057 if (abd_cmp_buf(zio->io_orig_abd, abuf->b_data,
3058 zio->io_orig_size) != 0)
3059 error = SET_ERROR(ENOENT);
3060 arc_buf_destroy(abuf, &abuf);
3064 return (error != 0);
3072 zio_ddt_child_write_ready(zio_t *zio)
3074 int p = zio->io_prop.zp_copies;
3075 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
3076 ddt_entry_t *dde = zio->io_private;
3077 ddt_phys_t *ddp = &dde->dde_phys[p];
3085 ASSERT(dde->dde_lead_zio[p] == zio);
3087 ddt_phys_fill(ddp, zio->io_bp);
3089 zio_link_t *zl = NULL;
3090 while ((pio = zio_walk_parents(zio, &zl)) != NULL)
3091 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
3097 zio_ddt_child_write_done(zio_t *zio)
3099 int p = zio->io_prop.zp_copies;
3100 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
3101 ddt_entry_t *dde = zio->io_private;
3102 ddt_phys_t *ddp = &dde->dde_phys[p];
3106 ASSERT(ddp->ddp_refcnt == 0);
3107 ASSERT(dde->dde_lead_zio[p] == zio);
3108 dde->dde_lead_zio[p] = NULL;
3110 if (zio->io_error == 0) {
3111 zio_link_t *zl = NULL;
3112 while (zio_walk_parents(zio, &zl) != NULL)
3113 ddt_phys_addref(ddp);
3115 ddt_phys_clear(ddp);
3122 zio_ddt_ditto_write_done(zio_t *zio)
3124 int p = DDT_PHYS_DITTO;
3125 ASSERTV(zio_prop_t *zp = &zio->io_prop);
3126 blkptr_t *bp = zio->io_bp;
3127 ddt_t *ddt = ddt_select(zio->io_spa, bp);
3128 ddt_entry_t *dde = zio->io_private;
3129 ddt_phys_t *ddp = &dde->dde_phys[p];
3130 ddt_key_t *ddk = &dde->dde_key;
3134 ASSERT(ddp->ddp_refcnt == 0);
3135 ASSERT(dde->dde_lead_zio[p] == zio);
3136 dde->dde_lead_zio[p] = NULL;
3138 if (zio->io_error == 0) {
3139 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
3140 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
3141 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
3142 if (ddp->ddp_phys_birth != 0)
3143 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
3144 ddt_phys_fill(ddp, bp);
3151 zio_ddt_write(zio_t *zio)
3153 spa_t *spa = zio->io_spa;
3154 blkptr_t *bp = zio->io_bp;
3155 uint64_t txg = zio->io_txg;
3156 zio_prop_t *zp = &zio->io_prop;
3157 int p = zp->zp_copies;
3161 ddt_t *ddt = ddt_select(spa, bp);
3165 ASSERT(BP_GET_DEDUP(bp));
3166 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
3167 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
3168 ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW)));
3171 dde = ddt_lookup(ddt, bp, B_TRUE);
3172 ddp = &dde->dde_phys[p];
3174 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
3176 * If we're using a weak checksum, upgrade to a strong checksum
3177 * and try again. If we're already using a strong checksum,
3178 * we can't resolve it, so just convert to an ordinary write.
3179 * (And automatically e-mail a paper to Nature?)
3181 if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
3182 ZCHECKSUM_FLAG_DEDUP)) {
3183 zp->zp_checksum = spa_dedup_checksum(spa);
3184 zio_pop_transforms(zio);
3185 zio->io_stage = ZIO_STAGE_OPEN;
3188 zp->zp_dedup = B_FALSE;
3190 zio->io_pipeline = ZIO_WRITE_PIPELINE;
3195 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
3196 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
3198 if (ditto_copies > ddt_ditto_copies_present(dde) &&
3199 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
3200 zio_prop_t czp = *zp;
3202 czp.zp_copies = ditto_copies;
3205 * If we arrived here with an override bp, we won't have run
3206 * the transform stack, so we won't have the data we need to
3207 * generate a child i/o. So, toss the override bp and restart.
3208 * This is safe, because using the override bp is just an
3209 * optimization; and it's rare, so the cost doesn't matter.
3211 if (zio->io_bp_override) {
3212 zio_pop_transforms(zio);
3213 zio->io_stage = ZIO_STAGE_OPEN;
3214 zio->io_pipeline = ZIO_WRITE_PIPELINE;
3215 zio->io_bp_override = NULL;
3221 dio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
3222 zio->io_orig_size, zio->io_orig_size, &czp, NULL, NULL,
3223 NULL, zio_ddt_ditto_write_done, dde, zio->io_priority,
3224 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
3226 zio_push_transform(dio, zio->io_abd, zio->io_size, 0, NULL);
3227 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
3230 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
3231 if (ddp->ddp_phys_birth != 0)
3232 ddt_bp_fill(ddp, bp, txg);
3233 if (dde->dde_lead_zio[p] != NULL)
3234 zio_add_child(zio, dde->dde_lead_zio[p]);
3236 ddt_phys_addref(ddp);
3237 } else if (zio->io_bp_override) {
3238 ASSERT(bp->blk_birth == txg);
3239 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
3240 ddt_phys_fill(ddp, bp);
3241 ddt_phys_addref(ddp);
3243 cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
3244 zio->io_orig_size, zio->io_orig_size, zp,
3245 zio_ddt_child_write_ready, NULL, NULL,
3246 zio_ddt_child_write_done, dde, zio->io_priority,
3247 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
3249 zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL);
3250 dde->dde_lead_zio[p] = cio;
3263 ddt_entry_t *freedde; /* for debugging */
3266 zio_ddt_free(zio_t *zio)
3268 spa_t *spa = zio->io_spa;
3269 blkptr_t *bp = zio->io_bp;
3270 ddt_t *ddt = ddt_select(spa, bp);
3274 ASSERT(BP_GET_DEDUP(bp));
3275 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3278 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
3280 ddp = ddt_phys_select(dde, bp);
3282 ddt_phys_decref(ddp);
3290 * ==========================================================================
3291 * Allocate and free blocks
3292 * ==========================================================================
3296 zio_io_to_allocate(spa_t *spa, int allocator)
3300 ASSERT(MUTEX_HELD(&spa->spa_alloc_locks[allocator]));
3302 zio = avl_first(&spa->spa_alloc_trees[allocator]);
3306 ASSERT(IO_IS_ALLOCATING(zio));
3309 * Try to place a reservation for this zio. If we're unable to
3310 * reserve then we throttle.
3312 ASSERT3U(zio->io_allocator, ==, allocator);
3313 if (!metaslab_class_throttle_reserve(zio->io_metaslab_class,
3314 zio->io_prop.zp_copies, zio->io_allocator, zio, 0)) {
3318 avl_remove(&spa->spa_alloc_trees[allocator], zio);
3319 ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE);
3325 zio_dva_throttle(zio_t *zio)
3327 spa_t *spa = zio->io_spa;
3329 metaslab_class_t *mc;
3331 /* locate an appropriate allocation class */
3332 mc = spa_preferred_class(spa, zio->io_size, zio->io_prop.zp_type,
3333 zio->io_prop.zp_level, zio->io_prop.zp_zpl_smallblk);
3335 if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE ||
3336 !mc->mc_alloc_throttle_enabled ||
3337 zio->io_child_type == ZIO_CHILD_GANG ||
3338 zio->io_flags & ZIO_FLAG_NODATA) {
3342 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3344 ASSERT3U(zio->io_queued_timestamp, >, 0);
3345 ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE);
3347 zbookmark_phys_t *bm = &zio->io_bookmark;
3349 * We want to try to use as many allocators as possible to help improve
3350 * performance, but we also want logically adjacent IOs to be physically
3351 * adjacent to improve sequential read performance. We chunk each object
3352 * into 2^20 block regions, and then hash based on the objset, object,
3353 * level, and region to accomplish both of these goals.
3355 zio->io_allocator = cityhash4(bm->zb_objset, bm->zb_object,
3356 bm->zb_level, bm->zb_blkid >> 20) % spa->spa_alloc_count;
3357 mutex_enter(&spa->spa_alloc_locks[zio->io_allocator]);
3358 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3359 zio->io_metaslab_class = mc;
3360 avl_add(&spa->spa_alloc_trees[zio->io_allocator], zio);
3361 nio = zio_io_to_allocate(spa, zio->io_allocator);
3362 mutex_exit(&spa->spa_alloc_locks[zio->io_allocator]);
3367 zio_allocate_dispatch(spa_t *spa, int allocator)
3371 mutex_enter(&spa->spa_alloc_locks[allocator]);
3372 zio = zio_io_to_allocate(spa, allocator);
3373 mutex_exit(&spa->spa_alloc_locks[allocator]);
3377 ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE);
3378 ASSERT0(zio->io_error);
3379 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE);
3383 zio_dva_allocate(zio_t *zio)
3385 spa_t *spa = zio->io_spa;
3386 metaslab_class_t *mc;
3387 blkptr_t *bp = zio->io_bp;
3391 if (zio->io_gang_leader == NULL) {
3392 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3393 zio->io_gang_leader = zio;
3396 ASSERT(BP_IS_HOLE(bp));
3397 ASSERT0(BP_GET_NDVAS(bp));
3398 ASSERT3U(zio->io_prop.zp_copies, >, 0);
3399 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
3400 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
3402 flags |= (zio->io_flags & ZIO_FLAG_FASTWRITE) ? METASLAB_FASTWRITE : 0;
3403 if (zio->io_flags & ZIO_FLAG_NODATA)
3404 flags |= METASLAB_DONT_THROTTLE;
3405 if (zio->io_flags & ZIO_FLAG_GANG_CHILD)
3406 flags |= METASLAB_GANG_CHILD;
3407 if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE)
3408 flags |= METASLAB_ASYNC_ALLOC;
3411 * if not already chosen, locate an appropriate allocation class
3413 mc = zio->io_metaslab_class;
3415 mc = spa_preferred_class(spa, zio->io_size,
3416 zio->io_prop.zp_type, zio->io_prop.zp_level,
3417 zio->io_prop.zp_zpl_smallblk);
3418 zio->io_metaslab_class = mc;
3421 error = metaslab_alloc(spa, mc, zio->io_size, bp,
3422 zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
3423 &zio->io_alloc_list, zio, zio->io_allocator);
3426 * Fallback to normal class when an alloc class is full
3428 if (error == ENOSPC && mc != spa_normal_class(spa)) {
3430 * If throttling, transfer reservation over to normal class.
3431 * The io_allocator slot can remain the same even though we
3432 * are switching classes.
3434 if (mc->mc_alloc_throttle_enabled &&
3435 (zio->io_flags & ZIO_FLAG_IO_ALLOCATING)) {
3436 metaslab_class_throttle_unreserve(mc,
3437 zio->io_prop.zp_copies, zio->io_allocator, zio);
3438 zio->io_flags &= ~ZIO_FLAG_IO_ALLOCATING;
3440 mc = spa_normal_class(spa);
3441 VERIFY(metaslab_class_throttle_reserve(mc,
3442 zio->io_prop.zp_copies, zio->io_allocator, zio,
3443 flags | METASLAB_MUST_RESERVE));
3445 mc = spa_normal_class(spa);
3447 zio->io_metaslab_class = mc;
3449 error = metaslab_alloc(spa, mc, zio->io_size, bp,
3450 zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
3451 &zio->io_alloc_list, zio, zio->io_allocator);
3455 zfs_dbgmsg("%s: metaslab allocation failure: zio %px, "
3456 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
3458 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
3459 return (zio_write_gang_block(zio));
3460 zio->io_error = error;
3467 zio_dva_free(zio_t *zio)
3469 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
3475 zio_dva_claim(zio_t *zio)
3479 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
3481 zio->io_error = error;
3487 * Undo an allocation. This is used by zio_done() when an I/O fails
3488 * and we want to give back the block we just allocated.
3489 * This handles both normal blocks and gang blocks.
3492 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
3494 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
3495 ASSERT(zio->io_bp_override == NULL);
3497 if (!BP_IS_HOLE(bp))
3498 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
3501 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
3502 zio_dva_unallocate(zio, gn->gn_child[g],
3503 &gn->gn_gbh->zg_blkptr[g]);
3509 * Try to allocate an intent log block. Return 0 on success, errno on failure.
3512 zio_alloc_zil(spa_t *spa, objset_t *os, uint64_t txg, blkptr_t *new_bp,
3513 uint64_t size, boolean_t *slog)
3516 zio_alloc_list_t io_alloc_list;
3518 ASSERT(txg > spa_syncing_txg(spa));
3520 metaslab_trace_init(&io_alloc_list);
3523 * Block pointer fields are useful to metaslabs for stats and debugging.
3524 * Fill in the obvious ones before calling into metaslab_alloc().
3526 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3527 BP_SET_PSIZE(new_bp, size);
3528 BP_SET_LEVEL(new_bp, 0);
3531 * When allocating a zil block, we don't have information about
3532 * the final destination of the block except the objset it's part
3533 * of, so we just hash the objset ID to pick the allocator to get
3536 error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1,
3537 txg, NULL, METASLAB_FASTWRITE, &io_alloc_list, NULL,
3538 cityhash4(0, 0, 0, os->os_dsl_dataset->ds_object) %
3539 spa->spa_alloc_count);
3543 error = metaslab_alloc(spa, spa_normal_class(spa), size,
3544 new_bp, 1, txg, NULL, METASLAB_FASTWRITE,
3545 &io_alloc_list, NULL, cityhash4(0, 0, 0,
3546 os->os_dsl_dataset->ds_object) % spa->spa_alloc_count);
3550 metaslab_trace_fini(&io_alloc_list);
3553 BP_SET_LSIZE(new_bp, size);
3554 BP_SET_PSIZE(new_bp, size);
3555 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
3556 BP_SET_CHECKSUM(new_bp,
3557 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
3558 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
3559 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3560 BP_SET_LEVEL(new_bp, 0);
3561 BP_SET_DEDUP(new_bp, 0);
3562 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
3565 * encrypted blocks will require an IV and salt. We generate
3566 * these now since we will not be rewriting the bp at
3569 if (os->os_encrypted) {
3570 uint8_t iv[ZIO_DATA_IV_LEN];
3571 uint8_t salt[ZIO_DATA_SALT_LEN];
3573 BP_SET_CRYPT(new_bp, B_TRUE);
3574 VERIFY0(spa_crypt_get_salt(spa,
3575 dmu_objset_id(os), salt));
3576 VERIFY0(zio_crypt_generate_iv(iv));
3578 zio_crypt_encode_params_bp(new_bp, salt, iv);
3581 zfs_dbgmsg("%s: zil block allocation failure: "
3582 "size %llu, error %d", spa_name(spa), size, error);
3589 * ==========================================================================
3590 * Read and write to physical devices
3591 * ==========================================================================
3595 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3596 * stops after this stage and will resume upon I/O completion.
3597 * However, there are instances where the vdev layer may need to
3598 * continue the pipeline when an I/O was not issued. Since the I/O
3599 * that was sent to the vdev layer might be different than the one
3600 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3601 * force the underlying vdev layers to call either zio_execute() or
3602 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3605 zio_vdev_io_start(zio_t *zio)
3607 vdev_t *vd = zio->io_vd;
3609 spa_t *spa = zio->io_spa;
3613 ASSERT(zio->io_error == 0);
3614 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
3617 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3618 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
3621 * The mirror_ops handle multiple DVAs in a single BP.
3623 vdev_mirror_ops.vdev_op_io_start(zio);
3627 ASSERT3P(zio->io_logical, !=, zio);
3628 if (zio->io_type == ZIO_TYPE_WRITE) {
3629 ASSERT(spa->spa_trust_config);
3632 * Note: the code can handle other kinds of writes,
3633 * but we don't expect them.
3635 if (zio->io_vd->vdev_removing) {
3636 ASSERT(zio->io_flags &
3637 (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL |
3638 ZIO_FLAG_RESILVER | ZIO_FLAG_INDUCE_DAMAGE));
3642 align = 1ULL << vd->vdev_top->vdev_ashift;
3644 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
3645 P2PHASE(zio->io_size, align) != 0) {
3646 /* Transform logical writes to be a full physical block size. */
3647 uint64_t asize = P2ROUNDUP(zio->io_size, align);
3648 abd_t *abuf = abd_alloc_sametype(zio->io_abd, asize);
3649 ASSERT(vd == vd->vdev_top);
3650 if (zio->io_type == ZIO_TYPE_WRITE) {
3651 abd_copy(abuf, zio->io_abd, zio->io_size);
3652 abd_zero_off(abuf, zio->io_size, asize - zio->io_size);
3654 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
3658 * If this is not a physical io, make sure that it is properly aligned
3659 * before proceeding.
3661 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
3662 ASSERT0(P2PHASE(zio->io_offset, align));
3663 ASSERT0(P2PHASE(zio->io_size, align));
3666 * For physical writes, we allow 512b aligned writes and assume
3667 * the device will perform a read-modify-write as necessary.
3669 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
3670 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
3673 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
3676 * If this is a repair I/O, and there's no self-healing involved --
3677 * that is, we're just resilvering what we expect to resilver --
3678 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3679 * This prevents spurious resilvering.
3681 * There are a few ways that we can end up creating these spurious
3684 * 1. A resilver i/o will be issued if any DVA in the BP has a
3685 * dirty DTL. The mirror code will issue resilver writes to
3686 * each DVA, including the one(s) that are not on vdevs with dirty
3689 * 2. With nested replication, which happens when we have a
3690 * "replacing" or "spare" vdev that's a child of a mirror or raidz.
3691 * For example, given mirror(replacing(A+B), C), it's likely that
3692 * only A is out of date (it's the new device). In this case, we'll
3693 * read from C, then use the data to resilver A+B -- but we don't
3694 * actually want to resilver B, just A. The top-level mirror has no
3695 * way to know this, so instead we just discard unnecessary repairs
3696 * as we work our way down the vdev tree.
3698 * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc.
3699 * The same logic applies to any form of nested replication: ditto
3700 * + mirror, RAID-Z + replacing, etc.
3702 * However, indirect vdevs point off to other vdevs which may have
3703 * DTL's, so we never bypass them. The child i/os on concrete vdevs
3704 * will be properly bypassed instead.
3706 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3707 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
3708 zio->io_txg != 0 && /* not a delegated i/o */
3709 vd->vdev_ops != &vdev_indirect_ops &&
3710 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
3711 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3712 zio_vdev_io_bypass(zio);
3716 if (vd->vdev_ops->vdev_op_leaf && (zio->io_type == ZIO_TYPE_READ ||
3717 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_TRIM)) {
3719 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio))
3722 if ((zio = vdev_queue_io(zio)) == NULL)
3725 if (!vdev_accessible(vd, zio)) {
3726 zio->io_error = SET_ERROR(ENXIO);
3730 zio->io_delay = gethrtime();
3733 vd->vdev_ops->vdev_op_io_start(zio);
3738 zio_vdev_io_done(zio_t *zio)
3740 vdev_t *vd = zio->io_vd;
3741 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
3742 boolean_t unexpected_error = B_FALSE;
3744 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3748 ASSERT(zio->io_type == ZIO_TYPE_READ ||
3749 zio->io_type == ZIO_TYPE_WRITE || zio->io_type == ZIO_TYPE_TRIM);
3752 zio->io_delay = gethrtime() - zio->io_delay;
3754 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
3756 vdev_queue_io_done(zio);
3758 if (zio->io_type == ZIO_TYPE_WRITE)
3759 vdev_cache_write(zio);
3761 if (zio_injection_enabled && zio->io_error == 0)
3762 zio->io_error = zio_handle_device_injections(vd, zio,
3765 if (zio_injection_enabled && zio->io_error == 0)
3766 zio->io_error = zio_handle_label_injection(zio, EIO);
3768 if (zio->io_error && zio->io_type != ZIO_TYPE_TRIM) {
3769 if (!vdev_accessible(vd, zio)) {
3770 zio->io_error = SET_ERROR(ENXIO);
3772 unexpected_error = B_TRUE;
3777 ops->vdev_op_io_done(zio);
3779 if (unexpected_error)
3780 VERIFY(vdev_probe(vd, zio) == NULL);
3786 * This function is used to change the priority of an existing zio that is
3787 * currently in-flight. This is used by the arc to upgrade priority in the
3788 * event that a demand read is made for a block that is currently queued
3789 * as a scrub or async read IO. Otherwise, the high priority read request
3790 * would end up having to wait for the lower priority IO.
3793 zio_change_priority(zio_t *pio, zio_priority_t priority)
3795 zio_t *cio, *cio_next;
3796 zio_link_t *zl = NULL;
3798 ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
3800 if (pio->io_vd != NULL && pio->io_vd->vdev_ops->vdev_op_leaf) {
3801 vdev_queue_change_io_priority(pio, priority);
3803 pio->io_priority = priority;
3806 mutex_enter(&pio->io_lock);
3807 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
3808 cio_next = zio_walk_children(pio, &zl);
3809 zio_change_priority(cio, priority);
3811 mutex_exit(&pio->io_lock);
3815 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3816 * disk, and use that to finish the checksum ereport later.
3819 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
3820 const abd_t *good_buf)
3822 /* no processing needed */
3823 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
3828 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
3830 void *abd = abd_alloc_sametype(zio->io_abd, zio->io_size);
3832 abd_copy(abd, zio->io_abd, zio->io_size);
3834 zcr->zcr_cbinfo = zio->io_size;
3835 zcr->zcr_cbdata = abd;
3836 zcr->zcr_finish = zio_vsd_default_cksum_finish;
3837 zcr->zcr_free = zio_abd_free;
3841 zio_vdev_io_assess(zio_t *zio)
3843 vdev_t *vd = zio->io_vd;
3845 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3849 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3850 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
3852 if (zio->io_vsd != NULL) {
3853 zio->io_vsd_ops->vsd_free(zio);
3857 if (zio_injection_enabled && zio->io_error == 0)
3858 zio->io_error = zio_handle_fault_injection(zio, EIO);
3861 * If the I/O failed, determine whether we should attempt to retry it.
3863 * On retry, we cut in line in the issue queue, since we don't want
3864 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3866 if (zio->io_error && vd == NULL &&
3867 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
3868 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
3869 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
3871 zio->io_flags |= ZIO_FLAG_IO_RETRY |
3872 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
3873 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
3874 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
3875 zio_requeue_io_start_cut_in_line);
3880 * If we got an error on a leaf device, convert it to ENXIO
3881 * if the device is not accessible at all.
3883 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3884 !vdev_accessible(vd, zio))
3885 zio->io_error = SET_ERROR(ENXIO);
3888 * If we can't write to an interior vdev (mirror or RAID-Z),
3889 * set vdev_cant_write so that we stop trying to allocate from it.
3891 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
3892 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
3893 vd->vdev_cant_write = B_TRUE;
3897 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
3898 * attempts will ever succeed. In this case we set a persistent
3899 * boolean flag so that we don't bother with it in the future.
3901 if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) &&
3902 zio->io_type == ZIO_TYPE_IOCTL &&
3903 zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL)
3904 vd->vdev_nowritecache = B_TRUE;
3907 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3909 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3910 zio->io_physdone != NULL) {
3911 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
3912 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
3913 zio->io_physdone(zio->io_logical);
3920 zio_vdev_io_reissue(zio_t *zio)
3922 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3923 ASSERT(zio->io_error == 0);
3925 zio->io_stage >>= 1;
3929 zio_vdev_io_redone(zio_t *zio)
3931 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
3933 zio->io_stage >>= 1;
3937 zio_vdev_io_bypass(zio_t *zio)
3939 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3940 ASSERT(zio->io_error == 0);
3942 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
3943 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
3947 * ==========================================================================
3948 * Encrypt and store encryption parameters
3949 * ==========================================================================
3954 * This function is used for ZIO_STAGE_ENCRYPT. It is responsible for
3955 * managing the storage of encryption parameters and passing them to the
3956 * lower-level encryption functions.
3959 zio_encrypt(zio_t *zio)
3961 zio_prop_t *zp = &zio->io_prop;
3962 spa_t *spa = zio->io_spa;
3963 blkptr_t *bp = zio->io_bp;
3964 uint64_t psize = BP_GET_PSIZE(bp);
3965 uint64_t dsobj = zio->io_bookmark.zb_objset;
3966 dmu_object_type_t ot = BP_GET_TYPE(bp);
3967 void *enc_buf = NULL;
3969 uint8_t salt[ZIO_DATA_SALT_LEN];
3970 uint8_t iv[ZIO_DATA_IV_LEN];
3971 uint8_t mac[ZIO_DATA_MAC_LEN];
3972 boolean_t no_crypt = B_FALSE;
3974 /* the root zio already encrypted the data */
3975 if (zio->io_child_type == ZIO_CHILD_GANG)
3978 /* only ZIL blocks are re-encrypted on rewrite */
3979 if (!IO_IS_ALLOCATING(zio) && ot != DMU_OT_INTENT_LOG)
3982 if (!(zp->zp_encrypt || BP_IS_ENCRYPTED(bp))) {
3983 BP_SET_CRYPT(bp, B_FALSE);
3987 /* if we are doing raw encryption set the provided encryption params */
3988 if (zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) {
3989 ASSERT0(BP_GET_LEVEL(bp));
3990 BP_SET_CRYPT(bp, B_TRUE);
3991 BP_SET_BYTEORDER(bp, zp->zp_byteorder);
3992 if (ot != DMU_OT_OBJSET)
3993 zio_crypt_encode_mac_bp(bp, zp->zp_mac);
3995 /* dnode blocks must be written out in the provided byteorder */
3996 if (zp->zp_byteorder != ZFS_HOST_BYTEORDER &&
3997 ot == DMU_OT_DNODE) {
3998 void *bswap_buf = zio_buf_alloc(psize);
3999 abd_t *babd = abd_get_from_buf(bswap_buf, psize);
4001 ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
4002 abd_copy_to_buf(bswap_buf, zio->io_abd, psize);
4003 dmu_ot_byteswap[DMU_OT_BYTESWAP(ot)].ob_func(bswap_buf,
4006 abd_take_ownership_of_buf(babd, B_TRUE);
4007 zio_push_transform(zio, babd, psize, psize, NULL);
4010 if (DMU_OT_IS_ENCRYPTED(ot))
4011 zio_crypt_encode_params_bp(bp, zp->zp_salt, zp->zp_iv);
4015 /* indirect blocks only maintain a cksum of the lower level MACs */
4016 if (BP_GET_LEVEL(bp) > 0) {
4017 BP_SET_CRYPT(bp, B_TRUE);
4018 VERIFY0(zio_crypt_do_indirect_mac_checksum_abd(B_TRUE,
4019 zio->io_orig_abd, BP_GET_LSIZE(bp), BP_SHOULD_BYTESWAP(bp),
4021 zio_crypt_encode_mac_bp(bp, mac);
4026 * Objset blocks are a special case since they have 2 256-bit MACs
4027 * embedded within them.
4029 if (ot == DMU_OT_OBJSET) {
4030 ASSERT0(DMU_OT_IS_ENCRYPTED(ot));
4031 ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
4032 BP_SET_CRYPT(bp, B_TRUE);
4033 VERIFY0(spa_do_crypt_objset_mac_abd(B_TRUE, spa, dsobj,
4034 zio->io_abd, psize, BP_SHOULD_BYTESWAP(bp)));
4038 /* unencrypted object types are only authenticated with a MAC */
4039 if (!DMU_OT_IS_ENCRYPTED(ot)) {
4040 BP_SET_CRYPT(bp, B_TRUE);
4041 VERIFY0(spa_do_crypt_mac_abd(B_TRUE, spa, dsobj,
4042 zio->io_abd, psize, mac));
4043 zio_crypt_encode_mac_bp(bp, mac);
4048 * Later passes of sync-to-convergence may decide to rewrite data
4049 * in place to avoid more disk reallocations. This presents a problem
4050 * for encryption because this constitutes rewriting the new data with
4051 * the same encryption key and IV. However, this only applies to blocks
4052 * in the MOS (particularly the spacemaps) and we do not encrypt the
4053 * MOS. We assert that the zio is allocating or an intent log write
4056 ASSERT(IO_IS_ALLOCATING(zio) || ot == DMU_OT_INTENT_LOG);
4057 ASSERT(BP_GET_LEVEL(bp) == 0 || ot == DMU_OT_INTENT_LOG);
4058 ASSERT(spa_feature_is_active(spa, SPA_FEATURE_ENCRYPTION));
4059 ASSERT3U(psize, !=, 0);
4061 enc_buf = zio_buf_alloc(psize);
4062 eabd = abd_get_from_buf(enc_buf, psize);
4063 abd_take_ownership_of_buf(eabd, B_TRUE);
4066 * For an explanation of what encryption parameters are stored
4067 * where, see the block comment in zio_crypt.c.
4069 if (ot == DMU_OT_INTENT_LOG) {
4070 zio_crypt_decode_params_bp(bp, salt, iv);
4072 BP_SET_CRYPT(bp, B_TRUE);
4075 /* Perform the encryption. This should not fail */
4076 VERIFY0(spa_do_crypt_abd(B_TRUE, spa, &zio->io_bookmark,
4077 BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp),
4078 salt, iv, mac, psize, zio->io_abd, eabd, &no_crypt));
4080 /* encode encryption metadata into the bp */
4081 if (ot == DMU_OT_INTENT_LOG) {
4083 * ZIL blocks store the MAC in the embedded checksum, so the
4084 * transform must always be applied.
4086 zio_crypt_encode_mac_zil(enc_buf, mac);
4087 zio_push_transform(zio, eabd, psize, psize, NULL);
4089 BP_SET_CRYPT(bp, B_TRUE);
4090 zio_crypt_encode_params_bp(bp, salt, iv);
4091 zio_crypt_encode_mac_bp(bp, mac);
4094 ASSERT3U(ot, ==, DMU_OT_DNODE);
4097 zio_push_transform(zio, eabd, psize, psize, NULL);
4105 * ==========================================================================
4106 * Generate and verify checksums
4107 * ==========================================================================
4110 zio_checksum_generate(zio_t *zio)
4112 blkptr_t *bp = zio->io_bp;
4113 enum zio_checksum checksum;
4117 * This is zio_write_phys().
4118 * We're either generating a label checksum, or none at all.
4120 checksum = zio->io_prop.zp_checksum;
4122 if (checksum == ZIO_CHECKSUM_OFF)
4125 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
4127 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
4128 ASSERT(!IO_IS_ALLOCATING(zio));
4129 checksum = ZIO_CHECKSUM_GANG_HEADER;
4131 checksum = BP_GET_CHECKSUM(bp);
4135 zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size);
4141 zio_checksum_verify(zio_t *zio)
4143 zio_bad_cksum_t info;
4144 blkptr_t *bp = zio->io_bp;
4147 ASSERT(zio->io_vd != NULL);
4151 * This is zio_read_phys().
4152 * We're either verifying a label checksum, or nothing at all.
4154 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
4157 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
4160 if ((error = zio_checksum_error(zio, &info)) != 0) {
4161 zio->io_error = error;
4162 if (error == ECKSUM &&
4163 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
4164 mutex_enter(&zio->io_vd->vdev_stat_lock);
4165 zio->io_vd->vdev_stat.vs_checksum_errors++;
4166 mutex_exit(&zio->io_vd->vdev_stat_lock);
4168 zfs_ereport_start_checksum(zio->io_spa,
4169 zio->io_vd, &zio->io_bookmark, zio,
4170 zio->io_offset, zio->io_size, NULL, &info);
4178 * Called by RAID-Z to ensure we don't compute the checksum twice.
4181 zio_checksum_verified(zio_t *zio)
4183 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
4187 * ==========================================================================
4188 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
4189 * An error of 0 indicates success. ENXIO indicates whole-device failure,
4190 * which may be transient (e.g. unplugged) or permanent. ECKSUM and EIO
4191 * indicate errors that are specific to one I/O, and most likely permanent.
4192 * Any other error is presumed to be worse because we weren't expecting it.
4193 * ==========================================================================
4196 zio_worst_error(int e1, int e2)
4198 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
4201 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
4202 if (e1 == zio_error_rank[r1])
4205 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
4206 if (e2 == zio_error_rank[r2])
4209 return (r1 > r2 ? e1 : e2);
4213 * ==========================================================================
4215 * ==========================================================================
4218 zio_ready(zio_t *zio)
4220 blkptr_t *bp = zio->io_bp;
4221 zio_t *pio, *pio_next;
4222 zio_link_t *zl = NULL;
4224 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT | ZIO_CHILD_DDT_BIT,
4229 if (zio->io_ready) {
4230 ASSERT(IO_IS_ALLOCATING(zio));
4231 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
4232 (zio->io_flags & ZIO_FLAG_NOPWRITE));
4233 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
4238 if (bp != NULL && bp != &zio->io_bp_copy)
4239 zio->io_bp_copy = *bp;
4241 if (zio->io_error != 0) {
4242 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
4244 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
4245 ASSERT(IO_IS_ALLOCATING(zio));
4246 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
4247 ASSERT(zio->io_metaslab_class != NULL);
4250 * We were unable to allocate anything, unreserve and
4251 * issue the next I/O to allocate.
4253 metaslab_class_throttle_unreserve(
4254 zio->io_metaslab_class, zio->io_prop.zp_copies,
4255 zio->io_allocator, zio);
4256 zio_allocate_dispatch(zio->io_spa, zio->io_allocator);
4260 mutex_enter(&zio->io_lock);
4261 zio->io_state[ZIO_WAIT_READY] = 1;
4262 pio = zio_walk_parents(zio, &zl);
4263 mutex_exit(&zio->io_lock);
4266 * As we notify zio's parents, new parents could be added.
4267 * New parents go to the head of zio's io_parent_list, however,
4268 * so we will (correctly) not notify them. The remainder of zio's
4269 * io_parent_list, from 'pio_next' onward, cannot change because
4270 * all parents must wait for us to be done before they can be done.
4272 for (; pio != NULL; pio = pio_next) {
4273 pio_next = zio_walk_parents(zio, &zl);
4274 zio_notify_parent(pio, zio, ZIO_WAIT_READY, NULL);
4277 if (zio->io_flags & ZIO_FLAG_NODATA) {
4278 if (BP_IS_GANG(bp)) {
4279 zio->io_flags &= ~ZIO_FLAG_NODATA;
4281 ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE);
4282 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
4286 if (zio_injection_enabled &&
4287 zio->io_spa->spa_syncing_txg == zio->io_txg)
4288 zio_handle_ignored_writes(zio);
4294 * Update the allocation throttle accounting.
4297 zio_dva_throttle_done(zio_t *zio)
4299 ASSERTV(zio_t *lio = zio->io_logical);
4300 zio_t *pio = zio_unique_parent(zio);
4301 vdev_t *vd = zio->io_vd;
4302 int flags = METASLAB_ASYNC_ALLOC;
4304 ASSERT3P(zio->io_bp, !=, NULL);
4305 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
4306 ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE);
4307 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
4309 ASSERT3P(vd, ==, vd->vdev_top);
4310 ASSERT(zio_injection_enabled || !(zio->io_flags & ZIO_FLAG_IO_RETRY));
4311 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
4312 ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING);
4313 ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE));
4314 ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA));
4317 * Parents of gang children can have two flavors -- ones that
4318 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
4319 * and ones that allocated the constituent blocks. The allocation
4320 * throttle needs to know the allocating parent zio so we must find
4323 if (pio->io_child_type == ZIO_CHILD_GANG) {
4325 * If our parent is a rewrite gang child then our grandparent
4326 * would have been the one that performed the allocation.
4328 if (pio->io_flags & ZIO_FLAG_IO_REWRITE)
4329 pio = zio_unique_parent(pio);
4330 flags |= METASLAB_GANG_CHILD;
4333 ASSERT(IO_IS_ALLOCATING(pio));
4334 ASSERT3P(zio, !=, zio->io_logical);
4335 ASSERT(zio->io_logical != NULL);
4336 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
4337 ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE);
4338 ASSERT(zio->io_metaslab_class != NULL);
4340 mutex_enter(&pio->io_lock);
4341 metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags,
4342 pio->io_allocator, B_TRUE);
4343 mutex_exit(&pio->io_lock);
4345 metaslab_class_throttle_unreserve(zio->io_metaslab_class, 1,
4346 pio->io_allocator, pio);
4349 * Call into the pipeline to see if there is more work that
4350 * needs to be done. If there is work to be done it will be
4351 * dispatched to another taskq thread.
4353 zio_allocate_dispatch(zio->io_spa, pio->io_allocator);
4357 zio_done(zio_t *zio)
4360 * Always attempt to keep stack usage minimal here since
4361 * we can be called recursively up to 19 levels deep.
4363 const uint64_t psize = zio->io_size;
4364 zio_t *pio, *pio_next;
4365 zio_link_t *zl = NULL;
4368 * If our children haven't all completed,
4369 * wait for them and then repeat this pipeline stage.
4371 if (zio_wait_for_children(zio, ZIO_CHILD_ALL_BITS, ZIO_WAIT_DONE)) {
4376 * If the allocation throttle is enabled, then update the accounting.
4377 * We only track child I/Os that are part of an allocating async
4378 * write. We must do this since the allocation is performed
4379 * by the logical I/O but the actual write is done by child I/Os.
4381 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING &&
4382 zio->io_child_type == ZIO_CHILD_VDEV) {
4383 ASSERT(zio->io_metaslab_class != NULL);
4384 ASSERT(zio->io_metaslab_class->mc_alloc_throttle_enabled);
4385 zio_dva_throttle_done(zio);
4389 * If the allocation throttle is enabled, verify that
4390 * we have decremented the refcounts for every I/O that was throttled.
4392 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
4393 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
4394 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
4395 ASSERT(zio->io_bp != NULL);
4397 metaslab_group_alloc_verify(zio->io_spa, zio->io_bp, zio,
4399 VERIFY(zfs_refcount_not_held(
4400 &zio->io_metaslab_class->mc_alloc_slots[zio->io_allocator],
4405 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
4406 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
4407 ASSERT(zio->io_children[c][w] == 0);
4409 if (zio->io_bp != NULL && !BP_IS_EMBEDDED(zio->io_bp)) {
4410 ASSERT(zio->io_bp->blk_pad[0] == 0);
4411 ASSERT(zio->io_bp->blk_pad[1] == 0);
4412 ASSERT(bcmp(zio->io_bp, &zio->io_bp_copy,
4413 sizeof (blkptr_t)) == 0 ||
4414 (zio->io_bp == zio_unique_parent(zio)->io_bp));
4415 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) &&
4416 zio->io_bp_override == NULL &&
4417 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
4418 ASSERT3U(zio->io_prop.zp_copies, <=,
4419 BP_GET_NDVAS(zio->io_bp));
4420 ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 ||
4421 (BP_COUNT_GANG(zio->io_bp) ==
4422 BP_GET_NDVAS(zio->io_bp)));
4424 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
4425 VERIFY(BP_EQUAL(zio->io_bp, &zio->io_bp_orig));
4429 * If there were child vdev/gang/ddt errors, they apply to us now.
4431 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
4432 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
4433 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
4436 * If the I/O on the transformed data was successful, generate any
4437 * checksum reports now while we still have the transformed data.
4439 if (zio->io_error == 0) {
4440 while (zio->io_cksum_report != NULL) {
4441 zio_cksum_report_t *zcr = zio->io_cksum_report;
4442 uint64_t align = zcr->zcr_align;
4443 uint64_t asize = P2ROUNDUP(psize, align);
4444 abd_t *adata = zio->io_abd;
4446 if (asize != psize) {
4447 adata = abd_alloc(asize, B_TRUE);
4448 abd_copy(adata, zio->io_abd, psize);
4449 abd_zero_off(adata, psize, asize - psize);
4452 zio->io_cksum_report = zcr->zcr_next;
4453 zcr->zcr_next = NULL;
4454 zcr->zcr_finish(zcr, adata);
4455 zfs_ereport_free_checksum(zcr);
4462 zio_pop_transforms(zio); /* note: may set zio->io_error */
4464 vdev_stat_update(zio, psize);
4467 * If this I/O is attached to a particular vdev is slow, exceeding
4468 * 30 seconds to complete, post an error described the I/O delay.
4469 * We ignore these errors if the device is currently unavailable.
4471 if (zio->io_delay >= MSEC2NSEC(zio_slow_io_ms)) {
4472 if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd)) {
4474 * We want to only increment our slow IO counters if
4475 * the IO is valid (i.e. not if the drive is removed).
4477 * zfs_ereport_post() will also do these checks, but
4478 * it can also ratelimit and have other failures, so we
4479 * need to increment the slow_io counters independent
4482 if (zfs_ereport_is_valid(FM_EREPORT_ZFS_DELAY,
4483 zio->io_spa, zio->io_vd, zio)) {
4484 mutex_enter(&zio->io_vd->vdev_stat_lock);
4485 zio->io_vd->vdev_stat.vs_slow_ios++;
4486 mutex_exit(&zio->io_vd->vdev_stat_lock);
4488 zfs_ereport_post(FM_EREPORT_ZFS_DELAY,
4489 zio->io_spa, zio->io_vd, &zio->io_bookmark,
4495 if (zio->io_error) {
4497 * If this I/O is attached to a particular vdev,
4498 * generate an error message describing the I/O failure
4499 * at the block level. We ignore these errors if the
4500 * device is currently unavailable.
4502 if (zio->io_error != ECKSUM && zio->io_vd != NULL &&
4503 !vdev_is_dead(zio->io_vd)) {
4504 mutex_enter(&zio->io_vd->vdev_stat_lock);
4505 if (zio->io_type == ZIO_TYPE_READ) {
4506 zio->io_vd->vdev_stat.vs_read_errors++;
4507 } else if (zio->io_type == ZIO_TYPE_WRITE) {
4508 zio->io_vd->vdev_stat.vs_write_errors++;
4510 mutex_exit(&zio->io_vd->vdev_stat_lock);
4512 zfs_ereport_post(FM_EREPORT_ZFS_IO, zio->io_spa,
4513 zio->io_vd, &zio->io_bookmark, zio, 0, 0);
4516 if ((zio->io_error == EIO || !(zio->io_flags &
4517 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
4518 zio == zio->io_logical) {
4520 * For logical I/O requests, tell the SPA to log the
4521 * error and generate a logical data ereport.
4523 spa_log_error(zio->io_spa, &zio->io_bookmark);
4524 zfs_ereport_post(FM_EREPORT_ZFS_DATA, zio->io_spa,
4525 NULL, &zio->io_bookmark, zio, 0, 0);
4529 if (zio->io_error && zio == zio->io_logical) {
4531 * Determine whether zio should be reexecuted. This will
4532 * propagate all the way to the root via zio_notify_parent().
4534 ASSERT(zio->io_vd == NULL && zio->io_bp != NULL);
4535 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
4537 if (IO_IS_ALLOCATING(zio) &&
4538 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
4539 if (zio->io_error != ENOSPC)
4540 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
4542 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4545 if ((zio->io_type == ZIO_TYPE_READ ||
4546 zio->io_type == ZIO_TYPE_FREE) &&
4547 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
4548 zio->io_error == ENXIO &&
4549 spa_load_state(zio->io_spa) == SPA_LOAD_NONE &&
4550 spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE)
4551 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4553 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
4554 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4557 * Here is a possibly good place to attempt to do
4558 * either combinatorial reconstruction or error correction
4559 * based on checksums. It also might be a good place
4560 * to send out preliminary ereports before we suspend
4566 * If there were logical child errors, they apply to us now.
4567 * We defer this until now to avoid conflating logical child
4568 * errors with errors that happened to the zio itself when
4569 * updating vdev stats and reporting FMA events above.
4571 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
4573 if ((zio->io_error || zio->io_reexecute) &&
4574 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
4575 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
4576 zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp);
4578 zio_gang_tree_free(&zio->io_gang_tree);
4581 * Godfather I/Os should never suspend.
4583 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
4584 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
4585 zio->io_reexecute &= ~ZIO_REEXECUTE_SUSPEND;
4587 if (zio->io_reexecute) {
4589 * This is a logical I/O that wants to reexecute.
4591 * Reexecute is top-down. When an i/o fails, if it's not
4592 * the root, it simply notifies its parent and sticks around.
4593 * The parent, seeing that it still has children in zio_done(),
4594 * does the same. This percolates all the way up to the root.
4595 * The root i/o will reexecute or suspend the entire tree.
4597 * This approach ensures that zio_reexecute() honors
4598 * all the original i/o dependency relationships, e.g.
4599 * parents not executing until children are ready.
4601 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
4603 zio->io_gang_leader = NULL;
4605 mutex_enter(&zio->io_lock);
4606 zio->io_state[ZIO_WAIT_DONE] = 1;
4607 mutex_exit(&zio->io_lock);
4610 * "The Godfather" I/O monitors its children but is
4611 * not a true parent to them. It will track them through
4612 * the pipeline but severs its ties whenever they get into
4613 * trouble (e.g. suspended). This allows "The Godfather"
4614 * I/O to return status without blocking.
4617 for (pio = zio_walk_parents(zio, &zl); pio != NULL;
4619 zio_link_t *remove_zl = zl;
4620 pio_next = zio_walk_parents(zio, &zl);
4622 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
4623 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
4624 zio_remove_child(pio, zio, remove_zl);
4626 * This is a rare code path, so we don't
4627 * bother with "next_to_execute".
4629 zio_notify_parent(pio, zio, ZIO_WAIT_DONE,
4634 if ((pio = zio_unique_parent(zio)) != NULL) {
4636 * We're not a root i/o, so there's nothing to do
4637 * but notify our parent. Don't propagate errors
4638 * upward since we haven't permanently failed yet.
4640 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
4641 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
4643 * This is a rare code path, so we don't bother with
4644 * "next_to_execute".
4646 zio_notify_parent(pio, zio, ZIO_WAIT_DONE, NULL);
4647 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
4649 * We'd fail again if we reexecuted now, so suspend
4650 * until conditions improve (e.g. device comes online).
4652 zio_suspend(zio->io_spa, zio, ZIO_SUSPEND_IOERR);
4655 * Reexecution is potentially a huge amount of work.
4656 * Hand it off to the otherwise-unused claim taskq.
4658 ASSERT(taskq_empty_ent(&zio->io_tqent));
4659 spa_taskq_dispatch_ent(zio->io_spa,
4660 ZIO_TYPE_CLAIM, ZIO_TASKQ_ISSUE,
4661 (task_func_t *)zio_reexecute, zio, 0,
4667 ASSERT(zio->io_child_count == 0);
4668 ASSERT(zio->io_reexecute == 0);
4669 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
4672 * Report any checksum errors, since the I/O is complete.
4674 while (zio->io_cksum_report != NULL) {
4675 zio_cksum_report_t *zcr = zio->io_cksum_report;
4676 zio->io_cksum_report = zcr->zcr_next;
4677 zcr->zcr_next = NULL;
4678 zcr->zcr_finish(zcr, NULL);
4679 zfs_ereport_free_checksum(zcr);
4682 if (zio->io_flags & ZIO_FLAG_FASTWRITE && zio->io_bp &&
4683 !BP_IS_HOLE(zio->io_bp) && !BP_IS_EMBEDDED(zio->io_bp) &&
4684 !(zio->io_flags & ZIO_FLAG_NOPWRITE)) {
4685 metaslab_fastwrite_unmark(zio->io_spa, zio->io_bp);
4689 * It is the responsibility of the done callback to ensure that this
4690 * particular zio is no longer discoverable for adoption, and as
4691 * such, cannot acquire any new parents.
4696 mutex_enter(&zio->io_lock);
4697 zio->io_state[ZIO_WAIT_DONE] = 1;
4698 mutex_exit(&zio->io_lock);
4701 * We are done executing this zio. We may want to execute a parent
4702 * next. See the comment in zio_notify_parent().
4704 zio_t *next_to_execute = NULL;
4706 for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) {
4707 zio_link_t *remove_zl = zl;
4708 pio_next = zio_walk_parents(zio, &zl);
4709 zio_remove_child(pio, zio, remove_zl);
4710 zio_notify_parent(pio, zio, ZIO_WAIT_DONE, &next_to_execute);
4713 if (zio->io_waiter != NULL) {
4714 mutex_enter(&zio->io_lock);
4715 zio->io_executor = NULL;
4716 cv_broadcast(&zio->io_cv);
4717 mutex_exit(&zio->io_lock);
4722 return (next_to_execute);
4726 * ==========================================================================
4727 * I/O pipeline definition
4728 * ==========================================================================
4730 static zio_pipe_stage_t *zio_pipeline[] = {
4738 zio_checksum_generate,
4754 zio_checksum_verify,
4762 * Compare two zbookmark_phys_t's to see which we would reach first in a
4763 * pre-order traversal of the object tree.
4765 * This is simple in every case aside from the meta-dnode object. For all other
4766 * objects, we traverse them in order (object 1 before object 2, and so on).
4767 * However, all of these objects are traversed while traversing object 0, since
4768 * the data it points to is the list of objects. Thus, we need to convert to a
4769 * canonical representation so we can compare meta-dnode bookmarks to
4770 * non-meta-dnode bookmarks.
4772 * We do this by calculating "equivalents" for each field of the zbookmark.
4773 * zbookmarks outside of the meta-dnode use their own object and level, and
4774 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
4775 * blocks this bookmark refers to) by multiplying their blkid by their span
4776 * (the number of L0 blocks contained within one block at their level).
4777 * zbookmarks inside the meta-dnode calculate their object equivalent
4778 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
4779 * level + 1<<31 (any value larger than a level could ever be) for their level.
4780 * This causes them to always compare before a bookmark in their object
4781 * equivalent, compare appropriately to bookmarks in other objects, and to
4782 * compare appropriately to other bookmarks in the meta-dnode.
4785 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
4786 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
4789 * These variables represent the "equivalent" values for the zbookmark,
4790 * after converting zbookmarks inside the meta dnode to their
4791 * normal-object equivalents.
4793 uint64_t zb1obj, zb2obj;
4794 uint64_t zb1L0, zb2L0;
4795 uint64_t zb1level, zb2level;
4797 if (zb1->zb_object == zb2->zb_object &&
4798 zb1->zb_level == zb2->zb_level &&
4799 zb1->zb_blkid == zb2->zb_blkid)
4802 IMPLY(zb1->zb_level > 0, ibs1 >= SPA_MINBLOCKSHIFT);
4803 IMPLY(zb2->zb_level > 0, ibs2 >= SPA_MINBLOCKSHIFT);
4806 * BP_SPANB calculates the span in blocks.
4808 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
4809 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
4811 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
4812 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4814 zb1level = zb1->zb_level + COMPARE_META_LEVEL;
4816 zb1obj = zb1->zb_object;
4817 zb1level = zb1->zb_level;
4820 if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
4821 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4823 zb2level = zb2->zb_level + COMPARE_META_LEVEL;
4825 zb2obj = zb2->zb_object;
4826 zb2level = zb2->zb_level;
4829 /* Now that we have a canonical representation, do the comparison. */
4830 if (zb1obj != zb2obj)
4831 return (zb1obj < zb2obj ? -1 : 1);
4832 else if (zb1L0 != zb2L0)
4833 return (zb1L0 < zb2L0 ? -1 : 1);
4834 else if (zb1level != zb2level)
4835 return (zb1level > zb2level ? -1 : 1);
4837 * This can (theoretically) happen if the bookmarks have the same object
4838 * and level, but different blkids, if the block sizes are not the same.
4839 * There is presently no way to change the indirect block sizes
4845 * This function checks the following: given that last_block is the place that
4846 * our traversal stopped last time, does that guarantee that we've visited
4847 * every node under subtree_root? Therefore, we can't just use the raw output
4848 * of zbookmark_compare. We have to pass in a modified version of
4849 * subtree_root; by incrementing the block id, and then checking whether
4850 * last_block is before or equal to that, we can tell whether or not having
4851 * visited last_block implies that all of subtree_root's children have been
4855 zbookmark_subtree_completed(const dnode_phys_t *dnp,
4856 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
4858 zbookmark_phys_t mod_zb = *subtree_root;
4860 ASSERT(last_block->zb_level == 0);
4862 /* The objset_phys_t isn't before anything. */
4867 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
4868 * data block size in sectors, because that variable is only used if
4869 * the bookmark refers to a block in the meta-dnode. Since we don't
4870 * know without examining it what object it refers to, and there's no
4871 * harm in passing in this value in other cases, we always pass it in.
4873 * We pass in 0 for the indirect block size shift because zb2 must be
4874 * level 0. The indirect block size is only used to calculate the span
4875 * of the bookmark, but since the bookmark must be level 0, the span is
4876 * always 1, so the math works out.
4878 * If you make changes to how the zbookmark_compare code works, be sure
4879 * to make sure that this code still works afterwards.
4881 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
4882 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,
4886 #if defined(_KERNEL)
4887 EXPORT_SYMBOL(zio_type_name);
4888 EXPORT_SYMBOL(zio_buf_alloc);
4889 EXPORT_SYMBOL(zio_data_buf_alloc);
4890 EXPORT_SYMBOL(zio_buf_free);
4891 EXPORT_SYMBOL(zio_data_buf_free);
4893 module_param(zio_slow_io_ms, int, 0644);
4894 MODULE_PARM_DESC(zio_slow_io_ms,
4895 "Max I/O completion time (milliseconds) before marking it as slow");
4897 module_param(zio_requeue_io_start_cut_in_line, int, 0644);
4898 MODULE_PARM_DESC(zio_requeue_io_start_cut_in_line, "Prioritize requeued I/O");
4900 module_param(zfs_sync_pass_deferred_free, int, 0644);
4901 MODULE_PARM_DESC(zfs_sync_pass_deferred_free,
4902 "Defer frees starting in this pass");
4904 module_param(zfs_sync_pass_dont_compress, int, 0644);
4905 MODULE_PARM_DESC(zfs_sync_pass_dont_compress,
4906 "Don't compress starting in this pass");
4908 module_param(zfs_sync_pass_rewrite, int, 0644);
4909 MODULE_PARM_DESC(zfs_sync_pass_rewrite,
4910 "Rewrite new bps starting in this pass");
4912 module_param(zio_dva_throttle_enabled, int, 0644);
4913 MODULE_PARM_DESC(zio_dva_throttle_enabled,
4914 "Throttle block allocations in the ZIO pipeline");
4916 module_param(zio_deadman_log_all, int, 0644);
4917 MODULE_PARM_DESC(zio_deadman_log_all,
4918 "Log all slow ZIOs, not just those with vdevs");