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.
25 /* Copyright (c) 2013 by Saso Kiselkov. All rights reserved. */
26 /* Copyright (c) 2013, Joyent, Inc. All rights reserved. */
27 /* Copyright 2016 Nexenta Systems, Inc. All rights reserved. */
30 #include <sys/dmu_impl.h>
31 #include <sys/dmu_tx.h>
33 #include <sys/dnode.h>
34 #include <sys/zfs_context.h>
35 #include <sys/dmu_objset.h>
36 #include <sys/dmu_traverse.h>
37 #include <sys/dsl_dataset.h>
38 #include <sys/dsl_dir.h>
39 #include <sys/dsl_pool.h>
40 #include <sys/dsl_synctask.h>
41 #include <sys/dsl_prop.h>
42 #include <sys/dmu_zfetch.h>
43 #include <sys/zfs_ioctl.h>
45 #include <sys/zio_checksum.h>
46 #include <sys/zio_compress.h>
48 #include <sys/zfeature.h>
51 #include <sys/racct.h>
53 #include <sys/zfs_znode.h>
57 * Enable/disable nopwrite feature.
59 int zfs_nopwrite_enabled = 1;
60 SYSCTL_DECL(_vfs_zfs);
61 SYSCTL_INT(_vfs_zfs, OID_AUTO, nopwrite_enabled, CTLFLAG_RDTUN,
62 &zfs_nopwrite_enabled, 0, "Enable nopwrite feature");
65 * Tunable to control percentage of dirtied blocks from frees in one TXG.
66 * After this threshold is crossed, additional dirty blocks from frees
67 * wait until the next TXG.
68 * A value of zero will disable this throttle.
70 uint32_t zfs_per_txg_dirty_frees_percent = 30;
71 SYSCTL_INT(_vfs_zfs, OID_AUTO, per_txg_dirty_frees_percent, CTLFLAG_RWTUN,
72 &zfs_per_txg_dirty_frees_percent, 0, "Percentage of dirtied blocks from frees in one txg");
74 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
75 { DMU_BSWAP_UINT8, TRUE, "unallocated" },
76 { DMU_BSWAP_ZAP, TRUE, "object directory" },
77 { DMU_BSWAP_UINT64, TRUE, "object array" },
78 { DMU_BSWAP_UINT8, TRUE, "packed nvlist" },
79 { DMU_BSWAP_UINT64, TRUE, "packed nvlist size" },
80 { DMU_BSWAP_UINT64, TRUE, "bpobj" },
81 { DMU_BSWAP_UINT64, TRUE, "bpobj header" },
82 { DMU_BSWAP_UINT64, TRUE, "SPA space map header" },
83 { DMU_BSWAP_UINT64, TRUE, "SPA space map" },
84 { DMU_BSWAP_UINT64, TRUE, "ZIL intent log" },
85 { DMU_BSWAP_DNODE, TRUE, "DMU dnode" },
86 { DMU_BSWAP_OBJSET, TRUE, "DMU objset" },
87 { DMU_BSWAP_UINT64, TRUE, "DSL directory" },
88 { DMU_BSWAP_ZAP, TRUE, "DSL directory child map"},
89 { DMU_BSWAP_ZAP, TRUE, "DSL dataset snap map" },
90 { DMU_BSWAP_ZAP, TRUE, "DSL props" },
91 { DMU_BSWAP_UINT64, TRUE, "DSL dataset" },
92 { DMU_BSWAP_ZNODE, TRUE, "ZFS znode" },
93 { DMU_BSWAP_OLDACL, TRUE, "ZFS V0 ACL" },
94 { DMU_BSWAP_UINT8, FALSE, "ZFS plain file" },
95 { DMU_BSWAP_ZAP, TRUE, "ZFS directory" },
96 { DMU_BSWAP_ZAP, TRUE, "ZFS master node" },
97 { DMU_BSWAP_ZAP, TRUE, "ZFS delete queue" },
98 { DMU_BSWAP_UINT8, FALSE, "zvol object" },
99 { DMU_BSWAP_ZAP, TRUE, "zvol prop" },
100 { DMU_BSWAP_UINT8, FALSE, "other uint8[]" },
101 { DMU_BSWAP_UINT64, FALSE, "other uint64[]" },
102 { DMU_BSWAP_ZAP, TRUE, "other ZAP" },
103 { DMU_BSWAP_ZAP, TRUE, "persistent error log" },
104 { DMU_BSWAP_UINT8, TRUE, "SPA history" },
105 { DMU_BSWAP_UINT64, TRUE, "SPA history offsets" },
106 { DMU_BSWAP_ZAP, TRUE, "Pool properties" },
107 { DMU_BSWAP_ZAP, TRUE, "DSL permissions" },
108 { DMU_BSWAP_ACL, TRUE, "ZFS ACL" },
109 { DMU_BSWAP_UINT8, TRUE, "ZFS SYSACL" },
110 { DMU_BSWAP_UINT8, TRUE, "FUID table" },
111 { DMU_BSWAP_UINT64, TRUE, "FUID table size" },
112 { DMU_BSWAP_ZAP, TRUE, "DSL dataset next clones"},
113 { DMU_BSWAP_ZAP, TRUE, "scan work queue" },
114 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group used" },
115 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group quota" },
116 { DMU_BSWAP_ZAP, TRUE, "snapshot refcount tags"},
117 { DMU_BSWAP_ZAP, TRUE, "DDT ZAP algorithm" },
118 { DMU_BSWAP_ZAP, TRUE, "DDT statistics" },
119 { DMU_BSWAP_UINT8, TRUE, "System attributes" },
120 { DMU_BSWAP_ZAP, TRUE, "SA master node" },
121 { DMU_BSWAP_ZAP, TRUE, "SA attr registration" },
122 { DMU_BSWAP_ZAP, TRUE, "SA attr layouts" },
123 { DMU_BSWAP_ZAP, TRUE, "scan translations" },
124 { DMU_BSWAP_UINT8, FALSE, "deduplicated block" },
125 { DMU_BSWAP_ZAP, TRUE, "DSL deadlist map" },
126 { DMU_BSWAP_UINT64, TRUE, "DSL deadlist map hdr" },
127 { DMU_BSWAP_ZAP, TRUE, "DSL dir clones" },
128 { DMU_BSWAP_UINT64, TRUE, "bpobj subobj" }
131 const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
132 { byteswap_uint8_array, "uint8" },
133 { byteswap_uint16_array, "uint16" },
134 { byteswap_uint32_array, "uint32" },
135 { byteswap_uint64_array, "uint64" },
136 { zap_byteswap, "zap" },
137 { dnode_buf_byteswap, "dnode" },
138 { dmu_objset_byteswap, "objset" },
139 { zfs_znode_byteswap, "znode" },
140 { zfs_oldacl_byteswap, "oldacl" },
141 { zfs_acl_byteswap, "acl" }
145 dmu_buf_hold_noread_by_dnode(dnode_t *dn, uint64_t offset,
146 void *tag, dmu_buf_t **dbp)
151 blkid = dbuf_whichblock(dn, 0, offset);
152 rw_enter(&dn->dn_struct_rwlock, RW_READER);
153 db = dbuf_hold(dn, blkid, tag);
154 rw_exit(&dn->dn_struct_rwlock);
158 return (SET_ERROR(EIO));
165 dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset,
166 void *tag, dmu_buf_t **dbp)
173 err = dnode_hold(os, object, FTAG, &dn);
176 blkid = dbuf_whichblock(dn, 0, offset);
177 rw_enter(&dn->dn_struct_rwlock, RW_READER);
178 db = dbuf_hold(dn, blkid, tag);
179 rw_exit(&dn->dn_struct_rwlock);
180 dnode_rele(dn, FTAG);
184 return (SET_ERROR(EIO));
192 dmu_buf_hold_by_dnode(dnode_t *dn, uint64_t offset,
193 void *tag, dmu_buf_t **dbp, int flags)
196 int db_flags = DB_RF_CANFAIL;
198 if (flags & DMU_READ_NO_PREFETCH)
199 db_flags |= DB_RF_NOPREFETCH;
201 err = dmu_buf_hold_noread_by_dnode(dn, offset, tag, dbp);
203 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
204 err = dbuf_read(db, NULL, db_flags);
215 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
216 void *tag, dmu_buf_t **dbp, int flags)
219 int db_flags = DB_RF_CANFAIL;
221 if (flags & DMU_READ_NO_PREFETCH)
222 db_flags |= DB_RF_NOPREFETCH;
224 err = dmu_buf_hold_noread(os, object, offset, tag, dbp);
226 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
227 err = dbuf_read(db, NULL, db_flags);
240 return (DN_MAX_BONUSLEN);
244 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx)
246 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
253 if (dn->dn_bonus != db) {
254 error = SET_ERROR(EINVAL);
255 } else if (newsize < 0 || newsize > db_fake->db_size) {
256 error = SET_ERROR(EINVAL);
258 dnode_setbonuslen(dn, newsize, tx);
267 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
269 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
276 if (!DMU_OT_IS_VALID(type)) {
277 error = SET_ERROR(EINVAL);
278 } else if (dn->dn_bonus != db) {
279 error = SET_ERROR(EINVAL);
281 dnode_setbonus_type(dn, type, tx);
290 dmu_get_bonustype(dmu_buf_t *db_fake)
292 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
294 dmu_object_type_t type;
298 type = dn->dn_bonustype;
305 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
310 error = dnode_hold(os, object, FTAG, &dn);
311 dbuf_rm_spill(dn, tx);
312 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
313 dnode_rm_spill(dn, tx);
314 rw_exit(&dn->dn_struct_rwlock);
315 dnode_rele(dn, FTAG);
320 * returns ENOENT, EIO, or 0.
323 dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp)
329 error = dnode_hold(os, object, FTAG, &dn);
333 rw_enter(&dn->dn_struct_rwlock, RW_READER);
334 if (dn->dn_bonus == NULL) {
335 rw_exit(&dn->dn_struct_rwlock);
336 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
337 if (dn->dn_bonus == NULL)
338 dbuf_create_bonus(dn);
342 /* as long as the bonus buf is held, the dnode will be held */
343 if (refcount_add(&db->db_holds, tag) == 1) {
344 VERIFY(dnode_add_ref(dn, db));
345 atomic_inc_32(&dn->dn_dbufs_count);
349 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
350 * hold and incrementing the dbuf count to ensure that dnode_move() sees
351 * a dnode hold for every dbuf.
353 rw_exit(&dn->dn_struct_rwlock);
355 dnode_rele(dn, FTAG);
357 VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH));
364 * returns ENOENT, EIO, or 0.
366 * This interface will allocate a blank spill dbuf when a spill blk
367 * doesn't already exist on the dnode.
369 * if you only want to find an already existing spill db, then
370 * dmu_spill_hold_existing() should be used.
373 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp)
375 dmu_buf_impl_t *db = NULL;
378 if ((flags & DB_RF_HAVESTRUCT) == 0)
379 rw_enter(&dn->dn_struct_rwlock, RW_READER);
381 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
383 if ((flags & DB_RF_HAVESTRUCT) == 0)
384 rw_exit(&dn->dn_struct_rwlock);
387 err = dbuf_read(db, NULL, flags);
396 dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
398 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
405 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
406 err = SET_ERROR(EINVAL);
408 rw_enter(&dn->dn_struct_rwlock, RW_READER);
410 if (!dn->dn_have_spill) {
411 err = SET_ERROR(ENOENT);
413 err = dmu_spill_hold_by_dnode(dn,
414 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
417 rw_exit(&dn->dn_struct_rwlock);
425 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
427 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
433 err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp);
440 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
441 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
442 * and can induce severe lock contention when writing to several files
443 * whose dnodes are in the same block.
446 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
447 boolean_t read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
450 uint64_t blkid, nblks, i;
455 ASSERT(length <= DMU_MAX_ACCESS);
458 * Note: We directly notify the prefetch code of this read, so that
459 * we can tell it about the multi-block read. dbuf_read() only knows
460 * about the one block it is accessing.
462 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT |
465 rw_enter(&dn->dn_struct_rwlock, RW_READER);
466 if (dn->dn_datablkshift) {
467 int blkshift = dn->dn_datablkshift;
468 nblks = (P2ROUNDUP(offset + length, 1ULL << blkshift) -
469 P2ALIGN(offset, 1ULL << blkshift)) >> blkshift;
471 if (offset + length > dn->dn_datablksz) {
472 zfs_panic_recover("zfs: accessing past end of object "
473 "%llx/%llx (size=%u access=%llu+%llu)",
474 (longlong_t)dn->dn_objset->
475 os_dsl_dataset->ds_object,
476 (longlong_t)dn->dn_object, dn->dn_datablksz,
477 (longlong_t)offset, (longlong_t)length);
478 rw_exit(&dn->dn_struct_rwlock);
479 return (SET_ERROR(EIO));
483 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
485 #if defined(_KERNEL) && defined(RACCT)
486 if (racct_enable && !read) {
488 racct_add_force(curproc, RACCT_WRITEBPS, length);
489 racct_add_force(curproc, RACCT_WRITEIOPS, nblks);
490 PROC_UNLOCK(curproc);
494 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
495 blkid = dbuf_whichblock(dn, 0, offset);
496 for (i = 0; i < nblks; i++) {
497 dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag);
499 rw_exit(&dn->dn_struct_rwlock);
500 dmu_buf_rele_array(dbp, nblks, tag);
502 return (SET_ERROR(EIO));
505 /* initiate async i/o */
507 (void) dbuf_read(db, zio, dbuf_flags);
510 curthread->td_ru.ru_oublock++;
515 if ((flags & DMU_READ_NO_PREFETCH) == 0 &&
516 DNODE_META_IS_CACHEABLE(dn) && length <= zfetch_array_rd_sz) {
517 dmu_zfetch(&dn->dn_zfetch, blkid, nblks,
518 read && DNODE_IS_CACHEABLE(dn));
520 rw_exit(&dn->dn_struct_rwlock);
522 /* wait for async i/o */
525 dmu_buf_rele_array(dbp, nblks, tag);
529 /* wait for other io to complete */
531 for (i = 0; i < nblks; i++) {
532 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
533 mutex_enter(&db->db_mtx);
534 while (db->db_state == DB_READ ||
535 db->db_state == DB_FILL)
536 cv_wait(&db->db_changed, &db->db_mtx);
537 if (db->db_state == DB_UNCACHED)
538 err = SET_ERROR(EIO);
539 mutex_exit(&db->db_mtx);
541 dmu_buf_rele_array(dbp, nblks, tag);
553 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
554 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
559 err = dnode_hold(os, object, FTAG, &dn);
563 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
564 numbufsp, dbpp, DMU_READ_PREFETCH);
566 dnode_rele(dn, FTAG);
572 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
573 uint64_t length, boolean_t read, void *tag, int *numbufsp,
576 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
582 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
583 numbufsp, dbpp, DMU_READ_PREFETCH);
590 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
593 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
598 for (i = 0; i < numbufs; i++) {
600 dbuf_rele(dbp[i], tag);
603 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
607 * Issue prefetch i/os for the given blocks. If level is greater than 0, the
608 * indirect blocks prefeteched will be those that point to the blocks containing
609 * the data starting at offset, and continuing to offset + len.
611 * Note that if the indirect blocks above the blocks being prefetched are not in
612 * cache, they will be asychronously read in.
615 dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset,
616 uint64_t len, zio_priority_t pri)
622 if (len == 0) { /* they're interested in the bonus buffer */
623 dn = DMU_META_DNODE(os);
625 if (object == 0 || object >= DN_MAX_OBJECT)
628 rw_enter(&dn->dn_struct_rwlock, RW_READER);
629 blkid = dbuf_whichblock(dn, level,
630 object * sizeof (dnode_phys_t));
631 dbuf_prefetch(dn, level, blkid, pri, 0);
632 rw_exit(&dn->dn_struct_rwlock);
637 * XXX - Note, if the dnode for the requested object is not
638 * already cached, we will do a *synchronous* read in the
639 * dnode_hold() call. The same is true for any indirects.
641 err = dnode_hold(os, object, FTAG, &dn);
645 rw_enter(&dn->dn_struct_rwlock, RW_READER);
647 * offset + len - 1 is the last byte we want to prefetch for, and offset
648 * is the first. Then dbuf_whichblk(dn, level, off + len - 1) is the
649 * last block we want to prefetch, and dbuf_whichblock(dn, level,
650 * offset) is the first. Then the number we need to prefetch is the
653 if (level > 0 || dn->dn_datablkshift != 0) {
654 nblks = dbuf_whichblock(dn, level, offset + len - 1) -
655 dbuf_whichblock(dn, level, offset) + 1;
657 nblks = (offset < dn->dn_datablksz);
661 blkid = dbuf_whichblock(dn, level, offset);
662 for (int i = 0; i < nblks; i++)
663 dbuf_prefetch(dn, level, blkid + i, pri, 0);
666 rw_exit(&dn->dn_struct_rwlock);
668 dnode_rele(dn, FTAG);
672 * Get the next "chunk" of file data to free. We traverse the file from
673 * the end so that the file gets shorter over time (if we crashes in the
674 * middle, this will leave us in a better state). We find allocated file
675 * data by simply searching the allocated level 1 indirects.
677 * On input, *start should be the first offset that does not need to be
678 * freed (e.g. "offset + length"). On return, *start will be the first
679 * offset that should be freed.
682 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum)
684 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
685 /* bytes of data covered by a level-1 indirect block */
687 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
689 ASSERT3U(minimum, <=, *start);
691 if (*start - minimum <= iblkrange * maxblks) {
695 ASSERT(ISP2(iblkrange));
697 for (uint64_t blks = 0; *start > minimum && blks < maxblks; blks++) {
701 * dnode_next_offset(BACKWARDS) will find an allocated L1
702 * indirect block at or before the input offset. We must
703 * decrement *start so that it is at the end of the region
707 err = dnode_next_offset(dn,
708 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
710 /* if there are no indirect blocks before start, we are done */
714 } else if (err != 0) {
718 /* set start to the beginning of this L1 indirect */
719 *start = P2ALIGN(*start, iblkrange);
721 if (*start < minimum)
727 * If this objset is of type OST_ZFS return true if vfs's unmounted flag is set,
728 * otherwise return false.
729 * Used below in dmu_free_long_range_impl() to enable abort when unmounting
733 dmu_objset_zfs_unmounting(objset_t *os)
736 if (dmu_objset_type(os) == DMU_OST_ZFS)
737 return (zfs_get_vfs_flag_unmounted(os));
743 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
746 uint64_t object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
748 uint64_t dirty_frees_threshold;
749 dsl_pool_t *dp = dmu_objset_pool(os);
751 if (offset >= object_size)
754 if (zfs_per_txg_dirty_frees_percent <= 100)
755 dirty_frees_threshold =
756 zfs_per_txg_dirty_frees_percent * zfs_dirty_data_max / 100;
758 dirty_frees_threshold = zfs_dirty_data_max / 4;
760 if (length == DMU_OBJECT_END || offset + length > object_size)
761 length = object_size - offset;
763 while (length != 0) {
764 uint64_t chunk_end, chunk_begin, chunk_len;
765 uint64_t long_free_dirty_all_txgs = 0;
768 if (dmu_objset_zfs_unmounting(dn->dn_objset))
769 return (SET_ERROR(EINTR));
771 chunk_end = chunk_begin = offset + length;
773 /* move chunk_begin backwards to the beginning of this chunk */
774 err = get_next_chunk(dn, &chunk_begin, offset);
777 ASSERT3U(chunk_begin, >=, offset);
778 ASSERT3U(chunk_begin, <=, chunk_end);
780 chunk_len = chunk_end - chunk_begin;
782 mutex_enter(&dp->dp_lock);
783 for (int t = 0; t < TXG_SIZE; t++) {
784 long_free_dirty_all_txgs +=
785 dp->dp_long_free_dirty_pertxg[t];
787 mutex_exit(&dp->dp_lock);
790 * To avoid filling up a TXG with just frees wait for
791 * the next TXG to open before freeing more chunks if
792 * we have reached the threshold of frees
794 if (dirty_frees_threshold != 0 &&
795 long_free_dirty_all_txgs >= dirty_frees_threshold) {
796 txg_wait_open(dp, 0);
800 tx = dmu_tx_create(os);
801 dmu_tx_hold_free(tx, dn->dn_object, chunk_begin, chunk_len);
804 * Mark this transaction as typically resulting in a net
805 * reduction in space used.
807 dmu_tx_mark_netfree(tx);
808 err = dmu_tx_assign(tx, TXG_WAIT);
814 mutex_enter(&dp->dp_lock);
815 dp->dp_long_free_dirty_pertxg[dmu_tx_get_txg(tx) & TXG_MASK] +=
817 mutex_exit(&dp->dp_lock);
818 DTRACE_PROBE3(free__long__range,
819 uint64_t, long_free_dirty_all_txgs, uint64_t, chunk_len,
820 uint64_t, dmu_tx_get_txg(tx));
821 dnode_free_range(dn, chunk_begin, chunk_len, tx);
830 dmu_free_long_range(objset_t *os, uint64_t object,
831 uint64_t offset, uint64_t length)
836 err = dnode_hold(os, object, FTAG, &dn);
839 err = dmu_free_long_range_impl(os, dn, offset, length);
842 * It is important to zero out the maxblkid when freeing the entire
843 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
844 * will take the fast path, and (b) dnode_reallocate() can verify
845 * that the entire file has been freed.
847 if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
850 dnode_rele(dn, FTAG);
855 dmu_free_long_object(objset_t *os, uint64_t object)
860 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
864 tx = dmu_tx_create(os);
865 dmu_tx_hold_bonus(tx, object);
866 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
867 dmu_tx_mark_netfree(tx);
868 err = dmu_tx_assign(tx, TXG_WAIT);
870 err = dmu_object_free(os, object, tx);
880 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
881 uint64_t size, dmu_tx_t *tx)
884 int err = dnode_hold(os, object, FTAG, &dn);
887 ASSERT(offset < UINT64_MAX);
888 ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
889 dnode_free_range(dn, offset, size, tx);
890 dnode_rele(dn, FTAG);
895 dmu_read_impl(dnode_t *dn, uint64_t offset, uint64_t size,
896 void *buf, uint32_t flags)
899 int numbufs, err = 0;
902 * Deal with odd block sizes, where there can't be data past the first
903 * block. If we ever do the tail block optimization, we will need to
904 * handle that here as well.
906 if (dn->dn_maxblkid == 0) {
907 int newsz = offset > dn->dn_datablksz ? 0 :
908 MIN(size, dn->dn_datablksz - offset);
909 bzero((char *)buf + newsz, size - newsz);
914 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
918 * NB: we could do this block-at-a-time, but it's nice
919 * to be reading in parallel.
921 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
922 TRUE, FTAG, &numbufs, &dbp, flags);
926 for (i = 0; i < numbufs; i++) {
929 dmu_buf_t *db = dbp[i];
933 bufoff = offset - db->db_offset;
934 tocpy = (int)MIN(db->db_size - bufoff, size);
936 bcopy((char *)db->db_data + bufoff, buf, tocpy);
940 buf = (char *)buf + tocpy;
942 dmu_buf_rele_array(dbp, numbufs, FTAG);
948 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
949 void *buf, uint32_t flags)
954 err = dnode_hold(os, object, FTAG, &dn);
958 err = dmu_read_impl(dn, offset, size, buf, flags);
959 dnode_rele(dn, FTAG);
964 dmu_read_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, void *buf,
967 return (dmu_read_impl(dn, offset, size, buf, flags));
971 dmu_write_impl(dmu_buf_t **dbp, int numbufs, uint64_t offset, uint64_t size,
972 const void *buf, dmu_tx_t *tx)
976 for (i = 0; i < numbufs; i++) {
979 dmu_buf_t *db = dbp[i];
983 bufoff = offset - db->db_offset;
984 tocpy = (int)MIN(db->db_size - bufoff, size);
986 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
988 if (tocpy == db->db_size)
989 dmu_buf_will_fill(db, tx);
991 dmu_buf_will_dirty(db, tx);
993 bcopy(buf, (char *)db->db_data + bufoff, tocpy);
995 if (tocpy == db->db_size)
996 dmu_buf_fill_done(db, tx);
1000 buf = (char *)buf + tocpy;
1005 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1006 const void *buf, dmu_tx_t *tx)
1014 VERIFY0(dmu_buf_hold_array(os, object, offset, size,
1015 FALSE, FTAG, &numbufs, &dbp));
1016 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1017 dmu_buf_rele_array(dbp, numbufs, FTAG);
1021 dmu_write_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size,
1022 const void *buf, dmu_tx_t *tx)
1030 VERIFY0(dmu_buf_hold_array_by_dnode(dn, offset, size,
1031 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH));
1032 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1033 dmu_buf_rele_array(dbp, numbufs, FTAG);
1037 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1046 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
1047 FALSE, FTAG, &numbufs, &dbp));
1049 for (i = 0; i < numbufs; i++) {
1050 dmu_buf_t *db = dbp[i];
1052 dmu_buf_will_not_fill(db, tx);
1054 dmu_buf_rele_array(dbp, numbufs, FTAG);
1058 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset,
1059 void *data, uint8_t etype, uint8_t comp, int uncompressed_size,
1060 int compressed_size, int byteorder, dmu_tx_t *tx)
1064 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES);
1065 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS);
1066 VERIFY0(dmu_buf_hold_noread(os, object, offset,
1069 dmu_buf_write_embedded(db,
1070 data, (bp_embedded_type_t)etype, (enum zio_compress)comp,
1071 uncompressed_size, compressed_size, byteorder, tx);
1073 dmu_buf_rele(db, FTAG);
1077 * DMU support for xuio
1079 kstat_t *xuio_ksp = NULL;
1082 dmu_xuio_init(xuio_t *xuio, int nblk)
1085 uio_t *uio = &xuio->xu_uio;
1087 uio->uio_iovcnt = nblk;
1088 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
1090 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
1092 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
1093 priv->iovp = uio->uio_iov;
1094 XUIO_XUZC_PRIV(xuio) = priv;
1096 if (XUIO_XUZC_RW(xuio) == UIO_READ)
1097 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
1099 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
1105 dmu_xuio_fini(xuio_t *xuio)
1107 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1108 int nblk = priv->cnt;
1110 kmem_free(priv->iovp, nblk * sizeof (iovec_t));
1111 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
1112 kmem_free(priv, sizeof (dmu_xuio_t));
1114 if (XUIO_XUZC_RW(xuio) == UIO_READ)
1115 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
1117 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
1121 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
1122 * and increase priv->next by 1.
1125 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
1128 uio_t *uio = &xuio->xu_uio;
1129 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1130 int i = priv->next++;
1132 ASSERT(i < priv->cnt);
1133 ASSERT(off + n <= arc_buf_lsize(abuf));
1134 iov = uio->uio_iov + i;
1135 iov->iov_base = (char *)abuf->b_data + off;
1137 priv->bufs[i] = abuf;
1142 dmu_xuio_cnt(xuio_t *xuio)
1144 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1149 dmu_xuio_arcbuf(xuio_t *xuio, int i)
1151 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1153 ASSERT(i < priv->cnt);
1154 return (priv->bufs[i]);
1158 dmu_xuio_clear(xuio_t *xuio, int i)
1160 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1162 ASSERT(i < priv->cnt);
1163 priv->bufs[i] = NULL;
1167 xuio_stat_init(void)
1169 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
1170 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
1171 KSTAT_FLAG_VIRTUAL);
1172 if (xuio_ksp != NULL) {
1173 xuio_ksp->ks_data = &xuio_stats;
1174 kstat_install(xuio_ksp);
1179 xuio_stat_fini(void)
1181 if (xuio_ksp != NULL) {
1182 kstat_delete(xuio_ksp);
1188 xuio_stat_wbuf_copied(void)
1190 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1194 xuio_stat_wbuf_nocopy(void)
1196 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
1201 dmu_read_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size)
1204 int numbufs, i, err;
1205 xuio_t *xuio = NULL;
1208 * NB: we could do this block-at-a-time, but it's nice
1209 * to be reading in parallel.
1211 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1212 TRUE, FTAG, &numbufs, &dbp, 0);
1217 if (uio->uio_extflg == UIO_XUIO)
1218 xuio = (xuio_t *)uio;
1221 for (i = 0; i < numbufs; i++) {
1224 dmu_buf_t *db = dbp[i];
1228 bufoff = uio->uio_loffset - db->db_offset;
1229 tocpy = (int)MIN(db->db_size - bufoff, size);
1232 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
1233 arc_buf_t *dbuf_abuf = dbi->db_buf;
1234 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
1235 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
1237 uio->uio_resid -= tocpy;
1238 uio->uio_loffset += tocpy;
1241 if (abuf == dbuf_abuf)
1242 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
1244 XUIOSTAT_BUMP(xuiostat_rbuf_copied);
1247 err = uiomove((char *)db->db_data + bufoff, tocpy,
1250 err = vn_io_fault_uiomove((char *)db->db_data + bufoff,
1259 dmu_buf_rele_array(dbp, numbufs, FTAG);
1265 * Read 'size' bytes into the uio buffer.
1266 * From object zdb->db_object.
1267 * Starting at offset uio->uio_loffset.
1269 * If the caller already has a dbuf in the target object
1270 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1271 * because we don't have to find the dnode_t for the object.
1274 dmu_read_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size)
1276 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1285 err = dmu_read_uio_dnode(dn, uio, size);
1292 * Read 'size' bytes into the uio buffer.
1293 * From the specified object
1294 * Starting at offset uio->uio_loffset.
1297 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
1305 err = dnode_hold(os, object, FTAG, &dn);
1309 err = dmu_read_uio_dnode(dn, uio, size);
1311 dnode_rele(dn, FTAG);
1317 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
1324 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1325 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1329 for (i = 0; i < numbufs; i++) {
1332 dmu_buf_t *db = dbp[i];
1336 bufoff = uio->uio_loffset - db->db_offset;
1337 tocpy = (int)MIN(db->db_size - bufoff, size);
1339 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1341 if (tocpy == db->db_size)
1342 dmu_buf_will_fill(db, tx);
1344 dmu_buf_will_dirty(db, tx);
1348 * XXX uiomove could block forever (eg. nfs-backed
1349 * pages). There needs to be a uiolockdown() function
1350 * to lock the pages in memory, so that uiomove won't
1353 err = uiomove((char *)db->db_data + bufoff, tocpy,
1356 err = vn_io_fault_uiomove((char *)db->db_data + bufoff, tocpy,
1360 if (tocpy == db->db_size)
1361 dmu_buf_fill_done(db, tx);
1369 dmu_buf_rele_array(dbp, numbufs, FTAG);
1374 * Write 'size' bytes from the uio buffer.
1375 * To object zdb->db_object.
1376 * Starting at offset uio->uio_loffset.
1378 * If the caller already has a dbuf in the target object
1379 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1380 * because we don't have to find the dnode_t for the object.
1383 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
1386 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1395 err = dmu_write_uio_dnode(dn, uio, size, tx);
1402 * Write 'size' bytes from the uio buffer.
1403 * To the specified object.
1404 * Starting at offset uio->uio_loffset.
1407 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1416 err = dnode_hold(os, object, FTAG, &dn);
1420 err = dmu_write_uio_dnode(dn, uio, size, tx);
1422 dnode_rele(dn, FTAG);
1429 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1430 page_t *pp, dmu_tx_t *tx)
1439 err = dmu_buf_hold_array(os, object, offset, size,
1440 FALSE, FTAG, &numbufs, &dbp);
1444 for (i = 0; i < numbufs; i++) {
1445 int tocpy, copied, thiscpy;
1447 dmu_buf_t *db = dbp[i];
1451 ASSERT3U(db->db_size, >=, PAGESIZE);
1453 bufoff = offset - db->db_offset;
1454 tocpy = (int)MIN(db->db_size - bufoff, size);
1456 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1458 if (tocpy == db->db_size)
1459 dmu_buf_will_fill(db, tx);
1461 dmu_buf_will_dirty(db, tx);
1463 for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1464 ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff);
1465 thiscpy = MIN(PAGESIZE, tocpy - copied);
1466 va = zfs_map_page(pp, S_READ);
1467 bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1468 zfs_unmap_page(pp, va);
1473 if (tocpy == db->db_size)
1474 dmu_buf_fill_done(db, tx);
1479 dmu_buf_rele_array(dbp, numbufs, FTAG);
1483 #else /* !illumos */
1486 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1487 vm_page_t *ma, dmu_tx_t *tx)
1497 err = dmu_buf_hold_array(os, object, offset, size,
1498 FALSE, FTAG, &numbufs, &dbp);
1502 for (i = 0; i < numbufs; i++) {
1503 int tocpy, copied, thiscpy;
1505 dmu_buf_t *db = dbp[i];
1509 ASSERT3U(db->db_size, >=, PAGESIZE);
1511 bufoff = offset - db->db_offset;
1512 tocpy = (int)MIN(db->db_size - bufoff, size);
1514 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1516 if (tocpy == db->db_size)
1517 dmu_buf_will_fill(db, tx);
1519 dmu_buf_will_dirty(db, tx);
1521 for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1522 ASSERT3U(ptoa((*ma)->pindex), ==, db->db_offset + bufoff);
1523 thiscpy = MIN(PAGESIZE, tocpy - copied);
1524 va = zfs_map_page(*ma, &sf);
1525 bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1531 if (tocpy == db->db_size)
1532 dmu_buf_fill_done(db, tx);
1537 dmu_buf_rele_array(dbp, numbufs, FTAG);
1542 dmu_read_pages(objset_t *os, uint64_t object, vm_page_t *ma, int count,
1543 int *rbehind, int *rahead, int last_size)
1552 int bufoff, pgoff, tocpy;
1556 ASSERT3U(ma[0]->pindex + count - 1, ==, ma[count - 1]->pindex);
1557 ASSERT(last_size <= PAGE_SIZE);
1559 err = dmu_buf_hold_array(os, object, IDX_TO_OFF(ma[0]->pindex),
1560 IDX_TO_OFF(count - 1) + last_size, TRUE, FTAG, &numbufs, &dbp);
1565 IMPLY(last_size < PAGE_SIZE, *rahead == 0);
1566 if (dbp[0]->db_offset != 0 || numbufs > 1) {
1567 for (i = 0; i < numbufs; i++) {
1568 ASSERT(ISP2(dbp[i]->db_size));
1569 ASSERT((dbp[i]->db_offset % dbp[i]->db_size) == 0);
1570 ASSERT3U(dbp[i]->db_size, ==, dbp[0]->db_size);
1575 vmobj = ma[0]->object;
1576 zfs_vmobject_wlock(vmobj);
1579 for (i = 0; i < *rbehind; i++) {
1580 m = vm_page_grab(vmobj, ma[0]->pindex - 1 - i,
1581 VM_ALLOC_NORMAL | VM_ALLOC_NOWAIT | VM_ALLOC_NOBUSY);
1584 if (m->valid != 0) {
1585 ASSERT3U(m->valid, ==, VM_PAGE_BITS_ALL);
1588 ASSERT(m->dirty == 0);
1589 ASSERT(!pmap_page_is_mapped(m));
1591 ASSERT(db->db_size > PAGE_SIZE);
1592 bufoff = IDX_TO_OFF(m->pindex) % db->db_size;
1593 va = zfs_map_page(m, &sf);
1594 bcopy((char *)db->db_data + bufoff, va, PAGESIZE);
1596 m->valid = VM_PAGE_BITS_ALL;
1598 if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1599 vm_page_activate(m);
1601 vm_page_deactivate(m);
1606 bufoff = IDX_TO_OFF(ma[0]->pindex) % db->db_size;
1608 for (mi = 0, di = 0; mi < count && di < numbufs; ) {
1611 vm_page_assert_xbusied(m);
1612 ASSERT(m->valid == 0);
1613 ASSERT(m->dirty == 0);
1614 ASSERT(!pmap_page_is_mapped(m));
1615 va = zfs_map_page(m, &sf);
1620 ASSERT3U(IDX_TO_OFF(m->pindex) + pgoff, ==,
1621 db->db_offset + bufoff);
1624 * We do not need to clamp the copy size by the file
1625 * size as the last block is zero-filled beyond the
1626 * end of file anyway.
1628 tocpy = MIN(db->db_size - bufoff, PAGESIZE - pgoff);
1629 bcopy((char *)db->db_data + bufoff, va + pgoff, tocpy);
1632 ASSERT(pgoff <= PAGESIZE);
1633 if (pgoff == PAGESIZE) {
1635 m->valid = VM_PAGE_BITS_ALL;
1642 ASSERT(bufoff <= db->db_size);
1643 if (bufoff == db->db_size) {
1644 ASSERT(di < numbufs);
1652 * Three possibilities:
1653 * - last requested page ends at a buffer boundary and , thus,
1654 * all pages and buffers have been iterated;
1655 * - all requested pages are filled, but the last buffer
1656 * has not been exhausted;
1657 * the read-ahead is possible only in this case;
1658 * - all buffers have been read, but the last page has not been
1660 * this is only possible if the file has only a single buffer
1661 * with a size that is not a multiple of the page size.
1664 ASSERT(di >= numbufs - 1);
1665 IMPLY(*rahead != 0, di == numbufs - 1);
1666 IMPLY(*rahead != 0, bufoff != 0);
1669 if (di == numbufs) {
1670 ASSERT(mi >= count - 1);
1671 ASSERT(*rahead == 0);
1672 IMPLY(pgoff == 0, mi == count);
1674 ASSERT(mi == count - 1);
1675 ASSERT((dbp[0]->db_size & PAGE_MASK) != 0);
1680 bzero(va + pgoff, PAGESIZE - pgoff);
1682 m->valid = VM_PAGE_BITS_ALL;
1685 for (i = 0; i < *rahead; i++) {
1686 m = vm_page_grab(vmobj, ma[count - 1]->pindex + 1 + i,
1687 VM_ALLOC_NORMAL | VM_ALLOC_NOWAIT | VM_ALLOC_NOBUSY);
1690 if (m->valid != 0) {
1691 ASSERT3U(m->valid, ==, VM_PAGE_BITS_ALL);
1694 ASSERT(m->dirty == 0);
1695 ASSERT(!pmap_page_is_mapped(m));
1697 ASSERT(db->db_size > PAGE_SIZE);
1698 bufoff = IDX_TO_OFF(m->pindex) % db->db_size;
1699 tocpy = MIN(db->db_size - bufoff, PAGESIZE);
1700 va = zfs_map_page(m, &sf);
1701 bcopy((char *)db->db_data + bufoff, va, tocpy);
1702 if (tocpy < PAGESIZE) {
1703 ASSERT(i == *rahead - 1);
1704 ASSERT((db->db_size & PAGE_MASK) != 0);
1705 bzero(va + tocpy, PAGESIZE - tocpy);
1708 m->valid = VM_PAGE_BITS_ALL;
1710 if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1711 vm_page_activate(m);
1713 vm_page_deactivate(m);
1717 zfs_vmobject_wunlock(vmobj);
1719 dmu_buf_rele_array(dbp, numbufs, FTAG);
1722 #endif /* illumos */
1723 #endif /* _KERNEL */
1726 * Allocate a loaned anonymous arc buffer.
1729 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1731 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1733 return (arc_loan_buf(db->db_objset->os_spa, B_FALSE, size));
1737 * Free a loaned arc buffer.
1740 dmu_return_arcbuf(arc_buf_t *buf)
1742 arc_return_buf(buf, FTAG);
1743 arc_buf_destroy(buf, FTAG);
1747 * When possible directly assign passed loaned arc buffer to a dbuf.
1748 * If this is not possible copy the contents of passed arc buf via
1752 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1755 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1758 uint32_t blksz = (uint32_t)arc_buf_lsize(buf);
1761 DB_DNODE_ENTER(dbuf);
1762 dn = DB_DNODE(dbuf);
1763 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1764 blkid = dbuf_whichblock(dn, 0, offset);
1765 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1766 rw_exit(&dn->dn_struct_rwlock);
1767 DB_DNODE_EXIT(dbuf);
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) {
1775 curthread->td_ru.ru_oublock++;
1779 racct_add_force(curproc, RACCT_WRITEBPS, blksz);
1780 racct_add_force(curproc, RACCT_WRITEIOPS, 1);
1781 PROC_UNLOCK(curproc);
1784 #endif /* _KERNEL */
1785 dbuf_assign_arcbuf(db, buf, tx);
1786 dbuf_rele(db, FTAG);
1791 /* compressed bufs must always be assignable to their dbuf */
1792 ASSERT3U(arc_get_compression(buf), ==, ZIO_COMPRESS_OFF);
1793 ASSERT(!(buf->b_flags & ARC_BUF_FLAG_COMPRESSED));
1795 DB_DNODE_ENTER(dbuf);
1796 dn = DB_DNODE(dbuf);
1798 object = dn->dn_object;
1799 DB_DNODE_EXIT(dbuf);
1801 dbuf_rele(db, FTAG);
1802 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1803 dmu_return_arcbuf(buf);
1804 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1809 dbuf_dirty_record_t *dsa_dr;
1810 dmu_sync_cb_t *dsa_done;
1817 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1819 dmu_sync_arg_t *dsa = varg;
1820 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1821 blkptr_t *bp = zio->io_bp;
1823 if (zio->io_error == 0) {
1824 if (BP_IS_HOLE(bp)) {
1826 * A block of zeros may compress to a hole, but the
1827 * block size still needs to be known for replay.
1829 BP_SET_LSIZE(bp, db->db_size);
1830 } else if (!BP_IS_EMBEDDED(bp)) {
1831 ASSERT(BP_GET_LEVEL(bp) == 0);
1838 dmu_sync_late_arrival_ready(zio_t *zio)
1840 dmu_sync_ready(zio, NULL, zio->io_private);
1845 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1847 dmu_sync_arg_t *dsa = varg;
1848 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1849 dmu_buf_impl_t *db = dr->dr_dbuf;
1851 mutex_enter(&db->db_mtx);
1852 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1853 if (zio->io_error == 0) {
1854 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1855 if (dr->dt.dl.dr_nopwrite) {
1856 blkptr_t *bp = zio->io_bp;
1857 blkptr_t *bp_orig = &zio->io_bp_orig;
1858 uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
1860 ASSERT(BP_EQUAL(bp, bp_orig));
1861 VERIFY(BP_EQUAL(bp, db->db_blkptr));
1862 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1863 ASSERT(zio_checksum_table[chksum].ci_flags &
1864 ZCHECKSUM_FLAG_NOPWRITE);
1866 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1867 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1868 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1871 * Old style holes are filled with all zeros, whereas
1872 * new-style holes maintain their lsize, type, level,
1873 * and birth time (see zio_write_compress). While we
1874 * need to reset the BP_SET_LSIZE() call that happened
1875 * in dmu_sync_ready for old style holes, we do *not*
1876 * want to wipe out the information contained in new
1877 * style holes. Thus, only zero out the block pointer if
1878 * it's an old style hole.
1880 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) &&
1881 dr->dt.dl.dr_overridden_by.blk_birth == 0)
1882 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1884 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1886 cv_broadcast(&db->db_changed);
1887 mutex_exit(&db->db_mtx);
1889 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1891 kmem_free(dsa, sizeof (*dsa));
1895 dmu_sync_late_arrival_done(zio_t *zio)
1897 blkptr_t *bp = zio->io_bp;
1898 dmu_sync_arg_t *dsa = zio->io_private;
1899 blkptr_t *bp_orig = &zio->io_bp_orig;
1901 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1902 ASSERT(!(zio->io_flags & ZIO_FLAG_NOPWRITE));
1903 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1904 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1905 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1906 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1909 dmu_tx_commit(dsa->dsa_tx);
1911 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1913 abd_put(zio->io_abd);
1914 kmem_free(dsa, sizeof (*dsa));
1918 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1919 zio_prop_t *zp, zbookmark_phys_t *zb)
1921 dmu_sync_arg_t *dsa;
1924 tx = dmu_tx_create(os);
1925 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1926 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1928 /* Make zl_get_data do txg_waited_synced() */
1929 return (SET_ERROR(EIO));
1933 * In order to prevent the zgd's lwb from being free'd prior to
1934 * dmu_sync_late_arrival_done() being called, we have to ensure
1935 * the lwb's "max txg" takes this tx's txg into account.
1937 zil_lwb_add_txg(zgd->zgd_lwb, dmu_tx_get_txg(tx));
1939 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1941 dsa->dsa_done = done;
1946 * Since we are currently syncing this txg, it's nontrivial to
1947 * determine what BP to nopwrite against, so we disable nopwrite.
1949 * When syncing, the db_blkptr is initially the BP of the previous
1950 * txg. We can not nopwrite against it because it will be changed
1951 * (this is similar to the non-late-arrival case where the dbuf is
1952 * dirty in a future txg).
1954 * Then dbuf_write_ready() sets bp_blkptr to the location we will write.
1955 * We can not nopwrite against it because although the BP will not
1956 * (typically) be changed, the data has not yet been persisted to this
1959 * Finally, when dbuf_write_done() is called, it is theoretically
1960 * possible to always nopwrite, because the data that was written in
1961 * this txg is the same data that we are trying to write. However we
1962 * would need to check that this dbuf is not dirty in any future
1963 * txg's (as we do in the normal dmu_sync() path). For simplicity, we
1964 * don't nopwrite in this case.
1966 zp->zp_nopwrite = B_FALSE;
1968 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1969 abd_get_from_buf(zgd->zgd_db->db_data, zgd->zgd_db->db_size),
1970 zgd->zgd_db->db_size, zgd->zgd_db->db_size, zp,
1971 dmu_sync_late_arrival_ready, NULL, NULL, dmu_sync_late_arrival_done,
1972 dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1978 * Intent log support: sync the block associated with db to disk.
1979 * N.B. and XXX: the caller is responsible for making sure that the
1980 * data isn't changing while dmu_sync() is writing it.
1984 * EEXIST: this txg has already been synced, so there's nothing to do.
1985 * The caller should not log the write.
1987 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1988 * The caller should not log the write.
1990 * EALREADY: this block is already in the process of being synced.
1991 * The caller should track its progress (somehow).
1993 * EIO: could not do the I/O.
1994 * The caller should do a txg_wait_synced().
1996 * 0: the I/O has been initiated.
1997 * The caller should log this blkptr in the done callback.
1998 * It is possible that the I/O will fail, in which case
1999 * the error will be reported to the done callback and
2000 * propagated to pio from zio_done().
2003 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
2005 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
2006 objset_t *os = db->db_objset;
2007 dsl_dataset_t *ds = os->os_dsl_dataset;
2008 dbuf_dirty_record_t *dr;
2009 dmu_sync_arg_t *dsa;
2010 zbookmark_phys_t zb;
2014 ASSERT(pio != NULL);
2017 SET_BOOKMARK(&zb, ds->ds_object,
2018 db->db.db_object, db->db_level, db->db_blkid);
2022 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
2026 * If we're frozen (running ziltest), we always need to generate a bp.
2028 if (txg > spa_freeze_txg(os->os_spa))
2029 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
2032 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
2033 * and us. If we determine that this txg is not yet syncing,
2034 * but it begins to sync a moment later, that's OK because the
2035 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
2037 mutex_enter(&db->db_mtx);
2039 if (txg <= spa_last_synced_txg(os->os_spa)) {
2041 * This txg has already synced. There's nothing to do.
2043 mutex_exit(&db->db_mtx);
2044 return (SET_ERROR(EEXIST));
2047 if (txg <= spa_syncing_txg(os->os_spa)) {
2049 * This txg is currently syncing, so we can't mess with
2050 * the dirty record anymore; just write a new log block.
2052 mutex_exit(&db->db_mtx);
2053 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
2056 dr = db->db_last_dirty;
2057 while (dr && dr->dr_txg != txg)
2062 * There's no dr for this dbuf, so it must have been freed.
2063 * There's no need to log writes to freed blocks, so we're done.
2065 mutex_exit(&db->db_mtx);
2066 return (SET_ERROR(ENOENT));
2069 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg);
2071 if (db->db_blkptr != NULL) {
2073 * We need to fill in zgd_bp with the current blkptr so that
2074 * the nopwrite code can check if we're writing the same
2075 * data that's already on disk. We can only nopwrite if we
2076 * are sure that after making the copy, db_blkptr will not
2077 * change until our i/o completes. We ensure this by
2078 * holding the db_mtx, and only allowing nopwrite if the
2079 * block is not already dirty (see below). This is verified
2080 * by dmu_sync_done(), which VERIFYs that the db_blkptr has
2083 *zgd->zgd_bp = *db->db_blkptr;
2087 * Assume the on-disk data is X, the current syncing data (in
2088 * txg - 1) is Y, and the current in-memory data is Z (currently
2091 * We usually want to perform a nopwrite if X and Z are the
2092 * same. However, if Y is different (i.e. the BP is going to
2093 * change before this write takes effect), then a nopwrite will
2094 * be incorrect - we would override with X, which could have
2095 * been freed when Y was written.
2097 * (Note that this is not a concern when we are nop-writing from
2098 * syncing context, because X and Y must be identical, because
2099 * all previous txgs have been synced.)
2101 * Therefore, we disable nopwrite if the current BP could change
2102 * before this TXG. There are two ways it could change: by
2103 * being dirty (dr_next is non-NULL), or by being freed
2104 * (dnode_block_freed()). This behavior is verified by
2105 * zio_done(), which VERIFYs that the override BP is identical
2106 * to the on-disk BP.
2110 if (dr->dr_next != NULL || dnode_block_freed(dn, db->db_blkid))
2111 zp.zp_nopwrite = B_FALSE;
2114 ASSERT(dr->dr_txg == txg);
2115 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
2116 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
2118 * We have already issued a sync write for this buffer,
2119 * or this buffer has already been synced. It could not
2120 * have been dirtied since, or we would have cleared the state.
2122 mutex_exit(&db->db_mtx);
2123 return (SET_ERROR(EALREADY));
2126 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
2127 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
2128 mutex_exit(&db->db_mtx);
2130 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
2132 dsa->dsa_done = done;
2136 zio_nowait(arc_write(pio, os->os_spa, txg,
2137 zgd->zgd_bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db),
2138 &zp, dmu_sync_ready, NULL, NULL, dmu_sync_done, dsa,
2139 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb));
2145 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
2151 err = dnode_hold(os, object, FTAG, &dn);
2154 err = dnode_set_blksz(dn, size, ibs, tx);
2155 dnode_rele(dn, FTAG);
2160 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
2166 * Send streams include each object's checksum function. This
2167 * check ensures that the receiving system can understand the
2168 * checksum function transmitted.
2170 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS);
2172 VERIFY0(dnode_hold(os, object, FTAG, &dn));
2173 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS);
2174 dn->dn_checksum = checksum;
2175 dnode_setdirty(dn, tx);
2176 dnode_rele(dn, FTAG);
2180 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
2186 * Send streams include each object's compression function. This
2187 * check ensures that the receiving system can understand the
2188 * compression function transmitted.
2190 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS);
2192 VERIFY0(dnode_hold(os, object, FTAG, &dn));
2193 dn->dn_compress = compress;
2194 dnode_setdirty(dn, tx);
2195 dnode_rele(dn, FTAG);
2198 int zfs_mdcomp_disable = 0;
2199 SYSCTL_INT(_vfs_zfs, OID_AUTO, mdcomp_disable, CTLFLAG_RWTUN,
2200 &zfs_mdcomp_disable, 0, "Disable metadata compression");
2203 * When the "redundant_metadata" property is set to "most", only indirect
2204 * blocks of this level and higher will have an additional ditto block.
2206 int zfs_redundant_metadata_most_ditto_level = 2;
2209 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
2211 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
2212 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
2214 enum zio_checksum checksum = os->os_checksum;
2215 enum zio_compress compress = os->os_compress;
2216 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
2217 boolean_t dedup = B_FALSE;
2218 boolean_t nopwrite = B_FALSE;
2219 boolean_t dedup_verify = os->os_dedup_verify;
2220 int copies = os->os_copies;
2223 * We maintain different write policies for each of the following
2226 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
2227 * 3. all other level 0 blocks
2230 if (zfs_mdcomp_disable) {
2231 compress = ZIO_COMPRESS_EMPTY;
2234 * XXX -- we should design a compression algorithm
2235 * that specializes in arrays of bps.
2237 compress = zio_compress_select(os->os_spa,
2238 ZIO_COMPRESS_ON, ZIO_COMPRESS_ON);
2242 * Metadata always gets checksummed. If the data
2243 * checksum is multi-bit correctable, and it's not a
2244 * ZBT-style checksum, then it's suitable for metadata
2245 * as well. Otherwise, the metadata checksum defaults
2248 if (!(zio_checksum_table[checksum].ci_flags &
2249 ZCHECKSUM_FLAG_METADATA) ||
2250 (zio_checksum_table[checksum].ci_flags &
2251 ZCHECKSUM_FLAG_EMBEDDED))
2252 checksum = ZIO_CHECKSUM_FLETCHER_4;
2254 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL ||
2255 (os->os_redundant_metadata ==
2256 ZFS_REDUNDANT_METADATA_MOST &&
2257 (level >= zfs_redundant_metadata_most_ditto_level ||
2258 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))))
2260 } else if (wp & WP_NOFILL) {
2264 * If we're writing preallocated blocks, we aren't actually
2265 * writing them so don't set any policy properties. These
2266 * blocks are currently only used by an external subsystem
2267 * outside of zfs (i.e. dump) and not written by the zio
2270 compress = ZIO_COMPRESS_OFF;
2271 checksum = ZIO_CHECKSUM_NOPARITY;
2273 compress = zio_compress_select(os->os_spa, dn->dn_compress,
2276 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
2277 zio_checksum_select(dn->dn_checksum, checksum) :
2281 * Determine dedup setting. If we are in dmu_sync(),
2282 * we won't actually dedup now because that's all
2283 * done in syncing context; but we do want to use the
2284 * dedup checkum. If the checksum is not strong
2285 * enough to ensure unique signatures, force
2288 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
2289 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
2290 if (!(zio_checksum_table[checksum].ci_flags &
2291 ZCHECKSUM_FLAG_DEDUP))
2292 dedup_verify = B_TRUE;
2296 * Enable nopwrite if we have secure enough checksum
2297 * algorithm (see comment in zio_nop_write) and
2298 * compression is enabled. We don't enable nopwrite if
2299 * dedup is enabled as the two features are mutually
2302 nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags &
2303 ZCHECKSUM_FLAG_NOPWRITE) &&
2304 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
2307 zp->zp_checksum = checksum;
2308 zp->zp_compress = compress;
2309 ASSERT3U(zp->zp_compress, !=, ZIO_COMPRESS_INHERIT);
2311 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
2312 zp->zp_level = level;
2313 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
2314 zp->zp_dedup = dedup;
2315 zp->zp_dedup_verify = dedup && dedup_verify;
2316 zp->zp_nopwrite = nopwrite;
2320 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
2326 * Sync any current changes before
2327 * we go trundling through the block pointers.
2329 err = dmu_object_wait_synced(os, object);
2334 err = dnode_hold(os, object, FTAG, &dn);
2339 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
2340 dnode_rele(dn, FTAG);
2346 * Given the ZFS object, if it contains any dirty nodes
2347 * this function flushes all dirty blocks to disk. This
2348 * ensures the DMU object info is updated. A more efficient
2349 * future version might just find the TXG with the maximum
2350 * ID and wait for that to be synced.
2353 dmu_object_wait_synced(objset_t *os, uint64_t object)
2358 error = dnode_hold(os, object, FTAG, &dn);
2363 for (i = 0; i < TXG_SIZE; i++) {
2364 if (list_link_active(&dn->dn_dirty_link[i])) {
2368 dnode_rele(dn, FTAG);
2369 if (i != TXG_SIZE) {
2370 txg_wait_synced(dmu_objset_pool(os), 0);
2377 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2381 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2382 mutex_enter(&dn->dn_mtx);
2386 doi->doi_data_block_size = dn->dn_datablksz;
2387 doi->doi_metadata_block_size = dn->dn_indblkshift ?
2388 1ULL << dn->dn_indblkshift : 0;
2389 doi->doi_type = dn->dn_type;
2390 doi->doi_bonus_type = dn->dn_bonustype;
2391 doi->doi_bonus_size = dn->dn_bonuslen;
2392 doi->doi_indirection = dn->dn_nlevels;
2393 doi->doi_checksum = dn->dn_checksum;
2394 doi->doi_compress = dn->dn_compress;
2395 doi->doi_nblkptr = dn->dn_nblkptr;
2396 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
2397 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
2398 doi->doi_fill_count = 0;
2399 for (int i = 0; i < dnp->dn_nblkptr; i++)
2400 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]);
2402 mutex_exit(&dn->dn_mtx);
2403 rw_exit(&dn->dn_struct_rwlock);
2407 * Get information on a DMU object.
2408 * If doi is NULL, just indicates whether the object exists.
2411 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
2414 int err = dnode_hold(os, object, FTAG, &dn);
2420 dmu_object_info_from_dnode(dn, doi);
2422 dnode_rele(dn, FTAG);
2427 * As above, but faster; can be used when you have a held dbuf in hand.
2430 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
2432 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2435 dmu_object_info_from_dnode(DB_DNODE(db), doi);
2440 * Faster still when you only care about the size.
2441 * This is specifically optimized for zfs_getattr().
2444 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
2445 u_longlong_t *nblk512)
2447 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2453 *blksize = dn->dn_datablksz;
2454 /* add 1 for dnode space */
2455 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
2456 SPA_MINBLOCKSHIFT) + 1;
2461 byteswap_uint64_array(void *vbuf, size_t size)
2463 uint64_t *buf = vbuf;
2464 size_t count = size >> 3;
2467 ASSERT((size & 7) == 0);
2469 for (i = 0; i < count; i++)
2470 buf[i] = BSWAP_64(buf[i]);
2474 byteswap_uint32_array(void *vbuf, size_t size)
2476 uint32_t *buf = vbuf;
2477 size_t count = size >> 2;
2480 ASSERT((size & 3) == 0);
2482 for (i = 0; i < count; i++)
2483 buf[i] = BSWAP_32(buf[i]);
2487 byteswap_uint16_array(void *vbuf, size_t size)
2489 uint16_t *buf = vbuf;
2490 size_t count = size >> 1;
2493 ASSERT((size & 1) == 0);
2495 for (i = 0; i < count; i++)
2496 buf[i] = BSWAP_16(buf[i]);
2501 byteswap_uint8_array(void *vbuf, size_t size)
2515 zio_compress_init();
2524 arc_fini(); /* arc depends on l2arc, so arc must go first */
2527 zio_compress_fini();