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, 2017 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.
31 #include <sys/dmu_impl.h>
32 #include <sys/dmu_tx.h>
34 #include <sys/dnode.h>
35 #include <sys/zfs_context.h>
36 #include <sys/dmu_objset.h>
37 #include <sys/dmu_traverse.h>
38 #include <sys/dsl_dataset.h>
39 #include <sys/dsl_dir.h>
40 #include <sys/dsl_pool.h>
41 #include <sys/dsl_synctask.h>
42 #include <sys/dsl_prop.h>
43 #include <sys/dmu_zfetch.h>
44 #include <sys/zfs_ioctl.h>
46 #include <sys/zio_checksum.h>
47 #include <sys/zio_compress.h>
49 #include <sys/zfeature.h>
51 #include <sys/trace_dmu.h>
52 #include <sys/zfs_rlock.h>
54 #include <sys/vmsystm.h>
55 #include <sys/zfs_znode.h>
59 * Enable/disable nopwrite feature.
61 int zfs_nopwrite_enabled = 1;
64 * Tunable to control percentage of dirtied blocks from frees in one TXG.
65 * After this threshold is crossed, additional dirty blocks from frees
66 * wait until the next TXG.
67 * A value of zero will disable this throttle.
69 unsigned long zfs_per_txg_dirty_frees_percent = 30;
72 * Enable/disable forcing txg sync when dirty in dmu_offset_next.
74 int zfs_dmu_offset_next_sync = 0;
77 * This can be used for testing, to ensure that certain actions happen
78 * while in the middle of a remap (which might otherwise complete too
81 int zfs_object_remap_one_indirect_delay_ticks = 0;
83 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
84 { DMU_BSWAP_UINT8, TRUE, FALSE, "unallocated" },
85 { DMU_BSWAP_ZAP, TRUE, FALSE, "object directory" },
86 { DMU_BSWAP_UINT64, TRUE, FALSE, "object array" },
87 { DMU_BSWAP_UINT8, TRUE, FALSE, "packed nvlist" },
88 { DMU_BSWAP_UINT64, TRUE, FALSE, "packed nvlist size" },
89 { DMU_BSWAP_UINT64, TRUE, FALSE, "bpobj" },
90 { DMU_BSWAP_UINT64, TRUE, FALSE, "bpobj header" },
91 { DMU_BSWAP_UINT64, TRUE, FALSE, "SPA space map header" },
92 { DMU_BSWAP_UINT64, TRUE, FALSE, "SPA space map" },
93 { DMU_BSWAP_UINT64, TRUE, TRUE, "ZIL intent log" },
94 { DMU_BSWAP_DNODE, TRUE, TRUE, "DMU dnode" },
95 { DMU_BSWAP_OBJSET, TRUE, FALSE, "DMU objset" },
96 { DMU_BSWAP_UINT64, TRUE, FALSE, "DSL directory" },
97 { DMU_BSWAP_ZAP, TRUE, FALSE, "DSL directory child map"},
98 { DMU_BSWAP_ZAP, TRUE, FALSE, "DSL dataset snap map" },
99 { DMU_BSWAP_ZAP, TRUE, FALSE, "DSL props" },
100 { DMU_BSWAP_UINT64, TRUE, FALSE, "DSL dataset" },
101 { DMU_BSWAP_ZNODE, TRUE, FALSE, "ZFS znode" },
102 { DMU_BSWAP_OLDACL, TRUE, TRUE, "ZFS V0 ACL" },
103 { DMU_BSWAP_UINT8, FALSE, TRUE, "ZFS plain file" },
104 { DMU_BSWAP_ZAP, TRUE, TRUE, "ZFS directory" },
105 { DMU_BSWAP_ZAP, TRUE, FALSE, "ZFS master node" },
106 { DMU_BSWAP_ZAP, TRUE, TRUE, "ZFS delete queue" },
107 { DMU_BSWAP_UINT8, FALSE, TRUE, "zvol object" },
108 { DMU_BSWAP_ZAP, TRUE, FALSE, "zvol prop" },
109 { DMU_BSWAP_UINT8, FALSE, TRUE, "other uint8[]" },
110 { DMU_BSWAP_UINT64, FALSE, TRUE, "other uint64[]" },
111 { DMU_BSWAP_ZAP, TRUE, FALSE, "other ZAP" },
112 { DMU_BSWAP_ZAP, TRUE, FALSE, "persistent error log" },
113 { DMU_BSWAP_UINT8, TRUE, FALSE, "SPA history" },
114 { DMU_BSWAP_UINT64, TRUE, FALSE, "SPA history offsets" },
115 { DMU_BSWAP_ZAP, TRUE, FALSE, "Pool properties" },
116 { DMU_BSWAP_ZAP, TRUE, FALSE, "DSL permissions" },
117 { DMU_BSWAP_ACL, TRUE, TRUE, "ZFS ACL" },
118 { DMU_BSWAP_UINT8, TRUE, TRUE, "ZFS SYSACL" },
119 { DMU_BSWAP_UINT8, TRUE, TRUE, "FUID table" },
120 { DMU_BSWAP_UINT64, TRUE, FALSE, "FUID table size" },
121 { DMU_BSWAP_ZAP, TRUE, FALSE, "DSL dataset next clones"},
122 { DMU_BSWAP_ZAP, TRUE, FALSE, "scan work queue" },
123 { DMU_BSWAP_ZAP, TRUE, TRUE, "ZFS user/group/project used" },
124 { DMU_BSWAP_ZAP, TRUE, TRUE, "ZFS user/group/project quota"},
125 { DMU_BSWAP_ZAP, TRUE, FALSE, "snapshot refcount tags"},
126 { DMU_BSWAP_ZAP, TRUE, FALSE, "DDT ZAP algorithm" },
127 { DMU_BSWAP_ZAP, TRUE, FALSE, "DDT statistics" },
128 { DMU_BSWAP_UINT8, TRUE, TRUE, "System attributes" },
129 { DMU_BSWAP_ZAP, TRUE, TRUE, "SA master node" },
130 { DMU_BSWAP_ZAP, TRUE, TRUE, "SA attr registration" },
131 { DMU_BSWAP_ZAP, TRUE, TRUE, "SA attr layouts" },
132 { DMU_BSWAP_ZAP, TRUE, FALSE, "scan translations" },
133 { DMU_BSWAP_UINT8, FALSE, TRUE, "deduplicated block" },
134 { DMU_BSWAP_ZAP, TRUE, FALSE, "DSL deadlist map" },
135 { DMU_BSWAP_UINT64, TRUE, FALSE, "DSL deadlist map hdr" },
136 { DMU_BSWAP_ZAP, TRUE, FALSE, "DSL dir clones" },
137 { DMU_BSWAP_UINT64, TRUE, FALSE, "bpobj subobj" }
140 const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
141 { byteswap_uint8_array, "uint8" },
142 { byteswap_uint16_array, "uint16" },
143 { byteswap_uint32_array, "uint32" },
144 { byteswap_uint64_array, "uint64" },
145 { zap_byteswap, "zap" },
146 { dnode_buf_byteswap, "dnode" },
147 { dmu_objset_byteswap, "objset" },
148 { zfs_znode_byteswap, "znode" },
149 { zfs_oldacl_byteswap, "oldacl" },
150 { zfs_acl_byteswap, "acl" }
154 dmu_buf_hold_noread_by_dnode(dnode_t *dn, uint64_t offset,
155 void *tag, dmu_buf_t **dbp)
160 blkid = dbuf_whichblock(dn, 0, offset);
161 rw_enter(&dn->dn_struct_rwlock, RW_READER);
162 db = dbuf_hold(dn, blkid, tag);
163 rw_exit(&dn->dn_struct_rwlock);
167 return (SET_ERROR(EIO));
174 dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset,
175 void *tag, dmu_buf_t **dbp)
182 err = dnode_hold(os, object, FTAG, &dn);
185 blkid = dbuf_whichblock(dn, 0, offset);
186 rw_enter(&dn->dn_struct_rwlock, RW_READER);
187 db = dbuf_hold(dn, blkid, tag);
188 rw_exit(&dn->dn_struct_rwlock);
189 dnode_rele(dn, FTAG);
193 return (SET_ERROR(EIO));
201 dmu_buf_hold_by_dnode(dnode_t *dn, uint64_t offset,
202 void *tag, dmu_buf_t **dbp, int flags)
205 int db_flags = DB_RF_CANFAIL;
207 if (flags & DMU_READ_NO_PREFETCH)
208 db_flags |= DB_RF_NOPREFETCH;
209 if (flags & DMU_READ_NO_DECRYPT)
210 db_flags |= DB_RF_NO_DECRYPT;
212 err = dmu_buf_hold_noread_by_dnode(dn, offset, tag, dbp);
214 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
215 err = dbuf_read(db, NULL, db_flags);
226 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
227 void *tag, dmu_buf_t **dbp, int flags)
230 int db_flags = DB_RF_CANFAIL;
232 if (flags & DMU_READ_NO_PREFETCH)
233 db_flags |= DB_RF_NOPREFETCH;
234 if (flags & DMU_READ_NO_DECRYPT)
235 db_flags |= DB_RF_NO_DECRYPT;
237 err = dmu_buf_hold_noread(os, object, offset, tag, dbp);
239 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
240 err = dbuf_read(db, NULL, db_flags);
253 return (DN_OLD_MAX_BONUSLEN);
257 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx)
259 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
266 if (dn->dn_bonus != db) {
267 error = SET_ERROR(EINVAL);
268 } else if (newsize < 0 || newsize > db_fake->db_size) {
269 error = SET_ERROR(EINVAL);
271 dnode_setbonuslen(dn, newsize, tx);
280 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
282 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
289 if (!DMU_OT_IS_VALID(type)) {
290 error = SET_ERROR(EINVAL);
291 } else if (dn->dn_bonus != db) {
292 error = SET_ERROR(EINVAL);
294 dnode_setbonus_type(dn, type, tx);
303 dmu_get_bonustype(dmu_buf_t *db_fake)
305 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
307 dmu_object_type_t type;
311 type = dn->dn_bonustype;
318 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
323 error = dnode_hold(os, object, FTAG, &dn);
324 dbuf_rm_spill(dn, tx);
325 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
326 dnode_rm_spill(dn, tx);
327 rw_exit(&dn->dn_struct_rwlock);
328 dnode_rele(dn, FTAG);
333 * returns ENOENT, EIO, or 0.
336 dmu_bonus_hold_impl(objset_t *os, uint64_t object, void *tag, uint32_t flags,
342 uint32_t db_flags = DB_RF_MUST_SUCCEED;
344 if (flags & DMU_READ_NO_PREFETCH)
345 db_flags |= DB_RF_NOPREFETCH;
346 if (flags & DMU_READ_NO_DECRYPT)
347 db_flags |= DB_RF_NO_DECRYPT;
349 error = dnode_hold(os, object, FTAG, &dn);
353 rw_enter(&dn->dn_struct_rwlock, RW_READER);
354 if (dn->dn_bonus == NULL) {
355 rw_exit(&dn->dn_struct_rwlock);
356 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
357 if (dn->dn_bonus == NULL)
358 dbuf_create_bonus(dn);
362 /* as long as the bonus buf is held, the dnode will be held */
363 if (refcount_add(&db->db_holds, tag) == 1) {
364 VERIFY(dnode_add_ref(dn, db));
365 atomic_inc_32(&dn->dn_dbufs_count);
369 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
370 * hold and incrementing the dbuf count to ensure that dnode_move() sees
371 * a dnode hold for every dbuf.
373 rw_exit(&dn->dn_struct_rwlock);
375 dnode_rele(dn, FTAG);
377 error = dbuf_read(db, NULL, db_flags);
379 dnode_evict_bonus(dn);
390 dmu_bonus_hold(objset_t *os, uint64_t obj, void *tag, dmu_buf_t **dbp)
392 return (dmu_bonus_hold_impl(os, obj, tag, DMU_READ_NO_PREFETCH, dbp));
396 * returns ENOENT, EIO, or 0.
398 * This interface will allocate a blank spill dbuf when a spill blk
399 * doesn't already exist on the dnode.
401 * if you only want to find an already existing spill db, then
402 * dmu_spill_hold_existing() should be used.
405 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp)
407 dmu_buf_impl_t *db = NULL;
410 if ((flags & DB_RF_HAVESTRUCT) == 0)
411 rw_enter(&dn->dn_struct_rwlock, RW_READER);
413 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
415 if ((flags & DB_RF_HAVESTRUCT) == 0)
416 rw_exit(&dn->dn_struct_rwlock);
420 return (SET_ERROR(EIO));
422 err = dbuf_read(db, NULL, flags);
433 dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
435 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
442 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
443 err = SET_ERROR(EINVAL);
445 rw_enter(&dn->dn_struct_rwlock, RW_READER);
447 if (!dn->dn_have_spill) {
448 err = SET_ERROR(ENOENT);
450 err = dmu_spill_hold_by_dnode(dn,
451 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
454 rw_exit(&dn->dn_struct_rwlock);
462 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
464 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
470 err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp);
477 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
478 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
479 * and can induce severe lock contention when writing to several files
480 * whose dnodes are in the same block.
483 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
484 boolean_t read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
487 uint64_t blkid, nblks, i;
492 ASSERT(length <= DMU_MAX_ACCESS);
495 * Note: We directly notify the prefetch code of this read, so that
496 * we can tell it about the multi-block read. dbuf_read() only knows
497 * about the one block it is accessing.
499 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT |
502 rw_enter(&dn->dn_struct_rwlock, RW_READER);
503 if (dn->dn_datablkshift) {
504 int blkshift = dn->dn_datablkshift;
505 nblks = (P2ROUNDUP(offset + length, 1ULL << blkshift) -
506 P2ALIGN(offset, 1ULL << blkshift)) >> blkshift;
508 if (offset + length > dn->dn_datablksz) {
509 zfs_panic_recover("zfs: accessing past end of object "
510 "%llx/%llx (size=%u access=%llu+%llu)",
511 (longlong_t)dn->dn_objset->
512 os_dsl_dataset->ds_object,
513 (longlong_t)dn->dn_object, dn->dn_datablksz,
514 (longlong_t)offset, (longlong_t)length);
515 rw_exit(&dn->dn_struct_rwlock);
516 return (SET_ERROR(EIO));
520 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
522 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
523 blkid = dbuf_whichblock(dn, 0, offset);
524 for (i = 0; i < nblks; i++) {
525 dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag);
527 rw_exit(&dn->dn_struct_rwlock);
528 dmu_buf_rele_array(dbp, nblks, tag);
530 return (SET_ERROR(EIO));
533 /* initiate async i/o */
535 (void) dbuf_read(db, zio, dbuf_flags);
539 if ((flags & DMU_READ_NO_PREFETCH) == 0 &&
540 DNODE_META_IS_CACHEABLE(dn) && length <= zfetch_array_rd_sz) {
541 dmu_zfetch(&dn->dn_zfetch, blkid, nblks,
542 read && DNODE_IS_CACHEABLE(dn));
544 rw_exit(&dn->dn_struct_rwlock);
546 /* wait for async i/o */
549 dmu_buf_rele_array(dbp, nblks, tag);
553 /* wait for other io to complete */
555 for (i = 0; i < nblks; i++) {
556 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
557 mutex_enter(&db->db_mtx);
558 while (db->db_state == DB_READ ||
559 db->db_state == DB_FILL)
560 cv_wait(&db->db_changed, &db->db_mtx);
561 if (db->db_state == DB_UNCACHED)
562 err = SET_ERROR(EIO);
563 mutex_exit(&db->db_mtx);
565 dmu_buf_rele_array(dbp, nblks, tag);
577 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
578 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
583 err = dnode_hold(os, object, FTAG, &dn);
587 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
588 numbufsp, dbpp, DMU_READ_PREFETCH);
590 dnode_rele(dn, FTAG);
596 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
597 uint64_t length, boolean_t read, void *tag, int *numbufsp,
600 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
606 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
607 numbufsp, dbpp, DMU_READ_PREFETCH);
614 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
617 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
622 for (i = 0; i < numbufs; i++) {
624 dbuf_rele(dbp[i], tag);
627 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
631 * Issue prefetch i/os for the given blocks. If level is greater than 0, the
632 * indirect blocks prefeteched will be those that point to the blocks containing
633 * the data starting at offset, and continuing to offset + len.
635 * Note that if the indirect blocks above the blocks being prefetched are not
636 * in cache, they will be asychronously read in.
639 dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset,
640 uint64_t len, zio_priority_t pri)
646 if (len == 0) { /* they're interested in the bonus buffer */
647 dn = DMU_META_DNODE(os);
649 if (object == 0 || object >= DN_MAX_OBJECT)
652 rw_enter(&dn->dn_struct_rwlock, RW_READER);
653 blkid = dbuf_whichblock(dn, level,
654 object * sizeof (dnode_phys_t));
655 dbuf_prefetch(dn, level, blkid, pri, 0);
656 rw_exit(&dn->dn_struct_rwlock);
661 * XXX - Note, if the dnode for the requested object is not
662 * already cached, we will do a *synchronous* read in the
663 * dnode_hold() call. The same is true for any indirects.
665 err = dnode_hold(os, object, FTAG, &dn);
669 rw_enter(&dn->dn_struct_rwlock, RW_READER);
671 * offset + len - 1 is the last byte we want to prefetch for, and offset
672 * is the first. Then dbuf_whichblk(dn, level, off + len - 1) is the
673 * last block we want to prefetch, and dbuf_whichblock(dn, level,
674 * offset) is the first. Then the number we need to prefetch is the
677 if (level > 0 || dn->dn_datablkshift != 0) {
678 nblks = dbuf_whichblock(dn, level, offset + len - 1) -
679 dbuf_whichblock(dn, level, offset) + 1;
681 nblks = (offset < dn->dn_datablksz);
685 blkid = dbuf_whichblock(dn, level, offset);
686 for (int i = 0; i < nblks; i++)
687 dbuf_prefetch(dn, level, blkid + i, pri, 0);
690 rw_exit(&dn->dn_struct_rwlock);
692 dnode_rele(dn, FTAG);
696 * Get the next "chunk" of file data to free. We traverse the file from
697 * the end so that the file gets shorter over time (if we crashes in the
698 * middle, this will leave us in a better state). We find allocated file
699 * data by simply searching the allocated level 1 indirects.
701 * On input, *start should be the first offset that does not need to be
702 * freed (e.g. "offset + length"). On return, *start will be the first
703 * offset that should be freed.
706 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum)
708 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
709 /* bytes of data covered by a level-1 indirect block */
711 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
713 ASSERT3U(minimum, <=, *start);
715 if (*start - minimum <= iblkrange * maxblks) {
719 ASSERT(ISP2(iblkrange));
721 for (uint64_t blks = 0; *start > minimum && blks < maxblks; blks++) {
725 * dnode_next_offset(BACKWARDS) will find an allocated L1
726 * indirect block at or before the input offset. We must
727 * decrement *start so that it is at the end of the region
731 err = dnode_next_offset(dn,
732 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
734 /* if there are no indirect blocks before start, we are done */
738 } else if (err != 0) {
742 /* set start to the beginning of this L1 indirect */
743 *start = P2ALIGN(*start, iblkrange);
745 if (*start < minimum)
751 * If this objset is of type OST_ZFS return true if vfs's unmounted flag is set,
752 * otherwise return false.
753 * Used below in dmu_free_long_range_impl() to enable abort when unmounting
757 dmu_objset_zfs_unmounting(objset_t *os)
760 if (dmu_objset_type(os) == DMU_OST_ZFS)
761 return (zfs_get_vfs_flag_unmounted(os));
767 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
768 uint64_t length, boolean_t raw)
770 uint64_t object_size;
772 uint64_t dirty_frees_threshold;
773 dsl_pool_t *dp = dmu_objset_pool(os);
776 return (SET_ERROR(EINVAL));
778 object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
779 if (offset >= object_size)
782 if (zfs_per_txg_dirty_frees_percent <= 100)
783 dirty_frees_threshold =
784 zfs_per_txg_dirty_frees_percent * zfs_dirty_data_max / 100;
786 dirty_frees_threshold = zfs_dirty_data_max / 4;
788 if (length == DMU_OBJECT_END || offset + length > object_size)
789 length = object_size - offset;
791 while (length != 0) {
792 uint64_t chunk_end, chunk_begin, chunk_len;
793 uint64_t long_free_dirty_all_txgs = 0;
796 if (dmu_objset_zfs_unmounting(dn->dn_objset))
797 return (SET_ERROR(EINTR));
799 chunk_end = chunk_begin = offset + length;
801 /* move chunk_begin backwards to the beginning of this chunk */
802 err = get_next_chunk(dn, &chunk_begin, offset);
805 ASSERT3U(chunk_begin, >=, offset);
806 ASSERT3U(chunk_begin, <=, chunk_end);
808 chunk_len = chunk_end - chunk_begin;
810 mutex_enter(&dp->dp_lock);
811 for (int t = 0; t < TXG_SIZE; t++) {
812 long_free_dirty_all_txgs +=
813 dp->dp_long_free_dirty_pertxg[t];
815 mutex_exit(&dp->dp_lock);
818 * To avoid filling up a TXG with just frees wait for
819 * the next TXG to open before freeing more chunks if
820 * we have reached the threshold of frees
822 if (dirty_frees_threshold != 0 &&
823 long_free_dirty_all_txgs >= dirty_frees_threshold) {
824 txg_wait_open(dp, 0);
828 tx = dmu_tx_create(os);
829 dmu_tx_hold_free(tx, dn->dn_object, chunk_begin, chunk_len);
832 * Mark this transaction as typically resulting in a net
833 * reduction in space used.
835 dmu_tx_mark_netfree(tx);
836 err = dmu_tx_assign(tx, TXG_WAIT);
842 mutex_enter(&dp->dp_lock);
843 dp->dp_long_free_dirty_pertxg[dmu_tx_get_txg(tx) & TXG_MASK] +=
845 mutex_exit(&dp->dp_lock);
846 DTRACE_PROBE3(free__long__range,
847 uint64_t, long_free_dirty_all_txgs, uint64_t, chunk_len,
848 uint64_t, dmu_tx_get_txg(tx));
849 dnode_free_range(dn, chunk_begin, chunk_len, tx);
851 /* if this is a raw free, mark the dirty record as such */
853 dbuf_dirty_record_t *dr = dn->dn_dbuf->db_last_dirty;
855 while (dr != NULL && dr->dr_txg > tx->tx_txg)
857 if (dr != NULL && dr->dr_txg == tx->tx_txg) {
858 dr->dt.dl.dr_raw = B_TRUE;
859 dn->dn_objset->os_next_write_raw
860 [tx->tx_txg & TXG_MASK] = B_TRUE;
872 dmu_free_long_range(objset_t *os, uint64_t object,
873 uint64_t offset, uint64_t length)
878 err = dnode_hold(os, object, FTAG, &dn);
881 err = dmu_free_long_range_impl(os, dn, offset, length, B_FALSE);
884 * It is important to zero out the maxblkid when freeing the entire
885 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
886 * will take the fast path, and (b) dnode_reallocate() can verify
887 * that the entire file has been freed.
889 if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
892 dnode_rele(dn, FTAG);
897 * This function is equivalent to dmu_free_long_range(), but also
898 * marks the new dirty record as a raw write.
901 dmu_free_long_range_raw(objset_t *os, uint64_t object,
902 uint64_t offset, uint64_t length)
907 err = dnode_hold(os, object, FTAG, &dn);
910 err = dmu_free_long_range_impl(os, dn, offset, length, B_TRUE);
913 * It is important to zero out the maxblkid when freeing the entire
914 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
915 * will take the fast path, and (b) dnode_reallocate() can verify
916 * that the entire file has been freed.
918 if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
921 dnode_rele(dn, FTAG);
926 dmu_free_long_object_impl(objset_t *os, uint64_t object, boolean_t raw)
931 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
935 tx = dmu_tx_create(os);
936 dmu_tx_hold_bonus(tx, object);
937 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
938 dmu_tx_mark_netfree(tx);
939 err = dmu_tx_assign(tx, TXG_WAIT);
942 err = dmu_object_dirty_raw(os, object, tx);
944 err = dmu_object_free(os, object, tx);
955 dmu_free_long_object(objset_t *os, uint64_t object)
957 return (dmu_free_long_object_impl(os, object, B_FALSE));
961 dmu_free_long_object_raw(objset_t *os, uint64_t object)
963 return (dmu_free_long_object_impl(os, object, B_TRUE));
968 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
969 uint64_t size, dmu_tx_t *tx)
972 int err = dnode_hold(os, object, FTAG, &dn);
975 ASSERT(offset < UINT64_MAX);
976 ASSERT(size == DMU_OBJECT_END || size <= UINT64_MAX - offset);
977 dnode_free_range(dn, offset, size, tx);
978 dnode_rele(dn, FTAG);
983 dmu_read_impl(dnode_t *dn, uint64_t offset, uint64_t size,
984 void *buf, uint32_t flags)
987 int numbufs, err = 0;
990 * Deal with odd block sizes, where there can't be data past the first
991 * block. If we ever do the tail block optimization, we will need to
992 * handle that here as well.
994 if (dn->dn_maxblkid == 0) {
995 uint64_t newsz = offset > dn->dn_datablksz ? 0 :
996 MIN(size, dn->dn_datablksz - offset);
997 bzero((char *)buf + newsz, size - newsz);
1002 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
1006 * NB: we could do this block-at-a-time, but it's nice
1007 * to be reading in parallel.
1009 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
1010 TRUE, FTAG, &numbufs, &dbp, flags);
1014 for (i = 0; i < numbufs; i++) {
1017 dmu_buf_t *db = dbp[i];
1021 bufoff = offset - db->db_offset;
1022 tocpy = MIN(db->db_size - bufoff, size);
1024 (void) memcpy(buf, (char *)db->db_data + bufoff, tocpy);
1028 buf = (char *)buf + tocpy;
1030 dmu_buf_rele_array(dbp, numbufs, FTAG);
1036 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1037 void *buf, uint32_t flags)
1042 err = dnode_hold(os, object, FTAG, &dn);
1046 err = dmu_read_impl(dn, offset, size, buf, flags);
1047 dnode_rele(dn, FTAG);
1052 dmu_read_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, void *buf,
1055 return (dmu_read_impl(dn, offset, size, buf, flags));
1059 dmu_write_impl(dmu_buf_t **dbp, int numbufs, uint64_t offset, uint64_t size,
1060 const void *buf, dmu_tx_t *tx)
1064 for (i = 0; i < numbufs; i++) {
1067 dmu_buf_t *db = dbp[i];
1071 bufoff = offset - db->db_offset;
1072 tocpy = MIN(db->db_size - bufoff, size);
1074 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1076 if (tocpy == db->db_size)
1077 dmu_buf_will_fill(db, tx);
1079 dmu_buf_will_dirty(db, tx);
1081 (void) memcpy((char *)db->db_data + bufoff, buf, tocpy);
1083 if (tocpy == db->db_size)
1084 dmu_buf_fill_done(db, tx);
1088 buf = (char *)buf + tocpy;
1093 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1094 const void *buf, dmu_tx_t *tx)
1102 VERIFY0(dmu_buf_hold_array(os, object, offset, size,
1103 FALSE, FTAG, &numbufs, &dbp));
1104 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1105 dmu_buf_rele_array(dbp, numbufs, FTAG);
1109 dmu_write_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size,
1110 const void *buf, dmu_tx_t *tx)
1118 VERIFY0(dmu_buf_hold_array_by_dnode(dn, offset, size,
1119 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH));
1120 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1121 dmu_buf_rele_array(dbp, numbufs, FTAG);
1125 dmu_object_remap_one_indirect(objset_t *os, dnode_t *dn,
1126 uint64_t last_removal_txg, uint64_t offset)
1128 uint64_t l1blkid = dbuf_whichblock(dn, 1, offset);
1131 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1132 dmu_buf_impl_t *dbuf = dbuf_hold_level(dn, 1, l1blkid, FTAG);
1133 ASSERT3P(dbuf, !=, NULL);
1136 * If the block hasn't been written yet, this default will ensure
1137 * we don't try to remap it.
1139 uint64_t birth = UINT64_MAX;
1140 ASSERT3U(last_removal_txg, !=, UINT64_MAX);
1141 if (dbuf->db_blkptr != NULL)
1142 birth = dbuf->db_blkptr->blk_birth;
1143 rw_exit(&dn->dn_struct_rwlock);
1146 * If this L1 was already written after the last removal, then we've
1147 * already tried to remap it.
1149 if (birth <= last_removal_txg &&
1150 dbuf_read(dbuf, NULL, DB_RF_MUST_SUCCEED) == 0 &&
1151 dbuf_can_remap(dbuf)) {
1152 dmu_tx_t *tx = dmu_tx_create(os);
1153 dmu_tx_hold_remap_l1indirect(tx, dn->dn_object);
1154 err = dmu_tx_assign(tx, TXG_WAIT);
1156 (void) dbuf_dirty(dbuf, tx);
1163 dbuf_rele(dbuf, FTAG);
1165 delay(zfs_object_remap_one_indirect_delay_ticks);
1171 * Remap all blockpointers in the object, if possible, so that they reference
1172 * only concrete vdevs.
1174 * To do this, iterate over the L0 blockpointers and remap any that reference
1175 * an indirect vdev. Note that we only examine L0 blockpointers; since we
1176 * cannot guarantee that we can remap all blockpointer anyways (due to split
1177 * blocks), we do not want to make the code unnecessarily complicated to
1178 * catch the unlikely case that there is an L1 block on an indirect vdev that
1179 * contains no indirect blockpointers.
1182 dmu_object_remap_indirects(objset_t *os, uint64_t object,
1183 uint64_t last_removal_txg)
1185 uint64_t offset, l1span;
1189 err = dnode_hold(os, object, FTAG, &dn);
1194 if (dn->dn_nlevels <= 1) {
1195 if (issig(JUSTLOOKING) && issig(FORREAL)) {
1196 err = SET_ERROR(EINTR);
1200 * If the dnode has no indirect blocks, we cannot dirty them.
1201 * We still want to remap the blkptr(s) in the dnode if
1202 * appropriate, so mark it as dirty.
1204 if (err == 0 && dnode_needs_remap(dn)) {
1205 dmu_tx_t *tx = dmu_tx_create(os);
1206 dmu_tx_hold_bonus(tx, dn->dn_object);
1207 if ((err = dmu_tx_assign(tx, TXG_WAIT)) == 0) {
1208 dnode_setdirty(dn, tx);
1215 dnode_rele(dn, FTAG);
1220 l1span = 1ULL << (dn->dn_indblkshift - SPA_BLKPTRSHIFT +
1221 dn->dn_datablkshift);
1223 * Find the next L1 indirect that is not a hole.
1225 while (dnode_next_offset(dn, 0, &offset, 2, 1, 0) == 0) {
1226 if (issig(JUSTLOOKING) && issig(FORREAL)) {
1227 err = SET_ERROR(EINTR);
1230 if ((err = dmu_object_remap_one_indirect(os, dn,
1231 last_removal_txg, offset)) != 0) {
1237 dnode_rele(dn, FTAG);
1242 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1251 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
1252 FALSE, FTAG, &numbufs, &dbp));
1254 for (i = 0; i < numbufs; i++) {
1255 dmu_buf_t *db = dbp[i];
1257 dmu_buf_will_not_fill(db, tx);
1259 dmu_buf_rele_array(dbp, numbufs, FTAG);
1263 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset,
1264 void *data, uint8_t etype, uint8_t comp, int uncompressed_size,
1265 int compressed_size, int byteorder, dmu_tx_t *tx)
1269 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES);
1270 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS);
1271 VERIFY0(dmu_buf_hold_noread(os, object, offset,
1274 dmu_buf_write_embedded(db,
1275 data, (bp_embedded_type_t)etype, (enum zio_compress)comp,
1276 uncompressed_size, compressed_size, byteorder, tx);
1278 dmu_buf_rele(db, FTAG);
1282 * DMU support for xuio
1284 kstat_t *xuio_ksp = NULL;
1286 typedef struct xuio_stats {
1287 /* loaned yet not returned arc_buf */
1288 kstat_named_t xuiostat_onloan_rbuf;
1289 kstat_named_t xuiostat_onloan_wbuf;
1290 /* whether a copy is made when loaning out a read buffer */
1291 kstat_named_t xuiostat_rbuf_copied;
1292 kstat_named_t xuiostat_rbuf_nocopy;
1293 /* whether a copy is made when assigning a write buffer */
1294 kstat_named_t xuiostat_wbuf_copied;
1295 kstat_named_t xuiostat_wbuf_nocopy;
1298 static xuio_stats_t xuio_stats = {
1299 { "onloan_read_buf", KSTAT_DATA_UINT64 },
1300 { "onloan_write_buf", KSTAT_DATA_UINT64 },
1301 { "read_buf_copied", KSTAT_DATA_UINT64 },
1302 { "read_buf_nocopy", KSTAT_DATA_UINT64 },
1303 { "write_buf_copied", KSTAT_DATA_UINT64 },
1304 { "write_buf_nocopy", KSTAT_DATA_UINT64 }
1307 #define XUIOSTAT_INCR(stat, val) \
1308 atomic_add_64(&xuio_stats.stat.value.ui64, (val))
1309 #define XUIOSTAT_BUMP(stat) XUIOSTAT_INCR(stat, 1)
1311 #ifdef HAVE_UIO_ZEROCOPY
1313 dmu_xuio_init(xuio_t *xuio, int nblk)
1316 uio_t *uio = &xuio->xu_uio;
1318 uio->uio_iovcnt = nblk;
1319 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
1321 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
1323 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
1324 priv->iovp = (iovec_t *)uio->uio_iov;
1325 XUIO_XUZC_PRIV(xuio) = priv;
1327 if (XUIO_XUZC_RW(xuio) == UIO_READ)
1328 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
1330 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
1336 dmu_xuio_fini(xuio_t *xuio)
1338 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1339 int nblk = priv->cnt;
1341 kmem_free(priv->iovp, nblk * sizeof (iovec_t));
1342 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
1343 kmem_free(priv, sizeof (dmu_xuio_t));
1345 if (XUIO_XUZC_RW(xuio) == UIO_READ)
1346 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
1348 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
1352 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
1353 * and increase priv->next by 1.
1356 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
1359 uio_t *uio = &xuio->xu_uio;
1360 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1361 int i = priv->next++;
1363 ASSERT(i < priv->cnt);
1364 ASSERT(off + n <= arc_buf_lsize(abuf));
1365 iov = (iovec_t *)uio->uio_iov + i;
1366 iov->iov_base = (char *)abuf->b_data + off;
1368 priv->bufs[i] = abuf;
1373 dmu_xuio_cnt(xuio_t *xuio)
1375 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1380 dmu_xuio_arcbuf(xuio_t *xuio, int i)
1382 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1384 ASSERT(i < priv->cnt);
1385 return (priv->bufs[i]);
1389 dmu_xuio_clear(xuio_t *xuio, int i)
1391 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1393 ASSERT(i < priv->cnt);
1394 priv->bufs[i] = NULL;
1396 #endif /* HAVE_UIO_ZEROCOPY */
1399 xuio_stat_init(void)
1401 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
1402 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
1403 KSTAT_FLAG_VIRTUAL);
1404 if (xuio_ksp != NULL) {
1405 xuio_ksp->ks_data = &xuio_stats;
1406 kstat_install(xuio_ksp);
1411 xuio_stat_fini(void)
1413 if (xuio_ksp != NULL) {
1414 kstat_delete(xuio_ksp);
1420 xuio_stat_wbuf_copied(void)
1422 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1426 xuio_stat_wbuf_nocopy(void)
1428 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
1433 dmu_read_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size)
1436 int numbufs, i, err;
1437 #ifdef HAVE_UIO_ZEROCOPY
1438 xuio_t *xuio = NULL;
1442 * NB: we could do this block-at-a-time, but it's nice
1443 * to be reading in parallel.
1445 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1446 TRUE, FTAG, &numbufs, &dbp, 0);
1450 for (i = 0; i < numbufs; i++) {
1453 dmu_buf_t *db = dbp[i];
1457 bufoff = uio->uio_loffset - db->db_offset;
1458 tocpy = MIN(db->db_size - bufoff, size);
1460 #ifdef HAVE_UIO_ZEROCOPY
1462 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
1463 arc_buf_t *dbuf_abuf = dbi->db_buf;
1464 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
1465 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
1467 uio->uio_resid -= tocpy;
1468 uio->uio_loffset += tocpy;
1471 if (abuf == dbuf_abuf)
1472 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
1474 XUIOSTAT_BUMP(xuiostat_rbuf_copied);
1477 err = uiomove((char *)db->db_data + bufoff, tocpy,
1484 dmu_buf_rele_array(dbp, numbufs, FTAG);
1490 * Read 'size' bytes into the uio buffer.
1491 * From object zdb->db_object.
1492 * Starting at offset uio->uio_loffset.
1494 * If the caller already has a dbuf in the target object
1495 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1496 * because we don't have to find the dnode_t for the object.
1499 dmu_read_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size)
1501 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1510 err = dmu_read_uio_dnode(dn, uio, size);
1517 * Read 'size' bytes into the uio buffer.
1518 * From the specified object
1519 * Starting at offset uio->uio_loffset.
1522 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
1530 err = dnode_hold(os, object, FTAG, &dn);
1534 err = dmu_read_uio_dnode(dn, uio, size);
1536 dnode_rele(dn, FTAG);
1542 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
1549 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1550 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1554 for (i = 0; i < numbufs; i++) {
1557 dmu_buf_t *db = dbp[i];
1561 bufoff = uio->uio_loffset - db->db_offset;
1562 tocpy = MIN(db->db_size - bufoff, size);
1564 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1566 if (tocpy == db->db_size)
1567 dmu_buf_will_fill(db, tx);
1569 dmu_buf_will_dirty(db, tx);
1572 * XXX uiomove could block forever (eg.nfs-backed
1573 * pages). There needs to be a uiolockdown() function
1574 * to lock the pages in memory, so that uiomove won't
1577 err = uiomove((char *)db->db_data + bufoff, tocpy,
1580 if (tocpy == db->db_size)
1581 dmu_buf_fill_done(db, tx);
1589 dmu_buf_rele_array(dbp, numbufs, FTAG);
1594 * Write 'size' bytes from the uio buffer.
1595 * To object zdb->db_object.
1596 * Starting at offset uio->uio_loffset.
1598 * If the caller already has a dbuf in the target object
1599 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1600 * because we don't have to find the dnode_t for the object.
1603 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
1606 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1615 err = dmu_write_uio_dnode(dn, uio, size, tx);
1622 * Write 'size' bytes from the uio buffer.
1623 * To the specified object.
1624 * Starting at offset uio->uio_loffset.
1627 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1636 err = dnode_hold(os, object, FTAG, &dn);
1640 err = dmu_write_uio_dnode(dn, uio, size, tx);
1642 dnode_rele(dn, FTAG);
1646 #endif /* _KERNEL */
1649 * Allocate a loaned anonymous arc buffer.
1652 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1654 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1656 return (arc_loan_buf(db->db_objset->os_spa, B_FALSE, size));
1660 * Free a loaned arc buffer.
1663 dmu_return_arcbuf(arc_buf_t *buf)
1665 arc_return_buf(buf, FTAG);
1666 arc_buf_destroy(buf, FTAG);
1670 dmu_convert_mdn_block_to_raw(objset_t *os, uint64_t firstobj,
1671 boolean_t byteorder, const uint8_t *salt, const uint8_t *iv,
1672 const uint8_t *mac, dmu_tx_t *tx)
1675 dmu_buf_t *handle = NULL;
1676 dmu_buf_impl_t *db = NULL;
1677 uint64_t offset = firstobj * DNODE_MIN_SIZE;
1678 uint64_t dsobj = dmu_objset_id(os);
1680 ret = dmu_buf_hold_by_dnode(DMU_META_DNODE(os), offset, FTAG, &handle,
1681 DMU_READ_PREFETCH | DMU_READ_NO_DECRYPT);
1685 dmu_buf_will_change_crypt_params(handle, tx);
1687 db = (dmu_buf_impl_t *)handle;
1688 ASSERT3P(db->db_buf, !=, NULL);
1689 ASSERT3U(dsobj, !=, 0);
1692 * This technically violates the assumption the dmu code makes
1693 * that dnode blocks are only released in syncing context.
1695 (void) arc_release(db->db_buf, db);
1696 arc_convert_to_raw(db->db_buf, dsobj, byteorder, DMU_OT_DNODE,
1699 dmu_buf_rele(handle, FTAG);
1705 dmu_copy_from_buf(objset_t *os, uint64_t object, uint64_t offset,
1706 dmu_buf_t *handle, dmu_tx_t *tx)
1708 dmu_buf_t *dst_handle;
1709 dmu_buf_impl_t *dstdb;
1710 dmu_buf_impl_t *srcdb = (dmu_buf_impl_t *)handle;
1713 boolean_t byteorder;
1714 uint8_t salt[ZIO_DATA_SALT_LEN];
1715 uint8_t iv[ZIO_DATA_IV_LEN];
1716 uint8_t mac[ZIO_DATA_MAC_LEN];
1718 ASSERT3P(srcdb->db_buf, !=, NULL);
1720 /* hold the db that we want to write to */
1721 VERIFY0(dmu_buf_hold(os, object, offset, FTAG, &dst_handle,
1722 DMU_READ_NO_DECRYPT));
1723 dstdb = (dmu_buf_impl_t *)dst_handle;
1724 datalen = arc_buf_size(srcdb->db_buf);
1726 /* allocated an arc buffer that matches the type of srcdb->db_buf */
1727 if (arc_is_encrypted(srcdb->db_buf)) {
1728 arc_get_raw_params(srcdb->db_buf, &byteorder, salt, iv, mac);
1729 abuf = arc_loan_raw_buf(os->os_spa, dmu_objset_id(os),
1730 byteorder, salt, iv, mac, DB_DNODE(dstdb)->dn_type,
1731 datalen, arc_buf_lsize(srcdb->db_buf),
1732 arc_get_compression(srcdb->db_buf));
1734 /* we won't get a compressed db back from dmu_buf_hold() */
1735 ASSERT3U(arc_get_compression(srcdb->db_buf),
1736 ==, ZIO_COMPRESS_OFF);
1737 abuf = arc_loan_buf(os->os_spa,
1738 DMU_OT_IS_METADATA(DB_DNODE(dstdb)->dn_type), datalen);
1741 ASSERT3U(datalen, ==, arc_buf_size(abuf));
1743 /* copy the data to the new buffer and assign it to the dstdb */
1744 bcopy(srcdb->db_buf->b_data, abuf->b_data, datalen);
1745 dbuf_assign_arcbuf(dstdb, abuf, tx);
1746 dmu_buf_rele(dst_handle, FTAG);
1750 * When possible directly assign passed loaned arc buffer to a dbuf.
1751 * If this is not possible copy the contents of passed arc buf via
1755 dmu_assign_arcbuf_by_dnode(dnode_t *dn, uint64_t offset, arc_buf_t *buf,
1759 objset_t *os = dn->dn_objset;
1760 uint64_t object = dn->dn_object;
1761 uint32_t blksz = (uint32_t)arc_buf_lsize(buf);
1764 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1765 blkid = dbuf_whichblock(dn, 0, offset);
1766 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1767 rw_exit(&dn->dn_struct_rwlock);
1770 * We can only assign if the offset is aligned, the arc buf is the
1771 * same size as the dbuf, and the dbuf is not metadata.
1773 if (offset == db->db.db_offset && blksz == db->db.db_size) {
1774 dbuf_assign_arcbuf(db, buf, tx);
1775 dbuf_rele(db, FTAG);
1777 /* compressed bufs must always be assignable to their dbuf */
1778 ASSERT3U(arc_get_compression(buf), ==, ZIO_COMPRESS_OFF);
1779 ASSERT(!(buf->b_flags & ARC_BUF_FLAG_COMPRESSED));
1781 dbuf_rele(db, FTAG);
1782 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1783 dmu_return_arcbuf(buf);
1784 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1789 dmu_assign_arcbuf_by_dbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1792 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1794 DB_DNODE_ENTER(dbuf);
1795 dmu_assign_arcbuf_by_dnode(DB_DNODE(dbuf), offset, buf, tx);
1796 DB_DNODE_EXIT(dbuf);
1800 dbuf_dirty_record_t *dsa_dr;
1801 dmu_sync_cb_t *dsa_done;
1808 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1810 dmu_sync_arg_t *dsa = varg;
1811 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1812 blkptr_t *bp = zio->io_bp;
1814 if (zio->io_error == 0) {
1815 if (BP_IS_HOLE(bp)) {
1817 * A block of zeros may compress to a hole, but the
1818 * block size still needs to be known for replay.
1820 BP_SET_LSIZE(bp, db->db_size);
1821 } else if (!BP_IS_EMBEDDED(bp)) {
1822 ASSERT(BP_GET_LEVEL(bp) == 0);
1829 dmu_sync_late_arrival_ready(zio_t *zio)
1831 dmu_sync_ready(zio, NULL, zio->io_private);
1836 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1838 dmu_sync_arg_t *dsa = varg;
1839 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1840 dmu_buf_impl_t *db = dr->dr_dbuf;
1842 mutex_enter(&db->db_mtx);
1843 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1844 if (zio->io_error == 0) {
1845 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1846 if (dr->dt.dl.dr_nopwrite) {
1847 blkptr_t *bp = zio->io_bp;
1848 blkptr_t *bp_orig = &zio->io_bp_orig;
1849 uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
1851 ASSERT(BP_EQUAL(bp, bp_orig));
1852 VERIFY(BP_EQUAL(bp, db->db_blkptr));
1853 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1854 VERIFY(zio_checksum_table[chksum].ci_flags &
1855 ZCHECKSUM_FLAG_NOPWRITE);
1857 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1858 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1859 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1862 * Old style holes are filled with all zeros, whereas
1863 * new-style holes maintain their lsize, type, level,
1864 * and birth time (see zio_write_compress). While we
1865 * need to reset the BP_SET_LSIZE() call that happened
1866 * in dmu_sync_ready for old style holes, we do *not*
1867 * want to wipe out the information contained in new
1868 * style holes. Thus, only zero out the block pointer if
1869 * it's an old style hole.
1871 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) &&
1872 dr->dt.dl.dr_overridden_by.blk_birth == 0)
1873 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1875 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1877 cv_broadcast(&db->db_changed);
1878 mutex_exit(&db->db_mtx);
1880 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1882 kmem_free(dsa, sizeof (*dsa));
1886 dmu_sync_late_arrival_done(zio_t *zio)
1888 blkptr_t *bp = zio->io_bp;
1889 dmu_sync_arg_t *dsa = zio->io_private;
1890 ASSERTV(blkptr_t *bp_orig = &zio->io_bp_orig);
1892 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1893 ASSERT(!(zio->io_flags & ZIO_FLAG_NOPWRITE));
1894 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1895 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1896 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1897 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1900 dmu_tx_commit(dsa->dsa_tx);
1902 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1904 abd_put(zio->io_abd);
1905 kmem_free(dsa, sizeof (*dsa));
1909 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1910 zio_prop_t *zp, zbookmark_phys_t *zb)
1912 dmu_sync_arg_t *dsa;
1915 tx = dmu_tx_create(os);
1916 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1917 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1919 /* Make zl_get_data do txg_waited_synced() */
1920 return (SET_ERROR(EIO));
1924 * In order to prevent the zgd's lwb from being free'd prior to
1925 * dmu_sync_late_arrival_done() being called, we have to ensure
1926 * the lwb's "max txg" takes this tx's txg into account.
1928 zil_lwb_add_txg(zgd->zgd_lwb, dmu_tx_get_txg(tx));
1930 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1932 dsa->dsa_done = done;
1937 * Since we are currently syncing this txg, it's nontrivial to
1938 * determine what BP to nopwrite against, so we disable nopwrite.
1940 * When syncing, the db_blkptr is initially the BP of the previous
1941 * txg. We can not nopwrite against it because it will be changed
1942 * (this is similar to the non-late-arrival case where the dbuf is
1943 * dirty in a future txg).
1945 * Then dbuf_write_ready() sets bp_blkptr to the location we will write.
1946 * We can not nopwrite against it because although the BP will not
1947 * (typically) be changed, the data has not yet been persisted to this
1950 * Finally, when dbuf_write_done() is called, it is theoretically
1951 * possible to always nopwrite, because the data that was written in
1952 * this txg is the same data that we are trying to write. However we
1953 * would need to check that this dbuf is not dirty in any future
1954 * txg's (as we do in the normal dmu_sync() path). For simplicity, we
1955 * don't nopwrite in this case.
1957 zp->zp_nopwrite = B_FALSE;
1959 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1960 abd_get_from_buf(zgd->zgd_db->db_data, zgd->zgd_db->db_size),
1961 zgd->zgd_db->db_size, zgd->zgd_db->db_size, zp,
1962 dmu_sync_late_arrival_ready, NULL, NULL, dmu_sync_late_arrival_done,
1963 dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1969 * Intent log support: sync the block associated with db to disk.
1970 * N.B. and XXX: the caller is responsible for making sure that the
1971 * data isn't changing while dmu_sync() is writing it.
1975 * EEXIST: this txg has already been synced, so there's nothing to do.
1976 * The caller should not log the write.
1978 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1979 * The caller should not log the write.
1981 * EALREADY: this block is already in the process of being synced.
1982 * The caller should track its progress (somehow).
1984 * EIO: could not do the I/O.
1985 * The caller should do a txg_wait_synced().
1987 * 0: the I/O has been initiated.
1988 * The caller should log this blkptr in the done callback.
1989 * It is possible that the I/O will fail, in which case
1990 * the error will be reported to the done callback and
1991 * propagated to pio from zio_done().
1994 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1996 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1997 objset_t *os = db->db_objset;
1998 dsl_dataset_t *ds = os->os_dsl_dataset;
1999 dbuf_dirty_record_t *dr;
2000 dmu_sync_arg_t *dsa;
2001 zbookmark_phys_t zb;
2005 ASSERT(pio != NULL);
2008 /* dbuf is within the locked range */
2009 ASSERT3U(db->db.db_offset, >=, zgd->zgd_rl->r_off);
2010 ASSERT3U(db->db.db_offset + db->db.db_size, <=,
2011 zgd->zgd_rl->r_off + zgd->zgd_rl->r_len);
2013 SET_BOOKMARK(&zb, ds->ds_object,
2014 db->db.db_object, db->db_level, db->db_blkid);
2018 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
2022 * If we're frozen (running ziltest), we always need to generate a bp.
2024 if (txg > spa_freeze_txg(os->os_spa))
2025 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
2028 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
2029 * and us. If we determine that this txg is not yet syncing,
2030 * but it begins to sync a moment later, that's OK because the
2031 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
2033 mutex_enter(&db->db_mtx);
2035 if (txg <= spa_last_synced_txg(os->os_spa)) {
2037 * This txg has already synced. There's nothing to do.
2039 mutex_exit(&db->db_mtx);
2040 return (SET_ERROR(EEXIST));
2043 if (txg <= spa_syncing_txg(os->os_spa)) {
2045 * This txg is currently syncing, so we can't mess with
2046 * the dirty record anymore; just write a new log block.
2048 mutex_exit(&db->db_mtx);
2049 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
2052 dr = db->db_last_dirty;
2053 while (dr && dr->dr_txg != txg)
2058 * There's no dr for this dbuf, so it must have been freed.
2059 * There's no need to log writes to freed blocks, so we're done.
2061 mutex_exit(&db->db_mtx);
2062 return (SET_ERROR(ENOENT));
2065 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg);
2067 if (db->db_blkptr != NULL) {
2069 * We need to fill in zgd_bp with the current blkptr so that
2070 * the nopwrite code can check if we're writing the same
2071 * data that's already on disk. We can only nopwrite if we
2072 * are sure that after making the copy, db_blkptr will not
2073 * change until our i/o completes. We ensure this by
2074 * holding the db_mtx, and only allowing nopwrite if the
2075 * block is not already dirty (see below). This is verified
2076 * by dmu_sync_done(), which VERIFYs that the db_blkptr has
2079 *zgd->zgd_bp = *db->db_blkptr;
2083 * Assume the on-disk data is X, the current syncing data (in
2084 * txg - 1) is Y, and the current in-memory data is Z (currently
2087 * We usually want to perform a nopwrite if X and Z are the
2088 * same. However, if Y is different (i.e. the BP is going to
2089 * change before this write takes effect), then a nopwrite will
2090 * be incorrect - we would override with X, which could have
2091 * been freed when Y was written.
2093 * (Note that this is not a concern when we are nop-writing from
2094 * syncing context, because X and Y must be identical, because
2095 * all previous txgs have been synced.)
2097 * Therefore, we disable nopwrite if the current BP could change
2098 * before this TXG. There are two ways it could change: by
2099 * being dirty (dr_next is non-NULL), or by being freed
2100 * (dnode_block_freed()). This behavior is verified by
2101 * zio_done(), which VERIFYs that the override BP is identical
2102 * to the on-disk BP.
2106 if (dr->dr_next != NULL || dnode_block_freed(dn, db->db_blkid))
2107 zp.zp_nopwrite = B_FALSE;
2110 ASSERT(dr->dr_txg == txg);
2111 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
2112 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
2114 * We have already issued a sync write for this buffer,
2115 * or this buffer has already been synced. It could not
2116 * have been dirtied since, or we would have cleared the state.
2118 mutex_exit(&db->db_mtx);
2119 return (SET_ERROR(EALREADY));
2122 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
2123 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
2124 mutex_exit(&db->db_mtx);
2126 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
2128 dsa->dsa_done = done;
2132 zio_nowait(arc_write(pio, os->os_spa, txg,
2133 zgd->zgd_bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db),
2134 &zp, dmu_sync_ready, NULL, NULL, dmu_sync_done, dsa,
2135 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb));
2141 dmu_object_set_nlevels(objset_t *os, uint64_t object, int nlevels, dmu_tx_t *tx)
2146 err = dnode_hold(os, object, FTAG, &dn);
2149 err = dnode_set_nlevels(dn, nlevels, tx);
2150 dnode_rele(dn, FTAG);
2155 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
2161 err = dnode_hold(os, object, FTAG, &dn);
2164 err = dnode_set_blksz(dn, size, ibs, tx);
2165 dnode_rele(dn, FTAG);
2170 dmu_object_set_maxblkid(objset_t *os, uint64_t object, uint64_t maxblkid,
2176 err = dnode_hold(os, object, FTAG, &dn);
2179 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2180 dnode_new_blkid(dn, maxblkid, tx, B_FALSE);
2181 rw_exit(&dn->dn_struct_rwlock);
2182 dnode_rele(dn, FTAG);
2187 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
2193 * Send streams include each object's checksum function. This
2194 * check ensures that the receiving system can understand the
2195 * checksum function transmitted.
2197 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS);
2199 VERIFY0(dnode_hold(os, object, FTAG, &dn));
2200 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS);
2201 dn->dn_checksum = checksum;
2202 dnode_setdirty(dn, tx);
2203 dnode_rele(dn, FTAG);
2207 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
2213 * Send streams include each object's compression function. This
2214 * check ensures that the receiving system can understand the
2215 * compression function transmitted.
2217 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS);
2219 VERIFY0(dnode_hold(os, object, FTAG, &dn));
2220 dn->dn_compress = compress;
2221 dnode_setdirty(dn, tx);
2222 dnode_rele(dn, FTAG);
2226 * Dirty an object and set the dirty record's raw flag. This is used
2227 * when writing raw data to an object that will not effect the
2228 * encryption parameters, specifically during raw receives.
2231 dmu_object_dirty_raw(objset_t *os, uint64_t object, dmu_tx_t *tx)
2236 err = dnode_hold(os, object, FTAG, &dn);
2239 dmu_buf_will_change_crypt_params((dmu_buf_t *)dn->dn_dbuf, tx);
2240 dnode_rele(dn, FTAG);
2245 * When the "redundant_metadata" property is set to "most", only indirect
2246 * blocks of this level and higher will have an additional ditto block.
2248 int zfs_redundant_metadata_most_ditto_level = 2;
2251 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
2253 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
2254 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
2256 enum zio_checksum checksum = os->os_checksum;
2257 enum zio_compress compress = os->os_compress;
2258 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
2259 boolean_t dedup = B_FALSE;
2260 boolean_t nopwrite = B_FALSE;
2261 boolean_t dedup_verify = os->os_dedup_verify;
2262 boolean_t encrypt = B_FALSE;
2263 int copies = os->os_copies;
2266 * We maintain different write policies for each of the following
2269 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
2270 * 3. all other level 0 blocks
2274 * XXX -- we should design a compression algorithm
2275 * that specializes in arrays of bps.
2277 compress = zio_compress_select(os->os_spa,
2278 ZIO_COMPRESS_ON, ZIO_COMPRESS_ON);
2281 * Metadata always gets checksummed. If the data
2282 * checksum is multi-bit correctable, and it's not a
2283 * ZBT-style checksum, then it's suitable for metadata
2284 * as well. Otherwise, the metadata checksum defaults
2287 if (!(zio_checksum_table[checksum].ci_flags &
2288 ZCHECKSUM_FLAG_METADATA) ||
2289 (zio_checksum_table[checksum].ci_flags &
2290 ZCHECKSUM_FLAG_EMBEDDED))
2291 checksum = ZIO_CHECKSUM_FLETCHER_4;
2293 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL ||
2294 (os->os_redundant_metadata ==
2295 ZFS_REDUNDANT_METADATA_MOST &&
2296 (level >= zfs_redundant_metadata_most_ditto_level ||
2297 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))))
2299 } else if (wp & WP_NOFILL) {
2303 * If we're writing preallocated blocks, we aren't actually
2304 * writing them so don't set any policy properties. These
2305 * blocks are currently only used by an external subsystem
2306 * outside of zfs (i.e. dump) and not written by the zio
2309 compress = ZIO_COMPRESS_OFF;
2310 checksum = ZIO_CHECKSUM_OFF;
2312 compress = zio_compress_select(os->os_spa, dn->dn_compress,
2315 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
2316 zio_checksum_select(dn->dn_checksum, checksum) :
2320 * Determine dedup setting. If we are in dmu_sync(),
2321 * we won't actually dedup now because that's all
2322 * done in syncing context; but we do want to use the
2323 * dedup checkum. If the checksum is not strong
2324 * enough to ensure unique signatures, force
2327 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
2328 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
2329 if (!(zio_checksum_table[checksum].ci_flags &
2330 ZCHECKSUM_FLAG_DEDUP))
2331 dedup_verify = B_TRUE;
2335 * Enable nopwrite if we have secure enough checksum
2336 * algorithm (see comment in zio_nop_write) and
2337 * compression is enabled. We don't enable nopwrite if
2338 * dedup is enabled as the two features are mutually
2341 nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags &
2342 ZCHECKSUM_FLAG_NOPWRITE) &&
2343 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
2347 * All objects in an encrypted objset are protected from modification
2348 * via a MAC. Encrypted objects store their IV and salt in the last DVA
2349 * in the bp, so we cannot use all copies. Encrypted objects are also
2350 * not subject to nopwrite since writing the same data will still
2351 * result in a new ciphertext. Only encrypted blocks can be dedup'd
2352 * to avoid ambiguity in the dedup code since the DDT does not store
2355 if (os->os_encrypted && (wp & WP_NOFILL) == 0) {
2358 if (DMU_OT_IS_ENCRYPTED(type)) {
2359 copies = MIN(copies, SPA_DVAS_PER_BP - 1);
2366 (type == DMU_OT_DNODE || type == DMU_OT_OBJSET)) {
2367 compress = ZIO_COMPRESS_EMPTY;
2371 zp->zp_compress = compress;
2372 zp->zp_checksum = checksum;
2373 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
2374 zp->zp_level = level;
2375 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
2376 zp->zp_dedup = dedup;
2377 zp->zp_dedup_verify = dedup && dedup_verify;
2378 zp->zp_nopwrite = nopwrite;
2379 zp->zp_encrypt = encrypt;
2380 zp->zp_byteorder = ZFS_HOST_BYTEORDER;
2381 bzero(zp->zp_salt, ZIO_DATA_SALT_LEN);
2382 bzero(zp->zp_iv, ZIO_DATA_IV_LEN);
2383 bzero(zp->zp_mac, ZIO_DATA_MAC_LEN);
2385 ASSERT3U(zp->zp_compress, !=, ZIO_COMPRESS_INHERIT);
2389 * This function is only called from zfs_holey_common() for zpl_llseek()
2390 * in order to determine the location of holes. In order to accurately
2391 * report holes all dirty data must be synced to disk. This causes extremely
2392 * poor performance when seeking for holes in a dirty file. As a compromise,
2393 * only provide hole data when the dnode is clean. When a dnode is dirty
2394 * report the dnode as having no holes which is always a safe thing to do.
2397 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
2401 boolean_t clean = B_TRUE;
2403 err = dnode_hold(os, object, FTAG, &dn);
2408 * Check if dnode is dirty
2410 for (i = 0; i < TXG_SIZE; i++) {
2411 if (multilist_link_active(&dn->dn_dirty_link[i])) {
2418 * If compatibility option is on, sync any current changes before
2419 * we go trundling through the block pointers.
2421 if (!clean && zfs_dmu_offset_next_sync) {
2423 dnode_rele(dn, FTAG);
2424 txg_wait_synced(dmu_objset_pool(os), 0);
2425 err = dnode_hold(os, object, FTAG, &dn);
2431 err = dnode_next_offset(dn,
2432 (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
2434 err = SET_ERROR(EBUSY);
2436 dnode_rele(dn, FTAG);
2442 __dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2444 dnode_phys_t *dnp = dn->dn_phys;
2446 doi->doi_data_block_size = dn->dn_datablksz;
2447 doi->doi_metadata_block_size = dn->dn_indblkshift ?
2448 1ULL << dn->dn_indblkshift : 0;
2449 doi->doi_type = dn->dn_type;
2450 doi->doi_bonus_type = dn->dn_bonustype;
2451 doi->doi_bonus_size = dn->dn_bonuslen;
2452 doi->doi_dnodesize = dn->dn_num_slots << DNODE_SHIFT;
2453 doi->doi_indirection = dn->dn_nlevels;
2454 doi->doi_checksum = dn->dn_checksum;
2455 doi->doi_compress = dn->dn_compress;
2456 doi->doi_nblkptr = dn->dn_nblkptr;
2457 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
2458 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
2459 doi->doi_fill_count = 0;
2460 for (int i = 0; i < dnp->dn_nblkptr; i++)
2461 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]);
2465 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2467 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2468 mutex_enter(&dn->dn_mtx);
2470 __dmu_object_info_from_dnode(dn, doi);
2472 mutex_exit(&dn->dn_mtx);
2473 rw_exit(&dn->dn_struct_rwlock);
2477 * Get information on a DMU object.
2478 * If doi is NULL, just indicates whether the object exists.
2481 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
2484 int err = dnode_hold(os, object, FTAG, &dn);
2490 dmu_object_info_from_dnode(dn, doi);
2492 dnode_rele(dn, FTAG);
2497 * As above, but faster; can be used when you have a held dbuf in hand.
2500 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
2502 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2505 dmu_object_info_from_dnode(DB_DNODE(db), doi);
2510 * Faster still when you only care about the size.
2511 * This is specifically optimized for zfs_getattr().
2514 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
2515 u_longlong_t *nblk512)
2517 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2523 *blksize = dn->dn_datablksz;
2524 /* add in number of slots used for the dnode itself */
2525 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
2526 SPA_MINBLOCKSHIFT) + dn->dn_num_slots;
2531 dmu_object_dnsize_from_db(dmu_buf_t *db_fake, int *dnsize)
2533 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2538 *dnsize = dn->dn_num_slots << DNODE_SHIFT;
2543 byteswap_uint64_array(void *vbuf, size_t size)
2545 uint64_t *buf = vbuf;
2546 size_t count = size >> 3;
2549 ASSERT((size & 7) == 0);
2551 for (i = 0; i < count; i++)
2552 buf[i] = BSWAP_64(buf[i]);
2556 byteswap_uint32_array(void *vbuf, size_t size)
2558 uint32_t *buf = vbuf;
2559 size_t count = size >> 2;
2562 ASSERT((size & 3) == 0);
2564 for (i = 0; i < count; i++)
2565 buf[i] = BSWAP_32(buf[i]);
2569 byteswap_uint16_array(void *vbuf, size_t size)
2571 uint16_t *buf = vbuf;
2572 size_t count = size >> 1;
2575 ASSERT((size & 1) == 0);
2577 for (i = 0; i < count; i++)
2578 buf[i] = BSWAP_16(buf[i]);
2583 byteswap_uint8_array(void *vbuf, size_t size)
2606 arc_fini(); /* arc depends on l2arc, so arc must go first */
2619 #if defined(_KERNEL) && defined(HAVE_SPL)
2620 EXPORT_SYMBOL(dmu_bonus_hold);
2621 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus);
2622 EXPORT_SYMBOL(dmu_buf_rele_array);
2623 EXPORT_SYMBOL(dmu_prefetch);
2624 EXPORT_SYMBOL(dmu_free_range);
2625 EXPORT_SYMBOL(dmu_free_long_range);
2626 EXPORT_SYMBOL(dmu_free_long_range_raw);
2627 EXPORT_SYMBOL(dmu_free_long_object);
2628 EXPORT_SYMBOL(dmu_free_long_object_raw);
2629 EXPORT_SYMBOL(dmu_read);
2630 EXPORT_SYMBOL(dmu_read_by_dnode);
2631 EXPORT_SYMBOL(dmu_write);
2632 EXPORT_SYMBOL(dmu_write_by_dnode);
2633 EXPORT_SYMBOL(dmu_prealloc);
2634 EXPORT_SYMBOL(dmu_object_info);
2635 EXPORT_SYMBOL(dmu_object_info_from_dnode);
2636 EXPORT_SYMBOL(dmu_object_info_from_db);
2637 EXPORT_SYMBOL(dmu_object_size_from_db);
2638 EXPORT_SYMBOL(dmu_object_dnsize_from_db);
2639 EXPORT_SYMBOL(dmu_object_set_nlevels);
2640 EXPORT_SYMBOL(dmu_object_set_blocksize);
2641 EXPORT_SYMBOL(dmu_object_set_maxblkid);
2642 EXPORT_SYMBOL(dmu_object_set_checksum);
2643 EXPORT_SYMBOL(dmu_object_set_compress);
2644 EXPORT_SYMBOL(dmu_write_policy);
2645 EXPORT_SYMBOL(dmu_sync);
2646 EXPORT_SYMBOL(dmu_request_arcbuf);
2647 EXPORT_SYMBOL(dmu_return_arcbuf);
2648 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dnode);
2649 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dbuf);
2650 EXPORT_SYMBOL(dmu_buf_hold);
2651 EXPORT_SYMBOL(dmu_ot);
2654 module_param(zfs_nopwrite_enabled, int, 0644);
2655 MODULE_PARM_DESC(zfs_nopwrite_enabled, "Enable NOP writes");
2657 module_param(zfs_per_txg_dirty_frees_percent, ulong, 0644);
2658 MODULE_PARM_DESC(zfs_per_txg_dirty_frees_percent,
2659 "percentage of dirtied blocks from frees in one TXG");
2661 module_param(zfs_dmu_offset_next_sync, int, 0644);
2662 MODULE_PARM_DESC(zfs_dmu_offset_next_sync,
2663 "Enable forcing txg sync to find holes");