4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2017 by Delphix. All rights reserved.
24 * Copyright (c) 2019 Datto Inc.
26 /* Copyright (c) 2013 by Saso Kiselkov. All rights reserved. */
27 /* Copyright (c) 2013, Joyent, Inc. All rights reserved. */
28 /* Copyright 2016 Nexenta Systems, Inc. All rights reserved. */
31 #include <sys/dmu_impl.h>
32 #include <sys/dmu_tx.h>
34 #include <sys/dnode.h>
35 #include <sys/zfs_context.h>
36 #include <sys/dmu_objset.h>
37 #include <sys/dmu_traverse.h>
38 #include <sys/dsl_dataset.h>
39 #include <sys/dsl_dir.h>
40 #include <sys/dsl_pool.h>
41 #include <sys/dsl_synctask.h>
42 #include <sys/dsl_prop.h>
43 #include <sys/dmu_zfetch.h>
44 #include <sys/zfs_ioctl.h>
46 #include <sys/zio_checksum.h>
47 #include <sys/zio_compress.h>
49 #include <sys/zfeature.h>
52 #include <sys/racct.h>
54 #include <sys/zfs_znode.h>
58 * Enable/disable nopwrite feature.
60 int zfs_nopwrite_enabled = 1;
61 SYSCTL_DECL(_vfs_zfs);
62 SYSCTL_INT(_vfs_zfs, OID_AUTO, nopwrite_enabled, CTLFLAG_RDTUN,
63 &zfs_nopwrite_enabled, 0, "Enable nopwrite feature");
66 * Tunable to control percentage of dirtied L1 blocks from frees allowed into
67 * one TXG. After this threshold is crossed, additional dirty blocks from frees
68 * will wait until the next TXG.
69 * A value of zero will disable this throttle.
71 uint32_t zfs_per_txg_dirty_frees_percent = 5;
72 SYSCTL_INT(_vfs_zfs, OID_AUTO, per_txg_dirty_frees_percent, CTLFLAG_RWTUN,
73 &zfs_per_txg_dirty_frees_percent, 0,
74 "Percentage of dirtied indirect blocks from frees allowed in one txg");
77 * This can be used for testing, to ensure that certain actions happen
78 * while in the middle of a remap (which might otherwise complete too
81 int zfs_object_remap_one_indirect_delay_ticks = 0;
83 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
84 { DMU_BSWAP_UINT8, TRUE, FALSE, "unallocated" },
85 { DMU_BSWAP_ZAP, TRUE, TRUE, "object directory" },
86 { DMU_BSWAP_UINT64, TRUE, TRUE, "object array" },
87 { DMU_BSWAP_UINT8, TRUE, FALSE, "packed nvlist" },
88 { DMU_BSWAP_UINT64, TRUE, FALSE, "packed nvlist size" },
89 { DMU_BSWAP_UINT64, TRUE, FALSE, "bpobj" },
90 { DMU_BSWAP_UINT64, TRUE, FALSE, "bpobj header" },
91 { DMU_BSWAP_UINT64, TRUE, FALSE, "SPA space map header" },
92 { DMU_BSWAP_UINT64, TRUE, FALSE, "SPA space map" },
93 { DMU_BSWAP_UINT64, TRUE, FALSE, "ZIL intent log" },
94 { DMU_BSWAP_DNODE, TRUE, FALSE, "DMU dnode" },
95 { DMU_BSWAP_OBJSET, TRUE, TRUE, "DMU objset" },
96 { DMU_BSWAP_UINT64, TRUE, TRUE, "DSL directory" },
97 { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL directory child map" },
98 { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL dataset snap map" },
99 { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL props" },
100 { DMU_BSWAP_UINT64, TRUE, TRUE, "DSL dataset" },
101 { DMU_BSWAP_ZNODE, TRUE, FALSE, "ZFS znode" },
102 { DMU_BSWAP_OLDACL, TRUE, FALSE, "ZFS V0 ACL" },
103 { DMU_BSWAP_UINT8, FALSE, FALSE, "ZFS plain file" },
104 { DMU_BSWAP_ZAP, TRUE, FALSE, "ZFS directory" },
105 { DMU_BSWAP_ZAP, TRUE, FALSE, "ZFS master node" },
106 { DMU_BSWAP_ZAP, TRUE, FALSE, "ZFS delete queue" },
107 { DMU_BSWAP_UINT8, FALSE, FALSE, "zvol object" },
108 { DMU_BSWAP_ZAP, TRUE, FALSE, "zvol prop" },
109 { DMU_BSWAP_UINT8, FALSE, FALSE, "other uint8[]" },
110 { DMU_BSWAP_UINT64, FALSE, FALSE, "other uint64[]" },
111 { DMU_BSWAP_ZAP, TRUE, FALSE, "other ZAP" },
112 { DMU_BSWAP_ZAP, TRUE, FALSE, "persistent error log" },
113 { DMU_BSWAP_UINT8, TRUE, FALSE, "SPA history" },
114 { DMU_BSWAP_UINT64, TRUE, FALSE, "SPA history offsets" },
115 { DMU_BSWAP_ZAP, TRUE, TRUE, "Pool properties" },
116 { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL permissions" },
117 { DMU_BSWAP_ACL, TRUE, FALSE, "ZFS ACL" },
118 { DMU_BSWAP_UINT8, TRUE, FALSE, "ZFS SYSACL" },
119 { DMU_BSWAP_UINT8, TRUE, FALSE, "FUID table" },
120 { DMU_BSWAP_UINT64, TRUE, FALSE, "FUID table size" },
121 { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL dataset next clones" },
122 { DMU_BSWAP_ZAP, TRUE, FALSE, "scan work queue" },
123 { DMU_BSWAP_ZAP, TRUE, FALSE, "ZFS user/group used" },
124 { DMU_BSWAP_ZAP, TRUE, FALSE, "ZFS user/group quota" },
125 { DMU_BSWAP_ZAP, TRUE, TRUE, "snapshot refcount tags" },
126 { DMU_BSWAP_ZAP, TRUE, FALSE, "DDT ZAP algorithm" },
127 { DMU_BSWAP_ZAP, TRUE, FALSE, "DDT statistics" },
128 { DMU_BSWAP_UINT8, TRUE, FALSE, "System attributes" },
129 { DMU_BSWAP_ZAP, TRUE, FALSE, "SA master node" },
130 { DMU_BSWAP_ZAP, TRUE, FALSE, "SA attr registration" },
131 { DMU_BSWAP_ZAP, TRUE, FALSE, "SA attr layouts" },
132 { DMU_BSWAP_ZAP, TRUE, FALSE, "scan translations" },
133 { DMU_BSWAP_UINT8, FALSE, FALSE, "deduplicated block" },
134 { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL deadlist map" },
135 { DMU_BSWAP_UINT64, TRUE, TRUE, "DSL deadlist map hdr" },
136 { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL dir clones" },
137 { DMU_BSWAP_UINT64, TRUE, FALSE, "bpobj subobj" }
140 const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
141 { byteswap_uint8_array, "uint8" },
142 { byteswap_uint16_array, "uint16" },
143 { byteswap_uint32_array, "uint32" },
144 { byteswap_uint64_array, "uint64" },
145 { zap_byteswap, "zap" },
146 { dnode_buf_byteswap, "dnode" },
147 { dmu_objset_byteswap, "objset" },
148 { zfs_znode_byteswap, "znode" },
149 { zfs_oldacl_byteswap, "oldacl" },
150 { zfs_acl_byteswap, "acl" }
154 dmu_buf_hold_noread_by_dnode(dnode_t *dn, uint64_t offset,
155 void *tag, dmu_buf_t **dbp)
160 blkid = dbuf_whichblock(dn, 0, offset);
161 rw_enter(&dn->dn_struct_rwlock, RW_READER);
162 db = dbuf_hold(dn, blkid, tag);
163 rw_exit(&dn->dn_struct_rwlock);
167 return (SET_ERROR(EIO));
174 dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset,
175 void *tag, dmu_buf_t **dbp)
182 err = dnode_hold(os, object, FTAG, &dn);
185 blkid = dbuf_whichblock(dn, 0, offset);
186 rw_enter(&dn->dn_struct_rwlock, RW_READER);
187 db = dbuf_hold(dn, blkid, tag);
188 rw_exit(&dn->dn_struct_rwlock);
189 dnode_rele(dn, FTAG);
193 return (SET_ERROR(EIO));
201 dmu_buf_hold_by_dnode(dnode_t *dn, uint64_t offset,
202 void *tag, dmu_buf_t **dbp, int flags)
205 int db_flags = DB_RF_CANFAIL;
207 if (flags & DMU_READ_NO_PREFETCH)
208 db_flags |= DB_RF_NOPREFETCH;
210 err = dmu_buf_hold_noread_by_dnode(dn, offset, tag, dbp);
212 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
213 err = dbuf_read(db, NULL, db_flags);
224 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
225 void *tag, dmu_buf_t **dbp, int flags)
228 int db_flags = DB_RF_CANFAIL;
230 if (flags & DMU_READ_NO_PREFETCH)
231 db_flags |= DB_RF_NOPREFETCH;
233 err = dmu_buf_hold_noread(os, object, offset, tag, dbp);
235 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
236 err = dbuf_read(db, NULL, db_flags);
249 return (DN_OLD_MAX_BONUSLEN);
253 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx)
255 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
262 if (dn->dn_bonus != db) {
263 error = SET_ERROR(EINVAL);
264 } else if (newsize < 0 || newsize > db_fake->db_size) {
265 error = SET_ERROR(EINVAL);
267 dnode_setbonuslen(dn, newsize, tx);
276 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
278 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
285 if (!DMU_OT_IS_VALID(type)) {
286 error = SET_ERROR(EINVAL);
287 } else if (dn->dn_bonus != db) {
288 error = SET_ERROR(EINVAL);
290 dnode_setbonus_type(dn, type, tx);
299 dmu_get_bonustype(dmu_buf_t *db_fake)
301 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
303 dmu_object_type_t type;
307 type = dn->dn_bonustype;
314 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
319 error = dnode_hold(os, object, FTAG, &dn);
320 dbuf_rm_spill(dn, tx);
321 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
322 dnode_rm_spill(dn, tx);
323 rw_exit(&dn->dn_struct_rwlock);
324 dnode_rele(dn, FTAG);
329 * returns ENOENT, EIO, or 0.
332 dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp)
338 error = dnode_hold(os, object, FTAG, &dn);
342 rw_enter(&dn->dn_struct_rwlock, RW_READER);
343 if (dn->dn_bonus == NULL) {
344 rw_exit(&dn->dn_struct_rwlock);
345 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
346 if (dn->dn_bonus == NULL)
347 dbuf_create_bonus(dn);
351 /* as long as the bonus buf is held, the dnode will be held */
352 if (refcount_add(&db->db_holds, tag) == 1) {
353 VERIFY(dnode_add_ref(dn, db));
354 atomic_inc_32(&dn->dn_dbufs_count);
358 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
359 * hold and incrementing the dbuf count to ensure that dnode_move() sees
360 * a dnode hold for every dbuf.
362 rw_exit(&dn->dn_struct_rwlock);
364 dnode_rele(dn, FTAG);
366 VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH));
373 * returns ENOENT, EIO, or 0.
375 * This interface will allocate a blank spill dbuf when a spill blk
376 * doesn't already exist on the dnode.
378 * if you only want to find an already existing spill db, then
379 * dmu_spill_hold_existing() should be used.
382 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp)
384 dmu_buf_impl_t *db = NULL;
387 if ((flags & DB_RF_HAVESTRUCT) == 0)
388 rw_enter(&dn->dn_struct_rwlock, RW_READER);
390 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
392 if ((flags & DB_RF_HAVESTRUCT) == 0)
393 rw_exit(&dn->dn_struct_rwlock);
396 err = dbuf_read(db, NULL, flags);
405 dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
407 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
414 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
415 err = SET_ERROR(EINVAL);
417 rw_enter(&dn->dn_struct_rwlock, RW_READER);
419 if (!dn->dn_have_spill) {
420 err = SET_ERROR(ENOENT);
422 err = dmu_spill_hold_by_dnode(dn,
423 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
426 rw_exit(&dn->dn_struct_rwlock);
434 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
436 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
442 err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp);
449 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
450 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
451 * and can induce severe lock contention when writing to several files
452 * whose dnodes are in the same block.
455 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
456 boolean_t read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
459 uint64_t blkid, nblks, i;
464 ASSERT(length <= DMU_MAX_ACCESS);
467 * Note: We directly notify the prefetch code of this read, so that
468 * we can tell it about the multi-block read. dbuf_read() only knows
469 * about the one block it is accessing.
471 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT |
474 rw_enter(&dn->dn_struct_rwlock, RW_READER);
475 if (dn->dn_datablkshift) {
476 int blkshift = dn->dn_datablkshift;
477 nblks = (P2ROUNDUP(offset + length, 1ULL << blkshift) -
478 P2ALIGN(offset, 1ULL << blkshift)) >> blkshift;
480 if (offset + length > dn->dn_datablksz) {
481 zfs_panic_recover("zfs: accessing past end of object "
482 "%llx/%llx (size=%u access=%llu+%llu)",
483 (longlong_t)dn->dn_objset->
484 os_dsl_dataset->ds_object,
485 (longlong_t)dn->dn_object, dn->dn_datablksz,
486 (longlong_t)offset, (longlong_t)length);
487 rw_exit(&dn->dn_struct_rwlock);
488 return (SET_ERROR(EIO));
492 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
494 #if defined(_KERNEL) && defined(RACCT)
495 if (racct_enable && !read) {
497 racct_add_force(curproc, RACCT_WRITEBPS, length);
498 racct_add_force(curproc, RACCT_WRITEIOPS, nblks);
499 PROC_UNLOCK(curproc);
503 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
504 blkid = dbuf_whichblock(dn, 0, offset);
505 for (i = 0; i < nblks; i++) {
506 dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag);
508 rw_exit(&dn->dn_struct_rwlock);
509 dmu_buf_rele_array(dbp, nblks, tag);
511 return (SET_ERROR(EIO));
514 /* initiate async i/o */
516 (void) dbuf_read(db, zio, dbuf_flags);
519 curthread->td_ru.ru_oublock++;
524 if ((flags & DMU_READ_NO_PREFETCH) == 0 &&
525 DNODE_META_IS_CACHEABLE(dn) && length <= zfetch_array_rd_sz) {
526 dmu_zfetch(&dn->dn_zfetch, blkid, nblks,
527 read && DNODE_IS_CACHEABLE(dn));
529 rw_exit(&dn->dn_struct_rwlock);
531 /* wait for async i/o */
534 dmu_buf_rele_array(dbp, nblks, tag);
538 /* wait for other io to complete */
540 for (i = 0; i < nblks; i++) {
541 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
542 mutex_enter(&db->db_mtx);
543 while (db->db_state == DB_READ ||
544 db->db_state == DB_FILL)
545 cv_wait(&db->db_changed, &db->db_mtx);
546 if (db->db_state == DB_UNCACHED)
547 err = SET_ERROR(EIO);
548 mutex_exit(&db->db_mtx);
550 dmu_buf_rele_array(dbp, nblks, tag);
562 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
563 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
568 err = dnode_hold(os, object, FTAG, &dn);
572 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
573 numbufsp, dbpp, DMU_READ_PREFETCH);
575 dnode_rele(dn, FTAG);
581 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
582 uint64_t length, boolean_t read, void *tag, int *numbufsp,
585 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
591 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
592 numbufsp, dbpp, DMU_READ_PREFETCH);
599 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
602 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
607 for (i = 0; i < numbufs; i++) {
609 dbuf_rele(dbp[i], tag);
612 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
616 * Issue prefetch i/os for the given blocks. If level is greater than 0, the
617 * indirect blocks prefeteched will be those that point to the blocks containing
618 * the data starting at offset, and continuing to offset + len.
620 * Note that if the indirect blocks above the blocks being prefetched are not in
621 * cache, they will be asychronously read in.
624 dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset,
625 uint64_t len, zio_priority_t pri)
631 if (len == 0) { /* they're interested in the bonus buffer */
632 dn = DMU_META_DNODE(os);
634 if (object == 0 || object >= DN_MAX_OBJECT)
637 rw_enter(&dn->dn_struct_rwlock, RW_READER);
638 blkid = dbuf_whichblock(dn, level,
639 object * sizeof (dnode_phys_t));
640 dbuf_prefetch(dn, level, blkid, pri, 0);
641 rw_exit(&dn->dn_struct_rwlock);
646 * XXX - Note, if the dnode for the requested object is not
647 * already cached, we will do a *synchronous* read in the
648 * dnode_hold() call. The same is true for any indirects.
650 err = dnode_hold(os, object, FTAG, &dn);
654 rw_enter(&dn->dn_struct_rwlock, RW_READER);
656 * offset + len - 1 is the last byte we want to prefetch for, and offset
657 * is the first. Then dbuf_whichblk(dn, level, off + len - 1) is the
658 * last block we want to prefetch, and dbuf_whichblock(dn, level,
659 * offset) is the first. Then the number we need to prefetch is the
662 if (level > 0 || dn->dn_datablkshift != 0) {
663 nblks = dbuf_whichblock(dn, level, offset + len - 1) -
664 dbuf_whichblock(dn, level, offset) + 1;
666 nblks = (offset < dn->dn_datablksz);
670 blkid = dbuf_whichblock(dn, level, offset);
671 for (int i = 0; i < nblks; i++)
672 dbuf_prefetch(dn, level, blkid + i, pri, 0);
675 rw_exit(&dn->dn_struct_rwlock);
677 dnode_rele(dn, FTAG);
681 * Get the next "chunk" of file data to free. We traverse the file from
682 * the end so that the file gets shorter over time (if we crashes in the
683 * middle, this will leave us in a better state). We find allocated file
684 * data by simply searching the allocated level 1 indirects.
686 * On input, *start should be the first offset that does not need to be
687 * freed (e.g. "offset + length"). On return, *start will be the first
688 * offset that should be freed and l1blks is set to the number of level 1
689 * indirect blocks found within the chunk.
692 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum, uint64_t *l1blks)
695 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
696 /* bytes of data covered by a level-1 indirect block */
698 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
700 ASSERT3U(minimum, <=, *start);
703 * Check if we can free the entire range assuming that all of the
704 * L1 blocks in this range have data. If we can, we use this
705 * worst case value as an estimate so we can avoid having to look
706 * at the object's actual data.
708 uint64_t total_l1blks =
709 (roundup(*start, iblkrange) - (minimum / iblkrange * iblkrange)) /
711 if (total_l1blks <= maxblks) {
712 *l1blks = total_l1blks;
716 ASSERT(ISP2(iblkrange));
718 for (blks = 0; *start > minimum && blks < maxblks; blks++) {
722 * dnode_next_offset(BACKWARDS) will find an allocated L1
723 * indirect block at or before the input offset. We must
724 * decrement *start so that it is at the end of the region
729 err = dnode_next_offset(dn,
730 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
732 /* if there are no indirect blocks before start, we are done */
736 } else if (err != 0) {
741 /* set start to the beginning of this L1 indirect */
742 *start = P2ALIGN(*start, iblkrange);
744 if (*start < minimum)
752 * If this objset is of type OST_ZFS return true if vfs's unmounted flag is set,
753 * otherwise return false.
754 * Used below in dmu_free_long_range_impl() to enable abort when unmounting
758 dmu_objset_zfs_unmounting(objset_t *os)
761 if (dmu_objset_type(os) == DMU_OST_ZFS)
762 return (zfs_get_vfs_flag_unmounted(os));
768 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
771 uint64_t object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
773 uint64_t dirty_frees_threshold;
774 dsl_pool_t *dp = dmu_objset_pool(os);
776 if (offset >= object_size)
779 if (zfs_per_txg_dirty_frees_percent <= 100)
780 dirty_frees_threshold =
781 zfs_per_txg_dirty_frees_percent * zfs_dirty_data_max / 100;
783 dirty_frees_threshold = zfs_dirty_data_max / 20;
785 if (length == DMU_OBJECT_END || offset + length > object_size)
786 length = object_size - offset;
788 while (length != 0) {
789 uint64_t chunk_end, chunk_begin, chunk_len;
793 if (dmu_objset_zfs_unmounting(dn->dn_objset))
794 return (SET_ERROR(EINTR));
796 chunk_end = chunk_begin = offset + length;
798 /* move chunk_begin backwards to the beginning of this chunk */
799 err = get_next_chunk(dn, &chunk_begin, offset, &l1blks);
802 ASSERT3U(chunk_begin, >=, offset);
803 ASSERT3U(chunk_begin, <=, chunk_end);
805 chunk_len = chunk_end - chunk_begin;
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 uint64_t txg = dmu_tx_get_txg(tx);
823 mutex_enter(&dp->dp_lock);
824 uint64_t long_free_dirty =
825 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK];
826 mutex_exit(&dp->dp_lock);
829 * To avoid filling up a TXG with just frees, wait for
830 * the next TXG to open before freeing more chunks if
831 * we have reached the threshold of frees.
833 if (dirty_frees_threshold != 0 &&
834 long_free_dirty >= dirty_frees_threshold) {
836 txg_wait_open(dp, 0);
841 * In order to prevent unnecessary write throttling, for each
842 * TXG, we track the cumulative size of L1 blocks being dirtied
843 * in dnode_free_range() below. We compare this number to a
844 * tunable threshold, past which we prevent new L1 dirty freeing
845 * blocks from being added into the open TXG. See
846 * dmu_free_long_range_impl() for details. The threshold
847 * prevents write throttle activation due to dirty freeing L1
848 * blocks taking up a large percentage of zfs_dirty_data_max.
850 mutex_enter(&dp->dp_lock);
851 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] +=
852 l1blks << dn->dn_indblkshift;
853 mutex_exit(&dp->dp_lock);
854 DTRACE_PROBE3(free__long__range,
855 uint64_t, long_free_dirty, uint64_t, chunk_len,
857 dnode_free_range(dn, chunk_begin, chunk_len, tx);
866 dmu_free_long_range(objset_t *os, uint64_t object,
867 uint64_t offset, uint64_t length)
872 err = dnode_hold(os, object, FTAG, &dn);
875 err = dmu_free_long_range_impl(os, dn, offset, length);
878 * It is important to zero out the maxblkid when freeing the entire
879 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
880 * will take the fast path, and (b) dnode_reallocate() can verify
881 * that the entire file has been freed.
883 if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
886 dnode_rele(dn, FTAG);
891 dmu_free_long_object(objset_t *os, uint64_t object)
896 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
900 tx = dmu_tx_create(os);
901 dmu_tx_hold_bonus(tx, object);
902 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
903 dmu_tx_mark_netfree(tx);
904 err = dmu_tx_assign(tx, TXG_WAIT);
906 err = dmu_object_free(os, object, tx);
916 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
917 uint64_t size, dmu_tx_t *tx)
920 int err = dnode_hold(os, object, FTAG, &dn);
923 ASSERT(offset < UINT64_MAX);
924 ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
925 dnode_free_range(dn, offset, size, tx);
926 dnode_rele(dn, FTAG);
931 dmu_read_impl(dnode_t *dn, uint64_t offset, uint64_t size,
932 void *buf, uint32_t flags)
935 int numbufs, err = 0;
938 * Deal with odd block sizes, where there can't be data past the first
939 * block. If we ever do the tail block optimization, we will need to
940 * handle that here as well.
942 if (dn->dn_maxblkid == 0) {
943 int newsz = offset > dn->dn_datablksz ? 0 :
944 MIN(size, dn->dn_datablksz - offset);
945 bzero((char *)buf + newsz, size - newsz);
950 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
954 * NB: we could do this block-at-a-time, but it's nice
955 * to be reading in parallel.
957 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
958 TRUE, FTAG, &numbufs, &dbp, flags);
962 for (i = 0; i < numbufs; i++) {
965 dmu_buf_t *db = dbp[i];
969 bufoff = offset - db->db_offset;
970 tocpy = (int)MIN(db->db_size - bufoff, size);
972 bcopy((char *)db->db_data + bufoff, buf, tocpy);
976 buf = (char *)buf + tocpy;
978 dmu_buf_rele_array(dbp, numbufs, FTAG);
984 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
985 void *buf, uint32_t flags)
990 err = dnode_hold(os, object, FTAG, &dn);
994 err = dmu_read_impl(dn, offset, size, buf, flags);
995 dnode_rele(dn, FTAG);
1000 dmu_read_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, void *buf,
1003 return (dmu_read_impl(dn, offset, size, buf, flags));
1007 dmu_write_impl(dmu_buf_t **dbp, int numbufs, uint64_t offset, uint64_t size,
1008 const void *buf, dmu_tx_t *tx)
1012 for (i = 0; i < numbufs; i++) {
1015 dmu_buf_t *db = dbp[i];
1019 bufoff = offset - db->db_offset;
1020 tocpy = (int)MIN(db->db_size - bufoff, size);
1022 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1024 if (tocpy == db->db_size)
1025 dmu_buf_will_fill(db, tx);
1027 dmu_buf_will_dirty(db, tx);
1029 bcopy(buf, (char *)db->db_data + bufoff, tocpy);
1031 if (tocpy == db->db_size)
1032 dmu_buf_fill_done(db, tx);
1036 buf = (char *)buf + tocpy;
1041 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1042 const void *buf, dmu_tx_t *tx)
1050 VERIFY0(dmu_buf_hold_array(os, object, offset, size,
1051 FALSE, FTAG, &numbufs, &dbp));
1052 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1053 dmu_buf_rele_array(dbp, numbufs, FTAG);
1057 dmu_write_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size,
1058 const void *buf, dmu_tx_t *tx)
1066 VERIFY0(dmu_buf_hold_array_by_dnode(dn, offset, size,
1067 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH));
1068 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1069 dmu_buf_rele_array(dbp, numbufs, FTAG);
1073 dmu_object_remap_one_indirect(objset_t *os, dnode_t *dn,
1074 uint64_t last_removal_txg, uint64_t offset)
1076 uint64_t l1blkid = dbuf_whichblock(dn, 1, offset);
1079 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1080 dmu_buf_impl_t *dbuf = dbuf_hold_level(dn, 1, l1blkid, FTAG);
1081 ASSERT3P(dbuf, !=, NULL);
1084 * If the block hasn't been written yet, this default will ensure
1085 * we don't try to remap it.
1087 uint64_t birth = UINT64_MAX;
1088 ASSERT3U(last_removal_txg, !=, UINT64_MAX);
1089 if (dbuf->db_blkptr != NULL)
1090 birth = dbuf->db_blkptr->blk_birth;
1091 rw_exit(&dn->dn_struct_rwlock);
1094 * If this L1 was already written after the last removal, then we've
1095 * already tried to remap it.
1097 if (birth <= last_removal_txg &&
1098 dbuf_read(dbuf, NULL, DB_RF_MUST_SUCCEED) == 0 &&
1099 dbuf_can_remap(dbuf)) {
1100 dmu_tx_t *tx = dmu_tx_create(os);
1101 dmu_tx_hold_remap_l1indirect(tx, dn->dn_object);
1102 err = dmu_tx_assign(tx, TXG_WAIT);
1104 (void) dbuf_dirty(dbuf, tx);
1111 dbuf_rele(dbuf, FTAG);
1113 delay(zfs_object_remap_one_indirect_delay_ticks);
1119 * Remap all blockpointers in the object, if possible, so that they reference
1120 * only concrete vdevs.
1122 * To do this, iterate over the L0 blockpointers and remap any that reference
1123 * an indirect vdev. Note that we only examine L0 blockpointers; since we
1124 * cannot guarantee that we can remap all blockpointer anyways (due to split
1125 * blocks), we do not want to make the code unnecessarily complicated to
1126 * catch the unlikely case that there is an L1 block on an indirect vdev that
1127 * contains no indirect blockpointers.
1130 dmu_object_remap_indirects(objset_t *os, uint64_t object,
1131 uint64_t last_removal_txg)
1133 uint64_t offset, l1span;
1137 err = dnode_hold(os, object, FTAG, &dn);
1142 if (dn->dn_nlevels <= 1) {
1143 if (issig(JUSTLOOKING) && issig(FORREAL)) {
1144 err = SET_ERROR(EINTR);
1148 * If the dnode has no indirect blocks, we cannot dirty them.
1149 * We still want to remap the blkptr(s) in the dnode if
1150 * appropriate, so mark it as dirty.
1152 if (err == 0 && dnode_needs_remap(dn)) {
1153 dmu_tx_t *tx = dmu_tx_create(os);
1154 dmu_tx_hold_bonus(tx, dn->dn_object);
1155 if ((err = dmu_tx_assign(tx, TXG_WAIT)) == 0) {
1156 dnode_setdirty(dn, tx);
1163 dnode_rele(dn, FTAG);
1168 l1span = 1ULL << (dn->dn_indblkshift - SPA_BLKPTRSHIFT +
1169 dn->dn_datablkshift);
1171 * Find the next L1 indirect that is not a hole.
1173 while (dnode_next_offset(dn, 0, &offset, 2, 1, 0) == 0) {
1174 if (issig(JUSTLOOKING) && issig(FORREAL)) {
1175 err = SET_ERROR(EINTR);
1178 if ((err = dmu_object_remap_one_indirect(os, dn,
1179 last_removal_txg, offset)) != 0) {
1185 dnode_rele(dn, FTAG);
1190 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1199 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
1200 FALSE, FTAG, &numbufs, &dbp));
1202 for (i = 0; i < numbufs; i++) {
1203 dmu_buf_t *db = dbp[i];
1205 dmu_buf_will_not_fill(db, tx);
1207 dmu_buf_rele_array(dbp, numbufs, FTAG);
1211 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset,
1212 void *data, uint8_t etype, uint8_t comp, int uncompressed_size,
1213 int compressed_size, int byteorder, dmu_tx_t *tx)
1217 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES);
1218 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS);
1219 VERIFY0(dmu_buf_hold_noread(os, object, offset,
1222 dmu_buf_write_embedded(db,
1223 data, (bp_embedded_type_t)etype, (enum zio_compress)comp,
1224 uncompressed_size, compressed_size, byteorder, tx);
1226 dmu_buf_rele(db, FTAG);
1230 * DMU support for xuio
1232 kstat_t *xuio_ksp = NULL;
1235 dmu_xuio_init(xuio_t *xuio, int nblk)
1238 uio_t *uio = &xuio->xu_uio;
1240 uio->uio_iovcnt = nblk;
1241 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
1243 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
1245 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
1246 priv->iovp = uio->uio_iov;
1247 XUIO_XUZC_PRIV(xuio) = priv;
1249 if (XUIO_XUZC_RW(xuio) == UIO_READ)
1250 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
1252 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
1258 dmu_xuio_fini(xuio_t *xuio)
1260 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1261 int nblk = priv->cnt;
1263 kmem_free(priv->iovp, nblk * sizeof (iovec_t));
1264 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
1265 kmem_free(priv, sizeof (dmu_xuio_t));
1267 if (XUIO_XUZC_RW(xuio) == UIO_READ)
1268 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
1270 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
1274 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
1275 * and increase priv->next by 1.
1278 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
1281 uio_t *uio = &xuio->xu_uio;
1282 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1283 int i = priv->next++;
1285 ASSERT(i < priv->cnt);
1286 ASSERT(off + n <= arc_buf_lsize(abuf));
1287 iov = uio->uio_iov + i;
1288 iov->iov_base = (char *)abuf->b_data + off;
1290 priv->bufs[i] = abuf;
1295 dmu_xuio_cnt(xuio_t *xuio)
1297 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1302 dmu_xuio_arcbuf(xuio_t *xuio, int i)
1304 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1306 ASSERT(i < priv->cnt);
1307 return (priv->bufs[i]);
1311 dmu_xuio_clear(xuio_t *xuio, int i)
1313 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1315 ASSERT(i < priv->cnt);
1316 priv->bufs[i] = NULL;
1320 xuio_stat_init(void)
1322 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
1323 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
1324 KSTAT_FLAG_VIRTUAL);
1325 if (xuio_ksp != NULL) {
1326 xuio_ksp->ks_data = &xuio_stats;
1327 kstat_install(xuio_ksp);
1332 xuio_stat_fini(void)
1334 if (xuio_ksp != NULL) {
1335 kstat_delete(xuio_ksp);
1341 xuio_stat_wbuf_copied(void)
1343 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1347 xuio_stat_wbuf_nocopy(void)
1349 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
1354 dmu_read_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size)
1357 int numbufs, i, err;
1358 xuio_t *xuio = NULL;
1361 * NB: we could do this block-at-a-time, but it's nice
1362 * to be reading in parallel.
1364 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1365 TRUE, FTAG, &numbufs, &dbp, 0);
1370 if (uio->uio_extflg == UIO_XUIO)
1371 xuio = (xuio_t *)uio;
1374 for (i = 0; i < numbufs; i++) {
1377 dmu_buf_t *db = dbp[i];
1381 bufoff = uio->uio_loffset - db->db_offset;
1382 tocpy = (int)MIN(db->db_size - bufoff, size);
1385 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
1386 arc_buf_t *dbuf_abuf = dbi->db_buf;
1387 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
1388 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
1390 uio->uio_resid -= tocpy;
1391 uio->uio_loffset += tocpy;
1394 if (abuf == dbuf_abuf)
1395 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
1397 XUIOSTAT_BUMP(xuiostat_rbuf_copied);
1400 err = uiomove((char *)db->db_data + bufoff, tocpy,
1403 err = vn_io_fault_uiomove((char *)db->db_data + bufoff,
1412 dmu_buf_rele_array(dbp, numbufs, FTAG);
1418 * Read 'size' bytes into the uio buffer.
1419 * From object zdb->db_object.
1420 * Starting at offset uio->uio_loffset.
1422 * If the caller already has a dbuf in the target object
1423 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1424 * because we don't have to find the dnode_t for the object.
1427 dmu_read_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size)
1429 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1438 err = dmu_read_uio_dnode(dn, uio, size);
1445 * Read 'size' bytes into the uio buffer.
1446 * From the specified object
1447 * Starting at offset uio->uio_loffset.
1450 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
1458 err = dnode_hold(os, object, FTAG, &dn);
1462 err = dmu_read_uio_dnode(dn, uio, size);
1464 dnode_rele(dn, FTAG);
1470 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
1477 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1478 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1482 for (i = 0; i < numbufs; i++) {
1485 dmu_buf_t *db = dbp[i];
1489 bufoff = uio->uio_loffset - db->db_offset;
1490 tocpy = (int)MIN(db->db_size - bufoff, size);
1492 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1494 if (tocpy == db->db_size)
1495 dmu_buf_will_fill(db, tx);
1497 dmu_buf_will_dirty(db, tx);
1501 * XXX uiomove could block forever (eg. nfs-backed
1502 * pages). There needs to be a uiolockdown() function
1503 * to lock the pages in memory, so that uiomove won't
1506 err = uiomove((char *)db->db_data + bufoff, tocpy,
1509 err = vn_io_fault_uiomove((char *)db->db_data + bufoff, tocpy,
1513 if (tocpy == db->db_size)
1514 dmu_buf_fill_done(db, tx);
1522 dmu_buf_rele_array(dbp, numbufs, FTAG);
1527 * Write 'size' bytes from the uio buffer.
1528 * To object zdb->db_object.
1529 * Starting at offset uio->uio_loffset.
1531 * If the caller already has a dbuf in the target object
1532 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1533 * because we don't have to find the dnode_t for the object.
1536 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
1539 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1548 err = dmu_write_uio_dnode(dn, uio, size, tx);
1555 * Write 'size' bytes from the uio buffer.
1556 * To the specified object.
1557 * Starting at offset uio->uio_loffset.
1560 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1569 err = dnode_hold(os, object, FTAG, &dn);
1573 err = dmu_write_uio_dnode(dn, uio, size, tx);
1575 dnode_rele(dn, FTAG);
1582 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1583 page_t *pp, dmu_tx_t *tx)
1592 err = dmu_buf_hold_array(os, object, offset, size,
1593 FALSE, FTAG, &numbufs, &dbp);
1597 for (i = 0; i < numbufs; i++) {
1598 int tocpy, copied, thiscpy;
1600 dmu_buf_t *db = dbp[i];
1604 ASSERT3U(db->db_size, >=, PAGESIZE);
1606 bufoff = offset - db->db_offset;
1607 tocpy = (int)MIN(db->db_size - bufoff, size);
1609 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1611 if (tocpy == db->db_size)
1612 dmu_buf_will_fill(db, tx);
1614 dmu_buf_will_dirty(db, tx);
1616 for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1617 ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff);
1618 thiscpy = MIN(PAGESIZE, tocpy - copied);
1619 va = zfs_map_page(pp, S_READ);
1620 bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1621 zfs_unmap_page(pp, va);
1626 if (tocpy == db->db_size)
1627 dmu_buf_fill_done(db, tx);
1632 dmu_buf_rele_array(dbp, numbufs, FTAG);
1636 #else /* !illumos */
1639 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1640 vm_page_t *ma, dmu_tx_t *tx)
1650 err = dmu_buf_hold_array(os, object, offset, size,
1651 FALSE, FTAG, &numbufs, &dbp);
1655 for (i = 0; i < numbufs; i++) {
1656 int tocpy, copied, thiscpy;
1658 dmu_buf_t *db = dbp[i];
1662 ASSERT3U(db->db_size, >=, PAGESIZE);
1664 bufoff = offset - db->db_offset;
1665 tocpy = (int)MIN(db->db_size - bufoff, size);
1667 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1669 if (tocpy == db->db_size)
1670 dmu_buf_will_fill(db, tx);
1672 dmu_buf_will_dirty(db, tx);
1674 for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1675 ASSERT3U(ptoa((*ma)->pindex), ==, db->db_offset + bufoff);
1676 thiscpy = MIN(PAGESIZE, tocpy - copied);
1677 va = zfs_map_page(*ma, &sf);
1678 bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1684 if (tocpy == db->db_size)
1685 dmu_buf_fill_done(db, tx);
1690 dmu_buf_rele_array(dbp, numbufs, FTAG);
1695 dmu_read_pages(objset_t *os, uint64_t object, vm_page_t *ma, int count,
1696 int *rbehind, int *rahead, int last_size)
1705 int bufoff, pgoff, tocpy;
1709 ASSERT3U(ma[0]->pindex + count - 1, ==, ma[count - 1]->pindex);
1710 ASSERT(last_size <= PAGE_SIZE);
1712 err = dmu_buf_hold_array(os, object, IDX_TO_OFF(ma[0]->pindex),
1713 IDX_TO_OFF(count - 1) + last_size, TRUE, FTAG, &numbufs, &dbp);
1718 IMPLY(last_size < PAGE_SIZE, *rahead == 0);
1719 if (dbp[0]->db_offset != 0 || numbufs > 1) {
1720 for (i = 0; i < numbufs; i++) {
1721 ASSERT(ISP2(dbp[i]->db_size));
1722 ASSERT((dbp[i]->db_offset % dbp[i]->db_size) == 0);
1723 ASSERT3U(dbp[i]->db_size, ==, dbp[0]->db_size);
1728 vmobj = ma[0]->object;
1729 zfs_vmobject_wlock(vmobj);
1732 for (i = 0; i < *rbehind; i++) {
1733 m = vm_page_grab(vmobj, ma[0]->pindex - 1 - i,
1734 VM_ALLOC_NORMAL | VM_ALLOC_NOWAIT |
1735 VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY);
1738 if (!vm_page_none_valid(m)) {
1739 ASSERT3U(m->valid, ==, VM_PAGE_BITS_ALL);
1743 ASSERT(m->dirty == 0);
1744 ASSERT(!pmap_page_is_mapped(m));
1746 ASSERT(db->db_size > PAGE_SIZE);
1747 bufoff = IDX_TO_OFF(m->pindex) % db->db_size;
1748 va = zfs_map_page(m, &sf);
1749 bcopy((char *)db->db_data + bufoff, va, PAGESIZE);
1753 if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1754 vm_page_activate(m);
1756 vm_page_deactivate(m);
1762 bufoff = IDX_TO_OFF(ma[0]->pindex) % db->db_size;
1764 for (mi = 0, di = 0; mi < count && di < numbufs; ) {
1767 if (m != bogus_page) {
1768 vm_page_assert_xbusied(m);
1769 ASSERT(vm_page_none_valid(m));
1770 ASSERT(m->dirty == 0);
1771 ASSERT(!pmap_page_is_mapped(m));
1772 va = zfs_map_page(m, &sf);
1778 if (m != bogus_page) {
1779 ASSERT3U(IDX_TO_OFF(m->pindex) + pgoff, ==,
1780 db->db_offset + bufoff);
1784 * We do not need to clamp the copy size by the file
1785 * size as the last block is zero-filled beyond the
1786 * end of file anyway.
1788 tocpy = MIN(db->db_size - bufoff, PAGESIZE - pgoff);
1789 if (m != bogus_page)
1790 bcopy((char *)db->db_data + bufoff, va + pgoff, tocpy);
1793 ASSERT(pgoff <= PAGESIZE);
1794 if (pgoff == PAGESIZE) {
1795 if (m != bogus_page) {
1805 ASSERT(bufoff <= db->db_size);
1806 if (bufoff == db->db_size) {
1807 ASSERT(di < numbufs);
1815 * Three possibilities:
1816 * - last requested page ends at a buffer boundary and , thus,
1817 * all pages and buffers have been iterated;
1818 * - all requested pages are filled, but the last buffer
1819 * has not been exhausted;
1820 * the read-ahead is possible only in this case;
1821 * - all buffers have been read, but the last page has not been
1823 * this is only possible if the file has only a single buffer
1824 * with a size that is not a multiple of the page size.
1827 ASSERT(di >= numbufs - 1);
1828 IMPLY(*rahead != 0, di == numbufs - 1);
1829 IMPLY(*rahead != 0, bufoff != 0);
1832 if (di == numbufs) {
1833 ASSERT(mi >= count - 1);
1834 ASSERT(*rahead == 0);
1835 IMPLY(pgoff == 0, mi == count);
1837 ASSERT(mi == count - 1);
1838 ASSERT((dbp[0]->db_size & PAGE_MASK) != 0);
1843 ASSERT(m != bogus_page);
1844 bzero(va + pgoff, PAGESIZE - pgoff);
1849 for (i = 0; i < *rahead; i++) {
1850 m = vm_page_grab(vmobj, ma[count - 1]->pindex + 1 + i,
1851 VM_ALLOC_NORMAL | VM_ALLOC_NOWAIT |
1852 VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY);
1855 if (!vm_page_none_valid(m)) {
1856 ASSERT3U(m->valid, ==, VM_PAGE_BITS_ALL);
1860 ASSERT(m->dirty == 0);
1861 ASSERT(!pmap_page_is_mapped(m));
1863 ASSERT(db->db_size > PAGE_SIZE);
1864 bufoff = IDX_TO_OFF(m->pindex) % db->db_size;
1865 tocpy = MIN(db->db_size - bufoff, PAGESIZE);
1866 va = zfs_map_page(m, &sf);
1867 bcopy((char *)db->db_data + bufoff, va, tocpy);
1868 if (tocpy < PAGESIZE) {
1869 ASSERT(i == *rahead - 1);
1870 ASSERT((db->db_size & PAGE_MASK) != 0);
1871 bzero(va + tocpy, PAGESIZE - tocpy);
1876 if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1877 vm_page_activate(m);
1879 vm_page_deactivate(m);
1884 zfs_vmobject_wunlock(vmobj);
1886 dmu_buf_rele_array(dbp, numbufs, FTAG);
1889 #endif /* illumos */
1890 #endif /* _KERNEL */
1893 * Allocate a loaned anonymous arc buffer.
1896 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1898 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1900 return (arc_loan_buf(db->db_objset->os_spa, B_FALSE, size));
1904 * Free a loaned arc buffer.
1907 dmu_return_arcbuf(arc_buf_t *buf)
1909 arc_return_buf(buf, FTAG);
1910 arc_buf_destroy(buf, FTAG);
1914 * When possible directly assign passed loaned arc buffer to a dbuf.
1915 * If this is not possible copy the contents of passed arc buf via
1919 dmu_assign_arcbuf_dnode(dnode_t *dn, uint64_t offset, arc_buf_t *buf,
1923 uint32_t blksz = (uint32_t)arc_buf_lsize(buf);
1926 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1927 blkid = dbuf_whichblock(dn, 0, offset);
1928 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1929 rw_exit(&dn->dn_struct_rwlock);
1932 * We can only assign if the offset is aligned, the arc buf is the
1933 * same size as the dbuf, and the dbuf is not metadata.
1935 if (offset == db->db.db_offset && blksz == db->db.db_size) {
1937 curthread->td_ru.ru_oublock++;
1941 racct_add_force(curproc, RACCT_WRITEBPS, blksz);
1942 racct_add_force(curproc, RACCT_WRITEIOPS, 1);
1943 PROC_UNLOCK(curproc);
1946 #endif /* _KERNEL */
1947 dbuf_assign_arcbuf(db, buf, tx);
1948 dbuf_rele(db, FTAG);
1953 /* compressed bufs must always be assignable to their dbuf */
1954 ASSERT3U(arc_get_compression(buf), ==, ZIO_COMPRESS_OFF);
1955 ASSERT(!(buf->b_flags & ARC_BUF_FLAG_COMPRESSED));
1958 object = dn->dn_object;
1960 dbuf_rele(db, FTAG);
1961 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1962 dmu_return_arcbuf(buf);
1963 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1968 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1971 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1973 DB_DNODE_ENTER(dbuf);
1974 dmu_assign_arcbuf_dnode(DB_DNODE(dbuf), offset, buf, tx);
1975 DB_DNODE_EXIT(dbuf);
1979 dbuf_dirty_record_t *dsa_dr;
1980 dmu_sync_cb_t *dsa_done;
1987 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1989 dmu_sync_arg_t *dsa = varg;
1990 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1991 blkptr_t *bp = zio->io_bp;
1993 if (zio->io_error == 0) {
1994 if (BP_IS_HOLE(bp)) {
1996 * A block of zeros may compress to a hole, but the
1997 * block size still needs to be known for replay.
1999 BP_SET_LSIZE(bp, db->db_size);
2000 } else if (!BP_IS_EMBEDDED(bp)) {
2001 ASSERT(BP_GET_LEVEL(bp) == 0);
2008 dmu_sync_late_arrival_ready(zio_t *zio)
2010 dmu_sync_ready(zio, NULL, zio->io_private);
2015 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
2017 dmu_sync_arg_t *dsa = varg;
2018 dbuf_dirty_record_t *dr = dsa->dsa_dr;
2019 dmu_buf_impl_t *db = dr->dr_dbuf;
2020 zgd_t *zgd = dsa->dsa_zgd;
2023 * Record the vdev(s) backing this blkptr so they can be flushed after
2024 * the writes for the lwb have completed.
2026 if (zio->io_error == 0) {
2027 zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp);
2030 mutex_enter(&db->db_mtx);
2031 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
2032 if (zio->io_error == 0) {
2033 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
2034 if (dr->dt.dl.dr_nopwrite) {
2035 blkptr_t *bp = zio->io_bp;
2036 blkptr_t *bp_orig = &zio->io_bp_orig;
2037 uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
2039 ASSERT(BP_EQUAL(bp, bp_orig));
2040 VERIFY(BP_EQUAL(bp, db->db_blkptr));
2041 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
2042 ASSERT(zio_checksum_table[chksum].ci_flags &
2043 ZCHECKSUM_FLAG_NOPWRITE);
2045 dr->dt.dl.dr_overridden_by = *zio->io_bp;
2046 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
2047 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
2050 * Old style holes are filled with all zeros, whereas
2051 * new-style holes maintain their lsize, type, level,
2052 * and birth time (see zio_write_compress). While we
2053 * need to reset the BP_SET_LSIZE() call that happened
2054 * in dmu_sync_ready for old style holes, we do *not*
2055 * want to wipe out the information contained in new
2056 * style holes. Thus, only zero out the block pointer if
2057 * it's an old style hole.
2059 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) &&
2060 dr->dt.dl.dr_overridden_by.blk_birth == 0)
2061 BP_ZERO(&dr->dt.dl.dr_overridden_by);
2063 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
2065 cv_broadcast(&db->db_changed);
2066 mutex_exit(&db->db_mtx);
2068 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
2070 kmem_free(dsa, sizeof (*dsa));
2074 dmu_sync_late_arrival_done(zio_t *zio)
2076 blkptr_t *bp = zio->io_bp;
2077 dmu_sync_arg_t *dsa = zio->io_private;
2078 blkptr_t *bp_orig = &zio->io_bp_orig;
2079 zgd_t *zgd = dsa->dsa_zgd;
2081 if (zio->io_error == 0) {
2083 * Record the vdev(s) backing this blkptr so they can be
2084 * flushed after the writes for the lwb have completed.
2086 zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp);
2088 if (!BP_IS_HOLE(bp)) {
2089 ASSERT(!(zio->io_flags & ZIO_FLAG_NOPWRITE));
2090 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
2091 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
2092 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
2093 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
2097 dmu_tx_commit(dsa->dsa_tx);
2099 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
2101 abd_put(zio->io_abd);
2102 kmem_free(dsa, sizeof (*dsa));
2106 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
2107 zio_prop_t *zp, zbookmark_phys_t *zb)
2109 dmu_sync_arg_t *dsa;
2112 tx = dmu_tx_create(os);
2113 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
2114 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
2116 /* Make zl_get_data do txg_waited_synced() */
2117 return (SET_ERROR(EIO));
2121 * In order to prevent the zgd's lwb from being free'd prior to
2122 * dmu_sync_late_arrival_done() being called, we have to ensure
2123 * the lwb's "max txg" takes this tx's txg into account.
2125 zil_lwb_add_txg(zgd->zgd_lwb, dmu_tx_get_txg(tx));
2127 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
2129 dsa->dsa_done = done;
2134 * Since we are currently syncing this txg, it's nontrivial to
2135 * determine what BP to nopwrite against, so we disable nopwrite.
2137 * When syncing, the db_blkptr is initially the BP of the previous
2138 * txg. We can not nopwrite against it because it will be changed
2139 * (this is similar to the non-late-arrival case where the dbuf is
2140 * dirty in a future txg).
2142 * Then dbuf_write_ready() sets bp_blkptr to the location we will write.
2143 * We can not nopwrite against it because although the BP will not
2144 * (typically) be changed, the data has not yet been persisted to this
2147 * Finally, when dbuf_write_done() is called, it is theoretically
2148 * possible to always nopwrite, because the data that was written in
2149 * this txg is the same data that we are trying to write. However we
2150 * would need to check that this dbuf is not dirty in any future
2151 * txg's (as we do in the normal dmu_sync() path). For simplicity, we
2152 * don't nopwrite in this case.
2154 zp->zp_nopwrite = B_FALSE;
2156 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
2157 abd_get_from_buf(zgd->zgd_db->db_data, zgd->zgd_db->db_size),
2158 zgd->zgd_db->db_size, zgd->zgd_db->db_size, zp,
2159 dmu_sync_late_arrival_ready, NULL, NULL, dmu_sync_late_arrival_done,
2160 dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
2166 * Intent log support: sync the block associated with db to disk.
2167 * N.B. and XXX: the caller is responsible for making sure that the
2168 * data isn't changing while dmu_sync() is writing it.
2172 * EEXIST: this txg has already been synced, so there's nothing to do.
2173 * The caller should not log the write.
2175 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
2176 * The caller should not log the write.
2178 * EALREADY: this block is already in the process of being synced.
2179 * The caller should track its progress (somehow).
2181 * EIO: could not do the I/O.
2182 * The caller should do a txg_wait_synced().
2184 * 0: the I/O has been initiated.
2185 * The caller should log this blkptr in the done callback.
2186 * It is possible that the I/O will fail, in which case
2187 * the error will be reported to the done callback and
2188 * propagated to pio from zio_done().
2191 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
2193 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
2194 objset_t *os = db->db_objset;
2195 dsl_dataset_t *ds = os->os_dsl_dataset;
2196 dbuf_dirty_record_t *dr;
2197 dmu_sync_arg_t *dsa;
2198 zbookmark_phys_t zb;
2202 ASSERT(pio != NULL);
2205 SET_BOOKMARK(&zb, ds->ds_object,
2206 db->db.db_object, db->db_level, db->db_blkid);
2210 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
2214 * If we're frozen (running ziltest), we always need to generate a bp.
2216 if (txg > spa_freeze_txg(os->os_spa))
2217 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
2220 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
2221 * and us. If we determine that this txg is not yet syncing,
2222 * but it begins to sync a moment later, that's OK because the
2223 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
2225 mutex_enter(&db->db_mtx);
2227 if (txg <= spa_last_synced_txg(os->os_spa)) {
2229 * This txg has already synced. There's nothing to do.
2231 mutex_exit(&db->db_mtx);
2232 return (SET_ERROR(EEXIST));
2235 if (txg <= spa_syncing_txg(os->os_spa)) {
2237 * This txg is currently syncing, so we can't mess with
2238 * the dirty record anymore; just write a new log block.
2240 mutex_exit(&db->db_mtx);
2241 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
2244 dr = db->db_last_dirty;
2245 while (dr && dr->dr_txg != txg)
2250 * There's no dr for this dbuf, so it must have been freed.
2251 * There's no need to log writes to freed blocks, so we're done.
2253 mutex_exit(&db->db_mtx);
2254 return (SET_ERROR(ENOENT));
2257 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg);
2259 if (db->db_blkptr != NULL) {
2261 * We need to fill in zgd_bp with the current blkptr so that
2262 * the nopwrite code can check if we're writing the same
2263 * data that's already on disk. We can only nopwrite if we
2264 * are sure that after making the copy, db_blkptr will not
2265 * change until our i/o completes. We ensure this by
2266 * holding the db_mtx, and only allowing nopwrite if the
2267 * block is not already dirty (see below). This is verified
2268 * by dmu_sync_done(), which VERIFYs that the db_blkptr has
2271 *zgd->zgd_bp = *db->db_blkptr;
2275 * Assume the on-disk data is X, the current syncing data (in
2276 * txg - 1) is Y, and the current in-memory data is Z (currently
2279 * We usually want to perform a nopwrite if X and Z are the
2280 * same. However, if Y is different (i.e. the BP is going to
2281 * change before this write takes effect), then a nopwrite will
2282 * be incorrect - we would override with X, which could have
2283 * been freed when Y was written.
2285 * (Note that this is not a concern when we are nop-writing from
2286 * syncing context, because X and Y must be identical, because
2287 * all previous txgs have been synced.)
2289 * Therefore, we disable nopwrite if the current BP could change
2290 * before this TXG. There are two ways it could change: by
2291 * being dirty (dr_next is non-NULL), or by being freed
2292 * (dnode_block_freed()). This behavior is verified by
2293 * zio_done(), which VERIFYs that the override BP is identical
2294 * to the on-disk BP.
2298 if (dr->dr_next != NULL || dnode_block_freed(dn, db->db_blkid))
2299 zp.zp_nopwrite = B_FALSE;
2302 ASSERT(dr->dr_txg == txg);
2303 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
2304 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
2306 * We have already issued a sync write for this buffer,
2307 * or this buffer has already been synced. It could not
2308 * have been dirtied since, or we would have cleared the state.
2310 mutex_exit(&db->db_mtx);
2311 return (SET_ERROR(EALREADY));
2314 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
2315 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
2316 mutex_exit(&db->db_mtx);
2318 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
2320 dsa->dsa_done = done;
2324 zio_nowait(arc_write(pio, os->os_spa, txg,
2325 zgd->zgd_bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db),
2326 &zp, dmu_sync_ready, NULL, NULL, dmu_sync_done, dsa,
2327 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb));
2333 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
2339 err = dnode_hold(os, object, FTAG, &dn);
2342 err = dnode_set_blksz(dn, size, ibs, tx);
2343 dnode_rele(dn, FTAG);
2348 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
2354 * Send streams include each object's checksum function. This
2355 * check ensures that the receiving system can understand the
2356 * checksum function transmitted.
2358 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS);
2360 VERIFY0(dnode_hold(os, object, FTAG, &dn));
2361 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS);
2362 dn->dn_checksum = checksum;
2363 dnode_setdirty(dn, tx);
2364 dnode_rele(dn, FTAG);
2368 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
2374 * Send streams include each object's compression function. This
2375 * check ensures that the receiving system can understand the
2376 * compression function transmitted.
2378 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS);
2380 VERIFY0(dnode_hold(os, object, FTAG, &dn));
2381 dn->dn_compress = compress;
2382 dnode_setdirty(dn, tx);
2383 dnode_rele(dn, FTAG);
2386 int zfs_mdcomp_disable = 0;
2387 SYSCTL_INT(_vfs_zfs, OID_AUTO, mdcomp_disable, CTLFLAG_RWTUN,
2388 &zfs_mdcomp_disable, 0, "Disable metadata compression");
2391 * When the "redundant_metadata" property is set to "most", only indirect
2392 * blocks of this level and higher will have an additional ditto block.
2394 int zfs_redundant_metadata_most_ditto_level = 2;
2397 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
2399 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
2400 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
2402 enum zio_checksum checksum = os->os_checksum;
2403 enum zio_compress compress = os->os_compress;
2404 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
2405 boolean_t dedup = B_FALSE;
2406 boolean_t nopwrite = B_FALSE;
2407 boolean_t dedup_verify = os->os_dedup_verify;
2408 int copies = os->os_copies;
2411 * We maintain different write policies for each of the following
2414 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
2415 * 3. all other level 0 blocks
2418 if (zfs_mdcomp_disable) {
2419 compress = ZIO_COMPRESS_EMPTY;
2422 * XXX -- we should design a compression algorithm
2423 * that specializes in arrays of bps.
2425 compress = zio_compress_select(os->os_spa,
2426 ZIO_COMPRESS_ON, ZIO_COMPRESS_ON);
2430 * Metadata always gets checksummed. If the data
2431 * checksum is multi-bit correctable, and it's not a
2432 * ZBT-style checksum, then it's suitable for metadata
2433 * as well. Otherwise, the metadata checksum defaults
2436 if (!(zio_checksum_table[checksum].ci_flags &
2437 ZCHECKSUM_FLAG_METADATA) ||
2438 (zio_checksum_table[checksum].ci_flags &
2439 ZCHECKSUM_FLAG_EMBEDDED))
2440 checksum = ZIO_CHECKSUM_FLETCHER_4;
2442 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL ||
2443 (os->os_redundant_metadata ==
2444 ZFS_REDUNDANT_METADATA_MOST &&
2445 (level >= zfs_redundant_metadata_most_ditto_level ||
2446 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))))
2448 } else if (wp & WP_NOFILL) {
2452 * If we're writing preallocated blocks, we aren't actually
2453 * writing them so don't set any policy properties. These
2454 * blocks are currently only used by an external subsystem
2455 * outside of zfs (i.e. dump) and not written by the zio
2458 compress = ZIO_COMPRESS_OFF;
2459 checksum = ZIO_CHECKSUM_NOPARITY;
2461 compress = zio_compress_select(os->os_spa, dn->dn_compress,
2464 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
2465 zio_checksum_select(dn->dn_checksum, checksum) :
2469 * Determine dedup setting. If we are in dmu_sync(),
2470 * we won't actually dedup now because that's all
2471 * done in syncing context; but we do want to use the
2472 * dedup checkum. If the checksum is not strong
2473 * enough to ensure unique signatures, force
2476 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
2477 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
2478 if (!(zio_checksum_table[checksum].ci_flags &
2479 ZCHECKSUM_FLAG_DEDUP))
2480 dedup_verify = B_TRUE;
2484 * Enable nopwrite if we have secure enough checksum
2485 * algorithm (see comment in zio_nop_write) and
2486 * compression is enabled. We don't enable nopwrite if
2487 * dedup is enabled as the two features are mutually
2490 nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags &
2491 ZCHECKSUM_FLAG_NOPWRITE) &&
2492 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
2495 zp->zp_checksum = checksum;
2496 zp->zp_compress = compress;
2497 ASSERT3U(zp->zp_compress, !=, ZIO_COMPRESS_INHERIT);
2499 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
2500 zp->zp_level = level;
2501 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
2502 zp->zp_dedup = dedup;
2503 zp->zp_dedup_verify = dedup && dedup_verify;
2504 zp->zp_nopwrite = nopwrite;
2508 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
2514 * Sync any current changes before
2515 * we go trundling through the block pointers.
2517 err = dmu_object_wait_synced(os, object);
2522 err = dnode_hold(os, object, FTAG, &dn);
2527 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
2528 dnode_rele(dn, FTAG);
2534 * Given the ZFS object, if it contains any dirty nodes
2535 * this function flushes all dirty blocks to disk. This
2536 * ensures the DMU object info is updated. A more efficient
2537 * future version might just find the TXG with the maximum
2538 * ID and wait for that to be synced.
2541 dmu_object_wait_synced(objset_t *os, uint64_t object)
2546 error = dnode_hold(os, object, FTAG, &dn);
2551 for (i = 0; i < TXG_SIZE; i++) {
2552 if (list_link_active(&dn->dn_dirty_link[i])) {
2556 dnode_rele(dn, FTAG);
2557 if (i != TXG_SIZE) {
2558 txg_wait_synced(dmu_objset_pool(os), 0);
2565 __dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2567 dnode_phys_t *dnp = dn->dn_phys;
2569 doi->doi_data_block_size = dn->dn_datablksz;
2570 doi->doi_metadata_block_size = dn->dn_indblkshift ?
2571 1ULL << dn->dn_indblkshift : 0;
2572 doi->doi_type = dn->dn_type;
2573 doi->doi_bonus_type = dn->dn_bonustype;
2574 doi->doi_bonus_size = dn->dn_bonuslen;
2575 doi->doi_dnodesize = dn->dn_num_slots << DNODE_SHIFT;
2576 doi->doi_indirection = dn->dn_nlevels;
2577 doi->doi_checksum = dn->dn_checksum;
2578 doi->doi_compress = dn->dn_compress;
2579 doi->doi_nblkptr = dn->dn_nblkptr;
2580 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
2581 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
2582 doi->doi_fill_count = 0;
2583 for (int i = 0; i < dnp->dn_nblkptr; i++)
2584 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]);
2588 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2590 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2591 mutex_enter(&dn->dn_mtx);
2593 __dmu_object_info_from_dnode(dn, doi);
2595 mutex_exit(&dn->dn_mtx);
2596 rw_exit(&dn->dn_struct_rwlock);
2600 * Get information on a DMU object.
2601 * If doi is NULL, just indicates whether the object exists.
2604 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
2607 int err = dnode_hold(os, object, FTAG, &dn);
2613 dmu_object_info_from_dnode(dn, doi);
2615 dnode_rele(dn, FTAG);
2620 * As above, but faster; can be used when you have a held dbuf in hand.
2623 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
2625 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2628 dmu_object_info_from_dnode(DB_DNODE(db), doi);
2633 * Faster still when you only care about the size.
2634 * This is specifically optimized for zfs_getattr().
2637 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
2638 u_longlong_t *nblk512)
2640 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2646 *blksize = dn->dn_datablksz;
2647 /* add in number of slots used for the dnode itself */
2648 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
2649 SPA_MINBLOCKSHIFT) + dn->dn_num_slots;
2654 dmu_object_dnsize_from_db(dmu_buf_t *db_fake, int *dnsize)
2656 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2661 *dnsize = dn->dn_num_slots << DNODE_SHIFT;
2666 byteswap_uint64_array(void *vbuf, size_t size)
2668 uint64_t *buf = vbuf;
2669 size_t count = size >> 3;
2672 ASSERT((size & 7) == 0);
2674 for (i = 0; i < count; i++)
2675 buf[i] = BSWAP_64(buf[i]);
2679 byteswap_uint32_array(void *vbuf, size_t size)
2681 uint32_t *buf = vbuf;
2682 size_t count = size >> 2;
2685 ASSERT((size & 3) == 0);
2687 for (i = 0; i < count; i++)
2688 buf[i] = BSWAP_32(buf[i]);
2692 byteswap_uint16_array(void *vbuf, size_t size)
2694 uint16_t *buf = vbuf;
2695 size_t count = size >> 1;
2698 ASSERT((size & 1) == 0);
2700 for (i = 0; i < count; i++)
2701 buf[i] = BSWAP_16(buf[i]);
2706 byteswap_uint8_array(void *vbuf, size_t size)
2720 zio_compress_init();
2729 arc_fini(); /* arc depends on l2arc, so arc must go first */
2732 zio_compress_fini();