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");
75 * This can be used for testing, to ensure that certain actions happen
76 * while in the middle of a remap (which might otherwise complete too
79 int zfs_object_remap_one_indirect_delay_ticks = 0;
81 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
82 { DMU_BSWAP_UINT8, TRUE, FALSE, "unallocated" },
83 { DMU_BSWAP_ZAP, TRUE, TRUE, "object directory" },
84 { DMU_BSWAP_UINT64, TRUE, TRUE, "object array" },
85 { DMU_BSWAP_UINT8, TRUE, FALSE, "packed nvlist" },
86 { DMU_BSWAP_UINT64, TRUE, FALSE, "packed nvlist size" },
87 { DMU_BSWAP_UINT64, TRUE, FALSE, "bpobj" },
88 { DMU_BSWAP_UINT64, TRUE, FALSE, "bpobj header" },
89 { DMU_BSWAP_UINT64, TRUE, FALSE, "SPA space map header" },
90 { DMU_BSWAP_UINT64, TRUE, FALSE, "SPA space map" },
91 { DMU_BSWAP_UINT64, TRUE, FALSE, "ZIL intent log" },
92 { DMU_BSWAP_DNODE, TRUE, FALSE, "DMU dnode" },
93 { DMU_BSWAP_OBJSET, TRUE, TRUE, "DMU objset" },
94 { DMU_BSWAP_UINT64, TRUE, TRUE, "DSL directory" },
95 { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL directory child map" },
96 { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL dataset snap map" },
97 { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL props" },
98 { DMU_BSWAP_UINT64, TRUE, TRUE, "DSL dataset" },
99 { DMU_BSWAP_ZNODE, TRUE, FALSE, "ZFS znode" },
100 { DMU_BSWAP_OLDACL, TRUE, FALSE, "ZFS V0 ACL" },
101 { DMU_BSWAP_UINT8, FALSE, FALSE, "ZFS plain file" },
102 { DMU_BSWAP_ZAP, TRUE, FALSE, "ZFS directory" },
103 { DMU_BSWAP_ZAP, TRUE, FALSE, "ZFS master node" },
104 { DMU_BSWAP_ZAP, TRUE, FALSE, "ZFS delete queue" },
105 { DMU_BSWAP_UINT8, FALSE, FALSE, "zvol object" },
106 { DMU_BSWAP_ZAP, TRUE, FALSE, "zvol prop" },
107 { DMU_BSWAP_UINT8, FALSE, FALSE, "other uint8[]" },
108 { DMU_BSWAP_UINT64, FALSE, FALSE, "other uint64[]" },
109 { DMU_BSWAP_ZAP, TRUE, FALSE, "other ZAP" },
110 { DMU_BSWAP_ZAP, TRUE, FALSE, "persistent error log" },
111 { DMU_BSWAP_UINT8, TRUE, FALSE, "SPA history" },
112 { DMU_BSWAP_UINT64, TRUE, FALSE, "SPA history offsets" },
113 { DMU_BSWAP_ZAP, TRUE, TRUE, "Pool properties" },
114 { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL permissions" },
115 { DMU_BSWAP_ACL, TRUE, FALSE, "ZFS ACL" },
116 { DMU_BSWAP_UINT8, TRUE, FALSE, "ZFS SYSACL" },
117 { DMU_BSWAP_UINT8, TRUE, FALSE, "FUID table" },
118 { DMU_BSWAP_UINT64, TRUE, FALSE, "FUID table size" },
119 { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL dataset next clones" },
120 { DMU_BSWAP_ZAP, TRUE, FALSE, "scan work queue" },
121 { DMU_BSWAP_ZAP, TRUE, FALSE, "ZFS user/group used" },
122 { DMU_BSWAP_ZAP, TRUE, FALSE, "ZFS user/group quota" },
123 { DMU_BSWAP_ZAP, TRUE, TRUE, "snapshot refcount tags" },
124 { DMU_BSWAP_ZAP, TRUE, FALSE, "DDT ZAP algorithm" },
125 { DMU_BSWAP_ZAP, TRUE, FALSE, "DDT statistics" },
126 { DMU_BSWAP_UINT8, TRUE, FALSE, "System attributes" },
127 { DMU_BSWAP_ZAP, TRUE, FALSE, "SA master node" },
128 { DMU_BSWAP_ZAP, TRUE, FALSE, "SA attr registration" },
129 { DMU_BSWAP_ZAP, TRUE, FALSE, "SA attr layouts" },
130 { DMU_BSWAP_ZAP, TRUE, FALSE, "scan translations" },
131 { DMU_BSWAP_UINT8, FALSE, FALSE, "deduplicated block" },
132 { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL deadlist map" },
133 { DMU_BSWAP_UINT64, TRUE, TRUE, "DSL deadlist map hdr" },
134 { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL dir clones" },
135 { DMU_BSWAP_UINT64, TRUE, FALSE, "bpobj subobj" }
138 const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
139 { byteswap_uint8_array, "uint8" },
140 { byteswap_uint16_array, "uint16" },
141 { byteswap_uint32_array, "uint32" },
142 { byteswap_uint64_array, "uint64" },
143 { zap_byteswap, "zap" },
144 { dnode_buf_byteswap, "dnode" },
145 { dmu_objset_byteswap, "objset" },
146 { zfs_znode_byteswap, "znode" },
147 { zfs_oldacl_byteswap, "oldacl" },
148 { zfs_acl_byteswap, "acl" }
152 dmu_buf_hold_noread_by_dnode(dnode_t *dn, uint64_t offset,
153 void *tag, dmu_buf_t **dbp)
158 blkid = dbuf_whichblock(dn, 0, offset);
159 rw_enter(&dn->dn_struct_rwlock, RW_READER);
160 db = dbuf_hold(dn, blkid, tag);
161 rw_exit(&dn->dn_struct_rwlock);
165 return (SET_ERROR(EIO));
172 dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset,
173 void *tag, dmu_buf_t **dbp)
180 err = dnode_hold(os, object, FTAG, &dn);
183 blkid = dbuf_whichblock(dn, 0, offset);
184 rw_enter(&dn->dn_struct_rwlock, RW_READER);
185 db = dbuf_hold(dn, blkid, tag);
186 rw_exit(&dn->dn_struct_rwlock);
187 dnode_rele(dn, FTAG);
191 return (SET_ERROR(EIO));
199 dmu_buf_hold_by_dnode(dnode_t *dn, uint64_t offset,
200 void *tag, dmu_buf_t **dbp, int flags)
203 int db_flags = DB_RF_CANFAIL;
205 if (flags & DMU_READ_NO_PREFETCH)
206 db_flags |= DB_RF_NOPREFETCH;
208 err = dmu_buf_hold_noread_by_dnode(dn, offset, tag, dbp);
210 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
211 err = dbuf_read(db, NULL, db_flags);
222 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
223 void *tag, dmu_buf_t **dbp, int flags)
226 int db_flags = DB_RF_CANFAIL;
228 if (flags & DMU_READ_NO_PREFETCH)
229 db_flags |= DB_RF_NOPREFETCH;
231 err = dmu_buf_hold_noread(os, object, offset, tag, dbp);
233 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
234 err = dbuf_read(db, NULL, db_flags);
247 return (DN_OLD_MAX_BONUSLEN);
251 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx)
253 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
260 if (dn->dn_bonus != db) {
261 error = SET_ERROR(EINVAL);
262 } else if (newsize < 0 || newsize > db_fake->db_size) {
263 error = SET_ERROR(EINVAL);
265 dnode_setbonuslen(dn, newsize, tx);
274 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
276 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
283 if (!DMU_OT_IS_VALID(type)) {
284 error = SET_ERROR(EINVAL);
285 } else if (dn->dn_bonus != db) {
286 error = SET_ERROR(EINVAL);
288 dnode_setbonus_type(dn, type, tx);
297 dmu_get_bonustype(dmu_buf_t *db_fake)
299 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
301 dmu_object_type_t type;
305 type = dn->dn_bonustype;
312 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
317 error = dnode_hold(os, object, FTAG, &dn);
318 dbuf_rm_spill(dn, tx);
319 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
320 dnode_rm_spill(dn, tx);
321 rw_exit(&dn->dn_struct_rwlock);
322 dnode_rele(dn, FTAG);
327 * returns ENOENT, EIO, or 0.
330 dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp)
336 error = dnode_hold(os, object, FTAG, &dn);
340 rw_enter(&dn->dn_struct_rwlock, RW_READER);
341 if (dn->dn_bonus == NULL) {
342 rw_exit(&dn->dn_struct_rwlock);
343 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
344 if (dn->dn_bonus == NULL)
345 dbuf_create_bonus(dn);
349 /* as long as the bonus buf is held, the dnode will be held */
350 if (refcount_add(&db->db_holds, tag) == 1) {
351 VERIFY(dnode_add_ref(dn, db));
352 atomic_inc_32(&dn->dn_dbufs_count);
356 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
357 * hold and incrementing the dbuf count to ensure that dnode_move() sees
358 * a dnode hold for every dbuf.
360 rw_exit(&dn->dn_struct_rwlock);
362 dnode_rele(dn, FTAG);
364 VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH));
371 * returns ENOENT, EIO, or 0.
373 * This interface will allocate a blank spill dbuf when a spill blk
374 * doesn't already exist on the dnode.
376 * if you only want to find an already existing spill db, then
377 * dmu_spill_hold_existing() should be used.
380 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp)
382 dmu_buf_impl_t *db = NULL;
385 if ((flags & DB_RF_HAVESTRUCT) == 0)
386 rw_enter(&dn->dn_struct_rwlock, RW_READER);
388 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
390 if ((flags & DB_RF_HAVESTRUCT) == 0)
391 rw_exit(&dn->dn_struct_rwlock);
394 err = dbuf_read(db, NULL, flags);
403 dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
405 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
412 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
413 err = SET_ERROR(EINVAL);
415 rw_enter(&dn->dn_struct_rwlock, RW_READER);
417 if (!dn->dn_have_spill) {
418 err = SET_ERROR(ENOENT);
420 err = dmu_spill_hold_by_dnode(dn,
421 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
424 rw_exit(&dn->dn_struct_rwlock);
432 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
434 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
440 err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp);
447 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
448 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
449 * and can induce severe lock contention when writing to several files
450 * whose dnodes are in the same block.
453 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
454 boolean_t read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
457 uint64_t blkid, nblks, i;
462 ASSERT(length <= DMU_MAX_ACCESS);
465 * Note: We directly notify the prefetch code of this read, so that
466 * we can tell it about the multi-block read. dbuf_read() only knows
467 * about the one block it is accessing.
469 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT |
472 rw_enter(&dn->dn_struct_rwlock, RW_READER);
473 if (dn->dn_datablkshift) {
474 int blkshift = dn->dn_datablkshift;
475 nblks = (P2ROUNDUP(offset + length, 1ULL << blkshift) -
476 P2ALIGN(offset, 1ULL << blkshift)) >> blkshift;
478 if (offset + length > dn->dn_datablksz) {
479 zfs_panic_recover("zfs: accessing past end of object "
480 "%llx/%llx (size=%u access=%llu+%llu)",
481 (longlong_t)dn->dn_objset->
482 os_dsl_dataset->ds_object,
483 (longlong_t)dn->dn_object, dn->dn_datablksz,
484 (longlong_t)offset, (longlong_t)length);
485 rw_exit(&dn->dn_struct_rwlock);
486 return (SET_ERROR(EIO));
490 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
492 #if defined(_KERNEL) && defined(RACCT)
493 if (racct_enable && !read) {
495 racct_add_force(curproc, RACCT_WRITEBPS, length);
496 racct_add_force(curproc, RACCT_WRITEIOPS, nblks);
497 PROC_UNLOCK(curproc);
501 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
502 blkid = dbuf_whichblock(dn, 0, offset);
503 for (i = 0; i < nblks; i++) {
504 dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag);
506 rw_exit(&dn->dn_struct_rwlock);
507 dmu_buf_rele_array(dbp, nblks, tag);
509 return (SET_ERROR(EIO));
512 /* initiate async i/o */
514 (void) dbuf_read(db, zio, dbuf_flags);
517 curthread->td_ru.ru_oublock++;
522 if ((flags & DMU_READ_NO_PREFETCH) == 0 &&
523 DNODE_META_IS_CACHEABLE(dn) && length <= zfetch_array_rd_sz) {
524 dmu_zfetch(&dn->dn_zfetch, blkid, nblks,
525 read && DNODE_IS_CACHEABLE(dn));
527 rw_exit(&dn->dn_struct_rwlock);
529 /* wait for async i/o */
532 dmu_buf_rele_array(dbp, nblks, tag);
536 /* wait for other io to complete */
538 for (i = 0; i < nblks; i++) {
539 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
540 mutex_enter(&db->db_mtx);
541 while (db->db_state == DB_READ ||
542 db->db_state == DB_FILL)
543 cv_wait(&db->db_changed, &db->db_mtx);
544 if (db->db_state == DB_UNCACHED)
545 err = SET_ERROR(EIO);
546 mutex_exit(&db->db_mtx);
548 dmu_buf_rele_array(dbp, nblks, tag);
560 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
561 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
566 err = dnode_hold(os, object, FTAG, &dn);
570 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
571 numbufsp, dbpp, DMU_READ_PREFETCH);
573 dnode_rele(dn, FTAG);
579 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
580 uint64_t length, boolean_t read, void *tag, int *numbufsp,
583 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
589 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
590 numbufsp, dbpp, DMU_READ_PREFETCH);
597 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
600 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
605 for (i = 0; i < numbufs; i++) {
607 dbuf_rele(dbp[i], tag);
610 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
614 * Issue prefetch i/os for the given blocks. If level is greater than 0, the
615 * indirect blocks prefeteched will be those that point to the blocks containing
616 * the data starting at offset, and continuing to offset + len.
618 * Note that if the indirect blocks above the blocks being prefetched are not in
619 * cache, they will be asychronously read in.
622 dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset,
623 uint64_t len, zio_priority_t pri)
629 if (len == 0) { /* they're interested in the bonus buffer */
630 dn = DMU_META_DNODE(os);
632 if (object == 0 || object >= DN_MAX_OBJECT)
635 rw_enter(&dn->dn_struct_rwlock, RW_READER);
636 blkid = dbuf_whichblock(dn, level,
637 object * sizeof (dnode_phys_t));
638 dbuf_prefetch(dn, level, blkid, pri, 0);
639 rw_exit(&dn->dn_struct_rwlock);
644 * XXX - Note, if the dnode for the requested object is not
645 * already cached, we will do a *synchronous* read in the
646 * dnode_hold() call. The same is true for any indirects.
648 err = dnode_hold(os, object, FTAG, &dn);
652 rw_enter(&dn->dn_struct_rwlock, RW_READER);
654 * offset + len - 1 is the last byte we want to prefetch for, and offset
655 * is the first. Then dbuf_whichblk(dn, level, off + len - 1) is the
656 * last block we want to prefetch, and dbuf_whichblock(dn, level,
657 * offset) is the first. Then the number we need to prefetch is the
660 if (level > 0 || dn->dn_datablkshift != 0) {
661 nblks = dbuf_whichblock(dn, level, offset + len - 1) -
662 dbuf_whichblock(dn, level, offset) + 1;
664 nblks = (offset < dn->dn_datablksz);
668 blkid = dbuf_whichblock(dn, level, offset);
669 for (int i = 0; i < nblks; i++)
670 dbuf_prefetch(dn, level, blkid + i, pri, 0);
673 rw_exit(&dn->dn_struct_rwlock);
675 dnode_rele(dn, FTAG);
679 * Get the next "chunk" of file data to free. We traverse the file from
680 * the end so that the file gets shorter over time (if we crashes in the
681 * middle, this will leave us in a better state). We find allocated file
682 * data by simply searching the allocated level 1 indirects.
684 * On input, *start should be the first offset that does not need to be
685 * freed (e.g. "offset + length"). On return, *start will be the first
686 * offset that should be freed.
689 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum)
691 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
692 /* bytes of data covered by a level-1 indirect block */
694 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
696 ASSERT3U(minimum, <=, *start);
698 if (*start - minimum <= iblkrange * maxblks) {
702 ASSERT(ISP2(iblkrange));
704 for (uint64_t blks = 0; *start > minimum && blks < maxblks; blks++) {
708 * dnode_next_offset(BACKWARDS) will find an allocated L1
709 * indirect block at or before the input offset. We must
710 * decrement *start so that it is at the end of the region
714 err = dnode_next_offset(dn,
715 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
717 /* if there are no indirect blocks before start, we are done */
721 } else if (err != 0) {
725 /* set start to the beginning of this L1 indirect */
726 *start = P2ALIGN(*start, iblkrange);
728 if (*start < minimum)
734 * If this objset is of type OST_ZFS return true if vfs's unmounted flag is set,
735 * otherwise return false.
736 * Used below in dmu_free_long_range_impl() to enable abort when unmounting
740 dmu_objset_zfs_unmounting(objset_t *os)
743 if (dmu_objset_type(os) == DMU_OST_ZFS)
744 return (zfs_get_vfs_flag_unmounted(os));
750 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
753 uint64_t object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
755 uint64_t dirty_frees_threshold;
756 dsl_pool_t *dp = dmu_objset_pool(os);
758 if (offset >= object_size)
761 if (zfs_per_txg_dirty_frees_percent <= 100)
762 dirty_frees_threshold =
763 zfs_per_txg_dirty_frees_percent * zfs_dirty_data_max / 100;
765 dirty_frees_threshold = zfs_dirty_data_max / 4;
767 if (length == DMU_OBJECT_END || offset + length > object_size)
768 length = object_size - offset;
770 while (length != 0) {
771 uint64_t chunk_end, chunk_begin, chunk_len;
772 uint64_t long_free_dirty_all_txgs = 0;
775 if (dmu_objset_zfs_unmounting(dn->dn_objset))
776 return (SET_ERROR(EINTR));
778 chunk_end = chunk_begin = offset + length;
780 /* move chunk_begin backwards to the beginning of this chunk */
781 err = get_next_chunk(dn, &chunk_begin, offset);
784 ASSERT3U(chunk_begin, >=, offset);
785 ASSERT3U(chunk_begin, <=, chunk_end);
787 chunk_len = chunk_end - chunk_begin;
789 mutex_enter(&dp->dp_lock);
790 for (int t = 0; t < TXG_SIZE; t++) {
791 long_free_dirty_all_txgs +=
792 dp->dp_long_free_dirty_pertxg[t];
794 mutex_exit(&dp->dp_lock);
797 * To avoid filling up a TXG with just frees wait for
798 * the next TXG to open before freeing more chunks if
799 * we have reached the threshold of frees
801 if (dirty_frees_threshold != 0 &&
802 long_free_dirty_all_txgs >= dirty_frees_threshold) {
803 txg_wait_open(dp, 0);
807 tx = dmu_tx_create(os);
808 dmu_tx_hold_free(tx, dn->dn_object, chunk_begin, chunk_len);
811 * Mark this transaction as typically resulting in a net
812 * reduction in space used.
814 dmu_tx_mark_netfree(tx);
815 err = dmu_tx_assign(tx, TXG_WAIT);
821 mutex_enter(&dp->dp_lock);
822 dp->dp_long_free_dirty_pertxg[dmu_tx_get_txg(tx) & TXG_MASK] +=
824 mutex_exit(&dp->dp_lock);
825 DTRACE_PROBE3(free__long__range,
826 uint64_t, long_free_dirty_all_txgs, uint64_t, chunk_len,
827 uint64_t, dmu_tx_get_txg(tx));
828 dnode_free_range(dn, chunk_begin, chunk_len, tx);
837 dmu_free_long_range(objset_t *os, uint64_t object,
838 uint64_t offset, uint64_t length)
843 err = dnode_hold(os, object, FTAG, &dn);
846 err = dmu_free_long_range_impl(os, dn, offset, length);
849 * It is important to zero out the maxblkid when freeing the entire
850 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
851 * will take the fast path, and (b) dnode_reallocate() can verify
852 * that the entire file has been freed.
854 if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
857 dnode_rele(dn, FTAG);
862 dmu_free_long_object(objset_t *os, uint64_t object)
867 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
871 tx = dmu_tx_create(os);
872 dmu_tx_hold_bonus(tx, object);
873 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
874 dmu_tx_mark_netfree(tx);
875 err = dmu_tx_assign(tx, TXG_WAIT);
877 err = dmu_object_free(os, object, tx);
887 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
888 uint64_t size, dmu_tx_t *tx)
891 int err = dnode_hold(os, object, FTAG, &dn);
894 ASSERT(offset < UINT64_MAX);
895 ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
896 dnode_free_range(dn, offset, size, tx);
897 dnode_rele(dn, FTAG);
902 dmu_read_impl(dnode_t *dn, uint64_t offset, uint64_t size,
903 void *buf, uint32_t flags)
906 int numbufs, err = 0;
909 * Deal with odd block sizes, where there can't be data past the first
910 * block. If we ever do the tail block optimization, we will need to
911 * handle that here as well.
913 if (dn->dn_maxblkid == 0) {
914 int newsz = offset > dn->dn_datablksz ? 0 :
915 MIN(size, dn->dn_datablksz - offset);
916 bzero((char *)buf + newsz, size - newsz);
921 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
925 * NB: we could do this block-at-a-time, but it's nice
926 * to be reading in parallel.
928 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
929 TRUE, FTAG, &numbufs, &dbp, flags);
933 for (i = 0; i < numbufs; i++) {
936 dmu_buf_t *db = dbp[i];
940 bufoff = offset - db->db_offset;
941 tocpy = (int)MIN(db->db_size - bufoff, size);
943 bcopy((char *)db->db_data + bufoff, buf, tocpy);
947 buf = (char *)buf + tocpy;
949 dmu_buf_rele_array(dbp, numbufs, FTAG);
955 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
956 void *buf, uint32_t flags)
961 err = dnode_hold(os, object, FTAG, &dn);
965 err = dmu_read_impl(dn, offset, size, buf, flags);
966 dnode_rele(dn, FTAG);
971 dmu_read_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, void *buf,
974 return (dmu_read_impl(dn, offset, size, buf, flags));
978 dmu_write_impl(dmu_buf_t **dbp, int numbufs, uint64_t offset, uint64_t size,
979 const void *buf, dmu_tx_t *tx)
983 for (i = 0; i < numbufs; i++) {
986 dmu_buf_t *db = dbp[i];
990 bufoff = offset - db->db_offset;
991 tocpy = (int)MIN(db->db_size - bufoff, size);
993 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
995 if (tocpy == db->db_size)
996 dmu_buf_will_fill(db, tx);
998 dmu_buf_will_dirty(db, tx);
1000 bcopy(buf, (char *)db->db_data + bufoff, tocpy);
1002 if (tocpy == db->db_size)
1003 dmu_buf_fill_done(db, tx);
1007 buf = (char *)buf + tocpy;
1012 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1013 const void *buf, dmu_tx_t *tx)
1021 VERIFY0(dmu_buf_hold_array(os, object, offset, size,
1022 FALSE, FTAG, &numbufs, &dbp));
1023 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1024 dmu_buf_rele_array(dbp, numbufs, FTAG);
1028 dmu_write_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size,
1029 const void *buf, dmu_tx_t *tx)
1037 VERIFY0(dmu_buf_hold_array_by_dnode(dn, offset, size,
1038 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH));
1039 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1040 dmu_buf_rele_array(dbp, numbufs, FTAG);
1044 dmu_object_remap_one_indirect(objset_t *os, dnode_t *dn,
1045 uint64_t last_removal_txg, uint64_t offset)
1047 uint64_t l1blkid = dbuf_whichblock(dn, 1, offset);
1050 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1051 dmu_buf_impl_t *dbuf = dbuf_hold_level(dn, 1, l1blkid, FTAG);
1052 ASSERT3P(dbuf, !=, NULL);
1055 * If the block hasn't been written yet, this default will ensure
1056 * we don't try to remap it.
1058 uint64_t birth = UINT64_MAX;
1059 ASSERT3U(last_removal_txg, !=, UINT64_MAX);
1060 if (dbuf->db_blkptr != NULL)
1061 birth = dbuf->db_blkptr->blk_birth;
1062 rw_exit(&dn->dn_struct_rwlock);
1065 * If this L1 was already written after the last removal, then we've
1066 * already tried to remap it.
1068 if (birth <= last_removal_txg &&
1069 dbuf_read(dbuf, NULL, DB_RF_MUST_SUCCEED) == 0 &&
1070 dbuf_can_remap(dbuf)) {
1071 dmu_tx_t *tx = dmu_tx_create(os);
1072 dmu_tx_hold_remap_l1indirect(tx, dn->dn_object);
1073 err = dmu_tx_assign(tx, TXG_WAIT);
1075 (void) dbuf_dirty(dbuf, tx);
1082 dbuf_rele(dbuf, FTAG);
1084 delay(zfs_object_remap_one_indirect_delay_ticks);
1090 * Remap all blockpointers in the object, if possible, so that they reference
1091 * only concrete vdevs.
1093 * To do this, iterate over the L0 blockpointers and remap any that reference
1094 * an indirect vdev. Note that we only examine L0 blockpointers; since we
1095 * cannot guarantee that we can remap all blockpointer anyways (due to split
1096 * blocks), we do not want to make the code unnecessarily complicated to
1097 * catch the unlikely case that there is an L1 block on an indirect vdev that
1098 * contains no indirect blockpointers.
1101 dmu_object_remap_indirects(objset_t *os, uint64_t object,
1102 uint64_t last_removal_txg)
1104 uint64_t offset, l1span;
1108 err = dnode_hold(os, object, FTAG, &dn);
1113 if (dn->dn_nlevels <= 1) {
1114 if (issig(JUSTLOOKING) && issig(FORREAL)) {
1115 err = SET_ERROR(EINTR);
1119 * If the dnode has no indirect blocks, we cannot dirty them.
1120 * We still want to remap the blkptr(s) in the dnode if
1121 * appropriate, so mark it as dirty.
1123 if (err == 0 && dnode_needs_remap(dn)) {
1124 dmu_tx_t *tx = dmu_tx_create(os);
1125 dmu_tx_hold_bonus(tx, dn->dn_object);
1126 if ((err = dmu_tx_assign(tx, TXG_WAIT)) == 0) {
1127 dnode_setdirty(dn, tx);
1134 dnode_rele(dn, FTAG);
1139 l1span = 1ULL << (dn->dn_indblkshift - SPA_BLKPTRSHIFT +
1140 dn->dn_datablkshift);
1142 * Find the next L1 indirect that is not a hole.
1144 while (dnode_next_offset(dn, 0, &offset, 2, 1, 0) == 0) {
1145 if (issig(JUSTLOOKING) && issig(FORREAL)) {
1146 err = SET_ERROR(EINTR);
1149 if ((err = dmu_object_remap_one_indirect(os, dn,
1150 last_removal_txg, offset)) != 0) {
1156 dnode_rele(dn, FTAG);
1161 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1170 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
1171 FALSE, FTAG, &numbufs, &dbp));
1173 for (i = 0; i < numbufs; i++) {
1174 dmu_buf_t *db = dbp[i];
1176 dmu_buf_will_not_fill(db, tx);
1178 dmu_buf_rele_array(dbp, numbufs, FTAG);
1182 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset,
1183 void *data, uint8_t etype, uint8_t comp, int uncompressed_size,
1184 int compressed_size, int byteorder, dmu_tx_t *tx)
1188 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES);
1189 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS);
1190 VERIFY0(dmu_buf_hold_noread(os, object, offset,
1193 dmu_buf_write_embedded(db,
1194 data, (bp_embedded_type_t)etype, (enum zio_compress)comp,
1195 uncompressed_size, compressed_size, byteorder, tx);
1197 dmu_buf_rele(db, FTAG);
1201 * DMU support for xuio
1203 kstat_t *xuio_ksp = NULL;
1206 dmu_xuio_init(xuio_t *xuio, int nblk)
1209 uio_t *uio = &xuio->xu_uio;
1211 uio->uio_iovcnt = nblk;
1212 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
1214 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
1216 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
1217 priv->iovp = uio->uio_iov;
1218 XUIO_XUZC_PRIV(xuio) = priv;
1220 if (XUIO_XUZC_RW(xuio) == UIO_READ)
1221 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
1223 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
1229 dmu_xuio_fini(xuio_t *xuio)
1231 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1232 int nblk = priv->cnt;
1234 kmem_free(priv->iovp, nblk * sizeof (iovec_t));
1235 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
1236 kmem_free(priv, sizeof (dmu_xuio_t));
1238 if (XUIO_XUZC_RW(xuio) == UIO_READ)
1239 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
1241 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
1245 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
1246 * and increase priv->next by 1.
1249 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
1252 uio_t *uio = &xuio->xu_uio;
1253 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1254 int i = priv->next++;
1256 ASSERT(i < priv->cnt);
1257 ASSERT(off + n <= arc_buf_lsize(abuf));
1258 iov = uio->uio_iov + i;
1259 iov->iov_base = (char *)abuf->b_data + off;
1261 priv->bufs[i] = abuf;
1266 dmu_xuio_cnt(xuio_t *xuio)
1268 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1273 dmu_xuio_arcbuf(xuio_t *xuio, int i)
1275 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1277 ASSERT(i < priv->cnt);
1278 return (priv->bufs[i]);
1282 dmu_xuio_clear(xuio_t *xuio, int i)
1284 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1286 ASSERT(i < priv->cnt);
1287 priv->bufs[i] = NULL;
1291 xuio_stat_init(void)
1293 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
1294 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
1295 KSTAT_FLAG_VIRTUAL);
1296 if (xuio_ksp != NULL) {
1297 xuio_ksp->ks_data = &xuio_stats;
1298 kstat_install(xuio_ksp);
1303 xuio_stat_fini(void)
1305 if (xuio_ksp != NULL) {
1306 kstat_delete(xuio_ksp);
1312 xuio_stat_wbuf_copied(void)
1314 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1318 xuio_stat_wbuf_nocopy(void)
1320 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
1325 dmu_read_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size)
1328 int numbufs, i, err;
1329 xuio_t *xuio = NULL;
1332 * NB: we could do this block-at-a-time, but it's nice
1333 * to be reading in parallel.
1335 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1336 TRUE, FTAG, &numbufs, &dbp, 0);
1341 if (uio->uio_extflg == UIO_XUIO)
1342 xuio = (xuio_t *)uio;
1345 for (i = 0; i < numbufs; i++) {
1348 dmu_buf_t *db = dbp[i];
1352 bufoff = uio->uio_loffset - db->db_offset;
1353 tocpy = (int)MIN(db->db_size - bufoff, size);
1356 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
1357 arc_buf_t *dbuf_abuf = dbi->db_buf;
1358 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
1359 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
1361 uio->uio_resid -= tocpy;
1362 uio->uio_loffset += tocpy;
1365 if (abuf == dbuf_abuf)
1366 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
1368 XUIOSTAT_BUMP(xuiostat_rbuf_copied);
1371 err = uiomove((char *)db->db_data + bufoff, tocpy,
1374 err = vn_io_fault_uiomove((char *)db->db_data + bufoff,
1383 dmu_buf_rele_array(dbp, numbufs, FTAG);
1389 * Read 'size' bytes into the uio buffer.
1390 * From object zdb->db_object.
1391 * Starting at offset uio->uio_loffset.
1393 * If the caller already has a dbuf in the target object
1394 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1395 * because we don't have to find the dnode_t for the object.
1398 dmu_read_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size)
1400 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1409 err = dmu_read_uio_dnode(dn, uio, size);
1416 * Read 'size' bytes into the uio buffer.
1417 * From the specified object
1418 * Starting at offset uio->uio_loffset.
1421 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
1429 err = dnode_hold(os, object, FTAG, &dn);
1433 err = dmu_read_uio_dnode(dn, uio, size);
1435 dnode_rele(dn, FTAG);
1441 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
1448 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1449 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1453 for (i = 0; i < numbufs; i++) {
1456 dmu_buf_t *db = dbp[i];
1460 bufoff = uio->uio_loffset - db->db_offset;
1461 tocpy = (int)MIN(db->db_size - bufoff, size);
1463 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1465 if (tocpy == db->db_size)
1466 dmu_buf_will_fill(db, tx);
1468 dmu_buf_will_dirty(db, tx);
1472 * XXX uiomove could block forever (eg. nfs-backed
1473 * pages). There needs to be a uiolockdown() function
1474 * to lock the pages in memory, so that uiomove won't
1477 err = uiomove((char *)db->db_data + bufoff, tocpy,
1480 err = vn_io_fault_uiomove((char *)db->db_data + bufoff, tocpy,
1484 if (tocpy == db->db_size)
1485 dmu_buf_fill_done(db, tx);
1493 dmu_buf_rele_array(dbp, numbufs, FTAG);
1498 * Write 'size' bytes from the uio buffer.
1499 * To object zdb->db_object.
1500 * Starting at offset uio->uio_loffset.
1502 * If the caller already has a dbuf in the target object
1503 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1504 * because we don't have to find the dnode_t for the object.
1507 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
1510 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1519 err = dmu_write_uio_dnode(dn, uio, size, tx);
1526 * Write 'size' bytes from the uio buffer.
1527 * To the specified object.
1528 * Starting at offset uio->uio_loffset.
1531 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1540 err = dnode_hold(os, object, FTAG, &dn);
1544 err = dmu_write_uio_dnode(dn, uio, size, tx);
1546 dnode_rele(dn, FTAG);
1553 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1554 page_t *pp, dmu_tx_t *tx)
1563 err = dmu_buf_hold_array(os, object, offset, size,
1564 FALSE, FTAG, &numbufs, &dbp);
1568 for (i = 0; i < numbufs; i++) {
1569 int tocpy, copied, thiscpy;
1571 dmu_buf_t *db = dbp[i];
1575 ASSERT3U(db->db_size, >=, PAGESIZE);
1577 bufoff = offset - db->db_offset;
1578 tocpy = (int)MIN(db->db_size - bufoff, size);
1580 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1582 if (tocpy == db->db_size)
1583 dmu_buf_will_fill(db, tx);
1585 dmu_buf_will_dirty(db, tx);
1587 for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1588 ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff);
1589 thiscpy = MIN(PAGESIZE, tocpy - copied);
1590 va = zfs_map_page(pp, S_READ);
1591 bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1592 zfs_unmap_page(pp, va);
1597 if (tocpy == db->db_size)
1598 dmu_buf_fill_done(db, tx);
1603 dmu_buf_rele_array(dbp, numbufs, FTAG);
1607 #else /* !illumos */
1610 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1611 vm_page_t *ma, dmu_tx_t *tx)
1621 err = dmu_buf_hold_array(os, object, offset, size,
1622 FALSE, FTAG, &numbufs, &dbp);
1626 for (i = 0; i < numbufs; i++) {
1627 int tocpy, copied, thiscpy;
1629 dmu_buf_t *db = dbp[i];
1633 ASSERT3U(db->db_size, >=, PAGESIZE);
1635 bufoff = offset - db->db_offset;
1636 tocpy = (int)MIN(db->db_size - bufoff, size);
1638 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1640 if (tocpy == db->db_size)
1641 dmu_buf_will_fill(db, tx);
1643 dmu_buf_will_dirty(db, tx);
1645 for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1646 ASSERT3U(ptoa((*ma)->pindex), ==, db->db_offset + bufoff);
1647 thiscpy = MIN(PAGESIZE, tocpy - copied);
1648 va = zfs_map_page(*ma, &sf);
1649 bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1655 if (tocpy == db->db_size)
1656 dmu_buf_fill_done(db, tx);
1661 dmu_buf_rele_array(dbp, numbufs, FTAG);
1666 dmu_read_pages(objset_t *os, uint64_t object, vm_page_t *ma, int count,
1667 int *rbehind, int *rahead, int last_size)
1676 int bufoff, pgoff, tocpy;
1680 ASSERT3U(ma[0]->pindex + count - 1, ==, ma[count - 1]->pindex);
1681 ASSERT(last_size <= PAGE_SIZE);
1683 err = dmu_buf_hold_array(os, object, IDX_TO_OFF(ma[0]->pindex),
1684 IDX_TO_OFF(count - 1) + last_size, TRUE, FTAG, &numbufs, &dbp);
1689 IMPLY(last_size < PAGE_SIZE, *rahead == 0);
1690 if (dbp[0]->db_offset != 0 || numbufs > 1) {
1691 for (i = 0; i < numbufs; i++) {
1692 ASSERT(ISP2(dbp[i]->db_size));
1693 ASSERT((dbp[i]->db_offset % dbp[i]->db_size) == 0);
1694 ASSERT3U(dbp[i]->db_size, ==, dbp[0]->db_size);
1699 vmobj = ma[0]->object;
1700 zfs_vmobject_wlock(vmobj);
1703 for (i = 0; i < *rbehind; i++) {
1704 m = vm_page_grab(vmobj, ma[0]->pindex - 1 - i,
1705 VM_ALLOC_NORMAL | VM_ALLOC_NOWAIT | VM_ALLOC_NOBUSY);
1708 if (m->valid != 0) {
1709 ASSERT3U(m->valid, ==, VM_PAGE_BITS_ALL);
1712 ASSERT(m->dirty == 0);
1713 ASSERT(!pmap_page_is_mapped(m));
1715 ASSERT(db->db_size > PAGE_SIZE);
1716 bufoff = IDX_TO_OFF(m->pindex) % db->db_size;
1717 va = zfs_map_page(m, &sf);
1718 bcopy((char *)db->db_data + bufoff, va, PAGESIZE);
1720 m->valid = VM_PAGE_BITS_ALL;
1722 if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1723 vm_page_activate(m);
1725 vm_page_deactivate(m);
1730 bufoff = IDX_TO_OFF(ma[0]->pindex) % db->db_size;
1732 for (mi = 0, di = 0; mi < count && di < numbufs; ) {
1735 if (m != bogus_page) {
1736 vm_page_assert_xbusied(m);
1737 ASSERT(m->valid == 0);
1738 ASSERT(m->dirty == 0);
1739 ASSERT(!pmap_page_is_mapped(m));
1740 va = zfs_map_page(m, &sf);
1746 if (m != bogus_page) {
1747 ASSERT3U(IDX_TO_OFF(m->pindex) + pgoff, ==,
1748 db->db_offset + bufoff);
1752 * We do not need to clamp the copy size by the file
1753 * size as the last block is zero-filled beyond the
1754 * end of file anyway.
1756 tocpy = MIN(db->db_size - bufoff, PAGESIZE - pgoff);
1757 if (m != bogus_page)
1758 bcopy((char *)db->db_data + bufoff, va + pgoff, tocpy);
1761 ASSERT(pgoff <= PAGESIZE);
1762 if (pgoff == PAGESIZE) {
1763 if (m != bogus_page) {
1765 m->valid = VM_PAGE_BITS_ALL;
1773 ASSERT(bufoff <= db->db_size);
1774 if (bufoff == db->db_size) {
1775 ASSERT(di < numbufs);
1783 * Three possibilities:
1784 * - last requested page ends at a buffer boundary and , thus,
1785 * all pages and buffers have been iterated;
1786 * - all requested pages are filled, but the last buffer
1787 * has not been exhausted;
1788 * the read-ahead is possible only in this case;
1789 * - all buffers have been read, but the last page has not been
1791 * this is only possible if the file has only a single buffer
1792 * with a size that is not a multiple of the page size.
1795 ASSERT(di >= numbufs - 1);
1796 IMPLY(*rahead != 0, di == numbufs - 1);
1797 IMPLY(*rahead != 0, bufoff != 0);
1800 if (di == numbufs) {
1801 ASSERT(mi >= count - 1);
1802 ASSERT(*rahead == 0);
1803 IMPLY(pgoff == 0, mi == count);
1805 ASSERT(mi == count - 1);
1806 ASSERT((dbp[0]->db_size & PAGE_MASK) != 0);
1811 ASSERT(m != bogus_page);
1812 bzero(va + pgoff, PAGESIZE - pgoff);
1814 m->valid = VM_PAGE_BITS_ALL;
1817 for (i = 0; i < *rahead; i++) {
1818 m = vm_page_grab(vmobj, ma[count - 1]->pindex + 1 + i,
1819 VM_ALLOC_NORMAL | VM_ALLOC_NOWAIT | VM_ALLOC_NOBUSY);
1822 if (m->valid != 0) {
1823 ASSERT3U(m->valid, ==, VM_PAGE_BITS_ALL);
1826 ASSERT(m->dirty == 0);
1827 ASSERT(!pmap_page_is_mapped(m));
1829 ASSERT(db->db_size > PAGE_SIZE);
1830 bufoff = IDX_TO_OFF(m->pindex) % db->db_size;
1831 tocpy = MIN(db->db_size - bufoff, PAGESIZE);
1832 va = zfs_map_page(m, &sf);
1833 bcopy((char *)db->db_data + bufoff, va, tocpy);
1834 if (tocpy < PAGESIZE) {
1835 ASSERT(i == *rahead - 1);
1836 ASSERT((db->db_size & PAGE_MASK) != 0);
1837 bzero(va + tocpy, PAGESIZE - tocpy);
1840 m->valid = VM_PAGE_BITS_ALL;
1842 if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1843 vm_page_activate(m);
1845 vm_page_deactivate(m);
1849 zfs_vmobject_wunlock(vmobj);
1851 dmu_buf_rele_array(dbp, numbufs, FTAG);
1854 #endif /* illumos */
1855 #endif /* _KERNEL */
1858 * Allocate a loaned anonymous arc buffer.
1861 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1863 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1865 return (arc_loan_buf(db->db_objset->os_spa, B_FALSE, size));
1869 * Free a loaned arc buffer.
1872 dmu_return_arcbuf(arc_buf_t *buf)
1874 arc_return_buf(buf, FTAG);
1875 arc_buf_destroy(buf, FTAG);
1879 * When possible directly assign passed loaned arc buffer to a dbuf.
1880 * If this is not possible copy the contents of passed arc buf via
1884 dmu_assign_arcbuf_dnode(dnode_t *dn, uint64_t offset, arc_buf_t *buf,
1888 uint32_t blksz = (uint32_t)arc_buf_lsize(buf);
1891 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1892 blkid = dbuf_whichblock(dn, 0, offset);
1893 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1894 rw_exit(&dn->dn_struct_rwlock);
1897 * We can only assign if the offset is aligned, the arc buf is the
1898 * same size as the dbuf, and the dbuf is not metadata.
1900 if (offset == db->db.db_offset && blksz == db->db.db_size) {
1902 curthread->td_ru.ru_oublock++;
1906 racct_add_force(curproc, RACCT_WRITEBPS, blksz);
1907 racct_add_force(curproc, RACCT_WRITEIOPS, 1);
1908 PROC_UNLOCK(curproc);
1911 #endif /* _KERNEL */
1912 dbuf_assign_arcbuf(db, buf, tx);
1913 dbuf_rele(db, FTAG);
1918 /* compressed bufs must always be assignable to their dbuf */
1919 ASSERT3U(arc_get_compression(buf), ==, ZIO_COMPRESS_OFF);
1920 ASSERT(!(buf->b_flags & ARC_BUF_FLAG_COMPRESSED));
1923 object = dn->dn_object;
1925 dbuf_rele(db, FTAG);
1926 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1927 dmu_return_arcbuf(buf);
1928 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1933 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1936 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1938 DB_DNODE_ENTER(dbuf);
1939 dmu_assign_arcbuf_dnode(DB_DNODE(dbuf), offset, buf, tx);
1940 DB_DNODE_EXIT(dbuf);
1944 dbuf_dirty_record_t *dsa_dr;
1945 dmu_sync_cb_t *dsa_done;
1952 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1954 dmu_sync_arg_t *dsa = varg;
1955 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1956 blkptr_t *bp = zio->io_bp;
1958 if (zio->io_error == 0) {
1959 if (BP_IS_HOLE(bp)) {
1961 * A block of zeros may compress to a hole, but the
1962 * block size still needs to be known for replay.
1964 BP_SET_LSIZE(bp, db->db_size);
1965 } else if (!BP_IS_EMBEDDED(bp)) {
1966 ASSERT(BP_GET_LEVEL(bp) == 0);
1973 dmu_sync_late_arrival_ready(zio_t *zio)
1975 dmu_sync_ready(zio, NULL, zio->io_private);
1980 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1982 dmu_sync_arg_t *dsa = varg;
1983 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1984 dmu_buf_impl_t *db = dr->dr_dbuf;
1986 mutex_enter(&db->db_mtx);
1987 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1988 if (zio->io_error == 0) {
1989 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1990 if (dr->dt.dl.dr_nopwrite) {
1991 blkptr_t *bp = zio->io_bp;
1992 blkptr_t *bp_orig = &zio->io_bp_orig;
1993 uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
1995 ASSERT(BP_EQUAL(bp, bp_orig));
1996 VERIFY(BP_EQUAL(bp, db->db_blkptr));
1997 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1998 ASSERT(zio_checksum_table[chksum].ci_flags &
1999 ZCHECKSUM_FLAG_NOPWRITE);
2001 dr->dt.dl.dr_overridden_by = *zio->io_bp;
2002 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
2003 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
2006 * Old style holes are filled with all zeros, whereas
2007 * new-style holes maintain their lsize, type, level,
2008 * and birth time (see zio_write_compress). While we
2009 * need to reset the BP_SET_LSIZE() call that happened
2010 * in dmu_sync_ready for old style holes, we do *not*
2011 * want to wipe out the information contained in new
2012 * style holes. Thus, only zero out the block pointer if
2013 * it's an old style hole.
2015 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) &&
2016 dr->dt.dl.dr_overridden_by.blk_birth == 0)
2017 BP_ZERO(&dr->dt.dl.dr_overridden_by);
2019 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
2021 cv_broadcast(&db->db_changed);
2022 mutex_exit(&db->db_mtx);
2024 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
2026 kmem_free(dsa, sizeof (*dsa));
2030 dmu_sync_late_arrival_done(zio_t *zio)
2032 blkptr_t *bp = zio->io_bp;
2033 dmu_sync_arg_t *dsa = zio->io_private;
2034 blkptr_t *bp_orig = &zio->io_bp_orig;
2036 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
2037 ASSERT(!(zio->io_flags & ZIO_FLAG_NOPWRITE));
2038 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
2039 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
2040 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
2041 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
2044 dmu_tx_commit(dsa->dsa_tx);
2046 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
2048 abd_put(zio->io_abd);
2049 kmem_free(dsa, sizeof (*dsa));
2053 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
2054 zio_prop_t *zp, zbookmark_phys_t *zb)
2056 dmu_sync_arg_t *dsa;
2059 tx = dmu_tx_create(os);
2060 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
2061 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
2063 /* Make zl_get_data do txg_waited_synced() */
2064 return (SET_ERROR(EIO));
2068 * In order to prevent the zgd's lwb from being free'd prior to
2069 * dmu_sync_late_arrival_done() being called, we have to ensure
2070 * the lwb's "max txg" takes this tx's txg into account.
2072 zil_lwb_add_txg(zgd->zgd_lwb, dmu_tx_get_txg(tx));
2074 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
2076 dsa->dsa_done = done;
2081 * Since we are currently syncing this txg, it's nontrivial to
2082 * determine what BP to nopwrite against, so we disable nopwrite.
2084 * When syncing, the db_blkptr is initially the BP of the previous
2085 * txg. We can not nopwrite against it because it will be changed
2086 * (this is similar to the non-late-arrival case where the dbuf is
2087 * dirty in a future txg).
2089 * Then dbuf_write_ready() sets bp_blkptr to the location we will write.
2090 * We can not nopwrite against it because although the BP will not
2091 * (typically) be changed, the data has not yet been persisted to this
2094 * Finally, when dbuf_write_done() is called, it is theoretically
2095 * possible to always nopwrite, because the data that was written in
2096 * this txg is the same data that we are trying to write. However we
2097 * would need to check that this dbuf is not dirty in any future
2098 * txg's (as we do in the normal dmu_sync() path). For simplicity, we
2099 * don't nopwrite in this case.
2101 zp->zp_nopwrite = B_FALSE;
2103 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
2104 abd_get_from_buf(zgd->zgd_db->db_data, zgd->zgd_db->db_size),
2105 zgd->zgd_db->db_size, zgd->zgd_db->db_size, zp,
2106 dmu_sync_late_arrival_ready, NULL, NULL, dmu_sync_late_arrival_done,
2107 dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
2113 * Intent log support: sync the block associated with db to disk.
2114 * N.B. and XXX: the caller is responsible for making sure that the
2115 * data isn't changing while dmu_sync() is writing it.
2119 * EEXIST: this txg has already been synced, so there's nothing to do.
2120 * The caller should not log the write.
2122 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
2123 * The caller should not log the write.
2125 * EALREADY: this block is already in the process of being synced.
2126 * The caller should track its progress (somehow).
2128 * EIO: could not do the I/O.
2129 * The caller should do a txg_wait_synced().
2131 * 0: the I/O has been initiated.
2132 * The caller should log this blkptr in the done callback.
2133 * It is possible that the I/O will fail, in which case
2134 * the error will be reported to the done callback and
2135 * propagated to pio from zio_done().
2138 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
2140 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
2141 objset_t *os = db->db_objset;
2142 dsl_dataset_t *ds = os->os_dsl_dataset;
2143 dbuf_dirty_record_t *dr;
2144 dmu_sync_arg_t *dsa;
2145 zbookmark_phys_t zb;
2149 ASSERT(pio != NULL);
2152 SET_BOOKMARK(&zb, ds->ds_object,
2153 db->db.db_object, db->db_level, db->db_blkid);
2157 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
2161 * If we're frozen (running ziltest), we always need to generate a bp.
2163 if (txg > spa_freeze_txg(os->os_spa))
2164 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
2167 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
2168 * and us. If we determine that this txg is not yet syncing,
2169 * but it begins to sync a moment later, that's OK because the
2170 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
2172 mutex_enter(&db->db_mtx);
2174 if (txg <= spa_last_synced_txg(os->os_spa)) {
2176 * This txg has already synced. There's nothing to do.
2178 mutex_exit(&db->db_mtx);
2179 return (SET_ERROR(EEXIST));
2182 if (txg <= spa_syncing_txg(os->os_spa)) {
2184 * This txg is currently syncing, so we can't mess with
2185 * the dirty record anymore; just write a new log block.
2187 mutex_exit(&db->db_mtx);
2188 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
2191 dr = db->db_last_dirty;
2192 while (dr && dr->dr_txg != txg)
2197 * There's no dr for this dbuf, so it must have been freed.
2198 * There's no need to log writes to freed blocks, so we're done.
2200 mutex_exit(&db->db_mtx);
2201 return (SET_ERROR(ENOENT));
2204 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg);
2206 if (db->db_blkptr != NULL) {
2208 * We need to fill in zgd_bp with the current blkptr so that
2209 * the nopwrite code can check if we're writing the same
2210 * data that's already on disk. We can only nopwrite if we
2211 * are sure that after making the copy, db_blkptr will not
2212 * change until our i/o completes. We ensure this by
2213 * holding the db_mtx, and only allowing nopwrite if the
2214 * block is not already dirty (see below). This is verified
2215 * by dmu_sync_done(), which VERIFYs that the db_blkptr has
2218 *zgd->zgd_bp = *db->db_blkptr;
2222 * Assume the on-disk data is X, the current syncing data (in
2223 * txg - 1) is Y, and the current in-memory data is Z (currently
2226 * We usually want to perform a nopwrite if X and Z are the
2227 * same. However, if Y is different (i.e. the BP is going to
2228 * change before this write takes effect), then a nopwrite will
2229 * be incorrect - we would override with X, which could have
2230 * been freed when Y was written.
2232 * (Note that this is not a concern when we are nop-writing from
2233 * syncing context, because X and Y must be identical, because
2234 * all previous txgs have been synced.)
2236 * Therefore, we disable nopwrite if the current BP could change
2237 * before this TXG. There are two ways it could change: by
2238 * being dirty (dr_next is non-NULL), or by being freed
2239 * (dnode_block_freed()). This behavior is verified by
2240 * zio_done(), which VERIFYs that the override BP is identical
2241 * to the on-disk BP.
2245 if (dr->dr_next != NULL || dnode_block_freed(dn, db->db_blkid))
2246 zp.zp_nopwrite = B_FALSE;
2249 ASSERT(dr->dr_txg == txg);
2250 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
2251 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
2253 * We have already issued a sync write for this buffer,
2254 * or this buffer has already been synced. It could not
2255 * have been dirtied since, or we would have cleared the state.
2257 mutex_exit(&db->db_mtx);
2258 return (SET_ERROR(EALREADY));
2261 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
2262 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
2263 mutex_exit(&db->db_mtx);
2265 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
2267 dsa->dsa_done = done;
2271 zio_nowait(arc_write(pio, os->os_spa, txg,
2272 zgd->zgd_bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db),
2273 &zp, dmu_sync_ready, NULL, NULL, dmu_sync_done, dsa,
2274 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb));
2280 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
2286 err = dnode_hold(os, object, FTAG, &dn);
2289 err = dnode_set_blksz(dn, size, ibs, tx);
2290 dnode_rele(dn, FTAG);
2295 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
2301 * Send streams include each object's checksum function. This
2302 * check ensures that the receiving system can understand the
2303 * checksum function transmitted.
2305 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS);
2307 VERIFY0(dnode_hold(os, object, FTAG, &dn));
2308 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS);
2309 dn->dn_checksum = checksum;
2310 dnode_setdirty(dn, tx);
2311 dnode_rele(dn, FTAG);
2315 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
2321 * Send streams include each object's compression function. This
2322 * check ensures that the receiving system can understand the
2323 * compression function transmitted.
2325 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS);
2327 VERIFY0(dnode_hold(os, object, FTAG, &dn));
2328 dn->dn_compress = compress;
2329 dnode_setdirty(dn, tx);
2330 dnode_rele(dn, FTAG);
2333 int zfs_mdcomp_disable = 0;
2334 SYSCTL_INT(_vfs_zfs, OID_AUTO, mdcomp_disable, CTLFLAG_RWTUN,
2335 &zfs_mdcomp_disable, 0, "Disable metadata compression");
2338 * When the "redundant_metadata" property is set to "most", only indirect
2339 * blocks of this level and higher will have an additional ditto block.
2341 int zfs_redundant_metadata_most_ditto_level = 2;
2344 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
2346 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
2347 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
2349 enum zio_checksum checksum = os->os_checksum;
2350 enum zio_compress compress = os->os_compress;
2351 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
2352 boolean_t dedup = B_FALSE;
2353 boolean_t nopwrite = B_FALSE;
2354 boolean_t dedup_verify = os->os_dedup_verify;
2355 int copies = os->os_copies;
2358 * We maintain different write policies for each of the following
2361 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
2362 * 3. all other level 0 blocks
2365 if (zfs_mdcomp_disable) {
2366 compress = ZIO_COMPRESS_EMPTY;
2369 * XXX -- we should design a compression algorithm
2370 * that specializes in arrays of bps.
2372 compress = zio_compress_select(os->os_spa,
2373 ZIO_COMPRESS_ON, ZIO_COMPRESS_ON);
2377 * Metadata always gets checksummed. If the data
2378 * checksum is multi-bit correctable, and it's not a
2379 * ZBT-style checksum, then it's suitable for metadata
2380 * as well. Otherwise, the metadata checksum defaults
2383 if (!(zio_checksum_table[checksum].ci_flags &
2384 ZCHECKSUM_FLAG_METADATA) ||
2385 (zio_checksum_table[checksum].ci_flags &
2386 ZCHECKSUM_FLAG_EMBEDDED))
2387 checksum = ZIO_CHECKSUM_FLETCHER_4;
2389 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL ||
2390 (os->os_redundant_metadata ==
2391 ZFS_REDUNDANT_METADATA_MOST &&
2392 (level >= zfs_redundant_metadata_most_ditto_level ||
2393 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))))
2395 } else if (wp & WP_NOFILL) {
2399 * If we're writing preallocated blocks, we aren't actually
2400 * writing them so don't set any policy properties. These
2401 * blocks are currently only used by an external subsystem
2402 * outside of zfs (i.e. dump) and not written by the zio
2405 compress = ZIO_COMPRESS_OFF;
2406 checksum = ZIO_CHECKSUM_NOPARITY;
2408 compress = zio_compress_select(os->os_spa, dn->dn_compress,
2411 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
2412 zio_checksum_select(dn->dn_checksum, checksum) :
2416 * Determine dedup setting. If we are in dmu_sync(),
2417 * we won't actually dedup now because that's all
2418 * done in syncing context; but we do want to use the
2419 * dedup checkum. If the checksum is not strong
2420 * enough to ensure unique signatures, force
2423 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
2424 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
2425 if (!(zio_checksum_table[checksum].ci_flags &
2426 ZCHECKSUM_FLAG_DEDUP))
2427 dedup_verify = B_TRUE;
2431 * Enable nopwrite if we have secure enough checksum
2432 * algorithm (see comment in zio_nop_write) and
2433 * compression is enabled. We don't enable nopwrite if
2434 * dedup is enabled as the two features are mutually
2437 nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags &
2438 ZCHECKSUM_FLAG_NOPWRITE) &&
2439 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
2442 zp->zp_checksum = checksum;
2443 zp->zp_compress = compress;
2444 ASSERT3U(zp->zp_compress, !=, ZIO_COMPRESS_INHERIT);
2446 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
2447 zp->zp_level = level;
2448 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
2449 zp->zp_dedup = dedup;
2450 zp->zp_dedup_verify = dedup && dedup_verify;
2451 zp->zp_nopwrite = nopwrite;
2455 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
2461 * Sync any current changes before
2462 * we go trundling through the block pointers.
2464 err = dmu_object_wait_synced(os, object);
2469 err = dnode_hold(os, object, FTAG, &dn);
2474 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
2475 dnode_rele(dn, FTAG);
2481 * Given the ZFS object, if it contains any dirty nodes
2482 * this function flushes all dirty blocks to disk. This
2483 * ensures the DMU object info is updated. A more efficient
2484 * future version might just find the TXG with the maximum
2485 * ID and wait for that to be synced.
2488 dmu_object_wait_synced(objset_t *os, uint64_t object)
2493 error = dnode_hold(os, object, FTAG, &dn);
2498 for (i = 0; i < TXG_SIZE; i++) {
2499 if (list_link_active(&dn->dn_dirty_link[i])) {
2503 dnode_rele(dn, FTAG);
2504 if (i != TXG_SIZE) {
2505 txg_wait_synced(dmu_objset_pool(os), 0);
2512 __dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2514 dnode_phys_t *dnp = dn->dn_phys;
2516 doi->doi_data_block_size = dn->dn_datablksz;
2517 doi->doi_metadata_block_size = dn->dn_indblkshift ?
2518 1ULL << dn->dn_indblkshift : 0;
2519 doi->doi_type = dn->dn_type;
2520 doi->doi_bonus_type = dn->dn_bonustype;
2521 doi->doi_bonus_size = dn->dn_bonuslen;
2522 doi->doi_dnodesize = dn->dn_num_slots << DNODE_SHIFT;
2523 doi->doi_indirection = dn->dn_nlevels;
2524 doi->doi_checksum = dn->dn_checksum;
2525 doi->doi_compress = dn->dn_compress;
2526 doi->doi_nblkptr = dn->dn_nblkptr;
2527 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
2528 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
2529 doi->doi_fill_count = 0;
2530 for (int i = 0; i < dnp->dn_nblkptr; i++)
2531 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]);
2535 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2537 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2538 mutex_enter(&dn->dn_mtx);
2540 __dmu_object_info_from_dnode(dn, doi);
2542 mutex_exit(&dn->dn_mtx);
2543 rw_exit(&dn->dn_struct_rwlock);
2547 * Get information on a DMU object.
2548 * If doi is NULL, just indicates whether the object exists.
2551 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
2554 int err = dnode_hold(os, object, FTAG, &dn);
2560 dmu_object_info_from_dnode(dn, doi);
2562 dnode_rele(dn, FTAG);
2567 * As above, but faster; can be used when you have a held dbuf in hand.
2570 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
2572 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2575 dmu_object_info_from_dnode(DB_DNODE(db), doi);
2580 * Faster still when you only care about the size.
2581 * This is specifically optimized for zfs_getattr().
2584 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
2585 u_longlong_t *nblk512)
2587 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2593 *blksize = dn->dn_datablksz;
2594 /* add in number of slots used for the dnode itself */
2595 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
2596 SPA_MINBLOCKSHIFT) + dn->dn_num_slots;
2601 dmu_object_dnsize_from_db(dmu_buf_t *db_fake, int *dnsize)
2603 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2608 *dnsize = dn->dn_num_slots << DNODE_SHIFT;
2613 byteswap_uint64_array(void *vbuf, size_t size)
2615 uint64_t *buf = vbuf;
2616 size_t count = size >> 3;
2619 ASSERT((size & 7) == 0);
2621 for (i = 0; i < count; i++)
2622 buf[i] = BSWAP_64(buf[i]);
2626 byteswap_uint32_array(void *vbuf, size_t size)
2628 uint32_t *buf = vbuf;
2629 size_t count = size >> 2;
2632 ASSERT((size & 3) == 0);
2634 for (i = 0; i < count; i++)
2635 buf[i] = BSWAP_32(buf[i]);
2639 byteswap_uint16_array(void *vbuf, size_t size)
2641 uint16_t *buf = vbuf;
2642 size_t count = size >> 1;
2645 ASSERT((size & 1) == 0);
2647 for (i = 0; i < count; i++)
2648 buf[i] = BSWAP_16(buf[i]);
2653 byteswap_uint8_array(void *vbuf, size_t size)
2667 zio_compress_init();
2676 arc_fini(); /* arc depends on l2arc, so arc must go first */
2679 zio_compress_fini();