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, 2018 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/zio_impl.h>
36 #include <sys/zio_compress.h>
37 #include <sys/zio_checksum.h>
38 #include <sys/dmu_objset.h>
41 #include <sys/blkptr.h>
42 #include <sys/zfeature.h>
43 #include <sys/dsl_scan.h>
44 #include <sys/metaslab_impl.h>
46 #include <sys/trace_zio.h>
48 #include <sys/dsl_crypt.h>
49 #include <sys/cityhash.h>
52 * ==========================================================================
53 * I/O type descriptions
54 * ==========================================================================
56 const char *zio_type_name[ZIO_TYPES] = {
58 * Note: Linux kernel thread name length is limited
59 * so these names will differ from upstream open zfs.
61 "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl"
64 int zio_dva_throttle_enabled = B_TRUE;
67 * ==========================================================================
69 * ==========================================================================
71 kmem_cache_t *zio_cache;
72 kmem_cache_t *zio_link_cache;
73 kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
74 kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
75 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
76 uint64_t zio_buf_cache_allocs[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
77 uint64_t zio_buf_cache_frees[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
80 /* Mark IOs as "slow" if they take longer than 30 seconds */
81 int zio_slow_io_ms = (30 * MILLISEC);
83 #define BP_SPANB(indblkshift, level) \
84 (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
85 #define COMPARE_META_LEVEL 0x80000000ul
87 * The following actions directly effect the spa's sync-to-convergence logic.
88 * The values below define the sync pass when we start performing the action.
89 * Care should be taken when changing these values as they directly impact
90 * spa_sync() performance. Tuning these values may introduce subtle performance
91 * pathologies and should only be done in the context of performance analysis.
92 * These tunables will eventually be removed and replaced with #defines once
93 * enough analysis has been done to determine optimal values.
95 * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
96 * regular blocks are not deferred.
98 int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
99 int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
100 int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
103 * An allocating zio is one that either currently has the DVA allocate
104 * stage set or will have it later in its lifetime.
106 #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
108 int zio_requeue_io_start_cut_in_line = 1;
111 int zio_buf_debug_limit = 16384;
113 int zio_buf_debug_limit = 0;
116 static inline void __zio_execute(zio_t *zio);
118 static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t);
124 vmem_t *data_alloc_arena = NULL;
126 zio_cache = kmem_cache_create("zio_cache",
127 sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
128 zio_link_cache = kmem_cache_create("zio_link_cache",
129 sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
132 * For small buffers, we want a cache for each multiple of
133 * SPA_MINBLOCKSIZE. For larger buffers, we want a cache
134 * for each quarter-power of 2.
136 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
137 size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
140 size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
142 #if defined(_ILP32) && defined(_KERNEL)
144 * Cache size limited to 1M on 32-bit platforms until ARC
145 * buffers no longer require virtual address space.
147 if (size > zfs_max_recordsize)
156 * If we are using watchpoints, put each buffer on its own page,
157 * to eliminate the performance overhead of trapping to the
158 * kernel when modifying a non-watched buffer that shares the
159 * page with a watched buffer.
161 if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
164 * Here's the problem - on 4K native devices in userland on
165 * Linux using O_DIRECT, buffers must be 4K aligned or I/O
166 * will fail with EINVAL, causing zdb (and others) to coredump.
167 * Since userland probably doesn't need optimized buffer caches,
168 * we just force 4K alignment on everything.
170 align = 8 * SPA_MINBLOCKSIZE;
172 if (size < PAGESIZE) {
173 align = SPA_MINBLOCKSIZE;
174 } else if (IS_P2ALIGNED(size, p2 >> 2)) {
181 (void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
182 zio_buf_cache[c] = kmem_cache_create(name, size,
183 align, NULL, NULL, NULL, NULL, NULL, cflags);
185 (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
186 zio_data_buf_cache[c] = kmem_cache_create(name, size,
187 align, NULL, NULL, NULL, NULL,
188 data_alloc_arena, cflags);
193 ASSERT(zio_buf_cache[c] != NULL);
194 if (zio_buf_cache[c - 1] == NULL)
195 zio_buf_cache[c - 1] = zio_buf_cache[c];
197 ASSERT(zio_data_buf_cache[c] != NULL);
198 if (zio_data_buf_cache[c - 1] == NULL)
199 zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
211 kmem_cache_t *last_cache = NULL;
212 kmem_cache_t *last_data_cache = NULL;
214 for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
217 * Cache size limited to 1M on 32-bit platforms until ARC
218 * buffers no longer require virtual address space.
220 if (((c + 1) << SPA_MINBLOCKSHIFT) > zfs_max_recordsize)
223 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
224 if (zio_buf_cache_allocs[c] != zio_buf_cache_frees[c])
225 (void) printf("zio_fini: [%d] %llu != %llu\n",
226 (int)((c + 1) << SPA_MINBLOCKSHIFT),
227 (long long unsigned)zio_buf_cache_allocs[c],
228 (long long unsigned)zio_buf_cache_frees[c]);
230 if (zio_buf_cache[c] != last_cache) {
231 last_cache = zio_buf_cache[c];
232 kmem_cache_destroy(zio_buf_cache[c]);
234 zio_buf_cache[c] = NULL;
236 if (zio_data_buf_cache[c] != last_data_cache) {
237 last_data_cache = zio_data_buf_cache[c];
238 kmem_cache_destroy(zio_data_buf_cache[c]);
240 zio_data_buf_cache[c] = NULL;
243 kmem_cache_destroy(zio_link_cache);
244 kmem_cache_destroy(zio_cache);
252 * ==========================================================================
253 * Allocate and free I/O buffers
254 * ==========================================================================
258 * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
259 * crashdump if the kernel panics, so use it judiciously. Obviously, it's
260 * useful to inspect ZFS metadata, but if possible, we should avoid keeping
261 * excess / transient data in-core during a crashdump.
264 zio_buf_alloc(size_t size)
266 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
268 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
269 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
270 atomic_add_64(&zio_buf_cache_allocs[c], 1);
273 return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
277 * Use zio_data_buf_alloc to allocate data. The data will not appear in a
278 * crashdump if the kernel panics. This exists so that we will limit the amount
279 * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
280 * of kernel heap dumped to disk when the kernel panics)
283 zio_data_buf_alloc(size_t size)
285 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
287 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
289 return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
293 zio_buf_free(void *buf, size_t size)
295 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
297 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
298 #if defined(ZFS_DEBUG) && !defined(_KERNEL)
299 atomic_add_64(&zio_buf_cache_frees[c], 1);
302 kmem_cache_free(zio_buf_cache[c], buf);
306 zio_data_buf_free(void *buf, size_t size)
308 size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
310 VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
312 kmem_cache_free(zio_data_buf_cache[c], buf);
316 zio_abd_free(void *abd, size_t size)
318 abd_free((abd_t *)abd);
322 * ==========================================================================
323 * Push and pop I/O transform buffers
324 * ==========================================================================
327 zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize,
328 zio_transform_func_t *transform)
330 zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
333 * Ensure that anyone expecting this zio to contain a linear ABD isn't
334 * going to get a nasty surprise when they try to access the data.
336 IMPLY(abd_is_linear(zio->io_abd), abd_is_linear(data));
338 zt->zt_orig_abd = zio->io_abd;
339 zt->zt_orig_size = zio->io_size;
340 zt->zt_bufsize = bufsize;
341 zt->zt_transform = transform;
343 zt->zt_next = zio->io_transform_stack;
344 zio->io_transform_stack = zt;
351 zio_pop_transforms(zio_t *zio)
355 while ((zt = zio->io_transform_stack) != NULL) {
356 if (zt->zt_transform != NULL)
357 zt->zt_transform(zio,
358 zt->zt_orig_abd, zt->zt_orig_size);
360 if (zt->zt_bufsize != 0)
361 abd_free(zio->io_abd);
363 zio->io_abd = zt->zt_orig_abd;
364 zio->io_size = zt->zt_orig_size;
365 zio->io_transform_stack = zt->zt_next;
367 kmem_free(zt, sizeof (zio_transform_t));
372 * ==========================================================================
373 * I/O transform callbacks for subblocks, decompression, and decryption
374 * ==========================================================================
377 zio_subblock(zio_t *zio, abd_t *data, uint64_t size)
379 ASSERT(zio->io_size > size);
381 if (zio->io_type == ZIO_TYPE_READ)
382 abd_copy(data, zio->io_abd, size);
386 zio_decompress(zio_t *zio, abd_t *data, uint64_t size)
388 if (zio->io_error == 0) {
389 void *tmp = abd_borrow_buf(data, size);
390 int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
391 zio->io_abd, tmp, zio->io_size, size);
392 abd_return_buf_copy(data, tmp, size);
394 if (zio_injection_enabled && ret == 0)
395 ret = zio_handle_fault_injection(zio, EINVAL);
398 zio->io_error = SET_ERROR(EIO);
403 zio_decrypt(zio_t *zio, abd_t *data, uint64_t size)
407 blkptr_t *bp = zio->io_bp;
408 spa_t *spa = zio->io_spa;
409 uint64_t dsobj = zio->io_bookmark.zb_objset;
410 uint64_t lsize = BP_GET_LSIZE(bp);
411 dmu_object_type_t ot = BP_GET_TYPE(bp);
412 uint8_t salt[ZIO_DATA_SALT_LEN];
413 uint8_t iv[ZIO_DATA_IV_LEN];
414 uint8_t mac[ZIO_DATA_MAC_LEN];
415 boolean_t no_crypt = B_FALSE;
417 ASSERT(BP_USES_CRYPT(bp));
418 ASSERT3U(size, !=, 0);
420 if (zio->io_error != 0)
424 * Verify the cksum of MACs stored in an indirect bp. It will always
425 * be possible to verify this since it does not require an encryption
428 if (BP_HAS_INDIRECT_MAC_CKSUM(bp)) {
429 zio_crypt_decode_mac_bp(bp, mac);
431 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF) {
433 * We haven't decompressed the data yet, but
434 * zio_crypt_do_indirect_mac_checksum() requires
435 * decompressed data to be able to parse out the MACs
436 * from the indirect block. We decompress it now and
437 * throw away the result after we are finished.
439 tmp = zio_buf_alloc(lsize);
440 ret = zio_decompress_data(BP_GET_COMPRESS(bp),
441 zio->io_abd, tmp, zio->io_size, lsize);
443 ret = SET_ERROR(EIO);
446 ret = zio_crypt_do_indirect_mac_checksum(B_FALSE,
447 tmp, lsize, BP_SHOULD_BYTESWAP(bp), mac);
448 zio_buf_free(tmp, lsize);
450 ret = zio_crypt_do_indirect_mac_checksum_abd(B_FALSE,
451 zio->io_abd, size, BP_SHOULD_BYTESWAP(bp), mac);
453 abd_copy(data, zio->io_abd, size);
455 if (zio_injection_enabled && ot != DMU_OT_DNODE && ret == 0) {
456 ret = zio_handle_decrypt_injection(spa,
457 &zio->io_bookmark, ot, ECKSUM);
466 * If this is an authenticated block, just check the MAC. It would be
467 * nice to separate this out into its own flag, but for the moment
468 * enum zio_flag is out of bits.
470 if (BP_IS_AUTHENTICATED(bp)) {
471 if (ot == DMU_OT_OBJSET) {
472 ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa,
473 dsobj, zio->io_abd, size, BP_SHOULD_BYTESWAP(bp));
475 zio_crypt_decode_mac_bp(bp, mac);
476 ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj,
477 zio->io_abd, size, mac);
478 if (zio_injection_enabled && ret == 0) {
479 ret = zio_handle_decrypt_injection(spa,
480 &zio->io_bookmark, ot, ECKSUM);
483 abd_copy(data, zio->io_abd, size);
491 zio_crypt_decode_params_bp(bp, salt, iv);
493 if (ot == DMU_OT_INTENT_LOG) {
494 tmp = abd_borrow_buf_copy(zio->io_abd, sizeof (zil_chain_t));
495 zio_crypt_decode_mac_zil(tmp, mac);
496 abd_return_buf(zio->io_abd, tmp, sizeof (zil_chain_t));
498 zio_crypt_decode_mac_bp(bp, mac);
501 ret = spa_do_crypt_abd(B_FALSE, spa, &zio->io_bookmark, BP_GET_TYPE(bp),
502 BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp), salt, iv, mac, size, data,
503 zio->io_abd, &no_crypt);
505 abd_copy(data, zio->io_abd, size);
513 /* assert that the key was found unless this was speculative */
514 ASSERT(ret != EACCES || (zio->io_flags & ZIO_FLAG_SPECULATIVE));
517 * If there was a decryption / authentication error return EIO as
518 * the io_error. If this was not a speculative zio, create an ereport.
521 zio->io_error = SET_ERROR(EIO);
522 if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) {
523 spa_log_error(spa, &zio->io_bookmark);
524 zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION,
525 spa, NULL, &zio->io_bookmark, zio, 0, 0);
533 * ==========================================================================
534 * I/O parent/child relationships and pipeline interlocks
535 * ==========================================================================
538 zio_walk_parents(zio_t *cio, zio_link_t **zl)
540 list_t *pl = &cio->io_parent_list;
542 *zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl);
546 ASSERT((*zl)->zl_child == cio);
547 return ((*zl)->zl_parent);
551 zio_walk_children(zio_t *pio, zio_link_t **zl)
553 list_t *cl = &pio->io_child_list;
555 ASSERT(MUTEX_HELD(&pio->io_lock));
557 *zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl);
561 ASSERT((*zl)->zl_parent == pio);
562 return ((*zl)->zl_child);
566 zio_unique_parent(zio_t *cio)
568 zio_link_t *zl = NULL;
569 zio_t *pio = zio_walk_parents(cio, &zl);
571 VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL);
576 zio_add_child(zio_t *pio, zio_t *cio)
578 zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
581 * Logical I/Os can have logical, gang, or vdev children.
582 * Gang I/Os can have gang or vdev children.
583 * Vdev I/Os can only have vdev children.
584 * The following ASSERT captures all of these constraints.
586 ASSERT3S(cio->io_child_type, <=, pio->io_child_type);
591 mutex_enter(&pio->io_lock);
592 mutex_enter(&cio->io_lock);
594 ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
596 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
597 pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
599 list_insert_head(&pio->io_child_list, zl);
600 list_insert_head(&cio->io_parent_list, zl);
602 pio->io_child_count++;
603 cio->io_parent_count++;
605 mutex_exit(&cio->io_lock);
606 mutex_exit(&pio->io_lock);
610 zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
612 ASSERT(zl->zl_parent == pio);
613 ASSERT(zl->zl_child == cio);
615 mutex_enter(&pio->io_lock);
616 mutex_enter(&cio->io_lock);
618 list_remove(&pio->io_child_list, zl);
619 list_remove(&cio->io_parent_list, zl);
621 pio->io_child_count--;
622 cio->io_parent_count--;
624 mutex_exit(&cio->io_lock);
625 mutex_exit(&pio->io_lock);
626 kmem_cache_free(zio_link_cache, zl);
630 zio_wait_for_children(zio_t *zio, uint8_t childbits, enum zio_wait_type wait)
632 boolean_t waiting = B_FALSE;
634 mutex_enter(&zio->io_lock);
635 ASSERT(zio->io_stall == NULL);
636 for (int c = 0; c < ZIO_CHILD_TYPES; c++) {
637 if (!(ZIO_CHILD_BIT_IS_SET(childbits, c)))
640 uint64_t *countp = &zio->io_children[c][wait];
643 ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN);
644 zio->io_stall = countp;
649 mutex_exit(&zio->io_lock);
653 __attribute__((always_inline))
655 zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait,
656 zio_t **next_to_executep)
658 uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
659 int *errorp = &pio->io_child_error[zio->io_child_type];
661 mutex_enter(&pio->io_lock);
662 if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
663 *errorp = zio_worst_error(*errorp, zio->io_error);
664 pio->io_reexecute |= zio->io_reexecute;
665 ASSERT3U(*countp, >, 0);
669 if (*countp == 0 && pio->io_stall == countp) {
670 zio_taskq_type_t type =
671 pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE :
673 pio->io_stall = NULL;
674 mutex_exit(&pio->io_lock);
677 * If we can tell the caller to execute this parent next, do
678 * so. Otherwise dispatch the parent zio as its own task.
680 * Having the caller execute the parent when possible reduces
681 * locking on the zio taskq's, reduces context switch
682 * overhead, and has no recursion penalty. Note that one
683 * read from disk typically causes at least 3 zio's: a
684 * zio_null(), the logical zio_read(), and then a physical
685 * zio. When the physical ZIO completes, we are able to call
686 * zio_done() on all 3 of these zio's from one invocation of
687 * zio_execute() by returning the parent back to
688 * zio_execute(). Since the parent isn't executed until this
689 * thread returns back to zio_execute(), the caller should do
692 * In other cases, dispatching the parent prevents
693 * overflowing the stack when we have deeply nested
694 * parent-child relationships, as we do with the "mega zio"
695 * of writes for spa_sync(), and the chain of ZIL blocks.
697 if (next_to_executep != NULL && *next_to_executep == NULL) {
698 *next_to_executep = pio;
700 zio_taskq_dispatch(pio, type, B_FALSE);
703 mutex_exit(&pio->io_lock);
708 zio_inherit_child_errors(zio_t *zio, enum zio_child c)
710 if (zio->io_child_error[c] != 0 && zio->io_error == 0)
711 zio->io_error = zio->io_child_error[c];
715 zio_bookmark_compare(const void *x1, const void *x2)
717 const zio_t *z1 = x1;
718 const zio_t *z2 = x2;
720 if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset)
722 if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset)
725 if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object)
727 if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object)
730 if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level)
732 if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level)
735 if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid)
737 if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid)
749 * ==========================================================================
750 * Create the various types of I/O (read, write, free, etc)
751 * ==========================================================================
754 zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
755 abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done,
756 void *private, zio_type_t type, zio_priority_t priority,
757 enum zio_flag flags, vdev_t *vd, uint64_t offset,
758 const zbookmark_phys_t *zb, enum zio_stage stage,
759 enum zio_stage pipeline)
763 ASSERT3U(psize, <=, SPA_MAXBLOCKSIZE);
764 ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0);
765 ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
767 ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
768 ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
769 ASSERT(vd || stage == ZIO_STAGE_OPEN);
771 IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW_COMPRESS) != 0);
773 zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
774 bzero(zio, sizeof (zio_t));
776 mutex_init(&zio->io_lock, NULL, MUTEX_NOLOCKDEP, NULL);
777 cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
779 list_create(&zio->io_parent_list, sizeof (zio_link_t),
780 offsetof(zio_link_t, zl_parent_node));
781 list_create(&zio->io_child_list, sizeof (zio_link_t),
782 offsetof(zio_link_t, zl_child_node));
783 metaslab_trace_init(&zio->io_alloc_list);
786 zio->io_child_type = ZIO_CHILD_VDEV;
787 else if (flags & ZIO_FLAG_GANG_CHILD)
788 zio->io_child_type = ZIO_CHILD_GANG;
789 else if (flags & ZIO_FLAG_DDT_CHILD)
790 zio->io_child_type = ZIO_CHILD_DDT;
792 zio->io_child_type = ZIO_CHILD_LOGICAL;
795 zio->io_bp = (blkptr_t *)bp;
796 zio->io_bp_copy = *bp;
797 zio->io_bp_orig = *bp;
798 if (type != ZIO_TYPE_WRITE ||
799 zio->io_child_type == ZIO_CHILD_DDT)
800 zio->io_bp = &zio->io_bp_copy; /* so caller can free */
801 if (zio->io_child_type == ZIO_CHILD_LOGICAL)
802 zio->io_logical = zio;
803 if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
804 pipeline |= ZIO_GANG_STAGES;
810 zio->io_private = private;
812 zio->io_priority = priority;
814 zio->io_offset = offset;
815 zio->io_orig_abd = zio->io_abd = data;
816 zio->io_orig_size = zio->io_size = psize;
817 zio->io_lsize = lsize;
818 zio->io_orig_flags = zio->io_flags = flags;
819 zio->io_orig_stage = zio->io_stage = stage;
820 zio->io_orig_pipeline = zio->io_pipeline = pipeline;
821 zio->io_pipeline_trace = ZIO_STAGE_OPEN;
823 zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
824 zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
827 zio->io_bookmark = *zb;
830 if (zio->io_metaslab_class == NULL)
831 zio->io_metaslab_class = pio->io_metaslab_class;
832 if (zio->io_logical == NULL)
833 zio->io_logical = pio->io_logical;
834 if (zio->io_child_type == ZIO_CHILD_GANG)
835 zio->io_gang_leader = pio->io_gang_leader;
836 zio_add_child(pio, zio);
839 taskq_init_ent(&zio->io_tqent);
845 zio_destroy(zio_t *zio)
847 metaslab_trace_fini(&zio->io_alloc_list);
848 list_destroy(&zio->io_parent_list);
849 list_destroy(&zio->io_child_list);
850 mutex_destroy(&zio->io_lock);
851 cv_destroy(&zio->io_cv);
852 kmem_cache_free(zio_cache, zio);
856 zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
857 void *private, enum zio_flag flags)
861 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
862 ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
863 ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
869 zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
871 return (zio_null(NULL, spa, NULL, done, private, flags));
875 zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
877 if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
878 zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
879 bp, (longlong_t)BP_GET_TYPE(bp));
881 if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
882 BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
883 zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
884 bp, (longlong_t)BP_GET_CHECKSUM(bp));
886 if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
887 BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
888 zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
889 bp, (longlong_t)BP_GET_COMPRESS(bp));
891 if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
892 zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
893 bp, (longlong_t)BP_GET_LSIZE(bp));
895 if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
896 zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
897 bp, (longlong_t)BP_GET_PSIZE(bp));
900 if (BP_IS_EMBEDDED(bp)) {
901 if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
902 zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
903 bp, (longlong_t)BPE_GET_ETYPE(bp));
908 * Do not verify individual DVAs if the config is not trusted. This
909 * will be done once the zio is executed in vdev_mirror_map_alloc.
911 if (!spa->spa_trust_config)
915 * Pool-specific checks.
917 * Note: it would be nice to verify that the blk_birth and
918 * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
919 * allows the birth time of log blocks (and dmu_sync()-ed blocks
920 * that are in the log) to be arbitrarily large.
922 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
923 uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
925 if (vdevid >= spa->spa_root_vdev->vdev_children) {
926 zfs_panic_recover("blkptr at %p DVA %u has invalid "
928 bp, i, (longlong_t)vdevid);
931 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
933 zfs_panic_recover("blkptr at %p DVA %u has invalid "
935 bp, i, (longlong_t)vdevid);
938 if (vd->vdev_ops == &vdev_hole_ops) {
939 zfs_panic_recover("blkptr at %p DVA %u has hole "
941 bp, i, (longlong_t)vdevid);
944 if (vd->vdev_ops == &vdev_missing_ops) {
946 * "missing" vdevs are valid during import, but we
947 * don't have their detailed info (e.g. asize), so
948 * we can't perform any more checks on them.
952 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
953 uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
955 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
956 if (offset + asize > vd->vdev_asize) {
957 zfs_panic_recover("blkptr at %p DVA %u has invalid "
959 bp, i, (longlong_t)offset);
965 zfs_dva_valid(spa_t *spa, const dva_t *dva, const blkptr_t *bp)
967 uint64_t vdevid = DVA_GET_VDEV(dva);
969 if (vdevid >= spa->spa_root_vdev->vdev_children)
972 vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid];
976 if (vd->vdev_ops == &vdev_hole_ops)
979 if (vd->vdev_ops == &vdev_missing_ops) {
983 uint64_t offset = DVA_GET_OFFSET(dva);
984 uint64_t asize = DVA_GET_ASIZE(dva);
987 asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
988 if (offset + asize > vd->vdev_asize)
995 zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
996 abd_t *data, uint64_t size, zio_done_func_t *done, void *private,
997 zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
1001 zfs_blkptr_verify(spa, bp);
1003 zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
1004 data, size, size, done, private,
1005 ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
1006 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
1007 ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
1013 zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
1014 abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp,
1015 zio_done_func_t *ready, zio_done_func_t *children_ready,
1016 zio_done_func_t *physdone, zio_done_func_t *done,
1017 void *private, zio_priority_t priority, enum zio_flag flags,
1018 const zbookmark_phys_t *zb)
1022 ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
1023 zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
1024 zp->zp_compress >= ZIO_COMPRESS_OFF &&
1025 zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
1026 DMU_OT_IS_VALID(zp->zp_type) &&
1027 zp->zp_level < 32 &&
1028 zp->zp_copies > 0 &&
1029 zp->zp_copies <= spa_max_replication(spa));
1031 zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private,
1032 ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
1033 ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
1034 ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
1036 zio->io_ready = ready;
1037 zio->io_children_ready = children_ready;
1038 zio->io_physdone = physdone;
1042 * Data can be NULL if we are going to call zio_write_override() to
1043 * provide the already-allocated BP. But we may need the data to
1044 * verify a dedup hit (if requested). In this case, don't try to
1045 * dedup (just take the already-allocated BP verbatim). Encrypted
1046 * dedup blocks need data as well so we also disable dedup in this
1050 (zio->io_prop.zp_dedup_verify || zio->io_prop.zp_encrypt)) {
1051 zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
1058 zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data,
1059 uint64_t size, zio_done_func_t *done, void *private,
1060 zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
1064 zio = zio_create(pio, spa, txg, bp, data, size, size, done, private,
1065 ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb,
1066 ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
1072 zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
1074 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
1075 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1076 ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
1077 ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
1080 * We must reset the io_prop to match the values that existed
1081 * when the bp was first written by dmu_sync() keeping in mind
1082 * that nopwrite and dedup are mutually exclusive.
1084 zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
1085 zio->io_prop.zp_nopwrite = nopwrite;
1086 zio->io_prop.zp_copies = copies;
1087 zio->io_bp_override = bp;
1091 zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
1094 zfs_blkptr_verify(spa, bp);
1097 * The check for EMBEDDED is a performance optimization. We
1098 * process the free here (by ignoring it) rather than
1099 * putting it on the list and then processing it in zio_free_sync().
1101 if (BP_IS_EMBEDDED(bp))
1103 metaslab_check_free(spa, bp);
1106 * Frees that are for the currently-syncing txg, are not going to be
1107 * deferred, and which will not need to do a read (i.e. not GANG or
1108 * DEDUP), can be processed immediately. Otherwise, put them on the
1109 * in-memory list for later processing.
1111 if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
1112 txg != spa->spa_syncing_txg ||
1113 spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
1114 bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
1116 VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 0)));
1121 zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1122 enum zio_flag flags)
1125 enum zio_stage stage = ZIO_FREE_PIPELINE;
1127 ASSERT(!BP_IS_HOLE(bp));
1128 ASSERT(spa_syncing_txg(spa) == txg);
1129 ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
1131 if (BP_IS_EMBEDDED(bp))
1132 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
1134 metaslab_check_free(spa, bp);
1136 dsl_scan_freed(spa, bp);
1139 * GANG and DEDUP blocks can induce a read (for the gang block header,
1140 * or the DDT), so issue them asynchronously so that this thread is
1143 if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
1144 stage |= ZIO_STAGE_ISSUE_ASYNC;
1146 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1147 BP_GET_PSIZE(bp), NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW,
1148 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
1154 zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
1155 zio_done_func_t *done, void *private, enum zio_flag flags)
1159 zfs_blkptr_verify(spa, bp);
1161 if (BP_IS_EMBEDDED(bp))
1162 return (zio_null(pio, spa, NULL, NULL, NULL, 0));
1165 * A claim is an allocation of a specific block. Claims are needed
1166 * to support immediate writes in the intent log. The issue is that
1167 * immediate writes contain committed data, but in a txg that was
1168 * *not* committed. Upon opening the pool after an unclean shutdown,
1169 * the intent log claims all blocks that contain immediate write data
1170 * so that the SPA knows they're in use.
1172 * All claims *must* be resolved in the first txg -- before the SPA
1173 * starts allocating blocks -- so that nothing is allocated twice.
1174 * If txg == 0 we just verify that the block is claimable.
1176 ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <,
1177 spa_min_claim_txg(spa));
1178 ASSERT(txg == spa_min_claim_txg(spa) || txg == 0);
1179 ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
1181 zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
1182 BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW,
1183 flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
1184 ASSERT0(zio->io_queued_timestamp);
1190 zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
1191 zio_done_func_t *done, void *private, enum zio_flag flags)
1196 if (vd->vdev_children == 0) {
1197 zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
1198 ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
1199 ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
1203 zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
1205 for (c = 0; c < vd->vdev_children; c++)
1206 zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
1207 done, private, flags));
1214 zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1215 abd_t *data, int checksum, zio_done_func_t *done, void *private,
1216 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1220 ASSERT(vd->vdev_children == 0);
1221 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1222 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1223 ASSERT3U(offset + size, <=, vd->vdev_psize);
1225 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1226 private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1227 offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
1229 zio->io_prop.zp_checksum = checksum;
1235 zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
1236 abd_t *data, int checksum, zio_done_func_t *done, void *private,
1237 zio_priority_t priority, enum zio_flag flags, boolean_t labels)
1241 ASSERT(vd->vdev_children == 0);
1242 ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
1243 offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
1244 ASSERT3U(offset + size, <=, vd->vdev_psize);
1246 zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
1247 private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd,
1248 offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
1250 zio->io_prop.zp_checksum = checksum;
1252 if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
1254 * zec checksums are necessarily destructive -- they modify
1255 * the end of the write buffer to hold the verifier/checksum.
1256 * Therefore, we must make a local copy in case the data is
1257 * being written to multiple places in parallel.
1259 abd_t *wbuf = abd_alloc_sametype(data, size);
1260 abd_copy(wbuf, data, size);
1262 zio_push_transform(zio, wbuf, size, size, NULL);
1269 * Create a child I/O to do some work for us.
1272 zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
1273 abd_t *data, uint64_t size, int type, zio_priority_t priority,
1274 enum zio_flag flags, zio_done_func_t *done, void *private)
1276 enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
1280 * vdev child I/Os do not propagate their error to the parent.
1281 * Therefore, for correct operation the caller *must* check for
1282 * and handle the error in the child i/o's done callback.
1283 * The only exceptions are i/os that we don't care about
1284 * (OPTIONAL or REPAIR).
1286 ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) ||
1289 if (type == ZIO_TYPE_READ && bp != NULL) {
1291 * If we have the bp, then the child should perform the
1292 * checksum and the parent need not. This pushes error
1293 * detection as close to the leaves as possible and
1294 * eliminates redundant checksums in the interior nodes.
1296 pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
1297 pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
1300 if (vd->vdev_ops->vdev_op_leaf) {
1301 ASSERT0(vd->vdev_children);
1302 offset += VDEV_LABEL_START_SIZE;
1305 flags |= ZIO_VDEV_CHILD_FLAGS(pio);
1308 * If we've decided to do a repair, the write is not speculative --
1309 * even if the original read was.
1311 if (flags & ZIO_FLAG_IO_REPAIR)
1312 flags &= ~ZIO_FLAG_SPECULATIVE;
1315 * If we're creating a child I/O that is not associated with a
1316 * top-level vdev, then the child zio is not an allocating I/O.
1317 * If this is a retried I/O then we ignore it since we will
1318 * have already processed the original allocating I/O.
1320 if (flags & ZIO_FLAG_IO_ALLOCATING &&
1321 (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) {
1322 ASSERT(pio->io_metaslab_class != NULL);
1323 ASSERT(pio->io_metaslab_class->mc_alloc_throttle_enabled);
1324 ASSERT(type == ZIO_TYPE_WRITE);
1325 ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE);
1326 ASSERT(!(flags & ZIO_FLAG_IO_REPAIR));
1327 ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) ||
1328 pio->io_child_type == ZIO_CHILD_GANG);
1330 flags &= ~ZIO_FLAG_IO_ALLOCATING;
1334 zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size,
1335 done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
1336 ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
1337 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
1339 zio->io_physdone = pio->io_physdone;
1340 if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
1341 zio->io_logical->io_phys_children++;
1347 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size,
1348 zio_type_t type, zio_priority_t priority, enum zio_flag flags,
1349 zio_done_func_t *done, void *private)
1353 ASSERT(vd->vdev_ops->vdev_op_leaf);
1355 zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
1356 data, size, size, done, private, type, priority,
1357 flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
1359 ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
1365 zio_flush(zio_t *zio, vdev_t *vd)
1367 zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
1369 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
1373 zio_shrink(zio_t *zio, uint64_t size)
1375 ASSERT3P(zio->io_executor, ==, NULL);
1376 ASSERT3U(zio->io_orig_size, ==, zio->io_size);
1377 ASSERT3U(size, <=, zio->io_size);
1380 * We don't shrink for raidz because of problems with the
1381 * reconstruction when reading back less than the block size.
1382 * Note, BP_IS_RAIDZ() assumes no compression.
1384 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1385 if (!BP_IS_RAIDZ(zio->io_bp)) {
1386 /* we are not doing a raw write */
1387 ASSERT3U(zio->io_size, ==, zio->io_lsize);
1388 zio->io_orig_size = zio->io_size = zio->io_lsize = size;
1393 * ==========================================================================
1394 * Prepare to read and write logical blocks
1395 * ==========================================================================
1399 zio_read_bp_init(zio_t *zio)
1401 blkptr_t *bp = zio->io_bp;
1403 BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
1405 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1407 if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
1408 zio->io_child_type == ZIO_CHILD_LOGICAL &&
1409 !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) {
1410 zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1411 psize, psize, zio_decompress);
1414 if (((BP_IS_PROTECTED(bp) && !(zio->io_flags & ZIO_FLAG_RAW_ENCRYPT)) ||
1415 BP_HAS_INDIRECT_MAC_CKSUM(bp)) &&
1416 zio->io_child_type == ZIO_CHILD_LOGICAL) {
1417 zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize),
1418 psize, psize, zio_decrypt);
1421 if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
1422 int psize = BPE_GET_PSIZE(bp);
1423 void *data = abd_borrow_buf(zio->io_abd, psize);
1425 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1426 decode_embedded_bp_compressed(bp, data);
1427 abd_return_buf_copy(zio->io_abd, data, psize);
1429 ASSERT(!BP_IS_EMBEDDED(bp));
1430 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1433 if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
1434 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1436 if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
1437 zio->io_flags |= ZIO_FLAG_DONT_CACHE;
1439 if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
1440 zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
1446 zio_write_bp_init(zio_t *zio)
1448 if (!IO_IS_ALLOCATING(zio))
1451 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1453 if (zio->io_bp_override) {
1454 blkptr_t *bp = zio->io_bp;
1455 zio_prop_t *zp = &zio->io_prop;
1457 ASSERT(bp->blk_birth != zio->io_txg);
1458 ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
1460 *bp = *zio->io_bp_override;
1461 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1463 if (BP_IS_EMBEDDED(bp))
1467 * If we've been overridden and nopwrite is set then
1468 * set the flag accordingly to indicate that a nopwrite
1469 * has already occurred.
1471 if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
1472 ASSERT(!zp->zp_dedup);
1473 ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum);
1474 zio->io_flags |= ZIO_FLAG_NOPWRITE;
1478 ASSERT(!zp->zp_nopwrite);
1480 if (BP_IS_HOLE(bp) || !zp->zp_dedup)
1483 ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags &
1484 ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify);
1486 if (BP_GET_CHECKSUM(bp) == zp->zp_checksum &&
1488 BP_SET_DEDUP(bp, 1);
1489 zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
1494 * We were unable to handle this as an override bp, treat
1495 * it as a regular write I/O.
1497 zio->io_bp_override = NULL;
1498 *bp = zio->io_bp_orig;
1499 zio->io_pipeline = zio->io_orig_pipeline;
1506 zio_write_compress(zio_t *zio)
1508 spa_t *spa = zio->io_spa;
1509 zio_prop_t *zp = &zio->io_prop;
1510 enum zio_compress compress = zp->zp_compress;
1511 blkptr_t *bp = zio->io_bp;
1512 uint64_t lsize = zio->io_lsize;
1513 uint64_t psize = zio->io_size;
1517 * If our children haven't all reached the ready stage,
1518 * wait for them and then repeat this pipeline stage.
1520 if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT |
1521 ZIO_CHILD_GANG_BIT, ZIO_WAIT_READY)) {
1525 if (!IO_IS_ALLOCATING(zio))
1528 if (zio->io_children_ready != NULL) {
1530 * Now that all our children are ready, run the callback
1531 * associated with this zio in case it wants to modify the
1532 * data to be written.
1534 ASSERT3U(zp->zp_level, >, 0);
1535 zio->io_children_ready(zio);
1538 ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
1539 ASSERT(zio->io_bp_override == NULL);
1541 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
1543 * We're rewriting an existing block, which means we're
1544 * working on behalf of spa_sync(). For spa_sync() to
1545 * converge, it must eventually be the case that we don't
1546 * have to allocate new blocks. But compression changes
1547 * the blocksize, which forces a reallocate, and makes
1548 * convergence take longer. Therefore, after the first
1549 * few passes, stop compressing to ensure convergence.
1551 pass = spa_sync_pass(spa);
1553 ASSERT(zio->io_txg == spa_syncing_txg(spa));
1554 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1555 ASSERT(!BP_GET_DEDUP(bp));
1557 if (pass >= zfs_sync_pass_dont_compress)
1558 compress = ZIO_COMPRESS_OFF;
1560 /* Make sure someone doesn't change their mind on overwrites */
1561 ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
1562 spa_max_replication(spa)) == BP_GET_NDVAS(bp));
1565 /* If it's a compressed write that is not raw, compress the buffer. */
1566 if (compress != ZIO_COMPRESS_OFF &&
1567 !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) {
1568 void *cbuf = zio_buf_alloc(lsize);
1569 psize = zio_compress_data(compress, zio->io_abd, cbuf, lsize);
1570 if (psize == 0 || psize == lsize) {
1571 compress = ZIO_COMPRESS_OFF;
1572 zio_buf_free(cbuf, lsize);
1573 } else if (!zp->zp_dedup && !zp->zp_encrypt &&
1574 psize <= BPE_PAYLOAD_SIZE &&
1575 zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
1576 spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
1577 encode_embedded_bp_compressed(bp,
1578 cbuf, compress, lsize, psize);
1579 BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
1580 BP_SET_TYPE(bp, zio->io_prop.zp_type);
1581 BP_SET_LEVEL(bp, zio->io_prop.zp_level);
1582 zio_buf_free(cbuf, lsize);
1583 bp->blk_birth = zio->io_txg;
1584 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1585 ASSERT(spa_feature_is_active(spa,
1586 SPA_FEATURE_EMBEDDED_DATA));
1590 * Round up compressed size up to the ashift
1591 * of the smallest-ashift device, and zero the tail.
1592 * This ensures that the compressed size of the BP
1593 * (and thus compressratio property) are correct,
1594 * in that we charge for the padding used to fill out
1597 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
1598 size_t rounded = (size_t)P2ROUNDUP(psize,
1599 1ULL << spa->spa_min_ashift);
1600 if (rounded >= lsize) {
1601 compress = ZIO_COMPRESS_OFF;
1602 zio_buf_free(cbuf, lsize);
1605 abd_t *cdata = abd_get_from_buf(cbuf, lsize);
1606 abd_take_ownership_of_buf(cdata, B_TRUE);
1607 abd_zero_off(cdata, psize, rounded - psize);
1609 zio_push_transform(zio, cdata,
1610 psize, lsize, NULL);
1615 * We were unable to handle this as an override bp, treat
1616 * it as a regular write I/O.
1618 zio->io_bp_override = NULL;
1619 *bp = zio->io_bp_orig;
1620 zio->io_pipeline = zio->io_orig_pipeline;
1622 } else if ((zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) != 0 &&
1623 zp->zp_type == DMU_OT_DNODE) {
1625 * The DMU actually relies on the zio layer's compression
1626 * to free metadnode blocks that have had all contained
1627 * dnodes freed. As a result, even when doing a raw
1628 * receive, we must check whether the block can be compressed
1631 psize = zio_compress_data(ZIO_COMPRESS_EMPTY,
1632 zio->io_abd, NULL, lsize);
1634 compress = ZIO_COMPRESS_OFF;
1636 ASSERT3U(psize, !=, 0);
1640 * The final pass of spa_sync() must be all rewrites, but the first
1641 * few passes offer a trade-off: allocating blocks defers convergence,
1642 * but newly allocated blocks are sequential, so they can be written
1643 * to disk faster. Therefore, we allow the first few passes of
1644 * spa_sync() to allocate new blocks, but force rewrites after that.
1645 * There should only be a handful of blocks after pass 1 in any case.
1647 if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
1648 BP_GET_PSIZE(bp) == psize &&
1649 pass >= zfs_sync_pass_rewrite) {
1650 VERIFY3U(psize, !=, 0);
1651 enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
1653 zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
1654 zio->io_flags |= ZIO_FLAG_IO_REWRITE;
1657 zio->io_pipeline = ZIO_WRITE_PIPELINE;
1661 if (zio->io_bp_orig.blk_birth != 0 &&
1662 spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
1663 BP_SET_LSIZE(bp, lsize);
1664 BP_SET_TYPE(bp, zp->zp_type);
1665 BP_SET_LEVEL(bp, zp->zp_level);
1666 BP_SET_BIRTH(bp, zio->io_txg, 0);
1668 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
1670 ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
1671 BP_SET_LSIZE(bp, lsize);
1672 BP_SET_TYPE(bp, zp->zp_type);
1673 BP_SET_LEVEL(bp, zp->zp_level);
1674 BP_SET_PSIZE(bp, psize);
1675 BP_SET_COMPRESS(bp, compress);
1676 BP_SET_CHECKSUM(bp, zp->zp_checksum);
1677 BP_SET_DEDUP(bp, zp->zp_dedup);
1678 BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
1680 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1681 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1682 ASSERT(!zp->zp_encrypt ||
1683 DMU_OT_IS_ENCRYPTED(zp->zp_type));
1684 zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
1686 if (zp->zp_nopwrite) {
1687 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
1688 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
1689 zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
1696 zio_free_bp_init(zio_t *zio)
1698 blkptr_t *bp = zio->io_bp;
1700 if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
1701 if (BP_GET_DEDUP(bp))
1702 zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
1705 ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy);
1711 * ==========================================================================
1712 * Execute the I/O pipeline
1713 * ==========================================================================
1717 zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
1719 spa_t *spa = zio->io_spa;
1720 zio_type_t t = zio->io_type;
1721 int flags = (cutinline ? TQ_FRONT : 0);
1724 * If we're a config writer or a probe, the normal issue and
1725 * interrupt threads may all be blocked waiting for the config lock.
1726 * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
1728 if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
1732 * A similar issue exists for the L2ARC write thread until L2ARC 2.0.
1734 if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
1738 * If this is a high priority I/O, then use the high priority taskq if
1741 if (zio->io_priority == ZIO_PRIORITY_NOW &&
1742 spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
1745 ASSERT3U(q, <, ZIO_TASKQ_TYPES);
1748 * NB: We are assuming that the zio can only be dispatched
1749 * to a single taskq at a time. It would be a grievous error
1750 * to dispatch the zio to another taskq at the same time.
1752 ASSERT(taskq_empty_ent(&zio->io_tqent));
1753 spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
1754 flags, &zio->io_tqent);
1758 zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
1760 kthread_t *executor = zio->io_executor;
1761 spa_t *spa = zio->io_spa;
1763 for (zio_type_t t = 0; t < ZIO_TYPES; t++) {
1764 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1766 for (i = 0; i < tqs->stqs_count; i++) {
1767 if (taskq_member(tqs->stqs_taskq[i], executor))
1776 zio_issue_async(zio_t *zio)
1778 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
1784 zio_interrupt(zio_t *zio)
1786 zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
1790 zio_delay_interrupt(zio_t *zio)
1793 * The timeout_generic() function isn't defined in userspace, so
1794 * rather than trying to implement the function, the zio delay
1795 * functionality has been disabled for userspace builds.
1800 * If io_target_timestamp is zero, then no delay has been registered
1801 * for this IO, thus jump to the end of this function and "skip" the
1802 * delay; issuing it directly to the zio layer.
1804 if (zio->io_target_timestamp != 0) {
1805 hrtime_t now = gethrtime();
1807 if (now >= zio->io_target_timestamp) {
1809 * This IO has already taken longer than the target
1810 * delay to complete, so we don't want to delay it
1811 * any longer; we "miss" the delay and issue it
1812 * directly to the zio layer. This is likely due to
1813 * the target latency being set to a value less than
1814 * the underlying hardware can satisfy (e.g. delay
1815 * set to 1ms, but the disks take 10ms to complete an
1819 DTRACE_PROBE2(zio__delay__miss, zio_t *, zio,
1825 hrtime_t diff = zio->io_target_timestamp - now;
1826 clock_t expire_at_tick = ddi_get_lbolt() +
1829 DTRACE_PROBE3(zio__delay__hit, zio_t *, zio,
1830 hrtime_t, now, hrtime_t, diff);
1832 if (NSEC_TO_TICK(diff) == 0) {
1833 /* Our delay is less than a jiffy - just spin */
1834 zfs_sleep_until(zio->io_target_timestamp);
1837 * Use taskq_dispatch_delay() in the place of
1838 * OpenZFS's timeout_generic().
1840 tid = taskq_dispatch_delay(system_taskq,
1841 (task_func_t *)zio_interrupt,
1842 zio, TQ_NOSLEEP, expire_at_tick);
1843 if (tid == TASKQID_INVALID) {
1845 * Couldn't allocate a task. Just
1846 * finish the zio without a delay.
1855 DTRACE_PROBE1(zio__delay__skip, zio_t *, zio);
1860 zio_deadman_impl(zio_t *pio)
1862 zio_t *cio, *cio_next;
1863 zio_link_t *zl = NULL;
1864 vdev_t *vd = pio->io_vd;
1866 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
1867 vdev_queue_t *vq = &vd->vdev_queue;
1868 zbookmark_phys_t *zb = &pio->io_bookmark;
1869 uint64_t delta = gethrtime() - pio->io_timestamp;
1870 uint64_t failmode = spa_get_deadman_failmode(pio->io_spa);
1872 zfs_dbgmsg("slow zio: zio=%p timestamp=%llu "
1873 "delta=%llu queued=%llu io=%llu "
1874 "path=%s last=%llu "
1875 "type=%d priority=%d flags=0x%x "
1876 "stage=0x%x pipeline=0x%x pipeline-trace=0x%x "
1877 "objset=%llu object=%llu level=%llu blkid=%llu "
1878 "offset=%llu size=%llu error=%d",
1879 pio, pio->io_timestamp,
1880 delta, pio->io_delta, pio->io_delay,
1881 vd->vdev_path, vq->vq_io_complete_ts,
1882 pio->io_type, pio->io_priority, pio->io_flags,
1883 pio->io_state, pio->io_pipeline, pio->io_pipeline_trace,
1884 zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
1885 pio->io_offset, pio->io_size, pio->io_error);
1886 zfs_ereport_post(FM_EREPORT_ZFS_DEADMAN,
1887 pio->io_spa, vd, zb, pio, 0, 0);
1889 if (failmode == ZIO_FAILURE_MODE_CONTINUE &&
1890 taskq_empty_ent(&pio->io_tqent)) {
1895 mutex_enter(&pio->io_lock);
1896 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
1897 cio_next = zio_walk_children(pio, &zl);
1898 zio_deadman_impl(cio);
1900 mutex_exit(&pio->io_lock);
1904 * Log the critical information describing this zio and all of its children
1905 * using the zfs_dbgmsg() interface then post deadman event for the ZED.
1908 zio_deadman(zio_t *pio, char *tag)
1910 spa_t *spa = pio->io_spa;
1911 char *name = spa_name(spa);
1913 if (!zfs_deadman_enabled || spa_suspended(spa))
1916 zio_deadman_impl(pio);
1918 switch (spa_get_deadman_failmode(spa)) {
1919 case ZIO_FAILURE_MODE_WAIT:
1920 zfs_dbgmsg("%s waiting for hung I/O to pool '%s'", tag, name);
1923 case ZIO_FAILURE_MODE_CONTINUE:
1924 zfs_dbgmsg("%s restarting hung I/O for pool '%s'", tag, name);
1927 case ZIO_FAILURE_MODE_PANIC:
1928 fm_panic("%s determined I/O to pool '%s' is hung.", tag, name);
1934 * Execute the I/O pipeline until one of the following occurs:
1935 * (1) the I/O completes; (2) the pipeline stalls waiting for
1936 * dependent child I/Os; (3) the I/O issues, so we're waiting
1937 * for an I/O completion interrupt; (4) the I/O is delegated by
1938 * vdev-level caching or aggregation; (5) the I/O is deferred
1939 * due to vdev-level queueing; (6) the I/O is handed off to
1940 * another thread. In all cases, the pipeline stops whenever
1941 * there's no CPU work; it never burns a thread in cv_wait_io().
1943 * There's no locking on io_stage because there's no legitimate way
1944 * for multiple threads to be attempting to process the same I/O.
1946 static zio_pipe_stage_t *zio_pipeline[];
1949 * zio_execute() is a wrapper around the static function
1950 * __zio_execute() so that we can force __zio_execute() to be
1951 * inlined. This reduces stack overhead which is important
1952 * because __zio_execute() is called recursively in several zio
1953 * code paths. zio_execute() itself cannot be inlined because
1954 * it is externally visible.
1957 zio_execute(zio_t *zio)
1959 fstrans_cookie_t cookie;
1961 cookie = spl_fstrans_mark();
1963 spl_fstrans_unmark(cookie);
1967 * Used to determine if in the current context the stack is sized large
1968 * enough to allow zio_execute() to be called recursively. A minimum
1969 * stack size of 16K is required to avoid needing to re-dispatch the zio.
1972 zio_execute_stack_check(zio_t *zio)
1974 #if !defined(HAVE_LARGE_STACKS)
1975 dsl_pool_t *dp = spa_get_dsl(zio->io_spa);
1977 /* Executing in txg_sync_thread() context. */
1978 if (dp && curthread == dp->dp_tx.tx_sync_thread)
1981 /* Pool initialization outside of zio_taskq context. */
1982 if (dp && spa_is_initializing(dp->dp_spa) &&
1983 !zio_taskq_member(zio, ZIO_TASKQ_ISSUE) &&
1984 !zio_taskq_member(zio, ZIO_TASKQ_ISSUE_HIGH))
1986 #endif /* HAVE_LARGE_STACKS */
1991 __attribute__((always_inline))
1993 __zio_execute(zio_t *zio)
1995 ASSERT3U(zio->io_queued_timestamp, >, 0);
1997 while (zio->io_stage < ZIO_STAGE_DONE) {
1998 enum zio_stage pipeline = zio->io_pipeline;
1999 enum zio_stage stage = zio->io_stage;
2001 zio->io_executor = curthread;
2003 ASSERT(!MUTEX_HELD(&zio->io_lock));
2004 ASSERT(ISP2(stage));
2005 ASSERT(zio->io_stall == NULL);
2009 } while ((stage & pipeline) == 0);
2011 ASSERT(stage <= ZIO_STAGE_DONE);
2014 * If we are in interrupt context and this pipeline stage
2015 * will grab a config lock that is held across I/O,
2016 * or may wait for an I/O that needs an interrupt thread
2017 * to complete, issue async to avoid deadlock.
2019 * For VDEV_IO_START, we cut in line so that the io will
2020 * be sent to disk promptly.
2022 if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
2023 zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
2024 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
2025 zio_requeue_io_start_cut_in_line : B_FALSE;
2026 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
2031 * If the current context doesn't have large enough stacks
2032 * the zio must be issued asynchronously to prevent overflow.
2034 if (zio_execute_stack_check(zio)) {
2035 boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
2036 zio_requeue_io_start_cut_in_line : B_FALSE;
2037 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
2041 zio->io_stage = stage;
2042 zio->io_pipeline_trace |= zio->io_stage;
2045 * The zio pipeline stage returns the next zio to execute
2046 * (typically the same as this one), or NULL if we should
2049 zio = zio_pipeline[highbit64(stage) - 1](zio);
2058 * ==========================================================================
2059 * Initiate I/O, either sync or async
2060 * ==========================================================================
2063 zio_wait(zio_t *zio)
2065 long timeout = MSEC_TO_TICK(zfs_deadman_ziotime_ms);
2068 ASSERT3S(zio->io_stage, ==, ZIO_STAGE_OPEN);
2069 ASSERT3P(zio->io_executor, ==, NULL);
2071 zio->io_waiter = curthread;
2072 ASSERT0(zio->io_queued_timestamp);
2073 zio->io_queued_timestamp = gethrtime();
2077 mutex_enter(&zio->io_lock);
2078 while (zio->io_executor != NULL) {
2079 error = cv_timedwait_io(&zio->io_cv, &zio->io_lock,
2080 ddi_get_lbolt() + timeout);
2082 if (zfs_deadman_enabled && error == -1 &&
2083 gethrtime() - zio->io_queued_timestamp >
2084 spa_deadman_ziotime(zio->io_spa)) {
2085 mutex_exit(&zio->io_lock);
2086 timeout = MSEC_TO_TICK(zfs_deadman_checktime_ms);
2087 zio_deadman(zio, FTAG);
2088 mutex_enter(&zio->io_lock);
2091 mutex_exit(&zio->io_lock);
2093 error = zio->io_error;
2100 zio_nowait(zio_t *zio)
2102 ASSERT3P(zio->io_executor, ==, NULL);
2104 if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
2105 zio_unique_parent(zio) == NULL) {
2109 * This is a logical async I/O with no parent to wait for it.
2110 * We add it to the spa_async_root_zio "Godfather" I/O which
2111 * will ensure they complete prior to unloading the pool.
2113 spa_t *spa = zio->io_spa;
2115 pio = spa->spa_async_zio_root[CPU_SEQID];
2118 zio_add_child(pio, zio);
2121 ASSERT0(zio->io_queued_timestamp);
2122 zio->io_queued_timestamp = gethrtime();
2127 * ==========================================================================
2128 * Reexecute, cancel, or suspend/resume failed I/O
2129 * ==========================================================================
2133 zio_reexecute(zio_t *pio)
2135 zio_t *cio, *cio_next;
2137 ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
2138 ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
2139 ASSERT(pio->io_gang_leader == NULL);
2140 ASSERT(pio->io_gang_tree == NULL);
2142 pio->io_flags = pio->io_orig_flags;
2143 pio->io_stage = pio->io_orig_stage;
2144 pio->io_pipeline = pio->io_orig_pipeline;
2145 pio->io_reexecute = 0;
2146 pio->io_flags |= ZIO_FLAG_REEXECUTED;
2147 pio->io_pipeline_trace = 0;
2149 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2150 pio->io_state[w] = 0;
2151 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
2152 pio->io_child_error[c] = 0;
2154 if (IO_IS_ALLOCATING(pio))
2155 BP_ZERO(pio->io_bp);
2158 * As we reexecute pio's children, new children could be created.
2159 * New children go to the head of pio's io_child_list, however,
2160 * so we will (correctly) not reexecute them. The key is that
2161 * the remainder of pio's io_child_list, from 'cio_next' onward,
2162 * cannot be affected by any side effects of reexecuting 'cio'.
2164 zio_link_t *zl = NULL;
2165 mutex_enter(&pio->io_lock);
2166 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
2167 cio_next = zio_walk_children(pio, &zl);
2168 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
2169 pio->io_children[cio->io_child_type][w]++;
2170 mutex_exit(&pio->io_lock);
2172 mutex_enter(&pio->io_lock);
2174 mutex_exit(&pio->io_lock);
2177 * Now that all children have been reexecuted, execute the parent.
2178 * We don't reexecute "The Godfather" I/O here as it's the
2179 * responsibility of the caller to wait on it.
2181 if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) {
2182 pio->io_queued_timestamp = gethrtime();
2188 zio_suspend(spa_t *spa, zio_t *zio, zio_suspend_reason_t reason)
2190 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
2191 fm_panic("Pool '%s' has encountered an uncorrectable I/O "
2192 "failure and the failure mode property for this pool "
2193 "is set to panic.", spa_name(spa));
2195 cmn_err(CE_WARN, "Pool '%s' has encountered an uncorrectable I/O "
2196 "failure and has been suspended.\n", spa_name(spa));
2198 zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL,
2201 mutex_enter(&spa->spa_suspend_lock);
2203 if (spa->spa_suspend_zio_root == NULL)
2204 spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
2205 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
2206 ZIO_FLAG_GODFATHER);
2208 spa->spa_suspended = reason;
2211 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
2212 ASSERT(zio != spa->spa_suspend_zio_root);
2213 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2214 ASSERT(zio_unique_parent(zio) == NULL);
2215 ASSERT(zio->io_stage == ZIO_STAGE_DONE);
2216 zio_add_child(spa->spa_suspend_zio_root, zio);
2219 mutex_exit(&spa->spa_suspend_lock);
2223 zio_resume(spa_t *spa)
2228 * Reexecute all previously suspended i/o.
2230 mutex_enter(&spa->spa_suspend_lock);
2231 spa->spa_suspended = ZIO_SUSPEND_NONE;
2232 cv_broadcast(&spa->spa_suspend_cv);
2233 pio = spa->spa_suspend_zio_root;
2234 spa->spa_suspend_zio_root = NULL;
2235 mutex_exit(&spa->spa_suspend_lock);
2241 return (zio_wait(pio));
2245 zio_resume_wait(spa_t *spa)
2247 mutex_enter(&spa->spa_suspend_lock);
2248 while (spa_suspended(spa))
2249 cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
2250 mutex_exit(&spa->spa_suspend_lock);
2254 * ==========================================================================
2257 * A gang block is a collection of small blocks that looks to the DMU
2258 * like one large block. When zio_dva_allocate() cannot find a block
2259 * of the requested size, due to either severe fragmentation or the pool
2260 * being nearly full, it calls zio_write_gang_block() to construct the
2261 * block from smaller fragments.
2263 * A gang block consists of a gang header (zio_gbh_phys_t) and up to
2264 * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
2265 * an indirect block: it's an array of block pointers. It consumes
2266 * only one sector and hence is allocatable regardless of fragmentation.
2267 * The gang header's bps point to its gang members, which hold the data.
2269 * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
2270 * as the verifier to ensure uniqueness of the SHA256 checksum.
2271 * Critically, the gang block bp's blk_cksum is the checksum of the data,
2272 * not the gang header. This ensures that data block signatures (needed for
2273 * deduplication) are independent of how the block is physically stored.
2275 * Gang blocks can be nested: a gang member may itself be a gang block.
2276 * Thus every gang block is a tree in which root and all interior nodes are
2277 * gang headers, and the leaves are normal blocks that contain user data.
2278 * The root of the gang tree is called the gang leader.
2280 * To perform any operation (read, rewrite, free, claim) on a gang block,
2281 * zio_gang_assemble() first assembles the gang tree (minus data leaves)
2282 * in the io_gang_tree field of the original logical i/o by recursively
2283 * reading the gang leader and all gang headers below it. This yields
2284 * an in-core tree containing the contents of every gang header and the
2285 * bps for every constituent of the gang block.
2287 * With the gang tree now assembled, zio_gang_issue() just walks the gang tree
2288 * and invokes a callback on each bp. To free a gang block, zio_gang_issue()
2289 * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
2290 * zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
2291 * zio_read_gang() is a wrapper around zio_read() that omits reading gang
2292 * headers, since we already have those in io_gang_tree. zio_rewrite_gang()
2293 * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
2294 * of the gang header plus zio_checksum_compute() of the data to update the
2295 * gang header's blk_cksum as described above.
2297 * The two-phase assemble/issue model solves the problem of partial failure --
2298 * what if you'd freed part of a gang block but then couldn't read the
2299 * gang header for another part? Assembling the entire gang tree first
2300 * ensures that all the necessary gang header I/O has succeeded before
2301 * starting the actual work of free, claim, or write. Once the gang tree
2302 * is assembled, free and claim are in-memory operations that cannot fail.
2304 * In the event that a gang write fails, zio_dva_unallocate() walks the
2305 * gang tree to immediately free (i.e. insert back into the space map)
2306 * everything we've allocated. This ensures that we don't get ENOSPC
2307 * errors during repeated suspend/resume cycles due to a flaky device.
2309 * Gang rewrites only happen during sync-to-convergence. If we can't assemble
2310 * the gang tree, we won't modify the block, so we can safely defer the free
2311 * (knowing that the block is still intact). If we *can* assemble the gang
2312 * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
2313 * each constituent bp and we can allocate a new block on the next sync pass.
2315 * In all cases, the gang tree allows complete recovery from partial failure.
2316 * ==========================================================================
2320 zio_gang_issue_func_done(zio_t *zio)
2322 abd_put(zio->io_abd);
2326 zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2332 return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset),
2333 BP_GET_PSIZE(bp), zio_gang_issue_func_done,
2334 NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2335 &pio->io_bookmark));
2339 zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2346 abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2347 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2348 gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL,
2349 pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
2352 * As we rewrite each gang header, the pipeline will compute
2353 * a new gang block header checksum for it; but no one will
2354 * compute a new data checksum, so we do that here. The one
2355 * exception is the gang leader: the pipeline already computed
2356 * its data checksum because that stage precedes gang assembly.
2357 * (Presently, nothing actually uses interior data checksums;
2358 * this is just good hygiene.)
2360 if (gn != pio->io_gang_leader->io_gang_tree) {
2361 abd_t *buf = abd_get_offset(data, offset);
2363 zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
2364 buf, BP_GET_PSIZE(bp));
2369 * If we are here to damage data for testing purposes,
2370 * leave the GBH alone so that we can detect the damage.
2372 if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
2373 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2375 zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
2376 abd_get_offset(data, offset), BP_GET_PSIZE(bp),
2377 zio_gang_issue_func_done, NULL, pio->io_priority,
2378 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2386 zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2389 return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
2390 ZIO_GANG_CHILD_FLAGS(pio)));
2395 zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data,
2398 return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
2399 NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
2402 static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
2411 static void zio_gang_tree_assemble_done(zio_t *zio);
2413 static zio_gang_node_t *
2414 zio_gang_node_alloc(zio_gang_node_t **gnpp)
2416 zio_gang_node_t *gn;
2418 ASSERT(*gnpp == NULL);
2420 gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
2421 gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
2428 zio_gang_node_free(zio_gang_node_t **gnpp)
2430 zio_gang_node_t *gn = *gnpp;
2432 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2433 ASSERT(gn->gn_child[g] == NULL);
2435 zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2436 kmem_free(gn, sizeof (*gn));
2441 zio_gang_tree_free(zio_gang_node_t **gnpp)
2443 zio_gang_node_t *gn = *gnpp;
2448 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++)
2449 zio_gang_tree_free(&gn->gn_child[g]);
2451 zio_gang_node_free(gnpp);
2455 zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
2457 zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
2458 abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE);
2460 ASSERT(gio->io_gang_leader == gio);
2461 ASSERT(BP_IS_GANG(bp));
2463 zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2464 zio_gang_tree_assemble_done, gn, gio->io_priority,
2465 ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
2469 zio_gang_tree_assemble_done(zio_t *zio)
2471 zio_t *gio = zio->io_gang_leader;
2472 zio_gang_node_t *gn = zio->io_private;
2473 blkptr_t *bp = zio->io_bp;
2475 ASSERT(gio == zio_unique_parent(zio));
2476 ASSERT(zio->io_child_count == 0);
2481 /* this ABD was created from a linear buf in zio_gang_tree_assemble */
2482 if (BP_SHOULD_BYTESWAP(bp))
2483 byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size);
2485 ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh);
2486 ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
2487 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2489 abd_put(zio->io_abd);
2491 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2492 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2493 if (!BP_IS_GANG(gbp))
2495 zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
2500 zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data,
2503 zio_t *gio = pio->io_gang_leader;
2506 ASSERT(BP_IS_GANG(bp) == !!gn);
2507 ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
2508 ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
2511 * If you're a gang header, your data is in gn->gn_gbh.
2512 * If you're a gang member, your data is in 'data' and gn == NULL.
2514 zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset);
2517 ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2519 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
2520 blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
2521 if (BP_IS_HOLE(gbp))
2523 zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data,
2525 offset += BP_GET_PSIZE(gbp);
2529 if (gn == gio->io_gang_tree)
2530 ASSERT3U(gio->io_size, ==, offset);
2537 zio_gang_assemble(zio_t *zio)
2539 blkptr_t *bp = zio->io_bp;
2541 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
2542 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2544 zio->io_gang_leader = zio;
2546 zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
2552 zio_gang_issue(zio_t *zio)
2554 blkptr_t *bp = zio->io_bp;
2556 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT, ZIO_WAIT_DONE)) {
2560 ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
2561 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2563 if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
2564 zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd,
2567 zio_gang_tree_free(&zio->io_gang_tree);
2569 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2575 zio_write_gang_member_ready(zio_t *zio)
2577 zio_t *pio = zio_unique_parent(zio);
2578 dva_t *cdva = zio->io_bp->blk_dva;
2579 dva_t *pdva = pio->io_bp->blk_dva;
2581 ASSERTV(zio_t *gio = zio->io_gang_leader);
2583 if (BP_IS_HOLE(zio->io_bp))
2586 ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
2588 ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
2589 ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
2590 ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
2591 ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
2592 ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
2594 mutex_enter(&pio->io_lock);
2595 for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
2596 ASSERT(DVA_GET_GANG(&pdva[d]));
2597 asize = DVA_GET_ASIZE(&pdva[d]);
2598 asize += DVA_GET_ASIZE(&cdva[d]);
2599 DVA_SET_ASIZE(&pdva[d], asize);
2601 mutex_exit(&pio->io_lock);
2605 zio_write_gang_done(zio_t *zio)
2608 * The io_abd field will be NULL for a zio with no data. The io_flags
2609 * will initially have the ZIO_FLAG_NODATA bit flag set, but we can't
2610 * check for it here as it is cleared in zio_ready.
2612 if (zio->io_abd != NULL)
2613 abd_put(zio->io_abd);
2617 zio_write_gang_block(zio_t *pio)
2619 spa_t *spa = pio->io_spa;
2620 metaslab_class_t *mc = spa_normal_class(spa);
2621 blkptr_t *bp = pio->io_bp;
2622 zio_t *gio = pio->io_gang_leader;
2624 zio_gang_node_t *gn, **gnpp;
2625 zio_gbh_phys_t *gbh;
2627 uint64_t txg = pio->io_txg;
2628 uint64_t resid = pio->io_size;
2630 int copies = gio->io_prop.zp_copies;
2634 boolean_t has_data = !(pio->io_flags & ZIO_FLAG_NODATA);
2637 * encrypted blocks need DVA[2] free so encrypted gang headers can't
2638 * have a third copy.
2640 gbh_copies = MIN(copies + 1, spa_max_replication(spa));
2641 if (gio->io_prop.zp_encrypt && gbh_copies >= SPA_DVAS_PER_BP)
2642 gbh_copies = SPA_DVAS_PER_BP - 1;
2644 int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER;
2645 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2646 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2649 flags |= METASLAB_ASYNC_ALLOC;
2650 VERIFY(zfs_refcount_held(&mc->mc_alloc_slots[pio->io_allocator],
2654 * The logical zio has already placed a reservation for
2655 * 'copies' allocation slots but gang blocks may require
2656 * additional copies. These additional copies
2657 * (i.e. gbh_copies - copies) are guaranteed to succeed
2658 * since metaslab_class_throttle_reserve() always allows
2659 * additional reservations for gang blocks.
2661 VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies,
2662 pio->io_allocator, pio, flags));
2665 error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE,
2666 bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags,
2667 &pio->io_alloc_list, pio, pio->io_allocator);
2669 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2670 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2674 * If we failed to allocate the gang block header then
2675 * we remove any additional allocation reservations that
2676 * we placed here. The original reservation will
2677 * be removed when the logical I/O goes to the ready
2680 metaslab_class_throttle_unreserve(mc,
2681 gbh_copies - copies, pio->io_allocator, pio);
2684 pio->io_error = error;
2689 gnpp = &gio->io_gang_tree;
2691 gnpp = pio->io_private;
2692 ASSERT(pio->io_ready == zio_write_gang_member_ready);
2695 gn = zio_gang_node_alloc(gnpp);
2697 bzero(gbh, SPA_GANGBLOCKSIZE);
2698 gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE);
2701 * Create the gang header.
2703 zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE,
2704 zio_write_gang_done, NULL, pio->io_priority,
2705 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2708 * Create and nowait the gang children.
2710 for (int g = 0; resid != 0; resid -= lsize, g++) {
2711 lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
2713 ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2715 zp.zp_checksum = gio->io_prop.zp_checksum;
2716 zp.zp_compress = ZIO_COMPRESS_OFF;
2717 zp.zp_type = DMU_OT_NONE;
2719 zp.zp_copies = gio->io_prop.zp_copies;
2720 zp.zp_dedup = B_FALSE;
2721 zp.zp_dedup_verify = B_FALSE;
2722 zp.zp_nopwrite = B_FALSE;
2723 zp.zp_encrypt = gio->io_prop.zp_encrypt;
2724 zp.zp_byteorder = gio->io_prop.zp_byteorder;
2725 bzero(zp.zp_salt, ZIO_DATA_SALT_LEN);
2726 bzero(zp.zp_iv, ZIO_DATA_IV_LEN);
2727 bzero(zp.zp_mac, ZIO_DATA_MAC_LEN);
2729 zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
2730 has_data ? abd_get_offset(pio->io_abd, pio->io_size -
2731 resid) : NULL, lsize, lsize, &zp,
2732 zio_write_gang_member_ready, NULL, NULL,
2733 zio_write_gang_done, &gn->gn_child[g], pio->io_priority,
2734 ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2736 if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
2737 ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
2741 * Gang children won't throttle but we should
2742 * account for their work, so reserve an allocation
2743 * slot for them here.
2745 VERIFY(metaslab_class_throttle_reserve(mc,
2746 zp.zp_copies, cio->io_allocator, cio, flags));
2752 * Set pio's pipeline to just wait for zio to finish.
2754 pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2757 * We didn't allocate this bp, so make sure it doesn't get unmarked.
2759 pio->io_flags &= ~ZIO_FLAG_FASTWRITE;
2767 * The zio_nop_write stage in the pipeline determines if allocating a
2768 * new bp is necessary. The nopwrite feature can handle writes in
2769 * either syncing or open context (i.e. zil writes) and as a result is
2770 * mutually exclusive with dedup.
2772 * By leveraging a cryptographically secure checksum, such as SHA256, we
2773 * can compare the checksums of the new data and the old to determine if
2774 * allocating a new block is required. Note that our requirements for
2775 * cryptographic strength are fairly weak: there can't be any accidental
2776 * hash collisions, but we don't need to be secure against intentional
2777 * (malicious) collisions. To trigger a nopwrite, you have to be able
2778 * to write the file to begin with, and triggering an incorrect (hash
2779 * collision) nopwrite is no worse than simply writing to the file.
2780 * That said, there are no known attacks against the checksum algorithms
2781 * used for nopwrite, assuming that the salt and the checksums
2782 * themselves remain secret.
2785 zio_nop_write(zio_t *zio)
2787 blkptr_t *bp = zio->io_bp;
2788 blkptr_t *bp_orig = &zio->io_bp_orig;
2789 zio_prop_t *zp = &zio->io_prop;
2791 ASSERT(BP_GET_LEVEL(bp) == 0);
2792 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
2793 ASSERT(zp->zp_nopwrite);
2794 ASSERT(!zp->zp_dedup);
2795 ASSERT(zio->io_bp_override == NULL);
2796 ASSERT(IO_IS_ALLOCATING(zio));
2799 * Check to see if the original bp and the new bp have matching
2800 * characteristics (i.e. same checksum, compression algorithms, etc).
2801 * If they don't then just continue with the pipeline which will
2802 * allocate a new bp.
2804 if (BP_IS_HOLE(bp_orig) ||
2805 !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags &
2806 ZCHECKSUM_FLAG_NOPWRITE) ||
2807 BP_IS_ENCRYPTED(bp) || BP_IS_ENCRYPTED(bp_orig) ||
2808 BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
2809 BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
2810 BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
2811 zp->zp_copies != BP_GET_NDVAS(bp_orig))
2815 * If the checksums match then reset the pipeline so that we
2816 * avoid allocating a new bp and issuing any I/O.
2818 if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
2819 ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags &
2820 ZCHECKSUM_FLAG_NOPWRITE);
2821 ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
2822 ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
2823 ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
2824 ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
2825 sizeof (uint64_t)) == 0);
2828 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2829 zio->io_flags |= ZIO_FLAG_NOPWRITE;
2836 * ==========================================================================
2838 * ==========================================================================
2841 zio_ddt_child_read_done(zio_t *zio)
2843 blkptr_t *bp = zio->io_bp;
2844 ddt_entry_t *dde = zio->io_private;
2846 zio_t *pio = zio_unique_parent(zio);
2848 mutex_enter(&pio->io_lock);
2849 ddp = ddt_phys_select(dde, bp);
2850 if (zio->io_error == 0)
2851 ddt_phys_clear(ddp); /* this ddp doesn't need repair */
2853 if (zio->io_error == 0 && dde->dde_repair_abd == NULL)
2854 dde->dde_repair_abd = zio->io_abd;
2856 abd_free(zio->io_abd);
2857 mutex_exit(&pio->io_lock);
2861 zio_ddt_read_start(zio_t *zio)
2863 blkptr_t *bp = zio->io_bp;
2865 ASSERT(BP_GET_DEDUP(bp));
2866 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2867 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2869 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2870 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2871 ddt_entry_t *dde = ddt_repair_start(ddt, bp);
2872 ddt_phys_t *ddp = dde->dde_phys;
2873 ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
2876 ASSERT(zio->io_vsd == NULL);
2879 if (ddp_self == NULL)
2882 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2883 if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
2885 ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
2887 zio_nowait(zio_read(zio, zio->io_spa, &blk,
2888 abd_alloc_for_io(zio->io_size, B_TRUE),
2889 zio->io_size, zio_ddt_child_read_done, dde,
2890 zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) |
2891 ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark));
2896 zio_nowait(zio_read(zio, zio->io_spa, bp,
2897 zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority,
2898 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
2904 zio_ddt_read_done(zio_t *zio)
2906 blkptr_t *bp = zio->io_bp;
2908 if (zio_wait_for_children(zio, ZIO_CHILD_DDT_BIT, ZIO_WAIT_DONE)) {
2912 ASSERT(BP_GET_DEDUP(bp));
2913 ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
2914 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2916 if (zio->io_child_error[ZIO_CHILD_DDT]) {
2917 ddt_t *ddt = ddt_select(zio->io_spa, bp);
2918 ddt_entry_t *dde = zio->io_vsd;
2920 ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
2924 zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
2925 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
2928 if (dde->dde_repair_abd != NULL) {
2929 abd_copy(zio->io_abd, dde->dde_repair_abd,
2931 zio->io_child_error[ZIO_CHILD_DDT] = 0;
2933 ddt_repair_done(ddt, dde);
2937 ASSERT(zio->io_vsd == NULL);
2943 zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
2945 spa_t *spa = zio->io_spa;
2946 boolean_t do_raw = !!(zio->io_flags & ZIO_FLAG_RAW);
2948 ASSERT(!(zio->io_bp_override && do_raw));
2951 * Note: we compare the original data, not the transformed data,
2952 * because when zio->io_bp is an override bp, we will not have
2953 * pushed the I/O transforms. That's an important optimization
2954 * because otherwise we'd compress/encrypt all dmu_sync() data twice.
2955 * However, we should never get a raw, override zio so in these
2956 * cases we can compare the io_abd directly. This is useful because
2957 * it allows us to do dedup verification even if we don't have access
2958 * to the original data (for instance, if the encryption keys aren't
2962 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2963 zio_t *lio = dde->dde_lead_zio[p];
2965 if (lio != NULL && do_raw) {
2966 return (lio->io_size != zio->io_size ||
2967 abd_cmp(zio->io_abd, lio->io_abd) != 0);
2968 } else if (lio != NULL) {
2969 return (lio->io_orig_size != zio->io_orig_size ||
2970 abd_cmp(zio->io_orig_abd, lio->io_orig_abd) != 0);
2974 for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
2975 ddt_phys_t *ddp = &dde->dde_phys[p];
2977 if (ddp->ddp_phys_birth != 0 && do_raw) {
2978 blkptr_t blk = *zio->io_bp;
2983 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
2984 psize = BP_GET_PSIZE(&blk);
2986 if (psize != zio->io_size)
2991 tmpabd = abd_alloc_for_io(psize, B_TRUE);
2993 error = zio_wait(zio_read(NULL, spa, &blk, tmpabd,
2994 psize, NULL, NULL, ZIO_PRIORITY_SYNC_READ,
2995 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
2996 ZIO_FLAG_RAW, &zio->io_bookmark));
2999 if (abd_cmp(tmpabd, zio->io_abd) != 0)
3000 error = SET_ERROR(ENOENT);
3005 return (error != 0);
3006 } else if (ddp->ddp_phys_birth != 0) {
3007 arc_buf_t *abuf = NULL;
3008 arc_flags_t aflags = ARC_FLAG_WAIT;
3009 blkptr_t blk = *zio->io_bp;
3012 ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
3014 if (BP_GET_LSIZE(&blk) != zio->io_orig_size)
3019 error = arc_read(NULL, spa, &blk,
3020 arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
3021 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
3022 &aflags, &zio->io_bookmark);
3025 if (abd_cmp_buf(zio->io_orig_abd, abuf->b_data,
3026 zio->io_orig_size) != 0)
3027 error = SET_ERROR(ENOENT);
3028 arc_buf_destroy(abuf, &abuf);
3032 return (error != 0);
3040 zio_ddt_child_write_ready(zio_t *zio)
3042 int p = zio->io_prop.zp_copies;
3043 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
3044 ddt_entry_t *dde = zio->io_private;
3045 ddt_phys_t *ddp = &dde->dde_phys[p];
3053 ASSERT(dde->dde_lead_zio[p] == zio);
3055 ddt_phys_fill(ddp, zio->io_bp);
3057 zio_link_t *zl = NULL;
3058 while ((pio = zio_walk_parents(zio, &zl)) != NULL)
3059 ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
3065 zio_ddt_child_write_done(zio_t *zio)
3067 int p = zio->io_prop.zp_copies;
3068 ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
3069 ddt_entry_t *dde = zio->io_private;
3070 ddt_phys_t *ddp = &dde->dde_phys[p];
3074 ASSERT(ddp->ddp_refcnt == 0);
3075 ASSERT(dde->dde_lead_zio[p] == zio);
3076 dde->dde_lead_zio[p] = NULL;
3078 if (zio->io_error == 0) {
3079 zio_link_t *zl = NULL;
3080 while (zio_walk_parents(zio, &zl) != NULL)
3081 ddt_phys_addref(ddp);
3083 ddt_phys_clear(ddp);
3090 zio_ddt_ditto_write_done(zio_t *zio)
3092 int p = DDT_PHYS_DITTO;
3093 ASSERTV(zio_prop_t *zp = &zio->io_prop);
3094 blkptr_t *bp = zio->io_bp;
3095 ddt_t *ddt = ddt_select(zio->io_spa, bp);
3096 ddt_entry_t *dde = zio->io_private;
3097 ddt_phys_t *ddp = &dde->dde_phys[p];
3098 ddt_key_t *ddk = &dde->dde_key;
3102 ASSERT(ddp->ddp_refcnt == 0);
3103 ASSERT(dde->dde_lead_zio[p] == zio);
3104 dde->dde_lead_zio[p] = NULL;
3106 if (zio->io_error == 0) {
3107 ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
3108 ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
3109 ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
3110 if (ddp->ddp_phys_birth != 0)
3111 ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
3112 ddt_phys_fill(ddp, bp);
3119 zio_ddt_write(zio_t *zio)
3121 spa_t *spa = zio->io_spa;
3122 blkptr_t *bp = zio->io_bp;
3123 uint64_t txg = zio->io_txg;
3124 zio_prop_t *zp = &zio->io_prop;
3125 int p = zp->zp_copies;
3129 ddt_t *ddt = ddt_select(spa, bp);
3133 ASSERT(BP_GET_DEDUP(bp));
3134 ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
3135 ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
3136 ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW)));
3139 dde = ddt_lookup(ddt, bp, B_TRUE);
3140 ddp = &dde->dde_phys[p];
3142 if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
3144 * If we're using a weak checksum, upgrade to a strong checksum
3145 * and try again. If we're already using a strong checksum,
3146 * we can't resolve it, so just convert to an ordinary write.
3147 * (And automatically e-mail a paper to Nature?)
3149 if (!(zio_checksum_table[zp->zp_checksum].ci_flags &
3150 ZCHECKSUM_FLAG_DEDUP)) {
3151 zp->zp_checksum = spa_dedup_checksum(spa);
3152 zio_pop_transforms(zio);
3153 zio->io_stage = ZIO_STAGE_OPEN;
3156 zp->zp_dedup = B_FALSE;
3158 zio->io_pipeline = ZIO_WRITE_PIPELINE;
3163 ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
3164 ASSERT(ditto_copies < SPA_DVAS_PER_BP);
3166 if (ditto_copies > ddt_ditto_copies_present(dde) &&
3167 dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
3168 zio_prop_t czp = *zp;
3170 czp.zp_copies = ditto_copies;
3173 * If we arrived here with an override bp, we won't have run
3174 * the transform stack, so we won't have the data we need to
3175 * generate a child i/o. So, toss the override bp and restart.
3176 * This is safe, because using the override bp is just an
3177 * optimization; and it's rare, so the cost doesn't matter.
3179 if (zio->io_bp_override) {
3180 zio_pop_transforms(zio);
3181 zio->io_stage = ZIO_STAGE_OPEN;
3182 zio->io_pipeline = ZIO_WRITE_PIPELINE;
3183 zio->io_bp_override = NULL;
3189 dio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
3190 zio->io_orig_size, zio->io_orig_size, &czp, NULL, NULL,
3191 NULL, zio_ddt_ditto_write_done, dde, zio->io_priority,
3192 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
3194 zio_push_transform(dio, zio->io_abd, zio->io_size, 0, NULL);
3195 dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
3198 if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
3199 if (ddp->ddp_phys_birth != 0)
3200 ddt_bp_fill(ddp, bp, txg);
3201 if (dde->dde_lead_zio[p] != NULL)
3202 zio_add_child(zio, dde->dde_lead_zio[p]);
3204 ddt_phys_addref(ddp);
3205 } else if (zio->io_bp_override) {
3206 ASSERT(bp->blk_birth == txg);
3207 ASSERT(BP_EQUAL(bp, zio->io_bp_override));
3208 ddt_phys_fill(ddp, bp);
3209 ddt_phys_addref(ddp);
3211 cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd,
3212 zio->io_orig_size, zio->io_orig_size, zp,
3213 zio_ddt_child_write_ready, NULL, NULL,
3214 zio_ddt_child_write_done, dde, zio->io_priority,
3215 ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
3217 zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL);
3218 dde->dde_lead_zio[p] = cio;
3231 ddt_entry_t *freedde; /* for debugging */
3234 zio_ddt_free(zio_t *zio)
3236 spa_t *spa = zio->io_spa;
3237 blkptr_t *bp = zio->io_bp;
3238 ddt_t *ddt = ddt_select(spa, bp);
3242 ASSERT(BP_GET_DEDUP(bp));
3243 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
3246 freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
3248 ddp = ddt_phys_select(dde, bp);
3250 ddt_phys_decref(ddp);
3258 * ==========================================================================
3259 * Allocate and free blocks
3260 * ==========================================================================
3264 zio_io_to_allocate(spa_t *spa, int allocator)
3268 ASSERT(MUTEX_HELD(&spa->spa_alloc_locks[allocator]));
3270 zio = avl_first(&spa->spa_alloc_trees[allocator]);
3274 ASSERT(IO_IS_ALLOCATING(zio));
3277 * Try to place a reservation for this zio. If we're unable to
3278 * reserve then we throttle.
3280 ASSERT3U(zio->io_allocator, ==, allocator);
3281 if (!metaslab_class_throttle_reserve(zio->io_metaslab_class,
3282 zio->io_prop.zp_copies, zio->io_allocator, zio, 0)) {
3286 avl_remove(&spa->spa_alloc_trees[allocator], zio);
3287 ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE);
3293 zio_dva_throttle(zio_t *zio)
3295 spa_t *spa = zio->io_spa;
3297 metaslab_class_t *mc;
3299 /* locate an appropriate allocation class */
3300 mc = spa_preferred_class(spa, zio->io_size, zio->io_prop.zp_type,
3301 zio->io_prop.zp_level, zio->io_prop.zp_zpl_smallblk);
3303 if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE ||
3304 !mc->mc_alloc_throttle_enabled ||
3305 zio->io_child_type == ZIO_CHILD_GANG ||
3306 zio->io_flags & ZIO_FLAG_NODATA) {
3310 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3312 ASSERT3U(zio->io_queued_timestamp, >, 0);
3313 ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE);
3315 zbookmark_phys_t *bm = &zio->io_bookmark;
3317 * We want to try to use as many allocators as possible to help improve
3318 * performance, but we also want logically adjacent IOs to be physically
3319 * adjacent to improve sequential read performance. We chunk each object
3320 * into 2^20 block regions, and then hash based on the objset, object,
3321 * level, and region to accomplish both of these goals.
3323 zio->io_allocator = cityhash4(bm->zb_objset, bm->zb_object,
3324 bm->zb_level, bm->zb_blkid >> 20) % spa->spa_alloc_count;
3325 mutex_enter(&spa->spa_alloc_locks[zio->io_allocator]);
3326 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3327 zio->io_metaslab_class = mc;
3328 avl_add(&spa->spa_alloc_trees[zio->io_allocator], zio);
3329 nio = zio_io_to_allocate(spa, zio->io_allocator);
3330 mutex_exit(&spa->spa_alloc_locks[zio->io_allocator]);
3335 zio_allocate_dispatch(spa_t *spa, int allocator)
3339 mutex_enter(&spa->spa_alloc_locks[allocator]);
3340 zio = zio_io_to_allocate(spa, allocator);
3341 mutex_exit(&spa->spa_alloc_locks[allocator]);
3345 ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE);
3346 ASSERT0(zio->io_error);
3347 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE);
3351 zio_dva_allocate(zio_t *zio)
3353 spa_t *spa = zio->io_spa;
3354 metaslab_class_t *mc;
3355 blkptr_t *bp = zio->io_bp;
3359 if (zio->io_gang_leader == NULL) {
3360 ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
3361 zio->io_gang_leader = zio;
3364 ASSERT(BP_IS_HOLE(bp));
3365 ASSERT0(BP_GET_NDVAS(bp));
3366 ASSERT3U(zio->io_prop.zp_copies, >, 0);
3367 ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
3368 ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
3370 flags |= (zio->io_flags & ZIO_FLAG_FASTWRITE) ? METASLAB_FASTWRITE : 0;
3371 if (zio->io_flags & ZIO_FLAG_NODATA)
3372 flags |= METASLAB_DONT_THROTTLE;
3373 if (zio->io_flags & ZIO_FLAG_GANG_CHILD)
3374 flags |= METASLAB_GANG_CHILD;
3375 if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE)
3376 flags |= METASLAB_ASYNC_ALLOC;
3379 * if not already chosen, locate an appropriate allocation class
3381 mc = zio->io_metaslab_class;
3383 mc = spa_preferred_class(spa, zio->io_size,
3384 zio->io_prop.zp_type, zio->io_prop.zp_level,
3385 zio->io_prop.zp_zpl_smallblk);
3386 zio->io_metaslab_class = mc;
3389 error = metaslab_alloc(spa, mc, zio->io_size, bp,
3390 zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
3391 &zio->io_alloc_list, zio, zio->io_allocator);
3394 * Fallback to normal class when an alloc class is full
3396 if (error == ENOSPC && mc != spa_normal_class(spa)) {
3398 * If throttling, transfer reservation over to normal class.
3399 * The io_allocator slot can remain the same even though we
3400 * are switching classes.
3402 if (mc->mc_alloc_throttle_enabled &&
3403 (zio->io_flags & ZIO_FLAG_IO_ALLOCATING)) {
3404 metaslab_class_throttle_unreserve(mc,
3405 zio->io_prop.zp_copies, zio->io_allocator, zio);
3406 zio->io_flags &= ~ZIO_FLAG_IO_ALLOCATING;
3408 mc = spa_normal_class(spa);
3409 VERIFY(metaslab_class_throttle_reserve(mc,
3410 zio->io_prop.zp_copies, zio->io_allocator, zio,
3411 flags | METASLAB_MUST_RESERVE));
3413 mc = spa_normal_class(spa);
3415 zio->io_metaslab_class = mc;
3417 error = metaslab_alloc(spa, mc, zio->io_size, bp,
3418 zio->io_prop.zp_copies, zio->io_txg, NULL, flags,
3419 &zio->io_alloc_list, zio, zio->io_allocator);
3423 zfs_dbgmsg("%s: metaslab allocation failure: zio %p, "
3424 "size %llu, error %d", spa_name(spa), zio, zio->io_size,
3426 if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
3427 return (zio_write_gang_block(zio));
3428 zio->io_error = error;
3435 zio_dva_free(zio_t *zio)
3437 metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
3443 zio_dva_claim(zio_t *zio)
3447 error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
3449 zio->io_error = error;
3455 * Undo an allocation. This is used by zio_done() when an I/O fails
3456 * and we want to give back the block we just allocated.
3457 * This handles both normal blocks and gang blocks.
3460 zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
3462 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
3463 ASSERT(zio->io_bp_override == NULL);
3465 if (!BP_IS_HOLE(bp))
3466 metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
3469 for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) {
3470 zio_dva_unallocate(zio, gn->gn_child[g],
3471 &gn->gn_gbh->zg_blkptr[g]);
3477 * Try to allocate an intent log block. Return 0 on success, errno on failure.
3480 zio_alloc_zil(spa_t *spa, objset_t *os, uint64_t txg, blkptr_t *new_bp,
3481 uint64_t size, boolean_t *slog)
3484 zio_alloc_list_t io_alloc_list;
3486 ASSERT(txg > spa_syncing_txg(spa));
3488 metaslab_trace_init(&io_alloc_list);
3491 * Block pointer fields are useful to metaslabs for stats and debugging.
3492 * Fill in the obvious ones before calling into metaslab_alloc().
3494 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3495 BP_SET_PSIZE(new_bp, size);
3496 BP_SET_LEVEL(new_bp, 0);
3499 * When allocating a zil block, we don't have information about
3500 * the final destination of the block except the objset it's part
3501 * of, so we just hash the objset ID to pick the allocator to get
3504 error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1,
3505 txg, NULL, METASLAB_FASTWRITE, &io_alloc_list, NULL,
3506 cityhash4(0, 0, 0, os->os_dsl_dataset->ds_object) %
3507 spa->spa_alloc_count);
3511 error = metaslab_alloc(spa, spa_normal_class(spa), size,
3512 new_bp, 1, txg, NULL, METASLAB_FASTWRITE,
3513 &io_alloc_list, NULL, cityhash4(0, 0, 0,
3514 os->os_dsl_dataset->ds_object) % spa->spa_alloc_count);
3518 metaslab_trace_fini(&io_alloc_list);
3521 BP_SET_LSIZE(new_bp, size);
3522 BP_SET_PSIZE(new_bp, size);
3523 BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
3524 BP_SET_CHECKSUM(new_bp,
3525 spa_version(spa) >= SPA_VERSION_SLIM_ZIL
3526 ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
3527 BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
3528 BP_SET_LEVEL(new_bp, 0);
3529 BP_SET_DEDUP(new_bp, 0);
3530 BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
3533 * encrypted blocks will require an IV and salt. We generate
3534 * these now since we will not be rewriting the bp at
3537 if (os->os_encrypted) {
3538 uint8_t iv[ZIO_DATA_IV_LEN];
3539 uint8_t salt[ZIO_DATA_SALT_LEN];
3541 BP_SET_CRYPT(new_bp, B_TRUE);
3542 VERIFY0(spa_crypt_get_salt(spa,
3543 dmu_objset_id(os), salt));
3544 VERIFY0(zio_crypt_generate_iv(iv));
3546 zio_crypt_encode_params_bp(new_bp, salt, iv);
3549 zfs_dbgmsg("%s: zil block allocation failure: "
3550 "size %llu, error %d", spa_name(spa), size, error);
3557 * ==========================================================================
3558 * Read and write to physical devices
3559 * ==========================================================================
3564 * Issue an I/O to the underlying vdev. Typically the issue pipeline
3565 * stops after this stage and will resume upon I/O completion.
3566 * However, there are instances where the vdev layer may need to
3567 * continue the pipeline when an I/O was not issued. Since the I/O
3568 * that was sent to the vdev layer might be different than the one
3569 * currently active in the pipeline (see vdev_queue_io()), we explicitly
3570 * force the underlying vdev layers to call either zio_execute() or
3571 * zio_interrupt() to ensure that the pipeline continues with the correct I/O.
3574 zio_vdev_io_start(zio_t *zio)
3576 vdev_t *vd = zio->io_vd;
3578 spa_t *spa = zio->io_spa;
3582 ASSERT(zio->io_error == 0);
3583 ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
3586 if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3587 spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
3590 * The mirror_ops handle multiple DVAs in a single BP.
3592 vdev_mirror_ops.vdev_op_io_start(zio);
3596 ASSERT3P(zio->io_logical, !=, zio);
3597 if (zio->io_type == ZIO_TYPE_WRITE) {
3598 ASSERT(spa->spa_trust_config);
3601 * Note: the code can handle other kinds of writes,
3602 * but we don't expect them.
3604 if (zio->io_vd->vdev_removing) {
3605 ASSERT(zio->io_flags &
3606 (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL |
3607 ZIO_FLAG_RESILVER | ZIO_FLAG_INDUCE_DAMAGE));
3611 align = 1ULL << vd->vdev_top->vdev_ashift;
3613 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
3614 P2PHASE(zio->io_size, align) != 0) {
3615 /* Transform logical writes to be a full physical block size. */
3616 uint64_t asize = P2ROUNDUP(zio->io_size, align);
3617 abd_t *abuf = abd_alloc_sametype(zio->io_abd, asize);
3618 ASSERT(vd == vd->vdev_top);
3619 if (zio->io_type == ZIO_TYPE_WRITE) {
3620 abd_copy(abuf, zio->io_abd, zio->io_size);
3621 abd_zero_off(abuf, zio->io_size, asize - zio->io_size);
3623 zio_push_transform(zio, abuf, asize, asize, zio_subblock);
3627 * If this is not a physical io, make sure that it is properly aligned
3628 * before proceeding.
3630 if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
3631 ASSERT0(P2PHASE(zio->io_offset, align));
3632 ASSERT0(P2PHASE(zio->io_size, align));
3635 * For physical writes, we allow 512b aligned writes and assume
3636 * the device will perform a read-modify-write as necessary.
3638 ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
3639 ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
3642 VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
3645 * If this is a repair I/O, and there's no self-healing involved --
3646 * that is, we're just resilvering what we expect to resilver --
3647 * then don't do the I/O unless zio's txg is actually in vd's DTL.
3648 * This prevents spurious resilvering.
3650 * There are a few ways that we can end up creating these spurious
3653 * 1. A resilver i/o will be issued if any DVA in the BP has a
3654 * dirty DTL. The mirror code will issue resilver writes to
3655 * each DVA, including the one(s) that are not on vdevs with dirty
3658 * 2. With nested replication, which happens when we have a
3659 * "replacing" or "spare" vdev that's a child of a mirror or raidz.
3660 * For example, given mirror(replacing(A+B), C), it's likely that
3661 * only A is out of date (it's the new device). In this case, we'll
3662 * read from C, then use the data to resilver A+B -- but we don't
3663 * actually want to resilver B, just A. The top-level mirror has no
3664 * way to know this, so instead we just discard unnecessary repairs
3665 * as we work our way down the vdev tree.
3667 * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc.
3668 * The same logic applies to any form of nested replication: ditto
3669 * + mirror, RAID-Z + replacing, etc.
3671 * However, indirect vdevs point off to other vdevs which may have
3672 * DTL's, so we never bypass them. The child i/os on concrete vdevs
3673 * will be properly bypassed instead.
3675 if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
3676 !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
3677 zio->io_txg != 0 && /* not a delegated i/o */
3678 vd->vdev_ops != &vdev_indirect_ops &&
3679 !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
3680 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
3681 zio_vdev_io_bypass(zio);
3685 if (vd->vdev_ops->vdev_op_leaf &&
3686 (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
3688 if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio))
3691 if ((zio = vdev_queue_io(zio)) == NULL)
3694 if (!vdev_accessible(vd, zio)) {
3695 zio->io_error = SET_ERROR(ENXIO);
3699 zio->io_delay = gethrtime();
3702 vd->vdev_ops->vdev_op_io_start(zio);
3707 zio_vdev_io_done(zio_t *zio)
3709 vdev_t *vd = zio->io_vd;
3710 vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
3711 boolean_t unexpected_error = B_FALSE;
3713 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3717 ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
3720 zio->io_delay = gethrtime() - zio->io_delay;
3722 if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
3724 vdev_queue_io_done(zio);
3726 if (zio->io_type == ZIO_TYPE_WRITE)
3727 vdev_cache_write(zio);
3729 if (zio_injection_enabled && zio->io_error == 0)
3730 zio->io_error = zio_handle_device_injections(vd, zio,
3733 if (zio_injection_enabled && zio->io_error == 0)
3734 zio->io_error = zio_handle_label_injection(zio, EIO);
3736 if (zio->io_error) {
3737 if (!vdev_accessible(vd, zio)) {
3738 zio->io_error = SET_ERROR(ENXIO);
3740 unexpected_error = B_TRUE;
3745 ops->vdev_op_io_done(zio);
3747 if (unexpected_error)
3748 VERIFY(vdev_probe(vd, zio) == NULL);
3754 * This function is used to change the priority of an existing zio that is
3755 * currently in-flight. This is used by the arc to upgrade priority in the
3756 * event that a demand read is made for a block that is currently queued
3757 * as a scrub or async read IO. Otherwise, the high priority read request
3758 * would end up having to wait for the lower priority IO.
3761 zio_change_priority(zio_t *pio, zio_priority_t priority)
3763 zio_t *cio, *cio_next;
3764 zio_link_t *zl = NULL;
3766 ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
3768 if (pio->io_vd != NULL && pio->io_vd->vdev_ops->vdev_op_leaf) {
3769 vdev_queue_change_io_priority(pio, priority);
3771 pio->io_priority = priority;
3774 mutex_enter(&pio->io_lock);
3775 for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) {
3776 cio_next = zio_walk_children(pio, &zl);
3777 zio_change_priority(cio, priority);
3779 mutex_exit(&pio->io_lock);
3783 * For non-raidz ZIOs, we can just copy aside the bad data read from the
3784 * disk, and use that to finish the checksum ereport later.
3787 zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
3788 const abd_t *good_buf)
3790 /* no processing needed */
3791 zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
3796 zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
3798 void *abd = abd_alloc_sametype(zio->io_abd, zio->io_size);
3800 abd_copy(abd, zio->io_abd, zio->io_size);
3802 zcr->zcr_cbinfo = zio->io_size;
3803 zcr->zcr_cbdata = abd;
3804 zcr->zcr_finish = zio_vsd_default_cksum_finish;
3805 zcr->zcr_free = zio_abd_free;
3809 zio_vdev_io_assess(zio_t *zio)
3811 vdev_t *vd = zio->io_vd;
3813 if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) {
3817 if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
3818 spa_config_exit(zio->io_spa, SCL_ZIO, zio);
3820 if (zio->io_vsd != NULL) {
3821 zio->io_vsd_ops->vsd_free(zio);
3825 if (zio_injection_enabled && zio->io_error == 0)
3826 zio->io_error = zio_handle_fault_injection(zio, EIO);
3829 * If the I/O failed, determine whether we should attempt to retry it.
3831 * On retry, we cut in line in the issue queue, since we don't want
3832 * compression/checksumming/etc. work to prevent our (cheap) IO reissue.
3834 if (zio->io_error && vd == NULL &&
3835 !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
3836 ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
3837 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
3839 zio->io_flags |= ZIO_FLAG_IO_RETRY |
3840 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
3841 zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
3842 zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
3843 zio_requeue_io_start_cut_in_line);
3848 * If we got an error on a leaf device, convert it to ENXIO
3849 * if the device is not accessible at all.
3851 if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3852 !vdev_accessible(vd, zio))
3853 zio->io_error = SET_ERROR(ENXIO);
3856 * If we can't write to an interior vdev (mirror or RAID-Z),
3857 * set vdev_cant_write so that we stop trying to allocate from it.
3859 if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
3860 vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
3861 vd->vdev_cant_write = B_TRUE;
3865 * If a cache flush returns ENOTSUP or ENOTTY, we know that no future
3866 * attempts will ever succeed. In this case we set a persistent bit so
3867 * that we don't bother with it in the future.
3869 if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) &&
3870 zio->io_type == ZIO_TYPE_IOCTL &&
3871 zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL)
3872 vd->vdev_nowritecache = B_TRUE;
3875 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
3877 if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
3878 zio->io_physdone != NULL) {
3879 ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
3880 ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
3881 zio->io_physdone(zio->io_logical);
3888 zio_vdev_io_reissue(zio_t *zio)
3890 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3891 ASSERT(zio->io_error == 0);
3893 zio->io_stage >>= 1;
3897 zio_vdev_io_redone(zio_t *zio)
3899 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
3901 zio->io_stage >>= 1;
3905 zio_vdev_io_bypass(zio_t *zio)
3907 ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
3908 ASSERT(zio->io_error == 0);
3910 zio->io_flags |= ZIO_FLAG_IO_BYPASS;
3911 zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
3915 * ==========================================================================
3916 * Encrypt and store encryption parameters
3917 * ==========================================================================
3922 * This function is used for ZIO_STAGE_ENCRYPT. It is responsible for
3923 * managing the storage of encryption parameters and passing them to the
3924 * lower-level encryption functions.
3927 zio_encrypt(zio_t *zio)
3929 zio_prop_t *zp = &zio->io_prop;
3930 spa_t *spa = zio->io_spa;
3931 blkptr_t *bp = zio->io_bp;
3932 uint64_t psize = BP_GET_PSIZE(bp);
3933 uint64_t dsobj = zio->io_bookmark.zb_objset;
3934 dmu_object_type_t ot = BP_GET_TYPE(bp);
3935 void *enc_buf = NULL;
3937 uint8_t salt[ZIO_DATA_SALT_LEN];
3938 uint8_t iv[ZIO_DATA_IV_LEN];
3939 uint8_t mac[ZIO_DATA_MAC_LEN];
3940 boolean_t no_crypt = B_FALSE;
3942 /* the root zio already encrypted the data */
3943 if (zio->io_child_type == ZIO_CHILD_GANG)
3946 /* only ZIL blocks are re-encrypted on rewrite */
3947 if (!IO_IS_ALLOCATING(zio) && ot != DMU_OT_INTENT_LOG)
3950 if (!(zp->zp_encrypt || BP_IS_ENCRYPTED(bp))) {
3951 BP_SET_CRYPT(bp, B_FALSE);
3955 /* if we are doing raw encryption set the provided encryption params */
3956 if (zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) {
3957 ASSERT0(BP_GET_LEVEL(bp));
3958 BP_SET_CRYPT(bp, B_TRUE);
3959 BP_SET_BYTEORDER(bp, zp->zp_byteorder);
3960 if (ot != DMU_OT_OBJSET)
3961 zio_crypt_encode_mac_bp(bp, zp->zp_mac);
3963 /* dnode blocks must be written out in the provided byteorder */
3964 if (zp->zp_byteorder != ZFS_HOST_BYTEORDER &&
3965 ot == DMU_OT_DNODE) {
3966 void *bswap_buf = zio_buf_alloc(psize);
3967 abd_t *babd = abd_get_from_buf(bswap_buf, psize);
3969 ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
3970 abd_copy_to_buf(bswap_buf, zio->io_abd, psize);
3971 dmu_ot_byteswap[DMU_OT_BYTESWAP(ot)].ob_func(bswap_buf,
3974 abd_take_ownership_of_buf(babd, B_TRUE);
3975 zio_push_transform(zio, babd, psize, psize, NULL);
3978 if (DMU_OT_IS_ENCRYPTED(ot))
3979 zio_crypt_encode_params_bp(bp, zp->zp_salt, zp->zp_iv);
3983 /* indirect blocks only maintain a cksum of the lower level MACs */
3984 if (BP_GET_LEVEL(bp) > 0) {
3985 BP_SET_CRYPT(bp, B_TRUE);
3986 VERIFY0(zio_crypt_do_indirect_mac_checksum_abd(B_TRUE,
3987 zio->io_orig_abd, BP_GET_LSIZE(bp), BP_SHOULD_BYTESWAP(bp),
3989 zio_crypt_encode_mac_bp(bp, mac);
3994 * Objset blocks are a special case since they have 2 256-bit MACs
3995 * embedded within them.
3997 if (ot == DMU_OT_OBJSET) {
3998 ASSERT0(DMU_OT_IS_ENCRYPTED(ot));
3999 ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
4000 BP_SET_CRYPT(bp, B_TRUE);
4001 VERIFY0(spa_do_crypt_objset_mac_abd(B_TRUE, spa, dsobj,
4002 zio->io_abd, psize, BP_SHOULD_BYTESWAP(bp)));
4006 /* unencrypted object types are only authenticated with a MAC */
4007 if (!DMU_OT_IS_ENCRYPTED(ot)) {
4008 BP_SET_CRYPT(bp, B_TRUE);
4009 VERIFY0(spa_do_crypt_mac_abd(B_TRUE, spa, dsobj,
4010 zio->io_abd, psize, mac));
4011 zio_crypt_encode_mac_bp(bp, mac);
4016 * Later passes of sync-to-convergence may decide to rewrite data
4017 * in place to avoid more disk reallocations. This presents a problem
4018 * for encryption because this consitutes rewriting the new data with
4019 * the same encryption key and IV. However, this only applies to blocks
4020 * in the MOS (particularly the spacemaps) and we do not encrypt the
4021 * MOS. We assert that the zio is allocating or an intent log write
4024 ASSERT(IO_IS_ALLOCATING(zio) || ot == DMU_OT_INTENT_LOG);
4025 ASSERT(BP_GET_LEVEL(bp) == 0 || ot == DMU_OT_INTENT_LOG);
4026 ASSERT(spa_feature_is_active(spa, SPA_FEATURE_ENCRYPTION));
4027 ASSERT3U(psize, !=, 0);
4029 enc_buf = zio_buf_alloc(psize);
4030 eabd = abd_get_from_buf(enc_buf, psize);
4031 abd_take_ownership_of_buf(eabd, B_TRUE);
4034 * For an explanation of what encryption parameters are stored
4035 * where, see the block comment in zio_crypt.c.
4037 if (ot == DMU_OT_INTENT_LOG) {
4038 zio_crypt_decode_params_bp(bp, salt, iv);
4040 BP_SET_CRYPT(bp, B_TRUE);
4043 /* Perform the encryption. This should not fail */
4044 VERIFY0(spa_do_crypt_abd(B_TRUE, spa, &zio->io_bookmark,
4045 BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp),
4046 salt, iv, mac, psize, zio->io_abd, eabd, &no_crypt));
4048 /* encode encryption metadata into the bp */
4049 if (ot == DMU_OT_INTENT_LOG) {
4051 * ZIL blocks store the MAC in the embedded checksum, so the
4052 * transform must always be applied.
4054 zio_crypt_encode_mac_zil(enc_buf, mac);
4055 zio_push_transform(zio, eabd, psize, psize, NULL);
4057 BP_SET_CRYPT(bp, B_TRUE);
4058 zio_crypt_encode_params_bp(bp, salt, iv);
4059 zio_crypt_encode_mac_bp(bp, mac);
4062 ASSERT3U(ot, ==, DMU_OT_DNODE);
4065 zio_push_transform(zio, eabd, psize, psize, NULL);
4073 * ==========================================================================
4074 * Generate and verify checksums
4075 * ==========================================================================
4078 zio_checksum_generate(zio_t *zio)
4080 blkptr_t *bp = zio->io_bp;
4081 enum zio_checksum checksum;
4085 * This is zio_write_phys().
4086 * We're either generating a label checksum, or none at all.
4088 checksum = zio->io_prop.zp_checksum;
4090 if (checksum == ZIO_CHECKSUM_OFF)
4093 ASSERT(checksum == ZIO_CHECKSUM_LABEL);
4095 if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
4096 ASSERT(!IO_IS_ALLOCATING(zio));
4097 checksum = ZIO_CHECKSUM_GANG_HEADER;
4099 checksum = BP_GET_CHECKSUM(bp);
4103 zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size);
4109 zio_checksum_verify(zio_t *zio)
4111 zio_bad_cksum_t info;
4112 blkptr_t *bp = zio->io_bp;
4115 ASSERT(zio->io_vd != NULL);
4119 * This is zio_read_phys().
4120 * We're either verifying a label checksum, or nothing at all.
4122 if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
4125 ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
4128 if ((error = zio_checksum_error(zio, &info)) != 0) {
4129 zio->io_error = error;
4130 if (error == ECKSUM &&
4131 !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
4132 zfs_ereport_start_checksum(zio->io_spa,
4133 zio->io_vd, &zio->io_bookmark, zio,
4134 zio->io_offset, zio->io_size, NULL, &info);
4142 * Called by RAID-Z to ensure we don't compute the checksum twice.
4145 zio_checksum_verified(zio_t *zio)
4147 zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
4151 * ==========================================================================
4152 * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
4153 * An error of 0 indicates success. ENXIO indicates whole-device failure,
4154 * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
4155 * indicate errors that are specific to one I/O, and most likely permanent.
4156 * Any other error is presumed to be worse because we weren't expecting it.
4157 * ==========================================================================
4160 zio_worst_error(int e1, int e2)
4162 static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
4165 for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
4166 if (e1 == zio_error_rank[r1])
4169 for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
4170 if (e2 == zio_error_rank[r2])
4173 return (r1 > r2 ? e1 : e2);
4177 * ==========================================================================
4179 * ==========================================================================
4182 zio_ready(zio_t *zio)
4184 blkptr_t *bp = zio->io_bp;
4185 zio_t *pio, *pio_next;
4186 zio_link_t *zl = NULL;
4188 if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT | ZIO_CHILD_DDT_BIT,
4193 if (zio->io_ready) {
4194 ASSERT(IO_IS_ALLOCATING(zio));
4195 ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
4196 (zio->io_flags & ZIO_FLAG_NOPWRITE));
4197 ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
4202 if (bp != NULL && bp != &zio->io_bp_copy)
4203 zio->io_bp_copy = *bp;
4205 if (zio->io_error != 0) {
4206 zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
4208 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
4209 ASSERT(IO_IS_ALLOCATING(zio));
4210 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
4211 ASSERT(zio->io_metaslab_class != NULL);
4214 * We were unable to allocate anything, unreserve and
4215 * issue the next I/O to allocate.
4217 metaslab_class_throttle_unreserve(
4218 zio->io_metaslab_class, zio->io_prop.zp_copies,
4219 zio->io_allocator, zio);
4220 zio_allocate_dispatch(zio->io_spa, zio->io_allocator);
4224 mutex_enter(&zio->io_lock);
4225 zio->io_state[ZIO_WAIT_READY] = 1;
4226 pio = zio_walk_parents(zio, &zl);
4227 mutex_exit(&zio->io_lock);
4230 * As we notify zio's parents, new parents could be added.
4231 * New parents go to the head of zio's io_parent_list, however,
4232 * so we will (correctly) not notify them. The remainder of zio's
4233 * io_parent_list, from 'pio_next' onward, cannot change because
4234 * all parents must wait for us to be done before they can be done.
4236 for (; pio != NULL; pio = pio_next) {
4237 pio_next = zio_walk_parents(zio, &zl);
4238 zio_notify_parent(pio, zio, ZIO_WAIT_READY, NULL);
4241 if (zio->io_flags & ZIO_FLAG_NODATA) {
4242 if (BP_IS_GANG(bp)) {
4243 zio->io_flags &= ~ZIO_FLAG_NODATA;
4245 ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE);
4246 zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
4250 if (zio_injection_enabled &&
4251 zio->io_spa->spa_syncing_txg == zio->io_txg)
4252 zio_handle_ignored_writes(zio);
4258 * Update the allocation throttle accounting.
4261 zio_dva_throttle_done(zio_t *zio)
4263 ASSERTV(zio_t *lio = zio->io_logical);
4264 zio_t *pio = zio_unique_parent(zio);
4265 vdev_t *vd = zio->io_vd;
4266 int flags = METASLAB_ASYNC_ALLOC;
4268 ASSERT3P(zio->io_bp, !=, NULL);
4269 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
4270 ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE);
4271 ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV);
4273 ASSERT3P(vd, ==, vd->vdev_top);
4274 ASSERT(zio_injection_enabled || !(zio->io_flags & ZIO_FLAG_IO_RETRY));
4275 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
4276 ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING);
4277 ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE));
4278 ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA));
4281 * Parents of gang children can have two flavors -- ones that
4282 * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set)
4283 * and ones that allocated the constituent blocks. The allocation
4284 * throttle needs to know the allocating parent zio so we must find
4287 if (pio->io_child_type == ZIO_CHILD_GANG) {
4289 * If our parent is a rewrite gang child then our grandparent
4290 * would have been the one that performed the allocation.
4292 if (pio->io_flags & ZIO_FLAG_IO_REWRITE)
4293 pio = zio_unique_parent(pio);
4294 flags |= METASLAB_GANG_CHILD;
4297 ASSERT(IO_IS_ALLOCATING(pio));
4298 ASSERT3P(zio, !=, zio->io_logical);
4299 ASSERT(zio->io_logical != NULL);
4300 ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR));
4301 ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE);
4302 ASSERT(zio->io_metaslab_class != NULL);
4304 mutex_enter(&pio->io_lock);
4305 metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags,
4306 pio->io_allocator, B_TRUE);
4307 mutex_exit(&pio->io_lock);
4309 metaslab_class_throttle_unreserve(zio->io_metaslab_class, 1,
4310 pio->io_allocator, pio);
4313 * Call into the pipeline to see if there is more work that
4314 * needs to be done. If there is work to be done it will be
4315 * dispatched to another taskq thread.
4317 zio_allocate_dispatch(zio->io_spa, pio->io_allocator);
4321 zio_done(zio_t *zio)
4324 * Always attempt to keep stack usage minimal here since
4325 * we can be called recurisvely up to 19 levels deep.
4327 const uint64_t psize = zio->io_size;
4328 zio_t *pio, *pio_next;
4329 zio_link_t *zl = NULL;
4332 * If our children haven't all completed,
4333 * wait for them and then repeat this pipeline stage.
4335 if (zio_wait_for_children(zio, ZIO_CHILD_ALL_BITS, ZIO_WAIT_DONE)) {
4340 * If the allocation throttle is enabled, then update the accounting.
4341 * We only track child I/Os that are part of an allocating async
4342 * write. We must do this since the allocation is performed
4343 * by the logical I/O but the actual write is done by child I/Os.
4345 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING &&
4346 zio->io_child_type == ZIO_CHILD_VDEV) {
4347 ASSERT(zio->io_metaslab_class != NULL);
4348 ASSERT(zio->io_metaslab_class->mc_alloc_throttle_enabled);
4349 zio_dva_throttle_done(zio);
4353 * If the allocation throttle is enabled, verify that
4354 * we have decremented the refcounts for every I/O that was throttled.
4356 if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) {
4357 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
4358 ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE);
4359 ASSERT(zio->io_bp != NULL);
4361 metaslab_group_alloc_verify(zio->io_spa, zio->io_bp, zio,
4363 VERIFY(zfs_refcount_not_held(
4364 &zio->io_metaslab_class->mc_alloc_slots[zio->io_allocator],
4369 for (int c = 0; c < ZIO_CHILD_TYPES; c++)
4370 for (int w = 0; w < ZIO_WAIT_TYPES; w++)
4371 ASSERT(zio->io_children[c][w] == 0);
4373 if (zio->io_bp != NULL && !BP_IS_EMBEDDED(zio->io_bp)) {
4374 ASSERT(zio->io_bp->blk_pad[0] == 0);
4375 ASSERT(zio->io_bp->blk_pad[1] == 0);
4376 ASSERT(bcmp(zio->io_bp, &zio->io_bp_copy,
4377 sizeof (blkptr_t)) == 0 ||
4378 (zio->io_bp == zio_unique_parent(zio)->io_bp));
4379 if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) &&
4380 zio->io_bp_override == NULL &&
4381 !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
4382 ASSERT3U(zio->io_prop.zp_copies, <=,
4383 BP_GET_NDVAS(zio->io_bp));
4384 ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 ||
4385 (BP_COUNT_GANG(zio->io_bp) ==
4386 BP_GET_NDVAS(zio->io_bp)));
4388 if (zio->io_flags & ZIO_FLAG_NOPWRITE)
4389 VERIFY(BP_EQUAL(zio->io_bp, &zio->io_bp_orig));
4393 * If there were child vdev/gang/ddt errors, they apply to us now.
4395 zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
4396 zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
4397 zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
4400 * If the I/O on the transformed data was successful, generate any
4401 * checksum reports now while we still have the transformed data.
4403 if (zio->io_error == 0) {
4404 while (zio->io_cksum_report != NULL) {
4405 zio_cksum_report_t *zcr = zio->io_cksum_report;
4406 uint64_t align = zcr->zcr_align;
4407 uint64_t asize = P2ROUNDUP(psize, align);
4408 abd_t *adata = zio->io_abd;
4410 if (asize != psize) {
4411 adata = abd_alloc(asize, B_TRUE);
4412 abd_copy(adata, zio->io_abd, psize);
4413 abd_zero_off(adata, psize, asize - psize);
4416 zio->io_cksum_report = zcr->zcr_next;
4417 zcr->zcr_next = NULL;
4418 zcr->zcr_finish(zcr, adata);
4419 zfs_ereport_free_checksum(zcr);
4426 zio_pop_transforms(zio); /* note: may set zio->io_error */
4428 vdev_stat_update(zio, psize);
4431 * If this I/O is attached to a particular vdev is slow, exceeding
4432 * 30 seconds to complete, post an error described the I/O delay.
4433 * We ignore these errors if the device is currently unavailable.
4435 if (zio->io_delay >= MSEC2NSEC(zio_slow_io_ms)) {
4436 if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd)) {
4438 * We want to only increment our slow IO counters if
4439 * the IO is valid (i.e. not if the drive is removed).
4441 * zfs_ereport_post() will also do these checks, but
4442 * it can also ratelimit and have other failures, so we
4443 * need to increment the slow_io counters independent
4446 if (zfs_ereport_is_valid(FM_EREPORT_ZFS_DELAY,
4447 zio->io_spa, zio->io_vd, zio)) {
4448 mutex_enter(&zio->io_vd->vdev_stat_lock);
4449 zio->io_vd->vdev_stat.vs_slow_ios++;
4450 mutex_exit(&zio->io_vd->vdev_stat_lock);
4452 zfs_ereport_post(FM_EREPORT_ZFS_DELAY,
4453 zio->io_spa, zio->io_vd, &zio->io_bookmark,
4459 if (zio->io_error) {
4461 * If this I/O is attached to a particular vdev,
4462 * generate an error message describing the I/O failure
4463 * at the block level. We ignore these errors if the
4464 * device is currently unavailable.
4466 if (zio->io_error != ECKSUM && zio->io_vd != NULL &&
4467 !vdev_is_dead(zio->io_vd))
4468 zfs_ereport_post(FM_EREPORT_ZFS_IO, zio->io_spa,
4469 zio->io_vd, &zio->io_bookmark, zio, 0, 0);
4471 if ((zio->io_error == EIO || !(zio->io_flags &
4472 (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
4473 zio == zio->io_logical) {
4475 * For logical I/O requests, tell the SPA to log the
4476 * error and generate a logical data ereport.
4478 spa_log_error(zio->io_spa, &zio->io_bookmark);
4479 zfs_ereport_post(FM_EREPORT_ZFS_DATA, zio->io_spa,
4480 NULL, &zio->io_bookmark, zio, 0, 0);
4484 if (zio->io_error && zio == zio->io_logical) {
4486 * Determine whether zio should be reexecuted. This will
4487 * propagate all the way to the root via zio_notify_parent().
4489 ASSERT(zio->io_vd == NULL && zio->io_bp != NULL);
4490 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
4492 if (IO_IS_ALLOCATING(zio) &&
4493 !(zio->io_flags & ZIO_FLAG_CANFAIL)) {
4494 if (zio->io_error != ENOSPC)
4495 zio->io_reexecute |= ZIO_REEXECUTE_NOW;
4497 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4500 if ((zio->io_type == ZIO_TYPE_READ ||
4501 zio->io_type == ZIO_TYPE_FREE) &&
4502 !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
4503 zio->io_error == ENXIO &&
4504 spa_load_state(zio->io_spa) == SPA_LOAD_NONE &&
4505 spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE)
4506 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4508 if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
4509 zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
4512 * Here is a possibly good place to attempt to do
4513 * either combinatorial reconstruction or error correction
4514 * based on checksums. It also might be a good place
4515 * to send out preliminary ereports before we suspend
4521 * If there were logical child errors, they apply to us now.
4522 * We defer this until now to avoid conflating logical child
4523 * errors with errors that happened to the zio itself when
4524 * updating vdev stats and reporting FMA events above.
4526 zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
4528 if ((zio->io_error || zio->io_reexecute) &&
4529 IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
4530 !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
4531 zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp);
4533 zio_gang_tree_free(&zio->io_gang_tree);
4536 * Godfather I/Os should never suspend.
4538 if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
4539 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
4540 zio->io_reexecute &= ~ZIO_REEXECUTE_SUSPEND;
4542 if (zio->io_reexecute) {
4544 * This is a logical I/O that wants to reexecute.
4546 * Reexecute is top-down. When an i/o fails, if it's not
4547 * the root, it simply notifies its parent and sticks around.
4548 * The parent, seeing that it still has children in zio_done(),
4549 * does the same. This percolates all the way up to the root.
4550 * The root i/o will reexecute or suspend the entire tree.
4552 * This approach ensures that zio_reexecute() honors
4553 * all the original i/o dependency relationships, e.g.
4554 * parents not executing until children are ready.
4556 ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
4558 zio->io_gang_leader = NULL;
4560 mutex_enter(&zio->io_lock);
4561 zio->io_state[ZIO_WAIT_DONE] = 1;
4562 mutex_exit(&zio->io_lock);
4565 * "The Godfather" I/O monitors its children but is
4566 * not a true parent to them. It will track them through
4567 * the pipeline but severs its ties whenever they get into
4568 * trouble (e.g. suspended). This allows "The Godfather"
4569 * I/O to return status without blocking.
4572 for (pio = zio_walk_parents(zio, &zl); pio != NULL;
4574 zio_link_t *remove_zl = zl;
4575 pio_next = zio_walk_parents(zio, &zl);
4577 if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
4578 (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
4579 zio_remove_child(pio, zio, remove_zl);
4581 * This is a rare code path, so we don't
4582 * bother with "next_to_execute".
4584 zio_notify_parent(pio, zio, ZIO_WAIT_DONE,
4589 if ((pio = zio_unique_parent(zio)) != NULL) {
4591 * We're not a root i/o, so there's nothing to do
4592 * but notify our parent. Don't propagate errors
4593 * upward since we haven't permanently failed yet.
4595 ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
4596 zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
4598 * This is a rare code path, so we don't bother with
4599 * "next_to_execute".
4601 zio_notify_parent(pio, zio, ZIO_WAIT_DONE, NULL);
4602 } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
4604 * We'd fail again if we reexecuted now, so suspend
4605 * until conditions improve (e.g. device comes online).
4607 zio_suspend(zio->io_spa, zio, ZIO_SUSPEND_IOERR);
4610 * Reexecution is potentially a huge amount of work.
4611 * Hand it off to the otherwise-unused claim taskq.
4613 ASSERT(taskq_empty_ent(&zio->io_tqent));
4614 spa_taskq_dispatch_ent(zio->io_spa,
4615 ZIO_TYPE_CLAIM, ZIO_TASKQ_ISSUE,
4616 (task_func_t *)zio_reexecute, zio, 0,
4622 ASSERT(zio->io_child_count == 0);
4623 ASSERT(zio->io_reexecute == 0);
4624 ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
4627 * Report any checksum errors, since the I/O is complete.
4629 while (zio->io_cksum_report != NULL) {
4630 zio_cksum_report_t *zcr = zio->io_cksum_report;
4631 zio->io_cksum_report = zcr->zcr_next;
4632 zcr->zcr_next = NULL;
4633 zcr->zcr_finish(zcr, NULL);
4634 zfs_ereport_free_checksum(zcr);
4637 if (zio->io_flags & ZIO_FLAG_FASTWRITE && zio->io_bp &&
4638 !BP_IS_HOLE(zio->io_bp) && !BP_IS_EMBEDDED(zio->io_bp) &&
4639 !(zio->io_flags & ZIO_FLAG_NOPWRITE)) {
4640 metaslab_fastwrite_unmark(zio->io_spa, zio->io_bp);
4644 * It is the responsibility of the done callback to ensure that this
4645 * particular zio is no longer discoverable for adoption, and as
4646 * such, cannot acquire any new parents.
4651 mutex_enter(&zio->io_lock);
4652 zio->io_state[ZIO_WAIT_DONE] = 1;
4653 mutex_exit(&zio->io_lock);
4656 * We are done executing this zio. We may want to execute a parent
4657 * next. See the comment in zio_notify_parent().
4659 zio_t *next_to_execute = NULL;
4661 for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) {
4662 zio_link_t *remove_zl = zl;
4663 pio_next = zio_walk_parents(zio, &zl);
4664 zio_remove_child(pio, zio, remove_zl);
4665 zio_notify_parent(pio, zio, ZIO_WAIT_DONE, &next_to_execute);
4668 if (zio->io_waiter != NULL) {
4669 mutex_enter(&zio->io_lock);
4670 zio->io_executor = NULL;
4671 cv_broadcast(&zio->io_cv);
4672 mutex_exit(&zio->io_lock);
4677 return (next_to_execute);
4681 * ==========================================================================
4682 * I/O pipeline definition
4683 * ==========================================================================
4685 static zio_pipe_stage_t *zio_pipeline[] = {
4693 zio_checksum_generate,
4709 zio_checksum_verify,
4717 * Compare two zbookmark_phys_t's to see which we would reach first in a
4718 * pre-order traversal of the object tree.
4720 * This is simple in every case aside from the meta-dnode object. For all other
4721 * objects, we traverse them in order (object 1 before object 2, and so on).
4722 * However, all of these objects are traversed while traversing object 0, since
4723 * the data it points to is the list of objects. Thus, we need to convert to a
4724 * canonical representation so we can compare meta-dnode bookmarks to
4725 * non-meta-dnode bookmarks.
4727 * We do this by calculating "equivalents" for each field of the zbookmark.
4728 * zbookmarks outside of the meta-dnode use their own object and level, and
4729 * calculate the level 0 equivalent (the first L0 blkid that is contained in the
4730 * blocks this bookmark refers to) by multiplying their blkid by their span
4731 * (the number of L0 blocks contained within one block at their level).
4732 * zbookmarks inside the meta-dnode calculate their object equivalent
4733 * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
4734 * level + 1<<31 (any value larger than a level could ever be) for their level.
4735 * This causes them to always compare before a bookmark in their object
4736 * equivalent, compare appropriately to bookmarks in other objects, and to
4737 * compare appropriately to other bookmarks in the meta-dnode.
4740 zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
4741 const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
4744 * These variables represent the "equivalent" values for the zbookmark,
4745 * after converting zbookmarks inside the meta dnode to their
4746 * normal-object equivalents.
4748 uint64_t zb1obj, zb2obj;
4749 uint64_t zb1L0, zb2L0;
4750 uint64_t zb1level, zb2level;
4752 if (zb1->zb_object == zb2->zb_object &&
4753 zb1->zb_level == zb2->zb_level &&
4754 zb1->zb_blkid == zb2->zb_blkid)
4758 * BP_SPANB calculates the span in blocks.
4760 zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
4761 zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
4763 if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
4764 zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4766 zb1level = zb1->zb_level + COMPARE_META_LEVEL;
4768 zb1obj = zb1->zb_object;
4769 zb1level = zb1->zb_level;
4772 if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
4773 zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
4775 zb2level = zb2->zb_level + COMPARE_META_LEVEL;
4777 zb2obj = zb2->zb_object;
4778 zb2level = zb2->zb_level;
4781 /* Now that we have a canonical representation, do the comparison. */
4782 if (zb1obj != zb2obj)
4783 return (zb1obj < zb2obj ? -1 : 1);
4784 else if (zb1L0 != zb2L0)
4785 return (zb1L0 < zb2L0 ? -1 : 1);
4786 else if (zb1level != zb2level)
4787 return (zb1level > zb2level ? -1 : 1);
4789 * This can (theoretically) happen if the bookmarks have the same object
4790 * and level, but different blkids, if the block sizes are not the same.
4791 * There is presently no way to change the indirect block sizes
4797 * This function checks the following: given that last_block is the place that
4798 * our traversal stopped last time, does that guarantee that we've visited
4799 * every node under subtree_root? Therefore, we can't just use the raw output
4800 * of zbookmark_compare. We have to pass in a modified version of
4801 * subtree_root; by incrementing the block id, and then checking whether
4802 * last_block is before or equal to that, we can tell whether or not having
4803 * visited last_block implies that all of subtree_root's children have been
4807 zbookmark_subtree_completed(const dnode_phys_t *dnp,
4808 const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
4810 zbookmark_phys_t mod_zb = *subtree_root;
4812 ASSERT(last_block->zb_level == 0);
4814 /* The objset_phys_t isn't before anything. */
4819 * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
4820 * data block size in sectors, because that variable is only used if
4821 * the bookmark refers to a block in the meta-dnode. Since we don't
4822 * know without examining it what object it refers to, and there's no
4823 * harm in passing in this value in other cases, we always pass it in.
4825 * We pass in 0 for the indirect block size shift because zb2 must be
4826 * level 0. The indirect block size is only used to calculate the span
4827 * of the bookmark, but since the bookmark must be level 0, the span is
4828 * always 1, so the math works out.
4830 * If you make changes to how the zbookmark_compare code works, be sure
4831 * to make sure that this code still works afterwards.
4833 return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
4834 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,
4838 #if defined(_KERNEL)
4839 EXPORT_SYMBOL(zio_type_name);
4840 EXPORT_SYMBOL(zio_buf_alloc);
4841 EXPORT_SYMBOL(zio_data_buf_alloc);
4842 EXPORT_SYMBOL(zio_buf_free);
4843 EXPORT_SYMBOL(zio_data_buf_free);
4845 module_param(zio_slow_io_ms, int, 0644);
4846 MODULE_PARM_DESC(zio_slow_io_ms,
4847 "Max I/O completion time (milliseconds) before marking it as slow");
4849 module_param(zio_requeue_io_start_cut_in_line, int, 0644);
4850 MODULE_PARM_DESC(zio_requeue_io_start_cut_in_line, "Prioritize requeued I/O");
4852 module_param(zfs_sync_pass_deferred_free, int, 0644);
4853 MODULE_PARM_DESC(zfs_sync_pass_deferred_free,
4854 "Defer frees starting in this pass");
4856 module_param(zfs_sync_pass_dont_compress, int, 0644);
4857 MODULE_PARM_DESC(zfs_sync_pass_dont_compress,
4858 "Don't compress starting in this pass");
4860 module_param(zfs_sync_pass_rewrite, int, 0644);
4861 MODULE_PARM_DESC(zfs_sync_pass_rewrite,
4862 "Rewrite new bps starting in this pass");
4864 module_param(zio_dva_throttle_enabled, int, 0644);
4865 MODULE_PARM_DESC(zio_dva_throttle_enabled,
4866 "Throttle block allocations in the ZIO pipeline");