4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or https://opensource.org/licenses/CDDL-1.0.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2020 by Delphix. All rights reserved.
24 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
25 * Copyright (c) 2013, Joyent, Inc. All rights reserved.
26 * Copyright (c) 2016, Nexenta Systems, Inc. All rights reserved.
27 * Copyright (c) 2015 by Chunwei Chen. All rights reserved.
28 * Copyright (c) 2019 Datto Inc.
29 * Copyright (c) 2019, Klara Inc.
30 * Copyright (c) 2019, Allan Jude
31 * Copyright (c) 2022 Hewlett Packard Enterprise Development LP.
32 * Copyright (c) 2021, 2022 by Pawel Jakub Dawidek
36 #include <sys/dmu_impl.h>
37 #include <sys/dmu_tx.h>
39 #include <sys/dnode.h>
40 #include <sys/zfs_context.h>
41 #include <sys/dmu_objset.h>
42 #include <sys/dmu_traverse.h>
43 #include <sys/dsl_dataset.h>
44 #include <sys/dsl_dir.h>
45 #include <sys/dsl_pool.h>
46 #include <sys/dsl_synctask.h>
47 #include <sys/dsl_prop.h>
48 #include <sys/dmu_zfetch.h>
49 #include <sys/zfs_ioctl.h>
51 #include <sys/zio_checksum.h>
52 #include <sys/zio_compress.h>
54 #include <sys/zfeature.h>
57 #include <sys/trace_zfs.h>
58 #include <sys/zfs_racct.h>
59 #include <sys/zfs_rlock.h>
61 #include <sys/vmsystm.h>
62 #include <sys/zfs_znode.h>
66 * Enable/disable nopwrite feature.
68 static int zfs_nopwrite_enabled = 1;
71 * Tunable to control percentage of dirtied L1 blocks from frees allowed into
72 * one TXG. After this threshold is crossed, additional dirty blocks from frees
73 * will wait until the next TXG.
74 * A value of zero will disable this throttle.
76 static uint_t zfs_per_txg_dirty_frees_percent = 30;
79 * Enable/disable forcing txg sync when dirty checking for holes with lseek().
80 * By default this is enabled to ensure accurate hole reporting, it can result
81 * in a significant performance penalty for lseek(SEEK_HOLE) heavy workloads.
82 * Disabling this option will result in holes never being reported in dirty
83 * files which is always safe.
85 static int zfs_dmu_offset_next_sync = 1;
88 * Limit the amount we can prefetch with one call to this amount. This
89 * helps to limit the amount of memory that can be used by prefetching.
90 * Larger objects should be prefetched a bit at a time.
93 uint_t dmu_prefetch_max = 8 * 1024 * 1024;
95 uint_t dmu_prefetch_max = 8 * SPA_MAXBLOCKSIZE;
98 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
99 {DMU_BSWAP_UINT8, TRUE, FALSE, FALSE, "unallocated" },
100 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "object directory" },
101 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "object array" },
102 {DMU_BSWAP_UINT8, TRUE, FALSE, FALSE, "packed nvlist" },
103 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "packed nvlist size" },
104 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "bpobj" },
105 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "bpobj header" },
106 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "SPA space map header" },
107 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "SPA space map" },
108 {DMU_BSWAP_UINT64, TRUE, FALSE, TRUE, "ZIL intent log" },
109 {DMU_BSWAP_DNODE, TRUE, FALSE, TRUE, "DMU dnode" },
110 {DMU_BSWAP_OBJSET, TRUE, TRUE, FALSE, "DMU objset" },
111 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "DSL directory" },
112 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL directory child map"},
113 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL dataset snap map" },
114 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL props" },
115 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "DSL dataset" },
116 {DMU_BSWAP_ZNODE, TRUE, FALSE, FALSE, "ZFS znode" },
117 {DMU_BSWAP_OLDACL, TRUE, FALSE, TRUE, "ZFS V0 ACL" },
118 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "ZFS plain file" },
119 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS directory" },
120 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "ZFS master node" },
121 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS delete queue" },
122 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "zvol object" },
123 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "zvol prop" },
124 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "other uint8[]" },
125 {DMU_BSWAP_UINT64, FALSE, FALSE, TRUE, "other uint64[]" },
126 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "other ZAP" },
127 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "persistent error log" },
128 {DMU_BSWAP_UINT8, TRUE, FALSE, FALSE, "SPA history" },
129 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "SPA history offsets" },
130 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "Pool properties" },
131 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL permissions" },
132 {DMU_BSWAP_ACL, TRUE, FALSE, TRUE, "ZFS ACL" },
133 {DMU_BSWAP_UINT8, TRUE, FALSE, TRUE, "ZFS SYSACL" },
134 {DMU_BSWAP_UINT8, TRUE, FALSE, TRUE, "FUID table" },
135 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "FUID table size" },
136 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL dataset next clones"},
137 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "scan work queue" },
138 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS user/group/project used" },
139 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS user/group/project quota"},
140 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "snapshot refcount tags"},
141 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "DDT ZAP algorithm" },
142 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "DDT statistics" },
143 {DMU_BSWAP_UINT8, TRUE, FALSE, TRUE, "System attributes" },
144 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "SA master node" },
145 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "SA attr registration" },
146 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "SA attr layouts" },
147 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "scan translations" },
148 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "deduplicated block" },
149 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL deadlist map" },
150 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "DSL deadlist map hdr" },
151 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL dir clones" },
152 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "bpobj subobj" }
155 dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
156 { byteswap_uint8_array, "uint8" },
157 { byteswap_uint16_array, "uint16" },
158 { byteswap_uint32_array, "uint32" },
159 { byteswap_uint64_array, "uint64" },
160 { zap_byteswap, "zap" },
161 { dnode_buf_byteswap, "dnode" },
162 { dmu_objset_byteswap, "objset" },
163 { zfs_znode_byteswap, "znode" },
164 { zfs_oldacl_byteswap, "oldacl" },
165 { zfs_acl_byteswap, "acl" }
169 dmu_buf_hold_noread_by_dnode(dnode_t *dn, uint64_t offset,
170 const void *tag, dmu_buf_t **dbp)
175 rw_enter(&dn->dn_struct_rwlock, RW_READER);
176 blkid = dbuf_whichblock(dn, 0, offset);
177 db = dbuf_hold(dn, blkid, tag);
178 rw_exit(&dn->dn_struct_rwlock);
182 return (SET_ERROR(EIO));
190 dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset,
191 const void *tag, dmu_buf_t **dbp)
198 err = dnode_hold(os, object, FTAG, &dn);
201 rw_enter(&dn->dn_struct_rwlock, RW_READER);
202 blkid = dbuf_whichblock(dn, 0, offset);
203 db = dbuf_hold(dn, blkid, tag);
204 rw_exit(&dn->dn_struct_rwlock);
205 dnode_rele(dn, FTAG);
209 return (SET_ERROR(EIO));
217 dmu_buf_hold_by_dnode(dnode_t *dn, uint64_t offset,
218 const void *tag, dmu_buf_t **dbp, int flags)
221 int db_flags = DB_RF_CANFAIL;
223 if (flags & DMU_READ_NO_PREFETCH)
224 db_flags |= DB_RF_NOPREFETCH;
225 if (flags & DMU_READ_NO_DECRYPT)
226 db_flags |= DB_RF_NO_DECRYPT;
228 err = dmu_buf_hold_noread_by_dnode(dn, offset, tag, dbp);
230 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
231 err = dbuf_read(db, NULL, db_flags);
242 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
243 const void *tag, dmu_buf_t **dbp, int flags)
246 int db_flags = DB_RF_CANFAIL;
248 if (flags & DMU_READ_NO_PREFETCH)
249 db_flags |= DB_RF_NOPREFETCH;
250 if (flags & DMU_READ_NO_DECRYPT)
251 db_flags |= DB_RF_NO_DECRYPT;
253 err = dmu_buf_hold_noread(os, object, offset, tag, dbp);
255 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
256 err = dbuf_read(db, NULL, db_flags);
269 return (DN_OLD_MAX_BONUSLEN);
273 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx)
275 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
282 if (dn->dn_bonus != db) {
283 error = SET_ERROR(EINVAL);
284 } else if (newsize < 0 || newsize > db_fake->db_size) {
285 error = SET_ERROR(EINVAL);
287 dnode_setbonuslen(dn, newsize, tx);
296 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
298 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
305 if (!DMU_OT_IS_VALID(type)) {
306 error = SET_ERROR(EINVAL);
307 } else if (dn->dn_bonus != db) {
308 error = SET_ERROR(EINVAL);
310 dnode_setbonus_type(dn, type, tx);
319 dmu_get_bonustype(dmu_buf_t *db_fake)
321 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
323 dmu_object_type_t type;
327 type = dn->dn_bonustype;
334 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
339 error = dnode_hold(os, object, FTAG, &dn);
340 dbuf_rm_spill(dn, tx);
341 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
342 dnode_rm_spill(dn, tx);
343 rw_exit(&dn->dn_struct_rwlock);
344 dnode_rele(dn, FTAG);
349 * Lookup and hold the bonus buffer for the provided dnode. If the dnode
350 * has not yet been allocated a new bonus dbuf a will be allocated.
351 * Returns ENOENT, EIO, or 0.
353 int dmu_bonus_hold_by_dnode(dnode_t *dn, const void *tag, dmu_buf_t **dbp,
358 uint32_t db_flags = DB_RF_MUST_SUCCEED;
360 if (flags & DMU_READ_NO_PREFETCH)
361 db_flags |= DB_RF_NOPREFETCH;
362 if (flags & DMU_READ_NO_DECRYPT)
363 db_flags |= DB_RF_NO_DECRYPT;
365 rw_enter(&dn->dn_struct_rwlock, RW_READER);
366 if (dn->dn_bonus == NULL) {
367 if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
368 rw_exit(&dn->dn_struct_rwlock);
369 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
371 if (dn->dn_bonus == NULL)
372 dbuf_create_bonus(dn);
376 /* as long as the bonus buf is held, the dnode will be held */
377 if (zfs_refcount_add(&db->db_holds, tag) == 1) {
378 VERIFY(dnode_add_ref(dn, db));
379 atomic_inc_32(&dn->dn_dbufs_count);
383 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
384 * hold and incrementing the dbuf count to ensure that dnode_move() sees
385 * a dnode hold for every dbuf.
387 rw_exit(&dn->dn_struct_rwlock);
389 error = dbuf_read(db, NULL, db_flags);
391 dnode_evict_bonus(dn);
402 dmu_bonus_hold(objset_t *os, uint64_t object, const void *tag, dmu_buf_t **dbp)
407 error = dnode_hold(os, object, FTAG, &dn);
411 error = dmu_bonus_hold_by_dnode(dn, tag, dbp, DMU_READ_NO_PREFETCH);
412 dnode_rele(dn, FTAG);
418 * returns ENOENT, EIO, or 0.
420 * This interface will allocate a blank spill dbuf when a spill blk
421 * doesn't already exist on the dnode.
423 * if you only want to find an already existing spill db, then
424 * dmu_spill_hold_existing() should be used.
427 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, const void *tag,
430 dmu_buf_impl_t *db = NULL;
433 if ((flags & DB_RF_HAVESTRUCT) == 0)
434 rw_enter(&dn->dn_struct_rwlock, RW_READER);
436 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
438 if ((flags & DB_RF_HAVESTRUCT) == 0)
439 rw_exit(&dn->dn_struct_rwlock);
443 return (SET_ERROR(EIO));
445 err = dbuf_read(db, NULL, flags);
456 dmu_spill_hold_existing(dmu_buf_t *bonus, const void *tag, dmu_buf_t **dbp)
458 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
465 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
466 err = SET_ERROR(EINVAL);
468 rw_enter(&dn->dn_struct_rwlock, RW_READER);
470 if (!dn->dn_have_spill) {
471 err = SET_ERROR(ENOENT);
473 err = dmu_spill_hold_by_dnode(dn,
474 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
477 rw_exit(&dn->dn_struct_rwlock);
485 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, uint32_t flags, const void *tag,
488 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
491 uint32_t db_flags = DB_RF_CANFAIL;
493 if (flags & DMU_READ_NO_DECRYPT)
494 db_flags |= DB_RF_NO_DECRYPT;
498 err = dmu_spill_hold_by_dnode(dn, db_flags, tag, dbp);
505 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
506 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
507 * and can induce severe lock contention when writing to several files
508 * whose dnodes are in the same block.
511 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
512 boolean_t read, const void *tag, int *numbufsp, dmu_buf_t ***dbpp,
516 zstream_t *zs = NULL;
517 uint64_t blkid, nblks, i;
521 boolean_t missed = B_FALSE;
523 ASSERT(!read || length <= DMU_MAX_ACCESS);
526 * Note: We directly notify the prefetch code of this read, so that
527 * we can tell it about the multi-block read. dbuf_read() only knows
528 * about the one block it is accessing.
530 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT |
533 if ((flags & DMU_READ_NO_DECRYPT) != 0)
534 dbuf_flags |= DB_RF_NO_DECRYPT;
536 rw_enter(&dn->dn_struct_rwlock, RW_READER);
537 if (dn->dn_datablkshift) {
538 int blkshift = dn->dn_datablkshift;
539 nblks = (P2ROUNDUP(offset + length, 1ULL << blkshift) -
540 P2ALIGN(offset, 1ULL << blkshift)) >> blkshift;
542 if (offset + length > dn->dn_datablksz) {
543 zfs_panic_recover("zfs: accessing past end of object "
544 "%llx/%llx (size=%u access=%llu+%llu)",
545 (longlong_t)dn->dn_objset->
546 os_dsl_dataset->ds_object,
547 (longlong_t)dn->dn_object, dn->dn_datablksz,
548 (longlong_t)offset, (longlong_t)length);
549 rw_exit(&dn->dn_struct_rwlock);
550 return (SET_ERROR(EIO));
554 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
557 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL,
559 blkid = dbuf_whichblock(dn, 0, offset);
560 if ((flags & DMU_READ_NO_PREFETCH) == 0) {
562 * Prepare the zfetch before initiating the demand reads, so
563 * that if multiple threads block on same indirect block, we
564 * base predictions on the original less racy request order.
566 zs = dmu_zfetch_prepare(&dn->dn_zfetch, blkid, nblks, read,
569 for (i = 0; i < nblks; i++) {
570 dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag);
573 dmu_zfetch_run(&dn->dn_zfetch, zs, missed,
576 rw_exit(&dn->dn_struct_rwlock);
577 dmu_buf_rele_array(dbp, nblks, tag);
580 return (SET_ERROR(EIO));
584 * Initiate async demand data read.
585 * We check the db_state after calling dbuf_read() because
586 * (1) dbuf_read() may change the state to CACHED due to a
587 * hit in the ARC, and (2) on a cache miss, a child will
588 * have been added to "zio" but not yet completed, so the
589 * state will not yet be CACHED.
592 if (i == nblks - 1 && blkid + i < dn->dn_maxblkid &&
593 offset + length < db->db.db_offset +
595 if (offset <= db->db.db_offset)
596 dbuf_flags |= DB_RF_PARTIAL_FIRST;
598 dbuf_flags |= DB_RF_PARTIAL_MORE;
600 (void) dbuf_read(db, zio, dbuf_flags);
601 if (db->db_state != DB_CACHED)
608 zfs_racct_write(length, nblks);
611 dmu_zfetch_run(&dn->dn_zfetch, zs, missed, B_TRUE);
612 rw_exit(&dn->dn_struct_rwlock);
615 /* wait for async read i/o */
618 dmu_buf_rele_array(dbp, nblks, tag);
622 /* wait for other io to complete */
623 for (i = 0; i < nblks; i++) {
624 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
625 mutex_enter(&db->db_mtx);
626 while (db->db_state == DB_READ ||
627 db->db_state == DB_FILL)
628 cv_wait(&db->db_changed, &db->db_mtx);
629 if (db->db_state == DB_UNCACHED)
630 err = SET_ERROR(EIO);
631 mutex_exit(&db->db_mtx);
633 dmu_buf_rele_array(dbp, nblks, tag);
645 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
646 uint64_t length, int read, const void *tag, int *numbufsp,
652 err = dnode_hold(os, object, FTAG, &dn);
656 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
657 numbufsp, dbpp, DMU_READ_PREFETCH);
659 dnode_rele(dn, FTAG);
665 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
666 uint64_t length, boolean_t read, const void *tag, int *numbufsp,
669 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
675 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
676 numbufsp, dbpp, DMU_READ_PREFETCH);
683 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, const void *tag)
686 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
691 for (i = 0; i < numbufs; i++) {
693 dbuf_rele(dbp[i], tag);
696 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
700 * Issue prefetch I/Os for the given blocks. If level is greater than 0, the
701 * indirect blocks prefetched will be those that point to the blocks containing
702 * the data starting at offset, and continuing to offset + len. If the range
703 * it too long, prefetch the first dmu_prefetch_max bytes as requested, while
704 * for the rest only a higher level, also fitting within dmu_prefetch_max. It
705 * should primarily help random reads, since for long sequential reads there is
706 * a speculative prefetcher.
708 * Note that if the indirect blocks above the blocks being prefetched are not
709 * in cache, they will be asynchronously read in. Dnode read by dnode_hold()
710 * is currently synchronous.
713 dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset,
714 uint64_t len, zio_priority_t pri)
718 if (dmu_prefetch_max == 0 || len == 0) {
719 dmu_prefetch_dnode(os, object, pri);
723 if (dnode_hold(os, object, FTAG, &dn) != 0)
726 dmu_prefetch_by_dnode(dn, level, offset, len, pri);
728 dnode_rele(dn, FTAG);
732 dmu_prefetch_by_dnode(dnode_t *dn, int64_t level, uint64_t offset,
733 uint64_t len, zio_priority_t pri)
735 int64_t level2 = level;
736 uint64_t start, end, start2, end2;
739 * Depending on len we may do two prefetches: blocks [start, end) at
740 * level, and following blocks [start2, end2) at higher level2.
742 rw_enter(&dn->dn_struct_rwlock, RW_READER);
743 if (dn->dn_datablkshift != 0) {
745 * The object has multiple blocks. Calculate the full range
746 * of blocks [start, end2) and then split it into two parts,
747 * so that the first [start, end) fits into dmu_prefetch_max.
749 start = dbuf_whichblock(dn, level, offset);
750 end2 = dbuf_whichblock(dn, level, offset + len - 1) + 1;
751 uint8_t ibs = dn->dn_indblkshift;
752 uint8_t bs = (level == 0) ? dn->dn_datablkshift : ibs;
753 uint_t limit = P2ROUNDUP(dmu_prefetch_max, 1 << bs) >> bs;
754 start2 = end = MIN(end2, start + limit);
757 * Find level2 where [start2, end2) fits into dmu_prefetch_max.
759 uint8_t ibps = ibs - SPA_BLKPTRSHIFT;
760 limit = P2ROUNDUP(dmu_prefetch_max, 1 << ibs) >> ibs;
763 start2 = P2ROUNDUP(start2, 1 << ibps) >> ibps;
764 end2 = P2ROUNDUP(end2, 1 << ibps) >> ibps;
765 } while (end2 - start2 > limit);
767 /* There is only one block. Prefetch it or nothing. */
768 start = start2 = end2 = 0;
769 end = start + (level == 0 && offset < dn->dn_datablksz);
772 for (uint64_t i = start; i < end; i++)
773 dbuf_prefetch(dn, level, i, pri, 0);
774 for (uint64_t i = start2; i < end2; i++)
775 dbuf_prefetch(dn, level2, i, pri, 0);
776 rw_exit(&dn->dn_struct_rwlock);
780 * Issue prefetch I/Os for the given object's dnode.
783 dmu_prefetch_dnode(objset_t *os, uint64_t object, zio_priority_t pri)
785 if (object == 0 || object >= DN_MAX_OBJECT)
788 dnode_t *dn = DMU_META_DNODE(os);
789 rw_enter(&dn->dn_struct_rwlock, RW_READER);
790 uint64_t blkid = dbuf_whichblock(dn, 0, object * sizeof (dnode_phys_t));
791 dbuf_prefetch(dn, 0, blkid, pri, 0);
792 rw_exit(&dn->dn_struct_rwlock);
796 * Get the next "chunk" of file data to free. We traverse the file from
797 * the end so that the file gets shorter over time (if we crashes in the
798 * middle, this will leave us in a better state). We find allocated file
799 * data by simply searching the allocated level 1 indirects.
801 * On input, *start should be the first offset that does not need to be
802 * freed (e.g. "offset + length"). On return, *start will be the first
803 * offset that should be freed and l1blks is set to the number of level 1
804 * indirect blocks found within the chunk.
807 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum, uint64_t *l1blks)
810 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
811 /* bytes of data covered by a level-1 indirect block */
812 uint64_t iblkrange = (uint64_t)dn->dn_datablksz *
813 EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
815 ASSERT3U(minimum, <=, *start);
818 * Check if we can free the entire range assuming that all of the
819 * L1 blocks in this range have data. If we can, we use this
820 * worst case value as an estimate so we can avoid having to look
821 * at the object's actual data.
823 uint64_t total_l1blks =
824 (roundup(*start, iblkrange) - (minimum / iblkrange * iblkrange)) /
826 if (total_l1blks <= maxblks) {
827 *l1blks = total_l1blks;
831 ASSERT(ISP2(iblkrange));
833 for (blks = 0; *start > minimum && blks < maxblks; blks++) {
837 * dnode_next_offset(BACKWARDS) will find an allocated L1
838 * indirect block at or before the input offset. We must
839 * decrement *start so that it is at the end of the region
844 err = dnode_next_offset(dn,
845 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
847 /* if there are no indirect blocks before start, we are done */
851 } else if (err != 0) {
856 /* set start to the beginning of this L1 indirect */
857 *start = P2ALIGN(*start, iblkrange);
859 if (*start < minimum)
867 * If this objset is of type OST_ZFS return true if vfs's unmounted flag is set,
868 * otherwise return false.
869 * Used below in dmu_free_long_range_impl() to enable abort when unmounting
872 dmu_objset_zfs_unmounting(objset_t *os)
875 if (dmu_objset_type(os) == DMU_OST_ZFS)
876 return (zfs_get_vfs_flag_unmounted(os));
884 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
887 uint64_t object_size;
889 uint64_t dirty_frees_threshold;
890 dsl_pool_t *dp = dmu_objset_pool(os);
893 return (SET_ERROR(EINVAL));
895 object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
896 if (offset >= object_size)
899 if (zfs_per_txg_dirty_frees_percent <= 100)
900 dirty_frees_threshold =
901 zfs_per_txg_dirty_frees_percent * zfs_dirty_data_max / 100;
903 dirty_frees_threshold = zfs_dirty_data_max / 20;
905 if (length == DMU_OBJECT_END || offset + length > object_size)
906 length = object_size - offset;
908 while (length != 0) {
909 uint64_t chunk_end, chunk_begin, chunk_len;
913 if (dmu_objset_zfs_unmounting(dn->dn_objset))
914 return (SET_ERROR(EINTR));
916 chunk_end = chunk_begin = offset + length;
918 /* move chunk_begin backwards to the beginning of this chunk */
919 err = get_next_chunk(dn, &chunk_begin, offset, &l1blks);
922 ASSERT3U(chunk_begin, >=, offset);
923 ASSERT3U(chunk_begin, <=, chunk_end);
925 chunk_len = chunk_end - chunk_begin;
927 tx = dmu_tx_create(os);
928 dmu_tx_hold_free(tx, dn->dn_object, chunk_begin, chunk_len);
931 * Mark this transaction as typically resulting in a net
932 * reduction in space used.
934 dmu_tx_mark_netfree(tx);
935 err = dmu_tx_assign(tx, TXG_WAIT);
941 uint64_t txg = dmu_tx_get_txg(tx);
943 mutex_enter(&dp->dp_lock);
944 uint64_t long_free_dirty =
945 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK];
946 mutex_exit(&dp->dp_lock);
949 * To avoid filling up a TXG with just frees, wait for
950 * the next TXG to open before freeing more chunks if
951 * we have reached the threshold of frees.
953 if (dirty_frees_threshold != 0 &&
954 long_free_dirty >= dirty_frees_threshold) {
955 DMU_TX_STAT_BUMP(dmu_tx_dirty_frees_delay);
957 txg_wait_open(dp, 0, B_TRUE);
962 * In order to prevent unnecessary write throttling, for each
963 * TXG, we track the cumulative size of L1 blocks being dirtied
964 * in dnode_free_range() below. We compare this number to a
965 * tunable threshold, past which we prevent new L1 dirty freeing
966 * blocks from being added into the open TXG. See
967 * dmu_free_long_range_impl() for details. The threshold
968 * prevents write throttle activation due to dirty freeing L1
969 * blocks taking up a large percentage of zfs_dirty_data_max.
971 mutex_enter(&dp->dp_lock);
972 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] +=
973 l1blks << dn->dn_indblkshift;
974 mutex_exit(&dp->dp_lock);
975 DTRACE_PROBE3(free__long__range,
976 uint64_t, long_free_dirty, uint64_t, chunk_len,
978 dnode_free_range(dn, chunk_begin, chunk_len, tx);
988 dmu_free_long_range(objset_t *os, uint64_t object,
989 uint64_t offset, uint64_t length)
994 err = dnode_hold(os, object, FTAG, &dn);
997 err = dmu_free_long_range_impl(os, dn, offset, length);
1000 * It is important to zero out the maxblkid when freeing the entire
1001 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
1002 * will take the fast path, and (b) dnode_reallocate() can verify
1003 * that the entire file has been freed.
1005 if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
1006 dn->dn_maxblkid = 0;
1008 dnode_rele(dn, FTAG);
1013 dmu_free_long_object(objset_t *os, uint64_t object)
1018 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
1022 tx = dmu_tx_create(os);
1023 dmu_tx_hold_bonus(tx, object);
1024 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
1025 dmu_tx_mark_netfree(tx);
1026 err = dmu_tx_assign(tx, TXG_WAIT);
1028 err = dmu_object_free(os, object, tx);
1038 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
1039 uint64_t size, dmu_tx_t *tx)
1042 int err = dnode_hold(os, object, FTAG, &dn);
1045 ASSERT(offset < UINT64_MAX);
1046 ASSERT(size == DMU_OBJECT_END || size <= UINT64_MAX - offset);
1047 dnode_free_range(dn, offset, size, tx);
1048 dnode_rele(dn, FTAG);
1053 dmu_read_impl(dnode_t *dn, uint64_t offset, uint64_t size,
1054 void *buf, uint32_t flags)
1057 int numbufs, err = 0;
1060 * Deal with odd block sizes, where there can't be data past the first
1061 * block. If we ever do the tail block optimization, we will need to
1062 * handle that here as well.
1064 if (dn->dn_maxblkid == 0) {
1065 uint64_t newsz = offset > dn->dn_datablksz ? 0 :
1066 MIN(size, dn->dn_datablksz - offset);
1067 memset((char *)buf + newsz, 0, size - newsz);
1072 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
1076 * NB: we could do this block-at-a-time, but it's nice
1077 * to be reading in parallel.
1079 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
1080 TRUE, FTAG, &numbufs, &dbp, flags);
1084 for (i = 0; i < numbufs; i++) {
1087 dmu_buf_t *db = dbp[i];
1091 bufoff = offset - db->db_offset;
1092 tocpy = MIN(db->db_size - bufoff, size);
1094 (void) memcpy(buf, (char *)db->db_data + bufoff, tocpy);
1098 buf = (char *)buf + tocpy;
1100 dmu_buf_rele_array(dbp, numbufs, FTAG);
1106 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1107 void *buf, uint32_t flags)
1112 err = dnode_hold(os, object, FTAG, &dn);
1116 err = dmu_read_impl(dn, offset, size, buf, flags);
1117 dnode_rele(dn, FTAG);
1122 dmu_read_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, void *buf,
1125 return (dmu_read_impl(dn, offset, size, buf, flags));
1129 dmu_write_impl(dmu_buf_t **dbp, int numbufs, uint64_t offset, uint64_t size,
1130 const void *buf, dmu_tx_t *tx)
1134 for (i = 0; i < numbufs; i++) {
1137 dmu_buf_t *db = dbp[i];
1141 bufoff = offset - db->db_offset;
1142 tocpy = MIN(db->db_size - bufoff, size);
1144 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1146 if (tocpy == db->db_size)
1147 dmu_buf_will_fill(db, tx, B_FALSE);
1149 dmu_buf_will_dirty(db, tx);
1151 (void) memcpy((char *)db->db_data + bufoff, buf, tocpy);
1153 if (tocpy == db->db_size)
1154 dmu_buf_fill_done(db, tx, B_FALSE);
1158 buf = (char *)buf + tocpy;
1163 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1164 const void *buf, dmu_tx_t *tx)
1172 VERIFY0(dmu_buf_hold_array(os, object, offset, size,
1173 FALSE, FTAG, &numbufs, &dbp));
1174 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1175 dmu_buf_rele_array(dbp, numbufs, FTAG);
1179 * Note: Lustre is an external consumer of this interface.
1182 dmu_write_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size,
1183 const void *buf, dmu_tx_t *tx)
1191 VERIFY0(dmu_buf_hold_array_by_dnode(dn, offset, size,
1192 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH));
1193 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1194 dmu_buf_rele_array(dbp, numbufs, FTAG);
1198 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1207 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
1208 FALSE, FTAG, &numbufs, &dbp));
1210 for (i = 0; i < numbufs; i++) {
1211 dmu_buf_t *db = dbp[i];
1213 dmu_buf_will_not_fill(db, tx);
1215 dmu_buf_rele_array(dbp, numbufs, FTAG);
1219 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset,
1220 void *data, uint8_t etype, uint8_t comp, int uncompressed_size,
1221 int compressed_size, int byteorder, dmu_tx_t *tx)
1225 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES);
1226 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS);
1227 VERIFY0(dmu_buf_hold_noread(os, object, offset,
1230 dmu_buf_write_embedded(db,
1231 data, (bp_embedded_type_t)etype, (enum zio_compress)comp,
1232 uncompressed_size, compressed_size, byteorder, tx);
1234 dmu_buf_rele(db, FTAG);
1238 dmu_redact(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1244 VERIFY0(dmu_buf_hold_array(os, object, offset, size, FALSE, FTAG,
1246 for (i = 0; i < numbufs; i++)
1247 dmu_buf_redact(dbp[i], tx);
1248 dmu_buf_rele_array(dbp, numbufs, FTAG);
1253 dmu_read_uio_dnode(dnode_t *dn, zfs_uio_t *uio, uint64_t size)
1256 int numbufs, i, err;
1259 * NB: we could do this block-at-a-time, but it's nice
1260 * to be reading in parallel.
1262 err = dmu_buf_hold_array_by_dnode(dn, zfs_uio_offset(uio), size,
1263 TRUE, FTAG, &numbufs, &dbp, 0);
1267 for (i = 0; i < numbufs; i++) {
1270 dmu_buf_t *db = dbp[i];
1274 bufoff = zfs_uio_offset(uio) - db->db_offset;
1275 tocpy = MIN(db->db_size - bufoff, size);
1277 err = zfs_uio_fault_move((char *)db->db_data + bufoff, tocpy,
1285 dmu_buf_rele_array(dbp, numbufs, FTAG);
1291 * Read 'size' bytes into the uio buffer.
1292 * From object zdb->db_object.
1293 * Starting at zfs_uio_offset(uio).
1295 * If the caller already has a dbuf in the target object
1296 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1297 * because we don't have to find the dnode_t for the object.
1300 dmu_read_uio_dbuf(dmu_buf_t *zdb, zfs_uio_t *uio, uint64_t size)
1302 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1311 err = dmu_read_uio_dnode(dn, uio, size);
1318 * Read 'size' bytes into the uio buffer.
1319 * From the specified object
1320 * Starting at offset zfs_uio_offset(uio).
1323 dmu_read_uio(objset_t *os, uint64_t object, zfs_uio_t *uio, uint64_t size)
1331 err = dnode_hold(os, object, FTAG, &dn);
1335 err = dmu_read_uio_dnode(dn, uio, size);
1337 dnode_rele(dn, FTAG);
1343 dmu_write_uio_dnode(dnode_t *dn, zfs_uio_t *uio, uint64_t size, dmu_tx_t *tx)
1350 err = dmu_buf_hold_array_by_dnode(dn, zfs_uio_offset(uio), size,
1351 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1355 for (i = 0; i < numbufs; i++) {
1358 dmu_buf_t *db = dbp[i];
1362 offset_t off = zfs_uio_offset(uio);
1363 bufoff = off - db->db_offset;
1364 tocpy = MIN(db->db_size - bufoff, size);
1366 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1368 if (tocpy == db->db_size)
1369 dmu_buf_will_fill(db, tx, B_TRUE);
1371 dmu_buf_will_dirty(db, tx);
1373 err = zfs_uio_fault_move((char *)db->db_data + bufoff,
1374 tocpy, UIO_WRITE, uio);
1376 if (tocpy == db->db_size && dmu_buf_fill_done(db, tx, err)) {
1377 /* The fill was reverted. Undo any uio progress. */
1378 zfs_uio_advance(uio, off - zfs_uio_offset(uio));
1387 dmu_buf_rele_array(dbp, numbufs, FTAG);
1392 * Write 'size' bytes from the uio buffer.
1393 * To object zdb->db_object.
1394 * Starting at offset zfs_uio_offset(uio).
1396 * If the caller already has a dbuf in the target object
1397 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1398 * because we don't have to find the dnode_t for the object.
1401 dmu_write_uio_dbuf(dmu_buf_t *zdb, zfs_uio_t *uio, uint64_t size,
1404 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1413 err = dmu_write_uio_dnode(dn, uio, size, tx);
1420 * Write 'size' bytes from the uio buffer.
1421 * To the specified object.
1422 * Starting at offset zfs_uio_offset(uio).
1425 dmu_write_uio(objset_t *os, uint64_t object, zfs_uio_t *uio, uint64_t size,
1434 err = dnode_hold(os, object, FTAG, &dn);
1438 err = dmu_write_uio_dnode(dn, uio, size, tx);
1440 dnode_rele(dn, FTAG);
1444 #endif /* _KERNEL */
1447 * Allocate a loaned anonymous arc buffer.
1450 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1452 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1454 return (arc_loan_buf(db->db_objset->os_spa, B_FALSE, size));
1458 * Free a loaned arc buffer.
1461 dmu_return_arcbuf(arc_buf_t *buf)
1463 arc_return_buf(buf, FTAG);
1464 arc_buf_destroy(buf, FTAG);
1468 * A "lightweight" write is faster than a regular write (e.g.
1469 * dmu_write_by_dnode() or dmu_assign_arcbuf_by_dnode()), because it avoids the
1470 * CPU cost of creating a dmu_buf_impl_t and arc_buf_[hdr_]_t. However, the
1471 * data can not be read or overwritten until the transaction's txg has been
1472 * synced. This makes it appropriate for workloads that are known to be
1473 * (temporarily) write-only, like "zfs receive".
1475 * A single block is written, starting at the specified offset in bytes. If
1476 * the call is successful, it returns 0 and the provided abd has been
1477 * consumed (the caller should not free it).
1480 dmu_lightweight_write_by_dnode(dnode_t *dn, uint64_t offset, abd_t *abd,
1481 const zio_prop_t *zp, zio_flag_t flags, dmu_tx_t *tx)
1483 dbuf_dirty_record_t *dr =
1484 dbuf_dirty_lightweight(dn, dbuf_whichblock(dn, 0, offset), tx);
1486 return (SET_ERROR(EIO));
1487 dr->dt.dll.dr_abd = abd;
1488 dr->dt.dll.dr_props = *zp;
1489 dr->dt.dll.dr_flags = flags;
1494 * When possible directly assign passed loaned arc buffer to a dbuf.
1495 * If this is not possible copy the contents of passed arc buf via
1499 dmu_assign_arcbuf_by_dnode(dnode_t *dn, uint64_t offset, arc_buf_t *buf,
1503 objset_t *os = dn->dn_objset;
1504 uint64_t object = dn->dn_object;
1505 uint32_t blksz = (uint32_t)arc_buf_lsize(buf);
1508 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1509 blkid = dbuf_whichblock(dn, 0, offset);
1510 db = dbuf_hold(dn, blkid, FTAG);
1511 rw_exit(&dn->dn_struct_rwlock);
1513 return (SET_ERROR(EIO));
1516 * We can only assign if the offset is aligned and the arc buf is the
1517 * same size as the dbuf.
1519 if (offset == db->db.db_offset && blksz == db->db.db_size) {
1520 zfs_racct_write(blksz, 1);
1521 dbuf_assign_arcbuf(db, buf, tx);
1522 dbuf_rele(db, FTAG);
1524 /* compressed bufs must always be assignable to their dbuf */
1525 ASSERT3U(arc_get_compression(buf), ==, ZIO_COMPRESS_OFF);
1526 ASSERT(!(buf->b_flags & ARC_BUF_FLAG_COMPRESSED));
1528 dbuf_rele(db, FTAG);
1529 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1530 dmu_return_arcbuf(buf);
1537 dmu_assign_arcbuf_by_dbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1541 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1543 DB_DNODE_ENTER(dbuf);
1544 err = dmu_assign_arcbuf_by_dnode(DB_DNODE(dbuf), offset, buf, tx);
1545 DB_DNODE_EXIT(dbuf);
1551 dbuf_dirty_record_t *dsa_dr;
1552 dmu_sync_cb_t *dsa_done;
1558 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1561 dmu_sync_arg_t *dsa = varg;
1562 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1563 blkptr_t *bp = zio->io_bp;
1565 if (zio->io_error == 0) {
1566 if (BP_IS_HOLE(bp)) {
1568 * A block of zeros may compress to a hole, but the
1569 * block size still needs to be known for replay.
1571 BP_SET_LSIZE(bp, db->db_size);
1572 } else if (!BP_IS_EMBEDDED(bp)) {
1573 ASSERT(BP_GET_LEVEL(bp) == 0);
1580 dmu_sync_late_arrival_ready(zio_t *zio)
1582 dmu_sync_ready(zio, NULL, zio->io_private);
1586 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1589 dmu_sync_arg_t *dsa = varg;
1590 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1591 dmu_buf_impl_t *db = dr->dr_dbuf;
1592 zgd_t *zgd = dsa->dsa_zgd;
1595 * Record the vdev(s) backing this blkptr so they can be flushed after
1596 * the writes for the lwb have completed.
1598 if (zio->io_error == 0) {
1599 zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp);
1602 mutex_enter(&db->db_mtx);
1603 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1604 if (zio->io_error == 0) {
1605 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1606 if (dr->dt.dl.dr_nopwrite) {
1607 blkptr_t *bp = zio->io_bp;
1608 blkptr_t *bp_orig = &zio->io_bp_orig;
1609 uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
1611 ASSERT(BP_EQUAL(bp, bp_orig));
1612 VERIFY(BP_EQUAL(bp, db->db_blkptr));
1613 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1614 VERIFY(zio_checksum_table[chksum].ci_flags &
1615 ZCHECKSUM_FLAG_NOPWRITE);
1617 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1618 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1619 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1622 * Old style holes are filled with all zeros, whereas
1623 * new-style holes maintain their lsize, type, level,
1624 * and birth time (see zio_write_compress). While we
1625 * need to reset the BP_SET_LSIZE() call that happened
1626 * in dmu_sync_ready for old style holes, we do *not*
1627 * want to wipe out the information contained in new
1628 * style holes. Thus, only zero out the block pointer if
1629 * it's an old style hole.
1631 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) &&
1632 BP_GET_LOGICAL_BIRTH(&dr->dt.dl.dr_overridden_by) == 0)
1633 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1635 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1637 cv_broadcast(&db->db_changed);
1638 mutex_exit(&db->db_mtx);
1640 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1642 kmem_free(dsa, sizeof (*dsa));
1646 dmu_sync_late_arrival_done(zio_t *zio)
1648 blkptr_t *bp = zio->io_bp;
1649 dmu_sync_arg_t *dsa = zio->io_private;
1650 zgd_t *zgd = dsa->dsa_zgd;
1652 if (zio->io_error == 0) {
1654 * Record the vdev(s) backing this blkptr so they can be
1655 * flushed after the writes for the lwb have completed.
1657 zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp);
1659 if (!BP_IS_HOLE(bp)) {
1660 blkptr_t *bp_orig __maybe_unused = &zio->io_bp_orig;
1661 ASSERT(!(zio->io_flags & ZIO_FLAG_NOPWRITE));
1662 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1663 ASSERT(BP_GET_LOGICAL_BIRTH(zio->io_bp) == zio->io_txg);
1664 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1665 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1669 dmu_tx_commit(dsa->dsa_tx);
1671 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1673 abd_free(zio->io_abd);
1674 kmem_free(dsa, sizeof (*dsa));
1678 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1679 zio_prop_t *zp, zbookmark_phys_t *zb)
1681 dmu_sync_arg_t *dsa;
1685 error = dbuf_read((dmu_buf_impl_t *)zgd->zgd_db, NULL,
1686 DB_RF_CANFAIL | DB_RF_NOPREFETCH);
1690 tx = dmu_tx_create(os);
1691 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1693 * This transaction does not produce any dirty data or log blocks, so
1694 * it should not be throttled. All other cases wait for TXG sync, by
1695 * which time the log block we are writing will be obsolete, so we can
1696 * skip waiting and just return error here instead.
1698 if (dmu_tx_assign(tx, TXG_NOWAIT | TXG_NOTHROTTLE) != 0) {
1700 /* Make zl_get_data do txg_waited_synced() */
1701 return (SET_ERROR(EIO));
1705 * In order to prevent the zgd's lwb from being free'd prior to
1706 * dmu_sync_late_arrival_done() being called, we have to ensure
1707 * the lwb's "max txg" takes this tx's txg into account.
1709 zil_lwb_add_txg(zgd->zgd_lwb, dmu_tx_get_txg(tx));
1711 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1713 dsa->dsa_done = done;
1718 * Since we are currently syncing this txg, it's nontrivial to
1719 * determine what BP to nopwrite against, so we disable nopwrite.
1721 * When syncing, the db_blkptr is initially the BP of the previous
1722 * txg. We can not nopwrite against it because it will be changed
1723 * (this is similar to the non-late-arrival case where the dbuf is
1724 * dirty in a future txg).
1726 * Then dbuf_write_ready() sets bp_blkptr to the location we will write.
1727 * We can not nopwrite against it because although the BP will not
1728 * (typically) be changed, the data has not yet been persisted to this
1731 * Finally, when dbuf_write_done() is called, it is theoretically
1732 * possible to always nopwrite, because the data that was written in
1733 * this txg is the same data that we are trying to write. However we
1734 * would need to check that this dbuf is not dirty in any future
1735 * txg's (as we do in the normal dmu_sync() path). For simplicity, we
1736 * don't nopwrite in this case.
1738 zp->zp_nopwrite = B_FALSE;
1740 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1741 abd_get_from_buf(zgd->zgd_db->db_data, zgd->zgd_db->db_size),
1742 zgd->zgd_db->db_size, zgd->zgd_db->db_size, zp,
1743 dmu_sync_late_arrival_ready, NULL, dmu_sync_late_arrival_done,
1744 dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1750 * Intent log support: sync the block associated with db to disk.
1751 * N.B. and XXX: the caller is responsible for making sure that the
1752 * data isn't changing while dmu_sync() is writing it.
1756 * EEXIST: this txg has already been synced, so there's nothing to do.
1757 * The caller should not log the write.
1759 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1760 * The caller should not log the write.
1762 * EALREADY: this block is already in the process of being synced.
1763 * The caller should track its progress (somehow).
1765 * EIO: could not do the I/O.
1766 * The caller should do a txg_wait_synced().
1768 * 0: the I/O has been initiated.
1769 * The caller should log this blkptr in the done callback.
1770 * It is possible that the I/O will fail, in which case
1771 * the error will be reported to the done callback and
1772 * propagated to pio from zio_done().
1775 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1777 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1778 objset_t *os = db->db_objset;
1779 dsl_dataset_t *ds = os->os_dsl_dataset;
1780 dbuf_dirty_record_t *dr, *dr_next;
1781 dmu_sync_arg_t *dsa;
1782 zbookmark_phys_t zb;
1786 ASSERT(pio != NULL);
1789 SET_BOOKMARK(&zb, ds->ds_object,
1790 db->db.db_object, db->db_level, db->db_blkid);
1794 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1798 * If we're frozen (running ziltest), we always need to generate a bp.
1800 if (txg > spa_freeze_txg(os->os_spa))
1801 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1804 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1805 * and us. If we determine that this txg is not yet syncing,
1806 * but it begins to sync a moment later, that's OK because the
1807 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1809 mutex_enter(&db->db_mtx);
1811 if (txg <= spa_last_synced_txg(os->os_spa)) {
1813 * This txg has already synced. There's nothing to do.
1815 mutex_exit(&db->db_mtx);
1816 return (SET_ERROR(EEXIST));
1819 if (txg <= spa_syncing_txg(os->os_spa)) {
1821 * This txg is currently syncing, so we can't mess with
1822 * the dirty record anymore; just write a new log block.
1824 mutex_exit(&db->db_mtx);
1825 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1828 dr = dbuf_find_dirty_eq(db, txg);
1832 * There's no dr for this dbuf, so it must have been freed.
1833 * There's no need to log writes to freed blocks, so we're done.
1835 mutex_exit(&db->db_mtx);
1836 return (SET_ERROR(ENOENT));
1839 dr_next = list_next(&db->db_dirty_records, dr);
1840 ASSERT(dr_next == NULL || dr_next->dr_txg < txg);
1842 if (db->db_blkptr != NULL) {
1844 * We need to fill in zgd_bp with the current blkptr so that
1845 * the nopwrite code can check if we're writing the same
1846 * data that's already on disk. We can only nopwrite if we
1847 * are sure that after making the copy, db_blkptr will not
1848 * change until our i/o completes. We ensure this by
1849 * holding the db_mtx, and only allowing nopwrite if the
1850 * block is not already dirty (see below). This is verified
1851 * by dmu_sync_done(), which VERIFYs that the db_blkptr has
1854 *zgd->zgd_bp = *db->db_blkptr;
1858 * Assume the on-disk data is X, the current syncing data (in
1859 * txg - 1) is Y, and the current in-memory data is Z (currently
1862 * We usually want to perform a nopwrite if X and Z are the
1863 * same. However, if Y is different (i.e. the BP is going to
1864 * change before this write takes effect), then a nopwrite will
1865 * be incorrect - we would override with X, which could have
1866 * been freed when Y was written.
1868 * (Note that this is not a concern when we are nop-writing from
1869 * syncing context, because X and Y must be identical, because
1870 * all previous txgs have been synced.)
1872 * Therefore, we disable nopwrite if the current BP could change
1873 * before this TXG. There are two ways it could change: by
1874 * being dirty (dr_next is non-NULL), or by being freed
1875 * (dnode_block_freed()). This behavior is verified by
1876 * zio_done(), which VERIFYs that the override BP is identical
1877 * to the on-disk BP.
1881 if (dr_next != NULL || dnode_block_freed(dn, db->db_blkid))
1882 zp.zp_nopwrite = B_FALSE;
1885 ASSERT(dr->dr_txg == txg);
1886 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1887 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1889 * We have already issued a sync write for this buffer,
1890 * or this buffer has already been synced. It could not
1891 * have been dirtied since, or we would have cleared the state.
1893 mutex_exit(&db->db_mtx);
1894 return (SET_ERROR(EALREADY));
1897 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1898 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1899 mutex_exit(&db->db_mtx);
1901 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1903 dsa->dsa_done = done;
1907 zio_nowait(arc_write(pio, os->os_spa, txg, zgd->zgd_bp,
1908 dr->dt.dl.dr_data, !DBUF_IS_CACHEABLE(db), dbuf_is_l2cacheable(db),
1909 &zp, dmu_sync_ready, NULL, dmu_sync_done, dsa,
1910 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb));
1916 dmu_object_set_nlevels(objset_t *os, uint64_t object, int nlevels, dmu_tx_t *tx)
1921 err = dnode_hold(os, object, FTAG, &dn);
1924 err = dnode_set_nlevels(dn, nlevels, tx);
1925 dnode_rele(dn, FTAG);
1930 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1936 err = dnode_hold(os, object, FTAG, &dn);
1939 err = dnode_set_blksz(dn, size, ibs, tx);
1940 dnode_rele(dn, FTAG);
1945 dmu_object_set_maxblkid(objset_t *os, uint64_t object, uint64_t maxblkid,
1951 err = dnode_hold(os, object, FTAG, &dn);
1954 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1955 dnode_new_blkid(dn, maxblkid, tx, B_FALSE, B_TRUE);
1956 rw_exit(&dn->dn_struct_rwlock);
1957 dnode_rele(dn, FTAG);
1962 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1968 * Send streams include each object's checksum function. This
1969 * check ensures that the receiving system can understand the
1970 * checksum function transmitted.
1972 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS);
1974 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1975 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS);
1976 dn->dn_checksum = checksum;
1977 dnode_setdirty(dn, tx);
1978 dnode_rele(dn, FTAG);
1982 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1988 * Send streams include each object's compression function. This
1989 * check ensures that the receiving system can understand the
1990 * compression function transmitted.
1992 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS);
1994 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1995 dn->dn_compress = compress;
1996 dnode_setdirty(dn, tx);
1997 dnode_rele(dn, FTAG);
2001 * When the "redundant_metadata" property is set to "most", only indirect
2002 * blocks of this level and higher will have an additional ditto block.
2004 static const int zfs_redundant_metadata_most_ditto_level = 2;
2007 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
2009 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
2010 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
2012 enum zio_checksum checksum = os->os_checksum;
2013 enum zio_compress compress = os->os_compress;
2014 uint8_t complevel = os->os_complevel;
2015 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
2016 boolean_t dedup = B_FALSE;
2017 boolean_t nopwrite = B_FALSE;
2018 boolean_t dedup_verify = os->os_dedup_verify;
2019 boolean_t encrypt = B_FALSE;
2020 int copies = os->os_copies;
2023 * We maintain different write policies for each of the following
2026 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
2027 * 3. all other level 0 blocks
2031 * XXX -- we should design a compression algorithm
2032 * that specializes in arrays of bps.
2034 compress = zio_compress_select(os->os_spa,
2035 ZIO_COMPRESS_ON, ZIO_COMPRESS_ON);
2038 * Metadata always gets checksummed. If the data
2039 * checksum is multi-bit correctable, and it's not a
2040 * ZBT-style checksum, then it's suitable for metadata
2041 * as well. Otherwise, the metadata checksum defaults
2044 if (!(zio_checksum_table[checksum].ci_flags &
2045 ZCHECKSUM_FLAG_METADATA) ||
2046 (zio_checksum_table[checksum].ci_flags &
2047 ZCHECKSUM_FLAG_EMBEDDED))
2048 checksum = ZIO_CHECKSUM_FLETCHER_4;
2050 switch (os->os_redundant_metadata) {
2051 case ZFS_REDUNDANT_METADATA_ALL:
2054 case ZFS_REDUNDANT_METADATA_MOST:
2055 if (level >= zfs_redundant_metadata_most_ditto_level ||
2056 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))
2059 case ZFS_REDUNDANT_METADATA_SOME:
2060 if (DMU_OT_IS_CRITICAL(type))
2063 case ZFS_REDUNDANT_METADATA_NONE:
2066 } else if (wp & WP_NOFILL) {
2070 * If we're writing preallocated blocks, we aren't actually
2071 * writing them so don't set any policy properties. These
2072 * blocks are currently only used by an external subsystem
2073 * outside of zfs (i.e. dump) and not written by the zio
2076 compress = ZIO_COMPRESS_OFF;
2077 checksum = ZIO_CHECKSUM_OFF;
2079 compress = zio_compress_select(os->os_spa, dn->dn_compress,
2081 complevel = zio_complevel_select(os->os_spa, compress,
2082 complevel, complevel);
2084 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
2085 zio_checksum_select(dn->dn_checksum, checksum) :
2089 * Determine dedup setting. If we are in dmu_sync(),
2090 * we won't actually dedup now because that's all
2091 * done in syncing context; but we do want to use the
2092 * dedup checksum. If the checksum is not strong
2093 * enough to ensure unique signatures, force
2096 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
2097 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
2098 if (!(zio_checksum_table[checksum].ci_flags &
2099 ZCHECKSUM_FLAG_DEDUP))
2100 dedup_verify = B_TRUE;
2104 * Enable nopwrite if we have secure enough checksum
2105 * algorithm (see comment in zio_nop_write) and
2106 * compression is enabled. We don't enable nopwrite if
2107 * dedup is enabled as the two features are mutually
2110 nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags &
2111 ZCHECKSUM_FLAG_NOPWRITE) &&
2112 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
2116 * All objects in an encrypted objset are protected from modification
2117 * via a MAC. Encrypted objects store their IV and salt in the last DVA
2118 * in the bp, so we cannot use all copies. Encrypted objects are also
2119 * not subject to nopwrite since writing the same data will still
2120 * result in a new ciphertext. Only encrypted blocks can be dedup'd
2121 * to avoid ambiguity in the dedup code since the DDT does not store
2124 if (os->os_encrypted && (wp & WP_NOFILL) == 0) {
2127 if (DMU_OT_IS_ENCRYPTED(type)) {
2128 copies = MIN(copies, SPA_DVAS_PER_BP - 1);
2135 (type == DMU_OT_DNODE || type == DMU_OT_OBJSET)) {
2136 compress = ZIO_COMPRESS_EMPTY;
2140 zp->zp_compress = compress;
2141 zp->zp_complevel = complevel;
2142 zp->zp_checksum = checksum;
2143 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
2144 zp->zp_level = level;
2145 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
2146 zp->zp_dedup = dedup;
2147 zp->zp_dedup_verify = dedup && dedup_verify;
2148 zp->zp_nopwrite = nopwrite;
2149 zp->zp_encrypt = encrypt;
2150 zp->zp_byteorder = ZFS_HOST_BYTEORDER;
2151 memset(zp->zp_salt, 0, ZIO_DATA_SALT_LEN);
2152 memset(zp->zp_iv, 0, ZIO_DATA_IV_LEN);
2153 memset(zp->zp_mac, 0, ZIO_DATA_MAC_LEN);
2154 zp->zp_zpl_smallblk = DMU_OT_IS_FILE(zp->zp_type) ?
2155 os->os_zpl_special_smallblock : 0;
2157 ASSERT3U(zp->zp_compress, !=, ZIO_COMPRESS_INHERIT);
2161 * Reports the location of data and holes in an object. In order to
2162 * accurately report holes all dirty data must be synced to disk. This
2163 * causes extremely poor performance when seeking for holes in a dirty file.
2164 * As a compromise, only provide hole data when the dnode is clean. When
2165 * a dnode is dirty report the dnode as having no holes by returning EBUSY
2166 * which is always safe to do.
2169 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
2172 int restarted = 0, err;
2175 err = dnode_hold(os, object, FTAG, &dn);
2179 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2181 if (dnode_is_dirty(dn)) {
2183 * If the zfs_dmu_offset_next_sync module option is enabled
2184 * then hole reporting has been requested. Dirty dnodes
2185 * must be synced to disk to accurately report holes.
2187 * Provided a RL_READER rangelock spanning 0-UINT64_MAX is
2188 * held by the caller only a single restart will be required.
2189 * We tolerate callers which do not hold the rangelock by
2190 * returning EBUSY and not reporting holes after one restart.
2192 if (zfs_dmu_offset_next_sync) {
2193 rw_exit(&dn->dn_struct_rwlock);
2194 dnode_rele(dn, FTAG);
2197 return (SET_ERROR(EBUSY));
2199 txg_wait_synced(dmu_objset_pool(os), 0);
2204 err = SET_ERROR(EBUSY);
2206 err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK |
2207 (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
2210 rw_exit(&dn->dn_struct_rwlock);
2211 dnode_rele(dn, FTAG);
2217 dmu_read_l0_bps(objset_t *os, uint64_t object, uint64_t offset, uint64_t length,
2218 blkptr_t *bps, size_t *nbpsp)
2220 dmu_buf_t **dbp, *dbuf;
2225 error = dmu_buf_hold_array(os, object, offset, length, FALSE, FTAG,
2228 if (error == ESRCH) {
2229 error = SET_ERROR(ENXIO);
2234 ASSERT3U(numbufs, <=, *nbpsp);
2236 for (int i = 0; i < numbufs; i++) {
2238 db = (dmu_buf_impl_t *)dbuf;
2240 mutex_enter(&db->db_mtx);
2242 if (!list_is_empty(&db->db_dirty_records)) {
2243 dbuf_dirty_record_t *dr;
2245 dr = list_head(&db->db_dirty_records);
2246 if (dr->dt.dl.dr_brtwrite) {
2248 * This is very special case where we clone a
2249 * block and in the same transaction group we
2250 * read its BP (most likely to clone the clone).
2252 bp = &dr->dt.dl.dr_overridden_by;
2255 * The block was modified in the same
2256 * transaction group.
2258 mutex_exit(&db->db_mtx);
2259 error = SET_ERROR(EAGAIN);
2266 mutex_exit(&db->db_mtx);
2270 * The file size was increased, but the block was never
2271 * written, otherwise we would either have the block
2272 * pointer or the dirty record and would not get here.
2273 * It is effectively a hole, so report it as such.
2279 * Make sure we clone only data blocks.
2281 if (BP_IS_METADATA(bp) && !BP_IS_HOLE(bp)) {
2282 error = SET_ERROR(EINVAL);
2287 * If the block was allocated in transaction group that is not
2288 * yet synced, we could clone it, but we couldn't write this
2289 * operation into ZIL, or it may be impossible to replay, since
2290 * the block may appear not yet allocated at that point.
2292 if (BP_GET_BIRTH(bp) > spa_freeze_txg(os->os_spa)) {
2293 error = SET_ERROR(EINVAL);
2296 if (BP_GET_BIRTH(bp) > spa_last_synced_txg(os->os_spa)) {
2297 error = SET_ERROR(EAGAIN);
2306 dmu_buf_rele_array(dbp, numbufs, FTAG);
2312 dmu_brt_clone(objset_t *os, uint64_t object, uint64_t offset, uint64_t length,
2313 dmu_tx_t *tx, const blkptr_t *bps, size_t nbps)
2316 dmu_buf_t **dbp, *dbuf;
2318 struct dirty_leaf *dl;
2319 dbuf_dirty_record_t *dr;
2321 int error = 0, i, numbufs;
2325 VERIFY0(dmu_buf_hold_array(os, object, offset, length, FALSE, FTAG,
2327 ASSERT3U(nbps, ==, numbufs);
2330 * Before we start cloning make sure that the dbufs sizes match new BPs
2331 * sizes. If they don't, that's a no-go, as we are not able to shrink
2334 for (i = 0; i < numbufs; i++) {
2336 db = (dmu_buf_impl_t *)dbuf;
2339 ASSERT0(db->db_level);
2340 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2341 ASSERT(db->db_blkid != DMU_SPILL_BLKID);
2343 if (!BP_IS_HOLE(bp) && BP_GET_LSIZE(bp) != dbuf->db_size) {
2344 error = SET_ERROR(EXDEV);
2349 for (i = 0; i < numbufs; i++) {
2351 db = (dmu_buf_impl_t *)dbuf;
2354 ASSERT0(db->db_level);
2355 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2356 ASSERT(db->db_blkid != DMU_SPILL_BLKID);
2357 ASSERT(BP_IS_HOLE(bp) || dbuf->db_size == BP_GET_LSIZE(bp));
2359 dmu_buf_will_clone(dbuf, tx);
2361 mutex_enter(&db->db_mtx);
2363 dr = list_head(&db->db_dirty_records);
2365 ASSERT3U(dr->dr_txg, ==, tx->tx_txg);
2367 dl->dr_overridden_by = *bp;
2368 if (!BP_IS_HOLE(bp) || BP_GET_LOGICAL_BIRTH(bp) != 0) {
2369 if (!BP_IS_EMBEDDED(bp)) {
2370 BP_SET_BIRTH(&dl->dr_overridden_by, dr->dr_txg,
2373 BP_SET_LOGICAL_BIRTH(&dl->dr_overridden_by,
2377 dl->dr_brtwrite = B_TRUE;
2378 dl->dr_override_state = DR_OVERRIDDEN;
2380 mutex_exit(&db->db_mtx);
2383 * When data in embedded into BP there is no need to create
2384 * BRT entry as there is no data block. Just copy the BP as
2385 * it contains the data.
2387 if (!BP_IS_HOLE(bp) && !BP_IS_EMBEDDED(bp)) {
2388 brt_pending_add(spa, bp, tx);
2392 dmu_buf_rele_array(dbp, numbufs, FTAG);
2398 __dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2400 dnode_phys_t *dnp = dn->dn_phys;
2402 doi->doi_data_block_size = dn->dn_datablksz;
2403 doi->doi_metadata_block_size = dn->dn_indblkshift ?
2404 1ULL << dn->dn_indblkshift : 0;
2405 doi->doi_type = dn->dn_type;
2406 doi->doi_bonus_type = dn->dn_bonustype;
2407 doi->doi_bonus_size = dn->dn_bonuslen;
2408 doi->doi_dnodesize = dn->dn_num_slots << DNODE_SHIFT;
2409 doi->doi_indirection = dn->dn_nlevels;
2410 doi->doi_checksum = dn->dn_checksum;
2411 doi->doi_compress = dn->dn_compress;
2412 doi->doi_nblkptr = dn->dn_nblkptr;
2413 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
2414 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
2415 doi->doi_fill_count = 0;
2416 for (int i = 0; i < dnp->dn_nblkptr; i++)
2417 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]);
2421 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2423 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2424 mutex_enter(&dn->dn_mtx);
2426 __dmu_object_info_from_dnode(dn, doi);
2428 mutex_exit(&dn->dn_mtx);
2429 rw_exit(&dn->dn_struct_rwlock);
2433 * Get information on a DMU object.
2434 * If doi is NULL, just indicates whether the object exists.
2437 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
2440 int err = dnode_hold(os, object, FTAG, &dn);
2446 dmu_object_info_from_dnode(dn, doi);
2448 dnode_rele(dn, FTAG);
2453 * As above, but faster; can be used when you have a held dbuf in hand.
2456 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
2458 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2461 dmu_object_info_from_dnode(DB_DNODE(db), doi);
2466 * Faster still when you only care about the size.
2469 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
2470 u_longlong_t *nblk512)
2472 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2478 *blksize = dn->dn_datablksz;
2479 /* add in number of slots used for the dnode itself */
2480 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
2481 SPA_MINBLOCKSHIFT) + dn->dn_num_slots;
2486 dmu_object_dnsize_from_db(dmu_buf_t *db_fake, int *dnsize)
2488 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2493 *dnsize = dn->dn_num_slots << DNODE_SHIFT;
2498 byteswap_uint64_array(void *vbuf, size_t size)
2500 uint64_t *buf = vbuf;
2501 size_t count = size >> 3;
2504 ASSERT((size & 7) == 0);
2506 for (i = 0; i < count; i++)
2507 buf[i] = BSWAP_64(buf[i]);
2511 byteswap_uint32_array(void *vbuf, size_t size)
2513 uint32_t *buf = vbuf;
2514 size_t count = size >> 2;
2517 ASSERT((size & 3) == 0);
2519 for (i = 0; i < count; i++)
2520 buf[i] = BSWAP_32(buf[i]);
2524 byteswap_uint16_array(void *vbuf, size_t size)
2526 uint16_t *buf = vbuf;
2527 size_t count = size >> 1;
2530 ASSERT((size & 1) == 0);
2532 for (i = 0; i < count; i++)
2533 buf[i] = BSWAP_16(buf[i]);
2537 byteswap_uint8_array(void *vbuf, size_t size)
2539 (void) vbuf, (void) size;
2560 arc_fini(); /* arc depends on l2arc, so arc must go first */
2572 EXPORT_SYMBOL(dmu_bonus_hold);
2573 EXPORT_SYMBOL(dmu_bonus_hold_by_dnode);
2574 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus);
2575 EXPORT_SYMBOL(dmu_buf_rele_array);
2576 EXPORT_SYMBOL(dmu_prefetch);
2577 EXPORT_SYMBOL(dmu_prefetch_by_dnode);
2578 EXPORT_SYMBOL(dmu_prefetch_dnode);
2579 EXPORT_SYMBOL(dmu_free_range);
2580 EXPORT_SYMBOL(dmu_free_long_range);
2581 EXPORT_SYMBOL(dmu_free_long_object);
2582 EXPORT_SYMBOL(dmu_read);
2583 EXPORT_SYMBOL(dmu_read_by_dnode);
2584 EXPORT_SYMBOL(dmu_write);
2585 EXPORT_SYMBOL(dmu_write_by_dnode);
2586 EXPORT_SYMBOL(dmu_prealloc);
2587 EXPORT_SYMBOL(dmu_object_info);
2588 EXPORT_SYMBOL(dmu_object_info_from_dnode);
2589 EXPORT_SYMBOL(dmu_object_info_from_db);
2590 EXPORT_SYMBOL(dmu_object_size_from_db);
2591 EXPORT_SYMBOL(dmu_object_dnsize_from_db);
2592 EXPORT_SYMBOL(dmu_object_set_nlevels);
2593 EXPORT_SYMBOL(dmu_object_set_blocksize);
2594 EXPORT_SYMBOL(dmu_object_set_maxblkid);
2595 EXPORT_SYMBOL(dmu_object_set_checksum);
2596 EXPORT_SYMBOL(dmu_object_set_compress);
2597 EXPORT_SYMBOL(dmu_offset_next);
2598 EXPORT_SYMBOL(dmu_write_policy);
2599 EXPORT_SYMBOL(dmu_sync);
2600 EXPORT_SYMBOL(dmu_request_arcbuf);
2601 EXPORT_SYMBOL(dmu_return_arcbuf);
2602 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dnode);
2603 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dbuf);
2604 EXPORT_SYMBOL(dmu_buf_hold);
2605 EXPORT_SYMBOL(dmu_ot);
2607 ZFS_MODULE_PARAM(zfs, zfs_, nopwrite_enabled, INT, ZMOD_RW,
2608 "Enable NOP writes");
2610 ZFS_MODULE_PARAM(zfs, zfs_, per_txg_dirty_frees_percent, UINT, ZMOD_RW,
2611 "Percentage of dirtied blocks from frees in one TXG");
2613 ZFS_MODULE_PARAM(zfs, zfs_, dmu_offset_next_sync, INT, ZMOD_RW,
2614 "Enable forcing txg sync to find holes");
2617 ZFS_MODULE_PARAM(zfs, , dmu_prefetch_max, UINT, ZMOD_RW,
2618 "Limit one prefetch call to this size");